Subversion Repositories HelenOS-doc

Rev

Rev 76 | Rev 79 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed

Rev Author Line No. Line
9 bondari 1
<?xml version="1.0" encoding="UTF-8"?>
11 bondari 2
<chapter id="mm">
3
  <?dbhtml filename="mm.html"?>
9 bondari 4
 
11 bondari 5
  <title>Memory management</title>
9 bondari 6
 
67 jermar 7
  <para>In previous chapters, this book described the scheduling subsystem as
8
  the creator of the impression that threads execute in parallel. The memory
9
  management subsystem, on the other hand, creates the impression that there
10
  is enough physical memory for the kernel and that userspace tasks have the
11
  entire address space only for themselves.</para>
64 jermar 12
 
26 bondari 13
  <section>
64 jermar 14
    <title>Physical memory management</title>
15
 
16
    <section id="zones_and_frames">
17
      <title>Zones and frames</title>
18
 
67 jermar 19
      <para>HelenOS represents continuous areas of physical memory in
20
      structures called frame zones (abbreviated as zones). Each zone contains
21
      information about the number of allocated and unallocated physical
22
      memory frames as well as the physical base address of the zone and
23
      number of frames contained in it. A zone also contains an array of frame
24
      structures describing each frame of the zone and, in the last, but not
25
      the least important, front, each zone is equipped with a buddy system
26
      that faciliates effective allocation of power-of-two sized block of
27
      frames.</para>
64 jermar 28
 
67 jermar 29
      <para>This organization of physical memory provides good preconditions
30
      for hot-plugging of more zones. There is also one currently unused zone
31
      attribute: <code>flags</code>. The attribute could be used to give a
32
      special meaning to some zones in the future.</para>
64 jermar 33
 
67 jermar 34
      <para>The zones are linked in a doubly-linked list. This might seem a
35
      bit ineffective because the zone list is walked everytime a frame is
36
      allocated or deallocated. However, this does not represent a significant
37
      performance problem as it is expected that the number of zones will be
38
      rather low. Moreover, most architectures merge all zones into
39
      one.</para>
40
 
76 palkovsky 41
      <para>Every physical memory frame in a zone, is described by a structure
42
      that contains number of references and other data used by buddy
67 jermar 43
      system.</para>
64 jermar 44
    </section>
45
 
46
    <section id="frame_allocator">
71 bondari 47
      <indexterm>
48
        <primary>frame allocator</primary>
49
      </indexterm>
50
 
64 jermar 51
      <title>Frame allocator</title>
52
 
67 jermar 53
      <para>The frame allocator satisfies kernel requests to allocate
54
      power-of-two sized blocks of physical memory. Because of zonal
55
      organization of physical memory, the frame allocator is always working
76 palkovsky 56
      within a context of a particular frame zone. In order to carry out the
67 jermar 57
      allocation requests, the frame allocator is tightly integrated with the
58
      buddy system belonging to the zone. The frame allocator is also
59
      responsible for updating information about the number of free and busy
60
      frames in the zone. <figure>
61
          <mediaobject id="frame_alloc">
77 bondari 62
            <imageobject role="eps">
63
              <imagedata fileref="images.vector/frame_alloc.eps" format="EPS" />
64
            </imageobject>
65
 
67 jermar 66
            <imageobject role="html">
67
              <imagedata fileref="images/frame_alloc.png" format="PNG" />
68
            </imageobject>
64 jermar 69
 
67 jermar 70
            <imageobject role="fop">
71
              <imagedata fileref="images.vector/frame_alloc.svg" format="SVG" />
72
            </imageobject>
73
          </mediaobject>
64 jermar 74
 
67 jermar 75
          <title>Frame allocator scheme.</title>
76
        </figure></para>
64 jermar 77
 
78
      <formalpara>
79
        <title>Allocation / deallocation</title>
80
 
67 jermar 81
        <para>Upon allocation request via function <code>frame_alloc</code>,
82
        the frame allocator first tries to find a zone that can satisfy the
83
        request (i.e. has the required amount of free frames). Once a suitable
84
        zone is found, the frame allocator uses the buddy allocator on the
85
        zone's buddy system to perform the allocation. During deallocation,
86
        which is triggered by a call to <code>frame_free</code>, the frame
87
        allocator looks up the respective zone that contains the frame being
88
        deallocated. Afterwards, it calls the buddy allocator again, this time
89
        to take care of deallocation within the zone's buddy system.</para>
64 jermar 90
      </formalpara>
91
    </section>
92
 
93
    <section id="buddy_allocator">
71 bondari 94
      <indexterm>
95
        <primary>buddy system</primary>
96
      </indexterm>
97
 
64 jermar 98
      <title>Buddy allocator</title>
99
 
67 jermar 100
      <para>In the buddy system, the memory is broken down into power-of-two
101
      sized naturally aligned blocks. These blocks are organized in an array
102
      of lists, in which the list with index i contains all unallocated blocks
103
      of size <mathphrase>2<superscript>i</superscript></mathphrase>. The
104
      index i is called the order of block. Should there be two adjacent
105
      equally sized blocks in the list i<mathphrase />(i.e. buddies), the
106
      buddy allocator would coalesce them and put the resulting block in list
107
      <mathphrase>i + 1</mathphrase>, provided that the resulting block would
108
      be naturally aligned. Similarily, when the allocator is asked to
109
      allocate a block of size
110
      <mathphrase>2<superscript>i</superscript></mathphrase>, it first tries
111
      to satisfy the request from the list with index i. If the request cannot
112
      be satisfied (i.e. the list i is empty), the buddy allocator will try to
113
      allocate and split a larger block from the list with index i + 1. Both
114
      of these algorithms are recursive. The recursion ends either when there
115
      are no blocks to coalesce in the former case or when there are no blocks
116
      that can be split in the latter case.</para>
64 jermar 117
 
67 jermar 118
      <para>This approach greatly reduces external fragmentation of memory and
119
      helps in allocating bigger continuous blocks of memory aligned to their
120
      size. On the other hand, the buddy allocator suffers increased internal
121
      fragmentation of memory and is not suitable for general kernel
122
      allocations. This purpose is better addressed by the <link
123
      linkend="slab">slab allocator</link>.<figure>
64 jermar 124
          <mediaobject id="buddy_alloc">
77 bondari 125
            <imageobject role="eps">
126
              <imagedata fileref="images.vector/buddy_alloc.eps" format="EPS" />
127
            </imageobject>
128
 
64 jermar 129
            <imageobject role="html">
130
              <imagedata fileref="images/buddy_alloc.png" format="PNG" />
131
            </imageobject>
132
 
133
            <imageobject role="fop">
134
              <imagedata fileref="images.vector/buddy_alloc.svg" format="SVG" />
135
            </imageobject>
136
          </mediaobject>
137
 
138
          <title>Buddy system scheme.</title>
67 jermar 139
        </figure></para>
64 jermar 140
 
141
      <section>
142
        <title>Implementation</title>
143
 
144
        <para>The buddy allocator is, in fact, an abstract framework wich can
145
        be easily specialized to serve one particular task. It knows nothing
146
        about the nature of memory it helps to allocate. In order to beat the
147
        lack of this knowledge, the buddy allocator exports an interface that
148
        each of its clients is required to implement. When supplied with an
149
        implementation of this interface, the buddy allocator can use
150
        specialized external functions to find a buddy for a block, split and
151
        coalesce blocks, manipulate block order and mark blocks busy or
67 jermar 152
        available.</para>
64 jermar 153
 
154
        <formalpara>
155
          <title>Data organization</title>
156
 
157
          <para>Each entity allocable by the buddy allocator is required to
158
          contain space for storing block order number and a link variable
159
          used to interconnect blocks within the same order.</para>
160
 
161
          <para>Whatever entities are allocated by the buddy allocator, the
162
          first entity within a block is used to represent the entire block.
163
          The first entity keeps the order of the whole block. Other entities
164
          within the block are assigned the magic value
165
          <constant>BUDDY_INNER_BLOCK</constant>. This is especially important
166
          for effective identification of buddies in a one-dimensional array
167
          because the entity that represents a potential buddy cannot be
168
          associated with <constant>BUDDY_INNER_BLOCK</constant> (i.e. if it
169
          is associated with <constant>BUDDY_INNER_BLOCK</constant> then it is
170
          not a buddy).</para>
171
        </formalpara>
172
      </section>
173
    </section>
174
 
175
    <section id="slab">
71 bondari 176
      <indexterm>
177
        <primary>slab allocator</primary>
178
      </indexterm>
179
 
70 jermar 180
      <title>Slab allocator</title>
64 jermar 181
 
67 jermar 182
      <para>The majority of memory allocation requests in the kernel is for
183
      small, frequently used data structures. The basic idea behind the slab
184
      allocator is that commonly used objects are preallocated in continuous
185
      areas of physical memory called slabs<footnote>
186
          <para>Slabs are in fact blocks of physical memory frames allocated
187
          from the frame allocator.</para>
188
        </footnote>. Whenever an object is to be allocated, the slab allocator
189
      returns the first available item from a suitable slab corresponding to
190
      the object type<footnote>
191
          <para>The mechanism is rather more complicated, see the next
192
          paragraph.</para>
193
        </footnote>. Due to the fact that the sizes of the requested and
194
      allocated object match, the slab allocator significantly reduces
195
      internal fragmentation.</para>
64 jermar 196
 
71 bondari 197
      <indexterm>
198
        <primary>slab allocator</primary>
199
 
72 bondari 200
        <secondary>- slab cache</secondary>
71 bondari 201
      </indexterm>
202
 
67 jermar 203
      <para>Slabs of one object type are organized in a structure called slab
204
      cache. There are ususally more slabs in the slab cache, depending on
205
      previous allocations. If the the slab cache runs out of available slabs,
206
      new slabs are allocated. In order to exploit parallelism and to avoid
207
      locking of shared spinlocks, slab caches can have variants of
208
      processor-private slabs called magazines. On each processor, there is a
209
      two-magazine cache. Full magazines that are not part of any
210
      per-processor magazine cache are stored in a global list of full
211
      magazines.</para>
64 jermar 212
 
71 bondari 213
      <indexterm>
214
        <primary>slab allocator</primary>
215
 
72 bondari 216
        <secondary>- magazine</secondary>
71 bondari 217
      </indexterm>
218
 
67 jermar 219
      <para>Each object begins its life in a slab. When it is allocated from
220
      there, the slab allocator calls a constructor that is registered in the
221
      respective slab cache. The constructor initializes and brings the object
222
      into a known state. The object is then used by the user. When the user
223
      later frees the object, the slab allocator puts it into a processor
71 bondari 224
      private <indexterm>
225
          <primary>slab allocator</primary>
226
 
72 bondari 227
          <secondary>- magazine</secondary>
71 bondari 228
        </indexterm>magazine cache, from where it can be precedently allocated
67 jermar 229
      again. Note that allocations satisfied from a magazine are already
230
      initialized by the constructor. When both of the processor cached
231
      magazines get full, the allocator will move one of the magazines to the
232
      list of full magazines. Similarily, when allocating from an empty
233
      processor magazine cache, the kernel will reload only one magazine from
234
      the list of full magazines. In other words, the slab allocator tries to
235
      keep the processor magazine cache only half-full in order to prevent
236
      thrashing when allocations and deallocations interleave on magazine
70 jermar 237
      boundaries. The advantage of this setup is that during most of the
238
      allocations, no global spinlock needs to be held.</para>
65 jermar 239
 
67 jermar 240
      <para>Should HelenOS run short of memory, it would start deallocating
241
      objects from magazines, calling slab cache destructor on them and
242
      putting them back into slabs. When a slab contanins no allocated object,
243
      it is immediately freed.</para>
66 bondari 244
 
71 bondari 245
      <para>
246
        <figure>
64 jermar 247
          <mediaobject id="slab_alloc">
77 bondari 248
            <imageobject role="eps">
249
              <imagedata fileref="images.vector/slab_alloc.eps" format="EPS" />
250
            </imageobject>
251
 
64 jermar 252
            <imageobject role="html">
253
              <imagedata fileref="images/slab_alloc.png" format="PNG" />
254
            </imageobject>
77 bondari 255
 
256
            <imageobject role="fop">
257
              <imagedata fileref="images.vector/slab_alloc.svg" format="SVG" />
258
            </imageobject>
64 jermar 259
          </mediaobject>
260
 
261
          <title>Slab allocator scheme.</title>
71 bondari 262
        </figure>
263
      </para>
64 jermar 264
 
67 jermar 265
      <section>
266
        <title>Implementation</title>
267
 
268
        <para>The slab allocator is closely modelled after OpenSolaris slab
71 bondari 269
        allocator by Jeff Bonwick and Jonathan Adams <xref
270
        linkend="Bonwick01" /> with the following exceptions:<itemizedlist>
271
            <listitem>empty slabs are immediately deallocated and</listitem>
66 bondari 272
 
273
            <listitem>
70 jermar 274
              <para>empty magazines are deallocated when not needed.</para>
66 bondari 275
            </listitem>
70 jermar 276
          </itemizedlist>The following features are not currently supported
277
        but would be easy to do: <itemizedlist>
71 bondari 278
            <listitem>cache coloring and</listitem>
64 jermar 279
 
71 bondari 280
            <listitem>dynamic magazine grow (different magazine sizes are
281
            already supported, but the allocation strategy would need to be
282
            adjusted).</listitem>
64 jermar 283
          </itemizedlist></para>
284
 
285
        <section>
286
          <title>Allocation/deallocation</title>
287
 
70 jermar 288
          <para>The following two paragraphs summarize and complete the
289
          description of the slab allocator operation (i.e.
290
          <code>slab_alloc</code> and <code>slab_free</code>
291
          operations).</para>
64 jermar 292
 
293
          <formalpara>
294
            <title>Allocation</title>
295
 
70 jermar 296
            <para><emphasis>Step 1.</emphasis> When an allocation request
297
            comes, the slab allocator checks availability of memory in the
298
            current magazine of the local processor magazine cache. If the
76 palkovsky 299
            available memory is there, the allocator just pops the object from
300
            magazine and returns it.</para>
64 jermar 301
 
70 jermar 302
            <para><emphasis>Step 2.</emphasis> If the current magazine in the
303
            processor magazine cache is empty, the allocator will attempt to
304
            swap it with the last magazine from the cache and return to the
305
            first step. If also the last magazine is empty, the algorithm will
306
            fall through to Step 3.</para>
64 jermar 307
 
70 jermar 308
            <para><emphasis>Step 3.</emphasis> Now the allocator is in the
309
            situation when both magazines in the processor magazine cache are
310
            empty. The allocator reloads one magazine from the shared list of
311
            full magazines. If the reload is successful (i.e. there are full
312
            magazines in the list), the algorithm continues with Step
313
            1.</para>
64 jermar 314
 
70 jermar 315
            <para><emphasis>Step 4.</emphasis> In this fail-safe step, an
316
            object is allocated from the conventional slab layer and a pointer
317
            to it is returned. If also the last magazine is full,</para>
64 jermar 318
          </formalpara>
319
 
320
          <formalpara>
321
            <title>Deallocation</title>
322
 
70 jermar 323
            <para><emphasis>Step 1.</emphasis> During a deallocation request,
324
            the slab allocator checks if the current magazine of the local
76 palkovsky 325
            processor magazine cache is not full. If it is, the pointer to the
70 jermar 326
            objects is just pushed into the magazine and the algorithm
327
            returns.</para>
64 jermar 328
 
70 jermar 329
            <para><emphasis>Step 2.</emphasis> If the current magazine is
330
            full, the allocator will attempt to swap it with the last magazine
331
            from the cache and return to the first step. If also the last
332
            magazine is empty, the algorithm will fall through to Step
333
            3.</para>
64 jermar 334
 
70 jermar 335
            <para><emphasis>Step 3.</emphasis> Now the allocator is in the
336
            situation when both magazines in the processor magazine cache are
76 palkovsky 337
            full. The allocator tries to allocate a new empty magazine and
338
            flush one of the full magazines to the shared list of full
339
            magazines. If it is successfull, the algoritm continues with Step
340
            1.</para>
341
 
342
            <para><emphasis>Step 4. </emphasis>In case of low memory condition
343
            when the allocation of empty magazine fails, the object is moved
344
            directly into slab. In the worst case object deallocation does not
345
            need to allocate any additional memory.</para>
64 jermar 346
          </formalpara>
347
        </section>
348
      </section>
349
    </section>
350
  </section>
351
 
352
  <section>
67 jermar 353
    <title>Virtual memory management</title>
9 bondari 354
 
67 jermar 355
    <section>
356
      <title>Introduction</title>
357
 
358
      <para>Virtual memory is a special memory management technique, used by
359
      kernel to achieve a bunch of mission critical goals. <itemizedlist>
360
          <listitem>
361
             Isolate each task from other tasks that are running on the system at the same time.
362
          </listitem>
363
 
364
          <listitem>
365
             Allow to allocate more memory, than is actual physical memory size of the machine.
366
          </listitem>
367
 
368
          <listitem>
369
             Allowing, in general, to load and execute two programs that are linked on the same address without complicated relocations.
370
          </listitem>
371
        </itemizedlist></para>
372
 
373
      <para><!--
374
                <para>
375
                        Address spaces. Address space area (B+ tree). Only for uspace. Set of syscalls (shrink/extend etc).
376
                        Special address space area type - device - prohibits shrink/extend syscalls to call on it.
377
                        Address space has link to mapping tables (hierarchical - per Address space, hash - global tables).
378
                </para>
379
 
380
--></para>
381
    </section>
382
 
383
    <section>
71 bondari 384
      <title>Address spaces</title>
70 jermar 385
 
71 bondari 386
      <section>
387
        <indexterm>
388
          <primary>address space</primary>
70 jermar 389
 
72 bondari 390
          <secondary>- area</secondary>
71 bondari 391
        </indexterm>
70 jermar 392
 
393
        <title>Address space areas</title>
67 jermar 394
 
395
        <para>Each address space consists of mutually disjunctive continuous
396
        address space areas. Address space area is precisely defined by its
397
        base address and the number of frames/pages is contains.</para>
398
 
399
        <para>Address space area , that define behaviour and permissions on
400
        the particular area. <itemizedlist>
71 bondari 401
            <listitem><emphasis>AS_AREA_READ</emphasis> flag indicates reading
402
            permission.</listitem>
67 jermar 403
 
71 bondari 404
            <listitem><emphasis>AS_AREA_WRITE</emphasis> flag indicates
405
            writing permission.</listitem>
67 jermar 406
 
71 bondari 407
            <listitem><emphasis>AS_AREA_EXEC</emphasis> flag indicates code
408
            execution permission. Some architectures do not support execution
409
            persmission restriction. In this case this flag has no
410
            effect.</listitem>
67 jermar 411
 
71 bondari 412
            <listitem><emphasis>AS_AREA_DEVICE</emphasis> marks area as mapped
413
            to the device memory.</listitem>
67 jermar 414
          </itemizedlist></para>
415
 
416
        <para>Kernel provides possibility tasks create/expand/shrink/share its
417
        address space via the set of syscalls.</para>
418
      </section>
419
 
420
      <section>
71 bondari 421
        <indexterm>
422
          <primary>address space</primary>
423
 
72 bondari 424
          <secondary>- ASID</secondary>
71 bondari 425
        </indexterm>
426
 
67 jermar 427
        <title>Address Space ID (ASID)</title>
428
 
76 palkovsky 429
        <para>Every task in the operating system has it's own view of the
430
        virtual memory. When performing context switch between different
431
        tasks, the kernel must switch the address space mapping as well. As
432
        modern processors perform very aggressive caching of virtual mappings,
433
        flushing the complete TLB on every context switch would be very
434
        inefficient. To avoid such performance penalty, some architectures
435
        introduce an address space identifier, which allows storing several
436
        different mappings inside TLB.</para>
67 jermar 437
 
76 palkovsky 438
        <para>HelenOS kernel can take advantage of this hardware support by
439
        having an ASID abstraction. I.e. on ia64 kernel ASID is derived from
440
        RID (region identifier) and on the mips32 kernel ASID is actually the
441
        hardware identifier. As expected, this ASID information record is the
442
        part of <emphasis>as_t</emphasis> structure.</para>
67 jermar 443
 
444
        <para>Due to the hardware limitations, hardware ASID has limited
445
        length from 8 bits on ia64 to 24 bits on mips32, which makes it
446
        impossible to use it as unique address space identifier for all tasks
447
        running in the system. In such situations special ASID stealing
448
        algoritm is used, which takes ASID from inactive task and assigns it
449
        to the active task.</para>
450
 
71 bondari 451
        <indexterm>
452
          <primary>address space</primary>
453
 
72 bondari 454
          <secondary>- ASID stealing</secondary>
71 bondari 455
        </indexterm>
456
 
457
        <para>
458
          <classname>ASID stealing algoritm here.</classname>
459
        </para>
67 jermar 460
      </section>
461
    </section>
462
 
71 bondari 463
    <section id="paging">
67 jermar 464
      <title>Virtual address translation</title>
465
 
71 bondari 466
      <section>
467
        <title>Introduction</title>
67 jermar 468
 
71 bondari 469
        <para>Virtual memory is usually using paged memory model, where
470
        virtual memory address space is divided into the
471
        <emphasis>pages</emphasis> (usually having size 4096 bytes) and
472
        physical memory is divided into the frames (same sized as a page, of
473
        course). Each page may be mapped to some frame and then, upon memory
474
        access to the virtual address, CPU performs <emphasis>address
475
        translation</emphasis> during the instruction execution. Non-existing
476
        mapping generates page fault exception, calling kernel exception
477
        handler, thus allowing kernel to manipulate rules of memory access.
478
        Information for pages mapping is stored by kernel in the <link
479
        linkend="page_tables">page tables</link></para>
480
 
481
        <indexterm>
482
          <primary>page tables</primary>
483
        </indexterm>
484
 
485
        <para>The majority of the architectures use multi-level page tables,
486
        which means need to access physical memory several times before
487
        getting physical address. This fact would make serios performance
488
        overhead in virtual memory management. To avoid this <link
489
        linkend="tlb">Traslation Lookaside Buffer (TLB)</link> is used.</para>
490
 
70 jermar 491
        <para>HelenOS kernel has two different approaches to the paging
492
        implementation: <emphasis>4 level page tables</emphasis> and
71 bondari 493
        <emphasis>global hash table</emphasis>, which are accessible via
70 jermar 494
        generic paging abstraction layer. Such different functionality was
495
        caused by the major architectural differences between supported
496
        platforms. This abstraction is implemented with help of the global
497
        structure of pointers to basic mapping functions
498
        <emphasis>page_mapping_operations</emphasis>. To achieve different
499
        functionality of page tables, corresponding layer must implement
500
        functions, declared in
501
        <emphasis>page_mapping_operations</emphasis></para>
67 jermar 502
 
71 bondari 503
        <para>Thanks to the abstract paging interface, there was a place left
504
        for more paging implementations (besides already implemented
73 bondari 505
        hieararchical page tables and hash table), for example <indexterm>
506
            <primary>B-tree</primary>
507
          </indexterm> B-Tree based page tables.</para>
71 bondari 508
      </section>
67 jermar 509
 
71 bondari 510
      <section id="page_tables">
511
        <indexterm>
512
          <primary>page tables</primary>
67 jermar 513
 
72 bondari 514
          <secondary>- hierarchical</secondary>
71 bondari 515
        </indexterm>
67 jermar 516
 
71 bondari 517
        <title>Hierarchical 4-level page tables</title>
67 jermar 518
 
71 bondari 519
        <para>Hierarchical 4-level page tables are the generalization of the
520
        hardware capabilities of most architectures. Each address space has
521
        its own page tables.<itemizedlist>
522
            <listitem>ia32 uses 2-level page tables, with full hardware
523
            support.</listitem>
67 jermar 524
 
71 bondari 525
            <listitem>amd64 uses 4-level page tables, also coming with full
526
            hardware support.</listitem>
67 jermar 527
 
71 bondari 528
            <listitem>mips and ppc32 have 2-level tables, software simulated
529
            support.</listitem>
530
          </itemizedlist></para>
67 jermar 531
      </section>
532
 
71 bondari 533
      <section>
534
        <indexterm>
535
          <primary>page tables</primary>
67 jermar 536
 
72 bondari 537
          <secondary>- hashing</secondary>
71 bondari 538
        </indexterm>
67 jermar 539
 
71 bondari 540
        <title>Global hash table</title>
67 jermar 541
 
71 bondari 542
        <para>Implementation of the global hash table was encouraged by the
543
        ia64 architecture support. One of the major differences between global
544
        hash table and hierarchical tables is that global hash table exists
545
        only once in the system and the hierarchical tables are maintained per
546
        address space.</para>
67 jermar 547
 
71 bondari 548
        <para>Thus, hash table contains information about all address spaces
549
        mappings in the system, so, the hash of an entry must contain
550
        information of both address space pointer or id and the virtual
551
        address of the page. Generic hash table implementation assumes that
552
        the addresses of the pointers to the address spaces are likely to be
553
        on the close addresses, so it uses least significant bits for hash;
554
        also it assumes that the virtual page addresses have roughly the same
555
        probability of occurring, so the least significant bits of VPN compose
556
        the hash index.</para>
67 jermar 557
 
74 bondari 558
        <para>Paging hash table uses generic hash table with collision chains
559
        (see the <link linkend="hashtables">Data Structures</link> chapter of
560
        this manual for details).</para>
70 jermar 561
      </section>
562
    </section>
67 jermar 563
 
71 bondari 564
    <section id="tlb">
565
      <indexterm>
566
        <primary>TLB</primary>
567
      </indexterm>
67 jermar 568
 
71 bondari 569
      <title>Translation Lookaside buffer</title>
67 jermar 570
 
71 bondari 571
      <para>Due to the extensive overhead during the page mapping lookup in
572
      the page tables, all architectures has fast assotiative cache memory
573
      built-in CPU. This memory called TLB stores recently used page table
574
      entries.</para>
575
 
576
      <section id="tlb_shootdown">
577
        <indexterm>
578
          <primary>TLB</primary>
579
 
72 bondari 580
          <secondary>- TLB shootdown</secondary>
71 bondari 581
        </indexterm>
582
 
583
        <title>TLB consistency. TLB shootdown algorithm.</title>
584
 
585
        <para>Operating system is responsible for keeping TLB consistent by
586
        invalidating the contents of TLB, whenever there is some change in
587
        page tables. Those changes may occur when page or group of pages were
588
        unmapped, mapping is changed or system switching active address space
589
        to schedule a new system task. Moreover, this invalidation operation
590
        must be done an all system CPUs because each CPU has its own
591
        independent TLB cache. Thus maintaining TLB consistency on SMP
592
        configuration as not as trivial task as it looks on the first glance.
593
        Naive solution would assume that is the CPU which wants to invalidate
594
        TLB will invalidate TLB caches on other CPUs. It is not possible on
595
        the most of the architectures, because of the simple fact - flushing
596
        TLB is allowed only on the local CPU and there is no possibility to
597
        access other CPUs' TLB caches, thus invalidate TLB remotely.</para>
598
 
599
        <para>Technique of remote invalidation of TLB entries is called "TLB
600
        shootdown". HelenOS uses a variation of the algorithm described by D.
601
        Black et al., "Translation Lookaside Buffer Consistency: A Software
602
        Approach," Proc. Third Int'l Conf. Architectural Support for
603
        Programming Languages and Operating Systems, 1989, pp. 113-122. <xref
604
        linkend="Black89" /></para>
605
 
606
        <para>As the situation demands, you will want partitial invalidation
607
        of TLB caches. In case of simple memory mapping change it is necessary
608
        to invalidate only one or more adjacent pages. In case if the
609
        architecture is aware of ASIDs, when kernel needs to dump some ASID to
610
        use by another task, it invalidates only entries from this particular
611
        address space. Final option of the TLB invalidation is the complete
612
        TLB cache invalidation, which is the operation that flushes all
613
        entries in TLB.</para>
614
 
615
        <para>TLB shootdown is performed in two phases.</para>
616
 
617
        <formalpara>
618
          <title>Phase 1.</title>
619
 
620
          <para>First, initiator locks a global TLB spinlock, then request is
621
          being put to the local request cache of every other CPU in the
622
          system protected by its spinlock. In case the cache is full, all
623
          requests in the cache are replaced by one request, indicating global
624
          TLB flush. Then the initiator thread sends an IPI message indicating
625
          the TLB shootdown request to the rest of the CPUs and waits actively
626
          until all CPUs confirm TLB invalidating action execution by setting
627
          up a special flag. After setting this flag this thread is blocked on
628
          the TLB spinlock, held by the initiator.</para>
629
        </formalpara>
630
 
631
        <formalpara>
632
          <title>Phase 2.</title>
633
 
634
          <para>All CPUs are waiting on the TLB spinlock to execute TLB
635
          invalidation action and have indicated their intention to the
636
          initiator. Initiator continues, cleaning up its TLB and releasing
637
          the global TLB spinlock. After this all other CPUs gain and
638
          immidiately release TLB spinlock and perform TLB invalidation
639
          actions.</para>
640
        </formalpara>
641
      </section>
67 jermar 642
    </section>
26 bondari 643
  </section>
11 bondari 644
</chapter>