ppc64: actually add sys_splice() to the system call table
[linux-2.6] / arch / powerpc / kernel / kprobes.c
1 /*
2  *  Kernel Probes (KProbes)
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17  *
18  * Copyright (C) IBM Corporation, 2002, 2004
19  *
20  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
21  *              Probes initial implementation ( includes contributions from
22  *              Rusty Russell).
23  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
24  *              interface to access function arguments.
25  * 2004-Nov     Ananth N Mavinakayanahalli <ananth@in.ibm.com> kprobes port
26  *              for PPC64
27  */
28
29 #include <linux/config.h>
30 #include <linux/kprobes.h>
31 #include <linux/ptrace.h>
32 #include <linux/preempt.h>
33 #include <linux/module.h>
34 #include <asm/cacheflush.h>
35 #include <asm/kdebug.h>
36 #include <asm/sstep.h>
37 #include <asm/uaccess.h>
38
39 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
40 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
41
42 int __kprobes arch_prepare_kprobe(struct kprobe *p)
43 {
44         int ret = 0;
45         kprobe_opcode_t insn = *p->addr;
46
47         if ((unsigned long)p->addr & 0x03) {
48                 printk("Attempt to register kprobe at an unaligned address\n");
49                 ret = -EINVAL;
50         } else if (IS_MTMSRD(insn) || IS_RFID(insn)) {
51                 printk("Cannot register a kprobe on rfid or mtmsrd\n");
52                 ret = -EINVAL;
53         }
54
55         /* insn must be on a special executable page on ppc64 */
56         if (!ret) {
57                 p->ainsn.insn = get_insn_slot();
58                 if (!p->ainsn.insn)
59                         ret = -ENOMEM;
60         }
61
62         if (!ret) {
63                 memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t));
64                 p->opcode = *p->addr;
65         }
66
67         return ret;
68 }
69
70 void __kprobes arch_arm_kprobe(struct kprobe *p)
71 {
72         *p->addr = BREAKPOINT_INSTRUCTION;
73         flush_icache_range((unsigned long) p->addr,
74                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
75 }
76
77 void __kprobes arch_disarm_kprobe(struct kprobe *p)
78 {
79         *p->addr = p->opcode;
80         flush_icache_range((unsigned long) p->addr,
81                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
82 }
83
84 void __kprobes arch_remove_kprobe(struct kprobe *p)
85 {
86         mutex_lock(&kprobe_mutex);
87         free_insn_slot(p->ainsn.insn);
88         mutex_unlock(&kprobe_mutex);
89 }
90
91 static inline void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
92 {
93         kprobe_opcode_t insn = *p->ainsn.insn;
94
95         regs->msr |= MSR_SE;
96
97         /* single step inline if it is a trap variant */
98         if (is_trap(insn))
99                 regs->nip = (unsigned long)p->addr;
100         else
101                 regs->nip = (unsigned long)p->ainsn.insn;
102 }
103
104 static inline void save_previous_kprobe(struct kprobe_ctlblk *kcb)
105 {
106         kcb->prev_kprobe.kp = kprobe_running();
107         kcb->prev_kprobe.status = kcb->kprobe_status;
108         kcb->prev_kprobe.saved_msr = kcb->kprobe_saved_msr;
109 }
110
111 static inline void restore_previous_kprobe(struct kprobe_ctlblk *kcb)
112 {
113         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
114         kcb->kprobe_status = kcb->prev_kprobe.status;
115         kcb->kprobe_saved_msr = kcb->prev_kprobe.saved_msr;
116 }
117
118 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs,
119                                 struct kprobe_ctlblk *kcb)
120 {
121         __get_cpu_var(current_kprobe) = p;
122         kcb->kprobe_saved_msr = regs->msr;
123 }
124
125 /* Called with kretprobe_lock held */
126 void __kprobes arch_prepare_kretprobe(struct kretprobe *rp,
127                                       struct pt_regs *regs)
128 {
129         struct kretprobe_instance *ri;
130
131         if ((ri = get_free_rp_inst(rp)) != NULL) {
132                 ri->rp = rp;
133                 ri->task = current;
134                 ri->ret_addr = (kprobe_opcode_t *)regs->link;
135
136                 /* Replace the return addr with trampoline addr */
137                 regs->link = (unsigned long)kretprobe_trampoline;
138                 add_rp_inst(ri);
139         } else {
140                 rp->nmissed++;
141         }
142 }
143
144 static inline int kprobe_handler(struct pt_regs *regs)
145 {
146         struct kprobe *p;
147         int ret = 0;
148         unsigned int *addr = (unsigned int *)regs->nip;
149         struct kprobe_ctlblk *kcb;
150
151         /*
152          * We don't want to be preempted for the entire
153          * duration of kprobe processing
154          */
155         preempt_disable();
156         kcb = get_kprobe_ctlblk();
157
158         /* Check we're not actually recursing */
159         if (kprobe_running()) {
160                 p = get_kprobe(addr);
161                 if (p) {
162                         kprobe_opcode_t insn = *p->ainsn.insn;
163                         if (kcb->kprobe_status == KPROBE_HIT_SS &&
164                                         is_trap(insn)) {
165                                 regs->msr &= ~MSR_SE;
166                                 regs->msr |= kcb->kprobe_saved_msr;
167                                 goto no_kprobe;
168                         }
169                         /* We have reentered the kprobe_handler(), since
170                          * another probe was hit while within the handler.
171                          * We here save the original kprobes variables and
172                          * just single step on the instruction of the new probe
173                          * without calling any user handlers.
174                          */
175                         save_previous_kprobe(kcb);
176                         set_current_kprobe(p, regs, kcb);
177                         kcb->kprobe_saved_msr = regs->msr;
178                         kprobes_inc_nmissed_count(p);
179                         prepare_singlestep(p, regs);
180                         kcb->kprobe_status = KPROBE_REENTER;
181                         return 1;
182                 } else {
183                         if (*addr != BREAKPOINT_INSTRUCTION) {
184                                 /* If trap variant, then it belongs not to us */
185                                 kprobe_opcode_t cur_insn = *addr;
186                                 if (is_trap(cur_insn))
187                                         goto no_kprobe;
188                                 /* The breakpoint instruction was removed by
189                                  * another cpu right after we hit, no further
190                                  * handling of this interrupt is appropriate
191                                  */
192                                 ret = 1;
193                                 goto no_kprobe;
194                         }
195                         p = __get_cpu_var(current_kprobe);
196                         if (p->break_handler && p->break_handler(p, regs)) {
197                                 goto ss_probe;
198                         }
199                 }
200                 goto no_kprobe;
201         }
202
203         p = get_kprobe(addr);
204         if (!p) {
205                 if (*addr != BREAKPOINT_INSTRUCTION) {
206                         /*
207                          * PowerPC has multiple variants of the "trap"
208                          * instruction. If the current instruction is a
209                          * trap variant, it could belong to someone else
210                          */
211                         kprobe_opcode_t cur_insn = *addr;
212                         if (is_trap(cur_insn))
213                                 goto no_kprobe;
214                         /*
215                          * The breakpoint instruction was removed right
216                          * after we hit it.  Another cpu has removed
217                          * either a probepoint or a debugger breakpoint
218                          * at this address.  In either case, no further
219                          * handling of this interrupt is appropriate.
220                          */
221                         ret = 1;
222                 }
223                 /* Not one of ours: let kernel handle it */
224                 goto no_kprobe;
225         }
226
227         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
228         set_current_kprobe(p, regs, kcb);
229         if (p->pre_handler && p->pre_handler(p, regs))
230                 /* handler has already set things up, so skip ss setup */
231                 return 1;
232
233 ss_probe:
234         prepare_singlestep(p, regs);
235         kcb->kprobe_status = KPROBE_HIT_SS;
236         return 1;
237
238 no_kprobe:
239         preempt_enable_no_resched();
240         return ret;
241 }
242
243 /*
244  * Function return probe trampoline:
245  *      - init_kprobes() establishes a probepoint here
246  *      - When the probed function returns, this probe
247  *              causes the handlers to fire
248  */
249 void kretprobe_trampoline_holder(void)
250 {
251         asm volatile(".global kretprobe_trampoline\n"
252                         "kretprobe_trampoline:\n"
253                         "nop\n");
254 }
255
256 /*
257  * Called when the probe at kretprobe trampoline is hit
258  */
259 int __kprobes trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
260 {
261         struct kretprobe_instance *ri = NULL;
262         struct hlist_head *head;
263         struct hlist_node *node, *tmp;
264         unsigned long flags, orig_ret_address = 0;
265         unsigned long trampoline_address =(unsigned long)&kretprobe_trampoline;
266
267         spin_lock_irqsave(&kretprobe_lock, flags);
268         head = kretprobe_inst_table_head(current);
269
270         /*
271          * It is possible to have multiple instances associated with a given
272          * task either because an multiple functions in the call path
273          * have a return probe installed on them, and/or more then one return
274          * return probe was registered for a target function.
275          *
276          * We can handle this because:
277          *     - instances are always inserted at the head of the list
278          *     - when multiple return probes are registered for the same
279          *       function, the first instance's ret_addr will point to the
280          *       real return address, and all the rest will point to
281          *       kretprobe_trampoline
282          */
283         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
284                 if (ri->task != current)
285                         /* another task is sharing our hash bucket */
286                         continue;
287
288                 if (ri->rp && ri->rp->handler)
289                         ri->rp->handler(ri, regs);
290
291                 orig_ret_address = (unsigned long)ri->ret_addr;
292                 recycle_rp_inst(ri);
293
294                 if (orig_ret_address != trampoline_address)
295                         /*
296                          * This is the real return address. Any other
297                          * instances associated with this task are for
298                          * other calls deeper on the call stack
299                          */
300                         break;
301         }
302
303         BUG_ON(!orig_ret_address || (orig_ret_address == trampoline_address));
304         regs->nip = orig_ret_address;
305
306         reset_current_kprobe();
307         spin_unlock_irqrestore(&kretprobe_lock, flags);
308         preempt_enable_no_resched();
309
310         /*
311          * By returning a non-zero value, we are telling
312          * kprobe_handler() that we don't want the post_handler
313          * to run (and have re-enabled preemption)
314          */
315         return 1;
316 }
317
318 /*
319  * Called after single-stepping.  p->addr is the address of the
320  * instruction whose first byte has been replaced by the "breakpoint"
321  * instruction.  To avoid the SMP problems that can occur when we
322  * temporarily put back the original opcode to single-step, we
323  * single-stepped a copy of the instruction.  The address of this
324  * copy is p->ainsn.insn.
325  */
326 static void __kprobes resume_execution(struct kprobe *p, struct pt_regs *regs)
327 {
328         int ret;
329         unsigned int insn = *p->ainsn.insn;
330
331         regs->nip = (unsigned long)p->addr;
332         ret = emulate_step(regs, insn);
333         if (ret == 0)
334                 regs->nip = (unsigned long)p->addr + 4;
335 }
336
337 static inline int post_kprobe_handler(struct pt_regs *regs)
338 {
339         struct kprobe *cur = kprobe_running();
340         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
341
342         if (!cur)
343                 return 0;
344
345         if ((kcb->kprobe_status != KPROBE_REENTER) && cur->post_handler) {
346                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
347                 cur->post_handler(cur, regs, 0);
348         }
349
350         resume_execution(cur, regs);
351         regs->msr |= kcb->kprobe_saved_msr;
352
353         /*Restore back the original saved kprobes variables and continue. */
354         if (kcb->kprobe_status == KPROBE_REENTER) {
355                 restore_previous_kprobe(kcb);
356                 goto out;
357         }
358         reset_current_kprobe();
359 out:
360         preempt_enable_no_resched();
361
362         /*
363          * if somebody else is singlestepping across a probe point, msr
364          * will have SE set, in which case, continue the remaining processing
365          * of do_debug, as if this is not a probe hit.
366          */
367         if (regs->msr & MSR_SE)
368                 return 0;
369
370         return 1;
371 }
372
373 static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
374 {
375         struct kprobe *cur = kprobe_running();
376         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
377         const struct exception_table_entry *entry;
378
379         switch(kcb->kprobe_status) {
380         case KPROBE_HIT_SS:
381         case KPROBE_REENTER:
382                 /*
383                  * We are here because the instruction being single
384                  * stepped caused a page fault. We reset the current
385                  * kprobe and the nip points back to the probe address
386                  * and allow the page fault handler to continue as a
387                  * normal page fault.
388                  */
389                 regs->nip = (unsigned long)cur->addr;
390                 regs->msr &= ~MSR_SE;
391                 regs->msr |= kcb->kprobe_saved_msr;
392                 if (kcb->kprobe_status == KPROBE_REENTER)
393                         restore_previous_kprobe(kcb);
394                 else
395                         reset_current_kprobe();
396                 preempt_enable_no_resched();
397                 break;
398         case KPROBE_HIT_ACTIVE:
399         case KPROBE_HIT_SSDONE:
400                 /*
401                  * We increment the nmissed count for accounting,
402                  * we can also use npre/npostfault count for accouting
403                  * these specific fault cases.
404                  */
405                 kprobes_inc_nmissed_count(cur);
406
407                 /*
408                  * We come here because instructions in the pre/post
409                  * handler caused the page_fault, this could happen
410                  * if handler tries to access user space by
411                  * copy_from_user(), get_user() etc. Let the
412                  * user-specified handler try to fix it first.
413                  */
414                 if (cur->fault_handler && cur->fault_handler(cur, regs, trapnr))
415                         return 1;
416
417                 /*
418                  * In case the user-specified fault handler returned
419                  * zero, try to fix up.
420                  */
421                 if ((entry = search_exception_tables(regs->nip)) != NULL) {
422                         regs->nip = entry->fixup;
423                         return 1;
424                 }
425
426                 /*
427                  * fixup_exception() could not handle it,
428                  * Let do_page_fault() fix it.
429                  */
430                 break;
431         default:
432                 break;
433         }
434         return 0;
435 }
436
437 /*
438  * Wrapper routine to for handling exceptions.
439  */
440 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
441                                        unsigned long val, void *data)
442 {
443         struct die_args *args = (struct die_args *)data;
444         int ret = NOTIFY_DONE;
445
446         if (args->regs && user_mode(args->regs))
447                 return ret;
448
449         switch (val) {
450         case DIE_BPT:
451                 if (kprobe_handler(args->regs))
452                         ret = NOTIFY_STOP;
453                 break;
454         case DIE_SSTEP:
455                 if (post_kprobe_handler(args->regs))
456                         ret = NOTIFY_STOP;
457                 break;
458         case DIE_PAGE_FAULT:
459                 /* kprobe_running() needs smp_processor_id() */
460                 preempt_disable();
461                 if (kprobe_running() &&
462                     kprobe_fault_handler(args->regs, args->trapnr))
463                         ret = NOTIFY_STOP;
464                 preempt_enable();
465                 break;
466         default:
467                 break;
468         }
469         return ret;
470 }
471
472 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
473 {
474         struct jprobe *jp = container_of(p, struct jprobe, kp);
475         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
476
477         memcpy(&kcb->jprobe_saved_regs, regs, sizeof(struct pt_regs));
478
479         /* setup return addr to the jprobe handler routine */
480         regs->nip = (unsigned long)(((func_descr_t *)jp->entry)->entry);
481         regs->gpr[2] = (unsigned long)(((func_descr_t *)jp->entry)->toc);
482
483         return 1;
484 }
485
486 void __kprobes jprobe_return(void)
487 {
488         asm volatile("trap" ::: "memory");
489 }
490
491 void __kprobes jprobe_return_end(void)
492 {
493 };
494
495 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
496 {
497         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
498
499         /*
500          * FIXME - we should ideally be validating that we got here 'cos
501          * of the "trap" in jprobe_return() above, before restoring the
502          * saved regs...
503          */
504         memcpy(regs, &kcb->jprobe_saved_regs, sizeof(struct pt_regs));
505         preempt_enable_no_resched();
506         return 1;
507 }
508
509 static struct kprobe trampoline_p = {
510         .addr = (kprobe_opcode_t *) &kretprobe_trampoline,
511         .pre_handler = trampoline_probe_handler
512 };
513
514 int __init arch_init_kprobes(void)
515 {
516         return register_kprobe(&trampoline_p);
517 }