Is code a safe thread?

[Nephi Immortal]

I customize operator new to create my own definition.
m_out_of_memory is a global variable of class Test. m_out_of_memory
can be set to true or false inside operator new function before Test()
is called to construct three data members: a, b, c.
If m_out_of_memory is to be true, then Test() is not called and all
data members cannot be initialized.
My one question is – is thread safe? What happen if two main
functions are running on two threads at the same time while accessing
m_out_of_memory?
First thread will report out of memory before second thread will skip
to initialize data members. Can you please confirm that it never
happened as long as thread safety is taken care?
Or…is m_out_of_memory stored in separate memory on each thread and is
never shared?
I choose this method to avoid using exception.

class Test
{
public:
Test() : a( 1 ), b( 2 ), c( 3 ) {
}

~Test() {
}

bool Is_Sufficient_Memory() const {
return ( m_out_of_memory == false ) ? true : false;
}

void* operator new( size_t size ) {
void* p = malloc( size );

m_out_of_memory = ( p == 0 ) ? true : false;

return p;
}

void operator delete( void* memory ) {
free( memory );
}

void Set_abc( int _a, int _b, int _c ) {
a = _a; b = _b; c = _c;
}

private:
static bool m_out_of_memory;
int a;
int b;
int c;
};

bool Test::m_out_of_memory = false;

int main()
{
Test* T = new Test();

if( T->Is_Sufficient_Memory()) {
T->Set_abc( 10, 20, 30 );
delete T;
}

return 0;
}

 

[Kai-Uwe Bux]

I customize operator new to create my own definition.
m_out_of_memory is a global variable of class Test. m_out_of_memory
can be set to true or false inside operator new function before Test()
is called to construct three data members: a, b, c.
If m_out_of_memory is to be true, then Test() is not called and all
data members cannot be initialized.
My one question is ? is thread safe? What happen if two main
functions are running on two threads at the same time while accessing
m_out_of_memory?
First thread will report out of memory before second thread will skip
to initialize data members. Can you please confirm that it never
happened as long as thread safety is taken care?

Well, it would be impossible to take care of thread safety with the
interface design presented below. The content of the shared static variable
is meant to indicate the status returned by the _previous_ allocation
attempt. Hence, the meaning of the variable would differ from thread to
thread (as they go through different allocation attempts and there is no
trans-thread meaning of "last"). No synchronization would change that. The
only way to deal with it, would be to stop any other threads when it tries
to allocate memory until the first thread doing so has checked the status of
_its_ last allocation attempt. Such synchronization is not possible with the
interface you present.

Or?is m_out_of_memory stored in separate memory on each thread and is
never shared?

Which data are shared depends on the particulars of the thread model. E.g.,
in linux you could use multiple processes or multiple threads. In the later
case, static data would be shared.

I choose this method to avoid using exception.

Aha. And what's wrong with exceptions here?

class Test
{
public:
Test() : a( 1 ), b( 2 ), c( 3 ) {
}

~Test() {
}

bool Is_Sufficient_Memory() const {
return ( m_out_of_memory == false ) ? true : false;
}

void* operator new( size_t size ) {
void* p = malloc( size );

m_out_of_memory = ( p == 0 ) ? true : false;

return p;
}

void operator delete( void* memory ) {
free( memory );
}

void Set_abc( int _a, int _b, int _c ) {
a = _a; b = _b; c = _c;
}

private:
static bool m_out_of_memory;
int a;
int b;
int c;
};

bool Test::m_out_of_memory = false;

int main()
{
Test* T = new Test();

if( T->Is_Sufficient_Memory()) {

Hm .....

If the allocation via new() above fails, then T is 0 here and dereferencing
the pointer is undefined behavior. And yes, accessing a member function
qualifies as dereferencing even when the member function only accesses
static data.

T->Set_abc( 10, 20, 30 );
delete T;
}

return 0;
}

One more question: why do not use a nothrow allocation function. In that
case, a new expression would indicate an allocation failure not via an
exception but by returning 0. You can test for that. No threading issues
arise.


Best,

Kai-Uwe Bux

 

[Kai-Uwe Bux]

Well, it would be impossible to take care of thread safety with the
interface design presented below. The content of the shared static
variable is meant to indicate the status returned by the _previous_
allocation attempt. Hence, the meaning of the variable would differ from
thread to thread (as they go through different allocation attempts and
there is no trans-thread meaning of "last"). No synchronization would
change that. The only way to deal with it, would be to stop any other
threads when it tries to allocate memory until the first thread doing so
has checked the status of _its_ last allocation attempt. Such
synchronization is not possible with the interface you present.

Oops, maybe my imagination was too limited. I could think of something like
this:

class Test {
...

void* operator new ( size_t size ) {
lock the static variable m_out_of_memory;
...
}

static
bool last_alloc_ok ( void ) {
bool result = ...
unlock the static variable m_out_of_memory
return result;
}

...

};

Note, however, that this intimately ties the observer function and the
allocation function. You now must use them strictly alternating in each
thread.

Also, you can leave the locking issues to the calling threads.

Neither solution seems to be appealing: they look much too error-prone.

 

[Nephi Immortal]

Well, it would be impossible to take care of thread safety with the
interface design presented below. The content of the shared static variable
is meant to indicate the status returned by the _previous_ allocation
attempt. Hence, the meaning of the variable would differ from thread to
thread (as they go through different allocation attempts and there is no
trans-thread meaning of "last"). No synchronization would change that. The
only way to deal with it, would be to stop any other threads when it tries
to allocate memory until the first thread doing so has checked the statusof
_its_ last allocation attempt. Such synchronization is not possible with the
interface you present.


Which data are shared depends on the particulars of the thread model. E.g..,
in linux you could use multiple processes or multiple threads. In the later
case, static data would be shared.


Aha. And what's wrong with exceptions here?

Well, I avoid exception for some reasons. Some programmers choose
exception, but I choose alternative method. I prefer to use error
message handler. Another reason, C++ Compiler is unable to inline
some member functions in release mode if exception is enabled. And
other reason, exception degrade performance.
Please do not discuss the topic – exceptions. It is not important

Hm .....

If the allocation via new() above fails, then  T is 0 here and dereferencing
the pointer is undefined behavior. And yes, accessing a member function
qualifies as dereferencing even when the member function only accesses
static data.

Of course, you are correct. ‘this’ pointer is undefined if T is
zero. Accessing static data member is not a problem. I wonder “if( T-

Is_Sufficient_Memory())” is undefined. I should declare static on

Is_Sufficient_Memory() to become global function.

One more question: why do not use a nothrow allocation function. In that
case, a new expression would indicate an allocation failure not via an
exception but by returning 0. You can test for that. No threading issues
arise.

I can test T to be zero if memory allocation failed.
Is_Sufficient_Memory() is preferable choice over “if( T == NULL )”
because of better readable code.
Calling operator new and operator delete hundred times can slow the
memory alloction. Invoke operator new to allocate huge array over 10
megabytes one time and then invoke operator delete to free it prior to
the end of main() function.
A huge array is declared static. Each class has its own non-static
data member as memory address pointer. Each time, class constructs
memory address pointer to point to the shared huge array. This allows
many classes to share one array. The data can be erased before
another class can reuse the same memory space. My method runs faster
to eliminate the need to use malloc() and free() frequently.
My code is carefully designed NOT to overrun buffers outside memory
address pointer and the buffer length.
You got the point right that this code is not concerned with thread
safety.


.......

 

[Kai-Uwe Bux]

Well, I avoid exception for some reasons. Some programmers choose
exception, but I choose alternative method. I prefer to use error
message handler. Another reason, C++ Compiler is unable to inline
some member functions in release mode if exception is enabled. And
other reason, exception degrade performance.
Please do not discuss the topic ? exceptions. It is not important


Of course, you are correct. ?this? pointer is undefined if T is
zero. Accessing static data member is not a problem. I wonder ?if( T-
Is_Sufficient_Memory() to become global function.


I can test T to be zero if memory allocation failed.
Is_Sufficient_Memory() is preferable choice over ?if( T == NULL )?
because of better readable code.

That is debatable. In particular, if you declare Is_Sufficient_Memory() as a
non-member function. In that case it would be

if ( T ) {
...
}

versus

if ( Test::Is_Sufficient_Memory() ) {
...
}

The second version makes it very explicit, what the test is about. However,
it suppresses any reference to the allocation whose success is being tested.

You could do:

static
bool Was_Successfully_Allocated ( Test* T ) {
return ( T != 0 );
}

and then use

if ( Test::Was_Successfully_Allocated( T ) ) {
...
}

I don't feel that readability of code is improved by making that dependency
implicit and hiding it behind a static variable that refers back to the last
allocation. The last allocation need not have happened in nearby code (at
least, there could be nothing but coding style to enforce it).

Calling operator new and operator delete hundred times can slow the
memory alloction. Invoke operator new to allocate huge array over 10
megabytes one time and then invoke operator delete to free it prior to
the end of main() function.
A huge array is declared static. Each class has its own non-static
data member as memory address pointer. Each time, class constructs
memory address pointer to point to the shared huge array. This allows
many classes to share one array. The data can be erased before
another class can reuse the same memory space. My method runs faster
to eliminate the need to use malloc() and free() frequently.
My code is carefully designed NOT to overrun buffers outside memory
address pointer and the buffer length.

Sounds like you are implementing some kind of pool allocator.

You got the point right that this code is not concerned with thread
safety.

Yes. And you got the point right that using a static variable will add
complexity once thread safety becomes an issue to be dealt with.


Best,

Kai-Uwe Bux