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 #if defined(__linux__)
71 #include <sys/syscall.h>
72 #include "perf_event.h"
73 #define USE_PERF_EVENTS 1
74 static int read_perf_mmap (MonoProfiler* prof, int cpu);
77 #define BUFFER_SIZE (4096 * 16)
78 static int nocalls = 0;
79 static int notraces = 0;
80 static int use_zip = 0;
81 static int do_report = 0;
82 static int do_heap_shot = 0;
83 static int max_call_depth = 100;
84 static int runtime_inited = 0;
85 static int command_port = 0;
86 static int heapshot_requested = 0;
87 static int sample_type = 0;
88 static int sample_freq = 0;
89 static int do_mono_sample = 0;
90 static int in_shutdown = 0;
91 static int do_debug = 0;
92 static int do_counters = 0;
94 /* For linux compile with:
95 * gcc -fPIC -shared -o libmono-profiler-log.so proflog.c utils.c -Wall -g -lz `pkg-config --cflags --libs mono-2`
96 * gcc -o mprof-report decode.c utils.c -Wall -g -lz -lrt -lpthread `pkg-config --cflags mono-2`
98 * For osx compile with:
99 * 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
100 * gcc -m32 -o mprof-report decode.c utils.c -Wall -g -lz -lrt -lpthread `pkg-config --cflags mono-2`
103 * sudo cp mprof-report /usr/local/bin
104 * sudo cp libmono-profiler-log.so /usr/local/lib
108 typedef struct _LogBuffer LogBuffer;
114 * The file is composed by a header followed by 0 or more buffers.
115 * Each buffer contains events that happened on a thread: for a given thread
116 * buffers that appear later in the file are guaranteed to contain events
117 * that happened later in time. Buffers from separate threads could be interleaved,
119 * Buffers are not required to be aligned.
122 * [id: 4 bytes] constant value: LOG_HEADER_ID
123 * [major: 1 byte] [minor: 1 byte] major and minor version of the log profiler
124 * [format: 1 byte] version of the data format for the rest of the file
125 * [ptrsize: 1 byte] size in bytes of a pointer in the profiled program
126 * [startup time: 8 bytes] time in milliseconds since the unix epoch when the program started
127 * [timer overhead: 4 bytes] approximate overhead in nanoseconds of the timer
128 * [flags: 4 bytes] file format flags, should be 0 for now
129 * [pid: 4 bytes] pid of the profiled process
130 * [port: 2 bytes] tcp port for server if != 0
131 * [sysid: 2 bytes] operating system and architecture identifier
133 * The multiple byte integers are in little-endian format.
136 * [buffer header] [event]*
137 * Buffers have a fixed-size header followed by 0 or more bytes of event data.
138 * Timing information and other values in the event data are usually stored
139 * as uleb128 or sleb128 integers. To save space, as noted for each item below,
140 * some data is represented as a difference between the actual value and
141 * either the last value of the same type (like for timing information) or
142 * as the difference from a value stored in a buffer header.
144 * For timing information the data is stored as uleb128, since timing
145 * increases in a monotonic way in each thread: the value is the number of
146 * nanoseconds to add to the last seen timing data in a buffer. The first value
147 * in a buffer will be calculated from the time_base field in the buffer head.
149 * Object or heap sizes are stored as uleb128.
150 * Pointer differences are stored as sleb128, instead.
152 * If an unexpected value is found, the rest of the buffer should be ignored,
153 * as generally the later values need the former to be interpreted correctly.
155 * buffer header format:
156 * [bufid: 4 bytes] constant value: BUF_ID
157 * [len: 4 bytes] size of the data following the buffer header
158 * [time_base: 8 bytes] time base in nanoseconds since an unspecified epoch
159 * [ptr_base: 8 bytes] base value for pointers
160 * [obj_base: 8 bytes] base value for object addresses
161 * [thread id: 8 bytes] system-specific thread ID (pthread_t for example)
162 * [method_base: 8 bytes] base value for MonoMethod pointers
165 * [extended info: upper 4 bits] [type: lower 4 bits] [data]*
166 * The data that follows depends on type and the extended info.
167 * Type is one of the enum values in proflog.h: TYPE_ALLOC, TYPE_GC,
168 * TYPE_METADATA, TYPE_METHOD, TYPE_EXCEPTION, TYPE_MONITOR, TYPE_HEAP.
169 * The extended info bits are interpreted based on type, see
170 * each individual event description below.
171 * strings are represented as a 0-terminated utf8 sequence.
174 * [flags: uleb128] must be 0
175 * [num: uleb128] number of frames following
176 * [frame: sleb128]* num MonoMethod pointers as differences from ptr_base
180 * exinfo: flags: TYPE_ALLOC_BT
181 * [time diff: uleb128] nanoseconds since last timing
182 * [ptr: sleb128] class as a byte difference from ptr_base
183 * [obj: sleb128] object address as a byte difference from obj_base
184 * [size: uleb128] size of the object in the heap
185 * If the TYPE_ALLOC_BT flag is set, a backtrace follows.
189 * exinfo: one of TYPE_GC_EVENT, TYPE_GC_RESIZE, TYPE_GC_MOVE, TYPE_GC_HANDLE_CREATED,
190 * TYPE_GC_HANDLE_DESTROYED
191 * [time diff: uleb128] nanoseconds since last timing
192 * if exinfo == TYPE_GC_RESIZE
193 * [heap_size: uleb128] new heap size
194 * if exinfo == TYPE_GC_EVENT
195 * [event type: uleb128] GC event (MONO_GC_EVENT_* from profiler.h)
196 * [generation: uleb128] GC generation event refers to
197 * if exinfo == TYPE_GC_MOVE
198 * [num_objects: uleb128] number of object moves that follow
199 * [objaddr: sleb128]+ num_objects object pointer differences from obj_base
200 * num is always an even number: the even items are the old
201 * addresses, the odd numbers are the respective new object addresses
202 * if exinfo == TYPE_GC_HANDLE_CREATED
203 * [handle_type: uleb128] GC handle type (System.Runtime.InteropServices.GCHandleType)
204 * upper bits reserved as flags
205 * [handle: uleb128] GC handle value
206 * [objaddr: sleb128] object pointer differences from obj_base
207 * if exinfo == TYPE_GC_HANDLE_DESTROYED
208 * [handle_type: uleb128] GC handle type (System.Runtime.InteropServices.GCHandleType)
209 * upper bits reserved as flags
210 * [handle: uleb128] GC handle value
212 * type metadata format:
213 * type: TYPE_METADATA
214 * exinfo: flags: TYPE_LOAD_ERR
215 * [time diff: uleb128] nanoseconds since last timing
216 * [mtype: byte] metadata type, one of: TYPE_CLASS, TYPE_IMAGE, TYPE_ASSEMBLY, TYPE_DOMAIN,
218 * [pointer: sleb128] pointer of the metadata type depending on mtype
219 * if mtype == TYPE_CLASS
220 * [image: sleb128] MonoImage* as a pointer difference from ptr_base
221 * [flags: uleb128] must be 0
222 * [name: string] full class name
223 * if mtype == TYPE_IMAGE
224 * [flags: uleb128] must be 0
225 * [name: string] image file name
226 * if mtype == TYPE_THREAD
227 * [flags: uleb128] must be 0
228 * [name: string] thread name
230 * type method format:
232 * exinfo: one of: TYPE_LEAVE, TYPE_ENTER, TYPE_EXC_LEAVE, TYPE_JIT
233 * [time diff: uleb128] nanoseconds since last timing
234 * [method: sleb128] MonoMethod* as a pointer difference from the last such
235 * pointer or the buffer method_base
236 * if exinfo == TYPE_JIT
237 * [code address: sleb128] pointer to the native code as a diff from ptr_base
238 * [code size: uleb128] size of the generated code
239 * [name: string] full method name
241 * type exception format:
242 * type: TYPE_EXCEPTION
243 * exinfo: TYPE_EXCEPTION_BT flag and one of: TYPE_THROW, TYPE_CLAUSE
244 * [time diff: uleb128] nanoseconds since last timing
245 * if exinfo.low3bits == TYPE_CLAUSE
246 * [clause type: uleb128] finally/catch/fault/filter
247 * [clause num: uleb128] the clause number in the method header
248 * [method: sleb128] MonoMethod* as a pointer difference from the last such
249 * pointer or the buffer method_base
250 * if exinfo.low3bits == TYPE_THROW
251 * [object: sleb128] the object that was thrown as a difference from obj_base
252 * If the TYPE_EXCEPTION_BT flag is set, a backtrace follows.
254 * type monitor format:
256 * exinfo: TYPE_MONITOR_BT flag and one of: MONO_PROFILER_MONITOR_(CONTENTION|FAIL|DONE)
257 * [time diff: uleb128] nanoseconds since last timing
258 * [object: sleb128] the lock object as a difference from obj_base
259 * if exinfo.low3bits == MONO_PROFILER_MONITOR_CONTENTION
260 * If the TYPE_MONITOR_BT flag is set, a backtrace follows.
264 * exinfo: one of TYPE_HEAP_START, TYPE_HEAP_END, TYPE_HEAP_OBJECT, TYPE_HEAP_ROOT
265 * if exinfo == TYPE_HEAP_START
266 * [time diff: uleb128] nanoseconds since last timing
267 * if exinfo == TYPE_HEAP_END
268 * [time diff: uleb128] nanoseconds since last timing
269 * if exinfo == TYPE_HEAP_OBJECT
270 * [object: sleb128] the object as a difference from obj_base
271 * [class: sleb128] the object MonoClass* as a difference from ptr_base
272 * [size: uleb128] size of the object on the heap
273 * [num_refs: uleb128] number of object references
274 * if (format version > 1) each referenced objref is preceded by a
275 * uleb128 encoded offset: the first offset is from the object address
276 * and each next offset is relative to the previous one
277 * [objrefs: sleb128]+ object referenced as a difference from obj_base
278 * The same object can appear multiple times, but only the first time
279 * with size != 0: in the other cases this data will only be used to
280 * provide additional referenced objects.
281 * if exinfo == TYPE_HEAP_ROOT
282 * [num_roots: uleb128] number of root references
283 * [num_gc: uleb128] number of major gcs
284 * [object: sleb128] the object as a difference from obj_base
285 * [root_type: uleb128] the root_type: MonoProfileGCRootType (profiler.h)
286 * [extra_info: uleb128] the extra_info value
287 * object, root_type and extra_info are repeated num_roots times
291 * exinfo: one of TYPE_SAMPLE_HIT, TYPE_SAMPLE_USYM, TYPE_SAMPLE_UBIN, TYPE_SAMPLE_COUNTERS_DESC, TYPE_SAMPLE_COUNTERS
292 * if exinfo == TYPE_SAMPLE_HIT
293 * [sample_type: uleb128] type of sample (SAMPLE_*)
294 * [timestamp: uleb128] nanoseconds since startup (note: different from other timestamps!)
295 * [count: uleb128] number of following instruction addresses
296 * [ip: sleb128]* instruction pointer as difference from ptr_base
297 * if (format_version > 5)
298 * [mbt_count: uleb128] number of managed backtrace info triplets (method + IL offset + native offset)
299 * [method: sleb128]* MonoMethod* as a pointer difference from the last such
300 * pointer or the buffer method_base (the first such method can be also indentified by ip, but this is not neccessarily true)
301 * [il_offset: sleb128]* IL offset inside method where the hit occurred
302 * [native_offset: sleb128]* native offset inside method where the hit occurred
303 * if exinfo == TYPE_SAMPLE_USYM
304 * [address: sleb128] symbol address as a difference from ptr_base
305 * [size: uleb128] symbol size (may be 0 if unknown)
306 * [name: string] symbol name
307 * if exinfo == TYPE_SAMPLE_UBIN
308 * [time diff: uleb128] nanoseconds since last timing
309 * [address: sleb128] address where binary has been loaded
310 * [offset: uleb128] file offset of mapping (the same file can be mapped multiple times)
311 * [size: uleb128] memory size
312 * [name: string] binary name
313 * if exinfo == TYPE_SAMPLE_COUNTERS_DESC
314 * [len: uleb128] number of counters
316 * [section: uleb128] section name of counter
317 * [name: string] name of counter
318 * [type: uleb128] type name of counter
319 * [unit: uleb128] unit name of counter
320 * [variance: uleb128] variance name of counter
321 * [index: uleb128] unique index of counter
322 * if exinfo == TYPE_SAMPLE_COUNTERS
323 * [timestamp: uleb128] sampling timestamp
325 * [index: uleb128] unique index of counter
328 * [type: uleb128] type of counter value
331 * [0: uleb128] 0 -> value is null
333 * [1: uleb128] 1 -> value is not null
334 * [value: string] counter value
336 * [value: uleb128/sleb128/double] counter value, can be sleb128, uleb128 or double (determined by using type)
344 uintptr_t method_base;
345 uintptr_t last_method;
348 unsigned char* data_end;
353 unsigned char buf [1];
357 ign_res (int G_GNUC_UNUSED unused, ...)
361 #define ENTER_LOG(lb,str) if ((lb)->locked) {ign_res (write(2, str, strlen(str))); ign_res (write(2, "\n", 1));return;} else {(lb)->locked++;}
362 #define EXIT_LOG(lb) (lb)->locked--;
364 typedef struct _StatBuffer StatBuffer;
373 typedef struct _BinaryObject BinaryObject;
375 struct _BinaryObject {
381 struct _MonoProfiler {
383 StatBuffer *stat_buffers;
385 #if defined (HAVE_SYS_ZLIB)
388 uint64_t startup_time;
390 int last_gc_gen_started;
395 pthread_t helper_thread;
397 BinaryObject *binary_objects;
401 #define TLS_SET(x,y) TlsSetValue(x, y)
402 #define TLS_GET(x) ((LogBuffer *) TlsGetValue(x))
403 #define TLS_INIT(x) x = TlsAlloc ()
404 static int tlsbuffer;
406 #define TLS_SET(x,y) x = y
409 static __thread LogBuffer* tlsbuffer = NULL;
411 #define TLS_SET(x,y) pthread_setspecific(x, y)
412 #define TLS_GET(x) ((LogBuffer *) pthread_getspecific(x))
413 #define TLS_INIT(x) pthread_key_create(&x, NULL)
414 static pthread_key_t tlsbuffer;
417 static void safe_dump (MonoProfiler *profiler, LogBuffer *logbuffer);
420 pstrdup (const char *s)
422 int len = strlen (s) + 1;
423 char *p = malloc (len);
429 create_stat_buffer (void)
431 StatBuffer* buf = alloc_buffer (BUFFER_SIZE);
432 buf->size = BUFFER_SIZE;
433 buf->data_end = (uintptr_t*)((unsigned char*)buf + buf->size);
434 buf->data = buf->buf;
441 LogBuffer* buf = alloc_buffer (BUFFER_SIZE);
442 buf->size = BUFFER_SIZE;
443 buf->time_base = current_time ();
444 buf->last_time = buf->time_base;
445 buf->data_end = (unsigned char*)buf + buf->size;
446 buf->data = buf->buf;
453 LogBuffer *logbuffer;
454 if (TLS_GET (tlsbuffer))
456 logbuffer = create_buffer ();
457 TLS_SET (tlsbuffer, logbuffer);
458 logbuffer->thread_id = thread_id ();
459 //printf ("thread %p at time %llu\n", (void*)logbuffer->thread_id, logbuffer->time_base);
463 ensure_logbuf (int bytes)
465 LogBuffer *old = TLS_GET (tlsbuffer);
466 if (old && old->data + bytes + 100 < old->data_end)
468 TLS_SET (tlsbuffer, NULL);
470 TLS_GET (tlsbuffer)->next = old;
472 TLS_GET (tlsbuffer)->call_depth = old->call_depth;
473 //printf ("new logbuffer\n");
474 return TLS_GET (tlsbuffer);
478 emit_byte (LogBuffer *logbuffer, int value)
480 logbuffer->data [0] = value;
482 assert (logbuffer->data <= logbuffer->data_end);
486 emit_value (LogBuffer *logbuffer, int value)
488 encode_uleb128 (value, logbuffer->data, &logbuffer->data);
489 assert (logbuffer->data <= logbuffer->data_end);
493 emit_time (LogBuffer *logbuffer, uint64_t value)
495 uint64_t tdiff = value - logbuffer->last_time;
497 if (value < logbuffer->last_time)
498 printf ("time went backwards\n");
499 //if (tdiff > 1000000)
500 // printf ("large time offset: %llu\n", tdiff);
502 encode_uleb128 (tdiff, logbuffer->data, &logbuffer->data);
503 /*if (tdiff != decode_uleb128 (p, &p))
504 printf ("incorrect encoding: %llu\n", tdiff);*/
505 logbuffer->last_time = value;
506 assert (logbuffer->data <= logbuffer->data_end);
510 emit_svalue (LogBuffer *logbuffer, int64_t value)
512 encode_sleb128 (value, logbuffer->data, &logbuffer->data);
513 assert (logbuffer->data <= logbuffer->data_end);
517 emit_uvalue (LogBuffer *logbuffer, uint64_t value)
519 encode_uleb128 (value, logbuffer->data, &logbuffer->data);
520 assert (logbuffer->data <= logbuffer->data_end);
524 emit_ptr (LogBuffer *logbuffer, void *ptr)
526 if (!logbuffer->ptr_base)
527 logbuffer->ptr_base = (uintptr_t)ptr;
528 emit_svalue (logbuffer, (intptr_t)ptr - logbuffer->ptr_base);
529 assert (logbuffer->data <= logbuffer->data_end);
533 emit_method (LogBuffer *logbuffer, void *method)
535 if (!logbuffer->method_base) {
536 logbuffer->method_base = (intptr_t)method;
537 logbuffer->last_method = (intptr_t)method;
539 encode_sleb128 ((intptr_t)((char*)method - (char*)logbuffer->last_method), logbuffer->data, &logbuffer->data);
540 logbuffer->last_method = (intptr_t)method;
541 assert (logbuffer->data <= logbuffer->data_end);
545 emit_obj (LogBuffer *logbuffer, void *ptr)
547 if (!logbuffer->obj_base)
548 logbuffer->obj_base = (uintptr_t)ptr >> 3;
549 emit_svalue (logbuffer, ((uintptr_t)ptr >> 3) - logbuffer->obj_base);
550 assert (logbuffer->data <= logbuffer->data_end);
554 emit_string (LogBuffer *logbuffer, const char *str, size_t size)
558 for (; i < size; i++) {
561 emit_byte (logbuffer, str [i]);
564 emit_byte (logbuffer, '\0');
568 emit_double (LogBuffer *logbuffer, double value)
571 unsigned char buffer[8];
572 memcpy (buffer, &value, 8);
573 #if G_BYTE_ORDER == G_BIG_ENDIAN
574 for (i = 7; i >= 0; i--)
576 for (i = 0; i < 8; i++)
578 emit_byte (logbuffer, buffer[i]);
582 write_int16 (char *buf, int32_t value)
585 for (i = 0; i < 2; ++i) {
593 write_int32 (char *buf, int32_t value)
596 for (i = 0; i < 4; ++i) {
604 write_int64 (char *buf, int64_t value)
607 for (i = 0; i < 8; ++i) {
615 dump_header (MonoProfiler *profiler)
619 p = write_int32 (p, LOG_HEADER_ID);
620 *p++ = LOG_VERSION_MAJOR;
621 *p++ = LOG_VERSION_MINOR;
622 *p++ = LOG_DATA_VERSION;
623 *p++ = sizeof (void*);
624 p = write_int64 (p, ((uint64_t)time (NULL)) * 1000); /* startup time */
625 p = write_int32 (p, get_timer_overhead ()); /* timer overhead */
626 p = write_int32 (p, 0); /* flags */
627 p = write_int32 (p, process_id ()); /* pid */
628 p = write_int16 (p, profiler->command_port); /* port */
629 p = write_int16 (p, 0); /* opsystem */
630 #if defined (HAVE_SYS_ZLIB)
631 if (profiler->gzfile) {
632 gzwrite (profiler->gzfile, hbuf, p - hbuf);
634 fwrite (hbuf, p - hbuf, 1, profiler->file);
637 fwrite (hbuf, p - hbuf, 1, profiler->file);
638 fflush (profiler->file);
643 dump_buffer (MonoProfiler *profiler, LogBuffer *buf)
648 dump_buffer (profiler, buf->next);
649 p = write_int32 (p, BUF_ID);
650 p = write_int32 (p, buf->data - buf->buf);
651 p = write_int64 (p, buf->time_base);
652 p = write_int64 (p, buf->ptr_base);
653 p = write_int64 (p, buf->obj_base);
654 p = write_int64 (p, buf->thread_id);
655 p = write_int64 (p, buf->method_base);
656 #if defined (HAVE_SYS_ZLIB)
657 if (profiler->gzfile) {
658 gzwrite (profiler->gzfile, hbuf, p - hbuf);
659 gzwrite (profiler->gzfile, buf->buf, buf->data - buf->buf);
662 fwrite (hbuf, p - hbuf, 1, profiler->file);
663 fwrite (buf->buf, buf->data - buf->buf, 1, profiler->file);
664 fflush (profiler->file);
665 #if defined (HAVE_SYS_ZLIB)
668 free_buffer (buf, buf->size);
672 process_requests (MonoProfiler *profiler)
674 if (heapshot_requested)
675 mono_gc_collect (mono_gc_max_generation ());
678 static void counters_init (MonoProfiler *profiler);
681 runtime_initialized (MonoProfiler *profiler)
684 #ifndef DISABLE_HELPER_THREAD
685 counters_init (profiler);
687 /* ensure the main thread data and startup are available soon */
688 safe_dump (profiler, ensure_logbuf (0));
692 * Can be called only at safe callback locations.
695 safe_dump (MonoProfiler *profiler, LogBuffer *logbuffer)
697 int cd = logbuffer->call_depth;
699 dump_buffer (profiler, TLS_GET (tlsbuffer));
701 TLS_SET (tlsbuffer, NULL);
703 TLS_GET (tlsbuffer)->call_depth = cd;
707 gc_reference (MonoObject *obj, MonoClass *klass, uintptr_t size, uintptr_t num, MonoObject **refs, uintptr_t *offsets, void *data)
710 uintptr_t last_offset = 0;
711 //const char *name = mono_class_get_name (klass);
712 LogBuffer *logbuffer = ensure_logbuf (20 + num * 8);
713 emit_byte (logbuffer, TYPE_HEAP_OBJECT | TYPE_HEAP);
714 emit_obj (logbuffer, obj);
715 emit_ptr (logbuffer, klass);
716 /* account for object alignment in the heap */
719 emit_value (logbuffer, size);
720 emit_value (logbuffer, num);
721 for (i = 0; i < num; ++i) {
722 emit_value (logbuffer, offsets [i] - last_offset);
723 last_offset = offsets [i];
724 emit_obj (logbuffer, refs [i]);
727 // printf ("obj: %p, klass: %s, refs: %d, size: %d\n", obj, name, (int)num, (int)size);
731 static unsigned int hs_mode_ms = 0;
732 static unsigned int hs_mode_gc = 0;
733 static unsigned int hs_mode_ondemand = 0;
734 static unsigned int gc_count = 0;
735 static uint64_t last_hs_time = 0;
738 heap_walk (MonoProfiler *profiler)
742 LogBuffer *logbuffer;
745 logbuffer = ensure_logbuf (10);
746 now = current_time ();
747 if (hs_mode_ms && (now - last_hs_time)/1000000 >= hs_mode_ms)
749 else if (hs_mode_gc && (gc_count % hs_mode_gc) == 0)
751 else if (hs_mode_ondemand)
752 do_walk = heapshot_requested;
753 else if (!hs_mode_ms && !hs_mode_gc && profiler->last_gc_gen_started == mono_gc_max_generation ())
758 heapshot_requested = 0;
759 emit_byte (logbuffer, TYPE_HEAP_START | TYPE_HEAP);
760 emit_time (logbuffer, now);
761 mono_gc_walk_heap (0, gc_reference, NULL);
762 logbuffer = ensure_logbuf (10);
763 now = current_time ();
764 emit_byte (logbuffer, TYPE_HEAP_END | TYPE_HEAP);
765 emit_time (logbuffer, now);
770 gc_event (MonoProfiler *profiler, MonoGCEvent ev, int generation) {
772 LogBuffer *logbuffer = ensure_logbuf (10);
773 now = current_time ();
774 ENTER_LOG (logbuffer, "gcevent");
775 emit_byte (logbuffer, TYPE_GC_EVENT | TYPE_GC);
776 emit_time (logbuffer, now);
777 emit_value (logbuffer, ev);
778 emit_value (logbuffer, generation);
779 /* to deal with nested gen1 after gen0 started */
780 if (ev == MONO_GC_EVENT_START) {
781 profiler->last_gc_gen_started = generation;
782 if (generation == mono_gc_max_generation ())
785 if (ev == MONO_GC_EVENT_PRE_START_WORLD)
786 heap_walk (profiler);
787 EXIT_LOG (logbuffer);
788 if (ev == MONO_GC_EVENT_POST_START_WORLD)
789 safe_dump (profiler, logbuffer);
790 //printf ("gc event %d for generation %d\n", ev, generation);
794 gc_resize (MonoProfiler *profiler, int64_t new_size) {
796 LogBuffer *logbuffer = ensure_logbuf (10);
797 now = current_time ();
798 ENTER_LOG (logbuffer, "gcresize");
799 emit_byte (logbuffer, TYPE_GC_RESIZE | TYPE_GC);
800 emit_time (logbuffer, now);
801 emit_value (logbuffer, new_size);
802 //printf ("gc resized to %lld\n", new_size);
803 EXIT_LOG (logbuffer);
806 #define MAX_FRAMES 32
809 MonoMethod* methods [MAX_FRAMES];
810 int32_t il_offsets [MAX_FRAMES];
811 int32_t native_offsets [MAX_FRAMES];
813 static int num_frames = MAX_FRAMES;
816 walk_stack (MonoMethod *method, int32_t native_offset, int32_t il_offset, mono_bool managed, void* data)
818 FrameData *frame = data;
819 if (method && frame->count < num_frames) {
820 frame->il_offsets [frame->count] = il_offset;
821 frame->native_offsets [frame->count] = native_offset;
822 frame->methods [frame->count++] = method;
823 //printf ("In %d %s at %d (native: %d)\n", frame->count, mono_method_get_name (method), il_offset, native_offset);
825 return frame->count == num_frames;
829 * a note about stack walks: they can cause more profiler events to fire,
830 * so we need to make sure they don't happen after we started emitting an
831 * event, hence the collect_bt/emit_bt split.
834 collect_bt (FrameData *data, gboolean async_safe)
838 mono_stack_walk_async_safe (walk_stack, data);
840 mono_stack_walk_no_il (walk_stack, data);
844 emit_bt (LogBuffer *logbuffer, FrameData *data)
846 /* FIXME: this is actually tons of data and we should
847 * just output it the first time and use an id the next
849 if (data->count > num_frames)
850 printf ("bad num frames: %d\n", data->count);
851 emit_value (logbuffer, 0); /* flags */
852 emit_value (logbuffer, data->count);
853 //if (*p != data.count) {
854 // 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);}
855 while (data->count) {
856 emit_ptr (logbuffer, data->methods [--data->count]);
861 gc_alloc (MonoProfiler *prof, MonoObject *obj, MonoClass *klass)
865 int do_bt = (nocalls && runtime_inited && !notraces)? TYPE_ALLOC_BT: 0;
867 LogBuffer *logbuffer;
868 len = mono_object_get_size (obj);
869 /* account for object alignment in the heap */
873 collect_bt (&data, FALSE);
874 logbuffer = ensure_logbuf (32 + MAX_FRAMES * 8);
875 now = current_time ();
876 ENTER_LOG (logbuffer, "gcalloc");
877 emit_byte (logbuffer, do_bt | TYPE_ALLOC);
878 emit_time (logbuffer, now);
879 emit_ptr (logbuffer, klass);
880 emit_obj (logbuffer, obj);
881 emit_value (logbuffer, len);
883 emit_bt (logbuffer, &data);
884 EXIT_LOG (logbuffer);
886 safe_dump (prof, logbuffer);
887 process_requests (prof);
888 //printf ("gc alloc %s at %p\n", mono_class_get_name (klass), obj);
892 gc_moves (MonoProfiler *prof, void **objects, int num)
896 LogBuffer *logbuffer = ensure_logbuf (10 + num * 8);
897 now = current_time ();
898 ENTER_LOG (logbuffer, "gcmove");
899 emit_byte (logbuffer, TYPE_GC_MOVE | TYPE_GC);
900 emit_time (logbuffer, now);
901 emit_value (logbuffer, num);
902 for (i = 0; i < num; ++i)
903 emit_obj (logbuffer, objects [i]);
904 //printf ("gc moved %d objects\n", num/2);
905 EXIT_LOG (logbuffer);
909 gc_roots (MonoProfiler *prof, int num, void **objects, int *root_types, uintptr_t *extra_info)
912 LogBuffer *logbuffer = ensure_logbuf (5 + num * 18);
913 ENTER_LOG (logbuffer, "gcroots");
914 emit_byte (logbuffer, TYPE_HEAP_ROOT | TYPE_HEAP);
915 emit_value (logbuffer, num);
916 emit_value (logbuffer, mono_gc_collection_count (mono_gc_max_generation ()));
917 for (i = 0; i < num; ++i) {
918 emit_obj (logbuffer, objects [i]);
919 emit_value (logbuffer, root_types [i]);
920 emit_value (logbuffer, extra_info [i]);
922 EXIT_LOG (logbuffer);
926 gc_handle (MonoProfiler *prof, int op, int type, uintptr_t handle, MonoObject *obj)
929 LogBuffer *logbuffer = ensure_logbuf (16);
930 now = current_time ();
931 ENTER_LOG (logbuffer, "gchandle");
932 if (op == MONO_PROFILER_GC_HANDLE_CREATED)
933 emit_byte (logbuffer, TYPE_GC_HANDLE_CREATED | TYPE_GC);
934 else if (op == MONO_PROFILER_GC_HANDLE_DESTROYED)
935 emit_byte (logbuffer, TYPE_GC_HANDLE_DESTROYED | TYPE_GC);
938 emit_time (logbuffer, now);
939 emit_value (logbuffer, type);
940 emit_value (logbuffer, handle);
941 if (op == MONO_PROFILER_GC_HANDLE_CREATED)
942 emit_obj (logbuffer, obj);
943 EXIT_LOG (logbuffer);
944 process_requests (prof);
948 push_nesting (char *p, MonoClass *klass)
953 nesting = mono_class_get_nesting_type (klass);
955 p = push_nesting (p, nesting);
959 name = mono_class_get_name (klass);
960 nspace = mono_class_get_namespace (klass);
963 p += strlen (nspace);
973 type_name (MonoClass *klass)
977 push_nesting (buf, klass);
978 p = malloc (strlen (buf) + 1);
984 image_loaded (MonoProfiler *prof, MonoImage *image, int result)
989 LogBuffer *logbuffer;
990 if (result != MONO_PROFILE_OK)
992 name = mono_image_get_filename (image);
993 nlen = strlen (name) + 1;
994 logbuffer = ensure_logbuf (16 + nlen);
995 now = current_time ();
996 ENTER_LOG (logbuffer, "image");
997 emit_byte (logbuffer, TYPE_END_LOAD | TYPE_METADATA);
998 emit_time (logbuffer, now);
999 emit_byte (logbuffer, TYPE_IMAGE);
1000 emit_ptr (logbuffer, image);
1001 emit_value (logbuffer, 0); /* flags */
1002 memcpy (logbuffer->data, name, nlen);
1003 logbuffer->data += nlen;
1004 //printf ("loaded image %p (%s)\n", image, name);
1005 EXIT_LOG (logbuffer);
1006 if (logbuffer->next)
1007 safe_dump (prof, logbuffer);
1008 process_requests (prof);
1012 class_loaded (MonoProfiler *prof, MonoClass *klass, int result)
1018 LogBuffer *logbuffer;
1019 if (result != MONO_PROFILE_OK)
1022 name = mono_type_get_name (mono_class_get_type (klass));
1024 name = type_name (klass);
1025 nlen = strlen (name) + 1;
1026 image = mono_class_get_image (klass);
1027 logbuffer = ensure_logbuf (24 + nlen);
1028 now = current_time ();
1029 ENTER_LOG (logbuffer, "class");
1030 emit_byte (logbuffer, TYPE_END_LOAD | TYPE_METADATA);
1031 emit_time (logbuffer, now);
1032 emit_byte (logbuffer, TYPE_CLASS);
1033 emit_ptr (logbuffer, klass);
1034 emit_ptr (logbuffer, image);
1035 emit_value (logbuffer, 0); /* flags */
1036 memcpy (logbuffer->data, name, nlen);
1037 logbuffer->data += nlen;
1038 //printf ("loaded class %p (%s)\n", klass, name);
1043 EXIT_LOG (logbuffer);
1044 if (logbuffer->next)
1045 safe_dump (prof, logbuffer);
1046 process_requests (prof);
1050 method_enter (MonoProfiler *prof, MonoMethod *method)
1053 LogBuffer *logbuffer = ensure_logbuf (16);
1054 if (logbuffer->call_depth++ > max_call_depth)
1056 now = current_time ();
1057 ENTER_LOG (logbuffer, "enter");
1058 emit_byte (logbuffer, TYPE_ENTER | TYPE_METHOD);
1059 emit_time (logbuffer, now);
1060 emit_method (logbuffer, method);
1061 EXIT_LOG (logbuffer);
1062 process_requests (prof);
1066 method_leave (MonoProfiler *prof, MonoMethod *method)
1069 LogBuffer *logbuffer = ensure_logbuf (16);
1070 if (--logbuffer->call_depth > max_call_depth)
1072 now = current_time ();
1073 ENTER_LOG (logbuffer, "leave");
1074 emit_byte (logbuffer, TYPE_LEAVE | TYPE_METHOD);
1075 emit_time (logbuffer, now);
1076 emit_method (logbuffer, method);
1077 EXIT_LOG (logbuffer);
1078 if (logbuffer->next)
1079 safe_dump (prof, logbuffer);
1080 process_requests (prof);
1084 method_exc_leave (MonoProfiler *prof, MonoMethod *method)
1087 LogBuffer *logbuffer;
1090 logbuffer = ensure_logbuf (16);
1091 if (--logbuffer->call_depth > max_call_depth)
1093 now = current_time ();
1094 ENTER_LOG (logbuffer, "eleave");
1095 emit_byte (logbuffer, TYPE_EXC_LEAVE | TYPE_METHOD);
1096 emit_time (logbuffer, now);
1097 emit_method (logbuffer, method);
1098 EXIT_LOG (logbuffer);
1099 process_requests (prof);
1103 method_jitted (MonoProfiler *prof, MonoMethod *method, MonoJitInfo* jinfo, int result)
1108 LogBuffer *logbuffer;
1109 if (result != MONO_PROFILE_OK)
1111 name = mono_method_full_name (method, 1);
1112 nlen = strlen (name) + 1;
1113 logbuffer = ensure_logbuf (32 + nlen);
1114 now = current_time ();
1115 ENTER_LOG (logbuffer, "jit");
1116 emit_byte (logbuffer, TYPE_JIT | TYPE_METHOD);
1117 emit_time (logbuffer, now);
1118 emit_method (logbuffer, method);
1119 emit_ptr (logbuffer, mono_jit_info_get_code_start (jinfo));
1120 emit_value (logbuffer, mono_jit_info_get_code_size (jinfo));
1121 memcpy (logbuffer->data, name, nlen);
1122 logbuffer->data += nlen;
1124 EXIT_LOG (logbuffer);
1125 if (logbuffer->next)
1126 safe_dump (prof, logbuffer);
1127 process_requests (prof);
1131 throw_exc (MonoProfiler *prof, MonoObject *object)
1133 int do_bt = (nocalls && runtime_inited && !notraces)? TYPE_EXCEPTION_BT: 0;
1136 LogBuffer *logbuffer;
1138 collect_bt (&data, FALSE);
1139 logbuffer = ensure_logbuf (16 + MAX_FRAMES * 8);
1140 now = current_time ();
1141 ENTER_LOG (logbuffer, "throw");
1142 emit_byte (logbuffer, do_bt | TYPE_EXCEPTION);
1143 emit_time (logbuffer, now);
1144 emit_obj (logbuffer, object);
1146 emit_bt (logbuffer, &data);
1147 EXIT_LOG (logbuffer);
1148 process_requests (prof);
1152 clause_exc (MonoProfiler *prof, MonoMethod *method, int clause_type, int clause_num)
1155 LogBuffer *logbuffer = ensure_logbuf (16);
1156 now = current_time ();
1157 ENTER_LOG (logbuffer, "clause");
1158 emit_byte (logbuffer, TYPE_EXCEPTION | TYPE_CLAUSE);
1159 emit_time (logbuffer, now);
1160 emit_value (logbuffer, clause_type);
1161 emit_value (logbuffer, clause_num);
1162 emit_method (logbuffer, method);
1163 EXIT_LOG (logbuffer);
1167 monitor_event (MonoProfiler *profiler, MonoObject *object, MonoProfilerMonitorEvent event)
1169 int do_bt = (nocalls && runtime_inited && !notraces && event == MONO_PROFILER_MONITOR_CONTENTION)? TYPE_MONITOR_BT: 0;
1172 LogBuffer *logbuffer;
1174 collect_bt (&data, FALSE);
1175 logbuffer = ensure_logbuf (16 + MAX_FRAMES * 8);
1176 now = current_time ();
1177 ENTER_LOG (logbuffer, "monitor");
1178 emit_byte (logbuffer, (event << 4) | do_bt | TYPE_MONITOR);
1179 emit_time (logbuffer, now);
1180 emit_obj (logbuffer, object);
1182 emit_bt (logbuffer, &data);
1183 EXIT_LOG (logbuffer);
1184 process_requests (profiler);
1188 thread_start (MonoProfiler *prof, uintptr_t tid)
1190 //printf ("thread start %p\n", (void*)tid);
1195 thread_end (MonoProfiler *prof, uintptr_t tid)
1198 if (TLS_GET (tlsbuffer))
1199 dump_buffer (prof, TLS_GET (tlsbuffer));
1201 TLS_SET (tlsbuffer, NULL);
1205 thread_name (MonoProfiler *prof, uintptr_t tid, const char *name)
1207 int len = strlen (name) + 1;
1209 LogBuffer *logbuffer;
1210 logbuffer = ensure_logbuf (10 + len);
1211 now = current_time ();
1212 ENTER_LOG (logbuffer, "tname");
1213 emit_byte (logbuffer, TYPE_METADATA);
1214 emit_time (logbuffer, now);
1215 emit_byte (logbuffer, TYPE_THREAD);
1216 emit_ptr (logbuffer, (void*)tid);
1217 emit_value (logbuffer, 0); /* flags */
1218 memcpy (logbuffer->data, name, len);
1219 logbuffer->data += len;
1220 EXIT_LOG (logbuffer);
1224 mono_sample_hit (MonoProfiler *profiler, unsigned char *ip, void *context)
1229 uintptr_t *data, *new_data, *old_data;
1235 now = current_time ();
1236 collect_bt (&bt_data, TRUE);
1237 elapsed = (now - profiler->startup_time) / 10000;
1241 snprintf (buf, sizeof (buf), "hit at %p in thread %p after %llu ms\n", ip, (void*)thread_id (), (unsigned long long int)elapsed/100);
1243 ign_res (write (2, buf, len));
1245 sbuf = profiler->stat_buffers;
1248 /* flush the buffer at 1 second intervals */
1249 if (sbuf->data > sbuf->buf && (elapsed - sbuf->buf [2]) > 100000) {
1252 /* overflow: 400 slots is a big enough number to reduce the chance of losing this event if many
1253 * threads hit this same spot at the same time
1255 if (timedout || (sbuf->data + 400 >= sbuf->data_end)) {
1256 StatBuffer *oldsb, *foundsb;
1257 sbuf = create_stat_buffer ();
1259 oldsb = profiler->stat_buffers;
1261 foundsb = InterlockedCompareExchangePointer ((void * volatile*)&profiler->stat_buffers, sbuf, oldsb);
1262 } while (foundsb != oldsb);
1264 ign_res (write (2, "overflow\n", 9));
1265 /* notify the helper thread */
1266 if (sbuf->next->next) {
1268 ign_res (write (profiler->pipes [1], &c, 1));
1270 ign_res (write (2, "notify\n", 7));
1274 old_data = sbuf->data;
1275 new_data = old_data + 4 + bt_data.count * 3;
1276 data = InterlockedCompareExchangePointer ((void * volatile*)&sbuf->data, new_data, old_data);
1277 } while (data != old_data);
1278 if (old_data >= sbuf->data_end)
1279 return; /* lost event */
1280 old_data [0] = 1 | (sample_type << 16) | (bt_data.count << 8);
1281 old_data [1] = thread_id ();
1282 old_data [2] = elapsed;
1283 old_data [3] = (uintptr_t)ip;
1284 for (i = 0; i < bt_data.count; ++i) {
1285 old_data [4+3*i] = (uintptr_t)bt_data.methods [i];
1286 old_data [4+3*i+1] = (uintptr_t)bt_data.il_offsets [i];
1287 old_data [4+3*i+2] = (uintptr_t)bt_data.native_offsets [i];
1291 static uintptr_t *code_pages = 0;
1292 static int num_code_pages = 0;
1293 static int size_code_pages = 0;
1294 #define CPAGE_SHIFT (9)
1295 #define CPAGE_SIZE (1 << CPAGE_SHIFT)
1296 #define CPAGE_MASK (~(CPAGE_SIZE - 1))
1297 #define CPAGE_ADDR(p) ((p) & CPAGE_MASK)
1300 add_code_page (uintptr_t *hash, uintptr_t hsize, uintptr_t page)
1303 uintptr_t start_pos;
1304 start_pos = (page >> CPAGE_SHIFT) % hsize;
1307 if (hash [i] && CPAGE_ADDR (hash [i]) == CPAGE_ADDR (page)) {
1309 } else if (!hash [i]) {
1316 } while (i != start_pos);
1317 /* should not happen */
1318 printf ("failed code page store\n");
1323 add_code_pointer (uintptr_t ip)
1326 if (num_code_pages * 2 >= size_code_pages) {
1328 uintptr_t old_size = size_code_pages;
1329 size_code_pages *= 2;
1330 if (size_code_pages == 0)
1331 size_code_pages = 16;
1332 n = calloc (sizeof (uintptr_t) * size_code_pages, 1);
1333 for (i = 0; i < old_size; ++i) {
1335 add_code_page (n, size_code_pages, code_pages [i]);
1341 num_code_pages += add_code_page (code_pages, size_code_pages, ip & CPAGE_MASK);
1344 #if defined(HAVE_DL_ITERATE_PHDR) && defined(ELFMAG0)
1346 dump_ubin (const char *filename, uintptr_t load_addr, uint64_t offset, uintptr_t size)
1349 LogBuffer *logbuffer;
1351 len = strlen (filename) + 1;
1352 now = current_time ();
1353 logbuffer = ensure_logbuf (20 + len);
1354 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_UBIN);
1355 emit_time (logbuffer, now);
1356 emit_svalue (logbuffer, load_addr);
1357 emit_uvalue (logbuffer, offset);
1358 emit_uvalue (logbuffer, size);
1359 memcpy (logbuffer->data, filename, len);
1360 logbuffer->data += len;
1365 dump_usym (const char *name, uintptr_t value, uintptr_t size)
1367 LogBuffer *logbuffer;
1369 len = strlen (name) + 1;
1370 logbuffer = ensure_logbuf (20 + len);
1371 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_USYM);
1372 emit_ptr (logbuffer, (void*)value);
1373 emit_value (logbuffer, size);
1374 memcpy (logbuffer->data, name, len);
1375 logbuffer->data += len;
1380 #if SIZEOF_VOID_P == 4
1381 #define ELF_WSIZE 32
1383 #define ELF_WSIZE 64
1386 #define ElfW(type) _ElfW (Elf, ELF_WSIZE, type)
1387 #define _ElfW(e,w,t) _ElfW_1 (e, w, _##t)
1388 #define _ElfW_1(e,w,t) e##w##t
1392 dump_elf_symbols (ElfW(Sym) *symbols, int num_symbols, const char *strtab, void *load_addr)
1395 for (i = 0; i < num_symbols; ++i) {
1397 sym = strtab + symbols [i].st_name;
1398 if (!symbols [i].st_name || !symbols [i].st_size || (symbols [i].st_info & 0xf) != STT_FUNC)
1400 //printf ("symbol %s at %d\n", sym, symbols [i].st_value);
1401 dump_usym (sym, (uintptr_t)load_addr + symbols [i].st_value, symbols [i].st_size);
1406 read_elf_symbols (MonoProfiler *prof, const char *filename, void *load_addr)
1413 ElfW(Shdr) *sheader;
1414 ElfW(Shdr) *shstrtabh;
1415 ElfW(Shdr) *symtabh = NULL;
1416 ElfW(Shdr) *strtabh = NULL;
1417 ElfW(Sym) *symbols = NULL;
1421 fd = open (filename, O_RDONLY);
1424 if (fstat (fd, &statb) != 0) {
1428 file_size = statb.st_size;
1429 data = mmap (NULL, file_size, PROT_READ, MAP_PRIVATE, fd, 0);
1431 if (data == MAP_FAILED)
1434 if (header->e_ident [EI_MAG0] != ELFMAG0 ||
1435 header->e_ident [EI_MAG1] != ELFMAG1 ||
1436 header->e_ident [EI_MAG2] != ELFMAG2 ||
1437 header->e_ident [EI_MAG3] != ELFMAG3 ) {
1438 munmap (data, file_size);
1441 sheader = (void*)((char*)data + header->e_shoff);
1442 shstrtabh = (void*)((char*)sheader + (header->e_shentsize * header->e_shstrndx));
1443 strtab = (const char*)data + shstrtabh->sh_offset;
1444 for (i = 0; i < header->e_shnum; ++i) {
1445 //printf ("section header: %d\n", sheader->sh_type);
1446 if (sheader->sh_type == SHT_SYMTAB) {
1448 strtabh = (void*)((char*)data + header->e_shoff + sheader->sh_link * header->e_shentsize);
1449 /*printf ("symtab section header: %d, .strstr: %d\n", i, sheader->sh_link);*/
1452 sheader = (void*)((char*)sheader + header->e_shentsize);
1454 if (!symtabh || !strtabh) {
1455 munmap (data, file_size);
1458 strtab = (const char*)data + strtabh->sh_offset;
1459 num_symbols = symtabh->sh_size / symtabh->sh_entsize;
1460 symbols = (void*)((char*)data + symtabh->sh_offset);
1461 dump_elf_symbols (symbols, num_symbols, strtab, load_addr);
1462 munmap (data, file_size);
1467 #if defined(HAVE_DL_ITERATE_PHDR) && defined(ELFMAG0)
1469 elf_dl_callback (struct dl_phdr_info *info, size_t size, void *data)
1471 MonoProfiler *prof = data;
1473 const char *filename;
1475 char *a = (void*)info->dlpi_addr;
1477 ElfW(Dyn) *dyn = NULL;
1478 ElfW(Sym) *symtab = NULL;
1479 ElfW(Word) *hash_table = NULL;
1480 ElfW(Ehdr) *header = NULL;
1481 const char* strtab = NULL;
1482 for (obj = prof->binary_objects; obj; obj = obj->next) {
1486 filename = info->dlpi_name;
1489 if (!info->dlpi_addr && !filename [0]) {
1490 int l = readlink ("/proc/self/exe", buf, sizeof (buf) - 1);
1496 obj = calloc (sizeof (BinaryObject), 1);
1497 obj->addr = (void*)info->dlpi_addr;
1498 obj->name = pstrdup (filename);
1499 obj->next = prof->binary_objects;
1500 prof->binary_objects = obj;
1501 //printf ("loaded file: %s at %p, segments: %d\n", filename, (void*)info->dlpi_addr, info->dlpi_phnum);
1503 for (i = 0; i < info->dlpi_phnum; ++i) {
1504 //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);
1505 if (info->dlpi_phdr[i].p_type == PT_LOAD && !header) {
1506 header = (ElfW(Ehdr)*)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1507 if (header->e_ident [EI_MAG0] != ELFMAG0 ||
1508 header->e_ident [EI_MAG1] != ELFMAG1 ||
1509 header->e_ident [EI_MAG2] != ELFMAG2 ||
1510 header->e_ident [EI_MAG3] != ELFMAG3 ) {
1513 dump_ubin (filename, info->dlpi_addr + info->dlpi_phdr[i].p_vaddr, info->dlpi_phdr[i].p_offset, info->dlpi_phdr[i].p_memsz);
1514 } else if (info->dlpi_phdr[i].p_type == PT_DYNAMIC) {
1515 dyn = (ElfW(Dyn) *)(info->dlpi_addr + info->dlpi_phdr[i].p_vaddr);
1518 if (read_elf_symbols (prof, filename, (void*)info->dlpi_addr))
1520 if (!info->dlpi_name || !info->dlpi_name[0])
1524 for (i = 0; dyn [i].d_tag != DT_NULL; ++i) {
1525 if (dyn [i].d_tag == DT_SYMTAB) {
1526 if (symtab && do_debug)
1527 printf ("multiple symtabs: %d\n", i);
1528 symtab = (ElfW(Sym) *)(a + dyn [i].d_un.d_ptr);
1529 } else if (dyn [i].d_tag == DT_HASH) {
1530 hash_table = (ElfW(Word) *)(a + dyn [i].d_un.d_ptr);
1531 } else if (dyn [i].d_tag == DT_STRTAB) {
1532 strtab = (const char*)(a + dyn [i].d_un.d_ptr);
1537 num_sym = hash_table [1];
1538 dump_elf_symbols (symtab, num_sym, strtab, (void*)info->dlpi_addr);
1543 load_binaries (MonoProfiler *prof)
1545 dl_iterate_phdr (elf_dl_callback, prof);
1550 load_binaries (MonoProfiler *prof)
1557 symbol_for (uintptr_t code)
1560 void *ip = (void*)code;
1562 if (dladdr (ip, &di)) {
1564 return di.dli_sname;
1567 names = backtrace_symbols (&ip, 1);
1569 const char* p = names [0];
1580 dump_unmanaged_coderefs (MonoProfiler *prof)
1583 const char* last_symbol;
1584 uintptr_t addr, page_end;
1586 if (load_binaries (prof))
1588 for (i = 0; i < size_code_pages; ++i) {
1590 if (!code_pages [i] || code_pages [i] & 1)
1593 addr = CPAGE_ADDR (code_pages [i]);
1594 page_end = addr + CPAGE_SIZE;
1595 code_pages [i] |= 1;
1596 /* we dump the symbols for the whole page */
1597 for (; addr < page_end; addr += 16) {
1598 sym = symbol_for (addr);
1599 if (sym && sym == last_symbol)
1604 dump_usym (sym, addr, 0); /* let's not guess the size */
1605 //printf ("found symbol at %p: %s\n", (void*)addr, sym);
1611 dump_sample_hits (MonoProfiler *prof, StatBuffer *sbuf, int recurse)
1614 LogBuffer *logbuffer;
1617 if (recurse && sbuf->next) {
1618 dump_sample_hits (prof, sbuf->next, 1);
1619 free_buffer (sbuf->next, sbuf->next->size);
1622 for (sample = sbuf->buf; sample < sbuf->data;) {
1624 int count = sample [0] & 0xff;
1625 int mbt_count = (sample [0] & 0xff00) >> 8;
1626 int type = sample [0] >> 16;
1627 if (sample + count + 3 + mbt_count * 3 > sbuf->data)
1629 logbuffer = ensure_logbuf (20 + count * 8);
1630 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
1631 emit_value (logbuffer, type);
1632 emit_uvalue (logbuffer, prof->startup_time + (uint64_t)sample [2] * (uint64_t)10000);
1633 emit_value (logbuffer, count);
1634 for (i = 0; i < count; ++i) {
1635 emit_ptr (logbuffer, (void*)sample [i + 3]);
1636 add_code_pointer (sample [i + 3]);
1638 sample += count + 3;
1639 /* new in data version 6 */
1640 emit_uvalue (logbuffer, mbt_count);
1641 for (i = 0; i < mbt_count; ++i) {
1642 emit_method (logbuffer, (void*)sample [i * 3]); /* method */
1643 emit_svalue (logbuffer, sample [i * 3 + 1]); /* il offset */
1644 emit_svalue (logbuffer, sample [i * 3 + 2]); /* native offset */
1646 sample += 3 * mbt_count;
1648 dump_unmanaged_coderefs (prof);
1652 #ifndef __NR_perf_event_open
1654 #define __NR_perf_event_open 364
1656 #define __NR_perf_event_open 241
1661 mono_cpu_count (void)
1664 #ifdef PLATFORM_ANDROID
1665 /* Android tries really hard to save power by powering off CPUs on SMP phones which
1666 * means the normal way to query cpu count returns a wrong value with userspace API.
1667 * Instead we use /sys entries to query the actual hardware CPU count.
1669 char buffer[8] = {'\0'};
1670 int present = open ("/sys/devices/system/cpu/present", O_RDONLY);
1671 /* Format of the /sys entry is a cpulist of indexes which in the case
1672 * of present is always of the form "0-(n-1)" when there is more than
1673 * 1 core, n being the number of CPU cores in the system. Otherwise
1674 * the value is simply 0
1676 if (present != -1 && read (present, (char*)buffer, sizeof (buffer)) > 3)
1677 count = strtol (((char*)buffer) + 2, NULL, 10);
1683 #ifdef _SC_NPROCESSORS_ONLN
1684 count = sysconf (_SC_NPROCESSORS_ONLN);
1691 size_t len = sizeof (int);
1694 if (sysctl (mib, 2, &count, &len, NULL, 0) == 0)
1701 GetSystemInfo (&info);
1702 return info.dwNumberOfProcessors;
1711 unsigned int prev_pos;
1713 struct perf_event_mmap_page *page_desc;
1716 static PerfData *perf_data = NULL;
1717 static int num_perf;
1718 #define PERF_PAGES_SHIFT 4
1719 static int num_pages = 1 << PERF_PAGES_SHIFT;
1720 static unsigned int mmap_mask;
1723 struct perf_event_header h;
1733 perf_event_syscall (struct perf_event_attr *attr, pid_t pid, int cpu, int group_fd, unsigned long flags)
1735 attr->size = PERF_ATTR_SIZE_VER0;
1736 //printf ("perf attr size: %d\n", attr->size);
1737 #if defined(__x86_64__)
1738 return syscall(/*__NR_perf_event_open*/ 298, attr, pid, cpu, group_fd, flags);
1739 #elif defined(__i386__)
1740 return syscall(/*__NR_perf_event_open*/ 336, attr, pid, cpu, group_fd, flags);
1741 #elif defined(__arm__)
1742 return syscall(/*__NR_perf_event_open*/ 364, attr, pid, cpu, group_fd, flags);
1749 setup_perf_map (PerfData *perf)
1751 perf->mmap_base = mmap (NULL, (num_pages + 1) * getpagesize (), PROT_READ|PROT_WRITE, MAP_SHARED, perf->perf_fd, 0);
1752 if (perf->mmap_base == MAP_FAILED) {
1754 printf ("failed mmap\n");
1757 perf->page_desc = perf->mmap_base;
1759 printf ("mmap version: %d\n", perf->page_desc->version);
1764 dump_perf_hits (MonoProfiler *prof, void *buf, int size)
1766 LogBuffer *logbuffer;
1767 void *end = (char*)buf + size;
1769 int pid = getpid ();
1775 if (pid != s->pid) {
1777 printf ("event for different pid: %d\n", s->pid);
1778 buf = (char*)buf + s->h.size;
1781 /*ip = (void*)s->ip;
1782 printf ("sample: %d, size: %d, ip: %p (%s), timestamp: %llu, nframes: %llu\n",
1783 s->h.type, s->h.size, ip, symbol_for (ip), s->timestamp, s->nframes);*/
1784 logbuffer = ensure_logbuf (20 + s->nframes * 8);
1785 emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
1786 emit_value (logbuffer, sample_type);
1787 emit_uvalue (logbuffer, s->timestamp - prof->startup_time);
1788 emit_value (logbuffer, 1); /* count */
1789 emit_ptr (logbuffer, (void*)(uintptr_t)s->ip);
1790 /* no support here yet for the managed backtrace */
1791 emit_uvalue (logbuffer, 0);
1792 add_code_pointer (s->ip);
1793 buf = (char*)buf + s->h.size;
1797 printf ("dumped %d samples\n", samples);
1798 dump_unmanaged_coderefs (prof);
1801 /* read events from the ring buffer */
1803 read_perf_mmap (MonoProfiler* prof, int cpu)
1805 PerfData *perf = perf_data + cpu;
1807 unsigned char *data = (unsigned char*)perf->mmap_base + getpagesize ();
1808 unsigned int head = perf->page_desc->data_head;
1812 mono_memory_read_barrier ();
1814 old = perf->prev_pos;
1818 printf ("lost mmap events: old: %d, head: %d\n", old, head);
1822 if ((old & mmap_mask) + size != (head & mmap_mask)) {
1823 buf = data + (old & mmap_mask);
1824 size = mmap_mask + 1 - (old & mmap_mask);
1826 /* size bytes at buf */
1828 printf ("found1 bytes of events: %d\n", size);
1829 dump_perf_hits (prof, buf, size);
1831 buf = data + (old & mmap_mask);
1833 /* size bytes at buf */
1835 printf ("found bytes of events: %d\n", size);
1836 dump_perf_hits (prof, buf, size);
1838 perf->prev_pos = old;
1839 perf->page_desc->data_tail = old;
1844 setup_perf_event_for_cpu (PerfData *perf, int cpu)
1846 struct perf_event_attr attr;
1847 memset (&attr, 0, sizeof (attr));
1848 attr.type = PERF_TYPE_HARDWARE;
1849 switch (sample_type) {
1850 case SAMPLE_CYCLES: attr.config = PERF_COUNT_HW_CPU_CYCLES; break;
1851 case SAMPLE_INSTRUCTIONS: attr.config = PERF_COUNT_HW_INSTRUCTIONS; break;
1852 case SAMPLE_CACHE_MISSES: attr.config = PERF_COUNT_HW_CACHE_MISSES; break;
1853 case SAMPLE_CACHE_REFS: attr.config = PERF_COUNT_HW_CACHE_REFERENCES; break;
1854 case SAMPLE_BRANCHES: attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; break;
1855 case SAMPLE_BRANCH_MISSES: attr.config = PERF_COUNT_HW_BRANCH_MISSES; break;
1856 default: attr.config = PERF_COUNT_HW_CPU_CYCLES; break;
1858 attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID | PERF_SAMPLE_PERIOD | PERF_SAMPLE_TIME;
1859 // attr.sample_type |= PERF_SAMPLE_CALLCHAIN;
1860 attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | PERF_FORMAT_TOTAL_TIME_RUNNING | PERF_FORMAT_ID;
1863 attr.sample_freq = sample_freq;
1865 perf->perf_fd = perf_event_syscall (&attr, getpid (), cpu, -1, 0);
1867 printf ("perf fd: %d, freq: %d, event: %llu\n", perf->perf_fd, sample_freq, attr.config);
1868 if (perf->perf_fd < 0) {
1869 if (perf->perf_fd == -EPERM) {
1870 fprintf (stderr, "Perf syscall denied, do \"echo 1 > /proc/sys/kernel/perf_event_paranoid\" as root to enable.\n");
1873 perror ("open perf event");
1877 if (!setup_perf_map (perf)) {
1878 close (perf->perf_fd);
1886 setup_perf_event (void)
1889 mmap_mask = num_pages * getpagesize () - 1;
1890 num_perf = mono_cpu_count ();
1891 perf_data = calloc (num_perf, sizeof (PerfData));
1892 for (i = 0; i < num_perf; ++i) {
1893 count += setup_perf_event_for_cpu (perf_data + i, i);
1902 #endif /* USE_PERF_EVENTS */
1904 #ifndef DISABLE_HELPER_THREAD
1906 typedef struct MonoCounterAgent {
1907 MonoCounter *counter;
1908 // MonoCounterAgent specific data :
1912 struct MonoCounterAgent *next;
1915 static MonoCounterAgent* counters;
1916 static gboolean counters_initialized = FALSE;
1917 static int counters_index = 1;
1920 counters_init_add_counter (MonoCounter *counter, gpointer data)
1922 MonoCounterAgent *agent, *item;
1924 for (agent = counters; agent; agent = agent->next) {
1925 if (agent->counter == counter)
1929 agent = malloc (sizeof (MonoCounterAgent));
1930 agent->counter = counter;
1931 agent->value = NULL;
1932 agent->value_size = 0;
1933 agent->index = counters_index++;
1949 counters_init (MonoProfiler *profiler)
1951 MonoCounterAgent *agent;
1952 LogBuffer *logbuffer;
1953 int size = 1 + 5, len = 0;
1955 mono_counters_foreach (counters_init_add_counter, NULL);
1957 for (agent = counters; agent; agent = agent->next) {
1958 size += strlen (mono_counter_get_name (agent->counter)) + 1 + 5 * 5;
1962 logbuffer = ensure_logbuf (size);
1964 ENTER_LOG (logbuffer, "counters");
1965 emit_byte (logbuffer, TYPE_SAMPLE_COUNTERS_DESC | TYPE_SAMPLE);
1966 emit_value (logbuffer, len);
1967 for (agent = counters; agent; agent = agent->next) {
1968 const char *name = mono_counter_get_name (agent->counter);
1969 emit_value (logbuffer, mono_counter_get_section (agent->counter));
1970 emit_string (logbuffer, name, strlen (name) + 1);
1971 emit_value (logbuffer, mono_counter_get_type (agent->counter));
1972 emit_value (logbuffer, mono_counter_get_unit (agent->counter));
1973 emit_value (logbuffer, mono_counter_get_variance (agent->counter));
1974 emit_value (logbuffer, agent->index);
1976 EXIT_LOG (logbuffer);
1978 counters_initialized = TRUE;
1982 counters_sample (MonoProfiler *profiler, uint64_t timestamp)
1984 MonoCounterAgent *agent;
1985 MonoCounter *counter;
1986 LogBuffer *logbuffer;
1992 if (!counters_initialized)
1996 buffer = calloc (1, buffer_size);
1999 for (agent = counters; agent; agent = agent->next)
2000 size += 10 * 2 + mono_counter_get_size (agent->counter);
2002 logbuffer = ensure_logbuf (size);
2004 ENTER_LOG (logbuffer, "counters");
2005 emit_byte (logbuffer, TYPE_SAMPLE_COUNTERS | TYPE_SAMPLE);
2006 emit_uvalue (logbuffer, timestamp);
2007 for (agent = counters; agent; agent = agent->next) {
2010 counter = agent->counter;
2012 size = mono_counter_get_size (counter);
2014 continue; // FIXME error
2015 } else if (size > buffer_size) {
2017 buffer = realloc (buffer, buffer_size);
2020 memset (buffer, 0, buffer_size);
2022 if (mono_counters_sample (counter, buffer, size) < 0)
2023 continue; // FIXME error
2025 type = mono_counter_get_type (counter);
2027 if (!agent->value) {
2028 agent->value = calloc (1, size);
2029 agent->value_size = size;
2031 if (type == MONO_COUNTER_STRING) {
2032 if (strcmp (agent->value, buffer) == 0)
2035 if (agent->value_size == size && memcmp (agent->value, buffer, size) == 0)
2040 emit_uvalue (logbuffer, agent->index);
2041 emit_uvalue (logbuffer, type);
2043 case MONO_COUNTER_INT:
2044 #if SIZEOF_VOID_P == 4
2045 case MONO_COUNTER_WORD:
2047 emit_svalue (logbuffer, *(int*)buffer - *(int*)agent->value);
2049 case MONO_COUNTER_UINT:
2050 emit_uvalue (logbuffer, *(guint*)buffer - *(guint*)agent->value);
2052 case MONO_COUNTER_TIME_INTERVAL:
2053 case MONO_COUNTER_LONG:
2054 #if SIZEOF_VOID_P == 8
2055 case MONO_COUNTER_WORD:
2057 emit_svalue (logbuffer, *(gint64*)buffer - *(gint64*)agent->value);
2059 case MONO_COUNTER_ULONG:
2060 emit_uvalue (logbuffer, *(guint64*)buffer - *(guint64*)agent->value);
2062 case MONO_COUNTER_DOUBLE:
2063 emit_double (logbuffer, *(double*)buffer);
2065 case MONO_COUNTER_STRING:
2067 emit_byte (logbuffer, 0);
2069 emit_byte (logbuffer, 1);
2070 emit_string (logbuffer, (char*)buffer, size);
2077 if (type == MONO_COUNTER_STRING && size > agent->value_size) {
2078 agent->value = realloc (agent->value, size);
2079 agent->value_size = size;
2083 memcpy (agent->value, buffer, size);
2087 emit_value (logbuffer, 0);
2088 EXIT_LOG (logbuffer);
2090 safe_dump (profiler, ensure_logbuf (0));
2093 #endif /* DISABLE_HELPER_THREAD */
2096 log_shutdown (MonoProfiler *prof)
2099 #ifndef DISABLE_HELPER_THREAD
2100 if (prof->command_port) {
2103 ign_res (write (prof->pipes [1], &c, 1));
2104 pthread_join (prof->helper_thread, &res);
2110 for (i = 0; i < num_perf; ++i)
2111 read_perf_mmap (prof, i);
2114 dump_sample_hits (prof, prof->stat_buffers, 1);
2116 if (TLS_GET (tlsbuffer))
2117 dump_buffer (prof, TLS_GET (tlsbuffer));
2118 TLS_SET (tlsbuffer, NULL);
2120 #if defined (HAVE_SYS_ZLIB)
2122 gzclose (prof->gzfile);
2124 if (prof->pipe_output)
2125 pclose (prof->file);
2127 fclose (prof->file);
2132 new_filename (const char* filename)
2134 time_t t = time (NULL);
2135 int pid = process_id ();
2140 int count_dates = 0;
2144 for (p = filename; *p; p++) {
2155 if (!count_dates && !count_pids)
2156 return pstrdup (filename);
2157 snprintf (pid_buf, sizeof (pid_buf), "%d", pid);
2159 snprintf (time_buf, sizeof (time_buf), "%d%02d%02d%02d%02d%02d",
2160 1900 + ts->tm_year, 1 + ts->tm_mon, ts->tm_mday, ts->tm_hour, ts->tm_min, ts->tm_sec);
2161 s_date = strlen (time_buf);
2162 s_pid = strlen (pid_buf);
2163 d = res = malloc (strlen (filename) + s_date * count_dates + s_pid * count_pids);
2164 for (p = filename; *p; p++) {
2171 strcpy (d, time_buf);
2174 } else if (*p == 'p') {
2175 strcpy (d, pid_buf);
2178 } else if (*p == '%') {
2190 #ifndef DISABLE_HELPER_THREAD
2192 helper_thread (void* arg)
2194 MonoProfiler* prof = arg;
2198 MonoThread *thread = NULL;
2199 uint64_t start, now;
2201 //fprintf (stderr, "Server listening\n");
2202 start = current_time ();
2203 command_socket = -1;
2209 FD_SET (prof->server_socket, &rfds);
2210 max_fd = prof->server_socket;
2211 FD_SET (prof->pipes [0], &rfds);
2212 if (max_fd < prof->pipes [0])
2213 max_fd = prof->pipes [0];
2214 if (command_socket >= 0) {
2215 FD_SET (command_socket, &rfds);
2216 if (max_fd < command_socket)
2217 max_fd = command_socket;
2222 for ( i = 0; i < num_perf; ++i) {
2223 if (perf_data [i].perf_fd < 0)
2225 FD_SET (perf_data [i].perf_fd, &rfds);
2226 if (max_fd < perf_data [i].perf_fd)
2227 max_fd = perf_data [i].perf_fd;
2231 now = current_time ();
2232 counters_sample (prof, (now - start) / 1000/ 1000);
2236 len = select (max_fd + 1, &rfds, NULL, NULL, &tv);
2242 g_warning ("Error in proflog server: %s", strerror (errno));
2246 if (FD_ISSET (prof->pipes [0], &rfds)) {
2248 int r = read (prof->pipes [0], &c, 1);
2249 if (r == 1 && c == 0) {
2250 StatBuffer *sbufbase = prof->stat_buffers;
2252 if (!sbufbase->next)
2254 sbuf = sbufbase->next->next;
2255 sbufbase->next->next = NULL;
2257 fprintf (stderr, "stat buffer dump\n");
2258 dump_sample_hits (prof, sbuf, 1);
2259 free_buffer (sbuf, sbuf->size);
2260 safe_dump (prof, ensure_logbuf (0));
2263 /* time to shut down */
2265 mono_thread_detach (thread);
2267 fprintf (stderr, "helper shutdown\n");
2271 for ( i = 0; i < num_perf; ++i) {
2272 if (perf_data [i].perf_fd < 0)
2274 if (FD_ISSET (perf_data [i].perf_fd, &rfds))
2275 read_perf_mmap (prof, i);
2279 safe_dump (prof, ensure_logbuf (0));
2285 for ( i = 0; i < num_perf; ++i) {
2286 if (perf_data [i].perf_fd < 0)
2288 if (FD_ISSET (perf_data [i].perf_fd, &rfds)) {
2289 read_perf_mmap (prof, i);
2290 safe_dump (prof, ensure_logbuf (0));
2295 if (command_socket >= 0 && FD_ISSET (command_socket, &rfds)) {
2296 len = read (command_socket, buf, sizeof (buf) - 1);
2300 close (command_socket);
2301 command_socket = -1;
2305 if (strcmp (buf, "heapshot\n") == 0) {
2306 heapshot_requested = 1;
2307 //fprintf (stderr, "perform heapshot\n");
2308 if (runtime_inited && !thread) {
2309 thread = mono_thread_attach (mono_get_root_domain ());
2310 /*fprintf (stderr, "attached\n");*/
2313 process_requests (prof);
2314 mono_thread_detach (thread);
2320 if (!FD_ISSET (prof->server_socket, &rfds)) {
2323 command_socket = accept (prof->server_socket, NULL, NULL);
2324 if (command_socket < 0)
2326 //fprintf (stderr, "Accepted connection\n");
2332 start_helper_thread (MonoProfiler* prof)
2334 struct sockaddr_in server_address;
2337 if (pipe (prof->pipes) < 0) {
2338 fprintf (stderr, "Cannot create pipe\n");
2341 prof->server_socket = socket (PF_INET, SOCK_STREAM, 0);
2342 if (prof->server_socket < 0) {
2343 fprintf (stderr, "Cannot create server socket\n");
2346 memset (&server_address, 0, sizeof (server_address));
2347 server_address.sin_family = AF_INET;
2348 server_address.sin_addr.s_addr = INADDR_ANY;
2349 server_address.sin_port = htons (prof->command_port);
2350 if (bind (prof->server_socket, (struct sockaddr *) &server_address, sizeof (server_address)) < 0) {
2351 fprintf (stderr, "Cannot bind server socket, port: %d: %s\n", prof->command_port, strerror (errno));
2352 close (prof->server_socket);
2355 if (listen (prof->server_socket, 1) < 0) {
2356 fprintf (stderr, "Cannot listen server socket\n");
2357 close (prof->server_socket);
2360 slen = sizeof (server_address);
2361 if (getsockname (prof->server_socket, (struct sockaddr *)&server_address, &slen) == 0) {
2362 prof->command_port = ntohs (server_address.sin_port);
2363 /*fprintf (stderr, "Assigned server port: %d\n", prof->command_port);*/
2366 r = pthread_create (&prof->helper_thread, NULL, helper_thread, prof);
2368 close (prof->server_socket);
2375 static MonoProfiler*
2376 create_profiler (const char *filename)
2380 int force_delete = 0;
2381 int need_helper_thread = 0;
2382 prof = calloc (1, sizeof (MonoProfiler));
2384 prof->command_port = command_port;
2385 if (filename && *filename == '-') {
2391 filename = "|mprof-report -";
2393 filename = "output.mlpd";
2394 nf = (char*)filename;
2396 nf = new_filename (filename);
2398 int s = strlen (nf) + 32;
2399 char *p = malloc (s);
2400 snprintf (p, s, "|mprof-report '--out=%s' -", nf);
2406 prof->file = popen (nf + 1, "w");
2407 prof->pipe_output = 1;
2408 } else if (*nf == '#') {
2409 int fd = strtol (nf + 1, NULL, 10);
2410 prof->file = fdopen (fd, "a");
2415 if ((f = fopen (nf, "r"))) {
2417 fprintf (stderr, "The Mono profiler won't overwrite existing filename: %s.\n", nf);
2418 fprintf (stderr, "Profiling disabled: use a different name or -FILENAME to force overwrite.\n");
2422 prof->file = fopen (nf, "wb");
2425 fprintf (stderr, "Cannot create profiler output: %s\n", nf);
2428 #if defined (HAVE_SYS_ZLIB)
2430 prof->gzfile = gzdopen (fileno (prof->file), "wb");
2433 if (sample_type && !do_mono_sample)
2434 need_helper_thread = setup_perf_event ();
2436 /* FIXME: warn if different freq or sample type */
2440 if (do_mono_sample) {
2441 prof->stat_buffers = create_stat_buffer ();
2442 need_helper_thread = 1;
2444 if (do_counters && !need_helper_thread) {
2445 need_helper_thread = 1;
2447 #ifndef DISABLE_HELPER_THREAD
2448 if (hs_mode_ondemand || need_helper_thread) {
2449 if (!start_helper_thread (prof))
2450 prof->command_port = 0;
2453 if (hs_mode_ondemand)
2454 fprintf (stderr, "Ondemand heapshot unavailable on this arch.\n");
2456 prof->startup_time = current_time ();
2464 printf ("Log profiler version %d.%d (format: %d)\n", LOG_VERSION_MAJOR, LOG_VERSION_MINOR, LOG_DATA_VERSION);
2465 printf ("Usage: mono --profile=log[:OPTION1[,OPTION2...]] program.exe\n");
2466 printf ("Options:\n");
2467 printf ("\thelp show this usage info\n");
2468 printf ("\t[no]alloc enable/disable recording allocation info\n");
2469 printf ("\t[no]calls enable/disable recording enter/leave method events\n");
2470 printf ("\theapshot[=MODE] record heap shot info (by default at each major collection)\n");
2471 printf ("\t MODE: every XXms milliseconds, every YYgc collections, ondemand\n");
2472 printf ("\tcounters sample counters every 1s\n");
2473 printf ("\tsample[=TYPE] use statistical sampling mode (by default cycles/1000)\n");
2474 printf ("\t TYPE: cycles,instr,cacherefs,cachemiss,branches,branchmiss\n");
2475 printf ("\t TYPE can be followed by /FREQUENCY\n");
2476 printf ("\ttime=fast use a faster (but more inaccurate) timer\n");
2477 printf ("\tmaxframes=NUM collect up to NUM stack frames\n");
2478 printf ("\tcalldepth=NUM ignore method events for call chain depth bigger than NUM\n");
2479 printf ("\toutput=FILENAME write the data to file FILENAME (-FILENAME to overwrite)\n");
2480 printf ("\toutput=|PROGRAM write the data to the stdin of PROGRAM\n");
2481 printf ("\t %%t is subtituted with date and time, %%p with the pid\n");
2482 printf ("\treport create a report instead of writing the raw data to a file\n");
2483 printf ("\tzip compress the output data\n");
2484 printf ("\tport=PORTNUM use PORTNUM for the listening command server\n");
2490 match_option (const char* p, const char *opt, char **rval)
2492 int len = strlen (opt);
2493 if (strncmp (p, opt, len) == 0) {
2495 if (p [len] == '=' && p [len + 1]) {
2496 const char *opt = p + len + 1;
2497 const char *end = strchr (opt, ',');
2505 val = malloc (l + 1);
2506 memcpy (val, opt, l);
2511 if (p [len] == 0 || p [len] == ',') {
2513 return p + len + (p [len] == ',');
2531 static const SampleMode sample_modes [] = {
2532 {"cycles", SAMPLE_CYCLES},
2533 {"instr", SAMPLE_INSTRUCTIONS},
2534 {"cachemiss", SAMPLE_CACHE_MISSES},
2535 {"cacherefs", SAMPLE_CACHE_REFS},
2536 {"branches", SAMPLE_BRANCHES},
2537 {"branchmiss", SAMPLE_BRANCH_MISSES},
2542 set_sample_mode (char* val, int allow_empty)
2545 char *maybe_freq = NULL;
2547 const SampleMode *smode = sample_modes;
2548 #ifndef USE_PERF_EVENTS
2551 if (allow_empty && !val) {
2552 sample_type = SAMPLE_CYCLES;
2556 if (strcmp (val, "mono") == 0) {
2558 sample_type = SAMPLE_CYCLES;
2562 for (smode = sample_modes; smode->name; smode++) {
2563 int l = strlen (smode->name);
2564 if (strncmp (val, smode->name, l) == 0) {
2565 sample_type = smode->sample_mode;
2566 maybe_freq = val + l;
2572 if (*maybe_freq == '/') {
2573 count = strtoul (maybe_freq + 1, &end, 10);
2574 if (maybe_freq + 1 == end)
2576 sample_freq = count;
2577 } else if (*maybe_freq != 0) {
2586 set_hsmode (char* val, int allow_empty)
2590 if (allow_empty && !val)
2592 if (strcmp (val, "ondemand") == 0) {
2593 hs_mode_ondemand = 1;
2597 count = strtoul (val, &end, 10);
2600 if (strcmp (end, "ms") == 0)
2602 else if (strcmp (end, "gc") == 0)
2610 * declaration to silence the compiler: this is the entry point that
2611 * mono will load from the shared library and call.
2614 mono_profiler_startup (const char *desc);
2617 mono_profiler_startup_log (const char *desc);
2620 * this is the entry point that will be used when the profiler
2621 * is embedded inside the main executable.
2624 mono_profiler_startup_log (const char *desc)
2626 mono_profiler_startup (desc);
2630 mono_profiler_startup (const char *desc)
2633 char *filename = NULL;
2637 int calls_enabled = 0;
2638 int allocs_enabled = 0;
2639 int events = MONO_PROFILE_GC|MONO_PROFILE_ALLOCATIONS|
2640 MONO_PROFILE_GC_MOVES|MONO_PROFILE_CLASS_EVENTS|MONO_PROFILE_THREADS|
2641 MONO_PROFILE_ENTER_LEAVE|MONO_PROFILE_JIT_COMPILATION|MONO_PROFILE_EXCEPTIONS|
2642 MONO_PROFILE_MONITOR_EVENTS|MONO_PROFILE_MODULE_EVENTS|MONO_PROFILE_GC_ROOTS;
2645 if (strncmp (p, "log", 3))
2650 for (; *p; p = opt) {
2656 if ((opt = match_option (p, "help", NULL)) != p) {
2660 if ((opt = match_option (p, "calls", NULL)) != p) {
2664 if ((opt = match_option (p, "nocalls", NULL)) != p) {
2665 events &= ~MONO_PROFILE_ENTER_LEAVE;
2669 if ((opt = match_option (p, "alloc", NULL)) != p) {
2673 if ((opt = match_option (p, "noalloc", NULL)) != p) {
2674 events &= ~MONO_PROFILE_ALLOCATIONS;
2677 if ((opt = match_option (p, "time", &val)) != p) {
2678 if (strcmp (val, "fast") == 0)
2680 else if (strcmp (val, "null") == 0)
2687 if ((opt = match_option (p, "report", NULL)) != p) {
2691 if ((opt = match_option (p, "debug", NULL)) != p) {
2695 if ((opt = match_option (p, "heapshot", &val)) != p) {
2696 events &= ~MONO_PROFILE_ALLOCATIONS;
2697 events &= ~MONO_PROFILE_ENTER_LEAVE;
2700 set_hsmode (val, 1);
2703 if ((opt = match_option (p, "sample", &val)) != p) {
2704 events &= ~MONO_PROFILE_ALLOCATIONS;
2705 events &= ~MONO_PROFILE_ENTER_LEAVE;
2707 set_sample_mode (val, 1);
2710 if ((opt = match_option (p, "hsmode", &val)) != p) {
2711 fprintf (stderr, "The hsmode profiler option is obsolete, use heapshot=MODE.\n");
2712 set_hsmode (val, 0);
2715 if ((opt = match_option (p, "zip", NULL)) != p) {
2719 if ((opt = match_option (p, "output", &val)) != p) {
2723 if ((opt = match_option (p, "port", &val)) != p) {
2725 command_port = strtoul (val, &end, 10);
2729 if ((opt = match_option (p, "maxframes", &val)) != p) {
2731 num_frames = strtoul (val, &end, 10);
2732 if (num_frames > MAX_FRAMES)
2733 num_frames = MAX_FRAMES;
2735 notraces = num_frames == 0;
2738 if ((opt = match_option (p, "calldepth", &val)) != p) {
2740 max_call_depth = strtoul (val, &end, 10);
2744 if ((opt = match_option (p, "counters", NULL)) != p) {
2753 if (calls_enabled) {
2754 events |= MONO_PROFILE_ENTER_LEAVE;
2758 events |= MONO_PROFILE_ALLOCATIONS;
2759 utils_init (fast_time);
2761 prof = create_profiler (filename);
2766 mono_profiler_install (prof, log_shutdown);
2767 mono_profiler_install_gc (gc_event, gc_resize);
2768 mono_profiler_install_allocation (gc_alloc);
2769 mono_profiler_install_gc_moves (gc_moves);
2770 mono_profiler_install_gc_roots (gc_handle, gc_roots);
2771 mono_profiler_install_class (NULL, class_loaded, NULL, NULL);
2772 mono_profiler_install_module (NULL, image_loaded, NULL, NULL);
2773 mono_profiler_install_thread (thread_start, thread_end);
2774 mono_profiler_install_thread_name (thread_name);
2775 mono_profiler_install_enter_leave (method_enter, method_leave);
2776 mono_profiler_install_jit_end (method_jitted);
2777 mono_profiler_install_exception (throw_exc, method_exc_leave, clause_exc);
2778 mono_profiler_install_monitor (monitor_event);
2779 mono_profiler_install_runtime_initialized (runtime_initialized);
2782 if (do_mono_sample && sample_type == SAMPLE_CYCLES) {
2783 events |= MONO_PROFILE_STATISTICAL;
2784 mono_profiler_install_statistical (mono_sample_hit);
2787 mono_profiler_set_events (events);
2789 TLS_INIT (tlsbuffer);