xref: /openbmc/linux/arch/x86/kernel/kprobes/opt.c (revision dfd4f649)
1 /*
2  *  Kernel Probes Jump Optimization (Optprobes)
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  * Copyright (C) Hitachi Ltd., 2012
20  */
21 #include <linux/kprobes.h>
22 #include <linux/ptrace.h>
23 #include <linux/string.h>
24 #include <linux/slab.h>
25 #include <linux/hardirq.h>
26 #include <linux/preempt.h>
27 #include <linux/extable.h>
28 #include <linux/kdebug.h>
29 #include <linux/kallsyms.h>
30 #include <linux/ftrace.h>
31 #include <linux/frame.h>
32 
33 #include <asm/text-patching.h>
34 #include <asm/cacheflush.h>
35 #include <asm/desc.h>
36 #include <asm/pgtable.h>
37 #include <linux/uaccess.h>
38 #include <asm/alternative.h>
39 #include <asm/insn.h>
40 #include <asm/debugreg.h>
41 #include <asm/set_memory.h>
42 #include <asm/sections.h>
43 #include <asm/nospec-branch.h>
44 
45 #include "common.h"
46 
47 unsigned long __recover_optprobed_insn(kprobe_opcode_t *buf, unsigned long addr)
48 {
49 	struct optimized_kprobe *op;
50 	struct kprobe *kp;
51 	long offs;
52 	int i;
53 
54 	for (i = 0; i < RELATIVEJUMP_SIZE; i++) {
55 		kp = get_kprobe((void *)addr - i);
56 		/* This function only handles jump-optimized kprobe */
57 		if (kp && kprobe_optimized(kp)) {
58 			op = container_of(kp, struct optimized_kprobe, kp);
59 			/* If op->list is not empty, op is under optimizing */
60 			if (list_empty(&op->list))
61 				goto found;
62 		}
63 	}
64 
65 	return addr;
66 found:
67 	/*
68 	 * If the kprobe can be optimized, original bytes which can be
69 	 * overwritten by jump destination address. In this case, original
70 	 * bytes must be recovered from op->optinsn.copied_insn buffer.
71 	 */
72 	if (probe_kernel_read(buf, (void *)addr,
73 		MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
74 		return 0UL;
75 
76 	if (addr == (unsigned long)kp->addr) {
77 		buf[0] = kp->opcode;
78 		memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
79 	} else {
80 		offs = addr - (unsigned long)kp->addr - 1;
81 		memcpy(buf, op->optinsn.copied_insn + offs, RELATIVE_ADDR_SIZE - offs);
82 	}
83 
84 	return (unsigned long)buf;
85 }
86 
87 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
88 static void synthesize_set_arg1(kprobe_opcode_t *addr, unsigned long val)
89 {
90 #ifdef CONFIG_X86_64
91 	*addr++ = 0x48;
92 	*addr++ = 0xbf;
93 #else
94 	*addr++ = 0xb8;
95 #endif
96 	*(unsigned long *)addr = val;
97 }
98 
99 asm (
100 			".pushsection .rodata\n"
101 			"optprobe_template_func:\n"
102 			".global optprobe_template_entry\n"
103 			"optprobe_template_entry:\n"
104 #ifdef CONFIG_X86_64
105 			/* We don't bother saving the ss register */
106 			"	pushq %rsp\n"
107 			"	pushfq\n"
108 			SAVE_REGS_STRING
109 			"	movq %rsp, %rsi\n"
110 			".global optprobe_template_val\n"
111 			"optprobe_template_val:\n"
112 			ASM_NOP5
113 			ASM_NOP5
114 			".global optprobe_template_call\n"
115 			"optprobe_template_call:\n"
116 			ASM_NOP5
117 			/* Move flags to rsp */
118 			"	movq 144(%rsp), %rdx\n"
119 			"	movq %rdx, 152(%rsp)\n"
120 			RESTORE_REGS_STRING
121 			/* Skip flags entry */
122 			"	addq $8, %rsp\n"
123 			"	popfq\n"
124 #else /* CONFIG_X86_32 */
125 			"	pushf\n"
126 			SAVE_REGS_STRING
127 			"	movl %esp, %edx\n"
128 			".global optprobe_template_val\n"
129 			"optprobe_template_val:\n"
130 			ASM_NOP5
131 			".global optprobe_template_call\n"
132 			"optprobe_template_call:\n"
133 			ASM_NOP5
134 			RESTORE_REGS_STRING
135 			"	addl $4, %esp\n"	/* skip cs */
136 			"	popf\n"
137 #endif
138 			".global optprobe_template_end\n"
139 			"optprobe_template_end:\n"
140 			".popsection\n");
141 
142 void optprobe_template_func(void);
143 STACK_FRAME_NON_STANDARD(optprobe_template_func);
144 
145 #define TMPL_MOVE_IDX \
146 	((long)optprobe_template_val - (long)optprobe_template_entry)
147 #define TMPL_CALL_IDX \
148 	((long)optprobe_template_call - (long)optprobe_template_entry)
149 #define TMPL_END_IDX \
150 	((long)optprobe_template_end - (long)optprobe_template_entry)
151 
152 #define INT3_SIZE sizeof(kprobe_opcode_t)
153 
154 /* Optimized kprobe call back function: called from optinsn */
155 static void
156 optimized_callback(struct optimized_kprobe *op, struct pt_regs *regs)
157 {
158 	/* This is possible if op is under delayed unoptimizing */
159 	if (kprobe_disabled(&op->kp))
160 		return;
161 
162 	preempt_disable();
163 	if (kprobe_running()) {
164 		kprobes_inc_nmissed_count(&op->kp);
165 	} else {
166 		struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
167 		/* Save skipped registers */
168 #ifdef CONFIG_X86_64
169 		regs->cs = __KERNEL_CS;
170 #else
171 		regs->cs = __KERNEL_CS | get_kernel_rpl();
172 		regs->gs = 0;
173 #endif
174 		regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
175 		regs->orig_ax = ~0UL;
176 
177 		__this_cpu_write(current_kprobe, &op->kp);
178 		kcb->kprobe_status = KPROBE_HIT_ACTIVE;
179 		opt_pre_handler(&op->kp, regs);
180 		__this_cpu_write(current_kprobe, NULL);
181 	}
182 	preempt_enable();
183 }
184 NOKPROBE_SYMBOL(optimized_callback);
185 
186 static int copy_optimized_instructions(u8 *dest, u8 *src, u8 *real)
187 {
188 	struct insn insn;
189 	int len = 0, ret;
190 
191 	while (len < RELATIVEJUMP_SIZE) {
192 		ret = __copy_instruction(dest + len, src + len, real + len, &insn);
193 		if (!ret || !can_boost(&insn, src + len))
194 			return -EINVAL;
195 		len += ret;
196 	}
197 	/* Check whether the address range is reserved */
198 	if (ftrace_text_reserved(src, src + len - 1) ||
199 	    alternatives_text_reserved(src, src + len - 1) ||
200 	    jump_label_text_reserved(src, src + len - 1))
201 		return -EBUSY;
202 
203 	return len;
204 }
205 
206 /* Check whether insn is indirect jump */
207 static int __insn_is_indirect_jump(struct insn *insn)
208 {
209 	return ((insn->opcode.bytes[0] == 0xff &&
210 		(X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
211 		insn->opcode.bytes[0] == 0xea);	/* Segment based jump */
212 }
213 
214 /* Check whether insn jumps into specified address range */
215 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
216 {
217 	unsigned long target = 0;
218 
219 	switch (insn->opcode.bytes[0]) {
220 	case 0xe0:	/* loopne */
221 	case 0xe1:	/* loope */
222 	case 0xe2:	/* loop */
223 	case 0xe3:	/* jcxz */
224 	case 0xe9:	/* near relative jump */
225 	case 0xeb:	/* short relative jump */
226 		break;
227 	case 0x0f:
228 		if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
229 			break;
230 		return 0;
231 	default:
232 		if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
233 			break;
234 		return 0;
235 	}
236 	target = (unsigned long)insn->next_byte + insn->immediate.value;
237 
238 	return (start <= target && target <= start + len);
239 }
240 
241 static int insn_is_indirect_jump(struct insn *insn)
242 {
243 	int ret = __insn_is_indirect_jump(insn);
244 
245 #ifdef CONFIG_RETPOLINE
246 	/*
247 	 * Jump to x86_indirect_thunk_* is treated as an indirect jump.
248 	 * Note that even with CONFIG_RETPOLINE=y, the kernel compiled with
249 	 * older gcc may use indirect jump. So we add this check instead of
250 	 * replace indirect-jump check.
251 	 */
252 	if (!ret)
253 		ret = insn_jump_into_range(insn,
254 				(unsigned long)__indirect_thunk_start,
255 				(unsigned long)__indirect_thunk_end -
256 				(unsigned long)__indirect_thunk_start);
257 #endif
258 	return ret;
259 }
260 
261 /* Decode whole function to ensure any instructions don't jump into target */
262 static int can_optimize(unsigned long paddr)
263 {
264 	unsigned long addr, size = 0, offset = 0;
265 	struct insn insn;
266 	kprobe_opcode_t buf[MAX_INSN_SIZE];
267 
268 	/* Lookup symbol including addr */
269 	if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
270 		return 0;
271 
272 	/*
273 	 * Do not optimize in the entry code due to the unstable
274 	 * stack handling and registers setup.
275 	 */
276 	if (((paddr >= (unsigned long)__entry_text_start) &&
277 	     (paddr <  (unsigned long)__entry_text_end)) ||
278 	    ((paddr >= (unsigned long)__irqentry_text_start) &&
279 	     (paddr <  (unsigned long)__irqentry_text_end)))
280 		return 0;
281 
282 	/* Check there is enough space for a relative jump. */
283 	if (size - offset < RELATIVEJUMP_SIZE)
284 		return 0;
285 
286 	/* Decode instructions */
287 	addr = paddr - offset;
288 	while (addr < paddr - offset + size) { /* Decode until function end */
289 		unsigned long recovered_insn;
290 		if (search_exception_tables(addr))
291 			/*
292 			 * Since some fixup code will jumps into this function,
293 			 * we can't optimize kprobe in this function.
294 			 */
295 			return 0;
296 		recovered_insn = recover_probed_instruction(buf, addr);
297 		if (!recovered_insn)
298 			return 0;
299 		kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
300 		insn_get_length(&insn);
301 		/* Another subsystem puts a breakpoint */
302 		if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
303 			return 0;
304 		/* Recover address */
305 		insn.kaddr = (void *)addr;
306 		insn.next_byte = (void *)(addr + insn.length);
307 		/* Check any instructions don't jump into target */
308 		if (insn_is_indirect_jump(&insn) ||
309 		    insn_jump_into_range(&insn, paddr + INT3_SIZE,
310 					 RELATIVE_ADDR_SIZE))
311 			return 0;
312 		addr += insn.length;
313 	}
314 
315 	return 1;
316 }
317 
318 /* Check optimized_kprobe can actually be optimized. */
319 int arch_check_optimized_kprobe(struct optimized_kprobe *op)
320 {
321 	int i;
322 	struct kprobe *p;
323 
324 	for (i = 1; i < op->optinsn.size; i++) {
325 		p = get_kprobe(op->kp.addr + i);
326 		if (p && !kprobe_disabled(p))
327 			return -EEXIST;
328 	}
329 
330 	return 0;
331 }
332 
333 /* Check the addr is within the optimized instructions. */
334 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
335 				 unsigned long addr)
336 {
337 	return ((unsigned long)op->kp.addr <= addr &&
338 		(unsigned long)op->kp.addr + op->optinsn.size > addr);
339 }
340 
341 /* Free optimized instruction slot */
342 static
343 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
344 {
345 	if (op->optinsn.insn) {
346 		free_optinsn_slot(op->optinsn.insn, dirty);
347 		op->optinsn.insn = NULL;
348 		op->optinsn.size = 0;
349 	}
350 }
351 
352 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
353 {
354 	__arch_remove_optimized_kprobe(op, 1);
355 }
356 
357 /*
358  * Copy replacing target instructions
359  * Target instructions MUST be relocatable (checked inside)
360  * This is called when new aggr(opt)probe is allocated or reused.
361  */
362 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op,
363 				  struct kprobe *__unused)
364 {
365 	u8 *buf = NULL, *slot;
366 	int ret, len;
367 	long rel;
368 
369 	if (!can_optimize((unsigned long)op->kp.addr))
370 		return -EILSEQ;
371 
372 	buf = kzalloc(MAX_OPTINSN_SIZE, GFP_KERNEL);
373 	if (!buf)
374 		return -ENOMEM;
375 
376 	op->optinsn.insn = slot = get_optinsn_slot();
377 	if (!slot) {
378 		ret = -ENOMEM;
379 		goto out;
380 	}
381 
382 	/*
383 	 * Verify if the address gap is in 2GB range, because this uses
384 	 * a relative jump.
385 	 */
386 	rel = (long)slot - (long)op->kp.addr + RELATIVEJUMP_SIZE;
387 	if (abs(rel) > 0x7fffffff) {
388 		ret = -ERANGE;
389 		goto err;
390 	}
391 
392 	/* Copy arch-dep-instance from template */
393 	memcpy(buf, optprobe_template_entry, TMPL_END_IDX);
394 
395 	/* Copy instructions into the out-of-line buffer */
396 	ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr,
397 					  slot + TMPL_END_IDX);
398 	if (ret < 0)
399 		goto err;
400 	op->optinsn.size = ret;
401 	len = TMPL_END_IDX + op->optinsn.size;
402 
403 	/* Set probe information */
404 	synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
405 
406 	/* Set probe function call */
407 	synthesize_relcall(buf + TMPL_CALL_IDX,
408 			   slot + TMPL_CALL_IDX, optimized_callback);
409 
410 	/* Set returning jmp instruction at the tail of out-of-line buffer */
411 	synthesize_reljump(buf + len, slot + len,
412 			   (u8 *)op->kp.addr + op->optinsn.size);
413 	len += RELATIVEJUMP_SIZE;
414 
415 	/* We have to use text_poke for instuction buffer because it is RO */
416 	text_poke(slot, buf, len);
417 	ret = 0;
418 out:
419 	kfree(buf);
420 	return ret;
421 
422 err:
423 	__arch_remove_optimized_kprobe(op, 0);
424 	goto out;
425 }
426 
427 /*
428  * Replace breakpoints (int3) with relative jumps.
429  * Caller must call with locking kprobe_mutex and text_mutex.
430  */
431 void arch_optimize_kprobes(struct list_head *oplist)
432 {
433 	struct optimized_kprobe *op, *tmp;
434 	u8 insn_buf[RELATIVEJUMP_SIZE];
435 
436 	list_for_each_entry_safe(op, tmp, oplist, list) {
437 		s32 rel = (s32)((long)op->optinsn.insn -
438 			((long)op->kp.addr + RELATIVEJUMP_SIZE));
439 
440 		WARN_ON(kprobe_disabled(&op->kp));
441 
442 		/* Backup instructions which will be replaced by jump address */
443 		memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
444 		       RELATIVE_ADDR_SIZE);
445 
446 		insn_buf[0] = RELATIVEJUMP_OPCODE;
447 		*(s32 *)(&insn_buf[1]) = rel;
448 
449 		text_poke_bp(op->kp.addr, insn_buf, RELATIVEJUMP_SIZE,
450 			     op->optinsn.insn);
451 
452 		list_del_init(&op->list);
453 	}
454 }
455 
456 /* Replace a relative jump with a breakpoint (int3).  */
457 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
458 {
459 	u8 insn_buf[RELATIVEJUMP_SIZE];
460 
461 	/* Set int3 to first byte for kprobes */
462 	insn_buf[0] = BREAKPOINT_INSTRUCTION;
463 	memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
464 	text_poke_bp(op->kp.addr, insn_buf, RELATIVEJUMP_SIZE,
465 		     op->optinsn.insn);
466 }
467 
468 /*
469  * Recover original instructions and breakpoints from relative jumps.
470  * Caller must call with locking kprobe_mutex.
471  */
472 extern void arch_unoptimize_kprobes(struct list_head *oplist,
473 				    struct list_head *done_list)
474 {
475 	struct optimized_kprobe *op, *tmp;
476 
477 	list_for_each_entry_safe(op, tmp, oplist, list) {
478 		arch_unoptimize_kprobe(op);
479 		list_move(&op->list, done_list);
480 	}
481 }
482 
483 int setup_detour_execution(struct kprobe *p, struct pt_regs *regs, int reenter)
484 {
485 	struct optimized_kprobe *op;
486 
487 	if (p->flags & KPROBE_FLAG_OPTIMIZED) {
488 		/* This kprobe is really able to run optimized path. */
489 		op = container_of(p, struct optimized_kprobe, kp);
490 		/* Detour through copied instructions */
491 		regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
492 		if (!reenter)
493 			reset_current_kprobe();
494 		return 1;
495 	}
496 	return 0;
497 }
498 NOKPROBE_SYMBOL(setup_detour_execution);
499