1/*
2 * Copyright (c) 2012, 2015, 2017 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 <vector>
46
47class Drainable;
48
49/**
50 * Object drain/handover states
51 *
52 * An object starts out in the Running state. When the simulator
53 * prepares to take a snapshot or prepares a CPU for handover, it
54 * calls the drain() method to transfer the object into the Draining
55 * or Drained state. If any object enters the Draining state
56 * (Drainable::drain() returning >0), simulation continues until it
57 * all objects have entered the Drained state.
58 *
59 * Before resuming simulation, the simulator calls resume() to
60 * transfer the object to the Running state. This in turn results in a
61 * call to drainResume() for all Drainable objects in the
62 * simulator. New Drainable objects may be created while resuming. In
63 * such cases, the new objects will be created in the Resuming state
64 * and later resumed.
65 *
66 * \note Even though the state of an object (visible to the rest of
67 * the world through Drainable::getState()) could be used to determine
68 * if all objects have entered the Drained state, the protocol is
69 * actually a bit more elaborate. See Drainable::drain() for details.
70 */
71enum class DrainState {
72    Running,  /** Running normally */
73    Draining, /** Draining buffers pending serialization/handover */
74    Drained,  /** Buffers drained, ready for serialization/handover */
75    Resuming, /** Transient state while the simulator is resuming */
76};
77
78/**
79 * This class coordinates draining of a System.
80 *
81 * When draining the simulator, we need to make sure that all
82 * Drainable objects within the system have ended up in the drained
83 * state before declaring the operation to be successful. This class
84 * keeps track of how many objects are still in the process of
85 * draining. Once it determines that all objects have drained their
86 * state, it exits the simulation loop.
87 *
88 * @note A System might not be completely drained even though the
89 * DrainManager has caused the simulation loop to exit. Draining needs
90 * to be restarted until all Drainable objects declare that they don't
91 * need further simulation to be completely drained. See Drainable for
92 * more information.
93 */
94class DrainManager
95{
96  private:
97    DrainManager();
98    DrainManager(DrainManager &) = delete;
99    ~DrainManager();
100
101  public:
102    /** Get the singleton DrainManager instance */
103    static DrainManager &instance() { return _instance; }
104
105    /**
106     * Try to drain the system.
107     *
108     * Try to drain the system and return true if all objects are in a
109     * the Drained state at which point the whole simulator is in a
110     * consistent state and ready for checkpointing or CPU
111     * handover. The simulation script must continue simulating until
112     * the simulation loop returns "Finished drain", at which point
113     * this method should be called again. This cycle should continue
114     * until this method returns true.
115     *
116     * @return true if all objects were drained successfully, false if
117     * more simulation is needed.
118     */
119    bool tryDrain();
120
121    /**
122     * Resume normal simulation in a Drained system.
123     */
124    void resume();
125
126    /**
127     * Run state fixups before a checkpoint restore operation
128     *
129     * The drain state of an object isn't stored in a checkpoint since
130     * the whole system is always going to be in the Drained state
131     * when the checkpoint is created. When the checkpoint is restored
132     * at a later stage, recreated objects will be in the Running
133     * state since the state isn't stored in checkpoints. This method
134     * performs state fixups on all Drainable objects and the
135     * DrainManager itself.
136     */
137    void preCheckpointRestore();
138
139    /** Check if the system is drained */
140    bool isDrained() const { return _state == DrainState::Drained; }
141
142    /** Get the simulators global drain state */
143    DrainState state() const { return _state; }
144
145    /**
146     * Notify the DrainManager that a Drainable object has finished
147     * draining.
148     */
149    void signalDrainDone();
150
151  public:
152    void registerDrainable(Drainable *obj);
153    void unregisterDrainable(Drainable *obj);
154
155  private:
156    /**
157     * Helper function to check if all Drainable objects are in a
158     * specific state.
159     */
160    bool allInState(DrainState state) const;
161
162    /**
163     * Thread-safe helper function to get the number of Drainable
164     * objects in a system.
165     */
166    size_t drainableCount() const;
167
168    /** Lock protecting the set of drainable objects */
169    mutable std::mutex globalLock;
170
171    /** Set of all drainable objects */
172    std::vector<Drainable *> _allDrainable;
173
174    /**
175     * Number of objects still draining. This is flagged atomic since
176     * it can be manipulated by SimObjects living in different
177     * threads.
178     */
179    std::atomic_uint _count;
180
181    /** Global simulator drain state */
182    DrainState _state;
183
184    /** Singleton instance of the drain manager */
185    static DrainManager _instance;
186};
187
188/**
189 * Interface for objects that might require draining before
190 * checkpointing.
191 *
192 * An object's internal state needs to be drained when creating a
193 * checkpoint, switching between CPU models, or switching between
194 * timing models. Once the internal state has been drained from
195 * <i>all</i> objects in the simulator, the objects are serialized to
196 * disc or the configuration change takes place. The process works as
197 * follows (see simulate.py for details):
198 *
199 * <ol>
200 * <li>DrainManager::tryDrain() calls Drainable::drain() for every
201 *     object in the system. Draining has completed if all of them
202 *     return true. Otherwise, the drain manager keeps track of the
203 *     objects that requested draining and waits for them to signal
204 *     that they are done draining using the signalDrainDone() method.
205 *
206 * <li>Continue simulation. When an object has finished draining its
207 *     internal state, it calls DrainManager::signalDrainDone() on the
208 *     manager. The drain manager keeps track of the objects that
209 *     haven't drained yet, simulation stops when the set of
210 *     non-drained objects becomes empty.
211 *
212 * <li>Check if any object still needs draining
213 *     (DrainManager::tryDrain()), if so repeat the process above.
214 *
215 * <li>Serialize objects, switch CPU model, or change timing model.
216 *
217 * <li>Call DrainManager::resume(), which in turn calls
218 *     Drainable::drainResume() for all objects, and then continue the
219 *     simulation.
220 * </ol>
221 *
222 */
223class Drainable
224{
225    friend class DrainManager;
226
227  protected:
228    Drainable();
229    virtual ~Drainable();
230
231    /**
232     * Notify an object that it needs to drain its state.
233     *
234     * If the object does not need further simulation to drain
235     * internal buffers, it returns DrainState::Drained and
236     * automatically switches to the Drained state. If the object
237     * needs more simulation, it returns DrainState::Draining and
238     * automatically enters the Draining state. Other return values
239     * are invalid.
240     *
241     * @note An object that has entered the Drained state can be
242     * disturbed by other objects in the system and consequently stop
243     * being drained. These perturbations are not visible in the drain
244     * state. The simulator therefore repeats the draining process
245     * until all objects return DrainState::Drained on the first call
246     * to drain().
247     *
248     * @return DrainState::Drained if the object is drained at this
249     * point in time, DrainState::Draining if it needs further
250     * simulation.
251     */
252    virtual DrainState drain() = 0;
253
254    /**
255     * Resume execution after a successful drain.
256     */
257    virtual void drainResume() {};
258
259    /**
260     * Signal that an object is drained
261     *
262     * This method is designed to be called whenever an object enters
263     * into a state where it is ready to be drained. The method is
264     * safe to call multiple times and there is no need to check that
265     * draining has been requested before calling this method.
266     */
267    void signalDrainDone() const {
268        switch (_drainState) {
269          case DrainState::Running:
270          case DrainState::Drained:
271          case DrainState::Resuming:
272            return;
273          case DrainState::Draining:
274            _drainState = DrainState::Drained;
275            _drainManager.signalDrainDone();
276            return;
277        }
278    }
279
280  public:
281    /** Return the current drain state of an object. */
282    DrainState drainState() const { return _drainState; }
283
284    /**
285     * Notify a child process of a fork.
286     *
287     * When calling fork in gem5, we need to ensure that resources
288     * shared between the parent and the child are consistent. This
289     * method is intended to be overloaded to handle that. For
290     * example, an object could use this method to re-open input files
291     * to get a separate file description with a private file offset.
292     *
293     * This method is only called in the child of the fork. The call
294     * takes place in a drained system.
295     */
296    virtual void notifyFork() {};
297
298  private:
299    /** DrainManager interface to request a drain operation */
300    DrainState dmDrain();
301    /** DrainManager interface to request a resume operation */
302    void dmDrainResume();
303
304    /** Convenience reference to the drain manager */
305    DrainManager &_drainManager;
306
307    /**
308     * Current drain state of the object. Needs to be mutable since
309     * objects need to be able to signal that they have transitioned
310     * into a Drained state even if the calling method is const.
311     */
312    mutable DrainState _drainState;
313};
314
315#endif
316