arch/powerpc/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 out to P.A. Semi, Inc for supplying me with a PPC64 box.
   7  *
   8  * Added function graph tracer code, taken from x86 that was written
   9  * by Frederic Weisbecker, and ported to PPC by Steven Rostedt.
  10  *
  11  */
  12
  13 #include <linux/spinlock.h>
  14 #include <linux/hardirq.h>
  15 #include <linux/uaccess.h>
  16 #include <linux/module.h>
  17 #include <linux/ftrace.h>
  18 #include <linux/percpu.h>
  19 #include <linux/init.h>
  20 #include <linux/list.h>
  21
  22 #include <asm/cacheflush.h>
  23 #include <asm/code-patching.h>
  24 #include <asm/ftrace.h>
  25 #include <asm/syscall.h>
  26
  27
  28 #ifdef CONFIG_DYNAMIC_FTRACE
  29 static unsigned int
  30 ftrace_call_replace(unsigned long ip, unsigned long addr, int link)
  31 {
  32         unsigned int op;
  33
  34         addr = ppc_function_entry((void *)addr);
  35
  36         /* if (link) set op to 'bl' else 'b' */
  37         op = create_branch((unsigned int *)ip, addr, link ? 1 : 0);
  38
  39         return op;
  40 }
  41
  42 static int
  43 ftrace_modify_code(unsigned long ip, unsigned int old, unsigned int new)
  44 {
  45         unsigned int replaced;
  46
  47         /*
  48          * Note: Due to modules and __init, code can
  49          *  disappear and change, we need to protect against faulting
  50          *  as well as code changing. We do this by using the
  51          *  probe_kernel_* functions.
  52          *
  53          * No real locking needed, this code is run through
  54          * kstop_machine, or before SMP starts.
  55          */
  56
  57         /* read the text we want to modify */
  58         if (probe_kernel_read(&replaced, (void *)ip, MCOUNT_INSN_SIZE))
  59                 return -EFAULT;
  60
  61         /* Make sure it is what we expect it to be */
  62         if (replaced != old)
  63                 return -EINVAL;
  64
  65         /* replace the text with the new text */
  66         if (patch_instruction((unsigned int *)ip, new))
  67                 return -EPERM;
  68
  69         return 0;
  70 }
  71
  72 /*
  73  * Helper functions that are the same for both PPC64 and PPC32.
  74  */
  75 static int test_24bit_addr(unsigned long ip, unsigned long addr)
  76 {
  77         addr = ppc_function_entry((void *)addr);
  78
  79         /* use the create_branch to verify that this offset can be branched */
  80         return create_branch((unsigned int *)ip, addr, 0);
  81 }
  82
  83 #ifdef CONFIG_MODULES
  84
  85 static int is_bl_op(unsigned int op)
  86 {
  87         return (op & 0xfc000003) == 0x48000001;
  88 }
  89
  90 static unsigned long find_bl_target(unsigned long ip, unsigned int op)
  91 {
  92         static int offset;
  93
  94         offset = (op & 0x03fffffc);
  95         /* make it signed */
  96         if (offset & 0x02000000)
  97                 offset |= 0xfe000000;
  98
  99         return ip + (long)offset;
 100 }
 101
 102 #ifdef CONFIG_PPC64
 103 static int
 104 __ftrace_make_nop(struct module *mod,
 105                   struct dyn_ftrace *rec, unsigned long addr)
 106 {
 107         unsigned int op;
 108         unsigned long ptr;
 109         unsigned long ip = rec->ip;
 110         void *tramp;
 111
 112         /* read where this goes */
 113         if (probe_kernel_read(&op, (void *)ip, sizeof(int)))
 114                 return -EFAULT;
 115
 116         /* Make sure that that this is still a 24bit jump */
 117         if (!is_bl_op(op)) {
 118                 printk(KERN_ERR "Not expected bl: opcode is %x\n", op);
 119                 return -EINVAL;
 120         }
 121
 122         /* lets find where the pointer goes */
 123         tramp = (void *)find_bl_target(ip, op);
 124
 125         pr_devel("ip:%lx jumps to %p", ip, tramp);
 126
 127         if (!is_module_trampoline(tramp)) {
 128                 printk(KERN_ERR "Not a trampoline\n");
 129                 return -EINVAL;
 130         }
 131
 132         if (module_trampoline_target(mod, tramp, &ptr)) {
 133                 printk(KERN_ERR "Failed to get trampoline target\n");
 134                 return -EFAULT;
 135         }
 136
 137         pr_devel("trampoline target %lx", ptr);
 138
 139         /* This should match what was called */
 140         if (ptr != ppc_function_entry((void *)addr)) {
 141                 printk(KERN_ERR "addr %lx does not match expected %lx\n",
 142                         ptr, ppc_function_entry((void *)addr));
 143                 return -EINVAL;
 144         }
 145
 146         /*
 147          * Our original call site looks like:
 148          *
 149          * bl <tramp>
 150          * ld r2,XX(r1)
 151          *
 152          * Milton Miller pointed out that we can not simply nop the branch.
 153          * If a task was preempted when calling a trace function, the nops
 154          * will remove the way to restore the TOC in r2 and the r2 TOC will
 155          * get corrupted.
 156          *
 157          * Use a b +8 to jump over the load.
 158          */
 159         op = 0x48000008;        /* b +8 */
 160
 161         if (patch_instruction((unsigned int *)ip, op))
 162                 return -EPERM;
 163
 164         return 0;
 165 }
 166
 167 #else /* !PPC64 */
 168 static int
 169 __ftrace_make_nop(struct module *mod,
 170                   struct dyn_ftrace *rec, unsigned long addr)
 171 {
 172         unsigned int op;
 173         unsigned int jmp[4];
 174         unsigned long ip = rec->ip;
 175         unsigned long tramp;
 176
 177         if (probe_kernel_read(&op, (void *)ip, MCOUNT_INSN_SIZE))
 178                 return -EFAULT;
 179
 180         /* Make sure that that this is still a 24bit jump */
 181         if (!is_bl_op(op)) {
 182                 printk(KERN_ERR "Not expected bl: opcode is %x\n", op);
 183                 return -EINVAL;
 184         }
 185
 186         /* lets find where the pointer goes */
 187         tramp = find_bl_target(ip, op);
 188
 189         /*
 190          * On PPC32 the trampoline looks like:
 191          *  0x3d, 0x80, 0x00, 0x00  lis r12,sym@ha
 192          *  0x39, 0x8c, 0x00, 0x00  addi r12,r12,sym@l
 193          *  0x7d, 0x89, 0x03, 0xa6  mtctr r12
 194          *  0x4e, 0x80, 0x04, 0x20  bctr
 195          */
 196
 197         pr_devel("ip:%lx jumps to %lx", ip, tramp);
 198
 199         /* Find where the trampoline jumps to */
 200         if (probe_kernel_read(jmp, (void *)tramp, sizeof(jmp))) {
 201                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 202                 return -EFAULT;
 203         }
 204
 205         pr_devel(" %08x %08x ", jmp[0], jmp[1]);
 206
 207         /* verify that this is what we expect it to be */
 208         if (((jmp[0] & 0xffff0000) != 0x3d800000) ||
 209             ((jmp[1] & 0xffff0000) != 0x398c0000) ||
 210             (jmp[2] != 0x7d8903a6) ||
 211             (jmp[3] != 0x4e800420)) {
 212                 printk(KERN_ERR "Not a trampoline\n");
 213                 return -EINVAL;
 214         }
 215
 216         tramp = (jmp[1] & 0xffff) |
 217                 ((jmp[0] & 0xffff) << 16);
 218         if (tramp & 0x8000)
 219                 tramp -= 0x10000;
 220
 221         pr_devel(" %lx ", tramp);
 222
 223         if (tramp != addr) {
 224                 printk(KERN_ERR
 225                        "Trampoline location %08lx does not match addr\n",
 226                        tramp);
 227                 return -EINVAL;
 228         }
 229
 230         op = PPC_INST_NOP;
 231
 232         if (patch_instruction((unsigned int *)ip, op))
 233                 return -EPERM;
 234
 235         return 0;
 236 }
 237 #endif /* PPC64 */
 238 #endif /* CONFIG_MODULES */
 239
 240 int ftrace_make_nop(struct module *mod,
 241                     struct dyn_ftrace *rec, unsigned long addr)
 242 {
 243         unsigned long ip = rec->ip;
 244         unsigned int old, new;
 245
 246         /*
 247          * If the calling address is more that 24 bits away,
 248          * then we had to use a trampoline to make the call.
 249          * Otherwise just update the call site.
 250          */
 251         if (test_24bit_addr(ip, addr)) {
 252                 /* within range */
 253                 old = ftrace_call_replace(ip, addr, 1);
 254                 new = PPC_INST_NOP;
 255                 return ftrace_modify_code(ip, old, new);
 256         }
 257
 258 #ifdef CONFIG_MODULES
 259         /*
 260          * Out of range jumps are called from modules.
 261          * We should either already have a pointer to the module
 262          * or it has been passed in.
 263          */
 264         if (!rec->arch.mod) {
 265                 if (!mod) {
 266                         printk(KERN_ERR "No module loaded addr=%lx\n",
 267                                addr);
 268                         return -EFAULT;
 269                 }
 270                 rec->arch.mod = mod;
 271         } else if (mod) {
 272                 if (mod != rec->arch.mod) {
 273                         printk(KERN_ERR
 274                                "Record mod %p not equal to passed in mod %p\n",
 275                                rec->arch.mod, mod);
 276                         return -EINVAL;
 277                 }
 278                 /* nothing to do if mod == rec->arch.mod */
 279         } else
 280                 mod = rec->arch.mod;
 281
 282         return __ftrace_make_nop(mod, rec, addr);
 283 #else
 284         /* We should not get here without modules */
 285         return -EINVAL;
 286 #endif /* CONFIG_MODULES */
 287 }
 288
 289 #ifdef CONFIG_MODULES
 290 #ifdef CONFIG_PPC64
 291 static int
 292 __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 293 {
 294         unsigned int op[2];
 295         void *ip = (void *)rec->ip;
 296
 297         /* read where this goes */
 298         if (probe_kernel_read(op, ip, sizeof(op)))
 299                 return -EFAULT;
 300
 301         /*
 302          * We expect to see:
 303          *
 304          * b +8
 305          * ld r2,XX(r1)
 306          *
 307          * The load offset is different depending on the ABI. For simplicity
 308          * just mask it out when doing the compare.
 309          */
 310         if ((op[0] != 0x48000008) || ((op[1] & 0xffff00000) != 0xe8410000)) {
 311                 printk(KERN_ERR "Unexpected call sequence: %x %x\n",
 312                         op[0], op[1]);
 313                 return -EINVAL;
 314         }
 315
 316         /* If we never set up a trampoline to ftrace_caller, then bail */
 317         if (!rec->arch.mod->arch.tramp) {
 318                 printk(KERN_ERR "No ftrace trampoline\n");
 319                 return -EINVAL;
 320         }
 321
 322         /* Ensure branch is within 24 bits */
 323         if (create_branch(ip, rec->arch.mod->arch.tramp, BRANCH_SET_LINK)) {
 324                 printk(KERN_ERR "Branch out of range");
 325                 return -EINVAL;
 326         }
 327
 328         if (patch_branch(ip, rec->arch.mod->arch.tramp, BRANCH_SET_LINK)) {
 329                 printk(KERN_ERR "REL24 out of range!\n");
 330                 return -EINVAL;
 331         }
 332
 333         return 0;
 334 }
 335 #else
 336 static int
 337 __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 338 {
 339         unsigned int op;
 340         unsigned long ip = rec->ip;
 341
 342         /* read where this goes */
 343         if (probe_kernel_read(&op, (void *)ip, MCOUNT_INSN_SIZE))
 344                 return -EFAULT;
 345
 346         /* It should be pointing to a nop */
 347         if (op != PPC_INST_NOP) {
 348                 printk(KERN_ERR "Expected NOP but have %x\n", op);
 349                 return -EINVAL;
 350         }
 351
 352         /* If we never set up a trampoline to ftrace_caller, then bail */
 353         if (!rec->arch.mod->arch.tramp) {
 354                 printk(KERN_ERR "No ftrace trampoline\n");
 355                 return -EINVAL;
 356         }
 357
 358         /* create the branch to the trampoline */
 359         op = create_branch((unsigned int *)ip,
 360                            rec->arch.mod->arch.tramp, BRANCH_SET_LINK);
 361         if (!op) {
 362                 printk(KERN_ERR "REL24 out of range!\n");
 363                 return -EINVAL;
 364         }
 365
 366         pr_devel("write to %lx\n", rec->ip);
 367
 368         if (patch_instruction((unsigned int *)ip, op))
 369                 return -EPERM;
 370
 371         return 0;
 372 }
 373 #endif /* CONFIG_PPC64 */
 374 #endif /* CONFIG_MODULES */
 375
 376 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 377 {
 378         unsigned long ip = rec->ip;
 379         unsigned int old, new;
 380
 381         /*
 382          * If the calling address is more that 24 bits away,
 383          * then we had to use a trampoline to make the call.
 384          * Otherwise just update the call site.
 385          */
 386         if (test_24bit_addr(ip, addr)) {
 387                 /* within range */
 388                 old = PPC_INST_NOP;
 389                 new = ftrace_call_replace(ip, addr, 1);
 390                 return ftrace_modify_code(ip, old, new);
 391         }
 392
 393 #ifdef CONFIG_MODULES
 394         /*
 395          * Out of range jumps are called from modules.
 396          * Being that we are converting from nop, it had better
 397          * already have a module defined.
 398          */
 399         if (!rec->arch.mod) {
 400                 printk(KERN_ERR "No module loaded\n");
 401                 return -EINVAL;
 402         }
 403
 404         return __ftrace_make_call(rec, addr);
 405 #else
 406         /* We should not get here without modules */
 407         return -EINVAL;
 408 #endif /* CONFIG_MODULES */
 409 }
 410
 411 int ftrace_update_ftrace_func(ftrace_func_t func)
 412 {
 413         unsigned long ip = (unsigned long)(&ftrace_call);
 414         unsigned int old, new;
 415         int ret;
 416
 417         old = *(unsigned int *)&ftrace_call;
 418         new = ftrace_call_replace(ip, (unsigned long)func, 1);
 419         ret = ftrace_modify_code(ip, old, new);
 420
 421         return ret;
 422 }
 423
 424 static int __ftrace_replace_code(struct dyn_ftrace *rec, int enable)
 425 {
 426         unsigned long ftrace_addr = (unsigned long)FTRACE_ADDR;
 427         int ret;
 428
 429         ret = ftrace_update_record(rec, enable);
 430
 431         switch (ret) {
 432         case FTRACE_UPDATE_IGNORE:
 433                 return 0;
 434         case FTRACE_UPDATE_MAKE_CALL:
 435                 return ftrace_make_call(rec, ftrace_addr);
 436         case FTRACE_UPDATE_MAKE_NOP:
 437                 return ftrace_make_nop(NULL, rec, ftrace_addr);
 438         }
 439
 440         return 0;
 441 }
 442
 443 void ftrace_replace_code(int enable)
 444 {
 445         struct ftrace_rec_iter *iter;
 446         struct dyn_ftrace *rec;
 447         int ret;
 448
 449         for (iter = ftrace_rec_iter_start(); iter;
 450              iter = ftrace_rec_iter_next(iter)) {
 451                 rec = ftrace_rec_iter_record(iter);
 452                 ret = __ftrace_replace_code(rec, enable);
 453                 if (ret) {
 454                         ftrace_bug(ret, rec->ip);
 455                         return;
 456                 }
 457         }
 458 }
 459
 460 void arch_ftrace_update_code(int command)
 461 {
 462         if (command & FTRACE_UPDATE_CALLS)
 463                 ftrace_replace_code(1);
 464         else if (command & FTRACE_DISABLE_CALLS)
 465                 ftrace_replace_code(0);
 466
 467         if (command & FTRACE_UPDATE_TRACE_FUNC)
 468                 ftrace_update_ftrace_func(ftrace_trace_function);
 469
 470         if (command & FTRACE_START_FUNC_RET)
 471                 ftrace_enable_ftrace_graph_caller();
 472         else if (command & FTRACE_STOP_FUNC_RET)
 473                 ftrace_disable_ftrace_graph_caller();
 474 }
 475
 476 int __init ftrace_dyn_arch_init(void)
 477 {
 478         return 0;
 479 }
 480 #endif /* CONFIG_DYNAMIC_FTRACE */
 481
 482 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 483
 484 #ifdef CONFIG_DYNAMIC_FTRACE
 485 extern void ftrace_graph_call(void);
 486 extern void ftrace_graph_stub(void);
 487
 488 int ftrace_enable_ftrace_graph_caller(void)
 489 {
 490         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 491         unsigned long addr = (unsigned long)(&ftrace_graph_caller);
 492         unsigned long stub = (unsigned long)(&ftrace_graph_stub);
 493         unsigned int old, new;
 494
 495         old = ftrace_call_replace(ip, stub, 0);
 496         new = ftrace_call_replace(ip, addr, 0);
 497
 498         return ftrace_modify_code(ip, old, new);
 499 }
 500
 501 int ftrace_disable_ftrace_graph_caller(void)
 502 {
 503         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 504         unsigned long addr = (unsigned long)(&ftrace_graph_caller);
 505         unsigned long stub = (unsigned long)(&ftrace_graph_stub);
 506         unsigned int old, new;
 507
 508         old = ftrace_call_replace(ip, addr, 0);
 509         new = ftrace_call_replace(ip, stub, 0);
 510
 511         return ftrace_modify_code(ip, old, new);
 512 }
 513 #endif /* CONFIG_DYNAMIC_FTRACE */
 514
 515 #ifdef CONFIG_PPC64
 516 extern void mod_return_to_handler(void);
 517 #endif
 518
 519 /*
 520  * Hook the return address and push it in the stack of return addrs
 521  * in current thread info.
 522  */
 523 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
 524 {
 525         unsigned long old;
 526         int faulted;
 527         struct ftrace_graph_ent trace;
 528         unsigned long return_hooker = (unsigned long)&return_to_handler;
 529
 530         if (unlikely(atomic_read(&current->tracing_graph_pause)))
 531                 return;
 532
 533 #ifdef CONFIG_PPC64
 534         /* non core kernel code needs to save and restore the TOC */
 535         if (REGION_ID(self_addr) != KERNEL_REGION_ID)
 536                 return_hooker = (unsigned long)&mod_return_to_handler;
 537 #endif
 538
 539         return_hooker = ppc_function_entry((void *)return_hooker);
 540
 541         /*
 542          * Protect against fault, even if it shouldn't
 543          * happen. This tool is too much intrusive to
 544          * ignore such a protection.
 545          */
 546         asm volatile(
 547                 "1: " PPC_LL "%[old], 0(%[parent])\n"
 548                 "2: " PPC_STL "%[return_hooker], 0(%[parent])\n"
 549                 "   li %[faulted], 0\n"
 550                 "3:\n"
 551
 552                 ".section .fixup, \"ax\"\n"
 553                 "4: li %[faulted], 1\n"
 554                 "   b 3b\n"
 555                 ".previous\n"
 556
 557                 ".section __ex_table,\"a\"\n"
 558                         PPC_LONG_ALIGN "\n"
 559                         PPC_LONG "1b,4b\n"
 560                         PPC_LONG "2b,4b\n"
 561                 ".previous"
 562
 563                 : [old] "=&r" (old), [faulted] "=r" (faulted)
 564                 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
 565                 : "memory"
 566         );
 567
 568         if (unlikely(faulted)) {
 569                 ftrace_graph_stop();
 570                 WARN_ON(1);
 571                 return;
 572         }
 573
 574         trace.func = self_addr;
 575         trace.depth = current->curr_ret_stack + 1;
 576
 577         /* Only trace if the calling function expects to */
 578         if (!ftrace_graph_entry(&trace)) {
 579                 *parent = old;
 580                 return;
 581         }
 582
 583         if (ftrace_push_return_trace(old, self_addr, &trace.depth, 0) == -EBUSY)
 584                 *parent = old;
 585 }
 586 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
 587
 588 #if defined(CONFIG_FTRACE_SYSCALLS) && defined(CONFIG_PPC64)
 589 unsigned long __init arch_syscall_addr(int nr)
 590 {
 591         return sys_call_table[nr*2];
 592 }
 593 #endif /* CONFIG_FTRACE_SYSCALLS && CONFIG_PPC64 */