General Threading Rules
In the IntelliJ Platform, code is executed on one of two thread types:
Event Dispatch Thread (EDT) - also known as the UI thread. It is used for updating the UI and performing changes in the IDE data model. Operations performed on EDT must be fast.
background threads - used for performing costly operations.
Read-Write Lock
In general, code-related data structures in the IntelliJ Platform are covered by a single readers-writer (RW) lock.
Access to the model must be performed in a read or write action for the following subsystems:
Read Access
Reading data is allowed from any thread.
Read operations need to be wrapped in a read action (RA) if not invoked via Application.invokeLater().
If invoked from a background thread or from EDT but via SwingUtilities.invokeLater(), it must be explicitly wrapped in a read action (RA).
Reading data is allowed from any thread.
Reading data from EDT does not require any special effort.
However, read operations performed from any other thread must be wrapped in a read action (RA).
The corresponding objects are not guaranteed to survive between several consecutive read actions. As a rule of thumb, whenever starting a read action, check if the PSI/VFS/project/module is still valid.
Read Action (RA) API
ReadActionrun()orcompute()
Write Access
Writing data is only allowed from EDT, and write operations always need to be wrapped in a write action (WA).
Modifying the model is only allowed from write-safe contexts, including user actions and SwingUtilities.invokeLater() calls from them (see Modality and invokeLater()).
You may not modify PSI, VFS, or project model from inside UI renderers or SwingUtilities.invokeLater() calls.
Write Action (WA) API
WriteActionrun()orcompute()
Modality and invokeLater()
To pass control from a background thread to EDT, instead of the standard SwingUtilities.invokeLater(), plugins should use ApplicationManager.getApplication().invokeLater(). The latter API allows specifying the modality state (ModalityState) for the call, i.e., the stack of modal dialogs under which the call is allowed to execute:
ModalityState.nonModal()/NON_MODAL
The operation will be executed after all modal dialogs are closed. If any of the open (unrelated) projects displays a per-project modal dialog, the action will be performed after the dialog is closed.
ModalityState.stateForComponent()
The operation can be executed when the topmost shown dialog is the one that contains the specified component or is one of its parent dialogs.
None Specified
ModalityState.defaultModalityState() will be used. This is the optimal choice in most cases that uses the current modality state when invoked from EDT. It has special handling for background processes started with ProgressManager: invokeLater() from such a process may run in the same dialog that the process started.
ModalityState.any()
The operation will be executed as soon as possible regardless of modal dialogs. Please note that modifying PSI, VFS, or project model is prohibited from such runnables.
If EDT activity needs to access file-based index (e.g., it's doing any project-wide PSI analysis, resolves references, etc.), please use DumbService.smartInvokeLater(). That way, it is run after all possible indexing processes have been completed.
Background Processes and ProcessCanceledException
Background progresses are managed by ProgressManager class, which has plenty of methods to execute the given code with a modal (dialog), non-modal (visible in the status bar), or invisible progress. In all cases, the code is executed on a background thread, which is associated with a ProgressIndicator object. The current thread's indicator can be retrieved any time via ProgressIndicatorProvider.getGlobalProgressIndicator().
For visible progresses, threads can use ProgressIndicator to notify the user about current status: e.g., set text or visual fraction of the work done.
Progress indicators also provide means to handle cancellation of background processes, either by the user (pressing the Cancel button) or from code (e.g., when the current operation becomes obsolete due to some changes in the project). The progress can be marked as canceled by calling ProgressIndicator.cancel(). The process reacts to this by calling ProgressIndicator.checkCanceled() (or ProgressManager.checkCanceled() if no indicator instances at hand). This call throws a special unchecked ProcessCanceledException (PCE) if the background process has been canceled.
All code working with PSI, or in other kinds of background processes, must be prepared for PCE being thrown at any point. This exception must never be logged but rethrown, and it'll be handled in the infrastructure that started the process. Use inspection Plugin DevKit | Code | 'ProcessCanceledException' handled incorrectly (2023.3).
The checkCanceled() should be called often enough to guarantee the process's smooth cancellation. PSI internals have a lot of checkCanceled() calls inside. If a process does lengthy non-PSI activity, insert explicit checkCanceled() calls so that it happens frequently, e.g., on each Nth loop iteration. Use inspection Plugin DevKit | Code | Cancellation check in loops (2023.1).
Disabling ProcessCanceledException
Throwing PCE from ProgressIndicator.checkCanceled() can be disabled for development (e.g., while debugging the code) by invoking:
These actions are available only if Internal Mode is enabled.
Read Action Cancellability
Background threads shouldn't take plain read actions for a long time. The reason is that if EDT needs a write action (e.g., the user types something), it must be acquired as soon as possible. Otherwise, the UI will freeze until all background threads have released their read actions.
The best-known approach is to cancel background read actions whenever there's a write action about to occur, and restart that background read action later from scratch. Editor highlighting, code completion, Goto Class/File/... actions all work like this.
To achieve that, the lengthy background operation is started with a ProgressIndicator, and a dedicated listener cancels that indicator when write action is initiated. The next time the background thread calls checkCanceled(), a PCE is thrown, and the thread should stop its operation (and finish the read action) as soon as possible.
There are two recommended ways of doing this:
If on EDT, call
ReadAction.nonBlocking()which returnsNonBlockingReadAction(NBRA)If already in a background thread, use
ProgressManager.runInReadActionWithWriteActionPriority()in a loop, until it passes or the whole activity becomes obsolete.
In both approaches, always check at the start of each read action if the objects are still valid, and if the whole operation still makes sense (i.e., not canceled by the user, the project isn't closed, etc.). With ReadAction.nonBlocking(), use expireWith() or expireWhen() for that.
If the activity has to access file-based index (e.g., it's doing any project-wide PSI analysis, resolves references, etc.), use ReadAction.nonBlocking(...).inSmartMode().
Avoiding UI Freezes
Do not Perform Long Operations in EDT
In particular, don't traverse Virtual File System, parse PSI, resolve references or query indexes/stubs.
There are still some cases when the platform itself invokes such expensive code (e.g., resolve in AnAction.update()), but these are being worked on. Meanwhile, please try to speed up what you can in your plugin as it will be generally beneficial and also improve background highlighting performance. For implementations of AnAction, plugin authors should specifically review the documentation of AnAction.getActionUpdateThread() in the Actions section as it describes how threading works for actions.
WriteActions currently have to happen on EDT. To speed them up, as much as possible should be moved out of the write action into a preparation step which can be then invoked in the background (e.g., using ReadAction.nonBlocking(), see above).
Event Listeners
Listeners must not perform any heavy operations. Ideally, they should only clear some caches.
It is also possible to schedule background processing of events, but be prepared that some new events might be delivered before the background processing starts — and thus the world might have changed by that moment or even in the middle of background processing. Consider using MergingUpdateQueue and ReadAction.nonBlocking() (see Read Action Cancellability) to mitigate these issues.
VFS Events
Massive batches of VFS events can be pre-processed in the background, see AsyncFileListener (2019.2 or later).