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 <trace/syscall.h>
23 #include <asm/cacheflush.h>
24 #include <asm/ftrace.h>
29 #ifdef CONFIG_DYNAMIC_FTRACE
31 int ftrace_arch_code_modify_prepare(void)
37 int ftrace_arch_code_modify_post_process(void)
43 union ftrace_code_union {
44 char code[MCOUNT_INSN_SIZE];
48 } __attribute__((packed));
51 static int ftrace_calc_offset(long ip, long addr)
53 return (int)(addr - ip);
56 static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
58 static union ftrace_code_union calc;
61 calc.offset = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
64 * No locking needed, this must be called via kstop_machine
65 * which in essence is like running on a uniprocessor machine.
71 * Modifying code must take extra care. On an SMP machine, if
72 * the code being modified is also being executed on another CPU
73 * that CPU will have undefined results and possibly take a GPF.
74 * We use kstop_machine to stop other CPUS from exectuing code.
75 * But this does not stop NMIs from happening. We still need
76 * to protect against that. We separate out the modification of
77 * the code to take care of this.
79 * Two buffers are added: An IP buffer and a "code" buffer.
81 * 1) Put the instruction pointer into the IP buffer
82 * and the new code into the "code" buffer.
83 * 2) Wait for any running NMIs to finish and set a flag that says
84 * we are modifying code, it is done in an atomic operation.
87 * 5) Wait for any running NMIs to finish.
89 * If an NMI is executed, the first thing it does is to call
90 * "ftrace_nmi_enter". This will check if the flag is set to write
91 * and if it is, it will write what is in the IP and "code" buffers.
93 * The trick is, it does not matter if everyone is writing the same
94 * content to the code location. Also, if a CPU is executing code
95 * it is OK to write to that code location if the contents being written
96 * are the same as what exists.
99 #define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */
100 static atomic_t nmi_running = ATOMIC_INIT(0);
101 static int mod_code_status; /* holds return value of text write */
102 static void *mod_code_ip; /* holds the IP to write to */
103 static void *mod_code_newcode; /* holds the text to write to the IP */
105 static unsigned nmi_wait_count;
106 static atomic_t nmi_update_count = ATOMIC_INIT(0);
108 int ftrace_arch_read_dyn_info(char *buf, int size)
112 r = snprintf(buf, size, "%u %u",
114 atomic_read(&nmi_update_count));
118 static void clear_mod_flag(void)
120 int old = atomic_read(&nmi_running);
123 int new = old & ~MOD_CODE_WRITE_FLAG;
128 old = atomic_cmpxchg(&nmi_running, old, new);
132 static void ftrace_mod_code(void)
135 * Yes, more than one CPU process can be writing to mod_code_status.
136 * (and the code itself)
137 * But if one were to fail, then they all should, and if one were
138 * to succeed, then they all should.
140 mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
143 /* if we fail, then kill any new writers */
148 void ftrace_nmi_enter(void)
150 if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) {
153 atomic_inc(&nmi_update_count);
155 /* Must have previous changes seen before executions */
159 void ftrace_nmi_exit(void)
161 /* Finish all executions before clearing nmi_running */
163 atomic_dec(&nmi_running);
166 static void wait_for_nmi_and_set_mod_flag(void)
168 if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG))
173 } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG));
178 static void wait_for_nmi(void)
180 if (!atomic_read(&nmi_running))
185 } while (atomic_read(&nmi_running));
191 do_ftrace_mod_code(unsigned long ip, void *new_code)
193 mod_code_ip = (void *)ip;
194 mod_code_newcode = new_code;
196 /* The buffers need to be visible before we let NMIs write them */
199 wait_for_nmi_and_set_mod_flag();
201 /* Make sure all running NMIs have finished before we write the code */
206 /* Make sure the write happens before clearing the bit */
212 return mod_code_status;
218 static unsigned char ftrace_nop[MCOUNT_INSN_SIZE];
220 static unsigned char *ftrace_nop_replace(void)
226 ftrace_modify_code(unsigned long ip, unsigned char *old_code,
227 unsigned char *new_code)
229 unsigned char replaced[MCOUNT_INSN_SIZE];
232 * Note: Due to modules and __init, code can
233 * disappear and change, we need to protect against faulting
234 * as well as code changing. We do this by using the
235 * probe_kernel_* functions.
237 * No real locking needed, this code is run through
238 * kstop_machine, or before SMP starts.
241 /* read the text we want to modify */
242 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
245 /* Make sure it is what we expect it to be */
246 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
249 /* replace the text with the new text */
250 if (do_ftrace_mod_code(ip, new_code))
258 int ftrace_make_nop(struct module *mod,
259 struct dyn_ftrace *rec, unsigned long addr)
261 unsigned char *new, *old;
262 unsigned long ip = rec->ip;
264 old = ftrace_call_replace(ip, addr);
265 new = ftrace_nop_replace();
267 return ftrace_modify_code(rec->ip, old, new);
270 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
272 unsigned char *new, *old;
273 unsigned long ip = rec->ip;
275 old = ftrace_nop_replace();
276 new = ftrace_call_replace(ip, addr);
278 return ftrace_modify_code(rec->ip, old, new);
281 int ftrace_update_ftrace_func(ftrace_func_t func)
283 unsigned long ip = (unsigned long)(&ftrace_call);
284 unsigned char old[MCOUNT_INSN_SIZE], *new;
287 memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
288 new = ftrace_call_replace(ip, (unsigned long)func);
289 ret = ftrace_modify_code(ip, old, new);
294 int __init ftrace_dyn_arch_init(void *data)
296 extern const unsigned char ftrace_test_p6nop[];
297 extern const unsigned char ftrace_test_nop5[];
298 extern const unsigned char ftrace_test_jmp[];
302 * There is no good nop for all x86 archs.
303 * We will default to using the P6_NOP5, but first we
304 * will test to make sure that the nop will actually
305 * work on this CPU. If it faults, we will then
306 * go to a lesser efficient 5 byte nop. If that fails
307 * we then just use a jmp as our nop. This isn't the most
308 * efficient nop, but we can not use a multi part nop
309 * since we would then risk being preempted in the middle
310 * of that nop, and if we enabled tracing then, it might
311 * cause a system crash.
313 * TODO: check the cpuid to determine the best nop.
317 "jmp ftrace_test_p6nop\n"
320 "nop\n" /* 2 byte jmp + 3 bytes */
325 ".byte 0x66,0x66,0x66,0x66,0x90\n"
327 ".section .fixup, \"ax\"\n"
329 " jmp ftrace_test_nop5\n"
333 _ASM_EXTABLE(ftrace_test_p6nop, 2b)
334 _ASM_EXTABLE(ftrace_test_nop5, 3b)
335 : "=r"(faulted) : "0" (faulted));
339 pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n");
340 memcpy(ftrace_nop, ftrace_test_p6nop, MCOUNT_INSN_SIZE);
343 pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n");
344 memcpy(ftrace_nop, ftrace_test_nop5, MCOUNT_INSN_SIZE);
347 pr_info("ftrace: converting mcount calls to jmp . + 5\n");
348 memcpy(ftrace_nop, ftrace_test_jmp, MCOUNT_INSN_SIZE);
352 /* The return code is retured via data */
353 *(unsigned long *)data = 0;
359 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
361 #ifdef CONFIG_DYNAMIC_FTRACE
362 extern void ftrace_graph_call(void);
364 static int ftrace_mod_jmp(unsigned long ip,
365 int old_offset, int new_offset)
367 unsigned char code[MCOUNT_INSN_SIZE];
369 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
372 if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
375 *(int *)(&code[1]) = new_offset;
377 if (do_ftrace_mod_code(ip, &code))
383 int ftrace_enable_ftrace_graph_caller(void)
385 unsigned long ip = (unsigned long)(&ftrace_graph_call);
386 int old_offset, new_offset;
388 old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
389 new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
391 return ftrace_mod_jmp(ip, old_offset, new_offset);
394 int ftrace_disable_ftrace_graph_caller(void)
396 unsigned long ip = (unsigned long)(&ftrace_graph_call);
397 int old_offset, new_offset;
399 old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
400 new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
402 return ftrace_mod_jmp(ip, old_offset, new_offset);
405 #endif /* !CONFIG_DYNAMIC_FTRACE */
408 * Hook the return address and push it in the stack of return addrs
409 * in current thread info.
411 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
415 struct ftrace_graph_ent trace;
416 unsigned long return_hooker = (unsigned long)
419 /* Nmi's are currently unsupported */
420 if (unlikely(in_nmi()))
423 if (unlikely(atomic_read(¤t->tracing_graph_pause)))
427 * Protect against fault, even if it shouldn't
428 * happen. This tool is too much intrusive to
429 * ignore such a protection.
432 "1: " _ASM_MOV " (%[parent]), %[old]\n"
433 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
434 " movl $0, %[faulted]\n"
437 ".section .fixup, \"ax\"\n"
438 "4: movl $1, %[faulted]\n"
445 : [old] "=r" (old), [faulted] "=r" (faulted)
446 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
450 if (unlikely(faulted)) {
456 if (ftrace_push_return_trace(old, self_addr, &trace.depth) == -EBUSY) {
461 trace.func = self_addr;
463 /* Only trace if the calling function expects to */
464 if (!ftrace_graph_entry(&trace)) {
465 current->curr_ret_stack--;
469 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
471 #ifdef CONFIG_FTRACE_SYSCALLS
473 extern unsigned long __start_syscalls_metadata[];
474 extern unsigned long __stop_syscalls_metadata[];
475 extern unsigned long *sys_call_table;
477 static struct syscall_metadata **syscalls_metadata;
479 static struct syscall_metadata *find_syscall_meta(unsigned long *syscall)
481 struct syscall_metadata *start;
482 struct syscall_metadata *stop;
483 char str[KSYM_SYMBOL_LEN];
486 start = (struct syscall_metadata *)__start_syscalls_metadata;
487 stop = (struct syscall_metadata *)__stop_syscalls_metadata;
488 kallsyms_lookup((unsigned long) syscall, NULL, NULL, NULL, str);
490 for ( ; start < stop; start++) {
491 if (start->name && !strcmp(start->name, str))
497 struct syscall_metadata *syscall_nr_to_meta(int nr)
499 if (!syscalls_metadata || nr >= FTRACE_SYSCALL_MAX || nr < 0)
502 return syscalls_metadata[nr];
505 void arch_init_ftrace_syscalls(void)
508 struct syscall_metadata *meta;
509 unsigned long **psys_syscall_table = &sys_call_table;
510 static atomic_t refs;
512 if (atomic_inc_return(&refs) != 1)
515 syscalls_metadata = kzalloc(sizeof(*syscalls_metadata) *
516 FTRACE_SYSCALL_MAX, GFP_KERNEL);
517 if (!syscalls_metadata) {
522 for (i = 0; i < FTRACE_SYSCALL_MAX; i++) {
523 meta = find_syscall_meta(psys_syscall_table[i]);
524 syscalls_metadata[i] = meta;
528 /* Paranoid: avoid overflow */