1= Migration = 2 3QEMU has code to load/save the state of the guest that it is running. 4These are two complementary operations. Saving the state just does 5that, saves the state for each device that the guest is running. 6Restoring a guest is just the opposite operation: we need to load the 7state of each device. 8 9For this to work, QEMU has to be launched with the same arguments the 10two times. I.e. it can only restore the state in one guest that has 11the same devices that the one it was saved (this last requirement can 12be relaxed a bit, but for now we can consider that configuration has 13to be exactly the same). 14 15Once that we are able to save/restore a guest, a new functionality is 16requested: migration. This means that QEMU is able to start in one 17machine and being "migrated" to another machine. I.e. being moved to 18another machine. 19 20Next was the "live migration" functionality. This is important 21because some guests run with a lot of state (specially RAM), and it 22can take a while to move all state from one machine to another. Live 23migration allows the guest to continue running while the state is 24transferred. Only while the last part of the state is transferred has 25the guest to be stopped. Typically the time that the guest is 26unresponsive during live migration is the low hundred of milliseconds 27(notice that this depends on a lot of things). 28 29=== Types of migration === 30 31Now that we have talked about live migration, there are several ways 32to do migration: 33 34- tcp migration: do the migration using tcp sockets 35- unix migration: do the migration using unix sockets 36- exec migration: do the migration using the stdin/stdout through a process. 37- fd migration: do the migration using an file descriptor that is 38 passed to QEMU. QEMU doesn't care how this file descriptor is opened. 39 40All these four migration protocols use the same infrastructure to 41save/restore state devices. This infrastructure is shared with the 42savevm/loadvm functionality. 43 44=== State Live Migration === 45 46This is used for RAM and block devices. It is not yet ported to vmstate. 47<Fill more information here> 48 49=== What is the common infrastructure === 50 51QEMU uses a QEMUFile abstraction to be able to do migration. Any type 52of migration that wants to use QEMU infrastructure has to create a 53QEMUFile with: 54 55QEMUFile *qemu_fopen_ops(void *opaque, 56 QEMUFilePutBufferFunc *put_buffer, 57 QEMUFileGetBufferFunc *get_buffer, 58 QEMUFileCloseFunc *close); 59 60The functions have the following functionality: 61 62This function writes a chunk of data to a file at the given position. 63The pos argument can be ignored if the file is only used for 64streaming. The handler should try to write all of the data it can. 65 66typedef int (QEMUFilePutBufferFunc)(void *opaque, const uint8_t *buf, 67 int64_t pos, int size); 68 69Read a chunk of data from a file at the given position. The pos argument 70can be ignored if the file is only be used for streaming. The number of 71bytes actually read should be returned. 72 73typedef int (QEMUFileGetBufferFunc)(void *opaque, uint8_t *buf, 74 int64_t pos, int size); 75 76Close a file and return an error code. 77 78typedef int (QEMUFileCloseFunc)(void *opaque); 79 80You can use any internal state that you need using the opaque void * 81pointer that is passed to all functions. 82 83The important functions for us are put_buffer()/get_buffer() that 84allow to write/read a buffer into the QEMUFile. 85 86=== How to save the state of one device === 87 88The state of a device is saved using intermediate buffers. There are 89some helper functions to assist this saving. 90 91There is a new concept that we have to explain here: device state 92version. When we migrate a device, we save/load the state as a series 93of fields. Some times, due to bugs or new functionality, we need to 94change the state to store more/different information. We use the 95version to identify each time that we do a change. Each version is 96associated with a series of fields saved. The save_state always saves 97the state as the newer version. But load_state sometimes is able to 98load state from an older version. 99 100=== Legacy way === 101 102This way is going to disappear as soon as all current users are ported to VMSTATE. 103 104Each device has to register two functions, one to save the state and 105another to load the state back. 106 107int register_savevm(DeviceState *dev, 108 const char *idstr, 109 int instance_id, 110 int version_id, 111 SaveStateHandler *save_state, 112 LoadStateHandler *load_state, 113 void *opaque); 114 115typedef void SaveStateHandler(QEMUFile *f, void *opaque); 116typedef int LoadStateHandler(QEMUFile *f, void *opaque, int version_id); 117 118The important functions for the device state format are the save_state 119and load_state. Notice that load_state receives a version_id 120parameter to know what state format is receiving. save_state doesn't 121have a version_id parameter because it always uses the latest version. 122 123=== VMState === 124 125The legacy way of saving/loading state of the device had the problem 126that we have to maintain two functions in sync. If we did one change 127in one of them and not in the other, we would get a failed migration. 128 129VMState changed the way that state is saved/loaded. Instead of using 130a function to save the state and another to load it, it was changed to 131a declarative way of what the state consisted of. Now VMState is able 132to interpret that definition to be able to load/save the state. As 133the state is declared only once, it can't go out of sync in the 134save/load functions. 135 136An example (from hw/input/pckbd.c) 137 138static const VMStateDescription vmstate_kbd = { 139 .name = "pckbd", 140 .version_id = 3, 141 .minimum_version_id = 3, 142 .fields = (VMStateField[]) { 143 VMSTATE_UINT8(write_cmd, KBDState), 144 VMSTATE_UINT8(status, KBDState), 145 VMSTATE_UINT8(mode, KBDState), 146 VMSTATE_UINT8(pending, KBDState), 147 VMSTATE_END_OF_LIST() 148 } 149}; 150 151We are declaring the state with name "pckbd". 152The version_id is 3, and the fields are 4 uint8_t in a KBDState structure. 153We registered this with: 154 155 vmstate_register(NULL, 0, &vmstate_kbd, s); 156 157Note: talk about how vmstate <-> qdev interact, and what the instance ids mean. 158 159You can search for VMSTATE_* macros for lots of types used in QEMU in 160include/hw/hw.h. 161 162=== More about versions === 163 164You can see that there are several version fields: 165 166- version_id: the maximum version_id supported by VMState for that device. 167- minimum_version_id: the minimum version_id that VMState is able to understand 168 for that device. 169- minimum_version_id_old: For devices that were not able to port to vmstate, we can 170 assign a function that knows how to read this old state. This field is 171 ignored if there is no load_state_old handler. 172 173So, VMState is able to read versions from minimum_version_id to 174version_id. And the function load_state_old() (if present) is able to 175load state from minimum_version_id_old to minimum_version_id. This 176function is deprecated and will be removed when no more users are left. 177 178=== Massaging functions === 179 180Sometimes, it is not enough to be able to save the state directly 181from one structure, we need to fill the correct values there. One 182example is when we are using kvm. Before saving the cpu state, we 183need to ask kvm to copy to QEMU the state that it is using. And the 184opposite when we are loading the state, we need a way to tell kvm to 185load the state for the cpu that we have just loaded from the QEMUFile. 186 187The functions to do that are inside a vmstate definition, and are called: 188 189- int (*pre_load)(void *opaque); 190 191 This function is called before we load the state of one device. 192 193- int (*post_load)(void *opaque, int version_id); 194 195 This function is called after we load the state of one device. 196 197- void (*pre_save)(void *opaque); 198 199 This function is called before we save the state of one device. 200 201Example: You can look at hpet.c, that uses the three function to 202 massage the state that is transferred. 203 204If you use memory API functions that update memory layout outside 205initialization (i.e., in response to a guest action), this is a strong 206indication that you need to call these functions in a post_load callback. 207Examples of such memory API functions are: 208 209 - memory_region_add_subregion() 210 - memory_region_del_subregion() 211 - memory_region_set_readonly() 212 - memory_region_set_enabled() 213 - memory_region_set_address() 214 - memory_region_set_alias_offset() 215 216=== Subsections === 217 218The use of version_id allows to be able to migrate from older versions 219to newer versions of a device. But not the other way around. This 220makes very complicated to fix bugs in stable branches. If we need to 221add anything to the state to fix a bug, we have to disable migration 222to older versions that don't have that bug-fix (i.e. a new field). 223 224But sometimes, that bug-fix is only needed sometimes, not always. For 225instance, if the device is in the middle of a DMA operation, it is 226using a specific functionality, .... 227 228It is impossible to create a way to make migration from any version to 229any other version to work. But we can do better than only allowing 230migration from older versions to newer ones. For that fields that are 231only needed sometimes, we add the idea of subsections. A subsection 232is "like" a device vmstate, but with a particularity, it has a Boolean 233function that tells if that values are needed to be sent or not. If 234this functions returns false, the subsection is not sent. 235 236On the receiving side, if we found a subsection for a device that we 237don't understand, we just fail the migration. If we understand all 238the subsections, then we load the state with success. 239 240One important note is that the post_load() function is called "after" 241loading all subsections, because a newer subsection could change same 242value that it uses. 243 244Example: 245 246static bool ide_drive_pio_state_needed(void *opaque) 247{ 248 IDEState *s = opaque; 249 250 return ((s->status & DRQ_STAT) != 0) 251 || (s->bus->error_status & BM_STATUS_PIO_RETRY); 252} 253 254const VMStateDescription vmstate_ide_drive_pio_state = { 255 .name = "ide_drive/pio_state", 256 .version_id = 1, 257 .minimum_version_id = 1, 258 .pre_save = ide_drive_pio_pre_save, 259 .post_load = ide_drive_pio_post_load, 260 .needed = ide_drive_pio_state_needed, 261 .fields = (VMStateField[]) { 262 VMSTATE_INT32(req_nb_sectors, IDEState), 263 VMSTATE_VARRAY_INT32(io_buffer, IDEState, io_buffer_total_len, 1, 264 vmstate_info_uint8, uint8_t), 265 VMSTATE_INT32(cur_io_buffer_offset, IDEState), 266 VMSTATE_INT32(cur_io_buffer_len, IDEState), 267 VMSTATE_UINT8(end_transfer_fn_idx, IDEState), 268 VMSTATE_INT32(elementary_transfer_size, IDEState), 269 VMSTATE_INT32(packet_transfer_size, IDEState), 270 VMSTATE_END_OF_LIST() 271 } 272}; 273 274const VMStateDescription vmstate_ide_drive = { 275 .name = "ide_drive", 276 .version_id = 3, 277 .minimum_version_id = 0, 278 .post_load = ide_drive_post_load, 279 .fields = (VMStateField[]) { 280 .... several fields .... 281 VMSTATE_END_OF_LIST() 282 }, 283 .subsections = (const VMStateDescription*[]) { 284 &vmstate_ide_drive_pio_state, 285 NULL 286 } 287}; 288 289Here we have a subsection for the pio state. We only need to 290save/send this state when we are in the middle of a pio operation 291(that is what ide_drive_pio_state_needed() checks). If DRQ_STAT is 292not enabled, the values on that fields are garbage and don't need to 293be sent. 294 295= Return path = 296 297In most migration scenarios there is only a single data path that runs 298from the source VM to the destination, typically along a single fd (although 299possibly with another fd or similar for some fast way of throwing pages across). 300 301However, some uses need two way communication; in particular the Postcopy 302destination needs to be able to request pages on demand from the source. 303 304For these scenarios there is a 'return path' from the destination to the source; 305qemu_file_get_return_path(QEMUFile* fwdpath) gives the QEMUFile* for the return 306path. 307 308 Source side 309 Forward path - written by migration thread 310 Return path - opened by main thread, read by return-path thread 311 312 Destination side 313 Forward path - read by main thread 314 Return path - opened by main thread, written by main thread AND postcopy 315 thread (protected by rp_mutex) 316 317= Postcopy = 318'Postcopy' migration is a way to deal with migrations that refuse to converge 319(or take too long to converge) its plus side is that there is an upper bound on 320the amount of migration traffic and time it takes, the down side is that during 321the postcopy phase, a failure of *either* side or the network connection causes 322the guest to be lost. 323 324In postcopy the destination CPUs are started before all the memory has been 325transferred, and accesses to pages that are yet to be transferred cause 326a fault that's translated by QEMU into a request to the source QEMU. 327 328Postcopy can be combined with precopy (i.e. normal migration) so that if precopy 329doesn't finish in a given time the switch is made to postcopy. 330 331=== Enabling postcopy === 332 333To enable postcopy, issue this command on the monitor prior to the 334start of migration: 335 336migrate_set_capability postcopy-ram on 337 338The normal commands are then used to start a migration, which is still 339started in precopy mode. Issuing: 340 341migrate_start_postcopy 342 343will now cause the transition from precopy to postcopy. 344It can be issued immediately after migration is started or any 345time later on. Issuing it after the end of a migration is harmless. 346 347Note: During the postcopy phase, the bandwidth limits set using 348migrate_set_speed is ignored (to avoid delaying requested pages that 349the destination is waiting for). 350 351=== Postcopy device transfer === 352 353Loading of device data may cause the device emulation to access guest RAM 354that may trigger faults that have to be resolved by the source, as such 355the migration stream has to be able to respond with page data *during* the 356device load, and hence the device data has to be read from the stream completely 357before the device load begins to free the stream up. This is achieved by 358'packaging' the device data into a blob that's read in one go. 359 360Source behaviour 361 362Until postcopy is entered the migration stream is identical to normal 363precopy, except for the addition of a 'postcopy advise' command at 364the beginning, to tell the destination that postcopy might happen. 365When postcopy starts the source sends the page discard data and then 366forms the 'package' containing: 367 368 Command: 'postcopy listen' 369 The device state 370 A series of sections, identical to the precopy streams device state stream 371 containing everything except postcopiable devices (i.e. RAM) 372 Command: 'postcopy run' 373 374The 'package' is sent as the data part of a Command: 'CMD_PACKAGED', and the 375contents are formatted in the same way as the main migration stream. 376 377During postcopy the source scans the list of dirty pages and sends them 378to the destination without being requested (in much the same way as precopy), 379however when a page request is received from the destination, the dirty page 380scanning restarts from the requested location. This causes requested pages 381to be sent quickly, and also causes pages directly after the requested page 382to be sent quickly in the hope that those pages are likely to be used 383by the destination soon. 384 385Destination behaviour 386 387Initially the destination looks the same as precopy, with a single thread 388reading the migration stream; the 'postcopy advise' and 'discard' commands 389are processed to change the way RAM is managed, but don't affect the stream 390processing. 391 392------------------------------------------------------------------------------ 393 1 2 3 4 5 6 7 394main -----DISCARD-CMD_PACKAGED ( LISTEN DEVICE DEVICE DEVICE RUN ) 395thread | | 396 | (page request) 397 | \___ 398 v \ 399listen thread: --- page -- page -- page -- page -- page -- 400 401 a b c 402------------------------------------------------------------------------------ 403 404On receipt of CMD_PACKAGED (1) 405 All the data associated with the package - the ( ... ) section in the 406diagram - is read into memory (into a QEMUSizedBuffer), and the main thread 407recurses into qemu_loadvm_state_main to process the contents of the package (2) 408which contains commands (3,6) and devices (4...) 409 410On receipt of 'postcopy listen' - 3 -(i.e. the 1st command in the package) 411a new thread (a) is started that takes over servicing the migration stream, 412while the main thread carries on loading the package. It loads normal 413background page data (b) but if during a device load a fault happens (5) the 414returned page (c) is loaded by the listen thread allowing the main threads 415device load to carry on. 416 417The last thing in the CMD_PACKAGED is a 'RUN' command (6) letting the destination 418CPUs start running. 419At the end of the CMD_PACKAGED (7) the main thread returns to normal running behaviour 420and is no longer used by migration, while the listen thread carries 421on servicing page data until the end of migration. 422 423=== Postcopy states === 424 425Postcopy moves through a series of states (see postcopy_state) from 426ADVISE->DISCARD->LISTEN->RUNNING->END 427 428 Advise: Set at the start of migration if postcopy is enabled, even 429 if it hasn't had the start command; here the destination 430 checks that its OS has the support needed for postcopy, and performs 431 setup to ensure the RAM mappings are suitable for later postcopy. 432 The destination will fail early in migration at this point if the 433 required OS support is not present. 434 (Triggered by reception of POSTCOPY_ADVISE command) 435 436 Discard: Entered on receipt of the first 'discard' command; prior to 437 the first Discard being performed, hugepages are switched off 438 (using madvise) to ensure that no new huge pages are created 439 during the postcopy phase, and to cause any huge pages that 440 have discards on them to be broken. 441 442 Listen: The first command in the package, POSTCOPY_LISTEN, switches 443 the destination state to Listen, and starts a new thread 444 (the 'listen thread') which takes over the job of receiving 445 pages off the migration stream, while the main thread carries 446 on processing the blob. With this thread able to process page 447 reception, the destination now 'sensitises' the RAM to detect 448 any access to missing pages (on Linux using the 'userfault' 449 system). 450 451 Running: POSTCOPY_RUN causes the destination to synchronise all 452 state and start the CPUs and IO devices running. The main 453 thread now finishes processing the migration package and 454 now carries on as it would for normal precopy migration 455 (although it can't do the cleanup it would do as it 456 finishes a normal migration). 457 458 End: The listen thread can now quit, and perform the cleanup of migration 459 state, the migration is now complete. 460 461=== Source side page maps === 462 463The source side keeps two bitmaps during postcopy; 'the migration bitmap' 464and 'unsent map'. The 'migration bitmap' is basically the same as in 465the precopy case, and holds a bit to indicate that page is 'dirty' - 466i.e. needs sending. During the precopy phase this is updated as the CPU 467dirties pages, however during postcopy the CPUs are stopped and nothing 468should dirty anything any more. 469 470The 'unsent map' is used for the transition to postcopy. It is a bitmap that 471has a bit cleared whenever a page is sent to the destination, however during 472the transition to postcopy mode it is combined with the migration bitmap 473to form a set of pages that: 474 a) Have been sent but then redirtied (which must be discarded) 475 b) Have not yet been sent - which also must be discarded to cause any 476 transparent huge pages built during precopy to be broken. 477 478Note that the contents of the unsentmap are sacrificed during the calculation 479of the discard set and thus aren't valid once in postcopy. The dirtymap 480is still valid and is used to ensure that no page is sent more than once. Any 481request for a page that has already been sent is ignored. Duplicate requests 482such as this can happen as a page is sent at about the same time the 483destination accesses it. 484 485