drain.hh revision 11360
1/*
2 * Copyright (c) 2012, 2015 ARM Limited
3 * All rights reserved
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder.  You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Redistribution and use in source and binary forms, with or without
15 * modification, are permitted provided that the following conditions are
16 * met: redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer;
18 * redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution;
21 * neither the name of the copyright holders nor the names of its
22 * contributors may be used to endorse or promote products derived from
23 * this software without specific prior written permission.
24 *
25 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
26 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
27 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
28 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
29 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
30 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
31 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
32 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
33 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
34 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
35 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 *
37 * Authors: Andreas Sandberg
38 */
39
40#ifndef __SIM_DRAIN_HH__
41#define __SIM_DRAIN_HH__
42
43#include <atomic>
44#include <mutex>
45#include <unordered_set>
46
47#include "base/flags.hh"
48
49class Drainable;
50
51#ifndef SWIG // SWIG doesn't support strongly typed enums
52/**
53 * Object drain/handover states
54 *
55 * An object starts out in the Running state. When the simulator
56 * prepares to take a snapshot or prepares a CPU for handover, it
57 * calls the drain() method to transfer the object into the Draining
58 * or Drained state. If any object enters the Draining state
59 * (Drainable::drain() returning >0), simulation continues until it
60 * all objects have entered the Drained state.
61 *
62 * Before resuming simulation, the simulator calls resume() to
63 * transfer the object to the Running state.
64 *
65 * \note Even though the state of an object (visible to the rest of
66 * the world through Drainable::getState()) could be used to determine
67 * if all objects have entered the Drained state, the protocol is
68 * actually a bit more elaborate. See Drainable::drain() for details.
69 */
70enum class DrainState {
71    Running,  /** Running normally */
72    Draining, /** Draining buffers pending serialization/handover */
73    Drained   /** Buffers drained, ready for serialization/handover */
74};
75#endif
76
77/**
78 * This class coordinates draining of a System.
79 *
80 * When draining the simulator, we need to make sure that all
81 * Drainable objects within the system have ended up in the drained
82 * state before declaring the operation to be successful. This class
83 * keeps track of how many objects are still in the process of
84 * draining. Once it determines that all objects have drained their
85 * state, it exits the simulation loop.
86 *
87 * @note A System might not be completely drained even though the
88 * DrainManager has caused the simulation loop to exit. Draining needs
89 * to be restarted until all Drainable objects declare that they don't
90 * need further simulation to be completely drained. See Drainable for
91 * more information.
92 */
93class DrainManager
94{
95  private:
96    DrainManager();
97#ifndef SWIG
98    DrainManager(DrainManager &) = delete;
99#endif
100    ~DrainManager();
101
102  public:
103    /** Get the singleton DrainManager instance */
104    static DrainManager &instance() { return _instance; }
105
106    /**
107     * Try to drain the system.
108     *
109     * Try to drain the system and return true if all objects are in a
110     * the Drained state at which point the whole simulator is in a
111     * consistent state and ready for checkpointing or CPU
112     * handover. The simulation script must continue simulating until
113     * the simulation loop returns "Finished drain", at which point
114     * this method should be called again. This cycle should continue
115     * until this method returns true.
116     *
117     * @return true if all objects were drained successfully, false if
118     * more simulation is needed.
119     */
120    bool tryDrain();
121
122    /**
123     * Resume normal simulation in a Drained system.
124     */
125    void resume();
126
127    /**
128     * Run state fixups before a checkpoint restore operation
129     *
130     * The drain state of an object isn't stored in a checkpoint since
131     * the whole system is always going to be in the Drained state
132     * when the checkpoint is created. When the checkpoint is restored
133     * at a later stage, recreated objects will be in the Running
134     * state since the state isn't stored in checkpoints. This method
135     * performs state fixups on all Drainable objects and the
136     * DrainManager itself.
137     */
138    void preCheckpointRestore();
139
140    /** Check if the system is drained */
141    bool isDrained() const { return _state == DrainState::Drained; }
142
143    /** Get the simulators global drain state */
144    DrainState state() const { return _state; }
145
146    /**
147     * Notify the DrainManager that a Drainable object has finished
148     * draining.
149     */
150    void signalDrainDone();
151
152  public:
153    void registerDrainable(Drainable *obj);
154    void unregisterDrainable(Drainable *obj);
155
156  private:
157    /**
158     * Thread-safe helper function to get the number of Drainable
159     * objects in a system.
160     */
161    size_t drainableCount() const;
162
163    /** Lock protecting the set of drainable objects */
164    mutable std::mutex globalLock;
165
166    /** Set of all drainable objects */
167    std::unordered_set<Drainable *> _allDrainable;
168
169    /**
170     * Number of objects still draining. This is flagged atomic since
171     * it can be manipulated by SimObjects living in different
172     * threads.
173     */
174    std::atomic_uint _count;
175
176    /** Global simulator drain state */
177    DrainState _state;
178
179    /** Singleton instance of the drain manager */
180    static DrainManager _instance;
181};
182
183/**
184 * Interface for objects that might require draining before
185 * checkpointing.
186 *
187 * An object's internal state needs to be drained when creating a
188 * checkpoint, switching between CPU models, or switching between
189 * timing models. Once the internal state has been drained from
190 * <i>all</i> objects in the simulator, the objects are serialized to
191 * disc or the configuration change takes place. The process works as
192 * follows (see simulate.py for details):
193 *
194 * <ol>
195 * <li>DrainManager::tryDrain() calls Drainable::drain() for every
196 *     object in the system. Draining has completed if all of them
197 *     return true. Otherwise, the drain manager keeps track of the
198 *     objects that requested draining and waits for them to signal
199 *     that they are done draining using the signalDrainDone() method.
200 *
201 * <li>Continue simulation. When an object has finished draining its
202 *     internal state, it calls DrainManager::signalDrainDone() on the
203 *     manager. The drain manager keeps track of the objects that
204 *     haven't drained yet, simulation stops when the set of
205 *     non-drained objects becomes empty.
206 *
207 * <li>Check if any object still needs draining
208 *     (DrainManager::tryDrain()), if so repeat the process above.
209 *
210 * <li>Serialize objects, switch CPU model, or change timing model.
211 *
212 * <li>Call DrainManager::resume(), which in turn calls
213 *     Drainable::drainResume() for all objects, and then continue the
214 *     simulation.
215 * </ol>
216 *
217 */
218class Drainable
219{
220    friend class DrainManager;
221
222  protected:
223    Drainable();
224    virtual ~Drainable();
225
226    /**
227     * Notify an object that it needs to drain its state.
228     *
229     * If the object does not need further simulation to drain
230     * internal buffers, it returns DrainState::Drained and
231     * automatically switches to the Drained state. If the object
232     * needs more simulation, it returns DrainState::Draining and
233     * automatically enters the Draining state. Other return values
234     * are invalid.
235     *
236     * @note An object that has entered the Drained state can be
237     * disturbed by other objects in the system and consequently stop
238     * being drained. These perturbations are not visible in the drain
239     * state. The simulator therefore repeats the draining process
240     * until all objects return DrainState::Drained on the first call
241     * to drain().
242     *
243     * @return DrainState::Drained if the object is drained at this
244     * point in time, DrainState::Draining if it needs further
245     * simulation.
246     */
247    virtual DrainState drain() = 0;
248
249    /**
250     * Resume execution after a successful drain.
251     */
252    virtual void drainResume() {};
253
254    /**
255     * Signal that an object is drained
256     *
257     * This method is designed to be called whenever an object enters
258     * into a state where it is ready to be drained. The method is
259     * safe to call multiple times and there is no need to check that
260     * draining has been requested before calling this method.
261     */
262    void signalDrainDone() const {
263        switch (_drainState) {
264          case DrainState::Running:
265          case DrainState::Drained:
266            return;
267          case DrainState::Draining:
268            _drainState = DrainState::Drained;
269            _drainManager.signalDrainDone();
270            return;
271        }
272    }
273
274  public:
275    /** Return the current drain state of an object. */
276    DrainState drainState() const { return _drainState; }
277
278    /**
279     * Notify a child process of a fork.
280     *
281     * When calling fork in gem5, we need to ensure that resources
282     * shared between the parent and the child are consistent. This
283     * method is intended to be overloaded to handle that. For
284     * example, an object could use this method to re-open input files
285     * to get a separate file description with a private file offset.
286     *
287     * This method is only called in the child of the fork. The call
288     * takes place in a drained system.
289     */
290    virtual void notifyFork() {};
291
292  private:
293    /** DrainManager interface to request a drain operation */
294    DrainState dmDrain();
295    /** DrainManager interface to request a resume operation */
296    void dmDrainResume();
297
298    /** Convenience reference to the drain manager */
299    DrainManager &_drainManager;
300
301    /**
302     * Current drain state of the object. Needs to be mutable since
303     * objects need to be able to signal that they have transitioned
304     * into a Drained state even if the calling method is const.
305     */
306    mutable DrainState _drainState;
307};
308
309#endif
310