2 * proflog.c: mono log profiler
5 * Paolo Molaro (lupus@ximian.com)
7 * Copyright 2010 Novell, Inc (http://www.novell.com)
8 * Copyright 2011 Xamarin Inc (http://www.xamarin.com)
12 #include <mono/metadata/profiler.h>
13 #include <mono/metadata/threads.h>
14 #include <mono/metadata/mono-gc.h>
15 #include <mono/metadata/debug-helpers.h>
16 #include <mono/utils/atomic.h>
17 #include <mono/utils/mono-membar.h>
18 #include <mono/utils/mono-counters.h>
28 #if defined(HOST_WIN32) || defined(DISABLE_SOCKETS)
29 #define DISABLE_HELPER_THREAD 1
38 #ifdef HAVE_EXECINFO_H
45 #ifndef DISABLE_HELPER_THREAD
46 #include <sys/types.h>
47 #include <sys/socket.h>
48 #include <netinet/in.h>
49 #include <sys/select.h>
58 #ifdef HAVE_SYS_STAT_H
65 #if defined (HAVE_SYS_ZLIB)
69 /* the architecture needs a memory fence */
70 #if defined(__linux__) && (defined(__i386__) || defined(__x86_64__) || defined(__arm__))
72 #include <sys/syscall.h>
73 #include "perf_event.h"
74 #define USE_PERF_EVENTS 1
75 static int read_perf_mmap (MonoProfiler* prof, int cpu);
78 #define BUFFER_SIZE (4096 * 16)
79 static int nocalls = 0;
80 static int notraces = 0;
81 static int use_zip = 0;
82 static int do_report = 0;
83 static int do_heap_shot = 0;
84 static int max_call_depth = 100;
85 static int runtime_inited = 0;
86 static int command_port = 0;
87 static int heapshot_requested = 0;
88 static int sample_type = 0;
89 static int sample_freq = 0;
90 static int do_mono_sample = 0;
91 static int in_shutdown = 0;
92 static int do_debug = 0;
93 static int do_counters = 0;
95 /* For linux compile with:
96 * gcc -fPIC -shared -o libmono-profiler-log.so proflog.c utils.c -Wall -g -lz `pkg-config --cflags --libs mono-2`
97 * gcc -o mprof-report decode.c utils.c -Wall -g -lz -lrt -lpthread `pkg-config --cflags mono-2`
99 * For osx compile with:
100 * 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
101 * gcc -m32 -o mprof-report decode.c utils.c -Wall -g -lz -lrt -lpthread `pkg-config --cflags mono-2`
104 * sudo cp mprof-report /usr/local/bin
105 * sudo cp libmono-profiler-log.so /usr/local/lib
109 typedef struct _LogBuffer LogBuffer;
115 * The file is composed by a header followed by 0 or more buffers.
116 * Each buffer contains events that happened on a thread: for a given thread
117 * buffers that appear later in the file are guaranteed to contain events
118 * that happened later in time. Buffers from separate threads could be interleaved,
120 * Buffers are not required to be aligned.
123 * [id: 4 bytes] constant value: LOG_HEADER_ID
124 * [major: 1 byte] [minor: 1 byte] major and minor version of the log profiler
125 * [format: 1 byte] version of the data format for the rest of the file
126 * [ptrsize: 1 byte] size in bytes of a pointer in the profiled program
127 * [startup time: 8 bytes] time in milliseconds since the unix epoch when the program started
128 * [timer overhead: 4 bytes] approximate overhead in nanoseconds of the timer
129 * [flags: 4 bytes] file format flags, should be 0 for now
130 * [pid: 4 bytes] pid of the profiled process
131 * [port: 2 bytes] tcp port for server if != 0
132 * [sysid: 2 bytes] operating system and architecture identifier
134 * The multiple byte integers are in little-endian format.
137 * [buffer header] [event]*
138 * Buffers have a fixed-size header followed by 0 or more bytes of event data.
139 * Timing information and other values in the event data are usually stored
140 * as uleb128 or sleb128 integers. To save space, as noted for each item below,
141 * some data is represented as a difference between the actual value and
142 * either the last value of the same type (like for timing information) or
143 * as the difference from a value stored in a buffer header.
145 * For timing information the data is stored as uleb128, since timing
146 * increases in a monotonic way in each thread: the value is the number of
147 * nanoseconds to add to the last seen timing data in a buffer. The first value
148 * in a buffer will be calculated from the time_base field in the buffer head.
150 * Object or heap sizes are stored as uleb128.
151 * Pointer differences are stored as sleb128, instead.
153 * If an unexpected value is found, the rest of the buffer should be ignored,
154 * as generally the later values need the former to be interpreted correctly.
156 * buffer header format:
157 * [bufid: 4 bytes] constant value: BUF_ID
158 * [len: 4 bytes] size of the data following the buffer header
159 * [time_base: 8 bytes] time base in nanoseconds since an unspecified epoch
160 * [ptr_base: 8 bytes] base value for pointers
161 * [obj_base: 8 bytes] base value for object addresses
162 * [thread id: 8 bytes] system-specific thread ID (pthread_t for example)
163 * [method_base: 8 bytes] base value for MonoMethod pointers
166 * [extended info: upper 4 bits] [type: lower 4 bits] [data]*
167 * The data that follows depends on type and the extended info.
168 * Type is one of the enum values in proflog.h: TYPE_ALLOC, TYPE_GC,
169 * TYPE_METADATA, TYPE_METHOD, TYPE_EXCEPTION, TYPE_MONITOR, TYPE_HEAP.
170 * The extended info bits are interpreted based on type, see
171 * each individual event description below.
172 * strings are represented as a 0-terminated utf8 sequence.
175 * [flags: uleb128] must be 0
176 * [num: uleb128] number of frames following
177 * [frame: sleb128]* num MonoMethod pointers as differences from ptr_base
181 * exinfo: flags: TYPE_ALLOC_BT
182 * [time diff: uleb128] nanoseconds since last timing
183 * [ptr: sleb128] class as a byte difference from ptr_base
184 * [obj: sleb128] object address as a byte difference from obj_base
185 * [size: uleb128] size of the object in the heap
186 * If the TYPE_ALLOC_BT flag is set, a backtrace follows.
190 * exinfo: one of TYPE_GC_EVENT, TYPE_GC_RESIZE, TYPE_GC_MOVE, TYPE_GC_HANDLE_CREATED,
191 * TYPE_GC_HANDLE_DESTROYED
192 * [time diff: uleb128] nanoseconds since last timing
193 * if exinfo == TYPE_GC_RESIZE
194 * [heap_size: uleb128] new heap size
195 * if exinfo == TYPE_GC_EVENT
196 * [event type: uleb128] GC event (MONO_GC_EVENT_* from profiler.h)
197 * [generation: uleb128] GC generation event refers to
198 * if exinfo == TYPE_GC_MOVE
199 * [num_objects: uleb128] number of object moves that follow
200 * [objaddr: sleb128]+ num_objects object pointer differences from obj_base
201 * num is always an even number: the even items are the old
202 * addresses, the odd numbers are the respective new object addresses
203 * if exinfo == TYPE_GC_HANDLE_CREATED
204 * [handle_type: uleb128] GC handle type (System.Runtime.InteropServices.GCHandleType)
205 * upper bits reserved as flags
206 * [handle: uleb128] GC handle value
207 * [objaddr: sleb128] object pointer differences from obj_base
208 * if exinfo == TYPE_GC_HANDLE_DESTROYED
209 * [handle_type: uleb128] GC handle type (System.Runtime.InteropServices.GCHandleType)
210 * upper bits reserved as flags
211 * [handle: uleb128] GC handle value
213 * type metadata format:
214 * type: TYPE_METADATA
215 * exinfo: flags: TYPE_LOAD_ERR
216 * [time diff: uleb128] nanoseconds since last timing
217 * [mtype: byte] metadata type, one of: TYPE_CLASS, TYPE_IMAGE, TYPE_ASSEMBLY, TYPE_DOMAIN,
219 * [pointer: sleb128] pointer of the metadata type depending on mtype
220 * if mtype == TYPE_CLASS
221 * [image: sleb128] MonoImage* as a pointer difference from ptr_base
222 * [flags: uleb128] must be 0
223 * [name: string] full class name
224 * if mtype == TYPE_IMAGE
225 * [flags: uleb128] must be 0
226 * [name: string] image file name
227 * if mtype == TYPE_THREAD
228 * [flags: uleb128] must be 0
229 * [name: string] thread name
231 * type method format:
233 * exinfo: one of: TYPE_LEAVE, TYPE_ENTER, TYPE_EXC_LEAVE, TYPE_JIT
234 * [time diff: uleb128] nanoseconds since last timing
235 * [method: sleb128] MonoMethod* as a pointer difference from the last such
236 * pointer or the buffer method_base
237 * if exinfo == TYPE_JIT
238 * [code address: sleb128] pointer to the native code as a diff from ptr_base
239 * [code size: uleb128] size of the generated code
240 * [name: string] full method name
242 * type exception format:
243 * type: TYPE_EXCEPTION
244 * exinfo: TYPE_EXCEPTION_BT flag and one of: TYPE_THROW, TYPE_CLAUSE
245 * [time diff: uleb128] nanoseconds since last timing
246 * if exinfo.low3bits == TYPE_CLAUSE
247 * [clause type: uleb128] finally/catch/fault/filter
248 * [clause num: uleb128] the clause number in the method header
249 * [method: sleb128] MonoMethod* as a pointer difference from the last such
250 * pointer or the buffer method_base
251 * if exinfo.low3bits == TYPE_THROW
252 * [object: sleb128] the object that was thrown as a difference from obj_base
253 * If the TYPE_EXCEPTION_BT flag is set, a backtrace follows.
255 * type monitor format:
257 * exinfo: TYPE_MONITOR_BT flag and one of: MONO_PROFILER_MONITOR_(CONTENTION|FAIL|DONE)
258 * [time diff: uleb128] nanoseconds since last timing
259 * [object: sleb128] the lock object as a difference from obj_base
260 * if exinfo.low3bits == MONO_PROFILER_MONITOR_CONTENTION
261 * If the TYPE_MONITOR_BT flag is set, a backtrace follows.
265 * exinfo: one of TYPE_HEAP_START, TYPE_HEAP_END, TYPE_HEAP_OBJECT, TYPE_HEAP_ROOT
266 * if exinfo == TYPE_HEAP_START
267 * [time diff: uleb128] nanoseconds since last timing
268 * if exinfo == TYPE_HEAP_END
269 * [time diff: uleb128] nanoseconds since last timing
270 * if exinfo == TYPE_HEAP_OBJECT
271 * [object: sleb128] the object as a difference from obj_base
272 * [class: sleb128] the object MonoClass* as a difference from ptr_base
273 * [size: uleb128] size of the object on the heap
274 * [num_refs: uleb128] number of object references
275 * if (format version > 1) each referenced objref is preceded by a
276 * uleb128 encoded offset: the first offset is from the object address
277 * and each next offset is relative to the previous one
278 * [objrefs: sleb128]+ object referenced as a difference from obj_base
279 * The same object can appear multiple times, but only the first time
280 * with size != 0: in the other cases this data will only be used to
281 * provide additional referenced objects.
282 * if exinfo == TYPE_HEAP_ROOT
283 * [num_roots: uleb128] number of root references
284 * [num_gc: uleb128] number of major gcs
285 * [object: sleb128] the object as a difference from obj_base
286 * [root_type: uleb128] the root_type: MonoProfileGCRootType (profiler.h)
287 * [extra_info: uleb128] the extra_info value
288 * object, root_type and extra_info are repeated num_roots times
292 * exinfo: one of TYPE_SAMPLE_HIT, TYPE_SAMPLE_USYM, TYPE_SAMPLE_UBIN, TYPE_SAMPLE_COUNTERS_DESC, TYPE_SAMPLE_COUNTERS
293 * if exinfo == TYPE_SAMPLE_HIT
294 * [sample_type: uleb128] type of sample (SAMPLE_*)
295 * [timestamp: uleb128] nanoseconds since startup (note: different from other timestamps!)
296 * [count: uleb128] number of following instruction addresses
297 * [ip: sleb128]* instruction pointer as difference from ptr_base
298 * if (format_version > 5)
299 * [mbt_count: uleb128] number of managed backtrace info triplets (method + IL offset + native offset)
300 * [method: sleb128]* MonoMethod* as a pointer difference from the last such
301 * pointer or the buffer method_base (the first such method can be also indentified by ip, but this is not neccessarily true)
302 * [il_offset: sleb128]* IL offset inside method where the hit occurred
303 * [native_offset: sleb128]* native offset inside method where the hit occurred
304 * if exinfo == TYPE_SAMPLE_USYM
305 * [address: sleb128] symbol address as a difference from ptr_base
306 * [size: uleb128] symbol size (may be 0 if unknown)
307 * [name: string] symbol name
308 * if exinfo == TYPE_SAMPLE_UBIN
309 * [time diff: uleb128] nanoseconds since last timing
310 * [address: sleb128] address where binary has been loaded
311 * [offset: uleb128] file offset of mapping (the same file can be mapped multiple times)
312 * [size: uleb128] memory size
313 * [name: string] binary name
314 * if exinfo == TYPE_SAMPLE_COUNTERS_DESC
315 * [len: uleb128] number of counters
317 * [section: uleb128] section name of counter
318 * [name: string] name of counter
319 * [type: uleb128] type name of counter
320 * [unit: uleb128] unit name of counter
321 * [variance: uleb128] variance name of counter
322 * [index: uleb128] unique index of counter
323 * if exinfo == TYPE_SAMPLE_COUNTERS
324 * [timestamp: uleb128] sampling timestamp
326 * [index: uleb128] unique index of counter
329 * [type: uleb128] type of counter value
332 * [0: uleb128] 0 -> value is null
334 * [1: uleb128] 1 -> value is not null
335 * [value: string] counter value
337 * [value: uleb128/sleb128/double] counter value, can be sleb128, uleb128 or double (determined by using type)
345 uintptr_t method_base;
346 uintptr_t last_method;
349 unsigned char* data_end;
354 unsigned char buf [1];
357 #define ENTER_LOG(lb,str) if ((lb)->locked) {write(2, str, strlen(str)); write(2, "\n", 1);return;} else {(lb)->locked++;}
358 #define EXIT_LOG(lb) (lb)->locked--;
360 typedef struct _StatBuffer StatBuffer;
369 typedef struct _BinaryObject BinaryObject;
371 struct _BinaryObject {
377 struct _MonoProfiler {
379 StatBuffer *stat_buffers;
381 #if defined (HAVE_SYS_ZLIB)
384 uint64_t startup_time;
386 int last_gc_gen_started;
391 pthread_t helper_thread;
393 BinaryObject *binary_objects;
397 #define TLS_SET(x,y) TlsSetValue(x, y)
398 #define TLS_GET(x) ((LogBuffer *) TlsGetValue(x))
399 #define TLS_INIT(x) x = TlsAlloc ()
400 static int tlsbuffer;
402 #define TLS_SET(x,y) x = y
405 static __thread LogBuffer* tlsbuffer = NULL;
407 #define TLS_SET(x,y) pthread_setspecific(x, y)
408 #define TLS_GET(x) ((LogBuffer *) pthread_getspecific(x))
409 #define TLS_INIT(x) pthread_key_create(&x, NULL)
410 static pthread_key_t tlsbuffer;
413 static void safe_dump (MonoProfiler *profiler, LogBuffer *logbuffer);
416 pstrdup (const char *s)
418 int len = strlen (s) + 1;
419 char *p = malloc (len);
425 create_stat_buffer (void)
427 StatBuffer* buf = alloc_buffer (BUFFER_SIZE);
428 buf->size = BUFFER_SIZE;
429 buf->data_end = (uintptr_t*)((unsigned char*)buf + buf->size);
430 buf->data = buf->buf;
437 LogBuffer* buf = alloc_buffer (BUFFER_SIZE);
438 buf->size = BUFFER_SIZE;
439 buf->time_base = current_time ();
440 buf->last_time = buf->time_base;
441 buf->data_end = (unsigned char*)buf + buf->size;
442 buf->data = buf->buf;
449 LogBuffer *logbuffer;
450 if (TLS_GET (tlsbuffer))
452 logbuffer = create_buffer ();
453 TLS_SET (tlsbuffer, logbuffer);
454 logbuffer->thread_id = thread_id ();
455 //printf ("thread %p at time %llu\n", (void*)logbuffer->thread_id, logbuffer->time_base);
459 ensure_logbuf (int bytes)
461 LogBuffer *old = TLS_GET (tlsbuffer);
462 if (old && old->data + bytes + 100 < old->data_end)
464 TLS_SET (tlsbuffer, NULL);
466 TLS_GET (tlsbuffer)->next = old;
468 TLS_GET (tlsbuffer)->call_depth = old->call_depth;
469 //printf ("new logbuffer\n");
470 return TLS_GET (tlsbuffer);
474 emit_byte (LogBuffer *logbuffer, int value)
476 logbuffer->data [0] = value;
478 assert (logbuffer->data <= logbuffer->data_end);
482 emit_value (LogBuffer *logbuffer, int value)
484 encode_uleb128 (value, logbuffer->data, &logbuffer->data);
485 assert (logbuffer->data <= logbuffer->data_end);
489 emit_time (LogBuffer *logbuffer, uint64_t value)
491 uint64_t tdiff = value - logbuffer->last_time;
493 if (value < logbuffer->last_time)
494 printf ("time went backwards\n");
495 //if (tdiff > 1000000)
496 // printf ("large time offset: %llu\n", tdiff);
498 encode_uleb128 (tdiff, logbuffer->data, &logbuffer->data);
499 /*if (tdiff != decode_uleb128 (p, &p))
500 printf ("incorrect encoding: %llu\n", tdiff);*/
501 logbuffer->last_time = value;
502 assert (logbuffer->data <= logbuffer->data_end);
506 emit_svalue (LogBuffer *logbuffer, int64_t value)
508 encode_sleb128 (value, logbuffer->data, &logbuffer->data);
509 assert (logbuffer->data <= logbuffer->data_end);
513 emit_uvalue (LogBuffer *logbuffer, uint64_t value)
515 encode_uleb128 (value, logbuffer->data, &logbuffer->data);
516 assert (logbuffer->data <= logbuffer->data_end);
520 emit_ptr (LogBuffer *logbuffer, void *ptr)
522 if (!logbuffer->ptr_base)
523 logbuffer->ptr_base = (uintptr_t)ptr;
524 emit_svalue (logbuffer, (intptr_t)ptr - logbuffer->ptr_base);
525 assert (logbuffer->data <= logbuffer->data_end);
529 emit_method (LogBuffer *logbuffer, void *method)
531 if (!logbuffer->method_base) {
532 logbuffer->method_base = (intptr_t)method;
533 logbuffer->last_method = (intptr_t)method;
535 encode_sleb128 ((intptr_t)((char*)method - (char*)logbuffer->last_method), logbuffer->data, &logbuffer->data);
536 logbuffer->last_method = (intptr_t)method;
537 assert (logbuffer->data <= logbuffer->data_end);
541 emit_obj (LogBuffer *logbuffer, void *ptr)
543 if (!logbuffer->obj_base)
544 logbuffer->obj_base = (uintptr_t)ptr >> 3;
545 emit_svalue (logbuffer, ((uintptr_t)ptr >> 3) - logbuffer->obj_base);
546 assert (logbuffer->data <= logbuffer->data_end);
550 emit_string (LogBuffer *logbuffer, const char *str, size_t size)
554 for (; i < size; i++) {
557 emit_byte (logbuffer, str [i]);
560 emit_byte (logbuffer, '\0');
564 emit_double (LogBuffer *logbuffer, double value)
567 unsigned char buffer[8];
568 memcpy (buffer, &value, 8);
569 #if G_BYTE_ORDER == G_BIG_ENDIAN
570 for (i = 7; i >= 0; i--)
572 for (i = 0; i < 8; i++)
574 emit_byte (logbuffer, buffer[i]);
578 write_int16 (char *buf, int32_t value)
581 for (i = 0; i < 2; ++i) {
589 write_int32 (char *buf, int32_t value)
592 for (i = 0; i < 4; ++i) {
600 write_int64 (char *buf, int64_t value)
603 for (i = 0; i < 8; ++i) {
611 dump_header (MonoProfiler *profiler)
615 p = write_int32 (p, LOG_HEADER_ID);
616 *p++ = LOG_VERSION_MAJOR;
617 *p++ = LOG_VERSION_MINOR;
618 *p++ = LOG_DATA_VERSION;
619 *p++ = sizeof (void*);
620 p = write_int64 (p, ((uint64_t)time (NULL)) * 1000); /* startup time */
621 p = write_int32 (p, get_timer_overhead ()); /* timer overhead */
622 p = write_int32 (p, 0); /* flags */
623 p = write_int32 (p, process_id ()); /* pid */
624 p = write_int16 (p, profiler->command_port); /* port */
625 p = write_int16 (p, 0); /* opsystem */
626 #if defined (HAVE_SYS_ZLIB)
627 if (profiler->gzfile) {
628 gzwrite (profiler->gzfile, hbuf, p - hbuf);
630 fwrite (hbuf, p - hbuf, 1, profiler->file);
633 fwrite (hbuf, p - hbuf, 1, profiler->file);
634 fflush (profiler->file);
639 dump_buffer (MonoProfiler *profiler, LogBuffer *buf)
644 dump_buffer (profiler, buf->next);
645 p = write_int32 (p, BUF_ID);
646 p = write_int32 (p, buf->data - buf->buf);
647 p = write_int64 (p, buf->time_base);
648 p = write_int64 (p, buf->ptr_base);
649 p = write_int64 (p, buf->obj_base);
650 p = write_int64 (p, buf->thread_id);
651 p = write_int64 (p, buf->method_base);
652 #if defined (HAVE_SYS_ZLIB)
653 if (profiler->gzfile) {
654 gzwrite (profiler->gzfile, hbuf, p - hbuf);
655 gzwrite (profiler->gzfile, buf->buf, buf->data - buf->buf);
658 fwrite (hbuf, p - hbuf, 1, profiler->file);
659 fwrite (buf->buf, buf->data - buf->buf, 1, profiler->file);
660 fflush (profiler->file);
661 #if defined (HAVE_SYS_ZLIB)
664 free_buffer (buf, buf->size);
668 process_requests (MonoProfiler *profiler)
670 if (heapshot_requested)
671 mono_gc_collect (mono_gc_max_generation ());
674 static void counters_init (MonoProfiler *profiler);
677 runtime_initialized (MonoProfiler *profiler)
680 #ifndef DISABLE_HELPER_THREAD
681 counters_init (profiler);
683 /* ensure the main thread data and startup are available soon */
684 safe_dump (profiler, ensure_logbuf (0));
688 * Can be called only at safe callback locations.
691 safe_dump (MonoProfiler *profiler, LogBuffer *logbuffer)
693 int cd = logbuffer->call_depth;
695 dump_buffer (profiler, TLS_GET (tlsbuffer));
697 TLS_SET (tlsbuffer, NULL);
699 TLS_GET (tlsbuffer)->call_depth = cd;
703 gc_reference (MonoObject *obj, MonoClass *klass, uintptr_t size, uintptr_t num, MonoObject **refs, uintptr_t *offsets, void *data)
706 uintptr_t last_offset = 0;
707 //const char *name = mono_class_get_name (klass);
708 LogBuffer *logbuffer = ensure_logbuf (20 + num * 8);
709 emit_byte (logbuffer, TYPE_HEAP_OBJECT | TYPE_HEAP);
710 emit_obj (logbuffer, obj);
711 emit_ptr (logbuffer, klass);
712 /* account for object alignment in the heap */
715 emit_value (logbuffer, size);
716 emit_value (logbuffer, num);
717 for (i = 0; i < num; ++i) {
718 emit_value (logbuffer, offsets [i] - last_offset);
719 last_offset = offsets [i];
720 emit_obj (logbuffer, refs [i]);
723 // printf ("obj: %p, klass: %s, refs: %d, size: %d\n", obj, name, (int)num, (int)size);
727 static unsigned int hs_mode_ms = 0;
728 static unsigned int hs_mode_gc = 0;
729 static unsigned int hs_mode_ondemand = 0;
730 static unsigned int gc_count = 0;
731 static uint64_t last_hs_time = 0;
734 heap_walk (MonoProfiler *profiler)
738 LogBuffer *logbuffer;
741 logbuffer = ensure_logbuf (10);
742 now = current_time ();
743 if (hs_mode_ms && (now - last_hs_time)/1000000 >= hs_mode_ms)
745 else if (hs_mode_gc && (gc_count % hs_mode_gc) == 0)
747 else if (hs_mode_ondemand)
748 do_walk = heapshot_requested;
749 else if (!hs_mode_ms && !hs_mode_gc && profiler->last_gc_gen_started == mono_gc_max_generation ())
754 heapshot_requested = 0;
755 emit_byte (logbuffer, TYPE_HEAP_START | TYPE_HEAP);
756 emit_time (logbuffer, now);
757 mono_gc_walk_heap (0, gc_reference, NULL);
758 logbuffer = ensure_logbuf (10);
759 now = current_time ();
760 emit_byte (logbuffer, TYPE_HEAP_END | TYPE_HEAP);
761 emit_time (logbuffer, now);
766 gc_event (MonoProfiler *profiler, MonoGCEvent ev, int generation) {
768 LogBuffer *logbuffer = ensure_logbuf (10);
769 now = current_time ();
770 ENTER_LOG (logbuffer, "gcevent");
771 emit_byte (logbuffer, TYPE_GC_EVENT | TYPE_GC);
772 emit_time (logbuffer, now);
773 emit_value (logbuffer, ev);
774 emit_value (logbuffer, generation);
775 /* to deal with nested gen1 after gen0 started */
776 if (ev == MONO_GC_EVENT_START) {
777 profiler->last_gc_gen_started = generation;
778 if (generation == mono_gc_max_generation ())
781 if (ev == MONO_GC_EVENT_PRE_START_WORLD)
782 heap_walk (profiler);
783 EXIT_LOG (logbuffer);
784 if (ev == MONO_GC_EVENT_POST_START_WORLD)
785 safe_dump (profiler, logbuffer);
786 //printf ("gc event %d for generation %d\n", ev, generation);
790 gc_resize (MonoProfiler *profiler, int64_t new_size) {
792 LogBuffer *logbuffer = ensure_logbuf (10);
793 now = current_time ();
794 ENTER_LOG (logbuffer, "gcresize");
795 emit_byte (logbuffer, TYPE_GC_RESIZE | TYPE_GC);
796 emit_time (logbuffer, now);
797 emit_value (logbuffer, new_size);
798 //printf ("gc resized to %lld\n", new_size);
799 EXIT_LOG (logbuffer);
802 #define MAX_FRAMES 16
805 MonoMethod* methods [MAX_FRAMES];
806 int32_t il_offsets [MAX_FRAMES];
807 int32_t native_offsets [MAX_FRAMES];
809 static int num_frames = MAX_FRAMES / 2;
812 walk_stack (MonoMethod *method, int32_t native_offset, int32_t il_offset, mono_bool managed, void* data)
814 FrameData *frame = data;
815 if (method && frame->count < num_frames) {
816 frame->il_offsets [frame->count] = il_offset;
817 frame->native_offsets [frame->count] = native_offset;
818 frame->methods [frame->count++] = method;
819 //printf ("In %d %s at %d (native: %d)\n", frame->count, mono_method_get_name (method), il_offset, native_offset);
821 return frame->count == num_frames;
825 * a note about stack walks: they can cause more profiler events to fire,
826 * so we need to make sure they don't happen after we started emitting an
827 * event, hence the collect_bt/emit_bt split.
830 collect_bt (FrameData *data)
833 mono_stack_walk_no_il (walk_stack, data);
837 emit_bt (LogBuffer *logbuffer, FrameData *data)
839 /* FIXME: this is actually tons of data and we should
840 * just output it the first time and use an id the next
842 if (data->count > num_frames)
843 printf ("bad num frames: %d\n", data->count);
844 emit_value (logbuffer, 0); /* flags */
845 emit_value (logbuffer, data->count);
846 //if (*p != data.count) {
847 // 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);}
848 while (data->count) {
849 emit_ptr (logbuffer, data->methods [--data->count]);
854 gc_alloc (MonoProfiler *prof, MonoObject *obj, MonoClass *klass)
858 int do_bt = (nocalls && runtime_inited && !notraces)? TYPE_ALLOC_BT: 0;
860 LogBuffer *logbuffer;
861 len = mono_object_get_size (obj);
862 /* account for object alignment in the heap */
867 logbuffer = ensure_logbuf (32 + MAX_FRAMES * 8);
868 now = current_time ();
869 ENTER_LOG (logbuffer, "gcalloc");
870 emit_byte (logbuffer, do_bt | TYPE_ALLOC);
871 emit_time (logbuffer, now);
872 emit_ptr (logbuffer, klass);
873 emit_obj (logbuffer, obj);
874 emit_value (logbuffer, len);
876 emit_bt (logbuffer, &data);
877 EXIT_LOG (logbuffer);
879 safe_dump (prof, logbuffer);
880 process_requests (prof);
881 //printf ("gc alloc %s at %p\n", mono_class_get_name (klass), obj);
885 gc_moves (MonoProfiler *prof, void **objects, int num)
889 LogBuffer *logbuffer = ensure_logbuf (10 + num * 8);
890 now = current_time ();
891 ENTER_LOG (logbuffer, "gcmove");
892 emit_byte (logbuffer, TYPE_GC_MOVE | TYPE_GC);
893 emit_time (logbuffer, now);
894 emit_value (logbuffer, num);
895 for (i = 0; i < num; ++i)
896 emit_obj (logbuffer, objects [i]);
897 //printf ("gc moved %d objects\n", num/2);
898 EXIT_LOG (logbuffer);
902 gc_roots (MonoProfiler *prof, int num, void **objects, int *root_types, uintptr_t *extra_info)
905 LogBuffer *logbuffer = ensure_logbuf (5 + num * 18);
906 ENTER_LOG (logbuffer, "gcroots");
907 emit_byte (logbuffer, TYPE_HEAP_ROOT | TYPE_HEAP);
908 emit_value (logbuffer, num);
909 emit_value (logbuffer, mono_gc_collection_count (mono_gc_max_generation ()));
910 for (i = 0; i < num; ++i) {
911 emit_obj (logbuffer, objects [i]);
912 emit_value (logbuffer, root_types [i]);
913 emit_value (logbuffer, extra_info [i]);
915 EXIT_LOG (logbuffer);
919 gc_handle (MonoProfiler *prof, int op, int type, uintptr_t handle, MonoObject *obj)
922 LogBuffer *logbuffer = ensure_logbuf (16);
923 now = current_time ();
924 ENTER_LOG (logbuffer, "gchandle");
925 if (op == MONO_PROFILER_GC_HANDLE_CREATED)
926 emit_byte (logbuffer, TYPE_GC_HANDLE_CREATED | TYPE_GC);
927 else if (op == MONO_PROFILER_GC_HANDLE_DESTROYED)
928 emit_byte (logbuffer, TYPE_GC_HANDLE_DESTROYED | TYPE_GC);
931 emit_time (logbuffer, now);
932 emit_value (logbuffer, type);
933 emit_value (logbuffer, handle);
934 if (op == MONO_PROFILER_GC_HANDLE_CREATED)
935 emit_obj (logbuffer, obj);
936 EXIT_LOG (logbuffer);
937 process_requests (prof);
941 push_nesting (char *p, MonoClass *klass)
946 nesting = mono_class_get_nesting_type (klass);
948 p = push_nesting (p, nesting);
952 name = mono_class_get_name (klass);
953 nspace = mono_class_get_namespace (klass);
956 p += strlen (nspace);
966 type_name (MonoClass *klass)
970 push_nesting (buf, klass);
971 p = malloc (strlen (buf) + 1);
977 image_loaded (MonoProfiler *prof, MonoImage *image, int result)
982 LogBuffer *logbuffer;
983 if (result != MONO_PROFILE_OK)
985 name = mono_image_get_filename (image);
986 nlen = strlen (name) + 1;
987 logbuffer = ensure_logbuf (16 + nlen);
988 now = current_time ();
989 ENTER_LOG (logbuffer, "image");
990 emit_byte (logbuffer, TYPE_END_LOAD | TYPE_METADATA);
991 emit_time (logbuffer, now);
992 emit_byte (logbuffer, TYPE_IMAGE);
993 emit_ptr (logbuffer, image);
994 emit_value (logbuffer, 0); /* flags */
995 memcpy (logbuffer->data, name, nlen);
996 logbuffer->data += nlen;
997 //printf ("loaded image %p (%s)\n", image, name);
998 EXIT_LOG (logbuffer);
1000 safe_dump (prof, logbuffer);
1001 process_requests (prof);
1005 class_loaded (MonoProfiler *prof, MonoClass *klass, int result)
1011 LogBuffer *logbuffer;
1012 if (result != MONO_PROFILE_OK)
1015 name = mono_type_get_name (mono_class_get_type (klass));
1017 name = type_name (klass);
1018 nlen = strlen (name) + 1;
1019 image = mono_class_get_image (klass);
1020 logbuffer = ensure_logbuf (24 + nlen);
1021 now = current_time ();
1022 ENTER_LOG (logbuffer, "class");
1023 emit_byte (logbuffer, TYPE_END_LOAD | TYPE_METADATA);
1024 emit_time (logbuffer, now);
1025 emit_byte (logbuffer, TYPE_CLASS);
1026 emit_ptr (logbuffer, klass);
1027 emit_ptr (logbuffer, image);
1028 emit_value (logbuffer, 0); /* flags */
1029 memcpy (logbuffer->data, name, nlen);
1030 logbuffer->data += nlen;
1031 //printf ("loaded class %p (%s)\n", klass, name);
1036 EXIT_LOG (logbuffer);
1037 if (logbuffer->next)
1038 safe_dump (prof, logbuffer);
1039 process_requests (prof);
1043 method_enter (MonoProfiler *prof, MonoMethod *method)
1046 LogBuffer *logbuffer = ensure_logbuf (16);
1047 if (logbuffer->call_depth++ > max_call_depth)
1049 now = current_time ();
1050 ENTER_LOG (logbuffer, "enter");
1051 emit_byte (logbuffer, TYPE_ENTER | TYPE_METHOD);
1052 emit_time (logbuffer, now);
1053 emit_method (logbuffer, method);
1054 EXIT_LOG (logbuffer);
1055 process_requests (prof);
1059 method_leave (MonoProfiler *prof, MonoMethod *method)
1062 LogBuffer *logbuffer = ensure_logbuf (16);
1063 if (--logbuffer->call_depth > max_call_depth)
1065 now = current_time ();
1066 ENTER_LOG (logbuffer, "leave");
1067 emit_byte (logbuffer, TYPE_LEAVE | TYPE_METHOD);
1068 emit_time (logbuffer, now);
1069 emit_method (logbuffer, method);
1070 EXIT_LOG (logbuffer);
1071 if (logbuffer->next)
1072 safe_dump (prof, logbuffer);
1073 process_requests (prof);
1077 method_exc_leave (MonoProfiler *prof, MonoMethod *method)
1080 LogBuffer *logbuffer;
1083 logbuffer = ensure_logbuf (16);
1084 if (--logbuffer->call_depth > max_call_depth)
1086 now = current_time ();
1087 ENTER_LOG (logbuffer, "eleave");
1088 emit_byte (logbuffer, TYPE_EXC_LEAVE | TYPE_METHOD);
1089 emit_time (logbuffer, now);
1090 emit_method (logbuffer, method);
1091 EXIT_LOG (logbuffer);
1092 process_requests (prof);
1096 method_jitted (MonoProfiler *prof, MonoMethod *method, MonoJitInfo* jinfo, int result)
1101 LogBuffer *logbuffer;
1102 if (result != MONO_PROFILE_OK)
1104 name = mono_method_full_name (method, 1);
1105 nlen = strlen (name) + 1;
1106 logbuffer = ensure_logbuf (32 + nlen);
1107 now = current_time ();
1108 ENTER_LOG (logbuffer, "jit");
1109 emit_byte (logbuffer, TYPE_JIT | TYPE_METHOD);
1110 emit_time (logbuffer, now);
1111 emit_method (logbuffer, method);
1112 emit_ptr (logbuffer, mono_jit_info_get_code_start (jinfo));
1113 emit_value (logbuffer, mono_jit_info_get_code_size (jinfo));
1114 memcpy (logbuffer->data, name, nlen);
1115 logbuffer->data += nlen;
1117 EXIT_LOG (logbuffer);
1118 if (logbuffer->next)
1119 safe_dump (prof, logbuffer);
1120 process_requests (prof);
1124 throw_exc (MonoProfiler *prof, MonoObject *object)
1126 int do_bt = (nocalls && runtime_inited && !notraces)? TYPE_EXCEPTION_BT: 0;
1129 LogBuffer *logbuffer;
1132 logbuffer = ensure_logbuf (16 + MAX_FRAMES * 8);
1133 now = current_time ();
1134 ENTER_LOG (logbuffer, "throw");
1135 emit_byte (logbuffer, do_bt | TYPE_EXCEPTION);
1136 emit_time (logbuffer, now);
1137 emit_obj (logbuffer, object);
1139 emit_bt (logbuffer, &data);
1140 EXIT_LOG (logbuffer);
1141 process_requests (prof);
1145 clause_exc (MonoProfiler *prof, MonoMethod *method, int clause_type, int clause_num)
1148 LogBuffer *logbuffer = ensure_logbuf (16);
1149 now = current_time ();
1150 ENTER_LOG (logbuffer, "clause");
1151 emit_byte (logbuffer, TYPE_EXCEPTION | TYPE_CLAUSE);
1152 emit_time (logbuffer, now);
1153 emit_value (logbuffer, clause_type);
1154 emit_value (logbuffer, clause_num);
1155 emit_method (logbuffer, method);
1156 EXIT_LOG (logbuffer);
1160 monitor_event (MonoProfiler *profiler, MonoObject *object, MonoProfilerMonitorEvent event)
1162 int do_bt = (nocalls && runtime_inited && !notraces && event == MONO_PROFILER_MONITOR_CONTENTION)? TYPE_MONITOR_BT: 0;
1165 LogBuffer *logbuffer;
1168 logbuffer = ensure_logbuf (16 + MAX_FRAMES * 8);
1169 now = current_time ();
1170 ENTER_LOG (logbuffer, "monitor");
1171 emit_byte (logbuffer, (event << 4) | do_bt | TYPE_MONITOR);
1172 emit_time (logbuffer, now);
1173 emit_obj (logbuffer, object);
1175 emit_bt (logbuffer, &data);
1176 EXIT_LOG (logbuffer);
1177 process_requests (profiler);
1181 thread_start (MonoProfiler *prof, uintptr_t tid)
1183 //printf ("thread start %p\n", (void*)tid);
1188 thread_end (MonoProfiler *prof, uintptr_t tid)
1191 if (TLS_GET (tlsbuffer))
1192 dump_buffer (prof, TLS_GET (tlsbuffer));
1194 TLS_SET (tlsbuffer, NULL);
1198 thread_name (MonoProfiler *prof, uintptr_t tid, const char *name)
1200 int len = strlen (name) + 1;
1202 LogBuffer *logbuffer;
1203 logbuffer = ensure_logbuf (10 + len);
1204 now = current_time ();
1205 ENTER_LOG (logbuffer, "tname");
1206 emit_byte (logbuffer, TYPE_METADATA);
1207 emit_time (logbuffer, now);
1208 emit_byte (logbuffer, TYPE_THREAD);
1209 emit_ptr (logbuffer, (void*)tid);
1210 emit_value (logbuffer, 0); /* flags */
1211 memcpy (logbuffer->data, name, len);
1212 logbuffer->data += len;
1213 EXIT_LOG (logbuffer);
1217 mono_sample_hit (MonoProfiler *profiler, unsigned char *ip, void *context)
1222 uintptr_t *data, *new_data, *old_data;
1228 now = current_time ();
1229 collect_bt (&bt_data);
1230 elapsed = (now - profiler->startup_time) / 10000;
1234 snprintf (buf, sizeof (buf), "hit at %p in thread %p after %llu ms\n", ip, (void*)thread_id (), (unsigned long long int)elapsed/100);
1236 write (2, buf, len);
1238 sbuf = profiler->stat_buffers;
1241 /* flush the buffer at 1 second intervals */
1242 if (sbuf->data > sbuf->buf && (elapsed - sbuf->buf [2]) > 100000) {
1245 /* overflow: 400 slots is a big enough number to reduce the chance of losing this event if many
1246 * threads hit this same spot at the same time
1248 if (timedout || (sbuf->data + 400 >= sbuf->data_end)) {
1249 StatBuffer *oldsb, *foundsb;
1250 sbuf = create_stat_buffer ();
1252 oldsb = profiler->stat_buffers;
1254 foundsb = InterlockedCompareExchangePointer ((volatile void**)&profiler->stat_buffers, sbuf, oldsb);
1255 } while (foundsb != oldsb);
1257 write (2, "overflow\n", 9);
1258 /* notify the helper thread */
1259 if (sbuf->next->next) {
1261 write (profiler->pipes [1], &c, 1);
1263 write (2, "notify\n", 7);
1267 old_data = sbuf->data;
1268 new_data = old_data + 4 + bt_data.count * 3;
1269 data = InterlockedCompareExchangePointer ((volatile void**)&sbuf->data, new_data, old_data);
1270 } while (data != old_data);
1271 if (old_data >= sbuf->data_end)
1272 return; /* lost event */
1273 old_data [0] = 1 | (sample_type << 16) | (bt_data.count << 8);
1274 old_data [1] = thread_id ();
1275 old_data [2] = elapsed;
1276 old_data [3] = (uintptr_t)ip;
1277 for (i = 0; i < bt_data.count; ++i) {
1278 old_data [4+3*i] = (uintptr_t)bt_data.methods [i];
1279 old_data [4+3*i+1] = (uintptr_t)bt_data.il_offsets [i];
1280 old_data [4+3*i+2] = (uintptr_t)bt_data.native_offsets [i];
1284 static uintptr_t *code_pages = 0;
1285 static int num_code_pages = 0;
1286 static int size_code_pages = 0;
1287 #define CPAGE_SHIFT (9)
1288 #define CPAGE_SIZE (1 << CPAGE_SHIFT)
1289 #define CPAGE_MASK (~(CPAGE_SIZE - 1))
1290 #define CPAGE_ADDR(p) ((p) & CPAGE_MASK)
1293 add_code_page (uintptr_t *hash, uintptr_t hsize, uintptr_t page)
1296 uintptr_t start_pos;
1297 start_pos = (page >> CPAGE_SHIFT) % hsize;
1300 if (hash [i] && CPAGE_ADDR (hash [i]) == CPAGE_ADDR (page)) {
1302 } else if (!hash [i]) {
1309 } while (i != start_pos);
1310 /* should not happen */
1311 printf ("failed code page store\n");
1316 add_code_pointer (uintptr_t ip)
1319 if (num_code_pages * 2 >= size_code_pages) {
1321 uintptr_t old_size = size_code_pages;
1322 size_code_pages *= 2;
1323 if (size_code_pages == 0)
1324 size_code_pages = 16;
1325 n = calloc (sizeof (uintptr_t) * size_code_pages, 1);
1326 for (i = 0; i < old_size; ++i) {
1328 add_code_page (n, size_code_pages, code_pages [i]);
1334 num_code_pages += add_code_page (code_pages, size_code_pages, ip & CPAGE_MASK);
1338 dump_ubin (const char *filename, uintptr_t load_addr, uint64_t offset, uintptr_t size)
1341 LogBuffer *logbuffer;
1343 len = strlen (filename) + 1;
1344 now = current_time ();
1345 logbuffer = ensure_logbuf (20 + len);
1346 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_UBIN);
1347 emit_time (logbuffer, now);
1348 emit_svalue (logbuffer, load_addr);
1349 emit_uvalue (logbuffer, offset);
1350 emit_uvalue (logbuffer, size);
1351 memcpy (logbuffer->data, filename, len);
1352 logbuffer->data += len;
1356 dump_usym (const char *name, uintptr_t value, uintptr_t size)
1358 LogBuffer *logbuffer;
1360 len = strlen (name) + 1;
1361 logbuffer = ensure_logbuf (20 + len);
1362 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_USYM);
1363 emit_ptr (logbuffer, (void*)value);
1364 emit_value (logbuffer, size);
1365 memcpy (logbuffer->data, name, len);
1366 logbuffer->data += len;
1371 #if SIZEOF_VOID_P == 4
1372 #define ELF_WSIZE 32
1374 #define ELF_WSIZE 64
1377 #define ElfW(type) _ElfW (Elf, ELF_WSIZE, type)
1378 #define _ElfW(e,w,t) _ElfW_1 (e, w, _##t)
1379 #define _ElfW_1(e,w,t) e##w##t
1383 dump_elf_symbols (ElfW(Sym) *symbols, int num_symbols, const char *strtab, void *load_addr)
1386 for (i = 0; i < num_symbols; ++i) {
1388 sym = strtab + symbols [i].st_name;
1389 if (!symbols [i].st_name || !symbols [i].st_size || (symbols [i].st_info & 0xf) != STT_FUNC)
1391 //printf ("symbol %s at %d\n", sym, symbols [i].st_value);
1392 dump_usym (sym, (uintptr_t)load_addr + symbols [i].st_value, symbols [i].st_size);
1397 read_elf_symbols (MonoProfiler *prof, const char *filename, void *load_addr)
1404 ElfW(Shdr) *sheader;
1405 ElfW(Shdr) *shstrtabh;
1406 ElfW(Shdr) *symtabh = NULL;
1407 ElfW(Shdr) *strtabh = NULL;
1408 ElfW(Sym) *symbols = NULL;
1412 fd = open (filename, O_RDONLY);
1415 if (fstat (fd, &statb) != 0) {
1419 file_size = statb.st_size;
1420 data = mmap (NULL, file_size, PROT_READ, MAP_PRIVATE, fd, 0);
1422 if (data == MAP_FAILED)
1425 if (header->e_ident [EI_MAG0] != ELFMAG0 ||
1426 header->e_ident [EI_MAG1] != ELFMAG1 ||
1427 header->e_ident [EI_MAG2] != ELFMAG2 ||
1428 header->e_ident [EI_MAG3] != ELFMAG3 ) {
1429 munmap (data, file_size);
1432 sheader = (void*)((char*)data + header->e_shoff);
1433 shstrtabh = (void*)((char*)sheader + (header->e_shentsize * header->e_shstrndx));
1434 strtab = (const char*)data + shstrtabh->sh_offset;
1435 for (i = 0; i < header->e_shnum; ++i) {
1436 //printf ("section header: %d\n", sheader->sh_type);
1437 if (sheader->sh_type == SHT_SYMTAB) {
1439 strtabh = (void*)((char*)data + header->e_shoff + sheader->sh_link * header->e_shentsize);
1440 /*printf ("symtab section header: %d, .strstr: %d\n", i, sheader->sh_link);*/
1443 sheader = (void*)((char*)sheader + header->e_shentsize);
1445 if (!symtabh || !strtabh) {
1446 munmap (data, file_size);
1449 strtab = (const char*)data + strtabh->sh_offset;
1450 num_symbols = symtabh->sh_size / symtabh->sh_entsize;
1451 symbols = (void*)((char*)data + symtabh->sh_offset);
1452 dump_elf_symbols (symbols, num_symbols, strtab, load_addr);
1453 munmap (data, file_size);
1458 #if defined(HAVE_DL_ITERATE_PHDR) && defined(ELFMAG0)
1460 elf_dl_callback (struct dl_phdr_info *info, size_t size, void *data)
1462 MonoProfiler *prof = data;
1464 const char *filename;
1466 char *a = (void*)info->dlpi_addr;
1468 ElfW(Dyn) *dyn = NULL;
1469 ElfW(Sym) *symtab = NULL;
1470 ElfW(Word) *hash_table = NULL;
1471 ElfW(Ehdr) *header = NULL;
1472 const char* strtab = NULL;
1473 for (obj = prof->binary_objects; obj; obj = obj->next) {
1477 filename = info->dlpi_name;
1478 if (!info->dlpi_addr && !filename [0]) {
1479 int l = readlink ("/proc/self/exe", buf, sizeof (buf) - 1);
1485 obj = calloc (sizeof (BinaryObject), 1);
1486 obj->addr = (void*)info->dlpi_addr;
1487 obj->name = pstrdup (filename);
1488 obj->next = prof->binary_objects;
1489 prof->binary_objects = obj;
1490 //printf ("loaded file: %s at %p, segments: %d\n", filename, (void*)info->dlpi_addr, info->dlpi_phnum);
1492 for (i = 0; i < info->dlpi_phnum; ++i) {
1493 //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);
1494 if (info->dlpi_phdr[i].p_type == PT_LOAD && !header) {
1495 header = (ElfW(Ehdr)*)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1496 if (header->e_ident [EI_MAG0] != ELFMAG0 ||
1497 header->e_ident [EI_MAG1] != ELFMAG1 ||
1498 header->e_ident [EI_MAG2] != ELFMAG2 ||
1499 header->e_ident [EI_MAG3] != ELFMAG3 ) {
1502 dump_ubin (filename, info->dlpi_addr + info->dlpi_phdr[i].p_vaddr, info->dlpi_phdr[i].p_offset, info->dlpi_phdr[i].p_memsz);
1503 } else if (info->dlpi_phdr[i].p_type == PT_DYNAMIC) {
1504 dyn = (ElfW(Dyn) *)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1507 if (read_elf_symbols (prof, filename, (void*)info->dlpi_addr))
1509 if (!info->dlpi_name || !info->dlpi_name[0])
1513 for (i = 0; dyn [i].d_tag != DT_NULL; ++i) {
1514 if (dyn [i].d_tag == DT_SYMTAB) {
1515 if (symtab && do_debug)
1516 printf ("multiple symtabs: %d\n", i);
1517 symtab = (ElfW(Sym) *)(a + dyn [i].d_un.d_ptr);
1518 } else if (dyn [i].d_tag == DT_HASH) {
1519 hash_table = (ElfW(Word) *)(a + dyn [i].d_un.d_ptr);
1520 } else if (dyn [i].d_tag == DT_STRTAB) {
1521 strtab = (const char*)(a + dyn [i].d_un.d_ptr);
1526 num_sym = hash_table [1];
1527 dump_elf_symbols (symtab, num_sym, strtab, (void*)info->dlpi_addr);
1532 load_binaries (MonoProfiler *prof)
1534 dl_iterate_phdr (elf_dl_callback, prof);
1539 load_binaries (MonoProfiler *prof)
1546 symbol_for (uintptr_t code)
1549 void *ip = (void*)code;
1551 if (dladdr (ip, &di)) {
1553 return di.dli_sname;
1556 names = backtrace_symbols (&ip, 1);
1558 const char* p = names [0];
1569 dump_unmanaged_coderefs (MonoProfiler *prof)
1572 const char* last_symbol;
1573 uintptr_t addr, page_end;
1575 if (load_binaries (prof))
1577 for (i = 0; i < size_code_pages; ++i) {
1579 if (!code_pages [i] || code_pages [i] & 1)
1582 addr = CPAGE_ADDR (code_pages [i]);
1583 page_end = addr + CPAGE_SIZE;
1584 code_pages [i] |= 1;
1585 /* we dump the symbols for the whole page */
1586 for (; addr < page_end; addr += 16) {
1587 sym = symbol_for (addr);
1588 if (sym && sym == last_symbol)
1593 dump_usym (sym, addr, 0); /* let's not guess the size */
1594 //printf ("found symbol at %p: %s\n", (void*)addr, sym);
1600 dump_sample_hits (MonoProfiler *prof, StatBuffer *sbuf, int recurse)
1603 LogBuffer *logbuffer;
1606 if (recurse && sbuf->next) {
1607 dump_sample_hits (prof, sbuf->next, 1);
1608 free_buffer (sbuf->next, sbuf->next->size);
1611 for (sample = sbuf->buf; sample < sbuf->data;) {
1613 int count = sample [0] & 0xff;
1614 int mbt_count = (sample [0] & 0xff00) >> 8;
1615 int type = sample [0] >> 16;
1616 if (sample + count + 3 + mbt_count * 3 > sbuf->data)
1618 logbuffer = ensure_logbuf (20 + count * 8);
1619 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
1620 emit_value (logbuffer, type);
1621 emit_uvalue (logbuffer, prof->startup_time + (uint64_t)sample [2] * (uint64_t)10000);
1622 emit_value (logbuffer, count);
1623 for (i = 0; i < count; ++i) {
1624 emit_ptr (logbuffer, (void*)sample [i + 3]);
1625 add_code_pointer (sample [i + 3]);
1627 sample += count + 3;
1628 /* new in data version 6 */
1629 emit_uvalue (logbuffer, mbt_count);
1630 for (i = 0; i < mbt_count; ++i) {
1631 emit_method (logbuffer, (void*)sample [i * 3]); /* method */
1632 emit_svalue (logbuffer, sample [i * 3 + 1]); /* il offset */
1633 emit_svalue (logbuffer, sample [i * 3 + 2]); /* native offset */
1635 sample += 3 * mbt_count;
1637 dump_unmanaged_coderefs (prof);
1641 #ifndef __NR_perf_event_open
1643 #define __NR_perf_event_open 364
1645 #define __NR_perf_event_open 241
1650 mono_cpu_count (void)
1653 #ifdef PLATFORM_ANDROID
1654 /* Android tries really hard to save power by powering off CPUs on SMP phones which
1655 * means the normal way to query cpu count returns a wrong value with userspace API.
1656 * Instead we use /sys entries to query the actual hardware CPU count.
1658 char buffer[8] = {'\0'};
1659 int present = open ("/sys/devices/system/cpu/present", O_RDONLY);
1660 /* Format of the /sys entry is a cpulist of indexes which in the case
1661 * of present is always of the form "0-(n-1)" when there is more than
1662 * 1 core, n being the number of CPU cores in the system. Otherwise
1663 * the value is simply 0
1665 if (present != -1 && read (present, (char*)buffer, sizeof (buffer)) > 3)
1666 count = strtol (((char*)buffer) + 2, NULL, 10);
1672 #ifdef _SC_NPROCESSORS_ONLN
1673 count = sysconf (_SC_NPROCESSORS_ONLN);
1680 size_t len = sizeof (int);
1683 if (sysctl (mib, 2, &count, &len, NULL, 0) == 0)
1690 GetSystemInfo (&info);
1691 return info.dwNumberOfProcessors;
1700 unsigned int prev_pos;
1702 struct perf_event_mmap_page *page_desc;
1705 static PerfData *perf_data = NULL;
1706 static int num_perf;
1707 #define PERF_PAGES_SHIFT 4
1708 static int num_pages = 1 << PERF_PAGES_SHIFT;
1709 static unsigned int mmap_mask;
1712 struct perf_event_header h;
1722 perf_event_syscall (struct perf_event_attr *attr, pid_t pid, int cpu, int group_fd, unsigned long flags)
1724 attr->size = PERF_ATTR_SIZE_VER0;
1725 //printf ("perf attr size: %d\n", attr->size);
1726 #if defined(__x86_64__)
1727 return syscall(/*__NR_perf_event_open*/ 298, attr, pid, cpu, group_fd, flags);
1728 #elif defined(__i386__)
1729 return syscall(/*__NR_perf_event_open*/ 336, attr, pid, cpu, group_fd, flags);
1730 #elif defined(__arm__)
1731 return syscall(/*__NR_perf_event_open*/ 364, attr, pid, cpu, group_fd, flags);
1738 setup_perf_map (PerfData *perf)
1740 perf->mmap_base = mmap (NULL, (num_pages + 1) * getpagesize (), PROT_READ|PROT_WRITE, MAP_SHARED, perf->perf_fd, 0);
1741 if (perf->mmap_base == MAP_FAILED) {
1743 printf ("failed mmap\n");
1746 perf->page_desc = perf->mmap_base;
1748 printf ("mmap version: %d\n", perf->page_desc->version);
1753 dump_perf_hits (MonoProfiler *prof, void *buf, int size)
1755 LogBuffer *logbuffer;
1756 void *end = (char*)buf + size;
1758 int pid = getpid ();
1764 if (pid != s->pid) {
1766 printf ("event for different pid: %d\n", s->pid);
1767 buf = (char*)buf + s->h.size;
1770 /*ip = (void*)s->ip;
1771 printf ("sample: %d, size: %d, ip: %p (%s), timestamp: %llu, nframes: %llu\n",
1772 s->h.type, s->h.size, ip, symbol_for (ip), s->timestamp, s->nframes);*/
1773 logbuffer = ensure_logbuf (20 + s->nframes * 8);
1774 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
1775 emit_value (logbuffer, sample_type);
1776 emit_uvalue (logbuffer, s->timestamp - prof->startup_time);
1777 emit_value (logbuffer, 1); /* count */
1778 emit_ptr (logbuffer, (void*)(uintptr_t)s->ip);
1779 /* no support here yet for the managed backtrace */
1780 emit_uvalue (logbuffer, 0);
1781 add_code_pointer (s->ip);
1782 buf = (char*)buf + s->h.size;
1786 printf ("dumped %d samples\n", samples);
1787 dump_unmanaged_coderefs (prof);
1790 /* read events from the ring buffer */
1792 read_perf_mmap (MonoProfiler* prof, int cpu)
1794 PerfData *perf = perf_data + cpu;
1796 unsigned char *data = (unsigned char*)perf->mmap_base + getpagesize ();
1797 unsigned int head = perf->page_desc->data_head;
1801 mono_memory_read_barrier ();
1803 old = perf->prev_pos;
1807 printf ("lost mmap events: old: %d, head: %d\n", old, head);
1811 if ((old & mmap_mask) + size != (head & mmap_mask)) {
1812 buf = data + (old & mmap_mask);
1813 size = mmap_mask + 1 - (old & mmap_mask);
1815 /* size bytes at buf */
1817 printf ("found1 bytes of events: %d\n", size);
1818 dump_perf_hits (prof, buf, size);
1820 buf = data + (old & mmap_mask);
1822 /* size bytes at buf */
1824 printf ("found bytes of events: %d\n", size);
1825 dump_perf_hits (prof, buf, size);
1827 perf->prev_pos = old;
1828 perf->page_desc->data_tail = old;
1833 setup_perf_event_for_cpu (PerfData *perf, int cpu)
1835 struct perf_event_attr attr;
1836 memset (&attr, 0, sizeof (attr));
1837 attr.type = PERF_TYPE_HARDWARE;
1838 switch (sample_type) {
1839 case SAMPLE_CYCLES: attr.config = PERF_COUNT_HW_CPU_CYCLES; break;
1840 case SAMPLE_INSTRUCTIONS: attr.config = PERF_COUNT_HW_INSTRUCTIONS; break;
1841 case SAMPLE_CACHE_MISSES: attr.config = PERF_COUNT_HW_CACHE_MISSES; break;
1842 case SAMPLE_CACHE_REFS: attr.config = PERF_COUNT_HW_CACHE_REFERENCES; break;
1843 case SAMPLE_BRANCHES: attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; break;
1844 case SAMPLE_BRANCH_MISSES: attr.config = PERF_COUNT_HW_BRANCH_MISSES; break;
1845 default: attr.config = PERF_COUNT_HW_CPU_CYCLES; break;
1847 attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_PERIOD | PERF_SAMPLE_TIME;
1848 // attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1849 attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING | PERF_FORMAT_ID;
1852 attr.sample_freq = sample_freq;
1854 perf->perf_fd = perf_event_syscall (&attr, getpid (), cpu, -1, 0);
1856 printf ("perf fd: %d, freq: %d, event: %llu\n", perf->perf_fd, sample_freq, attr.config);
1857 if (perf->perf_fd < 0) {
1858 if (perf->perf_fd == -EPERM) {
1859 fprintf (stderr, "Perf syscall denied, do \"echo 1 > /proc/sys/kernel/perf_event_paranoid\" as root to enable.\n");
1862 perror ("open perf event");
1866 if (!setup_perf_map (perf)) {
1867 close (perf->perf_fd);
1875 setup_perf_event (void)
1878 mmap_mask = num_pages * getpagesize () - 1;
1879 num_perf = mono_cpu_count ();
1880 perf_data = calloc (num_perf, sizeof (PerfData));
1881 for (i = 0; i < num_perf; ++i) {
1882 count += setup_perf_event_for_cpu (perf_data + i, i);
1891 #endif /* USE_PERF_EVENTS */
1893 #ifndef DISABLE_HELPER_THREAD
1895 typedef struct MonoCounterAgent {
1896 MonoCounter *counter;
1897 // MonoCounterAgent specific data :
1901 struct MonoCounterAgent *next;
1904 static MonoCounterAgent* counters;
1905 static gboolean counters_initialized = FALSE;
1906 static int counters_index = 1;
1909 counters_init_add_counter (MonoCounter *counter, gpointer data)
1911 MonoCounterAgent *agent, *item;
1913 for (agent = counters; agent; agent = agent->next) {
1914 if (agent->counter == counter)
1918 agent = malloc (sizeof (MonoCounterAgent));
1919 agent->counter = counter;
1920 agent->value = NULL;
1921 agent->value_size = 0;
1922 agent->index = counters_index++;
1938 counters_init (MonoProfiler *profiler)
1940 mono_counters_foreach (counters_init_add_counter, NULL);
1942 MonoCounterAgent *agent;
1943 LogBuffer *logbuffer;
1944 int size = 1 + 5, len = 0;
1946 for (agent = counters; agent; agent = agent->next) {
1947 size += strlen (mono_counter_get_name (agent->counter)) + 1 + 5 * 5;
1951 logbuffer = ensure_logbuf (size);
1953 ENTER_LOG (logbuffer, "counters");
1954 emit_byte (logbuffer, TYPE_SAMPLE_COUNTERS_DESC | TYPE_SAMPLE);
1955 emit_value (logbuffer, len);
1956 for (agent = counters; agent; agent = agent->next) {
1957 const char *name = mono_counter_get_name (agent->counter);
1958 emit_value (logbuffer, mono_counter_get_section (agent->counter));
1959 emit_string (logbuffer, name, strlen (name) + 1);
1960 emit_value (logbuffer, mono_counter_get_type (agent->counter));
1961 emit_value (logbuffer, mono_counter_get_unit (agent->counter));
1962 emit_value (logbuffer, mono_counter_get_variance (agent->counter));
1963 emit_value (logbuffer, agent->index);
1965 EXIT_LOG (logbuffer);
1967 counters_initialized = TRUE;
1971 counters_sample (MonoProfiler *profiler, uint64_t timestamp)
1973 MonoCounterAgent *agent;
1974 MonoCounter *counter;
1975 LogBuffer *logbuffer;
1981 if (!counters_initialized)
1985 buffer = calloc (1, buffer_size);
1988 for (agent = counters; agent; agent = agent->next)
1989 size += 10 * 2 + mono_counter_get_size (agent->counter);
1991 logbuffer = ensure_logbuf (size);
1993 ENTER_LOG (logbuffer, "counters");
1994 emit_byte (logbuffer, TYPE_SAMPLE_COUNTERS | TYPE_SAMPLE);
1995 emit_uvalue (logbuffer, timestamp);
1996 for (agent = counters; agent; agent = agent->next) {
1997 counter = agent->counter;
1999 size_t size = mono_counter_get_size (counter);
2001 continue; // FIXME error
2002 } else if (size > buffer_size) {
2004 buffer = realloc (buffer, buffer_size);
2007 memset (buffer, 0, buffer_size);
2009 if (mono_counters_sample (counter, buffer, size) < 0)
2010 continue; // FIXME error
2012 type = mono_counter_get_type (counter);
2014 if (!agent->value) {
2015 agent->value = calloc (1, size);
2016 agent->value_size = size;
2018 if (type == MONO_COUNTER_STRING) {
2019 if (strcmp (agent->value, buffer) == 0)
2022 if (agent->value_size == size && memcmp (agent->value, buffer, size) == 0)
2027 emit_uvalue (logbuffer, agent->index);
2028 emit_uvalue (logbuffer, type);
2030 case MONO_COUNTER_INT:
2031 #if SIZEOF_VOID_P == 4
2032 case MONO_COUNTER_WORD:
2034 emit_svalue (logbuffer, *(int*)buffer - *(int*)agent->value);
2036 case MONO_COUNTER_UINT:
2037 emit_uvalue (logbuffer, *(guint*)buffer - *(guint*)agent->value);
2039 case MONO_COUNTER_TIME_INTERVAL:
2040 case MONO_COUNTER_LONG:
2041 #if SIZEOF_VOID_P == 8
2042 case MONO_COUNTER_WORD:
2044 emit_svalue (logbuffer, *(gint64*)buffer - *(gint64*)agent->value);
2046 case MONO_COUNTER_ULONG:
2047 emit_uvalue (logbuffer, *(guint64*)buffer - *(guint64*)agent->value);
2049 case MONO_COUNTER_DOUBLE:
2050 emit_double (logbuffer, *(double*)buffer);
2052 case MONO_COUNTER_STRING:
2054 emit_byte (logbuffer, 0);
2056 emit_byte (logbuffer, 1);
2057 emit_string (logbuffer, (char*)buffer, size);
2064 if (type == MONO_COUNTER_STRING && size > agent->value_size) {
2065 agent->value = realloc (agent->value, size);
2066 agent->value_size = size;
2070 memcpy (agent->value, buffer, size);
2074 emit_value (logbuffer, 0);
2075 EXIT_LOG (logbuffer);
2077 safe_dump (profiler, ensure_logbuf (0));
2080 #endif /* DISABLE_HELPER_THREAD */
2083 log_shutdown (MonoProfiler *prof)
2086 #ifndef DISABLE_HELPER_THREAD
2087 if (prof->command_port) {
2090 write (prof->pipes [1], &c, 1);
2091 pthread_join (prof->helper_thread, &res);
2097 for (i = 0; i < num_perf; ++i)
2098 read_perf_mmap (prof, i);
2101 dump_sample_hits (prof, prof->stat_buffers, 1);
2103 if (TLS_GET (tlsbuffer))
2104 dump_buffer (prof, TLS_GET (tlsbuffer));
2105 TLS_SET (tlsbuffer, NULL);
2107 #if defined (HAVE_SYS_ZLIB)
2109 gzclose (prof->gzfile);
2111 if (prof->pipe_output)
2112 pclose (prof->file);
2114 fclose (prof->file);
2119 new_filename (const char* filename)
2121 time_t t = time (NULL);
2122 int pid = process_id ();
2127 int count_dates = 0;
2131 for (p = filename; *p; p++) {
2142 if (!count_dates && !count_pids)
2143 return pstrdup (filename);
2144 snprintf (pid_buf, sizeof (pid_buf), "%d", pid);
2146 snprintf (time_buf, sizeof (time_buf), "%d%02d%02d%02d%02d%02d",
2147 1900 + ts->tm_year, 1 + ts->tm_mon, ts->tm_mday, ts->tm_hour, ts->tm_min, ts->tm_sec);
2148 s_date = strlen (time_buf);
2149 s_pid = strlen (pid_buf);
2150 d = res = malloc (strlen (filename) + s_date * count_dates + s_pid * count_pids);
2151 for (p = filename; *p; p++) {
2158 strcpy (d, time_buf);
2161 } else if (*p == 'p') {
2162 strcpy (d, pid_buf);
2165 } else if (*p == '%') {
2177 #ifndef DISABLE_HELPER_THREAD
2179 helper_thread (void* arg)
2181 MonoProfiler* prof = arg;
2185 MonoThread *thread = NULL;
2186 uint64_t start, now;
2188 //fprintf (stderr, "Server listening\n");
2189 start = current_time ();
2190 command_socket = -1;
2196 FD_SET (prof->server_socket, &rfds);
2197 max_fd = prof->server_socket;
2198 FD_SET (prof->pipes [0], &rfds);
2199 if (max_fd < prof->pipes [0])
2200 max_fd = prof->pipes [0];
2201 if (command_socket >= 0) {
2202 FD_SET (command_socket, &rfds);
2203 if (max_fd < command_socket)
2204 max_fd = command_socket;
2209 for ( i = 0; i < num_perf; ++i) {
2210 if (perf_data [i].perf_fd < 0)
2212 FD_SET (perf_data [i].perf_fd, &rfds);
2213 if (max_fd < perf_data [i].perf_fd)
2214 max_fd = perf_data [i].perf_fd;
2218 now = current_time ();
2219 counters_sample (prof, (now - start) / 1000/ 1000);
2223 len = select (max_fd + 1, &rfds, NULL, NULL, &tv);
2229 g_warning ("Error in proflog server: %s", strerror (errno));
2233 if (FD_ISSET (prof->pipes [0], &rfds)) {
2235 int r = read (prof->pipes [0], &c, 1);
2236 if (r == 1 && c == 0) {
2237 StatBuffer *sbufbase = prof->stat_buffers;
2239 if (!sbufbase->next)
2241 sbuf = sbufbase->next->next;
2242 sbufbase->next->next = NULL;
2244 fprintf (stderr, "stat buffer dump\n");
2245 dump_sample_hits (prof, sbuf, 1);
2246 free_buffer (sbuf, sbuf->size);
2247 safe_dump (prof, ensure_logbuf (0));
2250 /* time to shut down */
2252 mono_thread_detach (thread);
2254 fprintf (stderr, "helper shutdown\n");
2258 for ( i = 0; i < num_perf; ++i) {
2259 if (perf_data [i].perf_fd < 0)
2261 if (FD_ISSET (perf_data [i].perf_fd, &rfds))
2262 read_perf_mmap (prof, i);
2266 safe_dump (prof, ensure_logbuf (0));
2272 for ( i = 0; i < num_perf; ++i) {
2273 if (perf_data [i].perf_fd < 0)
2275 if (FD_ISSET (perf_data [i].perf_fd, &rfds)) {
2276 read_perf_mmap (prof, i);
2277 safe_dump (prof, ensure_logbuf (0));
2282 if (command_socket >= 0 && FD_ISSET (command_socket, &rfds)) {
2283 len = read (command_socket, buf, sizeof (buf) - 1);
2287 close (command_socket);
2288 command_socket = -1;
2292 if (strcmp (buf, "heapshot\n") == 0) {
2293 heapshot_requested = 1;
2294 //fprintf (stderr, "perform heapshot\n");
2295 if (runtime_inited && !thread) {
2296 thread = mono_thread_attach (mono_get_root_domain ());
2297 /*fprintf (stderr, "attached\n");*/
2300 process_requests (prof);
2301 mono_thread_detach (thread);
2307 if (!FD_ISSET (prof->server_socket, &rfds)) {
2310 command_socket = accept (prof->server_socket, NULL, NULL);
2311 if (command_socket < 0)
2313 //fprintf (stderr, "Accepted connection\n");
2319 start_helper_thread (MonoProfiler* prof)
2321 struct sockaddr_in server_address;
2324 if (pipe (prof->pipes) < 0) {
2325 fprintf (stderr, "Cannot create pipe\n");
2328 prof->server_socket = socket (PF_INET, SOCK_STREAM, 0);
2329 if (prof->server_socket < 0) {
2330 fprintf (stderr, "Cannot create server socket\n");
2333 memset (&server_address, 0, sizeof (server_address));
2334 server_address.sin_family = AF_INET;
2335 server_address.sin_addr.s_addr = INADDR_ANY;
2336 server_address.sin_port = htons (prof->command_port);
2337 if (bind (prof->server_socket, (struct sockaddr *) &server_address, sizeof (server_address)) < 0) {
2338 fprintf (stderr, "Cannot bind server socket, port: %d: %s\n", prof->command_port, strerror (errno));
2339 close (prof->server_socket);
2342 if (listen (prof->server_socket, 1) < 0) {
2343 fprintf (stderr, "Cannot listen server socket\n");
2344 close (prof->server_socket);
2347 slen = sizeof (server_address);
2348 if (getsockname (prof->server_socket, (struct sockaddr *)&server_address, &slen) == 0) {
2349 prof->command_port = ntohs (server_address.sin_port);
2350 /*fprintf (stderr, "Assigned server port: %d\n", prof->command_port);*/
2353 r = pthread_create (&prof->helper_thread, NULL, helper_thread, prof);
2355 close (prof->server_socket);
2362 static MonoProfiler*
2363 create_profiler (const char *filename)
2367 int force_delete = 0;
2368 int need_helper_thread = 0;
2369 prof = calloc (1, sizeof (MonoProfiler));
2371 prof->command_port = command_port;
2372 if (filename && *filename == '-') {
2378 filename = "|mprof-report -";
2380 filename = "output.mlpd";
2381 nf = (char*)filename;
2383 nf = new_filename (filename);
2385 int s = strlen (nf) + 32;
2386 char *p = malloc (s);
2387 snprintf (p, s, "|mprof-report '--out=%s' -", nf);
2393 prof->file = popen (nf + 1, "w");
2394 prof->pipe_output = 1;
2395 } else if (*nf == '#') {
2396 int fd = strtol (nf + 1, NULL, 10);
2397 prof->file = fdopen (fd, "a");
2402 if ((f = fopen (nf, "r"))) {
2404 fprintf (stderr, "The Mono profiler won't overwrite existing filename: %s.\n", nf);
2405 fprintf (stderr, "Profiling disabled: use a different name or -FILENAME to force overwrite.\n");
2409 prof->file = fopen (nf, "wb");
2412 fprintf (stderr, "Cannot create profiler output: %s\n", nf);
2415 #if defined (HAVE_SYS_ZLIB)
2417 prof->gzfile = gzdopen (fileno (prof->file), "wb");
2420 if (sample_type && !do_mono_sample)
2421 need_helper_thread = setup_perf_event ();
2423 /* FIXME: warn if different freq or sample type */
2427 if (do_mono_sample) {
2428 prof->stat_buffers = create_stat_buffer ();
2429 need_helper_thread = 1;
2431 if (do_counters && !need_helper_thread) {
2432 need_helper_thread = 1;
2434 #ifndef DISABLE_HELPER_THREAD
2435 if (hs_mode_ondemand || need_helper_thread) {
2436 if (!start_helper_thread (prof))
2437 prof->command_port = 0;
2440 if (hs_mode_ondemand)
2441 fprintf (stderr, "Ondemand heapshot unavailable on this arch.\n");
2443 prof->startup_time = current_time ();
2451 printf ("Log profiler version %d.%d (format: %d)\n", LOG_VERSION_MAJOR, LOG_VERSION_MINOR, LOG_DATA_VERSION);
2452 printf ("Usage: mono --profile=log[:OPTION1[,OPTION2...]] program.exe\n");
2453 printf ("Options:\n");
2454 printf ("\thelp show this usage info\n");
2455 printf ("\t[no]alloc enable/disable recording allocation info\n");
2456 printf ("\t[no]calls enable/disable recording enter/leave method events\n");
2457 printf ("\theapshot[=MODE] record heap shot info (by default at each major collection)\n");
2458 printf ("\t MODE: every XXms milliseconds, every YYgc collections, ondemand\n");
2459 printf ("\tcounters sample counters every 1s\n");
2460 printf ("\tsample[=TYPE] use statistical sampling mode (by default cycles/1000)\n");
2461 printf ("\t TYPE: cycles,instr,cacherefs,cachemiss,branches,branchmiss\n");
2462 printf ("\t TYPE can be followed by /FREQUENCY\n");
2463 printf ("\ttime=fast use a faster (but more inaccurate) timer\n");
2464 printf ("\tmaxframes=NUM collect up to NUM stack frames\n");
2465 printf ("\tcalldepth=NUM ignore method events for call chain depth bigger than NUM\n");
2466 printf ("\toutput=FILENAME write the data to file FILENAME (-FILENAME to overwrite)\n");
2467 printf ("\toutput=|PROGRAM write the data to the stdin of PROGRAM\n");
2468 printf ("\t %%t is subtituted with date and time, %%p with the pid\n");
2469 printf ("\treport create a report instead of writing the raw data to a file\n");
2470 printf ("\tzip compress the output data\n");
2471 printf ("\tport=PORTNUM use PORTNUM for the listening command server\n");
2477 match_option (const char* p, const char *opt, char **rval)
2479 int len = strlen (opt);
2480 if (strncmp (p, opt, len) == 0) {
2482 if (p [len] == '=' && p [len + 1]) {
2483 const char *opt = p + len + 1;
2484 const char *end = strchr (opt, ',');
2492 val = malloc (l + 1);
2493 memcpy (val, opt, l);
2498 if (p [len] == 0 || p [len] == ',') {
2500 return p + len + (p [len] == ',');
2518 static const SampleMode sample_modes [] = {
2519 {"cycles", SAMPLE_CYCLES},
2520 {"instr", SAMPLE_INSTRUCTIONS},
2521 {"cachemiss", SAMPLE_CACHE_MISSES},
2522 {"cacherefs", SAMPLE_CACHE_REFS},
2523 {"branches", SAMPLE_BRANCHES},
2524 {"branchmiss", SAMPLE_BRANCH_MISSES},
2529 set_sample_mode (char* val, int allow_empty)
2532 char *maybe_freq = NULL;
2534 const SampleMode *smode = sample_modes;
2535 #ifndef USE_PERF_EVENTS
2538 if (allow_empty && !val) {
2539 sample_type = SAMPLE_CYCLES;
2543 if (strcmp (val, "mono") == 0) {
2545 sample_type = SAMPLE_CYCLES;
2549 for (smode = sample_modes; smode->name; smode++) {
2550 int l = strlen (smode->name);
2551 if (strncmp (val, smode->name, l) == 0) {
2552 sample_type = smode->sample_mode;
2553 maybe_freq = val + l;
2559 if (*maybe_freq == '/') {
2560 count = strtoul (maybe_freq + 1, &end, 10);
2561 if (maybe_freq + 1 == end)
2563 sample_freq = count;
2564 } else if (*maybe_freq != 0) {
2573 set_hsmode (char* val, int allow_empty)
2577 if (allow_empty && !val)
2579 if (strcmp (val, "ondemand") == 0) {
2580 hs_mode_ondemand = 1;
2584 count = strtoul (val, &end, 10);
2587 if (strcmp (end, "ms") == 0)
2589 else if (strcmp (end, "gc") == 0)
2597 * declaration to silence the compiler: this is the entry point that
2598 * mono will load from the shared library and call.
2601 mono_profiler_startup (const char *desc);
2604 mono_profiler_startup_log (const char *desc);
2607 * this is the entry point that will be used when the profiler
2608 * is embedded inside the main executable.
2611 mono_profiler_startup_log (const char *desc)
2613 mono_profiler_startup (desc);
2617 mono_profiler_startup (const char *desc)
2620 char *filename = NULL;
2624 int calls_enabled = 0;
2625 int allocs_enabled = 0;
2626 int events = MONO_PROFILE_GC|MONO_PROFILE_ALLOCATIONS|
2627 MONO_PROFILE_GC_MOVES|MONO_PROFILE_CLASS_EVENTS|MONO_PROFILE_THREADS|
2628 MONO_PROFILE_ENTER_LEAVE|MONO_PROFILE_JIT_COMPILATION|MONO_PROFILE_EXCEPTIONS|
2629 MONO_PROFILE_MONITOR_EVENTS|MONO_PROFILE_MODULE_EVENTS|MONO_PROFILE_GC_ROOTS;
2632 if (strncmp (p, "log", 3))
2637 for (; *p; p = opt) {
2643 if ((opt = match_option (p, "help", NULL)) != p) {
2647 if ((opt = match_option (p, "calls", NULL)) != p) {
2651 if ((opt = match_option (p, "nocalls", NULL)) != p) {
2652 events &= ~MONO_PROFILE_ENTER_LEAVE;
2656 if ((opt = match_option (p, "alloc", NULL)) != p) {
2660 if ((opt = match_option (p, "noalloc", NULL)) != p) {
2661 events &= ~MONO_PROFILE_ALLOCATIONS;
2664 if ((opt = match_option (p, "time", &val)) != p) {
2665 if (strcmp (val, "fast") == 0)
2667 else if (strcmp (val, "null") == 0)
2674 if ((opt = match_option (p, "report", NULL)) != p) {
2678 if ((opt = match_option (p, "debug", NULL)) != p) {
2682 if ((opt = match_option (p, "heapshot", &val)) != p) {
2683 events &= ~MONO_PROFILE_ALLOCATIONS;
2684 events &= ~MONO_PROFILE_ENTER_LEAVE;
2687 set_hsmode (val, 1);
2690 if ((opt = match_option (p, "sample", &val)) != p) {
2691 events &= ~MONO_PROFILE_ALLOCATIONS;
2692 events &= ~MONO_PROFILE_ENTER_LEAVE;
2694 set_sample_mode (val, 1);
2697 if ((opt = match_option (p, "hsmode", &val)) != p) {
2698 fprintf (stderr, "The hsmode profiler option is obsolete, use heapshot=MODE.\n");
2699 set_hsmode (val, 0);
2702 if ((opt = match_option (p, "zip", NULL)) != p) {
2706 if ((opt = match_option (p, "output", &val)) != p) {
2710 if ((opt = match_option (p, "port", &val)) != p) {
2712 command_port = strtoul (val, &end, 10);
2716 if ((opt = match_option (p, "maxframes", &val)) != p) {
2718 num_frames = strtoul (val, &end, 10);
2719 if (num_frames > MAX_FRAMES)
2720 num_frames = MAX_FRAMES;
2722 notraces = num_frames == 0;
2725 if ((opt = match_option (p, "calldepth", &val)) != p) {
2727 max_call_depth = strtoul (val, &end, 10);
2731 if ((opt = match_option (p, "counters", NULL)) != p) {
2740 if (calls_enabled) {
2741 events |= MONO_PROFILE_ENTER_LEAVE;
2745 events |= MONO_PROFILE_ALLOCATIONS;
2746 utils_init (fast_time);
2748 prof = create_profiler (filename);
2753 mono_profiler_install (prof, log_shutdown);
2754 mono_profiler_install_gc (gc_event, gc_resize);
2755 mono_profiler_install_allocation (gc_alloc);
2756 mono_profiler_install_gc_moves (gc_moves);
2757 mono_profiler_install_gc_roots (gc_handle, gc_roots);
2758 mono_profiler_install_class (NULL, class_loaded, NULL, NULL);
2759 mono_profiler_install_module (NULL, image_loaded, NULL, NULL);
2760 mono_profiler_install_thread (thread_start, thread_end);
2761 mono_profiler_install_thread_name (thread_name);
2762 mono_profiler_install_enter_leave (method_enter, method_leave);
2763 mono_profiler_install_jit_end (method_jitted);
2764 mono_profiler_install_exception (throw_exc, method_exc_leave, clause_exc);
2765 mono_profiler_install_monitor (monitor_event);
2766 mono_profiler_install_runtime_initialized (runtime_initialized);
2769 if (do_mono_sample && sample_type == SAMPLE_CYCLES) {
2770 events |= MONO_PROFILE_STATISTICAL;
2771 mono_profiler_install_statistical (mono_sample_hit);
2774 mono_profiler_set_events (events);
2776 TLS_INIT (tlsbuffer);