case study of the unexplained

Post on 20-Aug-2015

204 Views

Category:

Technology

2 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Case Study of Programmer Nightmares

Shannon’s Edition 20120624

What is the talk about

• Inspired by Mark Russinovich’s Presentation– Case of the Unexplained– http://technet.microsoft.com/en-us/sysinternals/b

b963887• Here are my cases

– Mainly fixing programming problem– Mostly C++, some interop & cross-platform.– Most are from my bad memory.– Sorry about the boring slides.

Steps to Debug a problem

1. There is no step 12. See Step 13. ???4. Profit

General Guidelines

• Reproducible test case.• Learn the tools.• Make a Wild A** Guess (WAG) on source• Persistent

– Grind through it.• Ask someone else to handle it. (NOT ME)

Case: WOMM

Problem: Debug vs Release

• Program is not drawing the circle around the cursor but is where the user clicks.

• Same class does both drawings, different location• Did work previously

Debug Optimized

Causes Optimization Problems

1. Undefined Behaviors1. Uninitialized Memory2. Overflows/underflows

2. Thread problems.3. Code or Data is wrong.….999.Complier Bug (not likely, see #1)1000.Hardware/OS/driver bug.

Step I took

• What’s changed.– Major merge with other branch.– Massive file and project settings changes.

• Build optimized with debug symbols & debug– Could jump around a lot– Local variables will not be present or wrong*– this pointer only valid on member function entry.

• Compare working/non-working objects

Found the Function

• Formula:

• D, E, F = 0, so have this, and verified all inputs.

1**

**

FhEwD

CyBxAI

CyBxAI **

Next Trick: Binary Search

• Turning on/off optimizations – Per Library– Per File– Per function– Per optimization

• Found, Global Optimization “Cause” problem– Last merge turned it on.– Turned it off. Everything works

Extremely Important Rule

• Unless you understand why the problem is fixed, its not fixed. The problem is likely still there just hidden better.

• Formula:

• D, E, F = 0, so have this, and verified all inputs

Missed something important

1

** CyBxAI

1**

**

FhEwD

CyBxAI

Lets talk this Out

• w & h were uninitialized, but can’t be it.• 0 time any number is 0.• w & h are number.

– Double Precision IEEE 754• IEEE 754 only contains number.

– Contains ±0, ±INF, … NaNFALSE

MAYBETRUE

FALSE

NaN is weird.

• Any operation with NaN results in NaN– *, +, -, /, sin, etc

• Most comparisons with NaN are false.– <, <=, >, ==, etc, so NaN == NaN is false

• Not equals is always true.– NaN != NaN is true.

• Multiple types– QNaN, SNaN

Case Close

• Should have trusted 1st guess.• Gave up too soon with a quick wrong fix.

Case: Works Everywhere Else

Problem

• 6-8 high priority bugs from FAT.• All bugs had the same pattern.

– Only occurred on Window 2000 box.– Display wrong converted values.– Works on XP, and 2003.

• It a cross-platform assign to Me.

Steps I took

• Start Debug Build of Integration Branch.• Get the release, and try to reproduce bug.

– Grabbed it from the build NFS share.– Didn’t “fail”

• Try the test box.– It “fails”, but can’t debug.– Copy it to dev box– It fails on my box.

WAG time

• Cosmic Rays corrupted the Executable. – No replacing them with debug build still had bug.

What Could It Be

• Diff installed w/ what should be there.– Should be No Differences

• Massive Differences.• Install CD didn’t have and Differences• I know what happened.

AAAAAAHH!!!Stupid Tester

Here is What Happened

• Tester skipped using the Install Win2K CD.– Didn’t want to walk to other end of hall.

• TAR-ed up NFS install shared.• FTP it over.• Used WinZip to untar file.

Why is WinZip Bad?

Case Closed

• The “table.dat” file was converted to windows newlines.– Doesn’t work properly like that.

• All Test Follow Proper Procedures.• Don’t take Short Cuts.

– Especially During FAT.

Case: Psychic Debugging

Problem: Phone Call

1. Got a phone call2. Developer describe the problem and steps

taken to track down the problem.3. Answer with the root cause and how to fix.

Now its time for the interactive part of this talk.Pretend you me, ….

Real Problem

• File parsing code incorrectly errors out.– Worked on following

• Windows 32/64-bits debug/release, • Irix 32/64-bits debug/release, • Solaris SPARC 32/64-bit debug/release• Linux 64-bit debug/release, 32-bit debug.

– Fails on Linux 32-bit x86 gcc optimize

What does the code do?

• Read text like file– Contains repeated floating point numbers.– Lots of other data between repeated number.

• Parses data into native types (int, double)• Validate Data is sane

– Number are with spec.– Repeated doubles are the same with != check.

• This step failed.

Code

double lat1 = atof(buff1);…double lat2 = atof(buff2);…if(lat1 != lat2) return -1;

I’m 95% certain of problemWrite down your answer now.

More info from the developer

Additional QA with develop

• Did they check input file is valid?• How did the developer track down it down?

– Printf debugging number same, but check failed.• Did adding/moving additional printf make

the problem go away?– This confirmed that I guessed right

YES

YES

Your Turn

• Failed 32-bit x86 optimized linux• Deal with C++ native double types

– uses != to compare them.• Adding some printfs made problem go away.

Who know what happened.

Additional Slide If No One Knows

• Root cause is 486• Specifically math co-processor• C++ doubles are 64-bits in memory• 486 math registers are 80-bits• Can’t store 80-bits in 64-bit• Round double when copied into memory.• Optimizer will speed up code

– Will attempt to reduce the # of memory copies.• Wait here until some guesses.

Here is what happened

• Function converted 1st string to 80-bit double• Compiler moved result into 64-bit on stack• Function conerted 2nd string to 80-bit double• Compiler got smart and kept it in 80-bits.• Loaded 1st 64-bit double into 80-bit register.• 2nd number has more precision so it didn’t

match.

Optimized ASM Code

call atof ; buff1 in eaxfstor [sp+20], ST(0)……call atof ; buff2 in eaxfload ST(1), [sp+20]fcmp ST(0), ST(1) ; compare 80 w/ 64-bits

jmpe +8 ; skip over next line if ==ret ; error

Case Close

• Changed to use strcmp instead.• Never directly compare double without a

tolerance.• Round errors will cause mathematically

impossible to happen. • Stupid 80-bits.

Case: Shoot Self in Foot

Problem: Crash with no reasons

• New developed code• Crashed on Solaris while calling constructor• No “obvious” problem with code

Code

class A { … A(A *d) { *this = d; } … A& operator=(const A &d) { … return *this; }};

Steps I took

• Build code on Windows.– Visual Studio Debugger is 10x nicer

• Got a helpful warning– warning C4717: ‘A::A’ : recursive on all control

paths, function will cause runtime stack overflow

Code Again

class A { A(A *d) { *this = d; } A& operator=(const A &d) {…}};

What the Compiler Does

class A { A(A *d) { A __tempA(d); *this->operator=(__tempA); } A& operator=(const A &d) {…}};

Solution #1

class A { A(A *d) { *this = *d; } A& operator=(const A &d) {…}};

Problem With Solution #1

• What does the following code doA d = NULL;

• Compile does this followingA d = A(NULL);

• Which crashes.• “A d = 0” also crashes.

Solution #2

class A { explicit A(A *d) { *this = *d; } A& operator=(const A &d) {…}};

C++ “Rule of 3” Solution

class A { A(const A &d) {…} ~A() {…} A& operator=(const A &d) {…}};

C++11 “Rule of 3,4, or 5” Solution

class A { A(const A &d) {…} A(A &&d) {…} ~A() {…} A& operator=(const A &d) {…} A& operator=(A &&d) {…}};

Case Close

• Pay Attention to compiler warnings.– This particular warning appear in 3 other places.

• Use Compiler that give better warnings.– CLANG/LLVM has the best error/warnings.

Case: “Random” Crashes

Problem: GUI randomly crashes

Java

Automagic JNI Junk

C++

Steps I took

• Build Debug– debug runtimes make it crash faster due to checks

• Use 2 Debugger Visual Studio & JBuilder• 4 hours of persistent.

Track it down, but no clue

• Java had valid pointer to C++ object.• Pressed button, & pointer no longer valid• Trick time.

Data Breakpoint.

• x86 has 4 hardware data breakpoints– Program runs at full speed.– 1 is reserved by OS

• Must take following form. (Old Info)– Memory address, length(must be 4).– 0x12345678,4

How to do it VS2010

• Step 1

How to do it VS2010

• Step 2

How to do it VS2010

• Step 3 Done

How to do it VS2010

• Step 4 See Results

BAM Data Changed

• Java GC – > finalizer – > Automagic JNI junk– > delete object

• Why, leaky abstraction.

Here is What Happened.

Java C++

ARRAY | | | | | | |AMJJArray

AMJJThing

Case Close

• Data Breakpoints Rule.• All Abstraction Leak

– Know how before proceeding.

That’s all for Now

Questions, Comment, etc.

top related