2 * Kernel Probes (KProbes)
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 * Copyright (C) IBM Corporation, 2002, 2004
20 * 2002-Oct Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
21 * Probes initial implementation ( includes contributions from
23 * 2004-July Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
24 * interface to access function arguments.
25 * 2004-Oct Jim Keniston <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
26 * <prasanna@in.ibm.com> adapted for x86_64 from i386.
27 * 2005-Mar Roland McGrath <roland@redhat.com>
28 * Fixed to handle %rip-relative addressing mode correctly.
29 * 2005-May Hien Nguyen <hien@us.ibm.com>, Jim Keniston
30 * <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
31 * <prasanna@in.ibm.com> added function-return probes.
32 * 2005-May Rusty Lynch <rusty.lynch@intel.com>
33 * Added function return probes functionality
34 * 2006-Feb Masami Hiramatsu <hiramatu@sdl.hitachi.co.jp> added
35 * kprobe-booster and kretprobe-booster for i386.
36 * 2007-Dec Masami Hiramatsu <mhiramat@redhat.com> added kprobe-booster
37 * and kretprobe-booster for x86-64
38 * 2007-Dec Masami Hiramatsu <mhiramat@redhat.com>, Arjan van de Ven
39 * <arjan@infradead.org> and Jim Keniston <jkenisto@us.ibm.com>
40 * unified x86 kprobes code.
42 #include <linux/kprobes.h>
43 #include <linux/ptrace.h>
44 #include <linux/string.h>
45 #include <linux/slab.h>
46 #include <linux/hardirq.h>
47 #include <linux/preempt.h>
48 #include <linux/sched/debug.h>
49 #include <linux/extable.h>
50 #include <linux/kdebug.h>
51 #include <linux/kallsyms.h>
52 #include <linux/ftrace.h>
53 #include <linux/frame.h>
54 #include <linux/kasan.h>
55 #include <linux/moduleloader.h>
57 #include <asm/text-patching.h>
58 #include <asm/cacheflush.h>
60 #include <asm/pgtable.h>
61 #include <linux/uaccess.h>
62 #include <asm/alternative.h>
64 #include <asm/debugreg.h>
65 #include <asm/set_memory.h>
69 void jprobe_return_end(void);
71 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
72 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
74 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
76 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
77 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
78 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
79 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
80 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
83 * Undefined/reserved opcodes, conditional jump, Opcode Extension
84 * Groups, and some special opcodes can not boost.
85 * This is non-const and volatile to keep gcc from statically
86 * optimizing it out, as variable_test_bit makes gcc think only
87 * *(unsigned long*) is used.
89 static volatile u32 twobyte_is_boostable[256 / 32] = {
90 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
91 /* ---------------------------------------------- */
92 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
93 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
94 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
95 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
96 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
97 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
98 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
99 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
100 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
101 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
102 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
103 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
104 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
105 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
106 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
107 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
108 /* ----------------------------------------------- */
109 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
113 struct kretprobe_blackpoint kretprobe_blacklist[] = {
114 {"__switch_to", }, /* This function switches only current task, but
115 doesn't switch kernel stack.*/
116 {NULL, NULL} /* Terminator */
119 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
121 static nokprobe_inline void
122 __synthesize_relative_insn(void *from, void *to, u8 op)
124 struct __arch_relative_insn {
129 insn = (struct __arch_relative_insn *)from;
130 insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
134 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
135 void synthesize_reljump(void *from, void *to)
137 __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
139 NOKPROBE_SYMBOL(synthesize_reljump);
141 /* Insert a call instruction at address 'from', which calls address 'to'.*/
142 void synthesize_relcall(void *from, void *to)
144 __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
146 NOKPROBE_SYMBOL(synthesize_relcall);
149 * Skip the prefixes of the instruction.
151 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
155 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
156 while (inat_is_legacy_prefix(attr)) {
158 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
161 if (inat_is_rex_prefix(attr))
166 NOKPROBE_SYMBOL(skip_prefixes);
169 * Returns non-zero if INSN is boostable.
170 * RIP relative instructions are adjusted at copying time in 64 bits mode
172 int can_boost(struct insn *insn, void *addr)
174 kprobe_opcode_t opcode;
176 if (search_exception_tables((unsigned long)addr))
177 return 0; /* Page fault may occur on this address. */
179 /* 2nd-byte opcode */
180 if (insn->opcode.nbytes == 2)
181 return test_bit(insn->opcode.bytes[1],
182 (unsigned long *)twobyte_is_boostable);
184 if (insn->opcode.nbytes != 1)
187 /* Can't boost Address-size override prefix */
188 if (unlikely(inat_is_address_size_prefix(insn->attr)))
191 opcode = insn->opcode.bytes[0];
193 switch (opcode & 0xf0) {
195 /* can't boost "bound" */
196 return (opcode != 0x62);
198 return 0; /* can't boost conditional jump */
200 return opcode != 0x9a; /* can't boost call far */
202 /* can't boost software-interruptions */
203 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
205 /* can boost AA* and XLAT */
206 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
208 /* can boost in/out and absolute jmps */
209 return ((opcode & 0x04) || opcode == 0xea);
211 /* clear and set flags are boostable */
212 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
214 /* CS override prefix and call are not boostable */
215 return (opcode != 0x2e && opcode != 0x9a);
220 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
225 kp = get_kprobe((void *)addr);
226 faddr = ftrace_location(addr);
228 * Addresses inside the ftrace location are refused by
229 * arch_check_ftrace_location(). Something went terribly wrong
230 * if such an address is checked here.
232 if (WARN_ON(faddr && faddr != addr))
235 * Use the current code if it is not modified by Kprobe
236 * and it cannot be modified by ftrace.
242 * Basically, kp->ainsn.insn has an original instruction.
243 * However, RIP-relative instruction can not do single-stepping
244 * at different place, __copy_instruction() tweaks the displacement of
245 * that instruction. In that case, we can't recover the instruction
246 * from the kp->ainsn.insn.
248 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
249 * of the first byte of the probed instruction, which is overwritten
250 * by int3. And the instruction at kp->addr is not modified by kprobes
251 * except for the first byte, we can recover the original instruction
252 * from it and kp->opcode.
254 * In case of Kprobes using ftrace, we do not have a copy of
255 * the original instruction. In fact, the ftrace location might
256 * be modified at anytime and even could be in an inconsistent state.
257 * Fortunately, we know that the original code is the ideal 5-byte
260 if (probe_kernel_read(buf, (void *)addr,
261 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
265 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
268 return (unsigned long)buf;
272 * Recover the probed instruction at addr for further analysis.
273 * Caller must lock kprobes by kprobe_mutex, or disable preemption
274 * for preventing to release referencing kprobes.
275 * Returns zero if the instruction can not get recovered (or access failed).
277 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
279 unsigned long __addr;
281 __addr = __recover_optprobed_insn(buf, addr);
285 return __recover_probed_insn(buf, addr);
288 /* Check if paddr is at an instruction boundary */
289 static int can_probe(unsigned long paddr)
291 unsigned long addr, __addr, offset = 0;
293 kprobe_opcode_t buf[MAX_INSN_SIZE];
295 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
298 /* Decode instructions */
299 addr = paddr - offset;
300 while (addr < paddr) {
302 * Check if the instruction has been modified by another
303 * kprobe, in which case we replace the breakpoint by the
304 * original instruction in our buffer.
305 * Also, jump optimization will change the breakpoint to
306 * relative-jump. Since the relative-jump itself is
307 * normally used, we just go through if there is no kprobe.
309 __addr = recover_probed_instruction(buf, addr);
312 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
313 insn_get_length(&insn);
316 * Another debugging subsystem might insert this breakpoint.
317 * In that case, we can't recover it.
319 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
324 return (addr == paddr);
328 * Returns non-zero if opcode modifies the interrupt flag.
330 static int is_IF_modifier(kprobe_opcode_t *insn)
333 insn = skip_prefixes(insn);
338 case 0xcf: /* iret/iretd */
339 case 0x9d: /* popf/popfd */
347 * Copy an instruction with recovering modified instruction by kprobes
348 * and adjust the displacement if the instruction uses the %rip-relative
350 * This returns the length of copied instruction, or 0 if it has an error.
352 int __copy_instruction(u8 *dest, u8 *src, struct insn *insn)
354 kprobe_opcode_t buf[MAX_INSN_SIZE];
355 unsigned long recovered_insn =
356 recover_probed_instruction(buf, (unsigned long)src);
358 if (!recovered_insn || !insn)
361 /* This can access kernel text if given address is not recovered */
362 if (probe_kernel_read(dest, (void *)recovered_insn, MAX_INSN_SIZE))
365 kernel_insn_init(insn, dest, MAX_INSN_SIZE);
366 insn_get_length(insn);
368 /* Another subsystem puts a breakpoint, failed to recover */
369 if (insn->opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
373 /* Only x86_64 has RIP relative instructions */
374 if (insn_rip_relative(insn)) {
378 * The copied instruction uses the %rip-relative addressing
379 * mode. Adjust the displacement for the difference between
380 * the original location of this instruction and the location
381 * of the copy that will actually be run. The tricky bit here
382 * is making sure that the sign extension happens correctly in
383 * this calculation, since we need a signed 32-bit result to
384 * be sign-extended to 64 bits when it's added to the %rip
385 * value and yield the same 64-bit result that the sign-
386 * extension of the original signed 32-bit displacement would
389 newdisp = (u8 *) src + (s64) insn->displacement.value
391 if ((s64) (s32) newdisp != newdisp) {
392 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
393 pr_err("\tSrc: %p, Dest: %p, old disp: %x\n",
394 src, dest, insn->displacement.value);
397 disp = (u8 *) dest + insn_offset_displacement(insn);
398 *(s32 *) disp = (s32) newdisp;
404 /* Prepare reljump right after instruction to boost */
405 static void prepare_boost(struct kprobe *p, struct insn *insn)
407 if (can_boost(insn, p->addr) &&
408 MAX_INSN_SIZE - insn->length >= RELATIVEJUMP_SIZE) {
410 * These instructions can be executed directly if it
411 * jumps back to correct address.
413 synthesize_reljump(p->ainsn.insn + insn->length,
414 p->addr + insn->length);
415 p->ainsn.boostable = true;
417 p->ainsn.boostable = false;
421 /* Recover page to RW mode before releasing it */
422 void free_insn_page(void *page)
424 set_memory_nx((unsigned long)page & PAGE_MASK, 1);
425 set_memory_rw((unsigned long)page & PAGE_MASK, 1);
426 module_memfree(page);
429 static int arch_copy_kprobe(struct kprobe *p)
434 set_memory_rw((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
436 /* Copy an instruction with recovering if other optprobe modifies it.*/
437 len = __copy_instruction(p->ainsn.insn, p->addr, &insn);
442 * __copy_instruction can modify the displacement of the instruction,
443 * but it doesn't affect boostable check.
445 prepare_boost(p, &insn);
447 set_memory_ro((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
449 /* Check whether the instruction modifies Interrupt Flag or not */
450 p->ainsn.if_modifier = is_IF_modifier(p->ainsn.insn);
452 /* Also, displacement change doesn't affect the first byte */
453 p->opcode = p->ainsn.insn[0];
458 int arch_prepare_kprobe(struct kprobe *p)
460 if (alternatives_text_reserved(p->addr, p->addr))
463 if (!can_probe((unsigned long)p->addr))
465 /* insn: must be on special executable page on x86. */
466 p->ainsn.insn = get_insn_slot();
470 return arch_copy_kprobe(p);
473 void arch_arm_kprobe(struct kprobe *p)
475 text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
478 void arch_disarm_kprobe(struct kprobe *p)
480 text_poke(p->addr, &p->opcode, 1);
483 void arch_remove_kprobe(struct kprobe *p)
486 free_insn_slot(p->ainsn.insn, p->ainsn.boostable);
487 p->ainsn.insn = NULL;
491 static nokprobe_inline void
492 save_previous_kprobe(struct kprobe_ctlblk *kcb)
494 kcb->prev_kprobe.kp = kprobe_running();
495 kcb->prev_kprobe.status = kcb->kprobe_status;
496 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
497 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
500 static nokprobe_inline void
501 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
503 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
504 kcb->kprobe_status = kcb->prev_kprobe.status;
505 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
506 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
509 static nokprobe_inline void
510 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
511 struct kprobe_ctlblk *kcb)
513 __this_cpu_write(current_kprobe, p);
514 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
515 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
516 if (p->ainsn.if_modifier)
517 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
520 static nokprobe_inline void clear_btf(void)
522 if (test_thread_flag(TIF_BLOCKSTEP)) {
523 unsigned long debugctl = get_debugctlmsr();
525 debugctl &= ~DEBUGCTLMSR_BTF;
526 update_debugctlmsr(debugctl);
530 static nokprobe_inline void restore_btf(void)
532 if (test_thread_flag(TIF_BLOCKSTEP)) {
533 unsigned long debugctl = get_debugctlmsr();
535 debugctl |= DEBUGCTLMSR_BTF;
536 update_debugctlmsr(debugctl);
540 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
542 unsigned long *sara = stack_addr(regs);
544 ri->ret_addr = (kprobe_opcode_t *) *sara;
546 /* Replace the return addr with trampoline addr */
547 *sara = (unsigned long) &kretprobe_trampoline;
549 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
551 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
552 struct kprobe_ctlblk *kcb, int reenter)
554 if (setup_detour_execution(p, regs, reenter))
557 #if !defined(CONFIG_PREEMPT)
558 if (p->ainsn.boostable && !p->post_handler) {
559 /* Boost up -- we can execute copied instructions directly */
561 reset_current_kprobe();
563 * Reentering boosted probe doesn't reset current_kprobe,
564 * nor set current_kprobe, because it doesn't use single
567 regs->ip = (unsigned long)p->ainsn.insn;
568 preempt_enable_no_resched();
573 save_previous_kprobe(kcb);
574 set_current_kprobe(p, regs, kcb);
575 kcb->kprobe_status = KPROBE_REENTER;
577 kcb->kprobe_status = KPROBE_HIT_SS;
578 /* Prepare real single stepping */
580 regs->flags |= X86_EFLAGS_TF;
581 regs->flags &= ~X86_EFLAGS_IF;
582 /* single step inline if the instruction is an int3 */
583 if (p->opcode == BREAKPOINT_INSTRUCTION)
584 regs->ip = (unsigned long)p->addr;
586 regs->ip = (unsigned long)p->ainsn.insn;
588 NOKPROBE_SYMBOL(setup_singlestep);
591 * We have reentered the kprobe_handler(), since another probe was hit while
592 * within the handler. We save the original kprobes variables and just single
593 * step on the instruction of the new probe without calling any user handlers.
595 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
596 struct kprobe_ctlblk *kcb)
598 switch (kcb->kprobe_status) {
599 case KPROBE_HIT_SSDONE:
600 case KPROBE_HIT_ACTIVE:
602 kprobes_inc_nmissed_count(p);
603 setup_singlestep(p, regs, kcb, 1);
606 /* A probe has been hit in the codepath leading up to, or just
607 * after, single-stepping of a probed instruction. This entire
608 * codepath should strictly reside in .kprobes.text section.
609 * Raise a BUG or we'll continue in an endless reentering loop
610 * and eventually a stack overflow.
612 printk(KERN_WARNING "Unrecoverable kprobe detected at %p.\n",
617 /* impossible cases */
624 NOKPROBE_SYMBOL(reenter_kprobe);
627 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
628 * remain disabled throughout this function.
630 int kprobe_int3_handler(struct pt_regs *regs)
632 kprobe_opcode_t *addr;
634 struct kprobe_ctlblk *kcb;
639 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
641 * We don't want to be preempted for the entire
642 * duration of kprobe processing. We conditionally
643 * re-enable preemption at the end of this function,
644 * and also in reenter_kprobe() and setup_singlestep().
648 kcb = get_kprobe_ctlblk();
649 p = get_kprobe(addr);
652 if (kprobe_running()) {
653 if (reenter_kprobe(p, regs, kcb))
656 set_current_kprobe(p, regs, kcb);
657 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
660 * If we have no pre-handler or it returned 0, we
661 * continue with normal processing. If we have a
662 * pre-handler and it returned non-zero, it prepped
663 * for calling the break_handler below on re-entry
664 * for jprobe processing, so get out doing nothing
667 if (!p->pre_handler || !p->pre_handler(p, regs))
668 setup_singlestep(p, regs, kcb, 0);
671 } else if (*addr != BREAKPOINT_INSTRUCTION) {
673 * The breakpoint instruction was removed right
674 * after we hit it. Another cpu has removed
675 * either a probepoint or a debugger breakpoint
676 * at this address. In either case, no further
677 * handling of this interrupt is appropriate.
678 * Back up over the (now missing) int3 and run
679 * the original instruction.
681 regs->ip = (unsigned long)addr;
682 preempt_enable_no_resched();
684 } else if (kprobe_running()) {
685 p = __this_cpu_read(current_kprobe);
686 if (p->break_handler && p->break_handler(p, regs)) {
687 if (!skip_singlestep(p, regs, kcb))
688 setup_singlestep(p, regs, kcb, 0);
691 } /* else: not a kprobe fault; let the kernel handle it */
693 preempt_enable_no_resched();
696 NOKPROBE_SYMBOL(kprobe_int3_handler);
699 * When a retprobed function returns, this code saves registers and
700 * calls trampoline_handler() runs, which calls the kretprobe's handler.
703 ".global kretprobe_trampoline\n"
704 ".type kretprobe_trampoline, @function\n"
705 "kretprobe_trampoline:\n"
707 /* We don't bother saving the ss register */
712 " call trampoline_handler\n"
713 /* Replace saved sp with true return address. */
714 " movq %rax, 152(%rsp)\n"
721 " call trampoline_handler\n"
722 /* Move flags to cs */
723 " movl 56(%esp), %edx\n"
724 " movl %edx, 52(%esp)\n"
725 /* Replace saved flags with true return address. */
726 " movl %eax, 56(%esp)\n"
731 ".size kretprobe_trampoline, .-kretprobe_trampoline\n"
733 NOKPROBE_SYMBOL(kretprobe_trampoline);
734 STACK_FRAME_NON_STANDARD(kretprobe_trampoline);
737 * Called from kretprobe_trampoline
739 __visible __used void *trampoline_handler(struct pt_regs *regs)
741 struct kretprobe_instance *ri = NULL;
742 struct hlist_head *head, empty_rp;
743 struct hlist_node *tmp;
744 unsigned long flags, orig_ret_address = 0;
745 unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
746 kprobe_opcode_t *correct_ret_addr = NULL;
748 INIT_HLIST_HEAD(&empty_rp);
749 kretprobe_hash_lock(current, &head, &flags);
750 /* fixup registers */
752 regs->cs = __KERNEL_CS;
754 regs->cs = __KERNEL_CS | get_kernel_rpl();
757 regs->ip = trampoline_address;
758 regs->orig_ax = ~0UL;
761 * It is possible to have multiple instances associated with a given
762 * task either because multiple functions in the call path have
763 * return probes installed on them, and/or more than one
764 * return probe was registered for a target function.
766 * We can handle this because:
767 * - instances are always pushed into the head of the list
768 * - when multiple return probes are registered for the same
769 * function, the (chronologically) first instance's ret_addr
770 * will be the real return address, and all the rest will
771 * point to kretprobe_trampoline.
773 hlist_for_each_entry(ri, head, hlist) {
774 if (ri->task != current)
775 /* another task is sharing our hash bucket */
778 orig_ret_address = (unsigned long)ri->ret_addr;
780 if (orig_ret_address != trampoline_address)
782 * This is the real return address. Any other
783 * instances associated with this task are for
784 * other calls deeper on the call stack
789 kretprobe_assert(ri, orig_ret_address, trampoline_address);
791 correct_ret_addr = ri->ret_addr;
792 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
793 if (ri->task != current)
794 /* another task is sharing our hash bucket */
797 orig_ret_address = (unsigned long)ri->ret_addr;
798 if (ri->rp && ri->rp->handler) {
799 __this_cpu_write(current_kprobe, &ri->rp->kp);
800 get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
801 ri->ret_addr = correct_ret_addr;
802 ri->rp->handler(ri, regs);
803 __this_cpu_write(current_kprobe, NULL);
806 recycle_rp_inst(ri, &empty_rp);
808 if (orig_ret_address != trampoline_address)
810 * This is the real return address. Any other
811 * instances associated with this task are for
812 * other calls deeper on the call stack
817 kretprobe_hash_unlock(current, &flags);
819 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
820 hlist_del(&ri->hlist);
823 return (void *)orig_ret_address;
825 NOKPROBE_SYMBOL(trampoline_handler);
828 * Called after single-stepping. p->addr is the address of the
829 * instruction whose first byte has been replaced by the "int 3"
830 * instruction. To avoid the SMP problems that can occur when we
831 * temporarily put back the original opcode to single-step, we
832 * single-stepped a copy of the instruction. The address of this
833 * copy is p->ainsn.insn.
835 * This function prepares to return from the post-single-step
836 * interrupt. We have to fix up the stack as follows:
838 * 0) Except in the case of absolute or indirect jump or call instructions,
839 * the new ip is relative to the copied instruction. We need to make
840 * it relative to the original instruction.
842 * 1) If the single-stepped instruction was pushfl, then the TF and IF
843 * flags are set in the just-pushed flags, and may need to be cleared.
845 * 2) If the single-stepped instruction was a call, the return address
846 * that is atop the stack is the address following the copied instruction.
847 * We need to make it the address following the original instruction.
849 * If this is the first time we've single-stepped the instruction at
850 * this probepoint, and the instruction is boostable, boost it: add a
851 * jump instruction after the copied instruction, that jumps to the next
852 * instruction after the probepoint.
854 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
855 struct kprobe_ctlblk *kcb)
857 unsigned long *tos = stack_addr(regs);
858 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
859 unsigned long orig_ip = (unsigned long)p->addr;
860 kprobe_opcode_t *insn = p->ainsn.insn;
863 insn = skip_prefixes(insn);
865 regs->flags &= ~X86_EFLAGS_TF;
867 case 0x9c: /* pushfl */
868 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
869 *tos |= kcb->kprobe_old_flags;
871 case 0xc2: /* iret/ret/lret */
876 case 0xea: /* jmp absolute -- ip is correct */
877 /* ip is already adjusted, no more changes required */
878 p->ainsn.boostable = true;
880 case 0xe8: /* call relative - Fix return addr */
881 *tos = orig_ip + (*tos - copy_ip);
884 case 0x9a: /* call absolute -- same as call absolute, indirect */
885 *tos = orig_ip + (*tos - copy_ip);
889 if ((insn[1] & 0x30) == 0x10) {
891 * call absolute, indirect
892 * Fix return addr; ip is correct.
893 * But this is not boostable
895 *tos = orig_ip + (*tos - copy_ip);
897 } else if (((insn[1] & 0x31) == 0x20) ||
898 ((insn[1] & 0x31) == 0x21)) {
900 * jmp near and far, absolute indirect
901 * ip is correct. And this is boostable
903 p->ainsn.boostable = true;
910 regs->ip += orig_ip - copy_ip;
915 NOKPROBE_SYMBOL(resume_execution);
918 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
919 * remain disabled throughout this function.
921 int kprobe_debug_handler(struct pt_regs *regs)
923 struct kprobe *cur = kprobe_running();
924 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
929 resume_execution(cur, regs, kcb);
930 regs->flags |= kcb->kprobe_saved_flags;
932 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
933 kcb->kprobe_status = KPROBE_HIT_SSDONE;
934 cur->post_handler(cur, regs, 0);
937 /* Restore back the original saved kprobes variables and continue. */
938 if (kcb->kprobe_status == KPROBE_REENTER) {
939 restore_previous_kprobe(kcb);
942 reset_current_kprobe();
944 preempt_enable_no_resched();
947 * if somebody else is singlestepping across a probe point, flags
948 * will have TF set, in which case, continue the remaining processing
949 * of do_debug, as if this is not a probe hit.
951 if (regs->flags & X86_EFLAGS_TF)
956 NOKPROBE_SYMBOL(kprobe_debug_handler);
958 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
960 struct kprobe *cur = kprobe_running();
961 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
963 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
964 /* This must happen on single-stepping */
965 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
966 kcb->kprobe_status != KPROBE_REENTER);
968 * We are here because the instruction being single
969 * stepped caused a page fault. We reset the current
970 * kprobe and the ip points back to the probe address
971 * and allow the page fault handler to continue as a
974 regs->ip = (unsigned long)cur->addr;
976 * Trap flag (TF) has been set here because this fault
977 * happened where the single stepping will be done.
978 * So clear it by resetting the current kprobe:
980 regs->flags &= ~X86_EFLAGS_TF;
983 * If the TF flag was set before the kprobe hit,
986 regs->flags |= kcb->kprobe_old_flags;
988 if (kcb->kprobe_status == KPROBE_REENTER)
989 restore_previous_kprobe(kcb);
991 reset_current_kprobe();
992 preempt_enable_no_resched();
993 } else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
994 kcb->kprobe_status == KPROBE_HIT_SSDONE) {
996 * We increment the nmissed count for accounting,
997 * we can also use npre/npostfault count for accounting
998 * these specific fault cases.
1000 kprobes_inc_nmissed_count(cur);
1003 * We come here because instructions in the pre/post
1004 * handler caused the page_fault, this could happen
1005 * if handler tries to access user space by
1006 * copy_from_user(), get_user() etc. Let the
1007 * user-specified handler try to fix it first.
1009 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1013 * In case the user-specified fault handler returned
1014 * zero, try to fix up.
1016 if (fixup_exception(regs, trapnr))
1020 * fixup routine could not handle it,
1021 * Let do_page_fault() fix it.
1027 NOKPROBE_SYMBOL(kprobe_fault_handler);
1030 * Wrapper routine for handling exceptions.
1032 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1035 struct die_args *args = data;
1036 int ret = NOTIFY_DONE;
1038 if (args->regs && user_mode(args->regs))
1041 if (val == DIE_GPF) {
1043 * To be potentially processing a kprobe fault and to
1044 * trust the result from kprobe_running(), we have
1045 * be non-preemptible.
1047 if (!preemptible() && kprobe_running() &&
1048 kprobe_fault_handler(args->regs, args->trapnr))
1053 NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1055 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1057 struct jprobe *jp = container_of(p, struct jprobe, kp);
1059 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1061 kcb->jprobe_saved_regs = *regs;
1062 kcb->jprobe_saved_sp = stack_addr(regs);
1063 addr = (unsigned long)(kcb->jprobe_saved_sp);
1066 * As Linus pointed out, gcc assumes that the callee
1067 * owns the argument space and could overwrite it, e.g.
1068 * tailcall optimization. So, to be absolutely safe
1069 * we also save and restore enough stack bytes to cover
1070 * the argument area.
1071 * Use __memcpy() to avoid KASAN stack out-of-bounds reports as we copy
1072 * raw stack chunk with redzones:
1074 __memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr, MIN_STACK_SIZE(addr));
1075 regs->flags &= ~X86_EFLAGS_IF;
1076 trace_hardirqs_off();
1077 regs->ip = (unsigned long)(jp->entry);
1080 * jprobes use jprobe_return() which skips the normal return
1081 * path of the function, and this messes up the accounting of the
1082 * function graph tracer to get messed up.
1084 * Pause function graph tracing while performing the jprobe function.
1086 pause_graph_tracing();
1089 NOKPROBE_SYMBOL(setjmp_pre_handler);
1091 void jprobe_return(void)
1093 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1095 /* Unpoison stack redzones in the frames we are going to jump over. */
1096 kasan_unpoison_stack_above_sp_to(kcb->jprobe_saved_sp);
1099 #ifdef CONFIG_X86_64
1100 " xchg %%rbx,%%rsp \n"
1102 " xchgl %%ebx,%%esp \n"
1105 " .globl jprobe_return_end\n"
1106 " jprobe_return_end: \n"
1108 (kcb->jprobe_saved_sp):"memory");
1110 NOKPROBE_SYMBOL(jprobe_return);
1111 NOKPROBE_SYMBOL(jprobe_return_end);
1113 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1115 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1116 u8 *addr = (u8 *) (regs->ip - 1);
1117 struct jprobe *jp = container_of(p, struct jprobe, kp);
1118 void *saved_sp = kcb->jprobe_saved_sp;
1120 if ((addr > (u8 *) jprobe_return) &&
1121 (addr < (u8 *) jprobe_return_end)) {
1122 if (stack_addr(regs) != saved_sp) {
1123 struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1125 "current sp %p does not match saved sp %p\n",
1126 stack_addr(regs), saved_sp);
1127 printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1128 show_regs(saved_regs);
1129 printk(KERN_ERR "Current registers\n");
1133 /* It's OK to start function graph tracing again */
1134 unpause_graph_tracing();
1135 *regs = kcb->jprobe_saved_regs;
1136 __memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1137 preempt_enable_no_resched();
1142 NOKPROBE_SYMBOL(longjmp_break_handler);
1144 bool arch_within_kprobe_blacklist(unsigned long addr)
1146 return (addr >= (unsigned long)__kprobes_text_start &&
1147 addr < (unsigned long)__kprobes_text_end) ||
1148 (addr >= (unsigned long)__entry_text_start &&
1149 addr < (unsigned long)__entry_text_end);
1152 int __init arch_init_kprobes(void)
1157 int arch_trampoline_kprobe(struct kprobe *p)