arch/x86/kernel/ftrace.c

   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 #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/init.h>
  18 #include <linux/list.h>
  19
  20 #include <asm/ftrace.h>
  21 #include <asm/nops.h>
  22
  23
  24 /* Long is fine, even if it is only 4 bytes ;-) */
  25 static unsigned long *ftrace_nop;
  26
  27 union ftrace_code_union {
  28         char code[MCOUNT_INSN_SIZE];
  29         struct {
  30                 char e8;
  31                 int offset;
  32         } __attribute__((packed));
  33 };
  34
  35
  36 static int notrace ftrace_calc_offset(long ip, long addr)
  37 {
  38         return (int)(addr - ip);
  39 }
  40
  41 notrace unsigned char *ftrace_nop_replace(void)
  42 {
  43         return (char *)ftrace_nop;
  44 }
  45
  46 notrace unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
  47 {
  48         static union ftrace_code_union calc;
  49
  50         calc.e8         = 0xe8;
  51         calc.offset     = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
  52
  53         /*
  54          * No locking needed, this must be called via kstop_machine
  55          * which in essence is like running on a uniprocessor machine.
  56          */
  57         return calc.code;
  58 }
  59
  60 notrace int
  61 ftrace_modify_code(unsigned long ip, unsigned char *old_code,
  62                    unsigned char *new_code)
  63 {
  64         unsigned char replaced[MCOUNT_INSN_SIZE];
  65
  66         /*
  67          * Note: Due to modules and __init, code can
  68          *  disappear and change, we need to protect against faulting
  69          *  as well as code changing.
  70          *
  71          * No real locking needed, this code is run through
  72          * kstop_machine, or before SMP starts.
  73          */
  74         if (__copy_from_user_inatomic(replaced, (char __user *)ip, MCOUNT_INSN_SIZE))
  75                 return 1;
  76
  77         if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
  78                 return 2;
  79
  80         WARN_ON_ONCE(__copy_to_user_inatomic((char __user *)ip, new_code,
  81                                     MCOUNT_INSN_SIZE));
  82
  83         sync_core();
  84
  85         return 0;
  86 }
  87
  88 notrace int ftrace_update_ftrace_func(ftrace_func_t func)
  89 {
  90         unsigned long ip = (unsigned long)(&ftrace_call);
  91         unsigned char old[MCOUNT_INSN_SIZE], *new;
  92         int ret;
  93
  94         memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
  95         new = ftrace_call_replace(ip, (unsigned long)func);
  96         ret = ftrace_modify_code(ip, old, new);
  97
  98         return ret;
  99 }
 100
 101 notrace int ftrace_mcount_set(unsigned long *data)
 102 {
 103         /* mcount is initialized as a nop */
 104         *data = 0;
 105         return 0;
 106 }
 107
 108 int __init ftrace_dyn_arch_init(void *data)
 109 {
 110         extern const unsigned char ftrace_test_p6nop[];
 111         extern const unsigned char ftrace_test_nop5[];
 112         extern const unsigned char ftrace_test_jmp[];
 113         int faulted = 0;
 114
 115         /*
 116          * There is no good nop for all x86 archs.
 117          * We will default to using the P6_NOP5, but first we
 118          * will test to make sure that the nop will actually
 119          * work on this CPU. If it faults, we will then
 120          * go to a lesser efficient 5 byte nop. If that fails
 121          * we then just use a jmp as our nop. This isn't the most
 122          * efficient nop, but we can not use a multi part nop
 123          * since we would then risk being preempted in the middle
 124          * of that nop, and if we enabled tracing then, it might
 125          * cause a system crash.
 126          *
 127          * TODO: check the cpuid to determine the best nop.
 128          */
 129         asm volatile (
 130                 "jmp ftrace_test_jmp\n"
 131                 /* This code needs to stay around */
 132                 ".section .text, \"ax\"\n"
 133                 "ftrace_test_jmp:"
 134                 "jmp ftrace_test_p6nop\n"
 135                 "nop\n"
 136                 "nop\n"
 137                 "nop\n"  /* 2 byte jmp + 3 bytes */
 138                 "ftrace_test_p6nop:"
 139                 P6_NOP5
 140                 "jmp 1f\n"
 141                 "ftrace_test_nop5:"
 142                 ".byte 0x66,0x66,0x66,0x66,0x90\n"
 143                 "jmp 1f\n"
 144                 ".previous\n"
 145                 "1:"
 146                 ".section .fixup, \"ax\"\n"
 147                 "2:     movl $1, %0\n"
 148                 "       jmp ftrace_test_nop5\n"
 149                 "3:     movl $2, %0\n"
 150                 "       jmp 1b\n"
 151                 ".previous\n"
 152                 _ASM_EXTABLE(ftrace_test_p6nop, 2b)
 153                 _ASM_EXTABLE(ftrace_test_nop5, 3b)
 154                 : "=r"(faulted) : "0" (faulted));
 155
 156         switch (faulted) {
 157         case 0:
 158                 pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n");
 159                 ftrace_nop = (unsigned long *)ftrace_test_p6nop;
 160                 break;
 161         case 1:
 162                 pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n");
 163                 ftrace_nop = (unsigned long *)ftrace_test_nop5;
 164                 break;
 165         case 2:
 166                 pr_info("ftrace: converting mcount calls to jmp . + 5\n");
 167                 ftrace_nop = (unsigned long *)ftrace_test_jmp;
 168                 break;
 169         }
 170
 171         /* The return code is retured via data */
 172         *(unsigned long *)data = 0;
 173
 174         return 0;
 175 }