2 * Kernel Probes (KProbes)
3 * arch/i386/kernel/kprobes.c
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
19 * Copyright (C) IBM Corporation, 2002, 2004
21 * 2002-Oct Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
22 * Probes initial implementation ( includes contributions from
24 * 2004-July Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
25 * interface to access function arguments.
26 * 2005-May Hien Nguyen <hien@us.ibm.com>, Jim Keniston
27 * <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
28 * <prasanna@in.ibm.com> added function-return probes.
31 #include <linux/config.h>
32 #include <linux/kprobes.h>
33 #include <linux/ptrace.h>
34 #include <linux/preempt.h>
35 #include <asm/cacheflush.h>
36 #include <asm/kdebug.h>
38 #include <asm/uaccess.h>
40 void jprobe_return_end(void);
42 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
43 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
45 /* insert a jmp code */
46 static __always_inline void set_jmp_op(void *from, void *to)
48 struct __arch_jmp_op {
51 } __attribute__((packed)) *jop;
52 jop = (struct __arch_jmp_op *)from;
53 jop->raddr = (long)(to) - ((long)(from) + 5);
54 jop->op = RELATIVEJUMP_INSTRUCTION;
58 * returns non-zero if opcodes can be boosted.
60 static __always_inline int can_boost(kprobe_opcode_t *opcodes)
62 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf) \
63 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
64 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
65 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
66 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
69 * Undefined/reserved opcodes, conditional jump, Opcode Extension
70 * Groups, and some special opcodes can not be boost.
72 static const unsigned long twobyte_is_boostable[256 / 32] = {
73 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
74 /* ------------------------------- */
75 W(0x00, 0,0,1,1,0,0,1,0,1,1,0,0,0,0,0,0)| /* 00 */
76 W(0x10, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 10 */
77 W(0x20, 1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0)| /* 20 */
78 W(0x30, 0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 30 */
79 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 40 */
80 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 50 */
81 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,0,0,1,1)| /* 60 */
82 W(0x70, 0,0,0,0,1,1,1,1,0,0,0,0,0,0,1,1), /* 70 */
83 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 80 */
84 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1), /* 90 */
85 W(0xa0, 1,1,0,1,1,1,0,0,1,1,0,1,1,1,0,1)| /* a0 */
86 W(0xb0, 1,1,1,1,1,1,1,1,0,0,0,1,1,1,1,1), /* b0 */
87 W(0xc0, 1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1)| /* c0 */
88 W(0xd0, 0,1,1,1,0,1,0,0,1,1,0,1,1,1,0,1), /* d0 */
89 W(0xe0, 0,1,1,0,0,1,0,0,1,1,0,1,1,1,0,1)| /* e0 */
90 W(0xf0, 0,1,1,1,0,1,0,0,1,1,1,0,1,1,1,0) /* f0 */
91 /* ------------------------------- */
92 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
95 kprobe_opcode_t opcode;
96 kprobe_opcode_t *orig_opcodes = opcodes;
98 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
100 opcode = *(opcodes++);
102 /* 2nd-byte opcode */
103 if (opcode == 0x0f) {
104 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
106 return test_bit(*opcodes, twobyte_is_boostable);
109 switch (opcode & 0xf0) {
111 if (0x63 < opcode && opcode < 0x67)
112 goto retry; /* prefixes */
113 /* can't boost Address-size override and bound */
114 return (opcode != 0x62 && opcode != 0x67);
116 return 0; /* can't boost conditional jump */
118 /* can't boost software-interruptions */
119 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
121 /* can boost AA* and XLAT */
122 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
124 /* can boost in/out and absolute jmps */
125 return ((opcode & 0x04) || opcode == 0xea);
127 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
128 goto retry; /* lock/rep(ne) prefix */
129 /* clear and set flags can be boost */
130 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
132 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
133 goto retry; /* prefixes */
134 /* can't boost CS override and call */
135 return (opcode != 0x2e && opcode != 0x9a);
140 * returns non-zero if opcode modifies the interrupt flag.
142 static int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
147 case 0xcf: /* iret/iretd */
148 case 0x9d: /* popf/popfd */
154 int __kprobes arch_prepare_kprobe(struct kprobe *p)
156 /* insn: must be on special executable page on i386. */
157 p->ainsn.insn = get_insn_slot();
161 memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
162 p->opcode = *p->addr;
163 if (can_boost(p->addr)) {
164 p->ainsn.boostable = 0;
166 p->ainsn.boostable = -1;
171 void __kprobes arch_arm_kprobe(struct kprobe *p)
173 *p->addr = BREAKPOINT_INSTRUCTION;
174 flush_icache_range((unsigned long) p->addr,
175 (unsigned long) p->addr + sizeof(kprobe_opcode_t));
178 void __kprobes arch_disarm_kprobe(struct kprobe *p)
180 *p->addr = p->opcode;
181 flush_icache_range((unsigned long) p->addr,
182 (unsigned long) p->addr + sizeof(kprobe_opcode_t));
185 void __kprobes arch_remove_kprobe(struct kprobe *p)
187 mutex_lock(&kprobe_mutex);
188 free_insn_slot(p->ainsn.insn);
189 mutex_unlock(&kprobe_mutex);
192 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
194 kcb->prev_kprobe.kp = kprobe_running();
195 kcb->prev_kprobe.status = kcb->kprobe_status;
196 kcb->prev_kprobe.old_eflags = kcb->kprobe_old_eflags;
197 kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
200 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
202 __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
203 kcb->kprobe_status = kcb->prev_kprobe.status;
204 kcb->kprobe_old_eflags = kcb->prev_kprobe.old_eflags;
205 kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
208 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
209 struct kprobe_ctlblk *kcb)
211 __get_cpu_var(current_kprobe) = p;
212 kcb->kprobe_saved_eflags = kcb->kprobe_old_eflags
213 = (regs->eflags & (TF_MASK | IF_MASK));
214 if (is_IF_modifier(p->opcode))
215 kcb->kprobe_saved_eflags &= ~IF_MASK;
218 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
220 regs->eflags |= TF_MASK;
221 regs->eflags &= ~IF_MASK;
222 /*single step inline if the instruction is an int3*/
223 if (p->opcode == BREAKPOINT_INSTRUCTION)
224 regs->eip = (unsigned long)p->addr;
226 regs->eip = (unsigned long)p->ainsn.insn;
229 /* Called with kretprobe_lock held */
230 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
231 struct pt_regs *regs)
233 unsigned long *sara = (unsigned long *)®s->esp;
234 struct kretprobe_instance *ri;
236 if ((ri = get_free_rp_inst(rp)) != NULL) {
239 ri->ret_addr = (kprobe_opcode_t *) *sara;
241 /* Replace the return addr with trampoline addr */
242 *sara = (unsigned long) &kretprobe_trampoline;
251 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
252 * remain disabled thorough out this function.
254 static int __kprobes kprobe_handler(struct pt_regs *regs)
258 kprobe_opcode_t *addr;
259 struct kprobe_ctlblk *kcb;
260 #ifdef CONFIG_PREEMPT
261 unsigned pre_preempt_count = preempt_count();
262 #endif /* CONFIG_PREEMPT */
264 addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
267 * We don't want to be preempted for the entire
268 * duration of kprobe processing
271 kcb = get_kprobe_ctlblk();
273 /* Check we're not actually recursing */
274 if (kprobe_running()) {
275 p = get_kprobe(addr);
277 if (kcb->kprobe_status == KPROBE_HIT_SS &&
278 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
279 regs->eflags &= ~TF_MASK;
280 regs->eflags |= kcb->kprobe_saved_eflags;
283 /* We have reentered the kprobe_handler(), since
284 * another probe was hit while within the handler.
285 * We here save the original kprobes variables and
286 * just single step on the instruction of the new probe
287 * without calling any user handlers.
289 save_previous_kprobe(kcb);
290 set_current_kprobe(p, regs, kcb);
291 kprobes_inc_nmissed_count(p);
292 prepare_singlestep(p, regs);
293 kcb->kprobe_status = KPROBE_REENTER;
296 if (*addr != BREAKPOINT_INSTRUCTION) {
297 /* The breakpoint instruction was removed by
298 * another cpu right after we hit, no further
299 * handling of this interrupt is appropriate
301 regs->eip -= sizeof(kprobe_opcode_t);
305 p = __get_cpu_var(current_kprobe);
306 if (p->break_handler && p->break_handler(p, regs)) {
313 p = get_kprobe(addr);
315 if (*addr != BREAKPOINT_INSTRUCTION) {
317 * The breakpoint instruction was removed right
318 * after we hit it. Another cpu has removed
319 * either a probepoint or a debugger breakpoint
320 * at this address. In either case, no further
321 * handling of this interrupt is appropriate.
322 * Back up over the (now missing) int3 and run
323 * the original instruction.
325 regs->eip -= sizeof(kprobe_opcode_t);
328 /* Not one of ours: let kernel handle it */
332 set_current_kprobe(p, regs, kcb);
333 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
335 if (p->pre_handler && p->pre_handler(p, regs))
336 /* handler has already set things up, so skip ss setup */
339 if (p->ainsn.boostable == 1 &&
340 #ifdef CONFIG_PREEMPT
341 !(pre_preempt_count) && /*
342 * This enables booster when the direct
343 * execution path aren't preempted.
345 #endif /* CONFIG_PREEMPT */
346 !p->post_handler && !p->break_handler ) {
347 /* Boost up -- we can execute copied instructions directly */
348 reset_current_kprobe();
349 regs->eip = (unsigned long)p->ainsn.insn;
350 preempt_enable_no_resched();
355 prepare_singlestep(p, regs);
356 kcb->kprobe_status = KPROBE_HIT_SS;
360 preempt_enable_no_resched();
365 * For function-return probes, init_kprobes() establishes a probepoint
366 * here. When a retprobed function returns, this probe is hit and
367 * trampoline_probe_handler() runs, calling the kretprobe's handler.
369 void __kprobes kretprobe_trampoline_holder(void)
371 asm volatile ( ".global kretprobe_trampoline\n"
372 "kretprobe_trampoline: \n"
374 /* skip cs, eip, orig_eax, es, ds */
384 " call trampoline_handler\n"
385 /* move eflags to cs */
386 " movl 48(%esp), %edx\n"
387 " movl %edx, 44(%esp)\n"
388 /* save true return address on eflags */
389 " movl %eax, 48(%esp)\n"
397 /* skip eip, orig_eax, es, ds */
404 * Called from kretprobe_trampoline
406 fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
408 struct kretprobe_instance *ri = NULL;
409 struct hlist_head *head;
410 struct hlist_node *node, *tmp;
411 unsigned long flags, orig_ret_address = 0;
412 unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
414 spin_lock_irqsave(&kretprobe_lock, flags);
415 head = kretprobe_inst_table_head(current);
418 * It is possible to have multiple instances associated with a given
419 * task either because an multiple functions in the call path
420 * have a return probe installed on them, and/or more then one return
421 * return probe was registered for a target function.
423 * We can handle this because:
424 * - instances are always inserted at the head of the list
425 * - when multiple return probes are registered for the same
426 * function, the first instance's ret_addr will point to the
427 * real return address, and all the rest will point to
428 * kretprobe_trampoline
430 hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
431 if (ri->task != current)
432 /* another task is sharing our hash bucket */
435 if (ri->rp && ri->rp->handler){
436 __get_cpu_var(current_kprobe) = &ri->rp->kp;
437 ri->rp->handler(ri, regs);
438 __get_cpu_var(current_kprobe) = NULL;
441 orig_ret_address = (unsigned long)ri->ret_addr;
444 if (orig_ret_address != trampoline_address)
446 * This is the real return address. Any other
447 * instances associated with this task are for
448 * other calls deeper on the call stack
453 BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
455 spin_unlock_irqrestore(&kretprobe_lock, flags);
457 return (void*)orig_ret_address;
461 * Called after single-stepping. p->addr is the address of the
462 * instruction whose first byte has been replaced by the "int 3"
463 * instruction. To avoid the SMP problems that can occur when we
464 * temporarily put back the original opcode to single-step, we
465 * single-stepped a copy of the instruction. The address of this
466 * copy is p->ainsn.insn.
468 * This function prepares to return from the post-single-step
469 * interrupt. We have to fix up the stack as follows:
471 * 0) Except in the case of absolute or indirect jump or call instructions,
472 * the new eip is relative to the copied instruction. We need to make
473 * it relative to the original instruction.
475 * 1) If the single-stepped instruction was pushfl, then the TF and IF
476 * flags are set in the just-pushed eflags, and may need to be cleared.
478 * 2) If the single-stepped instruction was a call, the return address
479 * that is atop the stack is the address following the copied instruction.
480 * We need to make it the address following the original instruction.
482 * This function also checks instruction size for preparing direct execution.
484 static void __kprobes resume_execution(struct kprobe *p,
485 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
487 unsigned long *tos = (unsigned long *)®s->esp;
488 unsigned long copy_eip = (unsigned long)p->ainsn.insn;
489 unsigned long orig_eip = (unsigned long)p->addr;
491 regs->eflags &= ~TF_MASK;
492 switch (p->ainsn.insn[0]) {
493 case 0x9c: /* pushfl */
494 *tos &= ~(TF_MASK | IF_MASK);
495 *tos |= kcb->kprobe_old_eflags;
497 case 0xc2: /* iret/ret/lret */
502 case 0xea: /* jmp absolute -- eip is correct */
503 /* eip is already adjusted, no more changes required */
504 p->ainsn.boostable = 1;
506 case 0xe8: /* call relative - Fix return addr */
507 *tos = orig_eip + (*tos - copy_eip);
509 case 0x9a: /* call absolute -- same as call absolute, indirect */
510 *tos = orig_eip + (*tos - copy_eip);
513 if ((p->ainsn.insn[1] & 0x30) == 0x10) {
515 * call absolute, indirect
516 * Fix return addr; eip is correct.
517 * But this is not boostable
519 *tos = orig_eip + (*tos - copy_eip);
521 } else if (((p->ainsn.insn[1] & 0x31) == 0x20) || /* jmp near, absolute indirect */
522 ((p->ainsn.insn[1] & 0x31) == 0x21)) { /* jmp far, absolute indirect */
523 /* eip is correct. And this is boostable */
524 p->ainsn.boostable = 1;
531 if (p->ainsn.boostable == 0) {
532 if ((regs->eip > copy_eip) &&
533 (regs->eip - copy_eip) + 5 < MAX_INSN_SIZE) {
535 * These instructions can be executed directly if it
536 * jumps back to correct address.
538 set_jmp_op((void *)regs->eip,
539 (void *)orig_eip + (regs->eip - copy_eip));
540 p->ainsn.boostable = 1;
542 p->ainsn.boostable = -1;
546 regs->eip = orig_eip + (regs->eip - copy_eip);
553 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
554 * remain disabled thoroughout this function.
556 static int __kprobes post_kprobe_handler(struct pt_regs *regs)
558 struct kprobe *cur = kprobe_running();
559 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
564 if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
565 kcb->kprobe_status = KPROBE_HIT_SSDONE;
566 cur->post_handler(cur, regs, 0);
569 resume_execution(cur, regs, kcb);
570 regs->eflags |= kcb->kprobe_saved_eflags;
572 /*Restore back the original saved kprobes variables and continue. */
573 if (kcb->kprobe_status == KPROBE_REENTER) {
574 restore_previous_kprobe(kcb);
577 reset_current_kprobe();
579 preempt_enable_no_resched();
582 * if somebody else is singlestepping across a probe point, eflags
583 * will have TF set, in which case, continue the remaining processing
584 * of do_debug, as if this is not a probe hit.
586 if (regs->eflags & TF_MASK)
592 static int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
594 struct kprobe *cur = kprobe_running();
595 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
597 switch(kcb->kprobe_status) {
601 * We are here because the instruction being single
602 * stepped caused a page fault. We reset the current
603 * kprobe and the eip points back to the probe address
604 * and allow the page fault handler to continue as a
607 regs->eip = (unsigned long)cur->addr;
608 regs->eflags |= kcb->kprobe_old_eflags;
609 if (kcb->kprobe_status == KPROBE_REENTER)
610 restore_previous_kprobe(kcb);
612 reset_current_kprobe();
613 preempt_enable_no_resched();
615 case KPROBE_HIT_ACTIVE:
616 case KPROBE_HIT_SSDONE:
618 * We increment the nmissed count for accounting,
619 * we can also use npre/npostfault count for accouting
620 * these specific fault cases.
622 kprobes_inc_nmissed_count(cur);
625 * We come here because instructions in the pre/post
626 * handler caused the page_fault, this could happen
627 * if handler tries to access user space by
628 * copy_from_user(), get_user() etc. Let the
629 * user-specified handler try to fix it first.
631 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
635 * In case the user-specified fault handler returned
636 * zero, try to fix up.
638 if (fixup_exception(regs))
642 * fixup_exception() could not handle it,
643 * Let do_page_fault() fix it.
653 * Wrapper routine to for handling exceptions.
655 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
656 unsigned long val, void *data)
658 struct die_args *args = (struct die_args *)data;
659 int ret = NOTIFY_DONE;
661 if (args->regs && user_mode_vm(args->regs))
666 if (kprobe_handler(args->regs))
670 if (post_kprobe_handler(args->regs))
675 /* kprobe_running() needs smp_processor_id() */
677 if (kprobe_running() &&
678 kprobe_fault_handler(args->regs, args->trapnr))
688 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
690 struct jprobe *jp = container_of(p, struct jprobe, kp);
692 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
694 kcb->jprobe_saved_regs = *regs;
695 kcb->jprobe_saved_esp = ®s->esp;
696 addr = (unsigned long)(kcb->jprobe_saved_esp);
699 * TBD: As Linus pointed out, gcc assumes that the callee
700 * owns the argument space and could overwrite it, e.g.
701 * tailcall optimization. So, to be absolutely safe
702 * we also save and restore enough stack bytes to cover
705 memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
706 MIN_STACK_SIZE(addr));
707 regs->eflags &= ~IF_MASK;
708 regs->eip = (unsigned long)(jp->entry);
712 void __kprobes jprobe_return(void)
714 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
716 asm volatile (" xchgl %%ebx,%%esp \n"
718 " .globl jprobe_return_end \n"
719 " jprobe_return_end: \n"
721 (kcb->jprobe_saved_esp):"memory");
724 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
726 struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
727 u8 *addr = (u8 *) (regs->eip - 1);
728 unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
729 struct jprobe *jp = container_of(p, struct jprobe, kp);
731 if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
732 if (®s->esp != kcb->jprobe_saved_esp) {
733 struct pt_regs *saved_regs =
734 container_of(kcb->jprobe_saved_esp,
735 struct pt_regs, esp);
736 printk("current esp %p does not match saved esp %p\n",
737 ®s->esp, kcb->jprobe_saved_esp);
738 printk("Saved registers for jprobe %p\n", jp);
739 show_registers(saved_regs);
740 printk("Current registers\n");
741 show_registers(regs);
744 *regs = kcb->jprobe_saved_regs;
745 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
746 MIN_STACK_SIZE(stack_addr));
747 preempt_enable_no_resched();
753 int __init arch_init_kprobes(void)