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 <asm/bitsperlong.h>
  35
  36 #include "trace.h"
  37 #include "trace_output.h"
  38
  39 #define MAX_TRACE_ARGS 128
  40 #define MAX_ARGSTR_LEN 63
  41 #define MAX_EVENT_NAME_LEN 64
  42 #define MAX_STRING_SIZE PATH_MAX
  43 #define KPROBE_EVENT_SYSTEM "kprobes"
  44
  45 /* Reserved field names */
  46 #define FIELD_STRING_IP "__probe_ip"
  47 #define FIELD_STRING_RETIP "__probe_ret_ip"
  48 #define FIELD_STRING_FUNC "__probe_func"
  49
  50 const char *reserved_field_names[] = {
  51         "common_type",
  52         "common_flags",
  53         "common_preempt_count",
  54         "common_pid",
  55         "common_tgid",
  56         "common_lock_depth",
  57         FIELD_STRING_IP,
  58         FIELD_STRING_RETIP,
  59         FIELD_STRING_FUNC,
  60 };
  61
  62 /* Printing function type */
  63 typedef int (*print_type_func_t)(struct trace_seq *, const char *, void *,
  64                                  void *);
  65 #define PRINT_TYPE_FUNC_NAME(type)      print_type_##type
  66 #define PRINT_TYPE_FMT_NAME(type)       print_type_format_##type
  67
  68 /* Printing  in basic type function template */
  69 #define DEFINE_BASIC_PRINT_TYPE_FUNC(type, fmt, cast)                   \
  70 static __kprobes int PRINT_TYPE_FUNC_NAME(type)(struct trace_seq *s,    \
  71                                                 const char *name,       \
  72                                                 void *data, void *ent)\
  73 {                                                                       \
  74         return trace_seq_printf(s, " %s=" fmt, name, (cast)*(type *)data);\
  75 }                                                                       \
  76 static const char PRINT_TYPE_FMT_NAME(type)[] = fmt;
  77
  78 DEFINE_BASIC_PRINT_TYPE_FUNC(u8, "%x", unsigned int)
  79 DEFINE_BASIC_PRINT_TYPE_FUNC(u16, "%x", unsigned int)
  80 DEFINE_BASIC_PRINT_TYPE_FUNC(u32, "%lx", unsigned long)
  81 DEFINE_BASIC_PRINT_TYPE_FUNC(u64, "%llx", unsigned long long)
  82 DEFINE_BASIC_PRINT_TYPE_FUNC(s8, "%d", int)
  83 DEFINE_BASIC_PRINT_TYPE_FUNC(s16, "%d", int)
  84 DEFINE_BASIC_PRINT_TYPE_FUNC(s32, "%ld", long)
  85 DEFINE_BASIC_PRINT_TYPE_FUNC(s64, "%lld", long long)
  86
  87 /* data_rloc: data relative location, compatible with u32 */
  88 #define make_data_rloc(len, roffs)      \
  89         (((u32)(len) << 16) | ((u32)(roffs) & 0xffff))
  90 #define get_rloc_len(dl)        ((u32)(dl) >> 16)
  91 #define get_rloc_offs(dl)       ((u32)(dl) & 0xffff)
  92
  93 static inline void *get_rloc_data(u32 *dl)
  94 {
  95         return (u8 *)dl + get_rloc_offs(*dl);
  96 }
  97
  98 /* For data_loc conversion */
  99 static inline void *get_loc_data(u32 *dl, void *ent)
 100 {
 101         return (u8 *)ent + get_rloc_offs(*dl);
 102 }
 103
 104 /*
 105  * Convert data_rloc to data_loc:
 106  *  data_rloc stores the offset from data_rloc itself, but data_loc
 107  *  stores the offset from event entry.
 108  */
 109 #define convert_rloc_to_loc(dl, offs)   ((u32)(dl) + (offs))
 110
 111 /* For defining macros, define string/string_size types */
 112 typedef u32 string;
 113 typedef u32 string_size;
 114
 115 /* Print type function for string type */
 116 static __kprobes int PRINT_TYPE_FUNC_NAME(string)(struct trace_seq *s,
 117                                                   const char *name,
 118                                                   void *data, void *ent)
 119 {
 120         int len = *(u32 *)data >> 16;
 121
 122         if (!len)
 123                 return trace_seq_printf(s, " %s=(fault)", name);
 124         else
 125                 return trace_seq_printf(s, " %s=\"%s\"", name,
 126                                         (const char *)get_loc_data(data, ent));
 127 }
 128 static const char PRINT_TYPE_FMT_NAME(string)[] = "\\\"%s\\\"";
 129
 130 /* Data fetch function type */
 131 typedef void (*fetch_func_t)(struct pt_regs *, void *, void *);
 132
 133 struct fetch_param {
 134         fetch_func_t    fn;
 135         void *data;
 136 };
 137
 138 static __kprobes void call_fetch(struct fetch_param *fprm,
 139                                  struct pt_regs *regs, void *dest)
 140 {
 141         return fprm->fn(regs, fprm->data, dest);
 142 }
 143
 144 #define FETCH_FUNC_NAME(method, type)   fetch_##method##_##type
 145 /*
 146  * Define macro for basic types - we don't need to define s* types, because
 147  * we have to care only about bitwidth at recording time.
 148  */
 149 #define DEFINE_BASIC_FETCH_FUNCS(method) \
 150 DEFINE_FETCH_##method(u8)               \
 151 DEFINE_FETCH_##method(u16)              \
 152 DEFINE_FETCH_##method(u32)              \
 153 DEFINE_FETCH_##method(u64)
 154
 155 #define CHECK_FETCH_FUNCS(method, fn)                   \
 156         (((FETCH_FUNC_NAME(method, u8) == fn) ||        \
 157           (FETCH_FUNC_NAME(method, u16) == fn) ||       \
 158           (FETCH_FUNC_NAME(method, u32) == fn) ||       \
 159           (FETCH_FUNC_NAME(method, u64) == fn) ||       \
 160           (FETCH_FUNC_NAME(method, string) == fn) ||    \
 161           (FETCH_FUNC_NAME(method, string_size) == fn)) \
 162          && (fn != NULL))
 163
 164 /* Data fetch function templates */
 165 #define DEFINE_FETCH_reg(type)                                          \
 166 static __kprobes void FETCH_FUNC_NAME(reg, type)(struct pt_regs *regs,  \
 167                                         void *offset, void *dest)       \
 168 {                                                                       \
 169         *(type *)dest = (type)regs_get_register(regs,                   \
 170                                 (unsigned int)((unsigned long)offset)); \
 171 }
 172 DEFINE_BASIC_FETCH_FUNCS(reg)
 173 /* No string on the register */
 174 #define fetch_reg_string NULL
 175 #define fetch_reg_string_size NULL
 176
 177 #define DEFINE_FETCH_stack(type)                                        \
 178 static __kprobes void FETCH_FUNC_NAME(stack, type)(struct pt_regs *regs,\
 179                                           void *offset, void *dest)     \
 180 {                                                                       \
 181         *(type *)dest = (type)regs_get_kernel_stack_nth(regs,           \
 182                                 (unsigned int)((unsigned long)offset)); \
 183 }
 184 DEFINE_BASIC_FETCH_FUNCS(stack)
 185 /* No string on the stack entry */
 186 #define fetch_stack_string NULL
 187 #define fetch_stack_string_size NULL
 188
 189 #define DEFINE_FETCH_retval(type)                                       \
 190 static __kprobes void FETCH_FUNC_NAME(retval, type)(struct pt_regs *regs,\
 191                                           void *dummy, void *dest)      \
 192 {                                                                       \
 193         *(type *)dest = (type)regs_return_value(regs);                  \
 194 }
 195 DEFINE_BASIC_FETCH_FUNCS(retval)
 196 /* No string on the retval */
 197 #define fetch_retval_string NULL
 198 #define fetch_retval_string_size NULL
 199
 200 #define DEFINE_FETCH_memory(type)                                       \
 201 static __kprobes void FETCH_FUNC_NAME(memory, type)(struct pt_regs *regs,\
 202                                           void *addr, void *dest)       \
 203 {                                                                       \
 204         type retval;                                                    \
 205         if (probe_kernel_address(addr, retval))                         \
 206                 *(type *)dest = 0;                                      \
 207         else                                                            \
 208                 *(type *)dest = retval;                                 \
 209 }
 210 DEFINE_BASIC_FETCH_FUNCS(memory)
 211 /*
 212  * Fetch a null-terminated string. Caller MUST set *(u32 *)dest with max
 213  * length and relative data location.
 214  */
 215 static __kprobes void FETCH_FUNC_NAME(memory, string)(struct pt_regs *regs,
 216                                                       void *addr, void *dest)
 217 {
 218         long ret;
 219         int maxlen = get_rloc_len(*(u32 *)dest);
 220         u8 *dst = get_rloc_data(dest);
 221         u8 *src = addr;
 222         mm_segment_t old_fs = get_fs();
 223         if (!maxlen)
 224                 return;
 225         /*
 226          * Try to get string again, since the string can be changed while
 227          * probing.
 228          */
 229         set_fs(KERNEL_DS);
 230         pagefault_disable();
 231         do
 232                 ret = __copy_from_user_inatomic(dst++, src++, 1);
 233         while (dst[-1] && ret == 0 && src - (u8 *)addr < maxlen);
 234         dst[-1] = '\0';
 235         pagefault_enable();
 236         set_fs(old_fs);
 237
 238         if (ret < 0) {  /* Failed to fetch string */
 239                 ((u8 *)get_rloc_data(dest))[0] = '\0';
 240                 *(u32 *)dest = make_data_rloc(0, get_rloc_offs(*(u32 *)dest));
 241         } else
 242                 *(u32 *)dest = make_data_rloc(src - (u8 *)addr,
 243                                               get_rloc_offs(*(u32 *)dest));
 244 }
 245 /* Return the length of string -- including null terminal byte */
 246 static __kprobes void FETCH_FUNC_NAME(memory, string_size)(struct pt_regs *regs,
 247                                                         void *addr, void *dest)
 248 {
 249         int ret, len = 0;
 250         u8 c;
 251         mm_segment_t old_fs = get_fs();
 252
 253         set_fs(KERNEL_DS);
 254         pagefault_disable();
 255         do {
 256                 ret = __copy_from_user_inatomic(&c, (u8 *)addr + len, 1);
 257                 len++;
 258         } while (c && ret == 0 && len < MAX_STRING_SIZE);
 259         pagefault_enable();
 260         set_fs(old_fs);
 261
 262         if (ret < 0)    /* Failed to check the length */
 263                 *(u32 *)dest = 0;
 264         else
 265                 *(u32 *)dest = len;
 266 }
 267
 268 /* Memory fetching by symbol */
 269 struct symbol_cache {
 270         char *symbol;
 271         long offset;
 272         unsigned long addr;
 273 };
 274
 275 static unsigned long update_symbol_cache(struct symbol_cache *sc)
 276 {
 277         sc->addr = (unsigned long)kallsyms_lookup_name(sc->symbol);
 278         if (sc->addr)
 279                 sc->addr += sc->offset;
 280         return sc->addr;
 281 }
 282
 283 static void free_symbol_cache(struct symbol_cache *sc)
 284 {
 285         kfree(sc->symbol);
 286         kfree(sc);
 287 }
 288
 289 static struct symbol_cache *alloc_symbol_cache(const char *sym, long offset)
 290 {
 291         struct symbol_cache *sc;
 292
 293         if (!sym || strlen(sym) == 0)
 294                 return NULL;
 295         sc = kzalloc(sizeof(struct symbol_cache), GFP_KERNEL);
 296         if (!sc)
 297                 return NULL;
 298
 299         sc->symbol = kstrdup(sym, GFP_KERNEL);
 300         if (!sc->symbol) {
 301                 kfree(sc);
 302                 return NULL;
 303         }
 304         sc->offset = offset;
 305
 306         update_symbol_cache(sc);
 307         return sc;
 308 }
 309
 310 #define DEFINE_FETCH_symbol(type)                                       \
 311 static __kprobes void FETCH_FUNC_NAME(symbol, type)(struct pt_regs *regs,\
 312                                           void *data, void *dest)       \
 313 {                                                                       \
 314         struct symbol_cache *sc = data;                                 \
 315         if (sc->addr)                                                   \
 316                 fetch_memory_##type(regs, (void *)sc->addr, dest);      \
 317         else                                                            \
 318                 *(type *)dest = 0;                                      \
 319 }
 320 DEFINE_BASIC_FETCH_FUNCS(symbol)
 321 DEFINE_FETCH_symbol(string)
 322 DEFINE_FETCH_symbol(string_size)
 323
 324 /* Dereference memory access function */
 325 struct deref_fetch_param {
 326         struct fetch_param orig;
 327         long offset;
 328 };
 329
 330 #define DEFINE_FETCH_deref(type)                                        \
 331 static __kprobes void FETCH_FUNC_NAME(deref, type)(struct pt_regs *regs,\
 332                                             void *data, void *dest)     \
 333 {                                                                       \
 334         struct deref_fetch_param *dprm = data;                          \
 335         unsigned long addr;                                             \
 336         call_fetch(&dprm->orig, regs, &addr);                           \
 337         if (addr) {                                                     \
 338                 addr += dprm->offset;                                   \
 339                 fetch_memory_##type(regs, (void *)addr, dest);          \
 340         } else                                                          \
 341                 *(type *)dest = 0;                                      \
 342 }
 343 DEFINE_BASIC_FETCH_FUNCS(deref)
 344 DEFINE_FETCH_deref(string)
 345 DEFINE_FETCH_deref(string_size)
 346
 347 static __kprobes void free_deref_fetch_param(struct deref_fetch_param *data)
 348 {
 349         if (CHECK_FETCH_FUNCS(deref, data->orig.fn))
 350                 free_deref_fetch_param(data->orig.data);
 351         else if (CHECK_FETCH_FUNCS(symbol, data->orig.fn))
 352                 free_symbol_cache(data->orig.data);
 353         kfree(data);
 354 }
 355
 356 /* Default (unsigned long) fetch type */
 357 #define __DEFAULT_FETCH_TYPE(t) u##t
 358 #define _DEFAULT_FETCH_TYPE(t) __DEFAULT_FETCH_TYPE(t)
 359 #define DEFAULT_FETCH_TYPE _DEFAULT_FETCH_TYPE(BITS_PER_LONG)
 360 #define DEFAULT_FETCH_TYPE_STR __stringify(DEFAULT_FETCH_TYPE)
 361
 362 /* Fetch types */
 363 enum {
 364         FETCH_MTD_reg = 0,
 365         FETCH_MTD_stack,
 366         FETCH_MTD_retval,
 367         FETCH_MTD_memory,
 368         FETCH_MTD_symbol,
 369         FETCH_MTD_deref,
 370         FETCH_MTD_END,
 371 };
 372
 373 #define ASSIGN_FETCH_FUNC(method, type) \
 374         [FETCH_MTD_##method] = FETCH_FUNC_NAME(method, type)
 375
 376 #define __ASSIGN_FETCH_TYPE(_name, ptype, ftype, _size, sign, _fmttype) \
 377         {.name = _name,                         \
 378          .size = _size,                                 \
 379          .is_signed = sign,                             \
 380          .print = PRINT_TYPE_FUNC_NAME(ptype),          \
 381          .fmt = PRINT_TYPE_FMT_NAME(ptype),             \
 382          .fmttype = _fmttype,                           \
 383          .fetch = {                                     \
 384 ASSIGN_FETCH_FUNC(reg, ftype),                          \
 385 ASSIGN_FETCH_FUNC(stack, ftype),                        \
 386 ASSIGN_FETCH_FUNC(retval, ftype),                       \
 387 ASSIGN_FETCH_FUNC(memory, ftype),                       \
 388 ASSIGN_FETCH_FUNC(symbol, ftype),                       \
 389 ASSIGN_FETCH_FUNC(deref, ftype),                        \
 390           }                                             \
 391         }
 392
 393 #define ASSIGN_FETCH_TYPE(ptype, ftype, sign)                   \
 394         __ASSIGN_FETCH_TYPE(#ptype, ptype, ftype, sizeof(ftype), sign, #ptype)
 395
 396 #define FETCH_TYPE_STRING 0
 397 #define FETCH_TYPE_STRSIZE 1
 398
 399 /* Fetch type information table */
 400 static const struct fetch_type {
 401         const char      *name;          /* Name of type */
 402         size_t          size;           /* Byte size of type */
 403         int             is_signed;      /* Signed flag */
 404         print_type_func_t       print;  /* Print functions */
 405         const char      *fmt;           /* Fromat string */
 406         const char      *fmttype;       /* Name in format file */
 407         /* Fetch functions */
 408         fetch_func_t    fetch[FETCH_MTD_END];
 409 } fetch_type_table[] = {
 410         /* Special types */
 411         [FETCH_TYPE_STRING] = __ASSIGN_FETCH_TYPE("string", string, string,
 412                                         sizeof(u32), 1, "__data_loc char[]"),
 413         [FETCH_TYPE_STRSIZE] = __ASSIGN_FETCH_TYPE("string_size", u32,
 414                                         string_size, sizeof(u32), 0, "u32"),
 415         /* Basic types */
 416         ASSIGN_FETCH_TYPE(u8,  u8,  0),
 417         ASSIGN_FETCH_TYPE(u16, u16, 0),
 418         ASSIGN_FETCH_TYPE(u32, u32, 0),
 419         ASSIGN_FETCH_TYPE(u64, u64, 0),
 420         ASSIGN_FETCH_TYPE(s8,  u8,  1),
 421         ASSIGN_FETCH_TYPE(s16, u16, 1),
 422         ASSIGN_FETCH_TYPE(s32, u32, 1),
 423         ASSIGN_FETCH_TYPE(s64, u64, 1),
 424 };
 425
 426 static const struct fetch_type *find_fetch_type(const char *type)
 427 {
 428         int i;
 429
 430         if (!type)
 431                 type = DEFAULT_FETCH_TYPE_STR;
 432
 433         for (i = 0; i < ARRAY_SIZE(fetch_type_table); i++)
 434                 if (strcmp(type, fetch_type_table[i].name) == 0)
 435                         return &fetch_type_table[i];
 436         return NULL;
 437 }
 438
 439 /* Special function : only accept unsigned long */
 440 static __kprobes void fetch_stack_address(struct pt_regs *regs,
 441                                           void *dummy, void *dest)
 442 {
 443         *(unsigned long *)dest = kernel_stack_pointer(regs);
 444 }
 445
 446 static fetch_func_t get_fetch_size_function(const struct fetch_type *type,
 447                                             fetch_func_t orig_fn)
 448 {
 449         int i;
 450
 451         if (type != &fetch_type_table[FETCH_TYPE_STRING])
 452                 return NULL;    /* Only string type needs size function */
 453         for (i = 0; i < FETCH_MTD_END; i++)
 454                 if (type->fetch[i] == orig_fn)
 455                         return fetch_type_table[FETCH_TYPE_STRSIZE].fetch[i];
 456
 457         WARN_ON(1);     /* This should not happen */
 458         return NULL;
 459 }
 460
 461 /**
 462  * Kprobe event core functions
 463  */
 464
 465 struct probe_arg {
 466         struct fetch_param      fetch;
 467         struct fetch_param      fetch_size;
 468         unsigned int            offset; /* Offset from argument entry */
 469         const char              *name;  /* Name of this argument */
 470         const char              *comm;  /* Command of this argument */
 471         const struct fetch_type *type;  /* Type of this argument */
 472 };
 473
 474 /* Flags for trace_probe */
 475 #define TP_FLAG_TRACE   1
 476 #define TP_FLAG_PROFILE 2
 477
 478 struct trace_probe {
 479         struct list_head        list;
 480         struct kretprobe        rp;     /* Use rp.kp for kprobe use */
 481         unsigned long           nhit;
 482         unsigned int            flags;  /* For TP_FLAG_* */
 483         const char              *symbol;        /* symbol name */
 484         struct ftrace_event_class       class;
 485         struct ftrace_event_call        call;
 486         ssize_t                 size;           /* trace entry size */
 487         unsigned int            nr_args;
 488         struct probe_arg        args[];
 489 };
 490
 491 #define SIZEOF_TRACE_PROBE(n)                   \
 492         (offsetof(struct trace_probe, args) +   \
 493         (sizeof(struct probe_arg) * (n)))
 494
 495
 496 static __kprobes int probe_is_return(struct trace_probe *tp)
 497 {
 498         return tp->rp.handler != NULL;
 499 }
 500
 501 static __kprobes const char *probe_symbol(struct trace_probe *tp)
 502 {
 503         return tp->symbol ? tp->symbol : "unknown";
 504 }
 505
 506 static int register_probe_event(struct trace_probe *tp);
 507 static void unregister_probe_event(struct trace_probe *tp);
 508
 509 static DEFINE_MUTEX(probe_lock);
 510 static LIST_HEAD(probe_list);
 511
 512 static int kprobe_dispatcher(struct kprobe *kp, struct pt_regs *regs);
 513 static int kretprobe_dispatcher(struct kretprobe_instance *ri,
 514                                 struct pt_regs *regs);
 515
 516 /* Check the name is good for event/group/fields */
 517 static int is_good_name(const char *name)
 518 {
 519         if (!isalpha(*name) && *name != '_')
 520                 return 0;
 521         while (*++name != '\0') {
 522                 if (!isalpha(*name) && !isdigit(*name) && *name != '_')
 523                         return 0;
 524         }
 525         return 1;
 526 }
 527
 528 /*
 529  * Allocate new trace_probe and initialize it (including kprobes).
 530  */
 531 static struct trace_probe *alloc_trace_probe(const char *group,
 532                                              const char *event,
 533                                              void *addr,
 534                                              const char *symbol,
 535                                              unsigned long offs,
 536                                              int nargs, int is_return)
 537 {
 538         struct trace_probe *tp;
 539         int ret = -ENOMEM;
 540
 541         tp = kzalloc(SIZEOF_TRACE_PROBE(nargs), GFP_KERNEL);
 542         if (!tp)
 543                 return ERR_PTR(ret);
 544
 545         if (symbol) {
 546                 tp->symbol = kstrdup(symbol, GFP_KERNEL);
 547                 if (!tp->symbol)
 548                         goto error;
 549                 tp->rp.kp.symbol_name = tp->symbol;
 550                 tp->rp.kp.offset = offs;
 551         } else
 552                 tp->rp.kp.addr = addr;
 553
 554         if (is_return)
 555                 tp->rp.handler = kretprobe_dispatcher;
 556         else
 557                 tp->rp.kp.pre_handler = kprobe_dispatcher;
 558
 559         if (!event || !is_good_name(event)) {
 560                 ret = -EINVAL;
 561                 goto error;
 562         }
 563
 564         tp->call.class = &tp->class;
 565         tp->call.name = kstrdup(event, GFP_KERNEL);
 566         if (!tp->call.name)
 567                 goto error;
 568
 569         if (!group || !is_good_name(group)) {
 570                 ret = -EINVAL;
 571                 goto error;
 572         }
 573
 574         tp->class.system = kstrdup(group, GFP_KERNEL);
 575         if (!tp->class.system)
 576                 goto error;
 577
 578         INIT_LIST_HEAD(&tp->list);
 579         return tp;
 580 error:
 581         kfree(tp->call.name);
 582         kfree(tp->symbol);
 583         kfree(tp);
 584         return ERR_PTR(ret);
 585 }
 586
 587 static void free_probe_arg(struct probe_arg *arg)
 588 {
 589         if (CHECK_FETCH_FUNCS(deref, arg->fetch.fn))
 590                 free_deref_fetch_param(arg->fetch.data);
 591         else if (CHECK_FETCH_FUNCS(symbol, arg->fetch.fn))
 592                 free_symbol_cache(arg->fetch.data);
 593         kfree(arg->name);
 594         kfree(arg->comm);
 595 }
 596
 597 static void free_trace_probe(struct trace_probe *tp)
 598 {
 599         int i;
 600
 601         for (i = 0; i < tp->nr_args; i++)
 602                 free_probe_arg(&tp->args[i]);
 603
 604         kfree(tp->call.class->system);
 605         kfree(tp->call.name);
 606         kfree(tp->symbol);
 607         kfree(tp);
 608 }
 609
 610 static struct trace_probe *find_probe_event(const char *event,
 611                                             const char *group)
 612 {
 613         struct trace_probe *tp;
 614
 615         list_for_each_entry(tp, &probe_list, list)
 616                 if (strcmp(tp->call.name, event) == 0 &&
 617                     strcmp(tp->call.class->system, group) == 0)
 618                         return tp;
 619         return NULL;
 620 }
 621
 622 /* Unregister a trace_probe and probe_event: call with locking probe_lock */
 623 static void unregister_trace_probe(struct trace_probe *tp)
 624 {
 625         if (probe_is_return(tp))
 626                 unregister_kretprobe(&tp->rp);
 627         else
 628                 unregister_kprobe(&tp->rp.kp);
 629         list_del(&tp->list);
 630         unregister_probe_event(tp);
 631 }
 632
 633 /* Register a trace_probe and probe_event */
 634 static int register_trace_probe(struct trace_probe *tp)
 635 {
 636         struct trace_probe *old_tp;
 637         int ret;
 638
 639         mutex_lock(&probe_lock);
 640
 641         /* register as an event */
 642         old_tp = find_probe_event(tp->call.name, tp->call.class->system);
 643         if (old_tp) {
 644                 /* delete old event */
 645                 unregister_trace_probe(old_tp);
 646                 free_trace_probe(old_tp);
 647         }
 648         ret = register_probe_event(tp);
 649         if (ret) {
 650                 pr_warning("Failed to register probe event(%d)\n", ret);
 651                 goto end;
 652         }
 653
 654         tp->rp.kp.flags |= KPROBE_FLAG_DISABLED;
 655         if (probe_is_return(tp))
 656                 ret = register_kretprobe(&tp->rp);
 657         else
 658                 ret = register_kprobe(&tp->rp.kp);
 659
 660         if (ret) {
 661                 pr_warning("Could not insert probe(%d)\n", ret);
 662                 if (ret == -EILSEQ) {
 663                         pr_warning("Probing address(0x%p) is not an "
 664                                    "instruction boundary.\n",
 665                                    tp->rp.kp.addr);
 666                         ret = -EINVAL;
 667                 }
 668                 unregister_probe_event(tp);
 669         } else
 670                 list_add_tail(&tp->list, &probe_list);
 671 end:
 672         mutex_unlock(&probe_lock);
 673         return ret;
 674 }
 675
 676 /* Split symbol and offset. */
 677 static int split_symbol_offset(char *symbol, unsigned long *offset)
 678 {
 679         char *tmp;
 680         int ret;
 681
 682         if (!offset)
 683                 return -EINVAL;
 684
 685         tmp = strchr(symbol, '+');
 686         if (tmp) {
 687                 /* skip sign because strict_strtol doesn't accept '+' */
 688                 ret = strict_strtoul(tmp + 1, 0, offset);
 689                 if (ret)
 690                         return ret;
 691                 *tmp = '\0';
 692         } else
 693                 *offset = 0;
 694         return 0;
 695 }
 696
 697 #define PARAM_MAX_ARGS 16
 698 #define PARAM_MAX_STACK (THREAD_SIZE / sizeof(unsigned long))
 699
 700 static int parse_probe_vars(char *arg, const struct fetch_type *t,
 701                             struct fetch_param *f, int is_return)
 702 {
 703         int ret = 0;
 704         unsigned long param;
 705
 706         if (strcmp(arg, "retval") == 0) {
 707                 if (is_return)
 708                         f->fn = t->fetch[FETCH_MTD_retval];
 709                 else
 710                         ret = -EINVAL;
 711         } else if (strncmp(arg, "stack", 5) == 0) {
 712                 if (arg[5] == '\0') {
 713                         if (strcmp(t->name, DEFAULT_FETCH_TYPE_STR) == 0)
 714                                 f->fn = fetch_stack_address;
 715                         else
 716                                 ret = -EINVAL;
 717                 } else if (isdigit(arg[5])) {
 718                         ret = strict_strtoul(arg + 5, 10, &param);
 719                         if (ret || param > PARAM_MAX_STACK)
 720                                 ret = -EINVAL;
 721                         else {
 722                                 f->fn = t->fetch[FETCH_MTD_stack];
 723                                 f->data = (void *)param;
 724                         }
 725                 } else
 726                         ret = -EINVAL;
 727         } else
 728                 ret = -EINVAL;
 729         return ret;
 730 }
 731
 732 /* Recursive argument parser */
 733 static int __parse_probe_arg(char *arg, const struct fetch_type *t,
 734                              struct fetch_param *f, int is_return)
 735 {
 736         int ret = 0;
 737         unsigned long param;
 738         long offset;
 739         char *tmp;
 740
 741         switch (arg[0]) {
 742         case '$':
 743                 ret = parse_probe_vars(arg + 1, t, f, is_return);
 744                 break;
 745         case '%':       /* named register */
 746                 ret = regs_query_register_offset(arg + 1);
 747                 if (ret >= 0) {
 748                         f->fn = t->fetch[FETCH_MTD_reg];
 749                         f->data = (void *)(unsigned long)ret;
 750                         ret = 0;
 751                 }
 752                 break;
 753         case '@':       /* memory or symbol */
 754                 if (isdigit(arg[1])) {
 755                         ret = strict_strtoul(arg + 1, 0, &param);
 756                         if (ret)
 757                                 break;
 758                         f->fn = t->fetch[FETCH_MTD_memory];
 759                         f->data = (void *)param;
 760                 } else {
 761                         ret = split_symbol_offset(arg + 1, &offset);
 762                         if (ret)
 763                                 break;
 764                         f->data = alloc_symbol_cache(arg + 1, offset);
 765                         if (f->data)
 766                                 f->fn = t->fetch[FETCH_MTD_symbol];
 767                 }
 768                 break;
 769         case '+':       /* deref memory */
 770                 arg++;  /* Skip '+', because strict_strtol() rejects it. */
 771         case '-':
 772                 tmp = strchr(arg, '(');
 773                 if (!tmp)
 774                         break;
 775                 *tmp = '\0';
 776                 ret = strict_strtol(arg, 0, &offset);
 777                 if (ret)
 778                         break;
 779                 arg = tmp + 1;
 780                 tmp = strrchr(arg, ')');
 781                 if (tmp) {
 782                         struct deref_fetch_param *dprm;
 783                         const struct fetch_type *t2 = find_fetch_type(NULL);
 784                         *tmp = '\0';
 785                         dprm = kzalloc(sizeof(struct deref_fetch_param),
 786                                        GFP_KERNEL);
 787                         if (!dprm)
 788                                 return -ENOMEM;
 789                         dprm->offset = offset;
 790                         ret = __parse_probe_arg(arg, t2, &dprm->orig,
 791                                                 is_return);
 792                         if (ret)
 793                                 kfree(dprm);
 794                         else {
 795                                 f->fn = t->fetch[FETCH_MTD_deref];
 796                                 f->data = (void *)dprm;
 797                         }
 798                 }
 799                 break;
 800         }
 801         if (!ret && !f->fn) {   /* Parsed, but do not find fetch method */
 802                 pr_info("%s type has no corresponding fetch method.\n",
 803                         t->name);
 804                 ret = -EINVAL;
 805         }
 806         return ret;
 807 }
 808
 809 /* String length checking wrapper */
 810 static int parse_probe_arg(char *arg, struct trace_probe *tp,
 811                            struct probe_arg *parg, int is_return)
 812 {
 813         const char *t;
 814         int ret;
 815
 816         if (strlen(arg) > MAX_ARGSTR_LEN) {
 817                 pr_info("Argument is too long.: %s\n",  arg);
 818                 return -ENOSPC;
 819         }
 820         parg->comm = kstrdup(arg, GFP_KERNEL);
 821         if (!parg->comm) {
 822                 pr_info("Failed to allocate memory for command '%s'.\n", arg);
 823                 return -ENOMEM;
 824         }
 825         t = strchr(parg->comm, ':');
 826         if (t) {
 827                 arg[t - parg->comm] = '\0';
 828                 t++;
 829         }
 830         parg->type = find_fetch_type(t);
 831         if (!parg->type) {
 832                 pr_info("Unsupported type: %s\n", t);
 833                 return -EINVAL;
 834         }
 835         parg->offset = tp->size;
 836         tp->size += parg->type->size;
 837         ret = __parse_probe_arg(arg, parg->type, &parg->fetch, is_return);
 838         if (ret >= 0) {
 839                 parg->fetch_size.fn = get_fetch_size_function(parg->type,
 840                                                               parg->fetch.fn);
 841                 parg->fetch_size.data = parg->fetch.data;
 842         }
 843         return ret;
 844 }
 845
 846 /* Return 1 if name is reserved or already used by another argument */
 847 static int conflict_field_name(const char *name,
 848                                struct probe_arg *args, int narg)
 849 {
 850         int i;
 851         for (i = 0; i < ARRAY_SIZE(reserved_field_names); i++)
 852                 if (strcmp(reserved_field_names[i], name) == 0)
 853                         return 1;
 854         for (i = 0; i < narg; i++)
 855                 if (strcmp(args[i].name, name) == 0)
 856                         return 1;
 857         return 0;
 858 }
 859
 860 static int create_trace_probe(int argc, char **argv)
 861 {
 862         /*
 863          * Argument syntax:
 864          *  - Add kprobe: p[:[GRP/]EVENT] KSYM[+OFFS]|KADDR [FETCHARGS]
 865          *  - Add kretprobe: r[:[GRP/]EVENT] KSYM[+0] [FETCHARGS]
 866          * Fetch args:
 867          *  $retval     : fetch return value
 868          *  $stack      : fetch stack address
 869          *  $stackN     : fetch Nth of stack (N:0-)
 870          *  @ADDR       : fetch memory at ADDR (ADDR should be in kernel)
 871          *  @SYM[+|-offs] : fetch memory at SYM +|- offs (SYM is a data symbol)
 872          *  %REG        : fetch register REG
 873          * Dereferencing memory fetch:
 874          *  +|-offs(ARG) : fetch memory at ARG +|- offs address.
 875          * Alias name of args:
 876          *  NAME=FETCHARG : set NAME as alias of FETCHARG.
 877          * Type of args:
 878          *  FETCHARG:TYPE : use TYPE instead of unsigned long.
 879          */
 880         struct trace_probe *tp;
 881         int i, ret = 0;
 882         int is_return = 0, is_delete = 0;
 883         char *symbol = NULL, *event = NULL, *group = NULL;
 884         char *arg;
 885         unsigned long offset = 0;
 886         void *addr = NULL;
 887         char buf[MAX_EVENT_NAME_LEN];
 888
 889         /* argc must be >= 1 */
 890         if (argv[0][0] == 'p')
 891                 is_return = 0;
 892         else if (argv[0][0] == 'r')
 893                 is_return = 1;
 894         else if (argv[0][0] == '-')
 895                 is_delete = 1;
 896         else {
 897                 pr_info("Probe definition must be started with 'p', 'r' or"
 898                         " '-'.\n");
 899                 return -EINVAL;
 900         }
 901
 902         if (argv[0][1] == ':') {
 903                 event = &argv[0][2];
 904                 if (strchr(event, '/')) {
 905                         group = event;
 906                         event = strchr(group, '/') + 1;
 907                         event[-1] = '\0';
 908                         if (strlen(group) == 0) {
 909                                 pr_info("Group name is not specified\n");
 910                                 return -EINVAL;
 911                         }
 912                 }
 913                 if (strlen(event) == 0) {
 914                         pr_info("Event name is not specified\n");
 915                         return -EINVAL;
 916                 }
 917         }
 918         if (!group)
 919                 group = KPROBE_EVENT_SYSTEM;
 920
 921         if (is_delete) {
 922                 if (!event) {
 923                         pr_info("Delete command needs an event name.\n");
 924                         return -EINVAL;
 925                 }
 926                 mutex_lock(&probe_lock);
 927                 tp = find_probe_event(event, group);
 928                 if (!tp) {
 929                         mutex_unlock(&probe_lock);
 930                         pr_info("Event %s/%s doesn't exist.\n", group, event);
 931                         return -ENOENT;
 932                 }
 933                 /* delete an event */
 934                 unregister_trace_probe(tp);
 935                 free_trace_probe(tp);
 936                 mutex_unlock(&probe_lock);
 937                 return 0;
 938         }
 939
 940         if (argc < 2) {
 941                 pr_info("Probe point is not specified.\n");
 942                 return -EINVAL;
 943         }
 944         if (isdigit(argv[1][0])) {
 945                 if (is_return) {
 946                         pr_info("Return probe point must be a symbol.\n");
 947                         return -EINVAL;
 948                 }
 949                 /* an address specified */
 950                 ret = strict_strtoul(&argv[1][0], 0, (unsigned long *)&addr);
 951                 if (ret) {
 952                         pr_info("Failed to parse address.\n");
 953                         return ret;
 954                 }
 955         } else {
 956                 /* a symbol specified */
 957                 symbol = argv[1];
 958                 /* TODO: support .init module functions */
 959                 ret = split_symbol_offset(symbol, &offset);
 960                 if (ret) {
 961                         pr_info("Failed to parse symbol.\n");
 962                         return ret;
 963                 }
 964                 if (offset && is_return) {
 965                         pr_info("Return probe must be used without offset.\n");
 966                         return -EINVAL;
 967                 }
 968         }
 969         argc -= 2; argv += 2;
 970
 971         /* setup a probe */
 972         if (!event) {
 973                 /* Make a new event name */
 974                 if (symbol)
 975                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c_%s_%ld",
 976                                  is_return ? 'r' : 'p', symbol, offset);
 977                 else
 978                         snprintf(buf, MAX_EVENT_NAME_LEN, "%c_0x%p",
 979                                  is_return ? 'r' : 'p', addr);
 980                 event = buf;
 981         }
 982         tp = alloc_trace_probe(group, event, addr, symbol, offset, argc,
 983                                is_return);
 984         if (IS_ERR(tp)) {
 985                 pr_info("Failed to allocate trace_probe.(%d)\n",
 986                         (int)PTR_ERR(tp));
 987                 return PTR_ERR(tp);
 988         }
 989
 990         /* parse arguments */
 991         ret = 0;
 992         for (i = 0; i < argc && i < MAX_TRACE_ARGS; i++) {
 993                 /* Increment count for freeing args in error case */
 994                 tp->nr_args++;
 995
 996                 /* Parse argument name */
 997                 arg = strchr(argv[i], '=');
 998                 if (arg) {
 999                         *arg++ = '\0';
1000                         tp->args[i].name = kstrdup(argv[i], GFP_KERNEL);
1001                 } else {
1002                         arg = argv[i];
1003                         /* If argument name is omitted, set "argN" */
1004                         snprintf(buf, MAX_EVENT_NAME_LEN, "arg%d", i + 1);
1005                         tp->args[i].name = kstrdup(buf, GFP_KERNEL);
1006                 }
1007
1008                 if (!tp->args[i].name) {
1009                         pr_info("Failed to allocate argument[%d] name.\n", i);
1010                         ret = -ENOMEM;
1011                         goto error;
1012                 }
1013
1014                 if (!is_good_name(tp->args[i].name)) {
1015                         pr_info("Invalid argument[%d] name: %s\n",
1016                                 i, tp->args[i].name);
1017                         ret = -EINVAL;
1018                         goto error;
1019                 }
1020
1021                 if (conflict_field_name(tp->args[i].name, tp->args, i)) {
1022                         pr_info("Argument[%d] name '%s' conflicts with "
1023                                 "another field.\n", i, argv[i]);
1024                         ret = -EINVAL;
1025                         goto error;
1026                 }
1027
1028                 /* Parse fetch argument */
1029                 ret = parse_probe_arg(arg, tp, &tp->args[i], is_return);
1030                 if (ret) {
1031                         pr_info("Parse error at argument[%d]. (%d)\n", i, ret);
1032                         goto error;
1033                 }
1034         }
1035
1036         ret = register_trace_probe(tp);
1037         if (ret)
1038                 goto error;
1039         return 0;
1040
1041 error:
1042         free_trace_probe(tp);
1043         return ret;
1044 }
1045
1046 static void cleanup_all_probes(void)
1047 {
1048         struct trace_probe *tp;
1049
1050         mutex_lock(&probe_lock);
1051         /* TODO: Use batch unregistration */
1052         while (!list_empty(&probe_list)) {
1053                 tp = list_entry(probe_list.next, struct trace_probe, list);
1054                 unregister_trace_probe(tp);
1055                 free_trace_probe(tp);
1056         }
1057         mutex_unlock(&probe_lock);
1058 }
1059
1060
1061 /* Probes listing interfaces */
1062 static void *probes_seq_start(struct seq_file *m, loff_t *pos)
1063 {
1064         mutex_lock(&probe_lock);
1065         return seq_list_start(&probe_list, *pos);
1066 }
1067
1068 static void *probes_seq_next(struct seq_file *m, void *v, loff_t *pos)
1069 {
1070         return seq_list_next(v, &probe_list, pos);
1071 }
1072
1073 static void probes_seq_stop(struct seq_file *m, void *v)
1074 {
1075         mutex_unlock(&probe_lock);
1076 }
1077
1078 static int probes_seq_show(struct seq_file *m, void *v)
1079 {
1080         struct trace_probe *tp = v;
1081         int i;
1082
1083         seq_printf(m, "%c", probe_is_return(tp) ? 'r' : 'p');
1084         seq_printf(m, ":%s/%s", tp->call.class->system, tp->call.name);
1085
1086         if (!tp->symbol)
1087                 seq_printf(m, " 0x%p", tp->rp.kp.addr);
1088         else if (tp->rp.kp.offset)
1089                 seq_printf(m, " %s+%u", probe_symbol(tp), tp->rp.kp.offset);
1090         else
1091                 seq_printf(m, " %s", probe_symbol(tp));
1092
1093         for (i = 0; i < tp->nr_args; i++)
1094                 seq_printf(m, " %s=%s", tp->args[i].name, tp->args[i].comm);
1095         seq_printf(m, "\n");
1096
1097         return 0;
1098 }
1099
1100 static const struct seq_operations probes_seq_op = {
1101         .start  = probes_seq_start,
1102         .next   = probes_seq_next,
1103         .stop   = probes_seq_stop,
1104         .show   = probes_seq_show
1105 };
1106
1107 static int probes_open(struct inode *inode, struct file *file)
1108 {
1109         if ((file->f_mode & FMODE_WRITE) &&
1110             (file->f_flags & O_TRUNC))
1111                 cleanup_all_probes();
1112
1113         return seq_open(file, &probes_seq_op);
1114 }
1115
1116 static int command_trace_probe(const char *buf)
1117 {
1118         char **argv;
1119         int argc = 0, ret = 0;
1120
1121         argv = argv_split(GFP_KERNEL, buf, &argc);
1122         if (!argv)
1123                 return -ENOMEM;
1124
1125         if (argc)
1126                 ret = create_trace_probe(argc, argv);
1127
1128         argv_free(argv);
1129         return ret;
1130 }
1131
1132 #define WRITE_BUFSIZE 128
1133
1134 static ssize_t probes_write(struct file *file, const char __user *buffer,
1135                             size_t count, loff_t *ppos)
1136 {
1137         char *kbuf, *tmp;
1138         int ret;
1139         size_t done;
1140         size_t size;
1141
1142         kbuf = kmalloc(WRITE_BUFSIZE, GFP_KERNEL);
1143         if (!kbuf)
1144                 return -ENOMEM;
1145
1146         ret = done = 0;
1147         while (done < count) {
1148                 size = count - done;
1149                 if (size >= WRITE_BUFSIZE)
1150                         size = WRITE_BUFSIZE - 1;
1151                 if (copy_from_user(kbuf, buffer + done, size)) {
1152                         ret = -EFAULT;
1153                         goto out;
1154                 }
1155                 kbuf[size] = '\0';
1156                 tmp = strchr(kbuf, '\n');
1157                 if (tmp) {
1158                         *tmp = '\0';
1159                         size = tmp - kbuf + 1;
1160                 } else if (done + size < count) {
1161                         pr_warning("Line length is too long: "
1162                                    "Should be less than %d.", WRITE_BUFSIZE);
1163                         ret = -EINVAL;
1164                         goto out;
1165                 }
1166                 done += size;
1167                 /* Remove comments */
1168                 tmp = strchr(kbuf, '#');
1169                 if (tmp)
1170                         *tmp = '\0';
1171
1172                 ret = command_trace_probe(kbuf);
1173                 if (ret)
1174                         goto out;
1175         }
1176         ret = done;
1177 out:
1178         kfree(kbuf);
1179         return ret;
1180 }
1181
1182 static const struct file_operations kprobe_events_ops = {
1183         .owner          = THIS_MODULE,
1184         .open           = probes_open,
1185         .read           = seq_read,
1186         .llseek         = seq_lseek,
1187         .release        = seq_release,
1188         .write          = probes_write,
1189 };
1190
1191 /* Probes profiling interfaces */
1192 static int probes_profile_seq_show(struct seq_file *m, void *v)
1193 {
1194         struct trace_probe *tp = v;
1195
1196         seq_printf(m, "  %-44s %15lu %15lu\n", tp->call.name, tp->nhit,
1197                    tp->rp.kp.nmissed);
1198
1199         return 0;
1200 }
1201
1202 static const struct seq_operations profile_seq_op = {
1203         .start  = probes_seq_start,
1204         .next   = probes_seq_next,
1205         .stop   = probes_seq_stop,
1206         .show   = probes_profile_seq_show
1207 };
1208
1209 static int profile_open(struct inode *inode, struct file *file)
1210 {
1211         return seq_open(file, &profile_seq_op);
1212 }
1213
1214 static const struct file_operations kprobe_profile_ops = {
1215         .owner          = THIS_MODULE,
1216         .open           = profile_open,
1217         .read           = seq_read,
1218         .llseek         = seq_lseek,
1219         .release        = seq_release,
1220 };
1221
1222 /* Sum up total data length for dynamic arraies (strings) */
1223 static __kprobes int __get_data_size(struct trace_probe *tp,
1224                                      struct pt_regs *regs)
1225 {
1226         int i, ret = 0;
1227         u32 len;
1228
1229         for (i = 0; i < tp->nr_args; i++)
1230                 if (unlikely(tp->args[i].fetch_size.fn)) {
1231                         call_fetch(&tp->args[i].fetch_size, regs, &len);
1232                         ret += len;
1233                 }
1234
1235         return ret;
1236 }
1237
1238 /* Store the value of each argument */
1239 static __kprobes void store_trace_args(int ent_size, struct trace_probe *tp,
1240                                        struct pt_regs *regs,
1241                                        u8 *data, int maxlen)
1242 {
1243         int i;
1244         u32 end = tp->size;
1245         u32 *dl;        /* Data (relative) location */
1246
1247         for (i = 0; i < tp->nr_args; i++) {
1248                 if (unlikely(tp->args[i].fetch_size.fn)) {
1249                         /*
1250                          * First, we set the relative location and
1251                          * maximum data length to *dl
1252                          */
1253                         dl = (u32 *)(data + tp->args[i].offset);
1254                         *dl = make_data_rloc(maxlen, end - tp->args[i].offset);
1255                         /* Then try to fetch string or dynamic array data */
1256                         call_fetch(&tp->args[i].fetch, regs, dl);
1257                         /* Reduce maximum length */
1258                         end += get_rloc_len(*dl);
1259                         maxlen -= get_rloc_len(*dl);
1260                         /* Trick here, convert data_rloc to data_loc */
1261                         *dl = convert_rloc_to_loc(*dl,
1262                                  ent_size + tp->args[i].offset);
1263                 } else
1264                         /* Just fetching data normally */
1265                         call_fetch(&tp->args[i].fetch, regs,
1266                                    data + tp->args[i].offset);
1267         }
1268 }
1269
1270 /* Kprobe handler */
1271 static __kprobes void kprobe_trace_func(struct kprobe *kp, struct pt_regs *regs)
1272 {
1273         struct trace_probe *tp = container_of(kp, struct trace_probe, rp.kp);
1274         struct kprobe_trace_entry_head *entry;
1275         struct ring_buffer_event *event;
1276         struct ring_buffer *buffer;
1277         int size, dsize, pc;
1278         unsigned long irq_flags;
1279         struct ftrace_event_call *call = &tp->call;
1280
1281         tp->nhit++;
1282
1283         local_save_flags(irq_flags);
1284         pc = preempt_count();
1285
1286         dsize = __get_data_size(tp, regs);
1287         size = sizeof(*entry) + tp->size + dsize;
1288
1289         event = trace_current_buffer_lock_reserve(&buffer, call->event.type,
1290                                                   size, irq_flags, pc);
1291         if (!event)
1292                 return;
1293
1294         entry = ring_buffer_event_data(event);
1295         entry->ip = (unsigned long)kp->addr;
1296         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1297
1298         if (!filter_current_check_discard(buffer, call, entry, event))
1299                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
1300 }
1301
1302 /* Kretprobe handler */
1303 static __kprobes void kretprobe_trace_func(struct kretprobe_instance *ri,
1304                                           struct pt_regs *regs)
1305 {
1306         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
1307         struct kretprobe_trace_entry_head *entry;
1308         struct ring_buffer_event *event;
1309         struct ring_buffer *buffer;
1310         int size, pc, dsize;
1311         unsigned long irq_flags;
1312         struct ftrace_event_call *call = &tp->call;
1313
1314         local_save_flags(irq_flags);
1315         pc = preempt_count();
1316
1317         dsize = __get_data_size(tp, regs);
1318         size = sizeof(*entry) + tp->size + dsize;
1319
1320         event = trace_current_buffer_lock_reserve(&buffer, call->event.type,
1321                                                   size, irq_flags, pc);
1322         if (!event)
1323                 return;
1324
1325         entry = ring_buffer_event_data(event);
1326         entry->func = (unsigned long)tp->rp.kp.addr;
1327         entry->ret_ip = (unsigned long)ri->ret_addr;
1328         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1329
1330         if (!filter_current_check_discard(buffer, call, entry, event))
1331                 trace_nowake_buffer_unlock_commit(buffer, event, irq_flags, pc);
1332 }
1333
1334 /* Event entry printers */
1335 enum print_line_t
1336 print_kprobe_event(struct trace_iterator *iter, int flags,
1337                    struct trace_event *event)
1338 {
1339         struct kprobe_trace_entry_head *field;
1340         struct trace_seq *s = &iter->seq;
1341         struct trace_probe *tp;
1342         u8 *data;
1343         int i;
1344
1345         field = (struct kprobe_trace_entry_head *)iter->ent;
1346         tp = container_of(event, struct trace_probe, call.event);
1347
1348         if (!trace_seq_printf(s, "%s: (", tp->call.name))
1349                 goto partial;
1350
1351         if (!seq_print_ip_sym(s, field->ip, flags | TRACE_ITER_SYM_OFFSET))
1352                 goto partial;
1353
1354         if (!trace_seq_puts(s, ")"))
1355                 goto partial;
1356
1357         data = (u8 *)&field[1];
1358         for (i = 0; i < tp->nr_args; i++)
1359                 if (!tp->args[i].type->print(s, tp->args[i].name,
1360                                              data + tp->args[i].offset, field))
1361                         goto partial;
1362
1363         if (!trace_seq_puts(s, "\n"))
1364                 goto partial;
1365
1366         return TRACE_TYPE_HANDLED;
1367 partial:
1368         return TRACE_TYPE_PARTIAL_LINE;
1369 }
1370
1371 enum print_line_t
1372 print_kretprobe_event(struct trace_iterator *iter, int flags,
1373                       struct trace_event *event)
1374 {
1375         struct kretprobe_trace_entry_head *field;
1376         struct trace_seq *s = &iter->seq;
1377         struct trace_probe *tp;
1378         u8 *data;
1379         int i;
1380
1381         field = (struct kretprobe_trace_entry_head *)iter->ent;
1382         tp = container_of(event, struct trace_probe, call.event);
1383
1384         if (!trace_seq_printf(s, "%s: (", tp->call.name))
1385                 goto partial;
1386
1387         if (!seq_print_ip_sym(s, field->ret_ip, flags | TRACE_ITER_SYM_OFFSET))
1388                 goto partial;
1389
1390         if (!trace_seq_puts(s, " <- "))
1391                 goto partial;
1392
1393         if (!seq_print_ip_sym(s, field->func, flags & ~TRACE_ITER_SYM_OFFSET))
1394                 goto partial;
1395
1396         if (!trace_seq_puts(s, ")"))
1397                 goto partial;
1398
1399         data = (u8 *)&field[1];
1400         for (i = 0; i < tp->nr_args; i++)
1401                 if (!tp->args[i].type->print(s, tp->args[i].name,
1402                                              data + tp->args[i].offset, field))
1403                         goto partial;
1404
1405         if (!trace_seq_puts(s, "\n"))
1406                 goto partial;
1407
1408         return TRACE_TYPE_HANDLED;
1409 partial:
1410         return TRACE_TYPE_PARTIAL_LINE;
1411 }
1412
1413 static int probe_event_enable(struct ftrace_event_call *call)
1414 {
1415         struct trace_probe *tp = (struct trace_probe *)call->data;
1416
1417         tp->flags |= TP_FLAG_TRACE;
1418         if (probe_is_return(tp))
1419                 return enable_kretprobe(&tp->rp);
1420         else
1421                 return enable_kprobe(&tp->rp.kp);
1422 }
1423
1424 static void probe_event_disable(struct ftrace_event_call *call)
1425 {
1426         struct trace_probe *tp = (struct trace_probe *)call->data;
1427
1428         tp->flags &= ~TP_FLAG_TRACE;
1429         if (!(tp->flags & (TP_FLAG_TRACE | TP_FLAG_PROFILE))) {
1430                 if (probe_is_return(tp))
1431                         disable_kretprobe(&tp->rp);
1432                 else
1433                         disable_kprobe(&tp->rp.kp);
1434         }
1435 }
1436
1437 #undef DEFINE_FIELD
1438 #define DEFINE_FIELD(type, item, name, is_signed)                       \
1439         do {                                                            \
1440                 ret = trace_define_field(event_call, #type, name,       \
1441                                          offsetof(typeof(field), item), \
1442                                          sizeof(field.item), is_signed, \
1443                                          FILTER_OTHER);                 \
1444                 if (ret)                                                \
1445                         return ret;                                     \
1446         } while (0)
1447
1448 static int kprobe_event_define_fields(struct ftrace_event_call *event_call)
1449 {
1450         int ret, i;
1451         struct kprobe_trace_entry_head field;
1452         struct trace_probe *tp = (struct trace_probe *)event_call->data;
1453
1454         DEFINE_FIELD(unsigned long, ip, FIELD_STRING_IP, 0);
1455         /* Set argument names as fields */
1456         for (i = 0; i < tp->nr_args; i++) {
1457                 ret = trace_define_field(event_call, tp->args[i].type->fmttype,
1458                                          tp->args[i].name,
1459                                          sizeof(field) + tp->args[i].offset,
1460                                          tp->args[i].type->size,
1461                                          tp->args[i].type->is_signed,
1462                                          FILTER_OTHER);
1463                 if (ret)
1464                         return ret;
1465         }
1466         return 0;
1467 }
1468
1469 static int kretprobe_event_define_fields(struct ftrace_event_call *event_call)
1470 {
1471         int ret, i;
1472         struct kretprobe_trace_entry_head field;
1473         struct trace_probe *tp = (struct trace_probe *)event_call->data;
1474
1475         DEFINE_FIELD(unsigned long, func, FIELD_STRING_FUNC, 0);
1476         DEFINE_FIELD(unsigned long, ret_ip, FIELD_STRING_RETIP, 0);
1477         /* Set argument names as fields */
1478         for (i = 0; i < tp->nr_args; i++) {
1479                 ret = trace_define_field(event_call, tp->args[i].type->fmttype,
1480                                          tp->args[i].name,
1481                                          sizeof(field) + tp->args[i].offset,
1482                                          tp->args[i].type->size,
1483                                          tp->args[i].type->is_signed,
1484                                          FILTER_OTHER);
1485                 if (ret)
1486                         return ret;
1487         }
1488         return 0;
1489 }
1490
1491 static int __set_print_fmt(struct trace_probe *tp, char *buf, int len)
1492 {
1493         int i;
1494         int pos = 0;
1495
1496         const char *fmt, *arg;
1497
1498         if (!probe_is_return(tp)) {
1499                 fmt = "(%lx)";
1500                 arg = "REC->" FIELD_STRING_IP;
1501         } else {
1502                 fmt = "(%lx <- %lx)";
1503                 arg = "REC->" FIELD_STRING_FUNC ", REC->" FIELD_STRING_RETIP;
1504         }
1505
1506         /* When len=0, we just calculate the needed length */
1507 #define LEN_OR_ZERO (len ? len - pos : 0)
1508
1509         pos += snprintf(buf + pos, LEN_OR_ZERO, "\"%s", fmt);
1510
1511         for (i = 0; i < tp->nr_args; i++) {
1512                 pos += snprintf(buf + pos, LEN_OR_ZERO, " %s=%s",
1513                                 tp->args[i].name, tp->args[i].type->fmt);
1514         }
1515
1516         pos += snprintf(buf + pos, LEN_OR_ZERO, "\", %s", arg);
1517
1518         for (i = 0; i < tp->nr_args; i++) {
1519                 if (strcmp(tp->args[i].type->name, "string") == 0)
1520                         pos += snprintf(buf + pos, LEN_OR_ZERO,
1521                                         ", __get_str(%s)",
1522                                         tp->args[i].name);
1523                 else
1524                         pos += snprintf(buf + pos, LEN_OR_ZERO, ", REC->%s",
1525                                         tp->args[i].name);
1526         }
1527
1528 #undef LEN_OR_ZERO
1529
1530         /* return the length of print_fmt */
1531         return pos;
1532 }
1533
1534 static int set_print_fmt(struct trace_probe *tp)
1535 {
1536         int len;
1537         char *print_fmt;
1538
1539         /* First: called with 0 length to calculate the needed length */
1540         len = __set_print_fmt(tp, NULL, 0);
1541         print_fmt = kmalloc(len + 1, GFP_KERNEL);
1542         if (!print_fmt)
1543                 return -ENOMEM;
1544
1545         /* Second: actually write the @print_fmt */
1546         __set_print_fmt(tp, print_fmt, len + 1);
1547         tp->call.print_fmt = print_fmt;
1548
1549         return 0;
1550 }
1551
1552 #ifdef CONFIG_PERF_EVENTS
1553
1554 /* Kprobe profile handler */
1555 static __kprobes void kprobe_perf_func(struct kprobe *kp,
1556                                          struct pt_regs *regs)
1557 {
1558         struct trace_probe *tp = container_of(kp, struct trace_probe, rp.kp);
1559         struct ftrace_event_call *call = &tp->call;
1560         struct kprobe_trace_entry_head *entry;
1561         struct hlist_head *head;
1562         int size, __size, dsize;
1563         int rctx;
1564
1565         dsize = __get_data_size(tp, regs);
1566         __size = sizeof(*entry) + tp->size + dsize;
1567         size = ALIGN(__size + sizeof(u32), sizeof(u64));
1568         size -= sizeof(u32);
1569         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
1570                      "profile buffer not large enough"))
1571                 return;
1572
1573         entry = perf_trace_buf_prepare(size, call->event.type, regs, &rctx);
1574         if (!entry)
1575                 return;
1576
1577         entry->ip = (unsigned long)kp->addr;
1578         memset(&entry[1], 0, dsize);
1579         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1580
1581         head = this_cpu_ptr(call->perf_events);
1582         perf_trace_buf_submit(entry, size, rctx, entry->ip, 1, regs, head);
1583 }
1584
1585 /* Kretprobe profile handler */
1586 static __kprobes void kretprobe_perf_func(struct kretprobe_instance *ri,
1587                                             struct pt_regs *regs)
1588 {
1589         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
1590         struct ftrace_event_call *call = &tp->call;
1591         struct kretprobe_trace_entry_head *entry;
1592         struct hlist_head *head;
1593         int size, __size, dsize;
1594         int rctx;
1595
1596         dsize = __get_data_size(tp, regs);
1597         __size = sizeof(*entry) + tp->size + dsize;
1598         size = ALIGN(__size + sizeof(u32), sizeof(u64));
1599         size -= sizeof(u32);
1600         if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
1601                      "profile buffer not large enough"))
1602                 return;
1603
1604         entry = perf_trace_buf_prepare(size, call->event.type, regs, &rctx);
1605         if (!entry)
1606                 return;
1607
1608         entry->func = (unsigned long)tp->rp.kp.addr;
1609         entry->ret_ip = (unsigned long)ri->ret_addr;
1610         store_trace_args(sizeof(*entry), tp, regs, (u8 *)&entry[1], dsize);
1611
1612         head = this_cpu_ptr(call->perf_events);
1613         perf_trace_buf_submit(entry, size, rctx, entry->ret_ip, 1, regs, head);
1614 }
1615
1616 static int probe_perf_enable(struct ftrace_event_call *call)
1617 {
1618         struct trace_probe *tp = (struct trace_probe *)call->data;
1619
1620         tp->flags |= TP_FLAG_PROFILE;
1621
1622         if (probe_is_return(tp))
1623                 return enable_kretprobe(&tp->rp);
1624         else
1625                 return enable_kprobe(&tp->rp.kp);
1626 }
1627
1628 static void probe_perf_disable(struct ftrace_event_call *call)
1629 {
1630         struct trace_probe *tp = (struct trace_probe *)call->data;
1631
1632         tp->flags &= ~TP_FLAG_PROFILE;
1633
1634         if (!(tp->flags & TP_FLAG_TRACE)) {
1635                 if (probe_is_return(tp))
1636                         disable_kretprobe(&tp->rp);
1637                 else
1638                         disable_kprobe(&tp->rp.kp);
1639         }
1640 }
1641 #endif  /* CONFIG_PERF_EVENTS */
1642
1643 static __kprobes
1644 int kprobe_register(struct ftrace_event_call *event, enum trace_reg type)
1645 {
1646         switch (type) {
1647         case TRACE_REG_REGISTER:
1648                 return probe_event_enable(event);
1649         case TRACE_REG_UNREGISTER:
1650                 probe_event_disable(event);
1651                 return 0;
1652
1653 #ifdef CONFIG_PERF_EVENTS
1654         case TRACE_REG_PERF_REGISTER:
1655                 return probe_perf_enable(event);
1656         case TRACE_REG_PERF_UNREGISTER:
1657                 probe_perf_disable(event);
1658                 return 0;
1659 #endif
1660         }
1661         return 0;
1662 }
1663
1664 static __kprobes
1665 int kprobe_dispatcher(struct kprobe *kp, struct pt_regs *regs)
1666 {
1667         struct trace_probe *tp = container_of(kp, struct trace_probe, rp.kp);
1668
1669         if (tp->flags & TP_FLAG_TRACE)
1670                 kprobe_trace_func(kp, regs);
1671 #ifdef CONFIG_PERF_EVENTS
1672         if (tp->flags & TP_FLAG_PROFILE)
1673                 kprobe_perf_func(kp, regs);
1674 #endif
1675         return 0;       /* We don't tweek kernel, so just return 0 */
1676 }
1677
1678 static __kprobes
1679 int kretprobe_dispatcher(struct kretprobe_instance *ri, struct pt_regs *regs)
1680 {
1681         struct trace_probe *tp = container_of(ri->rp, struct trace_probe, rp);
1682
1683         if (tp->flags & TP_FLAG_TRACE)
1684                 kretprobe_trace_func(ri, regs);
1685 #ifdef CONFIG_PERF_EVENTS
1686         if (tp->flags & TP_FLAG_PROFILE)
1687                 kretprobe_perf_func(ri, regs);
1688 #endif
1689         return 0;       /* We don't tweek kernel, so just return 0 */
1690 }
1691
1692 static struct trace_event_functions kretprobe_funcs = {
1693         .trace          = print_kretprobe_event
1694 };
1695
1696 static struct trace_event_functions kprobe_funcs = {
1697         .trace          = print_kprobe_event
1698 };
1699
1700 static int register_probe_event(struct trace_probe *tp)
1701 {
1702         struct ftrace_event_call *call = &tp->call;
1703         int ret;
1704
1705         /* Initialize ftrace_event_call */
1706         INIT_LIST_HEAD(&call->class->fields);
1707         if (probe_is_return(tp)) {
1708                 call->event.funcs = &kretprobe_funcs;
1709                 call->class->define_fields = kretprobe_event_define_fields;
1710         } else {
1711                 call->event.funcs = &kprobe_funcs;
1712                 call->class->define_fields = kprobe_event_define_fields;
1713         }
1714         if (set_print_fmt(tp) < 0)
1715                 return -ENOMEM;
1716         ret = register_ftrace_event(&call->event);
1717         if (!ret) {
1718                 kfree(call->print_fmt);
1719                 return -ENODEV;
1720         }
1721         call->flags = 0;
1722         call->class->reg = kprobe_register;
1723         call->data = tp;
1724         ret = trace_add_event_call(call);
1725         if (ret) {
1726                 pr_info("Failed to register kprobe event: %s\n", call->name);
1727                 kfree(call->print_fmt);
1728                 unregister_ftrace_event(&call->event);
1729         }
1730         return ret;
1731 }
1732
1733 static void unregister_probe_event(struct trace_probe *tp)
1734 {
1735         /* tp->event is unregistered in trace_remove_event_call() */
1736         trace_remove_event_call(&tp->call);
1737         kfree(tp->call.print_fmt);
1738 }
1739
1740 /* Make a debugfs interface for controling probe points */
1741 static __init int init_kprobe_trace(void)
1742 {
1743         struct dentry *d_tracer;
1744         struct dentry *entry;
1745
1746         d_tracer = tracing_init_dentry();
1747         if (!d_tracer)
1748                 return 0;
1749
1750         entry = debugfs_create_file("kprobe_events", 0644, d_tracer,
1751                                     NULL, &kprobe_events_ops);
1752
1753         /* Event list interface */
1754         if (!entry)
1755                 pr_warning("Could not create debugfs "
1756                            "'kprobe_events' entry\n");
1757
1758         /* Profile interface */
1759         entry = debugfs_create_file("kprobe_profile", 0444, d_tracer,
1760                                     NULL, &kprobe_profile_ops);
1761
1762         if (!entry)
1763                 pr_warning("Could not create debugfs "
1764                            "'kprobe_profile' entry\n");
1765         return 0;
1766 }
1767 fs_initcall(init_kprobe_trace);
1768
1769
1770 #ifdef CONFIG_FTRACE_STARTUP_TEST
1771
1772 static int kprobe_trace_selftest_target(int a1, int a2, int a3,
1773                                         int a4, int a5, int a6)
1774 {
1775         return a1 + a2 + a3 + a4 + a5 + a6;
1776 }
1777
1778 static __init int kprobe_trace_self_tests_init(void)
1779 {
1780         int ret, warn = 0;
1781         int (*target)(int, int, int, int, int, int);
1782         struct trace_probe *tp;
1783
1784         target = kprobe_trace_selftest_target;
1785
1786         pr_info("Testing kprobe tracing: ");
1787
1788         ret = command_trace_probe("p:testprobe kprobe_trace_selftest_target "
1789                                   "$stack $stack0 +0($stack)");
1790         if (WARN_ON_ONCE(ret)) {
1791                 pr_warning("error on probing function entry.\n");
1792                 warn++;
1793         } else {
1794                 /* Enable trace point */
1795                 tp = find_probe_event("testprobe", KPROBE_EVENT_SYSTEM);
1796                 if (WARN_ON_ONCE(tp == NULL)) {
1797                         pr_warning("error on getting new probe.\n");
1798                         warn++;
1799                 } else
1800                         probe_event_enable(&tp->call);
1801         }
1802
1803         ret = command_trace_probe("r:testprobe2 kprobe_trace_selftest_target "
1804                                   "$retval");
1805         if (WARN_ON_ONCE(ret)) {
1806                 pr_warning("error on probing function return.\n");
1807                 warn++;
1808         } else {
1809                 /* Enable trace point */
1810                 tp = find_probe_event("testprobe2", KPROBE_EVENT_SYSTEM);
1811                 if (WARN_ON_ONCE(tp == NULL)) {
1812                         pr_warning("error on getting new probe.\n");
1813                         warn++;
1814                 } else
1815                         probe_event_enable(&tp->call);
1816         }
1817
1818         if (warn)
1819                 goto end;
1820
1821         ret = target(1, 2, 3, 4, 5, 6);
1822
1823         ret = command_trace_probe("-:testprobe");
1824         if (WARN_ON_ONCE(ret)) {
1825                 pr_warning("error on deleting a probe.\n");
1826                 warn++;
1827         }
1828
1829         ret = command_trace_probe("-:testprobe2");
1830         if (WARN_ON_ONCE(ret)) {
1831                 pr_warning("error on deleting a probe.\n");
1832                 warn++;
1833         }
1834
1835 end:
1836         cleanup_all_probes();
1837         if (warn)
1838                 pr_cont("NG: Some tests are failed. Please check them.\n");
1839         else
1840                 pr_cont("OK\n");
1841         return 0;
1842 }
1843
1844 late_initcall(kprobe_trace_self_tests_init);
1845
1846 #endif