diff --git a/CHANGELOG.md b/CHANGELOG.md
index d5e3ff14..e986fe8c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -23,6 +23,8 @@ state is to list breaking changes.
 
 ### Fixed
 
+- Fixed a race condition in the VST3 GUI event loop on Linux. This could cause
+  panics with certain versions of Carla.
 - The CPAL backend now correctly handles situations where it receives fewer
   samples than configured.
 - Fixed the handling of multichannel audio in the CPAL backend.
diff --git a/src/wrapper/vst3/view.rs b/src/wrapper/vst3/view.rs
index 55601d7d..fd0fd255 100644
--- a/src/wrapper/vst3/view.rs
+++ b/src/wrapper/vst3/view.rs
@@ -225,12 +225,8 @@ impl<P: Vst3Plugin> RunLoopEventHandler<P> {
         self.tasks.push(task)?;
 
         // We need to use a Unix domain socket to let the host know to call our event handler. In
-        // theory eventfd would be more suitable here, but Ardour does not support that.
-        // XXX: This can technically lead to a race condition if the host is currently calling
-        //      `on_fd_is_set()` on another thread and the task has already been popped and executed
-        //      and this value has not yet been written to the socket. Doing it the other way around
-        //      gets you the other situation where the event handler could be run without the task
-        //      being posted yet. In practice this won't cause any issues however.
+        // theory eventfd would be more suitable here, but Ardour does not support that. This is
+        // read again in `Self::on_fd_is_set()`.
         let notify_value = 1i8;
         const NOTIFY_VALUE_SIZE: usize = std::mem::size_of::<i8>();
         assert_eq!(
@@ -472,21 +468,34 @@ impl<P: Vst3Plugin> IPlugViewContentScaleSupport for WrapperView<P> {
 #[cfg(target_os = "linux")]
 impl<P: Vst3Plugin> IEventHandler for RunLoopEventHandler<P> {
     unsafe fn on_fd_is_set(&self, _fd: FileDescriptor) {
+        // There should be a one-to-one correlation to bytes being written to `self.socket_read_fd`
+        // and events being pushed to `self.tasks`, but because the process of pushing a task and
+        // notifying this thread through the socket is not atomic we can't reliably just read a byte
+        // from this socket for every task we process. For instance, if `Self::post_task()` gets
+        // called while this loop is already running, it could happen that we pop and execute the
+        // task before the byte gets written to the socket. To avoid this, we'll clear the socket
+        // upfront, and then execute the tasks afterwards. If this situation does happen, then the
+        // worst thing that can happen is that this function is called a second time while
+        // `self.tasks()` is already empty.
+        let mut notify_value = [0; 32];
+        loop {
+            let read_result = libc::read(
+                self.socket_read_fd,
+                &mut notify_value as *mut _ as *mut c_void,
+                std::mem::size_of_val(&notify_value),
+            );
+
+            // If after the first loop the socket contains no more data, then the `read()` call will
+            // return -1 and `errno` will have been set to `EAGAIN`
+            if read_result <= 0 {
+                break;
+            }
+        }
+
         // This gets called from the host's UI thread because we wrote some bytes to the Unix domain
         // socket. We'll read that data from the socket again just to make REAPER happy.
         while let Some(task) = self.tasks.pop() {
             self.inner.execute(task, true);
-
-            let mut notify_value = 1i8;
-            const NOTIFY_VALUE_SIZE: usize = std::mem::size_of::<i8>();
-            assert_eq!(
-                libc::read(
-                    self.socket_read_fd,
-                    &mut notify_value as *mut _ as *mut c_void,
-                    NOTIFY_VALUE_SIZE
-                ),
-                NOTIFY_VALUE_SIZE as isize
-            );
         }
     }
 }