1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * arch/arm64/kernel/ftrace.c
4  *
5  * Copyright (C) 2013 Linaro Limited
6  * Author: AKASHI Takahiro <takahiro.akashi@linaro.org>
7  */
8 
9 #include <linux/ftrace.h>
10 #include <linux/module.h>
11 #include <linux/swab.h>
12 #include <linux/uaccess.h>
13 
14 #include <asm/cacheflush.h>
15 #include <asm/debug-monitors.h>
16 #include <asm/ftrace.h>
17 #include <asm/insn.h>
18 #include <asm/text-patching.h>
19 
20 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_ARGS
21 struct fregs_offset {
22 	const char *name;
23 	int offset;
24 };
25 
26 #define FREGS_OFFSET(n, field)					\
27 {								\
28 	.name = n,						\
29 	.offset = offsetof(struct __arch_ftrace_regs, field),	\
30 }
31 
32 static const struct fregs_offset fregs_offsets[] = {
33 	FREGS_OFFSET("x0", regs[0]),
34 	FREGS_OFFSET("x1", regs[1]),
35 	FREGS_OFFSET("x2", regs[2]),
36 	FREGS_OFFSET("x3", regs[3]),
37 	FREGS_OFFSET("x4", regs[4]),
38 	FREGS_OFFSET("x5", regs[5]),
39 	FREGS_OFFSET("x6", regs[6]),
40 	FREGS_OFFSET("x7", regs[7]),
41 	FREGS_OFFSET("x8", regs[8]),
42 
43 	FREGS_OFFSET("x29", fp),
44 	FREGS_OFFSET("x30", lr),
45 	FREGS_OFFSET("lr", lr),
46 
47 	FREGS_OFFSET("sp", sp),
48 	FREGS_OFFSET("pc", pc),
49 };
50 
51 int ftrace_regs_query_register_offset(const char *name)
52 {
53 	for (int i = 0; i < ARRAY_SIZE(fregs_offsets); i++) {
54 		const struct fregs_offset *roff = &fregs_offsets[i];
55 		if (!strcmp(roff->name, name))
56 			return roff->offset;
57 	}
58 
59 	return -EINVAL;
60 }
61 #endif
62 
63 unsigned long ftrace_call_adjust(unsigned long addr)
64 {
65 	/*
66 	 * When using mcount, addr is the address of the mcount call
67 	 * instruction, and no adjustment is necessary.
68 	 */
69 	if (!IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_ARGS))
70 		return addr;
71 
72 	/*
73 	 * When using patchable-function-entry without pre-function NOPS, addr
74 	 * is the address of the first NOP after the function entry point.
75 	 *
76 	 * The compiler has either generated:
77 	 *
78 	 * addr+00:	func:	NOP		// To be patched to MOV X9, LR
79 	 * addr+04:		NOP		// To be patched to BL <caller>
80 	 *
81 	 * Or:
82 	 *
83 	 * addr-04:		BTI	C
84 	 * addr+00:	func:	NOP		// To be patched to MOV X9, LR
85 	 * addr+04:		NOP		// To be patched to BL <caller>
86 	 *
87 	 * We must adjust addr to the address of the NOP which will be patched
88 	 * to `BL <caller>`, which is at `addr + 4` bytes in either case.
89 	 *
90 	 */
91 	if (!IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_CALL_OPS))
92 		return addr + AARCH64_INSN_SIZE;
93 
94 	/*
95 	 * When using patchable-function-entry with pre-function NOPs, addr is
96 	 * the address of the first pre-function NOP.
97 	 *
98 	 * Starting from an 8-byte aligned base, the compiler has either
99 	 * generated:
100 	 *
101 	 * addr+00:		NOP		// Literal (first 32 bits)
102 	 * addr+04:		NOP		// Literal (last 32 bits)
103 	 * addr+08:	func:	NOP		// To be patched to MOV X9, LR
104 	 * addr+12:		NOP		// To be patched to BL <caller>
105 	 *
106 	 * Or:
107 	 *
108 	 * addr+00:		NOP		// Literal (first 32 bits)
109 	 * addr+04:		NOP		// Literal (last 32 bits)
110 	 * addr+08:	func:	BTI	C
111 	 * addr+12:		NOP		// To be patched to MOV X9, LR
112 	 * addr+16:		NOP		// To be patched to BL <caller>
113 	 *
114 	 * We must adjust addr to the address of the NOP which will be patched
115 	 * to `BL <caller>`, which is at either addr+12 or addr+16 depending on
116 	 * whether there is a BTI.
117 	 */
118 
119 	if (!IS_ALIGNED(addr, sizeof(unsigned long))) {
120 		WARN_RATELIMIT(1, "Misaligned patch-site %pS\n",
121 			       (void *)(addr + 8));
122 		return 0;
123 	}
124 
125 	/* Skip the NOPs placed before the function entry point */
126 	addr += 2 * AARCH64_INSN_SIZE;
127 
128 	/* Skip any BTI */
129 	if (IS_ENABLED(CONFIG_ARM64_BTI_KERNEL)) {
130 		u32 insn = le32_to_cpu(*(__le32 *)addr);
131 
132 		if (aarch64_insn_is_bti(insn)) {
133 			addr += AARCH64_INSN_SIZE;
134 		} else if (insn != aarch64_insn_gen_nop()) {
135 			WARN_RATELIMIT(1, "unexpected insn in patch-site %pS: 0x%08x\n",
136 				       (void *)addr, insn);
137 		}
138 	}
139 
140 	/* Skip the first NOP after function entry */
141 	addr += AARCH64_INSN_SIZE;
142 
143 	return addr;
144 }
145 
146 /* Convert fentry_ip to the symbol address without kallsyms */
147 unsigned long arch_ftrace_get_symaddr(unsigned long fentry_ip)
148 {
149 	u32 insn;
150 
151 	/*
152 	 * When using patchable-function-entry without pre-function NOPS, ftrace
153 	 * entry is the address of the first NOP after the function entry point.
154 	 *
155 	 * The compiler has either generated:
156 	 *
157 	 * func+00:	func:	NOP		// To be patched to MOV X9, LR
158 	 * func+04:		NOP		// To be patched to BL <caller>
159 	 *
160 	 * Or:
161 	 *
162 	 * func-04:		BTI	C
163 	 * func+00:	func:	NOP		// To be patched to MOV X9, LR
164 	 * func+04:		NOP		// To be patched to BL <caller>
165 	 *
166 	 * The fentry_ip is the address of `BL <caller>` which is at `func + 4`
167 	 * bytes in either case.
168 	 */
169 	if (!IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_CALL_OPS))
170 		return fentry_ip - AARCH64_INSN_SIZE;
171 
172 	/*
173 	 * When using patchable-function-entry with pre-function NOPs, BTI is
174 	 * a bit different.
175 	 *
176 	 * func+00:	func:	NOP		// To be patched to MOV X9, LR
177 	 * func+04:		NOP		// To be patched to BL <caller>
178 	 *
179 	 * Or:
180 	 *
181 	 * func+00:	func:	BTI	C
182 	 * func+04:		NOP		// To be patched to MOV X9, LR
183 	 * func+08:		NOP		// To be patched to BL <caller>
184 	 *
185 	 * The fentry_ip is the address of `BL <caller>` which is at either
186 	 * `func + 4` or `func + 8` depends on whether there is a BTI.
187 	 */
188 
189 	/* If there is no BTI, the func address should be one instruction before. */
190 	if (!IS_ENABLED(CONFIG_ARM64_BTI_KERNEL))
191 		return fentry_ip - AARCH64_INSN_SIZE;
192 
193 	/* We want to be extra safe in case entry ip is on the page edge,
194 	 * but otherwise we need to avoid get_kernel_nofault()'s overhead.
195 	 */
196 	if ((fentry_ip & ~PAGE_MASK) < AARCH64_INSN_SIZE * 2) {
197 		if (get_kernel_nofault(insn, (u32 *)(fentry_ip - AARCH64_INSN_SIZE * 2)))
198 			return 0;
199 	} else {
200 		insn = *(u32 *)(fentry_ip - AARCH64_INSN_SIZE * 2);
201 	}
202 
203 	if (aarch64_insn_is_bti(le32_to_cpu((__le32)insn)))
204 		return fentry_ip - AARCH64_INSN_SIZE * 2;
205 
206 	return fentry_ip - AARCH64_INSN_SIZE;
207 }
208 
209 /*
210  * Replace a single instruction, which may be a branch or NOP.
211  * If @validate == true, a replaced instruction is checked against 'old'.
212  */
213 static int ftrace_modify_code(unsigned long pc, u32 old, u32 new,
214 			      bool validate)
215 {
216 	u32 replaced;
217 
218 	/*
219 	 * Note:
220 	 * We are paranoid about modifying text, as if a bug were to happen, it
221 	 * could cause us to read or write to someplace that could cause harm.
222 	 * Carefully read and modify the code with aarch64_insn_*() which uses
223 	 * probe_kernel_*(), and make sure what we read is what we expected it
224 	 * to be before modifying it.
225 	 */
226 	if (validate) {
227 		if (aarch64_insn_read((void *)pc, &replaced))
228 			return -EFAULT;
229 
230 		if (replaced != old)
231 			return -EINVAL;
232 	}
233 	if (aarch64_insn_patch_text_nosync((void *)pc, new))
234 		return -EPERM;
235 
236 	return 0;
237 }
238 
239 /*
240  * Replace tracer function in ftrace_caller()
241  */
242 int ftrace_update_ftrace_func(ftrace_func_t func)
243 {
244 	unsigned long pc;
245 	u32 new;
246 
247 	/*
248 	 * When using CALL_OPS, the function to call is associated with the
249 	 * call site, and we don't have a global function pointer to update.
250 	 */
251 	if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_CALL_OPS))
252 		return 0;
253 
254 	pc = (unsigned long)ftrace_call;
255 	new = aarch64_insn_gen_branch_imm(pc, (unsigned long)func,
256 					  AARCH64_INSN_BRANCH_LINK);
257 
258 	return ftrace_modify_code(pc, 0, new, false);
259 }
260 
261 static struct plt_entry *get_ftrace_plt(struct module *mod)
262 {
263 #ifdef CONFIG_MODULES
264 	struct plt_entry *plt = mod->arch.ftrace_trampolines;
265 
266 	return &plt[FTRACE_PLT_IDX];
267 #else
268 	return NULL;
269 #endif
270 }
271 
272 static bool reachable_by_bl(unsigned long addr, unsigned long pc)
273 {
274 	long offset = (long)addr - (long)pc;
275 
276 	return offset >= -SZ_128M && offset < SZ_128M;
277 }
278 
279 /*
280  * Find the address the callsite must branch to in order to reach '*addr'.
281  *
282  * Due to the limited range of 'BL' instructions, modules may be placed too far
283  * away to branch directly and must use a PLT.
284  *
285  * Returns true when '*addr' contains a reachable target address, or has been
286  * modified to contain a PLT address. Returns false otherwise.
287  */
288 static bool ftrace_find_callable_addr(struct dyn_ftrace *rec,
289 				      struct module *mod,
290 				      unsigned long *addr)
291 {
292 	unsigned long pc = rec->ip;
293 	struct plt_entry *plt;
294 
295 	/*
296 	 * If a custom trampoline is unreachable, rely on the ftrace_caller
297 	 * trampoline which knows how to indirectly reach that trampoline
298 	 * through ops->direct_call.
299 	 */
300 	if (*addr != FTRACE_ADDR && !reachable_by_bl(*addr, pc))
301 		*addr = FTRACE_ADDR;
302 
303 	/*
304 	 * When the target is within range of the 'BL' instruction, use 'addr'
305 	 * as-is and branch to that directly.
306 	 */
307 	if (reachable_by_bl(*addr, pc))
308 		return true;
309 
310 	/*
311 	 * When the target is outside of the range of a 'BL' instruction, we
312 	 * must use a PLT to reach it. We can only place PLTs for modules, and
313 	 * only when module PLT support is built-in.
314 	 */
315 	if (!IS_ENABLED(CONFIG_MODULES))
316 		return false;
317 
318 	/*
319 	 * 'mod' is only set at module load time, but if we end up
320 	 * dealing with an out-of-range condition, we can assume it
321 	 * is due to a module being loaded far away from the kernel.
322 	 *
323 	 * NOTE: __module_text_address() must be called within a RCU read
324 	 * section, but we can rely on ftrace_lock to ensure that 'mod'
325 	 * retains its validity throughout the remainder of this code.
326 	 */
327 	if (!mod) {
328 		guard(rcu)();
329 		mod = __module_text_address(pc);
330 	}
331 
332 	if (WARN_ON(!mod))
333 		return false;
334 
335 	plt = get_ftrace_plt(mod);
336 	if (!plt) {
337 		pr_err("ftrace: no module PLT for %ps\n", (void *)*addr);
338 		return false;
339 	}
340 
341 	*addr = (unsigned long)plt;
342 	return true;
343 }
344 
345 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_CALL_OPS
346 static const struct ftrace_ops *arm64_rec_get_ops(struct dyn_ftrace *rec)
347 {
348 	const struct ftrace_ops *ops = NULL;
349 
350 	if (rec->flags & FTRACE_FL_CALL_OPS_EN) {
351 		ops = ftrace_find_unique_ops(rec);
352 		WARN_ON_ONCE(!ops);
353 	}
354 
355 	if (!ops)
356 		ops = &ftrace_list_ops;
357 
358 	return ops;
359 }
360 
361 static int ftrace_rec_set_ops(const struct dyn_ftrace *rec,
362 			      const struct ftrace_ops *ops)
363 {
364 	unsigned long literal = ALIGN_DOWN(rec->ip - 12, 8);
365 	return aarch64_insn_write_literal_u64((void *)literal,
366 					      (unsigned long)ops);
367 }
368 
369 static int ftrace_rec_set_nop_ops(struct dyn_ftrace *rec)
370 {
371 	return ftrace_rec_set_ops(rec, &ftrace_nop_ops);
372 }
373 
374 static int ftrace_rec_update_ops(struct dyn_ftrace *rec)
375 {
376 	return ftrace_rec_set_ops(rec, arm64_rec_get_ops(rec));
377 }
378 #else
379 static int ftrace_rec_set_nop_ops(struct dyn_ftrace *rec) { return 0; }
380 static int ftrace_rec_update_ops(struct dyn_ftrace *rec) { return 0; }
381 #endif
382 
383 /*
384  * Turn on the call to ftrace_caller() in instrumented function
385  */
386 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
387 {
388 	unsigned long pc = rec->ip;
389 	u32 old, new;
390 	int ret;
391 
392 	ret = ftrace_rec_update_ops(rec);
393 	if (ret)
394 		return ret;
395 
396 	if (!ftrace_find_callable_addr(rec, NULL, &addr))
397 		return -EINVAL;
398 
399 	old = aarch64_insn_gen_nop();
400 	new = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
401 
402 	return ftrace_modify_code(pc, old, new, true);
403 }
404 
405 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_CALL_OPS
406 int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
407 		       unsigned long addr)
408 {
409 	unsigned long pc = rec->ip;
410 	u32 old, new;
411 	int ret;
412 
413 	ret = ftrace_rec_set_ops(rec, arm64_rec_get_ops(rec));
414 	if (ret)
415 		return ret;
416 
417 	if (!ftrace_find_callable_addr(rec, NULL, &old_addr))
418 		return -EINVAL;
419 	if (!ftrace_find_callable_addr(rec, NULL, &addr))
420 		return -EINVAL;
421 
422 	old = aarch64_insn_gen_branch_imm(pc, old_addr,
423 					  AARCH64_INSN_BRANCH_LINK);
424 	new = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
425 
426 	return ftrace_modify_code(pc, old, new, true);
427 }
428 #endif
429 
430 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_ARGS
431 /*
432  * The compiler has inserted two NOPs before the regular function prologue.
433  * All instrumented functions follow the AAPCS, so x0-x8 and x19-x30 are live,
434  * and x9-x18 are free for our use.
435  *
436  * At runtime we want to be able to swing a single NOP <-> BL to enable or
437  * disable the ftrace call. The BL requires us to save the original LR value,
438  * so here we insert a <MOV X9, LR> over the first NOP so the instructions
439  * before the regular prologue are:
440  *
441  * | Compiled | Disabled   | Enabled    |
442  * +----------+------------+------------+
443  * | NOP      | MOV X9, LR | MOV X9, LR |
444  * | NOP      | NOP        | BL <entry> |
445  *
446  * The LR value will be recovered by ftrace_caller, and restored into LR
447  * before returning to the regular function prologue. When a function is not
448  * being traced, the MOV is not harmful given x9 is not live per the AAPCS.
449  *
450  * Note: ftrace_process_locs() has pre-adjusted rec->ip to be the address of
451  * the BL.
452  */
453 int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec)
454 {
455 	unsigned long pc = rec->ip - AARCH64_INSN_SIZE;
456 	u32 old, new;
457 	int ret;
458 
459 	ret = ftrace_rec_set_nop_ops(rec);
460 	if (ret)
461 		return ret;
462 
463 	old = aarch64_insn_gen_nop();
464 	new = aarch64_insn_gen_move_reg(AARCH64_INSN_REG_9,
465 					AARCH64_INSN_REG_LR,
466 					AARCH64_INSN_VARIANT_64BIT);
467 	return ftrace_modify_code(pc, old, new, true);
468 }
469 #endif
470 
471 /*
472  * Turn off the call to ftrace_caller() in instrumented function
473  */
474 int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec,
475 		    unsigned long addr)
476 {
477 	unsigned long pc = rec->ip;
478 	u32 old = 0, new;
479 	int ret;
480 
481 	new = aarch64_insn_gen_nop();
482 
483 	ret = ftrace_rec_set_nop_ops(rec);
484 	if (ret)
485 		return ret;
486 
487 	/*
488 	 * When using mcount, callsites in modules may have been initalized to
489 	 * call an arbitrary module PLT (which redirects to the _mcount stub)
490 	 * rather than the ftrace PLT we'll use at runtime (which redirects to
491 	 * the ftrace trampoline). We can ignore the old PLT when initializing
492 	 * the callsite.
493 	 *
494 	 * Note: 'mod' is only set at module load time.
495 	 */
496 	if (!IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_ARGS) && mod)
497 		return aarch64_insn_patch_text_nosync((void *)pc, new);
498 
499 	if (!ftrace_find_callable_addr(rec, mod, &addr))
500 		return -EINVAL;
501 
502 	old = aarch64_insn_gen_branch_imm(pc, addr, AARCH64_INSN_BRANCH_LINK);
503 
504 	return ftrace_modify_code(pc, old, new, true);
505 }
506 
507 void arch_ftrace_update_code(int command)
508 {
509 	command |= FTRACE_MAY_SLEEP;
510 	ftrace_modify_all_code(command);
511 }
512 
513 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
514 /*
515  * function_graph tracer expects ftrace_return_to_handler() to be called
516  * on the way back to parent. For this purpose, this function is called
517  * in _mcount() or ftrace_caller() to replace return address (*parent) on
518  * the call stack to return_to_handler.
519  */
520 void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
521 			   unsigned long frame_pointer)
522 {
523 	unsigned long return_hooker = (unsigned long)&return_to_handler;
524 	unsigned long old;
525 
526 	if (unlikely(atomic_read(&current->tracing_graph_pause)))
527 		return;
528 
529 	/*
530 	 * Note:
531 	 * No protection against faulting at *parent, which may be seen
532 	 * on other archs. It's unlikely on AArch64.
533 	 */
534 	old = *parent;
535 
536 	if (!function_graph_enter(old, self_addr, frame_pointer,
537 	    (void *)frame_pointer)) {
538 		*parent = return_hooker;
539 	}
540 }
541 
542 #ifdef CONFIG_DYNAMIC_FTRACE_WITH_ARGS
543 void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
544 		       struct ftrace_ops *op, struct ftrace_regs *fregs)
545 {
546 	unsigned long return_hooker = (unsigned long)&return_to_handler;
547 	unsigned long frame_pointer = arch_ftrace_regs(fregs)->fp;
548 	unsigned long *parent = &arch_ftrace_regs(fregs)->lr;
549 	unsigned long old;
550 
551 	if (unlikely(atomic_read(&current->tracing_graph_pause)))
552 		return;
553 
554 	old = *parent;
555 
556 	if (!function_graph_enter_regs(old, ip, frame_pointer,
557 				       (void *)frame_pointer, fregs)) {
558 		*parent = return_hooker;
559 	}
560 }
561 #else
562 /*
563  * Turn on/off the call to ftrace_graph_caller() in ftrace_caller()
564  * depending on @enable.
565  */
566 static int ftrace_modify_graph_caller(bool enable)
567 {
568 	unsigned long pc = (unsigned long)&ftrace_graph_call;
569 	u32 branch, nop;
570 
571 	branch = aarch64_insn_gen_branch_imm(pc,
572 					     (unsigned long)ftrace_graph_caller,
573 					     AARCH64_INSN_BRANCH_NOLINK);
574 	nop = aarch64_insn_gen_nop();
575 
576 	if (enable)
577 		return ftrace_modify_code(pc, nop, branch, true);
578 	else
579 		return ftrace_modify_code(pc, branch, nop, true);
580 }
581 
582 int ftrace_enable_ftrace_graph_caller(void)
583 {
584 	return ftrace_modify_graph_caller(true);
585 }
586 
587 int ftrace_disable_ftrace_graph_caller(void)
588 {
589 	return ftrace_modify_graph_caller(false);
590 }
591 #endif /* CONFIG_DYNAMIC_FTRACE_WITH_ARGS */
592 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
593