9bef2c8dc12ca729853142a105cc201dba899dd7
[linux-2.6.git] / arch / x86_64 / kernel / kprobes.c
1 /*
2  *  Kernel Probes (KProbes)
3  *  arch/x86_64/kernel/kprobes.c
4  *
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.
9  *
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.
14  *
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.
18  *
19  * Copyright (C) IBM Corporation, 2002, 2004
20  *
21  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
22  *              Probes initial implementation ( includes contributions from
23  *              Rusty Russell).
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
32  */
33
34 #include <linux/config.h>
35 #include <linux/kprobes.h>
36 #include <linux/ptrace.h>
37 #include <linux/string.h>
38 #include <linux/slab.h>
39 #include <linux/preempt.h>
40
41 #include <asm/cacheflush.h>
42 #include <asm/pgtable.h>
43 #include <asm/kdebug.h>
44
45 static DECLARE_MUTEX(kprobe_mutex);
46 void jprobe_return_end(void);
47
48 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
49 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
50
51 /*
52  * returns non-zero if opcode modifies the interrupt flag.
53  */
54 static inline int is_IF_modifier(kprobe_opcode_t *insn)
55 {
56         switch (*insn) {
57         case 0xfa:              /* cli */
58         case 0xfb:              /* sti */
59         case 0xcf:              /* iret/iretd */
60         case 0x9d:              /* popf/popfd */
61                 return 1;
62         }
63
64         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
65                 return 1;
66         return 0;
67 }
68
69 int __kprobes arch_prepare_kprobe(struct kprobe *p)
70 {
71         /* insn: must be on special executable page on x86_64. */
72         down(&kprobe_mutex);
73         p->ainsn.insn = get_insn_slot();
74         up(&kprobe_mutex);
75         if (!p->ainsn.insn) {
76                 return -ENOMEM;
77         }
78         return 0;
79 }
80
81 /*
82  * Determine if the instruction uses the %rip-relative addressing mode.
83  * If it does, return the address of the 32-bit displacement word.
84  * If not, return null.
85  */
86 static inline s32 *is_riprel(u8 *insn)
87 {
88 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
89         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
90           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
91           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
92           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
93          << (row % 64))
94         static const u64 onebyte_has_modrm[256 / 64] = {
95                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
96                 /*      -------------------------------         */
97                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
98                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
99                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
100                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
101                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
102                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
103                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
104                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
105                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
106                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
107                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
108                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
109                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
110                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
111                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
112                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
113                 /*      -------------------------------         */
114                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
115         };
116         static const u64 twobyte_has_modrm[256 / 64] = {
117                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
118                 /*      -------------------------------         */
119                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
120                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
121                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
122                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
123                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
124                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
125                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
126                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
127                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
128                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
129                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
130                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
131                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
132                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
133                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
134                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
135                 /*      -------------------------------         */
136                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
137         };
138 #undef  W
139         int need_modrm;
140
141         /* Skip legacy instruction prefixes.  */
142         while (1) {
143                 switch (*insn) {
144                 case 0x66:
145                 case 0x67:
146                 case 0x2e:
147                 case 0x3e:
148                 case 0x26:
149                 case 0x64:
150                 case 0x65:
151                 case 0x36:
152                 case 0xf0:
153                 case 0xf3:
154                 case 0xf2:
155                         ++insn;
156                         continue;
157                 }
158                 break;
159         }
160
161         /* Skip REX instruction prefix.  */
162         if ((*insn & 0xf0) == 0x40)
163                 ++insn;
164
165         if (*insn == 0x0f) {    /* Two-byte opcode.  */
166                 ++insn;
167                 need_modrm = test_bit(*insn, twobyte_has_modrm);
168         } else {                /* One-byte opcode.  */
169                 need_modrm = test_bit(*insn, onebyte_has_modrm);
170         }
171
172         if (need_modrm) {
173                 u8 modrm = *++insn;
174                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
175                         /* Displacement follows ModRM byte.  */
176                         return (s32 *) ++insn;
177                 }
178         }
179
180         /* No %rip-relative addressing mode here.  */
181         return NULL;
182 }
183
184 void __kprobes arch_copy_kprobe(struct kprobe *p)
185 {
186         s32 *ripdisp;
187         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
188         ripdisp = is_riprel(p->ainsn.insn);
189         if (ripdisp) {
190                 /*
191                  * The copied instruction uses the %rip-relative
192                  * addressing mode.  Adjust the displacement for the
193                  * difference between the original location of this
194                  * instruction and the location of the copy that will
195                  * actually be run.  The tricky bit here is making sure
196                  * that the sign extension happens correctly in this
197                  * calculation, since we need a signed 32-bit result to
198                  * be sign-extended to 64 bits when it's added to the
199                  * %rip value and yield the same 64-bit result that the
200                  * sign-extension of the original signed 32-bit
201                  * displacement would have given.
202                  */
203                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
204                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
205                 *ripdisp = disp;
206         }
207         p->opcode = *p->addr;
208 }
209
210 void __kprobes arch_arm_kprobe(struct kprobe *p)
211 {
212         *p->addr = BREAKPOINT_INSTRUCTION;
213         flush_icache_range((unsigned long) p->addr,
214                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
215 }
216
217 void __kprobes arch_disarm_kprobe(struct kprobe *p)
218 {
219         *p->addr = p->opcode;
220         flush_icache_range((unsigned long) p->addr,
221                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
222 }
223
224 void __kprobes arch_remove_kprobe(struct kprobe *p)
225 {
226         down(&kprobe_mutex);
227         free_insn_slot(p->ainsn.insn);
228         up(&kprobe_mutex);
229 }
230
231 static inline void save_previous_kprobe(struct kprobe_ctlblk *kcb)
232 {
233         kcb->prev_kprobe.kp = kprobe_running();
234         kcb->prev_kprobe.status = kcb->kprobe_status;
235         kcb->prev_kprobe.old_rflags = kcb->kprobe_old_rflags;
236         kcb->prev_kprobe.saved_rflags = kcb->kprobe_saved_rflags;
237 }
238
239 static inline void restore_previous_kprobe(struct kprobe_ctlblk *kcb)
240 {
241         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
242         kcb->kprobe_status = kcb->prev_kprobe.status;
243         kcb->kprobe_old_rflags = kcb->prev_kprobe.old_rflags;
244         kcb->kprobe_saved_rflags = kcb->prev_kprobe.saved_rflags;
245 }
246
247 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
248                                 struct kprobe_ctlblk *kcb)
249 {
250         __get_cpu_var(current_kprobe) = p;
251         kcb->kprobe_saved_rflags = kcb->kprobe_old_rflags
252                 = (regs->eflags & (TF_MASK | IF_MASK));
253         if (is_IF_modifier(p->ainsn.insn))
254                 kcb->kprobe_saved_rflags &= ~IF_MASK;
255 }
256
257 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
258 {
259         regs->eflags |= TF_MASK;
260         regs->eflags &= ~IF_MASK;
261         /*single step inline if the instruction is an int3*/
262         if (p->opcode == BREAKPOINT_INSTRUCTION)
263                 regs->rip = (unsigned long)p->addr;
264         else
265                 regs->rip = (unsigned long)p->ainsn.insn;
266 }
267
268 /* Called with kretprobe_lock held */
269 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
270                                       struct pt_regs *regs)
271 {
272         unsigned long *sara = (unsigned long *)regs->rsp;
273         struct kretprobe_instance *ri;
274
275         if ((ri = get_free_rp_inst(rp)) != NULL) {
276                 ri->rp = rp;
277                 ri->task = current;
278                 ri->ret_addr = (kprobe_opcode_t *) *sara;
279
280                 /* Replace the return addr with trampoline addr */
281                 *sara = (unsigned long) &kretprobe_trampoline;
282
283                 add_rp_inst(ri);
284         } else {
285                 rp->nmissed++;
286         }
287 }
288
289 /*
290  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
291  * remain disabled thorough out this function.
292  */
293 int __kprobes kprobe_handler(struct pt_regs *regs)
294 {
295         struct kprobe *p;
296         int ret = 0;
297         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
298         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
299
300         /* Check we're not actually recursing */
301         if (kprobe_running()) {
302                 p = get_kprobe(addr);
303                 if (p) {
304                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
305                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
306                                 regs->eflags &= ~TF_MASK;
307                                 regs->eflags |= kcb->kprobe_saved_rflags;
308                                 goto no_kprobe;
309                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
310                                 /* TODO: Provide re-entrancy from
311                                  * post_kprobes_handler() and avoid exception
312                                  * stack corruption while single-stepping on
313                                  * the instruction of the new probe.
314                                  */
315                                 arch_disarm_kprobe(p);
316                                 regs->rip = (unsigned long)p->addr;
317                                 reset_current_kprobe();
318                                 ret = 1;
319                         } else {
320                                 /* We have reentered the kprobe_handler(), since
321                                  * another probe was hit while within the
322                                  * handler. We here save the original kprobe
323                                  * variables and just single step on instruction
324                                  * of the new probe without calling any user
325                                  * handlers.
326                                  */
327                                 save_previous_kprobe(kcb);
328                                 set_current_kprobe(p, regs, kcb);
329                                 p->nmissed++;
330                                 prepare_singlestep(p, regs);
331                                 kcb->kprobe_status = KPROBE_REENTER;
332                                 return 1;
333                         }
334                 } else {
335                         p = __get_cpu_var(current_kprobe);
336                         if (p->break_handler && p->break_handler(p, regs)) {
337                                 goto ss_probe;
338                         }
339                 }
340                 goto no_kprobe;
341         }
342
343         p = get_kprobe(addr);
344         if (!p) {
345                 if (*addr != BREAKPOINT_INSTRUCTION) {
346                         /*
347                          * The breakpoint instruction was removed right
348                          * after we hit it.  Another cpu has removed
349                          * either a probepoint or a debugger breakpoint
350                          * at this address.  In either case, no further
351                          * handling of this interrupt is appropriate.
352                          * Back up over the (now missing) int3 and run
353                          * the original instruction.
354                          */
355                         regs->rip = (unsigned long)addr;
356                         ret = 1;
357                 }
358                 /* Not one of ours: let kernel handle it */
359                 goto no_kprobe;
360         }
361
362         /*
363          * This preempt_disable() matches the preempt_enable_no_resched()
364          * in post_kprobe_handler()
365          */
366         preempt_disable();
367         set_current_kprobe(p, regs, kcb);
368         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
369
370         if (p->pre_handler && p->pre_handler(p, regs))
371                 /* handler has already set things up, so skip ss setup */
372                 return 1;
373
374 ss_probe:
375         prepare_singlestep(p, regs);
376         kcb->kprobe_status = KPROBE_HIT_SS;
377         return 1;
378
379 no_kprobe:
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;
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         spin_lock_irqsave(&kretprobe_lock, flags);
407         head = kretprobe_inst_table_head(current);
408
409         /*
410          * It is possible to have multiple instances associated with a given
411          * task either because an multiple functions in the call path
412          * have a return probe installed on them, and/or more then one return
413          * return probe was registered for a target function.
414          *
415          * We can handle this because:
416          *     - instances are always inserted at the head of the list
417          *     - when multiple return probes are registered for the same
418          *       function, the first instance's ret_addr will point to the
419          *       real return address, and all the rest will point to
420          *       kretprobe_trampoline
421          */
422         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
423                 if (ri->task != current)
424                         /* another task is sharing our hash bucket */
425                         continue;
426
427                 if (ri->rp && ri->rp->handler)
428                         ri->rp->handler(ri, regs);
429
430                 orig_ret_address = (unsigned long)ri->ret_addr;
431                 recycle_rp_inst(ri);
432
433                 if (orig_ret_address != trampoline_address)
434                         /*
435                          * This is the real return address. Any other
436                          * instances associated with this task are for
437                          * other calls deeper on the call stack
438                          */
439                         break;
440         }
441
442         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
443         regs->rip = orig_ret_address;
444
445         reset_current_kprobe();
446         spin_unlock_irqrestore(&kretprobe_lock, flags);
447         preempt_enable_no_resched();
448
449         /*
450          * By returning a non-zero value, we are telling
451          * kprobe_handler() that we have handled unlocking
452          * and re-enabling preemption
453          */
454         return 1;
455 }
456
457 /*
458  * Called after single-stepping.  p->addr is the address of the
459  * instruction whose first byte has been replaced by the "int 3"
460  * instruction.  To avoid the SMP problems that can occur when we
461  * temporarily put back the original opcode to single-step, we
462  * single-stepped a copy of the instruction.  The address of this
463  * copy is p->ainsn.insn.
464  *
465  * This function prepares to return from the post-single-step
466  * interrupt.  We have to fix up the stack as follows:
467  *
468  * 0) Except in the case of absolute or indirect jump or call instructions,
469  * the new rip is relative to the copied instruction.  We need to make
470  * it relative to the original instruction.
471  *
472  * 1) If the single-stepped instruction was pushfl, then the TF and IF
473  * flags are set in the just-pushed eflags, and may need to be cleared.
474  *
475  * 2) If the single-stepped instruction was a call, the return address
476  * that is atop the stack is the address following the copied instruction.
477  * We need to make it the address following the original instruction.
478  */
479 static void __kprobes resume_execution(struct kprobe *p,
480                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
481 {
482         unsigned long *tos = (unsigned long *)regs->rsp;
483         unsigned long next_rip = 0;
484         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
485         unsigned long orig_rip = (unsigned long)p->addr;
486         kprobe_opcode_t *insn = p->ainsn.insn;
487
488         /*skip the REX prefix*/
489         if (*insn >= 0x40 && *insn <= 0x4f)
490                 insn++;
491
492         switch (*insn) {
493         case 0x9c:              /* pushfl */
494                 *tos &= ~(TF_MASK | IF_MASK);
495                 *tos |= kcb->kprobe_old_rflags;
496                 break;
497         case 0xc3:              /* ret/lret */
498         case 0xcb:
499         case 0xc2:
500         case 0xca:
501                 regs->eflags &= ~TF_MASK;
502                 /* rip is already adjusted, no more changes required*/
503                 return;
504         case 0xe8:              /* call relative - Fix return addr */
505                 *tos = orig_rip + (*tos - copy_rip);
506                 break;
507         case 0xff:
508                 if ((*insn & 0x30) == 0x10) {
509                         /* call absolute, indirect */
510                         /* Fix return addr; rip is correct. */
511                         next_rip = regs->rip;
512                         *tos = orig_rip + (*tos - copy_rip);
513                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
514                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
515                         /* rip is correct. */
516                         next_rip = regs->rip;
517                 }
518                 break;
519         case 0xea:              /* jmp absolute -- rip is correct */
520                 next_rip = regs->rip;
521                 break;
522         default:
523                 break;
524         }
525
526         regs->eflags &= ~TF_MASK;
527         if (next_rip) {
528                 regs->rip = next_rip;
529         } else {
530                 regs->rip = orig_rip + (regs->rip - copy_rip);
531         }
532 }
533
534 int __kprobes post_kprobe_handler(struct pt_regs *regs)
535 {
536         struct kprobe *cur = kprobe_running();
537         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
538
539         if (!cur)
540                 return 0;
541
542         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
543                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
544                 cur->post_handler(cur, regs, 0);
545         }
546
547         resume_execution(cur, regs, kcb);
548         regs->eflags |= kcb->kprobe_saved_rflags;
549
550         /* Restore the original saved kprobes variables and continue. */
551         if (kcb->kprobe_status == KPROBE_REENTER) {
552                 restore_previous_kprobe(kcb);
553                 goto out;
554         }
555         reset_current_kprobe();
556 out:
557         preempt_enable_no_resched();
558
559         /*
560          * if somebody else is singlestepping across a probe point, eflags
561          * will have TF set, in which case, continue the remaining processing
562          * of do_debug, as if this is not a probe hit.
563          */
564         if (regs->eflags & TF_MASK)
565                 return 0;
566
567         return 1;
568 }
569
570 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
571 {
572         struct kprobe *cur = kprobe_running();
573         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
574
575         if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
576                 return 1;
577
578         if (kcb->kprobe_status & KPROBE_HIT_SS) {
579                 resume_execution(cur, regs, kcb);
580                 regs->eflags |= kcb->kprobe_old_rflags;
581
582                 reset_current_kprobe();
583                 preempt_enable_no_resched();
584         }
585         return 0;
586 }
587
588 /*
589  * Wrapper routine for handling exceptions.
590  */
591 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
592                                        unsigned long val, void *data)
593 {
594         struct die_args *args = (struct die_args *)data;
595         int ret = NOTIFY_DONE;
596
597         rcu_read_lock();
598         switch (val) {
599         case DIE_INT3:
600                 if (kprobe_handler(args->regs))
601                         ret = NOTIFY_STOP;
602                 break;
603         case DIE_DEBUG:
604                 if (post_kprobe_handler(args->regs))
605                         ret = NOTIFY_STOP;
606                 break;
607         case DIE_GPF:
608         case DIE_PAGE_FAULT:
609                 if (kprobe_running() &&
610                     kprobe_fault_handler(args->regs, args->trapnr))
611                         ret = NOTIFY_STOP;
612                 break;
613         default:
614                 break;
615         }
616         rcu_read_unlock();
617         return ret;
618 }
619
620 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
621 {
622         struct jprobe *jp = container_of(p, struct jprobe, kp);
623         unsigned long addr;
624         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
625
626         kcb->jprobe_saved_regs = *regs;
627         kcb->jprobe_saved_rsp = (long *) regs->rsp;
628         addr = (unsigned long)(kcb->jprobe_saved_rsp);
629         /*
630          * As Linus pointed out, gcc assumes that the callee
631          * owns the argument space and could overwrite it, e.g.
632          * tailcall optimization. So, to be absolutely safe
633          * we also save and restore enough stack bytes to cover
634          * the argument area.
635          */
636         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
637                         MIN_STACK_SIZE(addr));
638         regs->eflags &= ~IF_MASK;
639         regs->rip = (unsigned long)(jp->entry);
640         return 1;
641 }
642
643 void __kprobes jprobe_return(void)
644 {
645         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
646
647         asm volatile ("       xchg   %%rbx,%%rsp     \n"
648                       "       int3                      \n"
649                       "       .globl jprobe_return_end  \n"
650                       "       jprobe_return_end:        \n"
651                       "       nop                       \n"::"b"
652                       (kcb->jprobe_saved_rsp):"memory");
653 }
654
655 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
656 {
657         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
658         u8 *addr = (u8 *) (regs->rip - 1);
659         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
660         struct jprobe *jp = container_of(p, struct jprobe, kp);
661
662         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
663                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
664                         struct pt_regs *saved_regs =
665                             container_of(kcb->jprobe_saved_rsp,
666                                             struct pt_regs, rsp);
667                         printk("current rsp %p does not match saved rsp %p\n",
668                                (long *)regs->rsp, kcb->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);
673                         BUG();
674                 }
675                 *regs = kcb->jprobe_saved_regs;
676                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
677                        MIN_STACK_SIZE(stack_addr));
678                 return 1;
679         }
680         return 0;
681 }
682
683 static struct kprobe trampoline_p = {
684         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
685         .pre_handler = trampoline_probe_handler
686 };
687
688 int __init arch_init_kprobes(void)
689 {
690         return register_kprobe(&trampoline_p);
691 }