Data races in Java and static analysis

33 Pages·2008·0.41 MB·English

Data races in Java and static analysis
Fr´ed´eric Dabrowski
1INRIA, LANDE
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 1 / 33
Outline
1 Concurrency in Java
2 Static Detection of dataraces
{
Flanagan, Abadi & Freund
lock-based typing
Boyapati, Lee & Rinard
Points-to analysis : Naik & Aiken
(Points-to analysis + Type and effect system)
3 Conclusion and ongoing work
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 2 / 33
Concurrency in Java
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 3 / 33
Concurrency in Java
Concurrency model
Thread-based concurrency : shared memory (fields of shared objects)
lexically scoped locking construct : synchronized(x){. . .}
Preemptive scheduling (Interleaving semantics)
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 4 / 33
Concurrency in Java
Interleaving semantics
(small step) interleaving semantics
sequential semantics
′ ′
ti , Mem →seq t
i , Mem
t, Mem →seq t′, Mem′ {. . . , t
i , . . .}, Mem →inter {. . . , ti′, . . .}, Mem′
Problem : This semantics is incomplete with respect to the Java Memory
Model, unless you write well-synchronized programs
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 5 / 33
Concurrency in Java
Natural hypothesis : sequential consistency
Intuivively, sequential consistency means that all executions respect the
program order.
void mn(){
a a should not observe b
. . .
b
}
Problem : enforcing sequential consistency for all Java programs
makes many of the compiler/processor optimizations illegal.
Why ? some optimizations assume well-synchronized programs !
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 6 / 33
Concurrency in Java
Example : code reordering (cache mechanisms,...)
Original code Optimized code
C.f = C.g = 0 C.f = C.g = 0
1 : x = C.g; 3 : y = C.f ; 2 : C .f = 1; 4 : C .g = 1;
2 : C .f = 1; 4 : C .g = 1; 1 : x = C.g; 3 : y = C.f ;
{Perm(1, 2, 3, 4) | 1 < 2, 3 < 4} {Perm(1, 2, 3, 4) | 2 < 1, 4 < 3}
1 3 {x=0,y=0}
2 3 4 {x=0,y=1} original
4 1 2 {x=1,y=0}
3 1 {x=0,y=0}
2 4 1 3 {x=1,y=1} optimized
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 7 / 33
Concurrency in Java
Admissible behaviors (w.r.t the JMM)
Sequential consitency
Behaviors captured by the interleaving sem.
Java’s memory model is weak memory model
All executions of well-synchronized programs are sequentially consistent a.
aManson, Pugh & Adve : The Java Memory Model (Special Popl issue)
Programs must be well-synchronyzed
several static analysis depend on it
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 8 / 33
Concurrency in Java
Well-synchronized programs
(P1) : For all execution ( w.r.t the interleaving semantics), every
conflicting actions a and b are synchronized
synchronized(C){ synchronized(C){
(1 : x = C.g); (2 : C .f = 1)} (3 : y = C.f ); (4 : C .g = 1)}
lock(C) 1:x=C.g 2:C.f=1 unlock(C) lock(C) 3:y=C.f 4:C.y=1 unlock(C)
seq locking seq
lock(C) 3:y=C.f 4:C.y=1 unlock(C) lock(C) 1:x=C.g 2:C.f=1 unlock(C)
seq seq locking seq seq
compiler/jvm/jit : (P1) ⇒ every exec. is captured by the inter. sem.
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 9 / 33
Concurrency in Java
Happens-before relation
≺hb is the transitive closure of the following rules :
t 1 t
sequentiality a ≺ b
hb
{
1 t
t.start() ≺ a
hb
start/join synchronisation
t 1
a ≺ t.join()
hb
lock-based synchronisation
{
1
unlock(m) ≺ lock(m)
hb
1
write(x.f ) ≺ read(x.f ) (f volatile)
hb
Fr´ed´eric Dabrowski (INRIA, LANDE) Static race detection 10 / 33