diff --git a/core/client.js b/core/client.js
index a62b6f35..fa3d2845 100644
--- a/core/client.js
+++ b/core/client.js
@@ -137,7 +137,7 @@ function showFiles() {
 }
 
 function trace() {
-	window.open(`/speedscope/#profileURL=${encodeURIComponent('/trace')}&title=Tilde%20Friends`);
+	window.open(`/speedscope/#profileURL=${encodeURIComponent('/trace')}`);
 }
 
 function stats() {
diff --git a/src/ssb.c b/src/ssb.c
index 859818a6..eea00c48 100644
--- a/src/ssb.c
+++ b/src/ssb.c
@@ -2149,6 +2149,7 @@ sqlite3* tf_ssb_acquire_db_reader(tf_ssb_t* ssb)
 	{
 		sqlite3_open_v2(ssb->db_path, &db, SQLITE_OPEN_READONLY, NULL);
 		tf_ssb_db_init_reader(db);
+		tf_trace_sqlite(ssb->trace, db);
 	}
 	uv_mutex_unlock(&ssb->db_readers_lock);
 	return db;
diff --git a/src/trace.c b/src/trace.c
index 34bd83f0..3ccb835f 100644
--- a/src/trace.c
+++ b/src/trace.c
@@ -1,3 +1,7 @@
+#define _GNU_SOURCE
+#include <pthread.h>
+#undef _GNU_SOURCE
+
 #include "trace.h"
 
 #include "mem.h"
@@ -9,7 +13,9 @@
 #include <time.h>
 #include <unistd.h>
 
+#include <pthread.h>
 #include <sqlite3.h>
+#include <uv.h>
 
 #if !defined(_countof)
 #define _countof(a) ((int)(sizeof((a)) / sizeof(*(a))))
@@ -29,20 +35,31 @@ typedef struct _tf_trace_stack_t
 	tf_trace_stack_t* next;
 } tf_trace_stack_t;
 
+typedef struct _tf_trace_thread_t
+{
+	pthread_t id;
+	tf_trace_stack_t* stack;
+	char name[64];
+} tf_trace_thread_t;
+
 typedef struct _tf_trace_t
 {
 	char buffer[k_buffer_size];
 	char process_name[256];
 	int write_offset;
 
+	uv_mutex_t mutex;
 	tf_trace_write_callback_t* callback;
 	void* user_data;
 
-	tf_trace_stack_t* stack;
+	uv_rwlock_t threads_lock;
+	tf_trace_thread_t** threads;
+	int threads_count;
 } tf_trace_t;
 
 static void _trace_append(tf_trace_t* trace, const char* buffer, size_t size, void* user_data)
 {
+	uv_mutex_lock(&trace->mutex);
 	if (trace->write_offset + size + 2 >= k_buffer_size)
 	{
 		trace->buffer[trace->write_offset] = '\0';
@@ -55,6 +72,7 @@ static void _trace_append(tf_trace_t* trace, const char* buffer, size_t size, vo
 		trace->write_offset += size;
 		trace->buffer[trace->write_offset++] = '\n';
 	}
+	uv_mutex_unlock(&trace->mutex);
 }
 
 tf_trace_t* tf_trace_create()
@@ -62,17 +80,27 @@ tf_trace_t* tf_trace_create()
 	tf_trace_t* trace = tf_malloc(sizeof(tf_trace_t));
 	memset(trace, 0, sizeof(*trace));
 	trace->callback = _trace_append;
+	uv_mutex_init(&trace->mutex);
+	uv_rwlock_init(&trace->threads_lock);
 	return trace;
 }
 
 void tf_trace_destroy(tf_trace_t* trace)
 {
-	while (trace->stack)
+	for (int i = 0; i < trace->threads_count; i++)
 	{
-		tf_trace_stack_t* stack = trace->stack;
-		trace->stack = stack->next;
-		tf_free(stack);
+		tf_trace_thread_t* thread = trace->threads[i];
+		while (thread->stack)
+		{
+			tf_trace_stack_t* stack = thread->stack;
+			thread->stack = stack->next;
+			tf_free(stack);
+		}
+		tf_free(thread);
 	}
+	tf_free(trace->threads);
+	uv_rwlock_destroy(&trace->threads_lock);
+	uv_mutex_destroy(&trace->mutex);
 	tf_free(trace);
 }
 
@@ -152,6 +180,88 @@ static int _tf_trace_escape_name(char* out, size_t out_size, const char* name)
 	return p;
 }
 
+static tf_trace_thread_t* _tf_trace_get_thread(tf_trace_t* trace, pthread_t self)
+{
+	tf_trace_thread_t* found = NULL;
+	uv_rwlock_rdlock(&trace->threads_lock);
+	for (int i = 0; i < trace->threads_count; i++)
+	{
+		if (trace->threads[i]->id == self)
+		{
+			found = trace->threads[i];
+			break;
+		}
+	}
+	uv_rwlock_rdunlock(&trace->threads_lock);
+
+	if (!found)
+	{
+		uv_rwlock_wrlock(&trace->threads_lock);
+		/* Maybe it was added while we changed from rd to rw. */
+		for (int i = 0; i < trace->threads_count; i++)
+		{
+			if (trace->threads[i]->id == self)
+			{
+				found = trace->threads[i];
+				break;
+			}
+		}
+
+		if (!found)
+		{
+			found = tf_malloc(sizeof(tf_trace_thread_t));
+			*found = (tf_trace_thread_t)
+			{
+				.id = self,
+			};
+#if defined(__linux__)
+			pthread_getname_np(self, found->name, sizeof(found->name));
+#endif
+			trace->threads = tf_resize_vec(trace->threads, sizeof(tf_trace_thread_t*) * (trace->threads_count + 1));
+			trace->threads[trace->threads_count++] = found;
+		}
+		uv_rwlock_wrunlock(&trace->threads_lock);
+	}
+
+	return found;
+}
+
+static void _tf_push_stack(tf_trace_t* trace, pthread_t self, const char* name)
+{
+	tf_trace_thread_t* thread = _tf_trace_get_thread(trace, self);
+	if (!thread->stack || thread->stack->count + 1 > (int)_countof(thread->stack->names))
+	{
+		tf_trace_stack_t* stack = tf_malloc(sizeof(tf_trace_stack_t));
+		memset(stack, 0, sizeof(*stack));
+		stack->next = thread->stack;
+		thread->stack = stack;
+	}
+	tf_trace_stack_t* stack = thread->stack;
+	while (stack->count == 0 && stack->next && stack->next->count + 1 <= (int)_countof(thread->stack->names))
+	{
+		stack = stack->next;
+	}
+	stack->names[stack->count++] = name;
+
+}
+
+static const char* _tf_pop_stack(tf_trace_t* trace, pthread_t self)
+{
+	tf_trace_thread_t* thread = _tf_trace_get_thread(trace, self);
+	tf_trace_stack_t* stack = thread->stack;
+	while (stack && stack->count == 0)
+	{
+		stack = stack->next;
+	}
+	const char* name = NULL;
+	if (stack && stack->count > 0)
+	{
+		name = stack->names[stack->count - 1];
+		stack->count--;
+	}
+	return name;
+}
+
 void tf_trace_begin(tf_trace_t* trace, const char* name)
 {
 	if (!trace)
@@ -159,22 +269,11 @@ void tf_trace_begin(tf_trace_t* trace, const char* name)
 		return;
 	}
 
-	if (!trace->stack || trace->stack->count + 1 > _countof(trace->stack->names))
-	{
-		tf_trace_stack_t* stack = tf_malloc(sizeof(tf_trace_stack_t));
-		memset(stack, 0, sizeof(*stack));
-		stack->next = trace->stack;
-		trace->stack = stack;
-	}
-	tf_trace_stack_t* stack = trace->stack;
-	while (stack->count == 0 && stack->next && stack->next->count + 1 <= _countof(trace->stack->names))
-	{
-		stack = stack->next;
-	}
-	stack->names[stack->count++] = name;
+	pthread_t self = pthread_self();
+	_tf_push_stack(trace, self, name);
 
 	char line[1024];
-	int p = snprintf(line, sizeof(line), "{\"ph\": \"B\", \"pid\": %d, \"tid\": %" PRId64 ", \"ts\": %" PRId64 ", \"name\": \"", getpid(), (int64_t)pthread_self(), _trace_ts());
+	int p = snprintf(line, sizeof(line), "{\"ph\": \"B\", \"pid\": %d, \"tid\": \"0x%" PRIx64 "\", \"ts\": %" PRId64 ", \"name\": \"", getpid(), (int64_t)self, _trace_ts());
 	p += _tf_trace_escape_name(line + p, sizeof(line) - p, name);
 	p += snprintf(line + p, sizeof(line) - p, "\"},");
 	trace->callback(trace, line, p, trace->user_data);
@@ -187,24 +286,15 @@ void tf_trace_end(tf_trace_t* trace)
 		return;
 	}
 
-	tf_trace_stack_t* stack = trace->stack;
-	while (stack && stack->count == 0)
-	{
-		stack = stack->next;
-	}
-	const char* name = NULL;
-	if (stack && stack->count > 0)
-	{
-		name = stack->names[stack->count - 1];
-		stack->count--;
-	}
+	pthread_t self = pthread_self();
+	const char* name = _tf_pop_stack(trace, self);
 	if (!name)
 	{
 		return;
 	}
 
 	char line[1024];
-	int p = snprintf(line, sizeof(line), "{\"ph\": \"E\", \"pid\": %d, \"tid\": %" PRId64 ", \"ts\": %" PRId64 ", \"name\": \"", getpid(), (int64_t)pthread_self(), _trace_ts());
+	int p = snprintf(line, sizeof(line), "{\"ph\": \"E\", \"pid\": %d, \"tid\": \"0x%" PRIx64 "\", \"ts\": %" PRId64 ", \"name\": \"", getpid(), (int64_t)pthread_self(), _trace_ts());
 	p += _tf_trace_escape_name(line + p, sizeof(line) - p, name);
 	p += snprintf(line + p, sizeof(line) - p,  "\"},");
 	trace->callback(trace, line, p, trace->user_data);
@@ -219,6 +309,7 @@ char* tf_trace_export(tf_trace_t* trace)
 
 	static const int k_extra_size = 1024;
 	char* buffer = tf_malloc(k_buffer_size + k_extra_size);
+	uv_mutex_lock(&trace->mutex);
 	const char* newline = strchr(trace->buffer + trace->write_offset, '\n');
 	int begin = newline ? newline - trace->buffer : 0;
 	size_t size = 0;
@@ -226,10 +317,21 @@ char* tf_trace_export(tf_trace_t* trace)
 	if (*trace->process_name)
 	{
 		size += snprintf(buffer + size, k_buffer_size - size,
-			"{\"name\":\"process_name\",\"ph\":\"M\",\"pid\":%d,\"args\":{\"name\":\"%s\"}},",
+			"{\"ph\":\"M\",\"pid\":%d,\"name\":\"process_name\",\"args\":{\"name\":\"%s\"}},\n",
 			getpid(),
 			trace->process_name);
 	}
+	uv_rwlock_rdlock(&trace->threads_lock);
+	for (int i = 0; i < trace->threads_count; i++)
+	{
+		tf_trace_thread_t* thread = trace->threads[i];
+		size += snprintf(buffer + size, k_buffer_size - size,
+			"{\"ph\":\"M\",\"pid\":%d,\"tid\":\"0x%" PRIx64 "\",\"name\":\"thread_name\",\"args\":{\"name\":\"%s\"}},\n",
+			getpid(),
+			thread->id,
+			thread->name);
+	}
+	uv_rwlock_rdunlock(&trace->threads_lock);
 	if (begin)
 	{
 		size_t this_size = strlen(trace->buffer + begin);
@@ -238,14 +340,15 @@ char* tf_trace_export(tf_trace_t* trace)
 	}
 	memcpy(buffer + size, trace->buffer, trace->write_offset);
 	size += trace->write_offset;
+	uv_mutex_unlock(&trace->mutex);
 	if (size > 2 && buffer[size - 1] == '\n' && buffer[size - 2] == ',')
 	{
 		buffer[size - 2] = '\n';
 		size--;
 	}
 	size += snprintf(buffer + size, k_buffer_size - size, "]}\n");
-	buffer[size] = '\0';
 	assert(size < (size_t)k_buffer_size + k_extra_size);
+	buffer[size] = '\0';
 	return buffer;
 }