mmap() and actual physical allocation of the pages under Linux - linux-kernel

Question

When I do VirtualAlloc with MEM_RESERVE | MEM_COMMIT under Windows the virtually allocated pages actually don't get allocated physical pages immediately but the pages are first subtracted from the paging-file and then are mapped on demand as soon as there is a first access to the page (I've measured delays of >= 1.000 clock cycles for this mapping).
But what's with Linux ? When I have overcommitting it seems obvious to me that the physical pages get assigned immediately. But what's when I switch off overcommitting and have a mmap() ? Will the system just subtract the necessary space from the paging-file/-partition to have a backing if physical assignment would fail immediately ? I.e. are the pages then allocated dynamically like under Windows ?

Answer 1

Ok, so I wrote a little program that scans freshly mmap()-allocated pages
#if defined(_MSC_VER)
#include <Windows.h>
#elif defined(__unix__)
#include <sys/mman.h>
#endif
#include <iostream>
#include <chrono>
#include <atomic>
using namespace std;
using namespace chrono;
int main()
{
static size_t const SIZE = (size_t)8 << 30, N_PAGES = SIZE >> 12;
#if defined(_MSC_VER)
void *p = VirtualAlloc( nullptr, SIZE, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE );
#elif defined(__unix__)
void *p = mmap( nullptr, SIZE, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_SHARED, 0, 0 );
#endif
if( !p )
return EXIT_FAILURE;
auto scan = [&]()
{
auto start = high_resolution_clock::now();
for( atomic<char> *pc = (atomic<char> *)p, *pcEnd = pc + SIZE; pc < pcEnd; pc->store( 0, memory_order_relaxed ), pc += 0x1000 );
return (double)(int64_t)duration_cast<nanoseconds>( high_resolution_clock::now() - start ).count() / N_PAGES;
};
cout << scan() << endl;
cout << scan() << endl;
}
The first scan() is about 1.800ns per page on my Linux Ryzen 7 1800X PC, the latter is about 60ns. And my Ubuntu-computer has overcommitting enabled, so this would apply for non-overcommitting as well for sure. So Linux has the same behaviour like Windows here. My Windows-PC with a Ryzen Threadripper 3990X is significantly faster here; the first scan() is about 700ns per page.