SAFE public projects git trees. - safe/jmp/linux-2.6/blob - kernel/trace/trace_kprobe.c

   1 /*
   2  * kprobe based kernel tracer
   3  *
   4  * Created by Masami Hiramatsu <mhiramat@redhat.com>
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  18  */
  19
  20 #include <linux/module.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/kprobes.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/slab.h>
  25 #include <linux/smp.h>
  26 #include <linux/debugfs.h>
  27 #include <linux/types.h>
  28 #include <linux/string.h>
  29 #include <linux/ctype.h>
  30 #include <linux/ptrace.h>
  31
  32 #include "trace.h"
  33 #include "trace_output.h"
  34
  35 #define MAX_TRACE_ARGS 128
  36 #define MAX_ARGSTR_LEN 63
  37
  38 /* currently, trace_kprobe only supports X86. */
  39
  40 struct fetch_func {
  41         unsigned long (*func)(struct pt_regs *, void *);
  42         void *data;
  43 };
  44
  45 static __kprobes unsigned long call_fetch(struct fetch_func *f,
  46                                           struct pt_regs *regs)
  47 {
  48         return f->func(regs, f->data);
  49 }
  50
  51 /* fetch handlers */
  52 static __kprobes unsigned long fetch_register(struct pt_regs *regs,
  53                                               void *offset)
  54 {
  55         return regs_get_register(regs, (unsigned int)((unsigned long)offset));
  56 }
  57
  58 static __kprobes unsigned long fetch_stack(struct pt_regs *regs,
  59                                            void *num)
  60 {
  61         return regs_get_kernel_stack_nth(regs,
  62                                          (unsigned int)((unsigned long)num));
  63 }
  64
  65 static __kprobes unsigned long fetch_memory(struct pt_regs *regs, void *addr)
  66 {
  67         unsigned long retval;
  68
  69         if (probe_kernel_address(addr, retval))
  70                 return 0;
  71         return retval;
  72 }
  73
  74 static __kprobes unsigned long fetch_argument(struct pt_regs *regs, void *num)
  75 {
  76         return regs_get_argument_nth(regs, (unsigned int)((unsigned long)num));
  77 }
  78
  79 static __kprobes unsigned long fetch_retvalue(struct pt_regs *regs,
  80                                               void *dummy)
  81 {
  82         return regs_return_value(regs);
  83 }
  84
  85 static __kprobes unsigned long fetch_ip(struct pt_regs *regs, void *dummy)
  86 {
  87         return instruction_pointer(regs);
  88 }
  89
  90 static __kprobes unsigned long fetch_stack_address(struct pt_regs *regs,
  91                                                    void *dummy)
  92 {
  93         return kernel_stack_pointer(regs);
  94 }
  95
  96 /* Memory fetching by symbol */
  97 struct symbol_cache {
  98         char *symbol;
  99         long offset;
 100         unsigned long addr;
 101 };
 102
 103 static unsigned long update_symbol_cache(struct symbol_cache *sc)
 104 {
 105         sc->addr = (unsigned long)kallsyms_lookup_name(sc->symbol);
 106         if (sc->addr)
 107                 sc->addr += sc->offset;
 108         return sc->addr;
 109 }
 110
 111 static void free_symbol_cache(struct symbol_cache *sc)
 112 {
 113         kfree(sc->symbol);
 114         kfree(sc);
 115 }
 116
 117 static struct symbol_cache *alloc_symbol_cache(const char *sym, long offset)
 118 {
 119         struct symbol_cache *sc;
 120
 121         if (!sym || strlen(sym) == 0)
 122                 return NULL;
 123         sc = kzalloc(sizeof(struct symbol_cache), GFP_KERNEL);
 124         if (!sc)
 125                 return NULL;
 126
 127         sc->symbol = kstrdup(sym, GFP_KERNEL);
 128         if (!sc->symbol) {
 129                 kfree(sc);
 130                 return NULL;
 131         }
 132         sc->offset = offset;
 133
 134         update_symbol_cache(sc);
 135         return sc;
 136 }
 137
 138 static __kprobes unsigned long fetch_symbol(struct pt_regs *regs, void *data)
 139 {
 140         struct symbol_cache *sc = data;
 141
 142         if (sc->addr)
 143                 return fetch_memory(regs, (void *)sc->addr);
 144         else
 145                 return 0;
 146 }
 147
 148 /* Special indirect memory access interface */
 149 struct indirect_fetch_data {
 150         struct fetch_func orig;
 151         long offset;
 152 };
 153
 154 static __kprobes unsigned long fetch_indirect(struct pt_regs *regs, void *data)
 155 {
 156         struct indirect_fetch_data *ind = data;
 157         unsigned long addr;
 158
 159         addr = call_fetch(&ind->orig, regs);
 160         if (addr) {
 161                 addr += ind->offset;
 162                 return fetch_memory(regs, (void *)addr);
 163         } else
 164                 return 0;
 165 }
 166
 167 static __kprobes void free_indirect_fetch_data(struct indirect_fetch_data *data)
 168 {
 169         if (data->orig.func == fetch_indirect)
 170                 free_indirect_fetch_data(data->orig.data);
 171         else if (data->orig.func == fetch_symbol)
 172                 free_symbol_cache(data->orig.data);
 173         kfree(data);
 174 }
 175
 176 /**
 177  * kprobe_trace_core
 178  */
 179
 180 struct trace_probe {
 181         struct list_head        list;
 182         union {
 183                 struct kprobe           kp;
 184                 struct kretprobe        rp;
 185         };
 186         const char              *symbol;        /* symbol name */
 187         struct ftrace_event_call        call;
 188         unsigned int            nr_args;
 189         struct fetch_func       args[];
 190 };
 191
 192 #define SIZEOF_TRACE_PROBE(n)                   \
 193         (offsetof(struct trace_probe, args) +   \
 194         (sizeof(struct fetch_func) * (n)))
 195
 196 static int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs);
 197 static int kretprobe_trace_func(struct kretprobe_instance *ri,
 198                                 struct pt_regs *regs);
 199
 200 static __kprobes int probe_is_return(struct trace_probe *tp)
 201 {
 202         return (tp->rp.handler == kretprobe_trace_func);
 203 }
 204
 205 static __kprobes const char *probe_symbol(struct trace_probe *tp)
 206 {
 207         return tp->symbol ? tp->symbol : "unknown";
 208 }
 209
 210 static __kprobes long probe_offset(struct trace_probe *tp)
 211 {
 212         return (probe_is_return(tp)) ? tp->rp.kp.offset : tp->kp.offset;
 213 }
 214
 215 static __kprobes void *probe_address(struct trace_probe *tp)
 216 {
 217         return (probe_is_return(tp)) ? tp->rp.kp.addr : tp->kp.addr;
 218 }
 219
 220 static int trace_arg_string(char *buf, size_t n, struct fetch_func *ff)
 221 {
 222         int ret = -EINVAL;
 223
 224         if (ff->func == fetch_argument)
 225                 ret = snprintf(buf, n, "a%lu", (unsigned long)ff->data);
 226         else if (ff->func == fetch_register) {
 227                 const char *name;
 228                 name = regs_query_register_name((unsigned int)((long)ff->data));
 229                 ret = snprintf(buf, n, "%%%s", name);
 230         } else if (ff->func == fetch_stack)
 231                 ret = snprintf(buf, n, "s%lu", (unsigned long)ff->data);
 232         else if (ff->func == fetch_memory)
 233                 ret = snprintf(buf, n, "@0x%p", ff->data);
 234         else if (ff->func == fetch_symbol) {
 235                 struct symbol_cache *sc = ff->data;
 236                 ret = snprintf(buf, n, "@%s%+ld", sc->symbol, sc->offset);
 237         } else if (ff->func == fetch_retvalue)
 238                 ret = snprintf(buf, n, "rv");
 239         else if (ff->func == fetch_ip)
 240                 ret = snprintf(buf, n, "ra");
 241         else if (ff->func == fetch_stack_address)
 242                 ret = snprintf(buf, n, "sa");
 243         else if (ff->func == fetch_indirect) {
 244                 struct indirect_fetch_data *id = ff->data;
 245                 size_t l = 0;
 246                 ret = snprintf(buf, n, "%+ld(", id->offset);
 247                 if (ret >= n)
 248                         goto end;
 249                 l += ret;
 250                 ret = trace_arg_string(buf + l, n - l, &id->orig);
 251                 if (ret < 0)
 252                         goto end;
 253                 l += ret;
 254                 ret = snprintf(buf + l, n - l, ")");
 255                 ret += l;
 256         }
 257 end:
 258         if (ret >= n)
 259                 return -ENOSPC;
 260         return ret;
 261 }
 262
 263 static int register_probe_event(struct trace_probe *tp);
 264 static void unregister_probe_event(struct trace_probe *tp);
 265
 266 static DEFINE_MUTEX(probe_lock);
 267 static LIST_HEAD(probe_list);
 268
 269 static struct trace_probe *alloc_trace_probe(const char *symbol,
 270                                              const char *event, int nargs)
 271 {
 272         struct trace_probe *tp;
 273
 274         tp = kzalloc(SIZEOF_TRACE_PROBE(nargs), GFP_KERNEL);
 275         if (!tp)
 276                 return ERR_PTR(-ENOMEM);
 277
 278         if (symbol) {
 279                 tp->symbol = kstrdup(symbol, GFP_KERNEL);
 280                 if (!tp->symbol)
 281                         goto error;
 282         }
 283         if (event) {
 284                 tp->call.name = kstrdup(event, GFP_KERNEL);
 285                 if (!tp->call.name)
 286                         goto error;
 287         }
 288
 289         INIT_LIST_HEAD(&tp->list);
 290         return tp;
 291 error:
 292         kfree(tp->symbol);
 293         kfree(tp);
 294         return ERR_PTR(-ENOMEM);
 295 }
 296
 297 static void free_trace_probe(struct trace_probe *tp)
 298 {
 299         int i;
 300
 301         for (i = 0; i < tp->nr_args; i++)
 302                 if (tp->args[i].func == fetch_symbol)
 303                         free_symbol_cache(tp->args[i].data);
 304                 else if (tp->args[i].func == fetch_indirect)
 305                         free_indirect_fetch_data(tp->args[i].data);
 306
 307         kfree(tp->call.name);
 308         kfree(tp->symbol);
 309         kfree(tp);
 310 }
 311
 312 static struct trace_probe *find_probe_event(const char *event)
 313 {
 314         struct trace_probe *tp;
 315
 316         list_for_each_entry(tp, &probe_list, list)
 317                 if (tp->call.name && !strcmp(tp->call.name, event))
 318                         return tp;
 319         return NULL;
 320 }
 321
 322 static void __unregister_trace_probe(struct trace_probe *tp)
 323 {
 324         if (probe_is_return(tp))
 325                 unregister_kretprobe(&tp->rp);
 326         else
 327                 unregister_kprobe(&tp->kp);
 328 }
 329
 330 /* Unregister a trace_probe and probe_event: call with locking probe_lock */
 331 static void unregister_trace_probe(struct trace_probe *tp)
 332 {
 333         if (tp->call.name)
 334                 unregister_probe_event(tp);
 335         __unregister_trace_probe(tp);
 336         list_del(&tp->list);
 337 }
 338
 339 /* Register a trace_probe and probe_event */
 340 static int register_trace_probe(struct trace_probe *tp)
 341 {
 342         struct trace_probe *old_tp;
 343         int ret;
 344
 345         mutex_lock(&probe_lock);
 346
 347         if (probe_is_return(tp))
 348                 ret = register_kretprobe(&tp->rp);
 349         else
 350                 ret = register_kprobe(&tp->kp);
 351
 352         if (ret) {
 353                 pr_warning("Could not insert probe(%d)\n", ret);
 354                 if (ret == -EILSEQ) {
 355                         pr_warning("Probing address(0x%p) is not an "
 356                                    "instruction boundary.\n",
 357                                    probe_address(tp));
 358                         ret = -EINVAL;
 359                 }
 360                 goto end;
 361         }
 362         /* register as an event */
 363         if (tp->call.name) {
 364                 old_tp = find_probe_event(tp->call.name);
 365                 if (old_tp) {
 366                         /* delete old event */
 367                         unregister_trace_probe(old_tp);
 368                         free_trace_probe(old_tp);
 369                 }
 370                 ret = register_probe_event(tp);
 371                 if (ret) {
 372                         pr_warning("Faild to register probe event(%d)\n", ret);
 373                         __unregister_trace_probe(tp);
 374                 }
 375         }
 376         list_add_tail(&tp->list, &probe_list);
 377 end:
 378         mutex_unlock(&probe_lock);
 379         return ret;
 380 }
 381
 382 /* Split symbol and offset. */
 383 static int split_symbol_offset(char *symbol, long *offset)
 384 {
 385         char *tmp;
 386         int ret;
 387
 388         if (!offset)
 389                 return -EINVAL;
 390
 391         tmp = strchr(symbol, '+');
 392         if (!tmp)
 393                 tmp = strchr(symbol, '-');
 394
 395         if (tmp) {
 396                 /* skip sign because strict_strtol doesn't accept '+' */
 397                 ret = strict_strtol(tmp + 1, 0, offset);
 398                 if (ret)
 399                         return ret;
 400                 if (*tmp == '-')
 401                         *offset = -(*offset);
 402                 *tmp = '\0';
 403         } else
 404                 *offset = 0;
 405         return 0;
 406 }
 407
 408 #define PARAM_MAX_ARGS 16
 409 #define PARAM_MAX_STACK (THREAD_SIZE / sizeof(unsigned long))
 410
 411 static int parse_trace_arg(char *arg, struct fetch_func *ff, int is_return)
 412 {
 413         int ret = 0;
 414         unsigned long param;
 415         long offset;
 416         char *tmp;
 417
 418         switch (arg[0]) {
 419         case 'a':       /* argument */
 420                 ret = strict_strtoul(arg + 1, 10, &param);
 421                 if (ret || param > PARAM_MAX_ARGS)
 422                         ret = -EINVAL;
 423                 else {
 424                         ff->func = fetch_argument;
 425                         ff->data = (void *)param;
 426                 }
 427                 break;
 428         case 'r':       /* retval or retaddr */
 429                 if (is_return && arg[1] == 'v') {
 430                         ff->func = fetch_retvalue;
 431                         ff->data = NULL;
 432                 } else if (is_return && arg[1] == 'a') {
 433                         ff->func = fetch_ip;
 434                         ff->data = NULL;
 435                 } else
 436                         ret = -EINVAL;
 437                 break;
 438         case '%':       /* named register */
 439                 ret = regs_query_register_offset(arg + 1);
 440                 if (ret >= 0) {
 441                         ff->func = fetch_register;
 442                         ff->data = (void *)(unsigned long)ret;
 443                         ret = 0;
 444                 }
 445                 break;
 446         case 's':       /* stack */
 447                 if (arg[1] == 'a') {
 448                         ff->func = fetch_stack_address;
 449                         ff->data = NULL;
 450                 } else {
 451                         ret = strict_strtoul(arg + 1, 10, &param);
 452                         if (ret || param > PARAM_MAX_STACK)
 453                                 ret = -EINVAL;
 454                         else {
 455                                 ff->func = fetch_stack;
 456                                 ff->data = (void *)param;
 457                         }
 458                 }
 459                 break;
 460         case '@':       /* memory or symbol */
 461                 if (isdigit(arg[1])) {
 462                         ret = strict_strtoul(arg + 1, 0, &param);
 463                         if (ret)
 464                                 break;
 465                         ff->func = fetch_memory;
 466                         ff->data = (void *)param;
 467                 } else {
 468                         ret = split_symbol_offset(arg + 1, &offset);
 469                         if (ret)
 470                                 break;
 471                         ff->data = alloc_symbol_cache(arg + 1,
 472                                                               offset);
 473                         if (ff->data)
 474                                 ff->func = fetch_symbol;
 475                         else
 476                                 ret = -EINVAL;
 477                 }
 478                 break;
 479         case '+':       /* indirect memory */
 480         case '-':
 481                 tmp = strchr(arg, '(');
 482                 if (!tmp) {
 483                         ret = -EINVAL;
 484                         break;
 485                 }
 486                 *tmp = '\0';
 487                 ret = strict_strtol(arg + 1, 0, &offset);
 488                 if (ret)
 489                         break;
 490                 if (arg[0] == '-')
 491                         offset = -offset;
 492                 arg = tmp + 1;
 493                 tmp = strrchr(arg, ')');
 494                 if (tmp) {
 495                         struct indirect_fetch_data *id;
 496                         *tmp = '\0';
 497                         id = kzalloc(sizeof(struct indirect_fetch_data),
 498                                      GFP_KERNEL);
 499                         if (!id)
 500                                 return -ENOMEM;
 501                         id->offset = offset;
 502                         ret = parse_trace_arg(arg, &id->orig, is_return);
 503                         if (ret)
 504                                 kfree(id);
 505                         else {
 506                                 ff->func = fetch_indirect;
 507                                 ff->data = (void *)id;
 508                         }
 509                 } else
 510                         ret = -EINVAL;
 511                 break;
 512         default:
 513                 /* TODO: support custom handler */
 514                 ret = -EINVAL;
 515         }
 516         return ret;
 517 }
 518
 519 static int create_trace_probe(int argc, char **argv)
 520 {
 521         /*
 522          * Argument syntax:
 523          *  - Add kprobe: p[:EVENT] SYMBOL[+OFFS|-OFFS]|ADDRESS [FETCHARGS]
 524          *  - Add kretprobe: r[:EVENT] SYMBOL[+0] [FETCHARGS]
 525          * Fetch args:
 526          *  aN  : fetch Nth of function argument. (N:0-)
 527          *  rv  : fetch return value
 528          *  ra  : fetch return address
 529          *  sa  : fetch stack address
 530          *  sN  : fetch Nth of stack (N:0-)
 531          *  @ADDR       : fetch memory at ADDR (ADDR should be in kernel)
 532          *  @SYM[+|-offs] : fetch memory at SYM +|- offs (SYM is a data symbol)
 533          *  %REG        : fetch register REG
 534          * Indirect memory fetch:
 535          *  +|-offs(ARG) : fetch memory at ARG +|- offs address.
 536          */
 537         struct trace_probe *tp;
 538         struct kprobe *kp;
 539         int i, ret = 0;
 540         int is_return = 0;
 541         char *symbol = NULL, *event = NULL;
 542         long offset = 0;
 543         void *addr = NULL;
 544
 545         if (argc < 2)
 546                 return -EINVAL;
 547
 548         if (argv[0][0] == 'p')
 549                 is_return = 0;
 550         else if (argv[0][0] == 'r')
 551                 is_return = 1;
 552         else
 553                 return -EINVAL;
 554
 555         if (argv[0][1] == ':') {
 556                 event = &argv[0][2];
 557                 if (strlen(event) == 0) {
 558                         pr_info("Event name is not specifiled\n");
 559                         return -EINVAL;
 560                 }
 561         }
 562
 563         if (isdigit(argv[1][0])) {
 564                 if (is_return)
 565                         return -EINVAL;
 566                 /* an address specified */
 567                 ret = strict_strtoul(&argv[0][2], 0, (unsigned long *)&addr);
 568                 if (ret)
 569                         return ret;
 570         } else {
 571                 /* a symbol specified */
 572                 symbol = argv[1];
 573                 /* TODO: support .init module functions */
 574                 ret = split_symbol_offset(symbol, &offset);
 575                 if (ret)
 576                         return ret;
 577                 if (offset && is_return)
 578                         return -EINVAL;
 579         }
 580         argc -= 2; argv += 2;
 581
 582         /* setup a probe */
 583         tp = alloc_trace_probe(symbol, event, argc);
 584         if (IS_ERR(tp))
 585                 return PTR_ERR(tp);
 586
 587         if (is_return) {
 588                 kp = &tp->rp.kp;
 589                 tp->rp.handler = kretprobe_trace_func;
 590         } else {
 591                 kp = &tp->kp;
 592                 tp->kp.pre_handler = kprobe_trace_func;
 593         }
 594
 595         if (tp->symbol) {
 596                 kp->symbol_name = tp->symbol;
 597                 kp->offset = offset;
 598         } else
 599                 kp->addr = addr;
 600
 601         /* parse arguments */
 602         ret = 0;
 603         for (i = 0; i < argc && i < MAX_TRACE_ARGS; i++) {
 604                 if (strlen(argv[i]) > MAX_ARGSTR_LEN) {
 605                         pr_info("Argument%d(%s) is too long.\n", i, argv[i]);
 606                         ret = -ENOSPC;
 607                         goto error;
 608                 }
 609                 ret = parse_trace_arg(argv[i], &tp->args[i], is_return);
 610                 if (ret)
 611                         goto error;
 612         }
 613         tp->nr_args = i;
 614
 615         ret = register_trace_probe(tp);
 616         if (ret)
 617                 goto error;
 618         return 0;
 619
 620 error:
 621         free_trace_probe(tp);
 622         return ret;
 623 }
 624
 625 static void cleanup_all_probes(void)
 626 {
 627         struct trace_probe *tp;
 628
 629         mutex_lock(&probe_lock);
 630         /* TODO: Use batch unregistration */
 631         while (!list_empty(&probe_list)) {
 632                 tp = list_entry(probe_list.next, struct trace_probe, list);
 633                 unregister_trace_probe(tp);
 634                 free_trace_probe(tp);
 635         }
 636         mutex_unlock(&probe_lock);
 637 }
 638
 639
 640 /* Probes listing interfaces */
 641 static void *probes_seq_start(struct seq_file *m, loff_t *pos)
 642 {
 643         mutex_lock(&probe_lock);
 644         return seq_list_start(&probe_list, *pos);
 645 }
 646
 647 static void *probes_seq_next(struct seq_file *m, void *v, loff_t *pos)
 648 {
 649         return seq_list_next(v, &probe_list, pos);
 650 }
 651
 652 static void probes_seq_stop(struct seq_file *m, void *v)
 653 {
 654         mutex_unlock(&probe_lock);
 655 }
 656
 657 static int probes_seq_show(struct seq_file *m, void *v)
 658 {
 659         struct trace_probe *tp = v;
 660         int i, ret;
 661         char buf[MAX_ARGSTR_LEN + 1];
 662
 663         seq_printf(m, "%c", probe_is_return(tp) ? 'r' : 'p');
 664         if (tp->call.name)
 665                 seq_printf(m, ":%s", tp->call.name);
 666
 667         if (tp->symbol)
 668                 seq_printf(m, " %s%+ld", probe_symbol(tp), probe_offset(tp));
 669         else
 670                 seq_printf(m, " 0x%p", probe_address(tp));
 671
 672         for (i = 0; i < tp->nr_args; i++) {
 673                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 674                 if (ret < 0) {
 675                         pr_warning("Argument%d decoding error(%d).\n", i, ret);
 676                         return ret;
 677                 }
 678                 seq_printf(m, " %s", buf);
 679         }
 680         seq_printf(m, "\n");
 681         return 0;
 682 }
 683
 684 static const struct seq_operations probes_seq_op = {
 685         .start  = probes_seq_start,
 686         .next   = probes_seq_next,
 687         .stop   = probes_seq_stop,
 688         .show   = probes_seq_show
 689 };
 690
 691 static int probes_open(struct inode *inode, struct file *file)
 692 {
 693         if ((file->f_mode & FMODE_WRITE) &&
 694             (file->f_flags & O_TRUNC))
 695                 cleanup_all_probes();
 696
 697         return seq_open(file, &probes_seq_op);
 698 }
 699
 700 static int command_trace_probe(const char *buf)
 701 {
 702         char **argv;
 703         int argc = 0, ret = 0;
 704
 705         argv = argv_split(GFP_KERNEL, buf, &argc);
 706         if (!argv)
 707                 return -ENOMEM;
 708
 709         if (argc)
 710                 ret = create_trace_probe(argc, argv);
 711
 712         argv_free(argv);
 713         return ret;
 714 }
 715
 716 #define WRITE_BUFSIZE 128
 717
 718 static ssize_t probes_write(struct file *file, const char __user *buffer,
 719                             size_t count, loff_t *ppos)
 720 {
 721         char *kbuf, *tmp;
 722         int ret;
 723         size_t done;
 724         size_t size;
 725
 726         kbuf = kmalloc(WRITE_BUFSIZE, GFP_KERNEL);
 727         if (!kbuf)
 728                 return -ENOMEM;
 729
 730         ret = done = 0;
 731         while (done < count) {
 732                 size = count - done;
 733                 if (size >= WRITE_BUFSIZE)
 734                         size = WRITE_BUFSIZE - 1;
 735                 if (copy_from_user(kbuf, buffer + done, size)) {
 736                         ret = -EFAULT;
 737                         goto out;
 738                 }
 739                 kbuf[size] = '\0';
 740                 tmp = strchr(kbuf, '\n');
 741                 if (tmp) {
 742                         *tmp = '\0';
 743                         size = tmp - kbuf + 1;
 744                 } else if (done + size < count) {
 745                         pr_warning("Line length is too long: "
 746                                    "Should be less than %d.", WRITE_BUFSIZE);
 747                         ret = -EINVAL;
 748                         goto out;
 749                 }
 750                 done += size;
 751                 /* Remove comments */
 752                 tmp = strchr(kbuf, '#');
 753                 if (tmp)
 754                         *tmp = '\0';
 755
 756                 ret = command_trace_probe(kbuf);
 757                 if (ret)
 758                         goto out;
 759         }
 760         ret = done;
 761 out:
 762         kfree(kbuf);
 763         return ret;
 764 }
 765
 766 static const struct file_operations kprobe_events_ops = {
 767         .owner          = THIS_MODULE,
 768         .open           = probes_open,
 769         .read           = seq_read,
 770         .llseek         = seq_lseek,
 771         .release        = seq_release,
 772         .write          = probes_write,
 773 };
 774
 775 /* Kprobe handler */
 776 static __kprobes int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs)
 777 {
 778         struct trace_probe *tp = container_of(kp, struct trace_probe, kp);
 779         struct kprobe_trace_entry *entry;
 780         struct ring_buffer_event *event;
 781         int size, i, pc;
 782         unsigned long irq_flags;
 783         struct ftrace_event_call *call = &event_kprobe;
 784
 785         if (&tp->call.name)
 786                 call = &tp->call;
 787
 788         local_save_flags(irq_flags);
 789         pc = preempt_count();
 790
 791         size = SIZEOF_KPROBE_TRACE_ENTRY(tp->nr_args);
 792
 793         event = trace_current_buffer_lock_reserve(TRACE_KPROBE, size,
 794                                                   irq_flags, pc);
 795         if (!event)
 796                 return 0;
 797
 798         entry = ring_buffer_event_data(event);
 799         entry->nargs = tp->nr_args;
 800         entry->ip = (unsigned long)kp->addr;
 801         for (i = 0; i < tp->nr_args; i++)
 802                 entry->args[i] = call_fetch(&tp->args[i], regs);
 803
 804         if (!filter_current_check_discard(call, entry, event))
 805                 trace_nowake_buffer_unlock_commit(event, irq_flags, pc);
 806         return 0;
 807 }
 808
 809 /* Kretprobe handler */
 810 static __kprobes int kretprobe_trace_func(struct kretprobe_instance *ri,
 811                                           struct pt_regs *regs)
 812 {
 813         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
 814         struct kretprobe_trace_entry *entry;
 815         struct ring_buffer_event *event;
 816         int size, i, pc;
 817         unsigned long irq_flags;
 818         struct ftrace_event_call *call = &event_kretprobe;
 819
 820         if (&tp->call.name)
 821                 call = &tp->call;
 822
 823         local_save_flags(irq_flags);
 824         pc = preempt_count();
 825
 826         size = SIZEOF_KRETPROBE_TRACE_ENTRY(tp->nr_args);
 827
 828         event = trace_current_buffer_lock_reserve(TRACE_KRETPROBE, size,
 829                                                   irq_flags, pc);
 830         if (!event)
 831                 return 0;
 832
 833         entry = ring_buffer_event_data(event);
 834         entry->nargs = tp->nr_args;
 835         entry->func = (unsigned long)probe_address(tp);
 836         entry->ret_ip = (unsigned long)ri->ret_addr;
 837         for (i = 0; i < tp->nr_args; i++)
 838                 entry->args[i] = call_fetch(&tp->args[i], regs);
 839
 840         if (!filter_current_check_discard(call, entry, event))
 841                 trace_nowake_buffer_unlock_commit(event, irq_flags, pc);
 842
 843         return 0;
 844 }
 845
 846 /* Event entry printers */
 847 enum print_line_t
 848 print_kprobe_event(struct trace_iterator *iter, int flags)
 849 {
 850         struct kprobe_trace_entry *field;
 851         struct trace_seq *s = &iter->seq;
 852         int i;
 853
 854         trace_assign_type(field, iter->ent);
 855
 856         if (!seq_print_ip_sym(s, field->ip, flags | TRACE_ITER_SYM_OFFSET))
 857                 goto partial;
 858
 859         if (!trace_seq_puts(s, ":"))
 860                 goto partial;
 861
 862         for (i = 0; i < field->nargs; i++)
 863                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 864                         goto partial;
 865
 866         if (!trace_seq_puts(s, "\n"))
 867                 goto partial;
 868
 869         return TRACE_TYPE_HANDLED;
 870 partial:
 871         return TRACE_TYPE_PARTIAL_LINE;
 872 }
 873
 874 enum print_line_t
 875 print_kretprobe_event(struct trace_iterator *iter, int flags)
 876 {
 877         struct kretprobe_trace_entry *field;
 878         struct trace_seq *s = &iter->seq;
 879         int i;
 880
 881         trace_assign_type(field, iter->ent);
 882
 883         if (!seq_print_ip_sym(s, field->ret_ip, flags | TRACE_ITER_SYM_OFFSET))
 884                 goto partial;
 885
 886         if (!trace_seq_puts(s, " <- "))
 887                 goto partial;
 888
 889         if (!seq_print_ip_sym(s, field->func, flags & ~TRACE_ITER_SYM_OFFSET))
 890                 goto partial;
 891
 892         if (!trace_seq_puts(s, ":"))
 893                 goto partial;
 894
 895         for (i = 0; i < field->nargs; i++)
 896                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 897                         goto partial;
 898
 899         if (!trace_seq_puts(s, "\n"))
 900                 goto partial;
 901
 902         return TRACE_TYPE_HANDLED;
 903 partial:
 904         return TRACE_TYPE_PARTIAL_LINE;
 905 }
 906
 907 static struct trace_event kprobe_trace_event = {
 908         .type           = TRACE_KPROBE,
 909         .trace          = print_kprobe_event,
 910 };
 911
 912 static struct trace_event kretprobe_trace_event = {
 913         .type           = TRACE_KRETPROBE,
 914         .trace          = print_kretprobe_event,
 915 };
 916
 917 static int probe_event_enable(struct ftrace_event_call *call)
 918 {
 919         struct trace_probe *tp = (struct trace_probe *)call->data;
 920
 921         if (probe_is_return(tp))
 922                 return enable_kretprobe(&tp->rp);
 923         else
 924                 return enable_kprobe(&tp->kp);
 925 }
 926
 927 static void probe_event_disable(struct ftrace_event_call *call)
 928 {
 929         struct trace_probe *tp = (struct trace_probe *)call->data;
 930
 931         if (probe_is_return(tp))
 932                 disable_kretprobe(&tp->rp);
 933         else
 934                 disable_kprobe(&tp->kp);
 935 }
 936
 937 static int probe_event_raw_init(struct ftrace_event_call *event_call)
 938 {
 939         INIT_LIST_HEAD(&event_call->fields);
 940         init_preds(event_call);
 941         return 0;
 942 }
 943
 944 #undef DEFINE_FIELD
 945 #define DEFINE_FIELD(type, item, name, is_signed)                       \
 946         do {                                                            \
 947                 ret = trace_define_field(event_call, #type, name,       \
 948                                          offsetof(typeof(field), item), \
 949                                          sizeof(field.item), is_signed, \
 950                                          FILTER_OTHER);                 \
 951                 if (ret)                                                \
 952                         return ret;                                     \
 953         } while (0)
 954
 955 static int kprobe_event_define_fields(struct ftrace_event_call *event_call)
 956 {
 957         int ret, i;
 958         struct kprobe_trace_entry field;
 959         char buf[MAX_ARGSTR_LEN + 1];
 960         struct trace_probe *tp = (struct trace_probe *)event_call->data;
 961
 962         ret = trace_define_common_fields(event_call);
 963         if (!ret)
 964                 return ret;
 965
 966         DEFINE_FIELD(unsigned long, ip, "ip", 0);
 967         DEFINE_FIELD(int, nargs, "nargs", 1);
 968         for (i = 0; i < tp->nr_args; i++) {
 969                 /* Set argN as a field */
 970                 sprintf(buf, "arg%d", i);
 971                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
 972                 /* Set argument string as an alias field */
 973                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 974                 if (ret < 0)
 975                         return ret;
 976                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
 977         }
 978         return 0;
 979 }
 980
 981 static int kretprobe_event_define_fields(struct ftrace_event_call *event_call)
 982 {
 983         int ret, i;
 984         struct kretprobe_trace_entry field;
 985         char buf[MAX_ARGSTR_LEN + 1];
 986         struct trace_probe *tp = (struct trace_probe *)event_call->data;
 987
 988         ret = trace_define_common_fields(event_call);
 989         if (!ret)
 990                 return ret;
 991
 992         DEFINE_FIELD(unsigned long, func, "func", 0);
 993         DEFINE_FIELD(unsigned long, ret_ip, "ret_ip", 0);
 994         DEFINE_FIELD(int, nargs, "nargs", 1);
 995         for (i = 0; i < tp->nr_args; i++) {
 996                 /* Set argN as a field */
 997                 sprintf(buf, "arg%d", i);
 998                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
 999                 /* Set argument string as an alias field */
1000                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1001                 if (ret < 0)
1002                         return ret;
1003                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1004         }
1005         return 0;
1006 }
1007
1008 static int __probe_event_show_format(struct trace_seq *s,
1009                                      struct trace_probe *tp, const char *fmt,
1010                                      const char *arg)
1011 {
1012         int i, ret;
1013         char buf[MAX_ARGSTR_LEN + 1];
1014
1015         /* Show aliases */
1016         for (i = 0; i < tp->nr_args; i++) {
1017                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1018                 if (ret < 0)
1019                         return ret;
1020                 if (!trace_seq_printf(s, "\talias: %s;\toriginal: arg%d;\n",
1021                                       buf, i))
1022                         return 0;
1023         }
1024         /* Show format */
1025         if (!trace_seq_printf(s, "\nprint fmt: \"%s", fmt))
1026                 return 0;
1027
1028         for (i = 0; i < tp->nr_args; i++)
1029                 if (!trace_seq_puts(s, " 0x%lx"))
1030                         return 0;
1031
1032         if (!trace_seq_printf(s, "\", %s", arg))
1033                 return 0;
1034
1035         for (i = 0; i < tp->nr_args; i++)
1036                 if (!trace_seq_printf(s, ", arg%d", i))
1037                         return 0;
1038
1039         return trace_seq_puts(s, "\n");
1040 }
1041
1042 #undef SHOW_FIELD
1043 #define SHOW_FIELD(type, item, name)                                    \
1044         do {                                                            \
1045                 ret = trace_seq_printf(s, "\tfield: " #type " %s;\t"    \
1046                                 "offset:%u;tsize:%u;\n", name,          \
1047                                 (unsigned int)offsetof(typeof(field), item),\
1048                                 (unsigned int)sizeof(type));            \
1049                 if (!ret)                                               \
1050                         return 0;                                       \
1051         } while (0)
1052
1053 static int kprobe_event_show_format(struct ftrace_event_call *call,
1054                                     struct trace_seq *s)
1055 {
1056         struct kprobe_trace_entry field __attribute__((unused));
1057         int ret, i;
1058         char buf[8];
1059         struct trace_probe *tp = (struct trace_probe *)call->data;
1060
1061         SHOW_FIELD(unsigned long, ip, "ip");
1062         SHOW_FIELD(int, nargs, "nargs");
1063
1064         /* Show fields */
1065         for (i = 0; i < tp->nr_args; i++) {
1066                 sprintf(buf, "arg%d", i);
1067                 SHOW_FIELD(unsigned long, args[i], buf);
1068         }
1069         trace_seq_puts(s, "\n");
1070
1071         return __probe_event_show_format(s, tp, "%lx:", "ip");
1072 }
1073
1074 static int kretprobe_event_show_format(struct ftrace_event_call *call,
1075                                        struct trace_seq *s)
1076 {
1077         struct kretprobe_trace_entry field __attribute__((unused));
1078         int ret, i;
1079         char buf[8];
1080         struct trace_probe *tp = (struct trace_probe *)call->data;
1081
1082         SHOW_FIELD(unsigned long, func, "func");
1083         SHOW_FIELD(unsigned long, ret_ip, "ret_ip");
1084         SHOW_FIELD(int, nargs, "nargs");
1085
1086         /* Show fields */
1087         for (i = 0; i < tp->nr_args; i++) {
1088                 sprintf(buf, "arg%d", i);
1089                 SHOW_FIELD(unsigned long, args[i], buf);
1090         }
1091         trace_seq_puts(s, "\n");
1092
1093         return __probe_event_show_format(s, tp, "%lx <- %lx:",
1094                                           "func, ret_ip");
1095 }
1096
1097 static int register_probe_event(struct trace_probe *tp)
1098 {
1099         struct ftrace_event_call *call = &tp->call;
1100         int ret;
1101
1102         /* Initialize ftrace_event_call */
1103         call->system = "kprobes";
1104         if (probe_is_return(tp)) {
1105                 call->event = &kretprobe_trace_event;
1106                 call->id = TRACE_KRETPROBE;
1107                 call->raw_init = probe_event_raw_init;
1108                 call->show_format = kretprobe_event_show_format;
1109                 call->define_fields = kretprobe_event_define_fields;
1110         } else {
1111                 call->event = &kprobe_trace_event;
1112                 call->id = TRACE_KPROBE;
1113                 call->raw_init = probe_event_raw_init;
1114                 call->show_format = kprobe_event_show_format;
1115                 call->define_fields = kprobe_event_define_fields;
1116         }
1117         call->enabled = 1;
1118         call->regfunc = probe_event_enable;
1119         call->unregfunc = probe_event_disable;
1120         call->data = tp;
1121         ret = trace_add_event_call(call);
1122         if (ret)
1123                 pr_info("Failed to register kprobe event: %s\n", call->name);
1124         return ret;
1125 }
1126
1127 static void unregister_probe_event(struct trace_probe *tp)
1128 {
1129         /*
1130          * Prevent to unregister event itself because the event is shared
1131          * among other probes.
1132          */
1133         tp->call.event = NULL;
1134         trace_remove_event_call(&tp->call);
1135 }
1136
1137 /* Make a debugfs interface for controling probe points */
1138 static __init int init_kprobe_trace(void)
1139 {
1140         struct dentry *d_tracer;
1141         struct dentry *entry;
1142         int ret;
1143
1144         ret = register_ftrace_event(&kprobe_trace_event);
1145         if (!ret) {
1146                 pr_warning("Could not register kprobe_trace_event type.\n");
1147                 return 0;
1148         }
1149         ret = register_ftrace_event(&kretprobe_trace_event);
1150         if (!ret) {
1151                 pr_warning("Could not register kretprobe_trace_event type.\n");
1152                 return 0;
1153         }
1154
1155         d_tracer = tracing_init_dentry();
1156         if (!d_tracer)
1157                 return 0;
1158
1159         entry = debugfs_create_file("kprobe_events", 0644, d_tracer,
1160                                     NULL, &kprobe_events_ops);
1161
1162         if (!entry)
1163                 pr_warning("Could not create debugfs "
1164                            "'kprobe_events' entry\n");
1165         return 0;
1166 }
1167 fs_initcall(init_kprobe_trace);
1168
1169
1170 #ifdef CONFIG_FTRACE_STARTUP_TEST
1171
1172 static int kprobe_trace_selftest_target(int a1, int a2, int a3,
1173                                         int a4, int a5, int a6)
1174 {
1175         return a1 + a2 + a3 + a4 + a5 + a6;
1176 }
1177
1178 static __init int kprobe_trace_self_tests_init(void)
1179 {
1180         int ret;
1181         int (*target)(int, int, int, int, int, int);
1182
1183         target = kprobe_trace_selftest_target;
1184
1185         pr_info("Testing kprobe tracing: ");
1186
1187         ret = command_trace_probe("p:testprobe kprobe_trace_selftest_target "
1188                                   "a1 a2 a3 a4 a5 a6");
1189         if (WARN_ON_ONCE(ret))
1190                 pr_warning("error enabling function entry\n");
1191
1192         ret = command_trace_probe("r:testprobe2 kprobe_trace_selftest_target "
1193                                   "ra rv");
1194         if (WARN_ON_ONCE(ret))
1195                 pr_warning("error enabling function return\n");
1196
1197         ret = target(1, 2, 3, 4, 5, 6);
1198
1199         cleanup_all_probes();
1200
1201         pr_cont("OK\n");
1202         return 0;
1203 }
1204
1205 late_initcall(kprobe_trace_self_tests_init);
1206
1207 #endif