#include <mono/metadata/threads.h>
#include <mono/metadata/mono-gc.h>
#include <mono/metadata/debug-helpers.h>
+#include <mono/utils/atomic.h>
+#include <mono/utils/mono-membar.h>
+#include <mono/utils/mono-counters.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#endif
/* the architecture needs a memory fence */
-#if defined(__linux__) && (defined(__i386__) || defined(__x86_64__))
+#if defined(__linux__) && (defined(__i386__) || defined(__x86_64__) || defined(__arm__))
+#include <unistd.h>
+#include <sys/syscall.h>
#include "perf_event.h"
#define USE_PERF_EVENTS 1
-static int read_perf_mmap (MonoProfiler* prof);
+static int read_perf_mmap (MonoProfiler* prof, int cpu);
#endif
#define BUFFER_SIZE (4096 * 16)
static int do_mono_sample = 0;
static int in_shutdown = 0;
static int do_debug = 0;
+static int do_counters = 0;
/* For linux compile with:
* gcc -fPIC -shared -o libmono-profiler-log.so proflog.c utils.c -Wall -g -lz `pkg-config --cflags --libs mono-2`
*
* type sample format
* type: TYPE_SAMPLE
- * exinfo: one of TYPE_SAMPLE_HIT, TYPE_SAMPLE_USYM, TYPE_SAMPLE_UBIN
+ * exinfo: one of TYPE_SAMPLE_HIT, TYPE_SAMPLE_USYM, TYPE_SAMPLE_UBIN, TYPE_SAMPLE_COUNTERS_DESC, TYPE_SAMPLE_COUNTERS
* if exinfo == TYPE_SAMPLE_HIT
* [sample_type: uleb128] type of sample (SAMPLE_*)
* [timestamp: uleb128] nanoseconds since startup (note: different from other timestamps!)
* [offset: uleb128] file offset of mapping (the same file can be mapped multiple times)
* [size: uleb128] memory size
* [name: string] binary name
+ * if exinfo == TYPE_SAMPLE_COUNTERS_DESC
+ * [len: uleb128] number of counters
+ * for i = 0 to len
+ * [section: uleb128] section name of counter
+ * [name: string] name of counter
+ * [type: uleb128] type name of counter
+ * [unit: uleb128] unit name of counter
+ * [variance: uleb128] variance name of counter
+ * [index: uleb128] unique index of counter
+ * if exinfo == TYPE_SAMPLE_COUNTERS
+ * [timestamp: uleb128] sampling timestamp
+ * while true:
+ * [index: uleb128] unique index of counter
+ * if index == 0:
+ * break
+ * [size: uleb128] size of counter value
+ * [value: string/uleb128/sleb128/double] counter value, can be sleb128, uleb128, string or double (determined by using counter type)
*
*/
struct _LogBuffer {
static pthread_key_t tlsbuffer;
#endif
+static void safe_dump (MonoProfiler *profiler, LogBuffer *logbuffer);
+
static char*
pstrdup (const char *s)
{
assert (logbuffer->data <= logbuffer->data_end);
}
+static void
+emit_string (LogBuffer *logbuffer, const char *str, size_t size)
+{
+ size_t i = 0;
+ if (str) {
+ for (; i < size; i++) {
+ emit_byte (logbuffer, str [i]);
+ if (str[i] == '\0')
+ break;
+ }
+ }
+ if (!str || i == size)
+ emit_byte (logbuffer, '\0');
+}
+
+static void
+emit_double (LogBuffer *logbuffer, double value)
+{
+ int i;
+ unsigned char buffer[8];
+ memcpy (buffer, &value, 8);
+#if G_BYTE_ORDER == G_BIG_ENDIAN
+ for (i = 7; i >= 0; i--)
+#else
+ for (i = 0; i < 8; i++)
+#endif
+ emit_byte (logbuffer, buffer[i]);
+}
+
static char*
write_int16 (char *buf, int32_t value)
{
mono_gc_collect (mono_gc_max_generation ());
}
+static void counters_init (MonoProfiler *profiler);
+
static void
runtime_initialized (MonoProfiler *profiler)
{
runtime_inited = 1;
+#ifndef DISABLE_HELPER_THREAD
+ counters_init (profiler);
+#endif
+ /* ensure the main thread data and startup are available soon */
+ safe_dump (profiler, ensure_logbuf (0));
}
/*
EXIT_LOG (logbuffer);
}
-#ifndef HOST_WIN32
-#include "mono/io-layer/atomic.h"
-#endif
-#define cmp_exchange InterlockedCompareExchangePointer
-/*#else
-static void*
-cmp_exchange (volatile void **dest, void *exch, void *comp)
-{
- void *old;
- __asm__ __volatile__ ("lock; "
-#ifdef __x86_64__
- "cmpxchgq"
-#else
- "cmpxchgl"
-#endif
- " %2, %0"
- : "=m" (*dest), "=a" (old)
- : "r" (exch), "m" (*dest), "a" (comp));
- return old;
-}
-#endif
-*/
-
static void
mono_sample_hit (MonoProfiler *profiler, unsigned char *ip, void *context)
{
do {
old_data = sbuf->data;
new_data = old_data + 4;
- data = cmp_exchange ((volatile void**)&sbuf->data, new_data, old_data);
+ data = InterlockedCompareExchangePointer ((volatile void**)&sbuf->data, new_data, old_data);
} while (data != old_data);
if (old_data >= sbuf->data_end)
return; /* lost event */
logbuffer = ensure_logbuf (20 + count * 8);
emit_byte (logbuffer, TYPE_SAMPLE | TYPE_SAMPLE_HIT);
emit_value (logbuffer, type);
- emit_uvalue (logbuffer, (prof->startup_time + sample [2]) * 10000);
+ emit_uvalue (logbuffer, prof->startup_time + (uint64_t)sample [2] * (uint64_t)10000);
emit_value (logbuffer, count);
for (i = 0; i < count; ++i) {
emit_ptr (logbuffer, (void*)sample [i + 3]);
#if USE_PERF_EVENTS
#ifndef __NR_perf_event_open
+#ifdef __arm__
+#define __NR_perf_event_open 364
+#else
#define __NR_perf_event_open 241
#endif
+#endif
-static int perf_fd = -1;
-static void *mmap_base;
-static struct perf_event_mmap_page *page_desc = NULL;
-static int num_pages = 64;
+static int
+mono_cpu_count (void)
+{
+ int count = 0;
+#ifdef PLATFORM_ANDROID
+ /* Android tries really hard to save power by powering off CPUs on SMP phones which
+ * means the normal way to query cpu count returns a wrong value with userspace API.
+ * Instead we use /sys entries to query the actual hardware CPU count.
+ */
+ char buffer[8] = {'\0'};
+ int present = open ("/sys/devices/system/cpu/present", O_RDONLY);
+ /* Format of the /sys entry is a cpulist of indexes which in the case
+ * of present is always of the form "0-(n-1)" when there is more than
+ * 1 core, n being the number of CPU cores in the system. Otherwise
+ * the value is simply 0
+ */
+ if (present != -1 && read (present, (char*)buffer, sizeof (buffer)) > 3)
+ count = strtol (((char*)buffer) + 2, NULL, 10);
+ if (present != -1)
+ close (present);
+ if (count > 0)
+ return count + 1;
+#endif
+#ifdef _SC_NPROCESSORS_ONLN
+ count = sysconf (_SC_NPROCESSORS_ONLN);
+ if (count > 0)
+ return count;
+#endif
+#ifdef USE_SYSCTL
+ {
+ int mib [2];
+ size_t len = sizeof (int);
+ mib [0] = CTL_HW;
+ mib [1] = HW_NCPU;
+ if (sysctl (mib, 2, &count, &len, NULL, 0) == 0)
+ return count;
+ }
+#endif
+#ifdef HOST_WIN32
+ {
+ SYSTEM_INFO info;
+ GetSystemInfo (&info);
+ return info.dwNumberOfProcessors;
+ }
+#endif
+ /* FIXME: warn */
+ return 1;
+}
+
+typedef struct {
+ int perf_fd;
+ unsigned int prev_pos;
+ void *mmap_base;
+ struct perf_event_mmap_page *page_desc;
+} PerfData ;
+
+static PerfData *perf_data = NULL;
+static int num_perf;
+#define PERF_PAGES_SHIFT 4
+static int num_pages = 1 << PERF_PAGES_SHIFT;
static unsigned int mmap_mask;
-static unsigned int prev_pos = 0;
typedef struct {
struct perf_event_header h;
return syscall(/*__NR_perf_event_open*/ 298, attr, pid, cpu, group_fd, flags);
#elif defined(__i386__)
return syscall(/*__NR_perf_event_open*/ 336, attr, pid, cpu, group_fd, flags);
+#elif defined(__arm__)
+ return syscall(/*__NR_perf_event_open*/ 364, attr, pid, cpu, group_fd, flags);
#else
return -1;
#endif
}
static int
-setup_perf_map (void)
+setup_perf_map (PerfData *perf)
{
- mmap_mask = num_pages * getpagesize () - 1;
- mmap_base = mmap (NULL, (num_pages + 1) * getpagesize (), PROT_READ|PROT_WRITE, MAP_SHARED, perf_fd, 0);
- if (mmap_base == MAP_FAILED) {
+ perf->mmap_base = mmap (NULL, (num_pages + 1) * getpagesize (), PROT_READ|PROT_WRITE, MAP_SHARED, perf->perf_fd, 0);
+ if (perf->mmap_base == MAP_FAILED) {
if (do_debug)
printf ("failed mmap\n");
return 0;
}
- page_desc = mmap_base;
+ perf->page_desc = perf->mmap_base;
if (do_debug)
- printf ("mmap version: %d\n", page_desc->version);
+ printf ("mmap version: %d\n", perf->page_desc->version);
return 1;
}
/* read events from the ring buffer */
static int
-read_perf_mmap (MonoProfiler* prof)
+read_perf_mmap (MonoProfiler* prof, int cpu)
{
+ PerfData *perf = perf_data + cpu;
unsigned char *buf;
- unsigned char *data = (unsigned char*)mmap_base + getpagesize ();
- unsigned int head = page_desc->data_head;
+ unsigned char *data = (unsigned char*)perf->mmap_base + getpagesize ();
+ unsigned int head = perf->page_desc->data_head;
int diff, size;
unsigned int old;
-#if defined(__i386__)
- asm volatile("lock; addl $0,0(%%esp)":::"memory");
-#elif defined (__x86_64__)
- asm volatile("lfence":::"memory");
-#endif
+ mono_memory_read_barrier ();
- old = prev_pos;
+ old = perf->prev_pos;
diff = head - old;
if (diff < 0) {
if (do_debug)
printf ("found bytes of events: %d\n", size);
dump_perf_hits (prof, buf, size);
old += size;
- prev_pos = old;
- page_desc->data_tail = old;
+ perf->prev_pos = old;
+ perf->page_desc->data_tail = old;
return 0;
}
static int
-setup_perf_event (void)
+setup_perf_event_for_cpu (PerfData *perf, int cpu)
{
struct perf_event_attr attr;
memset (&attr, 0, sizeof (attr));
attr.freq = 1;
attr.sample_freq = sample_freq;
- perf_fd = perf_event_syscall (&attr, getpid (), -1, -1, 0);
+ perf->perf_fd = perf_event_syscall (&attr, getpid (), cpu, -1, 0);
if (do_debug)
- printf ("perf fd: %d, freq: %d, event: %llu\n", perf_fd, sample_freq, attr.config);
- if (perf_fd < 0) {
- if (perf_fd == -EPERM) {
+ printf ("perf fd: %d, freq: %d, event: %llu\n", perf->perf_fd, sample_freq, attr.config);
+ if (perf->perf_fd < 0) {
+ if (perf->perf_fd == -EPERM) {
fprintf (stderr, "Perf syscall denied, do \"echo 1 > /proc/sys/kernel/perf_event_paranoid\" as root to enable.\n");
} else {
if (do_debug)
}
return 0;
}
- if (!setup_perf_map ()) {
- close (perf_fd);
- perf_fd = -1;
+ if (!setup_perf_map (perf)) {
+ close (perf->perf_fd);
+ perf->perf_fd = -1;
return 0;
}
return 1;
}
+static int
+setup_perf_event (void)
+{
+ int i, count = 0;
+ mmap_mask = num_pages * getpagesize () - 1;
+ num_perf = mono_cpu_count ();
+ perf_data = calloc (num_perf, sizeof (PerfData));
+ for (i = 0; i < num_perf; ++i) {
+ count += setup_perf_event_for_cpu (perf_data + i, i);
+ }
+ if (count)
+ return 1;
+ free (perf_data);
+ perf_data = NULL;
+ return 0;
+}
+
#endif /* USE_PERF_EVENTS */
+#ifndef DISABLE_HELPER_THREAD
+
+typedef struct MonoCounterAgent {
+ MonoCounter *counter;
+ // MonoCounterAgent specific data :
+ void *value;
+ size_t value_size;
+ short index;
+ struct MonoCounterAgent *next;
+} MonoCounterAgent;
+
+static MonoCounterAgent* counters;
+static gboolean counters_initialized = FALSE;
+static int counters_index = 1;
+
+static mono_bool
+counters_init_add_counter (MonoCounter *counter, gpointer data)
+{
+ MonoCounterAgent *agent, *item;
+
+ for (agent = counters; agent; agent = agent->next) {
+ if (agent->counter == counter)
+ return TRUE;
+ }
+
+ agent = malloc (sizeof (MonoCounterAgent));
+ agent->counter = counter;
+ agent->value = NULL;
+ agent->value_size = 0;
+ agent->index = counters_index++;
+ agent->next = NULL;
+
+ if (!counters) {
+ counters = agent;
+ } else {
+ item = counters;
+ while (item->next)
+ item = item->next;
+ item->next = agent;
+ }
+
+ return TRUE;
+}
+
+static void
+counters_init (MonoProfiler *profiler)
+{
+ mono_counters_foreach (counters_init_add_counter, NULL);
+
+ MonoCounterAgent *agent;
+ LogBuffer *logbuffer;
+ int size = 1 + 5, len = 0;
+
+ for (agent = counters; agent; agent = agent->next) {
+ size += strlen (mono_counter_get_name (agent->counter)) + 1 + 5 * 5;
+ len += 1;
+ }
+
+ logbuffer = ensure_logbuf (size);
+
+ ENTER_LOG (logbuffer, "counters");
+ emit_byte (logbuffer, TYPE_SAMPLE_COUNTERS_DESC | TYPE_SAMPLE);
+ emit_value (logbuffer, len);
+ for (agent = counters; agent; agent = agent->next) {
+ const char *name = mono_counter_get_name (agent->counter);
+ emit_value (logbuffer, mono_counter_get_section (agent->counter));
+ emit_string (logbuffer, name, strlen (name) + 1);
+ emit_value (logbuffer, mono_counter_get_type (agent->counter));
+ emit_value (logbuffer, mono_counter_get_unit (agent->counter));
+ emit_value (logbuffer, mono_counter_get_variance (agent->counter));
+ emit_value (logbuffer, agent->index);
+ }
+ EXIT_LOG (logbuffer);
+
+ counters_initialized = TRUE;
+}
+
+static void
+counters_sample (MonoProfiler *profiler, uint64_t timestamp)
+{
+ MonoCounterAgent *agent;
+ MonoCounter *counter;
+ LogBuffer *logbuffer;
+ int type;
+ int buffer_size;
+ void *buffer;
+ int size;
+
+ if (!counters_initialized)
+ return;
+
+ buffer_size = 8;
+ buffer = calloc (1, buffer_size);
+
+ size = 1 + 10 + 5;
+ for (agent = counters; agent; agent = agent->next)
+ size += 10 * 2 + mono_counter_get_size (agent->counter);
+
+ logbuffer = ensure_logbuf (size);
+
+ ENTER_LOG (logbuffer, "counters");
+ emit_byte (logbuffer, TYPE_SAMPLE_COUNTERS | TYPE_SAMPLE);
+ emit_uvalue (logbuffer, timestamp);
+ for (agent = counters; agent; agent = agent->next) {
+ counter = agent->counter;
+
+ size_t size = mono_counter_get_size (counter);
+ if (size < 0) {
+ continue; // FIXME error
+ } else if (size > buffer_size) {
+ buffer_size = size;
+ buffer = realloc (buffer, buffer_size);
+ }
+
+ memset (buffer, 0, buffer_size);
+
+ if (mono_counters_sample (counter, buffer, size) < 0)
+ continue; // FIXME error
+
+ type = mono_counter_get_type (counter);
+
+ if (!agent->value) {
+ agent->value = calloc (1, size);
+ agent->value_size = size;
+ } else {
+ if (type == MONO_COUNTER_STRING) {
+ if (strncmp (agent->value, buffer, size) == 0)
+ continue;
+ } else {
+ if (agent->value_size == size && memcmp (agent->value, buffer, size) == 0)
+ continue;
+ }
+ }
+
+ emit_uvalue (logbuffer, agent->index);
+ emit_uvalue (logbuffer, type);
+ switch (type) {
+ case MONO_COUNTER_INT:
+#if SIZEOF_VOID_P == 4
+ case MONO_COUNTER_WORD:
+#endif
+ emit_svalue (logbuffer, *(int*)buffer - *(int*)agent->value);
+ break;
+ case MONO_COUNTER_UINT:
+ emit_uvalue (logbuffer, *(guint*)buffer - *(guint*)agent->value);
+ break;
+ case MONO_COUNTER_TIME_INTERVAL:
+ case MONO_COUNTER_LONG:
+#if SIZEOF_VOID_P == 8
+ case MONO_COUNTER_WORD:
+#endif
+ emit_svalue (logbuffer, *(gint64*)buffer - *(gint64*)agent->value);
+ break;
+ case MONO_COUNTER_ULONG:
+ emit_uvalue (logbuffer, *(guint64*)buffer - *(guint64*)agent->value);
+ break;
+ case MONO_COUNTER_DOUBLE:
+ emit_double (logbuffer, *(double*)buffer);
+ break;
+ case MONO_COUNTER_STRING:
+ if (size == 0)
+ emit_string (logbuffer, "(null)", 7);
+ else
+ emit_string (logbuffer, (char*)buffer, size);
+ break;
+ default:
+ assert (0);
+ }
+
+ if (type == MONO_COUNTER_STRING && size > agent->value_size) {
+ agent->value = realloc (agent->value, size);
+ agent->value_size = size;
+ }
+
+ if (size > 0)
+ memcpy (agent->value, buffer, size);
+ }
+ free (buffer);
+
+ emit_value (logbuffer, 0);
+ EXIT_LOG (logbuffer);
+}
+
+#endif /* DISABLE_HELPER_THREAD */
+
static void
log_shutdown (MonoProfiler *prof)
{
}
#endif
#if USE_PERF_EVENTS
- if (page_desc)
- read_perf_mmap (prof);
+ if (perf_data) {
+ int i;
+ for (i = 0; i < num_perf; ++i)
+ read_perf_mmap (prof, i);
+ }
#endif
dump_sample_hits (prof, prof->stat_buffers, 1);
take_lock ();
int len;
char buf [64];
MonoThread *thread = NULL;
+ GTimeVal start, now;
+
+ g_get_current_time (&start);
//fprintf (stderr, "Server listening\n");
command_socket = -1;
max_fd = command_socket;
}
#if USE_PERF_EVENTS
- if (perf_fd >= 0) {
- FD_SET (perf_fd, &rfds);
- if (max_fd < perf_fd)
- max_fd = perf_fd;
+ if (perf_data) {
+ int i;
+ for ( i = 0; i < num_perf; ++i) {
+ if (perf_data [i].perf_fd < 0)
+ continue;
+ FD_SET (perf_data [i].perf_fd, &rfds);
+ if (max_fd < perf_data [i].perf_fd)
+ max_fd = perf_data [i].perf_fd;
+ }
}
#endif
+ g_get_current_time (&now);
+ counters_sample (prof, (uint64_t)(now.tv_sec * 1000 + now.tv_usec / 1000) - (uint64_t)(start.tv_sec * 1000 + start.tv_usec / 1000));
+ safe_dump (prof, ensure_logbuf (0));
+
tv.tv_sec = 1;
tv.tv_usec = 0;
len = select (max_fd + 1, &rfds, NULL, NULL, &tv);
if (do_debug)
fprintf (stderr, "helper shutdown\n");
#if USE_PERF_EVENTS
- if (perf_fd >= 0)
- read_perf_mmap (prof);
+ if (perf_data) {
+ int i;
+ for ( i = 0; i < num_perf; ++i) {
+ if (perf_data [i].perf_fd < 0)
+ continue;
+ if (FD_ISSET (perf_data [i].perf_fd, &rfds))
+ read_perf_mmap (prof, i);
+ }
+ }
#endif
safe_dump (prof, ensure_logbuf (0));
return NULL;
}
#if USE_PERF_EVENTS
- if (perf_fd >= 0 && FD_ISSET (perf_fd, &rfds)) {
- read_perf_mmap (prof);
- safe_dump (prof, ensure_logbuf (0));
+ if (perf_data) {
+ int i;
+ for ( i = 0; i < num_perf; ++i) {
+ if (perf_data [i].perf_fd < 0)
+ continue;
+ if (FD_ISSET (perf_data [i].perf_fd, &rfds)) {
+ read_perf_mmap (prof, i);
+ safe_dump (prof, ensure_logbuf (0));
+ }
+ }
}
#endif
if (command_socket >= 0 && FD_ISSET (command_socket, &rfds)) {
#if USE_PERF_EVENTS
if (sample_type && !do_mono_sample)
need_helper_thread = setup_perf_event ();
- if (perf_fd < 0) {
+ if (!perf_data) {
/* FIXME: warn if different freq or sample type */
do_mono_sample = 1;
}
prof->stat_buffers = create_stat_buffer ();
need_helper_thread = 1;
}
+ if (do_counters && !need_helper_thread) {
+ need_helper_thread = 1;
+ }
#ifndef DISABLE_HELPER_THREAD
if (hs_mode_ondemand || need_helper_thread) {
if (!start_helper_thread (prof))
printf ("\t[no]calls enable/disable recording enter/leave method events\n");
printf ("\theapshot[=MODE] record heap shot info (by default at each major collection)\n");
printf ("\t MODE: every XXms milliseconds, every YYgc collections, ondemand\n");
+ printf ("\tcounters sample counters every 1s\n");
printf ("\tsample[=TYPE] use statistical sampling mode (by default cycles/1000)\n");
printf ("\t TYPE: cycles,instr,cacherefs,cachemiss,branches,branchmiss\n");
printf ("\t TYPE can be followed by /FREQUENCY\n");
free (val);
continue;
}
+ if ((opt = match_option (p, "counters", NULL)) != p) {
+ do_counters = 1;
+ continue;
+ }
if (opt == p) {
usage (0);
exit (0);
projects / mono.git / blobdiff