perf(sync): eliminate scheduling delay and increase download buffer

src/libsync/owncloudpropagator.cpp

@@ -892,7 +892,7 @@ void OwncloudPropagator::scheduleNextJob()
 {
     if (_jobScheduled) return; // don't schedule more than 1
     _jobScheduled = true;
-    QTimer::singleShot(3, this, &OwncloudPropagator::scheduleNextJobImpl);
+    QMetaObject::invokeMethod(this, &OwncloudPropagator::scheduleNextJobImpl, Qt::QueuedConnection);
 }
 
 void OwncloudPropagator::scheduleNextJobImpl()

src/libsync/propagatedownload.cpp

@@ -4,7 +4,7 @@
  * SPDX-License-Identifier: GPL-2.0-or-later
  */
 
 #include "config.h"
 #include "libsync/creds/abstractcredentials.h"
 #include "owncloudpropagator_p.h"
 #include "propagatedownload.h"
@@ -288,7 +288,7 @@
 {
     if (!reply())
         return;
-    int bufferSize = qMin(1024 * 8ll, reply()->bytesAvailable());
+    int bufferSize = qMin(1024 * 256ll, reply()->bytesAvailable());
     QByteArray buffer(bufferSize, Qt::Uninitialized);
 
     while (reply()->bytesAvailable() > 0 && _saveBodyToFile) {

test/CMakeLists.txt

@@ -149,6 +149,7 @@ endif()
 
 nextcloud_add_test(LongPath)
 nextcloud_add_benchmark(LargeSync)
+nextcloud_add_benchmark(ScheduleDelay)
 
 nextcloud_add_test(Account)
 nextcloud_add_test(Folder)

test/benchmarks/benchscheduledelay.cpp

@@ -0,0 +1,116 @@
+/*
+ * SPDX-FileCopyrightText: 2026 Nextcloud GmbH and Nextcloud contributors
+ * SPDX-License-Identifier: CC0-1.0
+ *
+ * This software is in the public domain, furnished "as is", without technical
+ * support, and with no warranty, express or implied, as to its usefulness for
+ * any purpose.
+ */
+
+#include <QCoreApplication>
+#include <QObject>
+#include <QTimer>
+#include <QEventLoop>
+#include <QElapsedTimer>
+#include <QtGlobal>
+
+#include <cstdio>
+
+// Reproduces the scheduling cost that OwncloudPropagator::scheduleNextJob()
+// used to pay. The propagator serializes scheduling passes with a single
+// "_jobScheduled" flag, so exactly one deferred pass is in flight at a time
+// and each of the N passes for an N-file sync pays the deferral latency
+// back-to-back.
+//
+// Old path: QTimer::singleShot(3, ...) -> 3 ms per pass.
+// New path: QMetaObject::invokeMethod(..., Qt::QueuedConnection) -> next
+//           event-loop iteration, no artificial delay.
+//
+// With rounds = 10000 the timer path takes ~30 s (10000 * 3 ms); the queued
+// path completes in a few milliseconds.
+
+namespace {
+constexpr int kRounds = 10000;
+constexpr int kDelayMs = 3; // value previously used in scheduleNextJob()
+}
+
+class ScheduleWorker : public QObject
+{
+    Q_OBJECT
+public:
+    enum class Mode { TimerSingleShot, QueuedInvoke };
+
+    explicit ScheduleWorker(Mode mode, QObject *parent = nullptr)
+        : QObject(parent)
+        , _mode(mode)
+    {
+    }
+
+    // Runs kRounds deferred scheduling passes through the configured
+    // mechanism and returns the elapsed wall-clock time in milliseconds.
+    qint64 run()
+    {
+        _count = 0;
+        QEventLoop loop;
+        _loop = &loop;
+        QElapsedTimer timer;
+        timer.start();
+        scheduleNext();
+        loop.exec();
+        _loop = nullptr;
+        return timer.elapsed();
+    }
+
+private:
+    void scheduleNext()
+    {
+        // Mirror the propagator's "one pass in flight" model: every pass
+        // schedules the next one through the deferral mechanism under test.
+        if (_mode == Mode::TimerSingleShot) {
+            QTimer::singleShot(kDelayMs, this, &ScheduleWorker::tick);
+        } else {
+            QMetaObject::invokeMethod(this, &ScheduleWorker::tick, Qt::QueuedConnection);
+        }
+    }
+
+private slots:
+    void tick()
+    {
+        if (++_count >= kRounds) {
+            if (_loop) {
+                _loop->quit();
+            }
+            return;
+        }
+        scheduleNext();
+    }
+
+private:
+    Mode _mode;
+    int _count = 0;
+    QEventLoop *_loop = nullptr;
+};
+
+int main(int argc, char *argv[])
+{
+    QCoreApplication app(argc, argv);
+
+    ScheduleWorker queued(ScheduleWorker::Mode::QueuedInvoke);
+    const auto queuedMs = queued.run();
+
+    ScheduleWorker timed(ScheduleWorker::Mode::TimerSingleShot);
+    const auto timerMs = timed.run();
+
+    std::printf("ROUNDS %d\n", kRounds);
+    std::printf("QTimer::singleShot(%dms): %lld ms\n", kDelayMs, static_cast<long long>(timerMs));
+    std::printf("QMetaObject::invokeMethod(QueuedConnection): %lld ms\n", static_cast<long long>(queuedMs));
+    std::printf("SAVED: %lld ms\n", static_cast<long long>(timerMs - queuedMs));
+    std::fflush(stdout);
+
+    // Sanity: both paths scheduled the same number of rounds; the queued path
+    // must be dramatically faster because it has no per-pass timer latency.
+    const bool ok = timerMs > queuedMs * 10;
+    return ok ? 0 : 1;
+}
+
+#include "benchscheduledelay.moc"