x86: Fix alternatives and kprobes to remap write-protected kernel text
[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 #include <linux/kdebug.h>
41
42 #include <asm/pgtable.h>
43 #include <asm/uaccess.h>
44 #include <asm/alternative.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         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
213 }
214
215 void __kprobes arch_disarm_kprobe(struct kprobe *p)
216 {
217         text_poke(p->addr, &p->opcode, 1);
218 }
219
220 void __kprobes arch_remove_kprobe(struct kprobe *p)
221 {
222         mutex_lock(&kprobe_mutex);
223         free_insn_slot(p->ainsn.insn, 0);
224         mutex_unlock(&kprobe_mutex);
225 }
226
227 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
228 {
229         kcb->prev_kprobe.kp = kprobe_running();
230         kcb->prev_kprobe.status = kcb->kprobe_status;
231         kcb->prev_kprobe.old_rflags = kcb->kprobe_old_rflags;
232         kcb->prev_kprobe.saved_rflags = kcb->kprobe_saved_rflags;
233 }
234
235 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
236 {
237         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
238         kcb->kprobe_status = kcb->prev_kprobe.status;
239         kcb->kprobe_old_rflags = kcb->prev_kprobe.old_rflags;
240         kcb->kprobe_saved_rflags = kcb->prev_kprobe.saved_rflags;
241 }
242
243 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
244                                 struct kprobe_ctlblk *kcb)
245 {
246         __get_cpu_var(current_kprobe) = p;
247         kcb->kprobe_saved_rflags = kcb->kprobe_old_rflags
248                 = (regs->eflags & (TF_MASK | IF_MASK));
249         if (is_IF_modifier(p->ainsn.insn))
250                 kcb->kprobe_saved_rflags &= ~IF_MASK;
251 }
252
253 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
254 {
255         regs->eflags |= TF_MASK;
256         regs->eflags &= ~IF_MASK;
257         /*single step inline if the instruction is an int3*/
258         if (p->opcode == BREAKPOINT_INSTRUCTION)
259                 regs->rip = (unsigned long)p->addr;
260         else
261                 regs->rip = (unsigned long)p->ainsn.insn;
262 }
263
264 /* Called with kretprobe_lock held */
265 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
266                                       struct pt_regs *regs)
267 {
268         unsigned long *sara = (unsigned long *)regs->rsp;
269
270         ri->ret_addr = (kprobe_opcode_t *) *sara;
271         /* Replace the return addr with trampoline addr */
272         *sara = (unsigned long) &kretprobe_trampoline;
273 }
274
275 int __kprobes kprobe_handler(struct pt_regs *regs)
276 {
277         struct kprobe *p;
278         int ret = 0;
279         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
280         struct kprobe_ctlblk *kcb;
281
282         /*
283          * We don't want to be preempted for the entire
284          * duration of kprobe processing
285          */
286         preempt_disable();
287         kcb = get_kprobe_ctlblk();
288
289         /* Check we're not actually recursing */
290         if (kprobe_running()) {
291                 p = get_kprobe(addr);
292                 if (p) {
293                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
294                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
295                                 regs->eflags &= ~TF_MASK;
296                                 regs->eflags |= kcb->kprobe_saved_rflags;
297                                 goto no_kprobe;
298                         } else if (kcb->kprobe_status == KPROBE_HIT_SSDONE) {
299                                 /* TODO: Provide re-entrancy from
300                                  * post_kprobes_handler() and avoid exception
301                                  * stack corruption while single-stepping on
302                                  * the instruction of the new probe.
303                                  */
304                                 arch_disarm_kprobe(p);
305                                 regs->rip = (unsigned long)p->addr;
306                                 reset_current_kprobe();
307                                 ret = 1;
308                         } else {
309                                 /* We have reentered the kprobe_handler(), since
310                                  * another probe was hit while within the
311                                  * handler. We here save the original kprobe
312                                  * variables and just single step on instruction
313                                  * of the new probe without calling any user
314                                  * handlers.
315                                  */
316                                 save_previous_kprobe(kcb);
317                                 set_current_kprobe(p, regs, kcb);
318                                 kprobes_inc_nmissed_count(p);
319                                 prepare_singlestep(p, regs);
320                                 kcb->kprobe_status = KPROBE_REENTER;
321                                 return 1;
322                         }
323                 } else {
324                         if (*addr != BREAKPOINT_INSTRUCTION) {
325                         /* The breakpoint instruction was removed by
326                          * another cpu right after we hit, no further
327                          * handling of this interrupt is appropriate
328                          */
329                                 regs->rip = (unsigned long)addr;
330                                 ret = 1;
331                                 goto no_kprobe;
332                         }
333                         p = __get_cpu_var(current_kprobe);
334                         if (p->break_handler && p->break_handler(p, regs)) {
335                                 goto ss_probe;
336                         }
337                 }
338                 goto no_kprobe;
339         }
340
341         p = get_kprobe(addr);
342         if (!p) {
343                 if (*addr != BREAKPOINT_INSTRUCTION) {
344                         /*
345                          * The breakpoint instruction was removed right
346                          * after we hit it.  Another cpu has removed
347                          * either a probepoint or a debugger breakpoint
348                          * at this address.  In either case, no further
349                          * handling of this interrupt is appropriate.
350                          * Back up over the (now missing) int3 and run
351                          * the original instruction.
352                          */
353                         regs->rip = (unsigned long)addr;
354                         ret = 1;
355                 }
356                 /* Not one of ours: let kernel handle it */
357                 goto no_kprobe;
358         }
359
360         set_current_kprobe(p, regs, kcb);
361         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
362
363         if (p->pre_handler && p->pre_handler(p, regs))
364                 /* handler has already set things up, so skip ss setup */
365                 return 1;
366
367 ss_probe:
368         prepare_singlestep(p, regs);
369         kcb->kprobe_status = KPROBE_HIT_SS;
370         return 1;
371
372 no_kprobe:
373         preempt_enable_no_resched();
374         return ret;
375 }
376
377 /*
378  * For function-return probes, init_kprobes() establishes a probepoint
379  * here. When a retprobed function returns, this probe is hit and
380  * trampoline_probe_handler() runs, calling the kretprobe's handler.
381  */
382  void kretprobe_trampoline_holder(void)
383  {
384         asm volatile (  ".global kretprobe_trampoline\n"
385                         "kretprobe_trampoline: \n"
386                         "nop\n");
387  }
388
389 /*
390  * Called when we hit the probe point at kretprobe_trampoline
391  */
392 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
393 {
394         struct kretprobe_instance *ri = NULL;
395         struct hlist_head *head, empty_rp;
396         struct hlist_node *node, *tmp;
397         unsigned long flags, orig_ret_address = 0;
398         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
399
400         INIT_HLIST_HEAD(&empty_rp);
401         spin_lock_irqsave(&kretprobe_lock, flags);
402         head = kretprobe_inst_table_head(current);
403
404         /*
405          * It is possible to have multiple instances associated with a given
406          * task either because an multiple functions in the call path
407          * have a return probe installed on them, and/or more then one return
408          * return probe was registered for a target function.
409          *
410          * We can handle this because:
411          *     - instances are always inserted at the head of the list
412          *     - when multiple return probes are registered for the same
413          *       function, the first instance's ret_addr will point to the
414          *       real return address, and all the rest will point to
415          *       kretprobe_trampoline
416          */
417         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
418                 if (ri->task != current)
419                         /* another task is sharing our hash bucket */
420                         continue;
421
422                 if (ri->rp && ri->rp->handler)
423                         ri->rp->handler(ri, regs);
424
425                 orig_ret_address = (unsigned long)ri->ret_addr;
426                 recycle_rp_inst(ri, &empty_rp);
427
428                 if (orig_ret_address != trampoline_address)
429                         /*
430                          * This is the real return address. Any other
431                          * instances associated with this task are for
432                          * other calls deeper on the call stack
433                          */
434                         break;
435         }
436
437         kretprobe_assert(ri, orig_ret_address, trampoline_address);
438         regs->rip = orig_ret_address;
439
440         reset_current_kprobe();
441         spin_unlock_irqrestore(&kretprobe_lock, flags);
442         preempt_enable_no_resched();
443
444         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
445                 hlist_del(&ri->hlist);
446                 kfree(ri);
447         }
448         /*
449          * By returning a non-zero value, we are telling
450          * kprobe_handler() that we don't want the post_handler
451          * to run (and have re-enabled preemption)
452          */
453         return 1;
454 }
455
456 /*
457  * Called after single-stepping.  p->addr is the address of the
458  * instruction whose first byte has been replaced by the "int 3"
459  * instruction.  To avoid the SMP problems that can occur when we
460  * temporarily put back the original opcode to single-step, we
461  * single-stepped a copy of the instruction.  The address of this
462  * copy is p->ainsn.insn.
463  *
464  * This function prepares to return from the post-single-step
465  * interrupt.  We have to fix up the stack as follows:
466  *
467  * 0) Except in the case of absolute or indirect jump or call instructions,
468  * the new rip is relative to the copied instruction.  We need to make
469  * it relative to the original instruction.
470  *
471  * 1) If the single-stepped instruction was pushfl, then the TF and IF
472  * flags are set in the just-pushed eflags, and may need to be cleared.
473  *
474  * 2) If the single-stepped instruction was a call, the return address
475  * that is atop the stack is the address following the copied instruction.
476  * We need to make it the address following the original instruction.
477  */
478 static void __kprobes resume_execution(struct kprobe *p,
479                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
480 {
481         unsigned long *tos = (unsigned long *)regs->rsp;
482         unsigned long next_rip = 0;
483         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
484         unsigned long orig_rip = (unsigned long)p->addr;
485         kprobe_opcode_t *insn = p->ainsn.insn;
486
487         /*skip the REX prefix*/
488         if (*insn >= 0x40 && *insn <= 0x4f)
489                 insn++;
490
491         switch (*insn) {
492         case 0x9c:              /* pushfl */
493                 *tos &= ~(TF_MASK | IF_MASK);
494                 *tos |= kcb->kprobe_old_rflags;
495                 break;
496         case 0xc3:              /* ret/lret */
497         case 0xcb:
498         case 0xc2:
499         case 0xca:
500                 regs->eflags &= ~TF_MASK;
501                 /* rip is already adjusted, no more changes required*/
502                 return;
503         case 0xe8:              /* call relative - Fix return addr */
504                 *tos = orig_rip + (*tos - copy_rip);
505                 break;
506         case 0xff:
507                 if ((insn[1] & 0x30) == 0x10) {
508                         /* call absolute, indirect */
509                         /* Fix return addr; rip is correct. */
510                         next_rip = regs->rip;
511                         *tos = orig_rip + (*tos - copy_rip);
512                 } else if (((insn[1] & 0x31) == 0x20) ||        /* jmp near, absolute indirect */
513                            ((insn[1] & 0x31) == 0x21)) {        /* jmp far, absolute indirect */
514                         /* rip is correct. */
515                         next_rip = regs->rip;
516                 }
517                 break;
518         case 0xea:              /* jmp absolute -- rip is correct */
519                 next_rip = regs->rip;
520                 break;
521         default:
522                 break;
523         }
524
525         regs->eflags &= ~TF_MASK;
526         if (next_rip) {
527                 regs->rip = next_rip;
528         } else {
529                 regs->rip = orig_rip + (regs->rip - copy_rip);
530         }
531 }
532
533 int __kprobes post_kprobe_handler(struct pt_regs *regs)
534 {
535         struct kprobe *cur = kprobe_running();
536         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
537
538         if (!cur)
539                 return 0;
540
541         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
542                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
543                 cur->post_handler(cur, regs, 0);
544         }
545
546         resume_execution(cur, regs, kcb);
547         regs->eflags |= kcb->kprobe_saved_rflags;
548
549         /* Restore the original saved kprobes variables and continue. */
550         if (kcb->kprobe_status == KPROBE_REENTER) {
551                 restore_previous_kprobe(kcb);
552                 goto out;
553         }
554         reset_current_kprobe();
555 out:
556         preempt_enable_no_resched();
557
558         /*
559          * if somebody else is singlestepping across a probe point, eflags
560          * will have TF set, in which case, continue the remaining processing
561          * of do_debug, as if this is not a probe hit.
562          */
563         if (regs->eflags & TF_MASK)
564                 return 0;
565
566         return 1;
567 }
568
569 int __kprobes kprobe_fault_handler(struct pt_regs *regs, int trapnr)
570 {
571         struct kprobe *cur = kprobe_running();
572         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
573         const struct exception_table_entry *fixup;
574
575         switch(kcb->kprobe_status) {
576         case KPROBE_HIT_SS:
577         case KPROBE_REENTER:
578                 /*
579                  * We are here because the instruction being single
580                  * stepped caused a page fault. We reset the current
581                  * kprobe and the rip points back to the probe address
582                  * and allow the page fault handler to continue as a
583                  * normal page fault.
584                  */
585                 regs->rip = (unsigned long)cur->addr;
586                 regs->eflags |= kcb->kprobe_old_rflags;
587                 if (kcb->kprobe_status == KPROBE_REENTER)
588                         restore_previous_kprobe(kcb);
589                 else
590                         reset_current_kprobe();
591                 preempt_enable_no_resched();
592                 break;
593         case KPROBE_HIT_ACTIVE:
594         case KPROBE_HIT_SSDONE:
595                 /*
596                  * We increment the nmissed count for accounting,
597                  * we can also use npre/npostfault count for accouting
598                  * these specific fault cases.
599                  */
600                 kprobes_inc_nmissed_count(cur);
601
602                 /*
603                  * We come here because instructions in the pre/post
604                  * handler caused the page_fault, this could happen
605                  * if handler tries to access user space by
606                  * copy_from_user(), get_user() etc. Let the
607                  * user-specified handler try to fix it first.
608                  */
609                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
610                         return 1;
611
612                 /*
613                  * In case the user-specified fault handler returned
614                  * zero, try to fix up.
615                  */
616                 fixup = search_exception_tables(regs->rip);
617                 if (fixup) {
618                         regs->rip = fixup->fixup;
619                         return 1;
620                 }
621
622                 /*
623                  * fixup() could not handle it,
624                  * Let do_page_fault() fix it.
625                  */
626                 break;
627         default:
628                 break;
629         }
630         return 0;
631 }
632
633 /*
634  * Wrapper routine for handling exceptions.
635  */
636 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
637                                        unsigned long val, void *data)
638 {
639         struct die_args *args = (struct die_args *)data;
640         int ret = NOTIFY_DONE;
641
642         if (args->regs && user_mode(args->regs))
643                 return ret;
644
645         switch (val) {
646         case DIE_INT3:
647                 if (kprobe_handler(args->regs))
648                         ret = NOTIFY_STOP;
649                 break;
650         case DIE_DEBUG:
651                 if (post_kprobe_handler(args->regs))
652                         ret = NOTIFY_STOP;
653                 break;
654         case DIE_GPF:
655         case DIE_PAGE_FAULT:
656                 /* kprobe_running() needs smp_processor_id() */
657                 preempt_disable();
658                 if (kprobe_running() &&
659                     kprobe_fault_handler(args->regs, args->trapnr))
660                         ret = NOTIFY_STOP;
661                 preempt_enable();
662                 break;
663         default:
664                 break;
665         }
666         return ret;
667 }
668
669 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
670 {
671         struct jprobe *jp = container_of(p, struct jprobe, kp);
672         unsigned long addr;
673         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
674
675         kcb->jprobe_saved_regs = *regs;
676         kcb->jprobe_saved_rsp = (long *) regs->rsp;
677         addr = (unsigned long)(kcb->jprobe_saved_rsp);
678         /*
679          * As Linus pointed out, gcc assumes that the callee
680          * owns the argument space and could overwrite it, e.g.
681          * tailcall optimization. So, to be absolutely safe
682          * we also save and restore enough stack bytes to cover
683          * the argument area.
684          */
685         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
686                         MIN_STACK_SIZE(addr));
687         regs->eflags &= ~IF_MASK;
688         regs->rip = (unsigned long)(jp->entry);
689         return 1;
690 }
691
692 void __kprobes jprobe_return(void)
693 {
694         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
695
696         asm volatile ("       xchg   %%rbx,%%rsp     \n"
697                       "       int3                      \n"
698                       "       .globl jprobe_return_end  \n"
699                       "       jprobe_return_end:        \n"
700                       "       nop                       \n"::"b"
701                       (kcb->jprobe_saved_rsp):"memory");
702 }
703
704 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
705 {
706         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
707         u8 *addr = (u8 *) (regs->rip - 1);
708         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_rsp);
709         struct jprobe *jp = container_of(p, struct jprobe, kp);
710
711         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
712                 if ((long *)regs->rsp != kcb->jprobe_saved_rsp) {
713                         struct pt_regs *saved_regs =
714                             container_of(kcb->jprobe_saved_rsp,
715                                             struct pt_regs, rsp);
716                         printk("current rsp %p does not match saved rsp %p\n",
717                                (long *)regs->rsp, kcb->jprobe_saved_rsp);
718                         printk("Saved registers for jprobe %p\n", jp);
719                         show_registers(saved_regs);
720                         printk("Current registers\n");
721                         show_registers(regs);
722                         BUG();
723                 }
724                 *regs = kcb->jprobe_saved_regs;
725                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
726                        MIN_STACK_SIZE(stack_addr));
727                 preempt_enable_no_resched();
728                 return 1;
729         }
730         return 0;
731 }
732
733 static struct kprobe trampoline_p = {
734         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
735         .pre_handler = trampoline_probe_handler
736 };
737
738 int __init arch_init_kprobes(void)
739 {
740         return register_kprobe(&trampoline_p);
741 }
742
743 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
744 {
745         if (p->addr == (kprobe_opcode_t *)&kretprobe_trampoline)
746                 return 1;
747
748         return 0;
749 }