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