I would like to build a memory allocator to manage a large chunk of non-shared, linear address space (e.g. 32GB). I need to run this on the Big-3 operating systems, but I can live with having OS-specific code.
Functionality desired:
- At start, acquire a 2^N byte contiguous region of virtual address space aligned on a 2^N byte boundary
- Initially, the entire region should be inaccessible (i.e. touching any address within this region should result in that OS's equivalent of a SEGV)
- Over time, from the base of this region (the lowest address), I will grow the portion that should be made accessible (i.e. allocate backing store, provision PTEs)
- It is acceptable, while running, to encounter a failure due to backing store exhaustion
Bonus points:
- It is desirable that, initially, no backing store be reserved
- Unprovisioning the entire region without releasing the address space
- Encourage use of larger pages and TLB entries
Questions:
- I know that I can implement this scheme using mmap on Linux. Is it doable on Win64 and MacOS?
- If yes, what primitives / system calls should I be looking at?
- How can I best achieve 2^N byte alignment when initially acquiring address space?