mono/profiler/proflog.c

   1 /*
   2  * proflog.c: mono log profiler
   3  *
   4  * Author:
   5  *   Paolo Molaro (lupus@ximian.com)
   6  *
   7  * Copyright 2010 Novell, Inc (http://www.novell.com)
   8  */
   9
  10 #include <config.h>
  11 #include <mono/metadata/profiler.h>
  12 #include <mono/metadata/threads.h>
  13 #include <mono/metadata/mono-gc.h>
  14 #include <mono/metadata/debug-helpers.h>
  15 #include <stdlib.h>
  16 #include <string.h>
  17 #include <assert.h>
  18 #ifdef HAVE_UNISTD_H
  19 #include <unistd.h>
  20 #endif
  21 #include <fcntl.h>
  22 #include <errno.h>
  23 #if defined(HOST_WIN32) || defined(DISABLE_SOCKETS)
  24 #define DISABLE_HELPER_THREAD 1
  25 #endif
  26
  27 #ifndef _GNU_SOURCE
  28 #define _GNU_SOURCE
  29 #endif
  30 #ifdef HAVE_DLFCN_H
  31 #include <dlfcn.h>
  32 #endif
  33 #ifdef HAVE_EXECINFO_H
  34 #include <execinfo.h>
  35 #endif
  36 #ifdef HAVE_LINK_H
  37 #include <link.h>
  38 #endif
  39
  40 #ifndef DISABLE_HELPER_THREAD
  41 #include <sys/types.h>
  42 #include <sys/socket.h>
  43 #include <netinet/in.h>
  44 #include <sys/select.h>
  45 #endif
  46
  47 #ifdef HOST_WIN32
  48 #include <windows.h>
  49 #else
  50 #include <pthread.h>
  51 #endif
  52
  53 #ifdef HAVE_SYS_STAT_H
  54 #include <sys/stat.h>
  55 #endif
  56
  57 #include "utils.c"
  58 #include "proflog.h"
  59
  60 #if defined (HAVE_SYS_ZLIB)
  61 #include <zlib.h>
  62 #endif
  63
  64 /* the architecture needs a memory fence */
  65 #if defined(__linux__) && (defined(__i386__) || defined(__x86_64__))
  66 #include "perf_event.h"
  67 #define USE_PERF_EVENTS 1
  68 static int read_perf_mmap (MonoProfiler* prof);
  69 #endif
  70
  71 #define BUFFER_SIZE (4096 * 16)
  72 static int nocalls = 0;
  73 static int notraces = 0;
  74 static int use_zip = 0;
  75 static int do_report = 0;
  76 static int do_heap_shot = 0;
  77 static int max_call_depth = 100;
  78 static int runtime_inited = 0;
  79 static int command_port = 0;
  80 static int heapshot_requested = 0;
  81 static int sample_type = 0;
  82 static int sample_freq = 0;
  83 static int do_mono_sample = 0;
  84 static int in_shutdown = 0;
  85 static int do_debug = 0;
  86
  87 /* For linux compile with:
  88  * gcc -fPIC -shared -o libmono-profiler-log.so proflog.c utils.c -Wall -g -lz `pkg-config --cflags --libs mono-2`
  89  * gcc -o mprof-report decode.c utils.c -Wall -g -lz -lrt -lpthread `pkg-config --cflags mono-2`
  90  *
  91  * For osx compile with:
  92  * gcc -m32 -Dmono_free=free shared -o libmono-profiler-log.dylib proflog.c utils.c -Wall -g -lz `pkg-config --cflags mono-2` -undefined suppress -flat_namespace
  93  * gcc -m32 -o mprof-report decode.c utils.c -Wall -g -lz -lrt -lpthread `pkg-config --cflags mono-2`
  94  *
  95  * Install with:
  96  * sudo cp mprof-report /usr/local/bin
  97  * sudo cp libmono-profiler-log.so /usr/local/lib
  98  * sudo ldconfig
  99  */
 100
 101 typedef struct _LogBuffer LogBuffer;
 102
 103 /*
 104  * file format:
 105  * [header] [buffer]*
 106  *
 107  * The file is composed by a header followed by 0 or more buffers.
 108  * Each buffer contains events that happened on a thread: for a given thread
 109  * buffers that appear later in the file are guaranteed to contain events
 110  * that happened later in time. Buffers from separate threads could be interleaved,
 111  * though.
 112  * Buffers are not required to be aligned.
 113  *
 114  * header format:
 115  * [id: 4 bytes] constant value: LOG_HEADER_ID
 116  * [major: 1 byte] [minor: 1 byte] major and minor version of the log profiler
 117  * [format: 1 byte] version of the data format for the rest of the file
 118  * [ptrsize: 1 byte] size in bytes of a pointer in the profiled program
 119  * [startup time: 8 bytes] time in milliseconds since the unix epoch when the program started
 120  * [timer overhead: 4 bytes] approximate overhead in nanoseconds of the timer
 121  * [flags: 4 bytes] file format flags, should be 0 for now
 122  * [pid: 4 bytes] pid of the profiled process
 123  * [port: 2 bytes] tcp port for server if != 0
 124  * [sysid: 2 bytes] operating system and architecture identifier
 125  *
 126  * The multiple byte integers are in little-endian format.
 127  *
 128  * buffer format:
 129  * [buffer header] [event]*
 130  * Buffers have a fixed-size header followed by 0 or more bytes of event data.
 131  * Timing information and other values in the event data are usually stored
 132  * as uleb128 or sleb128 integers. To save space, as noted for each item below,
 133  * some data is represented as a difference between the actual value and
 134  * either the last value of the same type (like for timing information) or
 135  * as the difference from a value stored in a buffer header.
 136  *
 137  * For timing information the data is stored as uleb128, since timing
 138  * increases in a monotonic way in each thread: the value is the number of
 139  * nanoseconds to add to the last seen timing data in a buffer. The first value
 140  * in a buffer will be calculated from the time_base field in the buffer head.
 141  *
 142  * Object or heap sizes are stored as uleb128.
 143  * Pointer differences are stored as sleb128, instead.
 144  *
 145  * If an unexpected value is found, the rest of the buffer should be ignored,
 146  * as generally the later values need the former to be interpreted correctly.
 147  *
 148  * buffer header format:
 149  * [bufid: 4 bytes] constant value: BUF_ID
 150  * [len: 4 bytes] size of the data following the buffer header
 151  * [time_base: 8 bytes] time base in nanoseconds since an unspecified epoch
 152  * [ptr_base: 8 bytes] base value for pointers
 153  * [obj_base: 8 bytes] base value for object addresses
 154  * [thread id: 8 bytes] system-specific thread ID (pthread_t for example)
 155  * [method_base: 8 bytes] base value for MonoMethod pointers
 156  *
 157  * event format:
 158  * [extended info: upper 4 bits] [type: lower 4 bits] [data]*
 159  * The data that follows depends on type and the extended info.
 160  * Type is one of the enum values in proflog.h: TYPE_ALLOC, TYPE_GC,
 161  * TYPE_METADATA, TYPE_METHOD, TYPE_EXCEPTION, TYPE_MONITOR, TYPE_HEAP.
 162  * The extended info bits are interpreted based on type, see
 163  * each individual event description below.
 164  * strings are represented as a 0-terminated utf8 sequence.
 165  *
 166  * backtrace format:
 167  * [flags: uleb128] must be 0
 168  * [num: uleb128] number of frames following
 169  * [frame: sleb128]* num MonoMethod pointers as differences from ptr_base
 170  *
 171  * type alloc format:
 172  * type: TYPE_ALLOC
 173  * exinfo: flags: TYPE_ALLOC_BT
 174  * [time diff: uleb128] nanoseconds since last timing
 175  * [ptr: sleb128] class as a byte difference from ptr_base
 176  * [obj: sleb128] object address as a byte difference from obj_base
 177  * [size: uleb128] size of the object in the heap
 178  * If the TYPE_ALLOC_BT flag is set, a backtrace follows.
 179  *
 180  * type GC format:
 181  * type: TYPE_GC
 182  * exinfo: one of TYPE_GC_EVENT, TYPE_GC_RESIZE, TYPE_GC_MOVE, TYPE_GC_HANDLE_CREATED,
 183  * TYPE_GC_HANDLE_DESTROYED
 184  * [time diff: uleb128] nanoseconds since last timing
 185  * if exinfo == TYPE_GC_RESIZE
 186  *      [heap_size: uleb128] new heap size
 187  * if exinfo == TYPE_GC_EVENT
 188  *      [event type: uleb128] GC event (MONO_GC_EVENT_* from profiler.h)
 189  *      [generation: uleb128] GC generation event refers to
 190  * if exinfo == TYPE_GC_MOVE
 191  *      [num_objects: uleb128] number of object moves that follow
 192  *      [objaddr: sleb128]+ num_objects object pointer differences from obj_base
 193  *      num is always an even number: the even items are the old
 194  *      addresses, the odd numbers are the respective new object addresses
 195  * if exinfo == TYPE_GC_HANDLE_CREATED
 196  *      [handle_type: uleb128] GC handle type (System.Runtime.InteropServices.GCHandleType)
 197  *      upper bits reserved as flags
 198  *      [handle: uleb128] GC handle value
 199  *      [objaddr: sleb128] object pointer differences from obj_base
 200  * if exinfo == TYPE_GC_HANDLE_DESTROYED
 201  *      [handle_type: uleb128] GC handle type (System.Runtime.InteropServices.GCHandleType)
 202  *      upper bits reserved as flags
 203  *      [handle: uleb128] GC handle value
 204  *
 205  * type metadata format:
 206  * type: TYPE_METADATA
 207  * exinfo: flags: TYPE_LOAD_ERR
 208  * [time diff: uleb128] nanoseconds since last timing
 209  * [mtype: byte] metadata type, one of: TYPE_CLASS, TYPE_IMAGE, TYPE_ASSEMBLY, TYPE_DOMAIN,
 210  * TYPE_THREAD
 211  * [pointer: sleb128] pointer of the metadata type depending on mtype
 212  * if mtype == TYPE_CLASS
 213  *      [image: sleb128] MonoImage* as a pointer difference from ptr_base
 214  *      [flags: uleb128] must be 0
 215  *      [name: string] full class name
 216  * if mtype == TYPE_IMAGE
 217  *      [flags: uleb128] must be 0
 218  *      [name: string] image file name
 219  * if mtype == TYPE_THREAD
 220  *      [flags: uleb128] must be 0
 221  *      [name: string] thread name
 222  *
 223  * type method format:
 224  * type: TYPE_METHOD
 225  * exinfo: one of: TYPE_LEAVE, TYPE_ENTER, TYPE_EXC_LEAVE, TYPE_JIT
 226  * [time diff: uleb128] nanoseconds since last timing
 227  * [method: sleb128] MonoMethod* as a pointer difference from the last such
 228  * pointer or the buffer method_base
 229  * if exinfo == TYPE_JIT
 230  *      [code address: sleb128] pointer to the native code as a diff from ptr_base
 231  *      [code size: uleb128] size of the generated code
 232  *      [name: string] full method name
 233  *
 234  * type exception format:
 235  * type: TYPE_EXCEPTION
 236  * exinfo: TYPE_EXCEPTION_BT flag and one of: TYPE_THROW, TYPE_CLAUSE
 237  * [time diff: uleb128] nanoseconds since last timing
 238  * if exinfo.low3bits == TYPE_CLAUSE
 239  *      [clause type: uleb128] finally/catch/fault/filter
 240  *      [clause num: uleb128] the clause number in the method header
 241  *      [method: sleb128] MonoMethod* as a pointer difference from the last such
 242  *      pointer or the buffer method_base
 243  * if exinfo.low3bits == TYPE_THROW
 244  *      [object: sleb128] the object that was thrown as a difference from obj_base
 245  *      If the TYPE_EXCEPTION_BT flag is set, a backtrace follows.
 246  *
 247  * type monitor format:
 248  * type: TYPE_MONITOR
 249  * exinfo: TYPE_MONITOR_BT flag and one of: MONO_PROFILER_MONITOR_(CONTENTION|FAIL|DONE)
 250  * [time diff: uleb128] nanoseconds since last timing
 251  * [object: sleb128] the lock object as a difference from obj_base
 252  * if exinfo.low3bits == MONO_PROFILER_MONITOR_CONTENTION
 253  *      If the TYPE_MONITOR_BT flag is set, a backtrace follows.
 254  *
 255  * type heap format
 256  * type: TYPE_HEAP
 257  * exinfo: one of TYPE_HEAP_START, TYPE_HEAP_END, TYPE_HEAP_OBJECT, TYPE_HEAP_ROOT
 258  * if exinfo == TYPE_HEAP_START
 259  *      [time diff: uleb128] nanoseconds since last timing
 260  * if exinfo == TYPE_HEAP_END
 261  *      [time diff: uleb128] nanoseconds since last timing
 262  * if exinfo == TYPE_HEAP_OBJECT
 263  *      [object: sleb128] the object as a difference from obj_base
 264  *      [class: sleb128] the object MonoClass* as a difference from ptr_base
 265  *      [size: uleb128] size of the object on the heap
 266  *      [num_refs: uleb128] number of object references
 267  *      if (format version > 1) each referenced objref is preceded by a
 268  *      uleb128 encoded offset: the first offset is from the object address
 269  *      and each next offset is relative to the previous one
 270  *      [objrefs: sleb128]+ object referenced as a difference from obj_base
 271  *      The same object can appear multiple times, but only the first time
 272  *      with size != 0: in the other cases this data will only be used to
 273  *      provide additional referenced objects.
 274  * if exinfo == TYPE_HEAP_ROOT
 275  *      [num_roots: uleb128] number of root references
 276  *      [num_gc: uleb128] number of major gcs
 277  *      [object: sleb128] the object as a difference from obj_base
 278  *      [root_type: uleb128] the root_type: MonoProfileGCRootType (profiler.h)
 279  *      [extra_info: uleb128] the extra_info value
 280  *      object, root_type_extra_info are repeated num_roots times
 281  *
 282  * type sample format
 283  * type: TYPE_SAMPLE
 284  * exinfo: one of TYPE_SAMPLE_HIT, TYPE_SAMPLE_USYM, TYPE_SAMPLE_UBIN
 285  * if exinfo == TYPE_SAMPLE_HIT
 286  *      [sample_type: uleb128] type of sample (SAMPLE_*)
 287  *      [timestamp: uleb128] nanoseconds since startup (note: different from other timestamps!)
 288  *      [count: uleb128] number of following instruction addresses
 289  *      [ip: sleb128]* instruction pointer as difference from ptr_base
 290  * if exinfo == TYPE_SAMPLE_USYM
 291  *      [address: sleb128] symbol address as a difference from ptr_base
 292  *      [size: uleb128] symbol size (may be 0 if unknown)
 293  *      [name: string] symbol name
 294  * if exinfo == TYPE_SAMPLE_UBIN
 295  *      [time diff: uleb128] nanoseconds since last timing
 296  *      [address: sleb128] address where binary has been loaded
 297  *      [offset: uleb128] file offset of mapping (the same file can be mapped multiple times)
 298  *      [size: uleb128] memory size
 299  *      [name: string] binary name
 300  *
 301  */
 302 struct _LogBuffer {
 303         LogBuffer *next;
 304         uint64_t time_base;
 305         uint64_t last_time;
 306         uintptr_t ptr_base;
 307         uintptr_t method_base;
 308         uintptr_t last_method;
 309         uintptr_t obj_base;
 310         uintptr_t thread_id;
 311         unsigned char* data_end;
 312         unsigned char* data;
 313         int locked;
 314         int size;
 315         int call_depth;
 316         unsigned char buf [1];
 317 };
 318
 319 #define ENTER_LOG(lb,str) if ((lb)->locked) {write(2, str, strlen(str)); write(2, "\n", 1);return;} else {(lb)->locked++;}
 320 #define EXIT_LOG(lb) (lb)->locked--;
 321
 322 typedef struct _StatBuffer StatBuffer;
 323 struct _StatBuffer {
 324         StatBuffer *next;
 325         uintptr_t size;
 326         uintptr_t *data_end;
 327         uintptr_t *data;
 328         uintptr_t buf [1];
 329 };
 330
 331 typedef struct _BinaryObject BinaryObject;
 332
 333 struct _BinaryObject {
 334         BinaryObject *next;
 335         void *addr;
 336         char *name;
 337 };
 338
 339 struct _MonoProfiler {
 340         LogBuffer *buffers;
 341         StatBuffer *stat_buffers;
 342         FILE* file;
 343 #if defined (HAVE_SYS_ZLIB)
 344         gzFile *gzfile;
 345 #endif
 346         uint64_t startup_time;
 347         int pipe_output;
 348         int last_gc_gen_started;
 349         int command_port;
 350         int server_socket;
 351         int pipes [2];
 352 #ifndef HOST_WIN32
 353         pthread_t helper_thread;
 354 #endif
 355         BinaryObject *binary_objects;
 356 };
 357
 358 #ifdef HOST_WIN32
 359 #define TLS_SET(x,y) TlsSetValue(x, y)
 360 #define TLS_GET(x) ((LogBuffer *) TlsGetValue(x))
 361 #define TLS_INIT(x) x = TlsAlloc ()
 362 static int tlsbuffer;
 363 #elif HAVE_KW_THREAD
 364 #define TLS_SET(x,y) x = y
 365 #define TLS_GET(x) x
 366 #define TLS_INIT(x)
 367 static __thread LogBuffer* tlsbuffer = NULL;
 368 #else
 369 #define TLS_SET(x,y) pthread_setspecific(x, y)
 370 #define TLS_GET(x) ((LogBuffer *) pthread_getspecific(x))
 371 #define TLS_INIT(x) pthread_key_create(&x, NULL)
 372 static pthread_key_t tlsbuffer;
 373 #endif
 374
 375 static char*
 376 pstrdup (const char *s)
 377 {
 378         int len = strlen (s) + 1;
 379         char *p = malloc (len);
 380         memcpy (p, s, len);
 381         return p;
 382 }
 383
 384 static StatBuffer*
 385 create_stat_buffer (void)
 386 {
 387         StatBuffer* buf = alloc_buffer (BUFFER_SIZE);
 388         buf->size = BUFFER_SIZE;
 389         buf->data_end = (uintptr_t*)((unsigned char*)buf + buf->size);
 390         buf->data = buf->buf;
 391         return buf;
 392 }
 393
 394 static LogBuffer*
 395 create_buffer (void)
 396 {
 397         LogBuffer* buf = alloc_buffer (BUFFER_SIZE);
 398         buf->size = BUFFER_SIZE;
 399         buf->time_base = current_time ();
 400         buf->last_time = buf->time_base;
 401         buf->data_end = (unsigned char*)buf + buf->size;
 402         buf->data = buf->buf;
 403         return buf;
 404 }
 405
 406 static void
 407 init_thread (void)
 408 {
 409         LogBuffer *logbuffer;
 410         if (TLS_GET (tlsbuffer))
 411                 return;
 412         logbuffer = create_buffer ();
 413         TLS_SET (tlsbuffer, logbuffer);
 414         logbuffer->thread_id = thread_id ();
 415         //printf ("thread %p at time %llu\n", (void*)logbuffer->thread_id, logbuffer->time_base);
 416 }
 417
 418 static LogBuffer*
 419 ensure_logbuf (int bytes)
 420 {
 421         LogBuffer *old = TLS_GET (tlsbuffer);
 422         if (old && old->data + bytes + 100 < old->data_end)
 423                 return old;
 424         TLS_SET (tlsbuffer, NULL);
 425         init_thread ();
 426         TLS_GET (tlsbuffer)->next = old;
 427         if (old)
 428                 TLS_GET (tlsbuffer)->call_depth = old->call_depth;
 429         //printf ("new logbuffer\n");
 430         return TLS_GET (tlsbuffer);
 431 }
 432
 433 static void
 434 emit_byte (LogBuffer *logbuffer, int value)
 435 {
 436         logbuffer->data [0] = value;
 437         logbuffer->data++;
 438         assert (logbuffer->data <= logbuffer->data_end);
 439 }
 440
 441 static void
 442 emit_value (LogBuffer *logbuffer, int value)
 443 {
 444         encode_uleb128 (value, logbuffer->data, &logbuffer->data);
 445         assert (logbuffer->data <= logbuffer->data_end);
 446 }
 447
 448 static void
 449 emit_time (LogBuffer *logbuffer, uint64_t value)
 450 {
 451         uint64_t tdiff = value - logbuffer->last_time;
 452         unsigned char *p;
 453         if (value < logbuffer->last_time)
 454                 printf ("time went backwards\n");
 455         //if (tdiff > 1000000)
 456         //      printf ("large time offset: %llu\n", tdiff);
 457         p = logbuffer->data;
 458         encode_uleb128 (tdiff, logbuffer->data, &logbuffer->data);
 459         /*if (tdiff != decode_uleb128 (p, &p))
 460                 printf ("incorrect encoding: %llu\n", tdiff);*/
 461         logbuffer->last_time = value;
 462         assert (logbuffer->data <= logbuffer->data_end);
 463 }
 464
 465 static void
 466 emit_svalue (LogBuffer *logbuffer, int64_t value)
 467 {
 468         encode_sleb128 (value, logbuffer->data, &logbuffer->data);
 469         assert (logbuffer->data <= logbuffer->data_end);
 470 }
 471
 472 static void
 473 emit_uvalue (LogBuffer *logbuffer, uint64_t value)
 474 {
 475         encode_uleb128 (value, logbuffer->data, &logbuffer->data);
 476         assert (logbuffer->data <= logbuffer->data_end);
 477 }
 478
 479 static void
 480 emit_ptr (LogBuffer *logbuffer, void *ptr)
 481 {
 482         if (!logbuffer->ptr_base)
 483                 logbuffer->ptr_base = (uintptr_t)ptr;
 484         emit_svalue (logbuffer, (intptr_t)ptr - logbuffer->ptr_base);
 485         assert (logbuffer->data <= logbuffer->data_end);
 486 }
 487
 488 static void
 489 emit_method (LogBuffer *logbuffer, void *method)
 490 {
 491         if (!logbuffer->method_base) {
 492                 logbuffer->method_base = (intptr_t)method;
 493                 logbuffer->last_method = (intptr_t)method;
 494         }
 495         encode_sleb128 ((intptr_t)((char*)method - (char*)logbuffer->last_method), logbuffer->data, &logbuffer->data);
 496         logbuffer->last_method = (intptr_t)method;
 497         assert (logbuffer->data <= logbuffer->data_end);
 498 }
 499
 500 static void
 501 emit_obj (LogBuffer *logbuffer, void *ptr)
 502 {
 503         if (!logbuffer->obj_base)
 504                 logbuffer->obj_base = (uintptr_t)ptr >> 3;
 505         emit_svalue (logbuffer, ((uintptr_t)ptr >> 3) - logbuffer->obj_base);
 506         assert (logbuffer->data <= logbuffer->data_end);
 507 }
 508
 509 static char*
 510 write_int16 (char *buf, int32_t value)
 511 {
 512         int i;
 513         for (i = 0; i < 2; ++i) {
 514                 buf [i] = value;
 515                 value >>= 8;
 516         }
 517         return buf + 2;
 518 }
 519
 520 static char*
 521 write_int32 (char *buf, int32_t value)
 522 {
 523         int i;
 524         for (i = 0; i < 4; ++i) {
 525                 buf [i] = value;
 526                 value >>= 8;
 527         }
 528         return buf + 4;
 529 }
 530
 531 static char*
 532 write_int64 (char *buf, int64_t value)
 533 {
 534         int i;
 535         for (i = 0; i < 8; ++i) {
 536                 buf [i] = value;
 537                 value >>= 8;
 538         }
 539         return buf + 8;
 540 }
 541
 542 static void
 543 dump_header (MonoProfiler *profiler)
 544 {
 545         char hbuf [128];
 546         char *p = hbuf;
 547         p = write_int32 (p, LOG_HEADER_ID);
 548         *p++ = LOG_VERSION_MAJOR;
 549         *p++ = LOG_VERSION_MINOR;
 550         *p++ = LOG_DATA_VERSION;
 551         *p++ = sizeof (void*);
 552         p = write_int64 (p, ((uint64_t)time (NULL)) * 1000); /* startup time */
 553         p = write_int32 (p, get_timer_overhead ()); /* timer overhead */
 554         p = write_int32 (p, 0); /* flags */
 555         p = write_int32 (p, process_id ()); /* pid */
 556         p = write_int16 (p, profiler->command_port); /* port */
 557         p = write_int16 (p, 0); /* opsystem */
 558 #if defined (HAVE_SYS_ZLIB)
 559         if (profiler->gzfile) {
 560                 gzwrite (profiler->gzfile, hbuf, p - hbuf);
 561         } else {
 562                 fwrite (hbuf, p - hbuf, 1, profiler->file);
 563         }
 564 #else
 565         fwrite (hbuf, p - hbuf, 1, profiler->file);
 566 #endif
 567 }
 568
 569 static void
 570 dump_buffer (MonoProfiler *profiler, LogBuffer *buf)
 571 {
 572         char hbuf [128];
 573         char *p = hbuf;
 574         if (buf->next)
 575                 dump_buffer (profiler, buf->next);
 576         p = write_int32 (p, BUF_ID);
 577         p = write_int32 (p, buf->data - buf->buf);
 578         p = write_int64 (p, buf->time_base);
 579         p = write_int64 (p, buf->ptr_base);
 580         p = write_int64 (p, buf->obj_base);
 581         p = write_int64 (p, buf->thread_id);
 582         p = write_int64 (p, buf->method_base);
 583 #if defined (HAVE_SYS_ZLIB)
 584         if (profiler->gzfile) {
 585                 gzwrite (profiler->gzfile, hbuf, p - hbuf);
 586                 gzwrite (profiler->gzfile, buf->buf, buf->data - buf->buf);
 587         } else {
 588 #endif
 589                 fwrite (hbuf, p - hbuf, 1, profiler->file);
 590                 fwrite (buf->buf, buf->data - buf->buf, 1, profiler->file);
 591 #if defined (HAVE_SYS_ZLIB)
 592         }
 593 #endif
 594         free_buffer (buf, buf->size);
 595 }
 596
 597 static void
 598 process_requests (MonoProfiler *profiler)
 599 {
 600         if (heapshot_requested)
 601                 mono_gc_collect (mono_gc_max_generation ());
 602 }
 603
 604 static void
 605 runtime_initialized (MonoProfiler *profiler)
 606 {
 607         runtime_inited = 1;
 608 }
 609
 610 /*
 611  * Can be called only at safe callback locations.
 612  */
 613 static void
 614 safe_dump (MonoProfiler *profiler, LogBuffer *logbuffer)
 615 {
 616         int cd = logbuffer->call_depth;
 617         take_lock ();
 618         dump_buffer (profiler, TLS_GET (tlsbuffer));
 619         release_lock ();
 620         TLS_SET (tlsbuffer, NULL);
 621         init_thread ();
 622         TLS_GET (tlsbuffer)->call_depth = cd;
 623 }
 624
 625 static int
 626 gc_reference (MonoObject *obj, MonoClass *klass, uintptr_t size, uintptr_t num, MonoObject **refs, uintptr_t *offsets, void *data)
 627 {
 628         int i;
 629         uintptr_t last_offset = 0;
 630         //const char *name = mono_class_get_name (klass);
 631         LogBuffer *logbuffer = ensure_logbuf (20 + num * 8);
 632         emit_byte (logbuffer, TYPE_HEAP_OBJECT | TYPE_HEAP);
 633         emit_obj (logbuffer, obj);
 634         emit_ptr (logbuffer, klass);
 635         /* account for object alignment in the heap */
 636         size += 7;
 637         size &= ~7;
 638         emit_value (logbuffer, size);
 639         emit_value (logbuffer, num);
 640         for (i = 0; i < num; ++i) {
 641                 emit_value (logbuffer, offsets [i] - last_offset);
 642                 last_offset = offsets [i];
 643                 emit_obj (logbuffer, refs [i]);
 644         }
 645         //if (num)
 646         //      printf ("obj: %p, klass: %s, refs: %d, size: %d\n", obj, name, (int)num, (int)size);
 647         return 0;
 648 }
 649
 650 static unsigned int hs_mode_ms = 0;
 651 static unsigned int hs_mode_gc = 0;
 652 static unsigned int hs_mode_ondemand = 0;
 653 static unsigned int gc_count = 0;
 654 static uint64_t last_hs_time = 0;
 655
 656 static void
 657 heap_walk (MonoProfiler *profiler)
 658 {
 659         int do_walk = 0;
 660         uint64_t now;
 661         LogBuffer *logbuffer;
 662         if (!do_heap_shot)
 663                 return;
 664         logbuffer = ensure_logbuf (10);
 665         now = current_time ();
 666         if (hs_mode_ms && (now - last_hs_time)/1000000 >= hs_mode_ms)
 667                 do_walk = 1;
 668         else if (hs_mode_gc && (gc_count % hs_mode_gc) == 0)
 669                 do_walk = 1;
 670         else if (hs_mode_ondemand && heapshot_requested)
 671                 do_walk = 1;
 672         else if (!hs_mode_ms && !hs_mode_gc && profiler->last_gc_gen_started == mono_gc_max_generation ())
 673                 do_walk = 1;
 674
 675         if (!do_walk)
 676                 return;
 677         heapshot_requested = 0;
 678         emit_byte (logbuffer, TYPE_HEAP_START | TYPE_HEAP);
 679         emit_time (logbuffer, now);
 680         mono_gc_walk_heap (0, gc_reference, NULL);
 681         logbuffer = ensure_logbuf (10);
 682         now = current_time ();
 683         emit_byte (logbuffer, TYPE_HEAP_END | TYPE_HEAP);
 684         emit_time (logbuffer, now);
 685         last_hs_time = now;
 686 }
 687
 688 static void
 689 gc_event (MonoProfiler *profiler, MonoGCEvent ev, int generation) {
 690         uint64_t now;
 691         LogBuffer *logbuffer = ensure_logbuf (10);
 692         now = current_time ();
 693         ENTER_LOG (logbuffer, "gcevent");
 694         emit_byte (logbuffer, TYPE_GC_EVENT | TYPE_GC);
 695         emit_time (logbuffer, now);
 696         emit_value (logbuffer, ev);
 697         emit_value (logbuffer, generation);
 698         /* to deal with nested gen1 after gen0 started */
 699         if (ev == MONO_GC_EVENT_START) {
 700                 profiler->last_gc_gen_started = generation;
 701                 if (generation == mono_gc_max_generation ())
 702                         gc_count++;
 703         }
 704         if (ev == MONO_GC_EVENT_PRE_START_WORLD)
 705                 heap_walk (profiler);
 706         EXIT_LOG (logbuffer);
 707         if (ev == MONO_GC_EVENT_POST_START_WORLD)
 708                 safe_dump (profiler, logbuffer);
 709         //printf ("gc event %d for generation %d\n", ev, generation);
 710 }
 711
 712 static void
 713 gc_resize (MonoProfiler *profiler, int64_t new_size) {
 714         uint64_t now;
 715         LogBuffer *logbuffer = ensure_logbuf (10);
 716         now = current_time ();
 717         ENTER_LOG (logbuffer, "gcresize");
 718         emit_byte (logbuffer, TYPE_GC_RESIZE | TYPE_GC);
 719         emit_time (logbuffer, now);
 720         emit_value (logbuffer, new_size);
 721         //printf ("gc resized to %lld\n", new_size);
 722         EXIT_LOG (logbuffer);
 723 }
 724
 725 #define MAX_FRAMES 16
 726 typedef struct {
 727         int count;
 728         MonoMethod* methods [MAX_FRAMES];
 729 } FrameData;
 730 static int num_frames = MAX_FRAMES / 2;
 731
 732 static mono_bool
 733 walk_stack (MonoMethod *method, int32_t native_offset, int32_t il_offset, mono_bool managed, void* data)
 734 {
 735         FrameData *frame = data;
 736         if (method && frame->count < num_frames) {
 737                 frame->methods [frame->count++] = method;
 738                 //printf ("In %d %s\n", frame->count, mono_method_get_name (method));
 739         }
 740         return frame->count == num_frames;
 741 }
 742
 743 /*
 744  * a note about stack walks: they can cause more profiler events to fire,
 745  * so we need to make sure they don't happen after we started emitting an
 746  * event, hence the collect_bt/emit_bt split.
 747  */
 748 static void
 749 collect_bt (FrameData *data)
 750 {
 751         data->count = 0;
 752         mono_stack_walk_no_il (walk_stack, data);
 753 }
 754
 755 static void
 756 emit_bt (LogBuffer *logbuffer, FrameData *data)
 757 {
 758         /* FIXME: this is actually tons of data and we should
 759          * just output it the first time and use an id the next
 760          */
 761         if (data->count > num_frames)
 762                 printf ("bad num frames: %d\n", data->count);
 763         emit_value (logbuffer, 0); /* flags */
 764         emit_value (logbuffer, data->count);
 765         //if (*p != data.count) {
 766         //      printf ("bad num frames enc at %d: %d -> %d\n", count, data.count, *p); printf ("frames end: %p->%p\n", p, logbuffer->data); exit(0);}
 767         while (data->count) {
 768                 emit_ptr (logbuffer, data->methods [--data->count]);
 769         }
 770 }
 771
 772 static void
 773 gc_alloc (MonoProfiler *prof, MonoObject *obj, MonoClass *klass)
 774 {
 775         uint64_t now;
 776         uintptr_t len;
 777         int do_bt = (nocalls && runtime_inited && !notraces)? TYPE_ALLOC_BT: 0;
 778         FrameData data;
 779         LogBuffer *logbuffer;
 780         len = mono_object_get_size (obj);
 781         /* account for object alignment in the heap */
 782         len += 7;
 783         len &= ~7;
 784         if (do_bt)
 785                 collect_bt (&data);
 786         logbuffer = ensure_logbuf (32 + MAX_FRAMES * 8);
 787         now = current_time ();
 788         ENTER_LOG (logbuffer, "gcalloc");
 789         emit_byte (logbuffer, do_bt | TYPE_ALLOC);
 790         emit_time (logbuffer, now);
 791         emit_ptr (logbuffer, klass);
 792         emit_obj (logbuffer, obj);
 793         emit_value (logbuffer, len);
 794         if (do_bt)
 795                 emit_bt (logbuffer, &data);
 796         EXIT_LOG (logbuffer);
 797         if (logbuffer->next)
 798                 safe_dump (prof, logbuffer);
 799         process_requests (prof);
 800         //printf ("gc alloc %s at %p\n", mono_class_get_name (klass), obj);
 801 }
 802
 803 static void
 804 gc_moves (MonoProfiler *prof, void **objects, int num)
 805 {
 806         int i;
 807         uint64_t now;
 808         LogBuffer *logbuffer = ensure_logbuf (10 + num * 8);
 809         now = current_time ();
 810         ENTER_LOG (logbuffer, "gcmove");
 811         emit_byte (logbuffer, TYPE_GC_MOVE | TYPE_GC);
 812         emit_time (logbuffer, now);
 813         emit_value (logbuffer, num);
 814         for (i = 0; i < num; ++i)
 815                 emit_obj (logbuffer, objects [i]);
 816         //printf ("gc moved %d objects\n", num/2);
 817         EXIT_LOG (logbuffer);
 818 }
 819
 820 static void
 821 gc_roots (MonoProfiler *prof, int num, void **objects, int *root_types, uintptr_t *extra_info)
 822 {
 823         int i;
 824         LogBuffer *logbuffer = ensure_logbuf (5 + num * 18);
 825         ENTER_LOG (logbuffer, "gcroots");
 826         emit_byte (logbuffer, TYPE_HEAP_ROOT | TYPE_HEAP);
 827         emit_value (logbuffer, num);
 828         emit_value (logbuffer, mono_gc_collection_count (mono_gc_max_generation ()));
 829         for (i = 0; i < num; ++i) {
 830                 emit_obj (logbuffer, objects [i]);
 831                 emit_value (logbuffer, root_types [i]);
 832                 emit_value (logbuffer, extra_info [i]);
 833         }
 834         EXIT_LOG (logbuffer);
 835 }
 836
 837 static void
 838 gc_handle (MonoProfiler *prof, int op, int type, uintptr_t handle, MonoObject *obj)
 839 {
 840         uint64_t now;
 841         LogBuffer *logbuffer = ensure_logbuf (16);
 842         now = current_time ();
 843         ENTER_LOG (logbuffer, "gchandle");
 844         if (op == MONO_PROFILER_GC_HANDLE_CREATED)
 845                 emit_byte (logbuffer, TYPE_GC_HANDLE_CREATED | TYPE_GC);
 846         else if (op == MONO_PROFILER_GC_HANDLE_DESTROYED)
 847                 emit_byte (logbuffer, TYPE_GC_HANDLE_DESTROYED | TYPE_GC);
 848         else
 849                 return;
 850         emit_time (logbuffer, now);
 851         emit_value (logbuffer, type);
 852         emit_value (logbuffer, handle);
 853         if (op == MONO_PROFILER_GC_HANDLE_CREATED)
 854                 emit_obj (logbuffer, obj);
 855         EXIT_LOG (logbuffer);
 856         process_requests (prof);
 857 }
 858
 859 static char*
 860 push_nesting (char *p, MonoClass *klass)
 861 {
 862         MonoClass *nesting;
 863         const char *name;
 864         const char *nspace;
 865         nesting = mono_class_get_nesting_type (klass);
 866         if (nesting) {
 867                 p = push_nesting (p, nesting);
 868                 *p++ = '/';
 869                 *p = 0;
 870         }
 871         name = mono_class_get_name (klass);
 872         nspace = mono_class_get_namespace (klass);
 873         if (*nspace) {
 874                 strcpy (p, nspace);
 875                 p += strlen (nspace);
 876                 *p++ = '.';
 877                 *p = 0;
 878         }
 879         strcpy (p, name);
 880         p += strlen (name);
 881         return p;
 882 }
 883
 884 static char*
 885 type_name (MonoClass *klass)
 886 {
 887         char buf [1024];
 888         char *p;
 889         push_nesting (buf, klass);
 890         p = malloc (strlen (buf) + 1);
 891         strcpy (p, buf);
 892         return p;
 893 }
 894
 895 static void
 896 image_loaded (MonoProfiler *prof, MonoImage *image, int result)
 897 {
 898         uint64_t now;
 899         const char *name;
 900         int nlen;
 901         LogBuffer *logbuffer;
 902         if (result != MONO_PROFILE_OK)
 903                 return;
 904         name = mono_image_get_filename (image);
 905         nlen = strlen (name) + 1;
 906         logbuffer = ensure_logbuf (16 + nlen);
 907         now = current_time ();
 908         ENTER_LOG (logbuffer, "image");
 909         emit_byte (logbuffer, TYPE_END_LOAD | TYPE_METADATA);
 910         emit_time (logbuffer, now);
 911         emit_byte (logbuffer, TYPE_IMAGE);
 912         emit_ptr (logbuffer, image);
 913         emit_value (logbuffer, 0); /* flags */
 914         memcpy (logbuffer->data, name, nlen);
 915         logbuffer->data += nlen;
 916         //printf ("loaded image %p (%s)\n", image, name);
 917         EXIT_LOG (logbuffer);
 918         if (logbuffer->next)
 919                 safe_dump (prof, logbuffer);
 920         process_requests (prof);
 921 }
 922
 923 static void
 924 class_loaded (MonoProfiler *prof, MonoClass *klass, int result)
 925 {
 926         uint64_t now;
 927         char *name;
 928         int nlen;
 929         MonoImage *image;
 930         LogBuffer *logbuffer;
 931         if (result != MONO_PROFILE_OK)
 932                 return;
 933         if (runtime_inited)
 934                 name = mono_type_get_name (mono_class_get_type (klass));
 935         else
 936                 name = type_name (klass);
 937         nlen = strlen (name) + 1;
 938         image = mono_class_get_image (klass);
 939         logbuffer = ensure_logbuf (24 + nlen);
 940         now = current_time ();
 941         ENTER_LOG (logbuffer, "class");
 942         emit_byte (logbuffer, TYPE_END_LOAD | TYPE_METADATA);
 943         emit_time (logbuffer, now);
 944         emit_byte (logbuffer, TYPE_CLASS);
 945         emit_ptr (logbuffer, klass);
 946         emit_ptr (logbuffer, image);
 947         emit_value (logbuffer, 0); /* flags */
 948         memcpy (logbuffer->data, name, nlen);
 949         logbuffer->data += nlen;
 950         //printf ("loaded class %p (%s)\n", klass, name);
 951         if (runtime_inited)
 952                 mono_free (name);
 953         else
 954                 free (name);
 955         EXIT_LOG (logbuffer);
 956         if (logbuffer->next)
 957                 safe_dump (prof, logbuffer);
 958         process_requests (prof);
 959 }
 960
 961 static void
 962 method_enter (MonoProfiler *prof, MonoMethod *method)
 963 {
 964         uint64_t now;
 965         LogBuffer *logbuffer = ensure_logbuf (16);
 966         if (logbuffer->call_depth++ > max_call_depth)
 967                 return;
 968         now = current_time ();
 969         ENTER_LOG (logbuffer, "enter");
 970         emit_byte (logbuffer, TYPE_ENTER | TYPE_METHOD);
 971         emit_time (logbuffer, now);
 972         emit_method (logbuffer, method);
 973         EXIT_LOG (logbuffer);
 974         process_requests (prof);
 975 }
 976
 977 static void
 978 method_leave (MonoProfiler *prof, MonoMethod *method)
 979 {
 980         uint64_t now;
 981         LogBuffer *logbuffer = ensure_logbuf (16);
 982         if (--logbuffer->call_depth > max_call_depth)
 983                 return;
 984         now = current_time ();
 985         ENTER_LOG (logbuffer, "leave");
 986         emit_byte (logbuffer, TYPE_LEAVE | TYPE_METHOD);
 987         emit_time (logbuffer, now);
 988         emit_method (logbuffer, method);
 989         EXIT_LOG (logbuffer);
 990         if (logbuffer->next)
 991                 safe_dump (prof, logbuffer);
 992         process_requests (prof);
 993 }
 994
 995 static void
 996 method_exc_leave (MonoProfiler *prof, MonoMethod *method)
 997 {
 998         uint64_t now;
 999         LogBuffer *logbuffer;
1000         if (nocalls)
1001                 return;
1002         logbuffer = ensure_logbuf (16);
1003         if (--logbuffer->call_depth > max_call_depth)
1004                 return;
1005         now = current_time ();
1006         ENTER_LOG (logbuffer, "eleave");
1007         emit_byte (logbuffer, TYPE_EXC_LEAVE | TYPE_METHOD);
1008         emit_time (logbuffer, now);
1009         emit_method (logbuffer, method);
1010         EXIT_LOG (logbuffer);
1011         process_requests (prof);
1012 }
1013
1014 static void
1015 method_jitted (MonoProfiler *prof, MonoMethod *method, MonoJitInfo* jinfo, int result)
1016 {
1017         uint64_t now;
1018         char *name;
1019         int nlen;
1020         LogBuffer *logbuffer;
1021         if (result != MONO_PROFILE_OK)
1022                 return;
1023         name = mono_method_full_name (method, 1);
1024         nlen = strlen (name) + 1;
1025         logbuffer = ensure_logbuf (32 + nlen);
1026         now = current_time ();
1027         ENTER_LOG (logbuffer, "jit");
1028         emit_byte (logbuffer, TYPE_JIT | TYPE_METHOD);
1029         emit_time (logbuffer, now);
1030         emit_method (logbuffer, method);
1031         emit_ptr (logbuffer, mono_jit_info_get_code_start (jinfo));
1032         emit_value (logbuffer, mono_jit_info_get_code_size (jinfo));
1033         memcpy (logbuffer->data, name, nlen);
1034         logbuffer->data += nlen;
1035         mono_free (name);
1036         EXIT_LOG (logbuffer);
1037         if (logbuffer->next)
1038                 safe_dump (prof, logbuffer);
1039         process_requests (prof);
1040 }
1041
1042 static void
1043 throw_exc (MonoProfiler *prof, MonoObject *object)
1044 {
1045         int do_bt = (nocalls && runtime_inited && !notraces)? TYPE_EXCEPTION_BT: 0;
1046         uint64_t now;
1047         FrameData data;
1048         LogBuffer *logbuffer;
1049         if (do_bt)
1050                 collect_bt (&data);
1051         logbuffer = ensure_logbuf (16 + MAX_FRAMES * 8);
1052         now = current_time ();
1053         ENTER_LOG (logbuffer, "throw");
1054         emit_byte (logbuffer, do_bt | TYPE_EXCEPTION);
1055         emit_time (logbuffer, now);
1056         emit_obj (logbuffer, object);
1057         if (do_bt)
1058                 emit_bt (logbuffer, &data);
1059         EXIT_LOG (logbuffer);
1060         process_requests (prof);
1061 }
1062
1063 static void
1064 clause_exc (MonoProfiler *prof, MonoMethod *method, int clause_type, int clause_num)
1065 {
1066         uint64_t now;
1067         LogBuffer *logbuffer = ensure_logbuf (16);
1068         now = current_time ();
1069         ENTER_LOG (logbuffer, "clause");
1070         emit_byte (logbuffer, TYPE_EXCEPTION | TYPE_CLAUSE);
1071         emit_time (logbuffer, now);
1072         emit_value (logbuffer, clause_type);
1073         emit_value (logbuffer, clause_num);
1074         emit_method (logbuffer, method);
1075         EXIT_LOG (logbuffer);
1076 }
1077
1078 static void
1079 monitor_event (MonoProfiler *profiler, MonoObject *object, MonoProfilerMonitorEvent event)
1080 {
1081         int do_bt = (nocalls && runtime_inited && !notraces && event == MONO_PROFILER_MONITOR_CONTENTION)? TYPE_MONITOR_BT: 0;
1082         uint64_t now;
1083         FrameData data;
1084         LogBuffer *logbuffer;
1085         if (do_bt)
1086                 collect_bt (&data);
1087         logbuffer = ensure_logbuf (16 + MAX_FRAMES * 8);
1088         now = current_time ();
1089         ENTER_LOG (logbuffer, "monitor");
1090         emit_byte (logbuffer, (event << 4) | do_bt | TYPE_MONITOR);
1091         emit_time (logbuffer, now);
1092         emit_obj (logbuffer, object);
1093         if (do_bt)
1094                 emit_bt (logbuffer, &data);
1095         EXIT_LOG (logbuffer);
1096         process_requests (profiler);
1097 }
1098
1099 static void
1100 thread_start (MonoProfiler *prof, uintptr_t tid)
1101 {
1102         //printf ("thread start %p\n", (void*)tid);
1103         init_thread ();
1104 }
1105
1106 static void
1107 thread_end (MonoProfiler *prof, uintptr_t tid)
1108 {
1109         take_lock ();
1110         if (TLS_GET (tlsbuffer))
1111                 dump_buffer (prof, TLS_GET (tlsbuffer));
1112         release_lock ();
1113         TLS_SET (tlsbuffer, NULL);
1114 }
1115
1116 static void
1117 thread_name (MonoProfiler *prof, uintptr_t tid, const char *name)
1118 {
1119         int len = strlen (name) + 1;
1120         uint64_t now;
1121         LogBuffer *logbuffer;
1122         logbuffer = ensure_logbuf (10 + len);
1123         now = current_time ();
1124         ENTER_LOG (logbuffer, "tname");
1125         emit_byte (logbuffer, TYPE_METADATA);
1126         emit_time (logbuffer, now);
1127         emit_byte (logbuffer, TYPE_THREAD);
1128         emit_ptr (logbuffer, (void*)tid);
1129         emit_value (logbuffer, 0); /* flags */
1130         memcpy (logbuffer->data, name, len);
1131         logbuffer->data += len;
1132         EXIT_LOG (logbuffer);
1133 }
1134
1135 #ifndef HOST_WIN32
1136 #include "mono/io-layer/atomic.h"
1137 #endif
1138 #define cmp_exchange InterlockedCompareExchangePointer
1139 /*#else
1140 static void*
1141 cmp_exchange (volatile void **dest, void *exch, void *comp)
1142 {
1143         void *old;
1144         __asm__ __volatile__ ("lock; "
1145 #ifdef __x86_64__
1146                 "cmpxchgq"
1147 #else
1148                 "cmpxchgl"
1149 #endif
1150                 " %2, %0"
1151                 : "=m" (*dest), "=a" (old)
1152                 : "r" (exch), "m" (*dest), "a" (comp));
1153         return old;
1154 }
1155 #endif
1156 */
1157
1158 static void
1159 mono_sample_hit (MonoProfiler *profiler, unsigned char *ip, void *context)
1160 {
1161         StatBuffer *sbuf;
1162         uint64_t now;
1163         uintptr_t *data, *new_data, *old_data;
1164         if (in_shutdown)
1165                 return;
1166         now = current_time ();
1167         if (do_debug) {
1168                 int len;
1169                 char buf [256];
1170                 snprintf (buf, sizeof (buf), "hit at %p in thread %p at %llu\n", ip, (void*)thread_id (), (unsigned long long int)now);
1171                 len = strlen (buf);
1172                 write (2, buf, len);
1173         }
1174         sbuf = profiler->stat_buffers;
1175         if (!sbuf)
1176                 return;
1177         /* overflow */
1178         if (sbuf->data + 400 >= sbuf->data_end) {
1179                 sbuf = create_stat_buffer ();
1180                 sbuf->next = profiler->stat_buffers;
1181                 profiler->stat_buffers = sbuf;
1182                 if (do_debug)
1183                         write (2, "overflow\n", 9);
1184                 /* notify the helper thread */
1185                 if (sbuf->next->next) {
1186                         char c = 0;
1187                         write (profiler->pipes [1], &c, 1);
1188                         if (do_debug)
1189                                 write (2, "notify\n", 7);
1190                 }
1191         }
1192         do {
1193                 old_data = sbuf->data;
1194                 new_data = old_data + 4;
1195                 data = cmp_exchange ((volatile void**)&sbuf->data, new_data, old_data);
1196         } while (data != old_data);
1197         if (old_data >= sbuf->data_end)
1198                 return; /* lost event */
1199         old_data [0] = 1 | (sample_type << 16);
1200         old_data [1] = thread_id ();
1201         old_data [2] = (now - profiler->startup_time) / 10000;
1202         old_data [3] = (uintptr_t)ip;
1203 }
1204
1205 static uintptr_t *code_pages = 0;
1206 static int num_code_pages = 0;
1207 static int size_code_pages = 0;
1208 #define CPAGE_SHIFT (9)
1209 #define CPAGE_SIZE (1 << CPAGE_SHIFT)
1210 #define CPAGE_MASK (~(CPAGE_SIZE - 1))
1211 #define CPAGE_ADDR(p) ((p) & CPAGE_MASK)
1212
1213 static uintptr_t
1214 add_code_page (uintptr_t *hash, uintptr_t hsize, uintptr_t page)
1215 {
1216         uintptr_t i;
1217         uintptr_t start_pos;
1218         start_pos = (page >> CPAGE_SHIFT) % hsize;
1219         i = start_pos;
1220         do {
1221                 if (hash [i] && CPAGE_ADDR (hash [i]) == CPAGE_ADDR (page)) {
1222                         return 0;
1223                 } else if (!hash [i]) {
1224                         hash [i] = page;
1225                         return 1;
1226                 }
1227                 /* wrap around */
1228                 if (++i == hsize)
1229                         i = 0;
1230         } while (i != start_pos);
1231         /* should not happen */
1232         printf ("failed code page store\n");
1233         return 0;
1234 }
1235
1236 static void
1237 add_code_pointer (uintptr_t ip)
1238 {
1239         uintptr_t i;
1240         if (num_code_pages * 2 >= size_code_pages) {
1241                 uintptr_t *n;
1242                 uintptr_t old_size = size_code_pages;
1243                 size_code_pages *= 2;
1244                 if (size_code_pages == 0)
1245                         size_code_pages = 16;
1246                 n = calloc (sizeof (uintptr_t) * size_code_pages, 1);
1247                 for (i = 0; i < old_size; ++i) {
1248                         if (code_pages [i])
1249                                 add_code_page (n, size_code_pages, code_pages [i]);
1250                 }
1251                 if (code_pages)
1252                         free (code_pages);
1253                 code_pages = n;
1254         }
1255         num_code_pages += add_code_page (code_pages, size_code_pages, ip & CPAGE_MASK);
1256 }
1257
1258 static void
1259 dump_ubin (const char *filename, uintptr_t load_addr, uint64_t offset, uintptr_t size)
1260 {
1261         uint64_t now;
1262         LogBuffer *logbuffer;
1263         int len;
1264         len = strlen (filename) + 1;
1265         now = current_time ();
1266         logbuffer = ensure_logbuf (20 + len);
1267         emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_UBIN);
1268         emit_time (logbuffer, now);
1269         emit_svalue (logbuffer, load_addr);
1270         emit_uvalue (logbuffer, offset);
1271         emit_uvalue (logbuffer, size);
1272         memcpy (logbuffer->data, filename, len);
1273         logbuffer->data += len;
1274 }
1275
1276 static void
1277 dump_usym (const char *name, uintptr_t value, uintptr_t size)
1278 {
1279         LogBuffer *logbuffer;
1280         int len;
1281         len = strlen (name) + 1;
1282         logbuffer = ensure_logbuf (20 + len);
1283         emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_USYM);
1284         emit_ptr (logbuffer, (void*)value);
1285         emit_value (logbuffer, size);
1286         memcpy (logbuffer->data, name, len);
1287         logbuffer->data += len;
1288 }
1289
1290 #ifdef ELFMAG0
1291
1292 #if SIZEOF_VOID_P == 4
1293 #define ELF_WSIZE 32
1294 #else
1295 #define ELF_WSIZE 64
1296 #endif
1297 #ifndef ElfW
1298 #define ElfW(type)      _ElfW (Elf, ELF_WSIZE, type)
1299 #define _ElfW(e,w,t)    _ElfW_1 (e, w, _##t)
1300 #define _ElfW_1(e,w,t)  e##w##t
1301 #endif
1302
1303 static void
1304 dump_elf_symbols (ElfW(Sym) *symbols, int num_symbols, const char *strtab, void *load_addr)
1305 {
1306         int i;
1307         for (i = 0; i < num_symbols; ++i) {
1308                 const char* sym;
1309                 sym =  strtab + symbols [i].st_name;
1310                 if (!symbols [i].st_name || !symbols [i].st_size || (symbols [i].st_info & 0xf) != STT_FUNC)
1311                         continue;
1312                 //printf ("symbol %s at %d\n", sym, symbols [i].st_value);
1313                 dump_usym (sym, (uintptr_t)load_addr + symbols [i].st_value, symbols [i].st_size);
1314         }
1315 }
1316
1317 static int
1318 read_elf_symbols (MonoProfiler *prof, const char *filename, void *load_addr)
1319 {
1320         int fd, i;
1321         void *data;
1322         struct stat statb;
1323         uint64_t file_size;
1324         ElfW(Ehdr) *header;
1325         ElfW(Shdr) *sheader;
1326         ElfW(Shdr) *shstrtabh;
1327         ElfW(Shdr) *symtabh = NULL;
1328         ElfW(Shdr) *strtabh = NULL;
1329         ElfW(Sym) *symbols = NULL;
1330         const char *strtab;
1331         int num_symbols;
1332
1333         fd = open (filename, O_RDONLY);
1334         if (fd < 0)
1335                 return 0;
1336         if (fstat (fd, &statb) != 0) {
1337                 close (fd);
1338                 return 0;
1339         }
1340         file_size = statb.st_size;
1341         data = mmap (NULL, file_size, PROT_READ, MAP_PRIVATE, fd, 0);
1342         close (fd);
1343         if (data == MAP_FAILED)
1344                 return 0;
1345         header = data;
1346         if (header->e_ident [EI_MAG0] != ELFMAG0 ||
1347                         header->e_ident [EI_MAG1] != ELFMAG1 ||
1348                         header->e_ident [EI_MAG2] != ELFMAG2 ||
1349                         header->e_ident [EI_MAG3] != ELFMAG3 ) {
1350                 munmap (data, file_size);
1351                 return 0;
1352         }
1353         sheader = (void*)((char*)data + header->e_shoff);
1354         shstrtabh = (void*)((char*)sheader + (header->e_shentsize * header->e_shstrndx));
1355         strtab = (const char*)data + shstrtabh->sh_offset;
1356         for (i = 0; i < header->e_shnum; ++i) {
1357                 //printf ("section header: %d\n", sheader->sh_type);
1358                 if (sheader->sh_type == SHT_SYMTAB) {
1359                         symtabh = sheader;
1360                         strtabh = (void*)((char*)data + header->e_shoff + sheader->sh_link * header->e_shentsize);
1361                         /*printf ("symtab section header: %d, .strstr: %d\n", i, sheader->sh_link);*/
1362                         break;
1363                 }
1364                 sheader = (void*)((char*)sheader + header->e_shentsize);
1365         }
1366         if (!symtabh || !strtabh) {
1367                 munmap (data, file_size);
1368                 return 0;
1369         }
1370         strtab = (const char*)data + strtabh->sh_offset;
1371         num_symbols = symtabh->sh_size / symtabh->sh_entsize;
1372         symbols = (void*)((char*)data + symtabh->sh_offset);
1373         dump_elf_symbols (symbols, num_symbols, strtab, load_addr);
1374         munmap (data, file_size);
1375         return 1;
1376 }
1377 #endif
1378
1379 #if defined(HAVE_DL_ITERATE_PHDR) && defined(ELFMAG0)
1380 static int
1381 elf_dl_callback (struct dl_phdr_info *info, size_t size, void *data)
1382 {
1383         MonoProfiler *prof = data;
1384         char buf [256];
1385         const char *filename;
1386         BinaryObject *obj;
1387         char *a = (void*)info->dlpi_addr;
1388         int i, num_sym;
1389         ElfW(Dyn) *dyn = NULL;
1390         ElfW(Sym) *symtab = NULL;
1391         ElfW(Word) *hash_table = NULL;
1392         ElfW(Ehdr) *header = NULL;
1393         const char* strtab = NULL;
1394         for (obj = prof->binary_objects; obj; obj = obj->next) {
1395                 if (obj->addr == a)
1396                         return 0;
1397         }
1398         filename = info->dlpi_name;
1399         if (!info->dlpi_addr && !filename [0]) {
1400                 int l = readlink ("/proc/self/exe", buf, sizeof (buf) - 1);
1401                 if (l > 0) {
1402                         buf [l] = 0;
1403                         filename = buf;
1404                 }
1405         }
1406         obj = calloc (sizeof (BinaryObject), 1);
1407         obj->addr = (void*)info->dlpi_addr;
1408         obj->name = pstrdup (filename);
1409         obj->next = prof->binary_objects;
1410         prof->binary_objects = obj;
1411         //printf ("loaded file: %s at %p, segments: %d\n", filename, (void*)info->dlpi_addr, info->dlpi_phnum);
1412         a = NULL;
1413         for (i = 0; i < info->dlpi_phnum; ++i) {
1414                 //printf ("segment type %d file offset: %d, size: %d\n", info->dlpi_phdr[i].p_type, info->dlpi_phdr[i].p_offset, info->dlpi_phdr[i].p_memsz);
1415                 if (info->dlpi_phdr[i].p_type == PT_LOAD && !header) {
1416                         header = (ElfW(Ehdr)*)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1417                         if (header->e_ident [EI_MAG0] != ELFMAG0 ||
1418                                         header->e_ident [EI_MAG1] != ELFMAG1 ||
1419                                         header->e_ident [EI_MAG2] != ELFMAG2 ||
1420                                         header->e_ident [EI_MAG3] != ELFMAG3 ) {
1421                                 header = NULL;
1422                         }
1423                         dump_ubin (filename, info->dlpi_addr + info->dlpi_phdr[i].p_vaddr, info->dlpi_phdr[i].p_offset, info->dlpi_phdr[i].p_memsz);
1424                 } else if (info->dlpi_phdr[i].p_type == PT_DYNAMIC) {
1425                         dyn = (ElfW(Dyn) *)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1426                 }
1427         }
1428         if (read_elf_symbols (prof, filename, (void*)info->dlpi_addr))
1429                 return 0;
1430         if (!info->dlpi_name || !info->dlpi_name[0])
1431                 return 0;
1432         if (!dyn)
1433                 return 0;
1434         for (i = 0; dyn [i].d_tag != DT_NULL; ++i) {
1435                 if (dyn [i].d_tag == DT_SYMTAB) {
1436                         if (symtab && do_debug)
1437                                 printf ("multiple symtabs: %d\n", i);
1438                         symtab = (ElfW(Sym) *)(a + dyn [i].d_un.d_ptr);
1439                 } else if (dyn [i].d_tag == DT_HASH) {
1440                         hash_table = (ElfW(Word) *)(a + dyn [i].d_un.d_ptr);
1441                 } else if (dyn [i].d_tag == DT_STRTAB) {
1442                         strtab = (const char*)(a + dyn [i].d_un.d_ptr);
1443                 }
1444         }
1445         if (!hash_table)
1446                 return 0;
1447         num_sym = hash_table [1];
1448         dump_elf_symbols (symtab, num_sym, strtab, (void*)info->dlpi_addr);
1449         return 0;
1450 }
1451
1452 static int
1453 load_binaries (MonoProfiler *prof)
1454 {
1455         dl_iterate_phdr (elf_dl_callback, prof);
1456         return 1;
1457 }
1458 #else
1459 static int
1460 load_binaries (MonoProfiler *prof)
1461 {
1462         return 0;
1463 }
1464 #endif
1465
1466 static const char*
1467 symbol_for (uintptr_t code)
1468 {
1469 #ifdef HAVE_DLADDR
1470         void *ip = (void*)code;
1471         Dl_info di;
1472         if (dladdr (ip, &di)) {
1473                 if (di.dli_sname)
1474                         return di.dli_sname;
1475         } else {
1476         /*      char **names;
1477                 names = backtrace_symbols (&ip, 1);
1478                 if (names) {
1479                         const char* p = names [0];
1480                         free (names);
1481                         return p;
1482                 }
1483                 */
1484         }
1485 #endif
1486         return NULL;
1487 }
1488
1489 static void
1490 dump_unmanaged_coderefs (MonoProfiler *prof)
1491 {
1492         int i;
1493         const char* last_symbol;
1494         uintptr_t addr, page_end;
1495
1496         if (load_binaries (prof))
1497                 return;
1498         for (i = 0; i < size_code_pages; ++i) {
1499                 const char* sym;
1500                 if (!code_pages [i] || code_pages [i] & 1)
1501                         continue;
1502                 last_symbol = NULL;
1503                 addr = CPAGE_ADDR (code_pages [i]);
1504                 page_end = addr + CPAGE_SIZE;
1505                 code_pages [i] |= 1;
1506                 /* we dump the symbols for the whole page */
1507                 for (; addr < page_end; addr += 16) {
1508                         sym = symbol_for (addr);
1509                         if (sym && sym == last_symbol)
1510                                 continue;
1511                         last_symbol = sym;
1512                         if (!sym)
1513                                 continue;
1514                         dump_usym (sym, addr, 0); /* let's not guess the size */
1515                         //printf ("found symbol at %p: %s\n", (void*)addr, sym);
1516                 }
1517         }
1518 }
1519
1520 static void
1521 dump_sample_hits (MonoProfiler *prof, StatBuffer *sbuf, int recurse)
1522 {
1523         uintptr_t *sample;
1524         LogBuffer *logbuffer;
1525         if (!sbuf)
1526                 return;
1527         if (recurse && sbuf->next) {
1528                 dump_sample_hits (prof, sbuf->next, 1);
1529                 free_buffer (sbuf->next, sbuf->next->size);
1530                 sbuf->next = NULL;
1531         }
1532         for (sample = sbuf->buf; sample < sbuf->data;) {
1533                 int i;
1534                 int count = sample [0] & 0xffff;
1535                 int type = sample [0] >> 16;
1536                 if (sample + count + 3 > sbuf->data)
1537                         break;
1538                 logbuffer = ensure_logbuf (20 + count * 8);
1539                 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
1540                 emit_value (logbuffer, type);
1541                 emit_uvalue (logbuffer, (prof->startup_time + sample [2]) * 10000);
1542                 emit_value (logbuffer, count);
1543                 for (i = 0; i < count; ++i) {
1544                         emit_ptr (logbuffer, (void*)sample [i + 3]);
1545                         add_code_pointer (sample [i + 3]);
1546                 }
1547                 sample += count + 3;
1548         }
1549         dump_unmanaged_coderefs (prof);
1550 }
1551
1552 #if USE_PERF_EVENTS
1553 #ifndef __NR_perf_event_open
1554 #define __NR_perf_event_open 241
1555 #endif
1556
1557 static int perf_fd = -1;
1558 static void *mmap_base;
1559 static struct perf_event_mmap_page *page_desc = NULL;
1560 static int num_pages = 64;
1561 static unsigned int mmap_mask;
1562 static unsigned int prev_pos = 0;
1563
1564 typedef struct {
1565         struct perf_event_header h;
1566         uint64_t ip;
1567         uint32_t pid;
1568         uint32_t tid;
1569         uint64_t timestamp;
1570         uint64_t period;
1571         uint64_t nframes;
1572 } PSample;
1573
1574 static int
1575 perf_event_syscall (struct perf_event_attr *attr, pid_t pid, int cpu, int group_fd, unsigned long flags)
1576 {
1577         attr->size = PERF_ATTR_SIZE_VER0;
1578         //printf ("perf attr size: %d\n", attr->size);
1579 #if defined(__x86_64__)
1580         return syscall(/*__NR_perf_event_open*/ 298, attr, pid, cpu, group_fd, flags);
1581 #elif defined(__i386__)
1582         return syscall(/*__NR_perf_event_open*/ 336, attr, pid, cpu, group_fd, flags);
1583 #else
1584         return -1;
1585 #endif
1586 }
1587
1588 static int
1589 setup_perf_map (void)
1590 {
1591         mmap_mask = num_pages * getpagesize () - 1;
1592         mmap_base = mmap (NULL, (num_pages + 1) * getpagesize (), PROT_READ|PROT_WRITE, MAP_SHARED, perf_fd, 0);
1593         if (mmap_base == MAP_FAILED) {
1594                 if (do_debug)
1595                         printf ("failed mmap\n");
1596                 return 0;
1597         }
1598         page_desc = mmap_base;
1599         if (do_debug)
1600                 printf ("mmap version: %d\n", page_desc->version);
1601         return 1;
1602 }
1603
1604 static void
1605 dump_perf_hits (MonoProfiler *prof, void *buf, int size)
1606 {
1607         LogBuffer *logbuffer;
1608         void *end = (char*)buf + size;
1609         int samples = 0;
1610         int pid = getpid ();
1611
1612         while (buf < end) {
1613                 PSample *s = buf;
1614                 if (s->h.size == 0)
1615                         break;
1616                 if (pid != s->pid) {
1617                         if (do_debug)
1618                                 printf ("event for different pid: %d\n", s->pid);
1619                         buf = (char*)buf + s->h.size;
1620                         continue;
1621                 }
1622                 /*ip = (void*)s->ip;
1623                 printf ("sample: %d, size: %d, ip: %p (%s), timestamp: %llu, nframes: %llu\n",
1624                         s->h.type, s->h.size, ip, symbol_for (ip), s->timestamp, s->nframes);*/
1625                 logbuffer = ensure_logbuf (20 + s->nframes * 8);
1626                 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
1627                 emit_value (logbuffer, sample_type);
1628                 emit_uvalue (logbuffer, s->timestamp - prof->startup_time);
1629                 emit_value (logbuffer, 1); /* count */
1630                 emit_ptr (logbuffer, (void*)(uintptr_t)s->ip);
1631                 add_code_pointer (s->ip);
1632                 buf = (char*)buf + s->h.size;
1633                 samples++;
1634         }
1635         if (do_debug)
1636                 printf ("dumped %d samples\n", samples);
1637         dump_unmanaged_coderefs (prof);
1638 }
1639
1640 /* read events from the ring buffer */
1641 static int
1642 read_perf_mmap (MonoProfiler* prof)
1643 {
1644         unsigned char *buf;
1645         unsigned char *data = (unsigned char*)mmap_base + getpagesize ();
1646         unsigned int head = page_desc->data_head;
1647         int diff, size;
1648         unsigned int old;
1649
1650 #if defined(__i386__)
1651         asm volatile("lock; addl $0,0(%%esp)":::"memory");
1652 #elif defined (__x86_64__)
1653         asm volatile("lfence":::"memory");
1654 #endif
1655
1656         old = prev_pos;
1657         diff = head - old;
1658         if (diff < 0) {
1659                 if (do_debug)
1660                         printf ("lost mmap events: old: %d, head: %d\n", old, head);
1661                 old = head;
1662         }
1663         size = head - old;
1664         if ((old & mmap_mask) + size != (head & mmap_mask)) {
1665                 buf = data + (old & mmap_mask);
1666                 size = mmap_mask + 1 - (old & mmap_mask);
1667                 old += size;
1668                 /* size bytes at buf */
1669                 if (do_debug)
1670                         printf ("found1 bytes of events: %d\n", size);
1671                 dump_perf_hits (prof, buf, size);
1672         }
1673         buf = data + (old & mmap_mask);
1674         size = head - old;
1675         /* size bytes at buf */
1676         if (do_debug)
1677                 printf ("found bytes of events: %d\n", size);
1678         dump_perf_hits (prof, buf, size);
1679         old += size;
1680         prev_pos = old;
1681         page_desc->data_tail = old;
1682         return 0;
1683 }
1684
1685 static int
1686 setup_perf_event (void)
1687 {
1688         struct perf_event_attr attr;
1689         memset (&attr, 0, sizeof (attr));
1690         attr.type = PERF_TYPE_HARDWARE;
1691         switch (sample_type) {
1692         case SAMPLE_CYCLES: attr.config = PERF_COUNT_HW_CPU_CYCLES; break;
1693         case SAMPLE_INSTRUCTIONS: attr.config = PERF_COUNT_HW_INSTRUCTIONS; break;
1694         case SAMPLE_CACHE_MISSES: attr.config = PERF_COUNT_HW_CACHE_MISSES; break;
1695         case SAMPLE_CACHE_REFS: attr.config = PERF_COUNT_HW_CACHE_REFERENCES; break;
1696         case SAMPLE_BRANCHES: attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; break;
1697         case SAMPLE_BRANCH_MISSES: attr.config = PERF_COUNT_HW_BRANCH_MISSES; break;
1698         default: attr.config = PERF_COUNT_HW_CPU_CYCLES; break;
1699         }
1700         attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_PERIOD | PERF_SAMPLE_TIME;
1701 //      attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1702         attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING | PERF_FORMAT_ID;
1703         attr.inherit = 1;
1704         attr.freq = 1;
1705         attr.sample_freq = sample_freq;
1706
1707         perf_fd = perf_event_syscall (&attr, getpid (), -1, -1, 0);
1708         if (do_debug)
1709                 printf ("perf fd: %d, freq: %d, event: %llu\n", perf_fd, sample_freq, attr.config);
1710         if (perf_fd < 0) {
1711                 if (perf_fd == -EPERM) {
1712                         fprintf (stderr, "Perf syscall denied, do \"echo 1 > /proc/sys/kernel/perf_event_paranoid\" as root to enable.\n");
1713                 } else {
1714                         if (do_debug)
1715                                 perror ("open perf event");
1716                 }
1717                 return 0;
1718         }
1719         if (!setup_perf_map ()) {
1720                 close (perf_fd);
1721                 perf_fd = -1;
1722                 return 0;
1723         }
1724         return 1;
1725 }
1726
1727 #endif /* USE_PERF_EVENTS */
1728
1729 static void
1730 log_shutdown (MonoProfiler *prof)
1731 {
1732         in_shutdown = 1;
1733 #ifndef DISABLE_HELPER_THREAD
1734         if (prof->command_port) {
1735                 char c = 1;
1736                 void *res;
1737                 write (prof->pipes [1], &c, 1);
1738                 pthread_join (prof->helper_thread, &res);
1739         }
1740 #endif
1741 #if USE_PERF_EVENTS
1742         if (page_desc)
1743                 read_perf_mmap (prof);
1744 #endif
1745         dump_sample_hits (prof, prof->stat_buffers, 1);
1746         take_lock ();
1747         if (TLS_GET (tlsbuffer))
1748                 dump_buffer (prof, TLS_GET (tlsbuffer));
1749         TLS_SET (tlsbuffer, NULL);
1750         release_lock ();
1751 #if defined (HAVE_SYS_ZLIB)
1752         if (prof->gzfile)
1753                 gzclose (prof->gzfile);
1754 #endif
1755         if (prof->pipe_output)
1756                 pclose (prof->file);
1757         else
1758                 fclose (prof->file);
1759         free (prof);
1760 }
1761
1762 static char*
1763 new_filename (const char* filename)
1764 {
1765         time_t t = time (NULL);
1766         int pid = process_id ();
1767         char pid_buf [16];
1768         char time_buf [16];
1769         char *res, *d;
1770         const char *p;
1771         int count_dates = 0;
1772         int count_pids = 0;
1773         int s_date, s_pid;
1774         struct tm *ts;
1775         for (p = filename; *p; p++) {
1776                 if (*p != '%')
1777                         continue;
1778                 p++;
1779                 if (*p == 't')
1780                         count_dates++;
1781                 else if (*p == 'p')
1782                         count_pids++;
1783                 else if (*p == 0)
1784                         break;
1785         }
1786         if (!count_dates && !count_pids)
1787                 return pstrdup (filename);
1788         snprintf (pid_buf, sizeof (pid_buf), "%d", pid);
1789         ts = gmtime (&t);
1790         snprintf (time_buf, sizeof (time_buf), "%d%02d%02d%02d%02d%02d",
1791                 1900 + ts->tm_year, 1 + ts->tm_mon, ts->tm_mday, ts->tm_hour, ts->tm_min, ts->tm_sec);
1792         s_date = strlen (time_buf);
1793         s_pid = strlen (pid_buf);
1794         d = res = malloc (strlen (filename) + s_date * count_dates + s_pid * count_pids);
1795         for (p = filename; *p; p++) {
1796                 if (*p != '%') {
1797                         *d++ = *p;
1798                         continue;
1799                 }
1800                 p++;
1801                 if (*p == 't') {
1802                         strcpy (d, time_buf);
1803                         d += s_date;
1804                         continue;
1805                 } else if (*p == 'p') {
1806                         strcpy (d, pid_buf);
1807                         d += s_pid;
1808                         continue;
1809                 } else if (*p == '%') {
1810                         *d++ = '%';
1811                         continue;
1812                 } else if (*p == 0)
1813                         break;
1814                 *d++ = '%';
1815                 *d++ = *p;
1816         }
1817         *d = 0;
1818         return res;
1819 }
1820
1821 #ifndef DISABLE_HELPER_THREAD
1822 static void*
1823 helper_thread (void* arg)
1824 {
1825         MonoProfiler* prof = arg;
1826         int command_socket;
1827         int len;
1828         char buf [64];
1829         MonoThread *thread = NULL;
1830
1831         //fprintf (stderr, "Server listening\n");
1832         command_socket = -1;
1833         while (1) {
1834                 fd_set rfds;
1835                 struct timeval tv;
1836                 int max_fd = -1;
1837                 FD_ZERO (&rfds);
1838                 FD_SET (prof->server_socket, &rfds);
1839                 max_fd = prof->server_socket;
1840                 FD_SET (prof->pipes [0], &rfds);
1841                 if (max_fd < prof->pipes [0])
1842                         max_fd = prof->pipes [0];
1843                 if (command_socket >= 0) {
1844                         FD_SET (command_socket, &rfds);
1845                         if (max_fd < command_socket)
1846                                 max_fd = command_socket;
1847                 }
1848 #if USE_PERF_EVENTS
1849                 if (perf_fd >= 0) {
1850                         FD_SET (perf_fd, &rfds);
1851                         if (max_fd < perf_fd)
1852                                 max_fd = perf_fd;
1853                 }
1854 #endif
1855                 tv.tv_sec = 1;
1856                 tv.tv_usec = 0;
1857                 len = select (max_fd + 1, &rfds, NULL, NULL, &tv);
1858                 if (FD_ISSET (prof->pipes [0], &rfds)) {
1859                         char c;
1860                         int r = read (prof->pipes [0], &c, 1);
1861                         if (r == 1 && c == 0) {
1862                                 StatBuffer *sbuf = prof->stat_buffers->next->next;
1863                                 prof->stat_buffers->next->next = NULL;
1864                                 if (do_debug)
1865                                         fprintf (stderr, "stat buffer dump\n");
1866                                 dump_sample_hits (prof, sbuf, 1);
1867                                 free_buffer (sbuf, sbuf->size);
1868                                 continue;
1869                         }
1870                         /* time to shut down */
1871                         if (thread)
1872                                 mono_thread_detach (thread);
1873                         if (do_debug)
1874                                 fprintf (stderr, "helper shutdown\n");
1875 #if USE_PERF_EVENTS
1876                         if (perf_fd >= 0)
1877                                 read_perf_mmap (prof);
1878 #endif
1879                         safe_dump (prof, ensure_logbuf (0));
1880                         return NULL;
1881                 }
1882 #if USE_PERF_EVENTS
1883                 if (perf_fd >= 0 && FD_ISSET (perf_fd, &rfds)) {
1884                         read_perf_mmap (prof);
1885                         safe_dump (prof, ensure_logbuf (0));
1886                 }
1887 #endif
1888                 if (command_socket >= 0 && FD_ISSET (command_socket, &rfds)) {
1889                         len = read (command_socket, buf, sizeof (buf) - 1);
1890                         if (len < 0)
1891                                 continue;
1892                         if (len == 0) {
1893                                 close (command_socket);
1894                                 command_socket = -1;
1895                                 continue;
1896                         }
1897                         buf [len] = 0;
1898                         if (strcmp (buf, "heapshot\n") == 0) {
1899                                 heapshot_requested = 1;
1900                                 //fprintf (stderr, "perform heapshot\n");
1901                                 if (runtime_inited && !thread) {
1902                                         thread = mono_thread_attach (mono_get_root_domain ());
1903                                         /*fprintf (stderr, "attached\n");*/
1904                                 }
1905                                 if (thread) {
1906                                         process_requests (prof);
1907                                         mono_thread_detach (thread);
1908                                         thread = NULL;
1909                                 }
1910                         }
1911                         continue;
1912                 }
1913                 if (!FD_ISSET (prof->server_socket, &rfds)) {
1914                         continue;
1915                 }
1916                 command_socket = accept (prof->server_socket, NULL, NULL);
1917                 if (command_socket < 0)
1918                         continue;
1919                 //fprintf (stderr, "Accepted connection\n");
1920         }
1921         return NULL;
1922 }
1923
1924 static int
1925 start_helper_thread (MonoProfiler* prof)
1926 {
1927         struct sockaddr_in server_address;
1928         int r;
1929         socklen_t slen;
1930         if (pipe (prof->pipes) < 0) {
1931                 fprintf (stderr, "Cannot create pipe\n");
1932                 return 0;
1933         }
1934         prof->server_socket = socket (PF_INET, SOCK_STREAM, 0);
1935         if (prof->server_socket < 0) {
1936                 fprintf (stderr, "Cannot create server socket\n");
1937                 return 0;
1938         }
1939         memset (&server_address, 0, sizeof (server_address));
1940         server_address.sin_family = AF_INET;
1941         server_address.sin_addr.s_addr = INADDR_ANY;
1942         server_address.sin_port = htons (prof->command_port);
1943         if (bind (prof->server_socket, (struct sockaddr *) &server_address, sizeof (server_address)) < 0) {
1944                 fprintf (stderr, "Cannot bind server socket, port: %d: %s\n", prof->command_port, strerror (errno));
1945                 close (prof->server_socket);
1946                 return 0;
1947         }
1948         if (listen (prof->server_socket, 1) < 0) {
1949                 fprintf (stderr, "Cannot listen server socket\n");
1950                 close (prof->server_socket);
1951                 return 0;
1952         }
1953         if (getsockname (prof->server_socket, (struct sockaddr *)&server_address, &slen) == 0) {
1954                 prof->command_port = ntohs (server_address.sin_port);
1955                 /*fprintf (stderr, "Assigned server port: %d\n", prof->command_port);*/
1956         }
1957
1958         r = pthread_create (&prof->helper_thread, NULL, helper_thread, prof);
1959         if (r) {
1960                 close (prof->server_socket);
1961                 return 0;
1962         }
1963         return 1;
1964 }
1965 #endif
1966
1967 static MonoProfiler*
1968 create_profiler (const char *filename)
1969 {
1970         MonoProfiler *prof;
1971         char *nf;
1972         int force_delete = 0;
1973         int need_helper_thread = 0;
1974         prof = calloc (1, sizeof (MonoProfiler));
1975
1976         prof->command_port = command_port;
1977         if (filename && *filename == '-') {
1978                 force_delete = 1;
1979                 filename++;
1980         }
1981         if (!filename) {
1982                 if (do_report)
1983                         filename = "|mprof-report -";
1984                 else
1985                         filename = "output.mlpd";
1986                 nf = (char*)filename;
1987         } else {
1988                 nf = new_filename (filename);
1989                 if (do_report) {
1990                         int s = strlen (nf) + 32;
1991                         char *p = malloc (s);
1992                         snprintf (p, s, "|mprof-report '--out=%s' -", nf);
1993                         free (nf);
1994                         nf = p;
1995                 }
1996         }
1997         if (*nf == '|') {
1998                 prof->file = popen (nf + 1, "w");
1999                 prof->pipe_output = 1;
2000         } else {
2001                 FILE *f;
2002                 if (force_delete)
2003                         unlink (nf);
2004                 if ((f = fopen (nf, "r"))) {
2005                         fclose (f);
2006                         fprintf (stderr, "The Mono profiler won't overwrite existing filename: %s.\n", nf);
2007                         fprintf (stderr, "Profiling disabled: use a different name or -FILENAME to force overwrite.\n");
2008                         free (prof);
2009                         return NULL;
2010                 }
2011                 prof->file = fopen (nf, "wb");
2012         }
2013         if (!prof->file) {
2014                 fprintf (stderr, "Cannot create profiler output: %s\n", nf);
2015                 exit (1);
2016         }
2017 #if defined (HAVE_SYS_ZLIB)
2018         if (use_zip)
2019                 prof->gzfile = gzdopen (fileno (prof->file), "wb");
2020 #endif
2021 #if USE_PERF_EVENTS
2022         if (sample_type && !do_mono_sample)
2023                 need_helper_thread = setup_perf_event ();
2024         if (perf_fd < 0) {
2025                 /* FIXME: warn if different freq or sample type */
2026                 do_mono_sample = 1;
2027         }
2028 #endif
2029         if (do_mono_sample) {
2030                 prof->stat_buffers = create_stat_buffer ();
2031                 need_helper_thread = 1;
2032         }
2033 #ifndef DISABLE_HELPER_THREAD
2034         if (hs_mode_ondemand || need_helper_thread) {
2035                 if (!start_helper_thread (prof))
2036                         prof->command_port = 0;
2037         }
2038 #else
2039         if (hs_mode_ondemand)
2040                 fprintf (stderr, "Ondemand heapshot unavailable on this arch.\n");
2041 #endif
2042         prof->startup_time = current_time ();
2043         dump_header (prof);
2044         return prof;
2045 }
2046
2047 static void
2048 usage (int do_exit)
2049 {
2050         printf ("Log profiler version %d.%d (format: %d)\n", LOG_VERSION_MAJOR, LOG_VERSION_MINOR, LOG_DATA_VERSION);
2051         printf ("Usage: mono --profile=log[:OPTION1[,OPTION2...]] program.exe\n");
2052         printf ("Options:\n");
2053         printf ("\thelp             show this usage info\n");
2054         printf ("\t[no]alloc        enable/disable recording allocation info\n");
2055         printf ("\t[no]calls        enable/disable recording enter/leave method events\n");
2056         printf ("\theapshot[=MODE]  record heap shot info (by default at each major collection)\n");
2057         printf ("\t                 MODE: every XXms milliseconds, every YYgc collections, ondemand\n");
2058         printf ("\tsample[=TYPE]    use statistical sampling mode (by default cycles/1000)\n");
2059         printf ("\t                 TYPE: cycles,instr,cacherefs,cachemiss,branches,branchmiss\n");
2060         printf ("\t                 TYPE can be followed by /FREQUENCY\n");
2061         printf ("\ttime=fast        use a faster (but more inaccurate) timer\n");
2062         printf ("\tmaxframes=NUM    collect up to NUM stack frames\n");
2063         printf ("\tcalldepth=NUM    ignore method events for call chain depth bigger than NUM\n");
2064         printf ("\toutput=FILENAME  write the data to file FILENAME (-FILENAME to overwrite)\n");
2065         printf ("\toutput=|PROGRAM  write the data to the stdin of PROGRAM\n");
2066         printf ("\t                 %%t is subtituted with date and time, %%p with the pid\n");
2067         printf ("\treport           create a report instead of writing the raw data to a file\n");
2068         printf ("\tzip              compress the output data\n");
2069         printf ("\tport=PORTNUM     use PORTNUM for the listening command server\n");
2070         if (do_exit)
2071                 exit (1);
2072 }
2073
2074 static const char*
2075 match_option (const char* p, const char *opt, char **rval)
2076 {
2077         int len = strlen (opt);
2078         if (strncmp (p, opt, len) == 0) {
2079                 if (rval) {
2080                         if (p [len] == '=' && p [len + 1]) {
2081                                 const char *opt = p + len + 1;
2082                                 const char *end = strchr (opt, ',');
2083                                 char *val;
2084                                 int l;
2085                                 if (end == NULL) {
2086                                         l = strlen (opt);
2087                                 } else {
2088                                         l = end - opt;
2089                                 }
2090                                 val = malloc (l + 1);
2091                                 memcpy (val, opt, l);
2092                                 val [l] = 0;
2093                                 *rval = val;
2094                                 return opt + l;
2095                         }
2096                         if (p [len] == 0 || p [len] == ',') {
2097                                 *rval = NULL;
2098                                 return p + len + (p [len] == ',');
2099                         }
2100                         usage (1);
2101                 } else {
2102                         if (p [len] == 0)
2103                                 return p + len;
2104                         if (p [len] == ',')
2105                                 return p + len + 1;
2106                 }
2107         }
2108         return p;
2109 }
2110
2111 typedef struct {
2112         const char *name;
2113         int sample_mode;
2114 } SampleMode;
2115
2116 static const SampleMode sample_modes [] = {
2117         {"cycles", SAMPLE_CYCLES},
2118         {"instr", SAMPLE_INSTRUCTIONS},
2119         {"cachemiss", SAMPLE_CACHE_MISSES},
2120         {"cacherefs", SAMPLE_CACHE_REFS},
2121         {"branches", SAMPLE_BRANCHES},
2122         {"branchmiss", SAMPLE_BRANCH_MISSES},
2123         {NULL, 0}
2124 };
2125
2126 static void
2127 set_sample_mode (char* val, int allow_empty)
2128 {
2129         char *end;
2130         char *maybe_freq = NULL;
2131         unsigned int count;
2132         const SampleMode *smode = sample_modes;
2133 #ifndef USE_PERF_EVENTS
2134         do_mono_sample = 1;
2135 #endif
2136         if (allow_empty && !val) {
2137                 sample_type = SAMPLE_CYCLES;
2138                 sample_freq = 1000;
2139                 return;
2140         }
2141         if (strcmp (val, "mono") == 0) {
2142                 do_mono_sample = 1;
2143                 sample_type = SAMPLE_CYCLES;
2144                 free (val);
2145                 return;
2146         }
2147         for (smode = sample_modes; smode->name; smode++) {
2148                 int l = strlen (smode->name);
2149                 if (strncmp (val, smode->name, l) == 0) {
2150                         sample_type = smode->sample_mode;
2151                         maybe_freq = val + l;
2152                         break;
2153                 }
2154         }
2155         if (!smode->name)
2156                 usage (1);
2157         if (*maybe_freq == '/') {
2158                 count = strtoul (maybe_freq + 1, &end, 10);
2159                 if (maybe_freq + 1 == end)
2160                         usage (1);
2161                 sample_freq = count;
2162         } else if (*maybe_freq != 0) {
2163                 usage (1);
2164         } else {
2165                 sample_freq = 1000;
2166         }
2167         free (val);
2168 }
2169
2170 static void
2171 set_hsmode (char* val, int allow_empty)
2172 {
2173         char *end;
2174         unsigned int count;
2175         if (allow_empty && !val)
2176                 return;
2177         if (strcmp (val, "ondemand") == 0) {
2178                 hs_mode_ondemand = 1;
2179                 free (val);
2180                 return;
2181         }
2182         count = strtoul (val, &end, 10);
2183         if (val == end)
2184                 usage (1);
2185         if (strcmp (end, "ms") == 0)
2186                 hs_mode_ms = count;
2187         else if (strcmp (end, "gc") == 0)
2188                 hs_mode_gc = count;
2189         else
2190                 usage (1);
2191         free (val);
2192 }
2193
2194 /*
2195  * declaration to silence the compiler: this is the entry point that
2196  * mono will load from the shared library and call.
2197  */
2198 extern void
2199 mono_profiler_startup (const char *desc);
2200
2201 void
2202 mono_profiler_startup (const char *desc)
2203 {
2204         MonoProfiler *prof;
2205         char *filename = NULL;
2206         const char *p;
2207         const char *opt;
2208         int fast_time = 0;
2209         int calls_enabled = 0;
2210         int allocs_enabled = 0;
2211         int events = MONO_PROFILE_GC|MONO_PROFILE_ALLOCATIONS|
2212                 MONO_PROFILE_GC_MOVES|MONO_PROFILE_CLASS_EVENTS|MONO_PROFILE_THREADS|
2213                 MONO_PROFILE_ENTER_LEAVE|MONO_PROFILE_JIT_COMPILATION|MONO_PROFILE_EXCEPTIONS|
2214                 MONO_PROFILE_MONITOR_EVENTS|MONO_PROFILE_MODULE_EVENTS|MONO_PROFILE_GC_ROOTS;
2215
2216         p = desc;
2217         if (strncmp (p, "log", 3))
2218                 usage (1);
2219         p += 3;
2220         if (*p == ':')
2221                 p++;
2222         for (; *p; p = opt) {
2223                 char *val;
2224                 if (*p == ',') {
2225                         opt = p + 1;
2226                         continue;
2227                 }
2228                 if ((opt = match_option (p, "help", NULL)) != p) {
2229                         usage (0);
2230                         continue;
2231                 }
2232                 if ((opt = match_option (p, "calls", NULL)) != p) {
2233                         calls_enabled = 1;
2234                         continue;
2235                 }
2236                 if ((opt = match_option (p, "nocalls", NULL)) != p) {
2237                         events &= ~MONO_PROFILE_ENTER_LEAVE;
2238                         nocalls = 1;
2239                         continue;
2240                 }
2241                 if ((opt = match_option (p, "alloc", NULL)) != p) {
2242                         allocs_enabled = 1;
2243                         continue;
2244                 }
2245                 if ((opt = match_option (p, "noalloc", NULL)) != p) {
2246                         events &= ~MONO_PROFILE_ALLOCATIONS;
2247                         continue;
2248                 }
2249                 if ((opt = match_option (p, "time", &val)) != p) {
2250                         if (strcmp (val, "fast") == 0)
2251                                 fast_time = 1;
2252                         else if (strcmp (val, "null") == 0)
2253                                 fast_time = 2;
2254                         else
2255                                 usage (1);
2256                         free (val);
2257                         continue;
2258                 }
2259                 if ((opt = match_option (p, "report", NULL)) != p) {
2260                         do_report = 1;
2261                         continue;
2262                 }
2263                 if ((opt = match_option (p, "debug", NULL)) != p) {
2264                         do_debug = 1;
2265                         continue;
2266                 }
2267                 if ((opt = match_option (p, "heapshot", &val)) != p) {
2268                         events &= ~MONO_PROFILE_ALLOCATIONS;
2269                         events &= ~MONO_PROFILE_ENTER_LEAVE;
2270                         nocalls = 1;
2271                         do_heap_shot = 1;
2272                         set_hsmode (val, 1);
2273                         continue;
2274                 }
2275                 if ((opt = match_option (p, "sample", &val)) != p) {
2276                         events &= ~MONO_PROFILE_ALLOCATIONS;
2277                         events &= ~MONO_PROFILE_ENTER_LEAVE;
2278                         nocalls = 1;
2279                         set_sample_mode (val, 1);
2280                         continue;
2281                 }
2282                 if ((opt = match_option (p, "hsmode", &val)) != p) {
2283                         fprintf (stderr, "The hsmode profiler option is obsolete, use heapshot=MODE.\n");
2284                         set_hsmode (val, 0);
2285                         continue;
2286                 }
2287                 if ((opt = match_option (p, "zip", NULL)) != p) {
2288                         use_zip = 1;
2289                         continue;
2290                 }
2291                 if ((opt = match_option (p, "output", &val)) != p) {
2292                         filename = val;
2293                         continue;
2294                 }
2295                 if ((opt = match_option (p, "port", &val)) != p) {
2296                         char *end;
2297                         command_port = strtoul (val, &end, 10);
2298                         free (val);
2299                         continue;
2300                 }
2301                 if ((opt = match_option (p, "maxframes", &val)) != p) {
2302                         char *end;
2303                         num_frames = strtoul (val, &end, 10);
2304                         if (num_frames > MAX_FRAMES)
2305                                 num_frames = MAX_FRAMES;
2306                         free (val);
2307                         notraces = num_frames == 0;
2308                         continue;
2309                 }
2310                 if ((opt = match_option (p, "calldepth", &val)) != p) {
2311                         char *end;
2312                         max_call_depth = strtoul (val, &end, 10);
2313                         free (val);
2314                         continue;
2315                 }
2316                 if (opt == p) {
2317                         usage (0);
2318                         exit (0);
2319                 }
2320         }
2321         if (calls_enabled) {
2322                 events |= MONO_PROFILE_ENTER_LEAVE;
2323                 nocalls = 0;
2324         }
2325         if (allocs_enabled)
2326                 events |= MONO_PROFILE_ALLOCATIONS;
2327         utils_init (fast_time);
2328
2329         prof = create_profiler (filename);
2330         if (!prof)
2331                 return;
2332         init_thread ();
2333
2334         mono_profiler_install (prof, log_shutdown);
2335         mono_profiler_install_gc (gc_event, gc_resize);
2336         mono_profiler_install_allocation (gc_alloc);
2337         mono_profiler_install_gc_moves (gc_moves);
2338         mono_profiler_install_gc_roots (gc_handle, gc_roots);
2339         mono_profiler_install_class (NULL, class_loaded, NULL, NULL);
2340         mono_profiler_install_module (NULL, image_loaded, NULL, NULL);
2341         mono_profiler_install_thread (thread_start, thread_end);
2342         mono_profiler_install_thread_name (thread_name);
2343         mono_profiler_install_enter_leave (method_enter, method_leave);
2344         mono_profiler_install_jit_end (method_jitted);
2345         mono_profiler_install_exception (throw_exc, method_exc_leave, clause_exc);
2346         mono_profiler_install_monitor (monitor_event);
2347         mono_profiler_install_runtime_initialized (runtime_initialized);
2348
2349
2350         if (do_mono_sample && sample_type == SAMPLE_CYCLES) {
2351                 events |= MONO_PROFILE_STATISTICAL;
2352                 mono_profiler_install_statistical (mono_sample_hit);
2353         }
2354
2355         mono_profiler_set_events (events);
2356
2357         TLS_INIT (tlsbuffer);
2358 }
2359