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/*
2 * Dynamic function tracing support.
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/slab.h>
21#include <linux/init.h>
22#include <linux/list.h>
23#include <linux/module.h>
24
25#include <trace/syscall.h>
26
27#include <asm/cacheflush.h>
28#include <asm/kprobes.h>
29#include <asm/ftrace.h>
30#include <asm/nops.h>
31
32#ifdef CONFIG_DYNAMIC_FTRACE
33
34int ftrace_arch_code_modify_prepare(void)
35{
36 set_kernel_text_rw();
37 set_all_modules_text_rw();
38 return 0;
39}
40
41int ftrace_arch_code_modify_post_process(void)
42{
43 set_all_modules_text_ro();
44 set_kernel_text_ro();
45 return 0;
46}
47
48union ftrace_code_union {
49 char code[MCOUNT_INSN_SIZE];
50 struct {
51 unsigned char e8;
52 int offset;
53 } __attribute__((packed));
54};
55
56static int ftrace_calc_offset(long ip, long addr)
57{
58 return (int)(addr - ip);
59}
60
61static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
62{
63 static union ftrace_code_union calc;
64
65 calc.e8 = 0xe8;
66 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
67
68 /*
69 * No locking needed, this must be called via kstop_machine
70 * which in essence is like running on a uniprocessor machine.
71 */
72 return calc.code;
73}
74
75static inline int
76within(unsigned long addr, unsigned long start, unsigned long end)
77{
78 return addr >= start && addr < end;
79}
80
81static unsigned long text_ip_addr(unsigned long ip)
82{
83 /*
84 * On x86_64, kernel text mappings are mapped read-only, so we use
85 * the kernel identity mapping instead of the kernel text mapping
86 * to modify the kernel text.
87 *
88 * For 32bit kernels, these mappings are same and we can use
89 * kernel identity mapping to modify code.
90 */
91 if (within(ip, (unsigned long)_text, (unsigned long)_etext))
92 ip = (unsigned long)__va(__pa_symbol(ip));
93
94 return ip;
95}
96
97static const unsigned char *ftrace_nop_replace(void)
98{
99 return ideal_nops[NOP_ATOMIC5];
100}
101
102static int
103ftrace_modify_code_direct(unsigned long ip, unsigned const char *old_code,
104 unsigned const char *new_code)
105{
106 unsigned char replaced[MCOUNT_INSN_SIZE];
107
108 ftrace_expected = old_code;
109
110 /*
111 * Note:
112 * We are paranoid about modifying text, as if a bug was to happen, it
113 * could cause us to read or write to someplace that could cause harm.
114 * Carefully read and modify the code with probe_kernel_*(), and make
115 * sure what we read is what we expected it to be before modifying it.
116 */
117
118 /* read the text we want to modify */
119 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
120 return -EFAULT;
121
122 /* Make sure it is what we expect it to be */
123 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
124 return -EINVAL;
125
126 ip = text_ip_addr(ip);
127
128 /* replace the text with the new text */
129 if (probe_kernel_write((void *)ip, new_code, MCOUNT_INSN_SIZE))
130 return -EPERM;
131
132 sync_core();
133
134 return 0;
135}
136
137int ftrace_make_nop(struct module *mod,
138 struct dyn_ftrace *rec, unsigned long addr)
139{
140 unsigned const char *new, *old;
141 unsigned long ip = rec->ip;
142
143 old = ftrace_call_replace(ip, addr);
144 new = ftrace_nop_replace();
145
146 /*
147 * On boot up, and when modules are loaded, the MCOUNT_ADDR
148 * is converted to a nop, and will never become MCOUNT_ADDR
149 * again. This code is either running before SMP (on boot up)
150 * or before the code will ever be executed (module load).
151 * We do not want to use the breakpoint version in this case,
152 * just modify the code directly.
153 */
154 if (addr == MCOUNT_ADDR)
155 return ftrace_modify_code_direct(rec->ip, old, new);
156
157 ftrace_expected = NULL;
158
159 /* Normal cases use add_brk_on_nop */
160 WARN_ONCE(1, "invalid use of ftrace_make_nop");
161 return -EINVAL;
162}
163
164int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
165{
166 unsigned const char *new, *old;
167 unsigned long ip = rec->ip;
168
169 old = ftrace_nop_replace();
170 new = ftrace_call_replace(ip, addr);
171
172 /* Should only be called when module is loaded */
173 return ftrace_modify_code_direct(rec->ip, old, new);
174}
175
176/*
177 * The modifying_ftrace_code is used to tell the breakpoint
178 * handler to call ftrace_int3_handler(). If it fails to
179 * call this handler for a breakpoint added by ftrace, then
180 * the kernel may crash.
181 *
182 * As atomic_writes on x86 do not need a barrier, we do not
183 * need to add smp_mb()s for this to work. It is also considered
184 * that we can not read the modifying_ftrace_code before
185 * executing the breakpoint. That would be quite remarkable if
186 * it could do that. Here's the flow that is required:
187 *
188 * CPU-0 CPU-1
189 *
190 * atomic_inc(mfc);
191 * write int3s
192 * <trap-int3> // implicit (r)mb
193 * if (atomic_read(mfc))
194 * call ftrace_int3_handler()
195 *
196 * Then when we are finished:
197 *
198 * atomic_dec(mfc);
199 *
200 * If we hit a breakpoint that was not set by ftrace, it does not
201 * matter if ftrace_int3_handler() is called or not. It will
202 * simply be ignored. But it is crucial that a ftrace nop/caller
203 * breakpoint is handled. No other user should ever place a
204 * breakpoint on an ftrace nop/caller location. It must only
205 * be done by this code.
206 */
207atomic_t modifying_ftrace_code __read_mostly;
208
209static int
210ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
211 unsigned const char *new_code);
212
213/*
214 * Should never be called:
215 * As it is only called by __ftrace_replace_code() which is called by
216 * ftrace_replace_code() that x86 overrides, and by ftrace_update_code()
217 * which is called to turn mcount into nops or nops into function calls
218 * but not to convert a function from not using regs to one that uses
219 * regs, which ftrace_modify_call() is for.
220 */
221int ftrace_modify_call(struct dyn_ftrace *rec, unsigned long old_addr,
222 unsigned long addr)
223{
224 WARN_ON(1);
225 ftrace_expected = NULL;
226 return -EINVAL;
227}
228
229static unsigned long ftrace_update_func;
230
231static int update_ftrace_func(unsigned long ip, void *new)
232{
233 unsigned char old[MCOUNT_INSN_SIZE];
234 int ret;
235
236 memcpy(old, (void *)ip, MCOUNT_INSN_SIZE);
237
238 ftrace_update_func = ip;
239 /* Make sure the breakpoints see the ftrace_update_func update */
240 smp_wmb();
241
242 /* See comment above by declaration of modifying_ftrace_code */
243 atomic_inc(&modifying_ftrace_code);
244
245 ret = ftrace_modify_code(ip, old, new);
246
247 atomic_dec(&modifying_ftrace_code);
248
249 return ret;
250}
251
252int ftrace_update_ftrace_func(ftrace_func_t func)
253{
254 unsigned long ip = (unsigned long)(&ftrace_call);
255 unsigned char *new;
256 int ret;
257
258 new = ftrace_call_replace(ip, (unsigned long)func);
259 ret = update_ftrace_func(ip, new);
260
261 /* Also update the regs callback function */
262 if (!ret) {
263 ip = (unsigned long)(&ftrace_regs_call);
264 new = ftrace_call_replace(ip, (unsigned long)func);
265 ret = update_ftrace_func(ip, new);
266 }
267
268 return ret;
269}
270
271static int is_ftrace_caller(unsigned long ip)
272{
273 if (ip == ftrace_update_func)
274 return 1;
275
276 return 0;
277}
278
279/*
280 * A breakpoint was added to the code address we are about to
281 * modify, and this is the handle that will just skip over it.
282 * We are either changing a nop into a trace call, or a trace
283 * call to a nop. While the change is taking place, we treat
284 * it just like it was a nop.
285 */
286int ftrace_int3_handler(struct pt_regs *regs)
287{
288 unsigned long ip;
289
290 if (WARN_ON_ONCE(!regs))
291 return 0;
292
293 ip = regs->ip - 1;
294 if (!ftrace_location(ip) && !is_ftrace_caller(ip))
295 return 0;
296
297 regs->ip += MCOUNT_INSN_SIZE - 1;
298
299 return 1;
300}
301
302static int ftrace_write(unsigned long ip, const char *val, int size)
303{
304 ip = text_ip_addr(ip);
305
306 if (probe_kernel_write((void *)ip, val, size))
307 return -EPERM;
308
309 return 0;
310}
311
312static int add_break(unsigned long ip, const char *old)
313{
314 unsigned char replaced[MCOUNT_INSN_SIZE];
315 unsigned char brk = BREAKPOINT_INSTRUCTION;
316
317 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
318 return -EFAULT;
319
320 ftrace_expected = old;
321
322 /* Make sure it is what we expect it to be */
323 if (memcmp(replaced, old, MCOUNT_INSN_SIZE) != 0)
324 return -EINVAL;
325
326 return ftrace_write(ip, &brk, 1);
327}
328
329static int add_brk_on_call(struct dyn_ftrace *rec, unsigned long addr)
330{
331 unsigned const char *old;
332 unsigned long ip = rec->ip;
333
334 old = ftrace_call_replace(ip, addr);
335
336 return add_break(rec->ip, old);
337}
338
339
340static int add_brk_on_nop(struct dyn_ftrace *rec)
341{
342 unsigned const char *old;
343
344 old = ftrace_nop_replace();
345
346 return add_break(rec->ip, old);
347}
348
349static int add_breakpoints(struct dyn_ftrace *rec, int enable)
350{
351 unsigned long ftrace_addr;
352 int ret;
353
354 ftrace_addr = ftrace_get_addr_curr(rec);
355
356 ret = ftrace_test_record(rec, enable);
357
358 switch (ret) {
359 case FTRACE_UPDATE_IGNORE:
360 return 0;
361
362 case FTRACE_UPDATE_MAKE_CALL:
363 /* converting nop to call */
364 return add_brk_on_nop(rec);
365
366 case FTRACE_UPDATE_MODIFY_CALL:
367 case FTRACE_UPDATE_MAKE_NOP:
368 /* converting a call to a nop */
369 return add_brk_on_call(rec, ftrace_addr);
370 }
371 return 0;
372}
373
374/*
375 * On error, we need to remove breakpoints. This needs to
376 * be done caefully. If the address does not currently have a
377 * breakpoint, we know we are done. Otherwise, we look at the
378 * remaining 4 bytes of the instruction. If it matches a nop
379 * we replace the breakpoint with the nop. Otherwise we replace
380 * it with the call instruction.
381 */
382static int remove_breakpoint(struct dyn_ftrace *rec)
383{
384 unsigned char ins[MCOUNT_INSN_SIZE];
385 unsigned char brk = BREAKPOINT_INSTRUCTION;
386 const unsigned char *nop;
387 unsigned long ftrace_addr;
388 unsigned long ip = rec->ip;
389
390 /* If we fail the read, just give up */
391 if (probe_kernel_read(ins, (void *)ip, MCOUNT_INSN_SIZE))
392 return -EFAULT;
393
394 /* If this does not have a breakpoint, we are done */
395 if (ins[0] != brk)
396 return 0;
397
398 nop = ftrace_nop_replace();
399
400 /*
401 * If the last 4 bytes of the instruction do not match
402 * a nop, then we assume that this is a call to ftrace_addr.
403 */
404 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0) {
405 /*
406 * For extra paranoidism, we check if the breakpoint is on
407 * a call that would actually jump to the ftrace_addr.
408 * If not, don't touch the breakpoint, we make just create
409 * a disaster.
410 */
411 ftrace_addr = ftrace_get_addr_new(rec);
412 nop = ftrace_call_replace(ip, ftrace_addr);
413
414 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) == 0)
415 goto update;
416
417 /* Check both ftrace_addr and ftrace_old_addr */
418 ftrace_addr = ftrace_get_addr_curr(rec);
419 nop = ftrace_call_replace(ip, ftrace_addr);
420
421 ftrace_expected = nop;
422
423 if (memcmp(&ins[1], &nop[1], MCOUNT_INSN_SIZE - 1) != 0)
424 return -EINVAL;
425 }
426
427 update:
428 return ftrace_write(ip, nop, 1);
429}
430
431static int add_update_code(unsigned long ip, unsigned const char *new)
432{
433 /* skip breakpoint */
434 ip++;
435 new++;
436 return ftrace_write(ip, new, MCOUNT_INSN_SIZE - 1);
437}
438
439static int add_update_call(struct dyn_ftrace *rec, unsigned long addr)
440{
441 unsigned long ip = rec->ip;
442 unsigned const char *new;
443
444 new = ftrace_call_replace(ip, addr);
445 return add_update_code(ip, new);
446}
447
448static int add_update_nop(struct dyn_ftrace *rec)
449{
450 unsigned long ip = rec->ip;
451 unsigned const char *new;
452
453 new = ftrace_nop_replace();
454 return add_update_code(ip, new);
455}
456
457static int add_update(struct dyn_ftrace *rec, int enable)
458{
459 unsigned long ftrace_addr;
460 int ret;
461
462 ret = ftrace_test_record(rec, enable);
463
464 ftrace_addr = ftrace_get_addr_new(rec);
465
466 switch (ret) {
467 case FTRACE_UPDATE_IGNORE:
468 return 0;
469
470 case FTRACE_UPDATE_MODIFY_CALL:
471 case FTRACE_UPDATE_MAKE_CALL:
472 /* converting nop to call */
473 return add_update_call(rec, ftrace_addr);
474
475 case FTRACE_UPDATE_MAKE_NOP:
476 /* converting a call to a nop */
477 return add_update_nop(rec);
478 }
479
480 return 0;
481}
482
483static int finish_update_call(struct dyn_ftrace *rec, unsigned long addr)
484{
485 unsigned long ip = rec->ip;
486 unsigned const char *new;
487
488 new = ftrace_call_replace(ip, addr);
489
490 return ftrace_write(ip, new, 1);
491}
492
493static int finish_update_nop(struct dyn_ftrace *rec)
494{
495 unsigned long ip = rec->ip;
496 unsigned const char *new;
497
498 new = ftrace_nop_replace();
499
500 return ftrace_write(ip, new, 1);
501}
502
503static int finish_update(struct dyn_ftrace *rec, int enable)
504{
505 unsigned long ftrace_addr;
506 int ret;
507
508 ret = ftrace_update_record(rec, enable);
509
510 ftrace_addr = ftrace_get_addr_new(rec);
511
512 switch (ret) {
513 case FTRACE_UPDATE_IGNORE:
514 return 0;
515
516 case FTRACE_UPDATE_MODIFY_CALL:
517 case FTRACE_UPDATE_MAKE_CALL:
518 /* converting nop to call */
519 return finish_update_call(rec, ftrace_addr);
520
521 case FTRACE_UPDATE_MAKE_NOP:
522 /* converting a call to a nop */
523 return finish_update_nop(rec);
524 }
525
526 return 0;
527}
528
529static void do_sync_core(void *data)
530{
531 sync_core();
532}
533
534static void run_sync(void)
535{
536 int enable_irqs = irqs_disabled();
537
538 /* We may be called with interrupts disbled (on bootup). */
539 if (enable_irqs)
540 local_irq_enable();
541 on_each_cpu(do_sync_core, NULL, 1);
542 if (enable_irqs)
543 local_irq_disable();
544}
545
546void ftrace_replace_code(int enable)
547{
548 struct ftrace_rec_iter *iter;
549 struct dyn_ftrace *rec;
550 const char *report = "adding breakpoints";
551 int count = 0;
552 int ret;
553
554 for_ftrace_rec_iter(iter) {
555 rec = ftrace_rec_iter_record(iter);
556
557 ret = add_breakpoints(rec, enable);
558 if (ret)
559 goto remove_breakpoints;
560 count++;
561 }
562
563 run_sync();
564
565 report = "updating code";
566 count = 0;
567
568 for_ftrace_rec_iter(iter) {
569 rec = ftrace_rec_iter_record(iter);
570
571 ret = add_update(rec, enable);
572 if (ret)
573 goto remove_breakpoints;
574 count++;
575 }
576
577 run_sync();
578
579 report = "removing breakpoints";
580 count = 0;
581
582 for_ftrace_rec_iter(iter) {
583 rec = ftrace_rec_iter_record(iter);
584
585 ret = finish_update(rec, enable);
586 if (ret)
587 goto remove_breakpoints;
588 count++;
589 }
590
591 run_sync();
592
593 return;
594
595 remove_breakpoints:
596 pr_warn("Failed on %s (%d):\n", report, count);
597 ftrace_bug(ret, rec);
598 for_ftrace_rec_iter(iter) {
599 rec = ftrace_rec_iter_record(iter);
600 /*
601 * Breakpoints are handled only when this function is in
602 * progress. The system could not work with them.
603 */
604 if (remove_breakpoint(rec))
605 BUG();
606 }
607 run_sync();
608}
609
610static int
611ftrace_modify_code(unsigned long ip, unsigned const char *old_code,
612 unsigned const char *new_code)
613{
614 int ret;
615
616 ret = add_break(ip, old_code);
617 if (ret)
618 goto out;
619
620 run_sync();
621
622 ret = add_update_code(ip, new_code);
623 if (ret)
624 goto fail_update;
625
626 run_sync();
627
628 ret = ftrace_write(ip, new_code, 1);
629 /*
630 * The breakpoint is handled only when this function is in progress.
631 * The system could not work if we could not remove it.
632 */
633 BUG_ON(ret);
634 out:
635 run_sync();
636 return ret;
637
638 fail_update:
639 /* Also here the system could not work with the breakpoint */
640 if (ftrace_write(ip, old_code, 1))
641 BUG();
642 goto out;
643}
644
645void arch_ftrace_update_code(int command)
646{
647 /* See comment above by declaration of modifying_ftrace_code */
648 atomic_inc(&modifying_ftrace_code);
649
650 ftrace_modify_all_code(command);
651
652 atomic_dec(&modifying_ftrace_code);
653}
654
655int __init ftrace_dyn_arch_init(void)
656{
657 return 0;
658}
659
660#if defined(CONFIG_X86_64) || defined(CONFIG_FUNCTION_GRAPH_TRACER)
661static unsigned char *ftrace_jmp_replace(unsigned long ip, unsigned long addr)
662{
663 static union ftrace_code_union calc;
664
665 /* Jmp not a call (ignore the .e8) */
666 calc.e8 = 0xe9;
667 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
668
669 /*
670 * ftrace external locks synchronize the access to the static variable.
671 */
672 return calc.code;
673}
674#endif
675
676/* Currently only x86_64 supports dynamic trampolines */
677#ifdef CONFIG_X86_64
678
679#ifdef CONFIG_MODULES
680#include <linux/moduleloader.h>
681/* Module allocation simplifies allocating memory for code */
682static inline void *alloc_tramp(unsigned long size)
683{
684 return module_alloc(size);
685}
686static inline void tramp_free(void *tramp)
687{
688 module_memfree(tramp);
689}
690#else
691/* Trampolines can only be created if modules are supported */
692static inline void *alloc_tramp(unsigned long size)
693{
694 return NULL;
695}
696static inline void tramp_free(void *tramp) { }
697#endif
698
699/* Defined as markers to the end of the ftrace default trampolines */
700extern void ftrace_regs_caller_end(void);
701extern void ftrace_epilogue(void);
702extern void ftrace_caller_op_ptr(void);
703extern void ftrace_regs_caller_op_ptr(void);
704
705/* movq function_trace_op(%rip), %rdx */
706/* 0x48 0x8b 0x15 <offset-to-ftrace_trace_op (4 bytes)> */
707#define OP_REF_SIZE 7
708
709/*
710 * The ftrace_ops is passed to the function callback. Since the
711 * trampoline only services a single ftrace_ops, we can pass in
712 * that ops directly.
713 *
714 * The ftrace_op_code_union is used to create a pointer to the
715 * ftrace_ops that will be passed to the callback function.
716 */
717union ftrace_op_code_union {
718 char code[OP_REF_SIZE];
719 struct {
720 char op[3];
721 int offset;
722 } __attribute__((packed));
723};
724
725static unsigned long
726create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
727{
728 unsigned const char *jmp;
729 unsigned long start_offset;
730 unsigned long end_offset;
731 unsigned long op_offset;
732 unsigned long offset;
733 unsigned long size;
734 unsigned long ip;
735 unsigned long *ptr;
736 void *trampoline;
737 /* 48 8b 15 <offset> is movq <offset>(%rip), %rdx */
738 unsigned const char op_ref[] = { 0x48, 0x8b, 0x15 };
739 union ftrace_op_code_union op_ptr;
740 int ret;
741
742 if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
743 start_offset = (unsigned long)ftrace_regs_caller;
744 end_offset = (unsigned long)ftrace_regs_caller_end;
745 op_offset = (unsigned long)ftrace_regs_caller_op_ptr;
746 } else {
747 start_offset = (unsigned long)ftrace_caller;
748 end_offset = (unsigned long)ftrace_epilogue;
749 op_offset = (unsigned long)ftrace_caller_op_ptr;
750 }
751
752 size = end_offset - start_offset;
753
754 /*
755 * Allocate enough size to store the ftrace_caller code,
756 * the jmp to ftrace_epilogue, as well as the address of
757 * the ftrace_ops this trampoline is used for.
758 */
759 trampoline = alloc_tramp(size + MCOUNT_INSN_SIZE + sizeof(void *));
760 if (!trampoline)
761 return 0;
762
763 *tramp_size = size + MCOUNT_INSN_SIZE + sizeof(void *);
764
765 /* Copy ftrace_caller onto the trampoline memory */
766 ret = probe_kernel_read(trampoline, (void *)start_offset, size);
767 if (WARN_ON(ret < 0)) {
768 tramp_free(trampoline);
769 return 0;
770 }
771
772 ip = (unsigned long)trampoline + size;
773
774 /* The trampoline ends with a jmp to ftrace_epilogue */
775 jmp = ftrace_jmp_replace(ip, (unsigned long)ftrace_epilogue);
776 memcpy(trampoline + size, jmp, MCOUNT_INSN_SIZE);
777
778 /*
779 * The address of the ftrace_ops that is used for this trampoline
780 * is stored at the end of the trampoline. This will be used to
781 * load the third parameter for the callback. Basically, that
782 * location at the end of the trampoline takes the place of
783 * the global function_trace_op variable.
784 */
785
786 ptr = (unsigned long *)(trampoline + size + MCOUNT_INSN_SIZE);
787 *ptr = (unsigned long)ops;
788
789 op_offset -= start_offset;
790 memcpy(&op_ptr, trampoline + op_offset, OP_REF_SIZE);
791
792 /* Are we pointing to the reference? */
793 if (WARN_ON(memcmp(op_ptr.op, op_ref, 3) != 0)) {
794 tramp_free(trampoline);
795 return 0;
796 }
797
798 /* Load the contents of ptr into the callback parameter */
799 offset = (unsigned long)ptr;
800 offset -= (unsigned long)trampoline + op_offset + OP_REF_SIZE;
801
802 op_ptr.offset = offset;
803
804 /* put in the new offset to the ftrace_ops */
805 memcpy(trampoline + op_offset, &op_ptr, OP_REF_SIZE);
806
807 /* ALLOC_TRAMP flags lets us know we created it */
808 ops->flags |= FTRACE_OPS_FL_ALLOC_TRAMP;
809
810 return (unsigned long)trampoline;
811}
812
813static unsigned long calc_trampoline_call_offset(bool save_regs)
814{
815 unsigned long start_offset;
816 unsigned long call_offset;
817
818 if (save_regs) {
819 start_offset = (unsigned long)ftrace_regs_caller;
820 call_offset = (unsigned long)ftrace_regs_call;
821 } else {
822 start_offset = (unsigned long)ftrace_caller;
823 call_offset = (unsigned long)ftrace_call;
824 }
825
826 return call_offset - start_offset;
827}
828
829void arch_ftrace_update_trampoline(struct ftrace_ops *ops)
830{
831 ftrace_func_t func;
832 unsigned char *new;
833 unsigned long offset;
834 unsigned long ip;
835 unsigned int size;
836 int ret;
837
838 if (ops->trampoline) {
839 /*
840 * The ftrace_ops caller may set up its own trampoline.
841 * In such a case, this code must not modify it.
842 */
843 if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
844 return;
845 } else {
846 ops->trampoline = create_trampoline(ops, &size);
847 if (!ops->trampoline)
848 return;
849 ops->trampoline_size = size;
850 }
851
852 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
853 ip = ops->trampoline + offset;
854
855 func = ftrace_ops_get_func(ops);
856
857 /* Do a safe modify in case the trampoline is executing */
858 new = ftrace_call_replace(ip, (unsigned long)func);
859 ret = update_ftrace_func(ip, new);
860
861 /* The update should never fail */
862 WARN_ON(ret);
863}
864
865/* Return the address of the function the trampoline calls */
866static void *addr_from_call(void *ptr)
867{
868 union ftrace_code_union calc;
869 int ret;
870
871 ret = probe_kernel_read(&calc, ptr, MCOUNT_INSN_SIZE);
872 if (WARN_ON_ONCE(ret < 0))
873 return NULL;
874
875 /* Make sure this is a call */
876 if (WARN_ON_ONCE(calc.e8 != 0xe8)) {
877 pr_warn("Expected e8, got %x\n", calc.e8);
878 return NULL;
879 }
880
881 return ptr + MCOUNT_INSN_SIZE + calc.offset;
882}
883
884void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
885 unsigned long frame_pointer);
886
887/*
888 * If the ops->trampoline was not allocated, then it probably
889 * has a static trampoline func, or is the ftrace caller itself.
890 */
891static void *static_tramp_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
892{
893 unsigned long offset;
894 bool save_regs = rec->flags & FTRACE_FL_REGS_EN;
895 void *ptr;
896
897 if (ops && ops->trampoline) {
898#ifdef CONFIG_FUNCTION_GRAPH_TRACER
899 /*
900 * We only know about function graph tracer setting as static
901 * trampoline.
902 */
903 if (ops->trampoline == FTRACE_GRAPH_ADDR)
904 return (void *)prepare_ftrace_return;
905#endif
906 return NULL;
907 }
908
909 offset = calc_trampoline_call_offset(save_regs);
910
911 if (save_regs)
912 ptr = (void *)FTRACE_REGS_ADDR + offset;
913 else
914 ptr = (void *)FTRACE_ADDR + offset;
915
916 return addr_from_call(ptr);
917}
918
919void *arch_ftrace_trampoline_func(struct ftrace_ops *ops, struct dyn_ftrace *rec)
920{
921 unsigned long offset;
922
923 /* If we didn't allocate this trampoline, consider it static */
924 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
925 return static_tramp_func(ops, rec);
926
927 offset = calc_trampoline_call_offset(ops->flags & FTRACE_OPS_FL_SAVE_REGS);
928 return addr_from_call((void *)ops->trampoline + offset);
929}
930
931void arch_ftrace_trampoline_free(struct ftrace_ops *ops)
932{
933 if (!ops || !(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
934 return;
935
936 tramp_free((void *)ops->trampoline);
937 ops->trampoline = 0;
938}
939
940#endif /* CONFIG_X86_64 */
941#endif /* CONFIG_DYNAMIC_FTRACE */
942
943#ifdef CONFIG_FUNCTION_GRAPH_TRACER
944
945#ifdef CONFIG_DYNAMIC_FTRACE
946extern void ftrace_graph_call(void);
947
948static int ftrace_mod_jmp(unsigned long ip, void *func)
949{
950 unsigned char *new;
951
952 new = ftrace_jmp_replace(ip, (unsigned long)func);
953
954 return update_ftrace_func(ip, new);
955}
956
957int ftrace_enable_ftrace_graph_caller(void)
958{
959 unsigned long ip = (unsigned long)(&ftrace_graph_call);
960
961 return ftrace_mod_jmp(ip, &ftrace_graph_caller);
962}
963
964int ftrace_disable_ftrace_graph_caller(void)
965{
966 unsigned long ip = (unsigned long)(&ftrace_graph_call);
967
968 return ftrace_mod_jmp(ip, &ftrace_stub);
969}
970
971#endif /* !CONFIG_DYNAMIC_FTRACE */
972
973/*
974 * Hook the return address and push it in the stack of return addrs
975 * in current thread info.
976 */
977void prepare_ftrace_return(unsigned long self_addr, unsigned long *parent,
978 unsigned long frame_pointer)
979{
980 unsigned long old;
981 int faulted;
982 struct ftrace_graph_ent trace;
983 unsigned long return_hooker = (unsigned long)
984 &return_to_handler;
985
986 if (unlikely(ftrace_graph_is_dead()))
987 return;
988
989 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
990 return;
991
992 /*
993 * Protect against fault, even if it shouldn't
994 * happen. This tool is too much intrusive to
995 * ignore such a protection.
996 */
997 asm volatile(
998 "1: " _ASM_MOV " (%[parent]), %[old]\n"
999 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
1000 " movl $0, %[faulted]\n"
1001 "3:\n"
1002
1003 ".section .fixup, \"ax\"\n"
1004 "4: movl $1, %[faulted]\n"
1005 " jmp 3b\n"
1006 ".previous\n"
1007
1008 _ASM_EXTABLE(1b, 4b)
1009 _ASM_EXTABLE(2b, 4b)
1010
1011 : [old] "=&r" (old), [faulted] "=r" (faulted)
1012 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
1013 : "memory"
1014 );
1015
1016 if (unlikely(faulted)) {
1017 ftrace_graph_stop();
1018 WARN_ON(1);
1019 return;
1020 }
1021
1022 trace.func = self_addr;
1023 trace.depth = current->curr_ret_stack + 1;
1024
1025 /* Only trace if the calling function expects to */
1026 if (!ftrace_graph_entry(&trace)) {
1027 *parent = old;
1028 return;
1029 }
1030
1031 if (ftrace_push_return_trace(old, self_addr, &trace.depth,
1032 frame_pointer, parent) == -EBUSY) {
1033 *parent = old;
1034 return;
1035 }
1036}
1037#endif /* CONFIG_FUNCTION_GRAPH_TRACER */