[PATCH] kprobes: enable booster on the preemptible kernel
[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/kprobes.h>
35 #include <linux/ptrace.h>
36 #include <linux/string.h>
37 #include <linux/slab.h>
38 #include <linux/preempt.h>
39 #include <linux/module.h>
40
41 #include <asm/cacheflush.h>
42 #include <asm/pgtable.h>
43 #include <asm/kdebug.h>
44 #include <asm/uaccess.h>
45
46 void jprobe_return_end(void);
47 static void __kprobes arch_copy_kprobe(struct kprobe *p);
48
49 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
50 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
51
52 /*
53  * returns non-zero if opcode modifies the interrupt flag.
54  */
55 static __always_inline int is_IF_modifier(kprobe_opcode_t *insn)
56 {
57         switch (*insn) {
58         case 0xfa:              /* cli */
59         case 0xfb:              /* sti */
60         case 0xcf:              /* iret/iretd */
61         case 0x9d:              /* popf/popfd */
62                 return 1;
63         }
64
65         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
66                 return 1;
67         return 0;
68 }
69
70 int __kprobes arch_prepare_kprobe(struct kprobe *p)
71 {
72         /* insn: must be on special executable page on x86_64. */
73         p->ainsn.insn = get_insn_slot();
74         if (!p->ainsn.insn) {
75                 return -ENOMEM;
76         }
77         arch_copy_kprobe(p);
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 s32 __kprobes *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 static 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         mutex_lock(&kprobe_mutex);
227         free_insn_slot(p->ainsn.insn, 0);
228         mutex_unlock(&kprobe_mutex);
229 }
230
231 static void __kprobes 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 void __kprobes 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 void __kprobes 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                 add_rp_inst(ri);
283         } else {
284                 rp->nmissed++;
285         }
286 }
287
288 int __kprobes kprobe_handler(struct pt_regs *regs)
289 {
290         struct kprobe *p;
291         int ret = 0;
292         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
293         struct kprobe_ctlblk *kcb;
294
295         /*
296          * We don't want to be preempted for the entire
297          * duration of kprobe processing
298          */
299         preempt_disable();
300         kcb = get_kprobe_ctlblk();
301
302         /* Check we're not actually recursing */
303         if (kprobe_running()) {
304                 p = get_kprobe(addr);
305                 if (p) {
306                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
307                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
308                                 regs->eflags &= ~TF_MASK;
309                                 regs->eflags |= kcb->kprobe_saved_rflags;
310                                 goto no_kprobe;
311                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
312                                 /* TODO: Provide re-entrancy from
313                                  * post_kprobes_handler() and avoid exception
314                                  * stack corruption while single-stepping on
315                                  * the instruction of the new probe.
316                                  */
317                                 arch_disarm_kprobe(p);
318                                 regs->rip = (unsigned long)p->addr;
319                                 reset_current_kprobe();
320                                 ret = 1;
321                         } else {
322                                 /* We have reentered the kprobe_handler(), since
323                                  * another probe was hit while within the
324                                  * handler. We here save the original kprobe
325                                  * variables and just single step on instruction
326                                  * of the new probe without calling any user
327                                  * handlers.
328                                  */
329                                 save_previous_kprobe(kcb);
330                                 set_current_kprobe(p, regs, kcb);
331                                 kprobes_inc_nmissed_count(p);
332                                 prepare_singlestep(p, regs);
333                                 kcb->kprobe_status = KPROBE_REENTER;
334                                 return 1;
335                         }
336                 } else {
337                         if (*addr != BREAKPOINT_INSTRUCTION) {
338                         /* The breakpoint instruction was removed by
339                          * another cpu right after we hit, no further
340                          * handling of this interrupt is appropriate
341                          */
342                                 regs->rip = (unsigned long)addr;
343                                 ret = 1;
344                                 goto no_kprobe;
345                         }
346                         p = __get_cpu_var(current_kprobe);
347                         if (p->break_handler && p->break_handler(p, regs)) {
348                                 goto ss_probe;
349                         }
350                 }
351                 goto no_kprobe;
352         }
353
354         p = get_kprobe(addr);
355         if (!p) {
356                 if (*addr != BREAKPOINT_INSTRUCTION) {
357                         /*
358                          * The breakpoint instruction was removed right
359                          * after we hit it.  Another cpu has removed
360                          * either a probepoint or a debugger breakpoint
361                          * at this address.  In either case, no further
362                          * handling of this interrupt is appropriate.
363                          * Back up over the (now missing) int3 and run
364                          * the original instruction.
365                          */
366                         regs->rip = (unsigned long)addr;
367                         ret = 1;
368                 }
369                 /* Not one of ours: let kernel handle it */
370                 goto no_kprobe;
371         }
372
373         set_current_kprobe(p, regs, kcb);
374         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
375
376         if (p->pre_handler && p->pre_handler(p, regs))
377                 /* handler has already set things up, so skip ss setup */
378                 return 1;
379
380 ss_probe:
381         prepare_singlestep(p, regs);
382         kcb->kprobe_status = KPROBE_HIT_SS;
383         return 1;
384
385 no_kprobe:
386         preempt_enable_no_resched();
387         return ret;
388 }
389
390 /*
391  * For function-return probes, init_kprobes() establishes a probepoint
392  * here. When a retprobed function returns, this probe is hit and
393  * trampoline_probe_handler() runs, calling the kretprobe's handler.
394  */
395  void kretprobe_trampoline_holder(void)
396  {
397         asm volatile (  ".global kretprobe_trampoline\n"
398                         "kretprobe_trampoline: \n"
399                         "nop\n");
400  }
401
402 /*
403  * Called when we hit the probe point at kretprobe_trampoline
404  */
405 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
406 {
407         struct kretprobe_instance *ri = NULL;
408         struct hlist_head *head, empty_rp;
409         struct hlist_node *node, *tmp;
410         unsigned long flags, orig_ret_address = 0;
411         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
412
413         INIT_HLIST_HEAD(&empty_rp);
414         spin_lock_irqsave(&kretprobe_lock, flags);
415         head = kretprobe_inst_table_head(current);
416
417         /*
418          * It is possible to have multiple instances associated with a given
419          * task either because an multiple functions in the call path
420          * have a return probe installed on them, and/or more then one return
421          * return probe was registered for a target function.
422          *
423          * We can handle this because:
424          *     - instances are always inserted at the head of the list
425          *     - when multiple return probes are registered for the same
426          *       function, the first instance's ret_addr will point to the
427          *       real return address, and all the rest will point to
428          *       kretprobe_trampoline
429          */
430         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
431                 if (ri->task != current)
432                         /* another task is sharing our hash bucket */
433                         continue;
434
435                 if (ri->rp && ri->rp->handler)
436                         ri->rp->handler(ri, regs);
437
438                 orig_ret_address = (unsigned long)ri->ret_addr;
439                 recycle_rp_inst(ri, &empty_rp);
440
441                 if (orig_ret_address != trampoline_address)
442                         /*
443                          * This is the real return address. Any other
444                          * instances associated with this task are for
445                          * other calls deeper on the call stack
446                          */
447                         break;
448         }
449
450         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
451         regs->rip = orig_ret_address;
452
453         reset_current_kprobe();
454         spin_unlock_irqrestore(&kretprobe_lock, flags);
455         preempt_enable_no_resched();
456
457         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
458                 hlist_del(&ri->hlist);
459                 kfree(ri);
460         }
461         /*
462          * By returning a non-zero value, we are telling
463          * kprobe_handler() that we don't want the post_handler
464          * to run (and have re-enabled preemption)
465          */
466         return 1;
467 }
468
469 /*
470  * Called after single-stepping.  p->addr is the address of the
471  * instruction whose first byte has been replaced by the "int 3"
472  * instruction.  To avoid the SMP problems that can occur when we
473  * temporarily put back the original opcode to single-step, we
474  * single-stepped a copy of the instruction.  The address of this
475  * copy is p->ainsn.insn.
476  *
477  * This function prepares to return from the post-single-step
478  * interrupt.  We have to fix up the stack as follows:
479  *
480  * 0) Except in the case of absolute or indirect jump or call instructions,
481  * the new rip is relative to the copied instruction.  We need to make
482  * it relative to the original instruction.
483  *
484  * 1) If the single-stepped instruction was pushfl, then the TF and IF
485  * flags are set in the just-pushed eflags, and may need to be cleared.
486  *
487  * 2) If the single-stepped instruction was a call, the return address
488  * that is atop the stack is the address following the copied instruction.
489  * We need to make it the address following the original instruction.
490  */
491 static void __kprobes resume_execution(struct kprobe *p,
492                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
493 {
494         unsigned long *tos = (unsigned long *)regs->rsp;
495         unsigned long next_rip = 0;
496         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
497         unsigned long orig_rip = (unsigned long)p->addr;
498         kprobe_opcode_t *insn = p->ainsn.insn;
499
500         /*skip the REX prefix*/
501         if (*insn >= 0x40 && *insn <= 0x4f)
502                 insn++;
503
504         switch (*insn) {
505         case 0x9c:              /* pushfl */
506                 *tos &= ~(TF_MASK | IF_MASK);
507                 *tos |= kcb->kprobe_old_rflags;
508                 break;
509         case 0xc3:              /* ret/lret */
510         case 0xcb:
511         case 0xc2:
512         case 0xca:
513                 regs->eflags &= ~TF_MASK;
514                 /* rip is already adjusted, no more changes required*/
515                 return;
516         case 0xe8:              /* call relative - Fix return addr */
517                 *tos = orig_rip + (*tos - copy_rip);
518                 break;
519         case 0xff:
520                 if ((insn[1] & 0x30) == 0x10) {
521                         /* call absolute, indirect */
522                         /* Fix return addr; rip is correct. */
523                         next_rip = regs->rip;
524                         *tos = orig_rip + (*tos - copy_rip);
525                 } else if (((insn[1] & 0x31) == 0x20) ||        /* jmp near, absolute indirect */
526                            ((insn[1] & 0x31) == 0x21)) {        /* jmp far, absolute indirect */
527                         /* rip is correct. */
528                         next_rip = regs->rip;
529                 }
530                 break;
531         case 0xea:              /* jmp absolute -- rip is correct */
532                 next_rip = regs->rip;
533                 break;
534         default:
535                 break;
536         }
537
538         regs->eflags &= ~TF_MASK;
539         if (next_rip) {
540                 regs->rip = next_rip;
541         } else {
542                 regs->rip = orig_rip + (regs->rip - copy_rip);
543         }
544 }
545
546 int __kprobes post_kprobe_handler(struct pt_regs *regs)
547 {
548         struct kprobe *cur = kprobe_running();
549         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
550
551         if (!cur)
552                 return 0;
553
554         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
555                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
556                 cur->post_handler(cur, regs, 0);
557         }
558
559         resume_execution(cur, regs, kcb);
560         regs->eflags |= kcb->kprobe_saved_rflags;
561
562         /* Restore the original saved kprobes variables and continue. */
563         if (kcb->kprobe_status == KPROBE_REENTER) {
564                 restore_previous_kprobe(kcb);
565                 goto out;
566         }
567         reset_current_kprobe();
568 out:
569         preempt_enable_no_resched();
570
571         /*
572          * if somebody else is singlestepping across a probe point, eflags
573          * will have TF set, in which case, continue the remaining processing
574          * of do_debug, as if this is not a probe hit.
575          */
576         if (regs->eflags & TF_MASK)
577                 return 0;
578
579         return 1;
580 }
581
582 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
583 {
584         struct kprobe *cur = kprobe_running();
585         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
586         const struct exception_table_entry *fixup;
587
588         switch(kcb->kprobe_status) {
589         case KPROBE_HIT_SS:
590         case KPROBE_REENTER:
591                 /*
592                  * We are here because the instruction being single
593                  * stepped caused a page fault. We reset the current
594                  * kprobe and the rip points back to the probe address
595                  * and allow the page fault handler to continue as a
596                  * normal page fault.
597                  */
598                 regs->rip = (unsigned long)cur->addr;
599                 regs->eflags |= kcb->kprobe_old_rflags;
600                 if (kcb->kprobe_status == KPROBE_REENTER)
601                         restore_previous_kprobe(kcb);
602                 else
603                         reset_current_kprobe();
604                 preempt_enable_no_resched();
605                 break;
606         case KPROBE_HIT_ACTIVE:
607         case KPROBE_HIT_SSDONE:
608                 /*
609                  * We increment the nmissed count for accounting,
610                  * we can also use npre/npostfault count for accouting
611                  * these specific fault cases.
612                  */
613                 kprobes_inc_nmissed_count(cur);
614
615                 /*
616                  * We come here because instructions in the pre/post
617                  * handler caused the page_fault, this could happen
618                  * if handler tries to access user space by
619                  * copy_from_user(), get_user() etc. Let the
620                  * user-specified handler try to fix it first.
621                  */
622                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
623                         return 1;
624
625                 /*
626                  * In case the user-specified fault handler returned
627                  * zero, try to fix up.
628                  */
629                 fixup = search_exception_tables(regs->rip);
630                 if (fixup) {
631                         regs->rip = fixup->fixup;
632                         return 1;
633                 }
634
635                 /*
636                  * fixup() could not handle it,
637                  * Let do_page_fault() fix it.
638                  */
639                 break;
640         default:
641                 break;
642         }
643         return 0;
644 }
645
646 /*
647  * Wrapper routine for handling exceptions.
648  */
649 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
650                                        unsigned long val, void *data)
651 {
652         struct die_args *args = (struct die_args *)data;
653         int ret = NOTIFY_DONE;
654
655         if (args->regs && user_mode(args->regs))
656                 return ret;
657
658         switch (val) {
659         case DIE_INT3:
660                 if (kprobe_handler(args->regs))
661                         ret = NOTIFY_STOP;
662                 break;
663         case DIE_DEBUG:
664                 if (post_kprobe_handler(args->regs))
665                         ret = NOTIFY_STOP;
666                 break;
667         case DIE_GPF:
668         case DIE_PAGE_FAULT:
669                 /* kprobe_running() needs smp_processor_id() */
670                 preempt_disable();
671                 if (kprobe_running() &&
672                     kprobe_fault_handler(args->regs, args->trapnr))
673                         ret = NOTIFY_STOP;
674                 preempt_enable();
675                 break;
676         default:
677                 break;
678         }
679         return ret;
680 }
681
682 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
683 {
684         struct jprobe *jp = container_of(p, struct jprobe, kp);
685         unsigned long addr;
686         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
687
688         kcb->jprobe_saved_regs = *regs;
689         kcb->jprobe_saved_rsp = (long *) regs->rsp;
690         addr = (unsigned long)(kcb->jprobe_saved_rsp);
691         /*
692          * As Linus pointed out, gcc assumes that the callee
693          * owns the argument space and could overwrite it, e.g.
694          * tailcall optimization. So, to be absolutely safe
695          * we also save and restore enough stack bytes to cover
696          * the argument area.
697          */
698         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
699                         MIN_STACK_SIZE(addr));
700         regs->eflags &= ~IF_MASK;
701         regs->rip = (unsigned long)(jp->entry);
702         return 1;
703 }
704
705 void __kprobes jprobe_return(void)
706 {
707         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
708
709         asm volatile ("       xchg   %%rbx,%%rsp     \n"
710                       "       int3                      \n"
711                       "       .globl jprobe_return_end  \n"
712                       "       jprobe_return_end:        \n"
713                       "       nop                       \n"::"b"
714                       (kcb->jprobe_saved_rsp):"memory");
715 }
716
717 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
718 {
719         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
720         u8 *addr = (u8 *) (regs->rip - 1);
721         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
722         struct jprobe *jp = container_of(p, struct jprobe, kp);
723
724         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
725                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
726                         struct pt_regs *saved_regs =
727                             container_of(kcb->jprobe_saved_rsp,
728                                             struct pt_regs, rsp);
729                         printk("current rsp %p does not match saved rsp %p\n",
730                                (long *)regs->rsp, kcb->jprobe_saved_rsp);
731                         printk("Saved registers for jprobe %p\n", jp);
732                         show_registers(saved_regs);
733                         printk("Current registers\n");
734                         show_registers(regs);
735                         BUG();
736                 }
737                 *regs = kcb->jprobe_saved_regs;
738                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
739                        MIN_STACK_SIZE(stack_addr));
740                 preempt_enable_no_resched();
741                 return 1;
742         }
743         return 0;
744 }
745
746 static struct kprobe trampoline_p = {
747         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
748         .pre_handler = trampoline_probe_handler
749 };
750
751 int __init arch_init_kprobes(void)
752 {
753         return register_kprobe(&trampoline_p);
754 }