Monthly Archives: January 2016

Creating a Windows Universal app to talk Bluetooth LE, save to SQLite and expose a REST service

The Goal

I’ve had a couple of TI SensorTags sitting on my shelf for a couple of years. These are the original ones, which have been superseded by smaller ones that have additional sensors for light and sound.

Sensor Tag with no caseSensor Tag with case

They are wonderful devices. They last for over a year on a watch battery, they talk Bluetooth LE, and they have loads of sensors including Temperature (both spot temperature of a nearby object, and overall ambient temperature), Gyroscope, Accelerometer, Magnetometer, Barometer, Humidity, etc.

Last, but not least, they cost less than US$30. Unless you actually enjoy wiring physical sensors into an Arduino or Raspberry PI, I think Sensor Tags are a great way to start collecting all kinds of information.

Rather than have a phone sitting talking Bluetooth LE, I decided I wanted to use a Mac Mini server that I have running Windows, which I could run continuously to capture, store and serve the sensor information.

My goal was to:

  • Create a Windows Universal App that talks Bluetooth LE to the Sensor Tag
  • Save the captured information to an SQLite database
  • Serve the captured information using REST (/GetTemperatures?start=201501010000&end=201701010000)

At each step I hit roadblocks, and the purpose of this blog post is to try to capture what I did to overcome them, in the hope that other people may benefit from my pain.

Although I’ve been mainly writing Java/Android, C, TypeScript and JavaScript over the last three years, I still retain a soft spot for C# and the associated tooling of Visual Studio and Resharper.

I really appreciate the C# syntax and associated features such as lambdas, and LINQ.

I wanted to try my hand at create a Windows app, to see how much I’d lost over the last few years.

Bluetooth LE, SensorTag and Windows Universal

I started off creating a new Windows Universal app in Visual Studio. I browsed the documentation, and found the classes associated with using Bluetooth LE. I liked the fact that my app would be able to run on desktops down to phones.

My initial code:

      _watcher = new BluetoothLEAdvertisementWatcher();
      _watcher.Received += BluetoothReceived;
      _watcher.Stopped += BluetoothStopped;

When I ran this, I got the following exception: onecoreuap\drivers\wdm\bluetooth\user\winrt\common\devicehandle.cpp(100)\ Windows.Devices.Bluetooth.dll!51D26D1B: (caller: 51D273AE) Exception(1) tid(13c0) 80070005 Access is denied.

Turns out I needed to add Bluetooth to my app’s capabilities by double-clicking the Package.appxmanifest file in the Solution Explorer, going to Capabilities and checking Bluetooth.

Enabling Bluetooth in Windows Universal App

Once that was done, I was able to look for the SensorTag’s Service UUID, and then check for the correct characteristics and enable the reception of the sensor’s data:

    const string BaseUuidStart = "f000";
    const string BaseUuidEnd = "-0451-4000-b000-000000000000";

    const string TempData = "aa01";
    const string TempConfig = "aa02";
    const string AccelData = "aa11";
    const string AccelConfig = "aa12";
    const string HumidData = "aa21";
    const string HumidConfig = "aa22";
    const string MagnetData = "aa31";
    const string MagnetConfig = "aa32";
    const string BaromData = "aa41";
    const string BaromConfig = "aa42";
    const string GyroData = "aa51";
    const string GyroConfig = "aa52";

    private bool _attaching;
    private readonly List<BluetoothLEDevice> _devices = new List<BluetoothLEDevice>();
    private readonly List<GattCharacteristic> _characteristics = new List<GattCharacteristic>();

    private async void BluetoothReceived(BluetoothLEAdvertisementWatcher sender,
      BluetoothLEAdvertisementReceivedEventArgs args) {
      if (_attaching) return;
      try {
        var device = await BluetoothLEDevice.FromBluetoothAddressAsync(args.BluetoothAddress);
        _attaching = true;
        device.ConnectionStatusChanged += DeviceConnectionStatusChanged;
        device.GattServicesChanged += DeviceGattServicesChanged;
        foreach (var service in device.GattServices) {
          var serviceUuid = service.Uuid.ToString().ToLowerInvariant();
          if (!serviceUuid.StartsWith(BaseUuidStart) || !serviceUuid.EndsWith(BaseUuidEnd)) {
          foreach (var characteristic in service.GetAllCharacteristics()) {
            var characteristicUuid = characteristic.Uuid.ToString().ToLowerInvariant();
            if (_characteristics.Any(c => c.Uuid.ToString() == characteristicUuid)) {
            var characteristicType = characteristicUuid.Substring(BaseUuidStart.Length, 4);
            switch (characteristicType) {
              case AccelData:
              case BaromData:
              case HumidData:
              case GyroData:
              case MagnetData:
              case TempData: {
                characteristic.ValueChanged += CharacteristicChanged;
                var status =
                  await characteristic.WriteClientCharacteristicConfigurationDescriptorAsync(
                Debug.WriteLine("Subscribed .... with status " + status);
              case AccelConfig:
              case BaromConfig:
              case HumidConfig:
              case GyroConfig:
              case MagnetConfig:
              case TempConfig: {
                var status = await characteristic.WriteValueAsync(new byte[] {1}.AsBuffer());

                Debug.WriteLine("Ignoring characteristic: " + characteristicType);
      catch (Exception ex) {
        Debug.WriteLine("got " + ex);

I used the Sensor Tag documentation to know about the GUIDs used for the services and characteristics.
I found I needed to press Advertise the button on the side of my Sensor Tag to get it to be seen.

Capturing the values was pretty easy, but I did hit one stumbling block which was the temperature. There is an algorithm described in the documentation as to how to transform the series of bytes received into the spot and ambient temperature in degrees Celsius. When I tried using it I got garbage values, but eventually found this C# example showing how they can be calculated:

    private async Task ProcessTempData(string bluetoothId, byte[] rawData) {
      // Extract ambiant temperature 
      var ambTemp = BitConverter.ToUInt16(rawData, 2)/128.0;

      // Extract object temperature 
      int twoByteValue = BitConverter.ToInt16(rawData, 0);
      var vobj2 = twoByteValue*0.00000015625;
      var tdie = ambTemp + 273.15;
      const double s0 = 5.593E-14; // Calibration factor 
      const double a1 = 1.75E-3;
      const double a2 = -1.678E-5;
      const double b0 = -2.94E-5;
      const double b1 = -5.7E-7;
      const double b2 = 4.63E-9;
      const double c2 = 13.4;
      const double tref = 298.15;
      var s = s0*(1 + a1*(tdie - tref) + a2*Math.Pow(tdie - tref, 2));
      var vos = b0 + b1*(tdie - tref) + b2*Math.Pow(tdie - tref, 2);
      var fObj = vobj2 - vos + c2*Math.Pow(vobj2 - vos, 2);
      var tObj = Math.Pow(Math.Pow(tdie, 4) + (fObj/s), .25);
      var objTemp = tObj - 273.15;

      await SaveTemperature(bluetoothId, ambTemp, objTemp);

SQLite and Windows Universal

Installing SQLite for Windows was pretty easy, but I couldn’t find clear, complete instructions. In short I used NuGet to install

  • SQLite.Net-PCL
  • SQLite.Net.Async-PCL
  • SQLite.Net.Core-PCL

Once I had this installed, I could define classes corresponding to the tables I wanted to create, such as:

  public class Temperature
    public int Id { get; set; }
    public DateTime Timestamp { get; set; }
    public string BluetoothId { get; set; }
    public double Ambient { get; set; }
    public double Spot { get; set; }

Then I could initialize the database:

    private SQLiteAsyncConnection _asyncConnection;
    private async Task InitializeDatabase() {
      Debug.WriteLine("Initializing database");
      var databasePath = Path.Combine(Windows.Storage.ApplicationData.Current.LocalFolder.Path, "sensortag.db");
      var connectionFactory = new Func<SQLiteConnectionWithLock>(() => new SQLiteConnectionWithLock(new SQLitePlatformWinRT(), new SQLiteConnectionString(databasePath, true)));
      _asyncConnection = new SQLiteAsyncConnection(connectionFactory);
      await _asyncConnection.CreateTablesAsync(typeof (Temperature));
      Debug.WriteLine("Initialized database");

And then write the data:

    private async Task SaveTemperature(string bluetoothId, double ambTemp, double objTemp) {
      var temperature = new Temperature {
        Timestamp = DateTime.Now,
        BluetoothId = bluetoothId,
        Ambient = ambTemp,
        Spot = objTemp
      Debug.WriteLine("Writing temperature");
      await _asyncConnection.InsertAsync(temperature);
      Debug.WriteLine("Wrote temperature");

It turns out this was wrong, though it is what was shown in the Stack Overflow posts I found. The reason that it is wrong is that it is creating a new database connection each time the factory lambda is invoked. When I used this code all would run fine for a while, until eventually I hit an SQLite Busy exception:

Exception thrown: &#039;SQLite.Net.SQLiteException&#039; in
SQLite.Net.SQLiteException: Busy
   at SQLite.Net.PreparedSqlLiteInsertCommand.ExecuteNonQuery(Object[] source)
   at SQLite.Net.SQLiteConnection.Insert(Object obj, String extra, Type objType)
   at SQLite.Net.SQLiteConnection.Insert(Object obj)
   at SQLite.Net.Async.SQLiteAsyncConnection.<>c__DisplayClass14_0.<InsertAsync>b__0()
   at System.Threading.Tasks.Task`1.InnerInvoke()
   at System.Threading.Tasks.Task.Execute()

The simple solution was to create a single database connection instance, and serve that, rather than continually serving new ones:

    private SQLiteAsyncConnection _asyncConnection;
    private SQLiteConnectionWithLock _sqliteConnectionWithLock;
    private async Task InitializeDatabase() {
      Debug.WriteLine("Initializing database");
      var databasePath = Path.Combine(Windows.Storage.ApplicationData.Current.LocalFolder.Path, "sensortag.db");
      _sqliteConnectionWithLock = new SQLiteConnectionWithLock(new SQLitePlatformWinRT(), new SQLiteConnectionString(databasePath, true));
      var connectionFactory = new Func<SQLiteConnectionWithLock>(() => _sqliteConnectionWithLock);
      _asyncConnection = new SQLiteAsyncConnection(connectionFactory);
      await _asyncConnection.CreateTablesAsync(typeof (Temperature));
      Debug.WriteLine("Initialized database");

Exposing a REST Service from Windows Universal

This was supposed to be trivially easy. I’ve done plenty of WCF in the past, and know how ridiculously straightforward it should be to expose a REST service from an app. Except that Windows Universal doesn’t currently support WCF.

I went searching and found Restup, currently in Beta, which aims to expose REST endpoints for Windows Universal apps.

I used NuGet to install it. I had to check the Include prerelease option because it was currently in beta.

Setting up was pretty easy:

    private async Task InitializeWebServer() {
      await InitializeDatabase();
      var webserver = new RestWebServer(); //defaults to 8800

      await webserver.StartServerAsync();
  class SensorTagService {
    private readonly SQLiteAsyncConnection _connection;

    public SensorTagService(SQLiteAsyncConnection sqLiteAsyncConnection) {
      _connection = sqLiteAsyncConnection;

    public async Task<GetResponse> GetTemperatures(string start, string end) {
      Debug.WriteLine("got temp request");

Note the escaping of the question mark in the UriFormat? I wanted to pass parameters to my endpoint, rather than use values that are part of the path, but all the RestUP examples showed values in the path. I eventually came up with this solution, however it may be unnecessary by the time you read this.

Once again the security model bit me, and I got the following exception:

An exception of type &#039;System.UnauthorizedAccessException&#039; occurred in but was not handled in user code
WinRT information: At least one of either InternetClientServer or PrivateNetworkClientServer capabilities is required to listen for or receive traffic
Additional information: Access is denied.

Once again I edited the app’s capabilities by double-clicking the Package.appxmanifest file in the Solution Explorer, going to Capabilities and checking

  • Internet (Client),
  • Internet (Client & Server) and
  • Private Networks (Client & Server) (so that I could use my service on my home network).

Accessing a local Windows Universal app from your web browser

Try as I might, I was not able to use my local Chrome browser to access my service. I resorted to using a totally separate machine to invoke my service. I used the CheckNetIsolation tool. I ensure that the Allow Network Loopback option was set for my project Visual Studio. I turned off my firewalls. Nothing!


The Bluetooth side of things was quite easy, but exposing a REST API was far too hard, despite the sterling work of Tom Kuijsten and the Restup project. Not being able to access my service locally was a complete pain – the Windows Universal restrictions on being able to be accessed from the local host seem strange – almost as though they are trying to stop you from building traditional apps that talk to Windows Universal apps …

In the end I’ll likely use the Windows Universal app to capture the SensorTag data via Bluetooth LE, and then create a Node.JS app to serve it over REST, sharing the same SQLite database, with code to handle retrying if the database is busy when inserting new values.

I’ll also push the data to a Node-RED instance to act on the data.

Radical surgery: Slimming Pebble apps down to run on Aplite

A long way to go

In December 2015, when first I released Powernoter, an unofficial Evernote client for the Pebble Watch, I initially targeted Pebble Time (codename Basalt), and Pebble Time Round (codename Chalk).

After all there was already the official Evernote Pebble app (which I also created) for the original Pebble (codename Aplite).

Then Pebble released a firmware update and SDK for the original Pebble which meant that I could easily release Powernoter for the original Pebble too, using the same SDK I’d already used.

This is the build log from the first time I built Powernoter targeting Aplite (the original Pebble), Basalt (Pebble Time) and Chalk (Pebble Time Round):

Total size of resources:        26461 bytes / 256KB
Total footprint in RAM:         25895 bytes / 64KB
Free RAM available (heap):      39641 bytes
Total size of resources:        26461 bytes / 256KB
Total footprint in RAM:         25943 bytes / 64KB
Free RAM available (heap):      39593 bytes
Total size of resources:        26341 bytes / 125KB
Total footprint in RAM:         23789 bytes / 24KB
Free RAM available (heap):      787 bytes

See the 787 bytes on the last line? That was how much free memory my app had before it even started running on an original Pebble. Before it created its first window or allocated memory to receive and send messages.

Although I successfully built Powernoter for Aplite, it couldn’t even start up, crashing immediately as it ran out of memory.

Not so verbose with the error messages

The first thing I did, was to run the pebble analyze-size command, which gave me a sense of where the memory was being used.

Like all good programmers, I very carefully and very consistently checked all OS calls for out of memory situations, and logged (very) verbose messages if I ran out of memory. Like this:

  bitmap_layer = bitmap_layer_create(image_layer_size);
  if(!bitmap_layer) {
    APP_LOG(APP_LOG_LEVEL_ERROR, "Couldn't allocate memory for the image");

All those strings had to be allocated somewhere. I went through my app and removed all those lovely descriptive messages. Instead I just logged the line number – that was enough to work out where it went wrong.

  bitmap_layer = bitmap_layer_create(image_layer_size);
  if(!bitmap_layer) {

I defined a couple of macros for Out Of Memory (OOM) situations:

#define OOM(s) log_oom(__FILE_NAME__, __LINE__, (int)s)
#define OOMCF() log_create_failed(__FILE_NAME__, __LINE__)
void log_create_failed(char* file, int line) {
  app_log(APP_LOG_LEVEL_DEBUG, file, line, "create failed %d free", (int)heap_bytes_free());

void log_oom(char* file, int line, int size) {
  app_log(APP_LOG_LEVEL_DEBUG, file, line, "oom %d, %d", size, (int)heap_bytes_free());

I also declared some handy logging macros, so that debug log strings were stripped out of shipping builds

#define LOG_MEM_START()
#define LOG_MEM_END()
#define LOG_FUNC_START(name)
#define LOG_FUNC_END(name)
#define LOG_DBG(fmt, args...)
#define LOG_ERR(fmt, args...) app_log(APP_LOG_LEVEL_ERROR, __FILE_NAME__, __LINE__, " ")
#define LOG_DBG(fmt, args...) app_log(APP_LOG_LEVEL_DEBUG, __FILE_NAME__, __LINE__, fmt, ## args)
#define LOG_MEM_START() app_log(APP_LOG_LEVEL_DEBUG, __FILE_NAME__, __LINE__, "start %d", (int)heap_bytes_free())
#define LOG_MEM_END() app_log(APP_LOG_LEVEL_DEBUG, __FILE_NAME__, __LINE__, "end %d", (int)heap_bytes_free())
#define LOG_FUNC_START(name) app_log(APP_LOG_LEVEL_DEBUG, __FILE_NAME__, __LINE__, "%s invoked", name)
#define LOG_FUNC_END(name) app_log(APP_LOG_LEVEL_DEBUG, __FILE_NAME__, __LINE__, "%s returning", name)
#define LOG_ERR(fmt, args...) app_log(APP_LOG_LEVEL_ERROR, __FILE_NAME__, __LINE__, fmt, ## args)

Use statics in moderation

Next I looked into how I was defining static variables. I like statics because they are only visible to the file in which they are declared: a primitive form of encapsulation. A typical C source file might have started with:

static CustomMenu* customMenu;
static CustomMenuItem* items;
static uint16_t itemCount;
static AppTimer *send_timeout_timer;
static NoteSelectedCallback noteSelectedCallback;

The types don’t matter (CustomMenu is my own class that does things like automatically scrolling long menu items).

What matters is that I have four pointers and a short declared as statics, meaning I have a whole chunk of memory statically allocated just for this one file.

Powernoter is not a small app … this multiplied by tens of files means that I had a load of memory statically allocated, which was never used unless the user was actually invoking the functionality represented by those files.

The solution was to move to a dynamically allocated memory:

typedef struct NoteList {
  CustomMenu *customMenu;
  CustomMenuItem *items;
  uint16_t itemCount;
  AppTimer *send_timeout_timer;
  NoteSelectedCallback noteSelectedCallback;
} NoteList;

I only allocate a NoteList when it is being used, and free it as soon as possible.

Omit needless code

Although the SDK includes definitions for things like DictationSession on Aplite, so that code can be compiled regardless of the platform (you do need to check return calls though), it made no sense to include that code at all. I #ifdefed whole chunks of code to reduce the app size:

static void dictation_session_callback(DictationSession *session, DictationSessionStatus status,
                                       char *transcription, void *context) {
  if(DictationSessionStatusSuccess == status) {
    if(!noteContext->waitingAnimation) {
      if(noteContext->customMenu) {
        layer_set_hidden(custom_menu_get_layer(noteContext->customMenu), true);

SUPPORTS_VOICE is my own macro:


Pebble have added a PBL_MICROPHONE macro so my use of SUPPORTS_VOICE is no longer necessary.

I did the same thing for animations and color support.

Although I think I am a decent enough software engineer, I am under few illusions as to my abilities as a designer, which is why I let you choose your very own foreground and background colors in Powernoter, except if you are running on an original Pebble, in which case all that code, including the color names, is #ifdefed out.

Be careful what you ask for (when calling app_message_xyz_maximum)

Once upon a time were were limited to 120 or so bytes per message sent between the watch and the phone. I wrote inordinately complex code to page menu items in dynamically from the phone to the watch so that you could scroll through infinitely long menus. Then Pebble gave us what we wanted, with massive (8Kish) message buffers.

When you only have a little memory free to start with, the last thing you want to do is go allocating 8K buffers. It won’t work.

My code to determine the size of the input buffer looks like this now:

#define MAX_INBOX_SIZE 512
#define MAX_INBOX_SIZE 4096

The LOW_MEMORY_DEVICE macro is set on Aplite only. Users on the original Pebble won’t see an enormous number of notes listed, or a lot of a note’s content, but at least they’ll see something.

Make long strings into Resources

There is an excellent Internationalization sample for the Pebble. Although Powernoter isn’t internationalized, there are no strings hardcoded in code … all strings are accessed via a single point. I include the strings in a single source file in the app, except for certain very long strings, such as the About page. These I load as resources from files:

static char* loadResource(uint32_t resourceId) {
  ResHandle handle = resource_get_handle(resourceId);
  size_t res_size = resource_size(handle);

  // Copy to buffer
  char* result = (char*)malloc(res_size + 1);
  if(!result) {
    result = (char*)malloc(1);
    if(result) {
      *result = '\0';
    return result;
  resource_load(handle, (uint8_t*)result, res_size);
  result[res_size] = '\0';
  return result;

Once I’m done with them, I free them as quickly as possible.


In case you were wondering, this is how things look right now:

Total size of resources:        27313 bytes / 256KB
Total footprint in RAM:         24244 bytes / 64KB
Free RAM available (heap):      41292 bytes
Total size of resources:        27313 bytes / 256KB
Total footprint in RAM:         24176 bytes / 64KB
Free RAM available (heap):      41360 bytes
Total size of resources:        13966 bytes / 125KB
Total footprint in RAM:         17353 bytes / 24KB
Free RAM available (heap):      7223 bytes

Getting from 787 bytes free to 7,223 bytes free, so that Powernoter can really run on Aplite involved many changes, some which I’d say were generally good practice (reducing statics and instead using structs which are allocated/freed), and some less so (removing error log messages).

In general I don’t think the code looks too unreadable as a result of supporting Aplite … certainly I’d prefer not to have as many #ifdefs sprinkled throughout my code as I have, but it’s not that bad.

You may also wish to check out this Pebble presentation on Pebble app memory usage.

One thing is for sure, the changes I had to make to Powernoter to get it to run on Aplite are nothing compared with the miracles the Pebble team pulled to get the original Pebble to support the same SDK as Pebble Time and Pebble Time Round.

About me

I’m an independent consultant and speaker, available for ad-hoc Pebble, Android and Android Wear and Tizen consulting and development.

If you like and use Powernoter, please consider supporting it.

On the other hand if something is missing or doesn’t work, check out this Trello board where you can comment to request enhancements or report bugs.