Beyond Breakpoints: A Tour of Dynamic Analysis

beyond breakpointsA Tour of Dynamic Analysis
Nathan Taylor, Fastly
http://nathan.dijkstracula.net
@dijkstracula

T. Ball: The Concept of Dynamic Analysis ESEC ‘99
“Dynamic analysis is the
analysis of the properties
of a running program […]
(usually through program
instrumentation).”

From Pin: Building Customized Program Analysis Tools with Dynamic Instrumentation

From ReVirt: Enabling Intrusion Analysis through Virtual-Machine Logging and Replay

classic use-cases
Debugging a program crash or race condition
Input to understand a new codebase
Analysis of obfuscated software or malware

who cares
Huge gulf between what tools practitioners use in
industry, and academic dynamic analysis work
Many techniques shared with other cool areas
like OS virtualization and compiler theory
Techniques have different performance tradeoffs;
knowing the lay of the land will help you choose
the right tool for your problem

dynamic vs static analysis
Static analysis is a conservative approximation of
runtime behavior
Programs are only partially known (dynamic linking;
user input)
Imprecise: consider a program that typechecks but still
has a bug
Language specifications often assume a single thread
of execution (or, don’t specify at all): “valid” programs
can still allow race conditions!

can’t prove correctness
but can demonstrate failure

today’s topics
Omniscient Debugging and State Tracking
Analyzing Concurrent Systems
Areas of Future Work

for each topic…
What open-source tooling exists?
How does it work under the hood?
What contributions has academia made?

J. Mickens: The Night Watch ;login Nov 2013
“Despair is when you’re
debugging a kernel driver and
you look at a memory dump and
you see that a pointer has a value
of 7. THERE IS NO
HARDWARE ARCHITECTURE
THAT IS ALIGNED ON 7.
Furthermore, 7 IS TOO SMALL
AND ONLY EVIL CODE WOULD
TRY TO ACCESS SMALL
NUMBER MEMORY. ”

line 9
old i: N/A
new i: 7
line 10
old i: 7
new i: 8
line 11
old j: N/A
new j: 5
line 12
old i: 8
new j: 13

Hofer et al. Design And Implementation of
a Backward-In-Time Debugger
NODe ‘06

King et al. Debugging Operating Systems
with Time-Traveling Virtual Machines
Usenix ‘05
“…operating systems run
for long periods of time,
such as weeks, months, or
even years.”

Usenix ‘05

Usenix ‘05
“Replay causes the virtual
machine to transition
through the same states as
it went through during the
original run”
“A VM can be replayed by
starting from a checkpoint,
then replaying […] the
network, keyboard, clock,
and timing of interrupts”

Usenix ‘05
“Taking checkpoints every minute
added less than 4% time overhead
and 1-5 MB/sec space overhead.”
“The logging added 3-12% in
running time and 2-85 KB/sec in
log space.”
“Taking checkpoints every 10
second added 16-33% overhead
and enabled reverse debugging
commands to complete in about
12 seconds”

GDB: linked list storage of every memory op
Easy to build and reason about
Both recording and replaying is very slow-
+
King (et al): Periodic state checkpointing
Easy to jump between large period of time
Can trade better fidelity for greater overhead
Have to replay execution between checkpoints
+
+
-
King (et al): Only records external, non-deterministic events
Reduces log size and improves performance
synchronizing external events becomes complicated
+
-

Improving
performance with
JIT compilation

compilation-based tooling
Rather than interrupt execution to
return control-flow to a debugger,
weave instrumentation into the
existing codebase at runtime
Allows for both better
performance and more flexible
analysis tools

Luk et al. Pin: Building Customized Program
Analysis Tools w/ Dynamic Instrumentation PLDI ‘05
“The goal is to provide an
implementation platform for
building program analysis tools”
“Its API allows a tool to insert
calls to instrumentation at
arbitrary locations in the
executable”
“Pin provides efficient
instrumentation by using a just-
in-time compiler to insert and
optimize code”

… 
4 byte write to 0x7fff572b3b1c
4 byte write to 0x7fff572b3b1c 
1 byte write to 0x7f86c0c04b00
1 byte write to 0x7f86c0c04b04
4 byte write to 0x7fff572b3b20
…

recordMemWrite(%rip,
&i);
&i);
&j);
&i);

movl

$0x7,
-‐0x8(%rbp)INS_AddInstrumentFunction()
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>
TRACE_AddInstrumentFunction()
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
RTN_AddInstrumentFunction()
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0x40052d
<baz>

leaveq

retq
push

%rbp

mov

%rsp,%rbp

mov

0x200b10(%rip),%rax

movl

$0x2a,(%rax)

pop

%rbp

retq
IMG_AddInstrumentFunction()

thread-sanitizer/blob/master/gcc/tree-tsan.c
static
void

instrument_gimple
(gimple_stmt_iterator
gsi)

{

unsigned
i;

gimple
stmt;

enum
gimple_code
gcode;

tree
rhs,
lhs;

stmt
=
gsi_stmt
(gsi);

gcode
=
gimple_code
(stmt);

if
(gcode
==
GIMPLE_ASSIGN)
{

/*
Handle
assignment
lhs
as
store.

*/

lhs
=
gimple_assign_lhs
(stmt);

instrument_expr
(gsi,
lhs,
1);

/*
Handle
operands
as
loads.

*/

for
(i
=
1;
i
<
gimple_num_ops
(stmt);
i++)
{

rhs
=
gimple_op
(stmt,
i);

instrument_expr
(gsi,
rhs,
0);

}

}

}
static
void

instrument_expr
(gimple_stmt_iterator
gsi,
tree
expr,
int
is_write)

{

enum
tree_code
tcode;

unsigned
fld_off,
fld_size;

tree
base,
rhs;

gimple
stmt;

gimple_seq
gs;

location_t
loc;

base
=
get_base_address
(expr);

if
(base
==
NULL_TREE
||
TREE_CODE
(base)
==
SSA_NAME

||
TREE_CODE
(base)
==
STRING_CST)

return;

tcode
=
TREE_CODE
(expr);

/*
Below
are
things
we
do
not
instrument

(no
possibility
of
races
or
not
implemented
yet).

*/

if
(/*
Compiler-‐emitted
artificial
variables.

*/

(DECL_P
(expr)
&&
DECL_ARTIFICIAL
(expr))

/*
The
var
does
not
live
in
memory
-‐>
no
possibility
of
races.

*/

||
(tcode
==
VAR_DECL

&&
TREE_ADDRESSABLE
(expr)
==
0

&&
TREE_STATIC
(expr)
==
0)

/*
Not
implemented.

*/

||
TREE_CODE
(TREE_TYPE
(expr))
==
RECORD_TYPE

||
tcode
==
CONSTRUCTOR

||
tcode
==
PARM_DECL

/*
Load
of
a
const
variable/parameter/field.

*/

||
is_load_of_const
(expr,
is_write))

return;

static
gimple_seq

instr_memory_access
(tree
expr,
int
is_write)

{

tree
addr_expr,
expr_type,
call_expr,
fdecl;

gimple_seq
gs;

unsigned
size;

gcc_assert
(is_gimple_addressable
(expr));

addr_expr
=
build_addr
(unshare_expr
(expr),
current_function_decl);

expr_type
=
TREE_TYPE
(expr);

while
(TREE_CODE
(expr_type)
==
ARRAY_TYPE)

expr_type
=
TREE_TYPE
(expr_type);

size
=
(TREE_INT_CST_LOW
(TYPE_SIZE
(expr_type)))
/
BITS_PER_UNIT;

fdecl
=
get_memory_access_decl
(is_write,
size);

call_expr
=
build_call_expr
(fdecl,
1,
addr_expr);

gs
=
NULL;

force_gimple_operand
(call_expr,
&gs,
true,
0);

return
gs;

static
tree

get_memory_access_decl
(int
is_write,
unsigned
size)

{

tree
typ,
*decl;

char
fname
[64];

static
tree
cache
[2][17];

is_write
=
!!is_write;

if
(size
<=
1)

size
=
1;

else
if
(size
<=
3)

size
=
2;

else
if
(size
<=
7)

size
=
4;

else
if
(size
<=
15)

size
=
8;

else

size
=
16;

decl
=
&cache[is_write][size];

if
(*decl
==
NULL)
{

snprintf(fname,
sizeof
fname,
"__tsan_%s%d",

is_write
?
"write"
:
"read",
size);

typ
=
build_function_type_list
(void_type_node,

ptr_type_node,
NULL_TREE);

*decl
=
build_func_decl
(typ,
fname);

}

return
*decl;

}

if
(tcode
==
COMPONENT_REF)
{

tree
field
=
TREE_OPERAND
(expr,
1);

if
(TREE_CODE
(field)
==
FIELD_DECL)
{

fld_off
=
TREE_INT_CST_LOW
(DECL_FIELD_BIT_OFFSET
(field));

fld_size
=
TREE_INT_CST_LOW
(DECL_SIZE
(field));

if
(((fld_off
%
BITS_PER_UNIT)
!=
0)

||
((fld_size
%
BITS_PER_UNIT)
!=
0))
{

/*
As
of
now
it
crashes
compilation.

TODO:
handle
bit-‐fields
as
if
touching
the
whole
field.

*/

return;

}

}

}

/*
TODO:
handle
other
cases

(FIELD_DECL,
MEM_REF,
ARRAY_RANGE_REF,
TARGET_MEM_REF,
ADDR_EXPR).

*/

if
(tcode
!=
ARRAY_REF

&&
tcode
!=
VAR_DECL
&&
tcode
!=
COMPONENT_REF

&&
tcode
!=
INDIRECT_REF
&&
tcode
!=
MEM_REF)

return;

func_mops++;

stmt
=
gsi_stmt
(gsi);

loc
=
gimple_location
(stmt);

rhs
=
is_vptr_store
(stmt,
expr,
is_write);

if
(rhs
==
NULL)

gs
=
instr_memory_access
(expr,
is_write);

else

gs
=
instr_vptr_update
(expr,
rhs);

set_location
(gs,
loc);

/*
Instrumentation
for
assignment
of
a
function
result

must
be
inserted
after
the
call.

Instrumentation
for

reads
of
function
arguments
must
be
inserted
before
the
call.

That's
because
the
call
can
contain
synchronization.

*/

if
(is_gimple_call
(stmt)
&&
is_write)

gsi_insert_seq_after
(&gsi,
gs,
GSI_NEW_STMT);

else

gsi_insert_seq_before
(&gsi,
gs,
GSI_SAME_STMT);

}

“Pin compiles from one ISA
directly into the same ISA
without going through an
intermediate format and
stored in a software-based
code cache”
“only code residing in the
code cache is executed - the
original code is never
executed.”

“Pin compiles from one ISA
directly into the same ISA
without going through an
intermediate format”
movl

$0x7,
-‐0x8(%rbp)
addl

$0x1,
-‐0x8(%rbp)
mov

-‐0x8(%rbp),
%eax
sub

$0x3,
%eax
mov

%eax,
-‐0x4(%rbp)
mov

$0x4(%rbp),
%eax
add

%eax,
-‐0x8(%rbp)
-‐0x8(%rbp))
-‐0x8(%rbp))
-‐0x4(%rbp))
-‐0x8(%rbp))
movl

$0x7,
-‐0x8(%rbp)
addl

$0x1,
-‐0x8(%rbp)
mov

-‐0x8(%rbp),
%eax
sub

$0x3,
%eax
mov

%eax,
-‐0x4(%rbp)
mov

$0x4(%rbp),
%eax
add

%eax,
-‐0x8(%rbp)

code rewriting
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0x40052d
<baz>

leaveq

retq
push

%rbp

mov

%rsp,%rbp

mov

0x200b10(%rip),%rax

movl

$0x2a,(%rax)

pop

%rbp

retq

jump rewriting
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0x40052d
<baz>

leaveq

retq
push

%rbp

mov

%rsp,%rbp

mov

0x200b10(%rip),%rax

movl

$0x2a,(%rax)

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%wax

callq

0xdeadbeef

pop

%rbp

retq
callq

0x400540
<bar>callq

0xdeadbeef

jump rewriting
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0x40052d
<baz>

leaveq

retq
push

%rbp

mov

%rsp,%rbp

mov

0x200b10(%rip),%rax

movl

$0x2a,(%rax)

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0xdeadbeef

leaveq

retq

jump rewriting
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0x40052d
<baz>

leaveq

retq
push

%rbp

mov

%rsp,%rbp

mov

0x200b10(%rip),%rax

movl

$0x2a,(%rax)

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

mov

$0x100,%edi

callq

0x400430
<malloc@plt>

mov

%rax,0x200ac8(%rip)

mov

$0x0,%eax

callq

0x400540
<bar>

pop

%rbp

retq
push

%rbp

mov

%rsp,%rbp

sub

$0x10,%rsp

movq

$0x0,-‐0x8(%rbp)

mov

-‐0x8(%rbp),%rax

add

$0x1c,%rax

mov

%rax,0x200ae9(%rip)

mov

$0x0,%eax

callq

0x40052d
<baz>

leaveq

retq
push

%rbp

mov

%rsp,%rbp

mov

0x200b10(%rip),%rax

movl

$0x2a,(%rax)

pop

%rbp

retq

stateless(ish) instrumentation
https://software.intel.com/sites/landingpage/pintool/docs/67254/Pin/html/index.html#EXAMPLES

Nethercote et al. Valgrind: A Framework for
Heavyweight Dynamic Binary Instrumentation PLDI ‘07
“which values are undefined (i.e.
uninitialised, or derived from
undefined values) and can thus
detect dangerous uses of
undefined values.”
“which values are tainted (i.e.
from an untrusted source,) and
can thus detect dangerous uses of
tainted values.”
“which word values are array
pointers, and from this can detect
bounds errors.”

One interesting group of
DBA tools are those that
use shadow values. These
tools shadow, purely in
software, every register and
memory value with another
value that says something
about it. We call these
shadow value tools.

<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>0x08048000
0x08048001
0x08048002
0xc000000
.
.
.

<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>
<value>0x08048000
0x08048001
0x08048002
0xc000000
.
.
.
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata>
<metadata> 0x08048000
0x08048001
0x08048002
0xc000000
.
.
.

0xc000000
memcheck/mc_main.c
set? set? set? set? set? set? set? set?

0xc000000
memcheck/mc_main.c
set? set? set? set? set? set? set? set?valid?

disassemble-and-
resynthesise: machine code is
converted to an IR in which each
instruction becomes one or more
IR operations. This IR is
instrumented and then converted
back to machine code.
copy-and-annotate: incoming
instructions are copied through
verbatim except for necessary
control flow changes. Tools use
the annotations to guide their
instrumentation.

“Pin provides no built-in support
for [shadow memory], so tools
must cope with the non-atomicity
of loads/stores and shadow
loads/stores themselves.” 
 
“Valgrind’s thread serialisation
and asynchronous signal
treatment frees shadow value
tools from having to deal with this
issue.”

“in Figure 7(b), Pin
outperforms both
DynamoRIO and
Valgrind by a significant
margin: on average,
Valgrind slows the
application down by 8.3
times […] and Pin by 2.5
times.”

Pin: a “copy and annotate” system
Lower overhead than the alternatives
Close-to-the-metal instrumentation API
Best for tracking control flow, rather than memory
+
+
-
Valgrind: a “disassemble and resynthsize” system
Richer API and shadow memory allows for per-memory word
metadata tracking
Greater runtime overhead
+
-

problem statement
“What values did a piece of memory
have over time?”
“How does a value propagate
through a running system?”

Chow et al. Understanding Data Lifetime via Whole
System Simulation USENIX Sec ‘04
TaintBochs tracks sensitive
data by “tainting” it at the
hardware level. Tainting
information is then
propagated across operating
system, language, and
application boundaries,
permitting analysis of
sensitive data handling at a
whole system level.

We have used TaintBochs
to analyze sensitive data
handling in several large,
real world applications.
Among these were Mozilla,
Apache, and Perl, which
are used to process
millions of passwords,
credit card numbers, etc.
on a daily basis.

“We must decide how
operations in the system
should affect shadow state.
If two registers A and B are
added, and one of them is
tainted, is the register where
the result are stored also
tainted? We refer to the
collective set of policies that
decide this as the
propagation policy.”

“ Our basic propagation
policy is simply that if
any byte of any input
value is tainted, then all
bytes of the output are
tainted. This policy
is clearly conservative
and errs on the side of
tainting too much.”

“ Sometimes tainted values are
used by instructions as indexes
into non-tainted memory (i.e. as
an index into a lookup table).”
For example, Linux routinely
remaps keyboard device data
through a lookup table before
sending keystrokes to user
programs. Thus, user programs
never directly see the data read in
from the keyboard device, only the
non tainted values they index in
the kernel’s key remapping
table.”

methodology
shadow memory to track
information flow
time-traveling debugging to
uncover relationship between
tainted memory

Slowinska & Bos: Pointless Tainting? Evaluating the
Practicality of Pointer Tainting Eurosys ‘09

Slowinska & Bos: Pointless Tainting? Evaluating the
Practicality of Pointer Tainting Eurosys ‘09
“…in-depth analysis of the
problems of pointer tainting
on real systems which shows
that it does not work against
malware spying on users’
behaviour”
“an analysis and evaluation
of all known fixes to the
problems that shows that
they all have serious
shortcomings.”

Usenix ‘05
“Replay causes the virtual
machine to transition through
the same states as it went
through during the original run”
“A virtual machine can be
replayed by starting from a
checkpoint, then replaying all
sources of nondeterminism […]
the network, keyboard, clock,
and timing of interrupts”
Sequential

problem statement
“how to build a dynamic analysis tool
that correctly runs in a concurrent
environment?”
“What challenges does concurrent
software have that dynamic analysis
could help with?”

concurrent record & replay
use locking operations as sequencing
points
use virtual memory page protection
(come up with ideas for custom
hardware)

“Valgrind serialises thread
execution with a thread locking
mechanism. Only the thread
holding the lock can run,”
“The kernel still chooses which
thread is to run next, but Valgrind
dictates when thread-switches
occur and prevents more than one
thread from running at a time.”
“How to best overcome this
problem remains an open
research question.”

Savage et al: Eraser: A Dynamic Data Race
Detector for Multithreaded Programs
SOSP ‘97

Savage et al: Eraser: A Dynamic Data Race
Detector for Multithreaded Programs
SOSP ‘97
“Eraser instruments
each call to acquire or
release a lock”

Dunlap et al: Execution Replay for
Multiprocessor Virtual Machines
VEE ‘08

VEE ‘08
“if two memory instructions on
different processors access the
same page, and one of them is a
write, there will be a CREW
event between the instructions
on each processor.”
• concurrent-read: All cpus
have read permission, but none
have write permission.
• exclusive-write: One cpu
(called the owner) has both read
and write permission; all virtual
cpus have no permission.

VEE ‘08
“we use hardware page
protections, enforced
by the memory
management unit
(MMU), which will
check each read and
write as the instruction
executes, and cause a
fault to the hypervisor
on any violation.”

CREW limitations
virtual memory hardware only
operates on 4k blocks
DMA not implemented in the paper
10x slowdown in the worst case!

race detection techiques
Lockset algorithm
Happens-before relation
Hybrid techniques

examples from https://github.com/google/sanitizers/wiki/ThreadSanitizerPopularDataRaces
int
var;

void
Thread1()
{

//
Runs
in
one
thread.

var++;

}

void
Thread2()
{

//
Runs
in
another
thread.

var++;

}

//
Ref()
and
Unref()
may
be
called
from
several
threads.

//
Last
Unref()
destroys
the
object.

class
RefCountedObject
{

...

public:

void
Ref()
{

ref_++;

//
Bug!

}

void
Unref()
{

if
(-‐-‐ref_
==
0)

//
Bug!
Need
to
use
atomic
decrement!

delete
this;

}

private:

int
ref_;

};

bool
done
=
false;

void
Thread1()
{

while
(!done)
{

do_something_useful_in_a_loop_1();

}

do_thread1_cleanup();

}

void
Thread2()
{

do_something_useful_2();

done
=
true;


}

bool
done
=
false;

void
Thread1()
{

while
(!done)
{

do_something_useful_in_a_loop_1();

}


}

void
Thread2()
{

done
=
true;

do_something_useful_2();


}

“Eraser instruments
each call to acquire
or release a lock”
“In order to reconstruct the
state of shared memory,
each processor must view
writes to shared memory
by other processors as
asynchronous events. We
therefore need to preserve
the order of execution
between the processors.”

Lamport et al: Time, Locks and the Ordering
of Events in a Distributed System
CACM ‘78

VEE ‘08
“In order to reconstruct the state
of shared memory, each
processor must view writes to
shared memory by other
processors as asynchronous
events. We therefore need to
preserve the order of execution
between the processors.”
“Only reads and writes to shared
memory need to be ordered with
respect to each other"

O’Callahan & Choi: Hybrid Dynamic Data
Race Detection
PPoPP ‘03
“Unfortunately violations of
the lockset hypothesis are not
always programming errors.
One common example is
programs which use
channels to pass objects
between threads in the style of
CSP [15].
In such programs thread
synchronization and mutual
exclusion are accomplished by
explicit signaling between
threads.”

Lockset - ensure >1 lock always protects shared me
Doesn’t need to observe a bug happening to find a race
False positives if different synchronization techniques used
+
-
Happens-Before - shared accesses must be separated by a sync
Causality def’n doesn’t require locks: fewer false positives
Observing races depends on execution order
Slower to run in practice than Lockset
+
-
-

Serebryany & Iskhodzhanov: ThreadSanitizer:
data race detection in practice WBIA ‘12

a hybrid approach
Full happens-before tracking for
synchronize operations
Lockset approach to determine
whether causally-related locksets are
non-empty

what have we seen
Time-Traveling Debugging for
reconstructing previous program state
JIT instrumentation for efficient control
flow modification
Shadow memory for efficient memory
introspection

the trace is out there
Novel and experimental hardware
Query languages and storage for large
trace data
Trace visualization techniques

Thanks
Slides, Citations,
and Acknowledgments:
https://github.com/dijkstracula/QConNYC2016
Nathan Taylor, Fastly
http://nathan.dijkstracula.net
@dijkstracula

Beyond Breakpoints: A Tour of Dynamic Analysis

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Beyond Breakpoints: A Tour of Dynamic Analysis (20)

More from Fastly (20)

Recently uploaded (20)

Beyond Breakpoints: A Tour of Dynamic Analysis