Merge branch 'paramwatcher' into watcher

This commit is contained in:
James Vecellio-Grant
2025-12-07 17:13:06 -08:00
committed by GitHub
3 changed files with 48 additions and 20 deletions
+25 -2
View File
@@ -35,7 +35,7 @@ The Linux implementation interacts directly with the kernel's `inotify` subsyste
* **Library Loading**: `libc = ctypes.CDLL('libc.so.6')` loads the standard C library to access system calls.
* **Initialization**: `inotify_init()` is called to create a new inotify instance, returning a file descriptor.
* **Watch Setup**: `inotify_add_watch(fd, path, mask)` registers the parameters directory. The mask includes `IN_MODIFY | IN_CREATE | IN_DELETE | IN_MOVED_TO | IN_CLOSE_WRITE` to capture all relevant file changes.
* **Watch Setup**: `inotify_add_watch(fd, path, mask)` registers the parameters directory. The mask includes `IN_MODIFY | IN_CREATE | IN_DELETE | IN_MOVED_TO | IN_CLOSE_WRITE` (Linux Kernel Organization, 2005) to capture all relevant file changes.
* **Event Loop**:
* **Polling**: `select.epoll()` is used to efficiently wait for activity on the file descriptor without busy-waiting.
* **Reading**: When events occur, `os.read(fd, 1024)` retrieves the raw binary event data.
@@ -48,6 +48,7 @@ The macOS implementation uses the `FSEvents` API from the `CoreServices` framewo
* **Framework Loading**: `ctypes.cdll.LoadLibrary` loads `CoreServices` and `CoreFoundation`.
* **Callback Definition**: `CFUNCTYPE` is used to define a C-compatible callback function. This function is invoked by the OS whenever a change occurs in the watched directory.
* **Stream Creation**: `FSEventStreamCreate` creates a stream for the target directory. The `kFSEventStreamCreateFlagFileEvents` flag is used to request file-level granularity where available.
* **Event Filtering**: The callback filters events using flags such as `kFSEventStreamEventFlagItemCreated` and `kFSEventStreamEventFlagItemModified` to ensure only relevant file changes trigger updates (Apple Inc., n.d.-c).
* **Scheduling**: `FSEventStreamScheduleWithRunLoop` attaches the stream to the current thread's run loop (Apple Inc., n.d.-b).
* **Execution**: `CFRunLoopRun()` starts the event loop. This passes control to the OS, which wakes the thread only when necessary.
* **Handling**: Inside the callback, the code iterates through the changed paths provided by the OS. It extracts the filename and calls `_trigger_callbacks` to invalidate the cache for that specific parameter.
@@ -81,14 +82,34 @@ While `ParamWatcher` offers superior performance for UI rendering, it presents s
* **Event Latency**: In high-load scenarios, `inotify` events may experience slight delays or coalescing compared to direct reads. However, for user interface applications, this latency (<10ms) is imperceptible.
* **Complexity**: The solution (the process singleton approach) requires managing a background thread and OS-specific event loops, increasing code complexity compared to the synchronous `Params::get` function.
## Alternative Architecture Considered: ZeroMQ Service (ZMQ)
During the development of `ParamWatcher`, a Client-Server architecture using ZMQ was evaluated. In this architecture, a single background service process would monitor file system events and publish changes over a ZMQ PUB socket to multiple client processes (SUB).
### Trade-off Analysis
| Metric | In-Process (Current) | ZMQ Service (Rejected) |
| :--- | :--- |:------------------------------------------------------|
| **Memory Usage** | Low (1 thread/process) | High (1 full Python process + ZMQ buffers per client) |
| **CPU Usage** | Low (Direct callback) | High (Serialization + TCP Stack + Deserialization) |
| **Latency** | Instant (<0.1ms) | Variable (TCP Loopback overhead) |
| **Scalability** | Limited by OS file handles | Limited by TCP ports/buffers |
| **Robustness** | Process-isolated failure | Single point of failure (Service crash affects all) |
### Decision Rationale
While the ZMQ approach offers better isolation and reduces the total number of OS file watchers (1 vs N), the overhead of inter-process communication (IPC) proved excessive for this use case.
* **Efficiency**: Even with 50+ processes, the memory footprint of 50 simple threads is significantly lower than the overhead of a dedicated Python service process plus the ZMQ context in every client.
* **Complexity**: The ZMQ architecture introduced synchronization challenges (e.g., service startup race conditions, "Address already in use" errors) that outweighed its benefits.
* **Performance**: The latency of serializing messages and passing them through the TCP stack is orders of magnitude higher than a direct function call within the same process memory space.
## Conclusion
Replacing polling mechanisms with event-driven caching shifts the computational load from kernel space (syscalls) to user space (RAM). This transition eliminates I/O overhead and UI stutters caused by garbage collection, resulting in a more responsive user experience.
Replacing polling mechanisms with event-driven caching shifts the computational load from kernel space (syscalls) to user space (RAM). This transition eliminates I/O overhead and UI stutters caused by garbage collection, resulting in a more responsive user experience. The In-Process Singleton approach was selected as the optimal balance between performance, complexity, and resource efficiency.
## References
Apple Inc. (n.d.-a). *File System Events*. Retrieved from https://developer.apple.com/documentation/coreservices/file_system_events
Apple Inc. (n.d.-b). *CFRunLoop*. Retrieved from https://developer.apple.com/documentation/corefoundation/cfrunloop
Apple Inc. (n.d.-c). *FSEventStreamEventFlags*. Retrieved from https://developer.apple.com/documentation/coreservices/1455361-fseventstreameventflags
Codezup. (2025). *Efficient File I/O in C++*. Retrieved from https://codezup.com/efficient-file-io-cpp-best-practices/
cppreference.com. (n.d.-a). *std::basic_ifstream*. Retrieved from https://en.cppreference.com/w/cpp/io/basic_ifstream
@@ -101,6 +122,8 @@ Linux man-pages. (2025-b). *read(2)*. Retrieved from https://man7.org/linux/man-
Linux man-pages. (2025 -c). *inotify(7)*. Retrieved from https://man7.org/linux/man-pages/man7/inotify.7.html
Linux Kernel Organization. (2005). *include/uapi/linux/inotify.h*. Retrieved from https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/uapi/linux/inotify.h
Python Software Foundation. (2025). *ctypes — A foreign function library for Python*. Retrieved from https://docs.python.org/3/library/ctypes.html
Gamma, E., Helm, R., Johnson, R., & Vlissides, J. (1994). *Design Patterns: Elements of Reusable Object-Oriented Software*. Addison-Wesley.
+20 -15
View File
@@ -61,10 +61,7 @@ class ParamWatcher(Params):
if callback not in self._callbacks:
self._callbacks.append(callback)
def _trigger_callbacks(self, path, mask):
if platform.system() == "Linux" and not (mask & (IN_MODIFY | IN_CREATE | IN_DELETE | IN_MOVED_TO | IN_CLOSE_WRITE)):
return
def _trigger_callbacks(self, path):
with self._lock:
if (now := time.monotonic()) - self._last_trigger.get(path, 0) < 0.1:
return
@@ -74,7 +71,7 @@ class ParamWatcher(Params):
for callback in self._callbacks:
try:
callback(path, mask)
callback(path)
except Exception:
cloudlog.exception("Param watcher callback failed")
@@ -110,13 +107,15 @@ class ParamWatcher(Params):
def _run_watcher(self):
system = platform.system()
try:
if system == "Linux":
self._run_linux()
elif system == "Darwin":
self._run_darwin()
except Exception:
traceback.print_exc()
while True:
try:
if system == "Linux":
self._run_linux()
elif system == "Darwin":
self._run_darwin()
except Exception:
cloudlog.exception("Param watcher crashed")
time.sleep(2)
def _run_linux(self):
path = Paths.params_root()
@@ -134,7 +133,8 @@ class ParamWatcher(Params):
i = 0
while i + 16 <= len(buffer):
_, mask, _, name_len = struct.unpack_from("iIII", buffer, i)
self._trigger_callbacks(buffer[i+16:i+16+name_len].rstrip(b"\0").decode(), mask)
if mask & (IN_MODIFY | IN_CREATE | IN_DELETE | IN_MOVED_TO | IN_CLOSE_WRITE):
self._trigger_callbacks(buffer[i+16:i+16+name_len].rstrip(b"\0").decode())
i += 16 + name_len
finally:
if 'poll' in locals():
@@ -149,6 +149,10 @@ class ParamWatcher(Params):
kCFAllocatorDefault = c_void_p(0)
kCFStringEncodingUTF8 = 0x08000100
kFSEventStreamCreateFlagFileEvents = 0x00000010
kFSEventStreamEventFlagItemCreated = 0x00000100
kFSEventStreamEventFlagItemRemoved = 0x00000200
kFSEventStreamEventFlagItemRenamed = 0x00000800
kFSEventStreamEventFlagItemModified = 0x00001000
CF.CFStringCreateWithCString.restype = c_void_p
CF.CFStringCreateWithCString.argtypes = [c_void_p, ctypes.c_char_p, c_uint32]
@@ -166,8 +170,9 @@ class ParamWatcher(Params):
flags_arr = ctypes.cast(flags, POINTER(c_uint32))
for i in range(num):
path = ctypes.cast(paths_arr[i], ctypes.c_char_p).value
if path:
self._trigger_callbacks(os.path.basename(path.decode('utf-8').rstrip('/')), flags_arr[i])
if path and (flags_arr[i] & (kFSEventStreamEventFlagItemCreated | kFSEventStreamEventFlagItemRemoved |
kFSEventStreamEventFlagItemRenamed | kFSEventStreamEventFlagItemModified)):
self._trigger_callbacks(os.path.basename(path.decode('utf-8').rstrip('/')))
except Exception:
traceback.print_exc()
@@ -6,7 +6,7 @@ import tracemalloc
from openpilot.common.params import Params
from openpilot.common.params_pyx import UnknownKeyName
from openpilot.sunnypilot.common.param_watcher import ParamWatcher, IN_MODIFY
from openpilot.sunnypilot.common.param_watcher import ParamWatcher
class TestParamWatcher:
@@ -90,8 +90,8 @@ class TestParamWatcher:
time.sleep(0.2)
event = threading.Event()
param_watcher.add_watcher(lambda key, mask: event.set())
param_watcher._trigger_callbacks(self.bytes_key, IN_MODIFY)
param_watcher.add_watcher(lambda key: event.set())
param_watcher._trigger_callbacks(self.bytes_key)
assert event.wait(timeout=2), "Callback not triggered"
self.params.put(self.bytes_key, b"new")