1. 1 Chao Wang, Yu Yang*, Aarti Gupta, and Ganesh Gopalakrishnan* NEC Laboratories America, Princeton, NJ * University of Utah, Salt Lake City, UT Dynamic Model Checking with Property Driven Pruning to Detect Race Conditions

2. 2 Motivation  Concurrent programs are hard to debug  Too many possible thread interleavings  Even for a given input  Data races – a representative type of concurrency bugs  e.g., among flaws in the Therac-25 radiation therapy machine  e.g., related to the 2003 North America Blackout  What’s a data race?  Multiple threads can simultaneously access a shared data variable  At least one is a write

3. 3 Related Work  Precisely detecting data races (or proving race-freedom) is hard  Simultaneous reachability  Previous efforts  Static checking (whole-program analysis)  [Flanagan et al 2002], [Engler & Ashcraft 2002], [Pratikakis et al 2006], [Voung et al 2007], [Kahlon et al 2007], …  Bogus warnings – too many of them!  Dynamic checking (on a particular execution trace)  Eraser [Savage et. al. 1997], Valgrind [Nethercote & Seward 2003], …  May miss real races; bogus warnings – may still appear  Classic model checking algorithms  Full coverage, but requires model building (non-trivial)  For example: pointers, rich data types, …

4. 4 Related Work (2)  (Stateless) dynamic model checking  e.g., Verisoft (Bell labs), CHESS (MSR), Inspect (U. of Utah)  Do not store the program states, but rely on a Depth-First Search to systematically explore all feasible thread schedules  Advantages  Run in the real environment  no bogus warnings  Full coverage for terminating programs  No missed data races  Disadvantages:  The search is inefficient – too many thread interleavings

5. 5 Related Work (3)  DPOR: Dynamic Partial Order Reduction  [Flanagan & Godefroid, POPL 2005]  Main idea: Remove redundant interleavings from each equivalence class of interleavings, provided that the representative has been checked  Still not good enough!  What if an entire equivalence class (of interleavings) is redundant  We need a property-specific reduction!  Remove redundant interleavings within each equivalence class  Remove redundant equivalence classes (w.r.t. the property)

6. 6 Outline Introduction and Related Work  Motivating Example  Set of Locksets  Modeling Unobserved Branches  Experiments  Conclusions

7. Motivating Example 7 Error trace: b1-b7, a1-a4, a5, b8-b9, {a6,b10} Where is the data race? Initial state: x=y=z=0

8. Motivating Example 8 Traces: a1-a4,a5-a8, a9-a11,b1-b7,b8-b11 a1-a4,a5-a8, b1-b7,a9-a11,b8-b11 a1-a4,a5-a8, b1-b7,b8-b11,a9-a11 a1-a4,…………………………………. …… Error: b1-b7, a1-a4, a5, b8-b9, {a6,b10} How would DPOR find it? … … it would take awhile. reduction

9. Motivating Example 9 Traces: a1-a4,a5-a8, a9-a11,b1-b7,b8-b11 a1-a4,a5-a8, b1-b7,a9-a11,b8-b11 a1-a4,a5-a8, b1-b7,b8-b11,a9-a11 a1-a4,………………………………….. …… Error: b1-b7, a1-a4, a5, b8-b9, {a6,b10} In this search sub-space, a9-a11 and b1-b11 run concurrently This sub-space does not have data race!!! How can we do better than that? … … lockset analysis of the sub-tree

10. Lockset Analysis: is the sub-space race-free? 10 In this search sub-space, a9-a11 and b1-b11 run concurrently For each variable access, compute the set of held locks (lockset) This sub-space does not have data race!!!

11. Identifying the locksets is a thread-local computation  scalable This reduction is beyond DPOR, but fits seamlessly with dynamic model checking Lockset Analysis: is the sub-space race-free? 11 ReceFreeSubSpace  prune away redundant equivalence classes

12. 12 Outline Introduction and Related Work  Motivating Example  Set of Locksets  Modeling Unobserved Branches  Experiments  Conclusions

13. Problem Statement  Given a trace and state Si, ask “whether all alternative traces with the same prefix (up to Si) are race free?” 13

14. Set of Locksets 14 Seg_i Seg_j For example, lsSet_x(seg_i) = { {f1}, {f2} } lsSet_x(seg_j) = { {f1,f2} }

15. Set of Locksets: it’s conservative! 15 Seg_i Seg_j RaceFreeSubSpace(S, si) If it reports a race  may be a real race if it reports race-free  indeed race-free When the subspace is race-free, we prune away all the related equivalence classes (of interleavings) Independent from (and potentially more powerful than) POR

16. 16 Outline Introduction and Related Work  Motivating Example  Set of Locksets  Modeling Unobserved Branches  Experiments  Conclusions

17. 17 The Missing Link (unobserved branches) In collecting lsSet_x(seg_i), we have to consider all feasible branches of (seg_i), which includes The observed path Unobserved paths (not-yet-executed) (we are talking about paths in a single thread)

18. Over-approximating Unobserved Branches 18 Our solution: 1.Use a priori static analysis to collect lock-info in all branches; 2.Instrument the source code program For both branches of every if-else statement, add calls to the following functions

19. Over-approximating Unobserved Branches 19 The Unobserved Branch What do we know? 1. it accesses variable x, with lockset {B} U ( {C}\{} ) = {B,C} 2. at the end, the held locks are {B} U ( {C}\{} ) = {B,C}

20. Over-approximating Unobserved Branches 20 The Unobserved Branch What do we know? 1. it accesses variable x, with lockset {B} U ( {C}\{} ) = {B,C} 2. at the end, the held locks are {B} U ( {C}\{} ) = {B,C}

21. Over-approximating Unobserved Branches 21 Our solution: 1.Use a priori static analysis to collect lock-info in all branches; 2.Instrument the source code program For both branches of every if—else statement, add calls to the following functions

22. 22 Outline Introduction and Related Work  Motivating Example  Set of Locksets  Modeling Unobserved Branches  Experiments  Conclusions

23. 23 Experiments  Compared the following methods  DPOR (implemented in Inspect)  DPOR + Property-Driven Pruning  Benchmark programs  Real Linux applications written in C using POSIX thread library  From public domain (sourceforge.net; freshmeat.org, etc.)  Fdrd2  Pfscan – file scanner  Aget – a ftp client for concurrently downloading segments of a large file  Bzip2smt – a multithreaded version of bzip

24. 24 Experiments

25. 25 Conclusions  We present a new pruning method for stateless model checking  Using a trace-based lockset analysis  The reduction (in thread interleavings) is property-specific, and is therefore is beyond POR  Significance  Our method scales much better to realistic programs  No bogus warnings, complete coverage  Future work  Extend the pruning method to handle more general safety properties (deadlock and assertion)