2 * Copyright (C) 2008 Matt Fleming <matt@console-pimps.org>
3 * Copyright (C) 2008 Paul Mundt <lethal@linux-sh.org>
5 * Code for replacing ftrace calls with jumps.
7 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com>
9 * Thanks goes to Ingo Molnar, for suggesting the idea.
10 * Mathieu Desnoyers, for suggesting postponing the modifications.
11 * Arjan van de Ven, for keeping me straight, and explaining to me
12 * the dangers of modifying code on the run.
14 #include <linux/uaccess.h>
15 #include <linux/ftrace.h>
16 #include <linux/string.h>
17 #include <linux/init.h>
19 #include <linux/kernel.h>
20 #include <asm/ftrace.h>
21 #include <asm/cacheflush.h>
22 #include <asm/unistd.h>
23 #include <trace/syscall.h>
25 #ifdef CONFIG_DYNAMIC_FTRACE
26 static unsigned char ftrace_replaced_code[MCOUNT_INSN_SIZE];
28 static unsigned char ftrace_nop[4];
30 * If we're trying to nop out a call to a function, we instead
31 * place a call to the address after the memory table.
34 * 8c011060: 02 d1 mov.l 8c01106c <a+0xc>,r1
35 * 8c011062: 22 4f sts.l pr,@-r15
36 * 8c011064: 02 c7 mova 8c011070 <a+0x10>,r0
37 * 8c011066: 2b 41 jmp @r1
38 * 8c011068: 2a 40 lds r0,pr
40 * 8c01106c: 68 24 .word 0x2468 <--- ip
41 * 8c01106e: 1d 8c .word 0x8c1d
42 * 8c011070: 26 4f lds.l @r15+,pr <--- ip + MCOUNT_INSN_SIZE
44 * We write 0x8c011070 to 0x8c01106c so that on entry to a() we branch
45 * past the _mcount call and continue executing code like normal.
47 static unsigned char *ftrace_nop_replace(unsigned long ip)
49 __raw_writel(ip + MCOUNT_INSN_SIZE, ftrace_nop);
53 static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
55 /* Place the address in the memory table. */
56 __raw_writel(addr, ftrace_replaced_code);
59 * No locking needed, this must be called via kstop_machine
60 * which in essence is like running on a uniprocessor machine.
62 return ftrace_replaced_code;
66 * Modifying code must take extra care. On an SMP machine, if
67 * the code being modified is also being executed on another CPU
68 * that CPU will have undefined results and possibly take a GPF.
69 * We use kstop_machine to stop other CPUS from exectuing code.
70 * But this does not stop NMIs from happening. We still need
71 * to protect against that. We separate out the modification of
72 * the code to take care of this.
74 * Two buffers are added: An IP buffer and a "code" buffer.
76 * 1) Put the instruction pointer into the IP buffer
77 * and the new code into the "code" buffer.
78 * 2) Wait for any running NMIs to finish and set a flag that says
79 * we are modifying code, it is done in an atomic operation.
82 * 5) Wait for any running NMIs to finish.
84 * If an NMI is executed, the first thing it does is to call
85 * "ftrace_nmi_enter". This will check if the flag is set to write
86 * and if it is, it will write what is in the IP and "code" buffers.
88 * The trick is, it does not matter if everyone is writing the same
89 * content to the code location. Also, if a CPU is executing code
90 * it is OK to write to that code location if the contents being written
91 * are the same as what exists.
93 #define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */
94 static atomic_t nmi_running = ATOMIC_INIT(0);
95 static int mod_code_status; /* holds return value of text write */
96 static void *mod_code_ip; /* holds the IP to write to */
97 static void *mod_code_newcode; /* holds the text to write to the IP */
99 static unsigned nmi_wait_count;
100 static atomic_t nmi_update_count = ATOMIC_INIT(0);
102 int ftrace_arch_read_dyn_info(char *buf, int size)
106 r = snprintf(buf, size, "%u %u",
108 atomic_read(&nmi_update_count));
112 static void clear_mod_flag(void)
114 int old = atomic_read(&nmi_running);
117 int new = old & ~MOD_CODE_WRITE_FLAG;
122 old = atomic_cmpxchg(&nmi_running, old, new);
126 static void ftrace_mod_code(void)
129 * Yes, more than one CPU process can be writing to mod_code_status.
130 * (and the code itself)
131 * But if one were to fail, then they all should, and if one were
132 * to succeed, then they all should.
134 mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
137 /* if we fail, then kill any new writers */
142 void ftrace_nmi_enter(void)
144 if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) {
147 atomic_inc(&nmi_update_count);
149 /* Must have previous changes seen before executions */
153 void ftrace_nmi_exit(void)
155 /* Finish all executions before clearing nmi_running */
157 atomic_dec(&nmi_running);
160 static void wait_for_nmi_and_set_mod_flag(void)
162 if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG))
167 } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG));
172 static void wait_for_nmi(void)
174 if (!atomic_read(&nmi_running))
179 } while (atomic_read(&nmi_running));
185 do_ftrace_mod_code(unsigned long ip, void *new_code)
187 mod_code_ip = (void *)ip;
188 mod_code_newcode = new_code;
190 /* The buffers need to be visible before we let NMIs write them */
193 wait_for_nmi_and_set_mod_flag();
195 /* Make sure all running NMIs have finished before we write the code */
200 /* Make sure the write happens before clearing the bit */
206 return mod_code_status;
209 static int ftrace_modify_code(unsigned long ip, unsigned char *old_code,
210 unsigned char *new_code)
212 unsigned char replaced[MCOUNT_INSN_SIZE];
216 * We are paranoid about modifying text, as if a bug was to happen, it
217 * could cause us to read or write to someplace that could cause harm.
218 * Carefully read and modify the code with probe_kernel_*(), and make
219 * sure what we read is what we expected it to be before modifying it.
222 /* read the text we want to modify */
223 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
226 /* Make sure it is what we expect it to be */
227 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
230 /* replace the text with the new text */
231 if (do_ftrace_mod_code(ip, new_code))
234 flush_icache_range(ip, ip + MCOUNT_INSN_SIZE);
239 int ftrace_update_ftrace_func(ftrace_func_t func)
241 unsigned long ip = (unsigned long)(&ftrace_call) + MCOUNT_INSN_OFFSET;
242 unsigned char old[MCOUNT_INSN_SIZE], *new;
244 memcpy(old, (unsigned char *)ip, MCOUNT_INSN_SIZE);
245 new = ftrace_call_replace(ip, (unsigned long)func);
247 return ftrace_modify_code(ip, old, new);
250 int ftrace_make_nop(struct module *mod,
251 struct dyn_ftrace *rec, unsigned long addr)
253 unsigned char *new, *old;
254 unsigned long ip = rec->ip;
256 old = ftrace_call_replace(ip, addr);
257 new = ftrace_nop_replace(ip);
259 return ftrace_modify_code(rec->ip, old, new);
262 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
264 unsigned char *new, *old;
265 unsigned long ip = rec->ip;
267 old = ftrace_nop_replace(ip);
268 new = ftrace_call_replace(ip, addr);
270 return ftrace_modify_code(rec->ip, old, new);
273 int __init ftrace_dyn_arch_init(void)
277 #endif /* CONFIG_DYNAMIC_FTRACE */
279 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
280 #ifdef CONFIG_DYNAMIC_FTRACE
281 extern void ftrace_graph_call(void);
283 static int ftrace_mod(unsigned long ip, unsigned long old_addr,
284 unsigned long new_addr)
286 unsigned char code[MCOUNT_INSN_SIZE];
288 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
291 if (old_addr != __raw_readl((unsigned long *)code))
294 __raw_writel(new_addr, ip);
298 int ftrace_enable_ftrace_graph_caller(void)
300 unsigned long ip, old_addr, new_addr;
302 ip = (unsigned long)(&ftrace_graph_call) + GRAPH_INSN_OFFSET;
303 old_addr = (unsigned long)(&skip_trace);
304 new_addr = (unsigned long)(&ftrace_graph_caller);
306 return ftrace_mod(ip, old_addr, new_addr);
309 int ftrace_disable_ftrace_graph_caller(void)
311 unsigned long ip, old_addr, new_addr;
313 ip = (unsigned long)(&ftrace_graph_call) + GRAPH_INSN_OFFSET;
314 old_addr = (unsigned long)(&ftrace_graph_caller);
315 new_addr = (unsigned long)(&skip_trace);
317 return ftrace_mod(ip, old_addr, new_addr);
319 #endif /* CONFIG_DYNAMIC_FTRACE */
322 * Hook the return address and push it in the stack of return addrs
323 * in the current thread info.
325 * This is the main routine for the function graph tracer. The function
326 * graph tracer essentially works like this:
328 * parent is the stack address containing self_addr's return address.
329 * We pull the real return address out of parent and store it in
330 * current's ret_stack. Then, we replace the return address on the stack
331 * with the address of return_to_handler. self_addr is the function that
334 * When self_addr returns, it will jump to return_to_handler which calls
335 * ftrace_return_to_handler. ftrace_return_to_handler will pull the real
336 * return address off of current's ret_stack and jump to it.
338 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
342 struct ftrace_graph_ent trace;
343 unsigned long return_hooker = (unsigned long)&return_to_handler;
345 if (unlikely(ftrace_graph_is_dead()))
348 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
352 * Protect against fault, even if it shouldn't
353 * happen. This tool is too much intrusive to
354 * ignore such a protection.
356 __asm__ __volatile__(
363 ".section .fixup, \"ax\" \n\t"
371 ".section __ex_table,\"a\" \n\t"
375 : "=&r" (old), "=r" (faulted)
376 : "r" (parent), "r" (return_hooker)
379 if (unlikely(faulted)) {
385 err = ftrace_push_return_trace(old, self_addr, &trace.depth, 0, NULL);
387 __raw_writel(old, parent);
391 trace.func = self_addr;
393 /* Only trace if the calling function expects to */
394 if (!ftrace_graph_entry(&trace)) {
395 current->curr_ret_stack--;
396 __raw_writel(old, parent);
399 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */