2 * Kernel Probes (KProbes)
3 * arch/x86_64/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 * 2004-Oct Jim Keniston <kenistoj@us.ibm.com> and Prasanna S Panchamukhi
27 * <prasanna@in.ibm.com> adapted for x86_64
28 * 2005-Mar Roland McGrath <roland@redhat.com>
29 * Fixed to handle %rip-relative addressing mode correctly.
30 * 2005-May Rusty Lynch <rusty.lynch@intel.com>
31 * Added function return probes functionality
34 #include <linux/config.h>
35 #include <linux/kprobes.h>
36 #include <linux/ptrace.h>
37 #include <linux/spinlock.h>
38 #include <linux/string.h>
39 #include <linux/slab.h>
40 #include <linux/preempt.h>
42 #include <asm/cacheflush.h>
43 #include <asm/pgtable.h>
44 #include <asm/kdebug.h>
46 static DECLARE_MUTEX(kprobe_mutex);
48 static struct kprobe *current_kprobe;
49 static unsigned long kprobe_status, kprobe_old_rflags, kprobe_saved_rflags;
50 static struct kprobe *kprobe_prev;
51 static unsigned long kprobe_status_prev, kprobe_old_rflags_prev, kprobe_saved_rflags_prev;
52 static struct pt_regs jprobe_saved_regs;
53 static long *jprobe_saved_rsp;
54 void jprobe_return_end(void);
56 /* copy of the kernel stack at the probe fire time */
57 static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE];
60 * returns non-zero if opcode modifies the interrupt flag.
62 static inline int is_IF_modifier(kprobe_opcode_t *insn)
67 case 0xcf: /* iret/iretd */
68 case 0x9d: /* popf/popfd */
72 if (*insn >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
77 int arch_prepare_kprobe(struct kprobe *p)
79 /* insn: must be on special executable page on x86_64. */
81 p->ainsn.insn = get_insn_slot();
90 * Determine if the instruction uses the %rip-relative addressing mode.
91 * If it does, return the address of the 32-bit displacement word.
92 * If not, return null.
94 static inline s32 *is_riprel(u8 *insn)
96 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf) \
97 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
98 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
99 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
100 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
102 static const u64 onebyte_has_modrm[256 / 64] = {
103 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
104 /* ------------------------------- */
105 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
106 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
107 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
108 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
109 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
110 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
111 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
112 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
113 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
114 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
115 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
116 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
117 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
118 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
119 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
120 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1) /* f0 */
121 /* ------------------------------- */
122 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
124 static const u64 twobyte_has_modrm[256 / 64] = {
125 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
126 /* ------------------------------- */
127 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
128 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
129 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
130 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
131 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
132 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
133 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
134 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
135 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
136 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
137 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
138 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
139 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
140 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
141 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
142 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0) /* ff */
143 /* ------------------------------- */
144 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
149 /* Skip legacy instruction prefixes. */
169 /* Skip REX instruction prefix. */
170 if ((*insn & 0xf0) == 0x40)
173 if (*insn == 0x0f) { /* Two-byte opcode. */
175 need_modrm = test_bit(*insn, twobyte_has_modrm);
176 } else { /* One-byte opcode. */
177 need_modrm = test_bit(*insn, onebyte_has_modrm);
182 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
183 /* Displacement follows ModRM byte. */
184 return (s32 *) ++insn;
188 /* No %rip-relative addressing mode here. */
192 void arch_copy_kprobe(struct kprobe *p)
195 memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
196 ripdisp = is_riprel(p->ainsn.insn);
199 * The copied instruction uses the %rip-relative
200 * addressing mode. Adjust the displacement for the
201 * difference between the original location of this
202 * instruction and the location of the copy that will
203 * actually be run. The tricky bit here is making sure
204 * that the sign extension happens correctly in this
205 * calculation, since we need a signed 32-bit result to
206 * be sign-extended to 64 bits when it's added to the
207 * %rip value and yield the same 64-bit result that the
208 * sign-extension of the original signed 32-bit
209 * displacement would have given.
211 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
212 BUG_ON((s64) (s32) disp != disp); /* Sanity check. */
215 p->opcode = *p->addr;
218 void arch_arm_kprobe(struct kprobe *p)
220 *p->addr = BREAKPOINT_INSTRUCTION;
221 flush_icache_range((unsigned long) p->addr,
222 (unsigned long) p->addr + sizeof(kprobe_opcode_t));
225 void arch_disarm_kprobe(struct kprobe *p)
227 *p->addr = p->opcode;
228 flush_icache_range((unsigned long) p->addr,
229 (unsigned long) p->addr + sizeof(kprobe_opcode_t));
232 void arch_remove_kprobe(struct kprobe *p)
235 free_insn_slot(p->ainsn.insn);
239 static inline void save_previous_kprobe(void)
241 kprobe_prev = current_kprobe;
242 kprobe_status_prev = kprobe_status;
243 kprobe_old_rflags_prev = kprobe_old_rflags;
244 kprobe_saved_rflags_prev = kprobe_saved_rflags;
247 static inline void restore_previous_kprobe(void)
249 current_kprobe = kprobe_prev;
250 kprobe_status = kprobe_status_prev;
251 kprobe_old_rflags = kprobe_old_rflags_prev;
252 kprobe_saved_rflags = kprobe_saved_rflags_prev;
255 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs)
258 kprobe_saved_rflags = kprobe_old_rflags
259 = (regs->eflags & (TF_MASK | IF_MASK));
260 if (is_IF_modifier(p->ainsn.insn))
261 kprobe_saved_rflags &= ~IF_MASK;
264 static void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
266 regs->eflags |= TF_MASK;
267 regs->eflags &= ~IF_MASK;
268 /*single step inline if the instruction is an int3*/
269 if (p->opcode == BREAKPOINT_INSTRUCTION)
270 regs->rip = (unsigned long)p->addr;
272 regs->rip = (unsigned long)p->ainsn.insn;
275 struct task_struct *arch_get_kprobe_task(void *ptr)
277 return ((struct thread_info *) (((unsigned long) ptr) &
278 (~(THREAD_SIZE -1))))->task;
281 void arch_prepare_kretprobe(struct kretprobe *rp, struct pt_regs *regs)
283 unsigned long *sara = (unsigned long *)regs->rsp;
284 struct kretprobe_instance *ri;
285 static void *orig_ret_addr;
288 * Save the return address when the return probe hits
289 * the first time, and use it to populate the (krprobe
290 * instance)->ret_addr for subsequent return probes at
291 * the same addrress since stack address would have
292 * the kretprobe_trampoline by then.
294 if (((void*) *sara) != kretprobe_trampoline)
295 orig_ret_addr = (void*) *sara;
297 if ((ri = get_free_rp_inst(rp)) != NULL) {
299 ri->stack_addr = sara;
300 ri->ret_addr = orig_ret_addr;
302 /* Replace the return addr with trampoline addr */
303 *sara = (unsigned long) &kretprobe_trampoline;
309 void arch_kprobe_flush_task(struct task_struct *tk)
311 struct kretprobe_instance *ri;
312 while ((ri = get_rp_inst_tsk(tk)) != NULL) {
313 *((unsigned long *)(ri->stack_addr)) =
314 (unsigned long) ri->ret_addr;
320 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
321 * remain disabled thorough out this function.
323 int kprobe_handler(struct pt_regs *regs)
327 kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
329 /* We're in an interrupt, but this is clear and BUG()-safe. */
332 /* Check we're not actually recursing */
333 if (kprobe_running()) {
334 /* We *are* holding lock here, so this is safe.
335 Disarm the probe we just hit, and ignore it. */
336 p = get_kprobe(addr);
338 if (kprobe_status == KPROBE_HIT_SS) {
339 regs->eflags &= ~TF_MASK;
340 regs->eflags |= kprobe_saved_rflags;
343 } else if (kprobe_status == KPROBE_HIT_SSDONE) {
344 /* TODO: Provide re-entrancy from
345 * post_kprobes_handler() and avoid exception
346 * stack corruption while single-stepping on
347 * the instruction of the new probe.
349 arch_disarm_kprobe(p);
350 regs->rip = (unsigned long)p->addr;
353 /* We have reentered the kprobe_handler(), since
354 * another probe was hit while within the
355 * handler. We here save the original kprobe
356 * variables and just single step on instruction
357 * of the new probe without calling any user
360 save_previous_kprobe();
361 set_current_kprobe(p, regs);
363 prepare_singlestep(p, regs);
364 kprobe_status = KPROBE_REENTER;
369 if (p->break_handler && p->break_handler(p, regs)) {
373 /* If it's not ours, can't be delete race, (we hold lock). */
378 p = get_kprobe(addr);
381 if (*addr != BREAKPOINT_INSTRUCTION) {
383 * The breakpoint instruction was removed right
384 * after we hit it. Another cpu has removed
385 * either a probepoint or a debugger breakpoint
386 * at this address. In either case, no further
387 * handling of this interrupt is appropriate.
391 /* Not one of ours: let kernel handle it */
395 kprobe_status = KPROBE_HIT_ACTIVE;
396 set_current_kprobe(p, regs);
398 if (p->pre_handler && p->pre_handler(p, regs))
399 /* handler has already set things up, so skip ss setup */
403 prepare_singlestep(p, regs);
404 kprobe_status = KPROBE_HIT_SS;
408 preempt_enable_no_resched();
413 * For function-return probes, init_kprobes() establishes a probepoint
414 * here. When a retprobed function returns, this probe is hit and
415 * trampoline_probe_handler() runs, calling the kretprobe's handler.
417 void kretprobe_trampoline_holder(void)
419 asm volatile ( ".global kretprobe_trampoline\n"
420 "kretprobe_trampoline: \n"
425 * Called when we hit the probe point at kretprobe_trampoline
427 int trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
429 struct task_struct *tsk;
430 struct kretprobe_instance *ri;
431 struct hlist_head *head;
432 struct hlist_node *node;
433 unsigned long *sara = (unsigned long *)regs->rsp - 1;
435 tsk = arch_get_kprobe_task(sara);
436 head = kretprobe_inst_table_head(tsk);
438 hlist_for_each_entry(ri, node, head, hlist) {
439 if (ri->stack_addr == sara && ri->rp) {
441 ri->rp->handler(ri, regs);
447 void trampoline_post_handler(struct kprobe *p, struct pt_regs *regs,
450 struct kretprobe_instance *ri;
451 /* RA already popped */
452 unsigned long *sara = ((unsigned long *)regs->rsp) - 1;
454 while ((ri = get_rp_inst(sara))) {
455 regs->rip = (unsigned long)ri->ret_addr;
458 regs->eflags &= ~TF_MASK;
462 * Called after single-stepping. p->addr is the address of the
463 * instruction whose first byte has been replaced by the "int 3"
464 * instruction. To avoid the SMP problems that can occur when we
465 * temporarily put back the original opcode to single-step, we
466 * single-stepped a copy of the instruction. The address of this
467 * copy is p->ainsn.insn.
469 * This function prepares to return from the post-single-step
470 * interrupt. We have to fix up the stack as follows:
472 * 0) Except in the case of absolute or indirect jump or call instructions,
473 * the new rip is relative to the copied instruction. We need to make
474 * it relative to the original instruction.
476 * 1) If the single-stepped instruction was pushfl, then the TF and IF
477 * flags are set in the just-pushed eflags, and may need to be cleared.
479 * 2) If the single-stepped instruction was a call, the return address
480 * that is atop the stack is the address following the copied instruction.
481 * We need to make it the address following the original instruction.
483 static void resume_execution(struct kprobe *p, struct pt_regs *regs)
485 unsigned long *tos = (unsigned long *)regs->rsp;
486 unsigned long next_rip = 0;
487 unsigned long copy_rip = (unsigned long)p->ainsn.insn;
488 unsigned long orig_rip = (unsigned long)p->addr;
489 kprobe_opcode_t *insn = p->ainsn.insn;
491 /*skip the REX prefix*/
492 if (*insn >= 0x40 && *insn <= 0x4f)
496 case 0x9c: /* pushfl */
497 *tos &= ~(TF_MASK | IF_MASK);
498 *tos |= kprobe_old_rflags;
500 case 0xc3: /* ret/lret */
504 regs->eflags &= ~TF_MASK;
505 /* rip is already adjusted, no more changes required*/
507 case 0xe8: /* call relative - Fix return addr */
508 *tos = orig_rip + (*tos - copy_rip);
511 if ((*insn & 0x30) == 0x10) {
512 /* call absolute, indirect */
513 /* Fix return addr; rip is correct. */
514 next_rip = regs->rip;
515 *tos = orig_rip + (*tos - copy_rip);
516 } else if (((*insn & 0x31) == 0x20) || /* jmp near, absolute indirect */
517 ((*insn & 0x31) == 0x21)) { /* jmp far, absolute indirect */
518 /* rip is correct. */
519 next_rip = regs->rip;
522 case 0xea: /* jmp absolute -- rip is correct */
523 next_rip = regs->rip;
529 regs->eflags &= ~TF_MASK;
531 regs->rip = next_rip;
533 regs->rip = orig_rip + (regs->rip - copy_rip);
538 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
539 * remain disabled thoroughout this function. And we hold kprobe lock.
541 int post_kprobe_handler(struct pt_regs *regs)
543 if (!kprobe_running())
546 if ((kprobe_status != KPROBE_REENTER) && current_kprobe->post_handler) {
547 kprobe_status = KPROBE_HIT_SSDONE;
548 current_kprobe->post_handler(current_kprobe, regs, 0);
551 if (current_kprobe->post_handler != trampoline_post_handler)
552 resume_execution(current_kprobe, regs);
553 regs->eflags |= kprobe_saved_rflags;
555 /* Restore the original saved kprobes variables and continue. */
556 if (kprobe_status == KPROBE_REENTER) {
557 restore_previous_kprobe();
563 preempt_enable_no_resched();
566 * if somebody else is singlestepping across a probe point, eflags
567 * will have TF set, in which case, continue the remaining processing
568 * of do_debug, as if this is not a probe hit.
570 if (regs->eflags & TF_MASK)
576 /* Interrupts disabled, kprobe_lock held. */
577 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
579 if (current_kprobe->fault_handler
580 && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
583 if (kprobe_status & KPROBE_HIT_SS) {
584 resume_execution(current_kprobe, regs);
585 regs->eflags |= kprobe_old_rflags;
588 preempt_enable_no_resched();
594 * Wrapper routine for handling exceptions.
596 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
599 struct die_args *args = (struct die_args *)data;
602 if (kprobe_handler(args->regs))
606 if (post_kprobe_handler(args->regs))
610 if (kprobe_running() &&
611 kprobe_fault_handler(args->regs, args->trapnr))
615 if (kprobe_running() &&
616 kprobe_fault_handler(args->regs, args->trapnr))
625 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
627 struct jprobe *jp = container_of(p, struct jprobe, kp);
630 jprobe_saved_regs = *regs;
631 jprobe_saved_rsp = (long *) regs->rsp;
632 addr = (unsigned long)jprobe_saved_rsp;
634 * As Linus pointed out, gcc assumes that the callee
635 * owns the argument space and could overwrite it, e.g.
636 * tailcall optimization. So, to be absolutely safe
637 * we also save and restore enough stack bytes to cover
640 memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
641 regs->eflags &= ~IF_MASK;
642 regs->rip = (unsigned long)(jp->entry);
646 void jprobe_return(void)
648 preempt_enable_no_resched();
649 asm volatile (" xchg %%rbx,%%rsp \n"
651 " .globl jprobe_return_end \n"
652 " jprobe_return_end: \n"
654 (jprobe_saved_rsp):"memory");
657 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
659 u8 *addr = (u8 *) (regs->rip - 1);
660 unsigned long stack_addr = (unsigned long)jprobe_saved_rsp;
661 struct jprobe *jp = container_of(p, struct jprobe, kp);
663 if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
664 if ((long *)regs->rsp != jprobe_saved_rsp) {
665 struct pt_regs *saved_regs =
666 container_of(jprobe_saved_rsp, struct pt_regs, rsp);
667 printk("current rsp %p does not match saved rsp %p\n",
668 (long *)regs->rsp, jprobe_saved_rsp);
669 printk("Saved registers for jprobe %p\n", jp);
670 show_registers(saved_regs);
671 printk("Current registers\n");
672 show_registers(regs);
675 *regs = jprobe_saved_regs;
676 memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
677 MIN_STACK_SIZE(stack_addr));