Robustness Evaluation of Operating Systems

Abstract

SummaryThe premise behind this thesis is the observation that Operating Systems (OS), being the foundation behind operations of computing systems, are complex entities and also subject to failures. Consequently, when they do fail, the impact is the loss of system service and the applications running thereon. While a multitude of sources for OS failures exist, device drivers are often identified as a prominent cause behind failures.In order to characterize the impact of driver failures, at both the OS and application levels, this thesis develops a framework for error propagation-based robustness profiling for an OS. The framework is first developed conceptually and then experimentally validated on a real OS, namely Windows CE .Net. The choice of Windows CE is driven by its representativeness for a multitude of OS's, as well as the ability to customize the OS components for particular needs.For experimental validation, fault injection is a prominent technique that can be used to simulate faults (or errors) in the system by inserting synthetic ones and study their effect. Three key questions with such a technique are where, what and when to inject faults. This thesis shows how injecting errors at the interface between drivers and the OS can be very effective in evaluating the effects driver faults can have.To quantify the OS's robustness, this thesis defines a series of error propagation measures, specifically tailored for device drivers. These measures allow for quantifying and comparing both individual services and device drivers on their susceptibility and diffusing abilities.This thesis compares three contemporary error models on their suitability for robustness evaluation. The classical bit-flip model is found to identify a higher number of severe failures in the system. It also identifies failures for more services than both other models, data type and fuzzing. However, its main drawback is that it requires substantially more injections than the other two. Fuzzing, even though not giving rise to as many failures is able to find new additional services with severe failures.A careful study of the injections performed with the bit-flip model shows that only a few bits are generally useful for identifying new services with robustness weaknesses. Consequently, a new composite model is proposed, combining the most effective bits of the bit-flip model with the fuzzing model's ability to identify new services, giving rise to new model without loss of important information and at the same time incurring a moderate number of injections.To answer the question of when to inject an error this thesis proposes a novel model of a driver's usage profile, focusing on high-level operations being carried out. It guides the injection of errors to instances when the driver is carrying out specific operations. Results from extensive fault injection experiments show that more service vulnerabilities can be discovered. Furthermore, a priori profiling of the drivers can show how effective the proposed approach will be. vi

Separation Kernel Robustness Testing: The XtratuM Case Study

“…Having defined the scope as a fault removal campaign for separation kernels, the next step is devising the system fault model. This is exercised through software fault injection and raises three main questions [10]:…”

“…This "dictionary" of interesting values is composed of values suggested in testing literature [14] and values that uncovered issues in previous tests. Since the greater majority of device drivers are coded in ANSI C [10], C-style data types are usually considered. As an example, possible test values for a signed integer data type would be: -1, 0, 1, MIN_INT and MAX_INT.…”

The impact of fault models on software robustness evaluations

“…We reuse the SWIFI framework of [15,18] and extend it to perform comparisons, as initiated in [24]. In the following section we extend the work presented in [18] and develop generalized and formally accurate definitions the informally specified metrics so that they can be used for (a) any robustness evaluation that aligns with the generic method described in Section 2.1 and (b) quantitative comparisons, i.e.…”

“…We evaluate the performance of four different fault models for a robustness evaluation of the CUE using the SWIFI tool from [15]. Three models have been implemented for extensive experimentation by our group [18,17].…”