5 QEMU has code to load/save the state of the guest that it is running.
6 These are two complementary operations. Saving the state just does
7 that, saves the state for each device that the guest is running.
8 Restoring a guest is just the opposite operation: we need to load the
11 For this to work, QEMU has to be launched with the same arguments the
12 two times. I.e. it can only restore the state in one guest that has
13 the same devices that the one it was saved (this last requirement can
14 be relaxed a bit, but for now we can consider that configuration has
15 to be exactly the same).
17 Once that we are able to save/restore a guest, a new functionality is
18 requested: migration. This means that QEMU is able to start in one
19 machine and being "migrated" to another machine. I.e. being moved to
22 Next was the "live migration" functionality. This is important
23 because some guests run with a lot of state (specially RAM), and it
24 can take a while to move all state from one machine to another. Live
25 migration allows the guest to continue running while the state is
26 transferred. Only while the last part of the state is transferred has
27 the guest to be stopped. Typically the time that the guest is
28 unresponsive during live migration is the low hundred of milliseconds
29 (notice that this depends on a lot of things).
34 The migration stream is normally just a byte stream that can be passed
37 - tcp migration: do the migration using tcp sockets
38 - unix migration: do the migration using unix sockets
39 - exec migration: do the migration using the stdin/stdout through a process.
40 - fd migration: do the migration using a file descriptor that is
41 passed to QEMU. QEMU doesn't care how this file descriptor is opened.
43 In addition, support is included for migration using RDMA, which
44 transports the page data using ``RDMA``, where the hardware takes care of
45 transporting the pages, and the load on the CPU is much lower. While the
46 internals of RDMA migration are a bit different, this isn't really visible
47 outside the RAM migration code.
49 All these migration protocols use the same infrastructure to
50 save/restore state devices. This infrastructure is shared with the
51 savevm/loadvm functionality.
56 The files, sockets or fd's that carry the migration stream are abstracted by
57 the ``QEMUFile`` type (see `migration/qemu-file.h`). In most cases this
58 is connected to a subtype of ``QIOChannel`` (see `io/`).
61 Saving the state of one device
62 ==============================
64 For most devices, the state is saved in a single call to the migration
65 infrastructure; these are *non-iterative* devices. The data for these
66 devices is sent at the end of precopy migration, when the CPUs are paused.
67 There are also *iterative* devices, which contain a very large amount of
68 data (e.g. RAM or large tables). See the iterative device section below.
70 General advice for device developers
71 ------------------------------------
73 - The migration state saved should reflect the device being modelled rather
74 than the way your implementation works. That way if you change the implementation
75 later the migration stream will stay compatible. That model may include
76 internal state that's not directly visible in a register.
78 - When saving a migration stream the device code may walk and check
79 the state of the device. These checks might fail in various ways (e.g.
80 discovering internal state is corrupt or that the guest has done something bad).
81 Consider carefully before asserting/aborting at this point, since the
82 normal response from users is that *migration broke their VM* since it had
83 apparently been running fine until then. In these error cases, the device
84 should log a message indicating the cause of error, and should consider
85 putting the device into an error state, allowing the rest of the VM to
88 - The migration might happen at an inconvenient point,
89 e.g. right in the middle of the guest reprogramming the device, during
90 guest reboot or shutdown or while the device is waiting for external IO.
91 It's strongly preferred that migrations do not fail in this situation,
92 since in the cloud environment migrations might happen automatically to
93 VMs that the administrator doesn't directly control.
95 - If you do need to fail a migration, ensure that sufficient information
96 is logged to identify what went wrong.
98 - The destination should treat an incoming migration stream as hostile
99 (which we do to varying degrees in the existing code). Check that offsets
100 into buffers and the like can't cause overruns. Fail the incoming migration
101 in the case of a corrupted stream like this.
103 - Take care with internal device state or behaviour that might become
104 migration version dependent. For example, the order of PCI capabilities
105 is required to stay constant across migration. Another example would
106 be that a special case handled by subsections (see below) might become
107 much more common if a default behaviour is changed.
109 - The state of the source should not be changed or destroyed by the
110 outgoing migration. Migrations timing out or being failed by
111 higher levels of management, or failures of the destination host are
112 not unusual, and in that case the VM is restarted on the source.
113 Note that the management layer can validly revert the migration
114 even though the QEMU level of migration has succeeded as long as it
115 does it before starting execution on the destination.
117 - Buses and devices should be able to explicitly specify addresses when
118 instantiated, and management tools should use those. For example,
119 when hot adding USB devices it's important to specify the ports
120 and addresses, since implicit ordering based on the command line order
121 may be different on the destination. This can result in the
122 device state being loaded into the wrong device.
127 Most device data can be described using the ``VMSTATE`` macros (mostly defined
128 in ``include/migration/vmstate.h``).
130 An example (from hw/input/pckbd.c)
134 static const VMStateDescription vmstate_kbd = {
137 .minimum_version_id = 3,
138 .fields = (VMStateField[]) {
139 VMSTATE_UINT8(write_cmd, KBDState),
140 VMSTATE_UINT8(status, KBDState),
141 VMSTATE_UINT8(mode, KBDState),
142 VMSTATE_UINT8(pending, KBDState),
143 VMSTATE_END_OF_LIST()
147 We are declaring the state with name "pckbd".
148 The `version_id` is 3, and the fields are 4 uint8_t in a KBDState structure.
149 We registered this with:
153 vmstate_register(NULL, 0, &vmstate_kbd, s);
155 For devices that are `qdev` based, we can register the device in the class
160 dc->vmsd = &vmstate_kbd_isa;
162 The VMState macros take care of ensuring that the device data section
163 is formatted portably (normally big endian) and make some compile time checks
164 against the types of the fields in the structures.
166 VMState macros can include other VMStateDescriptions to store substructures
167 (see ``VMSTATE_STRUCT_``), arrays (``VMSTATE_ARRAY_``) and variable length
168 arrays (``VMSTATE_VARRAY_``). Various other macros exist for special
171 Note that the format on the wire is still very raw; i.e. a VMSTATE_UINT32
172 ends up with a 4 byte bigendian representation on the wire; in the future
173 it might be possible to use a more structured format.
178 This way is going to disappear as soon as all current users are ported to VMSTATE;
179 although converting existing code can be tricky, and thus 'soon' is relative.
181 Each device has to register two functions, one to save the state and
182 another to load the state back.
186 int register_savevm_live(DeviceState *dev,
193 Two functions in the ``ops`` structure are the `save_state`
194 and `load_state` functions. Notice that `load_state` receives a version_id
195 parameter to know what state format is receiving. `save_state` doesn't
196 have a version_id parameter because it always uses the latest version.
198 Note that because the VMState macros still save the data in a raw
199 format, in many cases it's possible to replace legacy code
200 with a carefully constructed VMState description that matches the
201 byte layout of the existing code.
203 Changing migration data structures
204 ----------------------------------
206 When we migrate a device, we save/load the state as a series
207 of fields. Sometimes, due to bugs or new functionality, we need to
208 change the state to store more/different information. Changing the migration
209 state saved for a device can break migration compatibility unless
210 care is taken to use the appropriate techniques. In general QEMU tries
211 to maintain forward migration compatibility (i.e. migrating from
212 QEMU n->n+1) and there are users who benefit from backward compatibility
218 The most common structure change is adding new data, e.g. when adding
219 a newer form of device, or adding that state that you previously
220 forgot to migrate. This is best solved using a subsection.
222 A subsection is "like" a device vmstate, but with a particularity, it
223 has a Boolean function that tells if that values are needed to be sent
224 or not. If this functions returns false, the subsection is not sent.
225 Subsections have a unique name, that is looked for on the receiving
228 On the receiving side, if we found a subsection for a device that we
229 don't understand, we just fail the migration. If we understand all
230 the subsections, then we load the state with success. There's no check
231 that a subsection is loaded, so a newer QEMU that knows about a subsection
232 can (with care) load a stream from an older QEMU that didn't send
235 If the new data is only needed in a rare case, then the subsection
236 can be made conditional on that case and the migration will still
237 succeed to older QEMUs in most cases. This is OK for data that's
238 critical, but in some use cases it's preferred that the migration
239 should succeed even with the data missing. To support this the
240 subsection can be connected to a device property and from there
241 to a versioned machine type.
243 One important note is that the post_load() function is called "after"
244 loading all subsections, because a newer subsection could change same
245 value that it uses. A flag, and the combination of pre_load and post_load
246 can be used to detect whether a subsection was loaded, and to
247 fall back on default behaviour when the subsection isn't present.
253 static bool ide_drive_pio_state_needed(void *opaque)
255 IDEState *s = opaque;
257 return ((s->status & DRQ_STAT) != 0)
258 || (s->bus->error_status & BM_STATUS_PIO_RETRY);
261 const VMStateDescription vmstate_ide_drive_pio_state = {
262 .name = "ide_drive/pio_state",
264 .minimum_version_id = 1,
265 .pre_save = ide_drive_pio_pre_save,
266 .post_load = ide_drive_pio_post_load,
267 .needed = ide_drive_pio_state_needed,
268 .fields = (VMStateField[]) {
269 VMSTATE_INT32(req_nb_sectors, IDEState),
270 VMSTATE_VARRAY_INT32(io_buffer, IDEState, io_buffer_total_len, 1,
271 vmstate_info_uint8, uint8_t),
272 VMSTATE_INT32(cur_io_buffer_offset, IDEState),
273 VMSTATE_INT32(cur_io_buffer_len, IDEState),
274 VMSTATE_UINT8(end_transfer_fn_idx, IDEState),
275 VMSTATE_INT32(elementary_transfer_size, IDEState),
276 VMSTATE_INT32(packet_transfer_size, IDEState),
277 VMSTATE_END_OF_LIST()
281 const VMStateDescription vmstate_ide_drive = {
284 .minimum_version_id = 0,
285 .post_load = ide_drive_post_load,
286 .fields = (VMStateField[]) {
287 .... several fields ....
288 VMSTATE_END_OF_LIST()
290 .subsections = (const VMStateDescription*[]) {
291 &vmstate_ide_drive_pio_state,
296 Here we have a subsection for the pio state. We only need to
297 save/send this state when we are in the middle of a pio operation
298 (that is what ``ide_drive_pio_state_needed()`` checks). If DRQ_STAT is
299 not enabled, the values on that fields are garbage and don't need to
302 Connecting subsections to properties
303 ------------------------------------
305 Using a condition function that checks a 'property' to determine whether
306 to send a subsection allows backward migration compatibility when
307 new subsections are added, especially when combined with versioned
312 a) Add a new property using ``DEFINE_PROP_BOOL`` - e.g. support-foo and
314 b) Add an entry to the ``HW_COMPAT_`` for the previous version that sets
315 the property to false.
316 c) Add a static bool support_foo function that tests the property.
317 d) Add a subsection with a .needed set to the support_foo function
318 e) (potentially) Add a pre_load that sets up a default value for 'foo'
319 to be used if the subsection isn't loaded.
321 Now that subsection will not be generated when using an older
322 machine type and the migration stream will be accepted by older
325 Not sending existing elements
326 -----------------------------
328 Sometimes members of the VMState are no longer needed:
330 - removing them will break migration compatibility
332 - making them version dependent and bumping the version will break backward migration
335 Adding a dummy field into the migration stream is normally the best way to preserve
338 If the field really does need to be removed then:
340 a) Add a new property/compatibility/function in the same way for subsections above.
341 b) replace the VMSTATE macro with the _TEST version of the macro, e.g.:
343 ``VMSTATE_UINT32(foo, barstruct)``
347 ``VMSTATE_UINT32_TEST(foo, barstruct, pre_version_baz)``
349 Sometime in the future when we no longer care about the ancient versions these can be killed off.
350 Note that for backward compatibility it's important to fill in the structure with
351 data that the destination will understand.
353 Any difference in the predicates on the source and destination will end up
354 with different fields being enabled and data being loaded into the wrong
355 fields; for this reason conditional fields like this are very fragile.
360 Version numbers are intended for major incompatible changes to the
361 migration of a device, and using them breaks backward-migration
362 compatibility; in general most changes can be made by adding Subsections
363 (see above) or _TEST macros (see above) which won't break compatibility.
365 Each version is associated with a series of fields saved. The `save_state` always saves
366 the state as the newer version. But `load_state` sometimes is able to
367 load state from an older version.
369 You can see that there are several version fields:
371 - `version_id`: the maximum version_id supported by VMState for that device.
372 - `minimum_version_id`: the minimum version_id that VMState is able to understand
374 - `minimum_version_id_old`: For devices that were not able to port to vmstate, we can
375 assign a function that knows how to read this old state. This field is
376 ignored if there is no `load_state_old` handler.
378 VMState is able to read versions from minimum_version_id to
379 version_id. And the function ``load_state_old()`` (if present) is able to
380 load state from minimum_version_id_old to minimum_version_id. This
381 function is deprecated and will be removed when no more users are left.
383 There are *_V* forms of many ``VMSTATE_`` macros to load fields for version dependent fields,
388 VMSTATE_UINT16_V(ip_id, Slirp, 2),
390 only loads that field for versions 2 and newer.
392 Saving state will always create a section with the 'version_id' value
393 and thus can't be loaded by any older QEMU.
398 Sometimes, it is not enough to be able to save the state directly
399 from one structure, we need to fill the correct values there. One
400 example is when we are using kvm. Before saving the cpu state, we
401 need to ask kvm to copy to QEMU the state that it is using. And the
402 opposite when we are loading the state, we need a way to tell kvm to
403 load the state for the cpu that we have just loaded from the QEMUFile.
405 The functions to do that are inside a vmstate definition, and are called:
407 - ``int (*pre_load)(void *opaque);``
409 This function is called before we load the state of one device.
411 - ``int (*post_load)(void *opaque, int version_id);``
413 This function is called after we load the state of one device.
415 - ``int (*pre_save)(void *opaque);``
417 This function is called before we save the state of one device.
419 Example: You can look at hpet.c, that uses the three function to
420 massage the state that is transferred.
422 The ``VMSTATE_WITH_TMP`` macro may be useful when the migration
423 data doesn't match the stored device data well; it allows an
424 intermediate temporary structure to be populated with migration
425 data and then transferred to the main structure.
427 If you use memory API functions that update memory layout outside
428 initialization (i.e., in response to a guest action), this is a strong
429 indication that you need to call these functions in a `post_load` callback.
430 Examples of such memory API functions are:
432 - memory_region_add_subregion()
433 - memory_region_del_subregion()
434 - memory_region_set_readonly()
435 - memory_region_set_enabled()
436 - memory_region_set_address()
437 - memory_region_set_alias_offset()
439 Iterative device migration
440 --------------------------
442 Some devices, such as RAM, Block storage or certain platform devices,
443 have large amounts of data that would mean that the CPUs would be
444 paused for too long if they were sent in one section. For these
445 devices an *iterative* approach is taken.
447 The iterative devices generally don't use VMState macros
448 (although it may be possible in some cases) and instead use
449 qemu_put_*/qemu_get_* macros to read/write data to the stream. Specialist
450 versions exist for high bandwidth IO.
453 An iterative device must provide:
455 - A ``save_setup`` function that initialises the data structures and
456 transmits a first section containing information on the device. In the
457 case of RAM this transmits a list of RAMBlocks and sizes.
459 - A ``load_setup`` function that initialises the data structures on the
462 - A ``save_live_pending`` function that is called repeatedly and must
463 indicate how much more data the iterative data must save. The core
464 migration code will use this to determine when to pause the CPUs
465 and complete the migration.
467 - A ``save_live_iterate`` function (called after ``save_live_pending``
468 when there is significant data still to be sent). It should send
469 a chunk of data until the point that stream bandwidth limits tell it
470 to stop. Each call generates one section.
472 - A ``save_live_complete_precopy`` function that must transmit the
473 last section for the device containing any remaining data.
475 - A ``load_state`` function used to load sections generated by
476 any of the save functions that generate sections.
478 - ``cleanup`` functions for both save and load that are called
479 at the end of migration.
481 Note that the contents of the sections for iterative migration tend
482 to be open-coded by the devices; care should be taken in parsing
483 the results and structuring the stream to make them easy to validate.
488 There are cases in which the ordering of device loading matters; for
489 example in some systems where a device may assert an interrupt during loading,
490 if the interrupt controller is loaded later then it might lose the state.
492 Some ordering is implicitly provided by the order in which the machine
493 definition creates devices, however this is somewhat fragile.
495 The ``MigrationPriority`` enum provides a means of explicitly enforcing
496 ordering. Numerically higher priorities are loaded earlier.
497 The priority is set by setting the ``priority`` field of the top level
498 ``VMStateDescription`` for the device.
503 The stream tries to be word and endian agnostic, allowing migration between hosts
504 of different characteristics running the same VM.
510 - VM configuration section
515 Each section contains a device, or one iteration of a device save.
519 - ID string (First section of each device)
520 - instance id (First section of each device)
521 - version id (First section of each device)
525 - VM Description structure
526 Consisting of a JSON description of the contents for analysis only
528 The ``device data`` in each section consists of the data produced
529 by the code described above. For non-iterative devices they have a single
530 section; iterative devices have an initial and last section and a set
532 Note that there is very little checking by the common code of the integrity
533 of the ``device data`` contents, that's up to the devices themselves.
534 The ``footer mark`` provides a little bit of protection for the case where
535 the receiving side reads more or less data than expected.
537 The ``ID string`` is normally unique, having been formed from a bus name
538 and device address, PCI devices and storage devices hung off PCI controllers
539 fit this pattern well. Some devices are fixed single instances (e.g. "pc-ram").
540 Others (especially either older devices or system devices which for
541 some reason don't have a bus concept) make use of the ``instance id``
542 for otherwise identically named devices.
547 Only a unidirectional stream is required for normal migration, however a
548 ``return path`` can be created when bidirectional communication is desired.
549 This is primarily used by postcopy, but is also used to return a success
550 flag to the source at the end of migration.
552 ``qemu_file_get_return_path(QEMUFile* fwdpath)`` gives the QEMUFile* for the return
557 Forward path - written by migration thread
558 Return path - opened by main thread, read by return-path thread
562 Forward path - read by main thread
563 Return path - opened by main thread, written by main thread AND postcopy
564 thread (protected by rp_mutex)
569 'Postcopy' migration is a way to deal with migrations that refuse to converge
570 (or take too long to converge) its plus side is that there is an upper bound on
571 the amount of migration traffic and time it takes, the down side is that during
572 the postcopy phase, a failure of *either* side or the network connection causes
573 the guest to be lost.
575 In postcopy the destination CPUs are started before all the memory has been
576 transferred, and accesses to pages that are yet to be transferred cause
577 a fault that's translated by QEMU into a request to the source QEMU.
579 Postcopy can be combined with precopy (i.e. normal migration) so that if precopy
580 doesn't finish in a given time the switch is made to postcopy.
585 To enable postcopy, issue this command on the monitor (both source and
586 destination) prior to the start of migration:
588 ``migrate_set_capability postcopy-ram on``
590 The normal commands are then used to start a migration, which is still
591 started in precopy mode. Issuing:
593 ``migrate_start_postcopy``
595 will now cause the transition from precopy to postcopy.
596 It can be issued immediately after migration is started or any
597 time later on. Issuing it after the end of a migration is harmless.
599 Blocktime is a postcopy live migration metric, intended to show how
600 long the vCPU was in state of interruptable sleep due to pagefault.
601 That metric is calculated both for all vCPUs as overlapped value, and
602 separately for each vCPU. These values are calculated on destination
603 side. To enable postcopy blocktime calculation, enter following
604 command on destination monitor:
606 ``migrate_set_capability postcopy-blocktime on``
608 Postcopy blocktime can be retrieved by query-migrate qmp command.
609 postcopy-blocktime value of qmp command will show overlapped blocking
610 time for all vCPU, postcopy-vcpu-blocktime will show list of blocking
614 During the postcopy phase, the bandwidth limits set using
615 ``migrate_set_speed`` is ignored (to avoid delaying requested pages that
616 the destination is waiting for).
618 Postcopy device transfer
619 ------------------------
621 Loading of device data may cause the device emulation to access guest RAM
622 that may trigger faults that have to be resolved by the source, as such
623 the migration stream has to be able to respond with page data *during* the
624 device load, and hence the device data has to be read from the stream completely
625 before the device load begins to free the stream up. This is achieved by
626 'packaging' the device data into a blob that's read in one go.
631 Until postcopy is entered the migration stream is identical to normal
632 precopy, except for the addition of a 'postcopy advise' command at
633 the beginning, to tell the destination that postcopy might happen.
634 When postcopy starts the source sends the page discard data and then
635 forms the 'package' containing:
637 - Command: 'postcopy listen'
640 A series of sections, identical to the precopy streams device state stream
641 containing everything except postcopiable devices (i.e. RAM)
642 - Command: 'postcopy run'
644 The 'package' is sent as the data part of a Command: ``CMD_PACKAGED``, and the
645 contents are formatted in the same way as the main migration stream.
647 During postcopy the source scans the list of dirty pages and sends them
648 to the destination without being requested (in much the same way as precopy),
649 however when a page request is received from the destination, the dirty page
650 scanning restarts from the requested location. This causes requested pages
651 to be sent quickly, and also causes pages directly after the requested page
652 to be sent quickly in the hope that those pages are likely to be used
653 by the destination soon.
655 Destination behaviour
656 ---------------------
658 Initially the destination looks the same as precopy, with a single thread
659 reading the migration stream; the 'postcopy advise' and 'discard' commands
660 are processed to change the way RAM is managed, but don't affect the stream
665 ------------------------------------------------------------------------------
667 main -----DISCARD-CMD_PACKAGED ( LISTEN DEVICE DEVICE DEVICE RUN )
672 listen thread: --- page -- page -- page -- page -- page --
675 ------------------------------------------------------------------------------
677 - On receipt of ``CMD_PACKAGED`` (1)
679 All the data associated with the package - the ( ... ) section in the diagram -
680 is read into memory, and the main thread recurses into qemu_loadvm_state_main
681 to process the contents of the package (2) which contains commands (3,6) and
684 - On receipt of 'postcopy listen' - 3 -(i.e. the 1st command in the package)
686 a new thread (a) is started that takes over servicing the migration stream,
687 while the main thread carries on loading the package. It loads normal
688 background page data (b) but if during a device load a fault happens (5)
689 the returned page (c) is loaded by the listen thread allowing the main
690 threads device load to carry on.
692 - The last thing in the ``CMD_PACKAGED`` is a 'RUN' command (6)
694 letting the destination CPUs start running. At the end of the
695 ``CMD_PACKAGED`` (7) the main thread returns to normal running behaviour and
696 is no longer used by migration, while the listen thread carries on servicing
697 page data until the end of migration.
702 Postcopy moves through a series of states (see postcopy_state) from
703 ADVISE->DISCARD->LISTEN->RUNNING->END
707 Set at the start of migration if postcopy is enabled, even
708 if it hasn't had the start command; here the destination
709 checks that its OS has the support needed for postcopy, and performs
710 setup to ensure the RAM mappings are suitable for later postcopy.
711 The destination will fail early in migration at this point if the
712 required OS support is not present.
713 (Triggered by reception of POSTCOPY_ADVISE command)
717 Entered on receipt of the first 'discard' command; prior to
718 the first Discard being performed, hugepages are switched off
719 (using madvise) to ensure that no new huge pages are created
720 during the postcopy phase, and to cause any huge pages that
721 have discards on them to be broken.
725 The first command in the package, POSTCOPY_LISTEN, switches
726 the destination state to Listen, and starts a new thread
727 (the 'listen thread') which takes over the job of receiving
728 pages off the migration stream, while the main thread carries
729 on processing the blob. With this thread able to process page
730 reception, the destination now 'sensitises' the RAM to detect
731 any access to missing pages (on Linux using the 'userfault'
736 POSTCOPY_RUN causes the destination to synchronise all
737 state and start the CPUs and IO devices running. The main
738 thread now finishes processing the migration package and
739 now carries on as it would for normal precopy migration
740 (although it can't do the cleanup it would do as it
741 finishes a normal migration).
745 The listen thread can now quit, and perform the cleanup of migration
746 state, the migration is now complete.
748 Source side page maps
749 ---------------------
751 The source side keeps two bitmaps during postcopy; 'the migration bitmap'
752 and 'unsent map'. The 'migration bitmap' is basically the same as in
753 the precopy case, and holds a bit to indicate that page is 'dirty' -
754 i.e. needs sending. During the precopy phase this is updated as the CPU
755 dirties pages, however during postcopy the CPUs are stopped and nothing
756 should dirty anything any more.
758 The 'unsent map' is used for the transition to postcopy. It is a bitmap that
759 has a bit cleared whenever a page is sent to the destination, however during
760 the transition to postcopy mode it is combined with the migration bitmap
761 to form a set of pages that:
763 a) Have been sent but then redirtied (which must be discarded)
764 b) Have not yet been sent - which also must be discarded to cause any
765 transparent huge pages built during precopy to be broken.
767 Note that the contents of the unsentmap are sacrificed during the calculation
768 of the discard set and thus aren't valid once in postcopy. The dirtymap
769 is still valid and is used to ensure that no page is sent more than once. Any
770 request for a page that has already been sent is ignored. Duplicate requests
771 such as this can happen as a page is sent at about the same time the
772 destination accesses it.
774 Postcopy with hugepages
775 -----------------------
777 Postcopy now works with hugetlbfs backed memory:
779 a) The linux kernel on the destination must support userfault on hugepages.
780 b) The huge-page configuration on the source and destination VMs must be
781 identical; i.e. RAMBlocks on both sides must use the same page size.
782 c) Note that ``-mem-path /dev/hugepages`` will fall back to allocating normal
783 RAM if it doesn't have enough hugepages, triggering (b) to fail.
784 Using ``-mem-prealloc`` enforces the allocation using hugepages.
785 d) Care should be taken with the size of hugepage used; postcopy with 2MB
786 hugepages works well, however 1GB hugepages are likely to be problematic
787 since it takes ~1 second to transfer a 1GB hugepage across a 10Gbps link,
788 and until the full page is transferred the destination thread is blocked.
790 Postcopy with shared memory
791 ---------------------------
793 Postcopy migration with shared memory needs explicit support from the other
794 processes that share memory and from QEMU. There are restrictions on the type of
795 memory that userfault can support shared.
797 The Linux kernel userfault support works on `/dev/shm` memory and on `hugetlbfs`
798 (although the kernel doesn't provide an equivalent to `madvise(MADV_DONTNEED)`
799 for hugetlbfs which may be a problem in some configurations).
801 The vhost-user code in QEMU supports clients that have Postcopy support,
802 and the `vhost-user-bridge` (in `tests/`) and the DPDK package have changes
805 The client needs to open a userfaultfd and register the areas
806 of memory that it maps with userfault. The client must then pass the
807 userfaultfd back to QEMU together with a mapping table that allows
808 fault addresses in the clients address space to be converted back to
809 RAMBlock/offsets. The client's userfaultfd is added to the postcopy
810 fault-thread and page requests are made on behalf of the client by QEMU.
811 QEMU performs 'wake' operations on the client's userfaultfd to allow it
812 to continue after a page has arrived.
815 There are two future improvements that would be nice:
816 a) Some way to make QEMU ignorant of the addresses in the clients
818 b) Avoiding the need for QEMU to perform ufd-wake calls after the
821 Retro-fitting postcopy to existing clients is possible:
822 a) A mechanism is needed for the registration with userfault as above,
823 and the registration needs to be coordinated with the phases of
824 postcopy. In vhost-user extra messages are added to the existing
826 b) Any thread that can block due to guest memory accesses must be
827 identified and the implication understood; for example if the
828 guest memory access is made while holding a lock then all other
829 threads waiting for that lock will also be blocked.
834 Migration migrates the copies of RAM and ROM, and thus when running
835 on the destination it includes the firmware from the source. Even after
836 resetting a VM, the old firmware is used. Only once QEMU has been restarted
837 is the new firmware in use.
839 - Changes in firmware size can cause changes in the required RAMBlock size
840 to hold the firmware and thus migration can fail. In practice it's best
841 to pad firmware images to convenient powers of 2 with plenty of space
844 - Care should be taken with device emulation code so that newer
845 emulation code can work with older firmware to allow forward migration.
847 - Care should be taken with newer firmware so that backward migration
848 to older systems with older device emulation code will work.
850 In some cases it may be best to tie specific firmware versions to specific
851 versioned machine types to cut down on the combinations that will need
852 support. This is also useful when newer versions of firmware outgrow