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/module.h>
49 #include <linux/kdebug.h>
50 #include <linux/kallsyms.h>
51 #include <linux/ftrace.h>
52 #include <linux/moduleloader.h>
54 #include <asm/cacheflush.h>
56 #include <asm/pgtable.h>
57 #include <asm/uaccess.h>
58 #include <asm/alternative.h>
60 #include <asm/debugreg.h>
64 void jprobe_return_end(void);
66 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
67 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
69 #define stack_addr(regs) ((unsigned long *)kernel_stack_pointer(regs))
71 #define W(row, b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, ba, bb, bc, bd, be, bf)\
72 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
73 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
74 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
75 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
78 * Undefined/reserved opcodes, conditional jump, Opcode Extension
79 * Groups, and some special opcodes can not boost.
80 * This is non-const and volatile to keep gcc from statically
81 * optimizing it out, as variable_test_bit makes gcc think only
82 * *(unsigned long*) is used.
84 static volatile u32 twobyte_is_boostable[256 / 32] = {
85 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
86 /* ---------------------------------------------- */
87 W(0x00, 0, 0, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 0) | /* 00 */
88 W(0x10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1) , /* 10 */
89 W(0x20, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 20 */
90 W(0x30, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 30 */
91 W(0x40, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) | /* 40 */
92 W(0x50, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) , /* 50 */
93 W(0x60, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1) | /* 60 */
94 W(0x70, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1) , /* 70 */
95 W(0x80, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) | /* 80 */
96 W(0x90, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1) , /* 90 */
97 W(0xa0, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* a0 */
98 W(0xb0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1) , /* b0 */
99 W(0xc0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1) | /* c0 */
100 W(0xd0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) , /* d0 */
101 W(0xe0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1) | /* e0 */
102 W(0xf0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0) /* f0 */
103 /* ----------------------------------------------- */
104 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
108 struct kretprobe_blackpoint kretprobe_blacklist[] = {
109 {"__switch_to", }, /* This function switches only current task, but
110 doesn't switch kernel stack.*/
111 {NULL, NULL} /* Terminator */
114 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
116 static nokprobe_inline void
117 __synthesize_relative_insn(void *from, void *to, u8 op)
119 struct __arch_relative_insn {
124 insn = (struct __arch_relative_insn *)from;
125 insn->raddr = (s32)((long)(to) - ((long)(from) + 5));
129 /* Insert a jump instruction at address 'from', which jumps to address 'to'.*/
130 void synthesize_reljump(void *from, void *to)
132 __synthesize_relative_insn(from, to, RELATIVEJUMP_OPCODE);
134 NOKPROBE_SYMBOL(synthesize_reljump);
136 /* Insert a call instruction at address 'from', which calls address 'to'.*/
137 void synthesize_relcall(void *from, void *to)
139 __synthesize_relative_insn(from, to, RELATIVECALL_OPCODE);
141 NOKPROBE_SYMBOL(synthesize_relcall);
144 * Skip the prefixes of the instruction.
146 static kprobe_opcode_t *skip_prefixes(kprobe_opcode_t *insn)
150 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
151 while (inat_is_legacy_prefix(attr)) {
153 attr = inat_get_opcode_attribute((insn_byte_t)*insn);
156 if (inat_is_rex_prefix(attr))
161 NOKPROBE_SYMBOL(skip_prefixes);
164 * Returns non-zero if opcode is boostable.
165 * RIP relative instructions are adjusted at copying time in 64 bits mode
167 int can_boost(kprobe_opcode_t *opcodes, void *addr)
169 kprobe_opcode_t opcode;
170 kprobe_opcode_t *orig_opcodes = opcodes;
172 if (search_exception_tables((unsigned long)addr))
173 return 0; /* Page fault may occur on this address. */
176 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
178 opcode = *(opcodes++);
180 /* 2nd-byte opcode */
181 if (opcode == 0x0f) {
182 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
184 return test_bit(*opcodes,
185 (unsigned long *)twobyte_is_boostable);
188 switch (opcode & 0xf0) {
191 goto retry; /* REX prefix is boostable */
194 if (0x63 < opcode && opcode < 0x67)
195 goto retry; /* prefixes */
196 /* can't boost Address-size override and bound */
197 return (opcode != 0x62 && opcode != 0x67);
199 return 0; /* can't boost conditional jump */
201 return opcode != 0x9a; /* can't boost call far */
203 /* can't boost software-interruptions */
204 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
206 /* can boost AA* and XLAT */
207 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
209 /* can boost in/out and absolute jmps */
210 return ((opcode & 0x04) || opcode == 0xea);
212 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
213 goto retry; /* lock/rep(ne) prefix */
214 /* clear and set flags are boostable */
215 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
217 /* segment override prefixes are boostable */
218 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
219 goto retry; /* prefixes */
220 /* CS override prefix and call are not boostable */
221 return (opcode != 0x2e && opcode != 0x9a);
226 __recover_probed_insn(kprobe_opcode_t *buf, unsigned long addr)
231 kp = get_kprobe((void *)addr);
232 faddr = ftrace_location(addr);
234 * Addresses inside the ftrace location are refused by
235 * arch_check_ftrace_location(). Something went terribly wrong
236 * if such an address is checked here.
238 if (WARN_ON(faddr && faddr != addr))
241 * Use the current code if it is not modified by Kprobe
242 * and it cannot be modified by ftrace.
248 * Basically, kp->ainsn.insn has an original instruction.
249 * However, RIP-relative instruction can not do single-stepping
250 * at different place, __copy_instruction() tweaks the displacement of
251 * that instruction. In that case, we can't recover the instruction
252 * from the kp->ainsn.insn.
254 * On the other hand, in case on normal Kprobe, kp->opcode has a copy
255 * of the first byte of the probed instruction, which is overwritten
256 * by int3. And the instruction at kp->addr is not modified by kprobes
257 * except for the first byte, we can recover the original instruction
258 * from it and kp->opcode.
260 * In case of Kprobes using ftrace, we do not have a copy of
261 * the original instruction. In fact, the ftrace location might
262 * be modified at anytime and even could be in an inconsistent state.
263 * Fortunately, we know that the original code is the ideal 5-byte
266 memcpy(buf, (void *)addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
268 memcpy(buf, ideal_nops[NOP_ATOMIC5], 5);
271 return (unsigned long)buf;
275 * Recover the probed instruction at addr for further analysis.
276 * Caller must lock kprobes by kprobe_mutex, or disable preemption
277 * for preventing to release referencing kprobes.
278 * Returns zero if the instruction can not get recovered.
280 unsigned long recover_probed_instruction(kprobe_opcode_t *buf, unsigned long addr)
282 unsigned long __addr;
284 __addr = __recover_optprobed_insn(buf, addr);
288 return __recover_probed_insn(buf, addr);
291 /* Check if paddr is at an instruction boundary */
292 static int can_probe(unsigned long paddr)
294 unsigned long addr, __addr, offset = 0;
296 kprobe_opcode_t buf[MAX_INSN_SIZE];
298 if (!kallsyms_lookup_size_offset(paddr, NULL, &offset))
301 /* Decode instructions */
302 addr = paddr - offset;
303 while (addr < paddr) {
305 * Check if the instruction has been modified by another
306 * kprobe, in which case we replace the breakpoint by the
307 * original instruction in our buffer.
308 * Also, jump optimization will change the breakpoint to
309 * relative-jump. Since the relative-jump itself is
310 * normally used, we just go through if there is no kprobe.
312 __addr = recover_probed_instruction(buf, addr);
315 kernel_insn_init(&insn, (void *)__addr, MAX_INSN_SIZE);
316 insn_get_length(&insn);
319 * Another debugging subsystem might insert this breakpoint.
320 * In that case, we can't recover it.
322 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
327 return (addr == paddr);
331 * Returns non-zero if opcode modifies the interrupt flag.
333 static int is_IF_modifier(kprobe_opcode_t *insn)
336 insn = skip_prefixes(insn);
341 case 0xcf: /* iret/iretd */
342 case 0x9d: /* popf/popfd */
350 * Copy an instruction and adjust the displacement if the instruction
351 * uses the %rip-relative addressing mode.
352 * If it does, Return the address of the 32-bit displacement word.
353 * If not, return null.
354 * Only applicable to 64-bit x86.
356 int __copy_instruction(u8 *dest, u8 *src)
359 kprobe_opcode_t buf[MAX_INSN_SIZE];
361 unsigned long recovered_insn =
362 recover_probed_instruction(buf, (unsigned long)src);
366 kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
367 insn_get_length(&insn);
368 length = insn.length;
370 /* Another subsystem puts a breakpoint, failed to recover */
371 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
373 memcpy(dest, insn.kaddr, length);
375 /* We should not singlestep on the exception masking instructions */
376 if (insn_masking_exception(&insn))
380 if (insn_rip_relative(&insn)) {
383 kernel_insn_init(&insn, dest, length);
384 insn_get_displacement(&insn);
386 * The copied instruction uses the %rip-relative addressing
387 * mode. Adjust the displacement for the difference between
388 * the original location of this instruction and the location
389 * of the copy that will actually be run. The tricky bit here
390 * is making sure that the sign extension happens correctly in
391 * this calculation, since we need a signed 32-bit result to
392 * be sign-extended to 64 bits when it's added to the %rip
393 * value and yield the same 64-bit result that the sign-
394 * extension of the original signed 32-bit displacement would
397 newdisp = (u8 *) src + (s64) insn.displacement.value - (u8 *) dest;
398 if ((s64) (s32) newdisp != newdisp) {
399 pr_err("Kprobes error: new displacement does not fit into s32 (%llx)\n", newdisp);
402 disp = (u8 *) dest + insn_offset_displacement(&insn);
403 *(s32 *) disp = (s32) newdisp;
409 /* Recover page to RW mode before releasing it */
410 void free_insn_page(void *page)
412 set_memory_nx((unsigned long)page & PAGE_MASK, 1);
413 set_memory_rw((unsigned long)page & PAGE_MASK, 1);
414 module_memfree(page);
417 /* Prepare reljump right after instruction to boost */
418 static void prepare_boost(struct kprobe *p, int length)
420 if (can_boost(p->ainsn.insn, p->addr) &&
421 MAX_INSN_SIZE - length >= RELATIVEJUMP_SIZE) {
423 * These instructions can be executed directly if it
424 * jumps back to correct address.
426 synthesize_reljump(p->ainsn.insn + length, p->addr + length);
427 p->ainsn.boostable = 1;
429 p->ainsn.boostable = -1;
433 static int arch_copy_kprobe(struct kprobe *p)
437 set_memory_rw((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
439 /* Copy an instruction with recovering if other optprobe modifies it.*/
440 len = __copy_instruction(p->ainsn.insn, p->addr);
445 * __copy_instruction can modify the displacement of the instruction,
446 * but it doesn't affect boostable check.
448 prepare_boost(p, len);
450 set_memory_ro((unsigned long)p->ainsn.insn & PAGE_MASK, 1);
452 /* Check whether the instruction modifies Interrupt Flag or not */
453 p->ainsn.if_modifier = is_IF_modifier(p->ainsn.insn);
455 /* Also, displacement change doesn't affect the first byte */
456 p->opcode = p->ainsn.insn[0];
461 int arch_prepare_kprobe(struct kprobe *p)
463 if (alternatives_text_reserved(p->addr, p->addr))
466 if (!can_probe((unsigned long)p->addr))
468 /* insn: must be on special executable page on x86. */
469 p->ainsn.insn = get_insn_slot();
473 return arch_copy_kprobe(p);
476 void arch_arm_kprobe(struct kprobe *p)
478 text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
481 void arch_disarm_kprobe(struct kprobe *p)
483 text_poke(p->addr, &p->opcode, 1);
486 void arch_remove_kprobe(struct kprobe *p)
489 free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
490 p->ainsn.insn = NULL;
494 static nokprobe_inline void
495 save_previous_kprobe(struct kprobe_ctlblk *kcb)
497 kcb->prev_kprobe.kp = kprobe_running();
498 kcb->prev_kprobe.status = kcb->kprobe_status;
499 kcb->prev_kprobe.old_flags = kcb->kprobe_old_flags;
500 kcb->prev_kprobe.saved_flags = kcb->kprobe_saved_flags;
503 static nokprobe_inline void
504 restore_previous_kprobe(struct kprobe_ctlblk *kcb)
506 __this_cpu_write(current_kprobe, kcb->prev_kprobe.kp);
507 kcb->kprobe_status = kcb->prev_kprobe.status;
508 kcb->kprobe_old_flags = kcb->prev_kprobe.old_flags;
509 kcb->kprobe_saved_flags = kcb->prev_kprobe.saved_flags;
512 static nokprobe_inline void
513 set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
514 struct kprobe_ctlblk *kcb)
516 __this_cpu_write(current_kprobe, p);
517 kcb->kprobe_saved_flags = kcb->kprobe_old_flags
518 = (regs->flags & (X86_EFLAGS_TF | X86_EFLAGS_IF));
519 if (p->ainsn.if_modifier)
520 kcb->kprobe_saved_flags &= ~X86_EFLAGS_IF;
523 static nokprobe_inline void clear_btf(void)
525 if (test_thread_flag(TIF_BLOCKSTEP)) {
526 unsigned long debugctl = get_debugctlmsr();
528 debugctl &= ~DEBUGCTLMSR_BTF;
529 update_debugctlmsr(debugctl);
533 static nokprobe_inline void restore_btf(void)
535 if (test_thread_flag(TIF_BLOCKSTEP)) {
536 unsigned long debugctl = get_debugctlmsr();
538 debugctl |= DEBUGCTLMSR_BTF;
539 update_debugctlmsr(debugctl);
543 void arch_prepare_kretprobe(struct kretprobe_instance *ri, struct pt_regs *regs)
545 unsigned long *sara = stack_addr(regs);
547 ri->ret_addr = (kprobe_opcode_t *) *sara;
550 /* Replace the return addr with trampoline addr */
551 *sara = (unsigned long) &kretprobe_trampoline;
553 NOKPROBE_SYMBOL(arch_prepare_kretprobe);
555 static void setup_singlestep(struct kprobe *p, struct pt_regs *regs,
556 struct kprobe_ctlblk *kcb, int reenter)
558 if (setup_detour_execution(p, regs, reenter))
561 #if !defined(CONFIG_PREEMPT)
562 if (p->ainsn.boostable == 1 && !p->post_handler) {
563 /* Boost up -- we can execute copied instructions directly */
565 reset_current_kprobe();
567 * Reentering boosted probe doesn't reset current_kprobe,
568 * nor set current_kprobe, because it doesn't use single
571 regs->ip = (unsigned long)p->ainsn.insn;
572 preempt_enable_no_resched();
577 save_previous_kprobe(kcb);
578 set_current_kprobe(p, regs, kcb);
579 kcb->kprobe_status = KPROBE_REENTER;
581 kcb->kprobe_status = KPROBE_HIT_SS;
582 /* Prepare real single stepping */
584 regs->flags |= X86_EFLAGS_TF;
585 regs->flags &= ~X86_EFLAGS_IF;
586 /* single step inline if the instruction is an int3 */
587 if (p->opcode == BREAKPOINT_INSTRUCTION)
588 regs->ip = (unsigned long)p->addr;
590 regs->ip = (unsigned long)p->ainsn.insn;
592 NOKPROBE_SYMBOL(setup_singlestep);
595 * We have reentered the kprobe_handler(), since another probe was hit while
596 * within the handler. We save the original kprobes variables and just single
597 * step on the instruction of the new probe without calling any user handlers.
599 static int reenter_kprobe(struct kprobe *p, struct pt_regs *regs,
600 struct kprobe_ctlblk *kcb)
602 switch (kcb->kprobe_status) {
603 case KPROBE_HIT_SSDONE:
604 case KPROBE_HIT_ACTIVE:
606 kprobes_inc_nmissed_count(p);
607 setup_singlestep(p, regs, kcb, 1);
610 /* A probe has been hit in the codepath leading up to, or just
611 * after, single-stepping of a probed instruction. This entire
612 * codepath should strictly reside in .kprobes.text section.
613 * Raise a BUG or we'll continue in an endless reentering loop
614 * and eventually a stack overflow.
616 pr_err("Unrecoverable kprobe detected.\n");
620 /* impossible cases */
627 NOKPROBE_SYMBOL(reenter_kprobe);
630 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
631 * remain disabled throughout this function.
633 int kprobe_int3_handler(struct pt_regs *regs)
635 kprobe_opcode_t *addr;
637 struct kprobe_ctlblk *kcb;
642 addr = (kprobe_opcode_t *)(regs->ip - sizeof(kprobe_opcode_t));
644 * We don't want to be preempted for the entire
645 * duration of kprobe processing. We conditionally
646 * re-enable preemption at the end of this function,
647 * and also in reenter_kprobe() and setup_singlestep().
651 kcb = get_kprobe_ctlblk();
652 p = get_kprobe(addr);
655 if (kprobe_running()) {
656 if (reenter_kprobe(p, regs, kcb))
659 set_current_kprobe(p, regs, kcb);
660 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
663 * If we have no pre-handler or it returned 0, we
664 * continue with normal processing. If we have a
665 * pre-handler and it returned non-zero, it prepped
666 * for calling the break_handler below on re-entry
667 * for jprobe processing, so get out doing nothing
670 if (!p->pre_handler || !p->pre_handler(p, regs))
671 setup_singlestep(p, regs, kcb, 0);
674 } else if (*addr != BREAKPOINT_INSTRUCTION) {
676 * The breakpoint instruction was removed right
677 * after we hit it. Another cpu has removed
678 * either a probepoint or a debugger breakpoint
679 * at this address. In either case, no further
680 * handling of this interrupt is appropriate.
681 * Back up over the (now missing) int3 and run
682 * the original instruction.
684 regs->ip = (unsigned long)addr;
685 preempt_enable_no_resched();
687 } else if (kprobe_running()) {
688 p = __this_cpu_read(current_kprobe);
689 if (p->break_handler && p->break_handler(p, regs)) {
690 if (!skip_singlestep(p, regs, kcb))
691 setup_singlestep(p, regs, kcb, 0);
694 } /* else: not a kprobe fault; let the kernel handle it */
696 preempt_enable_no_resched();
699 NOKPROBE_SYMBOL(kprobe_int3_handler);
702 * When a retprobed function returns, this code saves registers and
703 * calls trampoline_handler() runs, which calls the kretprobe's handler.
705 static void __used kretprobe_trampoline_holder(void)
708 ".global kretprobe_trampoline\n"
709 "kretprobe_trampoline: \n"
711 /* We don't bother saving the ss register */
716 " call trampoline_handler\n"
717 /* Replace saved sp with true return address. */
718 " movq %rax, 152(%rsp)\n"
725 " call trampoline_handler\n"
726 /* Move flags to cs */
727 " movl 56(%esp), %edx\n"
728 " movl %edx, 52(%esp)\n"
729 /* Replace saved flags with true return address. */
730 " movl %eax, 56(%esp)\n"
736 NOKPROBE_SYMBOL(kretprobe_trampoline_holder);
737 NOKPROBE_SYMBOL(kretprobe_trampoline);
740 * Called from kretprobe_trampoline
742 __visible __used void *trampoline_handler(struct pt_regs *regs)
744 struct kretprobe_instance *ri = NULL;
745 struct hlist_head *head, empty_rp;
746 struct hlist_node *tmp;
747 unsigned long flags, orig_ret_address = 0;
748 unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
749 kprobe_opcode_t *correct_ret_addr = NULL;
751 bool skipped = false;
754 * Set a dummy kprobe for avoiding kretprobe recursion.
755 * Since kretprobe never run in kprobe handler, kprobe must not
756 * be running at this point.
760 INIT_HLIST_HEAD(&empty_rp);
761 kretprobe_hash_lock(current, &head, &flags);
762 /* fixup registers */
764 regs->cs = __KERNEL_CS;
765 /* On x86-64, we use pt_regs->sp for return address holder. */
766 frame_pointer = ®s->sp;
768 regs->cs = __KERNEL_CS | get_kernel_rpl();
770 /* On x86-32, we use pt_regs->flags for return address holder. */
771 frame_pointer = ®s->flags;
773 regs->ip = trampoline_address;
774 regs->orig_ax = ~0UL;
777 * It is possible to have multiple instances associated with a given
778 * task either because multiple functions in the call path have
779 * return probes installed on them, and/or more than one
780 * return probe was registered for a target function.
782 * We can handle this because:
783 * - instances are always pushed into the head of the list
784 * - when multiple return probes are registered for the same
785 * function, the (chronologically) first instance's ret_addr
786 * will be the real return address, and all the rest will
787 * point to kretprobe_trampoline.
789 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
790 if (ri->task != current)
791 /* another task is sharing our hash bucket */
794 * Return probes must be pushed on this hash list correct
795 * order (same as return order) so that it can be poped
796 * correctly. However, if we find it is pushed it incorrect
797 * order, this means we find a function which should not be
798 * probed, because the wrong order entry is pushed on the
799 * path of processing other kretprobe itself.
801 if (ri->fp != frame_pointer) {
803 pr_warn("kretprobe is stacked incorrectly. Trying to fixup.\n");
808 orig_ret_address = (unsigned long)ri->ret_addr;
810 pr_warn("%ps must be blacklisted because of incorrect kretprobe order\n",
813 if (orig_ret_address != trampoline_address)
815 * This is the real return address. Any other
816 * instances associated with this task are for
817 * other calls deeper on the call stack
822 kretprobe_assert(ri, orig_ret_address, trampoline_address);
824 correct_ret_addr = ri->ret_addr;
825 hlist_for_each_entry_safe(ri, tmp, head, hlist) {
826 if (ri->task != current)
827 /* another task is sharing our hash bucket */
829 if (ri->fp != frame_pointer)
832 orig_ret_address = (unsigned long)ri->ret_addr;
833 if (ri->rp && ri->rp->handler) {
834 __this_cpu_write(current_kprobe, &ri->rp->kp);
835 ri->ret_addr = correct_ret_addr;
836 ri->rp->handler(ri, regs);
837 __this_cpu_write(current_kprobe, &kprobe_busy);
840 recycle_rp_inst(ri, &empty_rp);
842 if (orig_ret_address != trampoline_address)
844 * This is the real return address. Any other
845 * instances associated with this task are for
846 * other calls deeper on the call stack
851 kretprobe_hash_unlock(current, &flags);
855 hlist_for_each_entry_safe(ri, tmp, &empty_rp, hlist) {
856 hlist_del(&ri->hlist);
859 return (void *)orig_ret_address;
861 NOKPROBE_SYMBOL(trampoline_handler);
864 * Called after single-stepping. p->addr is the address of the
865 * instruction whose first byte has been replaced by the "int 3"
866 * instruction. To avoid the SMP problems that can occur when we
867 * temporarily put back the original opcode to single-step, we
868 * single-stepped a copy of the instruction. The address of this
869 * copy is p->ainsn.insn.
871 * This function prepares to return from the post-single-step
872 * interrupt. We have to fix up the stack as follows:
874 * 0) Except in the case of absolute or indirect jump or call instructions,
875 * the new ip is relative to the copied instruction. We need to make
876 * it relative to the original instruction.
878 * 1) If the single-stepped instruction was pushfl, then the TF and IF
879 * flags are set in the just-pushed flags, and may need to be cleared.
881 * 2) If the single-stepped instruction was a call, the return address
882 * that is atop the stack is the address following the copied instruction.
883 * We need to make it the address following the original instruction.
885 * If this is the first time we've single-stepped the instruction at
886 * this probepoint, and the instruction is boostable, boost it: add a
887 * jump instruction after the copied instruction, that jumps to the next
888 * instruction after the probepoint.
890 static void resume_execution(struct kprobe *p, struct pt_regs *regs,
891 struct kprobe_ctlblk *kcb)
893 unsigned long *tos = stack_addr(regs);
894 unsigned long copy_ip = (unsigned long)p->ainsn.insn;
895 unsigned long orig_ip = (unsigned long)p->addr;
896 kprobe_opcode_t *insn = p->ainsn.insn;
899 insn = skip_prefixes(insn);
901 regs->flags &= ~X86_EFLAGS_TF;
903 case 0x9c: /* pushfl */
904 *tos &= ~(X86_EFLAGS_TF | X86_EFLAGS_IF);
905 *tos |= kcb->kprobe_old_flags;
907 case 0xc2: /* iret/ret/lret */
912 case 0xea: /* jmp absolute -- ip is correct */
913 /* ip is already adjusted, no more changes required */
914 p->ainsn.boostable = 1;
916 case 0xe8: /* call relative - Fix return addr */
917 *tos = orig_ip + (*tos - copy_ip);
920 case 0x9a: /* call absolute -- same as call absolute, indirect */
921 *tos = orig_ip + (*tos - copy_ip);
925 if ((insn[1] & 0x30) == 0x10) {
927 * call absolute, indirect
928 * Fix return addr; ip is correct.
929 * But this is not boostable
931 *tos = orig_ip + (*tos - copy_ip);
933 } else if (((insn[1] & 0x31) == 0x20) ||
934 ((insn[1] & 0x31) == 0x21)) {
936 * jmp near and far, absolute indirect
937 * ip is correct. And this is boostable
939 p->ainsn.boostable = 1;
946 regs->ip += orig_ip - copy_ip;
951 NOKPROBE_SYMBOL(resume_execution);
954 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
955 * remain disabled throughout this function.
957 int kprobe_debug_handler(struct pt_regs *regs)
959 struct kprobe *cur = kprobe_running();
960 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
965 resume_execution(cur, regs, kcb);
966 regs->flags |= kcb->kprobe_saved_flags;
968 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
969 kcb->kprobe_status = KPROBE_HIT_SSDONE;
970 cur->post_handler(cur, regs, 0);
973 /* Restore back the original saved kprobes variables and continue. */
974 if (kcb->kprobe_status == KPROBE_REENTER) {
975 restore_previous_kprobe(kcb);
978 reset_current_kprobe();
980 preempt_enable_no_resched();
983 * if somebody else is singlestepping across a probe point, flags
984 * will have TF set, in which case, continue the remaining processing
985 * of do_debug, as if this is not a probe hit.
987 if (regs->flags & X86_EFLAGS_TF)
992 NOKPROBE_SYMBOL(kprobe_debug_handler);
994 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
996 struct kprobe *cur = kprobe_running();
997 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
999 if (unlikely(regs->ip == (unsigned long)cur->ainsn.insn)) {
1000 /* This must happen on single-stepping */
1001 WARN_ON(kcb->kprobe_status != KPROBE_HIT_SS &&
1002 kcb->kprobe_status != KPROBE_REENTER);
1004 * We are here because the instruction being single
1005 * stepped caused a page fault. We reset the current
1006 * kprobe and the ip points back to the probe address
1007 * and allow the page fault handler to continue as a
1008 * normal page fault.
1010 regs->ip = (unsigned long)cur->addr;
1012 * Trap flag (TF) has been set here because this fault
1013 * happened where the single stepping will be done.
1014 * So clear it by resetting the current kprobe:
1016 regs->flags &= ~X86_EFLAGS_TF;
1018 * Since the single step (trap) has been cancelled,
1019 * we need to restore BTF here.
1024 * If the TF flag was set before the kprobe hit,
1027 regs->flags |= kcb->kprobe_old_flags;
1029 if (kcb->kprobe_status == KPROBE_REENTER)
1030 restore_previous_kprobe(kcb);
1032 reset_current_kprobe();
1033 preempt_enable_no_resched();
1034 } else if (kcb->kprobe_status == KPROBE_HIT_ACTIVE ||
1035 kcb->kprobe_status == KPROBE_HIT_SSDONE) {
1037 * We increment the nmissed count for accounting,
1038 * we can also use npre/npostfault count for accounting
1039 * these specific fault cases.
1041 kprobes_inc_nmissed_count(cur);
1044 * We come here because instructions in the pre/post
1045 * handler caused the page_fault, this could happen
1046 * if handler tries to access user space by
1047 * copy_from_user(), get_user() etc. Let the
1048 * user-specified handler try to fix it first.
1050 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
1054 * In case the user-specified fault handler returned
1055 * zero, try to fix up.
1057 if (fixup_exception(regs))
1061 * fixup routine could not handle it,
1062 * Let do_page_fault() fix it.
1068 NOKPROBE_SYMBOL(kprobe_fault_handler);
1071 * Wrapper routine for handling exceptions.
1073 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
1076 struct die_args *args = data;
1077 int ret = NOTIFY_DONE;
1079 if (args->regs && user_mode(args->regs))
1082 if (val == DIE_GPF) {
1084 * To be potentially processing a kprobe fault and to
1085 * trust the result from kprobe_running(), we have
1086 * be non-preemptible.
1088 if (!preemptible() && kprobe_running() &&
1089 kprobe_fault_handler(args->regs, args->trapnr))
1094 NOKPROBE_SYMBOL(kprobe_exceptions_notify);
1096 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
1098 struct jprobe *jp = container_of(p, struct jprobe, kp);
1100 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1102 kcb->jprobe_saved_regs = *regs;
1103 kcb->jprobe_saved_sp = stack_addr(regs);
1104 addr = (unsigned long)(kcb->jprobe_saved_sp);
1107 * As Linus pointed out, gcc assumes that the callee
1108 * owns the argument space and could overwrite it, e.g.
1109 * tailcall optimization. So, to be absolutely safe
1110 * we also save and restore enough stack bytes to cover
1111 * the argument area.
1113 memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
1114 MIN_STACK_SIZE(addr));
1115 regs->flags &= ~X86_EFLAGS_IF;
1116 trace_hardirqs_off();
1117 regs->ip = (unsigned long)(jp->entry);
1120 * jprobes use jprobe_return() which skips the normal return
1121 * path of the function, and this messes up the accounting of the
1122 * function graph tracer to get messed up.
1124 * Pause function graph tracing while performing the jprobe function.
1126 pause_graph_tracing();
1129 NOKPROBE_SYMBOL(setjmp_pre_handler);
1131 void jprobe_return(void)
1133 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1136 #ifdef CONFIG_X86_64
1137 " xchg %%rbx,%%rsp \n"
1139 " xchgl %%ebx,%%esp \n"
1142 " .globl jprobe_return_end\n"
1143 " jprobe_return_end: \n"
1145 (kcb->jprobe_saved_sp):"memory");
1147 NOKPROBE_SYMBOL(jprobe_return);
1148 NOKPROBE_SYMBOL(jprobe_return_end);
1150 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
1152 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
1153 u8 *addr = (u8 *) (regs->ip - 1);
1154 struct jprobe *jp = container_of(p, struct jprobe, kp);
1155 void *saved_sp = kcb->jprobe_saved_sp;
1157 if ((addr > (u8 *) jprobe_return) &&
1158 (addr < (u8 *) jprobe_return_end)) {
1159 if (stack_addr(regs) != saved_sp) {
1160 struct pt_regs *saved_regs = &kcb->jprobe_saved_regs;
1162 "current sp %p does not match saved sp %p\n",
1163 stack_addr(regs), saved_sp);
1164 printk(KERN_ERR "Saved registers for jprobe %p\n", jp);
1165 show_regs(saved_regs);
1166 printk(KERN_ERR "Current registers\n");
1170 /* It's OK to start function graph tracing again */
1171 unpause_graph_tracing();
1172 *regs = kcb->jprobe_saved_regs;
1173 memcpy(saved_sp, kcb->jprobes_stack, MIN_STACK_SIZE(saved_sp));
1174 preempt_enable_no_resched();
1179 NOKPROBE_SYMBOL(longjmp_break_handler);
1181 bool arch_within_kprobe_blacklist(unsigned long addr)
1183 return (addr >= (unsigned long)__kprobes_text_start &&
1184 addr < (unsigned long)__kprobes_text_end) ||
1185 (addr >= (unsigned long)__entry_text_start &&
1186 addr < (unsigned long)__entry_text_end);
1189 int __init arch_init_kprobes(void)
1194 int arch_trampoline_kprobe(struct kprobe *p)