Processor partitioning and hierarchical scheduling have been widely used for composing hard real-time systems on a shared hardware platform while preserving the timing requirements of the systems. Due to the safety critical nature of the hard real-time systems for deriving the sufficient partition size often conservative analysis is used. Applying the exact same analysis for deriving the partition sizes for soft real-time systems result in unnecessary processors overallocation and consequently waste of the CPU resource. In this paper, to address the problem of composing soft and hard real-time systems on a resource constrained shared hardware, we present a multi-level adaptive hierarchical scheduling framework. In our framework, we adapt the processor partition sizes of soft real-time systems according to their need at each time point by on-line monitoring their processor demand. Furthermore, we implement our adaptive framework in the Linux kernel and show the performance of our framework using a case-study.