2 * Code for replacing ftrace calls with jumps.
4 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
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.
12 #include <linux/spinlock.h>
13 #include <linux/hardirq.h>
14 #include <linux/uaccess.h>
15 #include <linux/ftrace.h>
16 #include <linux/percpu.h>
17 #include <linux/sched.h>
18 #include <linux/init.h>
19 #include <linux/list.h>
21 #include <asm/ftrace.h>
22 #include <linux/ftrace.h>
28 #ifdef CONFIG_FUNCTION_RET_TRACER
31 * These functions are picked from those used on
32 * this page for dynamic ftrace. They have been
33 * simplified to ignore all traces in NMI context.
35 static atomic_t in_nmi;
37 void ftrace_nmi_enter(void)
42 void ftrace_nmi_exit(void)
47 /* Add a function return address to the trace stack on thread info.*/
48 static int push_return_trace(unsigned long ret, unsigned long long time,
52 struct thread_info *ti = current_thread_info();
54 /* The return trace stack is full */
55 if (ti->curr_ret_stack == FTRACE_RET_STACK_SIZE - 1)
58 index = ++ti->curr_ret_stack;
59 ti->ret_stack[index].ret = ret;
60 ti->ret_stack[index].func = func;
61 ti->ret_stack[index].calltime = time;
66 /* Retrieve a function return address to the trace stack on thread info.*/
67 static void pop_return_trace(unsigned long *ret, unsigned long long *time,
72 struct thread_info *ti = current_thread_info();
73 index = ti->curr_ret_stack;
74 *ret = ti->ret_stack[index].ret;
75 *func = ti->ret_stack[index].func;
76 *time = ti->ret_stack[index].calltime;
81 * Send the trace to the ring-buffer.
82 * @return the original return address.
84 unsigned long ftrace_return_to_handler(void)
86 struct ftrace_retfunc trace;
87 pop_return_trace(&trace.ret, &trace.calltime, &trace.func);
88 trace.rettime = cpu_clock(raw_smp_processor_id());
89 ftrace_function_return(&trace);
95 * Hook the return address and push it in the stack of return addrs
96 * in current thread info.
98 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
101 unsigned long long calltime;
103 unsigned long return_hooker = (unsigned long)
106 /* Nmi's are currently unsupported */
107 if (atomic_read(&in_nmi))
111 * Protect against fault, even if it shouldn't
112 * happen. This tool is too much intrusive to
113 * ignore such a protection.
116 "1: movl (%[parent_old]), %[old]\n"
117 "2: movl %[return_hooker], (%[parent_replaced])\n"
118 " movl $0, %[faulted]\n"
120 ".section .fixup, \"ax\"\n"
121 "3: movl $1, %[faulted]\n"
124 ".section __ex_table, \"a\"\n"
129 : [parent_replaced] "=r" (parent), [old] "=r" (old),
130 [faulted] "=r" (faulted)
131 : [parent_old] "0" (parent), [return_hooker] "r" (return_hooker)
135 if (WARN_ON(faulted)) {
136 unregister_ftrace_return();
140 if (WARN_ON(!__kernel_text_address(old))) {
141 unregister_ftrace_return();
146 calltime = cpu_clock(raw_smp_processor_id());
148 if (push_return_trace(old, calltime, self_addr) == -EBUSY)
154 #ifdef CONFIG_DYNAMIC_FTRACE
156 union ftrace_code_union {
157 char code[MCOUNT_INSN_SIZE];
161 } __attribute__((packed));
164 static int ftrace_calc_offset(long ip, long addr)
166 return (int)(addr - ip);
169 static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
171 static union ftrace_code_union calc;
174 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
177 * No locking needed, this must be called via kstop_machine
178 * which in essence is like running on a uniprocessor machine.
184 * Modifying code must take extra care. On an SMP machine, if
185 * the code being modified is also being executed on another CPU
186 * that CPU will have undefined results and possibly take a GPF.
187 * We use kstop_machine to stop other CPUS from exectuing code.
188 * But this does not stop NMIs from happening. We still need
189 * to protect against that. We separate out the modification of
190 * the code to take care of this.
192 * Two buffers are added: An IP buffer and a "code" buffer.
194 * 1) Put the instruction pointer into the IP buffer
195 * and the new code into the "code" buffer.
196 * 2) Set a flag that says we are modifying code
197 * 3) Wait for any running NMIs to finish.
200 * 6) Wait for any running NMIs to finish.
202 * If an NMI is executed, the first thing it does is to call
203 * "ftrace_nmi_enter". This will check if the flag is set to write
204 * and if it is, it will write what is in the IP and "code" buffers.
206 * The trick is, it does not matter if everyone is writing the same
207 * content to the code location. Also, if a CPU is executing code
208 * it is OK to write to that code location if the contents being written
209 * are the same as what exists.
212 static atomic_t in_nmi = ATOMIC_INIT(0);
213 static int mod_code_status; /* holds return value of text write */
214 static int mod_code_write; /* set when NMI should do the write */
215 static void *mod_code_ip; /* holds the IP to write to */
216 static void *mod_code_newcode; /* holds the text to write to the IP */
218 static unsigned nmi_wait_count;
219 static atomic_t nmi_update_count = ATOMIC_INIT(0);
221 int ftrace_arch_read_dyn_info(char *buf, int size)
225 r = snprintf(buf, size, "%u %u",
227 atomic_read(&nmi_update_count));
231 static void ftrace_mod_code(void)
234 * Yes, more than one CPU process can be writing to mod_code_status.
235 * (and the code itself)
236 * But if one were to fail, then they all should, and if one were
237 * to succeed, then they all should.
239 mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
244 void ftrace_nmi_enter(void)
247 /* Must have in_nmi seen before reading write flag */
249 if (mod_code_write) {
251 atomic_inc(&nmi_update_count);
255 void ftrace_nmi_exit(void)
257 /* Finish all executions before clearing in_nmi */
262 static void wait_for_nmi(void)
266 while (atomic_read(&in_nmi)) {
276 do_ftrace_mod_code(unsigned long ip, void *new_code)
278 mod_code_ip = (void *)ip;
279 mod_code_newcode = new_code;
281 /* The buffers need to be visible before we let NMIs write them */
286 /* Make sure write bit is visible before we wait on NMIs */
291 /* Make sure all running NMIs have finished before we write the code */
296 /* Make sure the write happens before clearing the bit */
301 /* make sure NMIs see the cleared bit */
306 return mod_code_status;
312 static unsigned char ftrace_nop[MCOUNT_INSN_SIZE];
314 static unsigned char *ftrace_nop_replace(void)
320 ftrace_modify_code(unsigned long ip, unsigned char *old_code,
321 unsigned char *new_code)
323 unsigned char replaced[MCOUNT_INSN_SIZE];
326 * Note: Due to modules and __init, code can
327 * disappear and change, we need to protect against faulting
328 * as well as code changing. We do this by using the
329 * probe_kernel_* functions.
331 * No real locking needed, this code is run through
332 * kstop_machine, or before SMP starts.
335 /* read the text we want to modify */
336 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
339 /* Make sure it is what we expect it to be */
340 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
343 /* replace the text with the new text */
344 if (do_ftrace_mod_code(ip, new_code))
352 int ftrace_make_nop(struct module *mod,
353 struct dyn_ftrace *rec, unsigned long addr)
355 unsigned char *new, *old;
356 unsigned long ip = rec->ip;
358 old = ftrace_call_replace(ip, addr);
359 new = ftrace_nop_replace();
361 return ftrace_modify_code(rec->ip, old, new);
364 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
366 unsigned char *new, *old;
367 unsigned long ip = rec->ip;
369 old = ftrace_nop_replace();
370 new = ftrace_call_replace(ip, addr);
372 return ftrace_modify_code(rec->ip, old, new);
375 int ftrace_update_ftrace_func(ftrace_func_t func)
377 unsigned long ip = (unsigned long)(&ftrace_call);
378 unsigned char old[MCOUNT_INSN_SIZE], *new;
381 memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
382 new = ftrace_call_replace(ip, (unsigned long)func);
383 ret = ftrace_modify_code(ip, old, new);
388 int __init ftrace_dyn_arch_init(void *data)
390 extern const unsigned char ftrace_test_p6nop[];
391 extern const unsigned char ftrace_test_nop5[];
392 extern const unsigned char ftrace_test_jmp[];
396 * There is no good nop for all x86 archs.
397 * We will default to using the P6_NOP5, but first we
398 * will test to make sure that the nop will actually
399 * work on this CPU. If it faults, we will then
400 * go to a lesser efficient 5 byte nop. If that fails
401 * we then just use a jmp as our nop. This isn't the most
402 * efficient nop, but we can not use a multi part nop
403 * since we would then risk being preempted in the middle
404 * of that nop, and if we enabled tracing then, it might
405 * cause a system crash.
407 * TODO: check the cpuid to determine the best nop.
411 "jmp ftrace_test_p6nop\n"
414 "nop\n" /* 2 byte jmp + 3 bytes */
419 ".byte 0x66,0x66,0x66,0x66,0x90\n"
421 ".section .fixup, \"ax\"\n"
423 " jmp ftrace_test_nop5\n"
427 _ASM_EXTABLE(ftrace_test_p6nop, 2b)
428 _ASM_EXTABLE(ftrace_test_nop5, 3b)
429 : "=r"(faulted) : "0" (faulted));
433 pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n");
434 memcpy(ftrace_nop, ftrace_test_p6nop, MCOUNT_INSN_SIZE);
437 pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n");
438 memcpy(ftrace_nop, ftrace_test_nop5, MCOUNT_INSN_SIZE);
441 pr_info("ftrace: converting mcount calls to jmp . + 5\n");
442 memcpy(ftrace_nop, ftrace_test_jmp, MCOUNT_INSN_SIZE);
446 /* The return code is retured via data */
447 *(unsigned long *)data = 0;