TMPA-2017: Predicate Abstraction Based Configurable Method for Data Race Detection in Linux Kernel

TMPA-2017: Tools and Methods of Program Analysis
3-4 March, 2017, Hotel Holiday Inn Moscow Vinogradovo, Moscow

Predicate Abstraction Based Configurable Method for Data Race Detection in Linux Kernel
Pavel Andrianov, Vadim Mutilin,Alexey Khoroshilov, Institute for System Programming

For video follow the link: https://youtu.be/SxDVQ7lSTqc
Would like to know more?
Visit our website:
www.tmpaconf.org
www.exactprosystems.com/events/tmpa

Follow us:
https://www.linkedin.com/company/exactpro-systems-llc?trk=biz-companies-cym
https://twitter.com/exactpro

Transcript

1. Institute for System Programming of the Russian Academy of Sciences Predicate Abstraction Based Configurable Method for Data Race Detection in Linux Kernel Pavel Andrianov, Vadim Mutilin, Alexey Khoroshilov
2. 2 int global; Race Condition Thread 1 { ... global = 1; ... } Thread 2 { ... global = 2; ... } A situation, in which simultaneous accesses to the same memory location take place from several threads, one of the accesses is write
3. 3 Real Data Race drivers/net/wireless/marvell/libertas/libertas.ko disconnect: … kfree_skb(priv->currenttxskb); priv->currenttxskb = NULL; priv->tx_pending_len = 0; ... transmit: spin_lock(&priv->driver_lock, flags) if (priv->currenttxskb == NULL) return; … priv->currenttxskb->protocol = eth_type_trans(priv->currenttxskb, priv->dev); netif_rx(priv->currenttxskb); … spin_unlock(&priv->driver_lock, flags)
4. 4 Commit
5. 5 Motivation • Concurrency bugs make up 20% of all across the file systems (A Study of Linux File System Evolution, FAST'13) • Data race conditions make up 17% of all errors in the Linux kernel (Analysis of typical faults in Linux operating system drivers, Proceedings ISP RAN)
6. 6 Other Tools Fast and imprecise Precise, but slow Example: RELAY Example: Threader Difficult to adjust a tool to a particular task Adjustable analysis?
7. 7 Lockset Algorithm Potential data race is a situation, when accesses to the same shared data occur with disjoint sets of locks from two parallel threads, one access is write.
8. 8 Potential Race Condition … *a = 1; ... … mutex_lock(); *a = 1; mutex_unlock(); ... • A disjoint set of synchronization primitives • The same shared data • Accesses from different threads, which can be executed simultaneously • Real (reachable) paths
9. 9 Lightweight core algorithm Method overview A set of warnings Lockset algorithm Shared analysis Heavyweight extensions CEGAR Thread analysis Precise warnings Imprecise warnings
10. 10 Counter Example Guided Abstraction Refinement Error? Safe Counterexample Feasible? Abstraction Refinement Unsafe No Yes Yes No Solver Analysis Interpolation
11. 11 Reachability analysis based on predicate abstraction {},[] {},[] {lock}, [] {}, [] {lock},[] {}, [] {}, [] {}, [] {lock}, [] {lock}, [] {lock}, [] {}, [] {}, [] {}, [] int global; int func(int var) { if (var) { lock(); } global++; if (var) { unlock(); } } {}, []
12. 12 Reachability analysis based on predicate abstraction {} {} {lock}, [var != 0] {}, [var == 0] {lock} {}, [var != 0] {}, [var == 0] {} {lock}, [var != 0] {lock}, [var != 0] {lock}, [var != 0] {}, [var == 0] {}, [var == 0] {}, [var != 0] int global; int func(int var) { if (var) { lock(); } global++; if (var) { unlock(); } }
13. 13 Two Ways of Refinement Analysis Refinement Analysis Refinement
14. 14 Example of False Alarm adm8211_start(dev) adm8211_init_rings(dev) request_irq(adm8211_interrupt) dev->priv->tx_buffers[entry]->skb adm8211_interrupt(dev) dev->priv->tx_buffers[entry]->skb
15. 15 Example of False Alarm adm8211_start(dev) adm8211_interrupt(dev) request_irq(adm8211_interrupt) dev->priv->tx_buffers[entry]->skb dev->priv->tx_buffers[entry]->skb
16. 16 Example of Linux Driver module_init() catc_probe() catc_open() module_exit() usb_register_driver() register_netdev() catc_close() catc_disconnect() unregister_netdev() usb_deregister() usb_driver net_device
17. 17 Example of Model entry_point usb_driver handlers usb_register_driver usb_deregister() net_device handlers register_netdev() unregister_netdev()
18. 18 Анализ разделяемых данных struct my_struct { int *b; } *А; int func() { int *a; a = malloc(); If (undef_value) { A->b = a; } *a = 1; } Доступ к разделяемым данным – потенциальная гонка {} {a → local} {a → local} {a → shared} {a → shared} [undef_value ! = 0] [undef_value == 0]
19. 19 Анализ примитивов синхронизации int global; int func(int var) { if (var) { lock(); } global++; if (var) { unlock(); } } {} {} {lock} {} {lock} {} {lock} {} {lock} {lock} {} {} {} {}
20. 20 Thread Analysis int global; Int start() { global = 0; pthread_create(&thread, .., worker, ..); pthread_join(&thread); result = global; } {1.1} {1.1} {1.1, 2.1} {1.1} {1.1, 2.1} {1.1, 2.0} {1.1, 2.1} {1.1} {1.1} int worker() { global++; }
21. 21 Method Overview
22. 22 Results Unsafes Unknowns Safes Time, h Memory, Gb + Threads, + Refinement 5 61 51 3.2 8.1 - Threads, + Refinement 6 67 44 4.1 4.0 + Threads, - Refinement 27 57 49 2.3 8.2 - Threads, - Refinement 186 54 43 2.1 3.5 113 modules of OS Linux 4.5-rc1 subsystem drivers/net/wireless/
23. 23 2219 warnings at drivers/ • 2219 warnings = 270 unsafe drivers • 55% - imprecision of environment model • 10% - simple memory model • 10% - operations with lists • 10% - other inaccuracies in our analysis • 15% - true races • 290 true warnings = 32 bugs
24. 24 Conclusion • Flexible adjustment of the balance between resources and accuracy • Applicable to industry projects • Real race conditions are found
25. 25 Thank you! Questions?