Agner`s CPU blog

The current default implementation of malloc() is ptmalloc2 and it uses both sbrk() and mmap(), but in different scenarios:
- Small to medium allocations on the main thread use sbrk(), because it doesn't matter too much if some blocks get free()'d later on since the allocator can still reuse the leftover holes for allocations of the same size.
- Large allocations on the main thread use mmap(), because sbrk() can only reclaim free()'d memory in the exact reverse order that it was allocated, so that a large allocation can easily become impossible to reclaim if it's followed by small longer-lived allocations.
- Allocation on other threads use mmap() instead of sbrk(), because only mmap() is thread-safe. Also, using sbrk() would create a series of segments that are used by different threads on the heap, which would make it hard to reclaim memory until all the threads have finished running.