makeoptions DEBUG="-g" makeoptions COPTFLAGS="-O -pipe"
Chapter 7. Virtual Memory System
7.1. Management of Physical Memory
Physical memory is managed on a page-by-page basis through the
vm_page_t structure. Pages of physical memory are categorized through the placement of their respective
vm_page_t structures on one of several paging queues.
A page can be in a wired, active, inactive, cache, or free state. Except for the wired state, the page is typically placed in a doubly link list queue representing the state that it is in. Wired pages are not placed on any queue.
FreeBSD implements a more involved paging queue for cached and free pages in order to implement page coloring. Each of these states involves multiple queues arranged according to the size of the processor’s L1 and L2 caches. When a new page needs to be allocated, FreeBSD attempts to obtain one that is reasonably well aligned from the point of view of the L1 and L2 caches relative to the VM object the page is being allocated for.
Additionally, a page may be held with a reference count or locked with a busy count. The VM system also implements an "ultimate locked" state for a page using the PG_BUSY bit in the page’s flags.
In general terms, each of the paging queues operates in a LRU fashion. A page is typically placed in a wired or active state initially. When wired, the page is usually associated with a page table somewhere. The VM system ages the page by scanning pages in a more active paging queue (LRU) in order to move them to a less-active paging queue. Pages that get moved into the cache are still associated with a VM object but are candidates for immediate reuse. Pages in the free queue are truly free. FreeBSD attempts to minimize the number of pages in the free queue, but a certain minimum number of truly free pages must be maintained in order to accommodate page allocation at interrupt time.
If a process attempts to access a page that does not exist in its page table but does exist in one of the paging queues (such as the inactive or cache queues), a relatively inexpensive page reactivation fault occurs which causes the page to be reactivated. If the page does not exist in system memory at all, the process must block while the page is brought in from disk.
FreeBSD dynamically tunes its paging queues and attempts to maintain reasonable ratios of pages in the various queues as well as attempts to maintain a reasonable breakdown of clean versus dirty pages. The amount of rebalancing that occurs depends on the system’s memory load. This rebalancing is implemented by the pageout daemon and involves laundering dirty pages (syncing them with their backing store), noticing when pages are activity referenced (resetting their position in the LRU queues or moving them between queues), migrating pages between queues when the queues are out of balance, and so forth. FreeBSD’s VM system is willing to take a reasonable number of reactivation page faults to determine how active or how idle a page actually is. This leads to better decisions being made as to when to launder or swap-out a page.
7.2. The Unified Buffer Cache
FreeBSD implements the idea of a generic "VM object". VM objects can be associated with backing store of various typesunbacked, swap-backed, physical device-backed, or file-backed storage. Since the filesystem uses the same VM objects to manage in-core data relating to files, the result is a unified buffer cache.
VM objects can be shadowed. That is, they can be stacked on top of each other. For example, you might have a swap-backed VM object stacked on top of a file-backed VM object in order to implement a MAP_PRIVATE mmap()ing. This stacking is also used to implement various sharing properties, including copy-on-write, for forked address spaces.
It should be noted that a
vm_page_t can only be associated with one VM object at a time. The VM object shadowing implements the perceived sharing of the same page across multiple instances.
7.3. Filesystem I/O
vnode-backed VM objects, such as file-backed objects, generally need to maintain their own clean/dirty info independent from the VM system’s idea of clean/dirty. For example, when the VM system decides to synchronize a physical page to its backing store, the VM system needs to mark the page clean before the page is actually written to its backing store. Additionally, filesystems need to be able to map portions of a file or file metadata into KVM in order to operate on it.
The entities used to manage this are known as filesystem buffers,
struct buf's, or
bp's. When a filesystem needs to operate on a portion of a VM object, it typically maps part of the object into a struct buf and then maps the pages in the struct buf into KVM. In the same manner, disk I/O is typically issued by mapping portions of objects into buffer structures and then issuing the I/O on the buffer structures. The underlying vm_page_t’s are typically busied for the duration of the I/O. Filesystem buffers also have their own notion of being busy, which is useful to filesystem driver code which would rather operate on filesystem buffers instead of hard VM pages.
FreeBSD reserves a limited amount of KVM to hold mappings from struct bufs, but it should be made clear that this KVM is used solely to hold mappings and does not limit the ability to cache data. Physical data caching is strictly a function of
vm_page_t's, not filesystem buffers. However, since filesystem buffers are used to placehold I/O, they do inherently limit the amount of concurrent I/O possible. However, as there are usually a few thousand filesystem buffers available, this is not usually a problem.
7.4. Mapping Page Tables
FreeBSD separates the physical page table topology from the VM system. All hard per-process page tables can be reconstructed on the fly and are usually considered throwaway. Special page tables such as those managing KVM are typically permanently preallocated. These page tables are not throwaway.
FreeBSD associates portions of vm_objects with address ranges in virtual memory through
vm_entry_t structures. Page tables are directly synthesized from the
vm_object_t hierarchy. Recall that I mentioned that physical pages are only directly associated with a
vm_object; that is not quite true.
vm_page_t's are also linked into page tables that they are actively associated with. One
vm_page_t can be linked into several pmaps, as page tables are called. However, the hierarchical association holds, so all references to the same page in the same object reference the same
vm_page_t and thus give us buffer cache unification across the board.
7.5. KVM Memory Mapping
FreeBSD uses KVM to hold various kernel structures. The single largest entity held in KVM is the filesystem buffer cache. That is, mappings relating to
struct buf entities.
Unlike Linux, FreeBSD does not map all of physical memory into KVM. This means that FreeBSD can handle memory configurations up to 4G on 32 bit platforms. In fact, if the mmu were capable of it, FreeBSD could theoretically handle memory configurations up to 8TB on a 32 bit platform. However, since most 32 bit platforms are only capable of mapping 4GB of ram, this is a moot point.
KVM is managed through several mechanisms. The main mechanism used to manage KVM is the zone allocator. The zone allocator takes a chunk of KVM and splits it up into constant-sized blocks of memory in order to allocate a specific type of structure. You can use
vmstat -m to get an overview of current KVM utilization broken down by zone.
7.6. Tuning the FreeBSD VM System
A concerted effort has been made to make the FreeBSD kernel dynamically tune itself. Typically you do not need to mess with anything beyond the
NMBCLUSTERS kernel config options. That is, kernel compilation options specified in (typically) /usr/src/sys/i386/conf/CONFIG_FILE. A description of all available kernel configuration options can be found in /usr/src/sys/i386/conf/LINT.
In a large system configuration you may wish to increase
maxusers. Values typically range from 10 to 128. Note that raising
maxusers too high can cause the system to overflow available KVM resulting in unpredictable operation. It is better to leave
maxusers at some reasonable number and add other options, such as
NMBCLUSTERS, to increase specific resources.
If your system is going to use the network heavily, you may want to increase
NMBCLUSTERS. Typical values range from 1024 to 4096.
NBUF parameter is also traditionally used to scale the system. This parameter determines the amount of KVA the system can use to map filesystem buffers for I/O. Note that this parameter has nothing whatsoever to do with the unified buffer cache! This parameter is dynamically tuned in 3.0-CURRENT and later kernels and should generally not be adjusted manually. We recommend that you not try to specify an
NBUF parameter. Let the system pick it. Too small a value can result in extremely inefficient filesystem operation while too large a value can starve the page queues by causing too many pages to become wired down.
By default, FreeBSD kernels are not optimized. You can set debugging and optimization flags with the
makeoptions directive in the kernel configuration. Note that you should not use
-g unless you can accommodate the large (typically 7 MB+) kernels that result.
Sysctl provides a way to tune kernel parameters at run-time. You typically do not need to mess with any of the sysctl variables, especially the VM related ones.
Run time VM and system tuning is relatively straightforward. First, use Soft Updates on your UFS/FFS filesystems whenever possible. /usr/src/sys/ufs/ffs/README.softupdates contains instructions (and restrictions) on how to configure it.
Second, configure sufficient swap. You should have a swap partition configured on each physical disk, up to four, even on your "work" disks. You should have at least 2x the swap space as you have main memory, and possibly even more if you do not have a lot of memory. You should also size your swap partition based on the maximum memory configuration you ever intend to put on the machine so you do not have to repartition your disks later on. If you want to be able to accommodate a crash dump, your first swap partition must be at least as large as main memory and /var/crash must have sufficient free space to hold the dump.
NFS-based swap is perfectly acceptable on 4.X or later systems, but you must be aware that the NFS server will take the brunt of the paging load.