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.
32 #include <linux/config.h>
33 #include <linux/kprobes.h>
34 #include <linux/ptrace.h>
35 #include <linux/spinlock.h>
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/preempt.h>
39 #include <linux/moduleloader.h>
41 #include <asm/pgtable.h>
42 #include <asm/kdebug.h>
44 static DECLARE_MUTEX(kprobe_mutex);
46 /* kprobe_status settings */
47 #define KPROBE_HIT_ACTIVE 0x00000001
48 #define KPROBE_HIT_SS 0x00000002
50 static struct kprobe *current_kprobe;
51 static unsigned long kprobe_status, kprobe_old_rflags, kprobe_saved_rflags;
52 static struct pt_regs jprobe_saved_regs;
53 static long *jprobe_saved_rsp;
54 static kprobe_opcode_t *get_insn_slot(void);
55 static void free_insn_slot(kprobe_opcode_t *slot);
56 void jprobe_return_end(void);
58 /* copy of the kernel stack at the probe fire time */
59 static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE];
62 * returns non-zero if opcode modifies the interrupt flag.
64 static inline int is_IF_modifier(kprobe_opcode_t *insn)
69 case 0xcf: /* iret/iretd */
70 case 0x9d: /* popf/popfd */
74 if (*insn >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
79 int arch_prepare_kprobe(struct kprobe *p)
81 /* insn: must be on special executable page on x86_64. */
83 p->ainsn.insn = get_insn_slot();
92 * Determine if the instruction uses the %rip-relative addressing mode.
93 * If it does, return the address of the 32-bit displacement word.
94 * If not, return null.
96 static inline s32 *is_riprel(u8 *insn)
98 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf) \
99 (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) | \
100 (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) | \
101 (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) | \
102 (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf)) \
104 static const u64 onebyte_has_modrm[256 / 64] = {
105 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
106 /* ------------------------------- */
107 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
108 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
109 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
110 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
111 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
112 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
113 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
114 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
115 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
116 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
117 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
118 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
119 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
120 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
121 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
122 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1) /* f0 */
123 /* ------------------------------- */
124 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
126 static const u64 twobyte_has_modrm[256 / 64] = {
127 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
128 /* ------------------------------- */
129 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
130 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
131 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
132 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
133 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
134 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
135 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
136 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
137 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
138 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
139 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
140 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
141 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
142 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
143 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
144 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0) /* ff */
145 /* ------------------------------- */
146 /* 0 1 2 3 4 5 6 7 8 9 a b c d e f */
151 /* Skip legacy instruction prefixes. */
171 /* Skip REX instruction prefix. */
172 if ((*insn & 0xf0) == 0x40)
175 if (*insn == 0x0f) { /* Two-byte opcode. */
177 need_modrm = test_bit(*insn, twobyte_has_modrm);
178 } else { /* One-byte opcode. */
179 need_modrm = test_bit(*insn, onebyte_has_modrm);
184 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
185 /* Displacement follows ModRM byte. */
186 return (s32 *) ++insn;
190 /* No %rip-relative addressing mode here. */
194 void arch_copy_kprobe(struct kprobe *p)
197 memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
198 ripdisp = is_riprel(p->ainsn.insn);
201 * The copied instruction uses the %rip-relative
202 * addressing mode. Adjust the displacement for the
203 * difference between the original location of this
204 * instruction and the location of the copy that will
205 * actually be run. The tricky bit here is making sure
206 * that the sign extension happens correctly in this
207 * calculation, since we need a signed 32-bit result to
208 * be sign-extended to 64 bits when it's added to the
209 * %rip value and yield the same 64-bit result that the
210 * sign-extension of the original signed 32-bit
211 * displacement would have given.
213 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
214 BUG_ON((s64) (s32) disp != disp); /* Sanity check. */
219 void arch_remove_kprobe(struct kprobe *p)
222 free_insn_slot(p->ainsn.insn);
226 static inline void disarm_kprobe(struct kprobe *p, struct pt_regs *regs)
228 *p->addr = p->opcode;
229 regs->rip = (unsigned long)p->addr;
232 static void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
234 regs->eflags |= TF_MASK;
235 regs->eflags &= ~IF_MASK;
236 /*single step inline if the instruction is an int3*/
237 if (p->opcode == BREAKPOINT_INSTRUCTION)
238 regs->rip = (unsigned long)p->addr;
240 regs->rip = (unsigned long)p->ainsn.insn;
244 * Interrupts are disabled on entry as trap3 is an interrupt gate and they
245 * remain disabled thorough out this function.
247 int kprobe_handler(struct pt_regs *regs)
251 kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
253 /* We're in an interrupt, but this is clear and BUG()-safe. */
256 /* Check we're not actually recursing */
257 if (kprobe_running()) {
258 /* We *are* holding lock here, so this is safe.
259 Disarm the probe we just hit, and ignore it. */
260 p = get_kprobe(addr);
262 if (kprobe_status == KPROBE_HIT_SS) {
263 regs->eflags &= ~TF_MASK;
264 regs->eflags |= kprobe_saved_rflags;
268 disarm_kprobe(p, regs);
272 if (p->break_handler && p->break_handler(p, regs)) {
276 /* If it's not ours, can't be delete race, (we hold lock). */
281 p = get_kprobe(addr);
284 if (*addr != BREAKPOINT_INSTRUCTION) {
286 * The breakpoint instruction was removed right
287 * after we hit it. Another cpu has removed
288 * either a probepoint or a debugger breakpoint
289 * at this address. In either case, no further
290 * handling of this interrupt is appropriate.
294 /* Not one of ours: let kernel handle it */
298 kprobe_status = KPROBE_HIT_ACTIVE;
300 kprobe_saved_rflags = kprobe_old_rflags
301 = (regs->eflags & (TF_MASK | IF_MASK));
302 if (is_IF_modifier(p->ainsn.insn))
303 kprobe_saved_rflags &= ~IF_MASK;
305 if (p->pre_handler && p->pre_handler(p, regs))
306 /* handler has already set things up, so skip ss setup */
310 prepare_singlestep(p, regs);
311 kprobe_status = KPROBE_HIT_SS;
315 preempt_enable_no_resched();
320 * Called after single-stepping. p->addr is the address of the
321 * instruction whose first byte has been replaced by the "int 3"
322 * instruction. To avoid the SMP problems that can occur when we
323 * temporarily put back the original opcode to single-step, we
324 * single-stepped a copy of the instruction. The address of this
325 * copy is p->ainsn.insn.
327 * This function prepares to return from the post-single-step
328 * interrupt. We have to fix up the stack as follows:
330 * 0) Except in the case of absolute or indirect jump or call instructions,
331 * the new rip is relative to the copied instruction. We need to make
332 * it relative to the original instruction.
334 * 1) If the single-stepped instruction was pushfl, then the TF and IF
335 * flags are set in the just-pushed eflags, and may need to be cleared.
337 * 2) If the single-stepped instruction was a call, the return address
338 * that is atop the stack is the address following the copied instruction.
339 * We need to make it the address following the original instruction.
341 static void resume_execution(struct kprobe *p, struct pt_regs *regs)
343 unsigned long *tos = (unsigned long *)regs->rsp;
344 unsigned long next_rip = 0;
345 unsigned long copy_rip = (unsigned long)p->ainsn.insn;
346 unsigned long orig_rip = (unsigned long)p->addr;
347 kprobe_opcode_t *insn = p->ainsn.insn;
349 /*skip the REX prefix*/
350 if (*insn >= 0x40 && *insn <= 0x4f)
354 case 0x9c: /* pushfl */
355 *tos &= ~(TF_MASK | IF_MASK);
356 *tos |= kprobe_old_rflags;
358 case 0xe8: /* call relative - Fix return addr */
359 *tos = orig_rip + (*tos - copy_rip);
362 if ((*insn & 0x30) == 0x10) {
363 /* call absolute, indirect */
364 /* Fix return addr; rip is correct. */
365 next_rip = regs->rip;
366 *tos = orig_rip + (*tos - copy_rip);
367 } else if (((*insn & 0x31) == 0x20) || /* jmp near, absolute indirect */
368 ((*insn & 0x31) == 0x21)) { /* jmp far, absolute indirect */
369 /* rip is correct. */
370 next_rip = regs->rip;
373 case 0xea: /* jmp absolute -- rip is correct */
374 next_rip = regs->rip;
380 regs->eflags &= ~TF_MASK;
382 regs->rip = next_rip;
384 regs->rip = orig_rip + (regs->rip - copy_rip);
389 * Interrupts are disabled on entry as trap1 is an interrupt gate and they
390 * remain disabled thoroughout this function. And we hold kprobe lock.
392 int post_kprobe_handler(struct pt_regs *regs)
394 if (!kprobe_running())
397 if (current_kprobe->post_handler)
398 current_kprobe->post_handler(current_kprobe, regs, 0);
400 resume_execution(current_kprobe, regs);
401 regs->eflags |= kprobe_saved_rflags;
404 preempt_enable_no_resched();
407 * if somebody else is singlestepping across a probe point, eflags
408 * will have TF set, in which case, continue the remaining processing
409 * of do_debug, as if this is not a probe hit.
411 if (regs->eflags & TF_MASK)
417 /* Interrupts disabled, kprobe_lock held. */
418 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
420 if (current_kprobe->fault_handler
421 && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
424 if (kprobe_status & KPROBE_HIT_SS) {
425 resume_execution(current_kprobe, regs);
426 regs->eflags |= kprobe_old_rflags;
429 preempt_enable_no_resched();
435 * Wrapper routine for handling exceptions.
437 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
440 struct die_args *args = (struct die_args *)data;
443 if (kprobe_handler(args->regs))
447 if (post_kprobe_handler(args->regs))
451 if (kprobe_running() &&
452 kprobe_fault_handler(args->regs, args->trapnr))
456 if (kprobe_running() &&
457 kprobe_fault_handler(args->regs, args->trapnr))
466 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
468 struct jprobe *jp = container_of(p, struct jprobe, kp);
471 jprobe_saved_regs = *regs;
472 jprobe_saved_rsp = (long *) regs->rsp;
473 addr = (unsigned long)jprobe_saved_rsp;
475 * As Linus pointed out, gcc assumes that the callee
476 * owns the argument space and could overwrite it, e.g.
477 * tailcall optimization. So, to be absolutely safe
478 * we also save and restore enough stack bytes to cover
481 memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
482 regs->eflags &= ~IF_MASK;
483 regs->rip = (unsigned long)(jp->entry);
487 void jprobe_return(void)
489 preempt_enable_no_resched();
490 asm volatile (" xchg %%rbx,%%rsp \n"
492 " .globl jprobe_return_end \n"
493 " jprobe_return_end: \n"
495 (jprobe_saved_rsp):"memory");
498 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
500 u8 *addr = (u8 *) (regs->rip - 1);
501 unsigned long stack_addr = (unsigned long)jprobe_saved_rsp;
502 struct jprobe *jp = container_of(p, struct jprobe, kp);
504 if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
505 if ((long *)regs->rsp != jprobe_saved_rsp) {
506 struct pt_regs *saved_regs =
507 container_of(jprobe_saved_rsp, struct pt_regs, rsp);
508 printk("current rsp %p does not match saved rsp %p\n",
509 (long *)regs->rsp, jprobe_saved_rsp);
510 printk("Saved registers for jprobe %p\n", jp);
511 show_registers(saved_regs);
512 printk("Current registers\n");
513 show_registers(regs);
516 *regs = jprobe_saved_regs;
517 memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
518 MIN_STACK_SIZE(stack_addr));
525 * kprobe->ainsn.insn points to the copy of the instruction to be single-stepped.
526 * By default on x86_64, pages we get from kmalloc or vmalloc are not
527 * executable. Single-stepping an instruction on such a page yields an
528 * oops. So instead of storing the instruction copies in their respective
529 * kprobe objects, we allocate a page, map it executable, and store all the
530 * instruction copies there. (We can allocate additional pages if somebody
531 * inserts a huge number of probes.) Each page can hold up to INSNS_PER_PAGE
532 * instruction slots, each of which is MAX_INSN_SIZE*sizeof(kprobe_opcode_t)
535 #define INSNS_PER_PAGE (PAGE_SIZE/(MAX_INSN_SIZE*sizeof(kprobe_opcode_t)))
536 struct kprobe_insn_page {
537 struct hlist_node hlist;
538 kprobe_opcode_t *insns; /* page of instruction slots */
539 char slot_used[INSNS_PER_PAGE];
543 static struct hlist_head kprobe_insn_pages;
546 * get_insn_slot() - Find a slot on an executable page for an instruction.
547 * We allocate an executable page if there's no room on existing ones.
549 static kprobe_opcode_t *get_insn_slot(void)
551 struct kprobe_insn_page *kip;
552 struct hlist_node *pos;
554 hlist_for_each(pos, &kprobe_insn_pages) {
555 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
556 if (kip->nused < INSNS_PER_PAGE) {
558 for (i = 0; i < INSNS_PER_PAGE; i++) {
559 if (!kip->slot_used[i]) {
560 kip->slot_used[i] = 1;
562 return kip->insns + (i*MAX_INSN_SIZE);
565 /* Surprise! No unused slots. Fix kip->nused. */
566 kip->nused = INSNS_PER_PAGE;
570 /* All out of space. Need to allocate a new page. Use slot 0.*/
571 kip = kmalloc(sizeof(struct kprobe_insn_page), GFP_KERNEL);
577 * For the %rip-relative displacement fixups to be doable, we
578 * need our instruction copy to be within +/- 2GB of any data it
579 * might access via %rip. That is, within 2GB of where the
580 * kernel image and loaded module images reside. So we allocate
581 * a page in the module loading area.
583 kip->insns = module_alloc(PAGE_SIZE);
588 INIT_HLIST_NODE(&kip->hlist);
589 hlist_add_head(&kip->hlist, &kprobe_insn_pages);
590 memset(kip->slot_used, 0, INSNS_PER_PAGE);
591 kip->slot_used[0] = 1;
597 * free_insn_slot() - Free instruction slot obtained from get_insn_slot().
599 static void free_insn_slot(kprobe_opcode_t *slot)
601 struct kprobe_insn_page *kip;
602 struct hlist_node *pos;
604 hlist_for_each(pos, &kprobe_insn_pages) {
605 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
606 if (kip->insns <= slot
607 && slot < kip->insns+(INSNS_PER_PAGE*MAX_INSN_SIZE)) {
608 int i = (slot - kip->insns) / MAX_INSN_SIZE;
609 kip->slot_used[i] = 0;
611 if (kip->nused == 0) {
613 * Page is no longer in use. Free it unless
614 * it's the last one. We keep the last one
615 * so as not to have to set it up again the
616 * next time somebody inserts a probe.
618 hlist_del(&kip->hlist);
619 if (hlist_empty(&kprobe_insn_pages)) {
620 INIT_HLIST_NODE(&kip->hlist);
621 hlist_add_head(&kip->hlist,
624 module_free(NULL, kip->insns);