<?xml version="1.0" encoding="UTF-8"?>
<chapter id="mm">
<?dbhtml filename="mm.html"?>
<title>Memory management</title>
<para>In previous chapters, this book described the scheduling subsystem as
the creator of the impression that threads execute in parallel. The memory
management subsystem, on the other hand, creates the impression that there
is enough physical memory for the kernel and that userspace tasks have the
entire address space only for themselves.</para>
<section>
<title>Physical memory management</title>
<section id="zones_and_frames">
<title>Zones and frames</title>
<para>HelenOS represents continuous areas of physical memory in
structures called frame zones (abbreviated as zones). Each zone contains
information about the number of allocated and unallocated physical
memory frames as well as the physical base address of the zone and
number of frames contained in it. A zone also contains an array of frame
structures describing each frame of the zone and, in the last, but not
the least important, front, each zone is equipped with a buddy system
that faciliates effective allocation of power-of-two sized block of
frames.</para>
<para>This organization of physical memory provides good preconditions
for hot-plugging of more zones. There is also one currently unused zone
attribute: <code>flags</code>. The attribute could be used to give a
special meaning to some zones in the future.</para>
<para>The zones are linked in a doubly-linked list. This might seem a
bit ineffective because the zone list is walked everytime a frame is
allocated or deallocated. However, this does not represent a significant
performance problem as it is expected that the number of zones will be
rather low. Moreover, most architectures merge all zones into
one.</para>
<para>For each physical memory frame found in a zone, there is a frame
structure that contains number of references and data used by buddy
system.</para>
</section>
<section id="frame_allocator">
<title>Frame allocator</title>
<para>The frame allocator satisfies kernel requests to allocate
power-of-two sized blocks of physical memory. Because of zonal
organization of physical memory, the frame allocator is always working
within a context of some frame zone. In order to carry out the
allocation requests, the frame allocator is tightly integrated with the
buddy system belonging to the zone. The frame allocator is also
responsible for updating information about the number of free and busy
frames in the zone. <figure>
<mediaobject id="frame_alloc">
<imageobject role="html">
<imagedata fileref="images/frame_alloc.png" format="PNG" />
</imageobject>
<imageobject role="fop">
<imagedata fileref="images.vector/frame_alloc.svg" format="SVG" />
</imageobject>
</mediaobject>
<title>Frame allocator scheme.</title>
</figure></para>
<formalpara>
<title>Allocation / deallocation</title>
<para>Upon allocation request via function <code>frame_alloc</code>,
the frame allocator first tries to find a zone that can satisfy the
request (i.e. has the required amount of free frames). Once a suitable
zone is found, the frame allocator uses the buddy allocator on the
zone's buddy system to perform the allocation. During deallocation,
which is triggered by a call to <code>frame_free</code>, the frame
allocator looks up the respective zone that contains the frame being
deallocated. Afterwards, it calls the buddy allocator again, this time
to take care of deallocation within the zone's buddy system.</para>
</formalpara>
</section>
<section id="buddy_allocator">
<title>Buddy allocator</title>
<para>In the buddy system, the memory is broken down into power-of-two
sized naturally aligned blocks. These blocks are organized in an array
of lists, in which the list with index i contains all unallocated blocks
of size <mathphrase>2<superscript>i</superscript></mathphrase>. The
index i is called the order of block. Should there be two adjacent
equally sized blocks in the list i<mathphrase />(i.e. buddies), the
buddy allocator would coalesce them and put the resulting block in list
<mathphrase>i + 1</mathphrase>, provided that the resulting block would
be naturally aligned. Similarily, when the allocator is asked to
allocate a block of size
<mathphrase>2<superscript>i</superscript></mathphrase>, it first tries
to satisfy the request from the list with index i. If the request cannot
be satisfied (i.e. the list i is empty), the buddy allocator will try to
allocate and split a larger block from the list with index i + 1. Both
of these algorithms are recursive. The recursion ends either when there
are no blocks to coalesce in the former case or when there are no blocks
that can be split in the latter case.</para>
<para>This approach greatly reduces external fragmentation of memory and
helps in allocating bigger continuous blocks of memory aligned to their
size. On the other hand, the buddy allocator suffers increased internal
fragmentation of memory and is not suitable for general kernel
allocations. This purpose is better addressed by the <link
linkend="slab">slab allocator</link>.<figure>
<mediaobject id="buddy_alloc">
<imageobject role="html">
<imagedata fileref="images/buddy_alloc.png" format="PNG" />
</imageobject>
<imageobject role="fop">
<imagedata fileref="images.vector/buddy_alloc.svg" format="SVG" />
</imageobject>
</mediaobject>
<title>Buddy system scheme.</title>
</figure></para>
<section>
<title>Implementation</title>
<para>The buddy allocator is, in fact, an abstract framework wich can
be easily specialized to serve one particular task. It knows nothing
about the nature of memory it helps to allocate. In order to beat the
lack of this knowledge, the buddy allocator exports an interface that
each of its clients is required to implement. When supplied with an
implementation of this interface, the buddy allocator can use
specialized external functions to find a buddy for a block, split and
coalesce blocks, manipulate block order and mark blocks busy or
available.</para>
<formalpara>
<title>Data organization</title>
<para>Each entity allocable by the buddy allocator is required to
contain space for storing block order number and a link variable
used to interconnect blocks within the same order.</para>
<para>Whatever entities are allocated by the buddy allocator, the
first entity within a block is used to represent the entire block.
The first entity keeps the order of the whole block. Other entities
within the block are assigned the magic value
<constant>BUDDY_INNER_BLOCK</constant>. This is especially important
for effective identification of buddies in a one-dimensional array
because the entity that represents a potential buddy cannot be
associated with <constant>BUDDY_INNER_BLOCK</constant> (i.e. if it
is associated with <constant>BUDDY_INNER_BLOCK</constant> then it is
not a buddy).</para>
</formalpara>
</section>
</section>
<section id="slab">
<title>Slab allocator</title>
<para>The majority of memory allocation requests in the kernel is for
small, frequently used data structures. The basic idea behind the slab
allocator is that commonly used objects are preallocated in continuous
areas of physical memory called slabs<footnote>
<para>Slabs are in fact blocks of physical memory frames allocated
from the frame allocator.</para>
</footnote>. Whenever an object is to be allocated, the slab allocator
returns the first available item from a suitable slab corresponding to
the object type<footnote>
<para>The mechanism is rather more complicated, see the next
paragraph.</para>
</footnote>. Due to the fact that the sizes of the requested and
allocated object match, the slab allocator significantly reduces
internal fragmentation.</para>
<para>Slabs of one object type are organized in a structure called slab
cache. There are ususally more slabs in the slab cache, depending on
previous allocations. If the the slab cache runs out of available slabs,
new slabs are allocated. In order to exploit parallelism and to avoid
locking of shared spinlocks, slab caches can have variants of
processor-private slabs called magazines. On each processor, there is a
two-magazine cache. Full magazines that are not part of any
per-processor magazine cache are stored in a global list of full
magazines.</para>
<para>Each object begins its life in a slab. When it is allocated from
there, the slab allocator calls a constructor that is registered in the
respective slab cache. The constructor initializes and brings the object
into a known state. The object is then used by the user. When the user
later frees the object, the slab allocator puts it into a processor
private magazine cache, from where it can be precedently allocated
again. Note that allocations satisfied from a magazine are already
initialized by the constructor. When both of the processor cached
magazines get full, the allocator will move one of the magazines to the
list of full magazines. Similarily, when allocating from an empty
processor magazine cache, the kernel will reload only one magazine from
the list of full magazines. In other words, the slab allocator tries to
keep the processor magazine cache only half-full in order to prevent
thrashing when allocations and deallocations interleave on magazine
boundaries.</para>
<para>Should HelenOS run short of memory, it would start deallocating
objects from magazines, calling slab cache destructor on them and
putting them back into slabs. When a slab contanins no allocated object,
it is immediately freed.</para>
<para><figure>
<mediaobject id="slab_alloc">
<imageobject role="html">
<imagedata fileref="images/slab_alloc.png" format="PNG" />
</imageobject>
</mediaobject>
<title>Slab allocator scheme.</title>
</figure></para>
<section>
<title>Implementation</title>
<para>The slab allocator is closely modelled after OpenSolaris slab
allocator by Jeff Bonwick and Jonathan Adams with the following
exceptions:<itemizedlist>
<listitem>
empty slabs are immediately deallocated
</listitem>
<listitem>
<para>empty magazines are deallocated when not needed</para>
</listitem>
</itemizedlist> Following features are not currently supported but
would be easy to do: <itemizedlist>
<listitem>
cache coloring
</listitem>
<listitem>
dynamic magazine grow (different magazine sizes are already supported, but the allocation strategy would need to be adjusted)
</listitem>
</itemizedlist></para>
<section>
<title>Magazine layer</title>
<para>Due to the extensive bottleneck on SMP architures, caused by
global slab locking mechanism, making processing of all slab
allocation requests serialized, a new layer was introduced to the
classic slab allocator design. Slab allocator was extended to
support per-CPU caches 'magazines' to achieve good SMP scaling.
<termdef>Slab SMP perfromance bottleneck was resolved by introducing
a per-CPU caching scheme called as <glossterm>magazine
layer</glossterm></termdef>.</para>
<para>Magazine is a N-element cache of objects, so each magazine can
satisfy N allocations. Magazine behaves like a automatic weapon
magazine (LIFO, stack), so the allocation/deallocation become simple
push/pop pointer operation. Trick is that CPU does not access global
slab allocator data during the allocation from its magazine, thus
making possible parallel allocations between CPUs.</para>
<para>Implementation also requires adding another feature as the
CPU-bound magazine is actually a pair of magazines to avoid
thrashing when during allocation/deallocatiion of 1 item at the
magazine size boundary. LIFO order is enforced, which should avoid
fragmentation as much as possible.</para>
<para>Another important entity of magazine layer is the common full
magazine list (also called a depot), that stores full magazines that
may be used by any of the CPU magazine caches to reload active CPU
magazine. This list of magazines can be pre-filled with full
magazines during initialization, but in current implementation it is
filled during object deallocation, when CPU magazine becomes
full.</para>
<para>Slab allocator control structures are allocated from special
slabs, that are marked by special flag, indicating that it should
not be used for slab magazine layer. This is done to avoid possible
infinite recursions and deadlock during conventional slab allocaiton
requests.</para>
</section>
<section>
<title>Allocation/deallocation</title>
<para>Every cache contains list of full slabs and list of partialy
full slabs. Empty slabs are immediately freed (thrashing will be
avoided because of magazines).</para>
<para>The SLAB allocator allocates lots of space and does not free
it. When frame allocator fails to allocate the frame, it calls
slab_reclaim(). It tries 'light reclaim' first, then brutal reclaim.
The light reclaim releases slabs from cpu-shared magazine-list,
until at least 1 slab is deallocated in each cache (this algorithm
should probably change). The brutal reclaim removes all cached
objects, even from CPU-bound magazines.</para>
<formalpara>
<title>Allocation</title>
<para><emphasis>Step 1.</emphasis> When it comes to the allocation
request, slab allocator first of all checks availability of memory
in local CPU-bound magazine. If it is there, we would just "pop"
the CPU magazine and return the pointer to object.</para>
<para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is
empty, allocator will attempt to reload magazin, swapping it with
second CPU magazine and returns to the first step.</para>
<para><emphasis>Step 3.</emphasis> Now we are in the situation
when both CPU-bound magazines are empty, which makes allocator to
access shared full-magazines depot to reload CPU-bound magazines.
If reload is succesful (meaning there are full magazines in depot)
algoritm continues at Step 1.</para>
<para><emphasis>Step 4.</emphasis> Final step of the allocation.
In this step object is allocated from the conventional slab layer
and pointer is returned.</para>
</formalpara>
<formalpara>
<title>Deallocation</title>
<para><emphasis>Step 1.</emphasis> During deallocation request,
slab allocator will check if the local CPU-bound magazine is not
full. In this case we will just push the pointer to this
magazine.</para>
<para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is
full, allocator will attempt to reload magazin, swapping it with
second CPU magazine and returns to the first step.</para>
<para><emphasis>Step 3.</emphasis> Now we are in the situation
when both CPU-bound magazines are full, which makes allocator to
access shared full-magazines depot to put one of the magazines to
the depot and creating new empty magazine. Algoritm continues at
Step 1.</para>
</formalpara>
</section>
</section>
</section>
<!-- End of Physmem -->
</section>
<section>
<title>Virtual memory management</title>
<section>
<title>Introduction</title>
<para>Virtual memory is a special memory management technique, used by
kernel to achieve a bunch of mission critical goals. <itemizedlist>
<listitem>
Isolate each task from other tasks that are running on the system at the same time.
</listitem>
<listitem>
Allow to allocate more memory, than is actual physical memory size of the machine.
</listitem>
<listitem>
Allowing, in general, to load and execute two programs that are linked on the same address without complicated relocations.
</listitem>
</itemizedlist></para>
<para><!--
<para>
Address spaces. Address space area (B+ tree). Only for uspace. Set of syscalls (shrink/extend etc).
Special address space area type - device - prohibits shrink/extend syscalls to call on it.
Address space has link to mapping tables (hierarchical - per Address space, hash - global tables).
</para>
--></para>
</section>
<section>
<title>Address spaces</title>
<section>
<title>Address space areas</title>
<para>Each address space consists of mutually disjunctive continuous
address space areas. Address space area is precisely defined by its
base address and the number of frames/pages is contains.</para>
<para>Address space area , that define behaviour and permissions on
the particular area. <itemizedlist>
<listitem>
<emphasis>AS_AREA_READ</emphasis>
flag indicates reading permission.
</listitem>
<listitem>
<emphasis>AS_AREA_WRITE</emphasis>
flag indicates writing permission.
</listitem>
<listitem>
<emphasis>AS_AREA_EXEC</emphasis>
flag indicates code execution permission. Some architectures do not support execution persmission restriction. In this case this flag has no effect.
</listitem>
<listitem>
<emphasis>AS_AREA_DEVICE</emphasis>
marks area as mapped to the device memory.
</listitem>
</itemizedlist></para>
<para>Kernel provides possibility tasks create/expand/shrink/share its
address space via the set of syscalls.</para>
</section>
<section>
<title>Address Space ID (ASID)</title>
<para>When switching to the different task, kernel also require to
switch mappings to the different address space. In case TLB cannot
distinguish address space mappings, all mapping information in TLB
from the old address space must be flushed, which can create certain
uncessary overhead during the task switching. To avoid this, some
architectures have capability to segregate different address spaces on
hardware level introducing the address space identifier as a part of
TLB record, telling the virtual address space translation unit to
which address space this record is applicable.</para>
<para>HelenOS kernel can take advantage of this hardware supported
identifier by having an ASID abstraction which is somehow related to
the corresponding architecture identifier. I.e. on ia64 kernel ASID is
derived from RID (region identifier) and on the mips32 kernel ASID is
actually the hardware identifier. As expected, this ASID information
record is the part of <emphasis>as_t</emphasis> structure.</para>
<para>Due to the hardware limitations, hardware ASID has limited
length from 8 bits on ia64 to 24 bits on mips32, which makes it
impossible to use it as unique address space identifier for all tasks
running in the system. In such situations special ASID stealing
algoritm is used, which takes ASID from inactive task and assigns it
to the active task.</para>
<para><classname>ASID stealing algoritm here.</classname></para>
</section>
</section>
<section>
<title>Virtual address translation</title>
<section id="paging">
<title>Paging</title>
<section>
<title>Introduction</title>
<para>Virtual memory is usually using paged memory model, where
virtual memory address space is divided into the
<emphasis>pages</emphasis> (usually having size 4096 bytes) and
physical memory is divided into the frames (same sized as a page, of
course). Each page may be mapped to some frame and then, upon memory
access to the virtual address, CPU performs <emphasis>address
translation</emphasis> during the instruction execution.
Non-existing mapping generates page fault exception, calling kernel
exception handler, thus allowing kernel to manipulate rules of
memory access. Information for pages mapping is stored by kernel in
the <link linkend="page_tables">page tables</link></para>
<para>The majority of the architectures use multi-level page tables,
which means need to access physical memory several times before
getting physical address. This fact would make serios performance
overhead in virtual memory management. To avoid this <link
linkend="tlb">Traslation Lookaside Buffer (TLB)</link> is
used.</para>
<para>HelenOS kernel has two different approaches to the paging
implementation: <emphasis>4 level page tables</emphasis> and
<emphasis>global hash table</emphasis>, which are accessible via
generic paging abstraction layer. Such different functionality was
caused by the major architectural differences between supported
platforms. This abstraction is implemented with help of the global
structure of pointers to basic mapping functions
<emphasis>page_mapping_operations</emphasis>. To achieve different
functionality of page tables, corresponding layer must implement
functions, declared in
<emphasis>page_mapping_operations</emphasis></para>
<para>Thanks to the abstract paging interface, there was a place
left for more paging implementations (besides already implemented
hieararchical page tables and hash table), for example B-Tree based
page tables.</para>
</section>
<section>
<title>Hierarchical 4-level page tables</title>
<para>Hierarchical 4-level page tables are the generalization of the
hardware capabilities of most architectures. Each address space has
its own page tables.<itemizedlist>
<listitem>
ia32 uses 2-level page tables, with full hardware support.
</listitem>
<listitem>
amd64 uses 4-level page tables, also coming with full hardware support.
</listitem>
<listitem>
mips and ppc32 have 2-level tables, software simulated support.
</listitem>
</itemizedlist></para>
</section>
<section>
<title>Global hash table</title>
<para>Implementation of the global hash table was encouraged by the
ia64 architecture support. One of the major differences between
global hash table and hierarchical tables is that global hash table
exists only once in the system and the hierarchical tables are
maintained per address space.</para>
<para>Thus, hash table contains information about all address spaces
mappings in the system, so, the hash of an entry must contain
information of both address space pointer or id and the virtual
address of the page. Generic hash table implementation assumes that
the addresses of the pointers to the address spaces are likely to be
on the close addresses, so it uses least significant bits for hash;
also it assumes that the virtual page addresses have roughly the
same probability of occurring, so the least significant bits of VPN
compose the hash index.</para>
<para>Collision chains ...</para>
</section>
</section>
<section id="tlb">
<title>Translation Lookaside buffer</title>
<para>Due to the extensive overhead during the page mapping lookup in
the page tables, all architectures has fast assotiative cache memory
built-in CPU. This memory called TLB stores recently used page table
entries.</para>
<section id="tlb_shootdown">
<title>TLB consistency. TLB shootdown algorithm.</title>
<para>Operating system is responsible for keeping TLB consistent by
invalidating the contents of TLB, whenever there is some change in
page tables. Those changes may occur when page or group of pages
were unmapped, mapping is changed or system switching active address
space to schedule a new system task. Moreover, this invalidation
operation must be done an all system CPUs because each CPU has its
own independent TLB cache. Thus maintaining TLB consistency on SMP
configuration as not as trivial task as it looks on the first
glance. Naive solution would assume that is the CPU which wants to
invalidate TLB will invalidate TLB caches on other CPUs. It is not
possible on the most of the architectures, because of the simple
fact - flushing TLB is allowed only on the local CPU and there is no
possibility to access other CPUs' TLB caches, thus invalidate TLB
remotely.</para>
<para>Technique of remote invalidation of TLB entries is called "TLB
shootdown". HelenOS uses a variation of the algorithm described by
D. Black et al., "Translation Lookaside Buffer Consistency: A
Software Approach," Proc. Third Int'l Conf. Architectural Support
for Programming Languages and Operating Systems, 1989, pp.
113-122.</para>
<para>As the situation demands, you will want partitial invalidation
of TLB caches. In case of simple memory mapping change it is
necessary to invalidate only one or more adjacent pages. In case if
the architecture is aware of ASIDs, when kernel needs to dump some
ASID to use by another task, it invalidates only entries from this
particular address space. Final option of the TLB invalidation is
the complete TLB cache invalidation, which is the operation that
flushes all entries in TLB.</para>
<para>TLB shootdown is performed in two phases.</para>
<formalpara>
<title>Phase 1.</title>
<para>First, initiator locks a global TLB spinlock, then request
is being put to the local request cache of every other CPU in the
system protected by its spinlock. In case the cache is full, all
requests in the cache are replaced by one request, indicating
global TLB flush. Then the initiator thread sends an IPI message
indicating the TLB shootdown request to the rest of the CPUs and
waits actively until all CPUs confirm TLB invalidating action
execution by setting up a special flag. After setting this flag
this thread is blocked on the TLB spinlock, held by the
initiator.</para>
</formalpara>
<formalpara>
<title>Phase 2.</title>
<para>All CPUs are waiting on the TLB spinlock to execute TLB
invalidation action and have indicated their intention to the
initiator. Initiator continues, cleaning up its TLB and releasing
the global TLB spinlock. After this all other CPUs gain and
immidiately release TLB spinlock and perform TLB invalidation
actions.</para>
</formalpara>
</section>
</section>
</section>
</section>
</chapter>