1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2008 Matt Fleming <matt@console-pimps.org> 4 * Copyright (C) 2008 Paul Mundt <lethal@linux-sh.org> 5 * 6 * Code for replacing ftrace calls with jumps. 7 * 8 * Copyright (C) 2007-2008 Steven Rostedt <srostedt@redhat.com> 9 * 10 * Thanks goes to Ingo Molnar, for suggesting the idea. 11 * Mathieu Desnoyers, for suggesting postponing the modifications. 12 * Arjan van de Ven, for keeping me straight, and explaining to me 13 * the dangers of modifying code on the run. 14 */ 15 #include <linux/uaccess.h> 16 #include <linux/ftrace.h> 17 #include <linux/string.h> 18 #include <linux/init.h> 19 #include <linux/io.h> 20 #include <linux/kernel.h> 21 #include <asm/ftrace.h> 22 #include <asm/cacheflush.h> 23 #include <asm/unistd.h> 24 #include <trace/syscall.h> 25 26 #ifdef CONFIG_DYNAMIC_FTRACE 27 static unsigned char ftrace_replaced_code[MCOUNT_INSN_SIZE]; 28 29 static unsigned char ftrace_nop[4]; 30 /* 31 * If we're trying to nop out a call to a function, we instead 32 * place a call to the address after the memory table. 33 * 34 * 8c011060 <a>: 35 * 8c011060: 02 d1 mov.l 8c01106c <a+0xc>,r1 36 * 8c011062: 22 4f sts.l pr,@-r15 37 * 8c011064: 02 c7 mova 8c011070 <a+0x10>,r0 38 * 8c011066: 2b 41 jmp @r1 39 * 8c011068: 2a 40 lds r0,pr 40 * 8c01106a: 09 00 nop 41 * 8c01106c: 68 24 .word 0x2468 <--- ip 42 * 8c01106e: 1d 8c .word 0x8c1d 43 * 8c011070: 26 4f lds.l @r15+,pr <--- ip + MCOUNT_INSN_SIZE 44 * 45 * We write 0x8c011070 to 0x8c01106c so that on entry to a() we branch 46 * past the _mcount call and continue executing code like normal. 47 */ 48 static unsigned char *ftrace_nop_replace(unsigned long ip) 49 { 50 __raw_writel(ip + MCOUNT_INSN_SIZE, ftrace_nop); 51 return ftrace_nop; 52 } 53 54 static unsigned char *ftrace_call_replace(unsigned long ip, unsigned long addr) 55 { 56 /* Place the address in the memory table. */ 57 __raw_writel(addr, ftrace_replaced_code); 58 59 /* 60 * No locking needed, this must be called via kstop_machine 61 * which in essence is like running on a uniprocessor machine. 62 */ 63 return ftrace_replaced_code; 64 } 65 66 /* 67 * Modifying code must take extra care. On an SMP machine, if 68 * the code being modified is also being executed on another CPU 69 * that CPU will have undefined results and possibly take a GPF. 70 * We use kstop_machine to stop other CPUS from exectuing code. 71 * But this does not stop NMIs from happening. We still need 72 * to protect against that. We separate out the modification of 73 * the code to take care of this. 74 * 75 * Two buffers are added: An IP buffer and a "code" buffer. 76 * 77 * 1) Put the instruction pointer into the IP buffer 78 * and the new code into the "code" buffer. 79 * 2) Wait for any running NMIs to finish and set a flag that says 80 * we are modifying code, it is done in an atomic operation. 81 * 3) Write the code 82 * 4) clear the flag. 83 * 5) Wait for any running NMIs to finish. 84 * 85 * If an NMI is executed, the first thing it does is to call 86 * "ftrace_nmi_enter". This will check if the flag is set to write 87 * and if it is, it will write what is in the IP and "code" buffers. 88 * 89 * The trick is, it does not matter if everyone is writing the same 90 * content to the code location. Also, if a CPU is executing code 91 * it is OK to write to that code location if the contents being written 92 * are the same as what exists. 93 */ 94 #define MOD_CODE_WRITE_FLAG (1 << 31) /* set when NMI should do the write */ 95 static atomic_t nmi_running = ATOMIC_INIT(0); 96 static int mod_code_status; /* holds return value of text write */ 97 static void *mod_code_ip; /* holds the IP to write to */ 98 static void *mod_code_newcode; /* holds the text to write to the IP */ 99 100 static void clear_mod_flag(void) 101 { 102 int old = atomic_read(&nmi_running); 103 104 for (;;) { 105 int new = old & ~MOD_CODE_WRITE_FLAG; 106 107 if (old == new) 108 break; 109 110 old = atomic_cmpxchg(&nmi_running, old, new); 111 } 112 } 113 114 static void ftrace_mod_code(void) 115 { 116 /* 117 * Yes, more than one CPU process can be writing to mod_code_status. 118 * (and the code itself) 119 * But if one were to fail, then they all should, and if one were 120 * to succeed, then they all should. 121 */ 122 mod_code_status = probe_kernel_write(mod_code_ip, mod_code_newcode, 123 MCOUNT_INSN_SIZE); 124 125 /* if we fail, then kill any new writers */ 126 if (mod_code_status) 127 clear_mod_flag(); 128 } 129 130 void arch_ftrace_nmi_enter(void) 131 { 132 if (atomic_inc_return(&nmi_running) & MOD_CODE_WRITE_FLAG) { 133 smp_rmb(); 134 ftrace_mod_code(); 135 } 136 /* Must have previous changes seen before executions */ 137 smp_mb(); 138 } 139 140 void arch_ftrace_nmi_exit(void) 141 { 142 /* Finish all executions before clearing nmi_running */ 143 smp_mb(); 144 atomic_dec(&nmi_running); 145 } 146 147 static void wait_for_nmi_and_set_mod_flag(void) 148 { 149 if (!atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)) 150 return; 151 152 do { 153 cpu_relax(); 154 } while (atomic_cmpxchg(&nmi_running, 0, MOD_CODE_WRITE_FLAG)); 155 } 156 157 static void wait_for_nmi(void) 158 { 159 if (!atomic_read(&nmi_running)) 160 return; 161 162 do { 163 cpu_relax(); 164 } while (atomic_read(&nmi_running)); 165 } 166 167 static int 168 do_ftrace_mod_code(unsigned long ip, void *new_code) 169 { 170 mod_code_ip = (void *)ip; 171 mod_code_newcode = new_code; 172 173 /* The buffers need to be visible before we let NMIs write them */ 174 smp_mb(); 175 176 wait_for_nmi_and_set_mod_flag(); 177 178 /* Make sure all running NMIs have finished before we write the code */ 179 smp_mb(); 180 181 ftrace_mod_code(); 182 183 /* Make sure the write happens before clearing the bit */ 184 smp_mb(); 185 186 clear_mod_flag(); 187 wait_for_nmi(); 188 189 return mod_code_status; 190 } 191 192 static int 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: 199 * We are paranoid about modifying text, as if a bug was to happen, it 200 * could cause us to read or write to someplace that could cause harm. 201 * Carefully read and modify the code with probe_kernel_*(), and make 202 * sure what we read is what we expected it to be before modifying it. 203 */ 204 205 /* read the text we want to modify */ 206 if (probe_kernel_read(replaced, (void *)ip, MCOUNT_INSN_SIZE)) 207 return -EFAULT; 208 209 /* Make sure it is what we expect it to be */ 210 if (memcmp(replaced, old_code, MCOUNT_INSN_SIZE) != 0) 211 return -EINVAL; 212 213 /* replace the text with the new text */ 214 if (do_ftrace_mod_code(ip, new_code)) 215 return -EPERM; 216 217 flush_icache_range(ip, ip + MCOUNT_INSN_SIZE); 218 219 return 0; 220 } 221 222 int ftrace_update_ftrace_func(ftrace_func_t func) 223 { 224 unsigned long ip = (unsigned long)(&ftrace_call) + MCOUNT_INSN_OFFSET; 225 unsigned char old[MCOUNT_INSN_SIZE], *new; 226 227 memcpy(old, (unsigned char *)ip, MCOUNT_INSN_SIZE); 228 new = ftrace_call_replace(ip, (unsigned long)func); 229 230 return ftrace_modify_code(ip, old, new); 231 } 232 233 int ftrace_make_nop(struct module *mod, 234 struct dyn_ftrace *rec, unsigned long addr) 235 { 236 unsigned char *new, *old; 237 unsigned long ip = rec->ip; 238 239 old = ftrace_call_replace(ip, addr); 240 new = ftrace_nop_replace(ip); 241 242 return ftrace_modify_code(rec->ip, old, new); 243 } 244 245 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr) 246 { 247 unsigned char *new, *old; 248 unsigned long ip = rec->ip; 249 250 old = ftrace_nop_replace(ip); 251 new = ftrace_call_replace(ip, addr); 252 253 return ftrace_modify_code(rec->ip, old, new); 254 } 255 256 int __init ftrace_dyn_arch_init(void) 257 { 258 return 0; 259 } 260 #endif /* CONFIG_DYNAMIC_FTRACE */ 261 262 #ifdef CONFIG_FUNCTION_GRAPH_TRACER 263 #ifdef CONFIG_DYNAMIC_FTRACE 264 extern void ftrace_graph_call(void); 265 266 static int ftrace_mod(unsigned long ip, unsigned long old_addr, 267 unsigned long new_addr) 268 { 269 unsigned char code[MCOUNT_INSN_SIZE]; 270 271 if (probe_kernel_read(code, (void *)ip, MCOUNT_INSN_SIZE)) 272 return -EFAULT; 273 274 if (old_addr != __raw_readl((unsigned long *)code)) 275 return -EINVAL; 276 277 __raw_writel(new_addr, ip); 278 return 0; 279 } 280 281 int ftrace_enable_ftrace_graph_caller(void) 282 { 283 unsigned long ip, old_addr, new_addr; 284 285 ip = (unsigned long)(&ftrace_graph_call) + GRAPH_INSN_OFFSET; 286 old_addr = (unsigned long)(&skip_trace); 287 new_addr = (unsigned long)(&ftrace_graph_caller); 288 289 return ftrace_mod(ip, old_addr, new_addr); 290 } 291 292 int ftrace_disable_ftrace_graph_caller(void) 293 { 294 unsigned long ip, old_addr, new_addr; 295 296 ip = (unsigned long)(&ftrace_graph_call) + GRAPH_INSN_OFFSET; 297 old_addr = (unsigned long)(&ftrace_graph_caller); 298 new_addr = (unsigned long)(&skip_trace); 299 300 return ftrace_mod(ip, old_addr, new_addr); 301 } 302 #endif /* CONFIG_DYNAMIC_FTRACE */ 303 304 /* 305 * Hook the return address and push it in the stack of return addrs 306 * in the current thread info. 307 * 308 * This is the main routine for the function graph tracer. The function 309 * graph tracer essentially works like this: 310 * 311 * parent is the stack address containing self_addr's return address. 312 * We pull the real return address out of parent and store it in 313 * current's ret_stack. Then, we replace the return address on the stack 314 * with the address of return_to_handler. self_addr is the function that 315 * called mcount. 316 * 317 * When self_addr returns, it will jump to return_to_handler which calls 318 * ftrace_return_to_handler. ftrace_return_to_handler will pull the real 319 * return address off of current's ret_stack and jump to it. 320 */ 321 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr) 322 { 323 unsigned long old; 324 int faulted, err; 325 struct ftrace_graph_ent trace; 326 unsigned long return_hooker = (unsigned long)&return_to_handler; 327 328 if (unlikely(ftrace_graph_is_dead())) 329 return; 330 331 if (unlikely(atomic_read(¤t->tracing_graph_pause))) 332 return; 333 334 /* 335 * Protect against fault, even if it shouldn't 336 * happen. This tool is too much intrusive to 337 * ignore such a protection. 338 */ 339 __asm__ __volatile__( 340 "1: \n\t" 341 "mov.l @%2, %0 \n\t" 342 "2: \n\t" 343 "mov.l %3, @%2 \n\t" 344 "mov #0, %1 \n\t" 345 "3: \n\t" 346 ".section .fixup, \"ax\" \n\t" 347 "4: \n\t" 348 "mov.l 5f, %0 \n\t" 349 "jmp @%0 \n\t" 350 " mov #1, %1 \n\t" 351 ".balign 4 \n\t" 352 "5: .long 3b \n\t" 353 ".previous \n\t" 354 ".section __ex_table,\"a\" \n\t" 355 ".long 1b, 4b \n\t" 356 ".long 2b, 4b \n\t" 357 ".previous \n\t" 358 : "=&r" (old), "=r" (faulted) 359 : "r" (parent), "r" (return_hooker) 360 ); 361 362 if (unlikely(faulted)) { 363 ftrace_graph_stop(); 364 WARN_ON(1); 365 return; 366 } 367 368 err = ftrace_push_return_trace(old, self_addr, &trace.depth, 0, NULL); 369 if (err == -EBUSY) { 370 __raw_writel(old, parent); 371 return; 372 } 373 374 trace.func = self_addr; 375 376 /* Only trace if the calling function expects to */ 377 if (!ftrace_graph_entry(&trace)) { 378 current->curr_ret_stack--; 379 __raw_writel(old, parent); 380 } 381 } 382 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */ 383