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 #define MAX_EVENT_NAME_LEN 64
  38
  39 /* currently, trace_kprobe only supports X86. */
  40
  41 struct fetch_func {
  42         unsigned long (*func)(struct pt_regs *, void *);
  43         void *data;
  44 };
  45
  46 static __kprobes unsigned long call_fetch(struct fetch_func *f,
  47                                           struct pt_regs *regs)
  48 {
  49         return f->func(regs, f->data);
  50 }
  51
  52 /* fetch handlers */
  53 static __kprobes unsigned long fetch_register(struct pt_regs *regs,
  54                                               void *offset)
  55 {
  56         return regs_get_register(regs, (unsigned int)((unsigned long)offset));
  57 }
  58
  59 static __kprobes unsigned long fetch_stack(struct pt_regs *regs,
  60                                            void *num)
  61 {
  62         return regs_get_kernel_stack_nth(regs,
  63                                          (unsigned int)((unsigned long)num));
  64 }
  65
  66 static __kprobes unsigned long fetch_memory(struct pt_regs *regs, void *addr)
  67 {
  68         unsigned long retval;
  69
  70         if (probe_kernel_address(addr, retval))
  71                 return 0;
  72         return retval;
  73 }
  74
  75 static __kprobes unsigned long fetch_argument(struct pt_regs *regs, void *num)
  76 {
  77         return regs_get_argument_nth(regs, (unsigned int)((unsigned long)num));
  78 }
  79
  80 static __kprobes unsigned long fetch_retvalue(struct pt_regs *regs,
  81                                               void *dummy)
  82 {
  83         return regs_return_value(regs);
  84 }
  85
  86 static __kprobes unsigned long fetch_ip(struct pt_regs *regs, void *dummy)
  87 {
  88         return instruction_pointer(regs);
  89 }
  90
  91 static __kprobes unsigned long fetch_stack_address(struct pt_regs *regs,
  92                                                    void *dummy)
  93 {
  94         return kernel_stack_pointer(regs);
  95 }
  96
  97 /* Memory fetching by symbol */
  98 struct symbol_cache {
  99         char *symbol;
 100         long offset;
 101         unsigned long addr;
 102 };
 103
 104 static unsigned long update_symbol_cache(struct symbol_cache *sc)
 105 {
 106         sc->addr = (unsigned long)kallsyms_lookup_name(sc->symbol);
 107         if (sc->addr)
 108                 sc->addr += sc->offset;
 109         return sc->addr;
 110 }
 111
 112 static void free_symbol_cache(struct symbol_cache *sc)
 113 {
 114         kfree(sc->symbol);
 115         kfree(sc);
 116 }
 117
 118 static struct symbol_cache *alloc_symbol_cache(const char *sym, long offset)
 119 {
 120         struct symbol_cache *sc;
 121
 122         if (!sym || strlen(sym) == 0)
 123                 return NULL;
 124         sc = kzalloc(sizeof(struct symbol_cache), GFP_KERNEL);
 125         if (!sc)
 126                 return NULL;
 127
 128         sc->symbol = kstrdup(sym, GFP_KERNEL);
 129         if (!sc->symbol) {
 130                 kfree(sc);
 131                 return NULL;
 132         }
 133         sc->offset = offset;
 134
 135         update_symbol_cache(sc);
 136         return sc;
 137 }
 138
 139 static __kprobes unsigned long fetch_symbol(struct pt_regs *regs, void *data)
 140 {
 141         struct symbol_cache *sc = data;
 142
 143         if (sc->addr)
 144                 return fetch_memory(regs, (void *)sc->addr);
 145         else
 146                 return 0;
 147 }
 148
 149 /* Special indirect memory access interface */
 150 struct indirect_fetch_data {
 151         struct fetch_func orig;
 152         long offset;
 153 };
 154
 155 static __kprobes unsigned long fetch_indirect(struct pt_regs *regs, void *data)
 156 {
 157         struct indirect_fetch_data *ind = data;
 158         unsigned long addr;
 159
 160         addr = call_fetch(&ind->orig, regs);
 161         if (addr) {
 162                 addr += ind->offset;
 163                 return fetch_memory(regs, (void *)addr);
 164         } else
 165                 return 0;
 166 }
 167
 168 static __kprobes void free_indirect_fetch_data(struct indirect_fetch_data *data)
 169 {
 170         if (data->orig.func == fetch_indirect)
 171                 free_indirect_fetch_data(data->orig.data);
 172         else if (data->orig.func == fetch_symbol)
 173                 free_symbol_cache(data->orig.data);
 174         kfree(data);
 175 }
 176
 177 /**
 178  * kprobe_trace_core
 179  */
 180
 181 struct trace_probe {
 182         struct list_head        list;
 183         union {
 184                 struct kprobe           kp;
 185                 struct kretprobe        rp;
 186         };
 187         const char              *symbol;        /* symbol name */
 188         struct ftrace_event_call        call;
 189         unsigned int            nr_args;
 190         struct fetch_func       args[];
 191 };
 192
 193 #define SIZEOF_TRACE_PROBE(n)                   \
 194         (offsetof(struct trace_probe, args) +   \
 195         (sizeof(struct fetch_func) * (n)))
 196
 197 static int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs);
 198 static int kretprobe_trace_func(struct kretprobe_instance *ri,
 199                                 struct pt_regs *regs);
 200
 201 static __kprobes int probe_is_return(struct trace_probe *tp)
 202 {
 203         return (tp->rp.handler == kretprobe_trace_func);
 204 }
 205
 206 static __kprobes const char *probe_symbol(struct trace_probe *tp)
 207 {
 208         return tp->symbol ? tp->symbol : "unknown";
 209 }
 210
 211 static __kprobes long probe_offset(struct trace_probe *tp)
 212 {
 213         return (probe_is_return(tp)) ? tp->rp.kp.offset : tp->kp.offset;
 214 }
 215
 216 static __kprobes void *probe_address(struct trace_probe *tp)
 217 {
 218         return (probe_is_return(tp)) ? tp->rp.kp.addr : tp->kp.addr;
 219 }
 220
 221 static int trace_arg_string(char *buf, size_t n, struct fetch_func *ff)
 222 {
 223         int ret = -EINVAL;
 224
 225         if (ff->func == fetch_argument)
 226                 ret = snprintf(buf, n, "a%lu", (unsigned long)ff->data);
 227         else if (ff->func == fetch_register) {
 228                 const char *name;
 229                 name = regs_query_register_name((unsigned int)((long)ff->data));
 230                 ret = snprintf(buf, n, "%%%s", name);
 231         } else if (ff->func == fetch_stack)
 232                 ret = snprintf(buf, n, "s%lu", (unsigned long)ff->data);
 233         else if (ff->func == fetch_memory)
 234                 ret = snprintf(buf, n, "@0x%p", ff->data);
 235         else if (ff->func == fetch_symbol) {
 236                 struct symbol_cache *sc = ff->data;
 237                 ret = snprintf(buf, n, "@%s%+ld", sc->symbol, sc->offset);
 238         } else if (ff->func == fetch_retvalue)
 239                 ret = snprintf(buf, n, "rv");
 240         else if (ff->func == fetch_ip)
 241                 ret = snprintf(buf, n, "ra");
 242         else if (ff->func == fetch_stack_address)
 243                 ret = snprintf(buf, n, "sa");
 244         else if (ff->func == fetch_indirect) {
 245                 struct indirect_fetch_data *id = ff->data;
 246                 size_t l = 0;
 247                 ret = snprintf(buf, n, "%+ld(", id->offset);
 248                 if (ret >= n)
 249                         goto end;
 250                 l += ret;
 251                 ret = trace_arg_string(buf + l, n - l, &id->orig);
 252                 if (ret < 0)
 253                         goto end;
 254                 l += ret;
 255                 ret = snprintf(buf + l, n - l, ")");
 256                 ret += l;
 257         }
 258 end:
 259         if (ret >= n)
 260                 return -ENOSPC;
 261         return ret;
 262 }
 263
 264 static int register_probe_event(struct trace_probe *tp);
 265 static void unregister_probe_event(struct trace_probe *tp);
 266
 267 static DEFINE_MUTEX(probe_lock);
 268 static LIST_HEAD(probe_list);
 269
 270 static struct trace_probe *alloc_trace_probe(const char *symbol,
 271                                              const char *event, int nargs)
 272 {
 273         struct trace_probe *tp;
 274
 275         tp = kzalloc(SIZEOF_TRACE_PROBE(nargs), GFP_KERNEL);
 276         if (!tp)
 277                 return ERR_PTR(-ENOMEM);
 278
 279         if (symbol) {
 280                 tp->symbol = kstrdup(symbol, GFP_KERNEL);
 281                 if (!tp->symbol)
 282                         goto error;
 283         }
 284         if (!event)
 285                 goto error;
 286         tp->call.name = kstrdup(event, GFP_KERNEL);
 287         if (!tp->call.name)
 288                 goto error;
 289
 290         INIT_LIST_HEAD(&tp->list);
 291         return tp;
 292 error:
 293         kfree(tp->symbol);
 294         kfree(tp);
 295         return ERR_PTR(-ENOMEM);
 296 }
 297
 298 static void free_trace_probe(struct trace_probe *tp)
 299 {
 300         int i;
 301
 302         for (i = 0; i < tp->nr_args; i++)
 303                 if (tp->args[i].func == fetch_symbol)
 304                         free_symbol_cache(tp->args[i].data);
 305                 else if (tp->args[i].func == fetch_indirect)
 306                         free_indirect_fetch_data(tp->args[i].data);
 307
 308         kfree(tp->call.name);
 309         kfree(tp->symbol);
 310         kfree(tp);
 311 }
 312
 313 static struct trace_probe *find_probe_event(const char *event)
 314 {
 315         struct trace_probe *tp;
 316
 317         list_for_each_entry(tp, &probe_list, list)
 318                 if (!strcmp(tp->call.name, event))
 319                         return tp;
 320         return NULL;
 321 }
 322
 323 static void __unregister_trace_probe(struct trace_probe *tp)
 324 {
 325         if (probe_is_return(tp))
 326                 unregister_kretprobe(&tp->rp);
 327         else
 328                 unregister_kprobe(&tp->kp);
 329 }
 330
 331 /* Unregister a trace_probe and probe_event: call with locking probe_lock */
 332 static void unregister_trace_probe(struct trace_probe *tp)
 333 {
 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         old_tp = find_probe_event(tp->call.name);
 364         if (old_tp) {
 365                 /* delete old event */
 366                 unregister_trace_probe(old_tp);
 367                 free_trace_probe(old_tp);
 368         }
 369         ret = register_probe_event(tp);
 370         if (ret) {
 371                 pr_warning("Faild to register probe event(%d)\n", ret);
 372                 __unregister_trace_probe(tp);
 373         }
 374         list_add_tail(&tp->list, &probe_list);
 375 end:
 376         mutex_unlock(&probe_lock);
 377         return ret;
 378 }
 379
 380 /* Split symbol and offset. */
 381 static int split_symbol_offset(char *symbol, long *offset)
 382 {
 383         char *tmp;
 384         int ret;
 385
 386         if (!offset)
 387                 return -EINVAL;
 388
 389         tmp = strchr(symbol, '+');
 390         if (!tmp)
 391                 tmp = strchr(symbol, '-');
 392
 393         if (tmp) {
 394                 /* skip sign because strict_strtol doesn't accept '+' */
 395                 ret = strict_strtol(tmp + 1, 0, offset);
 396                 if (ret)
 397                         return ret;
 398                 if (*tmp == '-')
 399                         *offset = -(*offset);
 400                 *tmp = '\0';
 401         } else
 402                 *offset = 0;
 403         return 0;
 404 }
 405
 406 #define PARAM_MAX_ARGS 16
 407 #define PARAM_MAX_STACK (THREAD_SIZE / sizeof(unsigned long))
 408
 409 static int parse_trace_arg(char *arg, struct fetch_func *ff, int is_return)
 410 {
 411         int ret = 0;
 412         unsigned long param;
 413         long offset;
 414         char *tmp;
 415
 416         switch (arg[0]) {
 417         case 'a':       /* argument */
 418                 ret = strict_strtoul(arg + 1, 10, &param);
 419                 if (ret || param > PARAM_MAX_ARGS)
 420                         ret = -EINVAL;
 421                 else {
 422                         ff->func = fetch_argument;
 423                         ff->data = (void *)param;
 424                 }
 425                 break;
 426         case 'r':       /* retval or retaddr */
 427                 if (is_return && arg[1] == 'v') {
 428                         ff->func = fetch_retvalue;
 429                         ff->data = NULL;
 430                 } else if (is_return && arg[1] == 'a') {
 431                         ff->func = fetch_ip;
 432                         ff->data = NULL;
 433                 } else
 434                         ret = -EINVAL;
 435                 break;
 436         case '%':       /* named register */
 437                 ret = regs_query_register_offset(arg + 1);
 438                 if (ret >= 0) {
 439                         ff->func = fetch_register;
 440                         ff->data = (void *)(unsigned long)ret;
 441                         ret = 0;
 442                 }
 443                 break;
 444         case 's':       /* stack */
 445                 if (arg[1] == 'a') {
 446                         ff->func = fetch_stack_address;
 447                         ff->data = NULL;
 448                 } else {
 449                         ret = strict_strtoul(arg + 1, 10, &param);
 450                         if (ret || param > PARAM_MAX_STACK)
 451                                 ret = -EINVAL;
 452                         else {
 453                                 ff->func = fetch_stack;
 454                                 ff->data = (void *)param;
 455                         }
 456                 }
 457                 break;
 458         case '@':       /* memory or symbol */
 459                 if (isdigit(arg[1])) {
 460                         ret = strict_strtoul(arg + 1, 0, &param);
 461                         if (ret)
 462                                 break;
 463                         ff->func = fetch_memory;
 464                         ff->data = (void *)param;
 465                 } else {
 466                         ret = split_symbol_offset(arg + 1, &offset);
 467                         if (ret)
 468                                 break;
 469                         ff->data = alloc_symbol_cache(arg + 1,
 470                                                               offset);
 471                         if (ff->data)
 472                                 ff->func = fetch_symbol;
 473                         else
 474                                 ret = -EINVAL;
 475                 }
 476                 break;
 477         case '+':       /* indirect memory */
 478         case '-':
 479                 tmp = strchr(arg, '(');
 480                 if (!tmp) {
 481                         ret = -EINVAL;
 482                         break;
 483                 }
 484                 *tmp = '\0';
 485                 ret = strict_strtol(arg + 1, 0, &offset);
 486                 if (ret)
 487                         break;
 488                 if (arg[0] == '-')
 489                         offset = -offset;
 490                 arg = tmp + 1;
 491                 tmp = strrchr(arg, ')');
 492                 if (tmp) {
 493                         struct indirect_fetch_data *id;
 494                         *tmp = '\0';
 495                         id = kzalloc(sizeof(struct indirect_fetch_data),
 496                                      GFP_KERNEL);
 497                         if (!id)
 498                                 return -ENOMEM;
 499                         id->offset = offset;
 500                         ret = parse_trace_arg(arg, &id->orig, is_return);
 501                         if (ret)
 502                                 kfree(id);
 503                         else {
 504                                 ff->func = fetch_indirect;
 505                                 ff->data = (void *)id;
 506                         }
 507                 } else
 508                         ret = -EINVAL;
 509                 break;
 510         default:
 511                 /* TODO: support custom handler */
 512                 ret = -EINVAL;
 513         }
 514         return ret;
 515 }
 516
 517 static int create_trace_probe(int argc, char **argv)
 518 {
 519         /*
 520          * Argument syntax:
 521          *  - Add kprobe: p[:EVENT] SYMBOL[+OFFS|-OFFS]|ADDRESS [FETCHARGS]
 522          *  - Add kretprobe: r[:EVENT] SYMBOL[+0] [FETCHARGS]
 523          * Fetch args:
 524          *  aN  : fetch Nth of function argument. (N:0-)
 525          *  rv  : fetch return value
 526          *  ra  : fetch return address
 527          *  sa  : fetch stack address
 528          *  sN  : fetch Nth of stack (N:0-)
 529          *  @ADDR       : fetch memory at ADDR (ADDR should be in kernel)
 530          *  @SYM[+|-offs] : fetch memory at SYM +|- offs (SYM is a data symbol)
 531          *  %REG        : fetch register REG
 532          * Indirect memory fetch:
 533          *  +|-offs(ARG) : fetch memory at ARG +|- offs address.
 534          */
 535         struct trace_probe *tp;
 536         struct kprobe *kp;
 537         int i, ret = 0;
 538         int is_return = 0;
 539         char *symbol = NULL, *event = NULL;
 540         long offset = 0;
 541         void *addr = NULL;
 542
 543         if (argc < 2)
 544                 return -EINVAL;
 545
 546         if (argv[0][0] == 'p')
 547                 is_return = 0;
 548         else if (argv[0][0] == 'r')
 549                 is_return = 1;
 550         else
 551                 return -EINVAL;
 552
 553         if (argv[0][1] == ':') {
 554                 event = &argv[0][2];
 555                 if (strlen(event) == 0) {
 556                         pr_info("Event name is not specifiled\n");
 557                         return -EINVAL;
 558                 }
 559         }
 560
 561         if (isdigit(argv[1][0])) {
 562                 if (is_return)
 563                         return -EINVAL;
 564                 /* an address specified */
 565                 ret = strict_strtoul(&argv[0][2], 0, (unsigned long *)&addr);
 566                 if (ret)
 567                         return ret;
 568         } else {
 569                 /* a symbol specified */
 570                 symbol = argv[1];
 571                 /* TODO: support .init module functions */
 572                 ret = split_symbol_offset(symbol, &offset);
 573                 if (ret)
 574                         return ret;
 575                 if (offset && is_return)
 576                         return -EINVAL;
 577         }
 578         argc -= 2; argv += 2;
 579
 580         /* setup a probe */
 581         if (!event) {
 582                 /* Make a new event name */
 583                 char buf[MAX_EVENT_NAME_LEN];
 584                 if (symbol)
 585                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c@%s%+ld",
 586                                  is_return ? 'r' : 'p', symbol, offset);
 587                 else
 588                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c@0x%p",
 589                                  is_return ? 'r' : 'p', addr);
 590                 tp = alloc_trace_probe(symbol, buf, argc);
 591         } else
 592                 tp = alloc_trace_probe(symbol, event, argc);
 593         if (IS_ERR(tp))
 594                 return PTR_ERR(tp);
 595
 596         if (is_return) {
 597                 kp = &tp->rp.kp;
 598                 tp->rp.handler = kretprobe_trace_func;
 599         } else {
 600                 kp = &tp->kp;
 601                 tp->kp.pre_handler = kprobe_trace_func;
 602         }
 603
 604         if (tp->symbol) {
 605                 kp->symbol_name = tp->symbol;
 606                 kp->offset = offset;
 607         } else
 608                 kp->addr = addr;
 609
 610         /* parse arguments */
 611         ret = 0;
 612         for (i = 0; i < argc && i < MAX_TRACE_ARGS; i++) {
 613                 if (strlen(argv[i]) > MAX_ARGSTR_LEN) {
 614                         pr_info("Argument%d(%s) is too long.\n", i, argv[i]);
 615                         ret = -ENOSPC;
 616                         goto error;
 617                 }
 618                 ret = parse_trace_arg(argv[i], &tp->args[i], is_return);
 619                 if (ret)
 620                         goto error;
 621         }
 622         tp->nr_args = i;
 623
 624         ret = register_trace_probe(tp);
 625         if (ret)
 626                 goto error;
 627         return 0;
 628
 629 error:
 630         free_trace_probe(tp);
 631         return ret;
 632 }
 633
 634 static void cleanup_all_probes(void)
 635 {
 636         struct trace_probe *tp;
 637
 638         mutex_lock(&probe_lock);
 639         /* TODO: Use batch unregistration */
 640         while (!list_empty(&probe_list)) {
 641                 tp = list_entry(probe_list.next, struct trace_probe, list);
 642                 unregister_trace_probe(tp);
 643                 free_trace_probe(tp);
 644         }
 645         mutex_unlock(&probe_lock);
 646 }
 647
 648
 649 /* Probes listing interfaces */
 650 static void *probes_seq_start(struct seq_file *m, loff_t *pos)
 651 {
 652         mutex_lock(&probe_lock);
 653         return seq_list_start(&probe_list, *pos);
 654 }
 655
 656 static void *probes_seq_next(struct seq_file *m, void *v, loff_t *pos)
 657 {
 658         return seq_list_next(v, &probe_list, pos);
 659 }
 660
 661 static void probes_seq_stop(struct seq_file *m, void *v)
 662 {
 663         mutex_unlock(&probe_lock);
 664 }
 665
 666 static int probes_seq_show(struct seq_file *m, void *v)
 667 {
 668         struct trace_probe *tp = v;
 669         int i, ret;
 670         char buf[MAX_ARGSTR_LEN + 1];
 671
 672         seq_printf(m, "%c", probe_is_return(tp) ? 'r' : 'p');
 673         seq_printf(m, ":%s", tp->call.name);
 674
 675         if (tp->symbol)
 676                 seq_printf(m, " %s%+ld", probe_symbol(tp), probe_offset(tp));
 677         else
 678                 seq_printf(m, " 0x%p", probe_address(tp));
 679
 680         for (i = 0; i < tp->nr_args; i++) {
 681                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 682                 if (ret < 0) {
 683                         pr_warning("Argument%d decoding error(%d).\n", i, ret);
 684                         return ret;
 685                 }
 686                 seq_printf(m, " %s", buf);
 687         }
 688         seq_printf(m, "\n");
 689         return 0;
 690 }
 691
 692 static const struct seq_operations probes_seq_op = {
 693         .start  = probes_seq_start,
 694         .next   = probes_seq_next,
 695         .stop   = probes_seq_stop,
 696         .show   = probes_seq_show
 697 };
 698
 699 static int probes_open(struct inode *inode, struct file *file)
 700 {
 701         if ((file->f_mode & FMODE_WRITE) &&
 702             (file->f_flags & O_TRUNC))
 703                 cleanup_all_probes();
 704
 705         return seq_open(file, &probes_seq_op);
 706 }
 707
 708 static int command_trace_probe(const char *buf)
 709 {
 710         char **argv;
 711         int argc = 0, ret = 0;
 712
 713         argv = argv_split(GFP_KERNEL, buf, &argc);
 714         if (!argv)
 715                 return -ENOMEM;
 716
 717         if (argc)
 718                 ret = create_trace_probe(argc, argv);
 719
 720         argv_free(argv);
 721         return ret;
 722 }
 723
 724 #define WRITE_BUFSIZE 128
 725
 726 static ssize_t probes_write(struct file *file, const char __user *buffer,
 727                             size_t count, loff_t *ppos)
 728 {
 729         char *kbuf, *tmp;
 730         int ret;
 731         size_t done;
 732         size_t size;
 733
 734         kbuf = kmalloc(WRITE_BUFSIZE, GFP_KERNEL);
 735         if (!kbuf)
 736                 return -ENOMEM;
 737
 738         ret = done = 0;
 739         while (done < count) {
 740                 size = count - done;
 741                 if (size >= WRITE_BUFSIZE)
 742                         size = WRITE_BUFSIZE - 1;
 743                 if (copy_from_user(kbuf, buffer + done, size)) {
 744                         ret = -EFAULT;
 745                         goto out;
 746                 }
 747                 kbuf[size] = '\0';
 748                 tmp = strchr(kbuf, '\n');
 749                 if (tmp) {
 750                         *tmp = '\0';
 751                         size = tmp - kbuf + 1;
 752                 } else if (done + size < count) {
 753                         pr_warning("Line length is too long: "
 754                                    "Should be less than %d.", WRITE_BUFSIZE);
 755                         ret = -EINVAL;
 756                         goto out;
 757                 }
 758                 done += size;
 759                 /* Remove comments */
 760                 tmp = strchr(kbuf, '#');
 761                 if (tmp)
 762                         *tmp = '\0';
 763
 764                 ret = command_trace_probe(kbuf);
 765                 if (ret)
 766                         goto out;
 767         }
 768         ret = done;
 769 out:
 770         kfree(kbuf);
 771         return ret;
 772 }
 773
 774 static const struct file_operations kprobe_events_ops = {
 775         .owner          = THIS_MODULE,
 776         .open           = probes_open,
 777         .read           = seq_read,
 778         .llseek         = seq_lseek,
 779         .release        = seq_release,
 780         .write          = probes_write,
 781 };
 782
 783 /* Kprobe handler */
 784 static __kprobes int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs)
 785 {
 786         struct trace_probe *tp = container_of(kp, struct trace_probe, kp);
 787         struct kprobe_trace_entry *entry;
 788         struct ring_buffer_event *event;
 789         int size, i, pc;
 790         unsigned long irq_flags;
 791         struct ftrace_event_call *call = &tp->call;
 792
 793         local_save_flags(irq_flags);
 794         pc = preempt_count();
 795
 796         size = SIZEOF_KPROBE_TRACE_ENTRY(tp->nr_args);
 797
 798         event = trace_current_buffer_lock_reserve(TRACE_KPROBE, size,
 799                                                   irq_flags, pc);
 800         if (!event)
 801                 return 0;
 802
 803         entry = ring_buffer_event_data(event);
 804         entry->nargs = tp->nr_args;
 805         entry->ip = (unsigned long)kp->addr;
 806         for (i = 0; i < tp->nr_args; i++)
 807                 entry->args[i] = call_fetch(&tp->args[i], regs);
 808
 809         if (!filter_current_check_discard(call, entry, event))
 810                 trace_nowake_buffer_unlock_commit(event, irq_flags, pc);
 811         return 0;
 812 }
 813
 814 /* Kretprobe handler */
 815 static __kprobes int kretprobe_trace_func(struct kretprobe_instance *ri,
 816                                           struct pt_regs *regs)
 817 {
 818         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
 819         struct kretprobe_trace_entry *entry;
 820         struct ring_buffer_event *event;
 821         int size, i, pc;
 822         unsigned long irq_flags;
 823         struct ftrace_event_call *call = &tp->call;
 824
 825         local_save_flags(irq_flags);
 826         pc = preempt_count();
 827
 828         size = SIZEOF_KRETPROBE_TRACE_ENTRY(tp->nr_args);
 829
 830         event = trace_current_buffer_lock_reserve(TRACE_KRETPROBE, size,
 831                                                   irq_flags, pc);
 832         if (!event)
 833                 return 0;
 834
 835         entry = ring_buffer_event_data(event);
 836         entry->nargs = tp->nr_args;
 837         entry->func = (unsigned long)probe_address(tp);
 838         entry->ret_ip = (unsigned long)ri->ret_addr;
 839         for (i = 0; i < tp->nr_args; i++)
 840                 entry->args[i] = call_fetch(&tp->args[i], regs);
 841
 842         if (!filter_current_check_discard(call, entry, event))
 843                 trace_nowake_buffer_unlock_commit(event, irq_flags, pc);
 844
 845         return 0;
 846 }
 847
 848 /* Event entry printers */
 849 enum print_line_t
 850 print_kprobe_event(struct trace_iterator *iter, int flags)
 851 {
 852         struct kprobe_trace_entry *field;
 853         struct trace_seq *s = &iter->seq;
 854         int i;
 855
 856         trace_assign_type(field, iter->ent);
 857
 858         if (!seq_print_ip_sym(s, field->ip, flags | TRACE_ITER_SYM_OFFSET))
 859                 goto partial;
 860
 861         if (!trace_seq_puts(s, ":"))
 862                 goto partial;
 863
 864         for (i = 0; i < field->nargs; i++)
 865                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 866                         goto partial;
 867
 868         if (!trace_seq_puts(s, "\n"))
 869                 goto partial;
 870
 871         return TRACE_TYPE_HANDLED;
 872 partial:
 873         return TRACE_TYPE_PARTIAL_LINE;
 874 }
 875
 876 enum print_line_t
 877 print_kretprobe_event(struct trace_iterator *iter, int flags)
 878 {
 879         struct kretprobe_trace_entry *field;
 880         struct trace_seq *s = &iter->seq;
 881         int i;
 882
 883         trace_assign_type(field, iter->ent);
 884
 885         if (!seq_print_ip_sym(s, field->ret_ip, flags | TRACE_ITER_SYM_OFFSET))
 886                 goto partial;
 887
 888         if (!trace_seq_puts(s, " <- "))
 889                 goto partial;
 890
 891         if (!seq_print_ip_sym(s, field->func, flags & ~TRACE_ITER_SYM_OFFSET))
 892                 goto partial;
 893
 894         if (!trace_seq_puts(s, ":"))
 895                 goto partial;
 896
 897         for (i = 0; i < field->nargs; i++)
 898                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 899                         goto partial;
 900
 901         if (!trace_seq_puts(s, "\n"))
 902                 goto partial;
 903
 904         return TRACE_TYPE_HANDLED;
 905 partial:
 906         return TRACE_TYPE_PARTIAL_LINE;
 907 }
 908
 909 static struct trace_event kprobe_trace_event = {
 910         .type           = TRACE_KPROBE,
 911         .trace          = print_kprobe_event,
 912 };
 913
 914 static struct trace_event kretprobe_trace_event = {
 915         .type           = TRACE_KRETPROBE,
 916         .trace          = print_kretprobe_event,
 917 };
 918
 919 static int probe_event_enable(struct ftrace_event_call *call)
 920 {
 921         struct trace_probe *tp = (struct trace_probe *)call->data;
 922
 923         if (probe_is_return(tp))
 924                 return enable_kretprobe(&tp->rp);
 925         else
 926                 return enable_kprobe(&tp->kp);
 927 }
 928
 929 static void probe_event_disable(struct ftrace_event_call *call)
 930 {
 931         struct trace_probe *tp = (struct trace_probe *)call->data;
 932
 933         if (probe_is_return(tp))
 934                 disable_kretprobe(&tp->rp);
 935         else
 936                 disable_kprobe(&tp->kp);
 937 }
 938
 939 static int probe_event_raw_init(struct ftrace_event_call *event_call)
 940 {
 941         INIT_LIST_HEAD(&event_call->fields);
 942         init_preds(event_call);
 943         return 0;
 944 }
 945
 946 #undef DEFINE_FIELD
 947 #define DEFINE_FIELD(type, item, name, is_signed)                       \
 948         do {                                                            \
 949                 ret = trace_define_field(event_call, #type, name,       \
 950                                          offsetof(typeof(field), item), \
 951                                          sizeof(field.item), is_signed, \
 952                                          FILTER_OTHER);                 \
 953                 if (ret)                                                \
 954                         return ret;                                     \
 955         } while (0)
 956
 957 static int kprobe_event_define_fields(struct ftrace_event_call *event_call)
 958 {
 959         int ret, i;
 960         struct kprobe_trace_entry field;
 961         char buf[MAX_ARGSTR_LEN + 1];
 962         struct trace_probe *tp = (struct trace_probe *)event_call->data;
 963
 964         ret = trace_define_common_fields(event_call);
 965         if (!ret)
 966                 return ret;
 967
 968         DEFINE_FIELD(unsigned long, ip, "ip", 0);
 969         DEFINE_FIELD(int, nargs, "nargs", 1);
 970         for (i = 0; i < tp->nr_args; i++) {
 971                 /* Set argN as a field */
 972                 sprintf(buf, "arg%d", i);
 973                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
 974                 /* Set argument string as an alias field */
 975                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 976                 if (ret < 0)
 977                         return ret;
 978                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
 979         }
 980         return 0;
 981 }
 982
 983 static int kretprobe_event_define_fields(struct ftrace_event_call *event_call)
 984 {
 985         int ret, i;
 986         struct kretprobe_trace_entry field;
 987         char buf[MAX_ARGSTR_LEN + 1];
 988         struct trace_probe *tp = (struct trace_probe *)event_call->data;
 989
 990         ret = trace_define_common_fields(event_call);
 991         if (!ret)
 992                 return ret;
 993
 994         DEFINE_FIELD(unsigned long, func, "func", 0);
 995         DEFINE_FIELD(unsigned long, ret_ip, "ret_ip", 0);
 996         DEFINE_FIELD(int, nargs, "nargs", 1);
 997         for (i = 0; i < tp->nr_args; i++) {
 998                 /* Set argN as a field */
 999                 sprintf(buf, "arg%d", i);
1000                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1001                 /* Set argument string as an alias field */
1002                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1003                 if (ret < 0)
1004                         return ret;
1005                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1006         }
1007         return 0;
1008 }
1009
1010 static int __probe_event_show_format(struct trace_seq *s,
1011                                      struct trace_probe *tp, const char *fmt,
1012                                      const char *arg)
1013 {
1014         int i, ret;
1015         char buf[MAX_ARGSTR_LEN + 1];
1016
1017         /* Show aliases */
1018         for (i = 0; i < tp->nr_args; i++) {
1019                 ret = trace_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1020                 if (ret < 0)
1021                         return ret;
1022                 if (!trace_seq_printf(s, "\talias: %s;\toriginal: arg%d;\n",
1023                                       buf, i))
1024                         return 0;
1025         }
1026         /* Show format */
1027         if (!trace_seq_printf(s, "\nprint fmt: \"%s", fmt))
1028                 return 0;
1029
1030         for (i = 0; i < tp->nr_args; i++)
1031                 if (!trace_seq_puts(s, " 0x%lx"))
1032                         return 0;
1033
1034         if (!trace_seq_printf(s, "\", %s", arg))
1035                 return 0;
1036
1037         for (i = 0; i < tp->nr_args; i++)
1038                 if (!trace_seq_printf(s, ", arg%d", i))
1039                         return 0;
1040
1041         return trace_seq_puts(s, "\n");
1042 }
1043
1044 #undef SHOW_FIELD
1045 #define SHOW_FIELD(type, item, name)                                    \
1046         do {                                                            \
1047                 ret = trace_seq_printf(s, "\tfield: " #type " %s;\t"    \
1048                                 "offset:%u;tsize:%u;\n", name,          \
1049                                 (unsigned int)offsetof(typeof(field), item),\
1050                                 (unsigned int)sizeof(type));            \
1051                 if (!ret)                                               \
1052                         return 0;                                       \
1053         } while (0)
1054
1055 static int kprobe_event_show_format(struct ftrace_event_call *call,
1056                                     struct trace_seq *s)
1057 {
1058         struct kprobe_trace_entry field __attribute__((unused));
1059         int ret, i;
1060         char buf[8];
1061         struct trace_probe *tp = (struct trace_probe *)call->data;
1062
1063         SHOW_FIELD(unsigned long, ip, "ip");
1064         SHOW_FIELD(int, nargs, "nargs");
1065
1066         /* Show fields */
1067         for (i = 0; i < tp->nr_args; i++) {
1068                 sprintf(buf, "arg%d", i);
1069                 SHOW_FIELD(unsigned long, args[i], buf);
1070         }
1071         trace_seq_puts(s, "\n");
1072
1073         return __probe_event_show_format(s, tp, "%lx:", "ip");
1074 }
1075
1076 static int kretprobe_event_show_format(struct ftrace_event_call *call,
1077                                        struct trace_seq *s)
1078 {
1079         struct kretprobe_trace_entry field __attribute__((unused));
1080         int ret, i;
1081         char buf[8];
1082         struct trace_probe *tp = (struct trace_probe *)call->data;
1083
1084         SHOW_FIELD(unsigned long, func, "func");
1085         SHOW_FIELD(unsigned long, ret_ip, "ret_ip");
1086         SHOW_FIELD(int, nargs, "nargs");
1087
1088         /* Show fields */
1089         for (i = 0; i < tp->nr_args; i++) {
1090                 sprintf(buf, "arg%d", i);
1091                 SHOW_FIELD(unsigned long, args[i], buf);
1092         }
1093         trace_seq_puts(s, "\n");
1094
1095         return __probe_event_show_format(s, tp, "%lx <- %lx:",
1096                                           "func, ret_ip");
1097 }
1098
1099 static int register_probe_event(struct trace_probe *tp)
1100 {
1101         struct ftrace_event_call *call = &tp->call;
1102         int ret;
1103
1104         /* Initialize ftrace_event_call */
1105         call->system = "kprobes";
1106         if (probe_is_return(tp)) {
1107                 call->event = &kretprobe_trace_event;
1108                 call->id = TRACE_KRETPROBE;
1109                 call->raw_init = probe_event_raw_init;
1110                 call->show_format = kretprobe_event_show_format;
1111                 call->define_fields = kretprobe_event_define_fields;
1112         } else {
1113                 call->event = &kprobe_trace_event;
1114                 call->id = TRACE_KPROBE;
1115                 call->raw_init = probe_event_raw_init;
1116                 call->show_format = kprobe_event_show_format;
1117                 call->define_fields = kprobe_event_define_fields;
1118         }
1119         call->enabled = 1;
1120         call->regfunc = probe_event_enable;
1121         call->unregfunc = probe_event_disable;
1122         call->data = tp;
1123         ret = trace_add_event_call(call);
1124         if (ret)
1125                 pr_info("Failed to register kprobe event: %s\n", call->name);
1126         return ret;
1127 }
1128
1129 static void unregister_probe_event(struct trace_probe *tp)
1130 {
1131         /*
1132          * Prevent to unregister event itself because the event is shared
1133          * among other probes.
1134          */
1135         tp->call.event = NULL;
1136         trace_remove_event_call(&tp->call);
1137 }
1138
1139 /* Make a debugfs interface for controling probe points */
1140 static __init int init_kprobe_trace(void)
1141 {
1142         struct dentry *d_tracer;
1143         struct dentry *entry;
1144         int ret;
1145
1146         ret = register_ftrace_event(&kprobe_trace_event);
1147         if (!ret) {
1148                 pr_warning("Could not register kprobe_trace_event type.\n");
1149                 return 0;
1150         }
1151         ret = register_ftrace_event(&kretprobe_trace_event);
1152         if (!ret) {
1153                 pr_warning("Could not register kretprobe_trace_event type.\n");
1154                 return 0;
1155         }
1156
1157         d_tracer = tracing_init_dentry();
1158         if (!d_tracer)
1159                 return 0;
1160
1161         entry = debugfs_create_file("kprobe_events", 0644, d_tracer,
1162                                     NULL, &kprobe_events_ops);
1163
1164         if (!entry)
1165                 pr_warning("Could not create debugfs "
1166                            "'kprobe_events' entry\n");
1167         return 0;
1168 }
1169 fs_initcall(init_kprobe_trace);
1170
1171
1172 #ifdef CONFIG_FTRACE_STARTUP_TEST
1173
1174 static int kprobe_trace_selftest_target(int a1, int a2, int a3,
1175                                         int a4, int a5, int a6)
1176 {
1177         return a1 + a2 + a3 + a4 + a5 + a6;
1178 }
1179
1180 static __init int kprobe_trace_self_tests_init(void)
1181 {
1182         int ret;
1183         int (*target)(int, int, int, int, int, int);
1184
1185         target = kprobe_trace_selftest_target;
1186
1187         pr_info("Testing kprobe tracing: ");
1188
1189         ret = command_trace_probe("p:testprobe kprobe_trace_selftest_target "
1190                                   "a1 a2 a3 a4 a5 a6");
1191         if (WARN_ON_ONCE(ret))
1192                 pr_warning("error enabling function entry\n");
1193
1194         ret = command_trace_probe("r:testprobe2 kprobe_trace_selftest_target "
1195                                   "ra rv");
1196         if (WARN_ON_ONCE(ret))
1197                 pr_warning("error enabling function return\n");
1198
1199         ret = target(1, 2, 3, 4, 5, 6);
1200
1201         cleanup_all_probes();
1202
1203         pr_cont("OK\n");
1204         return 0;
1205 }
1206
1207 late_initcall(kprobe_trace_self_tests_init);
1208
1209 #endif