webcore: sweep worker-owned blob URLs on Worker shutdown

src/jsc/VirtualMachine.rs

@@ -1820,6 +1820,12 @@ pub struct RuntimeHooks {
     /// never lazily created. Called from `WebWorker::shutdown` / `global_exit`
     /// right after `close_all_socket_groups`.
     pub close_dns_for_terminate: fn(),
+    /// `ObjectURLRegistry::singleton().sweep_owner(owner)` — drop every
+    /// `URL.createObjectURL` entry registered by the given script-execution
+    /// context. Called from `WebWorker::shutdown` so a worker that never calls
+    /// `revokeObjectURL` does not pin its blob payloads for the lifetime of
+    /// the process. The registry lives in `bun_runtime::webcore` (forward-dep).
+    pub sweep_object_urls_for_owner: fn(owner: i32),
 }
 
 /// Canonical `EventLoopCtx` vtable for a `*mut VirtualMachine` owner — the JS

src/jsc/web_worker.rs

@@ -1313,6 +1313,12 @@ impl WebWorker {
             if let Some(rare) = vm.rare_data.as_deref_mut() {
                 rare.release_js_handles();
             }
+            if let Some(hooks) = runtime_hooks() {
+                // Drop this worker's URL.createObjectURL entries from the
+                // process-global registry; nothing else revokes them once the
+                // worker is gone.
+                (hooks.sweep_object_urls_for_owner)(vm.initial_script_execution_context_identifier);
+            }
             exit_code = i32::from(vm.exit_handler.exit_code);
             global_object = Some(vm.global);
         }

src/runtime/jsc_hooks.rs

@@ -1487,6 +1487,7 @@ pub(crate) static __BUN_RUNTIME_HOOKS: RuntimeHooks = RuntimeHooks {
     retroactively_report_discovered_tests,
     cancel_all_timers,
     close_dns_for_terminate,
+    sweep_object_urls_for_owner,
 };
 
 // ════════════════════════════════════════════════════════════════════════════
@@ -1636,6 +1637,12 @@ fn close_dns_for_terminate() {
     }
 }
 
+/// `RuntimeHooks::sweep_object_urls_for_owner` — drop every blob-URL registry
+/// entry whose recording context is `owner`.
+fn sweep_object_urls_for_owner(owner: i32) {
+    crate::webcore::object_url_registry::ObjectURLRegistry::singleton().sweep_owner(owner);
+}
+
 pub(crate) fn close_isolation_handles(vm: &mut VirtualMachine) {
     let state = runtime_state();
     if state.is_null() {

src/runtime/webcore/ObjectURLRegistry.rs

@@ -28,19 +28,23 @@ impl Default for ObjectURLRegistry {
 
 pub struct Entry {
     blob: Blob,
+    /// `VirtualMachine::initial_script_execution_context_identifier` of the
+    /// registering context; see [`ObjectURLRegistry::sweep_owner`].
+    owner: i32,
 }
 
-// `Entry` is auto-`Send`: its sole field is `Blob`, which already asserts
-// `Send + Sync` (see `webcore_types::Blob`). No `unsafe impl` needed.
+// `Entry` is auto-`Send`: `Blob` already asserts `Send + Sync`
+// (see `webcore_types::Blob`). No `unsafe impl` needed.
 const _: fn() = || {
     fn assert_send<T: Send>() {}
     assert_send::<Entry>();
 };
 
 impl Entry {
-    pub fn init(blob: &Blob) -> Box<Entry> {
+    pub fn init(blob: &Blob, owner: i32) -> Box<Entry> {
         Box::new(Entry {
             blob: blob.dupe_with_content_type(true),
+            owner,
         })
     }
 }
@@ -54,8 +58,9 @@ impl Drop for Entry {
 
 impl ObjectURLRegistry {
     pub fn register(&self, vm: &mut VirtualMachine, blob: &Blob) -> UUID {
+        let owner = vm.initial_script_execution_context_identifier;
         let uuid = vm.rare_data().next_uuid();
-        let entry = Entry::init(blob);
+        let entry = Entry::init(blob, owner);
 
         self.map.lock().insert(uuid.bytes, entry);
         uuid
@@ -97,6 +102,22 @@ impl ObjectURLRegistry {
         };
         self.map.lock().contains_key(&uuid.bytes)
     }
+
+    /// Drop every entry registered by `owner`. Called from
+    /// `WebWorker::shutdown` (via `RuntimeHooks`) so blob URLs minted inside a
+    /// worker are released when that worker exits. Entries from the main
+    /// context (or other still-live workers) are left untouched.
+    pub fn sweep_owner(&self, owner: i32) {
+        let mut map = self.map.lock();
+        let keys: Vec<[u8; 16]> = map
+            .iter()
+            .filter(|(_, e)| e.owner == owner)
+            .map(|(k, _)| *k)
+            .collect();
+        for k in keys {
+            let _ = map.remove(&k);
+        }
+    }
 }
 
 fn uuid_from_pathname(pathname: &[u8]) -> Option<UUID> {

test/js/web/workers/worker_blob.test.ts

@@ -1,4 +1,5 @@
 import { expect, test } from "bun:test";
+import { bunEnv, bunExe, isWindows } from "harness";
 
 test("Worker from a Blob", async () => {
   const worker = new Worker(
@@ -124,3 +125,92 @@ test("Worker on a revoked blob still works", async () => {
 
   expect(revoked).toBe("revoked.");
 });
+
+// The blob-URL registry is process-global. Before the per-owner sweep, an entry
+// registered inside a worker stayed in the map after the worker exited, and the
+// duped blob store pinned the payload for the lifetime of the process (roughly
+// 1 MB retained per MB minted per dead worker). After the sweep, a dead
+// worker's URL resolves like a revoked URL and the payload is released.
+test("blob URLs created inside a Worker are released when the Worker exits", async () => {
+  await using proc = Bun.spawn({
+    cmd: [
+      bunExe(),
+      "-e",
+      `
+        const workerBody = ${JSON.stringify(`
+          const urls = [];
+          for (let i = 0; i < 4; i++) {
+            const u8 = new Uint8Array(1 << 20);
+            u8.fill(65);
+            urls.push(URL.createObjectURL(new Blob([u8])));
+          }
+          postMessage(urls);
+        `)};
+        const src = URL.createObjectURL(
+          new Blob([workerBody], { type: "application/javascript" }),
+        );
+
+        async function runWorker() {
+          const w = new Worker(src);
+          const urls = await new Promise((resolve, reject) => {
+            w.onmessage = e => resolve(e.data);
+            w.onerror = e => reject(new Error("worker errored: " + e.message));
+          });
+          await new Promise(r => w.addEventListener("close", r, { once: true }));
+          return urls;
+        }
+
+        // Establish the allocator high-water mark before measuring.
+        for (let i = 0; i < 3; i++) await runWorker();
+        Bun.gc(true);
+        Bun.gc(true);
+
+        const rss0 = process.memoryUsage.rss();
+        let deadUrl;
+        for (let i = 0; i < 30; i++) {
+          const urls = await runWorker();
+          deadUrl ??= urls[0];
+        }
+        Bun.gc(true);
+        Bun.gc(true);
+        const growthMB = (process.memoryUsage.rss() - rss0) / 2 ** 20;
+
+        let deadUrlServed = false;
+        try {
+          await fetch(deadUrl);
+          deadUrlServed = true;
+        } catch {}
+
+        // The parent-minted worker-source URL must survive the sweep.
+        const parentUrlStillResolves = (await (await fetch(src)).text()) === workerBody;
+
+        console.log(JSON.stringify({ growthMB, deadUrlServed, parentUrlStillResolves }));
+      `,
+    ],
+    env: {
+      ...bunEnv,
+      // Under ASAN the 1 MB payload frees go into the quarantine (default
+      // quarantine_size_mb=256) instead of being returned, so the RSS delta
+      // measures the quarantine rather than the registry. Disable it so the
+      // threshold is meaningful on ASAN builds.
+      ASAN_OPTIONS: [bunEnv.ASAN_OPTIONS, "quarantine_size_mb=0"].filter(Boolean).join(":"),
+    },
+    stdout: "pipe",
+    stderr: "pipe",
+  });
+
+  const [stdout, stderr, exitCode] = await Promise.all([proc.stdout.text(), proc.stderr.text(), proc.exited]);
+
+  expect(stderr).toBe("");
+  const { growthMB, deadUrlServed, parentUrlStillResolves } = JSON.parse(stdout);
+  if (!isWindows) {
+    // 30 workers * 4 MB: when leaking, growth is ~120 MB; when swept, growth
+    // is allocator noise (typically under 15 MB even on debug builds). RSS
+    // does not drop after worker threads exit on Windows (per-thread mimalloc
+    // arenas stay committed), so the threshold is meaningless there.
+    expect(growthMB).toBeLessThan(40);
+  }
+  expect(deadUrlServed).toBe(false);
+  expect(parentUrlStillResolves).toBe(true);
+  expect(exitCode).toBe(0);
+}, 60_000);