Slideshow | A Practical Approach to Error Handling

class: middle

#A Practical Approach to Error Handling

---

#Introduction

- Errors can happen anywhere
- Want reliable program
- No time to write error handling

What do we do?

---

#Options for Error Handling

```cpp
file f("file.txt");
```

What happens if the file does not exist?

- return value
```cpp
file f;
bool bOk=f.open("file.txt");
if( !bOk ) {...}
```
  * not for ctor
--

- out parameter
```cpp
bool bOk;
file f("text.txt",bOk);
if( !bOk ) {...}
```

- clutter code with checks
  * can forget check - `[[nodiscard]]` for return values
  
---
#Options for Error Handling (2)

- status: bad flag on first failure

* single control path
  * good if checking at the very end is good enough
	  * writing a file - ok
      * reading a file - maybe not

* default for C++ iostreams
--

- monad
  * goal: same code path for success and error case
  * like `std::variant<result, error>` + utilities
  * P0323R11 `std::expected`
  
---
#Options for Error Handling: Exception

- exception
--

* Catch exception objects always by reference
     * Slicing
     * Copying of exception may throw -> `std::terminate`

```cpp
struct A {...};
struct B : A {...};

try {
	throw B();
*} catch( A a ) { // B gets sliced and copied into a
	...
	throw; // throws original B
};
```

---
#Options for Error Handling: Exception

- exception
  * Catch exception objects always by reference
     * Slicing
     * Copying of exception may throw -> `std::terminate`

```cpp
struct A {...};
struct B : A {...};

try {
	throw B();
*} catch( A const& a ) { // no slicing or copying
	...
	throw; // throws original B
};
```

---

#Options for Error Handling: Exception (2)

- work like multi-level return/goto
- add invisible code paths
  * one reason some code bases do not allow exceptions

```cpp
auto inc(int i)->int { // throw(char const*)
	if(3==i) throw "Hello";
	return i+1;
}

auto main()->int {
	try {
		int n=3;
		n=inc(n); // throw(char const*)
	} catch( char const* psz ) {
		std::cout << psz;
	}
	return 0;
}
```

---

#Options for Error Handling: Exception (2)

- work like multi-level return/goto
- add invisible code paths
  * one reason some code bases do not allow exceptions

```cpp
auto inc(int i)->int { // throw(char const*)
*   if(3==i) throw "Hello";
	return i+1;
}

auto main()->int {
	try {
		int n=3;
*       n=inc(n); // throw(char const*)
	} catch( char const* psz ) {
		std::cout << psz;
	}
	return 0;
}
```

---

#Options for Error Handling: Exception (3)

```cpp
auto inc(int i, char const* & pszException )->int {
	{
		if(3==i) {
			pszException="Hello";
*           goto exception;
		}
		return i+1;
	}
exception:
	return 0;
}
```

---

#Options for Error Handling: Exception (4)

```cpp
auto main()->int {
	char const* pszException=nullptr;
	{
		int n=3;
		n=inc(n,pszException);
*       if( pszException ) goto exception;
		return 0;
	}
exception:
	{
		std::cout << pszException;
		return 0;
	}
}
```

Stop whining! Of course must write exception-safe code!

---

#Exception Safety Guarantees
(not really exception-specific)

Part of function specification

- Never Fails

- Strong Exception Guarantee:
    * may fail (throw), but will restore program state to what it was before: transactional
	   * possible and desirable in library functions
	   * very hard in application code
			* usually too many state changes

--
			
- Basic Exception Guarantee:
	* may fail (throw), but will restore program to some valid state
		
---

#Basic Exception Safety Guarantee

Customer: "Hello, is this Microsoft Word support?
  I was writing a book. Suddenly, Word deleted
  everything."

Microsoft: "Oh, that's ok. Word only provides a basic
  exception guarantee."

Customer: "Oh, alright then, thank you very much and
  have a good day!"

---

#The Challenge

- Error handling is a lot of effort
  * in development
       * must be paranoid
	   * create a lot of extra code
  * in testing
       * many codepaths to test
	   * if you don't test them, they won't work
--

- Little customer gain
--

- So what do we do?

---
#So what do we do?

- Check everything
  * check every API call
       * one wrapper per error reporting method
	     * Windows: `GetLastError()`, `HRESULT`
		 * Unix: `errno`
  * `assert` aggressively
       * asserts stay in Release 
  * `noexcept` if caller does not handle exception
       * `std::terminate`, but unexpected exceptions will terminate anyway
	   * install handler with `std::set_terminate` for checking
--

- Assume everything works
--

- Goal:
  * keep set of code paths small
  * keep set of program states small

---
#If checks fail
- prio 1: collect as much information as possible
  * client: send report with memory dump home
  * server: halt thread and notify operator
--

- prio 2: carry on somehow
  * if check was critical, program behavior now undefined: no further reports
  * do not terminate when assertion fails
      * `assert`s can be wrong, too
  * if you need safety (nuclear powerplant, etc.), add at higher level
      * example: server stops processing request categories with too many pending requests

---
#Next: Homework

- Reproduce the error at home
--

- Add handling code only for errors that are reproducible
    * Otherwise you write
	    * error handlers that are never used
        * error handlers that are never tested, do the wrong thing

--
		
- 5% of handlers handle 95% of errors
	* Write high quality error handlers
		* Bad: show message box
		* Good: fix the problem

---
#Categories of Errors: Critical

- `nullptr` access
- API calls not expected to fail
  * not with this error code
- assertions
--

- "never happens"
  * no handler
  * like C++ undefined behavior: program is invalid
--

- Client: send report, disable future reports
- Server: notify operator, enter infinite loop (wait for debugger)

- Notify user only if false alarm unlikely
  * `assert`s may be wrong
  
---
#Categories of Errors: Untested

```cpp
auto RegisterFooHook(Foo foo) {
	errcode_t err=RegisterFoo(foo);
	if(err==SUCCESS) KeepTrackOfFoo(foo);
	return err;
}
```

- If `err` indicates error, does nothing, no error handling needed
--

- But no reproduction for `RegisterFoo` failing
- Effect on rest of the program?

- Client: send report, throttle future reports
  * in Debug: notify developer
- Server: send report

---
#Categories of Errors: Bad User Experience

- 3rd party bug
  * sometimes PowerPoint makes shape disappear
- Reproducible, supported and tested

- Not nice, users may complain
--

- Client/Server: only log, no report
  * to explain behavior if user calls

---
#Categories of Errors: Indication of broken environment

- Other add-in hooked same function as us
- OS reports space as default decimal separator
  * both fully supported by us
--

- Could still be cause of a problem
--

- Client during remote support: notify support engineer
  * maybe reason for support call

---
#Error Analysis
- Reports with memory dumps sent to server
  * automatically
  * if user opted out, user can send prepared email
--

- Error database
  * memory dumps opened in debugger
  * errors automatically categorized by file/line
  * details and memory dump accessible to devs
--

- Devs can mark errors as fixed
  * trigger automatic update
  * or send automatic email - magic!

---
#Cause Analysis

- Problem often related to customer environment
- Otherwise in-house testing would have found it
--

- Memory dumps have list of loaded modules (DLLs, dylibs)
- Can we identify module causing error?
  * or versions of module?
  
--

Report database with all reports
- 1 means has particular problem
- 0 means has different problem

`0 1 1 0 0 1 0 1 0 1 1 0` (6 occurrences among 12 reports)

`x - x - - x x - x - x -` Module A (with: 3/6, without: 3/6)

`- x x - x x - x x - - -` Module B (with: 4/6, without: 2/6)
--

- Module B responsible? Or chance?

---
# Minimum Description Length

- Compressing

`0 1 1 0 0 1 0 1 0 1 1 0` (6/12)

- Knowing if reports contain module B helps compressing?

`- x x - x x - x x - - -` Module B (with: 4/6, without: 2/6)
--

- perfect arithmetic compression (Laplacian estimator)
  * estimates probability `p` that report has particular problem
- all `p` in `[0,1]` equally likely

- no. bits to compress `N` bits with `K` ones:

`log [ (N+1) * (N over K) ]`

- no. bits becomes smaller if p is closer to 0 or 1:
  * 12 bits with 6 ones: 13.55 bits
  * 12 bits with no ones: 3.70 bits

---

#Compressing Reports

`0 0 1 0 0 1 0 1 0 1 1 0` (6/12)

`x - x - - x x - x - x -` Module A (with: 3/6, without: 3/6)

`- x x - x x - x x - - -` Module B (with: 4/6, without: 2/6)
--

- Compressing all reports together (6/12): 13.55 bits
--

- Make use of module A
  * choose module A over B: 1 bit
  * compressing all reports with A (3/6): 7.13 bits
  * compressing all reports without A (3/6): 7.13 bits
  * total: 15.26 bits - module A has nothing to do with problem

---

#Compressing Reports

`0 0 1 0 0 1 0 1 0 1 1 0` (6/12)

`x - x - - x x - x - x -` Module A (with: 3/6, without: 3/6)

`- x x - x x - x x - - -` Module B (with: 4/6, without: 2/6)
- Compressing all reports together (6/12): 13.55 bits
- Make use of module B
  * choose module B over A: 1 bit
  * compressing all reports with B (4/6): 6.71 bits
  * compressing all reports without B (2/6): 6.71 bits
  * total: 14.43 bits - still not relevant enough
  
---

#Compressing Reports

`0 0 1 0 0 1 0 1 0 1 1 0` (6/12)

`x - x - - x x - x - x -` Module A (with: 3/6, without: 3/6)

`- x x - x x - x x - - -` Module B (with: 4/6, without: 2/6)

`- x x - x x - x - x - -` Module C (with: 5/6, without: 1/6)
- Compressing all reports together (6/12): 13.55 bits
- Make use of module C
  * choose module C over A and B: log2(3) = 1.58 bits
  * compressing all reports with C (5/6): 5.39 bits
  * compressing all reports without C (1/6): 5.39 bits
  * total: 12.37 bits - relevant!
--

- More hypotheses make chance more likely
- Also works if certain module fixes problem
- Extend to module versions

---

#C++20 Contracts

- new language feature
- `assert` on steroids
- declarative function pre- and postconditions

```cpp
void push(int x, queue& q)
[[expects: !q.full()]]
[[ensures: !q.empty()]]
{
...
[[assert: q.is_valid()]]
...
}
```

---

#C++20 Contracts (2)

- When check contract?
  * debug
  * release
  * never

- What to do if contract violated?
  * terminate
  * carry on
  * report (what to whom?)

--
- removed from C++20 at last moment
- discussion will continue for C++23

---
#THANK YOU!

for attending.

And yes, we are recruiting:

`hr@think-cell.com`

![](/assets/en/career/talks/errorhandling/banner.jpg)

---

#A Very Special Class of Errors

```cpp
std::int32_t a=2 000 000 000;
std::int32_t b=a+a;
```

What is `b`?

Uuh, may overflow.
- Let's check for it!
```cpp
if( b<a ) {
	... treat overflow ...
}
```

Ok?

---

#Undefined Behavior (UB)

Example: int arithmetic overflow

If program contains undefined behavior, compiler
can do anything _with the whole program_!

- In particular, compiler may assume that UB never happens

```cpp
int i=...;
int j=i;
++i;
// code for overflow may be thrown out by compiler!
// if( i<j ) {
//		... treat overflow ...
// }
```

- This is not just theory, it happens _in practice_!
- `unsigned` types are different: modulo arithmetic

Share