kprobes: support kretprobe blacklist
[linux-2.6.git] / arch / x86 / kernel / kprobes_64.c
1 /*
2  *  Kernel Probes (KProbes)
3  *
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.
8  *
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.
13  *
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.
17  *
18  * Copyright (C) IBM Corporation, 2002, 2004
19  *
20  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
21  *              Probes initial implementation ( includes contributions from
22  *              Rusty Russell).
23  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
24  *              interface to access function arguments.
25  * 2004-Oct     Jim Keniston <kenistoj@us.ibm.com> and Prasanna S Panchamukhi
26  *              <prasanna@in.ibm.com> adapted for x86_64
27  * 2005-Mar     Roland McGrath <roland@redhat.com>
28  *              Fixed to handle %rip-relative addressing mode correctly.
29  * 2005-May     Rusty Lynch <rusty.lynch@intel.com>
30  *              Added function return probes functionality
31  */
32
33 #include <linux/kprobes.h>
34 #include <linux/ptrace.h>
35 #include <linux/string.h>
36 #include <linux/slab.h>
37 #include <linux/preempt.h>
38 #include <linux/module.h>
39 #include <linux/kdebug.h>
40
41 #include <asm/pgtable.h>
42 #include <asm/uaccess.h>
43 #include <asm/alternative.h>
44
45 void jprobe_return_end(void);
46 static void __kprobes arch_copy_kprobe(struct kprobe *p);
47
48 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
49 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
50
51 struct kretprobe_blackpoint kretprobe_blacklist[] = {
52         {"__switch_to", }, /* This function switches only current task, but
53                               doesn't switch kernel stack.*/
54         {NULL, NULL}    /* Terminator */
55 };
56 const int kretprobe_blacklist_size = ARRAY_SIZE(kretprobe_blacklist);
57
58 /*
59  * returns non-zero if opcode modifies the interrupt flag.
60  */
61 static __always_inline int is_IF_modifier(kprobe_opcode_t *insn)
62 {
63         switch (*insn) {
64         case 0xfa:              /* cli */
65         case 0xfb:              /* sti */
66         case 0xcf:              /* iret/iretd */
67         case 0x9d:              /* popf/popfd */
68                 return 1;
69         }
70
71         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
72                 return 1;
73         return 0;
74 }
75
76 int __kprobes arch_prepare_kprobe(struct kprobe *p)
77 {
78         /* insn: must be on special executable page on x86_64. */
79         p->ainsn.insn = get_insn_slot();
80         if (!p->ainsn.insn) {
81                 return -ENOMEM;
82         }
83         arch_copy_kprobe(p);
84         return 0;
85 }
86
87 /*
88  * Determine if the instruction uses the %rip-relative addressing mode.
89  * If it does, return the address of the 32-bit displacement word.
90  * If not, return null.
91  */
92 static s32 __kprobes *is_riprel(u8 *insn)
93 {
94 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
95         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
96           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
97           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
98           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
99          << (row % 64))
100         static const u64 onebyte_has_modrm[256 / 64] = {
101                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
102                 /*      -------------------------------         */
103                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
104                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
105                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
106                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
107                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
108                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
109                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
110                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
111                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
112                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
113                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
114                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
115                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
116                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
117                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
118                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
119                 /*      -------------------------------         */
120                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
121         };
122         static const u64 twobyte_has_modrm[256 / 64] = {
123                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
124                 /*      -------------------------------         */
125                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
126                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
127                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
128                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
129                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
130                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
131                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
132                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
133                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
134                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
135                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
136                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
137                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
138                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
139                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
140                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
141                 /*      -------------------------------         */
142                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
143         };
144 #undef  W
145         int need_modrm;
146
147         /* Skip legacy instruction prefixes.  */
148         while (1) {
149                 switch (*insn) {
150                 case 0x66:
151                 case 0x67:
152                 case 0x2e:
153                 case 0x3e:
154                 case 0x26:
155                 case 0x64:
156                 case 0x65:
157                 case 0x36:
158                 case 0xf0:
159                 case 0xf3:
160                 case 0xf2:
161                         ++insn;
162                         continue;
163                 }
164                 break;
165         }
166
167         /* Skip REX instruction prefix.  */
168         if ((*insn & 0xf0) == 0x40)
169                 ++insn;
170
171         if (*insn == 0x0f) {    /* Two-byte opcode.  */
172                 ++insn;
173                 need_modrm = test_bit(*insn, twobyte_has_modrm);
174         } else {                /* One-byte opcode.  */
175                 need_modrm = test_bit(*insn, onebyte_has_modrm);
176         }
177
178         if (need_modrm) {
179                 u8 modrm = *++insn;
180                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
181                         /* Displacement follows ModRM byte.  */
182                         return (s32 *) ++insn;
183                 }
184         }
185
186         /* No %rip-relative addressing mode here.  */
187         return NULL;
188 }
189
190 static void __kprobes arch_copy_kprobe(struct kprobe *p)
191 {
192         s32 *ripdisp;
193         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
194         ripdisp = is_riprel(p->ainsn.insn);
195         if (ripdisp) {
196                 /*
197                  * The copied instruction uses the %rip-relative
198                  * addressing mode.  Adjust the displacement for the
199                  * difference between the original location of this
200                  * instruction and the location of the copy that will
201                  * actually be run.  The tricky bit here is making sure
202                  * that the sign extension happens correctly in this
203                  * calculation, since we need a signed 32-bit result to
204                  * be sign-extended to 64 bits when it's added to the
205                  * %rip value and yield the same 64-bit result that the
206                  * sign-extension of the original signed 32-bit
207                  * displacement would have given.
208                  */
209                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
210                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
211                 *ripdisp = disp;
212         }
213         p->opcode = *p->addr;
214 }
215
216 void __kprobes arch_arm_kprobe(struct kprobe *p)
217 {
218         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
219 }
220
221 void __kprobes arch_disarm_kprobe(struct kprobe *p)
222 {
223         text_poke(p->addr, &p->opcode, 1);
224 }
225
226 void __kprobes arch_remove_kprobe(struct kprobe *p)
227 {
228         mutex_lock(&kprobe_mutex);
229         free_insn_slot(p->ainsn.insn, 0);
230         mutex_unlock(&kprobe_mutex);
231 }
232
233 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
234 {
235         kcb->prev_kprobe.kp = kprobe_running();
236         kcb->prev_kprobe.status = kcb->kprobe_status;
237         kcb->prev_kprobe.old_rflags = kcb->kprobe_old_rflags;
238         kcb->prev_kprobe.saved_rflags = kcb->kprobe_saved_rflags;
239 }
240
241 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
242 {
243         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
244         kcb->kprobe_status = kcb->prev_kprobe.status;
245         kcb->kprobe_old_rflags = kcb->prev_kprobe.old_rflags;
246         kcb->kprobe_saved_rflags = kcb->prev_kprobe.saved_rflags;
247 }
248
249 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
250                                 struct kprobe_ctlblk *kcb)
251 {
252         __get_cpu_var(current_kprobe) = p;
253         kcb->kprobe_saved_rflags = kcb->kprobe_old_rflags
254                 = (regs->eflags & (TF_MASK | IF_MASK));
255         if (is_IF_modifier(p->ainsn.insn))
256                 kcb->kprobe_saved_rflags &= ~IF_MASK;
257 }
258
259 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
260 {
261         regs->eflags |= TF_MASK;
262         regs->eflags &= ~IF_MASK;
263         /*single step inline if the instruction is an int3*/
264         if (p->opcode == BREAKPOINT_INSTRUCTION)
265                 regs->rip = (unsigned long)p->addr;
266         else
267                 regs->rip = (unsigned long)p->ainsn.insn;
268 }
269
270 /* Called with kretprobe_lock held */
271 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
272                                       struct pt_regs *regs)
273 {
274         unsigned long *sara = (unsigned long *)regs->rsp;
275
276         ri->ret_addr = (kprobe_opcode_t *) *sara;
277         /* Replace the return addr with trampoline addr */
278         *sara = (unsigned long) &kretprobe_trampoline;
279 }
280
281 int __kprobes kprobe_handler(struct pt_regs *regs)
282 {
283         struct kprobe *p;
284         int ret = 0;
285         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
286         struct kprobe_ctlblk *kcb;
287
288         /*
289          * We don't want to be preempted for the entire
290          * duration of kprobe processing
291          */
292         preempt_disable();
293         kcb = get_kprobe_ctlblk();
294
295         /* Check we're not actually recursing */
296         if (kprobe_running()) {
297                 p = get_kprobe(addr);
298                 if (p) {
299                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
300                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
301                                 regs->eflags &= ~TF_MASK;
302                                 regs->eflags |= kcb->kprobe_saved_rflags;
303                                 goto no_kprobe;
304                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
305                                 /* TODO: Provide re-entrancy from
306                                  * post_kprobes_handler() and avoid exception
307                                  * stack corruption while single-stepping on
308                                  * the instruction of the new probe.
309                                  */
310                                 arch_disarm_kprobe(p);
311                                 regs->rip = (unsigned long)p->addr;
312                                 reset_current_kprobe();
313                                 ret = 1;
314                         } else {
315                                 /* We have reentered the kprobe_handler(), since
316                                  * another probe was hit while within the
317                                  * handler. We here save the original kprobe
318                                  * variables and just single step on instruction
319                                  * of the new probe without calling any user
320                                  * handlers.
321                                  */
322                                 save_previous_kprobe(kcb);
323                                 set_current_kprobe(p, regs, kcb);
324                                 kprobes_inc_nmissed_count(p);
325                                 prepare_singlestep(p, regs);
326                                 kcb->kprobe_status = KPROBE_REENTER;
327                                 return 1;
328                         }
329                 } else {
330                         if (*addr != BREAKPOINT_INSTRUCTION) {
331                         /* The breakpoint instruction was removed by
332                          * another cpu right after we hit, no further
333                          * handling of this interrupt is appropriate
334                          */
335                                 regs->rip = (unsigned long)addr;
336                                 ret = 1;
337                                 goto no_kprobe;
338                         }
339                         p = __get_cpu_var(current_kprobe);
340                         if (p->break_handler && p->break_handler(p, regs)) {
341                                 goto ss_probe;
342                         }
343                 }
344                 goto no_kprobe;
345         }
346
347         p = get_kprobe(addr);
348         if (!p) {
349                 if (*addr != BREAKPOINT_INSTRUCTION) {
350                         /*
351                          * The breakpoint instruction was removed right
352                          * after we hit it.  Another cpu has removed
353                          * either a probepoint or a debugger breakpoint
354                          * at this address.  In either case, no further
355                          * handling of this interrupt is appropriate.
356                          * Back up over the (now missing) int3 and run
357                          * the original instruction.
358                          */
359                         regs->rip = (unsigned long)addr;
360                         ret = 1;
361                 }
362                 /* Not one of ours: let kernel handle it */
363                 goto no_kprobe;
364         }
365
366         set_current_kprobe(p, regs, kcb);
367         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
368
369         if (p->pre_handler && p->pre_handler(p, regs))
370                 /* handler has already set things up, so skip ss setup */
371                 return 1;
372
373 ss_probe:
374         prepare_singlestep(p, regs);
375         kcb->kprobe_status = KPROBE_HIT_SS;
376         return 1;
377
378 no_kprobe:
379         preempt_enable_no_resched();
380         return ret;
381 }
382
383 /*
384  * For function-return probes, init_kprobes() establishes a probepoint
385  * here. When a retprobed function returns, this probe is hit and
386  * trampoline_probe_handler() runs, calling the kretprobe's handler.
387  */
388  void kretprobe_trampoline_holder(void)
389  {
390         asm volatile (  ".global kretprobe_trampoline\n"
391                         "kretprobe_trampoline: \n"
392                         "nop\n");
393  }
394
395 /*
396  * Called when we hit the probe point at kretprobe_trampoline
397  */
398 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
399 {
400         struct kretprobe_instance *ri = NULL;
401         struct hlist_head *head, empty_rp;
402         struct hlist_node *node, *tmp;
403         unsigned long flags, orig_ret_address = 0;
404         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
405
406         INIT_HLIST_HEAD(&empty_rp);
407         spin_lock_irqsave(&kretprobe_lock, flags);
408         head = kretprobe_inst_table_head(current);
409
410         /*
411          * It is possible to have multiple instances associated with a given
412          * task either because an multiple functions in the call path
413          * have a return probe installed on them, and/or more then one return
414          * return probe was registered for a target function.
415          *
416          * We can handle this because:
417          *     - instances are always inserted at the head of the list
418          *     - when multiple return probes are registered for the same
419          *       function, the first instance's ret_addr will point to the
420          *       real return address, and all the rest will point to
421          *       kretprobe_trampoline
422          */
423         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
424                 if (ri->task != current)
425                         /* another task is sharing our hash bucket */
426                         continue;
427
428                 if (ri->rp && ri->rp->handler)
429                         ri->rp->handler(ri, regs);
430
431                 orig_ret_address = (unsigned long)ri->ret_addr;
432                 recycle_rp_inst(ri, &empty_rp);
433
434                 if (orig_ret_address != trampoline_address)
435                         /*
436                          * This is the real return address. Any other
437                          * instances associated with this task are for
438                          * other calls deeper on the call stack
439                          */
440                         break;
441         }
442
443         kretprobe_assert(ri, orig_ret_address, trampoline_address);
444         regs->rip = orig_ret_address;
445
446         reset_current_kprobe();
447         spin_unlock_irqrestore(&kretprobe_lock, flags);
448         preempt_enable_no_resched();
449
450         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
451                 hlist_del(&ri->hlist);
452                 kfree(ri);
453         }
454         /*
455          * By returning a non-zero value, we are telling
456          * kprobe_handler() that we don't want the post_handler
457          * to run (and have re-enabled preemption)
458          */
459         return 1;
460 }
461
462 /*
463  * Called after single-stepping.  p->addr is the address of the
464  * instruction whose first byte has been replaced by the "int 3"
465  * instruction.  To avoid the SMP problems that can occur when we
466  * temporarily put back the original opcode to single-step, we
467  * single-stepped a copy of the instruction.  The address of this
468  * copy is p->ainsn.insn.
469  *
470  * This function prepares to return from the post-single-step
471  * interrupt.  We have to fix up the stack as follows:
472  *
473  * 0) Except in the case of absolute or indirect jump or call instructions,
474  * the new rip is relative to the copied instruction.  We need to make
475  * it relative to the original instruction.
476  *
477  * 1) If the single-stepped instruction was pushfl, then the TF and IF
478  * flags are set in the just-pushed eflags, and may need to be cleared.
479  *
480  * 2) If the single-stepped instruction was a call, the return address
481  * that is atop the stack is the address following the copied instruction.
482  * We need to make it the address following the original instruction.
483  */
484 static void __kprobes resume_execution(struct kprobe *p,
485                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
486 {
487         unsigned long *tos = (unsigned long *)regs->rsp;
488         unsigned long next_rip = 0;
489         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
490         unsigned long orig_rip = (unsigned long)p->addr;
491         kprobe_opcode_t *insn = p->ainsn.insn;
492
493         /*skip the REX prefix*/
494         if (*insn >= 0x40 && *insn <= 0x4f)
495                 insn++;
496
497         switch (*insn) {
498         case 0x9c:              /* pushfl */
499                 *tos &= ~(TF_MASK | IF_MASK);
500                 *tos |= kcb->kprobe_old_rflags;
501                 break;
502         case 0xc3:              /* ret/lret */
503         case 0xcb:
504         case 0xc2:
505         case 0xca:
506                 regs->eflags &= ~TF_MASK;
507                 /* rip is already adjusted, no more changes required*/
508                 return;
509         case 0xe8:              /* call relative - Fix return addr */
510                 *tos = orig_rip + (*tos - copy_rip);
511                 break;
512         case 0xff:
513                 if ((insn[1] & 0x30) == 0x10) {
514                         /* call absolute, indirect */
515                         /* Fix return addr; rip is correct. */
516                         next_rip = regs->rip;
517                         *tos = orig_rip + (*tos - copy_rip);
518                 } else if (((insn[1] & 0x31) == 0x20) ||        /* jmp near, absolute indirect */
519                            ((insn[1] & 0x31) == 0x21)) {        /* jmp far, absolute indirect */
520                         /* rip is correct. */
521                         next_rip = regs->rip;
522                 }
523                 break;
524         case 0xea:              /* jmp absolute -- rip is correct */
525                 next_rip = regs->rip;
526                 break;
527         default:
528                 break;
529         }
530
531         regs->eflags &= ~TF_MASK;
532         if (next_rip) {
533                 regs->rip = next_rip;
534         } else {
535                 regs->rip = orig_rip + (regs->rip - copy_rip);
536         }
537 }
538
539 int __kprobes post_kprobe_handler(struct pt_regs *regs)
540 {
541         struct kprobe *cur = kprobe_running();
542         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
543
544         if (!cur)
545                 return 0;
546
547         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
548                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
549                 cur->post_handler(cur, regs, 0);
550         }
551
552         resume_execution(cur, regs, kcb);
553         regs->eflags |= kcb->kprobe_saved_rflags;
554 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT
555         if (raw_irqs_disabled_flags(regs->eflags))
556                 trace_hardirqs_off();
557         else
558                 trace_hardirqs_on();
559 #endif
560
561         /* Restore the original saved kprobes variables and continue. */
562         if (kcb->kprobe_status == KPROBE_REENTER) {
563                 restore_previous_kprobe(kcb);
564                 goto out;
565         }
566         reset_current_kprobe();
567 out:
568         preempt_enable_no_resched();
569
570         /*
571          * if somebody else is singlestepping across a probe point, eflags
572          * will have TF set, in which case, continue the remaining processing
573          * of do_debug, as if this is not a probe hit.
574          */
575         if (regs->eflags & TF_MASK)
576                 return 0;
577
578         return 1;
579 }
580
581 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
582 {
583         struct kprobe *cur = kprobe_running();
584         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
585         const struct exception_table_entry *fixup;
586
587         switch(kcb->kprobe_status) {
588         case KPROBE_HIT_SS:
589         case KPROBE_REENTER:
590                 /*
591                  * We are here because the instruction being single
592                  * stepped caused a page fault. We reset the current
593                  * kprobe and the rip points back to the probe address
594                  * and allow the page fault handler to continue as a
595                  * normal page fault.
596                  */
597                 regs->rip = (unsigned long)cur->addr;
598                 regs->eflags |= kcb->kprobe_old_rflags;
599                 if (kcb->kprobe_status == KPROBE_REENTER)
600                         restore_previous_kprobe(kcb);
601                 else
602                         reset_current_kprobe();
603                 preempt_enable_no_resched();
604                 break;
605         case KPROBE_HIT_ACTIVE:
606         case KPROBE_HIT_SSDONE:
607                 /*
608                  * We increment the nmissed count for accounting,
609                  * we can also use npre/npostfault count for accouting
610                  * these specific fault cases.
611                  */
612                 kprobes_inc_nmissed_count(cur);
613
614                 /*
615                  * We come here because instructions in the pre/post
616                  * handler caused the page_fault, this could happen
617                  * if handler tries to access user space by
618                  * copy_from_user(), get_user() etc. Let the
619                  * user-specified handler try to fix it first.
620                  */
621                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
622                         return 1;
623
624                 /*
625                  * In case the user-specified fault handler returned
626                  * zero, try to fix up.
627                  */
628                 fixup = search_exception_tables(regs->rip);
629                 if (fixup) {
630                         regs->rip = fixup->fixup;
631                         return 1;
632                 }
633
634                 /*
635                  * fixup() could not handle it,
636                  * Let do_page_fault() fix it.
637                  */
638                 break;
639         default:
640                 break;
641         }
642         return 0;
643 }
644
645 /*
646  * Wrapper routine for handling exceptions.
647  */
648 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
649                                        unsigned long val, void *data)
650 {
651         struct die_args *args = (struct die_args *)data;
652         int ret = NOTIFY_DONE;
653
654         if (args->regs && user_mode(args->regs))
655                 return ret;
656
657         switch (val) {
658         case DIE_INT3:
659                 if (kprobe_handler(args->regs))
660                         ret = NOTIFY_STOP;
661                 break;
662         case DIE_DEBUG:
663                 if (post_kprobe_handler(args->regs))
664                         ret = NOTIFY_STOP;
665                 break;
666         case DIE_GPF:
667                 /* kprobe_running() needs smp_processor_id() */
668                 preempt_disable();
669                 if (kprobe_running() &&
670                     kprobe_fault_handler(args->regs, args->trapnr))
671                         ret = NOTIFY_STOP;
672                 preempt_enable();
673                 break;
674         default:
675                 break;
676         }
677         return ret;
678 }
679
680 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
681 {
682         struct jprobe *jp = container_of(p, struct jprobe, kp);
683         unsigned long addr;
684         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
685
686         kcb->jprobe_saved_regs = *regs;
687         kcb->jprobe_saved_rsp = (long *) regs->rsp;
688         addr = (unsigned long)(kcb->jprobe_saved_rsp);
689         /*
690          * As Linus pointed out, gcc assumes that the callee
691          * owns the argument space and could overwrite it, e.g.
692          * tailcall optimization. So, to be absolutely safe
693          * we also save and restore enough stack bytes to cover
694          * the argument area.
695          */
696         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
697                         MIN_STACK_SIZE(addr));
698         regs->eflags &= ~IF_MASK;
699         trace_hardirqs_off();
700         regs->rip = (unsigned long)(jp->entry);
701         return 1;
702 }
703
704 void __kprobes jprobe_return(void)
705 {
706         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
707
708         asm volatile ("       xchg   %%rbx,%%rsp     \n"
709                       "       int3                      \n"
710                       "       .globl jprobe_return_end  \n"
711                       "       jprobe_return_end:        \n"
712                       "       nop                       \n"::"b"
713                       (kcb->jprobe_saved_rsp):"memory");
714 }
715
716 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
717 {
718         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
719         u8 *addr = (u8 *) (regs->rip - 1);
720         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
721         struct jprobe *jp = container_of(p, struct jprobe, kp);
722
723         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
724                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
725                         struct pt_regs *saved_regs =
726                             container_of(kcb->jprobe_saved_rsp,
727                                             struct pt_regs, rsp);
728                         printk("current rsp %p does not match saved rsp %p\n",
729                                (long *)regs->rsp, kcb->jprobe_saved_rsp);
730                         printk("Saved registers for jprobe %p\n", jp);
731                         show_registers(saved_regs);
732                         printk("Current registers\n");
733                         show_registers(regs);
734                         BUG();
735                 }
736                 *regs = kcb->jprobe_saved_regs;
737                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
738                        MIN_STACK_SIZE(stack_addr));
739                 preempt_enable_no_resched();
740                 return 1;
741         }
742         return 0;
743 }
744
745 static struct kprobe trampoline_p = {
746         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
747         .pre_handler = trampoline_probe_handler
748 };
749
750 int __init arch_init_kprobes(void)
751 {
752         return register_kprobe(&trampoline_p);
753 }
754
755 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
756 {
757         if (p->addr == (kprobe_opcode_t *)&kretprobe_trampoline)
758                 return 1;
759
760         return 0;
761 }