Causal-Consistent Reversible Debugginglanese/work/cina2015torino-revdeb-Mezzina.pdf ·...

Causal-Consistent Reversible Debugging

3rd Cina Meeting, Torino

Claudio Antares Mezzina(in collaboration with Elena Giachino and Ivan Lanese)

February 11, 2015

Contents

1 Introduction

2 The debugger

3 Implementation

4 Conclusions

Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)1/19

Roadmap

1 Introduction

2 The debugger

3 Implementation

4 Conclusions

Reversible Debugging

Jakob Englom, S4D 2012

Reverse debugging is the ability of a debugger to stop after afailure in a program has been observed and to go back into thehistory of the execution to uncover the reason of failure.

Implications

I Ability to execute an application both in forward andbackward way.

I Reproduce or keep track of the past of an execution.

Introduction Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)2/19

Reversibility and Debugging

Question

When a misbehaviour is detected, how one should proceed in orderto retrace the steps that led to the bug?

I Sequential setting: recursively undo the last action.

I Concurrent setting: there is not a clear understanding ofwhich is the last action.

Introduction Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)3/19

State of the Art for ConcurrentReversible Debugging

Non-deterministic replay

The execution is replayed non deterministically from the start (orfrom the a previous checkpoint) till the desired point.

Deterministic replay/ reverse-execute debugging

A log of the scheduling among threads is kept and then actions arereversed or replayed accordingly.

Introduction Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)4/19

Drawbacks

Non-deterministic replay:I Actions could get scheduled in a different order and hence the

bug may not be reproduced.

I Particularly difficult to reproduce concurrency problems (e.g.race conditions).

Deterministic replay / reverse execution:

I Also action in threads not related to the bug may be undone.

I If one among several indipendent threads causes the bug, andthis thread has been scheduled as first, then one has to undothe entire execution to find the bug.

Introduction Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)5/19

Our Approach: Causal-ConsistentReversibility

Actions are reversed respecting the causes:

I only actions that have caused no successive actions can beundone;

I concurrent actions can be reversed in any order;

I dependent actions are reversed starting from theconsequences.

Benefits

The programmer can easily individuate and undo the actions thatcaused a given misbehaviour.

Introduction Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)6/19

Roadmap

1 Introduction

2 The debugger

3 Implementation

4 Conclusions

Chosen language

I µOz: subset of the Oz language [Van Roy et al.]

I Functional language

I thread-based concurrencyI asynchronous communication via ports (channels)

I µOz advantages:

I well-know stack-based abstract machineI equipped with a causal-consistent reversible

semantics (from previous work)

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)7/19

Syntax

S ::=

skip empty stm

S1 S2 sequence

let x = v in S end var declaration

if x then S1 else S2 end conditional

thread S end thread creation

let x = c in S end procedure declaration

{x y} procedure call

let x = NewPort in S end port creation

{Send x y} send on a port

let x = { Receive y } in S end receive from a port

v ::= true | false | N simple values

c ::= proc {x} S end procedures

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)8/19

The Debugger Commands 1/2co

ntr

ol

forth (f) t (forward execution of one step of thread t)

run (runs the program)

rollvariable (rv) id (c-c undo of the creation of variable id)

rollsend (rs) id n (c-c undo of last n send to port id)

rollreceive (rr) id n (c-c undo of last n receive from port id)

rollthread (rt) t (c-c undo of the creation of thread t)

roll (r) t n (c-c undo of n steps of thread t)

back (b) t (backward execution of one step of

thread t (if possible))

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)9/19

The Debugger Commands 2/2ex

plo

re

list (l) (displays all the available threads)

store (s) (displays all the ids contained in the store)

print (p) id (shows the state of a thread, channel, or variable)

history (h) id (shows thread/channel computational history)

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)10/19

Example of execution

let a = true in (1)

let b = false in (2)

let chan = port in (3)

thread {send chan a}; skip; {send chan b} end; (4)

let y = {receive chan} in skip end (5)

end (6)

end (7)

end (8)

I at line (4) thread t1 is created from thread t0

I t1 fully executes, then t0 fully executes

I what should be the shape of t0 (and of the port) if t1 rolls of 3steps?

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)11/19

t0 let y = {receive chan} in skip end

t1 {send chan a}; skip; {send chan b}x ⊥

I t0 is automatically rolled-back enogh in order to release theread value a

I t0 rolled-back as little as possible (no domino effect)

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)12/19

Debugging Soundness

Properties

1. Every reduction step can be reversed.

2. Every state reached during debugging could have beenreached by forward-only execution from the initial state

Prop 1 ensures that the debugger can undo every forwar step,and, vice-versa, it can re-execute every step previously undone.

Prop 2 ensures that any sequence of debugging commands canonly lead to states which are part of the normal forward-onlycomputations.

The debugger Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)13/19

Roadmap

1 Introduction

2 The debugger

3 Implementation

4 Conclusions

Implementation

I Java based

I Interpreter of the µOz reversible semantics:

I forward and backward stepsI roll as controlled sequence of the backward stepsI rollvarialbe, rollthread, rollsend and rollreceive are

based on roll

I It keeps history and causality information to enablereversibiliby

http://www.cs.unibo.it/caredeb

Implementation Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)14/19

History and Causality Information

I The history of each thread.

I The history of each channel, containing:

I elements of the form (ta, i, v, tb, j)I ta sent a value v which has been received by tbI i and j are pointes to ta and tb send/receive instructions

I We also maintain the following mappings:

I var name → (thread name, i) pointing to the variablecreator (for rollvar)

I thread name → (thread name, i) pointing to the threadcreator (for rollvar)

I could be retrieved by inspecting histories, but storingthem is much more efficient

Implementation Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)15/19

Reversing: code snippet

private static void rollTill(HashMap <String , Integer > map) {

//map contains pairs <thread_name ,i>

Iterator <String > it = map.keySet (). iterator ();

while(it.hasNext ())

{

String id = it.next ();

int gamma = map.get(id);

// getGamma retrieves the next gamma in the history

while(gamma <= getGamma(id)){

try {

stepBack(id);

}

catch (WrongElementChannel e) {

rollTill(e.getDependencies ());}

catch (ChildMissingException e) {

rollEnd(e.getChild ());}

}

}

}

Implementation Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)16/19

Demo Time

Implementation Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)17/19

Roadmap

1 Introduction

2 The debugger

3 Implementation

4 Conclusions

Future Work

I Improve the debugges user experience

I GUII Eclipse plug-in

I Other form of causality analisys

I Move to more popular programing languages / models

I e.g. Java with Actors

I Causal Consistent Replay

Conclusions Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)18/19

Thank you!

Any Questions?

Conclusions Claudio Antares Mezzina (in collaboration with Elena Giachino and Ivan Lanese)19/19