Pin: Intel’s Dynamic Binary
Instrumentation Engine
Pin Tutorial
Intel Corporation
Presented By:
Tevi Devor
1
CGO ISPASS 2012
Software & Services Group
Legal Disclaimer
INFORMATION IN THIS DOCUMENT IS PROVIDED “AS IS”. NO LICENSE, EXPRESS OR IMPLIED, BY
ESTOPPEL OR OTHERWISE, ANY INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS
DOCUMENT. INTEL ASSUMES NO LIABILITY WHATSOEVER AND INTEL DISCLAIMS ANY EXPRESS OR
IMPLIED WARRANTY, RELATING TO THIS INFORMATION INCLUDING LIABILITY OR WARRANTIES
RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF
ANY PATENT, COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT.
Performance tests and ratings are measured using specific computer systems and/or components and
reflect the approximate performance of Intel products as measured by those tests. Any difference in
system hardware or software design or configuration may affect actual performance. Buyers should
consult other sources of information to evaluate the performance of systems or components they are
considering purchasing. For more information on performance tests and on the performance of Intel
products, reference www.intel.com/software/products.
Intel and the Intel logo are trademarks of Intel Corporation in the U.S. and other countries.
*Other names and brands may be claimed as the property of others.
Copyright © 2010. Intel Corporation.
Software & Services Group
2
Optimization Notice
Intel compilers, associated libraries and associated development tools may include or utilize options
that optimize for instruction sets that are available in both Intel and non-Intel microprocessors (for
example SIMD instruction sets), but do not optimize equally for non-Intel microprocessors. In addition,
certain compiler options for Intel compilers, including some that are not specific to Intel microarchitecture, are reserved for Intel microprocessors. For a detailed description of Intel compiler
options, including the instruction sets and specific microprocessors they implicate, please refer to the
“Intel Compiler User and Reference Guides” under “Compiler Options." Many library routines that are
part of Intel compiler products are more highly optimized for Intel microprocessors than for other
microprocessors. While the compilers and libraries in Intel compiler products offer optimizations for
both Intel and Intel-compatible microprocessors, depending on the options you select, your code and
other factors, you likely will get extra performance on Intel microprocessors.
Intel compilers, associated libraries and associated development tools may or may not optimize to the
same degree for non-Intel microprocessors for optimizations that are not unique to Intel
microprocessors. These optimizations include Intel® Streaming SIMD Extensions 2 (Intel® SSE2),
Intel® Streaming SIMD Extensions 3 (Intel® SSE3), and Supplemental Streaming SIMD Extensions 3
(Intel SSSE3) instruction sets and other optimizations. Intel does not guarantee the availability,
functionality, or effectiveness of any optimization on microprocessors not manufactured by
Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel
microprocessors.
While Intel believes our compilers and libraries are excellent choices to assist in obtaining the best
performance on Intel and non-Intel microprocessors, Intel recommends that you evaluate other
compilers and libraries to determine which best meet your requirements. We hope to win your business
by striving to offer the best performance of any compiler or library; please let us know if you find we do
not.
Notice revision #20110307
Software & Services Group
3
Agenda
• Part1: Introduction to Pin
• Part2: Larger Pin tools and writing efficient Pin
tools
• Part3: Deeper into Pin API
• Part4: Advanced Pin API
• Part5: Performance #s
Software & Services Group
4
Part1: Introduction to Pin
• Dynamic Binary Instrumentation
• Pin Capabilities
• Overview of how Pin works
• Sample Pin Tools
Software & Services Group
5
What Does “Pin” Stand For?
• Three Letter Acronyms @ Intel
– TLAs
– 263 possible TLAs
– 263 -1 are in use at Intel
– Only 1 is not approved for use at Intel
– Guess which one:
• Pin Is Not an acronym
• Pin is based on the post link optimizer Spike
– Pin is a small Spike
– Spike is EOL
http://www.cgo.org/cgo2004/papers/01_82_luk_ck.pdf
Software & Services Group
6
Which one of these
tevi.devor@intel.com
people is the Pin Performance Guru?
Software & Services Group
7
Instrumentation
A technique that inserts code into a program to collect
run-time information
 Program analysis : performance profiling, error detection,
capture & replay
 Architectural study : processor and cache simulation, trace
collection
• Source-Code Instrumentation
• Static Binary Instrumentation
• Dynamic Binary Instrumentation
 Instrument code just before it runs (Just In Time – JIT)
– No need to recompile or re-link
– Discover code at runtime
– Handle dynamically-generated code
Pin istoarunning
dynamic
binary
– Attach
processes
instrumentation engine
Software & Services Group
8
Advantages of Pin Instrumentation
• Programmable Instrumentation:
– Provides rich set of APIs to write, in C,C++,assembly, your own
instrumentation tools, called PinTools
– APIs are designed to maximize ease of use
– abstract away the underlying instruction set idiosyncrasies
• Multiplatform:
• Robust:
–
–
–
–
OSs
Windows, Linux
Architectures
IA-32, Intel64
Instruments real-life applications: Database, web browsers, …
Instruments multithreaded applications
Supports signals and exceptions, self modifying code…
If you can Run it – you can Pin it
• Efficient:
– Applies compiler optimizations on instrumentation code
Pin can be used to instrument all the user level code
in an application
Software & Services Group
9
Pin Instrumentation Capabilities
• Use Pin APIs to write PinTools that:
– Replace application functions with your own
– Call the original application function from within your replacement
function
– Fully examine any application instruction, insert a call to your
instrumenting function to be executed whenever that instruction
executes
– Pass parameters to your instrumenting function from a large set of
supported parameters
–
–
–
–
Register values (including IP), Register values by reference (for modification)
Memory addresses read/written by the instruction
Full register context
….
– Track function calls including syscalls and examine/change
arguments
– Track application threads
– Intercept signals
If Pina doesn’t
have it, you don’t want it
– Instrument
process tree
– Many other capabilities…
Software & Services Group
10
Usage of Pin at Intel
• Profiling and analysis products
– Intel Parallel Studio
– Amplifier (Performance Analysis)
– Lock and waits analysis
– Concurrency analysis
– Inspector (Correctness Analysis)
– Threading error detection (data race and deadlock)
– Memory error detection
GUI
Algorithm
PinTool
Pin
• Architectural research and enabling
– Emulating new instructions (Intel SDE)
– Trace generation
– Branch prediction and cache modeling
Software & Services Group
11
Example Pin-tools
Cache
Simulation
CMP$IM
MT Workload
Capture
& Deterministic
Replay
PinPlay
Simulation
Region
Selection
PinPoints
Instruction
Emulation
(new instructions)
SDE
Pin
Trace
Generation
pinLIT
SDE: http://software.intel.com/en-us/articles/intel-software-development-emulator
CMP$IM: http://www-mount.ece.umn.edu/~jjyi/MoBS/2008/program/02A-Jaleel.pdf
PinPlay: Paper presented at CGO2010 http://www.cgo.org/cgo2010/program.html
Software & Services Group
12
Pin Usage Outside Intel
• Popular and well supported
– 30,000+ downloads, 400+ citations
• Free DownLoad
– www.pintool.org
– Includes: Detailed user manual, source code for 100s of
Pin tools
• Pin User Group (PinHeads)
– http://tech.groups.yahoo.com/group/pinheads/
– Pin users and Pin developers answer questions
Software & Services Group
13
Launcher Process
PIN.EXE
Count
258743109
pin.exe
–tInvocation
inscount.dll
– gzip.exe input.txt
Pin
gzip.exe input.txt
Read a at
Trace
Application
Starting
firstfrom
application
IP Read
Code
PinTool
a Trace from Application
Codethat counts application
Source
Trace
exit branch
is
Start PINVM.DLL
Execution
of Trace
ends
instructions
executed, prints Count
Jit
adding
instrumentation
code
modified
to instrumentation
directly
branch
to
running
Jit
it,it,
adding
code
at Jit
endnext
LoadTrace to
from
inscount.dll
Call
into
PINVM.DLL
Destination
from
inscount.dll
Load
PINVM.DLL
inscount.dll
(firstAppIp,
trace
Encode
the
jitted
trace
the
and
run into
its
“inscount.dll”)
Encode
the
trace
theinto
Code
Code
Cache
Pass
in
app
IP
of
Trace’s
target
main()
Cache
Launcher
Execute Jitted code
WriteProcessMemory(BootRoutine,
BootData)
Resume
atand
BootRoutine
GetContext(&firstAppIp)
Inject
Pin BootRoutine
Data intosuspended)
application
SetContext(BootRoutineIp)
CreateProcess
(gzip.exe,
input.txt,
Boot Routine +
Data:
firstAppIp,
“Inscount.dll”
First
app
IP
Application Process
PIN.LIB
System Call
Dispatcher
Event
Dispatcher
Encoder
PINVM.DLL
Decoder
Application
Code and
Data
inscount.dll
Code
Cache
Thread
Dispatcher
NTDLL.DLL
app Ip of
Trace’s
target
Windows kernel
Software & Services Group
14
All code in this presentation is covered
by the following:
•
•
/*BEGIN_LEGAL
Intel Open Source License
•
•
•
•
•
Copyright (c) 2002-2010 Intel Corporation. All rights reserved.
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer. Redistributions
in binary form must reproduce the above copyright notice, this list of
conditions and the following disclaimer in the documentation and/or
other materials provided with the distribution. Neither the name of
the Intel Corporation nor the names of its contributors may be used to
endorse or promote products derived from this software without
specific prior written permission.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE INTEL OR
ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
END_LEGAL */
Software & Services Group
15
Instruction Counting Tool (inscount.dll)
#include "pin.h"
UINT64 icount = 0;
10pts:
For
20pts:
Instrumentation
routinethis
Analysis
routine
INS
is only
valid inside
the
Which
function
isroutine
executed
Called
during
of INS
Executes
eachjitting
time
jitted
Instrumentation
analysis
routine?
more
often?
INStruction
executes
void docount() { icount++; }
Execution time routine
void Instruction(INS ins, void *v) Jitting time
{
INS_InsertCall(ins, IPOINT_BEFORE,
(AFUNPTR)docount, IARG_END);
}
routine: Pin CallBack
void Fini(INT32 code, void *v)
{ std::cerr << "Count " << icount << endl; }
•
int main(int argc, char * argv[])
{
PIN_Init(argc, argv);
INS_AddInstrumentFunction(Instruction, 0);
PIN_AddFiniFunction(Fini, 0);
PIN_StartProgram(); // Never returns
return 0;
}
•
•
•
switch to pin stack
save registers
call docount
restore registers
inc icount
switch
to app stack
sub
$0xff, %edx
inc icount
cmp
%esi, %edx
save eflags
inc icount
restore eflags
jle
<L1>
inc icount
mov
0x1, %edi
Software & Services Group
16
Instrumentation vs. Analysis
•Concepts borrowed from the ATOM tool:
•Instrumentation routines define where
instrumentation is inserted
– e.g., before instruction
C Occurs when an instruction is being jitted
•Analysis routines define what to do when
instrumentation is activated
– e.g., increment counter
C Occurs every time an instruction is
executed
Software & Services Group
17
Launcher Process
pin.exe –t inscount.dll – gzip.exe input.txt
PIN.EXE
Read a Trace from Application Code
Jit it, adding instrumentation code
from inscount.dll
Launcher
Encode the Jitted trace into the
Code Cache
Boot Routine +
Data:
firstAppIp,
“Inscount.dll”
First
app
IP
Application Process
PIN.LIB
System Call
Dispatcher
Event
Dispatcher
Encoder
PINVM.DLL
Decoder
Application
Code and
Data
inscount.dll
Code
Cache
Thread
Dispatcher
NTDLL.DLL
Windows kernel
Software & Services Group
18
Trace
Trace
BBL#1
TK
FT
’
BBL#1
BBL#2
BBL#4
’
BBL#2
TK
BBL#3
Original
code
BBL# 5
Early Exit via Stub
BBL# 6
Early Exit via Stub
Trace Exit via Stub
FT
BBL# 7
BBL#3
• Trace: A sequence of continuous instructions, with
one entry point
• BBL: has one entry point and ends at first control
transfer instruction
Software & Services Group
19
ManualExamples/inscount2.cpp
#include "pin.H"
UINT64 icount = 0;
void PIN_FAST_ANALYSIS_CALL docount(INT32 c) { icount += c; }
void Trace(TRACE trace, void *v){// Pin Callback
for(BBL bbl = TRACE_BblHead(trace);
BBL_Valid(bbl);
bbl = BBL_Next(bbl))
BBL_InsertCall(bbl, IPOINT_ANYWHERE,
(AFUNPTR)docount, IARG_FAST_ANALYSIS_CALL,
IARG_UINT32, BBL_NumIns(bbl),
IARG_END);
}
void Fini(INT32 code, void *v) {// Pin Callback
fprintf(stderr, "Count %lld\n", icount);
}
int main(int argc, char * argv[]) {
PIN_Init(argc, argv);
TRACE_AddInstrumentFunction(Trace, 0);
PIN_AddFiniFunction(Fini, 0);
PIN_StartProgram();
return 0;
}
Software & Services Group
20
2
22
40
37
APP IP
0x77ec4600
0x77ec4603
0x77ec4609
0x77ec460d
cmp
jz
movzx
call
rax, rdx
0x77f1eac9
ecx, [rax+0x2]
0x77ef7870
Compiler
generated code
for docount
Application
Inlined
by PinTrace
r14 allocated by
Pin
How many BBLs in this
trace?
r15 allocated by
Pin
Points to
per-thread spill
area
21
20
58
2
9
52
29
57
37
50
30
12
40
22
61
17
0x001de0000
0x001de000a
0x001de0015
0x001de0018
0x001de001e
0x001de0028
0x001de002c
0x001de002e
0x001de0031
0x001de0039
0x001de003d
0x001de0048
0x001de004c
0x001de0051
0x001de0052
mov r14, 0xc5267d40
//inscount2.docount
add [r14], 0x2
//inscount2.docount
0x77ec4600 cmp rax, rdx
jz 0x1deffa0 L1
//patched in future
mov r14, 0xc5267d40
//inscount2.docount
mov [r15+0x60], rax
lahf
save
status
seto al
flags
mov [r15+0xd8], ax
mov rax, [r15+0x60]
add [r14], 0x2
//inscount2.docount
0x77ec4609 movzx edi, [rax+0x2] //ecx alloced to edi
push 0x77ec4612
//push retaddr
nop
jmp 0x1deffd0 L2
//patched in future
L2:
24 0x001deffd0 mov [r15+0x40], rsp // save app rsp
34 0x001deffd4 mov rsp, [r15+0x2d0] // switch to pin stack
66 0x001deffdb call [0x2f000000]// call VmEnter
// data used by VmEnter – pointed to by return-address of call
0x001deffe8_svc(VMSVC_XFER)
0x001defff0_sct(0x00065fb60)
// current register mapping
0x001defff8_iaddr(0x077ef7870)
// app target IP of
// call at 0x77ec460d
L1:
41 0x001deffa0 mov [r15+0x40], rsp // save app rsp
63 0x001deffa4 mov rsp, [r15+0x2d0] // switch to pin stack
56 0x001deffab call [0x2f000000]
// call VmEnter
// data used by VmEnter – pointed to by return-address of call
0x001deffb8_svc(VMSVC_XFER)
0x001deffc0_sct(0x00065f998) // current register mapping
Software & Services Group
0x001deffc8_iaddr(0x077f1eac9)// app target IP of jz at 0x77ec4603
SimpleExamples/inscount2_mt.cpp
#include "pin.H"
INT32 numThreads = 0;
const INT32 MaxNumThreads = 10000;
struct THREAD_DATA
Why is there NO synchronization?
{
UINT64 _count;
UINT8 _pad[56]; // guess why? }icount[MaxNumThreads];
// Analysis routine
VOID PIN_FAST_ANALYSIS_CALL docount(ADDRINT c, THREADID tid) { icount[tid]._count += c;}
// Pin Callback
VOID ThreadStart(THREADID threadid, CONTEXT *ctxt, INT32 flags, VOID *v){numThreads++;}
VOID Trace(TRACE trace, VOID *v) { // Jitting time routine: Pin Callback
for (BBL bbl = TRACE_BblHead(trace); BBL_Valid(bbl); bbl = BBL_Next(bbl))
BBL_InsertCall(bbl, IPOINT_ANYWHERE, (AFUNPTR)docount, IARG_FAST_ANALYSIS_CALL,
IARG_UINT32, BBL_NumIns(bbl), IARG_THREAD_ID, IARG_END);
}
VOID Fini(INT32 code, VOID *v){// Pin Callback
for (INT32 t=0; t<numThreads; t++)
printf ("InsCount[of thread#%d]= %d\n",t,icount[t]._count);
}
int main(int argc, char * argv[])
{
PIN_Init(argc, argv);
for (INT32 t=0; t<MaxNumThreads; t++) {icount[t]._count = 0;}
PIN_AddThreadStartFunction(ThreadStart, 0);
TRACE_AddInstrumentFunction(Trace, 0);
PIN_AddFiniFunction(Fini, 0);
PIN_StartProgram(); return 0;
}
Software & Services Group
22
Multi-Threading
• Pin supports multi-threading
– Application threads execute jitted code including
instrumentation code (inlined and not inlined), without any
serialization introduced by Pin
– Instrumentation code can use Pin and/or OS synchronization
constructs to introduce serialization if needed.
– Will see examples of this in Part4
– Pin provides APIs for thread local storage.
– Will see examples in Part3
– Pin callbacks are serialized
– Jitting is serialized
– Only one application thread can be jitting code at any time
Software & Services Group
23
Memory Read Logger Tool
#include "pin.h“#include <map>
std::map<ADDRINT, std::string> disAssemblyMap;
VOID ReadsMem (ADDRINT applicationIp, ADDRINT memoryAddressRead, UINT32 memoryReadSize) {
printf ("0x%x %s reads %d bytes of memory at 0x%x\n",
applicationIp, disAssemblyMap[applicationIp].c_str(),
memoryReadSize, memoryAddressRead);}
VOID Instruction(INS ins, void * v) {// Jitting time routine
// Pin Callback
if (INS_IsMemoryRead(ins))
{
Pin does full register
allocation during jitting.
disAssemblyMap[INS_Address(ins)] = INS_Disassemble(ins);
At the start
of jitting aReadsMem,
INS_InsertCall(ins, IPOINT_BEFORE,
(AFUNPTR)
trace, all registers accessed
IARG_INST_PTR,//
applicationcode,
IP
by the application
are
considered
to
be
virtual
IARG_MEMORYREAD_EA,
registers.
IARG_MEMORYREAD_SIZE,
IARG_BRANCH_TAKEN
IARG_BRANCH_TARGET_ADDR
Many other
IARGs
IARG_END);
IARG_REG_VALUE,
IARG_REG_REFERENCE,
<REG>
Work
in progress:
Pineach
has of these
Pin<REG>
allocates
IARG_CONTEXT
IARG_FUNCARG_ENTRYPOINT_VALUE,
<ARG#>
}
} registers
determined
that it
virtual
to either:
Is the branch taken (when executed)
The target address of a branch
can
overwrite
ecx
The
A “pointer”
value
of register
to awhen
register
REG
A (RTN
physical
register.
instruction
executed
(including
IARG_MAKE_ME_A_COFFEE
The value of ARGUMENT
#
of
the
function
instrumentation)
Handle
to
the
full
register
context
indirect branches)
int main(int argc,
char * argv[]) {
of the executing threadOr a per-thread register
spill area pointed to by
PIN_Init(argc, argv);
ebx (IA-32)
INS_AddInstrumentFunction(Instruction, 0);
r15 (Intel64)
PIN_StartProgram(); }
Switch to pin stack
push 4
push %eax
push 0x7f083de
call ReadsMem
Pop args off pin stack
Switch back to app stack
• inc DWORD_PTR[%eax]
Switch to pin stack
push 4
lea %ecx,[%esi]0x8
push %ecx
push 0x7f083e4
call ReadsMem
Pop args off pin stack
Switch back to app stack
• inc DWORD_PTR[%esi]0x8
Software & Services Group
24
Malloc Wrapping
#include "pin.H"
void * MallocWrapper( CONTEXT * ctxt, AFUNPTR pf_malloc, size_t size)
{ // Simulate out-of-memory every so often
void * res;
if (TimeForOutOfMem())
return (NULL);
PIN_CallApplicationFunction(ctxt, PIN_ThreadId(),
CALLINGSTD_DEFAULT, pf_malloc,
PIN_PARG(void *), &res, PIN_PARG(size_t), size);
return res; }
VOID ImageLoad(IMG img, VOID *v) { // Pin callback. Registered by IMG_AddInstrumentFunction
if (strstr(IMG_Name(img).c_str(), "libc.so") ||
The ImageLoad callback is called for each image
(exe, shared library) loaded
into the process
strstr(IMG_Name(img).c_str(), "MSVCR80") || strstr(IMG_Name(img).c_str(),
"MSVCR90"))
{
RTN mallocRtn = RTN_FindByName(img, "malloc");
It is called before any code in the loaded image
executed
PROTO protoMalloc = PROTO_Allocate( PIN_PARG(voidis*),
CALLINGSTD_DEFAULT,
"malloc", PIN_PARG(size_t), PIN_PARG_END() );
This is referred to ahead-of-time-instrumentation
}}
RTN_ReplaceSignature(mallocRtn, AFUNPTR(MallocWrapper),
IARG_PROTOTYPE, protoMalloc,
IARG_CONST_CONTEXT,
IARG_ORIG_FUNCPTR,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
int main(int argc, CHAR *argv[]) {
PIN_InitSymbols();
PIN_Init(argc,argv));
IMG_AddInstrumentFunction(ImageLoad, 0);
PIN_StartProgram();
}
This is rather expensive
Also has 2 synchroniztion
points
Software & Services Group
25
#include "pin.H“
ADDRINT mallocReturnIp = 0;
Cheaper Malloc Wrapping
VOID BeforeMalloc(ADDRINT returnIp, ADDRINT size)
{
mallocReturnIp = returnIp;
printf ("(tool) call to malloc for %5d bytes
call will return to %p\n", size, returnIp); }
int CheckReturn(ADDRINT sp, ADDRINT returnRegVal)
{ // is this a return from malloc
return (mallocReturnIp == *(reinterpret_cast<ADDRINT *>(sp))); }
VOID ProcessReturnFromMalloc(ADDRINT ip, ADDRINT returnVal)
{
printf ("(tool) return from malloc at %p
malloc returns %p\n", ip, returnVal);
mallocReturnIp = 0;
}
Functionisisinlinable
notisinlinable
Function
IPOINT_AFTER
NOT
guaranteed
to
find
all iff
Functionreturns
will{ beTRUE/FALSE
called
Function
returns
The IfCall function
{
static void Instruction(INS ins, void *v)
if( INS_IsRet(ins))
INS_InsertIfCall(ins, IPOINT_BEFORE, (AFUNPTR)CheckReturn,
(CheckReturn)
IARG_REG_VALUE, REG_STACK_PTR, IARG_END);
Returns TRUE
INS_InsertThenCall(ins, IPOINT_BEFORE, (AFUNPTR)ProcessReturnFromMalloc,
IARG_INST_PTR, IARG_FUNCRET_EXITPOINT_VALUE, IARG_END);
}}
VOID ImageLoad(IMG img, VOID *v) { // Pin callback. Registered by IMG_AddInstrumentFunction
if (strstr(IMG_Name(img).c_str(), "libc.so") || strstr(IMG_Name(img).c_str(), "MSVCR80”)) {
RTN mallocRtn = RTN_FindByName(img, "malloc");
RTN_Open(mallocRtn);
RTN_InsertCall(mallocRtn, IPOINT_BEFORE, AFUNPTR(BeforeMalloc),
IARG_RETURN_IP, IARG_FUNCARG_ENTRYPOINT_VALUE, 0, IARG_END);
RTN_Close(mallocRtn);
}}
int main(int argc, CHAR *argv[]) { PIN_InitSymbols();
26
IMG_AddInstrumentFunction(ImageLoad, 0);
INS_AddInstrumentFunction(Instruction, 0);
PIN_StartProgram();
}
PIN_Init(argc,argv);
Software & Services Group
Pin Probe-Mode
• Probe mode is a method of using Pin to instrument at the
function level only. Wrap, Replace, call Analysis function
before/after.
• Replacement or Wrapping function can call the replaced
(original) function.
• The application and the replacement routine are run natively
(not Jitted).
–
–
–
–
–
Faster than Jit-mode
Puts more responsibility on the tool writer.
Probes can only be placed on RTN boundaries
Must be inserted within the Image load callback.
Pin will automatically remove the probes when an image is unloaded.
• Many of the PIN APIs that are available in JIT mode are not
available in Probe mode.
Software & Services Group
27
A Sample Probe
– A probe is a jump instruction that overwrites
original instruction(s) in the application
– Instrumentation invoked with probes
– Pin copies/translates original bytes so probed (replaced)
functions can be called from the replacement function
Copy of entry point with
0x50000004: push
0x50000005: mov
0x50000007: push
0x50000008: push
0x50000009: jmp
original bytes:
%ebp
%esp,%ebp
%edi
%esi
0x400113d9
0x41481064: push %ebp
Entry
Original
point
function
overwritten
entrywith
point:
probe:
0x400113d4:
0x400113d5:
0x400113d7:
0x400113d8:
0x400113d9:
jmp
push
mov
push
push
push
0x41481064
%ebp
%esp,%ebp
%edi
%esi
%ebx
// tool wrapper func
::::::::::::::::::::
0x414827fe: call 0x50000004 // call original func
Software & Services Group
28
Malloc Wrapping Probe-Mode
#include "pin.H"
void * MallocWrapper(AFUNPTR pf_malloc,
size_t size)
{ // Simulate out-of-memory every so often
void * res;
if (TimeForOutOfMem())
return (NULL);
res = pf_malloc(size);
return res; }
VOID ImageLoad (IMG img, VOID *v) {
if (strstr(IMG_Name(img).c_str(), "libc.so") ||
strstr(IMG_Name(img).c_str(), "MSVCR80") || strstr(IMG_Name(img).c_str(), "MSVCR90"))
{
RTN mallocRtn = RTN_FindByName(img, "malloc");
if ( RTN_Valid(mallocRtn) &&
RTN_IsSafeForProbedReplacement(mallocRtn) )
{
PROTO proto_malloc = PROTO_Allocate(PIN_PARG(void *), CALLINGSTD_DEFAULT, "malloc",
PIN_PARG(size_t), PIN_PARG_END() );
} }}
RTN_ReplaceSignatureProbed (mallocRtn,
AFUNPTR(MallocWrapper),
IARG_PROTOTYPE, proto_malloc,
IARG_ORIG_FUNCPTR,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
int main(int argc, CHAR *argv[]) {
PIN_InitSymbols();
PIN_Init(argc,argv));
IMG_AddInstrumentFunction(ImageLoad, 0);
PIN_StartProgramProbed();
}
Software & Services Group
29
SDE
•SDE: A fast functional simulator for
applications with new instructions
–New instructions have been defined
–Compiler generates code with new
instructions
–What can be used to run the apps with
the new instructions?
–Use PinTool that emulates new instructions.
– vmovdqu ymm?, mem256
vmovdqu mem256, ymm?
– 16 new 256 bit ymm registers
– Read/Write ymm register from/to memory.
Software & Services Group
30
Launcher Process
PIN.EXE
pin.exe -t sde.dll -- gzip.exe input.txt
Decoder
New
can decode
is replaced
new
with
Readinstruction
a Trace
from
Application
Code
instructions
call
toadding
emulation
function in the
tool
Jit it,
instrumentation
code
gzip.exe compiled with
from sde.dll
compiler that generates new
Host
CPU
can
NOTtrace
execute
Encode
the
Jitted
intothem
the instructions
Code Cache
Launcher
Execute it
Boot Routine +
Data:
firstAppIp,
“sde.dll”
First
app
IP
Application Process
PIN.LIB
System Call
Dispatcher
Event
Dispatcher
Encoder
PINVM.DLL
Decoder
Application
Code and
Data
sde.dll
Code
Cache
Thread
Dispatcher
NTDLL.DLL
Windows kernel
Software & Services Group
31
#include "pin.H"
sde_emul.dll Schema
VOID EmVmovdquMem2Reg(unsigned int ymmDstRegNum, ADDRINT * ymmMemSrcPtr) {
PIN_SafeCopy(ymmRegs[ymmDstRegNum], ymmMemSrcPtr, 32); }
VOID EmVmovdquReg2Mem(int ymmSrcRegNum, ADDRINT * ymmMemDstPtr) {
PIN_SafeCopy(ymmMemDstPtr, ymmRegs[ymmRegNum], 32); }
VOID Instruction(INS ins, VOID *v) {
switch (INS_Opcode(ins)
{
:::::
case XED_ICLASS_VMOVDQU:
if (INS_IsMemoryRead(ins))
// vmovdqu ymm? <= mem256
INS_InsertCall(ins, IPOINT_BEFORE,
(AFUNPTR)EmVmovdquMem2Reg,
IARG_UINT32, REG(INS_OperandReg(ins, 0)) - REG_YMM0,
IARG_MEMORYREAD_EA,
IARG_END);
else if (INS_IsMemoryWrite(ins)) // vmovdqu mem256 <= ymm?
INS_InsertCall(ins, IPOINT_BEFORE,
(AFUNPTR)EmVmovdquReg2Mem,
IARG_UINT32, REG(INS_OperandReg(ins, 1)) - REG_YMM0,
IARG_MEMORYWRITE_EA,
IARG_END);
INS_DeleteIns(ins); //Processor does NOT execute this instruction
break;
}}
int main(int argc, CHAR *argv[]) {
PIN_Init(argc,argv));
INS_AddInstrumentFunction(Instruction, 0);
PIN_StartProgram(); }
Software & Services Group
32
fork
gzip
(Injectee)
Pin (Injectee)
Pin stack
Code to Save
MiniLoader
Code to Save
Pin Code
gzip
Code and
and
Data
Data
exitLoop = FALSE;
Linux
Invocation+Injection
pin –t
inscount.so
– gzip input.txt
gzip
input.txt
Ptrace TraceMe
Child
(Injector)
PinTool that counts application
while(!exitLoop){}
instructions executed, prints Count
Ptrace Injectee – Injectee Freezes
at end
MiniLoader
Injectee.exitLoop = TRUE;
MiniLoader
Ptrace continue (unFreezes Injectee)
execv(gzip);
// Injectee Freezes
Execution of Injector
resumes after execv(gzip)
in Injectee completes
Pin Code and
Data
Ptrace Copy (save, gzip.CodeSegment, sizeof(MiniLoader))
PtraceGetContext (gzip.OrigContext)
PtraceCopy (gzip.CodeSegment, MiniLoader, sizeof(MiniLoader))
MiniLoader
IP
Pin Code and
Data
Ptrace continue@MiniLoader (unFreezes Injectee)
MiniLoader loads Pin+Tool,
allocates Pin stack
Kill(SigTrace, Injector):
Freezes until Ptrace Cont
Wait for MiniLoader
complete (SigTrace from
Injectee)
gzip OrigCtxt
gzip OrigCtxt
Ptrace Copy (gzip.CodeSegment, save, sizeof(MiniLoader))
Ptrace Copy (gzip.pin.stack, gzip.OrigCtxt, sizeof (ctxt))
Ptrace SetContext (gzip.IP=pin, gzip.SP=pin.Stack)
Code to Save
Inscount2.so
Ptrace Detach
Software & Services Group
33
Part1 Summary
• Pin is Intel’s dynamic binary instrumentation engine
• Pin can be used to instrument all user level code
–
–
–
–
–
–
Windows, Linux
IA-32, Intel64, IA64
Product level robustness
Jit-Mode for full instrumentation: Thread, Function, Trace, BBL, Instruction
Probe-Mode for Function Replacement/Wrapping/Instrumentation only.
Pin supports multi-threading, no serialization of jitted application nor of instrumentation code
• Pin API makes Pin Tools easy to write
– Presented 6 full Pin tools, each one fit on 1 ppt slide
• Popular and well supported
– 30,000+ downloads, 400+ citations
• Free DownLoad
– www.pintool.org
– Includes: Detailed user manual, source code for 100s of Pin tools
• Pin User Group
– http://tech.groups.yahoo.com/group/pinheads/
– Pin users and Pin developers answer questions
Software & Services Group
34
Part2: Larger Pin tools and writing
efficient Pin tools
Software & Services Group
35
CMP$im – A CMP Cache Simulation Pin Tool
WORK LOAD
Modeling an 8-core CMP using CMP$im
Instrumentation Routines
PIN
• ThreadID
• Address,
Size
• Access
Type
ThreadID, Address, Size, Access Type
DL1 DL1 DL1 DL1 DL1 DL1 DL1 DL1
L2
L2
L2
L2
L2
L2
L2
L2
LLC
LLC
INTERCONNECT
Cache model
• Params to
configure # cache
levels, size,
threads/cache etc
36
LLC
LLC
LLC
LLC
LLC
LLC
PRIVATE LLC/SHARED BANKED LLC
Software & Services Group
CMP$im author: Aamer.Jaleel@intel.com
ANALYSIS
ROUTINES
INSTR
ROUTINES
•VOID Instruction(INS ins, VOID *v)
•{
•
if( INS_IsMemoryRead(ins) ) // If instruction reads
•
// from memory
•
INS_InsertCall(ins,
•
IPOINT_BEFORE, (AFUNPTR)MemoryReference,
•
IARG_THREAD_ID, IARG_MEMORYREAD_EA,
•
IARG_MEMORYREAD_SIZE, IARG_UINT32,
•
ACCESS_TYPE_LOAD, IARG_END);
• if( INS_IsMemoryWrite(ins) ) // If instructions writes
•
// to memory
•
INS_InsertCall(ins, IPOINT_BEFORE,
•
(AFUNPTR) MemoryReference,
•
IARG_THREAD_ID, IARG_MEMORYWRITE_EA,
•
IARG_MEMORYWRITE_SIZE, IARG_UINT32,
•
ACCESS_TYPE_STORE, IARG_END);
•}
Pin Tool
MAIN
CMP$im – Instrument Memory
References
Software & Services Group
37
CMP$im – Analyze Memory
References
•VOID MemoryReference(
•
int tid, ADDRINT addrStart, int size, int type)
•{
•
for(addr=addrStart; addr<(addrStart+size);
•
addr+=LINE_SIZE)
•
LookupHierarchy( tid, FIRST_LEVEL_CACHE, addr, type);
•}
•VOID LookupHierarchy(
•
int tid, int level, ADDRINT addr, int accessType) {
•
result = cacheHier[tid][cacheLevel]->Lookup(
•
addr, accessType );
•
if( result == CACHE_MISS ) {
•
if( level == LAST_LEVEL_CACHE ) return;
•
if( IsShared(level) ) AcquireLock(&lock[level], tid);
•
LookupHierarchy(tid, level+1, addr, accessType);
•
ReleaseLock(&lock[level]);
•
}
•}
38
INSTR
ROUTINES
CacheHierarchy[MAX_NUM_THREADS][MAX_NUM_LEVELS];
MAIN
•CACHE_t
ANALYSIS
ROUTINES
•#include “cache_model.h”
Synchronization
Software & Services Group
point
Intel Thread Checker
•Detect data races
Paul Petersen,
Zhiqiang Ma
•Instrumentation
– Memory operations
– Synchronization operations
•Analysis
– Use dynamic history of lock acquisition and release
to form a partial order of memory references
[Lamport 1978]
– Unordered read/write and write/write pairs to same
location are races
Software & Services Group
39
a documented data race in the
art benchmark is detected
Software & Services Group
40
40
PinPlay : Workload capture and
deterministic replay
•Problem : Multi-threaded programs are inherently nondeterministic making their analysis, simulation,
debugging very challenging
•Solution: PinPlay : A Pin-based framework for
capturing an execution of multi-threaded program and
App and input
replaying it deterministically under Pin
not needed
Application
input
Application
Pin
logger
41
PinPlay
LOGS
Pin
Replayer
once we have
the log
Deterministic
replay on any
machine
Harish Patil & Cristiano Pereira
Software & Services Group
Joint work with Brad Calder, UCSD
Logging to provide deterministic
behavior
•Start with checkpoint: memory image of code and
data
•A thread is deterministic if every loads sees either:
– Data from original checkpoint
– Or a value computed and stored on the thread
•Potential non-determinism when a load sees a
memory location written by an external agent
– Another thread
– Or system call, DMA, etc.
•Log these values with timestamps
Software & Services Group
42
Applying multi-threaded tracing to
software tools
•
Debugging. Customer interested in debugging tools
derived from PinPlay
–
–
–
Capture bug at customer, bring home log to debug
Capture multi-threaded “heisenbug”, replay multiple times
How: combine PinPlay tracing with transparent debugging
PinPlay
LOGS
Pin
Replayer
Debug
Agent
Pin debug agent enables
custom debugger commands
debugger
Standard
protocol
Software & Services Group
43
Reducing Instrumentation
Overhead
Total Overhead = Pin Overhead + Pintool Overhead
~5% for SPECfp and ~50% for SPECint
Pin team’s job is to minimize this
Usually much larger than pin overhead
Pintool writers can help minimize this!
Software & Services Group
44
Reducing the Pintool’s Overhead
Pintool’s Overhead
Instrumentation
Routines
Overhead
Analysis
Routines
Overhead
+
x Work required in the
Analysis Routine
Frequency of calling
an Analysis Routine
Work required for transiting
to Analysis Routine
+
Work done inside
Analysis Routine
Software & Services Group
45
Reducing Work in Analysis Routines
•Key: Shift computation from analysis routines to
instrumentation routines whenever possible
•This usually has the largest speedup
Software & Services Group
46
Counting control flow edges
jne
60
40
100
ret
40
40
call
jmp
60
jne
1
Software & Services Group
47
Edge Counting: a Slower Version
•...
•void docount2(ADDRINT src, ADDRINT dst, INT32 taken)
•{
•
COUNTER *pedg = Lookup(src, dst);
•
pedg->count += taken;
Analysis
•}
•void Instruction(INS ins, void *v) {
Instrumentation
•
if (INS_IsBranchOrCall(ins))
•
{
•
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)docount2,
•
IARG_INST_PTR, IARG_BRANCH_TARGET_ADDR,
•
IARG_BRANCH_TAKEN, IARG_END);
•
}
•}
•...
Software & Services Group
48
Edge Counting: a Faster Version
•void docount(COUNTER* pedge, INT32 taken) {
•
pedg->count += taken;
•}
•void docount2(ADDRINT src, ADDRINT dst, INT32 taken) {
•
COUNTER *pedg = Lookup(src, dst);
•
pedg->count += taken;
Analysis
•}
•void Instruction(INS ins, void *v) {
Instrumentation
•
if (INS_IsDirectBranchOrCall(ins)) {
•
COUNTER *pedg = Lookup(INS_Address(ins),
•
INS_DirectBranchOrCallTargetAddress(ins));
•
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR) docount,
•
IARG_ADDRINT, pedg, IARG_BRANCH_TAKEN, IARG_END);
•
} else
•
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR) docount2,
•
IARG_INST_PTR, IARG_BRANCH_TARGET_ADDR,
•
IARG_BRANCH_TAKEN, IARG_END);
•}
•…
Software & Services Group
49
Analysis Routines: Reduce Call
Frequency
•Key: Instrument at the largest granularity
whenever possible
Instead of inserting one call per instruction
Insert one call per basic block or trace
Software & Services Group
50
Slower Instruction Counting
counter++;
sub $0xff, %edx
counter++;
cmp %esi, %edx
counter++;
jle <L1>
counter++;
mov $0x1, %edi
counter++;
add $0x10, %eax
Software & Services Group
51
Faster Instruction Counting
Counting at BBL level
Counting at Trace level
counter += 3
sub $0xff, %edx
sub
$0xff, %edx
cmp
%esi, %edx
cmp
%esi, %edx
jle <L1>
counter += 2
mov $0x1, %edi
jle
<L1>
add
add $0x10, %eax
counter += 5
$0x10, %eax
mov
$0x1, %edi
counter+=3
L1
Software & Services Group
52
Reducing Work for Analysis Transitions
•Reduce number of arguments to analysis
routines
– Inline analysis routines
– Pass arguments in registers
– Instrumentation scheduling
Software & Services Group
53
Reduce Number of Arguments
•Eliminate arguments only used for debugging
•Instead of passing TRUE/FALSE, create 2 analysis
functions
– Instead of inserting a call to:
Analysis(BOOL val)
– Insert a call to one of these:
AnalysisTrue()
AnalysisFalse()
– IARG_CONTEXT is very expensive (> 10
arguments)
– Use the new IARG_CONST_CONTEXT
Software & Services Group
54
Inlining
Not-inlinable
Inlinable
int docount1(int i) {
int docount0(int i) {
if (i == 1000)
x[i]++
x[i]++;
return x[i];
}
return x[i];
}
Not-inlinable
int docount2(int i) {
x[i]++;
printf(“%d”, i);
Not-inlinable
void docount3() {
for(i=0;i<100;i++)
x[i]++;
Pin will inline analysis functions into
return x[i];
}
jitted application
code
}
Software & Services Group
55
Inlining
•Use the –log_inline invocation switch to record inlining decisions in pin.log
pin –log_inline –t mytool – app
•Look in pin.log
Analysis function (0x2a9651854c) from mytool.cpp:53
INLINED
Analysis function (0x2a9651858a) from mytool.cpp:178 NOT INLINED
The last instruction of the first BBL fetched is not a ret instruction
•Look at source or disassembly of the function in mytool.cpp at line 178
0x0000002a9651858a
0x0000002a9651858b
0x0000002a9651858e
0x0000002a96518595
0x0000002a96518597
0x0000002a9651859e
0x0000002a965185a4
push rbp
mov rbp, rsp
mov rax, qword ptr [rip+0x3ce2b3]
inc dword ptr [rax]
mov rax, qword ptr [rip+0x3ce2aa]
cmp dword ptr [rax], 0xf4240
jnz 0x11
– The function could not be inlined because it contains a control-flow changing
instruction (other than ret)
Software & Services Group
56
Conditional Inlining
• Inline a common scenario where the analysis
routine has a single “if-then”
– The “If” part is always executed
– The “then” part is rarely executed
– Useful cases:
1. “If” can be inlined, “Then” is not
2. “If” has small number of arguments, “then” has many arguments
(or IARG_CONST_CONTEXT)
• Pintool writer breaks analysis routine into two:
– INS_InsertIfCall (ins, …, (AFUNPTR)doif, …)
– INS_InsertThenCall (ins, …, (AFUNPTR)dothen, …)
Software & Services Group
57
IP-Sampling (a Slower Version)
const INT32 N = 10000; const INT32 M = 5000;
INT32 icount = N;
VOID IpSample(VOID* ip) {
--icount;
if (icount == 0) {
fprintf(trace, “%p\n”, ip);
icount = N + rand()%M; //icount is between <N, N+M>
}
}
VOID Instruction(INS ins, VOID *v) {
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)IpSample,
IARG_INST_PTR, IARG_END);
}
Software & Services Group
58
IP-Sampling (a Faster Version)
INT32 CountDown() {
--icount;
inlined
return (icount==0);
}
VOID PrintIp(VOID *ip) {
fprintf(trace, “%p\n”, ip);
not inlined
icount = N + rand()%M; //icount is between <N, N+M>
}
VOID Instruction(INS ins, VOID *v) {
// CountDown() is always called before an inst is executed
INS_InsertIfCall(ins, IPOINT_BEFORE, (AFUNPTR)CountDown,
IARG_END);
// PrintIp() is called only if the last call to CountDown()
// returns a non-zero value
INS_InsertThenCall(ins, IPOINT_BEFORE, (AFUNPTR)PrintIp,
IARG_INST_PTR, IARG_END);
}
Software & Services Group
59
Using Liveness Information
• Use -xyzzy –liveness 1
–
App does NOT have exception handler that examines
registers that are dead to the application
–
Perf gain mainly from NOT having to save app flags
Software & Services Group
60
Jitting time
• Jitting is expensive
–
Takes far more time to jit an instruction than to execute
a jitted instruction
• Portions of a workload where very many IPs are
being jitted, and executed a small number of
times
–
–
Jitting time dominates execution time
E.g.
–
–
–
startup of a GUI app
Compiler compiling a non-large file
Vs Loop executing a large number of times
–
Jitting time is amortized over execution time
Software & Services Group
61
Optimizing Your Pintools Summary
• Baseline Pin has fairly low overhead for non-jitting
portions of workloads (~5-20%)
• Adding instrumentation can increase overhead
significantly, but you can help!
1. Move work from analysis to instrumentation
routines
2. Explore larger granularity instrumentation
3. Explore conditional instrumentation
4. Understand when Pin can inline instrumentation
Software & Services Group
62
OS Specifics
Software & Services Group
63
Windows-specific Challenges (1/2)
• Handling system calls
– Pin must intercept system calls to regain control of the application
on return from the system
– Pin must monitor system calls to notify instrumentation when DLLs
are loaded/unloaded, threads are created/terminated, etc.
– System call interface is undocumented
– System call numbers potentially change with each system
build
• Handling exceptions and asynchronous interruptions
– To maintain control and notify instrumentation about control flow
changes Pin must intercept all transitions from kernel to user mode
– Windows is not designed to have an independent agent
interposed between the kernel and application
– The kernel dispatches interruptions via (undocumented) entry
points in ntdll.dll
Software & Services Group
The main obstacle: direct interface between user-level code and Windows kernel is undocumented
64
Windows-specific Challenges (2/2)
• Injection
– PIN VMM is a DLL that must be loaded into the address space of the
application to get initial control of the process
– Windows is not designed for proprietary loader
– Common practice: intercept control at the entry point of the
application
– Instrumentation can not observe initialization procedures in
staticly linked application DLLs
– Injection presented in the introduction is referred to as Late
Injection
– It misses the initialization procedures in staticly linked application DLLs
– Early injection is not trivial
• Isolation of instrumentation from the application
– Instrumentation runs in the same process as the application it is
observing
– Enabling C run-time in the instrumentation causes sharing of system
libraries (e.g. kernel32.dll) and their state with the application
Pin minimizes
its dependence onPin
Windows
system services in order to maximize Software
observability&and
achieve better
Services
Group
– To
be transparent,
must
isolation
65
– Preserve original state of system resources
Pin Architecture in Windows
PIN.EXE
PIN.EXE
CreateProcess
Injection Helper
Launcher
Symbol Server
DBGHELP.DLL
CreateProcess
Application Process
Application
Code and Data
PINTOOL.DLL
Shared memory
PINVM.DLL - Virtual Machine Monitor
Startup time
System Call
Emulator
System Gate
JIT Compiler
Code
Cache
Event
Thread
Dispatcher
Dispatcher
NTDLL.DLL
KERNEL32.DLL
Windows Kernel
Software & Services Group
66
Injection
• Injection is the procedure for loading the PINVM.DLL into the
address space of an application and gaining control of execution
• Other systems hook the entry point of the application
– Too late: initialization procedures in application DLLs can not be
instrumented
• For maximum observability, Pin should inject itself into a new
process as early as possible, however…
• Pin depends on some basic system services so it is not possible to
load PINVM.DLL until the loader and kernel32.dll have initialized
• The optimal injection point: just after initialization of
kernel32.dll
– Injection presented in the introduction is referred to as Late
Injection
– It misses the initialization procedures in staticly linked application DLLs
Software & Services Group
67
Launch of the Instrumented Process
pin –t pintool.dll -- application.exe
PIN.EXE
Application Process
Debugging API
Pin Boot Routine
PIN.EXE
PINVM.DLL
PINTOOL.DLL
 Create (suspended) application process

Attach to the application as a debugger
APPLICATION.EXE

Run the application process until kernel32.dll
is loaded and initialized
APPLICATION.DLL

Detach from the application process

Copy Boot Routine into the application
process and set PC to start of the routine
Pin Boot Routine
68

Load and start Pin VMM

PINVM.DLL
Load the instrumentation tool

Instrument and execute the application
KERNEL32.DLL
NTDLL.DLL
Windows Kernel
Software & Services Group
All application instructions are executed under Pin control
Handling System Calls
• Pin must manage the execution of system calls
– To regain control when the system returns to user mode with a modified
thread context
– To monitor and handle some important system events
– Loading DLLs, creation and termination of threads and processes, etc.
• Pin intercepts system call instructions, not Win32 APIs
– Pin instruments all modules in the user space, including system libraries
– Some applications use native API (NTDLL interface) directly, bypassing
Win32 API
– Win32 API layer is very wide, while system call instructions are easy to
discover
• Three steps in managing system calls:
– Detect a system call and redirect control to VMM
– Execute the system call on behalf of the application
– Regain control when the kernel returns to user with a new context
– The system may interrupt system call execution by asynchronous calls to
application procedures
Software & Services Group
69
System Call Interception
• Pin detects system call instructions when it generates traces
in the code cache
– IA-32: sysenter and int 2E; Intel64: syscall; etc.
– This is a static analysis, so the overhead is low
• Pin executes system calls in VMM, not in the code cache emits jump to VMM instead of the system call instruction
– Enables flushing the code cache while a system call blocks in the kernel
– VM lock is NOT held during the actual syscall
• Some system calls may affect Pin’s internal state. To handle
them properly, Pin must know the corresponding system call
numbers
– Windows system call numbers are unpublished and potentially change
with each system build
– Pin discovers system call numbers dynamically, on the early stage of the
injection process
– We trace the corresponding NTDLL functions until a system call instruction is
reached and then read the system call number from the EAX/RAX register
Software & Services Group
70
System Call Execution
• The System Call Emulator executes all “known” system calls that
may affect the VMM state, e.g. memory mappings, creation and
termination of threads and processes, etc.
• The remaining, unknown system calls are forwarded to the
System Gate
– Per-thread procedure that transparently executes system calls and regains
control upon return or interruption
– Fills/spills original context before/after system calls
– Recovers original context (PC) when a system call is interrupted
Original Code
System Call Emulator
System Gate
Notify Pintool before system call
sysenter
Is “known” system call?
N
Y
Code Cache
jmp VMM
Execute/Emulate
ReturnFromSystemCall:
Switch to the application context
int 2e
Switch to the Pin context
jmp ReturnFromSystemCall
Notify Pintool after system call
71
Software
& Services
Group
System Gate executes system calls “blindly”, assuming that each
of them
can arbitrarily
modify context and control flow (if interrupted)
User Procedure Calls (UPC)
• UPC is a control transfer from the kernel to a user-level
procedure
• Asynchronous procedure call (APC)
– Asynchronous events: file I/O completion, timer expiration
– Thread initialization APC signals start of a new thread
• Callback
– Asynchronous Windows GUI message
• Exception
– Access violation, illegal instruction, divide by zero, etc.
• Asynchronous events are not delivered immediately, but wait
in queue until the application invokes an interruptible
(alertable) system call
• Pin must intercept UPCs to maintain control of the
application and recover the original interruption
context (visible to the application)
Software & Services Group
72
UPC Interception
• The kernel dispatches UPCs through entry points in NTDLL.DLL
• To intercept UPCs, Pin overwrites the NTDLL entry points with trampolines that
jump to the Event Dispatcher in Pin
• When a UPC is intercepted, Pin recovers original interruption context in the UPC
frame prepared by the kernel
– JIT Compiler recovers context of exceptions that occurred in the code cache
– System Call Emulator recovers context of interrupted system calls
Code
Cache
Translated
KiUserApcDispatcher
Recover original PC
in the
APC frame
Translated
KiUserCallbackDispatcher
Save PC of the
interrupted
system call
Translated
KiUserExceptionDispatcher
Recover original
context in the
exception frame
Pin VMM
UPC Dispatcher
APC
NTDLL.DLL
KiUserApcDispatcher
Callback
Exception
KiUserCallbackDispatcher
KiUserExceptionDispatcher
Windows kernel
Software & Services Group
Pin intercepts all control transfers from the kernel to the user mode
73
Exceptions (1/2)
• Unlike APCs and callbacks that are queued and delivered at the
next alertable system call, exceptions are synchronous events
• Exceptions do not necessarily cause abnormal termination of the
process – the application may expect and handle exceptions
• Pin must provide exception handlers with the same exception
information that accompanies exceptions in the native
application
– Exception context, code and exception-specific parameters
• From the Pin’s perspective, there are three kinds (sources) of
exceptions in Windows applications:
– An attempt to fetch an invalid or inaccessible instruction
– An attempt to execute a faulting instruction
– Software exceptions generated by the application
Software & Services Group
74
Exceptions (2/2)
• Decoder (fetcher) of instructions raises an exception if it
encounters an invalid or inaccessible instruction
– When the kernel delivers this exception back to the user mode, Pin skips the
context translation because it sees original PC in the exception context
• Other hardware exceptions occur in the code cache
– Recovery of the original exception context is nontrivial due to register
allocation
– Pin retranslates the interrupted trace to get the virtual-physical register
bindings at the faulting point
– Optimization: small cache of register bindings for frequent exceptions
• Other hardware exceptions occur in the tool code
– Pin APIs for tool to manage it’s exceptions
• Application can generate software exceptions using Win32 API
– The exception context represents an original application state
– Context translation is not needed
Software & Services Group
Pin delivers precise exceptions to applications handlers
75
Multithreading Support
• Pin instruments and runs all threads of the application
from the first to the last user-mode instruction
– Attaches to a new thread when the system delivers the thread
initialization APC
– Maintains control until the thread exits
– Intercepts threads created by remote processes
• Pin’s threading activities are transparent to the
application
– Pin VMM serializes some of its operations (e.g. JIT compilation), but
never executes code of the application under Pin locks
– Except for initialization phase, Pin never acquires locks in system
libraries, e.g. loader lock or process heap lock
– Each thread has a shadow stack that is used by Pin VMM and Pintool
Software & Services Group
76
Thread-Local State
• Key elements of the Pin’s thread-local state:
– Spill area keeps values of spilled virtual registers
– JIT-compiled traces need fast access to spilled register values
– Pin steals one physical register to point to the spilling area
– TEB state keeps original thread-local state of system libraries,
e.g. last Win32 error value, stack limit
– C run-time routines may access/modify these values
– Need to preserve the original state while running in Pin VMM or
PinTool
– System call state contains information about active and
interrupted system calls in the thread
– The information is used to restore the original context on return
from the system
• Pin steals one TLS slot from the application to enable
fast access to the thread-local data in Pin VMM
Software & Services Group
77
Thread Suspension and Context Manipulation
• A thread can suspend another thread and read/modify its context
– SuspendThread(), GetThreadContext(), SetThreadContext()
• Pin must emulate the corresponding system calls to avoid
deadlocks and transparency issues
– Target thread may hold a Pin lock
– The thread context is not original
– Suspended traces disable flushing the code cache
• Solution: Force a thread to leave the code cache and wait
until the thread reaches a safe point
Safe point = no locks, not in the code cache, accessible original context
– Unlink the suspended trace from successors and let it enter VMM
– Block the thread in the safe VMM point or in the System Gate
– Use thread-local data to store and access the original context
associated with the safe point
Software & Services Group
78
Linux-specific Challenges (1/3)
• Handling system calls
– Pin must intercept system calls to regain control of the application
on return from the system
– Pin must monitor system calls to notify instrumentation when DLLs
are loaded/unloaded, threads are created/terminated, etc.
– Some system calls may behave differently on different Linux
distributions.
• Signal handling
– Pin must identify whether the signal originated from the application,
the tool or Pin itself.
– Pin cannot seem to interfere with the applications signal
mask.
Software & Services Group
79
Linux-specific Challenges (2/3)
• Injection
– Pin relies on the ptrace system call for injection. The ptrace system
call has known bugs in several Linux version.
– Some platforms do not allow tracing a parent application by a child.
In such cases the application is run on the child -> pid is changed.
• Isolation of instrumentation from the application
– Instrumentation runs in the same process as the application it is
observing.
– Pin and the application share the same physical segment registers. In
probe mode, this restricts the libc versions allowed.
Software & Services Group
80
Linux-specific Challenges (3/3)
• GLIBC
– Pin must emulate several libc services for several reasons. For example:
– Pin may run before libc is initialized e.g. during injection.
– Pin’s libc may “get confused”. For example: libc’s getpid wrapper function incorporates
a cache. The first call to getpid actually calls the getpid system call but any
subsequent calls will access the cache. Upon a fork, the cache is invalidated and the
next call to getpid will again call the getpid system call. Since the process has two
copies of libc, only the application’s cache is invalidated and Pin’s copy is stale. –
Software & Services Group
81
Handling System Calls
• Pin must manage the execution of system calls
– Pin must maintain control all the time
– System calls are executed inside pin and return to the application
– In most cases the system call is executed without the pin VM lock
– Certain system calls are emulated by pin (see below)
• System call emulation
– Pin detects if a system call needs emulation.
– Pin needs to know the attributes of each memory page for SMC support
– Therefore all system calls related to memory are emulated by pin
– Signal related system calls are emulated
– Creating of new threads and new child processes
– Setting/getting of the TLS segment registers
– Thread and process termination
Software & Services Group
82
Signal Handling
• Pin registers its own signal handlers for all signals, and saves
the application’s handlers.
• Pin must handle both synchronous and asynchronous signals.
• Asynchronous signals:
– These signals may be delivered “at will” so Pin waits for safe point to deliver
them.
– When such a signal arrives, Pin’s internal handler registers this signal, unlinks the
current trace and resumes execution from the code cache.
– At the trace’s exit point, the executing thread jumps to the VM, thus transferring control
over to Pin. The VM checks if there are pending signals and calls the application’s
original signal handlers for these signals (jitting them).
• Synchronous signals:
– These signals must be delivered immediately.
– They may originate from the application, the tool or Pin itself.
– Pin’s internal handler is called, it determines the origin of the signal and propagates the
signal delivery to the tool and application is necessary.
– If signal is delivered to the application, the application’s signal handler is jitted.
Software & Services Group
83
Multithreading Support
• Pin instruments and runs all threads of the application
from the first to the last user-mode instruction
– Attaches to the thread upon the first user-space instruction
– Maintains control until the thread exits
• Pin’s threading activities are transparent to the
application
– The Pin VM serializes some of its operations (e.g. JIT compilation),
but never executes code of the application under Pin locks
– Each thread has a shadow stack that is used by the Pin VM and the
Pintool
– Pin and pintools are prohibited from using the pthread library due to
conflicts with some internal structures. Therefore Pin provides its own
APIs for thread creation and control.
Software & Services Group
84
84
Thread-Local Storage (Linux)
• JIT mode – segment virtualization
– TLS is accessed via the fs (64 bit) or gs (32 bit) segment
register.
– Both the application and Pin share this register, but expect
different values.
– Pin emulates the application’s usage of the fs/gs register thus
isolating the application’s TLS for Pin’s.
• Probe mode – no TLS usage by Pin
– Probe mode does not enable the method described above.
– Pin uses its own version of GLIBC which does not use TLS.
Software & Services Group
85
Isolation/Windows
• Pin Tools are compiled to use the static CRT
• Pin on Windows does not separate DLLs loaded by
the tool from the application DLLs - it uses the
same system loader.
– The tool should not load any DLL that can be shared with
the application.
– The tool should avoid static links to any common DLL,
except for those listed in PIN_COMMON_LIBS (see
source\tools\ms.flags file).
Software & Services Group
86
Isolation/Windows
• Pin on Windows guarantees safe usage of C/C++
run-time services in Pin tools, including indirect
calls to Windows API through C run-time library.
– Any other use of Windows API in Pin tool is not guaranteed
to be safe
• Pin uses some base types that conflict with
Windows types. If you use "windows.h", you may
see compilation errors. So do:
namespace WINDOWS { #include <windows.h> }
Software & Services Group
87
Isolation/Linux
• Pin is injected in to address space and has its own
copy of the dynamic loader and runtime libraries
(GLIBC, etc).
• Pin uses a small library of CRT for direct calls to
system calls.
• The process has a single signals table (shared
among all threads), pin manages an internal signal
table and emulate all the system calls related to
signals.
Software & Services Group
88
Isolation/Linux
• pthread functions cannot be called from an analysis
or replacement routine
• Pintools on Linux need to take care when calling
standard C or C++ library routines from analysis or
replacement functions
– because the C and C++ libraries linked into Pintools are
not thread-safe
Software & Services Group
89
Part3: Deeper into Pin API
• Agenda
– memtrace_simple tool
– membuffer_simple tool
– branchbuffer_simple tool
– Symbols DebugInfo
– Probe-Mode
– Multi-Threading
– Taint analysis
Software & Services Group
90
memtrace_simple
• Tool code collects pairs of {appIP, memAddr} of
memory accessing instructions into a per-thread
buffer.
– Process when no more room in buffer
Collect pairs
of {appIP,
memAddr}
into buffer
Reset the
collection to
start at the
beginning of
the buffer
No more
room in
Buffer,
process the
entries in the
buffer
Software & Services Group
91
memtrace_simple
• Tool code must
– Instrument each memory accessing instruction
– Determine where in the buffer the {appIP, memAddr} of the
instruction should be written
– Determine when the buffer becomes full
• Will instrument instructions on Trace level – i.e.
TRACE_AddInstrumentFunction(Trace, 0);
– Not all instructions in the trace will necessarily execute
each time trace is executed – because of early exits.
• Will try to allocate, in the buffer, maximum space
needed by trace at the trace start – if not enough
space => buffer is full
Software & Services Group
92
memtrace_simple
• Instrumentation code for each memory accessing
instruction in the trace will write it’s
{appIP, memAddr} pair to a constant offset from
the start of the trace in the buffer.
– Empty pairs (those instructions that were NOT executed)
will be denoted by having an appIP==0.
Software & Services Group
93
memtrace_simple
Trace
If
endOf(Previous)TraceReg
+ TotalSizeOccupiedByTraceInBuffer > endOfBufferReg
Then
Buffer
Call BufferFull
endOfTraceReg += TotalSizeOccupiedByTraceInBuffer
Early Exit
appIP
memAddr
Trace Exit
appIP
memAddr
endOfTraceReg
TotalSize
Occupied
ByTraceIn
Buffer
appIP
Non memory access ins
Instrumentation code for following memory access ins
Memory access ins
memAddr
endOfBufferReg
Software & Services Group
94
memtrace_simple
Trace
If
endOf(Previous)TraceReg
+ TotalSizeOccupiedByTraceInBuffer > endOfBufferReg
Then
Call BufferFull
endOfTraceReg += TotalSizeOccupiedByTraceInBuffer
Buffer
appIP
endOfTraceReg
memAddr
appIP
memAddr
Early Exit
appIP
memAddr
Trace Exit
Non memory access ins
Instrumentation code for following memory access ins
Memory access ins
endOfTraceReg
TotalSize
Occupied
ByTraceIn
Buffer
endOfBufferReg
Software & Services Group
95
memtrace_simple
• Tool will:
– iterate thru all INSs of the Trace
– Record which ones need to be instrumented (access memory)
– Record the ins, the memop, the offset from start of the trace in the buffer
where the {appIP, memAddr} pair of this ins should be written
– Get a sum of the TotalSizeOccupiedByTraceInBuffer
– Insert the IF-THEN sequence at the beginning of the trace
– Insert the update of endOfTraceReg just after the IF-THEN
sequence
– iterate thru recorded (memory accessing) INSs of the Trace
– Insert the instrumentation code before each recorded memory
accessing instruction
– this is the code that writes the {appIP, memAddr} pair into the buffer at the
designated offset (from start of trace) for this INS.
– endOfTraceReg and endOfBufferReg are virtual registers allocated
by Pin to the Pin tool.
Software & Services Group
96
memtrace_simple
TLS_KEY appThreadRepresentitiveKey; // Pin TLS key
REG endOfTraceInBufferReg; // Pin virtual Reg that will hold the pointer to the end of the trace data in
// the buffer
REG endOfBufferReg; // Pin virtual Reg that will hold the pointer to the end of the buffer
struct MEMREF {
ADDRINT appIP;
ADDRINT memAddr;
} ; // structure of the {appIP, memAddr} pair of a memory accessing ins in the buffer
int main(int argc, char * argv[])
{
PIN_Init(argc,argv) ;
// Pin TLS slot for holding the object that represents the application thread
appThreadRepresentitiveKey = PIN_CreateThreadDataKey(0);
// get the registers to be used in each thread for managing the per-thread buffer
endOfTraceInBufferReg = PIN_ClaimToolRegister();
endOfBufferReg
= PIN_ClaimToolRegister();
TRACE_AddInstrumentFunction(TraceAnalysisCalls, 0);
PIN_AddThreadStartFunction(ThreadStart, 0);
PIN_AddThreadFiniFunction(ThreadFini, 0);
PIN_AddFiniFunction(Fini, 0);
}
PIN_StartProgram();
Software & Services Group
97
memtrace_simple
KNOB<UINT32> KnobNumBytesInBuffer(KNOB_MODE_WRITEONCE, "pintool", "num_bytes_in_buffer",
"0x100000", "number of bytes in buffer");
APP_THREAD_REPRESENTITVE::APP_THREAD_REPRESENTITVE(THREADID myTid) {
_buffer = new char[KnobNumBytesInBuffer.Value()]; // Allocate the buffer
_numBuffersFilled = 0;
_numElementsProcessed = 0;
_myTid = myTid;
}
char * APP_THREAD_REPRESENTITVE::Begin() { return _buffer; }
char * APP_THREAD_REPRESENTITVE:: End()
{ return _buffer + KnobNumBytesInBuffer.Value(); }
VOID ThreadStart(THREADID tid,
CONTEXT *ctxt,
INT32 flags, VOID *v) // Pin callback on thread creation
// There is a new APP_THREAD_REPRESENTITVE object for every thread
APP_THREAD_REPRESENTITVE * appThreadRepresentitive
= new APP_THREAD_REPRESENTITVE(tid);
{
// A thread will need to look up its APP_THREAD_REPRESENTITVE, so save pointer in Pin TLS
PIN_SetThreadData(appThreadRepresentitiveKey, appThreadRepresentitive, tid);
// Initialize endOfTraceInBufferReg to point at beginning of buffer
PIN_SetContextReg(ctxt, endOfTraceInBufferReg,
reinterpret_cast<ADDRINT>(appThreadRepresentitive->Begin()));
// Initialize endOfBufferReg to point at end of buffer
PIN_SetContextReg(ctxt, endOfBufferReg,
reinterpret_cast<ADDRINT>(appThreadRepresentitive->End()));
}
Software & Services Group
98
memtrace_simple / Trace Instrmnt
void TraceAnalysisCalls(TRACE trace, void *) /*TRACE_AddInstrumentFunction(TraceAnalysisCalls, 0)*/ {
// Go over all BBLs of the trace and for each BBL determine and record the INSs which need
// to be instrumented - i.e. the ins requires an analysis call
Now
insert
aof
call
the
function
Finally,
iterate
over
all of
theaccess
memory
CheckIfNoSpaceForTraceInBuffer
will
Record
all
INSs
intotrace
that
At the
start
the
trace:
Trace
level
instrumentation
AllocateSpaceForTraceInBuffer
which
TRACE_ANALYSIS_CALLS_NEEDED traceAnalysisCallsNeeded;
accessing
insert
a call
return
1 if INSs
therein
is the
NOTand
enough
room
memory
allocates
space
in the buffer
for the
for (BBL bbl = TRACE_BblHead(trace); BBL_Valid(bbl); bbl = BBL_Next(bbl))
InsertIf
call
to routine
to
the buffer
analysis
that records
in
the
instrumentations
of eachpair
of the
DetermineBBLAnalysisCalls(bbl, &traceAnalysisCallsNeeded);
CheckIfNoSpaceForTraceInBuffer
to
the
{appIp, memAddr}
into it’s
In
this ifcase
theisBufferFull
memory
accessing
INSs
infunction
theroom
tracewill
check
there
NOT
enough
into
allocated
space
in
the
buffer
be
called
to
process
all
pairs
in
the
write
their to
{appIp,
the buffer
insert memAddr}
the AppIp, pairs.
// If No memory accesses in this trace
and pair
set the
MemAddr
for each of the recorded
if (traceAnalysisCallsNeeded.NumAnalysisCallsNeeded() == 0) return;buffer
This
is
done
by adding size
of the
endOfTraceInBufferReg
to point
to the
INSs
space
needed
to
the
top
of
the
buffer
// APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer will
determine if there are NOT enough
endOfTraceInBufferReg
// available bytes in the buffer. If there are NOT then it returns TRUE and the BufferFull function is called
TRACE_InsertIfCall(trace, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg, // previous trace
IARG_REG_VALUE, endOfBufferReg,
IARG_UINT32, traceAnalysisCallsNeeded.TotalSizeOccupiedByTraceInBuffer(),
IARG_END);
TRACE_InsertThenCall(trace, IPOINT_BEFORE, AFUNPTR(APP_THREAD_REPRESENTITVE::BufferFull),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_THREAD_ID,
IARG_RETURN_REGS, endOfTraceInBufferReg,
IARG_END);
TRACE_InsertCall(trace, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::AllocateSpaceForTraceInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_UINT32, traceAnalysisCallsNeeded.TotalSizeOccupiedByTraceInBuffer(),
IARG_RETURN_REGS, endOfTraceInBufferReg,
IARG_END);
// Insert Analysis Calls for each INS on the trace that was recorded as needing one
traceAnalysisCallsNeeded.InsertAnalysisCalls();
}
Software & Services Group
99
memtrace_simple
class ANALYSIS_CALL_INFO
public:
ANALYSIS_CALL_INFO(INS ins, UINT32 offsetFromTraceStartInBuffer, UINT32 memop) :
_ins(ins),
_offsetFromTraceStartInBuffer(offsetFromTraceStartInBuffer), _memop (memop) {}
void InsertAnalysisCall(INT32 sizeofTraceInBuffer);
private:
INS _ins; INT32 _offsetFromTraceStartInBuffer; UINT32 _memop;
{
};
class TRACE_ANALYSIS_CALLS_NEEDED
{
public:
TRACE_ANALYSIS_CALLS_NEEDED() : _numAnalysisCallsNeeded(0), _currentOffsetFromTraceStartInBuffer(0) {}
UINT32 NumAnalysisCallsNeeded() const { return _numAnalysisCallsNeeded; }
UINT32 TotalSizeOccupiedByTraceInBuffer() const { return _currentOffsetFromTraceStartInBuffer; }
void RecordAnalysisCallNeeded(INS ins, UINT32 memop) {
_analysisCalls.push_back(ANALYSIS_CALL_INFO(ins, _currentOffsetFromTraceStartInBuffer,
memop));
Class
memory
Class for
for
recording
all
_currentOffsetFromTraceStartInBuffer += sizeof(MEMREF);
Called
forrecording
each BBLone
in the
Trace
accessing
INS that will
memory accessing
INSsbein the
_numAnalysisCallsNeeded++;
}
void InsertAnalysisCalls();
private:
INT32 _currentOffsetFromTraceStartInBuffer;
INT32 _numAnalysisCallsNeeded;
vector<ANALYSIS_CALL_INFO> _analysisCalls;
Records
each memory accessing
instrumented
Trace
INS in the BBL into vector of
ANALYSIS_CALL_INFO
};
void DetermineBBLAnalysisCalls (BBL bbl,
TRACE_ANALYSIS_CALLS_NEEDED * traceAnalysisCallsNeeded)
{
for (INS ins = BBL_InsHead(bbl); INS_Valid(ins); ins = INS_Next(ins))
{
// Iterate over each memory operand of the instruction.
for (UINT32 memOp = 0; memOp < INS_MemoryOperandCount(ins); memOp++)
// Record that an analysis call is needed, along with the info needed to generate the analysis
// call
traceAnalysisCallsNeeded->RecordAnalysisCallNeeded(ins, memOp);
}
}
•
100
Software & Services Group
memtrace_simple / Trace Instrmnt
void TraceAnalysisCalls(TRACE trace, void *) /*TRACE_AddInstrumentFunction(TraceAnalysisCalls, 0)*/ {
// Go over all BBLs of the trace and for each BBL determine and record the INSs which need
// to be instrumented - i.e. the ins requires an analysis call
TRACE_ANALYSIS_CALLS_NEEDED traceAnalysisCallsNeeded;
for (BBL bbl = TRACE_BblHead(trace); BBL_Valid(bbl); bbl = BBL_Next(bbl))
DetermineBBLAnalysisCalls(bbl, &traceAnalysisCallsNeeded);
// If No memory accesses in this trace
if (traceAnalysisCallsNeeded.NumAnalysisCallsNeeded() == 0) return;
// APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer will determine if there are NOT enough
// available bytes in the buffer. If there are NOT then it returns TRUE and the BufferFull function is called
TRACE_InsertIfCall(trace, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg, // previous trace
IARG_REG_VALUE, endOfBufferReg,
IARG_UINT32, traceAnalysisCallsNeeded.TotalSizeOccupiedByTraceInBuffer(),
IARG_END);
TRACE_InsertThenCall(trace, IPOINT_BEFORE, AFUNPTR(APP_THREAD_REPRESENTITVE::BufferFull),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_THREAD_ID,
IARG_RETURN_REGS, endOfTraceInBufferReg,
IARG_END);
TRACE_InsertCall(trace, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::AllocateSpaceForTraceInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_UINT32, traceAnalysisCallsNeeded.TotalSizeOccupiedByTraceInBuffer(),
IARG_RETURN_REGS, endOfTraceInBufferReg,
IARG_END);
// Insert Analysis Calls for each INS on the trace that was recorded as needing one
traceAnalysisCallsNeeded.InsertAnalysisCalls();
}
Software & Services Group
101
memtrace_simple
THEN call to process the buffer, and
set the endOfTraceInBufferReg to
the beginning of the buffer
static ADDRINT PIN_FAST_ANALYSIS_CALL
Inserted at beginning of each trace,
APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer ( // Pin
will
inline
this
just
AFTER
the
IF function
call
char * endOfPreviousTraceInBuffer,
Executed only when the IF function
char * bufferEnd,
returns 1
ADDRINT totalSizeOccupiedByTraceInBuffer)
{
return (endOfPreviousTraceInBuffer + totalSizeOccupiedByTraceInBuffer >= bufferEnd);
}
Function
inserted atifthe
beginning
static char * PIN_FAST_ANALYSIS_CALL
IF
call to determine
there
is NOT
of each room
trace,
just
after
thefor
enough
in
the
buffer
the
Analysis
functions
inserted
atTHEN
start
APP_THREAD_REPRESENTITVE::BufferFull ( // Pin will NOT inline this function
function.
{
appIp,
memAddr}
pairs
of
all
the
of
each
trace.
char *endOfTraceInBuffer,
memory
accessing
INSs
in
the
ADDRINT tid)
Executed each time trace executes.
trace.
{
Inlined by
PinExecuted at beginning
Inserted
and
// Get this thread’s APP_THREAD_REPRESENTITVE from the Pin TLS
of
each trace.
Allocates
space in the buffer for the
APP_THREAD_REPRESENTITVE * appThreadRepresentitive
{appIp,
memAddr}
pairs of all of
= static_cast<APP_THREAD_REPRESENTITVE*>
Inlined by Pin.
the
memory
accessing
INSs in the
(PIN_GetThreadData(appThreadRepresentitiveKey, tid));
appThreadRepresentitive->ProcessBuffer(endOfTraceInBuffer);
}
trace. 1 if there is NOT enough
Returns
room, 0 if there is
// After processing the buffer, move the endOfTraceInBuffer back to the beginning of the buffer
endOfTraceInBuffer = appThreadRepresentitive->Begin();
return endOfTraceInBuffer;
static char * PIN_FAST_ANALYSIS_CALL
APP_THREAD_REPRESENTITVE::AllocateSpaceForTraceInBuffer (// Pin will inline this function
{
}
char * endOfPreviousTraceInBuffer,
ADDRINT totalSizeOccupiedByTraceInBuffer)
return (endOfPreviousTraceInBuffer + totalSizeOccupiedByTraceInBuffer);
Software & Services Group
102
memtrace_simple / Trace Instrmnt
void TraceAnalysisCalls(TRACE trace, void *) /*TRACE_AddInstrumentFunction(TraceAnalysisCalls, 0)*/ {
// Go over all BBLs of the trace and for each BBL determine and record the INSs which need
// to be instrumented - i.e. the ins requires an analysis call
TRACE_ANALYSIS_CALLS_NEEDED traceAnalysisCallsNeeded;
for (BBL bbl = TRACE_BblHead(trace); BBL_Valid(bbl); bbl = BBL_Next(bbl))
DetermineBBLAnalysisCalls(bbl, &traceAnalysisCallsNeeded);
// If No memory accesses in this trace
if (traceAnalysisCallsNeeded.NumAnalysisCallsNeeded() == 0) return;
// APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer will determine if there are NOT enough
// available bytes in the buffer. If there are NOT then it returns TRUE and the BufferFull function is called
TRACE_InsertIfCall(trace, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg, // previous trace
IARG_REG_VALUE, endOfBufferReg,
IARG_UINT32, traceAnalysisCallsNeeded.TotalSizeOccupiedByTraceInBuffer(),
IARG_END);
TRACE_InsertThenCall(trace, IPOINT_BEFORE, AFUNPTR(APP_THREAD_REPRESENTITVE::BufferFull),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_THREAD_ID,
IARG_RETURN_REGS, endOfTraceInBufferReg,
IARG_END);
TRACE_InsertCall(trace, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::AllocateSpaceForTraceInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_UINT32, traceAnalysisCallsNeeded.TotalSizeOccupiedByTraceInBuffer(),
IARG_RETURN_REGS, endOfTraceInBufferReg,
IARG_END);
// Insert Analysis Calls for each INS on the trace that was recorded as needing one
traceAnalysisCallsNeeded.InsertAnalysisCalls();
}
Software & Services Group
103
memtrace_simple
static void PIN_FAST_ANALYSIS_CALL
APP_THREAD_REPRESENTITVE::RecordMEMREFInBuffer ( // Pin will inline this function
char* endOfTraceInBuffer, ADDRINT offsetFromEndOfTrace, ADDRINT appIp, ADDRINT memAddr)
{
*reinterpret_cast<ADDRINT*>(endOfTraceInBuffer+ offsetFromEndOfTrace) = appIp;
*reinterpret_cast<ADDRINT*>(endOfTraceInBuffer+ offsetFromEndOfTrace +sizeof(ADDRINT))
= memAddr;
}
void ANALYSIS_CALL_INFO::InsertAnalysisCall(INT32 sizeofTraceInBuffer)
{
/* the place in the buffer where the {appIp, memAddr} of this _ins should be recorded is
computed by: endOfTraceInBufferReg
-sizeofTraceInBuffer + _offsetFromTraceStartInBuffer(of this _ins) */
INS_InsertCall(_ins, IPOINT_BEFORE,
AFUNPTR(APP_THREAD_REPRESENTITVE::RecordMEMREFInBuffer),
IARG_FAST_ANALYSIS_CALL,
IARG_REG_VALUE, endOfTraceInBufferReg,
IARG_ADDRINT, ADDRINT(_offsetFromTraceStartInBuffer - sizeofTraceInBuffer),
IARG_INST_PTR,
IARG_MEMORYOP_EA, _memop,
IARG_END);
}
void TRACE_ANALYSIS_CALLS_NEEDED::InsertAnalysisCalls()
{// Iterate over the recorded ANALYSIS_CALL_INFO elements – insert the analysis call
for (vector<ANALYSIS_CALL_INFO>::iterator c = _analysisCalls.begin();
c != _analysisCalls.end();
c++)
c->InsertAnalysisCall(TotalSizeOccupiedByTraceInBuffer());
}
Software & Services Group
104
membuffer_simple
• Since managing a per-thread buffer is a necessity of a large
class of Pin tools: Provide Pin APIs to make it (more) easy.
• Pin Buffering API, abstracts away the need for a Pin tool to
manage per-thread buffers
• PIN_DefineTraceBuffer
– Define a per-thread buffer that each application trace can write
data to
• INS_InsertFillBuffer
– Instrumentation code is generated to write the desired data into
the buffer
– This code is inlined
• Tool defined BufferFull function, instrumentation code will
cause this function to be called when the buffer becomes full
Software & Services Group
105
membuffer_simple
• Pin Buffering API actually works somewhat different
than memtrace
– Instrumentation code will insert the data generated by an
INS into the buffer immediately after the data generated
by the previously executed instrumented INS
– Better buffer utilization
– Requires the instrumentation to update the next buffer location
to write to – this was not required in the memtrace
implementatio
– All this is invisible to the Pin tool writer
• membuffer_simple is a Pin tool that uses the Pin
Buffering API to do the same memory access
recording that memtrace_simple does
Software & Services Group
106
membuffer_simple
KNOB<UINT32> KnobNumPagesInBuffer(KNOB_MODE_WRITEONCE, "pintool",
"num_pages_in_buffer", "256", "number of pages in buffer");
// Struct of memory reference written to the buffer
struct MEMREF
ADDRINT appIP;
ADDRINT memAddr;
{
};
// The buffer ID returned by the one call to PIN_DefineTraceBuffer
BUFFER_ID bufId;
TLS_KEY appThreadRepresentitiveKey;
int main(int argc, char * argv[])
PIN_Init(argc,argv) ;
{
// Pin TLS slot for holding the object that represents an application thread
appThreadRepresentitiveKey = PIN_CreateThreadDataKey(0);
// Define the buffer that will be used – buffer is allocated to each thread when the thread starts
//running
bufId = PIN_DefineTraceBuffer(sizeof(struct MEMREF), KnobNumPagesInBuffer,
BufferFull, // This Pin tool function will be called when buffer is full
0);
INS_AddInstrumentFunction(Instruction, 0); // The Instruction function will use the Pin Buffering
// API to insert the instrumentation code that writes
// the MEMREF of a memory accessing INS into the buffer
PIN_AddThreadStartFunction(ThreadStart, 0);
PIN_AddThreadFiniFunction(ThreadFini, 0);
PIN_AddFiniFunction(Fini, 0);
PIN_StartProgram();
}
Software & Services Group
107
membuffer_simple
/*
* Pin generates code to call this function when a buffer fills up, and exceutes a callback
to this function when the thread exits
* Pin will NOT inline this function
* @param[in] id
buffer handle
* @param[in] tid
id of owning thread
* @param[in] ctxt
application context
* @param[in] buf
actual pointer to buffer
* @param[in] numElements number of records
* @param[in] v
callback value
* @return A pointer to the buffer to resume filling.
*/
VOID * BufferFull(BUFFER_ID id, THREADID tid, const CONTEXT *ctxt, VOID *buf,
UINT64 numElements, VOID *v)
{
// retrieve the APP_THREAD_REPRESENTITVE* of this thread from the Pin TLS
APP_THREAD_REPRESENTITVE * appThreadRepresentitive
= static_cast<APP_THREAD_REPRESENTITVE*>( PIN_GetThreadData(
appThreadRepresentitiveKey, tid ) );
appThreadRepresentitive->ProcessBuffer(buf, numElements);
return buf;
}}
Software & Services Group
108
membuffer_simple
VOID Instruction (INS ins, VOID *v)
{
UINT32 numMemOperands = INS_MemoryOperandCount(ins);
// Iterate over each memory operand of the instruction.
for (UINT32 memOp = 0; memOp < numMemOperands ; memOp++)
{ // Add the instrumentation code to write the appIP and memAddr
// of this memory operand into the buffer
// Pin will inline the code that writes to the buffer
INS_InsertFillBuffer(ins, IPOINT_BEFORE, bufId,
IARG_INST_PTR, offsetof(struct MEMREF, appIP),
IARG_MEMORYOP_EA, memOp,
offsetof(struct MEMREF, memAddr),
IARG_END);
}
}
Software & Services Group
109
branchbuffer_simple
• Use Pin Buffering API to collect a branch trace:
– For each executed branch instruction record:
– appIP of the branch instruction
– targetAddress of the branch instruction
– branchTaken boolean
Software & Services Group
110
branchbuffer_simple
KNOB<UINT32> KnobNumPagesInBuffer(KNOB_MODE_WRITEONCE, "pintool",
"num_pages_in_buffer", "256", "number of pages in
buffer");
struct BRANCH_INFO { // This is the structure of the data that will be written into the buffer
ADDRINT appIP;
ADDRINT targetAddress;
BOOL
branchTaken;
};
int main(int argc, char *argv[])
{
PIN_Init(argc,argv);
bufId
= PIN_DefineTraceBuffer(sizeof(BRANCH_INFO), KnobNumPagesInBuffer, BufferFull, 0);
// Register function to be called to instrument traces
TRACE_AddInstrumentFunction(Trace, 0);
// Register function to be called when the application exits
PIN_AddFiniFunction(Fini, 0);
}
// Start the program, never returns
PIN_StartProgram();
Software & Services Group
111
branchbuffer_simple
void Trace(TRACE tr, void* V) // TRACE_AddInstrumentFunction(Trace, 0);
{
for(BBL bbl = TRACE_BblHead(tr); BBL_Valid(bbl); bbl=BBL_Next(bbl))
{
if (INS_IsBranchOrCall(BBL_InsTail(bbl))) // The branch instruction, if it exists, will always
// be the last in the BBL
{
}
}
INS_InsertFillBuffer(BBL_InsTail(bbl),
IPOINT_BEFORE,
bufId,
IARG_INST_PTR, offsetof(BRANCH_INFO, appIP),
IARG_BRANCH_TARGET_ADDR, offsetof(BRANCH_INFO, targetAddress),
IARG_BRANCH_TAKEN, offsetof(BRANCH_INFO, branchTaken),
IARG_END);
}
Software & Services Group
112
Symbols
• PIN_InitSymbols()
– Pin will use whatever symbol information is available
–
–
–
–
Debug info in the app
Pdb files
Export Tables
On Windows uses dbghelp
– See PIN_InitSymbolsAlt() for more control over which
symbols will be used
• Use symbols to instrument/wrap/replace specific
functions
– wrap/replace: see malloc replacement examples in intro
• Access application debug information from a Pin
tool
Software & Services Group
113
Symbols: Instrument malloc and free
int main(int argc, char *argv[])
{
// Initialize pin symbol manager
PIN_InitSymbols();
// See also PIN_InitSymbolsAlt() for more control over which symbols are read
PIN_Init(argc,argv);
// Register the function ImageLoad to be called each time an image is loaded in the process
// This includes the process itself and all shared libraries it loads (implicitly or explicitly)
IMG_AddInstrumentFunction(ImageLoad, 0);
// Never returns
PIN_StartProgram();
}
Software & Services Group
114
Symbols: Instrument malloc and free
VOID ImageLoad(IMG img, VOID *v) // Pin Callback. IMG_AddInstrumentFunction(ImageLoad, 0);
{
// Instrument the malloc() and free() functions. Print the input argument
// of each malloc() or free(), and the return value of malloc().
RTN mallocRtn = RTN_FindByName(img, "_malloc"); // Find the malloc() function.
if (RTN_Valid(mallocRtn))
{
RTN_Open(mallocRtn);
// Instrument malloc() to print the input argument value and the return value.
RTN_InsertCall(mallocRtn, IPOINT_BEFORE, (AFUNPTR)MallocBefore,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
RTN_InsertCall(mallocRtn, IPOINT_AFTER, (AFUNPTR)MallocAfter,
IARG_FUNCRET_EXITPOINT_VALUE, IARG_END);
}
}
RTN_Close(mallocRtn);
RTN freeRtn = RTN_FindByName(img, "_free"); // Find the free() function.
if (RTN_Valid(freeRtn))
{
RTN_Open(freeRtn);
// Instrument free() to print the input argument value.
RTN_InsertCall(freeRtn, IPOINT_BEFORE, (AFUNPTR)FreeBefore,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
RTN_Close(freeRtn);
}
Software & Services Group
115
Symbols: Instrument malloc
Handling name-mangling and multiple
symbols at same address
VOID Image(IMG img, VOID *v) // IMG_AddInstrumentFunction(Image, 0);
{
// Walk through the symbols in the symbol table.
for (SYM sym = IMG_RegsymHead(img); SYM_Valid(sym); sym = SYM_Next(sym))
{
string undFuncName = PIN_UndecorateSymbolName(SYM_Name(sym), UNDECORATION_NAME_ONLY);
if (undFuncName == "malloc") // Find the malloc function.
{
RTN mallocRtn = RTN_FindByAddress(IMG_LowAddress(img) + SYM_Value(sym));
if (RTN_Valid(mallocRtn))
{
RTN_Open(mallocRtn);
// Instrument to print the input argument value and the return value.
RTN_InsertCall(mallocRtn, IPOINT_BEFORE, (AFUNPTR)MallocBefore,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
RTN_InsertCall(mallocRtn, IPOINT_AFTER, (AFUNPTR)MallocAfter,
IARG_FUNCRET_EXITPOINT_VALUE,
IARG_END);
}
}
}
}
RTN_Close(mallocRtn);
Software & Services Group
116
Symbols: Accessing Application Debug
Info from a Pin Tool:
Catch a Memory Overwrite
VOID Instruction(INS ins, VOID *v) // INS_AddInstrumentFunction(Instruction, 0);
{
UINT32 numMemOperands = INS_MemoryOperandCount(ins);
// Iterate over each memory operand of the instruction.
for (UINT32 memOp = 0; memOp < numMemOperands ; memOp++)
{
if (INS_MemoryOperandIsWritten(ins, memOp))
{ // Insert instrumentation code to catch a memory overwrite
INS_InsertIfCall (ins, IPOINT_BEFORE,
AFUNPTR(AnalyzeMemWrite),
IARG_FAST_ANALYSIS_CALL,
IARG_MEMORYOP_EA, memop,
IARG_MEMORYWRITE_SIZE,
IARG_END);
INS_InsertThenCall (ins, IPOINT_BEFORE,
AFUNPTR(MemoryOverWriteAt),
IARG_FAST_ANALYSIS_CALL,
IARG_INST_PTR,
IARG_MEMORYOP_EA, memop,
IARG_MEMORYWRITE_SIZE,
IARG_END);
}
}
}
Software & Services Group
117
Symbols: Accessing Application Debug
Info from a Pin Tool
KNOB<ADDRINT> KnobMemAddrBeingOverwritten(KNOB_MODE_WRITEONCE, "pintool",
"mem_overwrite_addr", "256", "overwritten memaddr");
static ADDRINT PIN_FAST_ANALYSIS_CALL
AnalyzeMemWrite ( // Pin will inline this function, it is the IF part
ADDRINT memWriteAddr, UINT32 numBytesWritten)
{
// return 1 if this memory write overwrites the address specified by
// KnobMemAddrBeingOverwritten
return (memWriteAddr<= KnobMemAddrBeingOverwritten &&
(memWriteAddr + numBytesWritten) > KnobMemAddrBeingOverwritten);
}
static VOID PIN_FAST_ANALYSIS_CALL
MemoryOverWriteAt ( // Pin will NOT inline this function, it is the THEN part
ADDRINT appIP, ADDRINT memWriteAddr, UINT32 numBytesWritten)
{
INT32 column, lineNum;
string fileName;
PIN_GetSourceLocation (appIP, &column, &line, &fileName);
}
printf ("overwrite of %p from instruction at %p originating from file %s line %d col %d\n",
KnobMemAddrBeingOverwritten, appIP, fileName.c_str(), lineNum, column);
printf (" writing %d bytes starting at %p\n", numBytesWritten, memWriteAddr);
Software & Services Group
118
Probe Mode
• JIT Mode
– Pin creates a modified copy of the application onthe-fly
– Original code never executes
More flexible, more common approach
• Probe Mode
– Pin modifies the original application instructions
– Inserts jumps to instrumentation code
(trampolines)
Lower overhead (less flexible) approach
Software & Services Group
119
Pin Probe-Mode
• Probe mode is a method of using Pin to instrument at the
function level only. Wrap, Replace, call Analysis function
before/after.
• Replacement or Wrapping function can call the replaced
(original) function.
• The application and the replacement routine are run natively
(not Jitted).
–
–
–
–
–
Faster than Jit-mode
Puts more responsibility on the tool writer.
Probes can only be placed on RTN boundaries
Must be inserted within the Image load callback.
Pin will automatically remove the probes when an image is unloaded.
• Many of the PIN APIs that are available in JIT mode are not
available in Probe mode.
Software & Services Group
120
A Sample Probe
– A probe is a jump instruction that overwrites
original instruction(s) in the application
– Instrumentation invoked with probes
– Pin copies/translates original bytes so probed (replaced)
functions can be called from the replacement function
Copy of entry point with
0x50000004: push
0x50000005: mov
0x50000007: push
0x50000008: push
0x50000009: jmp
original bytes:
%ebp
%esp,%ebp
%edi
%esi
0x400113d9
0x41481064: push %ebp
Entry
Original
point
function
overwritten
entrywith
point:
probe:
0x400113d4:
0x400113d5:
0x400113d7:
0x400113d8:
0x400113d9:
jmp
push
mov
push
push
push
0x41481064
%ebp
%esp,%ebp
%edi
%esi
%ebx
// tool wrapper func
::::::::::::::::::::
0x414827fe: call 0x50000004 // call original func
Software & Services Group
121
PinProbes Instrumentation
• Advantages:
– Low overhead – few percent
– Less intrusive – execute original code
– Leverages Pin:
– API
– Instrumentation engine
• Disadvantages:
– More tool writer responsibility
– Routine-level granularity (RTN)
Software & Services Group
122
Using Probes to Replace/Wrap a
Function
• RTN_ReplaceSignatureProbed() redirects all calls
to application routine rtn to the specified
replacementFunction
– Can add IARG_* types to be passed to the replacement
routine, including pointer to original function and
IARG_CONTEXT.
– Replacement function can call original function.
• To use:
– Must use PIN_StartProgramProbed()
– Application prototype is required
Software & Services Group
123
Malloc Replacement Probe-Mode
#include "pin.H"
void * MallocWrapper(AFUNPTR pf_malloc, size_t size)
{ // Simulate out-of-memory every so often
void * res;
if (TimeForOutOfMem())
return (NULL);
res = pf_malloc(size);
return res; }
VOID ImageLoad (IMG img, VOID *v) {
if (strstr(IMG_Name(img).c_str(), "libc.so") ||
strstr(IMG_Name(img).c_str(), "MSVCR80") || strstr(IMG_Name(img).c_str(), "MSVCR90"))
{
RTN mallocRtn = RTN_FindByName(img, "malloc");
if ( RTN_Valid(mallocRtn) && RTN_IsSafeForProbedReplacement(mallocRtn) )
{
PROTO proto_malloc = PROTO_Allocate(PIN_PARG(void *), CALLINGSTD_DEFAULT, "malloc",
PIN_PARG(size_t), PIN_PARG_END() );
} }}
RTN_ReplaceSignatureProbed(mallocRtn,
AFUNPTR(MallocWrapper),
IARG_PROTOTYPE, proto_malloc,
IARG_ORIG_FUNCPTR,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
int main(int argc, CHAR *argv[]) {
PIN_InitSymbols();
PIN_Init(argc,argv));
IMG_AddInstrumentFunction(ImageLoad, 0);
PIN_StartProgramProbed();
}
Software & Services Group
124
Using Probes to Call Analysis
Functions
• RTN_InsertCallProbed() invokes the analysis
routine before or after the specified rtn
– Use IPOINT_BEFORE or IPOINT_AFTER
– Pin may NOT be able to find all AFTER points on
the function when it is running in Probe-Mode
– PIN IARG_TYPEs are used for arguments
• To use:
– Must use PIN_StartProgramProbed()
– Application prototype is required for IPOINT_AFTER
Software & Services Group
125
Symbols: Instrument malloc
Handling name-mangling and multiple
symbols at same address Probe-Mode
VOID Image(IMG img, VOID *v) // IMG_AddInstrumentFunction(Image, 0);
{
// Walk through the symbols in the symbol table.
for (SYM sym = IMG_RegsymHead(img); SYM_Valid(sym); sym = SYM_Next(sym))
{
string undFuncName = PIN_UndecorateSymbolName(SYM_Name(sym), UNDECORATION_NAME_ONLY);
if (undFuncName == "malloc") // Find the malloc function.
{
RTN mallocRtn = RTN_FindByAddress(IMG_LowAddress(img) + SYM_Value(sym));
if (RTN_Valid(mallocRtn))
{
RTN_Open(mallocRtn);
PROTO proto_malloc = PROTO_Allocate(PIN_PARG(void *), CALLINGSTD_DEFAULT, "malloc",
PIN_PARG(size_t), PIN_PARG_END() );
// Instrument to print the input argument value and the return value.
RTN_InsertCallProbed(mallocRtn, IPOINT_BEFORE, (AFUNPTR)MallocBefore,
IARG_PROTOTYPE, proto_malloc,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_END);
RTN_InsertCallProbed(mallocRtn, IPOINT_AFTER, (AFUNPTR)MallocAfter,
IARG_PROTOTYPE, proto_malloc,
IARG_FUNCRET_EXITPOINT_VALUE,
IARG_END);
}
}
RTN_Close(mallocRtn);
}
}
Software & Services Group
126
Tool Writer Responsibilities
• No control flow into the instruction space where
probe is placed
– 6 bytes on IA-32, 7 bytes on Intel64, 1 bundle
on IA64
– Branch into “replaced” instructions will fail
– Probes at function entry point only
• Thread safety for insertion and deletion of probes
– During image load callback is safe
– Only loading thread has a handle to the image
• Replacement function has same behavior as
original
Software & Services Group
127
Multi-Threading
• Have shown a number of examples of Pin tools
supporting multi-threading
• Pin fully supports multi-threading
– Application threads execute jitted code including
instrumentation code (inlined and not inlined), without any
serialization introduced by Pin
– Instrumentation code can use Pin and/or OS synchronization
constructs to introduce serialization if needed.
– Will see examples of this in Part3
– System calls require serialized entry to the VM before and after
execution – BUT actual execution is NOT serialized
– Pin does NOT create any threads of it’s own
– Pin callbacks are serialized
– Including the BufferFull callback
– Jitting is serialized
– Only one application thread can be jitting code at any time
Software & Services Group
128
Multi-Threading
• Pin Tools, in Jit-Mode, can:
– Track Threads
– ThreadStart, ThreadFini callbacks
– IARG_THREAD_ID
– Use Pin TLS for thread-specific data
– Use Pin Locks to synchronize threads
– Create threads to do Pin Tool work
– Use Pin provided APIs to do this
– Otherwise these threads would be Jitted
– Details in Part4
Software & Services Group
129
Multi-Threading, Locking Guidelines
• Basic Rules
– If the tool acquires any locks in a Pin call-back, it must
release those locks before returning from that call-back.
– If the tool acquires any locks in an analysis routine, it must
release those locks before returning from the analysis
routine.
– If the tool calls a Pin API from a call-back, it should not
hold any tool locks when calling the API.
– If the tool calls a Pin API from an analysis routine, it may
need to acquire the Pin client lock first (see the
documentation for the API). The tool should not hold any
other locks when calling the API.
Software & Services Group
130
Multi-Threading, Locking Guidelines
• Advanced Rules
– If the tool acquires any locks in a Pin call-back, it must
release those locks before returning from that call-back.
– If the tool calls a Pin API from a call-back, it should not
hold any tool locks when calling the API.
– If the tool calls a Pin API from an analysis routine, it may
need to acquire the Pin client lock first (see the
documentation for the API). If the tool holds a tool lock L
while calling the API, that lock L must obey the following
sub-rule:
– The tool must not acquire lock L from any call-back. This
avoids a lock order inversion with respect to the Pin internal
locks.
Software & Services Group
131
Multi-Threading, Locking Guidelines
• Advanced Rules
– If the tool acquires any locks in an analysis routine, it must
release those locks before leaving the trace that contains
the analysis routine. Tools must expect that the trace may
exit “early” if an application instruction raises an
exception. Any lock L, which the tool might hold when the
application raises an exception, must obey the following
sub-rules:
– The tool must establish a call-back that executes when the
application raises an exception, and this call-back must release
lock L if it was acquired at the time of the exception. Tools
can use PIN_AddContextChangeFunction() to establish this
call-back.
– The tool must not acquire lock L from any call-back. This
avoids a lock order inversion with respect to the Pin internal
locks.
Software & Services Group
132
Taint Analysis
• For each instruction
– Identify source and destination operands
– Explicit, Implicit
– If SRC is tainted then set DEST is tainted
– If SRC isn’t tainted then set DEST isn’t tainted
• Sounds simple, right?
Software & Services Group
133
Taint Analysis
• Implicit operands
• Partial register taint
• Math instructions
• Logical instructions
• Exchange instructions
Software & Services Group
134
A simple taint analyzer
Set of Tainted Memory Addresses
Define
initial
taint
Fetch
next inst.
bffff081
bffff082
b64d4002
If src is
untainted
set dest
untainted
If src is
tainted
set dest
tainted
Tainted Registers
EAX
EDX
ESI
Software & Services Group
135
#include
#include
#include
#include
#include
#include
"pin.H"
<iostream>
<fstream>
<set>
<string.h>
"xed-iclass-enum.h"
set<ADDRINT> TaintedAddrs; // tainted memory addresses
bool TaintedRegs[REG_LAST]; // tainted registers
std::ofstream out;
// output file
KNOB<string> KnobOutputFile(KNOB_MODE_WRITEONCE, "pintool",
"o", "taint.out", "specify file name for the output file");
/*!
* Print out help message.
*/
INT32 Usage()
{
cerr << "This tool follows the taint defined by the first argument to " << endl <<
"the instrumented program command line and outputs details to a file" << endl ;
}
cerr << KNOB_BASE::StringKnobSummary() << endl;
return -1;
Software & Services Group
136
int main(int argc, char *argv[])
{
// Initialize PIN
PIN_InitSymbols();
if( PIN_Init(argc,argv) )
{
return Usage();
}
// Register function to be called to instrument traces
TRACE_AddInstrumentFunction(Trace, 0);
RTN_AddInstrumentFunction(Routine, 0);
// Register function to be called when the application exits
PIN_AddFiniFunction(Fini, 0);
// init output file
string fileName = KnobOutputFile.Value();
out.open(fileName.c_str());
// Start the program, never returns
PIN_StartProgram();
}
return 0;
Software & Services Group
137
/*!
* Routine instrumentation, called for every routine loaded
* this function adds a call to MainAddTaint on the main function
*/
VOID Routine(RTN rtn, VOID *v)
{
RTN_Open(rtn);
if (RTN_Name(rtn) == "main") //if this is the main function
{
RTN_InsertCall(rtn, IPOINT_BEFORE, (AFUNPTR)MainAddTaint,
IARG_FUNCARG_ENTRYPOINT_VALUE, 0,
IARG_FUNCARG_ENTRYPOINT_VALUE, 1,
IARG_END);
}
}
RTN_Close(rtn);
/*!
* Print out the taint analysis results.
* This function is called when the application exits.
*/
VOID Fini(INT32 code, VOID *v)
{
DumpTaint();
out.close();
}
Software & Services Group
138
VOID DumpTaint() {
out << "======================================" << endl;
out << "Tainted Memory: " << endl;
set<ADDRINT>::iterator it;
for ( it=TaintedAddrs.begin() ; it != TaintedAddrs.end(); it++ )
{
out << " " << *it;
}
out << endl << "***" << endl << "Tainted Regs:" << endl;
}
for (int i=0; i < REG_LAST; i++) {
if (TaintedRegs[i]) {
out << REG_StringShort((REG)i);
} }
out << "======================================" << endl;
// This function marks the contents of argv[1] as tainted
VOID MainAddTaint(unsigned int argc, char *argv[]) {
if (argc != 2) return;
int n = strlen(argv[1]);
ADDRINT taint = (ADDRINT)argv[1];
for (int i = 0; i < n; i++) TaintedAddrs.insert(taint + i);
}
DumpTaint();
Software & Services Group
139
// This function represents the case of a register copied to memory
void RegTaintMem(ADDRINT reg_r, ADDRINT mem_w) {
out << REG_StringShort((REG)reg_r) << " --> " << mem_w << endl;
}
if (TaintedRegs[reg_r]) {
TaintedAddrs.insert(mem_w);
}
else //reg not tainted --> mem not tainted
{
if (TaintedAddrs.count(mem_w)) { // if mem is already not tainted nothing to do
TaintedAddrs.erase(TaintedAddrs.find(mem_w));
}
}
// this function represents the case of a memory copied to register
void MemTaintReg(ADDRINT mem_r, ADDRINT reg_w, ADDRINT inst_addr) {
out << mem_r << " --> " << REG_StringShort((REG)reg_w) << endl;
}
if (TaintedAddrs.count(mem_r)) //count is either 0 or 1 for set
{
TaintedRegs[reg_w] = true;
}
else //mem is clean -> reg is cleaned
{
TaintedRegs[reg_w] = false;
}
Software & Services Group
140
// this function represents the case of a reg copied to another reg
void RegTaintReg(ADDRINT reg_r, ADDRINT reg_w)
{
out << REG_StringShort((REG)reg_r) << " --> " <<
REG_StringShort((REG)reg_w) << endl;
}
TaintedRegs[reg_w] = TaintedRegs[reg_r];
// this function represents the case of an immediate copied to a register
void ImmedCleanReg(ADDRINT reg_w)
{
out << "const --> " << REG_StringShort((REG)reg_w) << endl;
}
TaintedRegs[reg_w] = false;
// this function represents the case of an immediate copied to memory
void ImmedCleanMem(ADDRINT mem_w)
{
out << "const --> " << mem_w << endl;
}
if (TaintedAddrs.count(mem_w)) //if mem is not tainted nothing to do
{
TaintedAddrs.erase(TaintedAddrs.find(mem_w));
}
Software & Services Group
141
Helpers
// True if the instruction has an immediate operand
// meant to be called only from instrumentation routines
bool INS_has_immed(INS ins);
// returns the full name of the first register operand written
REG INS_get_write_reg(INS ins);
// returns the full name of the first register operand read
REG INS_get_read_reg(INS ins)
Software & Services Group
142
/*!
* This function checks for each instruction if it does a mov that can potentially
* transfer taint and if true adds the approriate analysis routine to check
* and propogate taint at run-time if needed
* This function is called every time a new trace is encountered.
*/
VOID Trace(TRACE trace, VOID *v) {
for (BBL bbl = TRACE_BblHead(trace); BBL_Valid(bbl); bbl = BBL_Next(bbl)) {
for (INS ins = BBL_InsHead(bbl); INS_Valid(ins); ins = INS_Next(ins)) {
if ( (INS_Opcode(ins) >= XED_ICLASS_MOV) &&
(INS_Opcode(ins) <= XED_ICLASS_MOVZX) ) {
if (INS_has_immed(ins)) {
if (INS_IsMemoryWrite(ins)) { //immed -> mem
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)ImmedCleanMem,
IARG_MEMORYOP_EA, 0,
IARG_END);
}
else //immed -> reg
{
REG insreg = INS_get_write_reg(ins);
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)ImmedCleanReg,
IARG_ADDRINT, (ADDRINT)insreg,
IARG_END);
}
} // end of if INS has immed
else if (INS_IsMemoryRead(ins)) //mem -> reg
Software & Services Group
143
else if (INS_IsMemoryRead(ins)) { //mem -> reg
//in this case we call MemTaintReg to copy the taint if relevant
REG insreg = INS_get_write_reg(ins);
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)MemTaintReg,
IARG_MEMORYOP_EA, 0,
IARG_ADDRINT, (ADDRINT)insreg, IARG_INST_PTR,
IARG_END);
}
else if (INS_IsMemoryWrite(ins)) { //reg -> mem
//in this case we call RegTaintMem to copy the taint if relevant
REG insreg = INS_get_read_reg(ins);
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)RegTaintMem,
IARG_ADDRINT, (ADDRINT)insreg,
IARG_MEMORYOP_EA, 0,
IARG_END);
}
else if (INS_RegR(ins, 0) != REG_INVALID()) { //reg -> reg
//in this case we call RegTaintReg
REG Rreg = INS_get_read_reg(ins);
REG Wreg = INS_get_write_reg(ins);
INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)RegTaintReg,
IARG_ADDRINT, (ADDRINT)Rreg,
IARG_ADDRINT, (ADDRINT)Wreg,
IARG_END);
}
else { out << "serious error?!\n" << endl; }
} // IF opcode is a MOV
} // For INS
} // For BBL
Software & Services Group
} // VOID Trace
144
Part3 Summary
• Saw Examples of
– Allocating Pin Registers for Pin Tool Use
– Pin IF-THEN instrumentation
– Changing register values in instrumentation code
– Changing register values in CONTEXT
– Knobs
– Pin TLS
– Pin Buffering API
– Using Symbol and Debug Info
– Probe-Mode
– Multi-Threading support
Software & Services Group
145
Part4: Advanced Pin API
“To boldly go where few PinHeads have
gone before…”
• Agenda
– membuffer_threadpool tool
– Using multiple buffers in the Pin Buffering API
– Using Pin Tool Threads
– Using Pin and OS locks to synchronize threads
– System call instrumentation
– Instrumenting a process tree
– CONTEXT* and IARG_CONST_CONTEXT, IARG_CONTEXT
– Managing Exceptions and Signals
– Accessing Decode API
– Pin Code-Cache API
– Transparent debugging, and extending the debugger
Software & Services Group
146
membuffer_threadpool
• Recall membuffer_simple:
– Uses Pin Buffering API
– One buffer for each thread
– Inlined call to INS_InsertFillBuffer writes instrumentation data
into the buffer
– Application threads execute jitted application and instrumentation code
– When buffer becomes full the Pin Tool defined BufferFull function
is called (by the application thread)
– Process the data in the buffer
– After the buffer is processed it is set to be re-filled from the top
– Application thread continues executing jitted application and instrumentation
code
Software & Services Group
147
membuffer_threadpool
• Improvement: Process buffers that become full asynchronously, allows
application code to continue executing while buffers are being processed.
– Pin Buffering API supports multiple buffers per-thread
– Each application thread will allocate a number of buffers.
– The buffers allocated by the thread can only be used by the allocating thread so:
– Each application thread will have a buffers-free list, holding all buffers that are not currently
full or being filled.
– Pin supports creating Pin Tool threads, these are NOT jitted and can be used to do
Pin Tool work asynchronously.
– A number of these threads will be created, their job is:
–
–
Process buffers that become full. These will be located on a global full-buffers list.
After processing, return them to the buffers-free list of the application thread that filled them
– Application threads execute jitted application code and instrumentation code – the
instrumentation code writes data into the buffers and when it detects that the buffer
is full calls the BufferFull callback.
– The BufferFull callback function will NOT process the buffer
– Remember it is executed by an application thread
– It places the buffer on the global full-buffers list
– It retrieves a free buffer from this application thread’s free buffer list and returns it as the next
buffer to fill.
Software & Services Group
148
membuffer_threadpool
Application thread
Pin Tool Processing thread
Buffer being filled
buffers-free list
buffers-full list
Application thread
buffers-free list
Buffer being filled
Buffer
becomes
full
BufferFull
function
executed
Buffer
becomes
full
BufferFull
function
executed
Buffer
Processing
finishes.
Buffer
returned to
owner’s
buffers-free
list
Pin Tool Processing thread
Software & Services Group
149
membuffer_threadpool
int main(int argc, char *argv[])
PIN_Init(argc,argv);
{
// Pin TLS slot for holding the object that represents an application thread
appThreadRepresentitiveKey = PIN_CreateThreadDataKey(0);
// Define the buffer that will be used –
bufId = PIN_DefineTraceBuffer(sizeof(struct MEMREF), KnobNumPagesInBuffer,
BufferFull, // This Pin tool function will be called when buffer is full
0);
TRACE_AddInstrumentFunction(Trace, 0); // add an instrumentation callback function
// add callbacks
PIN_AddThreadStartFunction(ThreadStart, 0);
PIN_AddThreadFiniFunction(ThreadFini, 0);
PIN_AddFiniFunction(Fini, 0);
PIN_AddFiniUnlockedFunction(FiniUnlocked, 0); // Used for Pin Tool thread termination
/* It is safe to create internal threads in the tool's main procedure and spawn new
* internal threads from existing ones. All other places, like Pin callbacks and
* analysis routines in application threads, are not safe for creating internal threads. */
// NOTE: These threads are NOT jitted, Need to discuss when the threads actually start running
for (int i=0; i<KnobNumProcessingThreads; i++)
{
THREADID threadId; PIN_THREAD_UID threadUid;
threadId
= PIN_SpawnInternalThread (BufferProcessingThread, NULL, 0, &threadUid);
RecordToolThreadCreated(threadUid); /* Used for Pin Tool thread termination */ }
PIN_StartProgram(); /* Start the program, never returns */
}
Software & Services Group
150
membuffer_threadpool
static void RecordToolThreadCreated (PIN_THREAD_UID threadUid)
{ // Record the unique ID of the Pin Tool thread
uidSet.insert(threadUid);
}
// The thread function of Pin Tool threads – this code runs natively: NO Jitting
static VOID BufferProcessingThread(VOID * arg)
{
processingThreadRunning = TRUE; // Indicate that thread has started running
THREADID myThreadId = PIN_ThreadId();
}
while (!doExit)
{
VOID *buf;
UINT64 numElements;
APP_THREAD_REPRESENTITVE *appThreadRepresentitive;
// Get full buffer from the full buffer list
fullBuffersListManager.GetBufferFromList(&buf ,&numElements,
&appThreadRepresentitive, myThreadId);
if (buf == NULL)
{ // this will happen at process termination time – when there are NO
ASSERTX(doExit);
// no buffers left to process
break;
}
// Process the full buffer
ProcessBuffer(buf, numElements, appThreadRepresentitive);
// Put the processed buffer back on the free buffer list of the application thread that owns it
appThreadRepresentitive->FreeBufferListManager()
->PutBufferOnList(buf, 0, appThreadRepresentitive, myThreadId);
}
Software & Services Group
151
membuffer_threadpool
/*!
* Called by, instrumentation code, when a buffer fills up, by Pin when the thread exits, so the buffer
can be processed
* Called in the context of the application thread
* @param[in] id
buffer handle
* @param[in] tid
id of owning thread
* @param[in] ctxt
application context
* @param[in] buf
actual pointer to buffer
* @param[in] numElements
number of records
* @param[in] v
callback value
* @return A pointer to the buffer to resume filling.
*/
VOID * BufferFull(BUFFER_ID id, THREADID tid, const CONTEXT *ctxt, VOID *buf,
UINT64 numElements, VOID *v)
{
// get the APP_THREAD_REPRESENTITVE of this app thread from the Pin TLS
APP_THREAD_REPRESENTITVE * appThreadRepresentitive
= static_cast<APP_THREAD_REPRESENTITVE*>( PIN_GetThreadData(
appThreadRepresentitiveKey, tid ) );
// Enqueue the full buffer, on the full-buffers list, and get the next buffer to fill, from this
// thread’s free buffer list
VOID *nextBuffToFill
= appThreadRepresentitive->EnqueFullAndGetNextToFill(buf, numElements);
}
return (nextBuffToFill);
Software & Services Group
152
membuffer_threadpool
VOID * APP_THREAD_REPRESENTITVE::EnqueFullAndGetNextToFill(VOID *fullBuf,
UINT64 numElements)
// cannot wait for Pin Tool threads to start running since this may cause deadlock
// because this app thread may be holding some OS resource that the Pin Tool
// thread needs to obtain in order to start - e.g. the LoaderLock
if ( !processingThreadRunning)
{ // process buffer in this app thread
ProcessBuffer(fullBuf, numElements, this);
return fullBuf;
}
if (!_buffersAllocated)
// now allocate the rest of the KnobNumBuffersPerAppThread buffers to be used
for (int i=0; i<KnobNumBuffersPerAppThread-1; i++)
_freeBufferListManager->PutBufferOnList(PIN_AllocateBuffer(bufId), 0, this, _myTid);
_buffersAllocated = TRUE;
{
{
}
// put the fullBuf on the full buffers list, on the Pin Tool processing
// threads will pick it from there, process it, and then put it on this app-thread's free buffer list
fullBuffersListManager.PutBufferOnList(fullBuf, numElements, this, _myTid);
// return the next buffer to fill.
// It is always taken from the free buffers list of this app thread. If the list is empty then this app
// thread will be blocked until one is placed there (by one of the Pin Tool buffer processing threads).
VOID *nextBufToFill;
UINT64 numElementsDummy;
APP_THREAD_REPRESENTITVE *appThreadRepresentitiveDummy;
_freeBufferListManager->GetBufferFromList(&nextBufToFill,
&numElementsDummy,
&appThreadRepresentitiveDummy,
_myTid);
ASSERTX(appThreadRepresentitiveDummy = this);
return nextBufToFill;
}
Software & Services Group
153
membuffer_threadpool
VOID Instruction (INS ins, VOID *v)
{
UINT32 numMemOperands = INS_MemoryOperandCount(ins);
// Iterate over each memory operand of the instruction.
for (UINT32 memOp = 0; memOp < numMemOperands ; memOp++)
{ // Add the instrumentation code to write the appIP and memAddr
// of this memory operand into the buffer
// Pin will inline the code that writes to the buffer
INS_InsertFillBuffer(ins, IPOINT_BEFORE, bufId,
IARG_INST_PTR, offsetof(struct MEMREF, appIP),
IARG_MEMORYOP_EA, memOp,
offsetof(struct MEMREF, memAddr),
IARG_END);
}
}
Software & Services Group
154
membuffer_threadpool
class BUFFER_LIST_MANAGER
{
public:
BUFFER_LIST_MANAGER();
VOID PutBufferOnList (VOID *buf, UINT64 numElements, APP_THREAD_REPRESENTITVE *appThreadRepresentitive,
THREADID tid)
// build the list element
BUFFER_LIST_ELEMENT bufferListElement;
bufferListElement.buf = buf;
bufferListElement.numElements = numElements;
bufferListElement.appThreadRepresentitive = appThreadRepresentitive;
GetLock(&_bufferListLock, tid+1);
// lock the list, using a Pin lock
_bufferList.push_back(bufferListElement); // insert the element at the end of the list
ReleaseLock(&_bufferListLock);
// unlock the list
WIND::ReleaseSemaphore(_bufferSem, 1, NULL); // signal that there is a buffer on the list
{
}
VOID GetBufferFromList (VOID **buf ,UINT64 *numElements, APP_THREAD_REPRESENTITVE **appThreadRepresentitive,
THREADID tid){
WIND::WaitForSingleObject (_bufferSem, INFINITE);
// wait until there is a buffer on the list
GetLock(&_bufferListLock, tid+1);
// lock the list
BUFFER_LIST_ELEMENT &bufferListElement = (_bufferList.front()); // retrieve the first element of the list
*buf = bufferListElement.buf;
*numElements = bufferListElement.numElements;
*appThreadRepresentitive = bufferListElement.appThreadRepresentitive;
_bufferList.pop_front();
// remove the first element from the list
ReleaseLock(&_bufferListLock);
// unlock the list
}
VOID SignalBufferSem() {WIND::ReleaseSemaphore(_bufferSem, 1, NULL);}
UINT32 NumBuffersOnList () { return (_bufferList.size());}
private:
struct BUFFER_LIST_ELEMENT // structure of an element of the buffer list
VOID *buf;
UINT64 numElements;
APP_THREAD_REPRESENTITVE *appThreadRepresentitive; // the application thread that owns this buffer
{
};
WIND::HANDLE _bufferSem; // counting semaphore, value is #of buffers on the list, value==0 => WaitForSingleObject
blocks
PIN_LOCK _bufferListLock; // Pin Lock
list<const BUFFER_LIST_ELEMENT> _bufferList;
};
Software & Services Group
155
membuffer_threadpool
VOID ThreadFini(THREADID tid, const CONTEXT *ctxt, INT32 code, VOID *v)
{
// get the APP_THREAD_REPRESENTITVE of this app thread from the Pin TLS
APP_THREAD_REPRESENTITVE * appThreadRepresentitive
= static_cast<APP_THREAD_REPRESENTITVE*>(PIN_GetThreadData(
appThreadRepresentitiveKey, tid));
// wait for all my buffers to be processed
while(appThreadRepresentitive->_freeBufferListManager->NumBuffersOnList() !=
KnobNumBuffersPerAppThread-1)
PIN_Sleep(1);
}
delete appThreadRepresentitive;
PIN_SetThreadData(appThreadRepresentitiveKey, 0, tid);
static VOID FiniUnlocked(INT32 code, VOID *v)
{
BOOL waitStatus;
INT32 threadExitCode;
doExit = TRUE; // indicate that process is exiting
// signal all the Pin Tool threads to wake up and recognize the exit
for (int i=0; i<KnobNumProcessingThreads; i++)
fullBuffersListManager.SignalBufferSem();
// Wait until all Pin Tool threads exit
for (set<PIN_THREAD_UID>::iterator it = uidSet.begin(); it != uidSet.end(); ++it)
waitStatus = PIN_WaitForThreadTermination(*it, PIN_INFINITE_TIMEOUT, &threadExitCode);
}
Software & Services Group
156
System Call Instrumentation
VOID SyscallEntry(THREADID threadIndex, CONTEXT *ctxt, SYSCALL_STANDARD std, VOID *v)
{
ADDRINT appIP = PIN_GetContextReg(ctxt, REG_INST_PTR);
}
printf ("syscall# %d at appIP %x param1 %x param2 %x param3 %x param4 %x param5 %x
param6 %x\n",
PIN_GetSyscallNumber(ctxt, std), appIP,
PIN_GetSyscallArgument(ctxt, std, 0), PIN_GetSyscallArgument(ctxt, std, 1),
PIN_GetSyscallArgument(ctxt, std, 2), PIN_GetSyscallArgument(ctxt, std, 3),
PIN_GetSyscallArgument(ctxt, std, 4), PIN_GetSyscallArgument(ctxt, std, 5));
VOID SyscallExit(THREADID threadIndex, CONTEXT *ctxt, SYSCALL_STANDARD std, VOID *v)
{
printf(" returns: %x\n", PIN_GetSyscallReturn(ctxt, std);
}
int main(int argc, char *argv[])
{
PIN_Init(argc, argv);
// Instrument system calls via these Pin Callbacks and not via analysis functions
PIN_AddSyscallEntryFunction (SyscallEntry, 0);
PIN_AddSyscallExitFunction (SyscallExit, 0);
}
PIN_StartProgram(); // Never returns
Software & Services Group
157
Instrumenting a Process Tree
• Process A creates Process B
– Process B creates Process C and D
– And so forth
• Can use Pin to instrument all or part of the processes of a
process tree
– Use the –follow_exevc Pin invocation switch to turn this on
• Can use different Pin modes (Jit or Probe) on the different
processes in the process tree.
• Can use different Pin Tools on the different processes of a
process tree.
• Architecture of processes in the process tree may be
intermixed: e.g. Process A is 32bit, Process B is 64 bit,
Process C is 64 bit, Process D is 32 bit…
Software & Services Group
158
Instrumenting a Process Tree
// If this Pin Callback returns FALSE, then the child process will run Natively
BOOL FollowChild(CHILD_PROCESS childProcess, VOID * userData)
{
BOOL res; INT appArgc;
CHAR const * const * appArgv;
OS_PROCESS_ID pid = CHILD_PROCESS_GetId(childProcess);
// Get the command line that child process will be Pinned with, these are the Pin invocation switches
// that were specified when this (parent) process was Pinned
CHILD_PROCESS_GetCommandLine(childProcess, &appArgc, &appArgv);
// The Pin invocation switches of the child can be made to order
INT pinArgc = 0;
CHAR const * pinArgv[20];
:::: Put values in pinArgv, Set pinArgc to be the number of entries in pinArgv that are to be used
CHILD_PROCESS_SetPinCommandLine(childProcess, pinArgc, pinArgv);
return TRUE; /* Specify Child process is to be Pinned */
}
int main(INT32 argc, CHAR **argv) {
PIN_Init(argc, argv);
cout << " Process is running on Pin in " << PIN_IsProbeMode() ? " Probe " : " Jit " << " mode "
// The FollowChild Pin Callback will be called when the application being Pinned is about to spawn
// child process
PIN_AddFollowChildProcessFunction (FollowChild, 0);
if ( PIN_IsProbeMode() )
PIN_StartProgramProbed(); // Never returns
else
PIN_StartProgram();
}
Software & Services Group
159
CONTEXT*,
IARG_CONST_CONTEXT, IARG_CONTEXT
• CONTEXT* is a Handle to the full register context of the
application at a particular point in the execution
– CONTEXT* can NOT be dereferenced. It is a handle to be passed
to Pin API functions
• CONTEXT* is passed by default to a number of Pin
Callback functions: e.g.
– ThreadStart
registered by PIN_AddThreadStartFunction
– BufferFull
registered by PIN_DefineTraceBuffer
– OnContextChange registered by PIN_AddContextChangeFunction
Software & Services Group
160
CONTEXT* …
• Pin API functions supplied to Get and Set registers within the
CONTEXT
• Have Pin API functions to Get and Set FP context
• Can request CONTEXT* be passed to an analysis function by
requesting IARG_(CONST)_CONTEXT
• Requesting IARG_CONTEXT
– The analysis function will NOT be inlined
– The passing of the CONTEXT* is time consuming
• Passing IARG_CONST_CONTEXT is ~4X faster than
IARG_CONTEXT
– Contents of CONTEXT* passed for IARG_CONST_CONTEXT can NOT be
changed
Software & Services Group
161
CONTEXT* …
• Changes made to the contents of a CONTEXT*
– IARG_CONTEXT
– Changes made will be visible in subsequent PIN API calls made from
within the nesting of the analysis function
– Changes made will NOT be visible in the application context after
return from the analysis function
– Passed to PIN Callbacks
– Changes made will be visible in both of above
Software & Services Group
162
Function Replacement with register change
#include "pin.H"
void *FunctionReplacer (
CONTEXT * ctxt,
AFUNPTR pf_malloc, size_t size)
{
void * res;
CONTEXT writableContext, * context = ctxt;
if (TimeForRegChange())
{
PIN_SaveContext(ctxt, &writableContext); // need to copy the ctxt into a writable context
context = & writableContext;
PIN_SetContextReg(context , REG_GAX, 1);
}
PIN_CallApplicationFunction(context , PIN_ThreadId(),
CALLINGSTD_DEFAULT, pf_malloc,
PIN_PARG(void *), &res, PIN_PARG(size_t), size);
return res; }
VOID ImageLoad(IMG img, VOID *v) { // Pin callback. Registered by IMG_AddInstrumentFunction
RTN rtn = RTN_FindByName(img, “Function");
PROTO proto = PROTO_Allocate( PIN_PARG(void *), CALLINGSTD_DEFAULT,
"proto", PIN_PARG(size_t), PIN_PARG_END() );
}
RTN_ReplaceSignature (rtn, AFUNPTR(FunctionReplacer), IARG_PROTOTYPE, proto,
IARG_CONST_CONTEXT,
IARG_ORIG_FUNCPTR, IARG_FUNCARG_ENTRYPOINT_VALUE, 0, IARG_END);
int main(int argc, CHAR *argv[]) {
PIN_InitSymbols();
PIN_Init(argc,argv));
IMG_AddInstrumentFunction(ImageLoad, 0);
PIN_StartProgram();
}
Software & Services Group
163
memtrace_simple
KNOB<UINT32> KnobNumBytesInBuffer(KNOB_MODE_WRITEONCE, "pintool", "num_bytes_in_buffer",
"0x100000", "number of bytes in buffer");
APP_THREAD_REPRESENTITVE::APP_THREAD_REPRESENTITVE(THREADID myTid) {
_buffer = new char[KnobNumBytesInBuffer.Value()]; // Allocate the buffer
_numBuffersFilled = 0;
_numElementsProcessed = 0;
_myTid = myTid;
}
char * APP_THREAD_REPRESENTITVE::Begin() { return _buffer; }
char * APP_THREAD_REPRESENTITVE:: End()
{ return _buffer + KnobNumBytesInBuffer.Value(); }
VOID ThreadStart(THREADID tid,
CONTEXT *ctxt,
INT32 flags, VOID *v) // Pin callback on thread creation
// There is a new APP_THREAD_REPRESENTITVE object for every thread
APP_THREAD_REPRESENTITVE * appThreadRepresentitive
= new APP_THREAD_REPRESENTITVE(tid);
{
// A thread will need to look up its APP_THREAD_REPRESENTITVE, so save pointer in Pin TLS
PIN_SetThreadData(appThreadRepresentitiveKey, appThreadRepresentitive, tid);
// Initialize endOfTraceInBufferReg to point at beginning of buffer
PIN_SetContextReg(ctxt, endOfTraceInBufferReg,
reinterpret_cast<ADDRINT>(appThreadRepresentitive->Begin()));
// Initialize endOfBufferReg to point at end of buffer
PIN_SetContextReg(ctxt, endOfBufferReg,
reinterpret_cast<ADDRINT>(appThreadRepresentitive->End()));
}
Software & Services Group
164
Managing Exceptions and Signals
Software & Services Group
165
Exceptions
• Catch Exceptions that occur in Pin Tool code
– Global exception handler
– PIN_AddInternalExceptionHandler
– Guard code section with exception handler
– PIN_TryStart
– PIN_TryEnd
Software & Services Group
166
Exceptions
VOID InstrumentDivide(INS ins, VOID* v)
{
if ((INS_Mnemonic(ins) == "DIV") &&
(INS_OperandIsReg(ins, 0)))
{ // Will Emulate div instruction with register operand
INS_InsertCall(ins,
IPOINT_BEFORE,
AFUNPTR(EmulateIntDivide),
IARG_REG_REFERENCE, REG_GDX,
IARG_REG_REFERENCE, REG_GAX,
IARG_REG_VALUE, REG(INS_OperandReg(ins, 0)),
IARG_CONST_CONTEXT,
IARG_THREAD_ID,
IARG_END);
INS_Delete(ins); // Delete the div instruction
}
int main(int argc, char * argv[])
{
PIN_Init(argc, argv);
INS_AddInstrumentFunction (InstrumentDivide, 0);
PIN_AddInternalExceptionHandler (GlobalHandler, NULL); // Registers a Global Exception Handler
PIN_StartProgram(); // Never returns
return 0;
}
Software & Services Group
167
Exceptions
EXCEPT_HANDLING_RESULT DivideHandler (THREADID tid, EXCEPTION_INFO * pExceptInfo,
PHYSICAL_CONTEXT * pPhysCtxt, // The context when the exception
// occurred
VOID *appContextArg
// The application context when the
// exception occurred
)
{
if(PIN_GetExceptionCode(pExceptInfo) == EXCEPTCODE_INT_DIVIDE_BY_ZERO)
{ // Divide by zero occurred in the code emulating the divide, use PIN_RaiseException to raise this exception
// at the appIP – for handling by the application
cout << " DivideHandler : Caught divide by zero." << PIN_ExceptionToString(pExceptInfo) << endl;
// Get the application IP where the exception occurred from the application context
CONTEXT * appCtxt = (CONTEXT *)appContextArg;
ADDRINT faultIp = PIN_GetContextReg (appCtxt, REG_INST_PTR);
// raise the exception at the application IP, so the application can handle it as it wants to
PIN_SetExceptionAddress (pExceptInfo, faultIp);
PIN_RaiseException (appCtxt, tid, pExceptInfo); // never returns
}
return EHR_CONTINUE_SEARCH;
}
VOID EmulateIntDivide(ADDRINT * pGdx, ADDRINT * pGax, ADDRINT divisor, CONTEXT * ctxt, THREADID tid)
{
PIN_TryStart(tid, DivideHandler, ctxt); // Register a Guard Code Section Exception Handler
UINT64 dividend = *pGdx;
dividend <<= 32;
dividend += *pGax;
*pGax = dividend / divisor;
*pGdx = dividend % divisor;
PIN_TryEnd(tid);
/* Guarded Code Section ends */
}
Software & Services Group
168
Exceptions
EXCEPT_HANDLING_RESULT GlobalHandler(THREADID threadIndex, EXCEPTION_INFO * pExceptInfo,
PHYSICAL_CONTEXT * pPhysCtxt, VOID *v)
{ // Any Exception occurring in Pin Tool, or Pin that is not in a Guarded Code Section will cause this function to be
// executed
cout << "GlobalHandler: Caught unexpected exception. " << PIN_ExceptionToString(pExceptInfo) << endl;
return EHR_UNHANDLED;
}
Software & Services Group
169
Exceptions, Monitoring Application
Exceptions
• PIN_AddContextChangeFunction
– Can monitor and change that application state at
application exceptions
int main(int argc, char **argv)
{
PIN_Init(argc, argv);
PIN_AddContextChangeFunction(OnContextChange, 0);
PIN_StartProgram();
}
Software & Services Group
170
Exceptions, Monitoring Application
Exceptions
static void OnContextChange (THREADID tid,
CONTEXT_CHANGE_REASON reason,
const CONTEXT *ctxtFrom // Application's register state at exception point
CONTEXT *ctxtTo,
// Application's register state delivered to handler
INT32 info,
VOID *v)
{
if (CONTEXT_CHANGE_REASON_SIGRETURN == reason
|| CONTEXT_CHANGE_REASON_APC == reason
|| CONTEXT_CHANGE_REASON_CALLBACK == reason
|| CONTEXT_CHANGE_REASON_FATALSIGNAL == reason
|| ctxtTo == NULL)
{ // don't want to handle these
return;
}
// CONTEXT_CHANGE_REASON_EXCEPTION
// change some register values in the context that the application will see at the handler
FPSTATE fpContextFromPin;
// change the bottom 4 bytes of xmm0
PIN_GetContextFPState (ctxtFrom, &fpContextFromPin);
fpContextFromPin.fxsave_legacy._xmm[3] = 'de';
fpContextFromPin.fxsave_legacy._xmm[2] = 'ad';
fpContextFromPin.fxsave_legacy._xmm[1] = 'be';
fpContextFromPin.fxsave_legacy._xmm[0] = 'ef';
PIN_SetContextFPState (ctxtTo,
&fpContextFromPin);
}
// change eax
PIN_SetContextReg(ctxtTo, REG_RAX, 0xbaadf00d);
Software & Services Group
171
Signals
• Establish an interceptor function for signals
delivered to the application
– Tools should never call sigaction() directly to handle
signals.
– function is called whenever the application receives the
requested signal, regardless of whether the application has
a handler for that signal.
– function can then decide whether the signal should be
forwarded to the application
Software & Services Group
172
Signals
• A tool can take over ownership of a signal in order
to:
– use the signal as an asynchronous communication
mechanism to the outside world.
– For example, if a tool intercepts SIGUSR1, a user of the tool could
send this signal and tell the tool to do something. In this usage
model, the tool may call PIN_UnblockSignal() so that it will receive the
signal even if the application attempts to block it.
– "squash" certain signals that the application generates.
– a tool that forces speculative execution in the application may want to
intercept and squash exceptions generated in the speculative code.
• A tool can set only one "intercept" handler for a
particular signal, so a new handler overwrites any
previous handler for the same signal. To disable a
handler, pass a NULL function pointer.
Software & Services Group
173
Signals
BOOL EnableInstrumentation = FALSE;
BOOL SignalHandler(THREADID, INT32, CONTEXT *, BOOL, const EXCEPTION_INFO *, void *)
{
// When tool receives the signal, enable instrumentation. Tool calls
// PIN_RemoveInstrumentation() to remove any existing instrumentation from Pin's code cache.
EnableInstrumentation = TRUE;
PIN_RemoveInstrumentation();
return FALSE; /* Tell Pin NOT to pass the signal to the application. */
VOID Trace(TRACE trace, VOID *)
if (!EnableInstrumentation)
return;
}
{
for (BBL bbl = TRACE_BblHead(trace); BBL_Valid(bbl); bbl = BBL_Next(bbl))
BBL_InsertCall(bbl, IPOINT_BEFORE, AFUNPTR(AnalysisFunc), IARG_INST_PTR, IARG_END);}
int main(int argc, char * argv[])
PIN_Init(argc, argv);
{
PIN_InterceptSignal(SIGUSR1, SignalHandler, 0); // Tool should really determine which signal is NOT in
// use by application
PIN_UnblockSignal(SIGUSR1, TRUE);
TRACE_AddInstrumentFunction(Trace, 0);
PIN_StartProgram();
}
Software & Services Group
174
Accessing the Decode API
• The decoder/encoder used is called XED
– http://www.pintool.org/docs/24110/Xed/html/
• Tool code can use the XED API
– E.g. decode an instruction inside an analysis routine.
Software & Services Group
175
Accessing the Decode API
extern "C" {
#include "xed-interface.h"
}
static VOID PIN_FAST_ANALYSIS_CALL
MemoryOverWriteAt ( // Pin will NOT inline this function, it is the THEN part
ADDRINT appIP, ADDRINT memWriteAddr, UINT32 numBytesWritten)
{
INT32 column, lineNum;
string fileName;
PIN_GetSourceLocation (appIP, &column, &line, &fileName);
static const xed_state_t dstate = { XED_MACHINE_MODE_LEGACY_32, XED_ADDRESS_WIDTH_32b};
xed_decoded_inst_t xedd;
xed_decoded_inst_zero_set_mode (&xedd, &dstate);
xed_error_enum_t xed_code = xed_decode (&xedd, reinterpret_cast<UINT8*>(appIP), 15);
char buf[256];
xed_decoded_inst_dump_intel_format(&xedd, buf, 256, appIP);
}
printf ("overwrite of %p from instruction at %p %s originating from file %s line %d col %d\n",
KnobMemAddrBeingOverwritten, appIP, buf, fileName.c_str(), lineNum, column);
printf (" writing %d bytes starting at %p\n", numBytesWritten, memWriteAddr);
Software & Services Group
176
Pin Code-Cache API
• The Code-Cache API allows a Pin Tool to:
– Inspect Pin's code cache and/or alter the code cache
replacement policy
– Assume full control of the code cache
– Remove all or selected traces from the code cache
– Monitor code cache activity, including start/end of
execution of code in the code cache
Software & Services Group
177
Pin Code-Cache API
VOID DoSmcCheck(VOID * traceAddr, VOID * traceCopyAddr, USIZE traceSize, CONTEXT * ctxP) {
if (memcmp(traceAddr, traceCopyAddr, traceSize) != 0) /* application code changed */
{
// the jitted trace is no longer valid
free(traceCopyAddr);
CODECACHE_InvalidateTraceAtProgramAddress((ADDRINT)traceAddr);
PIN_ExecuteAt(ctxP); /* Continue jited execution at this application trace */
} }
VOID InstrumentTrace(TRACE trace, VOID *v)
VOID * traceAddr;
VOID * traceCopyAddr;
{
USIZE traceSize;
traceAddr = (VOID *)TRACE_Address(trace); // The appIP of the start of the trace
traceSize = TRACE_Size(trace);
traceCopyAddr = malloc(traceSize);
// The size of the original application trace in bytes
if (traceCopyAddr != 0)
{
memcpy(traceCopyAddr, traceAddr, traceSize);
// Copy of original application code in trace
// Insert a call to DoSmcCheck before every trace
TRACE_InsertCall(trace, IPOINT_BEFORE, (AFUNPTR)DoSmcCheck,
IARG_PTR, traceAddr,
IARG_PTR, traceCopyAddr,
IARG_UINT32 , traceSize,
IARG_CONTEXT,
IARG_END);
}
}
int main(int argc, char * argv[]) {
PIN_Init(argc, argv);
TRACE_AddInstrumentFunction(InstrumentTrace, 0);
PIN_StartProgram();
}
Software & Services Group
178
Transparent debugging, and extending
the debugger
• Transparently debug the application while it is
running on Pin + Pin Tool
– PinADX: Customizable Debugging with Dynamic
Instrumentation CGO2012
• Use Pin Tool to enhance/extend the debugger
capabilities
– Watchpoint: Is order of magnitude faster when
implemented using Pin Tool
– See previous “Symbols: Accessing Application Debug Info from
a Pin Tool”
– Which branch is branching to address 0
– Easy to write a Pin Tool that implements this
Software & Services Group
179
Debug Application while Running
Pin
•Useful for Pin-based emulators
– User can debug application while emulating
•Provide advanced debugging features with Pin:
– Stack monitoring features
– Buffer overrun detection
– Reverse debugging
– Write your own debugger extension via Pin
Software & Services Group
180
Naïve Solution Won’t Work
Tool
?
Application
GDB
Pin
Pined process
• Why can’t we just debug normally?
– Debugger sees Pin state, not application state
– Pin recompiles application code
– Instructions wrong, registers wrong, PC wrong, …
Software & Services Group
181
Pin Debugger Interface
GDB remote
protocol (tcp)
Application
Debug
Tool
GDB
Pin
Agent
(unmodified)
Pin process
• GDB debugs application (not Pin itself)
• Leverage GDB remote protocol ABI
Software & Services Group
182
Debug the Application with Pin
1.
Run Pin with -appdebug
$ pin -appdebug -t tool.so -- ./application
Application stopped until continued from debugger.
Start GDB, then issue this command at the (gdb) prompt:
target remote :1234
2.
Start GDB, enter “target remote …”
$ gdb ./application
(gdb) target remote :1234
3.
Set breakpoints, etc. Continue with “cont”
(gdb) break main
(gdb) cont
Software & Services Group
183
Extending the Debugger
•Normal debugging with Pin useful but limited
•Extending the debugger:
– Add GDB commands via a Pin tool
– Stop at “semantic breakpoint” via
instrumentation
Software & Services Group
184
Pintool 4: Stack Debugger
$ pin -appdebug -t stack-debugger.so -./app
Commands implemented
$ gdb ./app
in Pintool
(gdb) target remote :1234
(gdb) monitor stackbreak 4000
Break when thread uses 4000 stack bytes.
(gdb) cont
Thread uses 4004 stack bytes.
[…]
(gdb) monitor stats
Maximum stack usage: 8560 bytes.
Software & Services Group
185
Stack-Debugger Instrumentation
•Thread Start:
StackBase = %esp;
MaxStack = 0;
After each stack-changing
instruction
•[…]
•sub $0x60, %esp
size = StackBase - %esp;
if (size > MaxStack) MaxStack = size;
if (size > StackLimit) TriggerBreakpoint();
•cmp %esi, %edx
•jle <L1>
Software & Services Group
186
ManualExamples/stack-debugger.cpp
instrumentation routine
VOID Instruction(INS ins, VOID *)
{
if (INS_RegWContain(ins, REG_STACK_PTR))
{
IPOINT where = (INS_HasFallThrough(ins)) ?
IPOINT_AFTER : IPOINT_TAKEN_BRANCH;
INS_InsertCall(ins, where, (AFUNPTR)OnStackChange,
IARG_REG_VALUE, REG_STACK_PTR,
IARG_THREAD_ID, IARG_CONST_CONTEXT, IARG_END);
}
Triggers debugger
}
breakpoint
VOID OnStackChange(ADDRINT sp, THREADID tid, CONTEXT *ctxt)
{
size_t size = StackBase - sp;
analysis routine
if (size > StackMax) StackMax = size;
Insert only
after >instructions
if (size
StackLimit) {
ostringstream os;
that write to
os%esp
<< "Thread uses " << size << " stack bytes.";
PIN_ApplicationBreakpoint(ctxt, tid, FALSE, os.str());
}
}
Software & Services Group
187
ManualExamples/stack-debugger.cpp
int main() {
[…]
PIN_AddDebugInterpreter(HandleDebugCommand, 0);
}
BOOL HandleDebugCommand(const string &cmd, string *result) {
if (cmd == "stats")
{
Hooks the GDB “monitor” command. E.g.:
ostringstream os;
(gdb) monitor
stats
os << "Maximum stack usage: " << StackMax << " bytes.\n";
text
“monitor”
(gdb)Receives
monitor
stackbreak
4000
*result
= after
os.str();
This string written to GDB session
return TRUE;
}
else if (cmd.find("stackbreak ") == 0) {
StackLimit = /* parse limit */;
ostringstream os;
os << "Break when thread uses " << StackLimit << " stack
bytes.";
*result = os.str();
return TRUE;
}
return FALSE;
// Unknown command
}
Software & Services Group
188
Other Debugger Tools
•
•
•
•
Breakpoint on buffer overrun
Debug from a recorded log file
Reverse debugging from a recording
Design your own custom debugger tool
Software & Services Group
189
Pin Debugger Internals
Software & Services Group
190
Pin Debugger Interface
GDB remote
protocol(tcp)
Application
Debug
Pin
Tool
GDB
Agent
(unmodified)
Pin process
Software & Services Group
191
Communication Details
Commands
• Read / write registers, memory
• Set breakpoints
• Continue, single-step, stop
GDB
Notifications
Debug
Pin
Agent
• Breakpoint triggered
• Caught signal
• Application exited
•
•
•
•
•
Very low level
Symbol processing in GDB
Expression evaluation in GDB
Reusing GDB’s remote debugging interface
Many debuggers have interface like this
Software & Services Group
192
Commands / Notifications
Virtualized
•Commands
• Read register -> read virtualized app register
• Set breakpoint -> set virtual breakpoint
• Single step -> step VM one app instruction
•Notifications
• Signal notification -> on virtual signal in VM
• Breakpoint notification -> on virtual BP in VM
Software & Services Group
193
Single Step
Original code
Code cache
1
1’
2
5
3
Execution stops in Pin
Waits for GDB to continue
6
4
Pin
do single-step
step
complete
notification
GDB
Software & Services Group
194
Breakpoint
Original code
Code cache
1
1’
2
2’
5
3
3’
6
BP
4
Execution stops in Pin
Waits for GDB to continue
Pin
set
continue
breakpoint
breakpoint
at
notification
4
GDB
Software & Services Group
195
Part4 Summary
• Boldly went where no pin head has gone before…
– Lived to tell the tail
– membuffer_threadpool tool
– Using multiple buffers in the Pin Buffering API
– Using Pin Tool Threads
– Using Pin and OS locks to synchronize threads
– System call instrumentation
– Instrumenting a process tree
– CONTEXT* and IARG_CONTEXT
– Managing Exceptions and Signals
– Decode API
– Pin Code-Cache API
– Transparent debugging, and extending the debugger
Software & Services Group
196
Part5 Performance #s
Software & Services Group
197
•Pin Performance (Windows)
Software & Services Group
198
•Pin Performance (Windows)
Software & Services Group
199
Scalability of Workloads on Pin Scalability
Native
CINEBENCH
POV-Ray
No Instrumentation
Speedup (Relative to Single Thread)
Speedup (Relative to Single Thread)
4.0X
3.5X
3.0X
2.5X
2.0X
1.5X
1.0X
0.5X
0.0X
4.0X
BBCount
3.5X
MemTrace
MemError
3.0X
2.5X
2.0X
1.5X
1.0X
0.5X
0.0X
2 threads
4 threads
2 threads
4 threads
• Pin VMM serialization does not impact scalability of the application
– Execution in the code cache is not serialized
• Scalability may drop due to limited memory bandwidth (MemTrace) or
contention for tool private data (MemError)
Software & Services Group
200
Kernel Interaction Overhead
Slowdown Relative to Native per Kernel Interaction
System
Calls
• assa
Exceptions
APCs
Callbacks
12X
10.5X
3X
Cost of a trip in Pin VMM for each system call is high
– ~3000 cycles in VMM vs. ~500 cycles for ring crossing
– Future work: a faster path in VMM for system calls
1.8X
Kernel Interaction Counts
Illustrator
System Calls
Excel
CINEBENCH
POV-Ray
1,659,298
658,683
101,700
75,313
Exceptions
1
0
0
0
APCs
6
6
24
24
Callbacks
73,062
68,767
961
7,682
Overhead vs. Total
Runtime
3.3%
2.8%
<1%
<1%
• Total overhead for handling kernel interactions is relatively low
– Kernel interactions are infrequent for majority of applications
Software & Services Group
201
Overall Summary
• Pin is Intel’s dynamic binary instrumentation engine
• Pin can be used to instrument all user level code
–
–
–
–
–
–
Windows, Linux
IA-32, Intel64, IA64
Product level robustness
Jit-Mode for full instrumentation: Thread, Function, Trace, BBL, Instruction
Probe-Mode for Function Replacement/Wrapping/Instrumentation only.
Pin supports multi-threading, no serialization of jitted application nor of instrumentation code
• Pin API makes Pin tools easy to write
–
Presented many tools, many fit on 1 ppt slide
• Pin performance is good
–
Pin APIs provide for writing efficient Pin tools
• Popular and well supported
–
30,000+ downloads, 400+ citations
• Free DownLoad
–
–
www.pintool.org
Includes: Detailed user manual, source code for 100s of Pin tools, tutorials
• Pin User Group
–
–
http://tech.groups.yahoo.com/group/pinheads/
Pin users and Pin developers answer questions
Software & Services Group
202