@hydranium/core 1.0.0-next.38 → 1.0.0-next.381

package/src/langium/workspace/file-system-provider.ts

@@ -9,17 +9,94 @@
 
 import { type Tracer } from '@hydranium/protocol';
 import { EmptyFileSystemProvider } from '@hydranium/langium';
-import type { URI } from '@hydranium/langium';
-import { type WritableFileSystemProvider } from '../../documents/ast-document-manager.js';
+import type { FileSystemNode, FileSystemProvider, URI } from '@hydranium/langium';
+import { type SelfSaveRegistry } from './self-save-registry.js';
 import { type LogNameOptions } from '../diagnostics/logger.js';
 import { type ServerSharedServicesMinimal } from '../shared-services.js';
-import { serveVirtualDocument } from './virtual-document.js';
 
 /**
- * Empty-filesystem default {@link WritableFileSystemProvider}. For
- * in-memory / browser / test / CLI hosts that don't have a real filesystem.
- * All writes are no-ops. This is the portable `.`-entry default; the
- * Node-backed `DefaultFileSystemProvider`
+ * File-system provider extension that supports writes. Langium's standard
+ * {@link FileSystemProvider} is read-only; `AstDocumentManager` needs to persist
+ * saves, so consumers wire in a provider that implements this superset.
+ *
+ * Implementations may optionally hold a reference to a
+ * {@link SelfSaveRegistry} so they can record their own write mtimes and
+ * help downstream file-watchers suppress echo events for those writes.
+ * The framework's `DefaultFileSystemProvider` reads it from
+ * `services.workspace.SelfSaveRegistry`; alternative implementations can do
+ * the same or leave the field unset.
+ */
+export interface WritableFileSystemProvider extends FileSystemProvider {
+   writeFile(uri: URI, content: string): Promise<void>;
+   /**
+    * Remove the file at `uri`: the name the URI gives, never the file a
+    * symlink there points at, leaving a parent directory it empties in place.
+    * A no-op when nothing is there; a directory is refused.
+    */
+   deleteFile(uri: URI): Promise<void>;
+   /**
+    * Last-modified time (ms) of the file at `uri`, or `undefined` if it can't
+    * be determined (missing file, or a provider with no disk — in-memory /
+    * browser). Optional: the only consumer is self-save echo suppression
+    * (`didChangeWatchedFiles`), which simply lets a change through when the
+    * mtime is unavailable. Distinct from Langium's `stat` (whose
+    * `FileSystemNode` carries no mtime). Keeping disk access on the provider
+    * seam is what lets the LSP update handler avoid a direct `node:fs` import.
+    */
+   mtimeMs?(uri: URI): Promise<number | undefined>;
+   /**
+    * Resolve `uri` to its real on-disk identity (symlinks collapsed, `..`/`.`
+    * walked, case-folded on case-insensitive filesystems), or `undefined` *iff
+    * the filesystem knows the path is absent*. The single source of the
+    * "nothing loadable here" signal `RealpathDocumentUriPolicy` turns into
+    * the synthetic-placeholder branch of `getOrCreateDocument`.
+    *
+    * Contract for implementers:
+    * - A `file:`-backed provider returns the resolved URI when the file exists,
+    *   and `undefined` when it cannot resolve the path (missing / unreadable).
+    * - A non-`file:` URI (or any URI the provider cannot stat) passes through
+    *   **unchanged** — it is treated as present, never reported absent.
+    * - A provider with no disk to check (in-memory / browser / empty) simply
+    *   omits this method; the framework's `FileSystemProviderRegistry` then
+    *   returns the URI unchanged, and `RealpathDocumentUriPolicy` answers as
+    *   the syntactic `DefaultDocumentUriPolicy` does.
+    *
+    * Synchronous (a `realpath` is a kernel-cached syscall) and optional — the
+    * only consumers are the document-identity policy's `canonicalUri`/`loadUri`.
+    * Keeping the syscall on the provider seam is what lets the policy stay
+    * browser-neutral (no `node:fs` import).
+    */
+   realpath?(uri: URI): URI | undefined;
+   /**
+    * The path `uri` names with every link followed, a link to nothing
+    * included, so that `RealpathDocumentUriPolicy.canonicalUri` keys a file
+    * reached through a link the same before it exists and once it does.
+    * Where {@link realpath} resolves, the same answer; otherwise what is
+    * missing joined as spelled under the real path of what exists. Optional,
+    * as `realpath` is: without it an absent path keys by its syntactic form.
+    */
+   resolveLinks?(uri: URI): URI;
+   /**
+    * Whether a file is at `uri`, where a miss counts only once it has lasted:
+    * on a disk where replacing a file can hide its name for a moment, a single
+    * {@link exists} cannot tell that from a deletion. Ask this before acting on
+    * a miss as a deletion; a caller expecting a miss, such as a create, asks
+    * {@link exists}, since here a real miss costs the whole wait.
+    *
+    * Optional: a provider whose misses are always final omits it, and a caller
+    * takes the answer of {@link exists} as settled.
+    */
+   existsSettled?(uri: URI): Promise<boolean>;
+   /** Registry the provider records its own writes against. Optional — providers without watcher integration may omit. */
+   readonly selfSaveRegistry?: SelfSaveRegistry;
+}
+
+/**
+ * Empty-filesystem {@link WritableFileSystemProvider}. For browser / test / CLI
+ * hosts that don't have a real filesystem. All writes and deletes are no-ops.
+ * On the portable `.` entry it is the
+ * {@link FileSystemProviderRegistry.host} when `context.fileSystemProvider`
+ * gives no writable provider; the Node-backed `DefaultFileSystemProvider`
  * lives in `@hydranium/core/node` (it pulls `node:fs`).
  */
 export class DefaultEmptyFileSystemProvider extends EmptyFileSystemProvider implements WritableFileSystemProvider {
@@ -37,16 +114,181 @@ export class DefaultEmptyFileSystemProvider extends EmptyFileSystemProvider impl
       // no-op
    }
 
-   // Serve virtual documents from the registered document (there is no disk to
-   // read); delegate everything else to the empty base (which throws). The
-   // base declares these param-less, so the param is optional here.
-   override readFile(uri?: URI): Promise<string> {
-      const served = uri !== undefined ? serveVirtualDocument(this.services, uri) : undefined;
-      return served !== undefined ? Promise.resolve(served) : super.readFile();
+   async deleteFile(_uri: URI): Promise<void> {
+      // no-op
+   }
+
+   // `async` so the base's refusal becomes a rejection: it throws
+   // SYNCHRONOUSLY from a method declared to return a promise, and a
+   // synchronous throw escapes the promise chain — `.catch()` never attaches
+   // and `Promise.all` dies while its argument array is still being built, so
+   // a caller reading several URIs cannot handle the miss it asked about. The
+   // base declares `readFile` and `readBinary` param-less, so the param is
+   // optional there.
+   override async readFile(_uri?: URI): Promise<string> {
+      return super.readFile();
+   }
+
+   override async readBinary(_uri?: URI): Promise<Uint8Array> {
+      return super.readBinary();
+   }
+
+   override async stat(uri: URI): Promise<FileSystemNode> {
+      return super.stat(uri);
+   }
+}
+
+/**
+ * The provider bound on `workspace.FileSystemProvider`: it dispatches each call
+ * by the URI's scheme to the provider registered for it in the shared
+ * `fileSystemProviders` group, and to the {@link host} for any scheme without
+ * one.
+ *
+ * So one provider answers for each scheme, and a host's provider needs to know
+ * only its own: the framework registers `virtual:` there, and an adopter adds a
+ * scheme by contributing to the group. The host takes every unclaimed scheme
+ * rather than `file:` alone, which keeps an in-memory workspace under `memory:`
+ * or any other scheme on the host's provider.
+ *
+ * `THost` types the host, so an adopter whose own provider has more members
+ * than the base declares the slot as `FileSystemProviderRegistry<MyProvider>`,
+ * binds a {@link DefaultFileSystemProviderRegistry} with that host, and reads
+ * those members through `services.workspace.FileSystemProvider.host`.
+ */
+export interface FileSystemProviderRegistry<
+   THost extends WritableFileSystemProvider = WritableFileSystemProvider
+> extends WritableFileSystemProvider {
+   /** The provider for every scheme no entry of `fileSystemProviders` claims. */
+   readonly host: THost;
+   /** The provider that answers for `uri`. */
+   providerFor(uri: URI): WritableFileSystemProvider;
+}
+
+/** Whether `provider` dispatches by scheme, as the provider bound on `workspace.FileSystemProvider` does. */
+export function isFileSystemProviderRegistry(provider: WritableFileSystemProvider): provider is FileSystemProviderRegistry {
+   return 'providerFor' in provider && typeof provider.providerFor === 'function';
+}
+
+/** Options for {@link DefaultFileSystemProviderRegistry}. */
+export interface FileSystemProviderRegistryOptions<
+   THost extends WritableFileSystemProvider = WritableFileSystemProvider
+> extends LogNameOptions {
+   /**
+    * The provider for every scheme no entry of `fileSystemProviders` claims.
+    * The framework passes the provider `context.fileSystemProvider` returns,
+    * or a {@link DefaultEmptyFileSystemProvider} when that is not writable.
+    */
+   host: THost;
+}
+
+/**
+ * The framework's {@link FileSystemProviderRegistry}.
+ *
+ * A directory is listed by the provider of its own scheme, so a workspace
+ * root's tree comes from that provider alone.
+ *
+ * `realpath`, `resolveLinks`, `mtimeMs` and `existsSettled` are always present
+ * and forward to the scheme's provider. Where that provider lacks one,
+ * `realpath` and `resolveLinks` return the URI unchanged, `mtimeMs` returns
+ * `undefined` and `existsSettled` answers as `exists`: the answers their
+ * callers already give a provider without them.
+ */
+export class DefaultFileSystemProviderRegistry<
+   THost extends WritableFileSystemProvider = WritableFileSystemProvider
+> implements FileSystemProviderRegistry<THost> {
+   readonly host: THost;
+   protected readonly tracer: Tracer;
+
+   constructor(
+      protected readonly services: ServerSharedServicesMinimal & { fileSystemProviders: Record<string, WritableFileSystemProvider> },
+      options: FileSystemProviderRegistryOptions<THost>
+   ) {
+      this.host = options.host;
+      this.tracer = services.Tracer.for(options.logName ?? this.constructor.name).trace('instantiated');
+   }
+
+   /**
+    * The group is read per call, not in the constructor, so a registered
+    * provider is built on first use and may read slots that reach back to this
+    * one. Only an OWN key of the group counts: the DI container resolves any
+    * key found on the prototype chain, so a scheme such as `constructor` would
+    * otherwise dispatch to a function.
+    */
+   providerFor(uri: URI): WritableFileSystemProvider {
+      const providers = this.services.fileSystemProviders;
+      return Object.hasOwn(providers, uri.scheme) ? providers[uri.scheme] : this.host;
+   }
+
+   // The async methods are `async` so a provider that throws synchronously from
+   // one still rejects, for the reason `DefaultEmptyFileSystemProvider` gives.
+   async writeFile(uri: URI, content: string): Promise<void> {
+      return this.providerFor(uri).writeFile(uri, content);
+   }
+
+   async deleteFile(uri: URI): Promise<void> {
+      return this.providerFor(uri).deleteFile(uri);
+   }
+
+   async mtimeMs(uri: URI): Promise<number | undefined> {
+      return this.providerFor(uri).mtimeMs?.(uri);
+   }
+
+   realpath(uri: URI): URI | undefined {
+      const provider = this.providerFor(uri);
+      return provider.realpath ? provider.realpath(uri) : uri;
+   }
+
+   resolveLinks(uri: URI): URI {
+      const provider = this.providerFor(uri);
+      if (provider.resolveLinks) {
+         return provider.resolveLinks(uri);
+      }
+      // A provider with `realpath` alone still keys a present path by it.
+      return provider.realpath?.(uri) ?? uri;
+   }
+
+   async existsSettled(uri: URI): Promise<boolean> {
+      const provider = this.providerFor(uri);
+      return provider.existsSettled ? provider.existsSettled(uri) : provider.exists(uri);
+   }
+
+   async stat(uri: URI): Promise<FileSystemNode> {
+      return this.providerFor(uri).stat(uri);
+   }
+
+   statSync(uri: URI): FileSystemNode {
+      return this.providerFor(uri).statSync(uri);
+   }
+
+   async exists(uri: URI): Promise<boolean> {
+      return this.providerFor(uri).exists(uri);
+   }
+
+   existsSync(uri: URI): boolean {
+      return this.providerFor(uri).existsSync(uri);
+   }
+
+   async readBinary(uri: URI): Promise<Uint8Array> {
+      return this.providerFor(uri).readBinary(uri);
+   }
+
+   readBinarySync(uri: URI): Uint8Array {
+      return this.providerFor(uri).readBinarySync(uri);
+   }
+
+   async readFile(uri: URI): Promise<string> {
+      return this.providerFor(uri).readFile(uri);
+   }
+
+   readFileSync(uri: URI): string {
+      return this.providerFor(uri).readFileSync(uri);
+   }
+
+   async readDirectory(uri: URI): Promise<FileSystemNode[]> {
+      return this.providerFor(uri).readDirectory(uri);
    }
 
-   override readFileSync(uri?: URI): string {
-      const served = uri !== undefined ? serveVirtualDocument(this.services, uri) : undefined;
-      return served !== undefined ? served : super.readFileSync();
+   readDirectorySync(uri: URI): FileSystemNode[] {
+      return this.providerFor(uri).readDirectorySync(uri);
    }
 }

package/src/langium/workspace/hydranium-langium-document-factory.ts

@@ -12,10 +12,18 @@ import {
    AstUtils,
    DefaultLangiumDocumentFactory,
    type LangiumDocument,
+   type Mutable,
    type ParseResult,
    type URI
 } from '@hydranium/langium';
+import { type CancellationToken } from 'vscode-languageserver-protocol';
+import { type TextDocument } from 'vscode-languageserver-textdocument';
+import { STALE_VERSION } from '@hydranium/protocol';
+import { type ModelLedger } from '../../documents/model-ledger.js';
+import { type CstResidencyService } from '../residency/cst-residency-service.js';
+import { type VersionSyncService } from '../../documents/version-sync-service.js';
 import { type ExtendedServiceRegistry } from '../service-registry.js';
+import { type ServerSharedServicesMinimal } from '../shared-services.js';
 import { type Serializer } from '../serialization/serializer.js';
 
 /** Structural view of the per-language serializer slot on the resolved services. */
@@ -57,6 +65,20 @@ export class HydraniumLangiumDocumentFactory extends DefaultLangiumDocumentFacto
     */
    declare protected readonly serviceRegistry: ExtendedServiceRegistry;
 
+   /** A thunk: a version-sync service whose constructor reads this factory would otherwise be a cycle. */
+   protected readonly versionSyncService: () => VersionSyncService;
+   /** A thunk: a ledger whose constructor reads this factory would otherwise be a cycle. */
+   protected readonly modelLedger: () => ModelLedger;
+   /** A thunk: the residency service resolves `LangiumDocuments`, which is built with this factory. */
+   protected readonly cstResidency: () => CstResidencyService;
+
+   constructor(services: ServerSharedServicesMinimal) {
+      super(services);
+      this.versionSyncService = () => services.workspace.VersionSyncService;
+      this.modelLedger = () => services.workspace.ModelLedger;
+      this.cstResidency = () => services.workspace.CstResidencyService;
+   }
+
    /**
     * Parse `text` for `uri` under the grammar `languageId` names, instead of
     * the one `uri` routes to.
@@ -95,6 +117,91 @@ export class HydraniumLangiumDocumentFactory extends DefaultLangiumDocumentFacto
       return this.createLangiumDocument<T>(parseResult, uri, undefined, text);
    }
 
+   protected override createLangiumDocument<T extends AstNode = AstNode>(
+      parseResult: ParseResult<T>,
+      uri: URI,
+      textDocument?: TextDocument,
+      text?: string
+   ): LangiumDocument<T> {
+      const document = super.createLangiumDocument(parseResult, uri, textDocument, text);
+      // Without `textDocument` Langium creates one at version 0 on first read,
+      // which needs a language for `uri` that a folder URI does not have.
+      const parsed = parseResult.value.$cstNode?.root.fullText ?? text ?? textDocument?.getText();
+      this.versionSyncService().modelProduced(document, { version: this.versionOfParsedText(uri, parsed, textDocument?.version ?? 0) });
+      return document;
+   }
+
+   /**
+    * The version is read before the parse, for the same reason as in
+    * {@link update}: the text document is the store's, which an edit updates
+    * in place while the parse runs.
+    */
+   protected override async createAsync<T extends AstNode = AstNode>(
+      uri: URI,
+      content: string | TextDocument,
+      cancelToken: CancellationToken
+   ): Promise<LangiumDocument<T>> {
+      const version = typeof content === 'string' ? undefined : content.version;
+      const document = await super.createAsync<T>(uri, content, cancelToken);
+      if (version !== undefined) {
+         this.versionSyncService().modelProduced(document, { version });
+      }
+      return document;
+   }
+
+   /**
+    * The store's version is read before Langium's `update`, which reads the
+    * store's text before its first await: read after, it can belong to an
+    * edit that arrived while the parse ran.
+    *
+    * Reconciled with the store here rather than in a `Parsed` listener, which
+    * a cancel skips, leaving the root on the version of its earlier text.
+    */
+   override async update<T extends AstNode = AstNode>(
+      document: Mutable<LangiumDocument<T>>,
+      cancellationToken: CancellationToken
+   ): Promise<LangiumDocument<T>> {
+      // Langium keeps the AST when the CST's text equals the new text, but a
+      // shed document has no CST and would get new AST nodes, leaving its
+      // dependents on the old ones. Grafted only when the new text is the text
+      // the AST was parsed from, so an edit on disk or in the store is still
+      // re-parsed. Not `isCstShed`: the builder has already reset the document
+      // to `Changed`. A refused or diverged graft re-parses, and a build after
+      // this one relinks the dependents: this build has prepared its
+      // documents, and a reset now would drop their build state.
+      const root = document.parseResult.value;
+      const parsedFrom = root.$cstNode === undefined ? this.modelLedger().textOf(root) : undefined;
+      if (parsedFrom !== undefined) {
+         const text = this.textDocuments?.get(document.uri.toString())?.getText() ?? (await this.fileSystemProvider.readFile(document.uri));
+         if (text === parsedFrom) {
+            this.cstResidency().rehydrate(document);
+            if (root.$cstNode === undefined) {
+               this.versionSyncService().rebuildWhileCurrent(document);
+            }
+         }
+      }
+      const stored = this.textDocuments?.get(document.uri.toString())?.version;
+      const updated = await super.update(document, cancellationToken);
+      const version =
+         stored ??
+         this.versionOfParsedText(
+            updated.uri,
+            this.modelLedger().textOf(updated.parseResult.value) ?? updated.textDocument.getText(),
+            updated.textDocument.version
+         );
+      this.versionSyncService().modelProduced(updated, { version });
+      return updated;
+   }
+
+   /**
+    * `version`, or {@link STALE_VERSION} when the store holds `uri` open with
+    * other text than `parsed`: text read from the file while an open landed.
+    */
+   protected versionOfParsedText(uri: URI, parsed: string | undefined, version: number): number {
+      const open = this.textDocuments?.get(uri.toString());
+      return open !== undefined && parsed !== undefined && open.getText() !== parsed ? STALE_VERSION : version;
+   }
+
    /**
     * Serialize `model` via the per-language serializer, or `undefined` if none
     * is bound, it fails, or it is async — `fromModel` is synchronous, so an

package/src/langium/workspace/hydranium-workspace-lock.ts

@@ -7,10 +7,31 @@
  * SPDX-License-Identifier: MIT
  ********************************************************************************/
 
+import { type Clock, type Tracer } from '@hydranium/protocol';
 import { DefaultWorkspaceLock, type MaybePromise } from '@hydranium/langium';
-import { type CancellationToken } from 'vscode-languageserver-protocol';
+import { type CancellationToken, type Disposable } from 'vscode-languageserver-protocol';
+import { type LogNameOptions } from '../diagnostics/logger.js';
+import { type ServerSharedServicesMinimal } from '../shared-services.js';
 import { runInWriteLockScope } from './write-lock-scope.js';
 
+/** Constructor options for {@link HydraniumWorkspaceLock}. */
+export interface HydraniumWorkspaceLockOptions extends LogNameOptions {
+   /**
+    * How long a write may keep running after a later write cancelled it before
+    * the lock logs a warning. Every write and read waits behind it meanwhile, so
+    * one that never ends stops the server. Defaults to 10 s; `0` turns the
+    * check off.
+    */
+   readonly stalledWriteWarnMs?: number;
+}
+
+const DEFAULT_STALLED_WRITE_WARN_MS = 10_000;
+
+/** The write action a {@link HydraniumWorkspaceLock} is running, with the timer its cancellation started. */
+export interface RunningWrite {
+   stallTimer?: Disposable;
+}
+
 /**
  * Langium's {@link DefaultWorkspaceLock} with the write action marked as a
  * write-lock scope, so anything it reaches can be asked whether the lock is
@@ -29,7 +50,70 @@ import { runInWriteLockScope } from './write-lock-scope.js';
  * would turn ordinary read-then-write sequences into false positives.
  */
 export class HydraniumWorkspaceLock extends DefaultWorkspaceLock {
+   /** Backs {@link writeCancellations}. */
+   protected cancelledWriteCount = 0;
+   protected readonly clock: Clock;
+   protected readonly tracer: Tracer;
+   protected readonly stalledWriteWarnMs: number;
+   protected runningWrite?: RunningWrite;
+
+   constructor(services: Pick<ServerSharedServicesMinimal, 'Clock' | 'Tracer'>, options: HydraniumWorkspaceLockOptions = {}) {
+      super();
+      this.clock = services.Clock;
+      this.tracer = services.Tracer.for(options.logName ?? 'WorkspaceLock');
+      this.stalledWriteWarnMs = options.stalledWriteWarnMs ?? DEFAULT_STALLED_WRITE_WARN_MS;
+   }
+
+   /**
+    * How many calls have cancelled the lock's latest write: every `write`,
+    * which cancels the write before it, and every `cancelWrite`. A caller that
+    * reads this right after queuing a write, and reads the same number later,
+    * holds a write nothing has cancelled, whether it is still queued or
+    * already running. Its token answers that only once the write runs, since
+    * a queued write has not been handed it yet.
+    */
+   get writeCancellations(): number {
+      return this.cancelledWriteCount;
+   }
+
    override write(action: (token: CancellationToken) => MaybePromise<void>): Promise<void> {
-      return super.write(token => runInWriteLockScope(() => action(token)));
+      return super.write(token => runInWriteLockScope(() => this.runWrite(action, token)));
+   }
+
+   /** Run `action` as {@link runningWrite}, so a cancellation while it runs can time it. */
+   protected async runWrite(action: (token: CancellationToken) => MaybePromise<void>, token: CancellationToken): Promise<void> {
+      const running: RunningWrite = {};
+      this.runningWrite = running;
+      try {
+         await action(token);
+      } finally {
+         running.stallTimer?.dispose();
+         if (this.runningWrite === running) {
+            this.runningWrite = undefined;
+         }
+      }
+   }
+
+   /**
+    * Counted here rather than in {@link write}, because Langium's `write` cancels
+    * through this method too. A write running when this is called has been
+    * cancelled, by this call or an earlier one, so its timer starts here.
+    */
+   override cancelWrite(): void {
+      this.cancelledWriteCount++;
+      const running = this.runningWrite;
+      if (running && !running.stallTimer && this.stalledWriteWarnMs > 0) {
+         running.stallTimer = this.clock.setTimer(() => this.reportStalledWrite(), this.stalledWriteWarnMs);
+      }
+      super.cancelWrite();
+   }
+
+   protected reportStalledWrite(): void {
+      this.tracer.warn(
+         `A write is still running ${this.stalledWriteWarnMs} ms after it was cancelled, and every write and read waits behind it. ` +
+            'Likely causes: a model-service write or rebuild from inside a build in a host with no write-lock scope tracker, ' +
+            'which waits for the build it runs in, or a write action that ignores its cancellation token. ' +
+            'See https://github.com/eclipse-emfcloud/hydranium/issues/245.'
+      );
    }
 }