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 (probe_kernel_write((void *)ip, &new, MCOUNT_INSN_SIZE))
  67                 return -EPERM;
  68
  69         flush_icache_range(ip, ip + 8);
  70
  71         return 0;
  72 }
  73
  74 /*
  75  * Helper functions that are the same for both PPC64 and PPC32.
  76  */
  77 static int test_24bit_addr(unsigned long ip, unsigned long addr)
  78 {
  79
  80         /* use the create_branch to verify that this offset can be branched */
  81         return create_branch((unsigned int *)ip, addr, 0);
  82 }
  83
  84 #ifdef CONFIG_MODULES
  85
  86 static int is_bl_op(unsigned int op)
  87 {
  88         return (op & 0xfc000003) == 0x48000001;
  89 }
  90
  91 static unsigned long find_bl_target(unsigned long ip, unsigned int op)
  92 {
  93         static int offset;
  94
  95         offset = (op & 0x03fffffc);
  96         /* make it signed */
  97         if (offset & 0x02000000)
  98                 offset |= 0xfe000000;
  99
 100         return ip + (long)offset;
 101 }
 102
 103 #ifdef CONFIG_PPC64
 104 static int
 105 __ftrace_make_nop(struct module *mod,
 106                   struct dyn_ftrace *rec, unsigned long addr)
 107 {
 108         unsigned int op;
 109         unsigned int jmp[5];
 110         unsigned long ptr;
 111         unsigned long ip = rec->ip;
 112         unsigned long tramp;
 113         int offset;
 114
 115         /* read where this goes */
 116         if (probe_kernel_read(&op, (void *)ip, sizeof(int)))
 117                 return -EFAULT;
 118
 119         /* Make sure that that this is still a 24bit jump */
 120         if (!is_bl_op(op)) {
 121                 printk(KERN_ERR "Not expected bl: opcode is %x\n", op);
 122                 return -EINVAL;
 123         }
 124
 125         /* lets find where the pointer goes */
 126         tramp = find_bl_target(ip, op);
 127
 128         /*
 129          * On PPC64 the trampoline looks like:
 130          * 0x3d, 0x82, 0x00, 0x00,    addis   r12,r2, <high>
 131          * 0x39, 0x8c, 0x00, 0x00,    addi    r12,r12, <low>
 132          *   Where the bytes 2,3,6 and 7 make up the 32bit offset
 133          *   to the TOC that holds the pointer.
 134          *   to jump to.
 135          * 0xf8, 0x41, 0x00, 0x28,    std     r2,40(r1)
 136          * 0xe9, 0x6c, 0x00, 0x20,    ld      r11,32(r12)
 137          *   The actually address is 32 bytes from the offset
 138          *   into the TOC.
 139          * 0xe8, 0x4c, 0x00, 0x28,    ld      r2,40(r12)
 140          */
 141
 142         pr_devel("ip:%lx jumps to %lx r2: %lx", ip, tramp, mod->arch.toc);
 143
 144         /* Find where the trampoline jumps to */
 145         if (probe_kernel_read(jmp, (void *)tramp, sizeof(jmp))) {
 146                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 147                 return -EFAULT;
 148         }
 149
 150         pr_devel(" %08x %08x", jmp[0], jmp[1]);
 151
 152         /* verify that this is what we expect it to be */
 153         if (((jmp[0] & 0xffff0000) != 0x3d820000) ||
 154             ((jmp[1] & 0xffff0000) != 0x398c0000) ||
 155             (jmp[2] != 0xf8410028) ||
 156             (jmp[3] != 0xe96c0020) ||
 157             (jmp[4] != 0xe84c0028)) {
 158                 printk(KERN_ERR "Not a trampoline\n");
 159                 return -EINVAL;
 160         }
 161
 162         /* The bottom half is signed extended */
 163         offset = ((unsigned)((unsigned short)jmp[0]) << 16) +
 164                 (int)((short)jmp[1]);
 165
 166         pr_devel(" %x ", offset);
 167
 168         /* get the address this jumps too */
 169         tramp = mod->arch.toc + offset + 32;
 170         pr_devel("toc: %lx", tramp);
 171
 172         if (probe_kernel_read(jmp, (void *)tramp, 8)) {
 173                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 174                 return -EFAULT;
 175         }
 176
 177         pr_devel(" %08x %08x\n", jmp[0], jmp[1]);
 178
 179         ptr = ((unsigned long)jmp[0] << 32) + jmp[1];
 180
 181         /* This should match what was called */
 182         if (ptr != ppc_function_entry((void *)addr)) {
 183                 printk(KERN_ERR "addr does not match %lx\n", ptr);
 184                 return -EINVAL;
 185         }
 186
 187         /*
 188          * We want to nop the line, but the next line is
 189          *  0xe8, 0x41, 0x00, 0x28   ld r2,40(r1)
 190          * This needs to be turned to a nop too.
 191          */
 192         if (probe_kernel_read(&op, (void *)(ip+4), MCOUNT_INSN_SIZE))
 193                 return -EFAULT;
 194
 195         if (op != 0xe8410028) {
 196                 printk(KERN_ERR "Next line is not ld! (%08x)\n", op);
 197                 return -EINVAL;
 198         }
 199
 200         /*
 201          * Milton Miller pointed out that we can not blindly do nops.
 202          * If a task was preempted when calling a trace function,
 203          * the nops will remove the way to restore the TOC in r2
 204          * and the r2 TOC will get corrupted.
 205          */
 206
 207         /*
 208          * Replace:
 209          *   bl <tramp>  <==== will be replaced with "b 1f"
 210          *   ld r2,40(r1)
 211          *  1:
 212          */
 213         op = 0x48000008;        /* b +8 */
 214
 215         if (probe_kernel_write((void *)ip, &op, MCOUNT_INSN_SIZE))
 216                 return -EPERM;
 217
 218
 219         flush_icache_range(ip, ip + 8);
 220
 221         return 0;
 222 }
 223
 224 #else /* !PPC64 */
 225 static int
 226 __ftrace_make_nop(struct module *mod,
 227                   struct dyn_ftrace *rec, unsigned long addr)
 228 {
 229         unsigned int op;
 230         unsigned int jmp[4];
 231         unsigned long ip = rec->ip;
 232         unsigned long tramp;
 233
 234         if (probe_kernel_read(&op, (void *)ip, MCOUNT_INSN_SIZE))
 235                 return -EFAULT;
 236
 237         /* Make sure that that this is still a 24bit jump */
 238         if (!is_bl_op(op)) {
 239                 printk(KERN_ERR "Not expected bl: opcode is %x\n", op);
 240                 return -EINVAL;
 241         }
 242
 243         /* lets find where the pointer goes */
 244         tramp = find_bl_target(ip, op);
 245
 246         /*
 247          * On PPC32 the trampoline looks like:
 248          *  0x3d, 0x60, 0x00, 0x00  lis r11,sym@ha
 249          *  0x39, 0x6b, 0x00, 0x00  addi r11,r11,sym@l
 250          *  0x7d, 0x69, 0x03, 0xa6  mtctr r11
 251          *  0x4e, 0x80, 0x04, 0x20  bctr
 252          */
 253
 254         pr_devel("ip:%lx jumps to %lx", ip, tramp);
 255
 256         /* Find where the trampoline jumps to */
 257         if (probe_kernel_read(jmp, (void *)tramp, sizeof(jmp))) {
 258                 printk(KERN_ERR "Failed to read %lx\n", tramp);
 259                 return -EFAULT;
 260         }
 261
 262         pr_devel(" %08x %08x ", jmp[0], jmp[1]);
 263
 264         /* verify that this is what we expect it to be */
 265         if (((jmp[0] & 0xffff0000) != 0x3d600000) ||
 266             ((jmp[1] & 0xffff0000) != 0x396b0000) ||
 267             (jmp[2] != 0x7d6903a6) ||
 268             (jmp[3] != 0x4e800420)) {
 269                 printk(KERN_ERR "Not a trampoline\n");
 270                 return -EINVAL;
 271         }
 272
 273         tramp = (jmp[1] & 0xffff) |
 274                 ((jmp[0] & 0xffff) << 16);
 275         if (tramp & 0x8000)
 276                 tramp -= 0x10000;
 277
 278         pr_devel(" %lx ", tramp);
 279
 280         if (tramp != addr) {
 281                 printk(KERN_ERR
 282                        "Trampoline location %08lx does not match addr\n",
 283                        tramp);
 284                 return -EINVAL;
 285         }
 286
 287         op = PPC_INST_NOP;
 288
 289         if (probe_kernel_write((void *)ip, &op, MCOUNT_INSN_SIZE))
 290                 return -EPERM;
 291
 292         flush_icache_range(ip, ip + 8);
 293
 294         return 0;
 295 }
 296 #endif /* PPC64 */
 297 #endif /* CONFIG_MODULES */
 298
 299 int ftrace_make_nop(struct module *mod,
 300                     struct dyn_ftrace *rec, unsigned long addr)
 301 {
 302         unsigned long ip = rec->ip;
 303         unsigned int old, new;
 304
 305         /*
 306          * If the calling address is more that 24 bits away,
 307          * then we had to use a trampoline to make the call.
 308          * Otherwise just update the call site.
 309          */
 310         if (test_24bit_addr(ip, addr)) {
 311                 /* within range */
 312                 old = ftrace_call_replace(ip, addr, 1);
 313                 new = PPC_INST_NOP;
 314                 return ftrace_modify_code(ip, old, new);
 315         }
 316
 317 #ifdef CONFIG_MODULES
 318         /*
 319          * Out of range jumps are called from modules.
 320          * We should either already have a pointer to the module
 321          * or it has been passed in.
 322          */
 323         if (!rec->arch.mod) {
 324                 if (!mod) {
 325                         printk(KERN_ERR "No module loaded addr=%lx\n",
 326                                addr);
 327                         return -EFAULT;
 328                 }
 329                 rec->arch.mod = mod;
 330         } else if (mod) {
 331                 if (mod != rec->arch.mod) {
 332                         printk(KERN_ERR
 333                                "Record mod %p not equal to passed in mod %p\n",
 334                                rec->arch.mod, mod);
 335                         return -EINVAL;
 336                 }
 337                 /* nothing to do if mod == rec->arch.mod */
 338         } else
 339                 mod = rec->arch.mod;
 340
 341         return __ftrace_make_nop(mod, rec, addr);
 342 #else
 343         /* We should not get here without modules */
 344         return -EINVAL;
 345 #endif /* CONFIG_MODULES */
 346 }
 347
 348 #ifdef CONFIG_MODULES
 349 #ifdef CONFIG_PPC64
 350 static int
 351 __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 352 {
 353         unsigned int op[2];
 354         unsigned long ip = rec->ip;
 355
 356         /* read where this goes */
 357         if (probe_kernel_read(op, (void *)ip, MCOUNT_INSN_SIZE * 2))
 358                 return -EFAULT;
 359
 360         /*
 361          * It should be pointing to two nops or
 362          *  b +8; ld r2,40(r1)
 363          */
 364         if (((op[0] != 0x48000008) || (op[1] != 0xe8410028)) &&
 365             ((op[0] != PPC_INST_NOP) || (op[1] != PPC_INST_NOP))) {
 366                 printk(KERN_ERR "Expected NOPs but have %x %x\n", op[0], op[1]);
 367                 return -EINVAL;
 368         }
 369
 370         /* If we never set up a trampoline to ftrace_caller, then bail */
 371         if (!rec->arch.mod->arch.tramp) {
 372                 printk(KERN_ERR "No ftrace trampoline\n");
 373                 return -EINVAL;
 374         }
 375
 376         /* create the branch to the trampoline */
 377         op[0] = create_branch((unsigned int *)ip,
 378                               rec->arch.mod->arch.tramp, BRANCH_SET_LINK);
 379         if (!op[0]) {
 380                 printk(KERN_ERR "REL24 out of range!\n");
 381                 return -EINVAL;
 382         }
 383
 384         /* ld r2,40(r1) */
 385         op[1] = 0xe8410028;
 386
 387         pr_devel("write to %lx\n", rec->ip);
 388
 389         if (probe_kernel_write((void *)ip, op, MCOUNT_INSN_SIZE * 2))
 390                 return -EPERM;
 391
 392         flush_icache_range(ip, ip + 8);
 393
 394         return 0;
 395 }
 396 #else
 397 static int
 398 __ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 399 {
 400         unsigned int op;
 401         unsigned long ip = rec->ip;
 402
 403         /* read where this goes */
 404         if (probe_kernel_read(&op, (void *)ip, MCOUNT_INSN_SIZE))
 405                 return -EFAULT;
 406
 407         /* It should be pointing to a nop */
 408         if (op != PPC_INST_NOP) {
 409                 printk(KERN_ERR "Expected NOP but have %x\n", op);
 410                 return -EINVAL;
 411         }
 412
 413         /* If we never set up a trampoline to ftrace_caller, then bail */
 414         if (!rec->arch.mod->arch.tramp) {
 415                 printk(KERN_ERR "No ftrace trampoline\n");
 416                 return -EINVAL;
 417         }
 418
 419         /* create the branch to the trampoline */
 420         op = create_branch((unsigned int *)ip,
 421                            rec->arch.mod->arch.tramp, BRANCH_SET_LINK);
 422         if (!op) {
 423                 printk(KERN_ERR "REL24 out of range!\n");
 424                 return -EINVAL;
 425         }
 426
 427         pr_devel("write to %lx\n", rec->ip);
 428
 429         if (probe_kernel_write((void *)ip, &op, MCOUNT_INSN_SIZE))
 430                 return -EPERM;
 431
 432         flush_icache_range(ip, ip + 8);
 433
 434         return 0;
 435 }
 436 #endif /* CONFIG_PPC64 */
 437 #endif /* CONFIG_MODULES */
 438
 439 int ftrace_make_call(struct dyn_ftrace *rec, unsigned long addr)
 440 {
 441         unsigned long ip = rec->ip;
 442         unsigned int old, new;
 443
 444         /*
 445          * If the calling address is more that 24 bits away,
 446          * then we had to use a trampoline to make the call.
 447          * Otherwise just update the call site.
 448          */
 449         if (test_24bit_addr(ip, addr)) {
 450                 /* within range */
 451                 old = PPC_INST_NOP;
 452                 new = ftrace_call_replace(ip, addr, 1);
 453                 return ftrace_modify_code(ip, old, new);
 454         }
 455
 456 #ifdef CONFIG_MODULES
 457         /*
 458          * Out of range jumps are called from modules.
 459          * Being that we are converting from nop, it had better
 460          * already have a module defined.
 461          */
 462         if (!rec->arch.mod) {
 463                 printk(KERN_ERR "No module loaded\n");
 464                 return -EINVAL;
 465         }
 466
 467         return __ftrace_make_call(rec, addr);
 468 #else
 469         /* We should not get here without modules */
 470         return -EINVAL;
 471 #endif /* CONFIG_MODULES */
 472 }
 473
 474 int ftrace_update_ftrace_func(ftrace_func_t func)
 475 {
 476         unsigned long ip = (unsigned long)(&ftrace_call);
 477         unsigned int old, new;
 478         int ret;
 479
 480         old = *(unsigned int *)&ftrace_call;
 481         new = ftrace_call_replace(ip, (unsigned long)func, 1);
 482         ret = ftrace_modify_code(ip, old, new);
 483
 484         return ret;
 485 }
 486
 487 int __init ftrace_dyn_arch_init(void *data)
 488 {
 489         /* caller expects data to be zero */
 490         unsigned long *p = data;
 491
 492         *p = 0;
 493
 494         return 0;
 495 }
 496 #endif /* CONFIG_DYNAMIC_FTRACE */
 497
 498 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
 499
 500 #ifdef CONFIG_DYNAMIC_FTRACE
 501 extern void ftrace_graph_call(void);
 502 extern void ftrace_graph_stub(void);
 503
 504 int ftrace_enable_ftrace_graph_caller(void)
 505 {
 506         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 507         unsigned long addr = (unsigned long)(&ftrace_graph_caller);
 508         unsigned long stub = (unsigned long)(&ftrace_graph_stub);
 509         unsigned int old, new;
 510
 511         old = ftrace_call_replace(ip, stub, 0);
 512         new = ftrace_call_replace(ip, addr, 0);
 513
 514         return ftrace_modify_code(ip, old, new);
 515 }
 516
 517 int ftrace_disable_ftrace_graph_caller(void)
 518 {
 519         unsigned long ip = (unsigned long)(&ftrace_graph_call);
 520         unsigned long addr = (unsigned long)(&ftrace_graph_caller);
 521         unsigned long stub = (unsigned long)(&ftrace_graph_stub);
 522         unsigned int old, new;
 523
 524         old = ftrace_call_replace(ip, addr, 0);
 525         new = ftrace_call_replace(ip, stub, 0);
 526
 527         return ftrace_modify_code(ip, old, new);
 528 }
 529 #endif /* CONFIG_DYNAMIC_FTRACE */
 530
 531 #ifdef CONFIG_PPC64
 532 extern void mod_return_to_handler(void);
 533 #endif
 534
 535 /*
 536  * Hook the return address and push it in the stack of return addrs
 537  * in current thread info.
 538  */
 539 void prepare_ftrace_return(unsigned long *parent, unsigned long self_addr)
 540 {
 541         unsigned long old;
 542         int faulted;
 543         struct ftrace_graph_ent trace;
 544         unsigned long return_hooker = (unsigned long)&return_to_handler;
 545
 546         if (unlikely(atomic_read(&current->tracing_graph_pause)))
 547                 return;
 548
 549 #ifdef CONFIG_PPC64
 550         /* non core kernel code needs to save and restore the TOC */
 551         if (REGION_ID(self_addr) != KERNEL_REGION_ID)
 552                 return_hooker = (unsigned long)&mod_return_to_handler;
 553 #endif
 554
 555         return_hooker = ppc_function_entry((void *)return_hooker);
 556
 557         /*
 558          * Protect against fault, even if it shouldn't
 559          * happen. This tool is too much intrusive to
 560          * ignore such a protection.
 561          */
 562         asm volatile(
 563                 "1: " PPC_LL "%[old], 0(%[parent])\n"
 564                 "2: " PPC_STL "%[return_hooker], 0(%[parent])\n"
 565                 "   li %[faulted], 0\n"
 566                 "3:\n"
 567
 568                 ".section .fixup, \"ax\"\n"
 569                 "4: li %[faulted], 1\n"
 570                 "   b 3b\n"
 571                 ".previous\n"
 572
 573                 ".section __ex_table,\"a\"\n"
 574                         PPC_LONG_ALIGN "\n"
 575                         PPC_LONG "1b,4b\n"
 576                         PPC_LONG "2b,4b\n"
 577                 ".previous"
 578
 579                 : [old] "=&r" (old), [faulted] "=r" (faulted)
 580                 : [parent] "r" (parent), [return_hooker] "r" (return_hooker)
 581                 : "memory"
 582         );
 583
 584         if (unlikely(faulted)) {
 585                 ftrace_graph_stop();
 586                 WARN_ON(1);
 587                 return;
 588         }
 589
 590         if (ftrace_push_return_trace(old, self_addr, &trace.depth, 0) == -EBUSY) {
 591                 *parent = old;
 592                 return;
 593         }
 594
 595         trace.func = self_addr;
 596
 597         /* Only trace if the calling function expects to */
 598         if (!ftrace_graph_entry(&trace)) {
 599                 current->curr_ret_stack--;
 600                 *parent = old;
 601         }
 602 }
 603 #endif /* CONFIG_FUNCTION_GRAPH_TRACER */
 604
 605 #if defined(CONFIG_FTRACE_SYSCALLS) && defined(CONFIG_PPC64)
 606 unsigned long __init arch_syscall_addr(int nr)
 607 {
 608         return sys_call_table[nr*2];
 609 }
 610 #endif /* CONFIG_FTRACE_SYSCALLS && CONFIG_PPC64 */