Loading...
1/*
2 * Code for replacing ftrace calls with jumps.
3 *
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
5 *
6 * Thanks goes to Ingo Molnar, for suggesting the idea.
7 * Mathieu Desnoyers, for suggesting postponing the modifications.
8 * Arjan van de Ven, for keeping me straight, and explaining to me
9 * the dangers of modifying code on the run.
10 */
11
12#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
13
14#include <linux/spinlock.h>
15#include <linux/hardirq.h>
16#include <linux/uaccess.h>
17#include <linux/ftrace.h>
18#include <linux/percpu.h>
19#include <linux/sched.h>
20#include <linux/init.h>
21#include <linux/list.h>
22#include <linux/module.h>
23
24#include <trace/syscall.h>
25
26#include <asm/cacheflush.h>
27#include <asm/ftrace.h>
28#include <asm/nops.h>
29#include <asm/nmi.h>
30
31
32#ifdef CONFIG_DYNAMIC_FTRACE
33
34/*
35 * modifying_code is set to notify NMIs that they need to use
36 * memory barriers when entering or exiting. But we don't want
37 * to burden NMIs with unnecessary memory barriers when code
38 * modification is not being done (which is most of the time).
39 *
40 * A mutex is already held when ftrace_arch_code_modify_prepare
41 * and post_process are called. No locks need to be taken here.
42 *
43 * Stop machine will make sure currently running NMIs are done
44 * and new NMIs will see the updated variable before we need
45 * to worry about NMIs doing memory barriers.
46 */
47static int modifying_code __read_mostly;
48static DEFINE_PER_CPU(int, save_modifying_code);
49
50int ftrace_arch_code_modify_prepare(void)
51{
52 set_kernel_text_rw();
53 set_all_modules_text_rw();
54 modifying_code = 1;
55 return 0;
56}
57
58int ftrace_arch_code_modify_post_process(void)
59{
60 modifying_code = 0;
61 set_all_modules_text_ro();
62 set_kernel_text_ro();
63 return 0;
64}
65
66union ftrace_code_union {
67 char code[MCOUNT_INSN_SIZE];
68 struct {
69 char e8;
70 int offset;
71 } __attribute__((packed));
72};
73
74static int ftrace_calc_offset(long ip, long addr)
75{
76 return (int)(addr - ip);
77}
78
79static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
80{
81 static union ftrace_code_union calc;
82
83 calc.e8 = 0xe8;
84 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
85
86 /*
87 * No locking needed, this must be called via kstop_machine
88 * which in essence is like running on a uniprocessor machine.
89 */
90 return calc.code;
91}
92
93/*
94 * Modifying code must take extra care. On an SMP machine, if
95 * the code being modified is also being executed on another CPU
96 * that CPU will have undefined results and possibly take a GPF.
97 * We use kstop_machine to stop other CPUS from exectuing code.
98 * But this does not stop NMIs from happening. We still need
99 * to protect against that. We separate out the modification of
100 * the code to take care of this.
101 *
102 * Two buffers are added: An IP buffer and a "code" buffer.
103 *
104 * 1) Put the instruction pointer into the IP buffer
105 * and the new code into the "code" buffer.
106 * 2) Wait for any running NMIs to finish and set a flag that says
107 * we are modifying code, it is done in an atomic operation.
108 * 3) Write the code
109 * 4) clear the flag.
110 * 5) Wait for any running NMIs to finish.
111 *
112 * If an NMI is executed, the first thing it does is to call
113 * "ftrace_nmi_enter". This will check if the flag is set to write
114 * and if it is, it will write what is in the IP and "code" buffers.
115 *
116 * The trick is, it does not matter if everyone is writing the same
117 * content to the code location. Also, if a CPU is executing code
118 * it is OK to write to that code location if the contents being written
119 * are the same as what exists.
120 */
121
122#define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */
123static atomic_t nmi_running = ATOMIC_INIT(0);
124static int mod_code_status; /* holds return value of text write */
125static void *mod_code_ip; /* holds the IP to write to */
126static const void *mod_code_newcode; /* holds the text to write to the IP */
127
128static unsigned nmi_wait_count;
129static atomic_t nmi_update_count = ATOMIC_INIT(0);
130
131int ftrace_arch_read_dyn_info(char *buf, int size)
132{
133 int r;
134
135 r = snprintf(buf, size, "%u %u",
136 nmi_wait_count,
137 atomic_read(&nmi_update_count));
138 return r;
139}
140
141static void clear_mod_flag(void)
142{
143 int old = atomic_read(&nmi_running);
144
145 for (;;) {
146 int new = old & ~MOD_CODE_WRITE_FLAG;
147
148 if (old == new)
149 break;
150
151 old = atomic_cmpxchg(&nmi_running, old, new);
152 }
153}
154
155static void ftrace_mod_code(void)
156{
157 /*
158 * Yes, more than one CPU process can be writing to mod_code_status.
159 * (and the code itself)
160 * But if one were to fail, then they all should, and if one were
161 * to succeed, then they all should.
162 */
163 mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
164 MCOUNT_INSN_SIZE);
165
166 /* if we fail, then kill any new writers */
167 if (mod_code_status)
168 clear_mod_flag();
169}
170
171void ftrace_nmi_enter(void)
172{
173 __this_cpu_write(save_modifying_code, modifying_code);
174
175 if (!__this_cpu_read(save_modifying_code))
176 return;
177
178 if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) {
179 smp_rmb();
180 ftrace_mod_code();
181 atomic_inc(&nmi_update_count);
182 }
183 /* Must have previous changes seen before executions */
184 smp_mb();
185}
186
187void ftrace_nmi_exit(void)
188{
189 if (!__this_cpu_read(save_modifying_code))
190 return;
191
192 /* Finish all executions before clearing nmi_running */
193 smp_mb();
194 atomic_dec(&nmi_running);
195}
196
197static void wait_for_nmi_and_set_mod_flag(void)
198{
199 if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG))
200 return;
201
202 do {
203 cpu_relax();
204 } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG));
205
206 nmi_wait_count++;
207}
208
209static void wait_for_nmi(void)
210{
211 if (!atomic_read(&nmi_running))
212 return;
213
214 do {
215 cpu_relax();
216 } while (atomic_read(&nmi_running));
217
218 nmi_wait_count++;
219}
220
221static inline int
222within(unsigned long addr, unsigned long start, unsigned long end)
223{
224 return addr >= start && addr < end;
225}
226
227static int
228do_ftrace_mod_code(unsigned long ip, const void *new_code)
229{
230 /*
231 * On x86_64, kernel text mappings are mapped read-only with
232 * CONFIG_DEBUG_RODATA. So we use the kernel identity mapping instead
233 * of the kernel text mapping to modify the kernel text.
234 *
235 * For 32bit kernels, these mappings are same and we can use
236 * kernel identity mapping to modify code.
237 */
238 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
239 ip = (unsigned long)__va(__pa(ip));
240
241 mod_code_ip = (void *)ip;
242 mod_code_newcode = new_code;
243
244 /* The buffers need to be visible before we let NMIs write them */
245 smp_mb();
246
247 wait_for_nmi_and_set_mod_flag();
248
249 /* Make sure all running NMIs have finished before we write the code */
250 smp_mb();
251
252 ftrace_mod_code();
253
254 /* Make sure the write happens before clearing the bit */
255 smp_mb();
256
257 clear_mod_flag();
258 wait_for_nmi();
259
260 return mod_code_status;
261}
262
263static const unsigned char *ftrace_nop_replace(void)
264{
265 return ideal_nops[NOP_ATOMIC5];
266}
267
268static int
269ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
270 unsigned const char *new_code)
271{
272 unsigned char replaced[MCOUNT_INSN_SIZE];
273
274 /*
275 * Note: Due to modules and __init, code can
276 * disappear and change, we need to protect against faulting
277 * as well as code changing. We do this by using the
278 * probe_kernel_* functions.
279 *
280 * No real locking needed, this code is run through
281 * kstop_machine, or before SMP starts.
282 */
283
284 /* read the text we want to modify */
285 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
286 return -EFAULT;
287
288 /* Make sure it is what we expect it to be */
289 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
290 return -EINVAL;
291
292 /* replace the text with the new text */
293 if (do_ftrace_mod_code(ip, new_code))
294 return -EPERM;
295
296 sync_core();
297
298 return 0;
299}
300
301int ftrace_make_nop(struct module *mod,
302 struct dyn_ftrace *rec, unsigned long addr)
303{
304 unsigned const char *new, *old;
305 unsigned long ip = rec->ip;
306
307 old = ftrace_call_replace(ip, addr);
308 new = ftrace_nop_replace();
309
310 return ftrace_modify_code(rec->ip, old, new);
311}
312
313int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
314{
315 unsigned const char *new, *old;
316 unsigned long ip = rec->ip;
317
318 old = ftrace_nop_replace();
319 new = ftrace_call_replace(ip, addr);
320
321 return ftrace_modify_code(rec->ip, old, new);
322}
323
324int ftrace_update_ftrace_func(ftrace_func_t func)
325{
326 unsigned long ip = (unsigned long)(&ftrace_call);
327 unsigned char old[MCOUNT_INSN_SIZE], *new;
328 int ret;
329
330 memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
331 new = ftrace_call_replace(ip, (unsigned long)func);
332 ret = ftrace_modify_code(ip, old, new);
333
334 return ret;
335}
336
337int __init ftrace_dyn_arch_init(void *data)
338{
339 /* The return code is retured via data */
340 *(unsigned long *)data = 0;
341
342 return 0;
343}
344#endif
345
346#ifdef CONFIG_FUNCTION_GRAPH_TRACER
347
348#ifdef CONFIG_DYNAMIC_FTRACE
349extern void ftrace_graph_call(void);
350
351static int ftrace_mod_jmp(unsigned long ip,
352 int old_offset, int new_offset)
353{
354 unsigned char code[MCOUNT_INSN_SIZE];
355
356 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
357 return -EFAULT;
358
359 if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
360 return -EINVAL;
361
362 *(int *)(&code[1]) = new_offset;
363
364 if (do_ftrace_mod_code(ip, &code))
365 return -EPERM;
366
367 return 0;
368}
369
370int ftrace_enable_ftrace_graph_caller(void)
371{
372 unsigned long ip = (unsigned long)(&ftrace_graph_call);
373 int old_offset, new_offset;
374
375 old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
376 new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
377
378 return ftrace_mod_jmp(ip, old_offset, new_offset);
379}
380
381int ftrace_disable_ftrace_graph_caller(void)
382{
383 unsigned long ip = (unsigned long)(&ftrace_graph_call);
384 int old_offset, new_offset;
385
386 old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
387 new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
388
389 return ftrace_mod_jmp(ip, old_offset, new_offset);
390}
391
392#endif /* !CONFIG_DYNAMIC_FTRACE */
393
394/*
395 * Hook the return address and push it in the stack of return addrs
396 * in current thread info.
397 */
398void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr,
399 unsigned long frame_pointer)
400{
401 unsigned long old;
402 int faulted;
403 struct ftrace_graph_ent trace;
404 unsigned long return_hooker = (unsigned long)
405 &return_to_handler;
406
407 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
408 return;
409
410 /*
411 * Protect against fault, even if it shouldn't
412 * happen. This tool is too much intrusive to
413 * ignore such a protection.
414 */
415 asm volatile(
416 "1: " _ASM_MOV " (%[parent]), %[old]\n"
417 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
418 " movl $0, %[faulted]\n"
419 "3:\n"
420
421 ".section .fixup, \"ax\"\n"
422 "4: movl $1, %[faulted]\n"
423 " jmp 3b\n"
424 ".previous\n"
425
426 _ASM_EXTABLE(1b, 4b)
427 _ASM_EXTABLE(2b, 4b)
428
429 : [old] "=&r" (old), [faulted] "=r" (faulted)
430 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
431 : "memory"
432 );
433
434 if (unlikely(faulted)) {
435 ftrace_graph_stop();
436 WARN_ON(1);
437 return;
438 }
439
440 trace.func = self_addr;
441 trace.depth = current->curr_ret_stack + 1;
442
443 /* Only trace if the calling function expects to */
444 if (!ftrace_graph_entry(&trace)) {
445 *parent = old;
446 return;
447 }
448
449 if (ftrace_push_return_trace(old, self_addr, &trace.depth,
450 frame_pointer) == -EBUSY) {
451 *parent = old;
452 return;
453 }
454}
455#endif /* CONFIG_FUNCTION_GRAPH_TRACER */
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Dynamic function tracing support.
4 *
5 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
6 *
7 * Thanks goes to Ingo Molnar, for suggesting the idea.
8 * Mathieu Desnoyers, for suggesting postponing the modifications.
9 * Arjan van de Ven, for keeping me straight, and explaining to me
10 * the dangers of modifying code on the run.
11 */
12
13#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14
15#include <linux/spinlock.h>
16#include <linux/hardirq.h>
17#include <linux/uaccess.h>
18#include <linux/ftrace.h>
19#include <linux/percpu.h>
20#include <linux/sched.h>
21#include <linux/slab.h>
22#include <linux/init.h>
23#include <linux/list.h>
24#include <linux/module.h>
25#include <linux/memory.h>
26#include <linux/vmalloc.h>
27#include <linux/set_memory.h>
28
29#include <trace/syscall.h>
30
31#include <asm/kprobes.h>
32#include <asm/ftrace.h>
33#include <asm/nops.h>
34#include <asm/text-patching.h>
35
36#ifdef CONFIG_DYNAMIC_FTRACE
37
38static int ftrace_poke_late = 0;
39
40void ftrace_arch_code_modify_prepare(void)
41 __acquires(&text_mutex)
42{
43 /*
44 * Need to grab text_mutex to prevent a race from module loading
45 * and live kernel patching from changing the text permissions while
46 * ftrace has it set to "read/write".
47 */
48 mutex_lock(&text_mutex);
49 ftrace_poke_late = 1;
50}
51
52void ftrace_arch_code_modify_post_process(void)
53 __releases(&text_mutex)
54{
55 /*
56 * ftrace_make_{call,nop}() may be called during
57 * module load, and we need to finish the text_poke_queue()
58 * that they do, here.
59 */
60 text_poke_finish();
61 ftrace_poke_late = 0;
62 mutex_unlock(&text_mutex);
63}
64
65static const char *ftrace_nop_replace(void)
66{
67 return x86_nops[5];
68}
69
70static const char *ftrace_call_replace(unsigned long ip, unsigned long addr)
71{
72 /*
73 * No need to translate into a callthunk. The trampoline does
74 * the depth accounting itself.
75 */
76 return text_gen_insn(CALL_INSN_OPCODE, (void *)ip, (void *)addr);
77}
78
79static int ftrace_verify_code(unsigned long ip, const char *old_code)
80{
81 char cur_code[MCOUNT_INSN_SIZE];
82
83 /*
84 * Note:
85 * We are paranoid about modifying text, as if a bug was to happen, it
86 * could cause us to read or write to someplace that could cause harm.
87 * Carefully read and modify the code with probe_kernel_*(), and make
88 * sure what we read is what we expected it to be before modifying it.
89 */
90 /* read the text we want to modify */
91 if (copy_from_kernel_nofault(cur_code, (void *)ip, MCOUNT_INSN_SIZE)) {
92 WARN_ON(1);
93 return -EFAULT;
94 }
95
96 /* Make sure it is what we expect it to be */
97 if (memcmp(cur_code, old_code, MCOUNT_INSN_SIZE) != 0) {
98 ftrace_expected = old_code;
99 WARN_ON(1);
100 return -EINVAL;
101 }
102
103 return 0;
104}
105
106/*
107 * Marked __ref because it calls text_poke_early() which is .init.text. That is
108 * ok because that call will happen early, during boot, when .init sections are
109 * still present.
110 */
111static int __ref
112ftrace_modify_code_direct(unsigned long ip, const char *old_code,
113 const char *new_code)
114{
115 int ret = ftrace_verify_code(ip, old_code);
116 if (ret)
117 return ret;
118
119 /* replace the text with the new text */
120 if (ftrace_poke_late)
121 text_poke_queue((void *)ip, new_code, MCOUNT_INSN_SIZE, NULL);
122 else
123 text_poke_early((void *)ip, new_code, MCOUNT_INSN_SIZE);
124 return 0;
125}
126
127int ftrace_make_nop(struct module *mod, struct dyn_ftrace *rec, unsigned long addr)
128{
129 unsigned long ip = rec->ip;
130 const char *new, *old;
131
132 old = ftrace_call_replace(ip, addr);
133 new = ftrace_nop_replace();
134
135 /*
136 * On boot up, and when modules are loaded, the MCOUNT_ADDR
137 * is converted to a nop, and will never become MCOUNT_ADDR
138 * again. This code is either running before SMP (on boot up)
139 * or before the code will ever be executed (module load).
140 * We do not want to use the breakpoint version in this case,
141 * just modify the code directly.
142 */
143 if (addr == MCOUNT_ADDR)
144 return ftrace_modify_code_direct(ip, old, new);
145
146 /*
147 * x86 overrides ftrace_replace_code -- this function will never be used
148 * in this case.
149 */
150 WARN_ONCE(1, "invalid use of ftrace_make_nop");
151 return -EINVAL;
152}
153
154int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
155{
156 unsigned long ip = rec->ip;
157 const char *new, *old;
158
159 old = ftrace_nop_replace();
160 new = ftrace_call_replace(ip, addr);
161
162 /* Should only be called when module is loaded */
163 return ftrace_modify_code_direct(rec->ip, old, new);
164}
165
166/*
167 * Should never be called:
168 * As it is only called by __ftrace_replace_code() which is called by
169 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
170 * which is called to turn mcount into nops or nops into function calls
171 * but not to convert a function from not using regs to one that uses
172 * regs, which ftrace_modify_call() is for.
173 */
174int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
175 unsigned long addr)
176{
177 WARN_ON(1);
178 return -EINVAL;
179}
180
181int ftrace_update_ftrace_func(ftrace_func_t func)
182{
183 unsigned long ip;
184 const char *new;
185
186 ip = (unsigned long)(&ftrace_call);
187 new = ftrace_call_replace(ip, (unsigned long)func);
188 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
189
190 ip = (unsigned long)(&ftrace_regs_call);
191 new = ftrace_call_replace(ip, (unsigned long)func);
192 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
193
194 return 0;
195}
196
197void ftrace_replace_code(int enable)
198{
199 struct ftrace_rec_iter *iter;
200 struct dyn_ftrace *rec;
201 const char *new, *old;
202 int ret;
203
204 for_ftrace_rec_iter(iter) {
205 rec = ftrace_rec_iter_record(iter);
206
207 switch (ftrace_test_record(rec, enable)) {
208 case FTRACE_UPDATE_IGNORE:
209 default:
210 continue;
211
212 case FTRACE_UPDATE_MAKE_CALL:
213 old = ftrace_nop_replace();
214 break;
215
216 case FTRACE_UPDATE_MODIFY_CALL:
217 case FTRACE_UPDATE_MAKE_NOP:
218 old = ftrace_call_replace(rec->ip, ftrace_get_addr_curr(rec));
219 break;
220 }
221
222 ret = ftrace_verify_code(rec->ip, old);
223 if (ret) {
224 ftrace_expected = old;
225 ftrace_bug(ret, rec);
226 ftrace_expected = NULL;
227 return;
228 }
229 }
230
231 for_ftrace_rec_iter(iter) {
232 rec = ftrace_rec_iter_record(iter);
233
234 switch (ftrace_test_record(rec, enable)) {
235 case FTRACE_UPDATE_IGNORE:
236 default:
237 continue;
238
239 case FTRACE_UPDATE_MAKE_CALL:
240 case FTRACE_UPDATE_MODIFY_CALL:
241 new = ftrace_call_replace(rec->ip, ftrace_get_addr_new(rec));
242 break;
243
244 case FTRACE_UPDATE_MAKE_NOP:
245 new = ftrace_nop_replace();
246 break;
247 }
248
249 text_poke_queue((void *)rec->ip, new, MCOUNT_INSN_SIZE, NULL);
250 ftrace_update_record(rec, enable);
251 }
252 text_poke_finish();
253}
254
255void arch_ftrace_update_code(int command)
256{
257 ftrace_modify_all_code(command);
258}
259
260/* Currently only x86_64 supports dynamic trampolines */
261#ifdef CONFIG_X86_64
262
263#ifdef CONFIG_MODULES
264#include <linux/moduleloader.h>
265/* Module allocation simplifies allocating memory for code */
266static inline void *alloc_tramp(unsigned long size)
267{
268 return module_alloc(size);
269}
270static inline void tramp_free(void *tramp)
271{
272 module_memfree(tramp);
273}
274#else
275/* Trampolines can only be created if modules are supported */
276static inline void *alloc_tramp(unsigned long size)
277{
278 return NULL;
279}
280static inline void tramp_free(void *tramp) { }
281#endif
282
283/* Defined as markers to the end of the ftrace default trampolines */
284extern void ftrace_regs_caller_end(void);
285extern void ftrace_caller_end(void);
286extern void ftrace_caller_op_ptr(void);
287extern void ftrace_regs_caller_op_ptr(void);
288extern void ftrace_regs_caller_jmp(void);
289
290/* movq function_trace_op(%rip), %rdx */
291/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
292#define OP_REF_SIZE 7
293
294/*
295 * The ftrace_ops is passed to the function callback. Since the
296 * trampoline only services a single ftrace_ops, we can pass in
297 * that ops directly.
298 *
299 * The ftrace_op_code_union is used to create a pointer to the
300 * ftrace_ops that will be passed to the callback function.
301 */
302union ftrace_op_code_union {
303 char code[OP_REF_SIZE];
304 struct {
305 char op[3];
306 int offset;
307 } __attribute__((packed));
308};
309
310#define RET_SIZE (IS_ENABLED(CONFIG_RETPOLINE) ? 5 : 1 + IS_ENABLED(CONFIG_SLS))
311
312static unsigned long
313create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
314{
315 unsigned long start_offset;
316 unsigned long end_offset;
317 unsigned long op_offset;
318 unsigned long call_offset;
319 unsigned long jmp_offset;
320 unsigned long offset;
321 unsigned long npages;
322 unsigned long size;
323 unsigned long *ptr;
324 void *trampoline;
325 void *ip, *dest;
326 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
327 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
328 unsigned const char retq[] = { RET_INSN_OPCODE, INT3_INSN_OPCODE };
329 union ftrace_op_code_union op_ptr;
330 int ret;
331
332 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
333 start_offset = (unsigned long)ftrace_regs_caller;
334 end_offset = (unsigned long)ftrace_regs_caller_end;
335 op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
336 call_offset = (unsigned long)ftrace_regs_call;
337 jmp_offset = (unsigned long)ftrace_regs_caller_jmp;
338 } else {
339 start_offset = (unsigned long)ftrace_caller;
340 end_offset = (unsigned long)ftrace_caller_end;
341 op_offset = (unsigned long)ftrace_caller_op_ptr;
342 call_offset = (unsigned long)ftrace_call;
343 jmp_offset = 0;
344 }
345
346 size = end_offset - start_offset;
347
348 /*
349 * Allocate enough size to store the ftrace_caller code,
350 * the iret , as well as the address of the ftrace_ops this
351 * trampoline is used for.
352 */
353 trampoline = alloc_tramp(size + RET_SIZE + sizeof(void *));
354 if (!trampoline)
355 return 0;
356
357 *tramp_size = size + RET_SIZE + sizeof(void *);
358 npages = DIV_ROUND_UP(*tramp_size, PAGE_SIZE);
359
360 /* Copy ftrace_caller onto the trampoline memory */
361 ret = copy_from_kernel_nofault(trampoline, (void *)start_offset, size);
362 if (WARN_ON(ret < 0))
363 goto fail;
364
365 ip = trampoline + size;
366 if (cpu_feature_enabled(X86_FEATURE_RETHUNK))
367 __text_gen_insn(ip, JMP32_INSN_OPCODE, ip, x86_return_thunk, JMP32_INSN_SIZE);
368 else
369 memcpy(ip, retq, sizeof(retq));
370
371 /* No need to test direct calls on created trampolines */
372 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
373 /* NOP the jnz 1f; but make sure it's a 2 byte jnz */
374 ip = trampoline + (jmp_offset - start_offset);
375 if (WARN_ON(*(char *)ip != 0x75))
376 goto fail;
377 ret = copy_from_kernel_nofault(ip, x86_nops[2], 2);
378 if (ret < 0)
379 goto fail;
380 }
381
382 /*
383 * The address of the ftrace_ops that is used for this trampoline
384 * is stored at the end of the trampoline. This will be used to
385 * load the third parameter for the callback. Basically, that
386 * location at the end of the trampoline takes the place of
387 * the global function_trace_op variable.
388 */
389
390 ptr = (unsigned long *)(trampoline + size + RET_SIZE);
391 *ptr = (unsigned long)ops;
392
393 op_offset -= start_offset;
394 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
395
396 /* Are we pointing to the reference? */
397 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0))
398 goto fail;
399
400 /* Load the contents of ptr into the callback parameter */
401 offset = (unsigned long)ptr;
402 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
403
404 op_ptr.offset = offset;
405
406 /* put in the new offset to the ftrace_ops */
407 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
408
409 /* put in the call to the function */
410 mutex_lock(&text_mutex);
411 call_offset -= start_offset;
412 /*
413 * No need to translate into a callthunk. The trampoline does
414 * the depth accounting before the call already.
415 */
416 dest = ftrace_ops_get_func(ops);
417 memcpy(trampoline + call_offset,
418 text_gen_insn(CALL_INSN_OPCODE, trampoline + call_offset, dest),
419 CALL_INSN_SIZE);
420 mutex_unlock(&text_mutex);
421
422 /* ALLOC_TRAMP flags lets us know we created it */
423 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
424
425 set_memory_rox((unsigned long)trampoline, npages);
426 return (unsigned long)trampoline;
427fail:
428 tramp_free(trampoline);
429 return 0;
430}
431
432void set_ftrace_ops_ro(void)
433{
434 struct ftrace_ops *ops;
435 unsigned long start_offset;
436 unsigned long end_offset;
437 unsigned long npages;
438 unsigned long size;
439
440 do_for_each_ftrace_op(ops, ftrace_ops_list) {
441 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
442 continue;
443
444 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
445 start_offset = (unsigned long)ftrace_regs_caller;
446 end_offset = (unsigned long)ftrace_regs_caller_end;
447 } else {
448 start_offset = (unsigned long)ftrace_caller;
449 end_offset = (unsigned long)ftrace_caller_end;
450 }
451 size = end_offset - start_offset;
452 size = size + RET_SIZE + sizeof(void *);
453 npages = DIV_ROUND_UP(size, PAGE_SIZE);
454 set_memory_ro((unsigned long)ops->trampoline, npages);
455 } while_for_each_ftrace_op(ops);
456}
457
458static unsigned long calc_trampoline_call_offset(bool save_regs)
459{
460 unsigned long start_offset;
461 unsigned long call_offset;
462
463 if (save_regs) {
464 start_offset = (unsigned long)ftrace_regs_caller;
465 call_offset = (unsigned long)ftrace_regs_call;
466 } else {
467 start_offset = (unsigned long)ftrace_caller;
468 call_offset = (unsigned long)ftrace_call;
469 }
470
471 return call_offset - start_offset;
472}
473
474void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
475{
476 ftrace_func_t func;
477 unsigned long offset;
478 unsigned long ip;
479 unsigned int size;
480 const char *new;
481
482 if (!ops->trampoline) {
483 ops->trampoline = create_trampoline(ops, &size);
484 if (!ops->trampoline)
485 return;
486 ops->trampoline_size = size;
487 return;
488 }
489
490 /*
491 * The ftrace_ops caller may set up its own trampoline.
492 * In such a case, this code must not modify it.
493 */
494 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
495 return;
496
497 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
498 ip = ops->trampoline + offset;
499 func = ftrace_ops_get_func(ops);
500
501 mutex_lock(&text_mutex);
502 /* Do a safe modify in case the trampoline is executing */
503 new = ftrace_call_replace(ip, (unsigned long)func);
504 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
505 mutex_unlock(&text_mutex);
506}
507
508/* Return the address of the function the trampoline calls */
509static void *addr_from_call(void *ptr)
510{
511 union text_poke_insn call;
512 int ret;
513
514 ret = copy_from_kernel_nofault(&call, ptr, CALL_INSN_SIZE);
515 if (WARN_ON_ONCE(ret < 0))
516 return NULL;
517
518 /* Make sure this is a call */
519 if (WARN_ON_ONCE(call.opcode != CALL_INSN_OPCODE)) {
520 pr_warn("Expected E8, got %x\n", call.opcode);
521 return NULL;
522 }
523
524 return ptr + CALL_INSN_SIZE + call.disp;
525}
526
527/*
528 * If the ops->trampoline was not allocated, then it probably
529 * has a static trampoline func, or is the ftrace caller itself.
530 */
531static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
532{
533 unsigned long offset;
534 bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
535 void *ptr;
536
537 if (ops && ops->trampoline) {
538#if !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS) && \
539 defined(CONFIG_FUNCTION_GRAPH_TRACER)
540 /*
541 * We only know about function graph tracer setting as static
542 * trampoline.
543 */
544 if (ops->trampoline == FTRACE_GRAPH_ADDR)
545 return (void *)prepare_ftrace_return;
546#endif
547 return NULL;
548 }
549
550 offset = calc_trampoline_call_offset(save_regs);
551
552 if (save_regs)
553 ptr = (void *)FTRACE_REGS_ADDR + offset;
554 else
555 ptr = (void *)FTRACE_ADDR + offset;
556
557 return addr_from_call(ptr);
558}
559
560void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
561{
562 unsigned long offset;
563
564 /* If we didn't allocate this trampoline, consider it static */
565 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
566 return static_tramp_func(ops, rec);
567
568 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
569 return addr_from_call((void *)ops->trampoline + offset);
570}
571
572void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
573{
574 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
575 return;
576
577 tramp_free((void *)ops->trampoline);
578 ops->trampoline = 0;
579}
580
581#endif /* CONFIG_X86_64 */
582#endif /* CONFIG_DYNAMIC_FTRACE */
583
584#ifdef CONFIG_FUNCTION_GRAPH_TRACER
585
586#if defined(CONFIG_DYNAMIC_FTRACE) && !defined(CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS)
587extern void ftrace_graph_call(void);
588static const char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
589{
590 return text_gen_insn(JMP32_INSN_OPCODE, (void *)ip, (void *)addr);
591}
592
593static int ftrace_mod_jmp(unsigned long ip, void *func)
594{
595 const char *new;
596
597 new = ftrace_jmp_replace(ip, (unsigned long)func);
598 text_poke_bp((void *)ip, new, MCOUNT_INSN_SIZE, NULL);
599 return 0;
600}
601
602int ftrace_enable_ftrace_graph_caller(void)
603{
604 unsigned long ip = (unsigned long)(&ftrace_graph_call);
605
606 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
607}
608
609int ftrace_disable_ftrace_graph_caller(void)
610{
611 unsigned long ip = (unsigned long)(&ftrace_graph_call);
612
613 return ftrace_mod_jmp(ip, &ftrace_stub);
614}
615#endif /* CONFIG_DYNAMIC_FTRACE && !CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS */
616
617/*
618 * Hook the return address and push it in the stack of return addrs
619 * in current thread info.
620 */
621void prepare_ftrace_return(unsigned long ip, unsigned long *parent,
622 unsigned long frame_pointer)
623{
624 unsigned long return_hooker = (unsigned long)&return_to_handler;
625 int bit;
626
627 /*
628 * When resuming from suspend-to-ram, this function can be indirectly
629 * called from early CPU startup code while the CPU is in real mode,
630 * which would fail miserably. Make sure the stack pointer is a
631 * virtual address.
632 *
633 * This check isn't as accurate as virt_addr_valid(), but it should be
634 * good enough for this purpose, and it's fast.
635 */
636 if (unlikely((long)__builtin_frame_address(0) >= 0))
637 return;
638
639 if (unlikely(ftrace_graph_is_dead()))
640 return;
641
642 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
643 return;
644
645 bit = ftrace_test_recursion_trylock(ip, *parent);
646 if (bit < 0)
647 return;
648
649 if (!function_graph_enter(*parent, ip, frame_pointer, parent))
650 *parent = return_hooker;
651
652 ftrace_test_recursion_unlock(bit);
653}
654
655#ifdef CONFIG_HAVE_DYNAMIC_FTRACE_WITH_ARGS
656void ftrace_graph_func(unsigned long ip, unsigned long parent_ip,
657 struct ftrace_ops *op, struct ftrace_regs *fregs)
658{
659 struct pt_regs *regs = &fregs->regs;
660 unsigned long *stack = (unsigned long *)kernel_stack_pointer(regs);
661
662 prepare_ftrace_return(ip, (unsigned long *)stack, 0);
663}
664#endif
665
666#endif /* CONFIG_FUNCTION_GRAPH_TRACER */