Difference between revisions of "Memory Management"

From D Wiki
Jump to: navigation, search
(Begin copying article from dlang.org)
 
(Finished migrating from dlang.org)
Line 8: Line 8:
  
 
The techniques for using them, as well as some advanced alternatives are:
 
The techniques for using them, as well as some advanced alternatives are:
 
* Strings (and Array) Copy-on-Write
 
* Real Time
 
* Smooth Operation
 
* Free Lists
 
* Reference Counting
 
* Explicit Class Instance Allocation
 
* Mark/Release
 
* RAII (Resource Acquisition Is Initialization)
 
* Allocating Class Instances On The Stack
 
* Allocating Uninitialized Arrays On The Stack
 
* Interrupt Service Routines
 
  
 
== Strings (and Array) Copy-on-Write ==
 
== Strings (and Array) Copy-on-Write ==
Line 71: Line 59:
  
 
Copy-on-write is the protocol implemented by array processing functions in the D Phobos runtime library.
 
Copy-on-write is the protocol implemented by array processing functions in the D Phobos runtime library.
 +
 +
== Real Time ==
 +
Real time programming means that a program must be able to guarantee a maximum latency, or time to complete an operation. With most memory allocation schemes, including malloc/free and garbage collection, the latency is theoretically not bound. The most reliable way to guarantee latency is to preallocate all data that will be needed by the time critical portion. If no calls to allocate memory are done, the GC will not run and so will not cause the maximum latency to be exceeded.
 +
 +
== Smooth Operation ==
 +
Related to real time programming is the need for a program to operate smoothly, without arbitrary pauses while the garbage collector stops everything to run a collection. An example of such a program would be an interactive shooter type game. Having the game play pause erratically, while not fatal to the program, can be annoying to the user.
 +
 +
There are several techniques to eliminate or mitigate the effect:
 +
 +
* Preallocate all data needed before the part of the code that needs to be smooth is run.
 +
* Manually run a GC collection cycle at points in program execution where it is already paused. An example of such a place would be where the program has just displayed a prompt for user input and the user has not responded yet. This reduces the odds that a collection cycle will be needed during the smooth code.
 +
* Call std.gc.disable() before the smooth code is run, and std.gc.enable() afterwards. This will cause the GC to favor allocating more memory instead of running a collection pass.
 +
 +
== Free Lists ==
 +
Free lists are a great way to accelerate access to a frequently allocated and discarded type. The idea is simple - instead of deallocating an object when done with it, put it on a free list. When allocating, pull one off the free list first.
 +
 +
<syntaxhighlight lang="D">
 +
class Foo
 +
{
 +
    static Foo freelist; // start of free list
 +
 +
    static Foo allocate()
 +
    {  Foo f;
 +
 +
if (freelist)
 +
{  f = freelist;
 +
    freelist = f.next;
 +
}
 +
else
 +
    f = new Foo();
 +
return f;
 +
    }
 +
 +
    static void deallocate(Foo f)
 +
    {
 +
f.next = freelist;
 +
freelist = f;
 +
    }
 +
 +
    Foo next; // for use by FooFreeList
 +
    ...
 +
}
 +
 +
void test()
 +
{
 +
    Foo f = Foo.allocate();
 +
    ...
 +
    Foo.deallocate(f);
 +
}
 +
</syntaxhighlight>
 +
 +
Such free list approaches can be very high performance.
 +
 +
* If used by multiple threads, the allocate() and deallocate() functions need to be synchronized.
 +
* The Foo constructor is not re-run by allocate() when allocating from the free list, so the allocator may need to reinitialize some of the members.
 +
* It is not necessary to practice RAII with this, since if any objects are not passed to deallocate() when done, because of a thrown exception, they'll eventually get picked up by the GC anyway.
 +
 +
== Reference Counting ==
 +
The idea behind reference counting is to include a count field in the object. Increment it for each additional reference to it, and decrement it whenever a reference to it ceases. When the count hits 0, the object can be deleted.
 +
 +
D doesn't provide any automated support for reference counting, it will have to be done explicitly.
 +
 +
[[COM_Programming | Win32 COM programming]] uses the members AddRef() and Release() to maintain the reference counts.
 +
 +
== Explicit Class Instance Allocation ==
 +
'''Note:''' [http://dlang.org/class.html#allocators Class allocators] and [http://dlang.org/class.html#deallocators deallocators] are deprecated in D2.
 +
 +
D provides a means of creating custom allocators and deallocators for class instances. Normally, these would be allocated on the garbage collected heap, and deallocated when the collector decides to run. For specialized purposes, this can be handled by creating NewDeclarations and DeleteDeclarations. For example, to allocate using the C runtime library's <code>malloc</code> and <code>free</code>:
 +
 +
<syntaxhighlight lang="D">
 +
import std.c.stdlib;
 +
import core.exception;
 +
import core.memory : GC;
 +
 +
class Foo
 +
{
 +
    new(size_t sz)
 +
    {
 +
        void* p;
 +
 +
        p = std.c.stdlib.malloc(sz);
 +
 +
        if (!p)
 +
            throw new OutOfMemoryError();
 +
 +
        GC.addRange(p, sz);
 +
        return p;
 +
    }
 +
 +
    delete(void* p)
 +
    {
 +
        if (p)
 +
        {
 +
            GC.removeRange(p);
 +
            std.c.stdlib.free(p);
 +
        }
 +
    }
 +
}
 +
</syntaxhighlight>
 +
 +
The critical features of new() are:
 +
 +
* new() does not have a return type specified, but it is defined to be void*. new() must return a void*.
 +
* If new() cannot allocate memory, it must not return null, but must throw an exception.
 +
* The pointer returned from new() must be to memory aligned to the default alignment. This is 8 on win32 systems.
 +
* The size parameter is needed in case the allocator is called from a class derived from Foo and is a larger size than Foo.
 +
* A null is not returned if storage cannot be allocated. Instead, an exception is thrown. Which exception gets thrown is up to the programmer, in this case, OutOfMemory() is.
 +
* When scanning memory for root pointers into the garbage collected heap, the static data segment and the stack are scanned automatically. The C heap is not. Therefore, if Foo or any class derived from Foo using the allocator contains any references to data allocated by the garbage collector, the GC needs to be notified. This is done with the std.gc.addRange() method.
 +
* No initialization of the memory is necessary, as code is automatically inserted after the call to new() to set the class instance members to their defaults and then the constructor (if any) is run.
 +
 +
The critical features of delete() are:
 +
 +
* The destructor (if any) has already been called on the argument p, so the data it points to should be assumed to be garbage.
 +
* The pointer p may be null.
 +
* If the GC was notified with std.gc.addRange(), a corresponding call to std.gc.removeRange() must happen in the deallocator.
 +
* If there is a delete(), there should be a corresponding new().
 +
 +
If memory is allocated using class specific allocators and deallocators, careful coding practices must be followed to avoid memory leaks and dangling references. In the presence of exceptions, it is particularly important to practice RAII to prevent memory leaks.
 +
 +
Custom allocators and deallocators can be done for structs and unions, too.
 +
 +
== Mark/Release ==
 +
Mark/Release is equivalent to a stack method of allocating and freeing memory. A ‘stack’ is created in memory. Objects are allocated by simply moving a pointer down the stack. Various points are ‘marked’, and then whole sections of memory are released simply by resetting the stack pointer back to a marked point.
 +
 +
<syntaxhighlight lang="D">
 +
import std.c.stdlib;
 +
import core.exception;
 +
import core.memory : GC;
 +
 +
class Foo
 +
{
 +
    static void[] buffer;
 +
    static int bufindex;
 +
    static const int bufsize = 100;
 +
 +
    static this()
 +
    {
 +
        void* p;
 +
        p = malloc(bufsize);
 +
 +
        if (p is null)
 +
            throw new OutOfMemoryError;
 +
 +
        GC.addRange(p, bufsize);
 +
        buffer = p[0 .. bufsize];
 +
    }
 +
 +
    static ~this()
 +
    {
 +
        if (buffer.length)
 +
        {
 +
            GC.removeRange(buffer.ptr);
 +
            free(buffer.ptr);
 +
            buffer = null;
 +
        }
 +
    }
 +
 +
    new(size_t sz)
 +
    {
 +
        void* p;
 +
 +
        p = &buffer[bufindex];
 +
        bufindex += sz;
 +
 +
        if (bufindex > buffer.length)
 +
            throw new OutOfMemoryError;
 +
 +
        return p;
 +
    }
 +
 +
    delete(void* p)
 +
    {
 +
        assert(0);
 +
    }
 +
 +
    static int mark()
 +
    {
 +
        return bufindex;
 +
    }
 +
 +
    static void release(int i)
 +
    {
 +
        bufindex = i;
 +
    }
 +
}
 +
 +
void main()
 +
{
 +
    int m  = Foo.mark();
 +
    Foo f1 = new Foo;          // allocate
 +
    Foo f2 = new Foo;          // allocate
 +
 +
    Foo.release(m);            // deallocate f1 and f2
 +
}
 +
</syntaxhighlight>
 +
 +
The allocation of buffer[] itself is added as a region to the GC, so there is no need for a separate call inside Foo.new() to do it.
 +
 +
== RAII (Resource Acquisition Is Initialization) ==
 +
RAII techniques can be useful in avoiding memory leaks when using explicit allocators and deallocators. Adding the [http://dlang.org/attribute.html#scope scope attribute] to such classes can help.
 +
 +
== Allocating Class Instances On The Stack ==
 +
Class instances are normally allocated on the garbage collected heap. However, if they:
 +
 +
* are allocated as local symbols in a function and
 +
* are allocated using <code>new</code> and
 +
* use <code>new</code> with no arguments (constructor arguments are allowed) and
 +
* have the <code>scope</code> storage class
 +
 +
then they are allocated on the stack. This is more efficient than doing an allocate/free cycle on the instance. But be careful that any reference to the object does not survive the return of the function.
 +
 +
<syntaxhighlight lang="D">
 +
class C { ... }
 +
 +
scope c = new C(); // c is allocated on the stack
 +
scope c2 = new C(5);    // allocated on stack
 +
scope c3 = new(5) C();  // allocated by a custom allocator
 +
</syntaxhighlight>
 +
 +
If the class has a destructor, then that destructor is guaranteed to be run when the class object goes out of scope, even if the scope is exited via an exception.
 +
 +
== Allocating Uninitialized Arrays On The Stack ==
 +
Arrays are always initialized in D. So, the following declaration:
 +
 +
<syntaxhighlight lang="D">
 +
void foo()
 +
{  byte[1024] buffer;
 +
 +
    fillBuffer(buffer);
 +
    ...
 +
}
 +
</syntaxhighlight>
 +
 +
will not be as fast as it might be since the buffer[] contents are always initialized. If careful profiling of the program shows that this initialization is a speed problem, it can be eliminated using a ''VoidInitializer'':
 +
 +
 +
<syntaxhighlight lang="D">
 +
void foo()
 +
{  byte[1024] buffer = void;
 +
 +
    fillBuffer(buffer);
 +
    ...
 +
}
 +
</syntaxhighlight>
 +
 +
Uninitialized data on the stack comes with some caveats that need to be carefully evaluated before using:
 +
 +
* The uninitialized data that is on the stack will get scanned by the garbage collector looking for any references to allocated memory. Since the uninitialized data consists of old D stack frames, it is highly likely that some of that garbage will look like references into the GC heap, and the GC memory will not get freed. This problem really does happen, and can be pretty frustrating to track down.
 +
* It's possible for a function to pass out of it a reference to data on that function's stack frame. By then allocating a new stack frame over the old data, and not initializing, the reference to the old data may still appear to be valid. The program will then behave erratically. Initializing all data on the stack frame will greatly increase the probability of forcing that bug into the open in a repeatable manner.
 +
* Uninitialized data can be a source of bugs and trouble, even when used correctly. One design goal of D is to improve reliability and portability by eliminating sources of undefined behavior, and uninitialized data is one huge source of undefined, unportable, erratic and unpredictable behavior. Hence this idiom should only be used after other opportunities for speed optimization are exhausted and if benchmarking shows that it really does speed up the overall execution.
 +
 +
== Interrupt Service Routines ==
 +
When the garbage collector does a collection pass, it must pause all running threads in order to scan their stacks and register contents for references to GC allocated objects. If an ISR (Interrupt Service Routine) thread is paused, this can break the program.
 +
 +
Therefore, the ISR thread should not be paused. Threads created with the [http://dlang.org/phobos/core_thread.html core.thread] functions will be paused. But threads created with C's <code>_beginthread()</code> or equivalent won't be, the GC won't know they exist.
 +
 +
For this to work successfully:
 +
 +
* The ISR thread cannot allocate any memory using the GC. This means that the global <code>new</code> cannot be used. Nor can dynamic arrays be resized, nor can any elements be added to associative arrays. Any use of the D runtime library should be examined for any possibility of allocating GC memory - or better yet, the ISR should not call any D runtime library functions at all.
 +
* The ISR cannot hold the sole reference to any GC allocated memory, otherwise the GC may free the memory while the ISR is still using it. The solution is to have one of the paused threads hold a reference to it too, or store a reference to it in global data.

Revision as of 15:51, 7 March 2014

Most non-trivial programs needs to allocate and free memory. Memory management techniques become more and more important as programs increase in complexity, size, and performance. D offers many options for managing memory.

The three primary methods for allocating memory in D are:

  • Static data, allocated in the default data segment.
  • Stack data, allocated on the CPU program stack.
  • Garbage collected data, allocated dynamically on the garbage collection heap.

The techniques for using them, as well as some advanced alternatives are:

Strings (and Array) Copy-on-Write

Consider the case of passing an array to a function, possibly modifying the contents of the array, and returning the modified array. As the contents of an array are accessed through a reference, a crucial issue is who owns the contents of the array? For example, a function to convert an array of characters to upper case:

char[] toupper(char[] s)
{
    int i;

    for (i = 0; i < s.length; i++)
    {
        char c = s[i];
        if ('a' <= c && c <= 'z')
            s[i] = cast(char)(c - ('a' - 'A'));
    }
    return s;
}

Note that the caller's version of s[] is also modified. This may be not at all what was intended, or worse, s[] may be a slice into a read-only section of memory.

If a copy of s[] was always made by toupper(), then that will unnecessarily consume time and memory for strings that are already all upper case.

The solution is to implement copy-on-write, which means that a copy is made only if the string needs to be modified. Some string processing languages do do this as the default behavior, but there is a huge cost to it. The string "abcdeF" will wind up being copied 5 times by the function. To get the maximum efficiency using the protocol, it'll have to be done explicitly in the code. Here's toupper() rewritten to implement copy-on-write in an efficient manner:

char[] toupper(char[] s)
{
    int changed;
    int i;

    changed = 0;
    for (i = 0; i < s.length; i++)
    {
        char c = s[i];
        if ('a' <= c && c <= 'z')
        {
            if (!changed)
            {
                char[] r = new char[s.length];
                r[] = s;
                s = r;
                changed = 1;
            }
            s[i] = cast(char)(c - ('a' - 'A'));
        }
    }
    return s;
}

Copy-on-write is the protocol implemented by array processing functions in the D Phobos runtime library.

Real Time

Real time programming means that a program must be able to guarantee a maximum latency, or time to complete an operation. With most memory allocation schemes, including malloc/free and garbage collection, the latency is theoretically not bound. The most reliable way to guarantee latency is to preallocate all data that will be needed by the time critical portion. If no calls to allocate memory are done, the GC will not run and so will not cause the maximum latency to be exceeded.

Smooth Operation

Related to real time programming is the need for a program to operate smoothly, without arbitrary pauses while the garbage collector stops everything to run a collection. An example of such a program would be an interactive shooter type game. Having the game play pause erratically, while not fatal to the program, can be annoying to the user.

There are several techniques to eliminate or mitigate the effect:

  • Preallocate all data needed before the part of the code that needs to be smooth is run.
  • Manually run a GC collection cycle at points in program execution where it is already paused. An example of such a place would be where the program has just displayed a prompt for user input and the user has not responded yet. This reduces the odds that a collection cycle will be needed during the smooth code.
  • Call std.gc.disable() before the smooth code is run, and std.gc.enable() afterwards. This will cause the GC to favor allocating more memory instead of running a collection pass.

Free Lists

Free lists are a great way to accelerate access to a frequently allocated and discarded type. The idea is simple - instead of deallocating an object when done with it, put it on a free list. When allocating, pull one off the free list first.

class Foo
{
    static Foo freelist;		// start of free list

    static Foo allocate()
    {   Foo f;

	if (freelist)
	{   f = freelist;
	    freelist = f.next;
	}
	else
	    f = new Foo();
	return f;
    }

    static void deallocate(Foo f)
    {
	f.next = freelist;
	freelist = f;
    }

    Foo next;		// for use by FooFreeList
    ...
}

void test()
{
    Foo f = Foo.allocate();
    ...
    Foo.deallocate(f);
}

Such free list approaches can be very high performance.

  • If used by multiple threads, the allocate() and deallocate() functions need to be synchronized.
  • The Foo constructor is not re-run by allocate() when allocating from the free list, so the allocator may need to reinitialize some of the members.
  • It is not necessary to practice RAII with this, since if any objects are not passed to deallocate() when done, because of a thrown exception, they'll eventually get picked up by the GC anyway.

Reference Counting

The idea behind reference counting is to include a count field in the object. Increment it for each additional reference to it, and decrement it whenever a reference to it ceases. When the count hits 0, the object can be deleted.

D doesn't provide any automated support for reference counting, it will have to be done explicitly.

Win32 COM programming uses the members AddRef() and Release() to maintain the reference counts.

Explicit Class Instance Allocation

Note: Class allocators and deallocators are deprecated in D2.

D provides a means of creating custom allocators and deallocators for class instances. Normally, these would be allocated on the garbage collected heap, and deallocated when the collector decides to run. For specialized purposes, this can be handled by creating NewDeclarations and DeleteDeclarations. For example, to allocate using the C runtime library's malloc and free:

import std.c.stdlib;
import core.exception;
import core.memory : GC;

class Foo
{
    new(size_t sz)
    {
        void* p;

        p = std.c.stdlib.malloc(sz);

        if (!p)
            throw new OutOfMemoryError();

        GC.addRange(p, sz);
        return p;
    }

    delete(void* p)
    {
        if (p)
        {
            GC.removeRange(p);
            std.c.stdlib.free(p);
        }
    }
}

The critical features of new() are:

  • new() does not have a return type specified, but it is defined to be void*. new() must return a void*.
  • If new() cannot allocate memory, it must not return null, but must throw an exception.
  • The pointer returned from new() must be to memory aligned to the default alignment. This is 8 on win32 systems.
  • The size parameter is needed in case the allocator is called from a class derived from Foo and is a larger size than Foo.
  • A null is not returned if storage cannot be allocated. Instead, an exception is thrown. Which exception gets thrown is up to the programmer, in this case, OutOfMemory() is.
  • When scanning memory for root pointers into the garbage collected heap, the static data segment and the stack are scanned automatically. The C heap is not. Therefore, if Foo or any class derived from Foo using the allocator contains any references to data allocated by the garbage collector, the GC needs to be notified. This is done with the std.gc.addRange() method.
  • No initialization of the memory is necessary, as code is automatically inserted after the call to new() to set the class instance members to their defaults and then the constructor (if any) is run.

The critical features of delete() are:

  • The destructor (if any) has already been called on the argument p, so the data it points to should be assumed to be garbage.
  • The pointer p may be null.
  • If the GC was notified with std.gc.addRange(), a corresponding call to std.gc.removeRange() must happen in the deallocator.
  • If there is a delete(), there should be a corresponding new().

If memory is allocated using class specific allocators and deallocators, careful coding practices must be followed to avoid memory leaks and dangling references. In the presence of exceptions, it is particularly important to practice RAII to prevent memory leaks.

Custom allocators and deallocators can be done for structs and unions, too.

Mark/Release

Mark/Release is equivalent to a stack method of allocating and freeing memory. A ‘stack’ is created in memory. Objects are allocated by simply moving a pointer down the stack. Various points are ‘marked’, and then whole sections of memory are released simply by resetting the stack pointer back to a marked point.

import std.c.stdlib;
import core.exception;
import core.memory : GC;

class Foo
{
    static void[] buffer;
    static int bufindex;
    static const int bufsize = 100;

    static this()
    {
        void* p;
        p = malloc(bufsize);

        if (p is null)
            throw new OutOfMemoryError;

        GC.addRange(p, bufsize);
        buffer = p[0 .. bufsize];
    }

    static ~this()
    {
        if (buffer.length)
        {
            GC.removeRange(buffer.ptr);
            free(buffer.ptr);
            buffer = null;
        }
    }

    new(size_t sz)
    {
        void* p;

        p = &buffer[bufindex];
        bufindex += sz;

        if (bufindex > buffer.length)
            throw new OutOfMemoryError;

        return p;
    }

    delete(void* p)
    {
        assert(0);
    }

    static int mark()
    {
        return bufindex;
    }

    static void release(int i)
    {
        bufindex = i;
    }
}

void main()
{
    int m  = Foo.mark();
    Foo f1 = new Foo;           // allocate
    Foo f2 = new Foo;           // allocate

    Foo.release(m);             // deallocate f1 and f2
}

The allocation of buffer[] itself is added as a region to the GC, so there is no need for a separate call inside Foo.new() to do it.

RAII (Resource Acquisition Is Initialization)

RAII techniques can be useful in avoiding memory leaks when using explicit allocators and deallocators. Adding the scope attribute to such classes can help.

Allocating Class Instances On The Stack

Class instances are normally allocated on the garbage collected heap. However, if they:

  • are allocated as local symbols in a function and
  • are allocated using new and
  • use new with no arguments (constructor arguments are allowed) and
  • have the scope storage class

then they are allocated on the stack. This is more efficient than doing an allocate/free cycle on the instance. But be careful that any reference to the object does not survive the return of the function.

class C { ... }

scope c = new C();	// c is allocated on the stack
scope c2 = new C(5);    // allocated on stack
scope c3 = new(5) C();  // allocated by a custom allocator

If the class has a destructor, then that destructor is guaranteed to be run when the class object goes out of scope, even if the scope is exited via an exception.

Allocating Uninitialized Arrays On The Stack

Arrays are always initialized in D. So, the following declaration:

void foo()
{   byte[1024] buffer;

    fillBuffer(buffer);
    ...
}

will not be as fast as it might be since the buffer[] contents are always initialized. If careful profiling of the program shows that this initialization is a speed problem, it can be eliminated using a VoidInitializer:


void foo()
{   byte[1024] buffer = void;

    fillBuffer(buffer);
    ...
}

Uninitialized data on the stack comes with some caveats that need to be carefully evaluated before using:

  • The uninitialized data that is on the stack will get scanned by the garbage collector looking for any references to allocated memory. Since the uninitialized data consists of old D stack frames, it is highly likely that some of that garbage will look like references into the GC heap, and the GC memory will not get freed. This problem really does happen, and can be pretty frustrating to track down.
  • It's possible for a function to pass out of it a reference to data on that function's stack frame. By then allocating a new stack frame over the old data, and not initializing, the reference to the old data may still appear to be valid. The program will then behave erratically. Initializing all data on the stack frame will greatly increase the probability of forcing that bug into the open in a repeatable manner.
  • Uninitialized data can be a source of bugs and trouble, even when used correctly. One design goal of D is to improve reliability and portability by eliminating sources of undefined behavior, and uninitialized data is one huge source of undefined, unportable, erratic and unpredictable behavior. Hence this idiom should only be used after other opportunities for speed optimization are exhausted and if benchmarking shows that it really does speed up the overall execution.

Interrupt Service Routines

When the garbage collector does a collection pass, it must pause all running threads in order to scan their stacks and register contents for references to GC allocated objects. If an ISR (Interrupt Service Routine) thread is paused, this can break the program.

Therefore, the ISR thread should not be paused. Threads created with the core.thread functions will be paused. But threads created with C's _beginthread() or equivalent won't be, the GC won't know they exist.

For this to work successfully:

  • The ISR thread cannot allocate any memory using the GC. This means that the global new cannot be used. Nor can dynamic arrays be resized, nor can any elements be added to associative arrays. Any use of the D runtime library should be examined for any possibility of allocating GC memory - or better yet, the ISR should not call any D runtime library functions at all.
  • The ISR cannot hold the sole reference to any GC allocated memory, otherwise the GC may free the memory while the ISR is still using it. The solution is to have one of the paused threads hold a reference to it too, or store a reference to it in global data.