x86: Fix alternatives and kprobes to remap write-protected kernel text
[linux-2.6.git] / arch / i386 / kernel / kprobes.c
1 /*
2  *  Kernel Probes (KProbes)
3  *  arch/i386/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  * 2005-May     Hien Nguyen <hien@us.ibm.com>, Jim Keniston
27  *              <jkenisto@us.ibm.com> and Prasanna S Panchamukhi
28  *              <prasanna@in.ibm.com> added function-return probes.
29  */
30
31 #include <linux/kprobes.h>
32 #include <linux/ptrace.h>
33 #include <linux/preempt.h>
34 #include <linux/kdebug.h>
35 #include <asm/cacheflush.h>
36 #include <asm/desc.h>
37 #include <asm/uaccess.h>
38 #include <asm/alternative.h>
39
40 void jprobe_return_end(void);
41
42 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
43 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
44
45 /* insert a jmp code */
46 static __always_inline void set_jmp_op(void *from, void *to)
47 {
48         struct __arch_jmp_op {
49                 char op;
50                 long raddr;
51         } __attribute__((packed)) *jop;
52         jop = (struct __arch_jmp_op *)from;
53         jop->raddr = (long)(to) - ((long)(from) + 5);
54         jop->op = RELATIVEJUMP_INSTRUCTION;
55 }
56
57 /*
58  * returns non-zero if opcodes can be boosted.
59  */
60 static __always_inline int can_boost(kprobe_opcode_t *opcodes)
61 {
62 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
63         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
64           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
65           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
66           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
67          << (row % 32))
68         /*
69          * Undefined/reserved opcodes, conditional jump, Opcode Extension
70          * Groups, and some special opcodes can not be boost.
71          */
72         static const unsigned long twobyte_is_boostable[256 / 32] = {
73                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
74                 /*      -------------------------------         */
75                 W(0x00, 0,0,1,1,0,0,1,0,1,1,0,0,0,0,0,0)| /* 00 */
76                 W(0x10, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 10 */
77                 W(0x20, 1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0)| /* 20 */
78                 W(0x30, 0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 30 */
79                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 40 */
80                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 50 */
81                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,0,0,1,1)| /* 60 */
82                 W(0x70, 0,0,0,0,1,1,1,1,0,0,0,0,0,0,1,1), /* 70 */
83                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 80 */
84                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1), /* 90 */
85                 W(0xa0, 1,1,0,1,1,1,0,0,1,1,0,1,1,1,0,1)| /* a0 */
86                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,0,1,1,1,1,1), /* b0 */
87                 W(0xc0, 1,1,0,0,0,0,0,0,1,1,1,1,1,1,1,1)| /* c0 */
88                 W(0xd0, 0,1,1,1,0,1,0,0,1,1,0,1,1,1,0,1), /* d0 */
89                 W(0xe0, 0,1,1,0,0,1,0,0,1,1,0,1,1,1,0,1)| /* e0 */
90                 W(0xf0, 0,1,1,1,0,1,0,0,1,1,1,0,1,1,1,0)  /* f0 */
91                 /*      -------------------------------         */
92                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
93         };
94 #undef W
95         kprobe_opcode_t opcode;
96         kprobe_opcode_t *orig_opcodes = opcodes;
97 retry:
98         if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
99                 return 0;
100         opcode = *(opcodes++);
101
102         /* 2nd-byte opcode */
103         if (opcode == 0x0f) {
104                 if (opcodes - orig_opcodes > MAX_INSN_SIZE - 1)
105                         return 0;
106                 return test_bit(*opcodes, twobyte_is_boostable);
107         }
108
109         switch (opcode & 0xf0) {
110         case 0x60:
111                 if (0x63 < opcode && opcode < 0x67)
112                         goto retry; /* prefixes */
113                 /* can't boost Address-size override and bound */
114                 return (opcode != 0x62 && opcode != 0x67);
115         case 0x70:
116                 return 0; /* can't boost conditional jump */
117         case 0xc0:
118                 /* can't boost software-interruptions */
119                 return (0xc1 < opcode && opcode < 0xcc) || opcode == 0xcf;
120         case 0xd0:
121                 /* can boost AA* and XLAT */
122                 return (opcode == 0xd4 || opcode == 0xd5 || opcode == 0xd7);
123         case 0xe0:
124                 /* can boost in/out and absolute jmps */
125                 return ((opcode & 0x04) || opcode == 0xea);
126         case 0xf0:
127                 if ((opcode & 0x0c) == 0 && opcode != 0xf1)
128                         goto retry; /* lock/rep(ne) prefix */
129                 /* clear and set flags can be boost */
130                 return (opcode == 0xf5 || (0xf7 < opcode && opcode < 0xfe));
131         default:
132                 if (opcode == 0x26 || opcode == 0x36 || opcode == 0x3e)
133                         goto retry; /* prefixes */
134                 /* can't boost CS override and call */
135                 return (opcode != 0x2e && opcode != 0x9a);
136         }
137 }
138
139 /*
140  * returns non-zero if opcode modifies the interrupt flag.
141  */
142 static int __kprobes is_IF_modifier(kprobe_opcode_t opcode)
143 {
144         switch (opcode) {
145         case 0xfa:              /* cli */
146         case 0xfb:              /* sti */
147         case 0xcf:              /* iret/iretd */
148         case 0x9d:              /* popf/popfd */
149                 return 1;
150         }
151         return 0;
152 }
153
154 int __kprobes arch_prepare_kprobe(struct kprobe *p)
155 {
156         /* insn: must be on special executable page on i386. */
157         p->ainsn.insn = get_insn_slot();
158         if (!p->ainsn.insn)
159                 return -ENOMEM;
160
161         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
162         p->opcode = *p->addr;
163         if (can_boost(p->addr)) {
164                 p->ainsn.boostable = 0;
165         } else {
166                 p->ainsn.boostable = -1;
167         }
168         return 0;
169 }
170
171 void __kprobes arch_arm_kprobe(struct kprobe *p)
172 {
173         text_poke(p->addr, ((unsigned char []){BREAKPOINT_INSTRUCTION}), 1);
174 }
175
176 void __kprobes arch_disarm_kprobe(struct kprobe *p)
177 {
178         text_poke(p->addr, &p->opcode, 1);
179 }
180
181 void __kprobes arch_remove_kprobe(struct kprobe *p)
182 {
183         mutex_lock(&kprobe_mutex);
184         free_insn_slot(p->ainsn.insn, (p->ainsn.boostable == 1));
185         mutex_unlock(&kprobe_mutex);
186 }
187
188 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
189 {
190         kcb->prev_kprobe.kp = kprobe_running();
191         kcb->prev_kprobe.status = kcb->kprobe_status;
192         kcb->prev_kprobe.old_eflags = kcb->kprobe_old_eflags;
193         kcb->prev_kprobe.saved_eflags = kcb->kprobe_saved_eflags;
194 }
195
196 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
197 {
198         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
199         kcb->kprobe_status = kcb->prev_kprobe.status;
200         kcb->kprobe_old_eflags = kcb->prev_kprobe.old_eflags;
201         kcb->kprobe_saved_eflags = kcb->prev_kprobe.saved_eflags;
202 }
203
204 static void __kprobes set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
205                                 struct kprobe_ctlblk *kcb)
206 {
207         __get_cpu_var(current_kprobe) = p;
208         kcb->kprobe_saved_eflags = kcb->kprobe_old_eflags
209                 = (regs->eflags & (TF_MASK | IF_MASK));
210         if (is_IF_modifier(p->opcode))
211                 kcb->kprobe_saved_eflags &= ~IF_MASK;
212 }
213
214 static void __kprobes prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
215 {
216         regs->eflags |= TF_MASK;
217         regs->eflags &= ~IF_MASK;
218         /*single step inline if the instruction is an int3*/
219         if (p->opcode == BREAKPOINT_INSTRUCTION)
220                 regs->eip = (unsigned long)p->addr;
221         else
222                 regs->eip = (unsigned long)p->ainsn.insn;
223 }
224
225 /* Called with kretprobe_lock held */
226 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
227                                       struct pt_regs *regs)
228 {
229         unsigned long *sara = (unsigned long *)&regs->esp;
230
231         ri->ret_addr = (kprobe_opcode_t *) *sara;
232
233         /* Replace the return addr with trampoline addr */
234         *sara = (unsigned long) &kretprobe_trampoline;
235 }
236
237 /*
238  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
239  * remain disabled thorough out this function.
240  */
241 static int __kprobes kprobe_handler(struct pt_regs *regs)
242 {
243         struct kprobe *p;
244         int ret = 0;
245         kprobe_opcode_t *addr;
246         struct kprobe_ctlblk *kcb;
247
248         addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t));
249
250         /*
251          * We don't want to be preempted for the entire
252          * duration of kprobe processing
253          */
254         preempt_disable();
255         kcb = get_kprobe_ctlblk();
256
257         /* Check we're not actually recursing */
258         if (kprobe_running()) {
259                 p = get_kprobe(addr);
260                 if (p) {
261                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
262                                 *p->ainsn.insn == BREAKPOINT_INSTRUCTION) {
263                                 regs->eflags &= ~TF_MASK;
264                                 regs->eflags |= kcb->kprobe_saved_eflags;
265                                 goto no_kprobe;
266                         }
267                         /* We have reentered the kprobe_handler(), since
268                          * another probe was hit while within the handler.
269                          * We here save the original kprobes variables and
270                          * just single step on the instruction of the new probe
271                          * without calling any user handlers.
272                          */
273                         save_previous_kprobe(kcb);
274                         set_current_kprobe(p, regs, kcb);
275                         kprobes_inc_nmissed_count(p);
276                         prepare_singlestep(p, regs);
277                         kcb->kprobe_status = KPROBE_REENTER;
278                         return 1;
279                 } else {
280                         if (*addr != BREAKPOINT_INSTRUCTION) {
281                         /* The breakpoint instruction was removed by
282                          * another cpu right after we hit, no further
283                          * handling of this interrupt is appropriate
284                          */
285                                 regs->eip -= sizeof(kprobe_opcode_t);
286                                 ret = 1;
287                                 goto no_kprobe;
288                         }
289                         p = __get_cpu_var(current_kprobe);
290                         if (p->break_handler && p->break_handler(p, regs)) {
291                                 goto ss_probe;
292                         }
293                 }
294                 goto no_kprobe;
295         }
296
297         p = get_kprobe(addr);
298         if (!p) {
299                 if (*addr != BREAKPOINT_INSTRUCTION) {
300                         /*
301                          * The breakpoint instruction was removed right
302                          * after we hit it.  Another cpu has removed
303                          * either a probepoint or a debugger breakpoint
304                          * at this address.  In either case, no further
305                          * handling of this interrupt is appropriate.
306                          * Back up over the (now missing) int3 and run
307                          * the original instruction.
308                          */
309                         regs->eip -= sizeof(kprobe_opcode_t);
310                         ret = 1;
311                 }
312                 /* Not one of ours: let kernel handle it */
313                 goto no_kprobe;
314         }
315
316         set_current_kprobe(p, regs, kcb);
317         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
318
319         if (p->pre_handler && p->pre_handler(p, regs))
320                 /* handler has already set things up, so skip ss setup */
321                 return 1;
322
323 ss_probe:
324 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PM)
325         if (p->ainsn.boostable == 1 && !p->post_handler){
326                 /* Boost up -- we can execute copied instructions directly */
327                 reset_current_kprobe();
328                 regs->eip = (unsigned long)p->ainsn.insn;
329                 preempt_enable_no_resched();
330                 return 1;
331         }
332 #endif
333         prepare_singlestep(p, regs);
334         kcb->kprobe_status = KPROBE_HIT_SS;
335         return 1;
336
337 no_kprobe:
338         preempt_enable_no_resched();
339         return ret;
340 }
341
342 /*
343  * For function-return probes, init_kprobes() establishes a probepoint
344  * here. When a retprobed function returns, this probe is hit and
345  * trampoline_probe_handler() runs, calling the kretprobe's handler.
346  */
347  void __kprobes kretprobe_trampoline_holder(void)
348  {
349         asm volatile ( ".global kretprobe_trampoline\n"
350                         "kretprobe_trampoline: \n"
351                         "       pushf\n"
352                         /* skip cs, eip, orig_eax */
353                         "       subl $12, %esp\n"
354                         "       pushl %fs\n"
355                         "       pushl %ds\n"
356                         "       pushl %es\n"
357                         "       pushl %eax\n"
358                         "       pushl %ebp\n"
359                         "       pushl %edi\n"
360                         "       pushl %esi\n"
361                         "       pushl %edx\n"
362                         "       pushl %ecx\n"
363                         "       pushl %ebx\n"
364                         "       movl %esp, %eax\n"
365                         "       call trampoline_handler\n"
366                         /* move eflags to cs */
367                         "       movl 52(%esp), %edx\n"
368                         "       movl %edx, 48(%esp)\n"
369                         /* save true return address on eflags */
370                         "       movl %eax, 52(%esp)\n"
371                         "       popl %ebx\n"
372                         "       popl %ecx\n"
373                         "       popl %edx\n"
374                         "       popl %esi\n"
375                         "       popl %edi\n"
376                         "       popl %ebp\n"
377                         "       popl %eax\n"
378                         /* skip eip, orig_eax, es, ds, fs */
379                         "       addl $20, %esp\n"
380                         "       popf\n"
381                         "       ret\n");
382 }
383
384 /*
385  * Called from kretprobe_trampoline
386  */
387 fastcall void *__kprobes trampoline_handler(struct pt_regs *regs)
388 {
389         struct kretprobe_instance *ri = NULL;
390         struct hlist_head *head, empty_rp;
391         struct hlist_node *node, *tmp;
392         unsigned long flags, orig_ret_address = 0;
393         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
394
395         INIT_HLIST_HEAD(&empty_rp);
396         spin_lock_irqsave(&kretprobe_lock, flags);
397         head = kretprobe_inst_table_head(current);
398         /* fixup registers */
399         regs->xcs = __KERNEL_CS | get_kernel_rpl();
400         regs->eip = trampoline_address;
401         regs->orig_eax = 0xffffffff;
402
403         /*
404          * It is possible to have multiple instances associated with a given
405          * task either because an multiple functions in the call path
406          * have a return probe installed on them, and/or more then one return
407          * return probe was registered for a target function.
408          *
409          * We can handle this because:
410          *     - instances are always inserted at the head of the list
411          *     - when multiple return probes are registered for the same
412          *       function, the first instance's ret_addr will point to the
413          *       real return address, and all the rest will point to
414          *       kretprobe_trampoline
415          */
416         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
417                 if (ri->task != current)
418                         /* another task is sharing our hash bucket */
419                         continue;
420
421                 if (ri->rp && ri->rp->handler){
422                         __get_cpu_var(current_kprobe) = &ri->rp->kp;
423                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
424                         ri->rp->handler(ri, regs);
425                         __get_cpu_var(current_kprobe) = NULL;
426                 }
427
428                 orig_ret_address = (unsigned long)ri->ret_addr;
429                 recycle_rp_inst(ri, &empty_rp);
430
431                 if (orig_ret_address != trampoline_address)
432                         /*
433                          * This is the real return address. Any other
434                          * instances associated with this task are for
435                          * other calls deeper on the call stack
436                          */
437                         break;
438         }
439
440         kretprobe_assert(ri, orig_ret_address, trampoline_address);
441         spin_unlock_irqrestore(&kretprobe_lock, flags);
442
443         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
444                 hlist_del(&ri->hlist);
445                 kfree(ri);
446         }
447         return (void*)orig_ret_address;
448 }
449
450 /*
451  * Called after single-stepping.  p->addr is the address of the
452  * instruction whose first byte has been replaced by the "int 3"
453  * instruction.  To avoid the SMP problems that can occur when we
454  * temporarily put back the original opcode to single-step, we
455  * single-stepped a copy of the instruction.  The address of this
456  * copy is p->ainsn.insn.
457  *
458  * This function prepares to return from the post-single-step
459  * interrupt.  We have to fix up the stack as follows:
460  *
461  * 0) Except in the case of absolute or indirect jump or call instructions,
462  * the new eip is relative to the copied instruction.  We need to make
463  * it relative to the original instruction.
464  *
465  * 1) If the single-stepped instruction was pushfl, then the TF and IF
466  * flags are set in the just-pushed eflags, and may need to be cleared.
467  *
468  * 2) If the single-stepped instruction was a call, the return address
469  * that is atop the stack is the address following the copied instruction.
470  * We need to make it the address following the original instruction.
471  *
472  * This function also checks instruction size for preparing direct execution.
473  */
474 static void __kprobes resume_execution(struct kprobe *p,
475                 struct pt_regs *regs, struct kprobe_ctlblk *kcb)
476 {
477         unsigned long *tos = (unsigned long *)&regs->esp;
478         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
479         unsigned long orig_eip = (unsigned long)p->addr;
480
481         regs->eflags &= ~TF_MASK;
482         switch (p->ainsn.insn[0]) {
483         case 0x9c:              /* pushfl */
484                 *tos &= ~(TF_MASK | IF_MASK);
485                 *tos |= kcb->kprobe_old_eflags;
486                 break;
487         case 0xc2:              /* iret/ret/lret */
488         case 0xc3:
489         case 0xca:
490         case 0xcb:
491         case 0xcf:
492         case 0xea:              /* jmp absolute -- eip is correct */
493                 /* eip is already adjusted, no more changes required */
494                 p->ainsn.boostable = 1;
495                 goto no_change;
496         case 0xe8:              /* call relative - Fix return addr */
497                 *tos = orig_eip + (*tos - copy_eip);
498                 break;
499         case 0x9a:              /* call absolute -- same as call absolute, indirect */
500                 *tos = orig_eip + (*tos - copy_eip);
501                 goto no_change;
502         case 0xff:
503                 if ((p->ainsn.insn[1] & 0x30) == 0x10) {
504                         /*
505                          * call absolute, indirect
506                          * Fix return addr; eip is correct.
507                          * But this is not boostable
508                          */
509                         *tos = orig_eip + (*tos - copy_eip);
510                         goto no_change;
511                 } else if (((p->ainsn.insn[1] & 0x31) == 0x20) ||       /* jmp near, absolute indirect */
512                            ((p->ainsn.insn[1] & 0x31) == 0x21)) {       /* jmp far, absolute indirect */
513                         /* eip is correct. And this is boostable */
514                         p->ainsn.boostable = 1;
515                         goto no_change;
516                 }
517         default:
518                 break;
519         }
520
521         if (p->ainsn.boostable == 0) {
522                 if ((regs->eip > copy_eip) &&
523                     (regs->eip - copy_eip) + 5 < MAX_INSN_SIZE) {
524                         /*
525                          * These instructions can be executed directly if it
526                          * jumps back to correct address.
527                          */
528                         set_jmp_op((void *)regs->eip,
529                                    (void *)orig_eip + (regs->eip - copy_eip));
530                         p->ainsn.boostable = 1;
531                 } else {
532                         p->ainsn.boostable = -1;
533                 }
534         }
535
536         regs->eip = orig_eip + (regs->eip - copy_eip);
537
538 no_change:
539         return;
540 }
541
542 /*
543  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
544  * remain disabled thoroughout this function.
545  */
546 static 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_eflags;
561
562         /*Restore back 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 static 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
587         switch(kcb->kprobe_status) {
588         case KPROBE_HIT_SS:
589         case KPROBE_REENTER:
590                 /*
591                  * We are here because the instruction being single
592                  * stepped caused a page fault. We reset the current
593                  * kprobe and the eip points back to the probe address
594                  * and allow the page fault handler to continue as a
595                  * normal page fault.
596                  */
597                 regs->eip = (unsigned long)cur->addr;
598                 regs->eflags |= kcb->kprobe_old_eflags;
599                 if (kcb->kprobe_status == KPROBE_REENTER)
600                         restore_previous_kprobe(kcb);
601                 else
602                         reset_current_kprobe();
603                 preempt_enable_no_resched();
604                 break;
605         case KPROBE_HIT_ACTIVE:
606         case KPROBE_HIT_SSDONE:
607                 /*
608                  * We increment the nmissed count for accounting,
609                  * we can also use npre/npostfault count for accouting
610                  * these specific fault cases.
611                  */
612                 kprobes_inc_nmissed_count(cur);
613
614                 /*
615                  * We come here because instructions in the pre/post
616                  * handler caused the page_fault, this could happen
617                  * if handler tries to access user space by
618                  * copy_from_user(), get_user() etc. Let the
619                  * user-specified handler try to fix it first.
620                  */
621                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
622                         return 1;
623
624                 /*
625                  * In case the user-specified fault handler returned
626                  * zero, try to fix up.
627                  */
628                 if (fixup_exception(regs))
629                         return 1;
630
631                 /*
632                  * fixup_exception() could not handle it,
633                  * Let do_page_fault() fix it.
634                  */
635                 break;
636         default:
637                 break;
638         }
639         return 0;
640 }
641
642 /*
643  * Wrapper routine to for handling exceptions.
644  */
645 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
646                                        unsigned long val, void *data)
647 {
648         struct die_args *args = (struct die_args *)data;
649         int ret = NOTIFY_DONE;
650
651         if (args->regs && user_mode_vm(args->regs))
652                 return ret;
653
654         switch (val) {
655         case DIE_INT3:
656                 if (kprobe_handler(args->regs))
657                         ret = NOTIFY_STOP;
658                 break;
659         case DIE_DEBUG:
660                 if (post_kprobe_handler(args->regs))
661                         ret = NOTIFY_STOP;
662                 break;
663         case DIE_GPF:
664         case DIE_PAGE_FAULT:
665                 /* kprobe_running() needs smp_processor_id() */
666                 preempt_disable();
667                 if (kprobe_running() &&
668                     kprobe_fault_handler(args->regs, args->trapnr))
669                         ret = NOTIFY_STOP;
670                 preempt_enable();
671                 break;
672         default:
673                 break;
674         }
675         return ret;
676 }
677
678 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
679 {
680         struct jprobe *jp = container_of(p, struct jprobe, kp);
681         unsigned long addr;
682         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
683
684         kcb->jprobe_saved_regs = *regs;
685         kcb->jprobe_saved_esp = &regs->esp;
686         addr = (unsigned long)(kcb->jprobe_saved_esp);
687
688         /*
689          * TBD: As Linus pointed out, gcc assumes that the callee
690          * owns the argument space and could overwrite it, e.g.
691          * tailcall optimization. So, to be absolutely safe
692          * we also save and restore enough stack bytes to cover
693          * the argument area.
694          */
695         memcpy(kcb->jprobes_stack, (kprobe_opcode_t *)addr,
696                         MIN_STACK_SIZE(addr));
697         regs->eflags &= ~IF_MASK;
698         regs->eip = (unsigned long)(jp->entry);
699         return 1;
700 }
701
702 void __kprobes jprobe_return(void)
703 {
704         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
705
706         asm volatile ("       xchgl   %%ebx,%%esp     \n"
707                       "       int3                      \n"
708                       "       .globl jprobe_return_end  \n"
709                       "       jprobe_return_end:        \n"
710                       "       nop                       \n"::"b"
711                       (kcb->jprobe_saved_esp):"memory");
712 }
713
714 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
715 {
716         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
717         u8 *addr = (u8 *) (regs->eip - 1);
718         unsigned long stack_addr = (unsigned long)(kcb->jprobe_saved_esp);
719         struct jprobe *jp = container_of(p, struct jprobe, kp);
720
721         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
722                 if (&regs->esp != kcb->jprobe_saved_esp) {
723                         struct pt_regs *saved_regs =
724                             container_of(kcb->jprobe_saved_esp,
725                                             struct pt_regs, esp);
726                         printk("current esp %p does not match saved esp %p\n",
727                                &regs->esp, kcb->jprobe_saved_esp);
728                         printk("Saved registers for jprobe %p\n", jp);
729                         show_registers(saved_regs);
730                         printk("Current registers\n");
731                         show_registers(regs);
732                         BUG();
733                 }
734                 *regs = kcb->jprobe_saved_regs;
735                 memcpy((kprobe_opcode_t *) stack_addr, kcb->jprobes_stack,
736                        MIN_STACK_SIZE(stack_addr));
737                 preempt_enable_no_resched();
738                 return 1;
739         }
740         return 0;
741 }
742
743 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
744 {
745         return 0;
746 }
747
748 int __init arch_init_kprobes(void)
749 {
750         return 0;
751 }