Kernel Memory Allocation 1

1
Chapter 12. Kernel Memory Allocation

• Introduction
• Resource Map Allocator
• Simple Power-of-Two Free Lists
• Mckusick-Karels Allocator
• Buddy System
• Mach-OSF/1 Zone Allocator
• Solaris 2.4 Slab Allocator

2
Introduction

• Page-level allocator
• Paging system
• Supports virtual memory system
• Kernel memory allocator
• Provides odd-sized buffers of memory to various
  kernel subsystems
• Kernel frequently needs chunks of memory of
  various sizes usually for short periods of time

3
Introduction (cont)

• Users of the kernel memory allocator
• pathname translation routine
• allocb( ) allocates STREAMs buffers of arbitrary
  size
• In SVR4, the kernel allocates many objects (proc
  structures, vnodes, file descriptor blocks)
  dynamically when needed

4
Introduction (cont)
physical memory
Page-level allocator
Paging system
kernel memory allocator
network buffers
proc structures
inodes, file descriptors
user processes
block buffer cache
5
Kernel Memory Allocator (KMA)

• Functional requirements
• Page-level allocator pre-allocates part of main
  memory to KMA, which must use this memory pool
  efficiently
• KMA also have to monitor which part of its pool
  are allocated and which are free
• Evaluation criteria
• Utilization factor
• KMA must be fast
• Both average and worst-cast latency are important
• Interaction with the paging system

6
Resource Map Allocator

• Resource map
• set of for free memory
• First fit, Best fit, Worst bit policies
• Advantages
• Easy to implement
• Not restricted to memory allocation
• Allocates the exact numbers of bytes requested
• Client is not constrained to release the exact
  region it has allocated
• Allocator coalesces adjacent free regions

7
Resource Map Allocator (cont)

• Drawbacks
• Map may become highly fragmented
• As the fragmentation increases, so does the size
  of the map
• Sorting for coalescing adjacent regions is
  expensive
• Linear search to find a free region

8
Simple Power-of-Two Free List
list headers
free buffers
allocated blocks
32
64
128
256
512
1024
9
Simple Power-of-Two Free List (cont)

• Used frequently to implement malloc( ), and free(
  ) in the user-level C library
• One-word header for each buffer
• Advantages
• Avoids the lengthy linear search of the resource
  map method
• Eliminates the fragmentation problem

10
Simple Power-of-Two Free List (cont)

• Drawbacks
• Rounding of requests to the next power of two
  results in poor memory utilization
• No provision for coalescing adjacent free buffer
  to satisfy larger requests
• No provision to return surplus free buffers to
  the page-level allocator

11
Mckusick-Karels Allocator
freelistaddr
free buffers
32
allocated blocks
64
128
256
512
32
512
64
F
32
128
F
32
32
256
64
F
2K
F
128
kmemsizes
freepages
12
Mckusick-Karels Allocator (cont)

• Used in 4.4BSD, Digital UNIX
• Requires that
• a set of contiguous pages
• all buffers belonging to the same page to be the
  same size
• States of each page
• Free
• corresponding element of kmemsizes contains a
  pointer to the element for the next free page
• Divided into buffers of a particular size
• kmemsizes element contains the size

13
Mckusick-Karels Allocator (cont)

• Part of a buffer that spanned multiple pages
• kmemsizes element corresponding to the first
  page of the buffer contains the buffer size
• Improvement over simple power-of-two
• Faster, wastes less memory, can handle large and
  small request efficiently
• Drawbacks
• No provision for moving memory from one list to
  another
• No way to return memory to the paging system

14
Buddy System

• Basic idea
• Combines free buffer coalescing with a
  power-of-two allocator
• When a buffer is split, each buffer is called the
  buddy of the other
• Advantages
• Coalescing adjacent free buffers
• Easy exchange of memory between the allocator and
  the paging system
• Disadvantages
• Performance degradation due to coalescing
• Program interface needs the size of the buffer

15
Buddy System (e.g.)
allocate(256) ? allocate(128) ? allocate(64)
bitmap
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
32
64
128
256
512
free list headers
0
256
384
448
512
1023
B
C
D
D
A
C
B
free
in use
A
16
Buddy System (e.g.)
? allocate(128) ? release(C, 128)
bitmap
1
1
1
1
1
1
1
1
0
0
0
0
1
1
0
0
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
32
64
128
256
512
free list headers
0
256
384
448
512
640
768
1023
B
C
D
D
F
F
E
C
B
E
A
A
17
Buddy System (e.g.)
? release(D, 64)
bitmap
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
32
64
128
256
512
free list headers
0
256
512
640
768
1023
B
B
F
F
E
C
E
A
18
SVR4 Lazy Buddy System

• Basic idea
• Defer coalescing until it becomes necessary, and
  then to coalesce as many buffers as possible
• Lazy coalescing
• N buffers, A buffers are active, L are locally
  free, G are globally free
• N A L G
• slack N - 2L - G A - L

19
SVR4 Lazy Buddy System (cont)

• Lazy state (slack 0)
• buffer consumption is in a steady state and
  coalescing is not necessary
• Reclaiming state (slack 1)
• consumption is borderline, coalescing is needed
• Accelerated state (slack 2 or more)
• consumption is not in a steady state, and the
  allocator must coalesce faster

20
Mach-OSF/1 Zone Allocator
active zones
free objects
free_elements
zone of zones
structzone
structzone
structzone
next_zone
first_zone
zone of zones
structobj1
structobj1
structobj1
last_zone
...
zone of zones
structobjn
struct objn
struct objn
21
Mach-OSF/1 Zone Allocator (cont)

• Basic idea
• Each class of dynamically allocated objects is
  assigned to its own size, which is simply a pool
  of free objects of that class
• Any single page is only used for one zone
• Free objects of each zone are maintained on a
  linked list, headed by a struct zone
• Allocation and release involve nothing more than
  removing objects from and returning objects to
  the free list
• If an allocation request finds the free list
  empty, it asks the page-level allocator for
  alloc( ) more bytes

22
Mach-OSF/1 Zone Allocator (cont)

• Garbage collection
• Free pages can be returned to the paging system
  and later recovered for other zones
• Analysis
• Zone allocator is fast and efficient
• No provision for releasing only part of the
  allocated object
• Zone objects are exactly the required size
• Garbage collector provides a mechanism for memory
  reuse

23
Hierarchical Allocator for Multiprocessors
per-pages freelists
Coalesce-to-Page layer
Global layer
global freelist
bucket list
target 3
main freelist
Per-CPU cache
aux freelist
target 3
24
Solaris 2.4 Slab Allocator

• Better performance and memory utilization than
  other implementation
• Main issues
• Object reuse
• advantage of caching and reusing the same object,
  rather than allocating and initializing arbitrary
  chunks of memory
• Hardware cache utilization
• Most kernel objects have their important,
  frequently accessed fields at the beginning of
  the object

25
Solaris 2.4 Slab Allocator

• Allocator footprint
• Footprint of a allocator is the portion of the
  hardware cache and the translation lookaside
  buffer (TLB) that is overwritten by the allocator
  itself
• Large footprint causes many cache and TLB misses,
  reducing the performance of the allocator
• Large footprint resource maps, buddy systems
• Smaller footprint Mckusick-Karels, zone
• Design
• slab allocator is a variant of the zone method
  and is organized as a collection of object caches

26
Solaris 2.4 Slab Allocator (cont)
page-level allocator
back end
vnode cache
proc cache
mbuf cache
msgb cache
...
front end
activevnodes
activeprocs
activembufs
activemsgbs
27
Solaris 2.4 Slab Allocator (cont)

• Implementation

slab data
coloring area (unused)
unused space
linked list of slabs in same cache
free
active
free
active
active
free
active
freelist
NULL
pointers
extra space for linkage
28
Solaris 2.4 Slab Allocator (cont)

• Analysis
• Space efficient
• Coalescing scheme results in better hardware
  cache and memory bus utilization
• Small footprint (only one slab for most requests)