kernel/trace/trace_kprobe.c

   1 /*
   2  * Kprobes-based tracing events
   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 #include <linux/perf_event.h>
  32 #include <linux/stringify.h>
  33 #include <linux/limits.h>
  34 #include <linux/uaccess.h>
  35 #include <asm/bitsperlong.h>
  36
  37 #include "trace.h"
  38 #include "trace_output.h"
  39
  40 #define MAX_TRACE_ARGS 128
  41 #define MAX_ARGSTR_LEN 63
  42 #define MAX_EVENT_NAME_LEN 64
  43 #define MAX_STRING_SIZE PATH_MAX
  44 #define KPROBE_EVENT_SYSTEM "kprobes"
  45
  46 /* Reserved field names */
  47 #define FIELD_STRING_IP "__probe_ip"
  48 #define FIELD_STRING_RETIP "__probe_ret_ip"
  49 #define FIELD_STRING_FUNC "__probe_func"
  50
  51 const char *reserved_field_names[] = {
  52         "common_type",
  53         "common_flags",
  54         "common_preempt_count",
  55         "common_pid",
  56         "common_tgid",
  57         "common_lock_depth",
  58         FIELD_STRING_IP,
  59         FIELD_STRING_RETIP,
  60         FIELD_STRING_FUNC,
  61 };
  62
  63 /* Printing function type */
  64 typedef int (*print_type_func_t)(struct trace_seq *, const char *, void *,
  65                                  void *);
  66 #define PRINT_TYPE_FUNC_NAME(type)      print_type_##type
  67 #define PRINT_TYPE_FMT_NAME(type)       print_type_format_##type
  68
  69 /* Printing  in basic type function template */
  70 #define DEFINE_BASIC_PRINT_TYPE_FUNC(type, fmt, cast)                   \
  71 static __kprobes int PRINT_TYPE_FUNC_NAME(type)(struct trace_seq *s,    \
  72                                                 const char *name,       \
  73                                                 void *data, void *ent)\
  74 {                                                                       \
  75         return trace_seq_printf(s, " %s=" fmt, name, (cast)*(type *)data);\
  76 }                                                                       \
  77 static const char PRINT_TYPE_FMT_NAME(type)[] = fmt;
  78
  79 DEFINE_BASIC_PRINT_TYPE_FUNC(u8, "%x", unsigned int)
  80 DEFINE_BASIC_PRINT_TYPE_FUNC(u16, "%x", unsigned int)
  81 DEFINE_BASIC_PRINT_TYPE_FUNC(u32, "%lx", unsigned long)
  82 DEFINE_BASIC_PRINT_TYPE_FUNC(u64, "%llx", unsigned long long)
  83 DEFINE_BASIC_PRINT_TYPE_FUNC(s8, "%d", int)
  84 DEFINE_BASIC_PRINT_TYPE_FUNC(s16, "%d", int)
  85 DEFINE_BASIC_PRINT_TYPE_FUNC(s32, "%ld", long)
  86 DEFINE_BASIC_PRINT_TYPE_FUNC(s64, "%lld", long long)
  87
  88 /* data_rloc: data relative location, compatible with u32 */
  89 #define make_data_rloc(len, roffs)      \
  90         (((u32)(len) << 16) | ((u32)(roffs) & 0xffff))
  91 #define get_rloc_len(dl)        ((u32)(dl) >> 16)
  92 #define get_rloc_offs(dl)       ((u32)(dl) & 0xffff)
  93
  94 static inline void *get_rloc_data(u32 *dl)
  95 {
  96         return (u8 *)dl + get_rloc_offs(*dl);
  97 }
  98
  99 /* For data_loc conversion */
 100 static inline void *get_loc_data(u32 *dl, void *ent)
 101 {
 102         return (u8 *)ent + get_rloc_offs(*dl);
 103 }
 104
 105 /*
 106  * Convert data_rloc to data_loc:
 107  *  data_rloc stores the offset from data_rloc itself, but data_loc
 108  *  stores the offset from event entry.
 109  */
 110 #define convert_rloc_to_loc(dl, offs)   ((u32)(dl) + (offs))
 111
 112 /* For defining macros, define string/string_size types */
 113 typedef u32 string;
 114 typedef u32 string_size;
 115
 116 /* Print type function for string type */
 117 static __kprobes int PRINT_TYPE_FUNC_NAME(string)(struct trace_seq *s,
 118                                                   const char *name,
 119                                                   void *data, void *ent)
 120 {
 121         int len = *(u32 *)data >> 16;
 122
 123         if (!len)
 124                 return trace_seq_printf(s, " %s=(fault)", name);
 125         else
 126                 return trace_seq_printf(s, " %s=\"%s\"", name,
 127                                         (const char *)get_loc_data(data, ent));
 128 }
 129 static const char PRINT_TYPE_FMT_NAME(string)[] = "\\\"%s\\\"";
 130
 131 /* Data fetch function type */
 132 typedef void (*fetch_func_t)(struct pt_regs *, void *, void *);
 133
 134 struct fetch_param {
 135         fetch_func_t    fn;
 136         void *data;
 137 };
 138
 139 static __kprobes void call_fetch(struct fetch_param *fprm,
 140                                  struct pt_regs *regs, void *dest)
 141 {
 142         return fprm->fn(regs, fprm->data, dest);
 143 }
 144
 145 #define FETCH_FUNC_NAME(method, type)   fetch_##method##_##type
 146 /*
 147  * Define macro for basic types - we don't need to define s* types, because
 148  * we have to care only about bitwidth at recording time.
 149  */
 150 #define DEFINE_BASIC_FETCH_FUNCS(method) \
 151 DEFINE_FETCH_##method(u8)               \
 152 DEFINE_FETCH_##method(u16)              \
 153 DEFINE_FETCH_##method(u32)              \
 154 DEFINE_FETCH_##method(u64)
 155
 156 #define CHECK_FETCH_FUNCS(method, fn)                   \
 157         (((FETCH_FUNC_NAME(method, u8) == fn) ||        \
 158           (FETCH_FUNC_NAME(method, u16) == fn) ||       \
 159           (FETCH_FUNC_NAME(method, u32) == fn) ||       \
 160           (FETCH_FUNC_NAME(method, u64) == fn) ||       \
 161           (FETCH_FUNC_NAME(method, string) == fn) ||    \
 162           (FETCH_FUNC_NAME(method, string_size) == fn)) \
 163          && (fn != NULL))
 164
 165 /* Data fetch function templates */
 166 #define DEFINE_FETCH_reg(type)                                          \
 167 static __kprobes void FETCH_FUNC_NAME(reg, type)(struct pt_regs *regs,  \
 168                                         void *offset, void *dest)       \
 169 {                                                                       \
 170         *(type *)dest = (type)regs_get_register(regs,                   \
 171                                 (unsigned int)((unsigned long)offset)); \
 172 }
 173 DEFINE_BASIC_FETCH_FUNCS(reg)
 174 /* No string on the register */
 175 #define fetch_reg_string NULL
 176 #define fetch_reg_string_size NULL
 177
 178 #define DEFINE_FETCH_stack(type)                                        \
 179 static __kprobes void FETCH_FUNC_NAME(stack, type)(struct pt_regs *regs,\
 180                                           void *offset, void *dest)     \
 181 {                                                                       \
 182         *(type *)dest = (type)regs_get_kernel_stack_nth(regs,           \
 183                                 (unsigned int)((unsigned long)offset)); \
 184 }
 185 DEFINE_BASIC_FETCH_FUNCS(stack)
 186 /* No string on the stack entry */
 187 #define fetch_stack_string NULL
 188 #define fetch_stack_string_size NULL
 189
 190 #define DEFINE_FETCH_retval(type)                                       \
 191 static __kprobes void FETCH_FUNC_NAME(retval, type)(struct pt_regs *regs,\
 192                                           void *dummy, void *dest)      \
 193 {                                                                       \
 194         *(type *)dest = (type)regs_return_value(regs);                  \
 195 }
 196 DEFINE_BASIC_FETCH_FUNCS(retval)
 197 /* No string on the retval */
 198 #define fetch_retval_string NULL
 199 #define fetch_retval_string_size NULL
 200
 201 #define DEFINE_FETCH_memory(type)                                       \
 202 static __kprobes void FETCH_FUNC_NAME(memory, type)(struct pt_regs *regs,\
 203                                           void *addr, void *dest)       \
 204 {                                                                       \
 205         type retval;                                                    \
 206         if (probe_kernel_address(addr, retval))                         \
 207                 *(type *)dest = 0;                                      \
 208         else                                                            \
 209                 *(type *)dest = retval;                                 \
 210 }
 211 DEFINE_BASIC_FETCH_FUNCS(memory)
 212 /*
 213  * Fetch a null-terminated string. Caller MUST set *(u32 *)dest with max
 214  * length and relative data location.
 215  */
 216 static __kprobes void FETCH_FUNC_NAME(memory, string)(struct pt_regs *regs,
 217                                                       void *addr, void *dest)
 218 {
 219         long ret;
 220         int maxlen = get_rloc_len(*(u32 *)dest);
 221         u8 *dst = get_rloc_data(dest);
 222         u8 *src = addr;
 223         mm_segment_t old_fs = get_fs();
 224         if (!maxlen)
 225                 return;
 226         /*
 227          * Try to get string again, since the string can be changed while
 228          * probing.
 229          */
 230         set_fs(KERNEL_DS);
 231         pagefault_disable();
 232         do
 233                 ret = __copy_from_user_inatomic(dst++, src++, 1);
 234         while (dst[-1] && ret == 0 && src - (u8 *)addr < maxlen);
 235         dst[-1] = '\0';
 236         pagefault_enable();
 237         set_fs(old_fs);
 238
 239         if (ret < 0) {  /* Failed to fetch string */
 240                 ((u8 *)get_rloc_data(dest))[0] = '\0';
 241                 *(u32 *)dest = make_data_rloc(0, get_rloc_offs(*(u32 *)dest));
 242         } else
 243                 *(u32 *)dest = make_data_rloc(src - (u8 *)addr,
 244                                               get_rloc_offs(*(u32 *)dest));
 245 }
 246 /* Return the length of string -- including null terminal byte */
 247 static __kprobes void FETCH_FUNC_NAME(memory, string_size)(struct pt_regs *regs,
 248                                                         void *addr, void *dest)
 249 {
 250         int ret, len = 0;
 251         u8 c;
 252         mm_segment_t old_fs = get_fs();
 253
 254         set_fs(KERNEL_DS);
 255         pagefault_disable();
 256         do {
 257                 ret = __copy_from_user_inatomic(&c, (u8 *)addr + len, 1);
 258                 len++;
 259         } while (c && ret == 0 && len < MAX_STRING_SIZE);
 260         pagefault_enable();
 261         set_fs(old_fs);
 262
 263         if (ret < 0)    /* Failed to check the length */
 264                 *(u32 *)dest = 0;
 265         else
 266                 *(u32 *)dest = len;
 267 }
 268
 269 /* Memory fetching by symbol */
 270 struct symbol_cache {
 271         char *symbol;
 272         long offset;
 273         unsigned long addr;
 274 };
 275
 276 static unsigned long update_symbol_cache(struct symbol_cache *sc)
 277 {
 278         sc->addr = (unsigned long)kallsyms_lookup_name(sc->symbol);
 279         if (sc->addr)
 280                 sc->addr += sc->offset;
 281         return sc->addr;
 282 }
 283
 284 static void free_symbol_cache(struct symbol_cache *sc)
 285 {
 286         kfree(sc->symbol);
 287         kfree(sc);
 288 }
 289
 290 static struct symbol_cache *alloc_symbol_cache(const char *sym, long offset)
 291 {
 292         struct symbol_cache *sc;
 293
 294         if (!sym || strlen(sym) == 0)
 295                 return NULL;
 296         sc = kzalloc(sizeof(struct symbol_cache), GFP_KERNEL);
 297         if (!sc)
 298                 return NULL;
 299
 300         sc->symbol = kstrdup(sym, GFP_KERNEL);
 301         if (!sc->symbol) {
 302                 kfree(sc);
 303                 return NULL;
 304         }
 305         sc->offset = offset;
 306
 307         update_symbol_cache(sc);
 308         return sc;
 309 }
 310
 311 #define DEFINE_FETCH_symbol(type)                                       \
 312 static __kprobes void FETCH_FUNC_NAME(symbol, type)(struct pt_regs *regs,\
 313                                           void *data, void *dest)       \
 314 {                                                                       \
 315         struct symbol_cache *sc = data;                                 \
 316         if (sc->addr)                                                   \
 317                 fetch_memory_##type(regs, (void *)sc->addr, dest);      \
 318         else                                                            \
 319                 *(type *)dest = 0;                                      \
 320 }
 321 DEFINE_BASIC_FETCH_FUNCS(symbol)
 322 DEFINE_FETCH_symbol(string)
 323 DEFINE_FETCH_symbol(string_size)
 324
 325 /* Dereference memory access function */
 326 struct deref_fetch_param {
 327         struct fetch_param orig;
 328         long offset;
 329 };
 330
 331 #define DEFINE_FETCH_deref(type)                                        \
 332 static __kprobes void FETCH_FUNC_NAME(deref, type)(struct pt_regs *regs,\
 333                                             void *data, void *dest)     \
 334 {                                                                       \
 335         struct deref_fetch_param *dprm = data;                          \
 336         unsigned long addr;                                             \
 337         call_fetch(&dprm->orig, regs, &addr);                           \
 338         if (addr) {                                                     \
 339                 addr += dprm->offset;                                   \
 340                 fetch_memory_##type(regs, (void *)addr, dest);          \
 341         } else                                                          \
 342                 *(type *)dest = 0;                                      \
 343 }
 344 DEFINE_BASIC_FETCH_FUNCS(deref)
 345 DEFINE_FETCH_deref(string)
 346 DEFINE_FETCH_deref(string_size)
 347
 348 static __kprobes void free_deref_fetch_param(struct deref_fetch_param *data)
 349 {
 350         if (CHECK_FETCH_FUNCS(deref, data->orig.fn))
 351                 free_deref_fetch_param(data->orig.data);
 352         else if (CHECK_FETCH_FUNCS(symbol, data->orig.fn))
 353                 free_symbol_cache(data->orig.data);
 354         kfree(data);
 355 }
 356
 357 /* Default (unsigned long) fetch type */
 358 #define __DEFAULT_FETCH_TYPE(t) u##t
 359 #define _DEFAULT_FETCH_TYPE(t) __DEFAULT_FETCH_TYPE(t)
 360 #define DEFAULT_FETCH_TYPE _DEFAULT_FETCH_TYPE(BITS_PER_LONG)
 361 #define DEFAULT_FETCH_TYPE_STR __stringify(DEFAULT_FETCH_TYPE)
 362
 363 /* Fetch types */
 364 enum {
 365         FETCH_MTD_reg = 0,
 366         FETCH_MTD_stack,
 367         FETCH_MTD_retval,
 368         FETCH_MTD_memory,
 369         FETCH_MTD_symbol,
 370         FETCH_MTD_deref,
 371         FETCH_MTD_END,
 372 };
 373
 374 #define ASSIGN_FETCH_FUNC(method, type) \
 375         [FETCH_MTD_##method] = FETCH_FUNC_NAME(method, type)
 376
 377 #define __ASSIGN_FETCH_TYPE(_name, ptype, ftype, _size, sign, _fmttype) \
 378         {.name = _name,                         \
 379          .size = _size,                                 \
 380          .is_signed = sign,                             \
 381          .print = PRINT_TYPE_FUNC_NAME(ptype),          \
 382          .fmt = PRINT_TYPE_FMT_NAME(ptype),             \
 383          .fmttype = _fmttype,                           \
 384          .fetch = {                                     \
 385 ASSIGN_FETCH_FUNC(reg, ftype),                          \
 386 ASSIGN_FETCH_FUNC(stack, ftype),                        \
 387 ASSIGN_FETCH_FUNC(retval, ftype),                       \
 388 ASSIGN_FETCH_FUNC(memory, ftype),                       \
 389 ASSIGN_FETCH_FUNC(symbol, ftype),                       \
 390 ASSIGN_FETCH_FUNC(deref, ftype),                        \
 391           }                                             \
 392         }
 393
 394 #define ASSIGN_FETCH_TYPE(ptype, ftype, sign)                   \
 395         __ASSIGN_FETCH_TYPE(#ptype, ptype, ftype, sizeof(ftype), sign, #ptype)
 396
 397 #define FETCH_TYPE_STRING 0
 398 #define FETCH_TYPE_STRSIZE 1
 399
 400 /* Fetch type information table */
 401 static const struct fetch_type {
 402         const char      *name;          /* Name of type */
 403         size_t          size;           /* Byte size of type */
 404         int             is_signed;      /* Signed flag */
 405         print_type_func_t       print;  /* Print functions */
 406         const char      *fmt;           /* Fromat string */
 407         const char      *fmttype;       /* Name in format file */
 408         /* Fetch functions */
 409         fetch_func_t    fetch[FETCH_MTD_END];
 410 } fetch_type_table[] = {
 411         /* Special types */
 412         [FETCH_TYPE_STRING] = __ASSIGN_FETCH_TYPE("string", string, string,
 413                                         sizeof(u32), 1, "__data_loc char[]"),
 414         [FETCH_TYPE_STRSIZE] = __ASSIGN_FETCH_TYPE("string_size", u32,
 415                                         string_size, sizeof(u32), 0, "u32"),
 416         /* Basic types */
 417         ASSIGN_FETCH_TYPE(u8,  u8,  0),
 418         ASSIGN_FETCH_TYPE(u16, u16, 0),
 419         ASSIGN_FETCH_TYPE(u32, u32, 0),
 420         ASSIGN_FETCH_TYPE(u64, u64, 0),
 421         ASSIGN_FETCH_TYPE(s8,  u8,  1),
 422         ASSIGN_FETCH_TYPE(s16, u16, 1),
 423         ASSIGN_FETCH_TYPE(s32, u32, 1),
 424         ASSIGN_FETCH_TYPE(s64, u64, 1),
 425 };
 426
 427 static const struct fetch_type *find_fetch_type(const char *type)
 428 {
 429         int i;
 430
 431         if (!type)
 432                 type = DEFAULT_FETCH_TYPE_STR;
 433
 434         for (i = 0; i < ARRAY_SIZE(fetch_type_table); i++)
 435                 if (strcmp(type, fetch_type_table[i].name) == 0)
 436                         return &fetch_type_table[i];
 437         return NULL;
 438 }
 439
 440 /* Special function : only accept unsigned long */
 441 static __kprobes void fetch_stack_address(struct pt_regs *regs,
 442                                           void *dummy, void *dest)
 443 {
 444         *(unsigned long *)dest = kernel_stack_pointer(regs);
 445 }
 446
 447 static fetch_func_t get_fetch_size_function(const struct fetch_type *type,
 448                                             fetch_func_t orig_fn)
 449 {
 450         int i;
 451
 452         if (type != &fetch_type_table[FETCH_TYPE_STRING])
 453                 return NULL;    /* Only string type needs size function */
 454         for (i = 0; i < FETCH_MTD_END; i++)
 455                 if (type->fetch[i] == orig_fn)
 456                         return fetch_type_table[FETCH_TYPE_STRSIZE].fetch[i];
 457
 458         WARN_ON(1);     /* This should not happen */
 459         return NULL;
 460 }
 461
 462 /**
 463  * Kprobe event core functions
 464  */
 465
 466 struct probe_arg {
 467         struct fetch_param      fetch;
 468         struct fetch_param      fetch_size;
 469         unsigned int            offset; /* Offset from argument entry */
 470         const char              *name;  /* Name of this argument */
 471         const char              *comm;  /* Command of this argument */
 472         const struct fetch_type *type;  /* Type of this argument */
 473 };
 474
 475 /* Flags for trace_probe */
 476 #define TP_FLAG_TRACE   1
 477 #define TP_FLAG_PROFILE 2
 478
 479 struct trace_probe {
 480         struct list_head        list;
 481         struct kretprobe        rp;     /* Use rp.kp for kprobe use */
 482         unsigned long           nhit;
 483         unsigned int            flags;  /* For TP_FLAG_* */
 484         const char              *symbol;        /* symbol name */
 485         struct ftrace_event_class       class;
 486         struct ftrace_event_call        call;
 487         ssize_t                 size;           /* trace entry size */
 488         unsigned int            nr_args;
 489         struct probe_arg        args[];
 490 };
 491
 492 #define SIZEOF_TRACE_PROBE(n)                   \
 493         (offsetof(struct trace_probe, args) +   \
 494         (sizeof(struct probe_arg) * (n)))
 495
 496
 497 static __kprobes int probe_is_return(struct trace_probe *tp)
 498 {
 499         return tp->rp.handler != NULL;
 500 }
 501
 502 static __kprobes const char *probe_symbol(struct trace_probe *tp)
 503 {
 504         return tp->symbol ? tp->symbol : "unknown";
 505 }
 506
 507 static int register_probe_event(struct trace_probe *tp);
 508 static void unregister_probe_event(struct trace_probe *tp);
 509
 510 static DEFINE_MUTEX(probe_lock);
 511 static LIST_HEAD(probe_list);
 512
 513 static int kprobe_dispatcher(struct kprobe *kp, struct pt_regs *regs);
 514 static int kretprobe_dispatcher(struct kretprobe_instance *ri,
 515                                 struct pt_regs *regs);
 516
 517 /* Check the name is good for event/group */
 518 static int check_event_name(const char *name)
 519 {
 520         if (!isalpha(*name) && *name != '_')
 521                 return 0;
 522         while (*++name != '\0') {
 523                 if (!isalpha(*name) && !isdigit(*name) && *name != '_')
 524                         return 0;
 525         }
 526         return 1;
 527 }
 528
 529 /*
 530  * Allocate new trace_probe and initialize it (including kprobes).
 531  */
 532 static struct trace_probe *alloc_trace_probe(const char *group,
 533                                              const char *event,
 534                                              void *addr,
 535                                              const char *symbol,
 536                                              unsigned long offs,
 537                                              int nargs, int is_return)
 538 {
 539         struct trace_probe *tp;
 540         int ret = -ENOMEM;
 541
 542         tp = kzalloc(SIZEOF_TRACE_PROBE(nargs), GFP_KERNEL);
 543         if (!tp)
 544                 return ERR_PTR(ret);
 545
 546         if (symbol) {
 547                 tp->symbol = kstrdup(symbol, GFP_KERNEL);
 548                 if (!tp->symbol)
 549                         goto error;
 550                 tp->rp.kp.symbol_name = tp->symbol;
 551                 tp->rp.kp.offset = offs;
 552         } else
 553                 tp->rp.kp.addr = addr;
 554
 555         if (is_return)
 556                 tp->rp.handler = kretprobe_dispatcher;
 557         else
 558                 tp->rp.kp.pre_handler = kprobe_dispatcher;
 559
 560         if (!event || !check_event_name(event)) {
 561                 ret = -EINVAL;
 562                 goto error;
 563         }
 564
 565         tp->call.class = &tp->class;
 566         tp->call.name = kstrdup(event, GFP_KERNEL);
 567         if (!tp->call.name)
 568                 goto error;
 569
 570         if (!group || !check_event_name(group)) {
 571                 ret = -EINVAL;
 572                 goto error;
 573         }
 574
 575         tp->class.system = kstrdup(group, GFP_KERNEL);
 576         if (!tp->class.system)
 577                 goto error;
 578
 579         INIT_LIST_HEAD(&tp->list);
 580         return tp;
 581 error:
 582         kfree(tp->call.name);
 583         kfree(tp->symbol);
 584         kfree(tp);
 585         return ERR_PTR(ret);
 586 }
 587
 588 static void free_probe_arg(struct probe_arg *arg)
 589 {
 590         if (CHECK_FETCH_FUNCS(deref, arg->fetch.fn))
 591                 free_deref_fetch_param(arg->fetch.data);
 592         else if (CHECK_FETCH_FUNCS(symbol, arg->fetch.fn))
 593                 free_symbol_cache(arg->fetch.data);
 594         kfree(arg->name);
 595         kfree(arg->comm);
 596 }
 597
 598 static void free_trace_probe(struct trace_probe *tp)
 599 {
 600         int i;
 601
 602         for (i = 0; i < tp->nr_args; i++)
 603                 free_probe_arg(&tp->args[i]);
 604
 605         kfree(tp->call.class->system);
 606         kfree(tp->call.name);
 607         kfree(tp->symbol);
 608         kfree(tp);
 609 }
 610
 611 static struct trace_probe *find_probe_event(const char *event,
 612                                             const char *group)
 613 {
 614         struct trace_probe *tp;
 615
 616         list_for_each_entry(tp, &probe_list, list)
 617                 if (strcmp(tp->call.name, event) == 0 &&
 618                     strcmp(tp->call.class->system, group) == 0)
 619                         return tp;
 620         return NULL;
 621 }
 622
 623 /* Unregister a trace_probe and probe_event: call with locking probe_lock */
 624 static void unregister_trace_probe(struct trace_probe *tp)
 625 {
 626         if (probe_is_return(tp))
 627                 unregister_kretprobe(&tp->rp);
 628         else
 629                 unregister_kprobe(&tp->rp.kp);
 630         list_del(&tp->list);
 631         unregister_probe_event(tp);
 632 }
 633
 634 /* Register a trace_probe and probe_event */
 635 static int register_trace_probe(struct trace_probe *tp)
 636 {
 637         struct trace_probe *old_tp;
 638         int ret;
 639
 640         mutex_lock(&probe_lock);
 641
 642         /* register as an event */
 643         old_tp = find_probe_event(tp->call.name, tp->call.class->system);
 644         if (old_tp) {
 645                 /* delete old event */
 646                 unregister_trace_probe(old_tp);
 647                 free_trace_probe(old_tp);
 648         }
 649         ret = register_probe_event(tp);
 650         if (ret) {
 651                 pr_warning("Faild to register probe event(%d)\n", ret);
 652                 goto end;
 653         }
 654
 655         tp->rp.kp.flags |= KPROBE_FLAG_DISABLED;
 656         if (probe_is_return(tp))
 657                 ret = register_kretprobe(&tp->rp);
 658         else
 659                 ret = register_kprobe(&tp->rp.kp);
 660
 661         if (ret) {
 662                 pr_warning("Could not insert probe(%d)\n", ret);
 663                 if (ret == -EILSEQ) {
 664                         pr_warning("Probing address(0x%p) is not an "
 665                                    "instruction boundary.\n",
 666                                    tp->rp.kp.addr);
 667                         ret = -EINVAL;
 668                 }
 669                 unregister_probe_event(tp);
 670         } else
 671                 list_add_tail(&tp->list, &probe_list);
 672 end:
 673         mutex_unlock(&probe_lock);
 674         return ret;
 675 }
 676
 677 /* Split symbol and offset. */
 678 static int split_symbol_offset(char *symbol, unsigned long *offset)
 679 {
 680         char *tmp;
 681         int ret;
 682
 683         if (!offset)
 684                 return -EINVAL;
 685
 686         tmp = strchr(symbol, '+');
 687         if (tmp) {
 688                 /* skip sign because strict_strtol doesn't accept '+' */
 689                 ret = strict_strtoul(tmp + 1, 0, offset);
 690                 if (ret)
 691                         return ret;
 692                 *tmp = '\0';
 693         } else
 694                 *offset = 0;
 695         return 0;
 696 }
 697
 698 #define PARAM_MAX_ARGS 16
 699 #define PARAM_MAX_STACK (THREAD_SIZE / sizeof(unsigned long))
 700
 701 static int parse_probe_vars(char *arg, const struct fetch_type *t,
 702                             struct fetch_param *f, int is_return)
 703 {
 704         int ret = 0;
 705         unsigned long param;
 706
 707         if (strcmp(arg, "retval") == 0) {
 708                 if (is_return)
 709                         f->fn = t->fetch[FETCH_MTD_retval];
 710                 else
 711                         ret = -EINVAL;
 712         } else if (strncmp(arg, "stack", 5) == 0) {
 713                 if (arg[5] == '\0') {
 714                         if (strcmp(t->name, DEFAULT_FETCH_TYPE_STR) == 0)
 715                                 f->fn = fetch_stack_address;
 716                         else
 717                                 ret = -EINVAL;
 718                 } else if (isdigit(arg[5])) {
 719                         ret = strict_strtoul(arg + 5, 10, &param);
 720                         if (ret || param > PARAM_MAX_STACK)
 721                                 ret = -EINVAL;
 722                         else {
 723                                 f->fn = t->fetch[FETCH_MTD_stack];
 724                                 f->data = (void *)param;
 725                         }
 726                 } else
 727                         ret = -EINVAL;
 728         } else
 729                 ret = -EINVAL;
 730         return ret;
 731 }
 732
 733 /* Recursive argument parser */
 734 static int __parse_probe_arg(char *arg, const struct fetch_type *t,
 735                              struct fetch_param *f, int is_return)
 736 {
 737         int ret = 0;
 738         unsigned long param;
 739         long offset;
 740         char *tmp;
 741
 742         switch (arg[0]) {
 743         case '$':
 744                 ret = parse_probe_vars(arg + 1, t, f, is_return);
 745                 break;
 746         case '%':       /* named register */
 747                 ret = regs_query_register_offset(arg + 1);
 748                 if (ret >= 0) {
 749                         f->fn = t->fetch[FETCH_MTD_reg];
 750                         f->data = (void *)(unsigned long)ret;
 751                         ret = 0;
 752                 }
 753                 break;
 754         case '@':       /* memory or symbol */
 755                 if (isdigit(arg[1])) {
 756                         ret = strict_strtoul(arg + 1, 0, &param);
 757                         if (ret)
 758                                 break;
 759                         f->fn = t->fetch[FETCH_MTD_memory];
 760                         f->data = (void *)param;
 761                 } else {
 762                         ret = split_symbol_offset(arg + 1, &offset);
 763                         if (ret)
 764                                 break;
 765                         f->data = alloc_symbol_cache(arg + 1, offset);
 766                         if (f->data)
 767                                 f->fn = t->fetch[FETCH_MTD_symbol];
 768                 }
 769                 break;
 770         case '+':       /* deref memory */
 771         case '-':
 772                 tmp = strchr(arg, '(');
 773                 if (!tmp)
 774                         break;
 775                 *tmp = '\0';
 776                 ret = strict_strtol(arg + 1, 0, &offset);
 777                 if (ret)
 778                         break;
 779                 if (arg[0] == '-')
 780                         offset = -offset;
 781                 arg = tmp + 1;
 782                 tmp = strrchr(arg, ')');
 783                 if (tmp) {
 784                         struct deref_fetch_param *dprm;
 785                         const struct fetch_type *t2 = find_fetch_type(NULL);
 786                         *tmp = '\0';
 787                         dprm = kzalloc(sizeof(struct deref_fetch_param),
 788                                        GFP_KERNEL);
 789                         if (!dprm)
 790                                 return -ENOMEM;
 791                         dprm->offset = offset;
 792                         ret = __parse_probe_arg(arg, t2, &dprm->orig,
 793                                                 is_return);
 794                         if (ret)
 795                                 kfree(dprm);
 796                         else {
 797                                 f->fn = t->fetch[FETCH_MTD_deref];
 798                                 f->data = (void *)dprm;
 799                         }
 800                 }
 801                 break;
 802         }
 803         if (!ret && !f->fn) {   /* Parsed, but do not find fetch method */
 804                 pr_info("%s type has no corresponding fetch method.\n",
 805                         t->name);
 806                 ret = -EINVAL;
 807         }
 808         return ret;
 809 }
 810
 811 /* String length checking wrapper */
 812 static int parse_probe_arg(char *arg, struct trace_probe *tp,
 813                            struct probe_arg *parg, int is_return)
 814 {
 815         const char *t;
 816         int ret;
 817
 818         if (strlen(arg) > MAX_ARGSTR_LEN) {
 819                 pr_info("Argument is too long.: %s\n",  arg);
 820                 return -ENOSPC;
 821         }
 822         parg->comm = kstrdup(arg, GFP_KERNEL);
 823         if (!parg->comm) {
 824                 pr_info("Failed to allocate memory for command '%s'.\n", arg);
 825                 return -ENOMEM;
 826         }
 827         t = strchr(parg->comm, ':');
 828         if (t) {
 829                 arg[t - parg->comm] = '\0';
 830                 t++;
 831         }
 832         parg->type = find_fetch_type(t);
 833         if (!parg->type) {
 834                 pr_info("Unsupported type: %s\n", t);
 835                 return -EINVAL;
 836         }
 837         parg->offset = tp->size;
 838         tp->size += parg->type->size;
 839         ret = __parse_probe_arg(arg, parg->type, &parg->fetch, is_return);
 840         if (ret >= 0) {
 841                 parg->fetch_size.fn = get_fetch_size_function(parg->type,
 842                                                               parg->fetch.fn);
 843                 parg->fetch_size.data = parg->fetch.data;
 844         }
 845         return ret;
 846 }
 847
 848 /* Return 1 if name is reserved or already used by another argument */
 849 static int conflict_field_name(const char *name,
 850                                struct probe_arg *args, int narg)
 851 {
 852         int i;
 853         for (i = 0; i < ARRAY_SIZE(reserved_field_names); i++)
 854                 if (strcmp(reserved_field_names[i], name) == 0)
 855                         return 1;
 856         for (i = 0; i < narg; i++)
 857                 if (strcmp(args[i].name, name) == 0)
 858                         return 1;
 859         return 0;
 860 }
 861
 862 static int create_trace_probe(int argc, char **argv)
 863 {
 864         /*
 865          * Argument syntax:
 866          *  - Add kprobe: p[:[GRP/]EVENT] KSYM[+OFFS]|KADDR [FETCHARGS]
 867          *  - Add kretprobe: r[:[GRP/]EVENT] KSYM[+0] [FETCHARGS]
 868          * Fetch args:
 869          *  $retval     : fetch return value
 870          *  $stack      : fetch stack address
 871          *  $stackN     : fetch Nth of stack (N:0-)
 872          *  @ADDR       : fetch memory at ADDR (ADDR should be in kernel)
 873          *  @SYM[+|-offs] : fetch memory at SYM +|- offs (SYM is a data symbol)
 874          *  %REG        : fetch register REG
 875          * Dereferencing memory fetch:
 876          *  +|-offs(ARG) : fetch memory at ARG +|- offs address.
 877          * Alias name of args:
 878          *  NAME=FETCHARG : set NAME as alias of FETCHARG.
 879          * Type of args:
 880          *  FETCHARG:TYPE : use TYPE instead of unsigned long.
 881          */
 882         struct trace_probe *tp;
 883         int i, ret = 0;
 884         int is_return = 0, is_delete = 0;
 885         char *symbol = NULL, *event = NULL, *group = NULL;
 886         char *arg, *tmp;
 887         unsigned long offset = 0;
 888         void *addr = NULL;
 889         char buf[MAX_EVENT_NAME_LEN];
 890
 891         /* argc must be >= 1 */
 892         if (argv[0][0] == 'p')
 893                 is_return = 0;
 894         else if (argv[0][0] == 'r')
 895                 is_return = 1;
 896         else if (argv[0][0] == '-')
 897                 is_delete = 1;
 898         else {
 899                 pr_info("Probe definition must be started with 'p', 'r' or"
 900                         " '-'.\n");
 901                 return -EINVAL;
 902         }
 903
 904         if (argv[0][1] == ':') {
 905                 event = &argv[0][2];
 906                 if (strchr(event, '/')) {
 907                         group = event;
 908                         event = strchr(group, '/') + 1;
 909                         event[-1] = '\0';
 910                         if (strlen(group) == 0) {
 911                                 pr_info("Group name is not specified\n");
 912                                 return -EINVAL;
 913                         }
 914                 }
 915                 if (strlen(event) == 0) {
 916                         pr_info("Event name is not specified\n");
 917                         return -EINVAL;
 918                 }
 919         }
 920         if (!group)
 921                 group = KPROBE_EVENT_SYSTEM;
 922
 923         if (is_delete) {
 924                 if (!event) {
 925                         pr_info("Delete command needs an event name.\n");
 926                         return -EINVAL;
 927                 }
 928                 mutex_lock(&probe_lock);
 929                 tp = find_probe_event(event, group);
 930                 if (!tp) {
 931                         mutex_unlock(&probe_lock);
 932                         pr_info("Event %s/%s doesn't exist.\n", group, event);
 933                         return -ENOENT;
 934                 }
 935                 /* delete an event */
 936                 unregister_trace_probe(tp);
 937                 free_trace_probe(tp);
 938                 mutex_unlock(&probe_lock);
 939                 return 0;
 940         }
 941
 942         if (argc < 2) {
 943                 pr_info("Probe point is not specified.\n");
 944                 return -EINVAL;
 945         }
 946         if (isdigit(argv[1][0])) {
 947                 if (is_return) {
 948                         pr_info("Return probe point must be a symbol.\n");
 949                         return -EINVAL;
 950                 }
 951                 /* an address specified */
 952                 ret = strict_strtoul(&argv[1][0], 0, (unsigned long *)&addr);
 953                 if (ret) {
 954                         pr_info("Failed to parse address.\n");
 955                         return ret;
 956                 }
 957         } else {
 958                 /* a symbol specified */
 959                 symbol = argv[1];
 960                 /* TODO: support .init module functions */
 961                 ret = split_symbol_offset(symbol, &offset);
 962                 if (ret) {
 963                         pr_info("Failed to parse symbol.\n");
 964                         return ret;
 965                 }
 966                 if (offset && is_return) {
 967                         pr_info("Return probe must be used without offset.\n");
 968                         return -EINVAL;
 969                 }
 970         }
 971         argc -= 2; argv += 2;
 972
 973         /* setup a probe */
 974         if (!event) {
 975                 /* Make a new event name */
 976                 if (symbol)
 977                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c_%s_%ld",
 978                                  is_return ? 'r' : 'p', symbol, offset);
 979                 else
 980                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c_0x%p",
 981                                  is_return ? 'r' : 'p', addr);
 982                 event = buf;
 983         }
 984         tp = alloc_trace_probe(group, event, addr, symbol, offset, argc,
 985                                is_return);
 986         if (IS_ERR(tp)) {
 987                 pr_info("Failed to allocate trace_probe.(%d)\n",
 988                         (int)PTR_ERR(tp));
 989                 return PTR_ERR(tp);
 990         }
 991
 992         /* parse arguments */
 993         ret = 0;
 994         for (i = 0; i < argc && i < MAX_TRACE_ARGS; i++) {
 995                 /* Increment count for freeing args in error case */
 996                 tp->nr_args++;
 997
 998                 /* Parse argument name */
 999                 arg = strchr(argv[i], '=');
1000                 if (arg) {
1001                         *arg++ = '\0';
1002                         tp->args[i].name = kstrdup(argv[i], GFP_KERNEL);
1003                 } else {
1004                         arg = argv[i];
1005                         /* If argument name is omitted, set "argN" */
1006                         snprintf(buf, MAX_EVENT_NAME_LEN, "arg%d", i + 1);
1007                         tp->args[i].name = kstrdup(buf, GFP_KERNEL);
1008                 }
1009
1010                 if (!tp->args[i].name) {
1011                         pr_info("Failed to allocate argument[%d] name.\n", i);
1012                         ret = -ENOMEM;
1013                         goto error;
1014                 }
1015                 tmp = strchr(tp->args[i].name, ':');
1016                 if (tmp)
1017                         *tmp = '_';     /* convert : to _ */
1018
1019                 if (conflict_field_name(tp->args[i].name, tp->args, i)) {
1020                         pr_info("Argument[%d] name '%s' conflicts with "
1021                                 "another field.\n", i, argv[i]);
1022                         ret = -EINVAL;
1023                         goto error;
1024                 }
1025
1026                 /* Parse fetch argument */
1027                 ret = parse_probe_arg(arg, tp, &tp->args[i], is_return);
1028                 if (ret) {
1029                         pr_info("Parse error at argument[%d]. (%d)\n", i, ret);
1030                         goto error;
1031                 }
1032         }
1033
1034         ret = register_trace_probe(tp);
1035         if (ret)
1036                 goto error;
1037         return 0;
1038
1039 error:
1040         free_trace_probe(tp);
1041         return ret;
1042 }
1043
1044 static void cleanup_all_probes(void)
1045 {
1046         struct trace_probe *tp;
1047
1048         mutex_lock(&probe_lock);
1049         /* TODO: Use batch unregistration */
1050         while (!list_empty(&probe_list)) {
1051                 tp = list_entry(probe_list.next, struct trace_probe, list);
1052                 unregister_trace_probe(tp);
1053                 free_trace_probe(tp);
1054         }
1055         mutex_unlock(&probe_lock);
1056 }
1057
1058
1059 /* Probes listing interfaces */
1060 static void *probes_seq_start(struct seq_file *m, loff_t *pos)
1061 {
1062         mutex_lock(&probe_lock);
1063         return seq_list_start(&probe_list, *pos);
1064 }
1065
1066 static void *probes_seq_next(struct seq_file *m, void *v, loff_t *pos)
1067 {
1068         return seq_list_next(v, &probe_list, pos);
1069 }
1070
1071 static void probes_seq_stop(struct seq_file *m, void *v)
1072 {
1073         mutex_unlock(&probe_lock);
1074 }
1075
1076 static int probes_seq_show(struct seq_file *m, void *v)
1077 {
1078         struct trace_probe *tp = v;
1079         int i;
1080
1081         seq_printf(m, "%c", probe_is_return(tp) ? 'r' : 'p');
1082         seq_printf(m, ":%s/%s", tp->call.class->system, tp->call.name);
1083
1084         if (!tp->symbol)
1085                 seq_printf(m, " 0x%p", tp->rp.kp.addr);
1086         else if (tp->rp.kp.offset)
1087                 seq_printf(m, " %s+%u", probe_symbol(tp), tp->rp.kp.offset);
1088         else
1089                 seq_printf(m, " %s", probe_symbol(tp));
1090
1091         for (i = 0; i < tp->nr_args; i++)
1092                 seq_printf(m, " %s=%s", tp->args[i].name, tp->args[i].comm);
1093         seq_printf(m, "\n");
1094
1095         return 0;
1096 }
1097
1098 static const struct seq_operations probes_seq_op = {
1099         .start  = probes_seq_start,
1100         .next   = probes_seq_next,
1101         .stop   = probes_seq_stop,
1102         .show   = probes_seq_show
1103 };
1104
1105 static int probes_open(struct inode *inode, struct file *file)
1106 {
1107         if ((file->f_mode & FMODE_WRITE) &&
1108             (file->f_flags & O_TRUNC))
1109                 cleanup_all_probes();
1110
1111         return seq_open(file, &probes_seq_op);
1112 }
1113
1114 static int command_trace_probe(const char *buf)
1115 {
1116         char **argv;
1117         int argc = 0, ret = 0;
1118
1119         argv = argv_split(GFP_KERNEL, buf, &argc);
1120         if (!argv)
1121                 return -ENOMEM;
1122
1123         if (argc)
1124                 ret = create_trace_probe(argc, argv);
1125
1126         argv_free(argv);
1127         return ret;
1128 }
1129
1130 #define WRITE_BUFSIZE 128
1131
1132 static ssize_t probes_write(struct file *file, const char __user *buffer,
1133                             size_t count, loff_t *ppos)
1134 {
1135         char *kbuf, *tmp;
1136         int ret;
1137         size_t done;
1138         size_t size;
1139
1140         kbuf = kmalloc(WRITE_BUFSIZE, GFP_KERNEL);
1141         if (!kbuf)
1142                 return -ENOMEM;
1143
1144         ret = done = 0;
1145         while (done < count) {
1146                 size = count - done;
1147                 if (size >= WRITE_BUFSIZE)
1148                         size = WRITE_BUFSIZE - 1;
1149                 if (copy_from_user(kbuf, buffer + done, size)) {
1150                         ret = -EFAULT;
1151                         goto out;
1152                 }
1153                 kbuf[size] = '\0';
1154                 tmp = strchr(kbuf, '\n');
1155                 if (tmp) {
1156                         *tmp = '\0';
1157                         size = tmp - kbuf + 1;
1158                 } else if (done + size < count) {
1159                         pr_warning("Line length is too long: "
1160                                    "Should be less than %d.", WRITE_BUFSIZE);
1161                         ret = -EINVAL;
1162                         goto out;
1163                 }
1164                 done += size;
1165                 /* Remove comments */
1166                 tmp = strchr(kbuf, '#');
1167                 if (tmp)
1168                         *tmp = '\0';
1169
1170                 ret = command_trace_probe(kbuf);
1171                 if (ret)
1172                         goto out;
1173         }
1174         ret = done;
1175 out:
1176         kfree(kbuf);
1177         return ret;
1178 }
1179
1180 static const struct file_operations kprobe_events_ops = {
1181         .owner          = THIS_MODULE,
1182         .open           = probes_open,
1183         .read           = seq_read,
1184         .llseek         = seq_lseek,
1185         .release        = seq_release,
1186         .write          = probes_write,
1187 };
1188
1189 /* Probes profiling interfaces */
1190 static int probes_profile_seq_show(struct seq_file *m, void *v)
1191 {
1192         struct trace_probe *tp = v;
1193
1194         seq_printf(m, "  %-44s %15lu %15lu\n", tp->call.name, tp->nhit,
1195                    tp->rp.kp.nmissed);
1196
1197         return 0;
1198 }
1199
1200 static const struct seq_operations profile_seq_op = {
1201         .start  = probes_seq_start,
1202         .next   = probes_seq_next,
1203         .stop   = probes_seq_stop,
1204         .show   = probes_profile_seq_show
1205 };
1206
1207 static int profile_open(struct inode *inode, struct file *file)
1208 {
1209         return seq_open(file, &profile_seq_op);
1210 }
1211
1212 static const struct file_operations kprobe_profile_ops = {
1213         .owner          = THIS_MODULE,
1214         .open           = profile_open,
1215         .read           = seq_read,
1216         .llseek         = seq_lseek,
1217         .release        = seq_release,
1218 };
1219
1220 /* Sum up total data length for dynamic arraies (strings) */
1221 static __kprobes int __get_data_size(struct trace_probe *tp,
1222                                      struct pt_regs *regs)
1223 {
1224         int i, ret = 0;
1225         u32 len;
1226
1227         for (i = 0; i < tp->nr_args; i++)
1228                 if (unlikely(tp->args[i].fetch_size.fn)) {
1229                         call_fetch(&tp->args[i].fetch_size, regs, &len);
1230                         ret += len;
1231                 }
1232
1233         return ret;
1234 }
1235
1236 /* Store the value of each argument */
1237 static __kprobes void store_trace_args(int ent_size, struct trace_probe *tp,
1238                                        struct pt_regs *regs,
1239                                        u8 *data, int maxlen)
1240 {
1241         int i;
1242         u32 end = tp->size;
1243         u32 *dl;        /* Data (relative) location */
1244
1245         for (i = 0; i < tp->nr_args; i++) {
1246                 if (unlikely(tp->args[i].fetch_size.fn)) {
1247                         /*
1248                          * First, we set the relative location and
1249                          * maximum data length to *dl
1250                          */
1251                         dl = (u32 *)(data + tp->args[i].offset);
1252                         *dl = make_data_rloc(maxlen, end - tp->args[i].offset);
1253                         /* Then try to fetch string or dynamic array data */
1254                         call_fetch(&tp->args[i].fetch, regs, dl);
1255                         /* Reduce maximum length */
1256                         end += get_rloc_len(*dl);
1257                         maxlen -= get_rloc_len(*dl);
1258                         /* Trick here, convert data_rloc to data_loc */
1259                         *dl = convert_rloc_to_loc(*dl,
1260                                  ent_size + tp->args[i].offset);
1261                 } else
1262                         /* Just fetching data normally */
1263                         call_fetch(&tp->args[i].fetch, regs,
1264                                    data + tp->args[i].offset);
1265         }
1266 }
1267
1268 /* Kprobe handler */
1269 static __kprobes void kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs)
1270 {
1271         struct trace_probe *tp = container_of(kp, struct trace_probe, rp.kp);
1272         struct kprobe_trace_entry_head *entry;
1273         struct ring_buffer_event *event;
1274         struct ring_buffer *buffer;
1275         int size, dsize, pc;
1276         unsigned long irq_flags;
1277         struct ftrace_event_call *call = &tp->call;
1278
1279         tp->nhit++;
1280
1281         local_save_flags(irq_flags);
1282         pc = preempt_count();
1283
1284         dsize = __get_data_size(tp, regs);
1285         size = sizeof(*entry) + tp->size + dsize;
1286
1287         event = trace_current_buffer_lock_reserve(&buffer, call->event.type,
1288                                                   size, irq_flags, pc);
1289         if (!event)
1290                 return;
1291
1292         entry = ring_buffer_event_data(event);
1293         entry->ip = (unsigned long)kp->addr;
1294         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1295
1296         if (!filter_current_check_discard(buffer, call, entry, event))
1297                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
1298 }
1299
1300 /* Kretprobe handler */
1301 static __kprobes void kretprobe_trace_func(struct kretprobe_instance *ri,
1302                                           struct pt_regs *regs)
1303 {
1304         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
1305         struct kretprobe_trace_entry_head *entry;
1306         struct ring_buffer_event *event;
1307         struct ring_buffer *buffer;
1308         int size, pc, dsize;
1309         unsigned long irq_flags;
1310         struct ftrace_event_call *call = &tp->call;
1311
1312         local_save_flags(irq_flags);
1313         pc = preempt_count();
1314
1315         dsize = __get_data_size(tp, regs);
1316         size = sizeof(*entry) + tp->size + dsize;
1317
1318         event = trace_current_buffer_lock_reserve(&buffer, call->event.type,
1319                                                   size, irq_flags, pc);
1320         if (!event)
1321                 return;
1322
1323         entry = ring_buffer_event_data(event);
1324         entry->func = (unsigned long)tp->rp.kp.addr;
1325         entry->ret_ip = (unsigned long)ri->ret_addr;
1326         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1327
1328         if (!filter_current_check_discard(buffer, call, entry, event))
1329                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
1330 }
1331
1332 /* Event entry printers */
1333 enum print_line_t
1334 print_kprobe_event(struct trace_iterator *iter, int flags,
1335                    struct trace_event *event)
1336 {
1337         struct kprobe_trace_entry_head *field;
1338         struct trace_seq *s = &iter->seq;
1339         struct trace_probe *tp;
1340         u8 *data;
1341         int i;
1342
1343         field = (struct kprobe_trace_entry_head *)iter->ent;
1344         tp = container_of(event, struct trace_probe, call.event);
1345
1346         if (!trace_seq_printf(s, "%s: (", tp->call.name))
1347                 goto partial;
1348
1349         if (!seq_print_ip_sym(s, field->ip, flags | TRACE_ITER_SYM_OFFSET))
1350                 goto partial;
1351
1352         if (!trace_seq_puts(s, ")"))
1353                 goto partial;
1354
1355         data = (u8 *)&field[1];
1356         for (i = 0; i < tp->nr_args; i++)
1357                 if (!tp->args[i].type->print(s, tp->args[i].name,
1358                                              data + tp->args[i].offset, field))
1359                         goto partial;
1360
1361         if (!trace_seq_puts(s, "\n"))
1362                 goto partial;
1363
1364         return TRACE_TYPE_HANDLED;
1365 partial:
1366         return TRACE_TYPE_PARTIAL_LINE;
1367 }
1368
1369 enum print_line_t
1370 print_kretprobe_event(struct trace_iterator *iter, int flags,
1371                       struct trace_event *event)
1372 {
1373         struct kretprobe_trace_entry_head *field;
1374         struct trace_seq *s = &iter->seq;
1375         struct trace_probe *tp;
1376         u8 *data;
1377         int i;
1378
1379         field = (struct kretprobe_trace_entry_head *)iter->ent;
1380         tp = container_of(event, struct trace_probe, call.event);
1381
1382         if (!trace_seq_printf(s, "%s: (", tp->call.name))
1383                 goto partial;
1384
1385         if (!seq_print_ip_sym(s, field->ret_ip, flags | TRACE_ITER_SYM_OFFSET))
1386                 goto partial;
1387
1388         if (!trace_seq_puts(s, " <- "))
1389                 goto partial;
1390
1391         if (!seq_print_ip_sym(s, field->func, flags & ~TRACE_ITER_SYM_OFFSET))
1392                 goto partial;
1393
1394         if (!trace_seq_puts(s, ")"))
1395                 goto partial;
1396
1397         data = (u8 *)&field[1];
1398         for (i = 0; i < tp->nr_args; i++)
1399                 if (!tp->args[i].type->print(s, tp->args[i].name,
1400                                              data + tp->args[i].offset, field))
1401                         goto partial;
1402
1403         if (!trace_seq_puts(s, "\n"))
1404                 goto partial;
1405
1406         return TRACE_TYPE_HANDLED;
1407 partial:
1408         return TRACE_TYPE_PARTIAL_LINE;
1409 }
1410
1411 static int probe_event_enable(struct ftrace_event_call *call)
1412 {
1413         struct trace_probe *tp = (struct trace_probe *)call->data;
1414
1415         tp->flags |= TP_FLAG_TRACE;
1416         if (probe_is_return(tp))
1417                 return enable_kretprobe(&tp->rp);
1418         else
1419                 return enable_kprobe(&tp->rp.kp);
1420 }
1421
1422 static void probe_event_disable(struct ftrace_event_call *call)
1423 {
1424         struct trace_probe *tp = (struct trace_probe *)call->data;
1425
1426         tp->flags &= ~TP_FLAG_TRACE;
1427         if (!(tp->flags & (TP_FLAG_TRACE | TP_FLAG_PROFILE))) {
1428                 if (probe_is_return(tp))
1429                         disable_kretprobe(&tp->rp);
1430                 else
1431                         disable_kprobe(&tp->rp.kp);
1432         }
1433 }
1434
1435 #undef DEFINE_FIELD
1436 #define DEFINE_FIELD(type, item, name, is_signed)                       \
1437         do {                                                            \
1438                 ret = trace_define_field(event_call, #type, name,       \
1439                                          offsetof(typeof(field), item), \
1440                                          sizeof(field.item), is_signed, \
1441                                          FILTER_OTHER);                 \
1442                 if (ret)                                                \
1443                         return ret;                                     \
1444         } while (0)
1445
1446 static int kprobe_event_define_fields(struct ftrace_event_call *event_call)
1447 {
1448         int ret, i;
1449         struct kprobe_trace_entry_head field;
1450         struct trace_probe *tp = (struct trace_probe *)event_call->data;
1451
1452         DEFINE_FIELD(unsigned long, ip, FIELD_STRING_IP, 0);
1453         /* Set argument names as fields */
1454         for (i = 0; i < tp->nr_args; i++) {
1455                 ret = trace_define_field(event_call, tp->args[i].type->fmttype,
1456                                          tp->args[i].name,
1457                                          sizeof(field) + tp->args[i].offset,
1458                                          tp->args[i].type->size,
1459                                          tp->args[i].type->is_signed,
1460                                          FILTER_OTHER);
1461                 if (ret)
1462                         return ret;
1463         }
1464         return 0;
1465 }
1466
1467 static int kretprobe_event_define_fields(struct ftrace_event_call *event_call)
1468 {
1469         int ret, i;
1470         struct kretprobe_trace_entry_head field;
1471         struct trace_probe *tp = (struct trace_probe *)event_call->data;
1472
1473         DEFINE_FIELD(unsigned long, func, FIELD_STRING_FUNC, 0);
1474         DEFINE_FIELD(unsigned long, ret_ip, FIELD_STRING_RETIP, 0);
1475         /* Set argument names as fields */
1476         for (i = 0; i < tp->nr_args; i++) {
1477                 ret = trace_define_field(event_call, tp->args[i].type->fmttype,
1478                                          tp->args[i].name,
1479                                          sizeof(field) + tp->args[i].offset,
1480                                          tp->args[i].type->size,
1481                                          tp->args[i].type->is_signed,
1482                                          FILTER_OTHER);
1483                 if (ret)
1484                         return ret;
1485         }
1486         return 0;
1487 }
1488
1489 static int __set_print_fmt(struct trace_probe *tp, char *buf, int len)
1490 {
1491         int i;
1492         int pos = 0;
1493
1494         const char *fmt, *arg;
1495
1496         if (!probe_is_return(tp)) {
1497                 fmt = "(%lx)";
1498                 arg = "REC->" FIELD_STRING_IP;
1499         } else {
1500                 fmt = "(%lx <- %lx)";
1501                 arg = "REC->" FIELD_STRING_FUNC ", REC->" FIELD_STRING_RETIP;
1502         }
1503
1504         /* When len=0, we just calculate the needed length */
1505 #define LEN_OR_ZERO (len ? len - pos : 0)
1506
1507         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"%s", fmt);
1508
1509         for (i = 0; i < tp->nr_args; i++) {
1510                 pos += snprintf(buf + pos, LEN_OR_ZERO, " %s=%s",
1511                                 tp->args[i].name, tp->args[i].type->fmt);
1512         }
1513
1514         pos += snprintf(buf + pos, LEN_OR_ZERO, "\", %s", arg);
1515
1516         for (i = 0; i < tp->nr_args; i++) {
1517                 if (strcmp(tp->args[i].type->name, "string") == 0)
1518                         pos += snprintf(buf + pos, LEN_OR_ZERO,
1519                                         ", __get_str(%s)",
1520                                         tp->args[i].name);
1521                 else
1522                         pos += snprintf(buf + pos, LEN_OR_ZERO, ", REC->%s",
1523                                         tp->args[i].name);
1524         }
1525
1526 #undef LEN_OR_ZERO
1527
1528         /* return the length of print_fmt */
1529         return pos;
1530 }
1531
1532 static int set_print_fmt(struct trace_probe *tp)
1533 {
1534         int len;
1535         char *print_fmt;
1536
1537         /* First: called with 0 length to calculate the needed length */
1538         len = __set_print_fmt(tp, NULL, 0);
1539         print_fmt = kmalloc(len + 1, GFP_KERNEL);
1540         if (!print_fmt)
1541                 return -ENOMEM;
1542
1543         /* Second: actually write the @print_fmt */
1544         __set_print_fmt(tp, print_fmt, len + 1);
1545         tp->call.print_fmt = print_fmt;
1546
1547         return 0;
1548 }
1549
1550 #ifdef CONFIG_PERF_EVENTS
1551
1552 /* Kprobe profile handler */
1553 static __kprobes void kprobe_perf_func(struct kprobe *kp,
1554                                          struct pt_regs *regs)
1555 {
1556         struct trace_probe *tp = container_of(kp, struct trace_probe, rp.kp);
1557         struct ftrace_event_call *call = &tp->call;
1558         struct kprobe_trace_entry_head *entry;
1559         struct hlist_head *head;
1560         int size, __size, dsize;
1561         int rctx;
1562
1563         dsize = __get_data_size(tp, regs);
1564         __size = sizeof(*entry) + tp->size + dsize;
1565         size = ALIGN(__size + sizeof(u32), sizeof(u64));
1566         size -= sizeof(u32);
1567         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
1568                      "profile buffer not large enough"))
1569                 return;
1570
1571         entry = perf_trace_buf_prepare(size, call->event.type, regs, &rctx);
1572         if (!entry)
1573                 return;
1574
1575         entry->ip = (unsigned long)kp->addr;
1576         memset(&entry[1], 0, dsize);
1577         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1578
1579         head = this_cpu_ptr(call->perf_events);
1580         perf_trace_buf_submit(entry, size, rctx, entry->ip, 1, regs, head);
1581 }
1582
1583 /* Kretprobe profile handler */
1584 static __kprobes void kretprobe_perf_func(struct kretprobe_instance *ri,
1585                                             struct pt_regs *regs)
1586 {
1587         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
1588         struct ftrace_event_call *call = &tp->call;
1589         struct kretprobe_trace_entry_head *entry;
1590         struct hlist_head *head;
1591         int size, __size, dsize;
1592         int rctx;
1593
1594         dsize = __get_data_size(tp, regs);
1595         __size = sizeof(*entry) + tp->size + dsize;
1596         size = ALIGN(__size + sizeof(u32), sizeof(u64));
1597         size -= sizeof(u32);
1598         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
1599                      "profile buffer not large enough"))
1600                 return;
1601
1602         entry = perf_trace_buf_prepare(size, call->event.type, regs, &rctx);
1603         if (!entry)
1604                 return;
1605
1606         entry->func = (unsigned long)tp->rp.kp.addr;
1607         entry->ret_ip = (unsigned long)ri->ret_addr;
1608         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1609
1610         head = this_cpu_ptr(call->perf_events);
1611         perf_trace_buf_submit(entry, size, rctx, entry->ret_ip, 1, regs, head);
1612 }
1613
1614 static int probe_perf_enable(struct ftrace_event_call *call)
1615 {
1616         struct trace_probe *tp = (struct trace_probe *)call->data;
1617
1618         tp->flags |= TP_FLAG_PROFILE;
1619
1620         if (probe_is_return(tp))
1621                 return enable_kretprobe(&tp->rp);
1622         else
1623                 return enable_kprobe(&tp->rp.kp);
1624 }
1625
1626 static void probe_perf_disable(struct ftrace_event_call *call)
1627 {
1628         struct trace_probe *tp = (struct trace_probe *)call->data;
1629
1630         tp->flags &= ~TP_FLAG_PROFILE;
1631
1632         if (!(tp->flags & TP_FLAG_TRACE)) {
1633                 if (probe_is_return(tp))
1634                         disable_kretprobe(&tp->rp);
1635                 else
1636                         disable_kprobe(&tp->rp.kp);
1637         }
1638 }
1639 #endif  /* CONFIG_PERF_EVENTS */
1640
1641 static __kprobes
1642 int kprobe_register(struct ftrace_event_call *event, enum trace_reg type)
1643 {
1644         switch (type) {
1645         case TRACE_REG_REGISTER:
1646                 return probe_event_enable(event);
1647         case TRACE_REG_UNREGISTER:
1648                 probe_event_disable(event);
1649                 return 0;
1650
1651 #ifdef CONFIG_PERF_EVENTS
1652         case TRACE_REG_PERF_REGISTER:
1653                 return probe_perf_enable(event);
1654         case TRACE_REG_PERF_UNREGISTER:
1655                 probe_perf_disable(event);
1656                 return 0;
1657 #endif
1658         }
1659         return 0;
1660 }
1661
1662 static __kprobes
1663 int kprobe_dispatcher(struct kprobe *kp, struct pt_regs *regs)
1664 {
1665         struct trace_probe *tp = container_of(kp, struct trace_probe, rp.kp);
1666
1667         if (tp->flags & TP_FLAG_TRACE)
1668                 kprobe_trace_func(kp, regs);
1669 #ifdef CONFIG_PERF_EVENTS
1670         if (tp->flags & TP_FLAG_PROFILE)
1671                 kprobe_perf_func(kp, regs);
1672 #endif
1673         return 0;       /* We don't tweek kernel, so just return 0 */
1674 }
1675
1676 static __kprobes
1677 int kretprobe_dispatcher(struct kretprobe_instance *ri, struct pt_regs *regs)
1678 {
1679         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
1680
1681         if (tp->flags & TP_FLAG_TRACE)
1682                 kretprobe_trace_func(ri, regs);
1683 #ifdef CONFIG_PERF_EVENTS
1684         if (tp->flags & TP_FLAG_PROFILE)
1685                 kretprobe_perf_func(ri, regs);
1686 #endif
1687         return 0;       /* We don't tweek kernel, so just return 0 */
1688 }
1689
1690 static struct trace_event_functions kretprobe_funcs = {
1691         .trace          = print_kretprobe_event
1692 };
1693
1694 static struct trace_event_functions kprobe_funcs = {
1695         .trace          = print_kprobe_event
1696 };
1697
1698 static int register_probe_event(struct trace_probe *tp)
1699 {
1700         struct ftrace_event_call *call = &tp->call;
1701         int ret;
1702
1703         /* Initialize ftrace_event_call */
1704         INIT_LIST_HEAD(&call->class->fields);
1705         if (probe_is_return(tp)) {
1706                 call->event.funcs = &kretprobe_funcs;
1707                 call->class->define_fields = kretprobe_event_define_fields;
1708         } else {
1709                 call->event.funcs = &kprobe_funcs;
1710                 call->class->define_fields = kprobe_event_define_fields;
1711         }
1712         if (set_print_fmt(tp) < 0)
1713                 return -ENOMEM;
1714         ret = register_ftrace_event(&call->event);
1715         if (!ret) {
1716                 kfree(call->print_fmt);
1717                 return -ENODEV;
1718         }
1719         call->flags = 0;
1720         call->class->reg = kprobe_register;
1721         call->data = tp;
1722         ret = trace_add_event_call(call);
1723         if (ret) {
1724                 pr_info("Failed to register kprobe event: %s\n", call->name);
1725                 kfree(call->print_fmt);
1726                 unregister_ftrace_event(&call->event);
1727         }
1728         return ret;
1729 }
1730
1731 static void unregister_probe_event(struct trace_probe *tp)
1732 {
1733         /* tp->event is unregistered in trace_remove_event_call() */
1734         trace_remove_event_call(&tp->call);
1735         kfree(tp->call.print_fmt);
1736 }
1737
1738 /* Make a debugfs interface for controling probe points */
1739 static __init int init_kprobe_trace(void)
1740 {
1741         struct dentry *d_tracer;
1742         struct dentry *entry;
1743
1744         d_tracer = tracing_init_dentry();
1745         if (!d_tracer)
1746                 return 0;
1747
1748         entry = debugfs_create_file("kprobe_events", 0644, d_tracer,
1749                                     NULL, &kprobe_events_ops);
1750
1751         /* Event list interface */
1752         if (!entry)
1753                 pr_warning("Could not create debugfs "
1754                            "'kprobe_events' entry\n");
1755
1756         /* Profile interface */
1757         entry = debugfs_create_file("kprobe_profile", 0444, d_tracer,
1758                                     NULL, &kprobe_profile_ops);
1759
1760         if (!entry)
1761                 pr_warning("Could not create debugfs "
1762                            "'kprobe_profile' entry\n");
1763         return 0;
1764 }
1765 fs_initcall(init_kprobe_trace);
1766
1767
1768 #ifdef CONFIG_FTRACE_STARTUP_TEST
1769
1770 static int kprobe_trace_selftest_target(int a1, int a2, int a3,
1771                                         int a4, int a5, int a6)
1772 {
1773         return a1 + a2 + a3 + a4 + a5 + a6;
1774 }
1775
1776 static __init int kprobe_trace_self_tests_init(void)
1777 {
1778         int ret, warn = 0;
1779         int (*target)(int, int, int, int, int, int);
1780         struct trace_probe *tp;
1781
1782         target = kprobe_trace_selftest_target;
1783
1784         pr_info("Testing kprobe tracing: ");
1785
1786         ret = command_trace_probe("p:testprobe kprobe_trace_selftest_target "
1787                                   "$stack $stack0 +0($stack)");
1788         if (WARN_ON_ONCE(ret)) {
1789                 pr_warning("error on probing function entry.\n");
1790                 warn++;
1791         } else {
1792                 /* Enable trace point */
1793                 tp = find_probe_event("testprobe", KPROBE_EVENT_SYSTEM);
1794                 if (WARN_ON_ONCE(tp == NULL)) {
1795                         pr_warning("error on getting new probe.\n");
1796                         warn++;
1797                 } else
1798                         probe_event_enable(&tp->call);
1799         }
1800
1801         ret = command_trace_probe("r:testprobe2 kprobe_trace_selftest_target "
1802                                   "$retval");
1803         if (WARN_ON_ONCE(ret)) {
1804                 pr_warning("error on probing function return.\n");
1805                 warn++;
1806         } else {
1807                 /* Enable trace point */
1808                 tp = find_probe_event("testprobe2", KPROBE_EVENT_SYSTEM);
1809                 if (WARN_ON_ONCE(tp == NULL)) {
1810                         pr_warning("error on getting new probe.\n");
1811                         warn++;
1812                 } else
1813                         probe_event_enable(&tp->call);
1814         }
1815
1816         if (warn)
1817                 goto end;
1818
1819         ret = target(1, 2, 3, 4, 5, 6);
1820
1821         ret = command_trace_probe("-:testprobe");
1822         if (WARN_ON_ONCE(ret)) {
1823                 pr_warning("error on deleting a probe.\n");
1824                 warn++;
1825         }
1826
1827         ret = command_trace_probe("-:testprobe2");
1828         if (WARN_ON_ONCE(ret)) {
1829                 pr_warning("error on deleting a probe.\n");
1830                 warn++;
1831         }
1832
1833 end:
1834         cleanup_all_probes();
1835         if (warn)
1836                 pr_cont("NG: Some tests are failed. Please check them.\n");
1837         else
1838                 pr_cont("OK\n");
1839         return 0;
1840 }
1841
1842 late_initcall(kprobe_trace_self_tests_init);
1843
1844 #endif