When handling async write error, if connection migration is turning on, HandleWriteError will be invoked, and a task will be posted to the message loop to complete migration on write error(actually when the task is executed, tasks will be posted to eventually write the packet to socket), HandleWriteError will return ERR_IO_PENDING synchronously. Instead of leaving the writer unblocked, we should block the writer so as to prevent new packets being written before WritePacketToSocket eventually executes. Failed to block the writer may cause write error triggered again or the cached packet goes out of order. 

void QuicChromiumPacketWriter::OnWriteComplete(int rv) {
  ...
  write_blocked_ = false;
  if (rv < 0) {
    if (MaybeRetryAfterWriteError(rv))
      return;

    // If write error, then call delegate's HandleWriteError, which
    // may be able to migrate and rewrite packet on a new socket.
    // HandleWriteError returns the outcome of that rewrite attempt.
    rv = delegate_->HandleWriteError(rv, std::move(packet_));            <------- MigrateSessionOnWriteError will not be executed when this returns.
    DCHECK(packet_ == nullptr);
    if (rv == ERR_IO_PENDING) 
      return;   <------------ writer become unblocked, if a write comes in, writer will write another packet again
}