Lightpanda: the headless browser designed for AI and automation
perf: v8 task wakeup
Karl Seguin43d384d676a2
diff --git a/src/browser/Browser.zig b/src/browser/Browser.zigindex 5f1257efe..c5b67033b 100644--- a/src/browser/Browser.zig+++ b/src/browser/Browser.zig@@ -34,6 +34,7 @@ const Selector = @import("webapi/selector/Selector.zig");const Geolocation = @import("webapi/geolocation/Geolocation.zig");const PermissionState = @import("webapi/Permissions.zig").State;+const log = lp.log;const ArenaPool = App.ArenaPool;const Allocator = std.mem.Allocator;@@ -104,6 +105,8 @@ fc_identity_pool: std.heap.MemoryPool(js.FinalizerCallback.Identity),// #2472).frame_id_gen: u32 = 0,+foreground_task_posted: std.atomic.Value(bool) = .init(false),+const InitOpts = struct {env: js.Env.InitOpts = .{},};@@ -146,6 +149,8 @@ pub fn init(self: *Browser, app: *App, opts: InitOpts) !void {.heartbeat = &self.http_client.heartbeat,};app.watchdog.register(&self.watchdog_entry);++ self.env.setForegroundTaskPostedCallback(onForegroundTaskPosted, self);}pub fn deinit(self: *Browser) void {@@ -158,6 +163,9 @@ pub fn deinit(self: *Browser) void {lp.metrics.js_heap_physical_bytes.add(-@as(i64, @intCast(self.last_reported_js_bytes)));self.last_reported_js_bytes = 0;+ // V8 workers can post until the isolate is gone; http_client is freed+ // after env.+ self.env.setForegroundTaskPostedCallback(null, null);self.env.deinit();// After env.deinit() the Isolate is gone, so no further weak finalizer can// fire — only now is it safe to free the pool backing their parameters.@@ -171,6 +179,24 @@ pub fn deinit(self: *Browser) void {self.selector_cache.deinit();}+// Callback from v8 when a foreground task is posted.+// !!Can be called from various threads!! (v8 worker threads)+fn onForegroundTaskPosted(ctx: ?*anyopaque, delay_in_seconds: f64) callconv(.c) void {+ if (delay_in_seconds > 0) {+ return;+ }+ const self: *Browser = @ptrCast(@alignCast(ctx.?));+ if (self.foreground_task_posted.swap(true, .acq_rel)) {+ // it was already woken up before+ return;+ }++ // wakeup is thread-safe+ self.http_client.handles.wakeup() catch |err| {+ log.err(.browser, "foreground task wakeup", .{ .err = err });+ };+}+// Wait out a watchdog scan before clearing its termination request.pub fn prepareForTeardown(self: *Browser) void {self.app.watchdog.unregister(&self.watchdog_entry);@@ -262,15 +288,15 @@ pub fn runMicrotasks(self: *Browser) void {self.env.runMicrotasks();}-pub fn runMacrotasks(self: *Browser) !bool {+pub fn runMacrotasks(self: *Browser) !void {const env = &self.env;try self.env.runMacrotasks();- const ran_platform_task = env.pumpMessageLoop();+ self.foreground_task_posted.store(false, .release);+ env.pumpMessageLoop();// either of the above could have queued more microtasksenv.runMicrotasks();- return ran_platform_task;}pub fn hasBackgroundTasks(self: *Browser) bool {diff --git a/src/browser/Runner.zig b/src/browser/Runner.zigindex c2fb72857..e648e89a8 100644--- a/src/browser/Runner.zig+++ b/src/browser/Runner.zig@@ -35,7 +35,6 @@ const Runner = @This();session: *Session,browser: *Browser,http_client: *HttpClient,-background_poll_ms: u32 = 0,pub const Opts = struct {};@@ -224,9 +223,8 @@ fn _tick(self: *Runner, comptime is_cdp: bool, timeout_ms: u32, conditions: []Waconst has_runnable_page = hasRunnablePage(session);- var ran_platform_task = false;if (has_runnable_page) {- ran_platform_task = try browser.runMacrotasks();+ try browser.runMacrotasks();}const activity = http_client.activity();@@ -319,18 +317,6 @@ fn _tick(self: *Runner, comptime is_cdp: bool, timeout_ms: u32, conditions: []Waif (has_runnable_page == false) {break :blk 200;}- if (browser.hasBackgroundTasks()) {- // if our last runMacrotasks() ran something and we now have- // a background, then don't linger in the http client waiting- // for I/O, instead, hurry back to run more tasks.- // Else, backoff to 10ms between runs.- // TODO: this is a temporary solution to ensuring background- // tasks are run promptly.The better solution is to have v8- // wakeup the http client when there's work to do.- self.background_poll_ms = if (ran_platform_task) 0 else @min(10, @max(1, self.background_poll_ms * 2));- // msToNextTask could be less than this, but 10ms drift is ok- break :blk self.background_poll_ms;- }break :blk browser.msToNextTask() orelse 200;};const ms_to_wait = @min(timeout_ms, ms_to_next_task);diff --git a/src/browser/js/Context.zig b/src/browser/js/Context.zigindex 2003b876c..dc09833e4 100644--- a/src/browser/js/Context.zig+++ b/src/browser/js/Context.zig@@ -220,7 +220,7 @@ pub fn deinit(self: *Context) void {env.isolate.notifyContextDisposed();// There can be other tasks associated with this context that we need to// purge while the context is still alive.- _ = env.pumpMessageLoop();+ env.pumpMessageLoop();v8.v8__MicrotaskQueue__DELETE(self.microtask_queue);}diff --git a/src/browser/js/Env.zig b/src/browser/js/Env.zigindex 799ed3b9f..edd99d786 100644--- a/src/browser/js/Env.zig+++ b/src/browser/js/Env.zig@@ -567,18 +567,18 @@ pub fn msToNextTask(self: *Env) ?u64 {return if (next_task == std.math.maxInt(u64)) null else next_task;}-pub fn pumpMessageLoop(self: *const Env) bool {+pub fn pumpMessageLoop(self: *const Env) void {var hs: v8.HandleScope = undefined;v8.v8__HandleScope__CONSTRUCT(&hs, self.isolate.handle);defer v8.v8__HandleScope__DESTRUCT(&hs);const isolate = self.isolate.handle;const platform = self.platform.handle;- var ran = false;- while (v8.v8__Platform__PumpMessageLoop(platform, isolate, false)) {- ran = true;- }- return ran;+ while (v8.v8__Platform__PumpMessageLoop(platform, isolate, false)) {}+}++pub fn setForegroundTaskPostedCallback(self: *const Env, callback: v8.ForegroundTaskPostedCallback, ctx: ?*anyopaque) void {+ v8.v8__Platform__SetForegroundTaskPostedCallback(self.platform.handle, self.isolate.handle, callback, ctx);}pub fn hasBackgroundTasks(self: *const Env) bool {diff --git a/src/browser/js/Local.zig b/src/browser/js/Local.zigindex f56bc9a99..1dc926183 100644--- a/src/browser/js/Local.zig+++ b/src/browser/js/Local.zig@@ -121,7 +121,7 @@ pub fn newCallback(pub fn runMacrotasks(self: *const Local) void {const env = self.ctx.env;- _ = env.pumpMessageLoop();+ env.pumpMessageLoop();env.runMicrotasks(); // macrotasks can cause microtasks to queue}