= Migration = QEMU has code to load/save the state of the guest that it is running. These are two complementary operations. Saving the state just does that, saves the state for each device that the guest is running. Restoring a guest is just the opposite operation: we need to load the state of each device. For this to work, QEMU has to be launched with the same arguments the two times. I.e. it can only restore the state in one guest that has the same devices that the one it was saved (this last requirement can be relaxed a bit, but for now we can consider that configuration has to be exactly the same). Once that we are able to save/restore a guest, a new functionality is requested: migration. This means that QEMU is able to start in one machine and being "migrated" to another machine. I.e. being moved to another machine. Next was the "live migration" functionality. This is important because some guests run with a lot of state (specially RAM), and it can take a while to move all state from one machine to another. Live migration allows the guest to continue running while the state is transferred. Only while the last part of the state is transferred has the guest to be stopped. Typically the time that the guest is unresponsive during live migration is the low hundred of milliseconds (notice that this depends on a lot of things). === Types of migration === Now that we have talked about live migration, there are several ways to do migration: - tcp migration: do the migration using tcp sockets - unix migration: do the migration using unix sockets - exec migration: do the migration using the stdin/stdout through a process. - fd migration: do the migration using an file descriptor that is passed to QEMU. QEMU doesn't care how this file descriptor is opened. All these four migration protocols use the same infrastructure to save/restore state devices. This infrastructure is shared with the savevm/loadvm functionality. === State Live Migration === This is used for RAM and block devices. It is not yet ported to vmstate. <Fill more information here> === What is the common infrastructure === QEMU uses a QEMUFile abstraction to be able to do migration. Any type of migration that wants to use QEMU infrastructure has to create a QEMUFile with: QEMUFile *qemu_fopen_ops(void *opaque, QEMUFilePutBufferFunc *put_buffer, QEMUFileGetBufferFunc *get_buffer, QEMUFileCloseFunc *close); The functions have the following functionality: This function writes a chunk of data to a file at the given position. The pos argument can be ignored if the file is only used for streaming. The handler should try to write all of the data it can. typedef int (QEMUFilePutBufferFunc)(void *opaque, const uint8_t *buf, int64_t pos, int size); Read a chunk of data from a file at the given position. The pos argument can be ignored if the file is only be used for streaming. The number of bytes actually read should be returned. typedef int (QEMUFileGetBufferFunc)(void *opaque, uint8_t *buf, int64_t pos, int size); Close a file and return an error code. typedef int (QEMUFileCloseFunc)(void *opaque); You can use any internal state that you need using the opaque void * pointer that is passed to all functions. The important functions for us are put_buffer()/get_buffer() that allow to write/read a buffer into the QEMUFile. === How to save the state of one device === The state of a device is saved using intermediate buffers. There are some helper functions to assist this saving. There is a new concept that we have to explain here: device state version. When we migrate a device, we save/load the state as a series of fields. Some times, due to bugs or new functionality, we need to change the state to store more/different information. We use the version to identify each time that we do a change. Each version is associated with a series of fields saved. The save_state always saves the state as the newer version. But load_state sometimes is able to load state from an older version. === Legacy way === This way is going to disappear as soon as all current users are ported to VMSTATE. Each device has to register two functions, one to save the state and another to load the state back. int register_savevm(DeviceState *dev, const char *idstr, int instance_id, int version_id, SaveStateHandler *save_state, LoadStateHandler *load_state, void *opaque); typedef void SaveStateHandler(QEMUFile *f, void *opaque); typedef int LoadStateHandler(QEMUFile *f, void *opaque, int version_id); The important functions for the device state format are the save_state and load_state. Notice that load_state receives a version_id parameter to know what state format is receiving. save_state doesn't have a version_id parameter because it always uses the latest version. === VMState === The legacy way of saving/loading state of the device had the problem that we have to maintain two functions in sync. If we did one change in one of them and not in the other, we would get a failed migration. VMState changed the way that state is saved/loaded. Instead of using a function to save the state and another to load it, it was changed to a declarative way of what the state consisted of. Now VMState is able to interpret that definition to be able to load/save the state. As the state is declared only once, it can't go out of sync in the save/load functions. An example (from hw/input/pckbd.c) static const VMStateDescription vmstate_kbd = { .name = "pckbd", .version_id = 3, .minimum_version_id = 3, .fields = (VMStateField[]) { VMSTATE_UINT8(write_cmd, KBDState), VMSTATE_UINT8(status, KBDState), VMSTATE_UINT8(mode, KBDState), VMSTATE_UINT8(pending, KBDState), VMSTATE_END_OF_LIST() } }; We are declaring the state with name "pckbd". The version_id is 3, and the fields are 4 uint8_t in a KBDState structure. We registered this with: vmstate_register(NULL, 0, &vmstate_kbd, s); Note: talk about how vmstate <-> qdev interact, and what the instance ids mean. You can search for VMSTATE_* macros for lots of types used in QEMU in include/hw/hw.h. === More about versions === Version numbers are intended for major incompatible changes to the migration of a device, and using them breaks backwards-migration compatibility; in general most changes can be made by adding Subsections (see below) or _TEST macros (see below) which won't break compatibility. You can see that there are several version fields: - version_id: the maximum version_id supported by VMState for that device. - minimum_version_id: the minimum version_id that VMState is able to understand for that device. - minimum_version_id_old: For devices that were not able to port to vmstate, we can assign a function that knows how to read this old state. This field is ignored if there is no load_state_old handler. So, VMState is able to read versions from minimum_version_id to version_id. And the function load_state_old() (if present) is able to load state from minimum_version_id_old to minimum_version_id. This function is deprecated and will be removed when no more users are left. Saving state will always create a section with the 'version_id' value and thus can't be loaded by any older QEMU. === Massaging functions === Sometimes, it is not enough to be able to save the state directly from one structure, we need to fill the correct values there. One example is when we are using kvm. Before saving the cpu state, we need to ask kvm to copy to QEMU the state that it is using. And the opposite when we are loading the state, we need a way to tell kvm to load the state for the cpu that we have just loaded from the QEMUFile. The functions to do that are inside a vmstate definition, and are called: - int (*pre_load)(void *opaque); This function is called before we load the state of one device. - int (*post_load)(void *opaque, int version_id); This function is called after we load the state of one device. - void (*pre_save)(void *opaque); This function is called before we save the state of one device. Example: You can look at hpet.c, that uses the three function to massage the state that is transferred. If you use memory API functions that update memory layout outside initialization (i.e., in response to a guest action), this is a strong indication that you need to call these functions in a post_load callback. Examples of such memory API functions are: - memory_region_add_subregion() - memory_region_del_subregion() - memory_region_set_readonly() - memory_region_set_enabled() - memory_region_set_address() - memory_region_set_alias_offset() === Subsections === The use of version_id allows to be able to migrate from older versions to newer versions of a device. But not the other way around. This makes very complicated to fix bugs in stable branches. If we need to add anything to the state to fix a bug, we have to disable migration to older versions that don't have that bug-fix (i.e. a new field). But sometimes, that bug-fix is only needed sometimes, not always. For instance, if the device is in the middle of a DMA operation, it is using a specific functionality, .... It is impossible to create a way to make migration from any version to any other version to work. But we can do better than only allowing migration from older versions to newer ones. For that fields that are only needed sometimes, we add the idea of subsections. A subsection is "like" a device vmstate, but with a particularity, it has a Boolean function that tells if that values are needed to be sent or not. If this functions returns false, the subsection is not sent. On the receiving side, if we found a subsection for a device that we don't understand, we just fail the migration. If we understand all the subsections, then we load the state with success. One important note is that the post_load() function is called "after" loading all subsections, because a newer subsection could change same value that it uses. Example: static bool ide_drive_pio_state_needed(void *opaque) { IDEState *s = opaque; return ((s->status & DRQ_STAT) != 0) || (s->bus->error_status & BM_STATUS_PIO_RETRY); } const VMStateDescription vmstate_ide_drive_pio_state = { .name = "ide_drive/pio_state", .version_id = 1, .minimum_version_id = 1, .pre_save = ide_drive_pio_pre_save, .post_load = ide_drive_pio_post_load, .needed = ide_drive_pio_state_needed, .fields = (VMStateField[]) { VMSTATE_INT32(req_nb_sectors, IDEState), VMSTATE_VARRAY_INT32(io_buffer, IDEState, io_buffer_total_len, 1, vmstate_info_uint8, uint8_t), VMSTATE_INT32(cur_io_buffer_offset, IDEState), VMSTATE_INT32(cur_io_buffer_len, IDEState), VMSTATE_UINT8(end_transfer_fn_idx, IDEState), VMSTATE_INT32(elementary_transfer_size, IDEState), VMSTATE_INT32(packet_transfer_size, IDEState), VMSTATE_END_OF_LIST() } }; const VMStateDescription vmstate_ide_drive = { .name = "ide_drive", .version_id = 3, .minimum_version_id = 0, .post_load = ide_drive_post_load, .fields = (VMStateField[]) { .... several fields .... VMSTATE_END_OF_LIST() }, .subsections = (const VMStateDescription*[]) { &vmstate_ide_drive_pio_state, NULL } }; Here we have a subsection for the pio state. We only need to save/send this state when we are in the middle of a pio operation (that is what ide_drive_pio_state_needed() checks). If DRQ_STAT is not enabled, the values on that fields are garbage and don't need to be sent. Using a condition function that checks a 'property' to determine whether to send a subsection allows backwards migration compatibility when new subsections are added. For example; a) Add a new property using DEFINE_PROP_BOOL - e.g. support-foo and default it to true. b) Add an entry to the HW_COMPAT_ for the previous version that sets the property to false. c) Add a static bool support_foo function that tests the property. d) Add a subsection with a .needed set to the support_foo function e) (potentially) Add a pre_load that sets up a default value for 'foo' to be used if the subsection isn't loaded. Now that subsection will not be generated when using an older machine type and the migration stream will be accepted by older QEMU versions. pre-load functions can be used to initialise state on the newer version so that they default to suitable values when loading streams created by older QEMU versions that do not generate the subsection. In some cases subsections are added for data that had been accidentally omitted by earlier versions; if the missing data causes the migration process to succeed but the guest to behave badly then it may be better to send the subsection and cause the migration to explicitly fail with the unknown subsection error. If the bad behaviour only happens with certain data values, making the subsection conditional on the data value (rather than the machine type) allows migrations to succeed in most cases. In general the preference is to tie the subsection to the machine type, and allow reliable migrations, unless the behaviour from omission of the subsection is really bad. = Not sending existing elements = Sometimes members of the VMState are no longer needed; removing them will break migration compatibility making them version dependent and bumping the version will break backwards migration compatibility. The best way is to: a) Add a new property/compatibility/function in the same way for subsections above. b) replace the VMSTATE macro with the _TEST version of the macro, e.g.: VMSTATE_UINT32(foo, barstruct) becomes VMSTATE_UINT32_TEST(foo, barstruct, pre_version_baz) Sometime in the future when we no longer care about the ancient versions these can be killed off. = Return path = In most migration scenarios there is only a single data path that runs from the source VM to the destination, typically along a single fd (although possibly with another fd or similar for some fast way of throwing pages across). However, some uses need two way communication; in particular the Postcopy destination needs to be able to request pages on demand from the source. For these scenarios there is a 'return path' from the destination to the source; qemu_file_get_return_path(QEMUFile* fwdpath) gives the QEMUFile* for the return path. Source side Forward path - written by migration thread Return path - opened by main thread, read by return-path thread Destination side Forward path - read by main thread Return path - opened by main thread, written by main thread AND postcopy thread (protected by rp_mutex) = Postcopy = 'Postcopy' migration is a way to deal with migrations that refuse to converge (or take too long to converge) its plus side is that there is an upper bound on the amount of migration traffic and time it takes, the down side is that during the postcopy phase, a failure of *either* side or the network connection causes the guest to be lost. In postcopy the destination CPUs are started before all the memory has been transferred, and accesses to pages that are yet to be transferred cause a fault that's translated by QEMU into a request to the source QEMU. Postcopy can be combined with precopy (i.e. normal migration) so that if precopy doesn't finish in a given time the switch is made to postcopy. === Enabling postcopy === To enable postcopy, issue this command on the monitor prior to the start of migration: migrate_set_capability postcopy-ram on The normal commands are then used to start a migration, which is still started in precopy mode. Issuing: migrate_start_postcopy will now cause the transition from precopy to postcopy. It can be issued immediately after migration is started or any time later on. Issuing it after the end of a migration is harmless. Note: During the postcopy phase, the bandwidth limits set using migrate_set_speed is ignored (to avoid delaying requested pages that the destination is waiting for). === Postcopy device transfer === Loading of device data may cause the device emulation to access guest RAM that may trigger faults that have to be resolved by the source, as such the migration stream has to be able to respond with page data *during* the device load, and hence the device data has to be read from the stream completely before the device load begins to free the stream up. This is achieved by 'packaging' the device data into a blob that's read in one go. Source behaviour Until postcopy is entered the migration stream is identical to normal precopy, except for the addition of a 'postcopy advise' command at the beginning, to tell the destination that postcopy might happen. When postcopy starts the source sends the page discard data and then forms the 'package' containing: Command: 'postcopy listen' The device state A series of sections, identical to the precopy streams device state stream containing everything except postcopiable devices (i.e. RAM) Command: 'postcopy run' The 'package' is sent as the data part of a Command: 'CMD_PACKAGED', and the contents are formatted in the same way as the main migration stream. During postcopy the source scans the list of dirty pages and sends them to the destination without being requested (in much the same way as precopy), however when a page request is received from the destination, the dirty page scanning restarts from the requested location. This causes requested pages to be sent quickly, and also causes pages directly after the requested page to be sent quickly in the hope that those pages are likely to be used by the destination soon. Destination behaviour Initially the destination looks the same as precopy, with a single thread reading the migration stream; the 'postcopy advise' and 'discard' commands are processed to change the way RAM is managed, but don't affect the stream processing. ------------------------------------------------------------------------------ 1 2 3 4 5 6 7 main -----DISCARD-CMD_PACKAGED ( LISTEN DEVICE DEVICE DEVICE RUN ) thread | | | (page request) | \___ v \ listen thread: --- page -- page -- page -- page -- page -- a b c ------------------------------------------------------------------------------ On receipt of CMD_PACKAGED (1) All the data associated with the package - the ( ... ) section in the diagram - is read into memory, and the main thread recurses into qemu_loadvm_state_main to process the contents of the package (2) which contains commands (3,6) and devices (4...) On receipt of 'postcopy listen' - 3 -(i.e. the 1st command in the package) a new thread (a) is started that takes over servicing the migration stream, while the main thread carries on loading the package. It loads normal background page data (b) but if during a device load a fault happens (5) the returned page (c) is loaded by the listen thread allowing the main threads device load to carry on. The last thing in the CMD_PACKAGED is a 'RUN' command (6) letting the destination CPUs start running. At the end of the CMD_PACKAGED (7) the main thread returns to normal running behaviour and is no longer used by migration, while the listen thread carries on servicing page data until the end of migration. === Postcopy states === Postcopy moves through a series of states (see postcopy_state) from ADVISE->DISCARD->LISTEN->RUNNING->END Advise: Set at the start of migration if postcopy is enabled, even if it hasn't had the start command; here the destination checks that its OS has the support needed for postcopy, and performs setup to ensure the RAM mappings are suitable for later postcopy. The destination will fail early in migration at this point if the required OS support is not present. (Triggered by reception of POSTCOPY_ADVISE command) Discard: Entered on receipt of the first 'discard' command; prior to the first Discard being performed, hugepages are switched off (using madvise) to ensure that no new huge pages are created during the postcopy phase, and to cause any huge pages that have discards on them to be broken. Listen: The first command in the package, POSTCOPY_LISTEN, switches the destination state to Listen, and starts a new thread (the 'listen thread') which takes over the job of receiving pages off the migration stream, while the main thread carries on processing the blob. With this thread able to process page reception, the destination now 'sensitises' the RAM to detect any access to missing pages (on Linux using the 'userfault' system). Running: POSTCOPY_RUN causes the destination to synchronise all state and start the CPUs and IO devices running. The main thread now finishes processing the migration package and now carries on as it would for normal precopy migration (although it can't do the cleanup it would do as it finishes a normal migration). End: The listen thread can now quit, and perform the cleanup of migration state, the migration is now complete. === Source side page maps === The source side keeps two bitmaps during postcopy; 'the migration bitmap' and 'unsent map'. The 'migration bitmap' is basically the same as in the precopy case, and holds a bit to indicate that page is 'dirty' - i.e. needs sending. During the precopy phase this is updated as the CPU dirties pages, however during postcopy the CPUs are stopped and nothing should dirty anything any more. The 'unsent map' is used for the transition to postcopy. It is a bitmap that has a bit cleared whenever a page is sent to the destination, however during the transition to postcopy mode it is combined with the migration bitmap to form a set of pages that: a) Have been sent but then redirtied (which must be discarded) b) Have not yet been sent - which also must be discarded to cause any transparent huge pages built during precopy to be broken. Note that the contents of the unsentmap are sacrificed during the calculation of the discard set and thus aren't valid once in postcopy. The dirtymap is still valid and is used to ensure that no page is sent more than once. Any request for a page that has already been sent is ignored. Duplicate requests such as this can happen as a page is sent at about the same time the destination accesses it. === Postcopy with hugepages === Postcopy now works with hugetlbfs backed memory: a) The linux kernel on the destination must support userfault on hugepages. b) The huge-page configuration on the source and destination VMs must be identical; i.e. RAMBlocks on both sides must use the same page size. c) Note that -mem-path /dev/hugepages will fall back to allocating normal RAM if it doesn't have enough hugepages, triggering (b) to fail. Using -mem-prealloc enforces the allocation using hugepages. d) Care should be taken with the size of hugepage used; postcopy with 2MB hugepages works well, however 1GB hugepages are likely to be problematic since it takes ~1 second to transfer a 1GB hugepage across a 10Gbps link, and until the full page is transferred the destination thread is blocked.