]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
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. | |
b94cce92 HN |
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. | |
1da177e4 LT |
29 | */ |
30 | ||
31 | #include <linux/config.h> | |
32 | #include <linux/kprobes.h> | |
33 | #include <linux/ptrace.h> | |
34 | #include <linux/spinlock.h> | |
35 | #include <linux/preempt.h> | |
36 | #include <asm/kdebug.h> | |
37 | #include <asm/desc.h> | |
38 | ||
39 | /* kprobe_status settings */ | |
40 | #define KPROBE_HIT_ACTIVE 0x00000001 | |
41 | #define KPROBE_HIT_SS 0x00000002 | |
42 | ||
43 | static struct kprobe *current_kprobe; | |
44 | static unsigned long kprobe_status, kprobe_old_eflags, kprobe_saved_eflags; | |
45 | static struct pt_regs jprobe_saved_regs; | |
46 | static long *jprobe_saved_esp; | |
47 | /* copy of the kernel stack at the probe fire time */ | |
48 | static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE]; | |
49 | void jprobe_return_end(void); | |
50 | ||
51 | /* | |
52 | * returns non-zero if opcode modifies the interrupt flag. | |
53 | */ | |
54 | static inline int is_IF_modifier(kprobe_opcode_t opcode) | |
55 | { | |
56 | switch (opcode) { | |
57 | case 0xfa: /* cli */ | |
58 | case 0xfb: /* sti */ | |
59 | case 0xcf: /* iret/iretd */ | |
60 | case 0x9d: /* popf/popfd */ | |
61 | return 1; | |
62 | } | |
63 | return 0; | |
64 | } | |
65 | ||
66 | int arch_prepare_kprobe(struct kprobe *p) | |
67 | { | |
68 | return 0; | |
69 | } | |
70 | ||
71 | void arch_copy_kprobe(struct kprobe *p) | |
72 | { | |
73 | memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE * sizeof(kprobe_opcode_t)); | |
74 | } | |
75 | ||
76 | void arch_remove_kprobe(struct kprobe *p) | |
77 | { | |
78 | } | |
79 | ||
80 | static inline void disarm_kprobe(struct kprobe *p, struct pt_regs *regs) | |
81 | { | |
82 | *p->addr = p->opcode; | |
83 | regs->eip = (unsigned long)p->addr; | |
84 | } | |
85 | ||
86 | static inline void prepare_singlestep(struct kprobe *p, struct pt_regs *regs) | |
87 | { | |
88 | regs->eflags |= TF_MASK; | |
89 | regs->eflags &= ~IF_MASK; | |
90 | /*single step inline if the instruction is an int3*/ | |
91 | if (p->opcode == BREAKPOINT_INSTRUCTION) | |
92 | regs->eip = (unsigned long)p->addr; | |
93 | else | |
94 | regs->eip = (unsigned long)&p->ainsn.insn; | |
95 | } | |
96 | ||
b94cce92 HN |
97 | struct task_struct *arch_get_kprobe_task(void *ptr) |
98 | { | |
99 | return ((struct thread_info *) (((unsigned long) ptr) & | |
100 | (~(THREAD_SIZE -1))))->task; | |
101 | } | |
102 | ||
103 | void arch_prepare_kretprobe(struct kretprobe *rp, struct pt_regs *regs) | |
104 | { | |
105 | unsigned long *sara = (unsigned long *)®s->esp; | |
106 | struct kretprobe_instance *ri; | |
107 | static void *orig_ret_addr; | |
108 | ||
109 | /* | |
110 | * Save the return address when the return probe hits | |
111 | * the first time, and use it to populate the (krprobe | |
112 | * instance)->ret_addr for subsequent return probes at | |
113 | * the same addrress since stack address would have | |
114 | * the kretprobe_trampoline by then. | |
115 | */ | |
116 | if (((void*) *sara) != kretprobe_trampoline) | |
117 | orig_ret_addr = (void*) *sara; | |
118 | ||
119 | if ((ri = get_free_rp_inst(rp)) != NULL) { | |
120 | ri->rp = rp; | |
121 | ri->stack_addr = sara; | |
122 | ri->ret_addr = orig_ret_addr; | |
123 | add_rp_inst(ri); | |
124 | /* Replace the return addr with trampoline addr */ | |
125 | *sara = (unsigned long) &kretprobe_trampoline; | |
126 | } else { | |
127 | rp->nmissed++; | |
128 | } | |
129 | } | |
130 | ||
131 | void arch_kprobe_flush_task(struct task_struct *tk, spinlock_t *kp_lock) | |
132 | { | |
133 | unsigned long flags = 0; | |
134 | struct kretprobe_instance *ri; | |
135 | spin_lock_irqsave(kp_lock, flags); | |
136 | while ((ri = get_rp_inst_tsk(tk)) != NULL) { | |
137 | *((unsigned long *)(ri->stack_addr)) = | |
138 | (unsigned long) ri->ret_addr; | |
139 | recycle_rp_inst(ri); | |
140 | } | |
141 | spin_unlock_irqrestore(kp_lock, flags); | |
142 | } | |
143 | ||
1da177e4 LT |
144 | /* |
145 | * Interrupts are disabled on entry as trap3 is an interrupt gate and they | |
146 | * remain disabled thorough out this function. | |
147 | */ | |
148 | static int kprobe_handler(struct pt_regs *regs) | |
149 | { | |
150 | struct kprobe *p; | |
151 | int ret = 0; | |
152 | kprobe_opcode_t *addr = NULL; | |
153 | unsigned long *lp; | |
154 | ||
155 | /* We're in an interrupt, but this is clear and BUG()-safe. */ | |
156 | preempt_disable(); | |
157 | /* Check if the application is using LDT entry for its code segment and | |
158 | * calculate the address by reading the base address from the LDT entry. | |
159 | */ | |
160 | if ((regs->xcs & 4) && (current->mm)) { | |
161 | lp = (unsigned long *) ((unsigned long)((regs->xcs >> 3) * 8) | |
162 | + (char *) current->mm->context.ldt); | |
163 | addr = (kprobe_opcode_t *) (get_desc_base(lp) + regs->eip - | |
164 | sizeof(kprobe_opcode_t)); | |
165 | } else { | |
166 | addr = (kprobe_opcode_t *)(regs->eip - sizeof(kprobe_opcode_t)); | |
167 | } | |
168 | /* Check we're not actually recursing */ | |
169 | if (kprobe_running()) { | |
170 | /* We *are* holding lock here, so this is safe. | |
171 | Disarm the probe we just hit, and ignore it. */ | |
172 | p = get_kprobe(addr); | |
173 | if (p) { | |
174 | if (kprobe_status == KPROBE_HIT_SS) { | |
175 | regs->eflags &= ~TF_MASK; | |
176 | regs->eflags |= kprobe_saved_eflags; | |
177 | unlock_kprobes(); | |
178 | goto no_kprobe; | |
179 | } | |
180 | disarm_kprobe(p, regs); | |
181 | ret = 1; | |
182 | } else { | |
183 | p = current_kprobe; | |
184 | if (p->break_handler && p->break_handler(p, regs)) { | |
185 | goto ss_probe; | |
186 | } | |
187 | } | |
188 | /* If it's not ours, can't be delete race, (we hold lock). */ | |
189 | goto no_kprobe; | |
190 | } | |
191 | ||
192 | lock_kprobes(); | |
193 | p = get_kprobe(addr); | |
194 | if (!p) { | |
195 | unlock_kprobes(); | |
196 | if (regs->eflags & VM_MASK) { | |
197 | /* We are in virtual-8086 mode. Return 0 */ | |
198 | goto no_kprobe; | |
199 | } | |
200 | ||
201 | if (*addr != BREAKPOINT_INSTRUCTION) { | |
202 | /* | |
203 | * The breakpoint instruction was removed right | |
204 | * after we hit it. Another cpu has removed | |
205 | * either a probepoint or a debugger breakpoint | |
206 | * at this address. In either case, no further | |
207 | * handling of this interrupt is appropriate. | |
208 | */ | |
209 | ret = 1; | |
210 | } | |
211 | /* Not one of ours: let kernel handle it */ | |
212 | goto no_kprobe; | |
213 | } | |
214 | ||
215 | kprobe_status = KPROBE_HIT_ACTIVE; | |
216 | current_kprobe = p; | |
217 | kprobe_saved_eflags = kprobe_old_eflags | |
218 | = (regs->eflags & (TF_MASK | IF_MASK)); | |
219 | if (is_IF_modifier(p->opcode)) | |
220 | kprobe_saved_eflags &= ~IF_MASK; | |
221 | ||
222 | if (p->pre_handler && p->pre_handler(p, regs)) | |
223 | /* handler has already set things up, so skip ss setup */ | |
224 | return 1; | |
225 | ||
226 | ss_probe: | |
227 | prepare_singlestep(p, regs); | |
228 | kprobe_status = KPROBE_HIT_SS; | |
229 | return 1; | |
230 | ||
231 | no_kprobe: | |
232 | preempt_enable_no_resched(); | |
233 | return ret; | |
234 | } | |
235 | ||
b94cce92 HN |
236 | /* |
237 | * For function-return probes, init_kprobes() establishes a probepoint | |
238 | * here. When a retprobed function returns, this probe is hit and | |
239 | * trampoline_probe_handler() runs, calling the kretprobe's handler. | |
240 | */ | |
241 | void kretprobe_trampoline_holder(void) | |
242 | { | |
243 | asm volatile ( ".global kretprobe_trampoline\n" | |
244 | "kretprobe_trampoline: \n" | |
245 | "nop\n"); | |
246 | } | |
247 | ||
248 | /* | |
249 | * Called when we hit the probe point at kretprobe_trampoline | |
250 | */ | |
251 | int trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs) | |
252 | { | |
253 | struct task_struct *tsk; | |
254 | struct kretprobe_instance *ri; | |
255 | struct hlist_head *head; | |
256 | struct hlist_node *node; | |
257 | unsigned long *sara = ((unsigned long *) ®s->esp) - 1; | |
258 | ||
259 | tsk = arch_get_kprobe_task(sara); | |
260 | head = kretprobe_inst_table_head(tsk); | |
261 | ||
262 | hlist_for_each_entry(ri, node, head, hlist) { | |
263 | if (ri->stack_addr == sara && ri->rp) { | |
264 | if (ri->rp->handler) | |
265 | ri->rp->handler(ri, regs); | |
266 | } | |
267 | } | |
268 | return 0; | |
269 | } | |
270 | ||
271 | void trampoline_post_handler(struct kprobe *p, struct pt_regs *regs, | |
272 | unsigned long flags) | |
273 | { | |
274 | struct kretprobe_instance *ri; | |
275 | /* RA already popped */ | |
276 | unsigned long *sara = ((unsigned long *)®s->esp) - 1; | |
277 | ||
278 | while ((ri = get_rp_inst(sara))) { | |
279 | regs->eip = (unsigned long)ri->ret_addr; | |
280 | recycle_rp_inst(ri); | |
281 | } | |
282 | regs->eflags &= ~TF_MASK; | |
283 | } | |
284 | ||
1da177e4 LT |
285 | /* |
286 | * Called after single-stepping. p->addr is the address of the | |
287 | * instruction whose first byte has been replaced by the "int 3" | |
288 | * instruction. To avoid the SMP problems that can occur when we | |
289 | * temporarily put back the original opcode to single-step, we | |
290 | * single-stepped a copy of the instruction. The address of this | |
291 | * copy is p->ainsn.insn. | |
292 | * | |
293 | * This function prepares to return from the post-single-step | |
294 | * interrupt. We have to fix up the stack as follows: | |
295 | * | |
296 | * 0) Except in the case of absolute or indirect jump or call instructions, | |
297 | * the new eip is relative to the copied instruction. We need to make | |
298 | * it relative to the original instruction. | |
299 | * | |
300 | * 1) If the single-stepped instruction was pushfl, then the TF and IF | |
301 | * flags are set in the just-pushed eflags, and may need to be cleared. | |
302 | * | |
303 | * 2) If the single-stepped instruction was a call, the return address | |
304 | * that is atop the stack is the address following the copied instruction. | |
305 | * We need to make it the address following the original instruction. | |
306 | */ | |
307 | static void resume_execution(struct kprobe *p, struct pt_regs *regs) | |
308 | { | |
309 | unsigned long *tos = (unsigned long *)®s->esp; | |
310 | unsigned long next_eip = 0; | |
311 | unsigned long copy_eip = (unsigned long)&p->ainsn.insn; | |
312 | unsigned long orig_eip = (unsigned long)p->addr; | |
313 | ||
314 | switch (p->ainsn.insn[0]) { | |
315 | case 0x9c: /* pushfl */ | |
316 | *tos &= ~(TF_MASK | IF_MASK); | |
317 | *tos |= kprobe_old_eflags; | |
318 | break; | |
0b9e2cac PP |
319 | case 0xc3: /* ret/lret */ |
320 | case 0xcb: | |
321 | case 0xc2: | |
322 | case 0xca: | |
323 | regs->eflags &= ~TF_MASK; | |
324 | /* eip is already adjusted, no more changes required*/ | |
325 | return; | |
1da177e4 LT |
326 | case 0xe8: /* call relative - Fix return addr */ |
327 | *tos = orig_eip + (*tos - copy_eip); | |
328 | break; | |
329 | case 0xff: | |
330 | if ((p->ainsn.insn[1] & 0x30) == 0x10) { | |
331 | /* call absolute, indirect */ | |
332 | /* Fix return addr; eip is correct. */ | |
333 | next_eip = regs->eip; | |
334 | *tos = orig_eip + (*tos - copy_eip); | |
335 | } else if (((p->ainsn.insn[1] & 0x31) == 0x20) || /* jmp near, absolute indirect */ | |
336 | ((p->ainsn.insn[1] & 0x31) == 0x21)) { /* jmp far, absolute indirect */ | |
337 | /* eip is correct. */ | |
338 | next_eip = regs->eip; | |
339 | } | |
340 | break; | |
341 | case 0xea: /* jmp absolute -- eip is correct */ | |
342 | next_eip = regs->eip; | |
343 | break; | |
344 | default: | |
345 | break; | |
346 | } | |
347 | ||
348 | regs->eflags &= ~TF_MASK; | |
349 | if (next_eip) { | |
350 | regs->eip = next_eip; | |
351 | } else { | |
352 | regs->eip = orig_eip + (regs->eip - copy_eip); | |
353 | } | |
354 | } | |
355 | ||
356 | /* | |
357 | * Interrupts are disabled on entry as trap1 is an interrupt gate and they | |
358 | * remain disabled thoroughout this function. And we hold kprobe lock. | |
359 | */ | |
360 | static inline int post_kprobe_handler(struct pt_regs *regs) | |
361 | { | |
362 | if (!kprobe_running()) | |
363 | return 0; | |
364 | ||
365 | if (current_kprobe->post_handler) | |
366 | current_kprobe->post_handler(current_kprobe, regs, 0); | |
367 | ||
b94cce92 HN |
368 | if (current_kprobe->post_handler != trampoline_post_handler) |
369 | resume_execution(current_kprobe, regs); | |
1da177e4 LT |
370 | regs->eflags |= kprobe_saved_eflags; |
371 | ||
372 | unlock_kprobes(); | |
373 | preempt_enable_no_resched(); | |
374 | ||
375 | /* | |
376 | * if somebody else is singlestepping across a probe point, eflags | |
377 | * will have TF set, in which case, continue the remaining processing | |
378 | * of do_debug, as if this is not a probe hit. | |
379 | */ | |
380 | if (regs->eflags & TF_MASK) | |
381 | return 0; | |
382 | ||
383 | return 1; | |
384 | } | |
385 | ||
386 | /* Interrupts disabled, kprobe_lock held. */ | |
387 | static inline int kprobe_fault_handler(struct pt_regs *regs, int trapnr) | |
388 | { | |
389 | if (current_kprobe->fault_handler | |
390 | && current_kprobe->fault_handler(current_kprobe, regs, trapnr)) | |
391 | return 1; | |
392 | ||
393 | if (kprobe_status & KPROBE_HIT_SS) { | |
394 | resume_execution(current_kprobe, regs); | |
395 | regs->eflags |= kprobe_old_eflags; | |
396 | ||
397 | unlock_kprobes(); | |
398 | preempt_enable_no_resched(); | |
399 | } | |
400 | return 0; | |
401 | } | |
402 | ||
403 | /* | |
404 | * Wrapper routine to for handling exceptions. | |
405 | */ | |
406 | int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val, | |
407 | void *data) | |
408 | { | |
409 | struct die_args *args = (struct die_args *)data; | |
410 | switch (val) { | |
411 | case DIE_INT3: | |
412 | if (kprobe_handler(args->regs)) | |
413 | return NOTIFY_STOP; | |
414 | break; | |
415 | case DIE_DEBUG: | |
416 | if (post_kprobe_handler(args->regs)) | |
417 | return NOTIFY_STOP; | |
418 | break; | |
419 | case DIE_GPF: | |
420 | if (kprobe_running() && | |
421 | kprobe_fault_handler(args->regs, args->trapnr)) | |
422 | return NOTIFY_STOP; | |
423 | break; | |
424 | case DIE_PAGE_FAULT: | |
425 | if (kprobe_running() && | |
426 | kprobe_fault_handler(args->regs, args->trapnr)) | |
427 | return NOTIFY_STOP; | |
428 | break; | |
429 | default: | |
430 | break; | |
431 | } | |
432 | return NOTIFY_DONE; | |
433 | } | |
434 | ||
435 | int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs) | |
436 | { | |
437 | struct jprobe *jp = container_of(p, struct jprobe, kp); | |
438 | unsigned long addr; | |
439 | ||
440 | jprobe_saved_regs = *regs; | |
441 | jprobe_saved_esp = ®s->esp; | |
442 | addr = (unsigned long)jprobe_saved_esp; | |
443 | ||
444 | /* | |
445 | * TBD: As Linus pointed out, gcc assumes that the callee | |
446 | * owns the argument space and could overwrite it, e.g. | |
447 | * tailcall optimization. So, to be absolutely safe | |
448 | * we also save and restore enough stack bytes to cover | |
449 | * the argument area. | |
450 | */ | |
451 | memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr)); | |
452 | regs->eflags &= ~IF_MASK; | |
453 | regs->eip = (unsigned long)(jp->entry); | |
454 | return 1; | |
455 | } | |
456 | ||
457 | void jprobe_return(void) | |
458 | { | |
459 | preempt_enable_no_resched(); | |
460 | asm volatile (" xchgl %%ebx,%%esp \n" | |
461 | " int3 \n" | |
462 | " .globl jprobe_return_end \n" | |
463 | " jprobe_return_end: \n" | |
464 | " nop \n"::"b" | |
465 | (jprobe_saved_esp):"memory"); | |
466 | } | |
467 | ||
468 | int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs) | |
469 | { | |
470 | u8 *addr = (u8 *) (regs->eip - 1); | |
471 | unsigned long stack_addr = (unsigned long)jprobe_saved_esp; | |
472 | struct jprobe *jp = container_of(p, struct jprobe, kp); | |
473 | ||
474 | if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) { | |
475 | if (®s->esp != jprobe_saved_esp) { | |
476 | struct pt_regs *saved_regs = | |
477 | container_of(jprobe_saved_esp, struct pt_regs, esp); | |
478 | printk("current esp %p does not match saved esp %p\n", | |
479 | ®s->esp, jprobe_saved_esp); | |
480 | printk("Saved registers for jprobe %p\n", jp); | |
481 | show_registers(saved_regs); | |
482 | printk("Current registers\n"); | |
483 | show_registers(regs); | |
484 | BUG(); | |
485 | } | |
486 | *regs = jprobe_saved_regs; | |
487 | memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack, | |
488 | MIN_STACK_SIZE(stack_addr)); | |
489 | return 1; | |
490 | } | |
491 | return 0; | |
492 | } |