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
  78         /* use the create_branch to verify that this offset can be branched */
  79         return create_branch((unsigned int *)ip, addr, 0);
  80 }
  81
  82 #ifdef CONFIG_MODULES
  83
  84 static int is_bl_op(unsigned int op)
  85 {
  86         return (op & 0xfc000003) == 0x48000001;
  87 }
  88
  89 static unsigned long find_bl_target(unsigned long ip, unsigned int op)
  90 {
  91         static int offset;
  92
  93         offset = (op & 0x03fffffc);
  94         /* make it signed */
  95         if (offset & 0x02000000)
  96                 offset |= 0xfe000000;
  97
  98         return ip + (long)offset;
  99 }
 100
 101 #ifdef CONFIG_PPC64
 102 static int
 103 __ftrace_make_nop(struct module *mod,
 104                   struct dyn_ftrace *rec, unsigned long addr)
 105 {
 106         unsigned int op;
 107         unsigned int jmp[5];
 108         unsigned long ptr;
 109         unsigned long ip = rec->ip;
 110         unsigned long tramp;
 111         int offset;
 112
 113         /* read where this goes */
 114         if (probe_kernel_read(&op, (void *)ip, sizeof(int)))
 115                 return -EFAULT;
 116
 117         /* Make sure that that this is still a 24bit jump */
 118         if (!is_bl_op(op)) {
 119                 printk(KERN_ERR "Not expected bl: opcode is %x\n", op);
 120                 return -EINVAL;
 121         }
 122
 123         /* lets find where the pointer goes */
 124         tramp = find_bl_target(ip, op);
 125
 126         /*
 127          * On PPC64 the trampoline looks like:
 128          * 0x3d, 0x82, 0x00, 0x00,    addis   r12,r2, <high>
 129          * 0x39, 0x8c, 0x00, 0x00,    addi    r12,r12, <low>
 130          *   Where the bytes 2,3,6 and 7 make up the 32bit offset
 131          *   to the TOC that holds the pointer.
 132          *   to jump to.
 133          * 0xf8, 0x41, 0x00, 0x28,    std     r2,40(r1)
 134          * 0xe9, 0x6c, 0x00, 0x20,    ld      r11,32(r12)
 135          *   The actually address is 32 bytes from the offset
 136          *   into the TOC.
 137          * 0xe8, 0x4c, 0x00, 0x28,    ld      r2,40(r12)
 138          */
 139
 140         pr_devel("ip:%lx jumps to %lx r2: %lx", ip, tramp, mod->arch.toc);
 141
 142         /* Find where the trampoline jumps to */
 143         if (probe_kernel_read(jmp, (void *)tramp, sizeof(jmp))) {
 144                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 145                 return -EFAULT;
 146         }
 147
 148         pr_devel(" %08x %08x", jmp[0], jmp[1]);
 149
 150         /* verify that this is what we expect it to be */
 151         if (((jmp[0] & 0xffff0000) != 0x3d820000) ||
 152             ((jmp[1] & 0xffff0000) != 0x398c0000) ||
 153             (jmp[2] != 0xf8410028) ||
 154             (jmp[3] != 0xe96c0020) ||
 155             (jmp[4] != 0xe84c0028)) {
 156                 printk(KERN_ERR "Not a trampoline\n");
 157                 return -EINVAL;
 158         }
 159
 160         /* The bottom half is signed extended */
 161         offset = ((unsigned)((unsigned short)jmp[0]) << 16) +
 162                 (int)((short)jmp[1]);
 163
 164         pr_devel(" %x ", offset);
 165
 166         /* get the address this jumps too */
 167         tramp = mod->arch.toc + offset + 32;
 168         pr_devel("toc: %lx", tramp);
 169
 170         if (probe_kernel_read(jmp, (void *)tramp, 8)) {
 171                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 172                 return -EFAULT;
 173         }
 174
 175         pr_devel(" %08x %08x\n", jmp[0], jmp[1]);
 176
 177         ptr = ((unsigned long)jmp[0] << 32) + jmp[1];
 178
 179         /* This should match what was called */
 180         if (ptr != ppc_function_entry((void *)addr)) {
 181                 printk(KERN_ERR "addr does not match %lx\n", ptr);
 182                 return -EINVAL;
 183         }
 184
 185         /*
 186          * We want to nop the line, but the next line is
 187          *  0xe8, 0x41, 0x00, 0x28   ld r2,40(r1)
 188          * This needs to be turned to a nop too.
 189          */
 190         if (probe_kernel_read(&op, (void *)(ip+4), MCOUNT_INSN_SIZE))
 191                 return -EFAULT;
 192
 193         if (op != 0xe8410028) {
 194                 printk(KERN_ERR "Next line is not ld! (%08x)\n", op);
 195                 return -EINVAL;
 196         }
 197
 198         /*
 199          * Milton Miller pointed out that we can not blindly do nops.
 200          * If a task was preempted when calling a trace function,
 201          * the nops will remove the way to restore the TOC in r2
 202          * and the r2 TOC will get corrupted.
 203          */
 204
 205         /*
 206          * Replace:
 207          *   bl <tramp>  <==== will be replaced with "b 1f"
 208          *   ld r2,40(r1)
 209          *  1:
 210          */
 211         op = 0x48000008;        /* b +8 */
 212
 213         if (patch_instruction((unsigned int *)ip, op))
 214                 return -EPERM;
 215
 216         return 0;
 217 }
 218
 219 #else /* !PPC64 */
 220 static int
 221 __ftrace_make_nop(struct module *mod,
 222                   struct dyn_ftrace *rec, unsigned long addr)
 223 {
 224         unsigned int op;
 225         unsigned int jmp[4];
 226         unsigned long ip = rec->ip;
 227         unsigned long tramp;
 228
 229         if (probe_kernel_read(&op, (void *)ip, MCOUNT_INSN_SIZE))
 230                 return -EFAULT;
 231
 232         /* Make sure that that this is still a 24bit jump */
 233         if (!is_bl_op(op)) {
 234                 printk(KERN_ERR "Not expected bl: opcode is %x\n", op);
 235                 return -EINVAL;
 236         }
 237
 238         /* lets find where the pointer goes */
 239         tramp = find_bl_target(ip, op);
 240
 241         /*
 242          * On PPC32 the trampoline looks like:
 243          *  0x3d, 0x80, 0x00, 0x00  lis r12,sym@ha
 244          *  0x39, 0x8c, 0x00, 0x00  addi r12,r12,sym@l
 245          *  0x7d, 0x89, 0x03, 0xa6  mtctr r12
 246          *  0x4e, 0x80, 0x04, 0x20  bctr
 247          */
 248
 249         pr_devel("ip:%lx jumps to %lx", ip, tramp);
 250
 251         /* Find where the trampoline jumps to */
 252         if (probe_kernel_read(jmp, (void *)tramp, sizeof(jmp))) {
 253                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 254                 return -EFAULT;
 255         }
 256
 257         pr_devel(" %08x %08x ", jmp[0], jmp[1]);
 258
 259         /* verify that this is what we expect it to be */
 260         if (((jmp[0] & 0xffff0000) != 0x3d800000) ||
 261             ((jmp[1] & 0xffff0000) != 0x398c0000) ||
 262             (jmp[2] != 0x7d8903a6) ||
 263             (jmp[3] != 0x4e800420)) {
 264                 printk(KERN_ERR "Not a trampoline\n");
 265                 return -EINVAL;
 266         }
 267
 268         tramp = (jmp[1] & 0xffff) |
 269                 ((jmp[0] & 0xffff) << 16);
 270         if (tramp & 0x8000)
 271                 tramp -= 0x10000;
 272
 273         pr_devel(" %lx ", tramp);
 274
 275         if (tramp != addr) {
 276                 printk(KERN_ERR
 277                        "Trampoline location %08lx does not match addr\n",
 278                        tramp);
 279                 return -EINVAL;
 280         }
 281
 282         op = PPC_INST_NOP;
 283
 284         if (patch_instruction((unsigned int *)ip, op))
 285                 return -EPERM;
 286
 287         return 0;
 288 }
 289 #endif /* PPC64 */
 290 #endif /* CONFIG_MODULES */
 291
 292 int ftrace_make_nop(struct module *mod,
 293                     struct dyn_ftrace *rec, unsigned long addr)
 294 {
 295         unsigned long ip = rec->ip;
 296         unsigned int old, new;
 297
 298         /*
 299          * If the calling address is more that 24 bits away,
 300          * then we had to use a trampoline to make the call.
 301          * Otherwise just update the call site.
 302          */
 303         if (test_24bit_addr(ip, addr)) {
 304                 /* within range */
 305                 old = ftrace_call_replace(ip, addr, 1);
 306                 new = PPC_INST_NOP;
 307                 return ftrace_modify_code(ip, old, new);
 308         }
 309
 310 #ifdef CONFIG_MODULES
 311         /*
 312          * Out of range jumps are called from modules.
 313          * We should either already have a pointer to the module
 314          * or it has been passed in.
 315          */
 316         if (!rec->arch.mod) {
 317                 if (!mod) {
 318                         printk(KERN_ERR "No module loaded addr=%lx\n",
 319                                addr);
 320                         return -EFAULT;
 321                 }
 322                 rec->arch.mod = mod;
 323         } else if (mod) {
 324                 if (mod != rec->arch.mod) {
 325                         printk(KERN_ERR
 326                                "Record mod %p not equal to passed in mod %p\n",
 327                                rec->arch.mod, mod);
 328                         return -EINVAL;
 329                 }
 330                 /* nothing to do if mod == rec->arch.mod */
 331         } else
 332                 mod = rec->arch.mod;
 333
 334         return __ftrace_make_nop(mod, rec, addr);
 335 #else
 336         /* We should not get here without modules */
 337         return -EINVAL;
 338 #endif /* CONFIG_MODULES */
 339 }
 340
 341 #ifdef CONFIG_MODULES
 342 #ifdef CONFIG_PPC64
 343 static int
 344 __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 345 {
 346         unsigned int op[2];
 347         unsigned long ip = rec->ip;
 348
 349         /* read where this goes */
 350         if (probe_kernel_read(op, (void *)ip, MCOUNT_INSN_SIZE * 2))
 351                 return -EFAULT;
 352
 353         /*
 354          * It should be pointing to two nops or
 355          *  b +8; ld r2,40(r1)
 356          */
 357         if (((op[0] != 0x48000008) || (op[1] != 0xe8410028)) &&
 358             ((op[0] != PPC_INST_NOP) || (op[1] != PPC_INST_NOP))) {
 359                 printk(KERN_ERR "Expected NOPs but have %x %x\n", op[0], op[1]);
 360                 return -EINVAL;
 361         }
 362
 363         /* If we never set up a trampoline to ftrace_caller, then bail */
 364         if (!rec->arch.mod->arch.tramp) {
 365                 printk(KERN_ERR "No ftrace trampoline\n");
 366                 return -EINVAL;
 367         }
 368
 369         /* create the branch to the trampoline */
 370         op[0] = create_branch((unsigned int *)ip,
 371                               rec->arch.mod->arch.tramp, BRANCH_SET_LINK);
 372         if (!op[0]) {
 373                 printk(KERN_ERR "REL24 out of range!\n");
 374                 return -EINVAL;
 375         }
 376
 377         /* ld r2,40(r1) */
 378         op[1] = 0xe8410028;
 379
 380         pr_devel("write to %lx\n", rec->ip);
 381
 382         if (probe_kernel_write((void *)ip, op, MCOUNT_INSN_SIZE * 2))
 383                 return -EPERM;
 384
 385         flush_icache_range(ip, ip + 8);
 386
 387         return 0;
 388 }
 389 #else
 390 static int
 391 __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 392 {
 393         unsigned int op;
 394         unsigned long ip = rec->ip;
 395
 396         /* read where this goes */
 397         if (probe_kernel_read(&op, (void *)ip, MCOUNT_INSN_SIZE))
 398                 return -EFAULT;
 399
 400         /* It should be pointing to a nop */
 401         if (op != PPC_INST_NOP) {
 402                 printk(KERN_ERR "Expected NOP but have %x\n", op);
 403                 return -EINVAL;
 404         }
 405
 406         /* If we never set up a trampoline to ftrace_caller, then bail */
 407         if (!rec->arch.mod->arch.tramp) {
 408                 printk(KERN_ERR "No ftrace trampoline\n");
 409                 return -EINVAL;
 410         }
 411
 412         /* create the branch to the trampoline */
 413         op = create_branch((unsigned int *)ip,
 414                            rec->arch.mod->arch.tramp, BRANCH_SET_LINK);
 415         if (!op) {
 416                 printk(KERN_ERR "REL24 out of range!\n");
 417                 return -EINVAL;
 418         }
 419
 420         pr_devel("write to %lx\n", rec->ip);
 421
 422         if (patch_instruction((unsigned int *)ip, op))
 423                 return -EPERM;
 424
 425         return 0;
 426 }
 427 #endif /* CONFIG_PPC64 */
 428 #endif /* CONFIG_MODULES */
 429
 430 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 431 {
 432         unsigned long ip = rec->ip;
 433         unsigned int old, new;
 434
 435         /*
 436          * If the calling address is more that 24 bits away,
 437          * then we had to use a trampoline to make the call.
 438          * Otherwise just update the call site.
 439          */
 440         if (test_24bit_addr(ip, addr)) {
 441                 /* within range */
 442                 old = PPC_INST_NOP;
 443                 new = ftrace_call_replace(ip, addr, 1);
 444                 return ftrace_modify_code(ip, old, new);
 445         }
 446
 447 #ifdef CONFIG_MODULES
 448         /*
 449          * Out of range jumps are called from modules.
 450          * Being that we are converting from nop, it had better
 451          * already have a module defined.
 452          */
 453         if (!rec->arch.mod) {
 454                 printk(KERN_ERR "No module loaded\n");
 455                 return -EINVAL;
 456         }
 457
 458         return __ftrace_make_call(rec, addr);
 459 #else
 460         /* We should not get here without modules */
 461         return -EINVAL;
 462 #endif /* CONFIG_MODULES */
 463 }
 464
 465 int ftrace_update_ftrace_func(ftrace_func_t func)
 466 {
 467         unsigned long ip = (unsigned long)(&ftrace_call);
 468         unsigned int old, new;
 469         int ret;
 470
 471         old = *(unsigned int *)&ftrace_call;
 472         new = ftrace_call_replace(ip, (unsigned long)func, 1);
 473         ret = ftrace_modify_code(ip, old, new);
 474
 475         return ret;
 476 }
 477
 478 static int __ftrace_replace_code(struct dyn_ftrace *rec, int enable)
 479 {
 480         unsigned long ftrace_addr = (unsigned long)FTRACE_ADDR;
 481         int ret;
 482
 483         ret = ftrace_update_record(rec, enable);
 484
 485         switch (ret) {
 486         case FTRACE_UPDATE_IGNORE:
 487                 return 0;
 488         case FTRACE_UPDATE_MAKE_CALL:
 489                 return ftrace_make_call(rec, ftrace_addr);
 490         case FTRACE_UPDATE_MAKE_NOP:
 491                 return ftrace_make_nop(NULL, rec, ftrace_addr);
 492         }
 493
 494         return 0;
 495 }
 496
 497 void ftrace_replace_code(int enable)
 498 {
 499         struct ftrace_rec_iter *iter;
 500         struct dyn_ftrace *rec;
 501         int ret;
 502
 503         for (iter = ftrace_rec_iter_start(); iter;
 504              iter = ftrace_rec_iter_next(iter)) {
 505                 rec = ftrace_rec_iter_record(iter);
 506                 ret = __ftrace_replace_code(rec, enable);
 507                 if (ret) {
 508                         ftrace_bug(ret, rec->ip);
 509                         return;
 510                 }
 511         }
 512 }
 513
 514 void arch_ftrace_update_code(int command)
 515 {
 516         if (command & FTRACE_UPDATE_CALLS)
 517                 ftrace_replace_code(1);
 518         else if (command & FTRACE_DISABLE_CALLS)
 519                 ftrace_replace_code(0);
 520
 521         if (command & FTRACE_UPDATE_TRACE_FUNC)
 522                 ftrace_update_ftrace_func(ftrace_trace_function);
 523
 524         if (command & FTRACE_START_FUNC_RET)
 525                 ftrace_enable_ftrace_graph_caller();
 526         else if (command & FTRACE_STOP_FUNC_RET)
 527                 ftrace_disable_ftrace_graph_caller();
 528 }
 529
 530 int __init ftrace_dyn_arch_init(void *data)
 531 {
 532         /* caller expects data to be zero */
 533         unsigned long *p = data;
 534
 535         *p = 0;
 536
 537         return 0;
 538 }
 539 #endif /* CONFIG_DYNAMIC_FTRACE */
 540
 541 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 542
 543 #ifdef CONFIG_DYNAMIC_FTRACE
 544 extern void ftrace_graph_call(void);
 545 extern void ftrace_graph_stub(void);
 546
 547 int ftrace_enable_ftrace_graph_caller(void)
 548 {
 549         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 550         unsigned long addr = (unsigned long)(&ftrace_graph_caller);
 551         unsigned long stub = (unsigned long)(&ftrace_graph_stub);
 552         unsigned int old, new;
 553
 554         old = ftrace_call_replace(ip, stub, 0);
 555         new = ftrace_call_replace(ip, addr, 0);
 556
 557         return ftrace_modify_code(ip, old, new);
 558 }
 559
 560 int ftrace_disable_ftrace_graph_caller(void)
 561 {
 562         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 563         unsigned long addr = (unsigned long)(&ftrace_graph_caller);
 564         unsigned long stub = (unsigned long)(&ftrace_graph_stub);
 565         unsigned int old, new;
 566
 567         old = ftrace_call_replace(ip, addr, 0);
 568         new = ftrace_call_replace(ip, stub, 0);
 569
 570         return ftrace_modify_code(ip, old, new);
 571 }
 572 #endif /* CONFIG_DYNAMIC_FTRACE */
 573
 574 #ifdef CONFIG_PPC64
 575 extern void mod_return_to_handler(void);
 576 #endif
 577
 578 /*
 579  * Hook the return address and push it in the stack of return addrs
 580  * in current thread info.
 581  */
 582 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
 583 {
 584         unsigned long old;
 585         int faulted;
 586         struct ftrace_graph_ent trace;
 587         unsigned long return_hooker = (unsigned long)&return_to_handler;
 588
 589         if (unlikely(atomic_read(&current->tracing_graph_pause)))
 590                 return;
 591
 592 #ifdef CONFIG_PPC64
 593         /* non core kernel code needs to save and restore the TOC */
 594         if (REGION_ID(self_addr) != KERNEL_REGION_ID)
 595                 return_hooker = (unsigned long)&mod_return_to_handler;
 596 #endif
 597
 598         return_hooker = ppc_function_entry((void *)return_hooker);
 599
 600         /*
 601          * Protect against fault, even if it shouldn't
 602          * happen. This tool is too much intrusive to
 603          * ignore such a protection.
 604          */
 605         asm volatile(
 606                 "1: " PPC_LL "%[old], 0(%[parent])\n"
 607                 "2: " PPC_STL "%[return_hooker], 0(%[parent])\n"
 608                 "   li %[faulted], 0\n"
 609                 "3:\n"
 610
 611                 ".section .fixup, \"ax\"\n"
 612                 "4: li %[faulted], 1\n"
 613                 "   b 3b\n"
 614                 ".previous\n"
 615
 616                 ".section __ex_table,\"a\"\n"
 617                         PPC_LONG_ALIGN "\n"
 618                         PPC_LONG "1b,4b\n"
 619                         PPC_LONG "2b,4b\n"
 620                 ".previous"
 621
 622                 : [old] "=&r" (old), [faulted] "=r" (faulted)
 623                 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
 624                 : "memory"
 625         );
 626
 627         if (unlikely(faulted)) {
 628                 ftrace_graph_stop();
 629                 WARN_ON(1);
 630                 return;
 631         }
 632
 633         trace.func = self_addr;
 634         trace.depth = current->curr_ret_stack + 1;
 635
 636         /* Only trace if the calling function expects to */
 637         if (!ftrace_graph_entry(&trace)) {
 638                 *parent = old;
 639                 return;
 640         }
 641
 642         if (ftrace_push_return_trace(old, self_addr, &trace.depth, 0) == -EBUSY)
 643                 *parent = old;
 644 }
 645 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
 646
 647 #if defined(CONFIG_FTRACE_SYSCALLS) && defined(CONFIG_PPC64)
 648 unsigned long __init arch_syscall_addr(int nr)
 649 {
 650         return sys_call_table[nr*2];
 651 }
 652 #endif /* CONFIG_FTRACE_SYSCALLS && CONFIG_PPC64 */