Virtual memory ============== FreeBSD's virtual memory system connects a process's address space to the objects that supply its data and the physical pages that currently hold it. The machine-independent implementation lives primarily in ``sys/vm``; architecture-specific ``pmap`` implementations manage hardware translations. An address-space mapping, a VM object, and a resident page have different lifetimes. Understanding those lifetimes is the starting point for following faults, copy-on-write, file mappings, and memory pressure. .. toctree:: :maxdepth: 1 address-spaces faults-and-reclaim locking api/index Finding the implementation -------------------------- .. list-table:: Source map :header-rows: 1 :widths: 35 65 * - Source - Responsibility * - ``vm_map.c``, ``vm_map.h`` - Address ranges, protections, mapping changes, and lookup. * - ``vm_object.c``, ``vm_object.h`` - Backing objects, resident-page indexing, and object lifetimes. * - ``vm_page.c``, ``vm_page.h`` - Physical-page metadata, allocation, busy state, and wiring. * - ``vm_fault.c`` - Resolving a virtual address to a page and installing a translation. * - ``vm_pager.c``, ``vm_pager.h`` - Dispatch to the object type's backing-store operations. * - ``vnode_pager.c``, ``swap_pager.c`` - File-backed and swap-backed object behavior. * - ``vm_pageout.c``, ``vm_pagequeue.h`` - Page aging, reclamation, and laundry processing. * - ``vm_phys.c``, ``vm_domainset.c``, ``vm_reserv.c`` - Physical allocation, NUMA policy, and page reservations. * - ``uma_core.c``, ``uma.h`` - The kernel's Universal Memory Allocator for object-sized allocations. All paths in this table are relative to ``sys/vm``. ``sys/vm/pmap.h`` declares the common translation interface; its implementations are in the architecture directories under ``sys``. Connections to other topics --------------------------- File-backed objects connect VM to :doc:`../vfs/index` through the vnode pager. Waiting for a busy page or memory allocation can block a thread, connecting VM to :doc:`../scheduler/index`. UMA allocates many kernel structures, but its zones and caches are a separate layer from user address-space mappings.