Subversion Repositories HelenOS-doc

Rev

Rev 169 | 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"?>
57 jermar 2
<chapter id="scheduling">
3
  <?dbhtml filename="scheduling.html"?>
9 bondari 4
 
57 jermar 5
  <title>Scheduling</title>
9 bondari 6
 
57 jermar 7
  <para>One of the key aims of the operating system is to create and support
8
  the impression that several activities are executing contemporarily. This is
9
  true for both uniprocessor as well as multiprocessor systems. In the case of
169 jermar 10
  multiprocessor systems, the activities are truly happening in parallel. The
57 jermar 11
  scheduler helps to materialize this impression by planning threads on as
142 jermar 12
  many processors as possible and, when this strategy reaches its limits, by
57 jermar 13
  quickly switching among threads executing on a single processor.</para>
14
 
15
  <section>
16
    <title>Contexts</title>
17
 
140 palkovsky 18
    <para>The term <emphasis>context</emphasis> refers to the set of processor
19
    resources that define the current state of the computation or the
20
    environment and the kernel understands it in several more or less narrow
21
    sences:</para>
57 jermar 22
 
23
    <itemizedlist>
24
      <listitem>
25
        <para>synchronous register context,</para>
26
      </listitem>
27
 
28
      <listitem>
29
        <para>asynchronous register context,</para>
30
      </listitem>
31
 
32
      <listitem>
33
        <para>FPU context and</para>
34
      </listitem>
35
 
36
      <listitem>
37
        <para>memory management context.</para>
38
      </listitem>
39
    </itemizedlist>
40
 
76 palkovsky 41
    <para>The most narrow sense refers to the the synchronous register
57 jermar 42
    context. It includes all the preserved registers as defined by the
43
    architecture. To highlight some, the program counter and stack pointer
76 palkovsky 44
    take part in the synchronous register context. These registers must be
45
    preserved across a procedure call and during synchronous context
46
    switches.</para>
57 jermar 47
 
48
    <para>The next type of the context understood by the kernel is the
49
    asynchronous register context. On an interrupt, the interrupted execution
50
    flow's state must be guaranteed to be eventually completely restored.
51
    Therefore the interrupt context includes, among other things, the scratch
52
    registers as defined by the architecture. As a special optimization and if
53
    certain conditions are met, it need not include the architecture's
54
    preserved registers. The condition mentioned in the previous sentence is
55
    that the low-level assembly language interrupt routines don't modify the
56
    preserved registers. The handlers usually call a higher-level C routine.
57
    The preserved registers are then saved on the stack by the compiler
58
    generated code of the higher-level function. In HelenOS, several
59
    architectures can be compiled with this optimization.</para>
60
 
61
    <para>Although the kernel does not do any floating point
62
    arithmetics<footnote>
63
        <para>Some architectures (e.g. ia64) inevitably use a fixed set of
64 jermar 64
        floating point registers to carry out their normal operations.</para>
57 jermar 65
      </footnote>, it must protect FPU context of userspace threads against
66
    destruction by other threads. Moreover, only a fraction of userspace
67
    programs use the floating point unit. HelenOS contains a generic framework
64 jermar 68
    for switching FPU context only when the switch is forced (i.e. a thread
69
    uses a floating point instruction and its FPU context is not loaded in the
70
    processor).</para>
57 jermar 71
 
72
    <para>The last member of the context family is the memory management
73
    context. It includes memory management registers that identify address
74
    spaces on hardware level (i.e. ASIDs and page tables pointers).</para>
75
 
9 bondari 76
    <section>
131 jermar 77
      <title>Synchronous Context Switches</title>
9 bondari 78
 
57 jermar 79
      <para>The scheduler, but also other pieces of the kernel, make heavy use
80
      of synchronous context switches, because it is a natural vehicle not
81
      only for changes in control flow, but also for switching between two
82
      kernel stacks. Two functions figure in a synchronous context switch
133 jermar 83
      implementation: <code>context_save()</code> and
84
      <code>context_restore()</code>. Note that these two functions break the
57 jermar 85
      natural perception of the linear C code execution flow starting at
86
      function's entry point and ending on one of the function's exit
87
      points.</para>
88
 
133 jermar 89
      <para>When the <code>context_save()</code> function is called, the
57 jermar 90
      synchronous context is saved in a memory structure passed to it. After
133 jermar 91
      executing <code>context_save()</code>, the caller is returned 1 as a
57 jermar 92
      return value. The execution of instructions continues as normally until
133 jermar 93
      <code>context_restore()</code> is called. For the caller, it seems like
57 jermar 94
      the call never returns<footnote>
95
          <para>Which might be a source of problems with variable liveliness
133 jermar 96
          after <code>context_restore()</code>.</para>
57 jermar 97
        </footnote>. Nevertheless, a synchronous register context, which is
133 jermar 98
      saved in a memory structure passed to <code>context_restore()</code>, is
57 jermar 99
      restored, thus transfering the control flow to the place of occurrence
133 jermar 100
      of the corresponding call to <code>context_save()</code>. From the
57 jermar 101
      perspective of the caller of the corresponding
133 jermar 102
      <code>context_save()</code>, it looks like a return from
103
      <code>context_save()</code>. However, this time a return value of 0 is
57 jermar 104
      returned.</para>
9 bondari 105
    </section>
57 jermar 106
  </section>
9 bondari 107
 
57 jermar 108
  <section>
58 jermar 109
    <title>Threads</title>
9 bondari 110
 
142 jermar 111
    <para>A thread is the basic executable entity with some code and a stack.
58 jermar 112
    While the code, implemented by a C language function, can be shared by
113
    several threads, the stack is always private to each instance of the
114
    thread. Each thread belongs to exactly one task through which it shares
115
    address space with its sibling threads. Threads that execute purely in the
116
    kernel don't have any userspace memory allocated. However, when a thread
117
    has ambitions to run in userspace, it must be allocated a userspace stack.
118
    The distinction between the purely kernel threads and threads running also
119
    in userspace is made by refering to the former group as to kernel threads
120
    and to the latter group as to userspace threads. Both kernel and userspace
121
    threads are visible to the scheduler and can become a subject of kernel
140 palkovsky 122
    preemption and thread migration anytime when preemption is not
123
    disabled.</para>
58 jermar 124
 
64 jermar 125
    <formalpara>
131 jermar 126
      <title>Thread States</title>
62 jermar 127
 
75 jermar 128
      <para>In each moment, a thread exists in one of six possible thread
140 palkovsky 129
      states. When the thread is created and first inserted into the
64 jermar 130
      scheduler's run queues or when a thread is migrated to a new processor,
140 palkovsky 131
      it is put into the <constant>Entering</constant> state. After some time
132
      elapses, the scheduler picks up the thread and starts executing it. A
133
      thread being currently executed on a processor is in the
75 jermar 134
      <constant>Running</constant> state. From there, the thread has three
135
      possibilities. It either runs until it is preemtped, in which case the
136
      state changes to <constant>Ready</constant>, goes to the
137
      <constant>Sleeping</constant> state by going to sleep or enters the
138
      <constant>Exiting</constant> state when it reaches termination. When the
139
      thread exits, its kernel structure usually stays in memory, until the
133 jermar 140
      thread is detached by another thread using <code>thread_detach()</code>
75 jermar 141
      function. Terminated but undetached threads are in the
171 jermar 142
      <constant>Lingering</constant> state. When the thread is detached or
75 jermar 143
      detaches itself during its life, it is destroyed in the
171 jermar 144
      <constant>Exiting</constant> state and the
145
      <constant>Lingering</constant> state is not reached.<figure float="1">
64 jermar 146
          <title>Transitions among thread states.</title>
61 jermar 147
 
64 jermar 148
          <mediaobject id="thread_states" xreflabel="">
87 bondari 149
            <imageobject role="pdf">
131 jermar 150
              <imagedata fileref="images/thread_states.pdf" format="PDF" />
77 bondari 151
            </imageobject>
152
 
64 jermar 153
            <imageobject role="html">
154
              <imagedata fileref="images/thread_states.png" format="PNG" />
155
            </imageobject>
62 jermar 156
 
64 jermar 157
            <imageobject role="fop">
131 jermar 158
              <imagedata fileref="images/thread_states.svg" format="SVG" />
64 jermar 159
            </imageobject>
160
          </mediaobject>
161
        </figure></para>
162
    </formalpara>
58 jermar 163
 
64 jermar 164
    <formalpara>
171 jermar 165
      <title>Fibrils</title>
58 jermar 166
 
64 jermar 167
      <para>HelenOS userspace layer knows even smaller units of execution.
171 jermar 168
      Each userspace thread can make use of an arbitrary number of fibrils.
169
      These fibrils have their own synchronous register context, userspace
170
      code and stack. They live their own life within the userspace thread and
171
      the scheduler does not have any idea about them because they are
172
      completely implemented by the userspace library. This implies several
173
      things:<itemizedlist>
64 jermar 174
          <listitem>
171 jermar 175
            <para>fibrils schedule themselves cooperatively within the time
176
            slice given to their userspace thread,</para>
64 jermar 177
          </listitem>
58 jermar 178
 
64 jermar 179
          <listitem>
171 jermar 180
            <para>fibrils share FPU context of their containing thread
64 jermar 181
            and</para>
182
          </listitem>
183
 
184
          <listitem>
171 jermar 185
            <para>all fibrils of one userspace thread block when one of them
186
            goes to sleep.</para>
64 jermar 187
          </listitem>
188
        </itemizedlist></para>
189
    </formalpara>
59 jermar 190
  </section>
58 jermar 191
 
59 jermar 192
  <section>
193
    <title>Scheduler</title>
194
 
195
    <section>
131 jermar 196
      <title>Run Queues</title>
59 jermar 197
 
198
      <para>There is an array of several run queues on each processor. The
199
      current version of HelenOS uses 16 run queues implemented by 16 doubly
200
      linked lists. Each of the run queues is associated with thread priority.
201
      The lower the run queue index in the array is, the higher is the
202
      priority of threads linked in that run queue and the shorter is the time
203
      in which those threads will execute. When kernel code wants to access
204
      the run queue, it must first acquire its lock.</para>
205
    </section>
206
 
207
    <section>
131 jermar 208
      <title>Scheduler Operation</title>
59 jermar 209
 
210
      <para>The scheduler is invoked either explicitly when a thread calls the
133 jermar 211
      <code>scheduler()</code> function (e.g. goes to sleep or merely wants to
59 jermar 212
      relinquish the processor for a while) or implicitly on a periodic basis
213
      when the generic clock interrupt preempts the current thread. After its
214
      invocation, the scheduler saves the synchronous register context of the
215
      current thread and switches to its private stack. Afterwards, a new
216
      thread is selected according to the scheduling policy. If there is no
217
      suitable thread, the processor is idle and no thread executes on it.
218
      Note that the act of switching to the private scheduler stack is
219
      essential. If the processor kept running using the stack of the
220
      preempted thread it could damage it because the old thread can be
221
      migrated to another processor and scheduled there. In the worst case
62 jermar 222
      scenario, two execution flows would be using the same stack.</para>
59 jermar 223
 
140 palkovsky 224
      <para>The scheduling policy is implemented in the function
159 jermar 225
      <code>find_best_thread()</code>. This function walks the processor run
59 jermar 226
      queues from lower towards higher indices and looks for a thread. If the
227
      visited run queue is empty, it simply searches the next run queue. If it
228
      is known in advance that there are no ready threads waiting for
133 jermar 229
      execution, <code>find_best_thread()</code> interruptibly halts the
59 jermar 230
      processor or busy waits until some threads arrive. This process repeats
133 jermar 231
      until <code>find_best_thread()</code> succeeds.</para>
59 jermar 232
 
233
      <para>After the best thread is chosen, the scheduler switches to the
234
      thread's task and memory management context. Finally, the saved
235
      synchronous register context is restored and the thread runs. Each
236
      scheduled thread is given a time slice depending on its priority (i.e.
237
      run queue). The higher priority, the shorter timeslice. To summarize,
238
      this policy schedules threads with high priorities more frequently but
239
      gives them smaller time slices. On the other hand, lower priority
240
      threads are scheduled less frequently, but run for longer periods of
241
      time.</para>
242
 
243
      <para>When a thread uses its entire time slice, it is preempted and put
244
      back into the run queue that immediately follows the previous run queue
245
      from which the thread ran. Threads that are woken up from a sleep are
246
      put into the biggest priority run queue. Low priority threads are
247
      therefore those that don't go to sleep so often and just occupy the
248
      processor.</para>
249
 
250
      <para>In order to avoid complete starvation of the low priority threads,
251
      from time to time, the scheduler will provide them with a bonus of one
252
      point priority increase. In other words, the scheduler will now and then
253
      move the entire run queues one level up.</para>
254
    </section>
255
 
256
    <section>
131 jermar 257
      <title>Processor Load Balancing</title>
59 jermar 258
 
259
      <para>Normally, for the sake of cache locality, threads are scheduled on
260
      one of the processors and don't leave it. Nevertheless, a situation in
261
      which one processor is heavily overloaded while others sit idle can
76 palkovsky 262
      occur. HelenOS deploys special kernel threads to help mitigate this
59 jermar 263
      problem. Each processor is associated with one load balancing thread
76 palkovsky 264
      called <code>kcpulb</code> that wakes up regularly to see whether its
265
      processor is underbalanced or not. If it is, the thread attempts to
59 jermar 266
      migrate threads from other overloaded processors to its own processor's
267
      run queues. When the job is done or there is no need for load balancing,
268
      the thread goes to sleep.</para>
269
 
270
      <para>The balancing threads operate very gently and try to migrate low
271
      priority threads first; one <code>kcpulb</code> never takes from one
272
      processor twice in a row. The load balancing threads as well as threads
273
      that were just stolen cannot be migrated. The <code>kcpulb</code>
274
      threads are wired to their processors and cannot be migrated whatsoever.
275
      The ordinary threads are protected only until they are
276
      rescheduled.</para>
277
    </section>
57 jermar 278
  </section>
279
</chapter>