1 Copyright (c) 2010-2015 Institute for System Programming
2 of the Russian Academy of Sciences.
4 This work is licensed under the terms of the GNU GPL, version 2 or later.
5 See the COPYING file in the top-level directory.
10 Record/replay functions are used for the deterministic replay of qemu execution.
11 Execution recording writes a non-deterministic events log, which can be later
12 used for replaying the execution anywhere and for unlimited number of times.
13 It also supports checkpointing for faster rewind to the specific replay moment.
14 Execution replaying reads the log and replays all non-deterministic events
15 including external input, hardware clocks, and interrupts.
17 Deterministic replay has the following features:
18 * Deterministically replays whole system execution and all contents of
19 the memory, state of the hardware devices, clocks, and screen of the VM.
20 * Writes execution log into the file for later replaying for multiple times
21 on different machines.
22 * Supports i386, x86_64, and ARM hardware platforms.
23 * Performs deterministic replay of all operations with keyboard and mouse
26 Usage of the record/replay:
27 * First, record the execution with the following command line:
29 -icount shift=7,rr=record,rrfile=replay.bin \
30 -drive file=disk.qcow2,if=none,snapshot,id=img-direct \
31 -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay \
32 -device ide-hd,drive=img-blkreplay \
33 -netdev user,id=net1 -device rtl8139,netdev=net1 \
34 -object filter-replay,id=replay,netdev=net1
35 * After recording, you can replay it by using another command line:
37 -icount shift=7,rr=replay,rrfile=replay.bin \
38 -drive file=disk.qcow2,if=none,snapshot,id=img-direct \
39 -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay \
40 -device ide-hd,drive=img-blkreplay \
41 -netdev user,id=net1 -device rtl8139,netdev=net1 \
42 -object filter-replay,id=replay,netdev=net1
43 The only difference with recording is changing the rr option
44 from record to replay.
45 * Block device images are not actually changed in the recording mode,
46 because all of the changes are written to the temporary overlay file.
47 This behavior is enabled by using blkreplay driver. It should be used
48 for every enabled block device, as described in 'Block devices' section.
49 * '-net none' option should be specified when network is not used,
50 because QEMU adds network card by default. When network is needed,
51 it should be configured explicitly with replay filter, as described
52 in 'Network devices' section.
53 * Interaction with audio devices and serial ports are recorded and replayed
54 automatically when such devices are enabled.
56 Academic papers with description of deterministic replay implementation:
57 http://www.computer.org/csdl/proceedings/csmr/2012/4666/00/4666a553-abs.html
58 http://dl.acm.org/citation.cfm?id=2786805.2803179
60 Modifications of qemu include:
61 * wrappers for clock and time functions to save their return values in the log
62 * saving different asynchronous events (e.g. system shutdown) into the log
63 * synchronization of the bottom halves execution
64 * synchronization of the threads from thread pool
65 * recording/replaying user input (mouse, keyboard, and microphone)
66 * adding internal checkpoints for cpu and io synchronization
67 * network filter for recording and replaying the packets
68 * block driver for making block layer deterministic
69 * serial port input record and replay
70 * recording of random numbers obtained from the external sources
72 Locking and thread synchronisation
73 ----------------------------------
75 Previously the synchronisation of the main thread and the vCPU thread
76 was ensured by the holding of the BQL. However the trend has been to
77 reduce the time the BQL was held across the system including under TCG
78 system emulation. As it is important that batches of events are kept
79 in sequence (e.g. expiring timers and checkpoints in the main thread
80 while instruction checkpoints are written by the vCPU thread) we need
81 another lock to keep things in lock-step. This role is now handled by
82 the replay_mutex_lock. It used to be held only for each event being
83 written but now it is held for a whole execution period. This results
84 in a deterministic ping-pong between the two main threads.
86 As the BQL is now a finer grained lock than the replay_lock it is almost
87 certainly a bug, and a source of deadlocks, to take the
88 replay_mutex_lock while the BQL is held. This is enforced by an assert.
89 While the unlocks are usually in the reverse order, this is not
90 necessary; you can drop the replay_lock while holding the BQL, without
91 doing a more complicated unlock_iothread/replay_unlock/lock_iothread
94 Non-deterministic events
95 ------------------------
97 Our record/replay system is based on saving and replaying non-deterministic
98 events (e.g. keyboard input) and simulating deterministic ones (e.g. reading
99 from HDD or memory of the VM). Saving only non-deterministic events makes
100 log file smaller and simulation faster.
102 The following non-deterministic data from peripheral devices is saved into
103 the log: mouse and keyboard input, network packets, audio controller input,
104 serial port input, and hardware clocks (they are non-deterministic
105 too, because their values are taken from the host machine). Inputs from
106 simulated hardware, memory of VM, software interrupts, and execution of
107 instructions are not saved into the log, because they are deterministic and
108 can be replayed by simulating the behavior of virtual machine starting from
111 We had to solve three tasks to implement deterministic replay: recording
112 non-deterministic events, replaying non-deterministic events, and checking
113 that there is no divergence between record and replay modes.
115 We changed several parts of QEMU to make event log recording and replaying.
116 Devices' models that have non-deterministic input from external devices were
117 changed to write every external event into the execution log immediately.
118 E.g. network packets are written into the log when they arrive into the virtual
121 All non-deterministic events are coming from these devices. But to
122 replay them we need to know at which moments they occur. We specify
123 these moments by counting the number of instructions executed between
124 every pair of consecutive events.
129 QEMU should work in icount mode to use record/replay feature. icount was
130 designed to allow deterministic execution in absence of external inputs
131 of the virtual machine. We also use icount to control the occurrence of the
132 non-deterministic events. The number of instructions elapsed from the last event
133 is written to the log while recording the execution. In replay mode we
134 can predict when to inject that event using the instruction counter.
139 Timers are used to execute callbacks from different subsystems of QEMU
140 at the specified moments of time. There are several kinds of timers:
141 * Real time clock. Based on host time and used only for callbacks that
142 do not change the virtual machine state. For this reason real time
143 clock and timers does not affect deterministic replay at all.
144 * Virtual clock. These timers run only during the emulation. In icount
145 mode virtual clock value is calculated using executed instructions counter.
146 That is why it is completely deterministic and does not have to be recorded.
147 * Host clock. This clock is used by device models that simulate real time
148 sources (e.g. real time clock chip). Host clock is the one of the sources
149 of non-determinism. Host clock read operations should be logged to
150 make the execution deterministic.
151 * Virtual real time clock. This clock is similar to real time clock but
152 it is used only for increasing virtual clock while virtual machine is
153 sleeping. Due to its nature it is also non-deterministic as the host clock
154 and has to be logged too.
159 Replaying of the execution of virtual machine is bound by sources of
160 non-determinism. These are inputs from clock and peripheral devices,
161 and QEMU thread scheduling. Thread scheduling affect on processing events
162 from timers, asynchronous input-output, and bottom halves.
164 Invocations of timers are coupled with clock reads and changing the state
165 of the virtual machine. Reads produce non-deterministic data taken from
166 host clock. And VM state changes should preserve their order. Their relative
167 order in replay mode must replicate the order of callbacks in record mode.
168 To preserve this order we use checkpoints. When a specific clock is processed
169 in record mode we save to the log special "checkpoint" event.
170 Checkpoints here do not refer to virtual machine snapshots. They are just
171 record/replay events used for synchronization.
173 QEMU in replay mode will try to invoke timers processing in random moment
174 of time. That's why we do not process a group of timers until the checkpoint
175 event will be read from the log. Such an event allows synchronizing CPU
176 execution and timer events.
178 Two other checkpoints govern the "warping" of the virtual clock.
179 While the virtual machine is idle, the virtual clock increments at
180 1 ns per *real time* nanosecond. This is done by setting up a timer
181 (called the warp timer) on the virtual real time clock, so that the
182 timer fires at the next deadline of the virtual clock; the virtual clock
183 is then incremented (which is called "warping" the virtual clock) as
184 soon as the timer fires or the CPUs need to go out of the idle state.
185 Two functions are used for this purpose; because these actions change
186 virtual machine state and must be deterministic, each of them creates a
187 checkpoint. qemu_start_warp_timer checks if the CPUs are idle and if so
188 starts accounting real time to virtual clock. qemu_account_warp_timer
189 is called when the CPUs get an interrupt or when the warp timer fires,
190 and it warps the virtual clock by the amount of real time that has passed
191 since qemu_start_warp_timer.
196 Disk I/O events are completely deterministic in our model, because
197 in both record and replay modes we start virtual machine from the same
198 disk state. But callbacks that virtual disk controller uses for reading and
199 writing the disk may occur at different moments of time in record and replay
202 Reading and writing requests are created by CPU thread of QEMU. Later these
203 requests proceed to block layer which creates "bottom halves". Bottom
204 halves consist of callback and its parameters. They are processed when
205 main loop locks the global mutex. These locks are not synchronized with
206 replaying process because main loop also processes the events that do not
207 affect the virtual machine state (like user interaction with monitor).
209 That is why we had to implement saving and replaying bottom halves callbacks
210 synchronously to the CPU execution. When the callback is about to execute
211 it is added to the queue in the replay module. This queue is written to the
212 log when its callbacks are executed. In replay mode callbacks are not processed
213 until the corresponding event is read from the events log file.
215 Sometimes the block layer uses asynchronous callbacks for its internal purposes
216 (like reading or writing VM snapshots or disk image cluster tables). In this
217 case bottom halves are not marked as "replayable" and do not saved
223 Block devices record/replay module intercepts calls of
224 bdrv coroutine functions at the top of block drivers stack.
225 To record and replay block operations the drive must be configured
227 -drive file=disk.qcow2,if=none,snapshot,id=img-direct
228 -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay
229 -device ide-hd,drive=img-blkreplay
231 blkreplay driver should be inserted between disk image and virtual driver
232 controller. Therefore all disk requests may be recorded and replayed.
234 All block completion operations are added to the queue in the coroutines.
235 Queue is flushed at checkpoints and information about processed requests
236 is recorded to the log. In replay phase the queue is matched with
237 events read from the log. Therefore block devices requests are processed
243 New VM snapshots may be created in replay mode. They can be used later
244 to recover the desired VM state. All VM states created in replay mode
245 are associated with the moment of time in the replay scenario.
246 After recovering the VM state replay will start from that position.
248 Default starting snapshot name may be specified with icount field
249 rrsnapshot as follows:
250 -icount shift=7,rr=record,rrfile=replay.bin,rrsnapshot=snapshot_name
252 This snapshot is created at start of recording and restored at start
253 of replaying. It also can be loaded while replaying to roll back
256 'snapshot' flag of the disk image must be removed to save the snapshots
257 in the overlay (or original image) instead of using the temporary overlay.
258 -drive file=disk.ovl,if=none,id=img-direct
259 -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay
260 -device ide-hd,drive=img-blkreplay
262 Use QEMU monitor to create additional snapshots. 'savevm <name>' command
263 created the snapshot and 'loadvm <name>' restores it. To prevent corruption
264 of the original disk image, use overlay files linked to the original images.
265 Therefore all new snapshots (including the starting one) will be saved in
266 overlays and the original image remains unchanged.
271 Record and replay for network interactions is performed with the network filter.
272 Each backend must have its own instance of the replay filter as follows:
273 -netdev user,id=net1 -device rtl8139,netdev=net1
274 -object filter-replay,id=replay,netdev=net1
276 Replay network filter is used to record and replay network packets. While
277 recording the virtual machine this filter puts all packets coming from
278 the outer world into the log. In replay mode packets from the log are
279 injected into the network device. All interactions with network backend
280 in replay mode are disabled.
285 Audio data is recorded and replay automatically. The command line for recording
286 and replaying must contain identical specifications of audio hardware, e.g.:
292 Serial ports input is recorded and replay automatically. The command lines
293 for recording and replaying must contain identical number of ports in record
294 and replay modes, but their backends may differ.
295 E.g., '-serial stdio' in record mode, and '-serial null' in replay mode.
300 Record/replay log consists of the header and the sequence of execution
301 events. The header includes 4-byte replay version id and 8-byte reserved
302 field. Version is updated every time replay log format changes to prevent
303 using replay log created by another build of qemu.
305 The sequence of the events describes virtual machine state changes.
306 It includes all non-deterministic inputs of VM, synchronization marks and
307 instruction counts used to correctly inject inputs at replay.
309 Synchronization marks (checkpoints) are used for synchronizing qemu threads
310 that perform operations with virtual hardware. These operations may change
311 system's state (e.g., change some register or generate interrupt) and
312 therefore should execute synchronously with CPU thread.
314 Every event in the log includes 1-byte event id and optional arguments.
315 When argument is an array, it is stored as 4-byte array length
316 and corresponding number of bytes with data.
317 Here is the list of events that are written into the log:
319 - EVENT_INSTRUCTION. Instructions executed since last event.
320 Argument: 4-byte number of executed instructions.
321 - EVENT_INTERRUPT. Used to synchronize interrupt processing.
322 - EVENT_EXCEPTION. Used to synchronize exception handling.
323 - EVENT_ASYNC. This is a group of events. They are always processed
324 together with checkpoints. When such an event is generated, it is
325 stored in the queue and processed only when checkpoint occurs.
326 Every such event is followed by 1-byte checkpoint id and 1-byte
327 async event id from the following list:
328 - REPLAY_ASYNC_EVENT_BH. Bottom-half callback. This event synchronizes
329 callbacks that affect virtual machine state, but normally called
331 Argument: 8-byte operation id.
332 - REPLAY_ASYNC_EVENT_INPUT. Input device event. Contains
333 parameters of keyboard and mouse input operations
334 (key press/release, mouse pointer movement).
335 Arguments: 9-16 bytes depending of input event.
336 - REPLAY_ASYNC_EVENT_INPUT_SYNC. Internal input synchronization event.
337 - REPLAY_ASYNC_EVENT_CHAR_READ. Character (e.g., serial port) device input
338 initiated by the sender.
339 Arguments: 1-byte character device id.
340 Array with bytes were read.
341 - REPLAY_ASYNC_EVENT_BLOCK. Block device operation. Used to synchronize
342 operations with disk and flash drives with CPU.
343 Argument: 8-byte operation id.
344 - REPLAY_ASYNC_EVENT_NET. Incoming network packet.
345 Arguments: 1-byte network adapter id.
347 Array with packet bytes.
348 - EVENT_SHUTDOWN. Occurs when user sends shutdown event to qemu,
349 e.g., by closing the window.
350 - EVENT_CHAR_WRITE. Used to synchronize character output operations.
351 Arguments: 4-byte output function return value.
352 4-byte offset in the output array.
353 - EVENT_CHAR_READ_ALL. Used to synchronize character input operations,
355 Argument: Array with bytes that were read.
356 - EVENT_CHAR_READ_ALL_ERROR. Unsuccessful character input operation,
358 Argument: 4-byte error code.
359 - EVENT_CLOCK + clock_id. Group of events for host clock read operations.
360 Argument: 8-byte clock value.
361 - EVENT_CHECKPOINT + checkpoint_id. Checkpoint for synchronization of
362 CPU, internal threads, and asynchronous input events. May be followed
363 by one or more EVENT_ASYNC events.
364 - EVENT_END. Last event in the log.