Subversion Repositories HelenOS-doc

Rev

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