Verdana Bold 30

Pin: Intel’s Dynamic Binary Instrumentation Engine Pin Tutorial Intel Corporation Presented By: Robert Cohn Tevi Devor CGO 2010 1 Software & Services Group 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 2 Part1: Introduction to Pin • Dynamic Binary Instrumentation • Pin Capabilities • Overview of how Pin works • Sample Pin Tools Software & Services Group 3 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 – Use dynamic code generation to make a less intrusive profile guided optimization and instrumentation system – Pin is a small Spike – Spike is EOL http://www.cgo.org/cgo2004/papers/01_82_luk_ck.pdf Software & Services Group 4 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 5 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: – Supports IA-32, Intel64, IA-64 – Supports Linux, Windows, MacOS • Robust: – – – – Can instrument real-life applications: Database, web browsers, … Can instrument multithreaded applications Supports signals and exceptions, self modifying code… If you can Run it – you can Pin it • Efficient: Pin can be used to instrument all the user level code – Applies compiler optimizations on instrumentation code in an application Software & Services Group 6 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, and 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 7 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 8 Example Pin-tools Cache Simulation CMP$IM Instruction Emulation MT Workload Capture & Deterministic Replay PinPlay (new instructions) SDE Pin Simulation Trace Region Generation Selection SDE: http://software.intel.com/en-us/articles/intel-software-development-emulator pinLIT PinPoints CMP$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 9 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 10 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 11 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 12 Instruction Counting Tool (inscount.dll) #include "pin.h" UINT64 icount = 0; Execution time routine void docount() { icount++; } void Instruction(INS ins, void *v) Jitting { INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)docount, IARG_END); } time 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 13 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 14 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 15 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){{ 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 16 2 22 40 37 17 APP IP 0x77ec4600 0x77ec4603 0x77ec4609 0x77ec460d cmp jz movzx call rax, rdx 0x77f1eac9 ecx, [rax+0x2] 0x77ef7870 20 0x001de0000 mov r14, 0xc5267d40 //inscount2.docount 58 0x001de000a add [r14], 0x2 //inscount2.docount 2 0x001de0015 0x77ec4600 cmp rax, rdx 9 0x001de0018 jz 0x1deffa0 (L1) //patched in future 52 0x001de001e mov r14, 0xc5267d40 //inscount2.docount 29 0x001de0028 mov [r15+0x60], rax 57 0x001de002c lahf save status 37 0x001de002e seto al flags 50 0x001de0031 mov [r15+0xd8], ax 30 0x001de0039 mov rax, [r15+0x60] 12 0x001de003d add [r14], 0x2 //inscount2.docount 40 0x001de0048 0x77ec4609 movzx edi, [rax+0x2] //ecx alloced to edi 22 0x001de004c push 0x77ec4612 //push retaddr 61 0x001de0051 nop 17 0x001de0052 jmp 0x1deffd0 (L2)//patched in future (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 0x001deffc8_iaddr(0x077f1eac9) // app target IP of jz (L2) // at 0x77ec4603 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 Software & Services Group // call at 0x77ec460d SimpleExamples/inscount2_mt.cpp #include "pin.H" INT32 numThreads = 0; const INT32 MaxNumThreads = 10000; struct THREAD_DATA { 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 ("Count[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 18 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 19 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)) { disAssemblyMap[INS_Address(ins)] = INS_Disassemble(ins); INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR) ReadsMem, IARG_INST_PTR,// application IP IARG_MEMORYREAD_EA, IARG_MEMORYREAD_SIZE, IARG_END); } } Pin has determined that it can overwrite ecx int main(int argc, char * argv[]) { PIN_Init(argc, argv); INS_AddInstrumentFunction(Instruction, 0); 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 PIN_StartProgram(); } Software & Services Group 20 Malloc Replacement #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") || strstr(IMG_Name(img).c_str(), "MSVCR80") || strstr(IMG_Name(img).c_str(), "MSVCR90")) { RTN mallocRtn = RTN_FindByName(img, "malloc"); PROTO protoMalloc = PROTO_Allocate( PIN_PARG(void *), CALLINGSTD_DEFAULT, "malloc", PIN_PARG(size_t), PIN_PARG_END() ); }} RTN_ReplaceSignature(mallocRtn, AFUNPTR(MallocWrapper), IARG_PROTOTYPE, protoMalloc, IARG_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 21 Pin Probe-Mode • Probe mode is a method of using Pin to wrap or replace application functions with functions in the tool. A jump instruction (probe), which redirects the flow of control to the replacement function is placed at the start of the specified function. • The bytes being overwritten are relocated, so that Pin can provide the replacement function with the address of the first relocated byte. This enables the replacement function to call the replaced (original) function. • In probe mode, the application and the replacement routine are run natively (not Jitted). This improves performance, but puts more responsibility on the tool writer. Probes can only be placed on RTN boundaries, and should 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 22 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 23 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 24 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 Execute it 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 25 #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 26 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 27 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 28 Part2: Larger Pin tools and writing efficient Pin tools Software & Services Group 29 CMP$im – A CMP Cache Simulation Pin Tool WORK LOAD Modeling an 8-core CMP using CMP$im Instrumentation Routines PIN • ThreadID • Address, Size • Access Type ThreadID, Address, Size, Access Type DL1 DL1 DL1 DL1 DL1 DL1 DL1 DL1 L2 L2 L2 L2 L2 L2 L2 L2 LLC LLC INTERCONNECT Cache model • Params to configure # cache levels, size, threads/cache etc 30 LLC LLC LLC LLC LLC LLC PRIVATE LLC/SHARED BANKED LLC Software & Services Group CMP$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$im – Instrument Memory References Software & Services Group 31 CMP$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]); • } •} 32 INSTR ROUTINES CacheHierarchy[MAX_NUM_THREADS][MAX_NUM_LEVELS]; MAIN •CACHE_t ANALYSIS ROUTINES •#include “cache_model.h” Synchronization Software & Services Group point 2MB LLC Cache Behavior – 4 Threads, AMMP 10 mil phase 0 Private Cache Miss Rate cumulative 0 Shared Cache Miss Rate • Miss Rate: • Shared caches have better hit rate when compared to private caches – Private Cache: 75% – Shared Cache: 50% Instruction Count (billions) Software & Services Group 33 Shared Refs & Shared Caches… Cache Miss % Total Accesses A 1 Thread 2 Thread 3 Thread (4 Threaded Run) B • Workloads have • Miss Rate Private LLC Miss Rate GeneNet – 16MB LLC Shared LLC 34 4 Thread different phases of execution Shared caches BETTER during phases when shared data is referenced frequently HPCA’06: Jaleel et al. Software & Services Group 34 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 35 a documented data race in the art benchmark is detected Software & Services Group 36 36 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 37 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 38 Example Program execution Thread T1 Thread T2 1: Store A DirEntry: [A:D] Last writer id: 1: Load F WAW T1: 2: Store A 3: Load F Last writer id: WAR 3: Store F T2: 2 DirEntry: [E:H] RAW 2: Load A 12 T2 T1 T1: 3 T2: T1 13 Last_writer SMO logs: T1 2 T2 2 T1 3 T2 3 T2 2 T1 1 Last access to the DirEntry Thread T2 cannot execute memory reference 2 until T1 executes its memory reference 1 Thread T1 cannot execute memory reference 2 until T2 executes its memory reference 2 Software & Services Group 39 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 40 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 41 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 42 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 43 Counting control flow edges jne 60 40 100 ret 40 40 call jmp 60 jne 1 Software & Services Group 44 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) { • if (INS_IsBranchOrCall(ins)) • { • INS_InsertCall(ins, IPOINT_BEFORE, (AFUNPTR)docount2, • IARG_INST_PTR, IARG_BRANCH_TARGET_ADDR, • IARG_BRANCH_TAKEN, IARG_END); • } •} •... Instrumentation Software & Services Group 45 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) { • 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); •} •… Instrumentation Software & Services Group 46 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 47 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 48 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 49 Reducing Work for Analysis Transitions •Reduce number of arguments to analysis routines – Inline analysis routines – Pass arguments in registers – Instrumentation scheduling Software & Services Group 50 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) Software & Services Group 51 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 52 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 53 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_CONTEXT) • Pintool writer breaks analysis routine into two: – INS_InsertIfCall (ins, …, (AFUNPTR)doif, …) – INS_InsertThenCall (ins, …, (AFUNPTR)dothen, …) Software & Services Group 54 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 55 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 56 Optimizing Your Pintools Summary • Baseline Pin has fairly low overhead (~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 57 Part3: Deeper into Pin API • Agenda – memtrace_simple tool – membuffer_simple tool – branchbuffer_simple tool – Symbols DebugInfo – Probe-Mode – Multi-Threading Software & Services Group 58 memtrace_simple • Tool code collects pairs of {appIP, memAddr} of memory accessing instructions into a per-thread buffer – Remember: It is the application thread(s) that execute the Pin Tool code. • When the buffer becomes full, tool code processes the entries in the buffer, resets the collection to start at the beginning of the buffer, then execution continues – and so-forth. • Is a representative of many memory-trace processing tools Software & Services Group 59 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 60 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 61 memtrace_simple Trace If endOf(Previous)TraceReg + TotalSizeOccupiedByTraceInBuffer > endOfBufferReg Then Buffer Call BufferFull endOfTraceReg += TotalSizeOccupiedByTraceInBuffer Early Exit appIP memAddr Trace Exit appIP memAddr endOfTraceReg TotalSizeO ccupiedBy TraceInBu ffer appIP Non memory access ins Instrumentation code for following memory access ins Memory access ins memAddr endOfBufferReg Software & Services Group 62 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 TotalSizeO ccupiedBy TraceInBu ffer endOfBufferReg Software & Services Group 63 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 64 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 65 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 66 memtrace_simple 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 67 memtrace_simple static ADDRINT PIN_FAST_ANALYSIS_CALL APP_THREAD_REPRESENTITVE::CheckIfNoSpaceForTraceInBuffer ( // Pin will inline this function char * endOfPreviousTraceInBuffer, char * bufferEnd, ADDRINT totalSizeOccupiedByTraceInBuffer) { return (endOfPreviousTraceInBuffer + totalSizeOccupiedByTraceInBuffer >= bufferEnd); } static char * PIN_FAST_ANALYSIS_CALL APP_THREAD_REPRESENTITVE::BufferFull ( // Pin will NOT inline this function char *endOfTraceInBuffer, ADDRINT tid) { // Get this thread’s APP_THREAD_REPRESENTITVE from the Pin TLS APP_THREAD_REPRESENTITVE * appThreadRepresentitive = static_cast<APP_THREAD_REPRESENTITVE*> (PIN_GetThreadData(appThreadRepresentitiveKey, tid)); appThreadRepresentitive->ProcessBuffer(endOfTraceInBuffer); } // 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 68 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)); _currentOffsetFromTraceStartInBuffer += sizeof(MEMREF); _numAnalysisCallsNeeded++; } void InsertAnalysisCalls(); private: INT32 _currentOffsetFromTraceStartInBuffer; INT32 _numAnalysisCallsNeeded; vector<ANALYSIS_CALL_INFO> _analysisCalls; }; 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); } } • 69 Software & Services Group 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 70 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 71 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 72 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 73 membuffer_simple /* * Pin Callback called, by application thread, when a buffer fills up, or 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 74 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 75 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 76 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 77 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 78 Symbols • PIN_InitSymbols() – Pin will use whatever symbol information is available – – – – Debug info in the app Pdb files Export Tables On Windows uses dbghelp • 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 79 Symbols: Instrument malloc and free int main(int argc, char *argv[]) { // Initialize pin symbol manager PIN_InitSymbols(); 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 80 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 81 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 82 Symbols: Accessing Application Debug Info from a Pin Tool 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 83 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 84 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 85 Pin Probe-Mode • Probe mode is a method of using Pin to wrap or replace application functions with functions in the tool. A jump instruction (probe), which redirects the flow of control to the replacement function is placed at the start of the specified function. • The bytes being overwritten are relocated, so that Pin can provide the replacement function with the address of the first relocated byte. This enables the replacement function to call the replaced (original) function. • In probe mode, the application and the replacement routine are run natively (not Jitted). This improves performance, but puts more responsibility on the tool writer. Probes can only be placed on RTN boundaries, and should 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 86 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 Original function entry point: Entry point overwritten with probe: push 0x41481064 %ebp 0x400113d4: jmp 0x400113d5: mov %esp,%ebp 0x400113d7: push %edi 0x400113d8: push %esi 0x400113d9: push %ebx // tool wrapper func :::::::::::::::::::: 0x414827fe: call 0x50000004 // call original func Software & Services Group 87 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 88 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 89 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 90 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 Software & Services Group 91 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 92 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 93 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 94 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 Part3 Software & Services Group 95 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 96 Part4: Advanced Pin API “To boldly go where no PinHead has 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_CONTEXT – Managing Exceptions and Signals – Accessing Decode API – Pin Code-Cache API – Transparent debugging, and extending the debugger Software & Services Group 97 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 callback 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 – All Pin callbacks are serialized – Only one buffer is being processed at any time Software & Services Group 98 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 99 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 100 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 101 membuffer_threadpool static void RecordToolThreadCreated (PIN_THREAD_UID threadUid) { // Record the unique ID of the Pin Tool thread BOOL insertStatus; insertStatus = (uidSet.insert(threadUid)).second; } // 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 102 membuffer_threadpool /*! Pin Callback * Called by, instrumentation code, when a buffer fills up, or 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 103 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 104 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 105 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 106 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 107 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 108 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 109 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 modified 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 110 CONTEXT* and IARG_CONTEXT • CONTEXT* is a Handle to the full register context of the application at a particular point in the execution • CONTEXT* is passed by default to a number of Pin Callback functions: e.g. – ThreadStart (registered by PIN_AddThreadStartFunction) – BufferFull (registered by PIN_DefineTraceBuffer) • Can request CONTEXT* be passed to an analysis function by requesting and IARG_CONTEXT Software & Services Group 111 CONTEXT* and IARG_CONTEXT • Passing IARG_CONTEXT to an analysis function has implications: – The analysis function will NOT be inlined – The passing of the IARG_CONTEXT is time consuming • CONTEXT* can NOT be dereferenced. It is a handle to be passed to Pin API functions • Pin API functions supplied to Get and Set registers within the CONTEXT – Set has no affect on CONTEXT* passed into analysis function (via IARG_CONTEXT request) • Have seen examples of both Get and Set • Have Pin API functions to Get and Set FP context Software & Services Group 112 Managing Exceptions and Signals Software & Services Group 113 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 114 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_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 115 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 116 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 117 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 118 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; } // 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 119 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 120 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 121 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 122 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 123 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 124 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 125 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 126 Transparent debugging, and extending the debugger • Transparently debug the application while it is running on Pin + Pin Tool – Have detailed explanation in the Pin User Manual • 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 127 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 128 Part5 Performance #s Software & Services Group 129 Performance Applications SpecINT SPEC CPU2000 SPEC integer benchmarks SpecFP SPEC CPU2000 SPEC floating point benchmarks Cinebench MAXON CINEBENCH 10 Graphics performance benchmarks Povray POV-Ray Ray tracing RealOne RealOne Player 1.0 Media player Illustrator Adobe Illustrator 9.1 Graphics design Dreamweaver Adobe Dreamweaver 9 Website design Director Macromedia Director 9 3D games, demos MediaEncoder Microsoft Media Encoder 9 Audio and video capturing Word, Excel, PowerPoint, Access, Outlook Microsoft Office XP Office applications Instrumentation BBCount Lightweight Counts executed basic blocks MemTrace Middleweight Records memory references MemError (Intel Parallel Inspector) Heavyweight Detects memory leaks, uninitialized variables, etc. Workloads SpecINT, SpecFP Reference input Cinebench, Povray Rendering an image (scalable) Other GUI applications Proprietary Visual Test scripts Software & Services Group 130 Pin in Windows vs. Pin in Linux Without Instrumentation SPEC CPU2000 32bit without Instrumentation 4.0X Windows Linux Slowdown (Relative to Native) 3.5X 3.0X 2.5X 2.0X 1.5X 1.0X 0.5X as EO p i M EA N G FP - ar eq t ua ke fa ce re c am m p lu ca s fm a3 d si xt ra ck p tw o EO lf M EA w up N w is e sw im m gr id ap pl u m es a ga gl e IN TG bz i ex vo rt ga p e pe on rlb m k vp r gc c m cf cr af ty pa rs er gz i p 0.0X • For CPU bound applications Pin shows nearly the same performance in Windows and Linux • SPECINT overhead is higher than SPECFP - more control instructions and low-tripcount code 131 & Services Group Pin uses the same binary translationSoftware technique on Windows and Linux Pin in Windows vs. Pin in Linux With Instrumentation SPEC CPU2000 32bit with inscount2 Linux 7X Slowdown (Relative to Native) Windows 6X 5X 4X 3X 2X 1X as EO p i M EA N G FP - a eq rt ua ke fa ce re c am m p lu ca s fm a3 d si xt ra ck ga p vo rte x bz ip IN tw To G EO lf M E wu AN pw is e sw im m gr id ap pl u m es a ga gl e e pe on rlb m k vp r gc c m cf cr af ty pa rs er gz i p 0X • Instrumentation overhead dominates the translation overhead Software & Services Group 132 Pin in Windows vs. Pin in Linux With Instrumentation SPEC CPU2000 32bit with MemTrace 12X Windows Linux 11X Slowdown (Relative to Native) 10X 9X 8X 7X 6X 5X 4X 3X 2X 1X as EO p i M EA N G FP - ar eq t ua ke fa ce re c am m p lu ca s fm a3 d si xt ra ck ga p vo rt ex bz ip IN t Tw o G EO lf M EA w up N w is e sw im m gr id ap pl u m es a ga gl e e pe on rlb m k se r ty pa r cf cr af m vp r gc c gz i p 0X • Instrumentation overhead dominates the translation overhead Software & Services Group 133 Windows Applications with Light and Middleweight Instrumentation No Instrumentation BBCount 9.0X MemTrace Slowdown (Relative to Native) 8.0X 7.0X 6.0X 5.0X 4.0X 3.0X 2.0X 1.0X RealOne MediaEncoder Director Dreamweaver Outlook Access PowerPoint Word Excel Illustrator Povray Cinebench SpecFP SpecINT 0.0X • #instructions per branch shrinks => BBCount overhead grows • %memory instructions grows => MemTrace overhead grows Software & Services Group 134 Windows Applications with Heavyweight Instrumentation No Instrumentation MemError Slowdown (Relative to Native) 25.0X 20.0X 15.0X 10.0X 5.0X Excel Illustrator Povray Cinebench SpecFP SpecINT 0.0X • The code translation overhead in Pin is much smaller than the overhead of non-trivial analysis in tools • MemError overhead is proportional to the number of memory accesses Software & Services Group 135 Pin Performance on Scalable Workloads 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 136 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 137 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 • Pin User Group – – http://tech.groups.yahoo.com/group/pinheads/ Pin users and Pin developers answer questions Software & Services Group 138

Verdana Bold 30

Related documents

Products

Support

Verdana Bold 30

Related documents

Add this document to collection(s)

Add this document to saved

Suggest us how to improve StudyLib