Analysis of typical faults in Linux operating system drivers

• 發表在: Труды Института системного программирования РАН
• Main Authors: V. S. Mutilin, E. M. Novikov, A. V. Khoroshilov
• 格式: Article
• 語言: 英语
• 出版: Russian Academy of Sciences, Ivannikov Institute for System Programming 2018-10-01
• 主題: операционная система; ядро; драйвер; правило взаимодействия; классификация ошибок
• 在線閱讀: https://ispranproceedings.elpub.ru/jour/article/view/1017

Fast evolution of the Linux operating system kernel and drivers, developed by a big programmers community distributed all over the world, led to nowadays there is not a common base of rules that completely describe a correct interaction between drivers and the kernel. This is an obstacle both for programmers that do not have expert knowledge in all peculiarities of the given interaction, and for development and application of tools that could find corresponding typical faults in the automatic way. The given paper presents a method to detect and to classify typical faults and corresponding rules. This method is based on analysis of changes, made to Linux operating system drivers.The paper gives results of the method application to stable versions of the Linux kernel from 2.6.35 till 3.0 starting from October 26, 2010 till October 26, 2011. We analyzed in total 1503 unique commits to drivers, marked 396 (about 27%) of them as fixes of typical faults and provided a classification and a distribution by classes for these typical faults. We distinguished 3 classes of typical faults: generic (faults all C programs are subjected to), specific (faults related to misuses of the Linux kernel API) and synchronization (faults related to parallel execution). We found that specific faults constitute about 50% of all typical faults. Then each typical fault was ascribed to one of 21 first-level subclasses: 12, 7 and 2 for specific, generic and synchronization classes correspondingly. We found that more than 50% of typical faults correspond to 5 first-level subclasses. More than 80% of typical faults correspond to 14 first-level subclasses. The most frequent faults were race conditions during parallel execution – they account for about 17% of all typical faults. Second and third places with about 9% were occupied by leaks of specific resources and null pointer deference.