About Virtual Memory

Virtual memory allows an operating system to escape the limitations of physical RAM. The virtual memory manager creates a logical address space (or “virtual” address space) for each process and divides it up into uniformly-sized chunks of memory called pages. The processor and its memory management unit (MMU) maintain a page table to map pages in the program’s logical address space to hardware addresses in the computer’s RAM. When a program’s code accesses an address in memory, the MMU uses the page table to translate the specified logical address into the actual hardware memory address. This translation occurs automatically and is transparent to the running application.

As far as a program is concerned, addresses in its logical address space are always available. However, if an application accesses an address on a memory page that is not currently in physical RAM, a page fault occurs. When that happens, the virtual memory system invokes a special page-fault handler to respond to the fault immediately. The page-fault handler stops the currently executing code, locates a free page of physical memory, loads the page containing the needed data from disk, updates the page table, and then returns control to the program’s code, which can then access the memory address normally. This process is known as paging.

If there are no free pages available in physical memory, the handler must first release an existing page to make room for the new page. How the system release pages depends on the platform. In OS X, the virtual memory system often writes pages to the backing store. The backing store is a disk-based repository containing a copy of the memory pages used by a given process. Moving data from physical memory to the backing store is called paging out (or “swapping out”); moving data from the backing store back in to physical memory is called paging in (or “swapping in”). In iOS, there is no backing store and so pages are are never paged out to disk, but read-only pages are still be paged in from disk as needed.

In OS X and in earlier versions of iOS, the size of a page is 4 kilobytes. In later versions of iOS, A7- and A8-based systems expose 16-kilobyte pages to the 64-bit userspace backed by 4-kilobyte physical pages, while A9 systems expose 16-kilobyte pages backed by 16-kilobyte physical pages. These sizes determine how many kilobytes the system reads from disk when a page fault occurs. Disk thrashing can occur when the system spends a disproportionate amount of time handling page faults and reading and writing pages, rather than executing code for a program.

Paging of any kind, and disk thrashing in particular, affects performance negatively because it forces the system to spend a lot of time reading and writing to disk. Reading a page in from the backing store takes a significant amount of time and is much slower than reading directly from RAM. If the system has to write a page to disk before it can read another page from disk, the performance impact is even worse.

Details of the Virtual Memory System

The logical address space of a process consists of mapped regions of memory. Each mapped memory region contains a known number of virtual memory pages. Each region has specific attributes controlling such things as inheritance (portions of the region may be mapped from “parent” regions), write-protection, and whether it is wired (that is, it cannot be paged out). Because regions contain a known number of pages, they are page-aligned, meaning the starting address of the region is also the starting address of a page and the ending address also defines the end of a page.

The kernel associates a VM object with each region of the logical address space. The kernel uses VM objects to track and manage the resident and nonresident pages of the associated regions. A region can map to part of the backing store or to a memory-mapped file in the file system. Each VM object contains a map that associates regions with either the default pager or the vnode pager. The default pager is a system manager that manages the nonresident virtual memory pages in the backing store and fetches those pages when requested. The vnode pager implements memory-mapped file access. The vnode pager uses the paging mechanism to provide a window directly into a file. This mechanism lets you read and write portions of the file as if they were located in memory.

In addition to mapping regions to either the default or vnode pager, a VM object may also map regions to another VM object. The kernel uses this self referencing technique to implement copy-on-write regions. Copy-on-write regions allow different processes (or multiple blocks of code within a process) to share a page as long as none of them write to that page. When a process attempts to write to the page, a copy of the page is created in the logical address space of the process doing the writing. From that point forward, the writing process maintains its own separate copy of the page, which it can write to at any time. Copy-on-write regions let the system share large quantities of data efficiently in memory while still letting processes manipulate those pages directly (and safely) if needed. These types of regions are most commonly used for the data pages loaded from system frameworks.

Each VM object contains several fields, as shown in Table 1.

If the VM object is involved in a copy-on-write (vm_copy) operation, the shadow and copy fields may point to other VM objects. Otherwise both fields are usually NULL.

Wired Memory

Wired memory (also called resident memory) stores kernel code and data structures that must never be paged out to disk. Applications, frameworks, and other user-level software cannot allocate wired memory. However, they can affect how much wired memory exists at any time. For example, an application that creates threads and ports implicitly allocates wired memory for the required kernel resources that are associated with them.

Table 2 lists some of the wired-memory costs for application-generated entities.

As you can see, every thread, process, and library contributes to the resident footprint of the system. In addition to your application using wired memory, however, the kernel itself requires wired memory for the following entities:

• VM objects

• the virtual memory buffer cache

• I/O buffer caches

• drivers

Wired data structures are also associated with the physical page and map tables used to store virtual-memory mapping information, Both of these entities scale with the amount of available physical memory. Consequently, when you add memory to a system, the amount of wired memory increases even if nothing else changes. When a computer is first booted into the Finder, with no other applications running, wired memory can consume approximately 14 megabytes of a 64 megabyte system and 17 megabytes of a 128 megabyte system.

Wired memory pages are not immediately moved back to the free list when they become invalid. Instead they are “garbage collected” when the free-page count falls below the threshold that triggers page out events.

Page Lists in the Kernel

The kernel maintains and queries three system-wide lists of physical memory pages:

• The active list contains pages that are currently mapped into memory and have been recently accessed.

• The inactive list contains pages that are currently resident in physical memory but have not been accessed recently. These pages contain valid data but may be removed from memory at any time.

• The free list contains pages of physical memory that are not associated with any address space of VM object. These pages are available for immediate use by any process that needs them.

When the number of pages on the free list falls below a threshold (determined by the size of physical memory), the pager attempts to balance the queues. It does this by pulling pages from the inactive list. If a page has been accessed recently, it is reactivated and placed on the end of the active list. In OS X, if an inactive page contains data that has not been written to the backing store recently, its contents must be paged out to disk before it can be placed on the free list. (In iOS, modified but inactive pages must remain in memory and be cleaned up by the application that owns them.) If an inactive page has not been modified and is not permanently resident (wired), it is stolen (any current virtual mappings to it are destroyed) and added to the free list. Once the free list size exceeds the target threshold, the pager rests.

The kernel moves pages from the active list to the inactive list if they are not accessed; it moves pages from the inactive list to the active list on a soft fault (see Paging In Process). When virtual pages are swapped out, the associated physical pages are placed in the free list. Also, when processes explicitly free memory, the kernel moves the affected pages to the free list.

Paging Out Process

In OS X, when the number of pages in the free list dips below a computed threshold, the kernel reclaims physical pages for the free list by swapping inactive pages out of memory. To do this, the kernel iterates all resident pages in the active and inactive lists, performing the following steps:

1. If a page in the active list is not recently touched, it is moved to the inactive list.

2. If a page in the inactive list is not recently touched, the kernel finds the page’s VM object.

3. If the VM object has never been paged before, the kernel calls an initialization routine that creates and assigns a default pager object.

4. The VM object’s default pager attempts to write the page out to the backing store.

5. If the pager succeeds, the kernel frees the physical memory occupied by the page and moves the page from the inactive to the free list.

Paging In Process

The final phase of virtual memory management moves pages into physical memory, either from the backing store or from the file containing the page data. A memory access fault initiates the page-in process. A memory access fault occurs when code tries to access data at a virtual address that is not mapped to physical memory. There are two kinds of faults:

• A soft fault occurs when the page of the referenced address is resident in physical memory but is currently not mapped into the address space of this process.

• A hard fault occurs when the page of the referenced address is not in physical memory but is swapped out to backing store (or is available from a mapped file). This is what is typically known as a page fault.

When any type of fault occurs, the kernel locates the map entry and VM object for the accessed region. The kernel then goes through the VM object’s list of resident pages. If the desired page is in the list of resident pages, the kernel generates a soft fault. If the page is not in the list of resident pages, it generates a hard fault.

For soft faults, the kernel maps the physical memory containing the pages to the virtual address space of the process. The kernel then marks the specific page as active. If the fault involved a write operation, the page is also marked as modified so that it will be written to backing store if it needs to be freed later.

For hard faults, the VM object’s pager finds the page in the backing store or from the file on disk, depending on the type of pager. After making the appropriate adjustments to the map information, the pager moves the page into physical memory and places the page on the active list. As with a soft fault, if the fault involved a write operation, the page is marked as modified.