- Overview of the SunOS 5.4 kernel memory allocator that is based on a set of object-caching primitives that reduce the cost of allocating complex objects by retaining their state between uses.
- The primitives prove equally effective for managing stateless memory (e.g. data pages and temp buffers) because they are space-effcient and fast.
- The allocator's object caches respond dynamically to global memory pressure, and employ an object-coloring scheme that improves the system's overall cache utilization and bus balance.
- The allocator also has several statistical and debugging features that can detect a wide range of problems throughout the system.
Motivation
- A fast kernel memory allocator is essential: allocation and freeing of objects are common kernel operations.
- Usually the cost of initializing and destroying the object exceeds the cost of allocating and freeing memory for it.
- Performance gains can be improved by caching frequently used objects so that their basic structure is preserved between uses.
The Case for Object Caching in the Central Allocator
Object caching can be implemented without any help from the central allocator - any subsystem can have a private implementation. However, there are several disadvantages to this approach:
- There is a natural tension between an object cache, which wants to keep memory, and the rest of the system, which wants that memory back. Privately-managed caches cannot handle this tension sensibly. They have limited insight into the system's overall memory needs and no insight into each other's needs. Similarly, the rest of the system has no knowledge of the existence of these caches and hence has no way to ``pull memory from them.
- Since private caches bypass the central allocator, they also bypass any accounting mechanisms and debugging features that allocator may possess. This makes the operating system more difficult to monitor and debug.
- Having many instances of the same solution to a common problem increases kernel code size and maintenance costs.
Object Caching
The idea behind object caching is to preserve the invariant portion of an object's initial state - its constructed state - between uses, so it does not have to be destroyed and recreated every time the object is used.
Object caching is important because the cost of constructing an object can be significantly higher than the cost of allocating memory for it. Object caching helps dealing with objects that are frequently allocated and freed. An object's embedded locks, condition variables, reference counts, lists of other objects, and read-only data all generally qualify as constructed state.
The design of an object cache is straightforward:
To allocate an object:
if (there's an object in the cache) {
take it (no construction required);
} else {
allocate memory;
construct the object;
}
To free an object:
return it to the cache (no destruction required);
To reclaim memory from the cache:
take some objects from the cache; destroy the objects; free the underlying memory;
An object's constructed state must be initialized only once - when the object is first brought into the cache. Once the cache is populated, allocating and freeing objects are fast, trivial operations.
Object caching requires a greater degree of cooperation between the allocator and its clients than the standard kmem_alloc(9F) / kmem_free(9F) interface allows. The next section develops an interface to support constructed object caching in the central allocator.
Another beneficial side-effect of object caching is that it reduces the instruction-cache footprint of the code that uses cached objects by moving the rarely-executed construction and destruction code out of the hot path.
Object Cache Interface
The interface presented here follows from two observations:
A. Descriptions of objects (name, size, alignment, constructor, and destructor) belong in the clients - not in the central allocator. The allocator should not just "know" that sizeof (struct inode) is a useful pool size, for example. Such assumptions are brittle and cannot anticipate the needs of third-party device drivers, streams modules and file systems.
B. Memory management policies belong in the central allocator - not in its clients. The clients just want to allocate and free objects quickly. They shouldn't have to worry about how to manage the underlying memory efficiently.
It follows from A that object cache creation must be client-driven and must include a full specification of the objects:
struct kmem_cache *kmem_cache_create(char *name, size_t size, int align, void (*constructor)(void *, size_t), void (*destructor)(void *, size_t));
Creates a cache of objects, each of size size, aligned on an align boundary. The alignment will always be rounded up to the minimum allowable value, so align can be zero whenever no special alignment is required. name identifies the cache for statistics and debugging. constructor is a function that constructs (that is, performs the one-time initialization of) objects in the cache; destructor undoes this, if applicable. The constructor and destructor take a size argument so that they can support families of similar caches, e.g. streams messages. kmem_cache_create returns an opaque descriptor for accessing the cache.
Tt follows from B that clients should need just two simple functions to allocate and free objects:
void *kmem_cache_alloc(struct kmem_cache *cp, int flags);
Gets an object from the cache. The object will be in its constructed state. flags is either KM_SLEEP or KM_NOSLEEP, indicating whether it's acceptable to wait for memory if none is currently available.
void kmem_cache_free(struct kmem_cache *cp, void *buf);
Returns an object to the cache. The object must still be in its constructed state.
Finally, if a cache is no longer needed the client can destroy it:
void kmem_cache_destroy(struct kmem_cache *cp);
Destroys the cache and reclaims all associated resources. All allocated objects must have been returned to the cache.
This interface allows us to build a flexible allocator that is ideally suited to the needs of its clients. In this sense it is a "custom" allocator. However, it does not have to be built with compile-time knowledge of its clients as most custom allocators do, nor does it have to keep guessing as in the adaptive-fit methods. Rather, the object-cache interface allows clients to specify the allocation services they need on the fly.
Slab Allocator Implementation
The implementation details of the SunOS 5.4 kernel memory allocator, or "slab allocator".
Caches
Each cache has a front end and back end which are designed to be as decoupled as possible:
back end front end -------- --------- --------------- kmem_cache_grow() --> | | --> kmem_cache_alloc() | cache | kmem_cache_reap() <-- | | <-- kmem_cache_free() ---------------
The front end is the public interface to the allocator. It moves objects to and from the cache, calling into the back end when it needs more objects.
The back end manages the flow of real memory through the cache. The influx routine (kmem_cache_grow()) gets memory from the VM system, makes objects out of it, and feeds those objects into the cache. The outflux routine (kmem_cache_reap()) is invoked by the VM system when it wants some of that memory back - e.g., at the onset of paging. Note that all back-end activity is triggered solely by memory pressure. Memory flows in when the cache needs more objects and flows back out when the rest of the system needs more pages; there are no arbitrary limits or watermarks. Hysteresis control is provided by a working-set algorithm.
The slab allocator is not a monolithic entity, but rather is a loose confederation of independent caches. The caches have no shared state, so the allocator can employ per-cache locking instead of protecting the entire arena (kernel heap) with one global lock. Per-cache locking improves scalability by allowing any number of distinct caches to be accessed simultaneously.
Each cache maintains its own statistics - total allocations, number of allocated and free buffers, etc. These per-cache statistics provide insight into overall system behavior. They indicate which parts of the system consume the most memory and help to identify memory leaks. They also indicate the activity level in various subsystems, to the extent that allocator traffic is an accurate approximation. (Streams message allocation is a direct measure of streams activity, for example.)
The slab allocator is operationally similar to the "CustoMalloc" [Grunwald93A], "QuickFit" [Weinstock88], and "Zone" [VanSciver88] allocators, all of which maintain distinct freelists of the most commonly requested buffer sizes. The Grunwald and Weinstock papers each demonstrate that a customized segregated-storage allocator - one that has a priori knowledge of the most common allocation sizes - is usually optimal in both space and time. The slab allocator is in this category, but has the advantage that its customizations are client-driven at run time rather than being hard-coded at compile time. (This is also true of the Zone allocator.)
The standard non-caching allocation routines, kmem_alloc(9F) and kmem_free(9F), use object caches internally. At startup, the system creates a set of about 30 caches ranging in size from 8 bytes to 9K in roughly 10-20% increments. kmem_alloc() simply performs a kmem_cache_alloc() from the nearest-size cache. Allocations larger than 9K, which are rare, are handled directly by the back-end page supplier.
Slabs
The slab is the primary unit of currency in the slab allocator. When the allocator needs to grow a cache, for example, it acquires an entire slab of objects at once. Similarly, the allocator reclaims unused memory (shrinks a cache) by relinquishing a complete slab.
A slab consists of one or more pages of virtually contiguous memory carved up into equal-size chunks, with a reference count indicating how many of those chunks have been allocated. The benefits of using this simple data structure to manage the arena are somewhat striking:
- Reclaiming unused memory is trivial. When the slab reference count goes to zero the associated pages can be returned to the VM system. Thus a simple reference count replaces the complex trees, bitmaps, and coalescing algorithms found in most other allocators [Knuth68, Korn85, Standish80].
- Allocating and freeing memory are fast, constant-time operations. All we have to do is move an object to or from a freelist and update a reference count.
- Severe external fragmentation (unused buffers on the freelist) is unlikely. Over time, many allocators develop an accumulation of small, unusable buffers. This occurs as the allocator splits existing free buffers to satisfy smaller requests. For example, the right sequence of 32-byte and 40-byte allocations may result in a large accumulation of free 8-byte buffers - even though no 8-byte buffers are ever requested [Standish80]. A segregated-storage allocator cannot suffer this fate, since the only way to populate its 8-byte freelist is to actually allocate and free 8-byte buffers. Any sequence of 32-byte and 40-byte allocations - no matter how complex - can only result in population of the 32-byte and 40-byte freelists. Since prior allocation is a good predictor of future allocation [Weinstock88] these buffers are likely to be used again. The other reason that slabs reduce external fragmentation is that all objects in a slab are of the same type, so they have the same lifetime distribution. (The generic caches that back kmem_alloc() are a notable exception, but they constitute a relatively small fraction of the arena in SunOS 5.4 - all of the major consumers of memory now use kmem_cache_alloc().) The resulting segregation of short-lived and long-lived objects at slab granularity reduces the likelihood of an entire page being held hostage due to a single long-lived allocation [Barrett93, Hanson90].
- Internal fragmentation (per-buffer wasted space) is minimal. Each buffer is exactly the right size (namely, the cache's object size), so the only wasted space is the unused portion at the end of the slab. For example, assuming 4096-byte pages, the slabs in a 400-byte object cache would each contain 10 buffers, with 96 bytes left over. We can view this as equivalent 9.6 bytes of wasted space per 400-byte buffer, or 2.4% internal fragmentation.
In general, if a slab contains n buffers, then the internal fragmentation is at most 1/n; thus the allocator can actually control the amount of internal fragmentation by controlling the slab size. However, larger slabs are more likely to cause external fragmentation, since the probability of being able to reclaim a slab decreases as the number of buffers per slab increases. The SunOS 5.4 implementation limits internal fragmentation to 12.5% (1/8), since this was found to be the empirical sweet-spot in the trade-off between internal and external fragmentation.
- Slab Layout - Logical
The contents of each slab are managed by a kmem_slab data structure that maintains the slab's linkage in the cache, its reference count, and its list of free buffers. In turn, each buffer in the slab is managed by a kmem_bufctl structure that holds the freelist linkage, buffer address, and a back-pointer to the controlling slab. Pictorially, a slab looks like this (bufctl-to-slab back-pointers not shown):
-------- | kmem | | slab | -------- | | V -------- -------- -------- | kmem | -----> | kmem | -----> | kmem | |bufctl| |bufctl| |bufctl| -------- -------- -------- | | | | | | V V V -------------------------------------------------------- | | | | | | buf | buf | buf |unused| | | | | | -------------------------------------------------------- |<---------------- one or more pages ----------------->|
- Slab Layout for Small Objects
For objects smaller than 1/8 of a page, a slab is built by allocating a page, placing the slab data at the end, and dividing the rest into equal-size buffers:
------------------------ -------------------------------------- | | | | | | un- | kmem | | buf | buf | ... | buf | buf | used | slab | | | | | | | | data | ------------------------ -------------------------------------- |<------------------------- one page -------------------------->|
Each buffer serves as its own bufctl while on the freelist. Only the linkage is actually needed, since everything else is computable. These are essential optimizations for small buffers - otherwise we would end up allocating almost as much memory for bufctls as for the buffers themselves.
The freelist linkage resides at the end of the buffer, rather than the beginning, to facilitate debugging. This is driven by the empirical observation that the beginning of a data structure is typically more active than the end. If a buffer is modified after being freed, the problem is easier to diagnose if the heap structure (freelist linkage) is still intact.
The allocator reserves an additional word for constructed objects so that the linkage doesn't overwrite any constructed state.
- Slab Layout for Large Objects
The above scheme is efficient for small objects, but not for large ones. It could fit only one 2K buffer on a 4K page because of the embedded slab data. Moreover, with large (multi-page) slabs we lose the ability to determine the slab data address from the buffer address. Therefore, for large objects the physical layout is identical to the logical layout. The required slab and bufctl data structures come from their own (small-object!) caches. A per-cache self-scaling hash table provides buffer-to-bufctl conversion.
Freelist Management
Each cache maintains a circular, doubly-linked list of all its slabs. The slab list is partially sorted, in that the empty slabs (all buffers allocated) come first, followed by the partial slabs (some buffers allocated, some free), and finally the complete slabs (all buffers free, refcnt == 0). The cache's freelist pointer points to its first non-empty slab. Each slab, in turn, has its own freelist of available buffers. This two-level freelist structure simplifies memory reclaiming. When the allocator reclaims a slab it doesn't have to unlink each buffer from the cache's freelist - it just unlinks the slab.
Reclaiming Memory
When kmem_cache_free() sees that the slab reference count is zero, it does not immediately reclaim the memory. Instead, it just moves the slab to the tail of the freelist where all the complete slabs reside. This ensures that no complete slab will be broken up unless all partial slabs have been depleted.
When the system runs low on memory it asks the allocator to liberate as much memory as it can. The allocator obliges, but retains a 15-second working set of recently-used slabs to prevent thrashing. Measurements indicate that system performance is fairly insensitive to the slab working-set interval. Presumably this is because the two extremes - zero working set (reclaim all complete slabs on demand) and infinite working-set (never reclaim anything) - are both reasonable, albeit suboptimal, policies.
Impact of Buffer Address Distribution on Cache Utilization
The address distribution of mid-size buffers can affect the system's overall cache utilization. In particular, power-of-two allocators - where all buffers are 2n bytes and are 2n-byte aligned - are pessimal. (Such allocators are common because they are easy to implement. For example, 4.4BSD and SVr4 both employ power-of-two methods [McKusick88, Lee89].) Suppose, for example, that every inode (~=300 bytes) is assigned a 512-byte buffer, 512-byte aligned, and that only the first dozen fields of an inode (48 bytes) are frequently referenced. Then the majority of inode-related memory traffic will be at addresses between 0 and 47 modulo 512. Thus the cache lines near 512-byte boundaries will be heavily loaded while the rest lie fallow. In effect only 9% (48/512) of the cache will be usable by inodes. Fully-associative caches would not suffer this problem, but current hardware trends are toward simpler rather than more complex caches.
Of course, there's nothing special about inodes. The kernel contains many other mid-size data structures (e.g. 100-500 bytes) with the same essential qualities: there are many of them, they contain only a few heavily used fields, and those fields are grouped together at or near the beginning of the structure. This artifact of the way data structures evolve has not previously been recognized as an important factor in allocator design.
Contributions
Impact of Buffer Address Distribution on Bus Balance
On a machine that interleaves memory across multiple main buses, the effects described above also have a significant impact on bus utilization. The SPARCcenter 2000, for example, employs 256-byte interleaving across two main buses [Cekleov92]. Continuing the example above, we see that any power-of-two allocator maps the first half of every inode (the hot part) to bus 0 and the second half to bus 1. Thus almost all inode-related cache misses are serviced by bus 0. The situation is exacerbated by an inflated miss rate, since all of the inodes are fighting over a small fraction of the cache.
These effects can be dramatic. On a SPARCcenter 2000 running LADDIS under a SunOS 5.4 development kernel, replacing the old allocator (a power-of-two buddy-system [Lee89]) with the slab allocator reduced bus imbalance from 43% to just 17%. In addition, the primary cache miss rate dropped by 13%.
Slab Coloring
The slab allocator incorporates a simple coloring scheme that distributes buffers evenly throughout the cache, resulting in excellent cache utilization and bus balance. The concept is simple: each time a new slab is created, the buffer addresses start at a slightly different offset (color) from the slab base (which is always page-aligned). For example, for a cache of 200-byte objects with 8-byte alignment, the first slab's buffers would be at addresses 0, 200, 400, ... relative to the slab base. The next slab's buffers would be at offsets 8, 208, 408, ... and so on. The maximum slab color is determined by the amount of unused space in the slab. In this example, assuming 4K pages, we can fit 20 200-byte buffers in a 4096-byte slab. The buffers consume 4000 bytes, the kmem_slab data consumes 32 bytes, and the remaining 64 bytes are available for coloring. Thus the maximum slab color is 64, and the slab color sequence is 0, 8, 16, 24, 32, 40, 48, 56, 64, 0, 8, ...
One particularly nice property of this coloring scheme is that mid-size power-of-two buffers receive the maximum amount of coloring, since they are the worst-fitting. For example, while 128 bytes goes perfectly into 4096, it goes near-pessimally into 4096 - 32, which is what's actually available (because of the embedded slab data).
Arena Management
An allocator's arena management strategy determines its dynamic cache footprint. These strategies fall into three broad categories: sequential-fit methods, buddy methods, and segregated-storage methods [Standish80].
A sequential-fit allocator must typically search several nodes to find a good-fitting buffer. Such methods are, by nature, condemned to a large cache footprint: they have to examine a significant number of nodes that are generally nowhere near each other. This causes not only cache misses, but TLB misses as well. The coalescing stages of buddy-system allocators [Knuth68, Lee89] have similar properties.
A segregated-storage allocator, such as the slab allocator, maintains separate freelists for different buffer sizes. These allocators generally have good cache locality because allocating a buffer is so simple. All the allocator has to do is determine the freelist (by computation, by table lookup, or by having it supplied as an argument) and take a buffer from it. Freeing a buffer is similarly straightforward. There are only a handful of pointers to load, so the cache footprint is small.
The slab allocator has the additional advantage that for small to mid-size buffers, most of the relevant information - the slab data, bufctls, and buffers themselves - resides on a single page. Thus a single TLB entry covers most of the action.
Auditing
In audit mode the allocator records its activity in a circular transaction log. It stores this information in an extended version of the bufctl structure that includes the thread pointer, hi-res timestamp, and stack trace of the transaction. When corruption is detected by any of the other methods, the previous owners of the affected buffer (the likely suspects) can be determined.
Freed-Address Verification
The buffer-to-bufctl hash table employed by large-object caches can be used as a debugging feature: if the hash lookup in kmem_cache_free() fails, then the caller must be attempting to free a bogus address. The allocator can verify all freed addresses by changing the ``large object threshold to zero.
Detecting Use of Freed Memory
When an object is freed, the allocator applies its destructor and fills it with the pattern 0xdeadbeef. The next time that object is allocated, the allocator verifies that it still contains the deadbeef pattern. It then fills the object with 0xbaddcafe and applies its constructor. The deadbeef and baddcafe patterns are chosen to be readily human-recognizable in a debugging session. They represent freed memory and uninitialized data, respectively.
Redzone Checking
Redzone checking detects writes past the end of a buffer. The allocator checks for redzone violations by adding a guard word to the end of each buffer and verifying that it is unmodified when the buffer is freed.
Synchronous Unmapping
Normally, the slab working-set algorithm retains complete slabs for a while. In synchronous-unmapping mode the allocator destroys complete slabs immediately. kmem_slab_destroy() returns the underlying memory to the back-end page supplier, which unmaps the page(s). Any subsequent reference to any object in that slab will cause a kernel data fault.
Page-per-buffer Mode
In page-per-buffer mode each buffer is given an entire page (or pages) so that every buffer can be unmapped when it is freed. The slab allocator implements this by increasing the alignment for all caches to the system page size. (This feature requires an obscene amount of physical memory.)
Leak Detection
The timestamps provided by auditing make it easy to implement a crude kernel memory leak detector at user level. All the user-level program has to do is periodically scan the arena (via /dev/kmem), looking for the appearance of new, persistent allocations. For example, any buffer that was allocated an hour ago and is still allocated now is a possible leak.
Performance
- Speed Comparison: On a SPARCstation-2 the time required to allocate and free a buffer under the various allocators. Any good segregated-storage implementation should achieve excellent performance. The SunOS 4.1.3 allocator, like most sequential-fit methods, is comparatively slow and quite variable. The benefits of object caching on some of the most frequent allocations.
- Memory Utilization Comparison: An allocator generally consumes more memory than its clients actually request due to imperfect fits (internal fragmentation), unused buffers on the freelist (external fragmentation), and the overhead of the allocator's internal data structures. The ratio of memory requested to memory consumed is the allocator's memory utilization. The complementary ratio is the memory wastage or total fragmentation. Good memory utilization is essential, since the kernel heap consumes physical memory. An allocator's space efficiency is harder to characterize than its speed because it is workload-dependent.
- System boot. This measures the system's memory utilization at the console login prompt after rebooting.
- A brief spike in load, generated by the following trivial program:
fork(); fork(); fork(); fork(); fork(); fork(); fork(); fork(); fd = socket(AF_UNIX, SOCK_STREAM, 0); sleep(60); close(fd);
This creates 256 processes, each of which creates a socket. This causes a temporary surge in demand for a variety of kernel data structures.
- Find. This is another trivial spike-generator:
find /usr -mount -exec file {} \;
- Kenbus. This is a standard timesharing benchmark. Kenbus generates a large amount of concurrent activity, creating large demand for both user and kernel memory.
- Overall System Performance
Future Work
Managing Other Types of Memory
The slab allocator gets its pages from segkmem via the routines kmem_getpages() and kmem_freepages(); it assumes nothing about the underlying segment driver, resource maps, translation setup, etc. Since the allocator respects this firewall, it would be trivial to plug in alternate back-end page suppliers. The "getpages" and "freepages" routines could be supplied as additional arguments to kmem_cache_create(). This would allow us to manage multiple types of memory (e.g. normal kernel memory, device memory, pageable kernel memory, NVRAM, etc.) with a single allocator.
Per-Processor Memory Allocation
The per-processor allocation techniques of McKenney and Slingwine [McKenney93] would fit nicely on top of the slab allocator. They define a four-layer allocation hierarchy of decreasing speed and locality: per-CPU, global, coalesce-to-page, and coalesce-to-VM-block. The latter three correspond closely to the slab allocator's front-end, back-end, and page-supplier layers, respectively. Even in the absence of lock contention, small per-processor freelists could improve performance by eliminating locking costs and reducing invalidation traffic.
User-level Applications
The slab allocator could also be used as a user-level memory allocator. The back-end page supplier could be mmap(2) or sbrk(2).
Conclusions
The slab allocator is a simple, fast, and space-efficient kernel memory allocator. The object-cache interface upon which it is based reduces the cost of allocating and freeing complex objects and enables the allocator to segregate objects by size and lifetime distribution. Slabs take advantage of object size and lifetime segregation to reduce internal and external fragmentation, respectively. Slabs also simplify reclaiming by using a simple reference count instead of coalescing. The slab allocator establishes a push/pull relationship between its clients and the VM system, eliminating the need for arbitrary limits or watermarks to govern reclaiming. The allocator's coloring scheme distributes buffers evenly throughout the cache, improving the system's overall cache utilization and bus balance. In several important areas, the slab allocator provides measurably better system performance.
