Open-source Versus Commercial Software: A Quantitative Comparison

Transcription

1 Open-source Versus Commercial Software: A Quantitative Comparison Rix Groenboom Reasoning NL BV [email protected]

2 Agenda About Reasoning The Study Inspection Results Analysis Conclusions New results Input for Discussion April Reasoning, Inc 2

3 About Reasoning Reasoning provides an automated inspection service for organizations that develop software Enables building better software in less time and at lower cost Support C, C++, and Java Have inspected over 1B LOC April Reasoning, Inc 3

4 The Study First, some background: Proponents of Open-Source software have long claimed their code is of higher quality Power of peer review Root cause analysis on site enables easier fix Commercial software vendors have long claimed their code is of higher quality Market focused Defined processes for development and testing April Reasoning, Inc 4

5 The Study Why do the study? Our customers wanted to understand the real differences between Open-Source software and Commercial software Reasoning is uniquely positioned to provide this information When was the study performed? Open-source inspection = December 2002 Commercial inspections = Throughout 2002 April Reasoning, Inc 5

6 Used Software Inspection AKA Peer Review Implicit in Extreme Programming Examination of source code to: Detect defects Trace code to requirements Check coding standards April Reasoning, Inc 6

7 Value of Inspections Increased Reliability, Reduced Cost and Time-to-Market Without ASI Software Development Lifecycle Design Code Code Integrate Integrate Test $$$ With ASI Automated Inspections Savings April Reasoning, Inc 7

8 Types of Defects Searched for NULL Pointer Dereference Out of Bounds Array Access Memory Leak Uninitialized Variable Bad Deallocation Using an Automated Software Inspection (ASI) methodology. April Reasoning, Inc 8

9 Architecture ASI Technology Source Code Language Parser Computation Analysis Graph (CAG) Builder Summarize CAGs by Function & Level Detection Rules Symbolic Evaluation Using Value Lattices Potential Violations Reports Defect Review Potential Defects Feasible Path Analysis April Reasoning, Inc 9

10 Project Scope Open Source implementation of TCP/IP in version of Linux Kernel Five commercial implementations of TCP/IP in commercial, general purpose operating systems and telecommunications equipment Why TCP/IP? Well-defined set of published requirements Implementations have been in existence for several years Publicly available conformance tests April Reasoning, Inc 10

11 Can You Spot the Defect? static int sock_fasync(int fd, struct file *filp, int on) { struct fasync_struct *fa, *fna=null, **prev; struct socket *sock; struct sock *sk; If this is true if (on) {.. this memory leaks fna=(struct fasync_struct *)kmalloc(sizeof(struct fasync_struct), GFP_KERNEL); if(fna==null) return -ENOMEM; } sock = socki_lookup(filp->f_dentry->d_inode); if ((sk=sock->sk) == NULL) return -EINVAL;.. and this true April Reasoning, Inc 11

12 DEFECT CLASS: Memory Leak LOCATION: src\linux \net\socket.c : 750 Actual Report DESCRIPTION Local variable fna, declared on line 735, is assigned a pointer to a block of memory allocated by kmalloc on line 741. No other pointer refers to this memory block, so it is inaccessible (still allocated, but unreachable) once fna goes out of scope after line 750. PRECONDITIONS The conditional expression (on) on line 739 evaluates to true AND The conditional expression (fna==null) on line 742 evaluates to false AND The conditional expression ((sk=sock->sk) == NULL) on line 749 evaluates to true. CODE FRAGMENT 733 static int sock_fasync(int fd, struct file *filp, int on) 734 { 735 st ruct f async_st ruc t *f a, *f na=null, **pre v; 736 st r uct socke t *s ock; 737 st r uct sock *s k; i f ( on) 740 { 741 fna=(struct fasync_struct *)kmalloc(sizeof(struct fasync_stru GF P_ KERNEL) ; 742 i f ( f na==null) 743 return -ENOMEM; 744 } sock = socki _l ookup( f i l p- >f _dent r y- >d_i node) ; i f ( ( sk=s ock- >s k) == NULL) 750 re t urn - EI NVAL; l ock_sock( sk) ; pr ev=&( sock- >f as ync _l i st ) ; for (fa=*prev; fa!=null; prev=&fa->fa_next, fa=*prev) 757 i f ( f a- >f a_f i l e==f i l p) 758 br ea k; i f ( on) April Reasoning, Inc 12

13 Defect Overview Memory Leak Null Pointer Dereference Bad Deallocation Out of Bounds Array Access Uninitialized Variable Totals Total Defects in 5 Commercial Implementations Total Defects in Linux Kernel April Reasoning, Inc 13

14 Feedback Linux Developer On OOB access: Nope, not wrong, the table is indexed off-by-one. If you were right, rtnetlink would simply not work. On NPD: In the cases where SKB is NULL, opt is never NULL, check the two callers. On ML: Fixed in the subsequent release. April Reasoning, Inc 14

15 Feedback Commercial Developer For the Linux community to just shrug these off with "well the kernel works so it must be ok" doesn't really cut it. I think the NULL dereference checks should be added, and definitely the out of bounds array checking [..] For example, the out of bounds array reference could start causing a problem by just rearranging the order variables are declared. April Reasoning, Inc 15

16 Fix Rates Reported Repaired % Commercial Implementations Linux Kernel April Reasoning, Inc 16

17 Metrics Comparison D e fe c t D e n s ity Comparison Defects per KLSC W ors tt h ird M iddle Third B est Third 0.00 Industry Average Linux Kernel C om m ercial TC P /IP Linux stack: 0.10 Commercial stack: 0.55 April Reasoning, Inc 17

18 Preliminary Hypothesis Number of Defects? Commercial TCP/IP Peer Review Linux TCP/IP More Research Needed Time Initial defect densities of Open Source vs. Commercial? Defect removal rates of Open Source projects April Reasoning, Inc 18

19 Conclusions Open-source is not inherently worse More research is required Code inspections still find critical defects in extremely well tested software April Reasoning, Inc 19

20 New results Apache test Httpd 2.1-dev (development version 01/31/03) Will determine, track, and report on defect density through lifecycle Results on Java implementation Tomcat Jakarta Results on (soon) What else would you like to see? April Reasoning, Inc 20

21 New results Apache implementation Httpd 2.1-dev (development version from 01/31/03) 31 Defects in 59 KLOC; density 0.53 Normal defect density; non typical defect distribution April Reasoning, Inc 21

22 New results Java inspections: Out of Bounds Array Access NULL pointer dereferences String Comparison Jakarta Tomcat : 11 Defects in 71 KLOC; density 0.15 (Java average commercial software is 0.16). Wait for feedback from the community Next inspection will include resource leaks April Reasoning, Inc 22

23 Towards discussion OK, nice hobby horse for Reasoning and others for research on test efficiency: See SPIder News for a Stanford study on Linux Les Hatton has compared Linux and CMM level Ownership: Who finds the bugs? Who fixes the bugs? Open source: Will commercial end-user will actually make changes to the source? Who will actually review the source (e.g. for quality of the algorithms / security)? April Reasoning, Inc 23

24 Discussion!!! April Reasoning, Inc 24