1. Efficient and Flexible Architectural Support for Dynamic Monitoring YUANYUAN ZHOU, PIN ZHOU, FENG QIN, WEI LIU, & JOSEP TORRELLAS UIUC

2. Outline Background iWatcher Functionality iWatcher Design Performance Conclusion

3. Static or Dynamic Monitoring? Static Monitoring – Needs annotation, programmer work – Difficult for unsafe languages (C, C++) Dynamic Monitoring – Large instrumentation cost – Significant slowdown, performance loss Dynamic is stronger than Static Monitoring – Dynamic based on actual execution path

4. Code or Location Controlled Dynamic Monitoring? Code-Controlled Monitoring – Monitoring performed by special instructions – Assertions & Dynamic Checkers belong here – No hardware support needed Location-Controlled Monitoring – Monitoring performed only when program accesses watched memory locations by any way – Hardware support is usually required – iWatcher and hardware-assisted watchpoints

5. iWatcher Functionality Flexible and low-overhead dynamic monitoring With hardware support – Without expensive exceptions – The program has its own internal light-weight exception handler, the monitoring function When a watched memory address is accessed, the monitoring function is automatically executed.

6. iWatcher Functionality (cont) If the check of the monitoring action fails, then: – Report, simply report error (non-interactive) – Break, raise a hardware exception, switching control to the debugger – Rollback, revert to a safe checkpoint For the same address, more than one monitors may be watching.

7. iWatcher – Software Level int x, *p;/* assume invariant: x = 1 */ iWatcherOn(&x, sizeof(int), READWRITE, BreakMode, &MonitorX, &x, 1);... p = foo(); /* a bug: p points to x incorrectly*/ *p = 5; /* line A: a triggering access */ z = Array[x]; /* line B: a triggering access */... iWatcherOff(&x, sizeof(int), READWRITE, &MonitorX); bool MonitorX(int *x, int value){ return (*x == value); }

8. Modest Hardware Support (?)

9. How to monitor a location? When iWatcherOn() is called – Add monitoring function to (software) CheckTable – If size < LargeRegion → all words are transferred to L2 cache and tagged update L1 if necessary – If size > LargeRegion → the entire area is tagged in the Range Watch Table (RWT) If RWT full, proceed as if size < LargeRegion

10. How to monitor a location? (cont) If a word is evicted from L2, store the watch bits (if valid) in Victim WatchFlag Table VWT – If VWT full, O/S support (rare) When the word is restored, copy the watch bits from VWT When iWatcherOff is called: – Remove monitoring function from Check Table – If no monitors are watching this area, update VWT, RWT, L1 and L2 bits as necessary.

11. How to detect a triggering access? Out of Order Execution, Pipelining → – Not all instructions will commit For each Load/Store – Check if valid entry exists in RWT – Bring word and WatchFlag from cache (load) or prefetch word to cache and get WatchFlag (store) – Store the flags in the ReOrder Buffer (ROB) – Upon retirement of instruction (if it retires), jump to the monitor, if bits are set.

12. How to Trigger Monitoring Functions? When a triggering access is detected – Save processor status and jump to Main_Check_Function Register – The monitor scans the CheckTable and calls serially all monitors that: Watch this address For this access mode – For performance, the Thread-Level Speculation (TLS) mechanism may be used.

13. Executing Monitoring Functions

15. Comparison to Other Approaches

16. Performance Compared to Valgrind 4-179% overhead, 25-169x less than Valgrind

17. Performance with/without TLS Up to 30% reduction in two cases

18. Performance varying the fraction of triggering loads and TLS

19. Performance varying the size of monitoring function and TLS Above 4 contexts there is no significant improvement

20. Conclusion Some Hardware Changes <180% overhead if 20% of loads are monitored Detects most bugs – Buffer Overflow – Memory Leaks – Access to non-allocated or non-initialized – …