Subversion Repositories HelenOS-doc

Rev

Rev 39 | Rev 46 | Go to most recent revision | Blame | Compare with Previous | Last modification | View Log | Download | RSS feed

  1. <?xml version="1.0" encoding="UTF-8"?>
  2. <chapter id="mm">
  3.   <?dbhtml filename="mm.html"?>
  4.  
  5.   <title>Memory management</title>
  6.  
  7.   <section>
  8.     <title>Virtual memory management</title>
  9.  
  10.     <section>
  11.       <title>Introduction</title>
  12.  
  13.       <para>Virtual memory is a special memory management technique, used by
  14.       kernel to achieve a bunch of mission critical goals. <itemizedlist>
  15.           <listitem>
  16.              Isolate each task from other tasks that are running on the system at the same time.
  17.           </listitem>
  18.  
  19.           <listitem>
  20.              Allow to allocate more memory, than is actual physical memory size of the machine.
  21.           </listitem>
  22.  
  23.           <listitem>
  24.              Allowing, in general, to load and execute two programs that are linked on the same address without complicated relocations.
  25.           </listitem>
  26.         </itemizedlist></para>
  27.  
  28.       <para><!--
  29.  
  30.                 TLB shootdown ASID/ASID:PAGE/ALL.
  31.                 TLB shootdown requests can come in asynchroniously
  32.                 so there is a cache of TLB shootdown requests. Upon cache overflow TLB shootdown ALL is executed
  33.  
  34.  
  35.                 <para>
  36.                         Address spaces. Address space area (B+ tree). Only for uspace. Set of syscalls (shrink/extend etc).
  37.                         Special address space area type - device - prohibits shrink/extend syscalls to call on it.
  38.                         Address space has link to mapping tables (hierarchical - per Address space, hash - global tables).
  39.                 </para>
  40.  
  41. --></para>
  42.     </section>
  43.  
  44.     <section>
  45.       <title>Paging</title>
  46.  
  47.       <para>Virtual memory is usually using paged memory model, where virtual
  48.       memory address space is divided into the <emphasis>pages</emphasis>
  49.       (usually having size 4096 bytes) and physical memory is divided into the
  50.       frames (same sized as a page, of course). Each page may be mapped to
  51.       some frame and then, upon memory access to the virtual address, CPU
  52.       performs <emphasis>address translation</emphasis> during the instruction
  53.       execution. Non-existing mapping generates page fault exception, calling
  54.       kernel exception handler, thus allowing kernel to manipulate rules of
  55.       memory access. Information for pages mapping is stored by kernel in the
  56.       <link linkend="page_tables">page tables</link></para>
  57.  
  58.       <para>The majority of the architectures use multi-level page tables,
  59.       which means need to access physical memory several times before getting
  60.       physical address. This fact would make serios performance overhead in
  61.       virtual memory management. To avoid this <link linkend="tlb">Traslation
  62.       Lookaside Buffer (TLB)</link> is used.</para>
  63.  
  64.       <para>At the moment HelenOS does not support swapping.</para>
  65.  
  66.       <para>- pouzivame vypadky stranky k alokaci ramcu on-demand v ramci
  67.       as_area - na architekturach, ktere to podporuji, podporujeme non-exec
  68.       stranky</para>
  69.     </section>
  70.  
  71.     <section>
  72.       <title>Address spaces</title>
  73.  
  74.       <section>
  75.         <title>Address spaces and areas</title>
  76.  
  77.         <para>- adresovy prostor se sklada z tzv. address space areas
  78.         usporadanych v B+stromu; tyto areas popisuji vyuzivane casti
  79.         adresoveho prostoru patrici do user address space. Kazda cast je dana
  80.         svoji bazovou adresou, velikosti a flagy (rwx/dd).</para>
  81.  
  82.         <para>- uzivatelske thready maji moznost manipulovat se svym adresovym
  83.         prostorem (vytvaret/resizovat/sdilet) as_areas pomoci syscallu</para>
  84.       </section>
  85.  
  86.       <section>
  87.         <title>Address Space ID (ASID)</title>
  88.  
  89.         <para>- nektery hardware umoznuje rozlisit ruzne adresove prostory od
  90.         sebe (cilem je maximalizovat vyuziti TLB); dela to tak, ze s kazdou
  91.         polozkou TLB/strankovacich tabulek sdruzi identifikator adresoveho
  92.         prostoru (ASID, RID, ppc32 ???). Tyto id mivaji ruznou sirku: 8-bitu
  93.         az 24-bitu (kolik ma ppc32?)</para>
  94.  
  95.         <para>- kernel tomu rozumi a sam pouziva abstrakci ASIDu (na ia64 to
  96.         je napr. cislo odvozene od RIDu, na mips32 to je ASID samotny);
  97.         existence ASIDu je nutnou podminkou pouziti _global_ page hash table
  98.         mechanismu.</para>
  99.  
  100.         <para>- na vsech arch. plati, ze asidu je mnohem mene, nez teoreticky
  101.         pocet soucasne bezicich tasku ~ adresovych prostoru, takze je
  102.         implementovan mechanismus, ktery umoznuje jednomu adresovemu prostoru
  103.         ASID odebrat a pridelit ho jinemu</para>
  104.  
  105.         <para>- vztah task ~ adresovy prostor: teoreticky existuje moznost, ze
  106.         je adresovy prostor sdilen vice tasky, avsak tuto moznost nepouzivame
  107.         a neni ani nijak osetrena. Tim padem plati, ze kazdy task ma vlastni
  108.         adresovy prostor</para>
  109.       </section>
  110.     </section>
  111.  
  112.     <section>
  113.       <title>Virtual address translation</title>
  114.  
  115.       <section id="page_tables">
  116.         <title>Page tables</title>
  117.  
  118.         <para>HelenOS kernel has two different approaches to the paging
  119.         implementation: <emphasis>4 level page tables</emphasis> and
  120.         <emphasis>global hash tables</emphasis>, which are accessible via
  121.         generic paging abstraction layer. This division was caused by the
  122.         major architectural differences between different platforms.</para>
  123.  
  124.         <formalpara>
  125.           <title>4-level page tables</title>
  126.  
  127.           <para>4-level page tables are the generalization of the hardware
  128.           capabilities of the certain platforms. <itemizedlist>
  129.               <listitem>
  130.                  ia32 uses 2-level page tables, with full hardware support.
  131.               </listitem>
  132.  
  133.               <listitem>
  134.                  amd64 uses 4-level page tables, also coming with full hardware support.
  135.               </listitem>
  136.  
  137.               <listitem>
  138.                  mips and ppc32 have 2-level tables, software simulated support.
  139.               </listitem>
  140.             </itemizedlist></para>
  141.         </formalpara>
  142.  
  143.         <formalpara>
  144.           <title>Global hash tables</title>
  145.  
  146.           <para>- global page hash table: existuje jen jedna v celem systemu
  147.           (vyuziva ji ia64), pozn. ia64 ma zatim vypnuty VHPT. Pouziva se
  148.           genericke hash table s oddelenymi collision chains</para>
  149.         </formalpara>
  150.  
  151.         <para>Thanks to the abstract paging interface, there is possibility
  152.         left have more paging implementations, for example B-Tree page
  153.         tables.</para>
  154.       </section>
  155.  
  156.       <section id="tlb">
  157.         <title>Translation Lookaside buffer</title>
  158.  
  159.         <para>- TLB cachuji informace ve strankovacich tabulkach; alternativne
  160.         se lze na strankovaci tabulky (ci ruzne hw rozsireni [e.g. VHPT, ppc32
  161.         hw hash table]) divat jako na velke TLB</para>
  162.  
  163.         <para>- pri modifikaci mapovani nebo odstraneni mapovani ze
  164.         strankovacich tabulek je potreba zajistit konsistenci TLB a techto
  165.         tabulek; nutne delat na vsech CPU; na to mame zjednodusenou verzi TLB
  166.         shootdown mechanismu; je to variace na algoritmus popsany zde: D.
  167.         Black et al., "Translation Lookaside Buffer Consistency: A Software
  168.        Approach," Proc. Third Int'l Conf. Architectural Support for
  169.        Programming Languages and Operating Systems, 1989, pp. 113-122.</para>
  170.  
  171.        <para>- nutno poznamenat, ze existuji odlehcenejsi verze TLB shootdown
  172.        algoritmu</para>
  173.      </section>
  174.    </section>
  175.  </section>
  176.  
  177.  <!-- End of VM -->
  178.  
  179.  <section>
  180.    <!-- Phys mem -->
  181.  
  182.    <title>Physical memory management</title>
  183.  
  184.    <section id="zones_and_frames">
  185.      <title>Zones and frames</title>
  186.  
  187.      <para><!--graphic fileref="images/mm2.png" /--><!--graphic fileref="images/buddy_alloc.svg" format="SVG" /--></para>
  188.  
  189.      <para>On some architectures not whole physical memory is available for
  190.      conventional usage. This limitations require from kernel to maintain a
  191.      table of available and unavailable ranges of physical memory addresses.
  192.      Main idea of zones is in creating memory zone entity, that is a
  193.      continuous chunk of memory available for allocation. If some chunk is
  194.      not available, we simply do not put it in any zone.</para>
  195.  
  196.      <para>Zone is also serves for informational purposes, containing
  197.      information about number of free and busy frames. Physical memory
  198.      allocation is also done inside the certain zone. Allocation of zone
  199.      frame must be organized by the <link linkend="frame_allocator">frame
  200.      allocator</link> associated with the zone.</para>
  201.  
  202.      <para>Some of the architectures (mips32, ppc32) have only one zone, that
  203.      covers whole physical memory, and the others (like ia32) may have
  204.      multiple zones. Information about zones on current machine is stored in
  205.      BIOS hardware tables or can be hardcoded into kernel during compile
  206.      time.</para>
  207.    </section>
  208.  
  209.    <section id="frame_allocator">
  210.      <title>Frame allocator</title>
  211.  
  212.      <para><mediaobject id="frame_alloc">
  213.          <imageobject role="html">
  214.            <imagedata fileref="images/frame_alloc.png" format="PNG" />
  215.          </imageobject>
  216.  
  217.          <imageobject role="fop">
  218.            <imagedata fileref="images.vector/frame_alloc.svg" format="SVG" />
  219.          </imageobject>
  220.        </mediaobject></para>
  221.  
  222.      <formalpara>
  223.        <title>Overview</title>
  224.  
  225.        <para>Frame allocator provides physical memory allocation for the
  226.        kernel. Because of zonal organization of physical memory, frame
  227.        allocator is always working in context of some zone, thus making
  228.        impossible to allocate a piece of memory, which lays in different
  229.        zone, which cannot happen, because two adjacent zones can be merged
  230.        into one. Frame allocator is also being responsible to update
  231.        information on the number of free/busy frames in zone. Physical memory
  232.        allocation inside one <link linkend="zones_and_frames">memory
  233.        zone</link> is being handled by an instance of <link
  234.        linkend="buddy_allocator">buddy allocator</link> tailored to allocate
  235.        blocks of physical memory frames.</para>
  236.      </formalpara>
  237.  
  238.      <formalpara>
  239.        <title>Allocation / deallocation</title>
  240.  
  241.        <para>Upon allocation request, frame allocator tries to find first
  242.        zone, that can satisfy the incoming request (has required amount of
  243.        free frames to allocate). During deallocation, frame allocator needs
  244.        to find zone, that contain deallocated frame. This approach could
  245.        bring up two potential problems: <itemizedlist>
  246.            <listitem>
  247.               Linear search of zones does not any good to performance, but number of zones is not expected to be high. And if yes, list of zones can be replaced with more time-efficient B-tree.
  248.            </listitem>
  249.  
  250.            <listitem>
  251.               Quickly find out if zone contains required number of frames to allocate and if this chunk of memory is properly aligned. This issue is perfectly solved bu the buddy allocator.
  252.            </listitem>
  253.          </itemizedlist></para>
  254.      </formalpara>
  255.    </section>
  256.  
  257.    <section id="buddy_allocator">
  258.      <title>Buddy allocator</title>
  259.  
  260.      <section>
  261.        <title>Overview</title>
  262.  
  263.        <para><mediaobject id="buddy_alloc">
  264.            <imageobject role="html">
  265.              <imagedata fileref="images/buddy_alloc.png" format="PNG" />
  266.            </imageobject>
  267.  
  268.            <imageobject role="fop">
  269.              <imagedata fileref="images.vector/buddy_alloc.svg" format="SVG" />
  270.            </imageobject>
  271.          </mediaobject></para>
  272.  
  273.        <para>In the buddy allocator, the memory is broken down into
  274.        power-of-two sized naturally aligned blocks. These blocks are
  275.        organized in an array of lists, in which the list with index i
  276.        contains all unallocated blocks of size
  277.        <mathphrase>2<superscript>i</superscript></mathphrase>. The index i is
  278.        called the order of block. Should there be two adjacent equally sized
  279.        blocks in the list i<mathphrase> </mathphrase>(i.e. buddies), the
  280.        buddy allocator would coalesce them and put the resulting block in
  281.        list <mathphrase>i + 1</mathphrase>, provided that the resulting block
  282.        would be naturally aligned. Similarily, when the allocator is asked to
  283.        allocate a block of size
  284.        <mathphrase>2<superscript>i</superscript></mathphrase>, it first tries
  285.        to satisfy the request from the list with index i. If the request
  286.        cannot be satisfied (i.e. the list i is empty), the buddy allocator
  287.        will try to allocate and split a larger block from the list with index
  288.        i + 1. Both of these algorithms are recursive. The recursion ends
  289.        either when there are no blocks to coalesce in the former case or when
  290.        there are no blocks that can be split in the latter case.</para>
  291.  
  292.        <!--graphic fileref="images/mm1.png" format="EPS" /-->
  293.  
  294.        <para>This approach greatly reduces external fragmentation of memory
  295.        and helps in allocating bigger continuous blocks of memory aligned to
  296.        their size. On the other hand, the buddy allocator suffers increased
  297.        internal fragmentation of memory and is not suitable for general
  298.        kernel allocations. This purpose is better addressed by the <link
  299.        linkend="slab">slab allocator</link>.</para>
  300.      </section>
  301.  
  302.      <section>
  303.        <title>Implementation</title>
  304.  
  305.        <para>The buddy allocator is, in fact, an abstract framework wich can
  306.        be easily specialized to serve one particular task. It knows nothing
  307.        about the nature of memory it helps to allocate. In order to beat the
  308.        lack of this knowledge, the buddy allocator exports an interface that
  309.        each of its clients is required to implement. When supplied with an
  310.        implementation of this interface, the buddy allocator can use
  311.        specialized external functions to find a buddy for a block, split and
  312.        coalesce blocks, manipulate block order and mark blocks busy or
  313.        available. For precise documentation of this interface, refer to
  314.        <emphasis>"HelenOS Generic Kernel Reference Manual"</emphasis>.</para>
  315.  
  316.        <formalpara>
  317.          <title>Data organization</title>
  318.  
  319.          <para>Each entity allocable by the buddy allocator is required to
  320.          contain space for storing block order number and a link variable
  321.          used to interconnect blocks within the same order.</para>
  322.  
  323.          <para>Whatever entities are allocated by the buddy allocator, the
  324.          first entity within a block is used to represent the entire block.
  325.          The first entity keeps the order of the whole block. Other entities
  326.          within the block are assigned the magic value
  327.          <constant>BUDDY_INNER_BLOCK</constant>. This is especially important
  328.          for effective identification of buddies in a one-dimensional array
  329.          because the entity that represents a potential buddy cannot be
  330.          associated with <constant>BUDDY_INNER_BLOCK</constant> (i.e. if it
  331.          is associated with <constant>BUDDY_INNER_BLOCK</constant> then it is
  332.          not a buddy).</para>
  333.  
  334.          <para>The buddy allocator always uses the first frame to represent
  335.          the frame block. This frame contains <varname>buddy_order</varname>
  336.          variable to provide information about the block size it actually
  337.          represents (
  338.          <mathphrase>2<superscript>buddy_order</superscript></mathphrase>
  339.          frames block). Other frames in block have this value set to magic
  340.          <constant>BUDDY_INNER_BLOCK</constant> that is much greater than
  341.          buddy <varname>max_order</varname> value.</para>
  342.  
  343.          <para>Each <varname>frame_t</varname> also contains pointer member
  344.          to hold frame structure in the linked list inside one order.</para>
  345.        </formalpara>
  346.  
  347.        <formalpara>
  348.          <title>Allocation algorithm</title>
  349.  
  350.          <para>Upon <mathphrase>2<superscript>i</superscript></mathphrase>
  351.          frames block allocation request, allocator checks if there are any
  352.          blocks available at the order list <varname>i</varname>. If yes,
  353.          removes block from order list and returns its address. If no,
  354.          recursively allocates
  355.          <mathphrase>2<superscript>i+1</superscript></mathphrase> frame
  356.          block, splits it into two
  357.          <mathphrase>2<superscript>i</superscript></mathphrase> frame blocks.
  358.          Then adds one of the blocks to the <varname>i</varname> order list
  359.          and returns address of another.</para>
  360.        </formalpara>
  361.  
  362.        <formalpara>
  363.          <title>Deallocation algorithm</title>
  364.  
  365.          <para>Check if block has so called buddy (another free
  366.          <mathphrase>2<superscript>i</superscript></mathphrase> frame block
  367.          that can be linked with freed block into the
  368.          <mathphrase>2<superscript>i+1</superscript></mathphrase> block).
  369.          Technically, buddy is a odd/even block for even/odd block
  370.          respectively. Plus we can put an extra requirement, that resulting
  371.          block must be aligned to its size. This requirement guarantees
  372.          natural block alignment for the blocks coming out the allocation
  373.          system.</para>
  374.  
  375.          <para>Using direct pointer arithmetics,
  376.          <varname>frame_t::ref_count</varname> and
  377.          <varname>frame_t::buddy_order</varname> variables, finding buddy is
  378.          done at constant time.</para>
  379.        </formalpara>
  380.      </section>
  381.    </section>
  382.  
  383.    <section id="slab">
  384.      <title>Slab allocator</title>
  385.  
  386.      <section>
  387.        <title>Overview</title>
  388.  
  389.        <para><termdef><glossterm>Slab</glossterm> represents a contiguous
  390.        piece of memory, usually made of several physically contiguous
  391.        pages.</termdef> <termdef><glossterm>Slab cache</glossterm> consists
  392.        of one or more slabs.</termdef></para>
  393.  
  394.        <para>The majority of memory allocation requests in the kernel are for
  395.        small, frequently used data structures. For this purpose the slab
  396.        allocator is a perfect solution. The basic idea behind the slab
  397.        allocator is to have lists of commonly used objects available packed
  398.        into pages. This avoids the overhead of allocating and destroying
  399.        commonly used types of objects such threads, virtual memory structures
  400.        etc. Also due to the exact allocated size matching, slab allocation
  401.        completely eliminates internal fragmentation issue.</para>
  402.      </section>
  403.  
  404.      <section>
  405.        <title>Implementation</title>
  406.  
  407.        <para><mediaobject id="slab_alloc">
  408.            <imageobject role="html">
  409.              <imagedata fileref="images/slab_alloc.png" format="PNG" />
  410.            </imageobject>
  411.  
  412.            <imageobject role="fop">
  413.              <imagedata fileref="images.vector/slab_alloc.svg" format="SVG" />
  414.            </imageobject>
  415.          </mediaobject></para>
  416.  
  417.        <para>The SLAB allocator is closely modelled after <ulink
  418.        url="http://www.usenix.org/events/usenix01/full_papers/bonwick/bonwick_html/">
  419.        OpenSolaris SLAB allocator by Jeff Bonwick and Jonathan Adams </ulink>
  420.        with the following exceptions: <itemizedlist>
  421.            <listitem>
  422.               empty SLABS are deallocated immediately (in Linux they are kept in linked list, in Solaris ???)
  423.            </listitem>
  424.  
  425.            <listitem>
  426.               empty magazines are deallocated when not needed (in Solaris they are held in linked list in slab cache)
  427.            </listitem>
  428.          </itemizedlist> Following features are not currently supported but
  429.        would be easy to do: <itemizedlist>
  430.            <listitem>
  431.               - cache coloring
  432.            </listitem>
  433.  
  434.            <listitem>
  435.               - dynamic magazine grow (different magazine sizes are already supported, but we would need to adjust allocation strategy)
  436.            </listitem>
  437.          </itemizedlist></para>
  438.  
  439.        <section>
  440.          <title>Magazine layer</title>
  441.  
  442.          <para>Due to the extensive bottleneck on SMP architures, caused by
  443.          global SLAB locking mechanism, making processing of all slab
  444.          allocation requests serialized, a new layer was introduced to the
  445.          classic slab allocator design. Slab allocator was extended to
  446.          support per-CPU caches 'magazines' to achieve good SMP scaling.
  447.          <termdef>Slab SMP perfromance bottleneck was resolved by introducing
  448.          a per-CPU caching scheme called as <glossterm>magazine
  449.          layer</glossterm></termdef>.</para>
  450.  
  451.          <para>Magazine is a N-element cache of objects, so each magazine can
  452.          satisfy N allocations. Magazine behaves like a automatic weapon
  453.          magazine (LIFO, stack), so the allocation/deallocation become simple
  454.          push/pop pointer operation. Trick is that CPU does not access global
  455.          slab allocator data during the allocation from its magazine, thus
  456.          making possible parallel allocations between CPUs.</para>
  457.  
  458.          <para>Implementation also requires adding another feature as the
  459.          CPU-bound magazine is actually a pair of magazines to avoid
  460.          thrashing when during allocation/deallocatiion of 1 item at the
  461.          magazine size boundary. LIFO order is enforced, which should avoid
  462.          fragmentation as much as possible.</para>
  463.  
  464.          <para>Another important entity of magazine layer is a full magazine
  465.          depot, that stores full magazines which are used by any of the CPU
  466.          magazine caches to reload active CPU magazine. Magazine depot can be
  467.          pre-filled with full magazines during initialization, but in current
  468.          implementation it is filled during object deallocation, when CPU
  469.          magazine becomes full.</para>
  470.  
  471.          <para>Slab allocator control structures are allocated from special
  472.          slabs, that are marked by special flag, indicating that it should
  473.          not be used for slab magazine layer. This is done to avoid possible
  474.          infinite recursions and deadlock during conventional slab allocaiton
  475.          requests.</para>
  476.        </section>
  477.  
  478.        <section>
  479.          <title>Allocation/deallocation</title>
  480.  
  481.          <para>Every cache contains list of full slabs and list of partialy
  482.          full slabs. Empty slabs are immediately freed (thrashing will be
  483.          avoided because of magazines).</para>
  484.  
  485.          <para>The SLAB allocator allocates lots of space and does not free
  486.          it. When frame allocator fails to allocate the frame, it calls
  487.          slab_reclaim(). It tries 'light reclaim' first, then brutal reclaim.
  488.          The light reclaim releases slabs from cpu-shared magazine-list,
  489.          until at least 1 slab is deallocated in each cache (this algorithm
  490.          should probably change). The brutal reclaim removes all cached
  491.          objects, even from CPU-bound magazines.</para>
  492.  
  493.          <formalpara>
  494.            <title>Allocation</title>
  495.  
  496.            <para><emphasis>Step 1.</emphasis> When it comes to the allocation
  497.            request, slab allocator first of all checks availability of memory
  498.            in local CPU-bound magazine. If it is there, we would just "pop"
  499.            the CPU magazine and return the pointer to object.</para>
  500.  
  501.            <para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is
  502.            empty, allocator will attempt to reload magazin, swapping it with
  503.            second CPU magazine and returns to the first step.</para>
  504.  
  505.            <para><emphasis>Step 3.</emphasis> Now we are in the situation
  506.            when both CPU-bound magazines are empty, which makes allocator to
  507.            access shared full-magazines depot to reload CPU-bound magazines.
  508.            If reload is succesful (meaning there are full magazines in depot)
  509.            algoritm continues at Step 1.</para>
  510.  
  511.            <para><emphasis>Step 4.</emphasis> Final step of the allocation.
  512.            In this step object is allocated from the conventional slab layer
  513.            and pointer is returned.</para>
  514.          </formalpara>
  515.  
  516.          <formalpara>
  517.            <title>Deallocation</title>
  518.  
  519.            <para><emphasis>Step 1.</emphasis> During deallocation request,
  520.            slab allocator will check if the local CPU-bound magazine is not
  521.            full. In this case we will just push the pointer to this
  522.            magazine.</para>
  523.  
  524.            <para><emphasis>Step 2.</emphasis> If the CPU-bound magazine is
  525.            full, allocator will attempt to reload magazin, swapping it with
  526.            second CPU magazine and returns to the first step.</para>
  527.  
  528.            <para><emphasis>Step 3.</emphasis> Now we are in the situation
  529.            when both CPU-bound magazines are full, which makes allocator to
  530.            access shared full-magazines depot to put one of the magazines to
  531.            the depot and creating new empty magazine. Algoritm continues at
  532.            Step 1.</para>
  533.          </formalpara>
  534.        </section>
  535.      </section>
  536.    </section>
  537.  
  538.    <!-- End of Physmem -->
  539.  </section>
  540.  
  541.  <section>
  542.    <title>Memory sharing</title>
  543.  
  544.    <para>Not implemented yet(?)</para>
  545.  </section>
  546. </chapter>