std.experimental.allocator.building_blocks.ascending_page_allocator

Source

std/experimental/allocator/building_blocks/ascending_page_allocator.d

struct AscendingPageAllocator;

AscendingPageAllocator is a fast and safe allocator that rounds all allocations to multiples of the system's page size. It reserves a range of virtual addresses (using mmap on Posix and VirtualAlloc on Windows) and allocates memory at consecutive virtual addresses.

When a chunk of memory is requested, the allocator finds a range of virtual pages that satisfy the requested size, changing their protection to read/write using OS primitives (mprotect and VirtualProtect, respectively). The physical memory is allocated on demand, when the pages are accessed.

Deallocation removes any read/write permissions from the target pages and notifies the OS to reclaim the physical memory, while keeping the virtual memory.

Because the allocator does not reuse memory, any dangling references to deallocated memory will always result in deterministically crashing the process.

See Also:

Simple Fast and Safe Manual Memory Management for the general approach.

Examples:

import core.memory : pageSize;

size_t numPages = 100;
void[] buf;
void[] prevBuf = null;
AscendingPageAllocator a = AscendingPageAllocator(numPages * pageSize);

foreach (i; 0 .. numPages)
{
    // Allocation is rounded up to page size
    buf = a.allocate(pageSize - 100);
    writeln(buf.length); // pageSize - 100

    // Allocations are served at increasing addresses
    if (prevBuf)
        writeln(prevBuf.ptr + pageSize); // buf.ptr

    assert(a.deallocate(buf));
    prevBuf = buf;
}

nothrow @nogc this(size_t n);

Rounds the mapping size to the next multiple of the page size and calls the OS primitive responsible for creating memory mappings: mmap on POSIX and VirtualAlloc on Windows.

Parameters:

size_t n

mapping size in bytes

nothrow @nogc size_t goodAllocSize(size_t n);

Rounds the requested size to the next multiple of the page size.

nothrow @nogc void deallocate(void[] b);

Decommit all physical memory associated with the buffer given as parameter, but keep the range of virtual addresses.

On POSIX systems deallocate calls mmap with `MAP_FIXED' a second time to decommit the memory. On Windows, it uses VirtualFree with MEM_DECOMMIT.

nothrow @nogc Ternary owns(void[] buf);

Returns Ternary.yes if the passed buffer is inside the range of virtual adresses. Does not guarantee that the passed buffer is still valid.

nothrow @nogc bool deallocateAll();

Removes the memory mapping causing all physical memory to be decommited and the virtual address space to be reclaimed.

nothrow @nogc size_t getAvailableSize();

Returns the available size for further allocations in bytes.

nothrow @nogc void[] allocate(size_t n);

Rounds the allocation size to the next multiple of the page size. The allocation only reserves a range of virtual pages but the actual physical memory is allocated on demand, when accessing the memory.

Parameters:

size_t n

Bytes to allocate

Returns:

null on failure or if the requested size exceeds the remaining capacity.

nothrow @nogc void[] alignedAllocate(size_t n, uint a);

Rounds the allocation size to the next multiple of the page size. The allocation only reserves a range of virtual pages but the actual physical memory is allocated on demand, when accessing the memory.

The allocated memory is aligned to the specified alignment a.

Parameters:

size_t n	Bytes to allocate
uint a	Alignment

Returns:

null on failure or if the requested size exceeds the remaining capacity.

nothrow @nogc bool expand(ref void[] b, size_t delta);

If the passed buffer is not the last allocation, then delta can be at most the number of bytes left on the last page. Otherwise, we can expand the last allocation until the end of the virtual address range.

nothrow @nogc Ternary empty();

Returns Ternary.yes if the allocator does not contain any alive objects and Ternary.no otherwise.

struct SharedAscendingPageAllocator;

SharedAscendingPageAllocator is the threadsafe version of AscendingPageAllocator.

Examples:

import core.memory : pageSize;
import core.thread : ThreadGroup;

enum numThreads = 100;
shared SharedAscendingPageAllocator a = SharedAscendingPageAllocator(pageSize * numThreads);

void fun()
{
    void[] b = a.allocate(pageSize);
    writeln(b.length); // pageSize

    assert(a.deallocate(b));
}

auto tg = new ThreadGroup;
foreach (i; 0 .. numThreads)
{
    tg.create(&fun);
}
tg.joinAll();

shared nothrow @nogc this(size_t n);

Rounds the mapping size to the next multiple of the page size and calls the OS primitive responsible for creating memory mappings: mmap on POSIX and VirtualAlloc on Windows.

Parameters:

size_t n

mapping size in bytes

shared nothrow @nogc size_t goodAllocSize(size_t n);

Rounds the requested size to the next multiple of the page size.

shared nothrow @nogc void deallocate(void[] b);

Decommit all physical memory associated with the buffer given as parameter, but keep the range of virtual addresses.

On POSIX systems deallocate calls mmap with `MAP_FIXED' a second time to decommit the memory. On Windows, it uses VirtualFree with MEM_DECOMMIT.

shared nothrow @nogc Ternary owns(void[] buf);

Returns Ternary.yes if the passed buffer is inside the range of virtual adresses. Does not guarantee that the passed buffer is still valid.

shared nothrow @nogc bool deallocateAll();

Removes the memory mapping causing all physical memory to be decommited and the virtual address space to be reclaimed.

shared nothrow @nogc size_t getAvailableSize();

Returns the available size for further allocations in bytes.

shared nothrow @nogc void[] allocate(size_t n);

Rounds the allocation size to the next multiple of the page size. The allocation only reserves a range of virtual pages but the actual physical memory is allocated on demand, when accessing the memory.

Parameters:

size_t n

Bytes to allocate

Returns:

null on failure or if the requested size exceeds the remaining capacity.

shared nothrow @nogc void[] alignedAllocate(size_t n, uint a);

Rounds the allocation size to the next multiple of the page size. The allocation only reserves a range of virtual pages but the actual physical memory is allocated on demand, when accessing the memory.

The allocated memory is aligned to the specified alignment a.

Parameters:

size_t n	Bytes to allocate
uint a	Alignment

Returns:

null on failure or if the requested size exceeds the remaining capacity.

shared nothrow @nogc bool expand(ref void[] b, size_t delta);

If the passed buffer is not the last allocation, then delta can be at most the number of bytes left on the last page. Otherwise, we can expand the last allocation until the end of the virtual address range.