fix: potential crash in very rare data-race scenario

src/api-wrappers/HelperExtensions.swift

@@ -47,7 +47,7 @@ extension Array {
         let group = DispatchGroup()
         for element in self {
             BackgroundWork.globalSemaphore.wait()
-            BackgroundWork.uiDisplayQueue.async(group: group) {
+            BackgroundWork.mainQueueConcurrentWorkQueue.async(group: group) {
                 group.enter()
                 fn(element)
                 BackgroundWork.globalSemaphore.signal()

src/logic/Application.swift

@@ -50,7 +50,7 @@ class Application: NSObject {
 
     func observeNewWindows() {
         if runningApplication.isFinishedLaunching && runningApplication.activationPolicy != .prohibited {
-            retryUntilTimeout({ [weak self] in
+            retryAxCallUntilTimeout({ [weak self] in
                 guard let self = self else { return }
                 if let windows_ = try self.axUiElement!.windows(), windows_.count > 0 {
                     // bug in macOS: sometimes the OS returns multiple duplicate windows (e.g. Mail.app starting at login)
@@ -92,15 +92,17 @@ class Application: NSObject {
     private func observeEvents() {
         guard let axObserver = axObserver else { return }
         for notification in Application.notifications(runningApplication) {
-            retryUntilTimeout({ [weak self] in
+            retryAxCallUntilTimeout({ [weak self] in
                 guard let self = self else { return }
-                try self.axUiElement!.subscribeToNotification(axObserver, notification, { [weak self] in
+                try self.axUiElement!.subscribeToNotification(axObserver, notification, {
-                    guard let self = self else { return }
+                    DispatchQueue.main.async { [weak self] in
-                    // some apps have `isFinishedLaunching == true` but are actually not finished, and will return .cannotComplete
+                        guard let self = self else { return }
-                    // we consider them ready when the first subscription succeeds, and list their windows again at that point
+                        // some apps have `isFinishedLaunching == true` but are actually not finished, and will return .cannotComplete
-                    if !self.isReallyFinishedLaunching {
+                        // we consider them ready when the first subscription succeeds, and list their windows again at that point
-                        self.isReallyFinishedLaunching = true
+                        if !self.isReallyFinishedLaunching {
-                        self.observeNewWindows()
+                            self.isReallyFinishedLaunching = true
+                            self.observeNewWindows()
+                        }
                     }
                 }, self.runningApplication)
             })

src/logic/BackgroundWork.swift

@@ -2,13 +2,16 @@ import Foundation
 
 // queues and dedicated threads to observe background events such as keyboard inputs, or accessibility events
 class BackgroundWork {
-    static let uiDisplayQueue = DispatchQueue.globalConcurrent("mainQueueConcurrentWork", .userInteractive)
+    static let mainQueueConcurrentWorkQueue = DispatchQueue.globalConcurrent("mainQueueConcurrentWorkQueue", .userInteractive)
-    static let accessibilityCommandsQueue = DispatchQueue.globalConcurrent("accessibilityCommands", .userInteractive)
+    static let accessibilityCommandsQueue = DispatchQueue.globalConcurrent("accessibilityCommandsQueue", .userInteractive)
-    static let accessibilityEventsThread = BackgroundThreadWithRunLoop("accessibilityEvents")
+    static let axCallsQueue = DispatchQueue.globalConcurrent("axCallsQueue", .userInteractive)
-    static let keyboardEventsThread = BackgroundThreadWithRunLoop("keyboardEvents")
+    static let accessibilityEventsThread = BackgroundThreadWithRunLoop("accessibilityEventsThread")
+    static let keyboardEventsThread = BackgroundThreadWithRunLoop("keyboardEventsThread")
 
     // we cap concurrent tasks to .processorCount to avoid thread explosion on the .global queue
     static let globalSemaphore = DispatchSemaphore(value: ProcessInfo.processInfo.processorCount)
+    // AX calls may block for a long time. We use a big semaphore to avoid blocking the main thread in case the window server is busy
+    static let axCallsGlobalSemaphore = DispatchSemaphore(value: 100)
 
     // swift static variables are lazy; we artificially force the threads to init
     static func start() {
@@ -22,12 +25,12 @@ extension DispatchQueue {
         return DispatchQueue(label: label, target: .global(qos: qos.qosClass))
     }
 
-    func asyncWithCap(_ deadline: DispatchTime? = nil, _ fn: @escaping () -> Void) {
+    func asyncWithCap(_ deadline: DispatchTime? = nil, semaphore: DispatchSemaphore = BackgroundWork.globalSemaphore, _ fn: @escaping () -> Void) {
         let block = {
             fn()
-            BackgroundWork.globalSemaphore.signal()
+            semaphore.signal()
         }
-        BackgroundWork.globalSemaphore.wait()
+        semaphore.wait()
         if let deadline = deadline {
             asyncAfter(deadline: deadline, execute: block)
         } else {

src/logic/Window.swift

@@ -55,7 +55,7 @@ class Window {
         AXObserverCreate(application.runningApplication.processIdentifier, axObserverCallback, &axObserver)
         guard let axObserver = axObserver else { return }
         for notification in Window.notifications {
-            retryUntilTimeout({ [weak self] in
+            retryAxCallUntilTimeout({ [weak self] in
                 guard let self = self else { return }
                 try self.axUiElement.subscribeToNotification(axObserver, notification, nil, nil, self.cgWindowId)
             })

src/logic/events/AccessibilityEvents.swift

@@ -2,18 +2,19 @@ import Cocoa
 
 func axObserverCallback(observer: AXObserver, element: AXUIElement, notificationName: CFString, _: UnsafeMutableRawPointer?) -> Void {
     let type = notificationName as String
-    retryUntilTimeout({ try handleEvent(type, element) })
+    retryAxCallUntilTimeout({ try handleEvent(type, element) })
 }
 
-func retryUntilTimeout(_ fn: @escaping () throws -> Void, _ startTime: DispatchTime = DispatchTime.now()) {
+// if the window server is busy, it may not reply to AX calls. We retry right before the call times-out and returns a bogus value
-    DispatchQueue.global(qos: .userInteractive).async {
+func retryAxCallUntilTimeout(_ fn: @escaping () throws -> Void, _ startTime: DispatchTime = DispatchTime.now()) {
+    BackgroundWork.axCallsQueue.asyncWithCap(semaphore: BackgroundWork.axCallsGlobalSemaphore) {
         do {
             try fn()
         } catch {
             let timePassedInSeconds = Double(DispatchTime.now().uptimeNanoseconds - startTime.uptimeNanoseconds) / 1_000_000_000
             if timePassedInSeconds < Double(AXUIElement.globalTimeoutInSeconds) {
-                DispatchQueue.global(qos: .userInteractive).asyncAfter(deadline: .now() + .milliseconds(10)) {
+                BackgroundWork.axCallsQueue.asyncWithCap(.now() + .milliseconds(10), semaphore: BackgroundWork.axCallsGlobalSemaphore) {
-                    retryUntilTimeout(fn, startTime)
+                    retryAxCallUntilTimeout(fn, startTime)
                 }
             }
         }