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/sched.h>
  18 #include <linux/init.h>
  19 #include <linux/list.h>
  20
  21 #include <asm/ftrace.h>
  22 #include <linux/ftrace.h>
  23 #include <asm/nops.h>
  24 #include <asm/nmi.h>
  25
  26
  27 #ifdef CONFIG_DYNAMIC_FTRACE
  28
  29 union ftrace_code_union {
  30         char code[MCOUNT_INSN_SIZE];
  31         struct {
  32                 char e8;
  33                 int offset;
  34         } __attribute__((packed));
  35 };
  36
  37 static int ftrace_calc_offset(long ip, long addr)
  38 {
  39         return (int)(addr - ip);
  40 }
  41
  42 static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr)
  43 {
  44         static union ftrace_code_union calc;
  45
  46         calc.e8         = 0xe8;
  47         calc.offset     = ftrace_calc_offset(ip + MCOUNT_INSN_SIZE, addr);
  48
  49         /*
  50          * No locking needed, this must be called via kstop_machine
  51          * which in essence is like running on a uniprocessor machine.
  52          */
  53         return calc.code;
  54 }
  55
  56 /*
  57  * Modifying code must take extra care. On an SMP machine, if
  58  * the code being modified is also being executed on another CPU
  59  * that CPU will have undefined results and possibly take a GPF.
  60  * We use kstop_machine to stop other CPUS from exectuing code.
  61  * But this does not stop NMIs from happening. We still need
  62  * to protect against that. We separate out the modification of
  63  * the code to take care of this.
  64  *
  65  * Two buffers are added: An IP buffer and a "code" buffer.
  66  *
  67  * 1) Put the instruction pointer into the IP buffer
  68  *    and the new code into the "code" buffer.
  69  * 2) Set a flag that says we are modifying code
  70  * 3) Wait for any running NMIs to finish.
  71  * 4) Write the code
  72  * 5) clear the flag.
  73  * 6) Wait for any running NMIs to finish.
  74  *
  75  * If an NMI is executed, the first thing it does is to call
  76  * "ftrace_nmi_enter". This will check if the flag is set to write
  77  * and if it is, it will write what is in the IP and "code" buffers.
  78  *
  79  * The trick is, it does not matter if everyone is writing the same
  80  * content to the code location. Also, if a CPU is executing code
  81  * it is OK to write to that code location if the contents being written
  82  * are the same as what exists.
  83  */
  84
  85 static atomic_t in_nmi = ATOMIC_INIT(0);
  86 static int mod_code_status;             /* holds return value of text write */
  87 static int mod_code_write;              /* set when NMI should do the write */
  88 static void *mod_code_ip;               /* holds the IP to write to */
  89 static void *mod_code_newcode;          /* holds the text to write to the IP */
  90
  91 static unsigned nmi_wait_count;
  92 static atomic_t nmi_update_count = ATOMIC_INIT(0);
  93
  94 int ftrace_arch_read_dyn_info(char *buf, int size)
  95 {
  96         int r;
  97
  98         r = snprintf(buf, size, "%u %u",
  99                      nmi_wait_count,
 100                      atomic_read(&nmi_update_count));
 101         return r;
 102 }
 103
 104 static void ftrace_mod_code(void)
 105 {
 106         /*
 107          * Yes, more than one CPU process can be writing to mod_code_status.
 108          *    (and the code itself)
 109          * But if one were to fail, then they all should, and if one were
 110          * to succeed, then they all should.
 111          */
 112         mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode,
 113                                              MCOUNT_INSN_SIZE);
 114 }
 115
 116 void ftrace_nmi_enter(void)
 117 {
 118         atomic_inc(&in_nmi);
 119         /* Must have in_nmi seen before reading write flag */
 120         smp_mb();
 121         if (mod_code_write) {
 122                 ftrace_mod_code();
 123                 atomic_inc(&nmi_update_count);
 124         }
 125 }
 126
 127 void ftrace_nmi_exit(void)
 128 {
 129         /* Finish all executions before clearing in_nmi */
 130         smp_wmb();
 131         atomic_dec(&in_nmi);
 132 }
 133
 134 static void wait_for_nmi(void)
 135 {
 136         int waited = 0;
 137
 138         while (atomic_read(&in_nmi)) {
 139                 waited = 1;
 140                 cpu_relax();
 141         }
 142
 143         if (waited)
 144                 nmi_wait_count++;
 145 }
 146
 147 static int
 148 do_ftrace_mod_code(unsigned long ip, void *new_code)
 149 {
 150         mod_code_ip = (void *)ip;
 151         mod_code_newcode = new_code;
 152
 153         /* The buffers need to be visible before we let NMIs write them */
 154         smp_wmb();
 155
 156         mod_code_write = 1;
 157
 158         /* Make sure write bit is visible before we wait on NMIs */
 159         smp_mb();
 160
 161         wait_for_nmi();
 162
 163         /* Make sure all running NMIs have finished before we write the code */
 164         smp_mb();
 165
 166         ftrace_mod_code();
 167
 168         /* Make sure the write happens before clearing the bit */
 169         smp_wmb();
 170
 171         mod_code_write = 0;
 172
 173         /* make sure NMIs see the cleared bit */
 174         smp_mb();
 175
 176         wait_for_nmi();
 177
 178         return mod_code_status;
 179 }
 180
 181
 182
 183
 184 static unsigned char ftrace_nop[MCOUNT_INSN_SIZE];
 185
 186 static unsigned char *ftrace_nop_replace(void)
 187 {
 188         return ftrace_nop;
 189 }
 190
 191 static int
 192 ftrace_modify_code(unsigned long ip, unsigned char *old_code,
 193                    unsigned char *new_code)
 194 {
 195         unsigned char replaced[MCOUNT_INSN_SIZE];
 196
 197         /*
 198          * Note: Due to modules and __init, code can
 199          *  disappear and change, we need to protect against faulting
 200          *  as well as code changing. We do this by using the
 201          *  probe_kernel_* functions.
 202          *
 203          * No real locking needed, this code is run through
 204          * kstop_machine, or before SMP starts.
 205          */
 206
 207         /* read the text we want to modify */
 208         if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE))
 209                 return -EFAULT;
 210
 211         /* Make sure it is what we expect it to be */
 212         if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0)
 213                 return -EINVAL;
 214
 215         /* replace the text with the new text */
 216         if (do_ftrace_mod_code(ip, new_code))
 217                 return -EPERM;
 218
 219         sync_core();
 220
 221         return 0;
 222 }
 223
 224 int ftrace_make_nop(struct module *mod,
 225                     struct dyn_ftrace *rec, unsigned long addr)
 226 {
 227         unsigned char *new, *old;
 228         unsigned long ip = rec->ip;
 229
 230         old = ftrace_call_replace(ip, addr);
 231         new = ftrace_nop_replace();
 232
 233         return ftrace_modify_code(rec->ip, old, new);
 234 }
 235
 236 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 237 {
 238         unsigned char *new, *old;
 239         unsigned long ip = rec->ip;
 240
 241         old = ftrace_nop_replace();
 242         new = ftrace_call_replace(ip, addr);
 243
 244         return ftrace_modify_code(rec->ip, old, new);
 245 }
 246
 247 int ftrace_update_ftrace_func(ftrace_func_t func)
 248 {
 249         unsigned long ip = (unsigned long)(&ftrace_call);
 250         unsigned char old[MCOUNT_INSN_SIZE], *new;
 251         int ret;
 252
 253         memcpy(old, &ftrace_call, MCOUNT_INSN_SIZE);
 254         new = ftrace_call_replace(ip, (unsigned long)func);
 255         ret = ftrace_modify_code(ip, old, new);
 256
 257         return ret;
 258 }
 259
 260 int __init ftrace_dyn_arch_init(void *data)
 261 {
 262         extern const unsigned char ftrace_test_p6nop[];
 263         extern const unsigned char ftrace_test_nop5[];
 264         extern const unsigned char ftrace_test_jmp[];
 265         int faulted = 0;
 266
 267         /*
 268          * There is no good nop for all x86 archs.
 269          * We will default to using the P6_NOP5, but first we
 270          * will test to make sure that the nop will actually
 271          * work on this CPU. If it faults, we will then
 272          * go to a lesser efficient 5 byte nop. If that fails
 273          * we then just use a jmp as our nop. This isn't the most
 274          * efficient nop, but we can not use a multi part nop
 275          * since we would then risk being preempted in the middle
 276          * of that nop, and if we enabled tracing then, it might
 277          * cause a system crash.
 278          *
 279          * TODO: check the cpuid to determine the best nop.
 280          */
 281         asm volatile (
 282                 "ftrace_test_jmp:"
 283                 "jmp ftrace_test_p6nop\n"
 284                 "nop\n"
 285                 "nop\n"
 286                 "nop\n"  /* 2 byte jmp + 3 bytes */
 287                 "ftrace_test_p6nop:"
 288                 P6_NOP5
 289                 "jmp 1f\n"
 290                 "ftrace_test_nop5:"
 291                 ".byte 0x66,0x66,0x66,0x66,0x90\n"
 292                 "1:"
 293                 ".section .fixup, \"ax\"\n"
 294                 "2:     movl $1, %0\n"
 295                 "       jmp ftrace_test_nop5\n"
 296                 "3:     movl $2, %0\n"
 297                 "       jmp 1b\n"
 298                 ".previous\n"
 299                 _ASM_EXTABLE(ftrace_test_p6nop, 2b)
 300                 _ASM_EXTABLE(ftrace_test_nop5, 3b)
 301                 : "=r"(faulted) : "0" (faulted));
 302
 303         switch (faulted) {
 304         case 0:
 305                 pr_info("ftrace: converting mcount calls to 0f 1f 44 00 00\n");
 306                 memcpy(ftrace_nop, ftrace_test_p6nop, MCOUNT_INSN_SIZE);
 307                 break;
 308         case 1:
 309                 pr_info("ftrace: converting mcount calls to 66 66 66 66 90\n");
 310                 memcpy(ftrace_nop, ftrace_test_nop5, MCOUNT_INSN_SIZE);
 311                 break;
 312         case 2:
 313                 pr_info("ftrace: converting mcount calls to jmp . + 5\n");
 314                 memcpy(ftrace_nop, ftrace_test_jmp, MCOUNT_INSN_SIZE);
 315                 break;
 316         }
 317
 318         /* The return code is retured via data */
 319         *(unsigned long *)data = 0;
 320
 321         return 0;
 322 }
 323 #endif
 324
 325 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 326
 327 #ifdef CONFIG_DYNAMIC_FTRACE
 328 extern void ftrace_graph_call(void);
 329
 330 static int ftrace_mod_jmp(unsigned long ip,
 331                           int old_offset, int new_offset)
 332 {
 333         unsigned char code[MCOUNT_INSN_SIZE];
 334
 335         if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE))
 336                 return -EFAULT;
 337
 338         if (code[0] != 0xe9 || old_offset != *(int *)(&code[1]))
 339                 return -EINVAL;
 340
 341         *(int *)(&code[1]) = new_offset;
 342
 343         if (do_ftrace_mod_code(ip, &code))
 344                 return -EPERM;
 345
 346         return 0;
 347 }
 348
 349 int ftrace_enable_ftrace_graph_caller(void)
 350 {
 351         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 352         int old_offset, new_offset;
 353
 354         old_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
 355         new_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
 356
 357         return ftrace_mod_jmp(ip, old_offset, new_offset);
 358 }
 359
 360 int ftrace_disable_ftrace_graph_caller(void)
 361 {
 362         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 363         int old_offset, new_offset;
 364
 365         old_offset = (unsigned long)(&ftrace_graph_caller) - (ip + MCOUNT_INSN_SIZE);
 366         new_offset = (unsigned long)(&ftrace_stub) - (ip + MCOUNT_INSN_SIZE);
 367
 368         return ftrace_mod_jmp(ip, old_offset, new_offset);
 369 }
 370
 371 #else /* CONFIG_DYNAMIC_FTRACE */
 372
 373 /*
 374  * These functions are picked from those used on
 375  * this page for dynamic ftrace. They have been
 376  * simplified to ignore all traces in NMI context.
 377  */
 378 static atomic_t in_nmi;
 379
 380 void ftrace_nmi_enter(void)
 381 {
 382         atomic_inc(&in_nmi);
 383 }
 384
 385 void ftrace_nmi_exit(void)
 386 {
 387         atomic_dec(&in_nmi);
 388 }
 389
 390 #endif /* !CONFIG_DYNAMIC_FTRACE */
 391
 392 /*
 393  * Hook the return address and push it in the stack of return addrs
 394  * in current thread info.
 395  */
 396 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
 397 {
 398         unsigned long old;
 399         unsigned long long calltime;
 400         int faulted;
 401         struct ftrace_graph_ent trace;
 402         unsigned long return_hooker = (unsigned long)
 403                                 &return_to_handler;
 404
 405         /* Nmi's are currently unsupported */
 406         if (unlikely(atomic_read(&in_nmi)))
 407                 return;
 408
 409         if (unlikely(atomic_read(&current->tracing_graph_pause)))
 410                 return;
 411
 412         /*
 413          * Protect against fault, even if it shouldn't
 414          * happen. This tool is too much intrusive to
 415          * ignore such a protection.
 416          */
 417         asm volatile(
 418                 "1: " _ASM_MOV " (%[parent]), %[old]\n"
 419                 "2: " _ASM_MOV " %[return_hooker], (%[parent])\n"
 420                 "   movl $0, %[faulted]\n"
 421                 "3:\n"
 422
 423                 ".section .fixup, \"ax\"\n"
 424                 "4: movl $1, %[faulted]\n"
 425                 "   jmp 3b\n"
 426                 ".previous\n"
 427
 428                 _ASM_EXTABLE(1b, 4b)
 429                 _ASM_EXTABLE(2b, 4b)
 430
 431                 : [old] "=r" (old), [faulted] "=r" (faulted)
 432                 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
 433                 : "memory"
 434         );
 435
 436         if (unlikely(faulted)) {
 437                 ftrace_graph_stop();
 438                 WARN_ON(1);
 439                 return;
 440         }
 441
 442         if (unlikely(!__kernel_text_address(old))) {
 443                 ftrace_graph_stop();
 444                 *parent = old;
 445                 WARN_ON(1);
 446                 return;
 447         }
 448
 449         calltime = cpu_clock(raw_smp_processor_id());
 450
 451         if (ftrace_push_return_trace(old, calltime,
 452                                 self_addr, &trace.depth) == -EBUSY) {
 453                 *parent = old;
 454                 return;
 455         }
 456
 457         trace.func = self_addr;
 458
 459         /* Only trace if the calling function expects to */
 460         if (!ftrace_graph_entry(&trace)) {
 461                 current->curr_ret_stack--;
 462                 *parent = old;
 463         }
 464 }
 465 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */