1/* 2 * Copyright (C) 2008 Matt Fleming <matt@console-pimps.org> 3 * Copyright (C) 2008 Paul Mundt <lethal@linux-sh.org> 4 * 5 * Code for replacing ftrace calls with jumps. 6 * 7 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com> 8 * 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. 13 */ 14#include <linux/uaccess.h> 15#include <linux/ftrace.h> 16#include <linux/string.h> 17#include <linux/init.h> 18#include <linux/io.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> 24 25#ifdef CONFIG_DYNAMIC_FTRACE 26static unsigned char ftrace_replaced_code[MCOUNT_INSN_SIZE]; 27 28static unsigned char ftrace_nop[4]; 29/* 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. 32 * 33 * 8c011060 <a>: 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 39 * 8c01106a: 09 00 nop 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 43 * 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. 46 */ 47static unsigned char *ftrace_nop_replace(unsigned long ip) 48{ 49 __raw_writel(ip + MCOUNT_INSN_SIZE, ftrace_nop); 50 return ftrace_nop; 51} 52 53static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) 54{ 55 /* Place the address in the memory table. */ 56 __raw_writel(addr, ftrace_replaced_code); 57 58 /* 59 * No locking needed, this must be called via kstop_machine 60 * which in essence is like running on a uniprocessor machine. 61 */ 62 return ftrace_replaced_code; 63} 64 65/* 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. 73 * 74 * Two buffers are added: An IP buffer and a "code" buffer. 75 * 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. 80 * 3) Write the code 81 * 4) clear the flag. 82 * 5) Wait for any running NMIs to finish. 83 * 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. 87 * 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. 92 */ 93#define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */ 94static atomic_t nmi_running = ATOMIC_INIT(0); 95static int mod_code_status; /* holds return value of text write */ 96static void *mod_code_ip; /* holds the IP to write to */ 97static void *mod_code_newcode; /* holds the text to write to the IP */ 98 99static unsigned nmi_wait_count; 100static atomic_t nmi_update_count = ATOMIC_INIT(0); 101 102int ftrace_arch_read_dyn_info(char *buf, int size) 103{ 104 int r; 105 106 r = snprintf(buf, size, "%u %u", 107 nmi_wait_count, 108 atomic_read(&nmi_update_count)); 109 return r; 110} 111 112static void clear_mod_flag(void) 113{ 114 int old = atomic_read(&nmi_running); 115 116 for (;;) { 117 int new = old & ~MOD_CODE_WRITE_FLAG; 118 119 if (old == new) 120 break; 121 122 old = atomic_cmpxchg(&nmi_running, old, new); 123 } 124} 125 126static void ftrace_mod_code(void) 127{ 128 /* 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. 133 */ 134 mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode, 135 MCOUNT_INSN_SIZE); 136 137 /* if we fail, then kill any new writers */ 138 if (mod_code_status) 139 clear_mod_flag(); 140} 141 142void ftrace_nmi_enter(void) 143{ 144 if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) { 145 smp_rmb(); 146 ftrace_mod_code(); 147 atomic_inc(&nmi_update_count); 148 } 149 /* Must have previous changes seen before executions */ 150 smp_mb(); 151} 152 153void ftrace_nmi_exit(void) 154{ 155 /* Finish all executions before clearing nmi_running */ 156 smp_mb(); 157 atomic_dec(&nmi_running); 158} 159 160static void wait_for_nmi_and_set_mod_flag(void) 161{ 162 if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)) 163 return; 164 165 do { 166 cpu_relax(); 167 } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)); 168 169 nmi_wait_count++; 170} 171 172static void wait_for_nmi(void) 173{ 174 if (!atomic_read(&nmi_running)) 175 return; 176 177 do { 178 cpu_relax(); 179 } while (atomic_read(&nmi_running)); 180 181 nmi_wait_count++; 182} 183 184static int 185do_ftrace_mod_code(unsigned long ip, void *new_code) 186{ 187 mod_code_ip = (void *)ip; 188 mod_code_newcode = new_code; 189 190 /* The buffers need to be visible before we let NMIs write them */ 191 smp_mb(); 192 193 wait_for_nmi_and_set_mod_flag(); 194 195 /* Make sure all running NMIs have finished before we write the code */ 196 smp_mb(); 197 198 ftrace_mod_code(); 199 200 /* Make sure the write happens before clearing the bit */ 201 smp_mb(); 202 203 clear_mod_flag(); 204 wait_for_nmi(); 205 206 return mod_code_status; 207} 208 209static int ftrace_modify_code(unsigned long ip, unsigned char *old_code, 210 unsigned char *new_code) 211{ 212 unsigned char replaced[MCOUNT_INSN_SIZE]; 213 214 /* 215 * Note: Due to modules and __init, code can 216 * disappear and change, we need to protect against faulting 217 * as well as code changing. We do this by using the 218 * probe_kernel_* functions. 219 * 220 * No real locking needed, this code is run through 221 * kstop_machine, or before SMP starts. 222 */ 223 224 /* read the text we want to modify */ 225 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) 226 return -EFAULT; 227 228 /* Make sure it is what we expect it to be */ 229 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) 230 return -EINVAL; 231 232 /* replace the text with the new text */ 233 if (do_ftrace_mod_code(ip, new_code)) 234 return -EPERM; 235 236 flush_icache_range(ip, ip + MCOUNT_INSN_SIZE); 237 238 return 0; 239} 240 241int ftrace_update_ftrace_func(ftrace_func_t func) 242{ 243 unsigned long ip = (unsigned long)(&ftrace_call) + MCOUNT_INSN_OFFSET; 244 unsigned char old[MCOUNT_INSN_SIZE], *new; 245 246 memcpy(old, (unsigned char *)ip, MCOUNT_INSN_SIZE); 247 new = ftrace_call_replace(ip, (unsigned long)func); 248 249 return ftrace_modify_code(ip, old, new); 250} 251 252int ftrace_make_nop(struct module *mod, 253 struct dyn_ftrace *rec, unsigned long addr) 254{ 255 unsigned char *new, *old; 256 unsigned long ip = rec->ip; 257 258 old = ftrace_call_replace(ip, addr); 259 new = ftrace_nop_replace(ip); 260 261 return ftrace_modify_code(rec->ip, old, new); 262} 263 264int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) 265{ 266 unsigned char *new, *old; 267 unsigned long ip = rec->ip; 268 269 old = ftrace_nop_replace(ip); 270 new = ftrace_call_replace(ip, addr); 271 272 return ftrace_modify_code(rec->ip, old, new); 273} 274 275int __init ftrace_dyn_arch_init(void) 276{ 277 return 0; 278} 279#endif /* CONFIG_DYNAMIC_FTRACE */ 280 281#ifdef CONFIG_FUNCTION_GRAPH_TRACER 282#ifdef CONFIG_DYNAMIC_FTRACE 283extern void ftrace_graph_call(void); 284 285static int ftrace_mod(unsigned long ip, unsigned long old_addr, 286 unsigned long new_addr) 287{ 288 unsigned char code[MCOUNT_INSN_SIZE]; 289 290 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE)) 291 return -EFAULT; 292 293 if (old_addr != __raw_readl((unsigned long *)code)) 294 return -EINVAL; 295 296 __raw_writel(new_addr, ip); 297 return 0; 298} 299 300int ftrace_enable_ftrace_graph_caller(void) 301{ 302 unsigned long ip, old_addr, new_addr; 303 304 ip = (unsigned long)(&ftrace_graph_call) + GRAPH_INSN_OFFSET; 305 old_addr = (unsigned long)(&skip_trace); 306 new_addr = (unsigned long)(&ftrace_graph_caller); 307 308 return ftrace_mod(ip, old_addr, new_addr); 309} 310 311int ftrace_disable_ftrace_graph_caller(void) 312{ 313 unsigned long ip, old_addr, new_addr; 314 315 ip = (unsigned long)(&ftrace_graph_call) + GRAPH_INSN_OFFSET; 316 old_addr = (unsigned long)(&ftrace_graph_caller); 317 new_addr = (unsigned long)(&skip_trace); 318 319 return ftrace_mod(ip, old_addr, new_addr); 320} 321#endif /* CONFIG_DYNAMIC_FTRACE */ 322 323/* 324 * Hook the return address and push it in the stack of return addrs 325 * in the current thread info. 326 * 327 * This is the main routine for the function graph tracer. The function 328 * graph tracer essentially works like this: 329 * 330 * parent is the stack address containing self_addr's return address. 331 * We pull the real return address out of parent and store it in 332 * current's ret_stack. Then, we replace the return address on the stack 333 * with the address of return_to_handler. self_addr is the function that 334 * called mcount. 335 * 336 * When self_addr returns, it will jump to return_to_handler which calls 337 * ftrace_return_to_handler. ftrace_return_to_handler will pull the real 338 * return address off of current's ret_stack and jump to it. 339 */ 340void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr) 341{ 342 unsigned long old; 343 int faulted, err; 344 struct ftrace_graph_ent trace; 345 unsigned long return_hooker = (unsigned long)&return_to_handler; 346 347 if (unlikely(ftrace_graph_is_dead())) 348 return; 349 350 if (unlikely(atomic_read(¤t->tracing_graph_pause))) 351 return; 352 353 /* 354 * Protect against fault, even if it shouldn't 355 * happen. This tool is too much intrusive to 356 * ignore such a protection. 357 */ 358 __asm__ __volatile__( 359 "1: \n\t" 360 "mov.l @%2, %0 \n\t" 361 "2: \n\t" 362 "mov.l %3, @%2 \n\t" 363 "mov #0, %1 \n\t" 364 "3: \n\t" 365 ".section .fixup, \"ax\" \n\t" 366 "4: \n\t" 367 "mov.l 5f, %0 \n\t" 368 "jmp @%0 \n\t" 369 " mov #1, %1 \n\t" 370 ".balign 4 \n\t" 371 "5: .long 3b \n\t" 372 ".previous \n\t" 373 ".section __ex_table,\"a\" \n\t" 374 ".long 1b, 4b \n\t" 375 ".long 2b, 4b \n\t" 376 ".previous \n\t" 377 : "=&r" (old), "=r" (faulted) 378 : "r" (parent), "r" (return_hooker) 379 ); 380 381 if (unlikely(faulted)) { 382 ftrace_graph_stop(); 383 WARN_ON(1); 384 return; 385 } 386 387 err = ftrace_push_return_trace(old, self_addr, &trace.depth, 0); 388 if (err == -EBUSY) { 389 __raw_writel(old, parent); 390 return; 391 } 392 393 trace.func = self_addr; 394 395 /* Only trace if the calling function expects to */ 396 if (!ftrace_graph_entry(&trace)) { 397 current->curr_ret_stack--; 398 __raw_writel(old, parent); 399 } 400} 401#endif /* CONFIG_FUNCTION_GRAPH_TRACER */ 402