1 ==== Summary ==== 2 3 android.frameworks.sensorservice (a] 1.0 is a package that mimics the sensors API in 4 NDK (sensor.h). It includes a subset of these APIs and introduces a few 5 adaptations. 6 7 === Design Details === 8 9 - ISensorManager 10 ISensorMangaer includes member functions that adapts the ASensorManager_* 11 series in NDK. An instance of ISensorManager must be able to 12 - retrieve sensors 13 - create direct report channel 14 - create event queue 15 16 - IDirectReportChannel 17 IDirectReportChannel corresponds to a channel ID, an integer obtained in 18 ASensorManager_createSharedMemoryDirectChannel and 19 ASensorManager_createHardwareBufferDirectChannel. An instance of 20 IDirectReportChannel must also destroy it against the sensor manager. An 21 IDirectReportChannel must be able to configure itself (an adaptation to 22 ASensorManager_configureDirectReport). The implementation must also call 23 ASensorManager_destroyEventQueue on destruction of IDirectReportChannel. 24 25 Usage typically looks like this (transaction errors are not handled): 26 27 sp<ISensorManager> manager = ISensorManager::getService(); 28 int32_t sensorHandle; 29 manager->getDefaultSensor(SensorType::GYROSCOPE, 30 [&sensorHandle] (const auto &info) { 31 sensorHandle = info.sensorHandle; 32 }); 33 hidl_memory mem; 34 const uint64_t size = 4096; 35 ::android::hidl::memory::V1_0::IAllocator::getService()->allocate(size, 36 [&](auto, const auto &m) { mem = m; }); 37 if (!mem.handle()) { 38 /* error handling */ 39 } 40 sp<IDirectChannel> chan; 41 Result res; 42 manager->createAshmemDirectChannel(mem, size, 43 [&chan, &res] (const auto &c, auto r) { 44 chan = c; res = r; 45 }); 46 if (res != Result::OK) { /* error handling */ } 47 chan->configure(sensorHandle, RateLevel::FAST, [&](auto token, auto result) { 48 if (result != Result::OK) { 49 /* error handling */ 50 } 51 }); 52 53 /* obtain sensor events from shared memory */ 54 55 chan->configure(sensorHandle, RateLevel::STOP, [&](auto token, auto result) { 56 if (result != Result::OK) { 57 /* error handling */ 58 } 59 }); 60 61 /* 62 * Free the channel. 63 * kernel calls decStrong() on server side implementation of IDirectChannel, 64 * hence resources are freed as well. 65 */ 66 chan = nullptr; 67 68 - IEventQueue, IEventQueueCallback 69 IEventQueue includes member functions that adapts some of the 70 ASensorEventQueue_* seeries in NDK. An instance of IEventQueue must be able to 71 - enable selected sensors (adapts ASensorEventQueue_registerSensor) 72 - disable selected sensors (adapts ASensorEventQueue_disableSensor) 73 74 The implementation must free all resources related to this IEventQueue instance 75 and call ASensorManager_destroyEventQueue on destruction of IEventQueue. 76 77 Unlike NDK ASensorEventQueue_hasEvents and ASensorEventQueue_getEvents, which 78 implies a poll model for sensor events, IEventQueue uses a push model by using 79 callbacks. When creating an event queue, client must provide an instance of 80 IEventQueueCallback. The implementation of IEventQueue must either use a global 81 looper or create a new looper to call on ASensorManager_createEventQueue. The 82 server implementation must use this looper to constantly poll for events, then 83 invoke the callback when any event is fired. 84 85 IEventQueueCallback.onEvent is designed to be oneway, because the server should 86 not wait for the client to finish handling the event. The callback 87 should finish in a predictably short time, and should not block or run for an 88 extended period of time. The callbacks can be invoked in a very high frequency; 89 a long running callback means delay in handling of subsequent events and filling 90 up the (kernel binder buffer) memory space of the client process, eventually the 91 server sees a transaction error when issuing the callback. It is up to the 92 client to be configured single-threaded or multi-threaded to handle these 93 callbacks. 94 - Single-threaded clients receive events in the correct order in the same 95 thread. 96 - Multi-threaded clients receive events in the correct order but in 97 different threads; it is the clients' responsibility to deal with 98 concurrency issues and handle events in the expected order to avoid race 99 conditions. 100 101 Usage typically looks like this (transaction errors are not handled): 102 103 struct Callback : IEventQueueCallback { 104 Return<void> onEvent(const Event &e) { 105 /* handle sensor event e */ 106 } 107 }; 108 sp<ISensorManager> manager = ISensorManager::getService(); 109 int32_t sensorHandle; 110 manager->getDefaultSensor(SensorType::GYROSCOPE, 111 [&sensorHandle] (const auto &info) { 112 sensorHandle = info.sensorHandle; 113 }); 114 sp<IEventQueue> queue; 115 Result res; 116 manager->createEventQueue(new Callback(), 117 [&queue, &res] (const auto &q, auto r) { 118 queue = q; res = r; 119 }); 120 /* Server side implementation of IEventQueue holds a strong reference to 121 * the callback. */ 122 if (res != Result::OK) { /* error handling */ } 123 124 if (q->enableSensor(sensorHandle, 125 20000 /* sample period */, 0 /* latency */) != Result::OK) { 126 /* error handling */ 127 } 128 129 /* start receiving events via onEvent */ 130 131 if (q->disableSensor(sensorHandle) != Result::OK) { 132 /* error handling */ 133 } 134 135 /* 136 * Free the event queue. 137 * kernel calls decStrong() on server side implementation of IEventQueue, 138 * hence resources (including the callback) are freed as well. 139 */ 140 queue = nullptr; 141