Simplifying Failure: Finding Bugs in Programs Effectively
470 likes | 575 Vues
Learn efficient methods for bug detection in programs through simplification of test cases to identify failures. Explore DDMin algorithm, case studies, and practical applications in software testing.
Simplifying Failure: Finding Bugs in Programs Effectively
E N D
Presentation Transcript
Vikas , Purdue Simplifying failure Inducing input
Problem • when testing a program • some test case fails • how to find the bug ?? • problems!!! • test input is huge – 800 lines say • do a binary search manually ? • how much time ?
Problem - example • Mozilla web browser • 1999 • bugzilla listed more than 370 bug reports • faced imminent doom !! • opened a program 'Bug a Thorn' • who ever finds bugs will get a prize • Finding bug is a SERIOUS PROBLEM
Problem – example • one of the bug • Mozilla crashed after 95 user actions • It was not able to print a particular html page of 895 lines • simplified to 3 user actions • html page simplified to 1 line • Now – 895 lines of input simplified to just 1 line !!
How to solve? • “Delta Debugging algorithm“ • take a failing test case • and take a passing test case • simplify tat and produce a minimal test case • the simplified test case still produces the failure
Conflicting issues • Decomposing a specific bug report into simple test case • A bug report should be as specific as possible • on the other hand test case should be as simple as possible • Test case simplification does both! • allows for short problem descriptions • subsumes all details in bug report
Delta Debugging – how • Define what is a successful test case • Feed with a failing test case • Ddmin simplifies it by successive testing • Stop when a minimal test case is reached • Now removing any single input entity would cause the failure to disappear
Analogous example-flight simulation • problem – flight crashes few seconds after take off , how do we find the bug?? • repeat the situation over again and again under changed circumstances • find out what is relevant and what is not relevant • eg , leave the passenger seats – still crashes • eg , leave the coffee vending machine – still crashes ! • hence both of them are irrelevant
DDMin • Not only minimises the failing input • Also maximises the the passing input • Not only limited to html input, character input nor to program input • can be applied to all circumstances that can make a program crash or those which will affect the program execution
Assumptions – reasons for failure • program code • data from storage or input devices • the programs environment • the specific hardware • the operating system • “All the above are called circumstances”
Changes that cause failure • We are interested only in the changeble circumstances • These changeable circumstances make up the program input in most of the cases
What is the change ?? ( delta ) - decomposition? • No specific way to get the changes . • html example • delta can be a single character • can be a single tag • can be a single line also • HOW TO DECOMPOSE THEM??
Definitions - Test cases and tests • According to POSIX starndart for testing • the test can succeed • the test can fail • the test produce intermediate result • We need a function 'rtest' that takes a program run and gives one of the above output.
Complexity of DDMin • Complexity of DDMin is Cx2 + 3Cx • Worst case 2 phases • 1. When every test has an unresolved input • then we go till the maximum granularity of Cx • # of tests to be carried out is 2+4+8+.... + 2 Cx • ~= 4Cx • 2. When testing the last complement fails (∆n) • results in Cx-1 calls • Total : 2(Cx-1)+2(Cx-2)+.....+2=~Cx2-Cx • Add up everything :- 4Cx + Cx2 - Cx
Minimizing algorithm 2 8 16
Case Study1 – GCC gets a fatal signal! - run in WYNOT • a program bug.c crashes when compiled with gcc • but the program crashes only with some optimization options given • does not crash with all optimizations enabled! • code is 755 characters – each character a component ( hence , may have a lot of useless C code )
Case study 1 size of Z = 1 Z[1] will segfault
Minimizing the test case 177 100 77
Minimizing Gcc options gcc -o -fforce-addr bug.c
Case study – Mozilla crashing • One of the bug report in Mozilla firefox • Following operations cause Mozilla to crash • Start Mozilla • Go to bugzilla.mozilla.org • Print to file setting the margins to .50 • Once its done printing do the exact same thing on the same file ( /var/tmp/netscape.ps ) • This causes browser to crash with a segfault
Mozilla crashes • Mozilla input consists of two items • 1 . The sequence of input events • ie the succession of mouse motions • pressed keys and clicked buttons • used XLAB to capture – 711 actions • 2. HTML code of the erroneous www page
Mozilla crashes • out of 711 actions – only 95 were user actions , rest were notifications by X server • out of 95 user actions only 3 are left after 82 test runs . • Invoke Print dialog • Press mouse button 1 on the print button • release mouse button 1
Mozilla crashes 95 user actions 82 runs , 3 user actions
Mozilla crashes - run 896 lines !!! 58 runs , 1 line
Mozilla crashes ! - 1 line • <SELECT_NAME=”priority”_MULTIPLE_SIZE=7> - is the culprit • or its just <SELECT> • in other words , the bug report is now just • Create an HTML page containing “<SELECT>” • Load the page and print it using Alt+P command • the browser crashes with a segmentation fault • or – printing the <SELECT> crashes !!
Minimizing fuzz • bart Miller and his team examined the robustness of UNIX utilities by sending fuzz input ( a huge number of random characters ) • found that 40% of the basic programs crashed or went into infinite loops • ddmin algorithm was tested on fuzz input sequences • for NROFF, TROFF, GLEX, CRTPLOT , UL,UNITS
Simplifying failure inducing input • The 3 case studies discussed show that • the larger the size of the simplified input , the higher is the number of tests required • because determining 1-minimality of a test case with n entities req atleast n tests • because each individual entity is removed and tested • for flex , the number of tests vary upto 104 for low precision to 36,000 for high precision .
other approaches ? • Simply stop the process when a certain time limit is reached • Simply stop the process when the input test case is reduced by a certain extent • Better approach is 'Isolation' • Find one relevant part of the test case .removing this particular part makes the failure go away. • Simplifying meant that – the simplified test case had all the relevant parts
Isolating example 7 tests!!
Simplyfying 26 tests
Future work • Domain specific methods • knowledge about the input structure can very much enhance the performance • for ex – valid program inputs are described by grammars , would be nice to rely on such grammars • can exclude syntactically invalid inputs