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 long 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, long *offset)
 384 {
 385         char *tmp;
 386         int ret;
 387
 388         if (!offset)
 389                 return -EINVAL;
 390
 391         tmp = strchr(symbol, '+');
 392         if (!tmp)
 393                 tmp = strchr(symbol, '-');
 394
 395         if (tmp) {
 396                 /* skip sign because strict_strtol doesn't accept '+' */
 397                 ret = strict_strtol(tmp + 1, 0, offset);
 398                 if (ret)
 399                         return ret;
 400                 if (*tmp == '-')
 401                         *offset = -(*offset);
 402                 *tmp = '\0';
 403         } else
 404                 *offset = 0;
 405         return 0;
 406 }
 407
 408 #define PARAM_MAX_ARGS 16
 409 #define PARAM_MAX_STACK (THREAD_SIZE / sizeof(unsigned long))
 410
 411 static int parse_probe_arg(char *arg, struct fetch_func *ff, int is_return)
 412 {
 413         int ret = 0;
 414         unsigned long param;
 415         long offset;
 416         char *tmp;
 417
 418         switch (arg[0]) {
 419         case 'a':       /* argument */
 420                 ret = strict_strtoul(arg + 1, 10, &param);
 421                 if (ret || param > PARAM_MAX_ARGS)
 422                         ret = -EINVAL;
 423                 else {
 424                         ff->func = fetch_argument;
 425                         ff->data = (void *)param;
 426                 }
 427                 break;
 428         case 'r':       /* retval or retaddr */
 429                 if (is_return && arg[1] == 'v') {
 430                         ff->func = fetch_retvalue;
 431                         ff->data = NULL;
 432                 } else if (is_return && arg[1] == 'a') {
 433                         ff->func = fetch_ip;
 434                         ff->data = NULL;
 435                 } else
 436                         ret = -EINVAL;
 437                 break;
 438         case '%':       /* named register */
 439                 ret = regs_query_register_offset(arg + 1);
 440                 if (ret >= 0) {
 441                         ff->func = fetch_register;
 442                         ff->data = (void *)(unsigned long)ret;
 443                         ret = 0;
 444                 }
 445                 break;
 446         case 's':       /* stack */
 447                 if (arg[1] == 'a') {
 448                         ff->func = fetch_stack_address;
 449                         ff->data = NULL;
 450                 } else {
 451                         ret = strict_strtoul(arg + 1, 10, &param);
 452                         if (ret || param > PARAM_MAX_STACK)
 453                                 ret = -EINVAL;
 454                         else {
 455                                 ff->func = fetch_stack;
 456                                 ff->data = (void *)param;
 457                         }
 458                 }
 459                 break;
 460         case '@':       /* memory or symbol */
 461                 if (isdigit(arg[1])) {
 462                         ret = strict_strtoul(arg + 1, 0, &param);
 463                         if (ret)
 464                                 break;
 465                         ff->func = fetch_memory;
 466                         ff->data = (void *)param;
 467                 } else {
 468                         ret = split_symbol_offset(arg + 1, &offset);
 469                         if (ret)
 470                                 break;
 471                         ff->data = alloc_symbol_cache(arg + 1,
 472                                                               offset);
 473                         if (ff->data)
 474                                 ff->func = fetch_symbol;
 475                         else
 476                                 ret = -EINVAL;
 477                 }
 478                 break;
 479         case '+':       /* indirect memory */
 480         case '-':
 481                 tmp = strchr(arg, '(');
 482                 if (!tmp) {
 483                         ret = -EINVAL;
 484                         break;
 485                 }
 486                 *tmp = '\0';
 487                 ret = strict_strtol(arg + 1, 0, &offset);
 488                 if (ret)
 489                         break;
 490                 if (arg[0] == '-')
 491                         offset = -offset;
 492                 arg = tmp + 1;
 493                 tmp = strrchr(arg, ')');
 494                 if (tmp) {
 495                         struct indirect_fetch_data *id;
 496                         *tmp = '\0';
 497                         id = kzalloc(sizeof(struct indirect_fetch_data),
 498                                      GFP_KERNEL);
 499                         if (!id)
 500                                 return -ENOMEM;
 501                         id->offset = offset;
 502                         ret = parse_probe_arg(arg, &id->orig, is_return);
 503                         if (ret)
 504                                 kfree(id);
 505                         else {
 506                                 ff->func = fetch_indirect;
 507                                 ff->data = (void *)id;
 508                         }
 509                 } else
 510                         ret = -EINVAL;
 511                 break;
 512         default:
 513                 /* TODO: support custom handler */
 514                 ret = -EINVAL;
 515         }
 516         return ret;
 517 }
 518
 519 static int create_trace_probe(int argc, char **argv)
 520 {
 521         /*
 522          * Argument syntax:
 523          *  - Add kprobe: p[:EVENT] SYMBOL[+OFFS|-OFFS]|ADDRESS [FETCHARGS]
 524          *  - Add kretprobe: r[:EVENT] SYMBOL[+0] [FETCHARGS]
 525          * Fetch args:
 526          *  aN  : fetch Nth of function argument. (N:0-)
 527          *  rv  : fetch return value
 528          *  ra  : fetch return address
 529          *  sa  : fetch stack address
 530          *  sN  : fetch Nth of stack (N:0-)
 531          *  @ADDR       : fetch memory at ADDR (ADDR should be in kernel)
 532          *  @SYM[+|-offs] : fetch memory at SYM +|- offs (SYM is a data symbol)
 533          *  %REG        : fetch register REG
 534          * Indirect memory fetch:
 535          *  +|-offs(ARG) : fetch memory at ARG +|- offs address.
 536          */
 537         struct trace_probe *tp;
 538         struct kprobe *kp;
 539         int i, ret = 0;
 540         int is_return = 0;
 541         char *symbol = NULL, *event = NULL;
 542         long offset = 0;
 543         void *addr = NULL;
 544
 545         if (argc < 2)
 546                 return -EINVAL;
 547
 548         if (argv[0][0] == 'p')
 549                 is_return = 0;
 550         else if (argv[0][0] == 'r')
 551                 is_return = 1;
 552         else
 553                 return -EINVAL;
 554
 555         if (argv[0][1] == ':') {
 556                 event = &argv[0][2];
 557                 if (strlen(event) == 0) {
 558                         pr_info("Event name is not specifiled\n");
 559                         return -EINVAL;
 560                 }
 561         }
 562
 563         if (isdigit(argv[1][0])) {
 564                 if (is_return)
 565                         return -EINVAL;
 566                 /* an address specified */
 567                 ret = strict_strtoul(&argv[0][2], 0, (unsigned long *)&addr);
 568                 if (ret)
 569                         return ret;
 570         } else {
 571                 /* a symbol specified */
 572                 symbol = argv[1];
 573                 /* TODO: support .init module functions */
 574                 ret = split_symbol_offset(symbol, &offset);
 575                 if (ret)
 576                         return ret;
 577                 if (offset && is_return)
 578                         return -EINVAL;
 579         }
 580         argc -= 2; argv += 2;
 581
 582         /* setup a probe */
 583         if (!event) {
 584                 /* Make a new event name */
 585                 char buf[MAX_EVENT_NAME_LEN];
 586                 if (symbol)
 587                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c@%s%+ld",
 588                                  is_return ? 'r' : 'p', symbol, offset);
 589                 else
 590                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c@0x%p",
 591                                  is_return ? 'r' : 'p', addr);
 592                 tp = alloc_trace_probe(symbol, buf, argc);
 593         } else
 594                 tp = alloc_trace_probe(symbol, event, argc);
 595         if (IS_ERR(tp))
 596                 return PTR_ERR(tp);
 597
 598         if (is_return) {
 599                 kp = &tp->rp.kp;
 600                 tp->rp.handler = kretprobe_trace_func;
 601         } else {
 602                 kp = &tp->kp;
 603                 tp->kp.pre_handler = kprobe_trace_func;
 604         }
 605
 606         if (tp->symbol) {
 607                 kp->symbol_name = tp->symbol;
 608                 kp->offset = offset;
 609         } else
 610                 kp->addr = addr;
 611
 612         /* parse arguments */
 613         ret = 0;
 614         for (i = 0; i < argc && i < MAX_TRACE_ARGS; i++) {
 615                 if (strlen(argv[i]) > MAX_ARGSTR_LEN) {
 616                         pr_info("Argument%d(%s) is too long.\n", i, argv[i]);
 617                         ret = -ENOSPC;
 618                         goto error;
 619                 }
 620                 ret = parse_probe_arg(argv[i], &tp->args[i], is_return);
 621                 if (ret)
 622                         goto error;
 623         }
 624         tp->nr_args = i;
 625
 626         ret = register_trace_probe(tp);
 627         if (ret)
 628                 goto error;
 629         return 0;
 630
 631 error:
 632         free_trace_probe(tp);
 633         return ret;
 634 }
 635
 636 static void cleanup_all_probes(void)
 637 {
 638         struct trace_probe *tp;
 639
 640         mutex_lock(&probe_lock);
 641         /* TODO: Use batch unregistration */
 642         while (!list_empty(&probe_list)) {
 643                 tp = list_entry(probe_list.next, struct trace_probe, list);
 644                 unregister_trace_probe(tp);
 645                 free_trace_probe(tp);
 646         }
 647         mutex_unlock(&probe_lock);
 648 }
 649
 650
 651 /* Probes listing interfaces */
 652 static void *probes_seq_start(struct seq_file *m, loff_t *pos)
 653 {
 654         mutex_lock(&probe_lock);
 655         return seq_list_start(&probe_list, *pos);
 656 }
 657
 658 static void *probes_seq_next(struct seq_file *m, void *v, loff_t *pos)
 659 {
 660         return seq_list_next(v, &probe_list, pos);
 661 }
 662
 663 static void probes_seq_stop(struct seq_file *m, void *v)
 664 {
 665         mutex_unlock(&probe_lock);
 666 }
 667
 668 static int probes_seq_show(struct seq_file *m, void *v)
 669 {
 670         struct trace_probe *tp = v;
 671         int i, ret;
 672         char buf[MAX_ARGSTR_LEN + 1];
 673
 674         seq_printf(m, "%c", probe_is_return(tp) ? 'r' : 'p');
 675         seq_printf(m, ":%s", tp->call.name);
 676
 677         if (tp->symbol)
 678                 seq_printf(m, " %s%+ld", probe_symbol(tp), probe_offset(tp));
 679         else
 680                 seq_printf(m, " 0x%p", probe_address(tp));
 681
 682         for (i = 0; i < tp->nr_args; i++) {
 683                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
 684                 if (ret < 0) {
 685                         pr_warning("Argument%d decoding error(%d).\n", i, ret);
 686                         return ret;
 687                 }
 688                 seq_printf(m, " %s", buf);
 689         }
 690         seq_printf(m, "\n");
 691         return 0;
 692 }
 693
 694 static const struct seq_operations probes_seq_op = {
 695         .start  = probes_seq_start,
 696         .next   = probes_seq_next,
 697         .stop   = probes_seq_stop,
 698         .show   = probes_seq_show
 699 };
 700
 701 static int probes_open(struct inode *inode, struct file *file)
 702 {
 703         if ((file->f_mode & FMODE_WRITE) &&
 704             (file->f_flags & O_TRUNC))
 705                 cleanup_all_probes();
 706
 707         return seq_open(file, &probes_seq_op);
 708 }
 709
 710 static int command_trace_probe(const char *buf)
 711 {
 712         char **argv;
 713         int argc = 0, ret = 0;
 714
 715         argv = argv_split(GFP_KERNEL, buf, &argc);
 716         if (!argv)
 717                 return -ENOMEM;
 718
 719         if (argc)
 720                 ret = create_trace_probe(argc, argv);
 721
 722         argv_free(argv);
 723         return ret;
 724 }
 725
 726 #define WRITE_BUFSIZE 128
 727
 728 static ssize_t probes_write(struct file *file, const char __user *buffer,
 729                             size_t count, loff_t *ppos)
 730 {
 731         char *kbuf, *tmp;
 732         int ret;
 733         size_t done;
 734         size_t size;
 735
 736         kbuf = kmalloc(WRITE_BUFSIZE, GFP_KERNEL);
 737         if (!kbuf)
 738                 return -ENOMEM;
 739
 740         ret = done = 0;
 741         while (done < count) {
 742                 size = count - done;
 743                 if (size >= WRITE_BUFSIZE)
 744                         size = WRITE_BUFSIZE - 1;
 745                 if (copy_from_user(kbuf, buffer + done, size)) {
 746                         ret = -EFAULT;
 747                         goto out;
 748                 }
 749                 kbuf[size] = '\0';
 750                 tmp = strchr(kbuf, '\n');
 751                 if (tmp) {
 752                         *tmp = '\0';
 753                         size = tmp - kbuf + 1;
 754                 } else if (done + size < count) {
 755                         pr_warning("Line length is too long: "
 756                                    "Should be less than %d.", WRITE_BUFSIZE);
 757                         ret = -EINVAL;
 758                         goto out;
 759                 }
 760                 done += size;
 761                 /* Remove comments */
 762                 tmp = strchr(kbuf, '#');
 763                 if (tmp)
 764                         *tmp = '\0';
 765
 766                 ret = command_trace_probe(kbuf);
 767                 if (ret)
 768                         goto out;
 769         }
 770         ret = done;
 771 out:
 772         kfree(kbuf);
 773         return ret;
 774 }
 775
 776 static const struct file_operations kprobe_events_ops = {
 777         .owner          = THIS_MODULE,
 778         .open           = probes_open,
 779         .read           = seq_read,
 780         .llseek         = seq_lseek,
 781         .release        = seq_release,
 782         .write          = probes_write,
 783 };
 784
 785 /* Probes profiling interfaces */
 786 static int probes_profile_seq_show(struct seq_file *m, void *v)
 787 {
 788         struct trace_probe *tp = v;
 789
 790         seq_printf(m, "  %-44s %15lu %15lu\n", tp->call.name, tp->nhit,
 791                    probe_is_return(tp) ? tp->rp.kp.nmissed : tp->kp.nmissed);
 792
 793         return 0;
 794 }
 795
 796 static const struct seq_operations profile_seq_op = {
 797         .start  = probes_seq_start,
 798         .next   = probes_seq_next,
 799         .stop   = probes_seq_stop,
 800         .show   = probes_profile_seq_show
 801 };
 802
 803 static int profile_open(struct inode *inode, struct file *file)
 804 {
 805         return seq_open(file, &profile_seq_op);
 806 }
 807
 808 static const struct file_operations kprobe_profile_ops = {
 809         .owner          = THIS_MODULE,
 810         .open           = profile_open,
 811         .read           = seq_read,
 812         .llseek         = seq_lseek,
 813         .release        = seq_release,
 814 };
 815
 816 /* Kprobe handler */
 817 static __kprobes int kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs)
 818 {
 819         struct trace_probe *tp = container_of(kp, struct trace_probe, kp);
 820         struct kprobe_trace_entry *entry;
 821         struct ring_buffer_event *event;
 822         struct ring_buffer *buffer;
 823         int size, i, pc;
 824         unsigned long irq_flags;
 825         struct ftrace_event_call *call = &tp->call;
 826
 827         tp->nhit++;
 828
 829         local_save_flags(irq_flags);
 830         pc = preempt_count();
 831
 832         size = SIZEOF_KPROBE_TRACE_ENTRY(tp->nr_args);
 833
 834         event = trace_current_buffer_lock_reserve(&buffer, call->id, size,
 835                                                   irq_flags, pc);
 836         if (!event)
 837                 return 0;
 838
 839         entry = ring_buffer_event_data(event);
 840         entry->nargs = tp->nr_args;
 841         entry->ip = (unsigned long)kp->addr;
 842         for (i = 0; i < tp->nr_args; i++)
 843                 entry->args[i] = call_fetch(&tp->args[i], regs);
 844
 845         if (!filter_current_check_discard(buffer, call, entry, event))
 846                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
 847         return 0;
 848 }
 849
 850 /* Kretprobe handler */
 851 static __kprobes int kretprobe_trace_func(struct kretprobe_instance *ri,
 852                                           struct pt_regs *regs)
 853 {
 854         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
 855         struct kretprobe_trace_entry *entry;
 856         struct ring_buffer_event *event;
 857         struct ring_buffer *buffer;
 858         int size, i, pc;
 859         unsigned long irq_flags;
 860         struct ftrace_event_call *call = &tp->call;
 861
 862         local_save_flags(irq_flags);
 863         pc = preempt_count();
 864
 865         size = SIZEOF_KRETPROBE_TRACE_ENTRY(tp->nr_args);
 866
 867         event = trace_current_buffer_lock_reserve(&buffer, call->id, size,
 868                                                   irq_flags, pc);
 869         if (!event)
 870                 return 0;
 871
 872         entry = ring_buffer_event_data(event);
 873         entry->nargs = tp->nr_args;
 874         entry->func = (unsigned long)probe_address(tp);
 875         entry->ret_ip = (unsigned long)ri->ret_addr;
 876         for (i = 0; i < tp->nr_args; i++)
 877                 entry->args[i] = call_fetch(&tp->args[i], regs);
 878
 879         if (!filter_current_check_discard(buffer, call, entry, event))
 880                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
 881
 882         return 0;
 883 }
 884
 885 /* Event entry printers */
 886 enum print_line_t
 887 print_kprobe_event(struct trace_iterator *iter, int flags)
 888 {
 889         struct kprobe_trace_entry *field;
 890         struct trace_seq *s = &iter->seq;
 891         int i;
 892
 893         field = (struct kprobe_trace_entry *)iter->ent;
 894
 895         if (!seq_print_ip_sym(s, field->ip, flags | TRACE_ITER_SYM_OFFSET))
 896                 goto partial;
 897
 898         if (!trace_seq_puts(s, ":"))
 899                 goto partial;
 900
 901         for (i = 0; i < field->nargs; i++)
 902                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 903                         goto partial;
 904
 905         if (!trace_seq_puts(s, "\n"))
 906                 goto partial;
 907
 908         return TRACE_TYPE_HANDLED;
 909 partial:
 910         return TRACE_TYPE_PARTIAL_LINE;
 911 }
 912
 913 enum print_line_t
 914 print_kretprobe_event(struct trace_iterator *iter, int flags)
 915 {
 916         struct kretprobe_trace_entry *field;
 917         struct trace_seq *s = &iter->seq;
 918         int i;
 919
 920         field = (struct kretprobe_trace_entry *)iter->ent;
 921
 922         if (!seq_print_ip_sym(s, field->ret_ip, flags | TRACE_ITER_SYM_OFFSET))
 923                 goto partial;
 924
 925         if (!trace_seq_puts(s, " <- "))
 926                 goto partial;
 927
 928         if (!seq_print_ip_sym(s, field->func, flags & ~TRACE_ITER_SYM_OFFSET))
 929                 goto partial;
 930
 931         if (!trace_seq_puts(s, ":"))
 932                 goto partial;
 933
 934         for (i = 0; i < field->nargs; i++)
 935                 if (!trace_seq_printf(s, " 0x%lx", field->args[i]))
 936                         goto partial;
 937
 938         if (!trace_seq_puts(s, "\n"))
 939                 goto partial;
 940
 941         return TRACE_TYPE_HANDLED;
 942 partial:
 943         return TRACE_TYPE_PARTIAL_LINE;
 944 }
 945
 946 static int probe_event_enable(struct ftrace_event_call *call)
 947 {
 948         struct trace_probe *tp = (struct trace_probe *)call->data;
 949
 950         if (probe_is_return(tp))
 951                 return enable_kretprobe(&tp->rp);
 952         else
 953                 return enable_kprobe(&tp->kp);
 954 }
 955
 956 static void probe_event_disable(struct ftrace_event_call *call)
 957 {
 958         struct trace_probe *tp = (struct trace_probe *)call->data;
 959
 960         if (probe_is_return(tp))
 961                 disable_kretprobe(&tp->rp);
 962         else
 963                 disable_kprobe(&tp->kp);
 964 }
 965
 966 static int probe_event_raw_init(struct ftrace_event_call *event_call)
 967 {
 968         INIT_LIST_HEAD(&event_call->fields);
 969
 970         return 0;
 971 }
 972
 973 #undef DEFINE_FIELD
 974 #define DEFINE_FIELD(type, item, name, is_signed)                       \
 975         do {                                                            \
 976                 ret = trace_define_field(event_call, #type, name,       \
 977                                          offsetof(typeof(field), item), \
 978                                          sizeof(field.item), is_signed, \
 979                                          FILTER_OTHER);                 \
 980                 if (ret)                                                \
 981                         return ret;                                     \
 982         } while (0)
 983
 984 static int kprobe_event_define_fields(struct ftrace_event_call *event_call)
 985 {
 986         int ret, i;
 987         struct kprobe_trace_entry field;
 988         char buf[MAX_ARGSTR_LEN + 1];
 989         struct trace_probe *tp = (struct trace_probe *)event_call->data;
 990
 991         ret = trace_define_common_fields(event_call);
 992         if (!ret)
 993                 return ret;
 994
 995         DEFINE_FIELD(unsigned long, ip, "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 = probe_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 kretprobe_event_define_fields(struct ftrace_event_call *event_call)
1011 {
1012         int ret, i;
1013         struct kretprobe_trace_entry field;
1014         char buf[MAX_ARGSTR_LEN + 1];
1015         struct trace_probe *tp = (struct trace_probe *)event_call->data;
1016
1017         ret = trace_define_common_fields(event_call);
1018         if (!ret)
1019                 return ret;
1020
1021         DEFINE_FIELD(unsigned long, func, "func", 0);
1022         DEFINE_FIELD(unsigned long, ret_ip, "ret_ip", 0);
1023         DEFINE_FIELD(int, nargs, "nargs", 1);
1024         for (i = 0; i < tp->nr_args; i++) {
1025                 /* Set argN as a field */
1026                 sprintf(buf, "arg%d", i);
1027                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1028                 /* Set argument string as an alias field */
1029                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1030                 if (ret < 0)
1031                         return ret;
1032                 DEFINE_FIELD(unsigned long, args[i], buf, 0);
1033         }
1034         return 0;
1035 }
1036
1037 static int __probe_event_show_format(struct trace_seq *s,
1038                                      struct trace_probe *tp, const char *fmt,
1039                                      const char *arg)
1040 {
1041         int i, ret;
1042         char buf[MAX_ARGSTR_LEN + 1];
1043
1044         /* Show aliases */
1045         for (i = 0; i < tp->nr_args; i++) {
1046                 ret = probe_arg_string(buf, MAX_ARGSTR_LEN, &tp->args[i]);
1047                 if (ret < 0)
1048                         return ret;
1049                 if (!trace_seq_printf(s, "\talias: %s;\toriginal: arg%d;\n",
1050                                       buf, i))
1051                         return 0;
1052         }
1053         /* Show format */
1054         if (!trace_seq_printf(s, "\nprint fmt: \"%s", fmt))
1055                 return 0;
1056
1057         for (i = 0; i < tp->nr_args; i++)
1058                 if (!trace_seq_puts(s, " 0x%lx"))
1059                         return 0;
1060
1061         if (!trace_seq_printf(s, "\", %s", arg))
1062                 return 0;
1063
1064         for (i = 0; i < tp->nr_args; i++)
1065                 if (!trace_seq_printf(s, ", arg%d", i))
1066                         return 0;
1067
1068         return trace_seq_puts(s, "\n");
1069 }
1070
1071 #undef SHOW_FIELD
1072 #define SHOW_FIELD(type, item, name)                                    \
1073         do {                                                            \
1074                 ret = trace_seq_printf(s, "\tfield: " #type " %s;\t"    \
1075                                 "offset:%u;\tsize:%u;\n", name,         \
1076                                 (unsigned int)offsetof(typeof(field), item),\
1077                                 (unsigned int)sizeof(type));            \
1078                 if (!ret)                                               \
1079                         return 0;                                       \
1080         } while (0)
1081
1082 static int kprobe_event_show_format(struct ftrace_event_call *call,
1083                                     struct trace_seq *s)
1084 {
1085         struct kprobe_trace_entry field __attribute__((unused));
1086         int ret, i;
1087         char buf[8];
1088         struct trace_probe *tp = (struct trace_probe *)call->data;
1089
1090         SHOW_FIELD(unsigned long, ip, "ip");
1091         SHOW_FIELD(int, nargs, "nargs");
1092
1093         /* Show fields */
1094         for (i = 0; i < tp->nr_args; i++) {
1095                 sprintf(buf, "arg%d", i);
1096                 SHOW_FIELD(unsigned long, args[i], buf);
1097         }
1098         trace_seq_puts(s, "\n");
1099
1100         return __probe_event_show_format(s, tp, "%lx:", "ip");
1101 }
1102
1103 static int kretprobe_event_show_format(struct ftrace_event_call *call,
1104                                        struct trace_seq *s)
1105 {
1106         struct kretprobe_trace_entry field __attribute__((unused));
1107         int ret, i;
1108         char buf[8];
1109         struct trace_probe *tp = (struct trace_probe *)call->data;
1110
1111         SHOW_FIELD(unsigned long, func, "func");
1112         SHOW_FIELD(unsigned long, ret_ip, "ret_ip");
1113         SHOW_FIELD(int, nargs, "nargs");
1114
1115         /* Show fields */
1116         for (i = 0; i < tp->nr_args; i++) {
1117                 sprintf(buf, "arg%d", i);
1118                 SHOW_FIELD(unsigned long, args[i], buf);
1119         }
1120         trace_seq_puts(s, "\n");
1121
1122         return __probe_event_show_format(s, tp, "%lx <- %lx:",
1123                                           "func, ret_ip");
1124 }
1125
1126 static int register_probe_event(struct trace_probe *tp)
1127 {
1128         struct ftrace_event_call *call = &tp->call;
1129         int ret;
1130
1131         /* Initialize ftrace_event_call */
1132         call->system = "kprobes";
1133         if (probe_is_return(tp)) {
1134                 tp->event.trace = print_kretprobe_event;
1135                 call->raw_init = probe_event_raw_init;
1136                 call->show_format = kretprobe_event_show_format;
1137                 call->define_fields = kretprobe_event_define_fields;
1138         } else {
1139                 tp->event.trace = print_kprobe_event;
1140                 call->raw_init = probe_event_raw_init;
1141                 call->show_format = kprobe_event_show_format;
1142                 call->define_fields = kprobe_event_define_fields;
1143         }
1144         call->event = &tp->event;
1145         call->id = register_ftrace_event(&tp->event);
1146         if (!call->id)
1147                 return -ENODEV;
1148         call->enabled = 1;
1149         call->regfunc = probe_event_enable;
1150         call->unregfunc = probe_event_disable;
1151         call->data = tp;
1152         ret = trace_add_event_call(call);
1153         if (ret) {
1154                 pr_info("Failed to register kprobe event: %s\n", call->name);
1155                 unregister_ftrace_event(&tp->event);
1156         }
1157         return ret;
1158 }
1159
1160 static void unregister_probe_event(struct trace_probe *tp)
1161 {
1162         /* tp->event is unregistered in trace_remove_event_call() */
1163         trace_remove_event_call(&tp->call);
1164 }
1165
1166 /* Make a debugfs interface for controling probe points */
1167 static __init int init_kprobe_trace(void)
1168 {
1169         struct dentry *d_tracer;
1170         struct dentry *entry;
1171
1172         d_tracer = tracing_init_dentry();
1173         if (!d_tracer)
1174                 return 0;
1175
1176         entry = debugfs_create_file("kprobe_events", 0644, d_tracer,
1177                                     NULL, &kprobe_events_ops);
1178
1179         /* Event list interface */
1180         if (!entry)
1181                 pr_warning("Could not create debugfs "
1182                            "'kprobe_events' entry\n");
1183
1184         /* Profile interface */
1185         entry = debugfs_create_file("kprobe_profile", 0444, d_tracer,
1186                                     NULL, &kprobe_profile_ops);
1187
1188         if (!entry)
1189                 pr_warning("Could not create debugfs "
1190                            "'kprobe_profile' entry\n");
1191         return 0;
1192 }
1193 fs_initcall(init_kprobe_trace);
1194
1195
1196 #ifdef CONFIG_FTRACE_STARTUP_TEST
1197
1198 static int kprobe_trace_selftest_target(int a1, int a2, int a3,
1199                                         int a4, int a5, int a6)
1200 {
1201         return a1 + a2 + a3 + a4 + a5 + a6;
1202 }
1203
1204 static __init int kprobe_trace_self_tests_init(void)
1205 {
1206         int ret;
1207         int (*target)(int, int, int, int, int, int);
1208
1209         target = kprobe_trace_selftest_target;
1210
1211         pr_info("Testing kprobe tracing: ");
1212
1213         ret = command_trace_probe("p:testprobe kprobe_trace_selftest_target "
1214                                   "a1 a2 a3 a4 a5 a6");
1215         if (WARN_ON_ONCE(ret))
1216                 pr_warning("error enabling function entry\n");
1217
1218         ret = command_trace_probe("r:testprobe2 kprobe_trace_selftest_target "
1219                                   "ra rv");
1220         if (WARN_ON_ONCE(ret))
1221                 pr_warning("error enabling function return\n");
1222
1223         ret = target(1, 2, 3, 4, 5, 6);
1224
1225         cleanup_all_probes();
1226
1227         pr_cont("OK\n");
1228         return 0;
1229 }
1230
1231 late_initcall(kprobe_trace_self_tests_init);
1232
1233 #endif