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