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         unsigned long           nhit;
 188         const char              *symbol;        /* symbol name */
 189         struct ftrace_event_call        call;
 190         struct trace_event              event;
 191         unsigned int            nr_args;
 192         struct fetch_func       args[];
 193 };
 194
 195 #define SIZEOF_TRACE_PROBE(n)                   \
 196         (offsetof(struct trace_probe, args) +   \
 197         (sizeof(struct fetch_func) * (n)))
 198
 199 static int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs);
 200 static int kretprobe_trace_func(struct kretprobe_instance *ri,
 201                                 struct pt_regs *regs);
 202
 203 static __kprobes int probe_is_return(struct trace_probe *tp)
 204 {
 205         return (tp->rp.handler == kretprobe_trace_func);
 206 }
 207
 208 static __kprobes const char *probe_symbol(struct trace_probe *tp)
 209 {
 210         return tp->symbol ? tp->symbol : "unknown";
 211 }
 212
 213 static __kprobes unsigned int probe_offset(struct trace_probe *tp)
 214 {
 215         return (probe_is_return(tp)) ? tp->rp.kp.offset : tp->kp.offset;
 216 }
 217
 218 static __kprobes void *probe_address(struct trace_probe *tp)
 219 {
 220         return (probe_is_return(tp)) ? tp->rp.kp.addr : tp->kp.addr;
 221 }
 222
 223 static int probe_arg_string(char *buf, size_t n, struct fetch_func *ff)
 224 {
 225         int ret = -EINVAL;
 226
 227         if (ff->func == fetch_argument)
 228                 ret = snprintf(buf, n, "a%lu", (unsigned long)ff->data);
 229         else if (ff->func == fetch_register) {
 230                 const char *name;
 231                 name = regs_query_register_name((unsigned int)((long)ff->data));
 232                 ret = snprintf(buf, n, "%%%s", name);
 233         } else if (ff->func == fetch_stack)
 234                 ret = snprintf(buf, n, "s%lu", (unsigned long)ff->data);
 235         else if (ff->func == fetch_memory)
 236                 ret = snprintf(buf, n, "@0x%p", ff->data);
 237         else if (ff->func == fetch_symbol) {
 238                 struct symbol_cache *sc = ff->data;
 239                 ret = snprintf(buf, n, "@%s%+ld", sc->symbol, sc->offset);
 240         } else if (ff->func == fetch_retvalue)
 241                 ret = snprintf(buf, n, "rv");
 242         else if (ff->func == fetch_ip)
 243                 ret = snprintf(buf, n, "ra");
 244         else if (ff->func == fetch_stack_address)
 245                 ret = snprintf(buf, n, "sa");
 246         else if (ff->func == fetch_indirect) {
 247                 struct indirect_fetch_data *id = ff->data;
 248                 size_t l = 0;
 249                 ret = snprintf(buf, n, "%+ld(", id->offset);
 250                 if (ret >= n)
 251                         goto end;
 252                 l += ret;
 253                 ret = probe_arg_string(buf + l, n - l, &id->orig);
 254                 if (ret < 0)
 255                         goto end;
 256                 l += ret;
 257                 ret = snprintf(buf + l, n - l, ")");
 258                 ret += l;
 259         }
 260 end:
 261         if (ret >= n)
 262                 return -ENOSPC;
 263         return ret;
 264 }
 265
 266 static int register_probe_event(struct trace_probe *tp);
 267 static void unregister_probe_event(struct trace_probe *tp);
 268
 269 static DEFINE_MUTEX(probe_lock);
 270 static LIST_HEAD(probe_list);
 271
 272 static struct trace_probe *alloc_trace_probe(const char *symbol,
 273                                              const char *event, int nargs)
 274 {
 275         struct trace_probe *tp;
 276
 277         tp = kzalloc(SIZEOF_TRACE_PROBE(nargs), GFP_KERNEL);
 278         if (!tp)
 279                 return ERR_PTR(-ENOMEM);
 280
 281         if (symbol) {
 282                 tp->symbol = kstrdup(symbol, GFP_KERNEL);
 283                 if (!tp->symbol)
 284                         goto error;
 285         }
 286         if (!event)
 287                 goto error;
 288         tp->call.name = kstrdup(event, GFP_KERNEL);
 289         if (!tp->call.name)
 290                 goto error;
 291
 292         INIT_LIST_HEAD(&tp->list);
 293         return tp;
 294 error:
 295         kfree(tp->symbol);
 296         kfree(tp);
 297         return ERR_PTR(-ENOMEM);
 298 }
 299
 300 static void free_trace_probe(struct trace_probe *tp)
 301 {
 302         int i;
 303
 304         for (i = 0; i < tp->nr_args; i++)
 305                 if (tp->args[i].func == fetch_symbol)
 306                         free_symbol_cache(tp->args[i].data);
 307                 else if (tp->args[i].func == fetch_indirect)
 308                         free_indirect_fetch_data(tp->args[i].data);
 309
 310         kfree(tp->call.name);
 311         kfree(tp->symbol);
 312         kfree(tp);
 313 }
 314
 315 static struct trace_probe *find_probe_event(const char *event)
 316 {
 317         struct trace_probe *tp;
 318
 319         list_for_each_entry(tp, &probe_list, list)
 320                 if (!strcmp(tp->call.name, event))
 321                         return tp;
 322         return NULL;
 323 }
 324
 325 static void __unregister_trace_probe(struct trace_probe *tp)
 326 {
 327         if (probe_is_return(tp))
 328                 unregister_kretprobe(&tp->rp);
 329         else
 330                 unregister_kprobe(&tp->kp);
 331 }
 332
 333 /* Unregister a trace_probe and probe_event: call with locking probe_lock */
 334 static void unregister_trace_probe(struct trace_probe *tp)
 335 {
 336         unregister_probe_event(tp);
 337         __unregister_trace_probe(tp);
 338         list_del(&tp->list);
 339 }
 340
 341 /* Register a trace_probe and probe_event */
 342 static int register_trace_probe(struct trace_probe *tp)
 343 {
 344         struct trace_probe *old_tp;
 345         int ret;
 346
 347         mutex_lock(&probe_lock);
 348
 349         if (probe_is_return(tp))
 350                 ret = register_kretprobe(&tp->rp);
 351         else
 352                 ret = register_kprobe(&tp->kp);
 353
 354         if (ret) {
 355                 pr_warning("Could not insert probe(%d)\n", ret);
 356                 if (ret == -EILSEQ) {
 357                         pr_warning("Probing address(0x%p) is not an "
 358                                    "instruction boundary.\n",
 359                                    probe_address(tp));
 360                         ret = -EINVAL;
 361                 }
 362                 goto end;
 363         }
 364         /* register as an event */
 365         old_tp = find_probe_event(tp->call.name);
 366         if (old_tp) {
 367                 /* delete old event */
 368                 unregister_trace_probe(old_tp);
 369                 free_trace_probe(old_tp);
 370         }
 371         ret = register_probe_event(tp);
 372         if (ret) {
 373                 pr_warning("Faild to register probe event(%d)\n", ret);
 374                 __unregister_trace_probe(tp);
 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, unsigned 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                 /* skip sign because strict_strtol doesn't accept '+' */
 394                 ret = strict_strtoul(tmp + 1, 0, offset);
 395                 if (ret)
 396                         return ret;
 397                 *tmp = '\0';
 398         } else
 399                 *offset = 0;
 400         return 0;
 401 }
 402
 403 #define PARAM_MAX_ARGS 16
 404 #define PARAM_MAX_STACK (THREAD_SIZE / sizeof(unsigned long))
 405
 406 static int parse_probe_arg(char *arg, struct fetch_func *ff, int is_return)
 407 {
 408         int ret = 0;
 409         unsigned long param;
 410         long offset;
 411         char *tmp;
 412
 413         switch (arg[0]) {
 414         case 'a':       /* argument */
 415                 ret = strict_strtoul(arg + 1, 10, &param);
 416                 if (ret || param > PARAM_MAX_ARGS)
 417                         ret = -EINVAL;
 418                 else {
 419                         ff->func = fetch_argument;
 420                         ff->data = (void *)param;
 421                 }
 422                 break;
 423         case 'r':       /* retval or retaddr */
 424                 if (is_return && arg[1] == 'v') {
 425                         ff->func = fetch_retvalue;
 426                         ff->data = NULL;
 427                 } else if (is_return && arg[1] == 'a') {
 428                         ff->func = fetch_ip;
 429                         ff->data = NULL;
 430                 } else
 431                         ret = -EINVAL;
 432                 break;
 433         case '%':       /* named register */
 434                 ret = regs_query_register_offset(arg + 1);
 435                 if (ret >= 0) {
 436                         ff->func = fetch_register;
 437                         ff->data = (void *)(unsigned long)ret;
 438                         ret = 0;
 439                 }
 440                 break;
 441         case 's':       /* stack */
 442                 if (arg[1] == 'a') {
 443                         ff->func = fetch_stack_address;
 444                         ff->data = NULL;
 445                 } else {
 446                         ret = strict_strtoul(arg + 1, 10, &param);
 447                         if (ret || param > PARAM_MAX_STACK)
 448                                 ret = -EINVAL;
 449                         else {
 450                                 ff->func = fetch_stack;
 451                                 ff->data = (void *)param;
 452                         }
 453                 }
 454                 break;
 455         case '@':       /* memory or symbol */
 456                 if (isdigit(arg[1])) {
 457                         ret = strict_strtoul(arg + 1, 0, &param);
 458                         if (ret)
 459                                 break;
 460                         ff->func = fetch_memory;
 461                         ff->data = (void *)param;
 462                 } else {
 463                         ret = split_symbol_offset(arg + 1, &offset);
 464                         if (ret)
 465                                 break;
 466                         ff->data = alloc_symbol_cache(arg + 1,
 467                                                               offset);
 468                         if (ff->data)
 469                                 ff->func = fetch_symbol;
 470                         else
 471                                 ret = -EINVAL;
 472                 }
 473                 break;
 474         case '+':       /* indirect memory */
 475         case '-':
 476                 tmp = strchr(arg, '(');
 477                 if (!tmp) {
 478                         ret = -EINVAL;
 479                         break;
 480                 }
 481                 *tmp = '\0';
 482                 ret = strict_strtol(arg + 1, 0, &offset);
 483                 if (ret)
 484                         break;
 485                 if (arg[0] == '-')
 486                         offset = -offset;
 487                 arg = tmp + 1;
 488                 tmp = strrchr(arg, ')');
 489                 if (tmp) {
 490                         struct indirect_fetch_data *id;
 491                         *tmp = '\0';
 492                         id = kzalloc(sizeof(struct indirect_fetch_data),
 493                                      GFP_KERNEL);
 494                         if (!id)
 495                                 return -ENOMEM;
 496                         id->offset = offset;
 497                         ret = parse_probe_arg(arg, &id->orig, is_return);
 498                         if (ret)
 499                                 kfree(id);
 500                         else {
 501                                 ff->func = fetch_indirect;
 502                                 ff->data = (void *)id;
 503                         }
 504                 } else
 505                         ret = -EINVAL;
 506                 break;
 507         default:
 508                 /* TODO: support custom handler */
 509                 ret = -EINVAL;
 510         }
 511         return ret;
 512 }
 513
 514 static int create_trace_probe(int argc, char **argv)
 515 {
 516         /*
 517          * Argument syntax:
 518          *  - Add kprobe: p[:EVENT] SYMBOL[+OFFS]|ADDRESS [FETCHARGS]
 519          *  - Add kretprobe: r[:EVENT] SYMBOL[+0] [FETCHARGS]
 520          * Fetch args:
 521          *  aN  : fetch Nth of function argument. (N:0-)
 522          *  rv  : fetch return value
 523          *  ra  : fetch return address
 524          *  sa  : fetch stack address
 525          *  sN  : fetch Nth of stack (N:0-)
 526          *  @ADDR       : fetch memory at ADDR (ADDR should be in kernel)
 527          *  @SYM[+|-offs] : fetch memory at SYM +|- offs (SYM is a data symbol)
 528          *  %REG        : fetch register REG
 529          * Indirect memory fetch:
 530          *  +|-offs(ARG) : fetch memory at ARG +|- offs address.
 531          */
 532         struct trace_probe *tp;
 533         struct kprobe *kp;
 534         int i, ret = 0;
 535         int is_return = 0;
 536         char *symbol = NULL, *event = NULL;
 537         unsigned long offset = 0;
 538         void *addr = NULL;
 539
 540         if (argc < 2)
 541                 return -EINVAL;
 542
 543         if (argv[0][0] == 'p')
 544                 is_return = 0;
 545         else if (argv[0][0] == 'r')
 546                 is_return = 1;
 547         else
 548                 return -EINVAL;
 549
 550         if (argv[0][1] == ':') {
 551                 event = &argv[0][2];
 552                 if (strlen(event) == 0) {
 553                         pr_info("Event name is not specifiled\n");
 554                         return -EINVAL;
 555                 }
 556         }
 557
 558         if (isdigit(argv[1][0])) {
 559                 if (is_return)
 560                         return -EINVAL;
 561                 /* an address specified */
 562                 ret = strict_strtoul(&argv[0][2], 0, (unsigned long *)&addr);
 563                 if (ret)
 564                         return ret;
 565         } else {
 566                 /* a symbol specified */
 567                 symbol = argv[1];
 568                 /* TODO: support .init module functions */
 569                 ret = split_symbol_offset(symbol, &offset);
 570                 if (ret)
 571                         return ret;
 572                 if (offset && is_return)
 573                         return -EINVAL;
 574         }
 575         argc -= 2; argv += 2;
 576
 577         /* setup a probe */
 578         if (!event) {
 579                 /* Make a new event name */
 580                 char buf[MAX_EVENT_NAME_LEN];
 581                 if (symbol)
 582                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c@%s%+ld",
 583                                  is_return ? 'r' : 'p', symbol, offset);
 584                 else
 585                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c@0x%p",
 586                                  is_return ? 'r' : 'p', addr);
 587                 tp = alloc_trace_probe(symbol, buf, argc);
 588         } else
 589                 tp = alloc_trace_probe(symbol, event, argc);
 590         if (IS_ERR(tp))
 591                 return PTR_ERR(tp);
 592
 593         if (is_return) {
 594                 kp = &tp->rp.kp;
 595                 tp->rp.handler = kretprobe_trace_func;
 596         } else {
 597                 kp = &tp->kp;
 598                 tp->kp.pre_handler = kprobe_trace_func;
 599         }
 600
 601         if (tp->symbol) {
 602                 kp->symbol_name = tp->symbol;
 603                 kp->offset = (unsigned int)offset;
 604         } else
 605                 kp->addr = addr;
 606
 607         /* parse arguments */
 608         ret = 0;
 609         for (i = 0; i < argc && i < MAX_TRACE_ARGS; i++) {
 610                 if (strlen(argv[i]) > MAX_ARGSTR_LEN) {
 611                         pr_info("Argument%d(%s) is too long.\n", i, argv[i]);
 612                         ret = -ENOSPC;
 613                         goto error;
 614                 }
 615                 ret = parse_probe_arg(argv[i], &tp->args[i], is_return);
 616                 if (ret)
 617                         goto error;
 618         }
 619         tp->nr_args = i;
 620
 621         ret = register_trace_probe(tp);
 622         if (ret)
 623                 goto error;
 624         return 0;
 625
 626 error:
 627         free_trace_probe(tp);
 628         return ret;
 629 }
 630
 631 static void cleanup_all_probes(void)
 632 {
 633         struct trace_probe *tp;
 634
 635         mutex_lock(&probe_lock);
 636         /* TODO: Use batch unregistration */
 637         while (!list_empty(&probe_list)) {
 638                 tp = list_entry(probe_list.next, struct trace_probe, list);
 639                 unregister_trace_probe(tp);
 640                 free_trace_probe(tp);
 641         }
 642         mutex_unlock(&probe_lock);
 643 }
 644
 645
 646 /* Probes listing interfaces */
 647 static void *probes_seq_start(struct seq_file *m, loff_t *pos)
 648 {
 649         mutex_lock(&probe_lock);
 650         return seq_list_start(&probe_list, *pos);
 651 }
 652
 653 static void *probes_seq_next(struct seq_file *m, void *v, loff_t *pos)
 654 {
 655         return seq_list_next(v, &probe_list, pos);
 656 }
 657
 658 static void probes_seq_stop(struct seq_file *m, void *v)
 659 {
 660         mutex_unlock(&probe_lock);
 661 }
 662
 663 static int probes_seq_show(struct seq_file *m, void *v)
 664 {
 665         struct trace_probe *tp = v;
 666         int i, ret;
 667         char buf[MAX_ARGSTR_LEN + 1];
 668
 669         seq_printf(m, "%c", probe_is_return(tp) ? 'r' : 'p');
 670         seq_printf(m, ":%s", tp->call.name);
 671
 672         if (tp->symbol)
 673                 seq_printf(m, " %s+%u", probe_symbol(tp), probe_offset(tp));
 674         else
 675                 seq_printf(m, " 0x%p", probe_address(tp));
 676
 677         for (i = 0; i < tp->nr_args; i++) {
 678                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 679                 if (ret < 0) {
 680                         pr_warning("Argument%d decoding error(%d).\n", i, ret);
 681                         return ret;
 682                 }
 683                 seq_printf(m, " %s", buf);
 684         }
 685         seq_printf(m, "\n");
 686         return 0;
 687 }
 688
 689 static const struct seq_operations probes_seq_op = {
 690         .start  = probes_seq_start,
 691         .next   = probes_seq_next,
 692         .stop   = probes_seq_stop,
 693         .show   = probes_seq_show
 694 };
 695
 696 static int probes_open(struct inode *inode, struct file *file)
 697 {
 698         if ((file->f_mode & FMODE_WRITE) &&
 699             (file->f_flags & O_TRUNC))
 700                 cleanup_all_probes();
 701
 702         return seq_open(file, &probes_seq_op);
 703 }
 704
 705 static int command_trace_probe(const char *buf)
 706 {
 707         char **argv;
 708         int argc = 0, ret = 0;
 709
 710         argv = argv_split(GFP_KERNEL, buf, &argc);
 711         if (!argv)
 712                 return -ENOMEM;
 713
 714         if (argc)
 715                 ret = create_trace_probe(argc, argv);
 716
 717         argv_free(argv);
 718         return ret;
 719 }
 720
 721 #define WRITE_BUFSIZE 128
 722
 723 static ssize_t probes_write(struct file *file, const char __user *buffer,
 724                             size_t count, loff_t *ppos)
 725 {
 726         char *kbuf, *tmp;
 727         int ret;
 728         size_t done;
 729         size_t size;
 730
 731         kbuf = kmalloc(WRITE_BUFSIZE, GFP_KERNEL);
 732         if (!kbuf)
 733                 return -ENOMEM;
 734
 735         ret = done = 0;
 736         while (done < count) {
 737                 size = count - done;
 738                 if (size >= WRITE_BUFSIZE)
 739                         size = WRITE_BUFSIZE - 1;
 740                 if (copy_from_user(kbuf, buffer + done, size)) {
 741                         ret = -EFAULT;
 742                         goto out;
 743                 }
 744                 kbuf[size] = '\0';
 745                 tmp = strchr(kbuf, '\n');
 746                 if (tmp) {
 747                         *tmp = '\0';
 748                         size = tmp - kbuf + 1;
 749                 } else if (done + size < count) {
 750                         pr_warning("Line length is too long: "
 751                                    "Should be less than %d.", WRITE_BUFSIZE);
 752                         ret = -EINVAL;
 753                         goto out;
 754                 }
 755                 done += size;
 756                 /* Remove comments */
 757                 tmp = strchr(kbuf, '#');
 758                 if (tmp)
 759                         *tmp = '\0';
 760
 761                 ret = command_trace_probe(kbuf);
 762                 if (ret)
 763                         goto out;
 764         }
 765         ret = done;
 766 out:
 767         kfree(kbuf);
 768         return ret;
 769 }
 770
 771 static const struct file_operations kprobe_events_ops = {
 772         .owner          = THIS_MODULE,
 773         .open           = probes_open,
 774         .read           = seq_read,
 775         .llseek         = seq_lseek,
 776         .release        = seq_release,
 777         .write          = probes_write,
 778 };
 779
 780 /* Probes profiling interfaces */
 781 static int probes_profile_seq_show(struct seq_file *m, void *v)
 782 {
 783         struct trace_probe *tp = v;
 784
 785         seq_printf(m, "  %-44s %15lu %15lu\n", tp->call.name, tp->nhit,
 786                    probe_is_return(tp) ? tp->rp.kp.nmissed : tp->kp.nmissed);
 787
 788         return 0;
 789 }
 790
 791 static const struct seq_operations profile_seq_op = {
 792         .start  = probes_seq_start,
 793         .next   = probes_seq_next,
 794         .stop   = probes_seq_stop,
 795         .show   = probes_profile_seq_show
 796 };
 797
 798 static int profile_open(struct inode *inode, struct file *file)
 799 {
 800         return seq_open(file, &profile_seq_op);
 801 }
 802
 803 static const struct file_operations kprobe_profile_ops = {
 804         .owner          = THIS_MODULE,
 805         .open           = profile_open,
 806         .read           = seq_read,
 807         .llseek         = seq_lseek,
 808         .release        = seq_release,
 809 };
 810
 811 /* Kprobe handler */
 812 static __kprobes int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs)
 813 {
 814         struct trace_probe *tp = container_of(kp, struct trace_probe, kp);
 815         struct kprobe_trace_entry *entry;
 816         struct ring_buffer_event *event;
 817         struct ring_buffer *buffer;
 818         int size, i, pc;
 819         unsigned long irq_flags;
 820         struct ftrace_event_call *call = &tp->call;
 821
 822         tp->nhit++;
 823
 824         local_save_flags(irq_flags);
 825         pc = preempt_count();
 826
 827         size = SIZEOF_KPROBE_TRACE_ENTRY(tp->nr_args);
 828
 829         event = trace_current_buffer_lock_reserve(&buffer, call->id, size,
 830                                                   irq_flags, pc);
 831         if (!event)
 832                 return 0;
 833
 834         entry = ring_buffer_event_data(event);
 835         entry->nargs = tp->nr_args;
 836         entry->ip = (unsigned long)kp->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(buffer, call, entry, event))
 841                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
 842         return 0;
 843 }
 844
 845 /* Kretprobe handler */
 846 static __kprobes int kretprobe_trace_func(struct kretprobe_instance *ri,
 847                                           struct pt_regs *regs)
 848 {
 849         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
 850         struct kretprobe_trace_entry *entry;
 851         struct ring_buffer_event *event;
 852         struct ring_buffer *buffer;
 853         int size, i, pc;
 854         unsigned long irq_flags;
 855         struct ftrace_event_call *call = &tp->call;
 856
 857         local_save_flags(irq_flags);
 858         pc = preempt_count();
 859
 860         size = SIZEOF_KRETPROBE_TRACE_ENTRY(tp->nr_args);
 861
 862         event = trace_current_buffer_lock_reserve(&buffer, call->id, size,
 863                                                   irq_flags, pc);
 864         if (!event)
 865                 return 0;
 866
 867         entry = ring_buffer_event_data(event);
 868         entry->nargs = tp->nr_args;
 869         entry->func = (unsigned long)probe_address(tp);
 870         entry->ret_ip = (unsigned long)ri->ret_addr;
 871         for (i = 0; i < tp->nr_args; i++)
 872                 entry->args[i] = call_fetch(&tp->args[i], regs);
 873
 874         if (!filter_current_check_discard(buffer, call, entry, event))
 875                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
 876
 877         return 0;
 878 }
 879
 880 /* Event entry printers */
 881 enum print_line_t
 882 print_kprobe_event(struct trace_iterator *iter, int flags)
 883 {
 884         struct kprobe_trace_entry *field;
 885         struct trace_seq *s = &iter->seq;
 886         int i;
 887
 888         field = (struct kprobe_trace_entry *)iter->ent;
 889
 890         if (!seq_print_ip_sym(s, field->ip, flags | TRACE_ITER_SYM_OFFSET))
 891                 goto partial;
 892
 893         if (!trace_seq_puts(s, ":"))
 894                 goto partial;
 895
 896         for (i = 0; i < field->nargs; i++)
 897                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 898                         goto partial;
 899
 900         if (!trace_seq_puts(s, "\n"))
 901                 goto partial;
 902
 903         return TRACE_TYPE_HANDLED;
 904 partial:
 905         return TRACE_TYPE_PARTIAL_LINE;
 906 }
 907
 908 enum print_line_t
 909 print_kretprobe_event(struct trace_iterator *iter, int flags)
 910 {
 911         struct kretprobe_trace_entry *field;
 912         struct trace_seq *s = &iter->seq;
 913         int i;
 914
 915         field = (struct kretprobe_trace_entry *)iter->ent;
 916
 917         if (!seq_print_ip_sym(s, field->ret_ip, flags | TRACE_ITER_SYM_OFFSET))
 918                 goto partial;
 919
 920         if (!trace_seq_puts(s, " <- "))
 921                 goto partial;
 922
 923         if (!seq_print_ip_sym(s, field->func, flags & ~TRACE_ITER_SYM_OFFSET))
 924                 goto partial;
 925
 926         if (!trace_seq_puts(s, ":"))
 927                 goto partial;
 928
 929         for (i = 0; i < field->nargs; i++)
 930                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 931                         goto partial;
 932
 933         if (!trace_seq_puts(s, "\n"))
 934                 goto partial;
 935
 936         return TRACE_TYPE_HANDLED;
 937 partial:
 938         return TRACE_TYPE_PARTIAL_LINE;
 939 }
 940
 941 static int probe_event_enable(struct ftrace_event_call *call)
 942 {
 943         struct trace_probe *tp = (struct trace_probe *)call->data;
 944
 945         if (probe_is_return(tp))
 946                 return enable_kretprobe(&tp->rp);
 947         else
 948                 return enable_kprobe(&tp->kp);
 949 }
 950
 951 static void probe_event_disable(struct ftrace_event_call *call)
 952 {
 953         struct trace_probe *tp = (struct trace_probe *)call->data;
 954
 955         if (probe_is_return(tp))
 956                 disable_kretprobe(&tp->rp);
 957         else
 958                 disable_kprobe(&tp->kp);
 959 }
 960
 961 static int probe_event_raw_init(struct ftrace_event_call *event_call)
 962 {
 963         INIT_LIST_HEAD(&event_call->fields);
 964
 965         return 0;
 966 }
 967
 968 #undef DEFINE_FIELD
 969 #define DEFINE_FIELD(type, item, name, is_signed)                       \
 970         do {                                                            \
 971                 ret = trace_define_field(event_call, #type, name,       \
 972                                          offsetof(typeof(field), item), \
 973                                          sizeof(field.item), is_signed, \
 974                                          FILTER_OTHER);                 \
 975                 if (ret)                                                \
 976                         return ret;                                     \
 977         } while (0)
 978
 979 static int kprobe_event_define_fields(struct ftrace_event_call *event_call)
 980 {
 981         int ret, i;
 982         struct kprobe_trace_entry field;
 983         char buf[MAX_ARGSTR_LEN + 1];
 984         struct trace_probe *tp = (struct trace_probe *)event_call->data;
 985
 986         ret = trace_define_common_fields(event_call);
 987         if (!ret)
 988                 return ret;
 989
 990         DEFINE_FIELD(unsigned long, ip, "ip", 0);
 991         DEFINE_FIELD(int, nargs, "nargs", 1);
 992         for (i = 0; i < tp->nr_args; i++) {
 993                 /* Set argN as a field */
 994                 sprintf(buf, "arg%d", i);
 995                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
 996                 /* Set argument string as an alias field */
 997                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 998                 if (ret < 0)
 999                         return ret;
1000                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1001         }
1002         return 0;
1003 }
1004
1005 static int kretprobe_event_define_fields(struct ftrace_event_call *event_call)
1006 {
1007         int ret, i;
1008         struct kretprobe_trace_entry field;
1009         char buf[MAX_ARGSTR_LEN + 1];
1010         struct trace_probe *tp = (struct trace_probe *)event_call->data;
1011
1012         ret = trace_define_common_fields(event_call);
1013         if (!ret)
1014                 return ret;
1015
1016         DEFINE_FIELD(unsigned long, func, "func", 0);
1017         DEFINE_FIELD(unsigned long, ret_ip, "ret_ip", 0);
1018         DEFINE_FIELD(int, nargs, "nargs", 1);
1019         for (i = 0; i < tp->nr_args; i++) {
1020                 /* Set argN as a field */
1021                 sprintf(buf, "arg%d", i);
1022                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1023                 /* Set argument string as an alias field */
1024                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1025                 if (ret < 0)
1026                         return ret;
1027                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1028         }
1029         return 0;
1030 }
1031
1032 static int __probe_event_show_format(struct trace_seq *s,
1033                                      struct trace_probe *tp, const char *fmt,
1034                                      const char *arg)
1035 {
1036         int i, ret;
1037         char buf[MAX_ARGSTR_LEN + 1];
1038
1039         /* Show aliases */
1040         for (i = 0; i < tp->nr_args; i++) {
1041                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1042                 if (ret < 0)
1043                         return ret;
1044                 if (!trace_seq_printf(s, "\talias: %s;\toriginal: arg%d;\n",
1045                                       buf, i))
1046                         return 0;
1047         }
1048         /* Show format */
1049         if (!trace_seq_printf(s, "\nprint fmt: \"%s", fmt))
1050                 return 0;
1051
1052         for (i = 0; i < tp->nr_args; i++)
1053                 if (!trace_seq_puts(s, " 0x%lx"))
1054                         return 0;
1055
1056         if (!trace_seq_printf(s, "\", %s", arg))
1057                 return 0;
1058
1059         for (i = 0; i < tp->nr_args; i++)
1060                 if (!trace_seq_printf(s, ", arg%d", i))
1061                         return 0;
1062
1063         return trace_seq_puts(s, "\n");
1064 }
1065
1066 #undef SHOW_FIELD
1067 #define SHOW_FIELD(type, item, name)                                    \
1068         do {                                                            \
1069                 ret = trace_seq_printf(s, "\tfield: " #type " %s;\t"    \
1070                                 "offset:%u;\tsize:%u;\n", name,         \
1071                                 (unsigned int)offsetof(typeof(field), item),\
1072                                 (unsigned int)sizeof(type));            \
1073                 if (!ret)                                               \
1074                         return 0;                                       \
1075         } while (0)
1076
1077 static int kprobe_event_show_format(struct ftrace_event_call *call,
1078                                     struct trace_seq *s)
1079 {
1080         struct kprobe_trace_entry field __attribute__((unused));
1081         int ret, i;
1082         char buf[8];
1083         struct trace_probe *tp = (struct trace_probe *)call->data;
1084
1085         SHOW_FIELD(unsigned long, ip, "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:", "ip");
1096 }
1097
1098 static int kretprobe_event_show_format(struct ftrace_event_call *call,
1099                                        struct trace_seq *s)
1100 {
1101         struct kretprobe_trace_entry field __attribute__((unused));
1102         int ret, i;
1103         char buf[8];
1104         struct trace_probe *tp = (struct trace_probe *)call->data;
1105
1106         SHOW_FIELD(unsigned long, func, "func");
1107         SHOW_FIELD(unsigned long, ret_ip, "ret_ip");
1108         SHOW_FIELD(int, nargs, "nargs");
1109
1110         /* Show fields */
1111         for (i = 0; i < tp->nr_args; i++) {
1112                 sprintf(buf, "arg%d", i);
1113                 SHOW_FIELD(unsigned long, args[i], buf);
1114         }
1115         trace_seq_puts(s, "\n");
1116
1117         return __probe_event_show_format(s, tp, "%lx <- %lx:",
1118                                           "func, ret_ip");
1119 }
1120
1121 static int register_probe_event(struct trace_probe *tp)
1122 {
1123         struct ftrace_event_call *call = &tp->call;
1124         int ret;
1125
1126         /* Initialize ftrace_event_call */
1127         call->system = "kprobes";
1128         if (probe_is_return(tp)) {
1129                 tp->event.trace = print_kretprobe_event;
1130                 call->raw_init = probe_event_raw_init;
1131                 call->show_format = kretprobe_event_show_format;
1132                 call->define_fields = kretprobe_event_define_fields;
1133         } else {
1134                 tp->event.trace = print_kprobe_event;
1135                 call->raw_init = probe_event_raw_init;
1136                 call->show_format = kprobe_event_show_format;
1137                 call->define_fields = kprobe_event_define_fields;
1138         }
1139         call->event = &tp->event;
1140         call->id = register_ftrace_event(&tp->event);
1141         if (!call->id)
1142                 return -ENODEV;
1143         call->enabled = 1;
1144         call->regfunc = probe_event_enable;
1145         call->unregfunc = probe_event_disable;
1146         call->data = tp;
1147         ret = trace_add_event_call(call);
1148         if (ret) {
1149                 pr_info("Failed to register kprobe event: %s\n", call->name);
1150                 unregister_ftrace_event(&tp->event);
1151         }
1152         return ret;
1153 }
1154
1155 static void unregister_probe_event(struct trace_probe *tp)
1156 {
1157         /* tp->event is unregistered in trace_remove_event_call() */
1158         trace_remove_event_call(&tp->call);
1159 }
1160
1161 /* Make a debugfs interface for controling probe points */
1162 static __init int init_kprobe_trace(void)
1163 {
1164         struct dentry *d_tracer;
1165         struct dentry *entry;
1166
1167         d_tracer = tracing_init_dentry();
1168         if (!d_tracer)
1169                 return 0;
1170
1171         entry = debugfs_create_file("kprobe_events", 0644, d_tracer,
1172                                     NULL, &kprobe_events_ops);
1173
1174         /* Event list interface */
1175         if (!entry)
1176                 pr_warning("Could not create debugfs "
1177                            "'kprobe_events' entry\n");
1178
1179         /* Profile interface */
1180         entry = debugfs_create_file("kprobe_profile", 0444, d_tracer,
1181                                     NULL, &kprobe_profile_ops);
1182
1183         if (!entry)
1184                 pr_warning("Could not create debugfs "
1185                            "'kprobe_profile' entry\n");
1186         return 0;
1187 }
1188 fs_initcall(init_kprobe_trace);
1189
1190
1191 #ifdef CONFIG_FTRACE_STARTUP_TEST
1192
1193 static int kprobe_trace_selftest_target(int a1, int a2, int a3,
1194                                         int a4, int a5, int a6)
1195 {
1196         return a1 + a2 + a3 + a4 + a5 + a6;
1197 }
1198
1199 static __init int kprobe_trace_self_tests_init(void)
1200 {
1201         int ret;
1202         int (*target)(int, int, int, int, int, int);
1203
1204         target = kprobe_trace_selftest_target;
1205
1206         pr_info("Testing kprobe tracing: ");
1207
1208         ret = command_trace_probe("p:testprobe kprobe_trace_selftest_target "
1209                                   "a1 a2 a3 a4 a5 a6");
1210         if (WARN_ON_ONCE(ret))
1211                 pr_warning("error enabling function entry\n");
1212
1213         ret = command_trace_probe("r:testprobe2 kprobe_trace_selftest_target "
1214                                   "ra rv");
1215         if (WARN_ON_ONCE(ret))
1216                 pr_warning("error enabling function return\n");
1217
1218         ret = target(1, 2, 3, 4, 5, 6);
1219
1220         cleanup_all_probes();
1221
1222         pr_cont("OK\n");
1223         return 0;
1224 }
1225
1226 late_initcall(kprobe_trace_self_tests_init);
1227
1228 #endif