4 * Copyright IBM, Corp. 2008
7 * Anthony Liguori <aliguori@us.ibm.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Contributions after 2012-01-13 are licensed under the terms of the
13 * GNU GPL, version 2 or (at your option) any later version.
16 #include "qemu/osdep.h"
17 #include "qemu/cutils.h"
18 #include "qemu/error-report.h"
19 #include "qemu/main-loop.h"
20 #include "migration/blocker.h"
25 #include "sysemu/runstate.h"
26 #include "sysemu/sysemu.h"
27 #include "sysemu/cpu-throttle.h"
30 #include "ram-compress.h"
31 #include "migration/global_state.h"
32 #include "migration/misc.h"
33 #include "migration.h"
34 #include "migration-stats.h"
36 #include "qemu-file.h"
38 #include "migration/vmstate.h"
39 #include "block/block.h"
40 #include "qapi/error.h"
41 #include "qapi/clone-visitor.h"
42 #include "qapi/qapi-visit-migration.h"
43 #include "qapi/qapi-visit-sockets.h"
44 #include "qapi/qapi-commands-migration.h"
45 #include "qapi/qapi-events-migration.h"
46 #include "qapi/qmp/qerror.h"
47 #include "qapi/qmp/qnull.h"
50 #include "postcopy-ram.h"
51 #include "qemu/thread.h"
53 #include "exec/target_page.h"
54 #include "io/channel-buffer.h"
55 #include "io/channel-tls.h"
56 #include "migration/colo.h"
57 #include "hw/boards.h"
58 #include "monitor/monitor.h"
59 #include "net/announce.h"
60 #include "qemu/queue.h"
62 #include "threadinfo.h"
63 #include "qemu/yank.h"
64 #include "sysemu/cpus.h"
65 #include "yank_functions.h"
66 #include "sysemu/qtest.h"
68 #include "sysemu/dirtylimit.h"
70 static NotifierList migration_state_notifiers
=
71 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers
);
73 /* Messages sent on the return path from destination to source */
74 enum mig_rp_message_type
{
75 MIG_RP_MSG_INVALID
= 0, /* Must be 0 */
76 MIG_RP_MSG_SHUT
, /* sibling will not send any more RP messages */
77 MIG_RP_MSG_PONG
, /* Response to a PING; data (seq: be32 ) */
79 MIG_RP_MSG_REQ_PAGES_ID
, /* data (start: be64, len: be32, id: string) */
80 MIG_RP_MSG_REQ_PAGES
, /* data (start: be64, len: be32) */
81 MIG_RP_MSG_RECV_BITMAP
, /* send recved_bitmap back to source */
82 MIG_RP_MSG_RESUME_ACK
, /* tell source that we are ready to resume */
83 MIG_RP_MSG_SWITCHOVER_ACK
, /* Tell source it's OK to do switchover */
88 /* When we add fault tolerance, we could have several
89 migrations at once. For now we don't need to add
90 dynamic creation of migration */
92 static MigrationState
*current_migration
;
93 static MigrationIncomingState
*current_incoming
;
95 static GSList
*migration_blockers
;
97 static bool migration_object_check(MigrationState
*ms
, Error
**errp
);
98 static int migration_maybe_pause(MigrationState
*s
,
99 int *current_active_state
,
101 static void migrate_fd_cancel(MigrationState
*s
);
102 static int close_return_path_on_source(MigrationState
*s
);
104 static bool migration_needs_multiple_sockets(void)
106 return migrate_multifd() || migrate_postcopy_preempt();
109 static bool uri_supports_multi_channels(const char *uri
)
111 return strstart(uri
, "tcp:", NULL
) || strstart(uri
, "unix:", NULL
) ||
112 strstart(uri
, "vsock:", NULL
);
116 migration_channels_and_uri_compatible(const char *uri
, Error
**errp
)
118 if (migration_needs_multiple_sockets() &&
119 !uri_supports_multi_channels(uri
)) {
120 error_setg(errp
, "Migration requires multi-channel URIs (e.g. tcp)");
127 static gint
page_request_addr_cmp(gconstpointer ap
, gconstpointer bp
)
129 uintptr_t a
= (uintptr_t) ap
, b
= (uintptr_t) bp
;
131 return (a
> b
) - (a
< b
);
134 void migration_object_init(void)
136 /* This can only be called once. */
137 assert(!current_migration
);
138 current_migration
= MIGRATION_OBJ(object_new(TYPE_MIGRATION
));
141 * Init the migrate incoming object as well no matter whether
142 * we'll use it or not.
144 assert(!current_incoming
);
145 current_incoming
= g_new0(MigrationIncomingState
, 1);
146 current_incoming
->state
= MIGRATION_STATUS_NONE
;
147 current_incoming
->postcopy_remote_fds
=
148 g_array_new(FALSE
, TRUE
, sizeof(struct PostCopyFD
));
149 qemu_mutex_init(¤t_incoming
->rp_mutex
);
150 qemu_mutex_init(¤t_incoming
->postcopy_prio_thread_mutex
);
151 qemu_event_init(¤t_incoming
->main_thread_load_event
, false);
152 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_dst
, 0);
153 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fault
, 0);
154 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fast_load
, 0);
155 qemu_sem_init(¤t_incoming
->postcopy_qemufile_dst_done
, 0);
157 qemu_mutex_init(¤t_incoming
->page_request_mutex
);
158 qemu_cond_init(¤t_incoming
->page_request_cond
);
159 current_incoming
->page_requested
= g_tree_new(page_request_addr_cmp
);
161 migration_object_check(current_migration
, &error_fatal
);
165 dirty_bitmap_mig_init();
168 void migration_cancel(const Error
*error
)
171 migrate_set_error(current_migration
, error
);
173 if (migrate_dirty_limit()) {
174 qmp_cancel_vcpu_dirty_limit(false, -1, NULL
);
176 migrate_fd_cancel(current_migration
);
179 void migration_shutdown(void)
182 * When the QEMU main thread exit, the COLO thread
183 * may wait a semaphore. So, we should wakeup the
184 * COLO thread before migration shutdown.
188 * Cancel the current migration - that will (eventually)
189 * stop the migration using this structure
191 migration_cancel(NULL
);
192 object_unref(OBJECT(current_migration
));
195 * Cancel outgoing migration of dirty bitmaps. It should
196 * at least unref used block nodes.
198 dirty_bitmap_mig_cancel_outgoing();
201 * Cancel incoming migration of dirty bitmaps. Dirty bitmaps
202 * are non-critical data, and their loss never considered as
205 dirty_bitmap_mig_cancel_incoming();
209 MigrationState
*migrate_get_current(void)
211 /* This can only be called after the object created. */
212 assert(current_migration
);
213 return current_migration
;
216 MigrationIncomingState
*migration_incoming_get_current(void)
218 assert(current_incoming
);
219 return current_incoming
;
222 void migration_incoming_transport_cleanup(MigrationIncomingState
*mis
)
224 if (mis
->socket_address_list
) {
225 qapi_free_SocketAddressList(mis
->socket_address_list
);
226 mis
->socket_address_list
= NULL
;
229 if (mis
->transport_cleanup
) {
230 mis
->transport_cleanup(mis
->transport_data
);
231 mis
->transport_data
= mis
->transport_cleanup
= NULL
;
235 void migration_incoming_state_destroy(void)
237 struct MigrationIncomingState
*mis
= migration_incoming_get_current();
239 multifd_load_cleanup();
240 compress_threads_load_cleanup();
242 if (mis
->to_src_file
) {
243 /* Tell source that we are done */
244 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
245 qemu_fclose(mis
->to_src_file
);
246 mis
->to_src_file
= NULL
;
249 if (mis
->from_src_file
) {
250 migration_ioc_unregister_yank_from_file(mis
->from_src_file
);
251 qemu_fclose(mis
->from_src_file
);
252 mis
->from_src_file
= NULL
;
254 if (mis
->postcopy_remote_fds
) {
255 g_array_free(mis
->postcopy_remote_fds
, TRUE
);
256 mis
->postcopy_remote_fds
= NULL
;
259 migration_incoming_transport_cleanup(mis
);
260 qemu_event_reset(&mis
->main_thread_load_event
);
262 if (mis
->page_requested
) {
263 g_tree_destroy(mis
->page_requested
);
264 mis
->page_requested
= NULL
;
267 if (mis
->postcopy_qemufile_dst
) {
268 migration_ioc_unregister_yank_from_file(mis
->postcopy_qemufile_dst
);
269 qemu_fclose(mis
->postcopy_qemufile_dst
);
270 mis
->postcopy_qemufile_dst
= NULL
;
273 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
276 static void migrate_generate_event(int new_state
)
278 if (migrate_events()) {
279 qapi_event_send_migration(new_state
);
284 * Send a message on the return channel back to the source
287 static int migrate_send_rp_message(MigrationIncomingState
*mis
,
288 enum mig_rp_message_type message_type
,
289 uint16_t len
, void *data
)
293 trace_migrate_send_rp_message((int)message_type
, len
);
294 QEMU_LOCK_GUARD(&mis
->rp_mutex
);
297 * It's possible that the file handle got lost due to network
300 if (!mis
->to_src_file
) {
305 qemu_put_be16(mis
->to_src_file
, (unsigned int)message_type
);
306 qemu_put_be16(mis
->to_src_file
, len
);
307 qemu_put_buffer(mis
->to_src_file
, data
, len
);
308 qemu_fflush(mis
->to_src_file
);
310 /* It's possible that qemu file got error during sending */
311 ret
= qemu_file_get_error(mis
->to_src_file
);
316 /* Request one page from the source VM at the given start address.
317 * rb: the RAMBlock to request the page in
318 * Start: Address offset within the RB
319 * Len: Length in bytes required - must be a multiple of pagesize
321 int migrate_send_rp_message_req_pages(MigrationIncomingState
*mis
,
322 RAMBlock
*rb
, ram_addr_t start
)
324 uint8_t bufc
[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
325 size_t msglen
= 12; /* start + len */
326 size_t len
= qemu_ram_pagesize(rb
);
327 enum mig_rp_message_type msg_type
;
331 *(uint64_t *)bufc
= cpu_to_be64((uint64_t)start
);
332 *(uint32_t *)(bufc
+ 8) = cpu_to_be32((uint32_t)len
);
335 * We maintain the last ramblock that we requested for page. Note that we
336 * don't need locking because this function will only be called within the
337 * postcopy ram fault thread.
339 if (rb
!= mis
->last_rb
) {
342 rbname
= qemu_ram_get_idstr(rb
);
343 rbname_len
= strlen(rbname
);
345 assert(rbname_len
< 256);
347 bufc
[msglen
++] = rbname_len
;
348 memcpy(bufc
+ msglen
, rbname
, rbname_len
);
349 msglen
+= rbname_len
;
350 msg_type
= MIG_RP_MSG_REQ_PAGES_ID
;
352 msg_type
= MIG_RP_MSG_REQ_PAGES
;
355 return migrate_send_rp_message(mis
, msg_type
, msglen
, bufc
);
358 int migrate_send_rp_req_pages(MigrationIncomingState
*mis
,
359 RAMBlock
*rb
, ram_addr_t start
, uint64_t haddr
)
361 void *aligned
= (void *)(uintptr_t)ROUND_DOWN(haddr
, qemu_ram_pagesize(rb
));
362 bool received
= false;
364 WITH_QEMU_LOCK_GUARD(&mis
->page_request_mutex
) {
365 received
= ramblock_recv_bitmap_test_byte_offset(rb
, start
);
366 if (!received
&& !g_tree_lookup(mis
->page_requested
, aligned
)) {
368 * The page has not been received, and it's not yet in the page
369 * request list. Queue it. Set the value of element to 1, so that
370 * things like g_tree_lookup() will return TRUE (1) when found.
372 g_tree_insert(mis
->page_requested
, aligned
, (gpointer
)1);
373 qatomic_inc(&mis
->page_requested_count
);
374 trace_postcopy_page_req_add(aligned
, mis
->page_requested_count
);
379 * If the page is there, skip sending the message. We don't even need the
380 * lock because as long as the page arrived, it'll be there forever.
386 return migrate_send_rp_message_req_pages(mis
, rb
, start
);
389 static bool migration_colo_enabled
;
390 bool migration_incoming_colo_enabled(void)
392 return migration_colo_enabled
;
395 void migration_incoming_disable_colo(void)
397 ram_block_discard_disable(false);
398 migration_colo_enabled
= false;
401 int migration_incoming_enable_colo(void)
403 #ifndef CONFIG_REPLICATION
404 error_report("ENABLE_COLO command come in migration stream, but COLO "
405 "module is not built in");
409 if (!migrate_colo()) {
410 error_report("ENABLE_COLO command come in migration stream, but c-colo "
411 "capability is not set");
415 if (ram_block_discard_disable(true)) {
416 error_report("COLO: cannot disable RAM discard");
419 migration_colo_enabled
= true;
423 void migrate_add_address(SocketAddress
*address
)
425 MigrationIncomingState
*mis
= migration_incoming_get_current();
427 QAPI_LIST_PREPEND(mis
->socket_address_list
,
428 QAPI_CLONE(SocketAddress
, address
));
431 static void qemu_start_incoming_migration(const char *uri
, Error
**errp
)
433 const char *p
= NULL
;
434 MigrationIncomingState
*mis
= migration_incoming_get_current();
436 /* URI is not suitable for migration? */
437 if (!migration_channels_and_uri_compatible(uri
, errp
)) {
441 migrate_set_state(&mis
->state
, MIGRATION_STATUS_NONE
,
442 MIGRATION_STATUS_SETUP
);
444 if (strstart(uri
, "tcp:", &p
) ||
445 strstart(uri
, "unix:", NULL
) ||
446 strstart(uri
, "vsock:", NULL
)) {
447 socket_start_incoming_migration(p
? p
: uri
, errp
);
449 } else if (strstart(uri
, "rdma:", &p
)) {
450 if (migrate_compress()) {
451 error_setg(errp
, "RDMA and compression can't be used together");
454 if (migrate_xbzrle()) {
455 error_setg(errp
, "RDMA and XBZRLE can't be used together");
458 if (migrate_multifd()) {
459 error_setg(errp
, "RDMA and multifd can't be used together");
462 rdma_start_incoming_migration(p
, errp
);
464 } else if (strstart(uri
, "exec:", &p
)) {
465 exec_start_incoming_migration(p
, errp
);
466 } else if (strstart(uri
, "fd:", &p
)) {
467 fd_start_incoming_migration(p
, errp
);
468 } else if (strstart(uri
, "file:", &p
)) {
469 file_start_incoming_migration(p
, errp
);
471 error_setg(errp
, "unknown migration protocol: %s", uri
);
475 static void process_incoming_migration_bh(void *opaque
)
477 Error
*local_err
= NULL
;
478 MigrationIncomingState
*mis
= opaque
;
480 /* If capability late_block_activate is set:
481 * Only fire up the block code now if we're going to restart the
482 * VM, else 'cont' will do it.
483 * This causes file locking to happen; so we don't want it to happen
484 * unless we really are starting the VM.
486 if (!migrate_late_block_activate() ||
487 (autostart
&& (!global_state_received() ||
488 global_state_get_runstate() == RUN_STATE_RUNNING
))) {
489 /* Make sure all file formats throw away their mutable metadata.
490 * If we get an error here, just don't restart the VM yet. */
491 bdrv_activate_all(&local_err
);
493 error_report_err(local_err
);
500 * This must happen after all error conditions are dealt with and
501 * we're sure the VM is going to be running on this host.
503 qemu_announce_self(&mis
->announce_timer
, migrate_announce_params());
505 multifd_load_shutdown();
507 dirty_bitmap_mig_before_vm_start();
509 if (!global_state_received() ||
510 global_state_get_runstate() == RUN_STATE_RUNNING
) {
514 runstate_set(RUN_STATE_PAUSED
);
516 } else if (migration_incoming_colo_enabled()) {
517 migration_incoming_disable_colo();
520 runstate_set(global_state_get_runstate());
523 * This must happen after any state changes since as soon as an external
524 * observer sees this event they might start to prod at the VM assuming
527 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
528 MIGRATION_STATUS_COMPLETED
);
529 qemu_bh_delete(mis
->bh
);
530 migration_incoming_state_destroy();
533 static void coroutine_fn
534 process_incoming_migration_co(void *opaque
)
536 MigrationIncomingState
*mis
= migration_incoming_get_current();
540 assert(mis
->from_src_file
);
542 if (compress_threads_load_setup(mis
->from_src_file
)) {
543 error_report("Failed to setup decompress threads");
547 mis
->largest_page_size
= qemu_ram_pagesize_largest();
548 postcopy_state_set(POSTCOPY_INCOMING_NONE
);
549 migrate_set_state(&mis
->state
, MIGRATION_STATUS_SETUP
,
550 MIGRATION_STATUS_ACTIVE
);
552 mis
->loadvm_co
= qemu_coroutine_self();
553 ret
= qemu_loadvm_state(mis
->from_src_file
);
554 mis
->loadvm_co
= NULL
;
556 ps
= postcopy_state_get();
557 trace_process_incoming_migration_co_end(ret
, ps
);
558 if (ps
!= POSTCOPY_INCOMING_NONE
) {
559 if (ps
== POSTCOPY_INCOMING_ADVISE
) {
561 * Where a migration had postcopy enabled (and thus went to advise)
562 * but managed to complete within the precopy period, we can use
565 postcopy_ram_incoming_cleanup(mis
);
566 } else if (ret
>= 0) {
568 * Postcopy was started, cleanup should happen at the end of the
571 trace_process_incoming_migration_co_postcopy_end_main();
574 /* Else if something went wrong then just fall out of the normal exit */
578 error_report("load of migration failed: %s", strerror(-ret
));
582 if (colo_incoming_co() < 0) {
586 mis
->bh
= qemu_bh_new(process_incoming_migration_bh
, mis
);
587 qemu_bh_schedule(mis
->bh
);
590 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
591 MIGRATION_STATUS_FAILED
);
592 qemu_fclose(mis
->from_src_file
);
594 multifd_load_cleanup();
595 compress_threads_load_cleanup();
601 * migration_incoming_setup: Setup incoming migration
602 * @f: file for main migration channel
603 * @errp: where to put errors
605 * Returns: %true on success, %false on error.
607 static bool migration_incoming_setup(QEMUFile
*f
, Error
**errp
)
609 MigrationIncomingState
*mis
= migration_incoming_get_current();
611 if (!mis
->from_src_file
) {
612 mis
->from_src_file
= f
;
614 qemu_file_set_blocking(f
, false);
618 void migration_incoming_process(void)
620 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
, NULL
);
621 qemu_coroutine_enter(co
);
624 /* Returns true if recovered from a paused migration, otherwise false */
625 static bool postcopy_try_recover(void)
627 MigrationIncomingState
*mis
= migration_incoming_get_current();
629 if (mis
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
630 /* Resumed from a paused postcopy migration */
632 /* This should be set already in migration_incoming_setup() */
633 assert(mis
->from_src_file
);
634 /* Postcopy has standalone thread to do vm load */
635 qemu_file_set_blocking(mis
->from_src_file
, true);
637 /* Re-configure the return path */
638 mis
->to_src_file
= qemu_file_get_return_path(mis
->from_src_file
);
640 migrate_set_state(&mis
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
641 MIGRATION_STATUS_POSTCOPY_RECOVER
);
644 * Here, we only wake up the main loading thread (while the
645 * rest threads will still be waiting), so that we can receive
646 * commands from source now, and answer it if needed. The
647 * rest threads will be woken up afterwards until we are sure
648 * that source is ready to reply to page requests.
650 qemu_sem_post(&mis
->postcopy_pause_sem_dst
);
657 void migration_fd_process_incoming(QEMUFile
*f
, Error
**errp
)
659 if (!migration_incoming_setup(f
, errp
)) {
662 if (postcopy_try_recover()) {
665 migration_incoming_process();
669 * Returns true when we want to start a new incoming migration process,
672 static bool migration_should_start_incoming(bool main_channel
)
674 /* Multifd doesn't start unless all channels are established */
675 if (migrate_multifd()) {
676 return migration_has_all_channels();
679 /* Preempt channel only starts when the main channel is created */
680 if (migrate_postcopy_preempt()) {
685 * For all the rest types of migration, we should only reach here when
686 * it's the main channel that's being created, and we should always
687 * proceed with this channel.
689 assert(main_channel
);
693 void migration_ioc_process_incoming(QIOChannel
*ioc
, Error
**errp
)
695 MigrationIncomingState
*mis
= migration_incoming_get_current();
696 Error
*local_err
= NULL
;
698 bool default_channel
= true;
699 uint32_t channel_magic
= 0;
702 if (migrate_multifd() && !migrate_postcopy_ram() &&
703 qio_channel_has_feature(ioc
, QIO_CHANNEL_FEATURE_READ_MSG_PEEK
)) {
705 * With multiple channels, it is possible that we receive channels
706 * out of order on destination side, causing incorrect mapping of
707 * source channels on destination side. Check channel MAGIC to
708 * decide type of channel. Please note this is best effort, postcopy
709 * preempt channel does not send any magic number so avoid it for
710 * postcopy live migration. Also tls live migration already does
711 * tls handshake while initializing main channel so with tls this
712 * issue is not possible.
714 ret
= migration_channel_read_peek(ioc
, (void *)&channel_magic
,
715 sizeof(channel_magic
), &local_err
);
718 error_propagate(errp
, local_err
);
722 default_channel
= (channel_magic
== cpu_to_be32(QEMU_VM_FILE_MAGIC
));
724 default_channel
= !mis
->from_src_file
;
727 if (multifd_load_setup(errp
) != 0) {
728 error_setg(errp
, "Failed to setup multifd channels");
732 if (default_channel
) {
733 f
= qemu_file_new_input(ioc
);
735 if (!migration_incoming_setup(f
, errp
)) {
739 /* Multiple connections */
740 assert(migration_needs_multiple_sockets());
741 if (migrate_multifd()) {
742 multifd_recv_new_channel(ioc
, &local_err
);
744 assert(migrate_postcopy_preempt());
745 f
= qemu_file_new_input(ioc
);
746 postcopy_preempt_new_channel(mis
, f
);
749 error_propagate(errp
, local_err
);
754 if (migration_should_start_incoming(default_channel
)) {
755 /* If it's a recovery, we're done */
756 if (postcopy_try_recover()) {
759 migration_incoming_process();
764 * @migration_has_all_channels: We have received all channels that we need
766 * Returns true when we have got connections to all the channels that
767 * we need for migration.
769 bool migration_has_all_channels(void)
771 MigrationIncomingState
*mis
= migration_incoming_get_current();
773 if (!mis
->from_src_file
) {
777 if (migrate_multifd()) {
778 return multifd_recv_all_channels_created();
781 if (migrate_postcopy_preempt()) {
782 return mis
->postcopy_qemufile_dst
!= NULL
;
788 int migrate_send_rp_switchover_ack(MigrationIncomingState
*mis
)
790 return migrate_send_rp_message(mis
, MIG_RP_MSG_SWITCHOVER_ACK
, 0, NULL
);
794 * Send a 'SHUT' message on the return channel with the given value
795 * to indicate that we've finished with the RP. Non-0 value indicates
798 void migrate_send_rp_shut(MigrationIncomingState
*mis
,
803 buf
= cpu_to_be32(value
);
804 migrate_send_rp_message(mis
, MIG_RP_MSG_SHUT
, sizeof(buf
), &buf
);
808 * Send a 'PONG' message on the return channel with the given value
809 * (normally in response to a 'PING')
811 void migrate_send_rp_pong(MigrationIncomingState
*mis
,
816 buf
= cpu_to_be32(value
);
817 migrate_send_rp_message(mis
, MIG_RP_MSG_PONG
, sizeof(buf
), &buf
);
820 void migrate_send_rp_recv_bitmap(MigrationIncomingState
*mis
,
828 * First, we send the header part. It contains only the len of
829 * idstr, and the idstr itself.
831 len
= strlen(block_name
);
833 memcpy(buf
+ 1, block_name
, len
);
835 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
836 error_report("%s: MSG_RP_RECV_BITMAP only used for recovery",
841 migrate_send_rp_message(mis
, MIG_RP_MSG_RECV_BITMAP
, len
+ 1, buf
);
844 * Next, we dump the received bitmap to the stream.
846 * TODO: currently we are safe since we are the only one that is
847 * using the to_src_file handle (fault thread is still paused),
848 * and it's ok even not taking the mutex. However the best way is
849 * to take the lock before sending the message header, and release
850 * the lock after sending the bitmap.
852 qemu_mutex_lock(&mis
->rp_mutex
);
853 res
= ramblock_recv_bitmap_send(mis
->to_src_file
, block_name
);
854 qemu_mutex_unlock(&mis
->rp_mutex
);
856 trace_migrate_send_rp_recv_bitmap(block_name
, res
);
859 void migrate_send_rp_resume_ack(MigrationIncomingState
*mis
, uint32_t value
)
863 buf
= cpu_to_be32(value
);
864 migrate_send_rp_message(mis
, MIG_RP_MSG_RESUME_ACK
, sizeof(buf
), &buf
);
868 * Return true if we're already in the middle of a migration
869 * (i.e. any of the active or setup states)
871 bool migration_is_setup_or_active(int state
)
874 case MIGRATION_STATUS_ACTIVE
:
875 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
876 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
877 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
878 case MIGRATION_STATUS_SETUP
:
879 case MIGRATION_STATUS_PRE_SWITCHOVER
:
880 case MIGRATION_STATUS_DEVICE
:
881 case MIGRATION_STATUS_WAIT_UNPLUG
:
882 case MIGRATION_STATUS_COLO
:
891 bool migration_is_running(int state
)
894 case MIGRATION_STATUS_ACTIVE
:
895 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
896 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
897 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
898 case MIGRATION_STATUS_SETUP
:
899 case MIGRATION_STATUS_PRE_SWITCHOVER
:
900 case MIGRATION_STATUS_DEVICE
:
901 case MIGRATION_STATUS_WAIT_UNPLUG
:
902 case MIGRATION_STATUS_CANCELLING
:
911 static bool migrate_show_downtime(MigrationState
*s
)
913 return (s
->state
== MIGRATION_STATUS_COMPLETED
) || migration_in_postcopy();
916 static void populate_time_info(MigrationInfo
*info
, MigrationState
*s
)
918 info
->has_status
= true;
919 info
->has_setup_time
= true;
920 info
->setup_time
= s
->setup_time
;
922 if (s
->state
== MIGRATION_STATUS_COMPLETED
) {
923 info
->has_total_time
= true;
924 info
->total_time
= s
->total_time
;
926 info
->has_total_time
= true;
927 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) -
931 if (migrate_show_downtime(s
)) {
932 info
->has_downtime
= true;
933 info
->downtime
= s
->downtime
;
935 info
->has_expected_downtime
= true;
936 info
->expected_downtime
= s
->expected_downtime
;
940 static void populate_ram_info(MigrationInfo
*info
, MigrationState
*s
)
942 size_t page_size
= qemu_target_page_size();
944 info
->ram
= g_malloc0(sizeof(*info
->ram
));
945 info
->ram
->transferred
= stat64_get(&mig_stats
.transferred
);
946 info
->ram
->total
= ram_bytes_total();
947 info
->ram
->duplicate
= stat64_get(&mig_stats
.zero_pages
);
948 /* legacy value. It is not used anymore */
949 info
->ram
->skipped
= 0;
950 info
->ram
->normal
= stat64_get(&mig_stats
.normal_pages
);
951 info
->ram
->normal_bytes
= info
->ram
->normal
* page_size
;
952 info
->ram
->mbps
= s
->mbps
;
953 info
->ram
->dirty_sync_count
=
954 stat64_get(&mig_stats
.dirty_sync_count
);
955 info
->ram
->dirty_sync_missed_zero_copy
=
956 stat64_get(&mig_stats
.dirty_sync_missed_zero_copy
);
957 info
->ram
->postcopy_requests
=
958 stat64_get(&mig_stats
.postcopy_requests
);
959 info
->ram
->page_size
= page_size
;
960 info
->ram
->multifd_bytes
= stat64_get(&mig_stats
.multifd_bytes
);
961 info
->ram
->pages_per_second
= s
->pages_per_second
;
962 info
->ram
->precopy_bytes
= stat64_get(&mig_stats
.precopy_bytes
);
963 info
->ram
->downtime_bytes
= stat64_get(&mig_stats
.downtime_bytes
);
964 info
->ram
->postcopy_bytes
= stat64_get(&mig_stats
.postcopy_bytes
);
966 if (migrate_xbzrle()) {
967 info
->xbzrle_cache
= g_malloc0(sizeof(*info
->xbzrle_cache
));
968 info
->xbzrle_cache
->cache_size
= migrate_xbzrle_cache_size();
969 info
->xbzrle_cache
->bytes
= xbzrle_counters
.bytes
;
970 info
->xbzrle_cache
->pages
= xbzrle_counters
.pages
;
971 info
->xbzrle_cache
->cache_miss
= xbzrle_counters
.cache_miss
;
972 info
->xbzrle_cache
->cache_miss_rate
= xbzrle_counters
.cache_miss_rate
;
973 info
->xbzrle_cache
->encoding_rate
= xbzrle_counters
.encoding_rate
;
974 info
->xbzrle_cache
->overflow
= xbzrle_counters
.overflow
;
977 if (migrate_compress()) {
978 info
->compression
= g_malloc0(sizeof(*info
->compression
));
979 info
->compression
->pages
= compression_counters
.pages
;
980 info
->compression
->busy
= compression_counters
.busy
;
981 info
->compression
->busy_rate
= compression_counters
.busy_rate
;
982 info
->compression
->compressed_size
=
983 compression_counters
.compressed_size
;
984 info
->compression
->compression_rate
=
985 compression_counters
.compression_rate
;
988 if (cpu_throttle_active()) {
989 info
->has_cpu_throttle_percentage
= true;
990 info
->cpu_throttle_percentage
= cpu_throttle_get_percentage();
993 if (s
->state
!= MIGRATION_STATUS_COMPLETED
) {
994 info
->ram
->remaining
= ram_bytes_remaining();
995 info
->ram
->dirty_pages_rate
=
996 stat64_get(&mig_stats
.dirty_pages_rate
);
999 if (migrate_dirty_limit() && dirtylimit_in_service()) {
1000 info
->has_dirty_limit_throttle_time_per_round
= true;
1001 info
->dirty_limit_throttle_time_per_round
=
1002 dirtylimit_throttle_time_per_round();
1004 info
->has_dirty_limit_ring_full_time
= true;
1005 info
->dirty_limit_ring_full_time
= dirtylimit_ring_full_time();
1009 static void populate_disk_info(MigrationInfo
*info
)
1011 if (blk_mig_active()) {
1012 info
->disk
= g_malloc0(sizeof(*info
->disk
));
1013 info
->disk
->transferred
= blk_mig_bytes_transferred();
1014 info
->disk
->remaining
= blk_mig_bytes_remaining();
1015 info
->disk
->total
= blk_mig_bytes_total();
1019 static void fill_source_migration_info(MigrationInfo
*info
)
1021 MigrationState
*s
= migrate_get_current();
1022 int state
= qatomic_read(&s
->state
);
1023 GSList
*cur_blocker
= migration_blockers
;
1025 info
->blocked_reasons
= NULL
;
1028 * There are two types of reasons a migration might be blocked;
1029 * a) devices marked in VMState as non-migratable, and
1030 * b) Explicit migration blockers
1031 * We need to add both of them here.
1033 qemu_savevm_non_migratable_list(&info
->blocked_reasons
);
1035 while (cur_blocker
) {
1036 QAPI_LIST_PREPEND(info
->blocked_reasons
,
1037 g_strdup(error_get_pretty(cur_blocker
->data
)));
1038 cur_blocker
= g_slist_next(cur_blocker
);
1040 info
->has_blocked_reasons
= info
->blocked_reasons
!= NULL
;
1043 case MIGRATION_STATUS_NONE
:
1044 /* no migration has happened ever */
1045 /* do not overwrite destination migration status */
1047 case MIGRATION_STATUS_SETUP
:
1048 info
->has_status
= true;
1049 info
->has_total_time
= false;
1051 case MIGRATION_STATUS_ACTIVE
:
1052 case MIGRATION_STATUS_CANCELLING
:
1053 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1054 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1055 case MIGRATION_STATUS_DEVICE
:
1056 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1057 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1058 /* TODO add some postcopy stats */
1059 populate_time_info(info
, s
);
1060 populate_ram_info(info
, s
);
1061 populate_disk_info(info
);
1062 migration_populate_vfio_info(info
);
1064 case MIGRATION_STATUS_COLO
:
1065 info
->has_status
= true;
1066 /* TODO: display COLO specific information (checkpoint info etc.) */
1068 case MIGRATION_STATUS_COMPLETED
:
1069 populate_time_info(info
, s
);
1070 populate_ram_info(info
, s
);
1071 migration_populate_vfio_info(info
);
1073 case MIGRATION_STATUS_FAILED
:
1074 info
->has_status
= true;
1076 case MIGRATION_STATUS_CANCELLED
:
1077 info
->has_status
= true;
1079 case MIGRATION_STATUS_WAIT_UNPLUG
:
1080 info
->has_status
= true;
1083 info
->status
= state
;
1085 QEMU_LOCK_GUARD(&s
->error_mutex
);
1087 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1091 static void fill_destination_migration_info(MigrationInfo
*info
)
1093 MigrationIncomingState
*mis
= migration_incoming_get_current();
1095 if (mis
->socket_address_list
) {
1096 info
->has_socket_address
= true;
1097 info
->socket_address
=
1098 QAPI_CLONE(SocketAddressList
, mis
->socket_address_list
);
1101 switch (mis
->state
) {
1102 case MIGRATION_STATUS_NONE
:
1104 case MIGRATION_STATUS_SETUP
:
1105 case MIGRATION_STATUS_CANCELLING
:
1106 case MIGRATION_STATUS_CANCELLED
:
1107 case MIGRATION_STATUS_ACTIVE
:
1108 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1109 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1110 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1111 case MIGRATION_STATUS_FAILED
:
1112 case MIGRATION_STATUS_COLO
:
1113 info
->has_status
= true;
1115 case MIGRATION_STATUS_COMPLETED
:
1116 info
->has_status
= true;
1117 fill_destination_postcopy_migration_info(info
);
1120 info
->status
= mis
->state
;
1123 MigrationInfo
*qmp_query_migrate(Error
**errp
)
1125 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
1127 fill_destination_migration_info(info
);
1128 fill_source_migration_info(info
);
1133 void qmp_migrate_start_postcopy(Error
**errp
)
1135 MigrationState
*s
= migrate_get_current();
1137 if (!migrate_postcopy()) {
1138 error_setg(errp
, "Enable postcopy with migrate_set_capability before"
1139 " the start of migration");
1143 if (s
->state
== MIGRATION_STATUS_NONE
) {
1144 error_setg(errp
, "Postcopy must be started after migration has been"
1149 * we don't error if migration has finished since that would be racy
1150 * with issuing this command.
1152 qatomic_set(&s
->start_postcopy
, true);
1155 /* shared migration helpers */
1157 void migrate_set_state(int *state
, int old_state
, int new_state
)
1159 assert(new_state
< MIGRATION_STATUS__MAX
);
1160 if (qatomic_cmpxchg(state
, old_state
, new_state
) == old_state
) {
1161 trace_migrate_set_state(MigrationStatus_str(new_state
));
1162 migrate_generate_event(new_state
);
1166 static void migrate_fd_cleanup(MigrationState
*s
)
1168 qemu_bh_delete(s
->cleanup_bh
);
1169 s
->cleanup_bh
= NULL
;
1171 g_free(s
->hostname
);
1173 json_writer_free(s
->vmdesc
);
1176 qemu_savevm_state_cleanup();
1178 if (s
->to_dst_file
) {
1181 trace_migrate_fd_cleanup();
1182 qemu_mutex_unlock_iothread();
1183 if (s
->migration_thread_running
) {
1184 qemu_thread_join(&s
->thread
);
1185 s
->migration_thread_running
= false;
1187 qemu_mutex_lock_iothread();
1189 multifd_save_cleanup();
1190 qemu_mutex_lock(&s
->qemu_file_lock
);
1191 tmp
= s
->to_dst_file
;
1192 s
->to_dst_file
= NULL
;
1193 qemu_mutex_unlock(&s
->qemu_file_lock
);
1195 * Close the file handle without the lock to make sure the
1196 * critical section won't block for long.
1198 migration_ioc_unregister_yank_from_file(tmp
);
1203 * We already cleaned up to_dst_file, so errors from the return
1204 * path might be due to that, ignore them.
1206 close_return_path_on_source(s
);
1208 assert(!migration_is_active(s
));
1210 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1211 migrate_set_state(&s
->state
, MIGRATION_STATUS_CANCELLING
,
1212 MIGRATION_STATUS_CANCELLED
);
1216 /* It is used on info migrate. We can't free it */
1217 error_report_err(error_copy(s
->error
));
1219 notifier_list_notify(&migration_state_notifiers
, s
);
1220 block_cleanup_parameters();
1221 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1224 static void migrate_fd_cleanup_schedule(MigrationState
*s
)
1227 * Ref the state for bh, because it may be called when
1228 * there're already no other refs
1230 object_ref(OBJECT(s
));
1231 qemu_bh_schedule(s
->cleanup_bh
);
1234 static void migrate_fd_cleanup_bh(void *opaque
)
1236 MigrationState
*s
= opaque
;
1237 migrate_fd_cleanup(s
);
1238 object_unref(OBJECT(s
));
1241 void migrate_set_error(MigrationState
*s
, const Error
*error
)
1243 QEMU_LOCK_GUARD(&s
->error_mutex
);
1245 s
->error
= error_copy(error
);
1249 bool migrate_has_error(MigrationState
*s
)
1251 /* The lock is not helpful here, but still follow the rule */
1252 QEMU_LOCK_GUARD(&s
->error_mutex
);
1253 return qatomic_read(&s
->error
);
1256 static void migrate_error_free(MigrationState
*s
)
1258 QEMU_LOCK_GUARD(&s
->error_mutex
);
1260 error_free(s
->error
);
1265 static void migrate_fd_error(MigrationState
*s
, const Error
*error
)
1267 trace_migrate_fd_error(error_get_pretty(error
));
1268 assert(s
->to_dst_file
== NULL
);
1269 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1270 MIGRATION_STATUS_FAILED
);
1271 migrate_set_error(s
, error
);
1274 static void migrate_fd_cancel(MigrationState
*s
)
1278 trace_migrate_fd_cancel();
1280 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1281 if (s
->rp_state
.from_dst_file
) {
1282 /* shutdown the rp socket, so causing the rp thread to shutdown */
1283 qemu_file_shutdown(s
->rp_state
.from_dst_file
);
1288 old_state
= s
->state
;
1289 if (!migration_is_running(old_state
)) {
1292 /* If the migration is paused, kick it out of the pause */
1293 if (old_state
== MIGRATION_STATUS_PRE_SWITCHOVER
) {
1294 qemu_sem_post(&s
->pause_sem
);
1296 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_CANCELLING
);
1297 } while (s
->state
!= MIGRATION_STATUS_CANCELLING
);
1300 * If we're unlucky the migration code might be stuck somewhere in a
1301 * send/write while the network has failed and is waiting to timeout;
1302 * if we've got shutdown(2) available then we can force it to quit.
1304 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1305 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1306 if (s
->to_dst_file
) {
1307 qemu_file_shutdown(s
->to_dst_file
);
1311 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& s
->block_inactive
) {
1312 Error
*local_err
= NULL
;
1314 bdrv_activate_all(&local_err
);
1316 error_report_err(local_err
);
1318 s
->block_inactive
= false;
1323 void add_migration_state_change_notifier(Notifier
*notify
)
1325 notifier_list_add(&migration_state_notifiers
, notify
);
1328 void remove_migration_state_change_notifier(Notifier
*notify
)
1330 notifier_remove(notify
);
1333 bool migration_in_setup(MigrationState
*s
)
1335 return s
->state
== MIGRATION_STATUS_SETUP
;
1338 bool migration_has_finished(MigrationState
*s
)
1340 return s
->state
== MIGRATION_STATUS_COMPLETED
;
1343 bool migration_has_failed(MigrationState
*s
)
1345 return (s
->state
== MIGRATION_STATUS_CANCELLED
||
1346 s
->state
== MIGRATION_STATUS_FAILED
);
1349 bool migration_in_postcopy(void)
1351 MigrationState
*s
= migrate_get_current();
1354 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1355 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1356 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1363 bool migration_in_postcopy_after_devices(MigrationState
*s
)
1365 return migration_in_postcopy() && s
->postcopy_after_devices
;
1368 bool migration_in_incoming_postcopy(void)
1370 PostcopyState ps
= postcopy_state_get();
1372 return ps
>= POSTCOPY_INCOMING_DISCARD
&& ps
< POSTCOPY_INCOMING_END
;
1375 bool migration_incoming_postcopy_advised(void)
1377 PostcopyState ps
= postcopy_state_get();
1379 return ps
>= POSTCOPY_INCOMING_ADVISE
&& ps
< POSTCOPY_INCOMING_END
;
1382 bool migration_in_bg_snapshot(void)
1384 MigrationState
*s
= migrate_get_current();
1386 return migrate_background_snapshot() &&
1387 migration_is_setup_or_active(s
->state
);
1390 bool migration_is_idle(void)
1392 MigrationState
*s
= current_migration
;
1399 case MIGRATION_STATUS_NONE
:
1400 case MIGRATION_STATUS_CANCELLED
:
1401 case MIGRATION_STATUS_COMPLETED
:
1402 case MIGRATION_STATUS_FAILED
:
1404 case MIGRATION_STATUS_SETUP
:
1405 case MIGRATION_STATUS_CANCELLING
:
1406 case MIGRATION_STATUS_ACTIVE
:
1407 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1408 case MIGRATION_STATUS_COLO
:
1409 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1410 case MIGRATION_STATUS_DEVICE
:
1411 case MIGRATION_STATUS_WAIT_UNPLUG
:
1413 case MIGRATION_STATUS__MAX
:
1414 g_assert_not_reached();
1420 bool migration_is_active(MigrationState
*s
)
1422 return (s
->state
== MIGRATION_STATUS_ACTIVE
||
1423 s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1426 int migrate_init(MigrationState
*s
, Error
**errp
)
1430 ret
= qemu_savevm_state_prepare(errp
);
1436 * Reinitialise all migration state, except
1437 * parameters/capabilities that the user set, and
1442 s
->to_dst_file
= NULL
;
1443 s
->state
= MIGRATION_STATUS_NONE
;
1444 s
->rp_state
.from_dst_file
= NULL
;
1445 s
->rp_state
.error
= false;
1447 s
->pages_per_second
= 0.0;
1449 s
->expected_downtime
= 0;
1451 s
->start_postcopy
= false;
1452 s
->postcopy_after_devices
= false;
1453 s
->migration_thread_running
= false;
1454 error_free(s
->error
);
1459 migrate_set_state(&s
->state
, MIGRATION_STATUS_NONE
, MIGRATION_STATUS_SETUP
);
1461 s
->start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1463 s
->vm_old_state
= -1;
1464 s
->iteration_initial_bytes
= 0;
1465 s
->threshold_size
= 0;
1466 s
->switchover_acked
= false;
1467 s
->rdma_migration
= false;
1469 * set mig_stats memory to zero for a new migration
1471 memset(&mig_stats
, 0, sizeof(mig_stats
));
1472 migration_reset_vfio_bytes_transferred();
1477 int migrate_add_blocker_internal(Error
*reason
, Error
**errp
)
1479 /* Snapshots are similar to migrations, so check RUN_STATE_SAVE_VM too. */
1480 if (runstate_check(RUN_STATE_SAVE_VM
) || !migration_is_idle()) {
1481 error_propagate_prepend(errp
, error_copy(reason
),
1482 "disallowing migration blocker "
1483 "(migration/snapshot in progress) for: ");
1487 migration_blockers
= g_slist_prepend(migration_blockers
, reason
);
1491 int migrate_add_blocker(Error
*reason
, Error
**errp
)
1493 if (only_migratable
) {
1494 error_propagate_prepend(errp
, error_copy(reason
),
1495 "disallowing migration blocker "
1496 "(--only-migratable) for: ");
1500 return migrate_add_blocker_internal(reason
, errp
);
1503 void migrate_del_blocker(Error
*reason
)
1505 migration_blockers
= g_slist_remove(migration_blockers
, reason
);
1508 void qmp_migrate_incoming(const char *uri
, Error
**errp
)
1510 Error
*local_err
= NULL
;
1511 static bool once
= true;
1514 error_setg(errp
, "The incoming migration has already been started");
1517 if (!runstate_check(RUN_STATE_INMIGRATE
)) {
1518 error_setg(errp
, "'-incoming' was not specified on the command line");
1522 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1526 qemu_start_incoming_migration(uri
, &local_err
);
1529 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1530 error_propagate(errp
, local_err
);
1537 void qmp_migrate_recover(const char *uri
, Error
**errp
)
1539 MigrationIncomingState
*mis
= migration_incoming_get_current();
1542 * Don't even bother to use ERRP_GUARD() as it _must_ always be set by
1543 * callers (no one should ignore a recover failure); if there is, it's a
1544 * programming error.
1548 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1549 error_setg(errp
, "Migrate recover can only be run "
1550 "when postcopy is paused.");
1554 /* If there's an existing transport, release it */
1555 migration_incoming_transport_cleanup(mis
);
1558 * Note that this call will never start a real migration; it will
1559 * only re-setup the migration stream and poke existing migration
1560 * to continue using that newly established channel.
1562 qemu_start_incoming_migration(uri
, errp
);
1565 void qmp_migrate_pause(Error
**errp
)
1567 MigrationState
*ms
= migrate_get_current();
1568 MigrationIncomingState
*mis
= migration_incoming_get_current();
1571 if (ms
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
1572 /* Source side, during postcopy */
1573 qemu_mutex_lock(&ms
->qemu_file_lock
);
1574 if (ms
->to_dst_file
) {
1575 ret
= qemu_file_shutdown(ms
->to_dst_file
);
1577 qemu_mutex_unlock(&ms
->qemu_file_lock
);
1579 error_setg(errp
, "Failed to pause source migration");
1584 if (mis
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
1585 ret
= qemu_file_shutdown(mis
->from_src_file
);
1587 error_setg(errp
, "Failed to pause destination migration");
1592 error_setg(errp
, "migrate-pause is currently only supported "
1593 "during postcopy-active state");
1596 bool migration_is_blocked(Error
**errp
)
1598 if (qemu_savevm_state_blocked(errp
)) {
1602 if (migration_blockers
) {
1603 error_propagate(errp
, error_copy(migration_blockers
->data
));
1610 /* Returns true if continue to migrate, or false if error detected */
1611 static bool migrate_prepare(MigrationState
*s
, bool blk
, bool blk_inc
,
1612 bool resume
, Error
**errp
)
1614 Error
*local_err
= NULL
;
1617 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1618 error_setg(errp
, "Cannot resume if there is no "
1619 "paused migration");
1624 * Postcopy recovery won't work well with release-ram
1625 * capability since release-ram will drop the page buffer as
1626 * long as the page is put into the send buffer. So if there
1627 * is a network failure happened, any page buffers that have
1628 * not yet reached the destination VM but have already been
1629 * sent from the source VM will be lost forever. Let's refuse
1630 * the client from resuming such a postcopy migration.
1631 * Luckily release-ram was designed to only be used when src
1632 * and destination VMs are on the same host, so it should be
1635 if (migrate_release_ram()) {
1636 error_setg(errp
, "Postcopy recovery cannot work "
1637 "when release-ram capability is set");
1641 /* This is a resume, skip init status */
1645 if (migration_is_running(s
->state
)) {
1646 error_setg(errp
, QERR_MIGRATION_ACTIVE
);
1650 if (runstate_check(RUN_STATE_INMIGRATE
)) {
1651 error_setg(errp
, "Guest is waiting for an incoming migration");
1655 if (runstate_check(RUN_STATE_POSTMIGRATE
)) {
1656 error_setg(errp
, "Can't migrate the vm that was paused due to "
1657 "previous migration");
1661 if (migration_is_blocked(errp
)) {
1665 if (blk
|| blk_inc
) {
1666 if (migrate_colo()) {
1667 error_setg(errp
, "No disk migration is required in COLO mode");
1670 if (migrate_block() || migrate_block_incremental()) {
1671 error_setg(errp
, "Command options are incompatible with "
1672 "current migration capabilities");
1675 if (!migrate_cap_set(MIGRATION_CAPABILITY_BLOCK
, true, &local_err
)) {
1676 error_propagate(errp
, local_err
);
1679 s
->must_remove_block_options
= true;
1683 migrate_set_block_incremental(true);
1686 if (migrate_init(s
, errp
)) {
1693 void qmp_migrate(const char *uri
, bool has_blk
, bool blk
,
1694 bool has_inc
, bool inc
, bool has_detach
, bool detach
,
1695 bool has_resume
, bool resume
, Error
**errp
)
1697 bool resume_requested
;
1698 Error
*local_err
= NULL
;
1699 MigrationState
*s
= migrate_get_current();
1700 const char *p
= NULL
;
1702 /* URI is not suitable for migration? */
1703 if (!migration_channels_and_uri_compatible(uri
, errp
)) {
1707 resume_requested
= has_resume
&& resume
;
1708 if (!migrate_prepare(s
, has_blk
&& blk
, has_inc
&& inc
,
1709 resume_requested
, errp
)) {
1710 /* Error detected, put into errp */
1714 if (!resume_requested
) {
1715 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1720 if (strstart(uri
, "tcp:", &p
) ||
1721 strstart(uri
, "unix:", NULL
) ||
1722 strstart(uri
, "vsock:", NULL
)) {
1723 socket_start_outgoing_migration(s
, p
? p
: uri
, &local_err
);
1725 } else if (strstart(uri
, "rdma:", &p
)) {
1726 rdma_start_outgoing_migration(s
, p
, &local_err
);
1728 } else if (strstart(uri
, "exec:", &p
)) {
1729 exec_start_outgoing_migration(s
, p
, &local_err
);
1730 } else if (strstart(uri
, "fd:", &p
)) {
1731 fd_start_outgoing_migration(s
, p
, &local_err
);
1732 } else if (strstart(uri
, "file:", &p
)) {
1733 file_start_outgoing_migration(s
, p
, &local_err
);
1735 error_setg(&local_err
, QERR_INVALID_PARAMETER_VALUE
, "uri",
1736 "a valid migration protocol");
1737 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1738 MIGRATION_STATUS_FAILED
);
1739 block_cleanup_parameters();
1743 if (!resume_requested
) {
1744 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1746 migrate_fd_error(s
, local_err
);
1747 error_propagate(errp
, local_err
);
1752 void qmp_migrate_cancel(Error
**errp
)
1754 migration_cancel(NULL
);
1757 void qmp_migrate_continue(MigrationStatus state
, Error
**errp
)
1759 MigrationState
*s
= migrate_get_current();
1760 if (s
->state
!= state
) {
1761 error_setg(errp
, "Migration not in expected state: %s",
1762 MigrationStatus_str(s
->state
));
1765 qemu_sem_post(&s
->pause_sem
);
1768 /* migration thread support */
1770 * Something bad happened to the RP stream, mark an error
1771 * The caller shall print or trace something to indicate why
1773 static void mark_source_rp_bad(MigrationState
*s
)
1775 s
->rp_state
.error
= true;
1778 void migration_rp_wait(MigrationState
*s
)
1780 qemu_sem_wait(&s
->rp_state
.rp_sem
);
1783 void migration_rp_kick(MigrationState
*s
)
1785 qemu_sem_post(&s
->rp_state
.rp_sem
);
1788 static struct rp_cmd_args
{
1789 ssize_t len
; /* -1 = variable */
1792 [MIG_RP_MSG_INVALID
] = { .len
= -1, .name
= "INVALID" },
1793 [MIG_RP_MSG_SHUT
] = { .len
= 4, .name
= "SHUT" },
1794 [MIG_RP_MSG_PONG
] = { .len
= 4, .name
= "PONG" },
1795 [MIG_RP_MSG_REQ_PAGES
] = { .len
= 12, .name
= "REQ_PAGES" },
1796 [MIG_RP_MSG_REQ_PAGES_ID
] = { .len
= -1, .name
= "REQ_PAGES_ID" },
1797 [MIG_RP_MSG_RECV_BITMAP
] = { .len
= -1, .name
= "RECV_BITMAP" },
1798 [MIG_RP_MSG_RESUME_ACK
] = { .len
= 4, .name
= "RESUME_ACK" },
1799 [MIG_RP_MSG_SWITCHOVER_ACK
] = { .len
= 0, .name
= "SWITCHOVER_ACK" },
1800 [MIG_RP_MSG_MAX
] = { .len
= -1, .name
= "MAX" },
1804 * Process a request for pages received on the return path,
1805 * We're allowed to send more than requested (e.g. to round to our page size)
1806 * and we don't need to send pages that have already been sent.
1808 static void migrate_handle_rp_req_pages(MigrationState
*ms
, const char* rbname
,
1809 ram_addr_t start
, size_t len
)
1811 long our_host_ps
= qemu_real_host_page_size();
1813 trace_migrate_handle_rp_req_pages(rbname
, start
, len
);
1816 * Since we currently insist on matching page sizes, just sanity check
1817 * we're being asked for whole host pages.
1819 if (!QEMU_IS_ALIGNED(start
, our_host_ps
) ||
1820 !QEMU_IS_ALIGNED(len
, our_host_ps
)) {
1821 error_report("%s: Misaligned page request, start: " RAM_ADDR_FMT
1822 " len: %zd", __func__
, start
, len
);
1823 mark_source_rp_bad(ms
);
1827 if (ram_save_queue_pages(rbname
, start
, len
)) {
1828 mark_source_rp_bad(ms
);
1832 static int migrate_handle_rp_recv_bitmap(MigrationState
*s
, char *block_name
)
1834 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
1837 error_report("%s: invalid block name '%s'", __func__
, block_name
);
1841 /* Fetch the received bitmap and refresh the dirty bitmap */
1842 return ram_dirty_bitmap_reload(s
, block
);
1845 static int migrate_handle_rp_resume_ack(MigrationState
*s
, uint32_t value
)
1847 trace_source_return_path_thread_resume_ack(value
);
1849 if (value
!= MIGRATION_RESUME_ACK_VALUE
) {
1850 error_report("%s: illegal resume_ack value %"PRIu32
,
1855 /* Now both sides are active. */
1856 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_RECOVER
,
1857 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1859 /* Notify send thread that time to continue send pages */
1860 migration_rp_kick(s
);
1866 * Release ms->rp_state.from_dst_file (and postcopy_qemufile_src if
1867 * existed) in a safe way.
1869 static void migration_release_dst_files(MigrationState
*ms
)
1873 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
1875 * Reset the from_dst_file pointer first before releasing it, as we
1876 * can't block within lock section
1878 file
= ms
->rp_state
.from_dst_file
;
1879 ms
->rp_state
.from_dst_file
= NULL
;
1883 * Do the same to postcopy fast path socket too if there is. No
1884 * locking needed because this qemufile should only be managed by
1885 * return path thread.
1887 if (ms
->postcopy_qemufile_src
) {
1888 migration_ioc_unregister_yank_from_file(ms
->postcopy_qemufile_src
);
1889 qemu_file_shutdown(ms
->postcopy_qemufile_src
);
1890 qemu_fclose(ms
->postcopy_qemufile_src
);
1891 ms
->postcopy_qemufile_src
= NULL
;
1898 * Handles messages sent on the return path towards the source VM
1901 static void *source_return_path_thread(void *opaque
)
1903 MigrationState
*ms
= opaque
;
1904 QEMUFile
*rp
= ms
->rp_state
.from_dst_file
;
1905 uint16_t header_len
, header_type
;
1907 uint32_t tmp32
, sibling_error
;
1908 ram_addr_t start
= 0; /* =0 to silence warning */
1909 size_t len
= 0, expected_len
;
1912 trace_source_return_path_thread_entry();
1913 rcu_register_thread();
1915 while (!ms
->rp_state
.error
&& !qemu_file_get_error(rp
) &&
1916 migration_is_setup_or_active(ms
->state
)) {
1917 trace_source_return_path_thread_loop_top();
1918 header_type
= qemu_get_be16(rp
);
1919 header_len
= qemu_get_be16(rp
);
1921 if (qemu_file_get_error(rp
)) {
1922 mark_source_rp_bad(ms
);
1926 if (header_type
>= MIG_RP_MSG_MAX
||
1927 header_type
== MIG_RP_MSG_INVALID
) {
1928 error_report("RP: Received invalid message 0x%04x length 0x%04x",
1929 header_type
, header_len
);
1930 mark_source_rp_bad(ms
);
1934 if ((rp_cmd_args
[header_type
].len
!= -1 &&
1935 header_len
!= rp_cmd_args
[header_type
].len
) ||
1936 header_len
> sizeof(buf
)) {
1937 error_report("RP: Received '%s' message (0x%04x) with"
1938 "incorrect length %d expecting %zu",
1939 rp_cmd_args
[header_type
].name
, header_type
, header_len
,
1940 (size_t)rp_cmd_args
[header_type
].len
);
1941 mark_source_rp_bad(ms
);
1945 /* We know we've got a valid header by this point */
1946 res
= qemu_get_buffer(rp
, buf
, header_len
);
1947 if (res
!= header_len
) {
1948 error_report("RP: Failed reading data for message 0x%04x"
1949 " read %d expected %d",
1950 header_type
, res
, header_len
);
1951 mark_source_rp_bad(ms
);
1955 /* OK, we have the message and the data */
1956 switch (header_type
) {
1957 case MIG_RP_MSG_SHUT
:
1958 sibling_error
= ldl_be_p(buf
);
1959 trace_source_return_path_thread_shut(sibling_error
);
1960 if (sibling_error
) {
1961 error_report("RP: Sibling indicated error %d", sibling_error
);
1962 mark_source_rp_bad(ms
);
1965 * We'll let the main thread deal with closing the RP
1966 * we could do a shutdown(2) on it, but we're the only user
1967 * anyway, so there's nothing gained.
1971 case MIG_RP_MSG_PONG
:
1972 tmp32
= ldl_be_p(buf
);
1973 trace_source_return_path_thread_pong(tmp32
);
1974 qemu_sem_post(&ms
->rp_state
.rp_pong_acks
);
1977 case MIG_RP_MSG_REQ_PAGES
:
1978 start
= ldq_be_p(buf
);
1979 len
= ldl_be_p(buf
+ 8);
1980 migrate_handle_rp_req_pages(ms
, NULL
, start
, len
);
1983 case MIG_RP_MSG_REQ_PAGES_ID
:
1984 expected_len
= 12 + 1; /* header + termination */
1986 if (header_len
>= expected_len
) {
1987 start
= ldq_be_p(buf
);
1988 len
= ldl_be_p(buf
+ 8);
1989 /* Now we expect an idstr */
1990 tmp32
= buf
[12]; /* Length of the following idstr */
1991 buf
[13 + tmp32
] = '\0';
1992 expected_len
+= tmp32
;
1994 if (header_len
!= expected_len
) {
1995 error_report("RP: Req_Page_id with length %d expecting %zd",
1996 header_len
, expected_len
);
1997 mark_source_rp_bad(ms
);
2000 migrate_handle_rp_req_pages(ms
, (char *)&buf
[13], start
, len
);
2003 case MIG_RP_MSG_RECV_BITMAP
:
2004 if (header_len
< 1) {
2005 error_report("%s: missing block name", __func__
);
2006 mark_source_rp_bad(ms
);
2009 /* Format: len (1B) + idstr (<255B). This ends the idstr. */
2010 buf
[buf
[0] + 1] = '\0';
2011 if (migrate_handle_rp_recv_bitmap(ms
, (char *)(buf
+ 1))) {
2012 mark_source_rp_bad(ms
);
2017 case MIG_RP_MSG_RESUME_ACK
:
2018 tmp32
= ldl_be_p(buf
);
2019 if (migrate_handle_rp_resume_ack(ms
, tmp32
)) {
2020 mark_source_rp_bad(ms
);
2025 case MIG_RP_MSG_SWITCHOVER_ACK
:
2026 ms
->switchover_acked
= true;
2027 trace_source_return_path_thread_switchover_acked();
2036 if (qemu_file_get_error(rp
)) {
2037 trace_source_return_path_thread_bad_end();
2038 mark_source_rp_bad(ms
);
2041 trace_source_return_path_thread_end();
2042 rcu_unregister_thread();
2046 static int open_return_path_on_source(MigrationState
*ms
)
2048 ms
->rp_state
.from_dst_file
= qemu_file_get_return_path(ms
->to_dst_file
);
2049 if (!ms
->rp_state
.from_dst_file
) {
2053 trace_open_return_path_on_source();
2055 qemu_thread_create(&ms
->rp_state
.rp_thread
, "return path",
2056 source_return_path_thread
, ms
, QEMU_THREAD_JOINABLE
);
2057 ms
->rp_state
.rp_thread_created
= true;
2059 trace_open_return_path_on_source_continue();
2064 static int close_return_path_on_source(MigrationState
*ms
)
2068 if (!ms
->rp_state
.rp_thread_created
) {
2072 trace_migration_return_path_end_before();
2075 * If this is a normal exit then the destination will send a SHUT
2076 * and the rp_thread will exit, however if there's an error we
2077 * need to cause it to exit. shutdown(2), if we have it, will
2078 * cause it to unblock if it's stuck waiting for the destination.
2080 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2081 if (ms
->to_dst_file
&& ms
->rp_state
.from_dst_file
&&
2082 qemu_file_get_error(ms
->to_dst_file
)) {
2083 qemu_file_shutdown(ms
->rp_state
.from_dst_file
);
2087 trace_await_return_path_close_on_source_joining();
2088 qemu_thread_join(&ms
->rp_state
.rp_thread
);
2089 ms
->rp_state
.rp_thread_created
= false;
2090 trace_await_return_path_close_on_source_close();
2092 ret
= ms
->rp_state
.error
;
2093 ms
->rp_state
.error
= false;
2095 migration_release_dst_files(ms
);
2097 trace_migration_return_path_end_after(ret
);
2102 migration_wait_main_channel(MigrationState
*ms
)
2104 /* Wait until one PONG message received */
2105 qemu_sem_wait(&ms
->rp_state
.rp_pong_acks
);
2109 * Switch from normal iteration to postcopy
2110 * Returns non-0 on error
2112 static int postcopy_start(MigrationState
*ms
, Error
**errp
)
2115 QIOChannelBuffer
*bioc
;
2117 int64_t time_at_stop
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2118 uint64_t bandwidth
= migrate_max_postcopy_bandwidth();
2119 bool restart_block
= false;
2120 int cur_state
= MIGRATION_STATUS_ACTIVE
;
2122 if (migrate_postcopy_preempt()) {
2123 migration_wait_main_channel(ms
);
2124 if (postcopy_preempt_establish_channel(ms
)) {
2125 migrate_set_state(&ms
->state
, ms
->state
, MIGRATION_STATUS_FAILED
);
2130 if (!migrate_pause_before_switchover()) {
2131 migrate_set_state(&ms
->state
, MIGRATION_STATUS_ACTIVE
,
2132 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2135 trace_postcopy_start();
2136 qemu_mutex_lock_iothread();
2137 trace_postcopy_start_set_run();
2139 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
, NULL
);
2140 global_state_store();
2141 ret
= vm_stop_force_state(RUN_STATE_FINISH_MIGRATE
);
2146 ret
= migration_maybe_pause(ms
, &cur_state
,
2147 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2152 ret
= bdrv_inactivate_all();
2156 restart_block
= true;
2159 * Cause any non-postcopiable, but iterative devices to
2160 * send out their final data.
2162 qemu_savevm_state_complete_precopy(ms
->to_dst_file
, true, false);
2165 * in Finish migrate and with the io-lock held everything should
2166 * be quiet, but we've potentially still got dirty pages and we
2167 * need to tell the destination to throw any pages it's already received
2170 if (migrate_postcopy_ram()) {
2171 ram_postcopy_send_discard_bitmap(ms
);
2175 * send rest of state - note things that are doing postcopy
2176 * will notice we're in POSTCOPY_ACTIVE and not actually
2177 * wrap their state up here
2179 migration_rate_set(bandwidth
);
2180 if (migrate_postcopy_ram()) {
2181 /* Ping just for debugging, helps line traces up */
2182 qemu_savevm_send_ping(ms
->to_dst_file
, 2);
2186 * While loading the device state we may trigger page transfer
2187 * requests and the fd must be free to process those, and thus
2188 * the destination must read the whole device state off the fd before
2189 * it starts processing it. Unfortunately the ad-hoc migration format
2190 * doesn't allow the destination to know the size to read without fully
2191 * parsing it through each devices load-state code (especially the open
2192 * coded devices that use get/put).
2193 * So we wrap the device state up in a package with a length at the start;
2194 * to do this we use a qemu_buf to hold the whole of the device state.
2196 bioc
= qio_channel_buffer_new(4096);
2197 qio_channel_set_name(QIO_CHANNEL(bioc
), "migration-postcopy-buffer");
2198 fb
= qemu_file_new_output(QIO_CHANNEL(bioc
));
2199 object_unref(OBJECT(bioc
));
2202 * Make sure the receiver can get incoming pages before we send the rest
2205 qemu_savevm_send_postcopy_listen(fb
);
2207 qemu_savevm_state_complete_precopy(fb
, false, false);
2208 if (migrate_postcopy_ram()) {
2209 qemu_savevm_send_ping(fb
, 3);
2212 qemu_savevm_send_postcopy_run(fb
);
2214 /* <><> end of stuff going into the package */
2216 /* Last point of recovery; as soon as we send the package the destination
2217 * can open devices and potentially start running.
2218 * Lets just check again we've not got any errors.
2220 ret
= qemu_file_get_error(ms
->to_dst_file
);
2222 error_setg(errp
, "postcopy_start: Migration stream errored (pre package)");
2226 restart_block
= false;
2228 /* Now send that blob */
2229 if (qemu_savevm_send_packaged(ms
->to_dst_file
, bioc
->data
, bioc
->usage
)) {
2234 /* Send a notify to give a chance for anything that needs to happen
2235 * at the transition to postcopy and after the device state; in particular
2236 * spice needs to trigger a transition now
2238 ms
->postcopy_after_devices
= true;
2239 notifier_list_notify(&migration_state_notifiers
, ms
);
2241 ms
->downtime
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) - time_at_stop
;
2243 qemu_mutex_unlock_iothread();
2245 if (migrate_postcopy_ram()) {
2247 * Although this ping is just for debug, it could potentially be
2248 * used for getting a better measurement of downtime at the source.
2250 qemu_savevm_send_ping(ms
->to_dst_file
, 4);
2253 if (migrate_release_ram()) {
2254 ram_postcopy_migrated_memory_release(ms
);
2257 ret
= qemu_file_get_error(ms
->to_dst_file
);
2259 error_setg(errp
, "postcopy_start: Migration stream errored");
2260 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2261 MIGRATION_STATUS_FAILED
);
2264 trace_postcopy_preempt_enabled(migrate_postcopy_preempt());
2271 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2272 MIGRATION_STATUS_FAILED
);
2273 if (restart_block
) {
2274 /* A failure happened early enough that we know the destination hasn't
2275 * accessed block devices, so we're safe to recover.
2277 Error
*local_err
= NULL
;
2279 bdrv_activate_all(&local_err
);
2281 error_report_err(local_err
);
2284 qemu_mutex_unlock_iothread();
2289 * migration_maybe_pause: Pause if required to by
2290 * migrate_pause_before_switchover called with the iothread locked
2291 * Returns: 0 on success
2293 static int migration_maybe_pause(MigrationState
*s
,
2294 int *current_active_state
,
2297 if (!migrate_pause_before_switchover()) {
2301 /* Since leaving this state is not atomic with posting the semaphore
2302 * it's possible that someone could have issued multiple migrate_continue
2303 * and the semaphore is incorrectly positive at this point;
2304 * the docs say it's undefined to reinit a semaphore that's already
2305 * init'd, so use timedwait to eat up any existing posts.
2307 while (qemu_sem_timedwait(&s
->pause_sem
, 1) == 0) {
2308 /* This block intentionally left blank */
2312 * If the migration is cancelled when it is in the completion phase,
2313 * the migration state is set to MIGRATION_STATUS_CANCELLING.
2314 * So we don't need to wait a semaphore, otherwise we would always
2315 * wait for the 'pause_sem' semaphore.
2317 if (s
->state
!= MIGRATION_STATUS_CANCELLING
) {
2318 qemu_mutex_unlock_iothread();
2319 migrate_set_state(&s
->state
, *current_active_state
,
2320 MIGRATION_STATUS_PRE_SWITCHOVER
);
2321 qemu_sem_wait(&s
->pause_sem
);
2322 migrate_set_state(&s
->state
, MIGRATION_STATUS_PRE_SWITCHOVER
,
2324 *current_active_state
= new_state
;
2325 qemu_mutex_lock_iothread();
2328 return s
->state
== new_state
? 0 : -EINVAL
;
2331 static int migration_completion_precopy(MigrationState
*s
,
2332 int *current_active_state
)
2336 qemu_mutex_lock_iothread();
2337 s
->downtime_start
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2338 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
, NULL
);
2340 s
->vm_old_state
= runstate_get();
2341 global_state_store();
2343 ret
= vm_stop_force_state(RUN_STATE_FINISH_MIGRATE
);
2344 trace_migration_completion_vm_stop(ret
);
2349 ret
= migration_maybe_pause(s
, current_active_state
,
2350 MIGRATION_STATUS_DEVICE
);
2356 * Inactivate disks except in COLO, and track that we have done so in order
2357 * to remember to reactivate them if migration fails or is cancelled.
2359 s
->block_inactive
= !migrate_colo();
2360 migration_rate_set(RATE_LIMIT_DISABLED
);
2361 ret
= qemu_savevm_state_complete_precopy(s
->to_dst_file
, false,
2364 qemu_mutex_unlock_iothread();
2368 static void migration_completion_postcopy(MigrationState
*s
)
2370 trace_migration_completion_postcopy_end();
2372 qemu_mutex_lock_iothread();
2373 qemu_savevm_state_complete_postcopy(s
->to_dst_file
);
2374 qemu_mutex_unlock_iothread();
2377 * Shutdown the postcopy fast path thread. This is only needed when dest
2378 * QEMU binary is old (7.1/7.2). QEMU 8.0+ doesn't need this.
2380 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
2381 postcopy_preempt_shutdown_file(s
);
2384 trace_migration_completion_postcopy_end_after_complete();
2387 static void migration_completion_failed(MigrationState
*s
,
2388 int current_active_state
)
2390 if (s
->block_inactive
&& (s
->state
== MIGRATION_STATUS_ACTIVE
||
2391 s
->state
== MIGRATION_STATUS_DEVICE
)) {
2393 * If not doing postcopy, vm_start() will be called: let's
2394 * regain control on images.
2396 Error
*local_err
= NULL
;
2398 qemu_mutex_lock_iothread();
2399 bdrv_activate_all(&local_err
);
2401 error_report_err(local_err
);
2403 s
->block_inactive
= false;
2405 qemu_mutex_unlock_iothread();
2408 migrate_set_state(&s
->state
, current_active_state
,
2409 MIGRATION_STATUS_FAILED
);
2413 * migration_completion: Used by migration_thread when there's not much left.
2414 * The caller 'breaks' the loop when this returns.
2416 * @s: Current migration state
2418 static void migration_completion(MigrationState
*s
)
2421 int current_active_state
= s
->state
;
2423 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2424 ret
= migration_completion_precopy(s
, ¤t_active_state
);
2425 } else if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2426 migration_completion_postcopy(s
);
2435 if (close_return_path_on_source(s
)) {
2439 if (qemu_file_get_error(s
->to_dst_file
)) {
2440 trace_migration_completion_file_err();
2444 if (migrate_colo() && s
->state
== MIGRATION_STATUS_ACTIVE
) {
2445 /* COLO does not support postcopy */
2446 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
2447 MIGRATION_STATUS_COLO
);
2449 migrate_set_state(&s
->state
, current_active_state
,
2450 MIGRATION_STATUS_COMPLETED
);
2456 migration_completion_failed(s
, current_active_state
);
2460 * bg_migration_completion: Used by bg_migration_thread when after all the
2461 * RAM has been saved. The caller 'breaks' the loop when this returns.
2463 * @s: Current migration state
2465 static void bg_migration_completion(MigrationState
*s
)
2467 int current_active_state
= s
->state
;
2469 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2471 * By this moment we have RAM content saved into the migration stream.
2472 * The next step is to flush the non-RAM content (device state)
2473 * right after the ram content. The device state has been stored into
2474 * the temporary buffer before RAM saving started.
2476 qemu_put_buffer(s
->to_dst_file
, s
->bioc
->data
, s
->bioc
->usage
);
2477 qemu_fflush(s
->to_dst_file
);
2478 } else if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
2482 if (qemu_file_get_error(s
->to_dst_file
)) {
2483 trace_migration_completion_file_err();
2487 migrate_set_state(&s
->state
, current_active_state
,
2488 MIGRATION_STATUS_COMPLETED
);
2492 migrate_set_state(&s
->state
, current_active_state
,
2493 MIGRATION_STATUS_FAILED
);
2496 typedef enum MigThrError
{
2497 /* No error detected */
2498 MIG_THR_ERR_NONE
= 0,
2499 /* Detected error, but resumed successfully */
2500 MIG_THR_ERR_RECOVERED
= 1,
2501 /* Detected fatal error, need to exit */
2502 MIG_THR_ERR_FATAL
= 2,
2505 static int postcopy_resume_handshake(MigrationState
*s
)
2507 qemu_savevm_send_postcopy_resume(s
->to_dst_file
);
2509 while (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2510 migration_rp_wait(s
);
2513 if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2520 /* Return zero if success, or <0 for error */
2521 static int postcopy_do_resume(MigrationState
*s
)
2526 * Call all the resume_prepare() hooks, so that modules can be
2527 * ready for the migration resume.
2529 ret
= qemu_savevm_state_resume_prepare(s
);
2531 error_report("%s: resume_prepare() failure detected: %d",
2537 * If preempt is enabled, re-establish the preempt channel. Note that
2538 * we do it after resume prepare to make sure the main channel will be
2539 * created before the preempt channel. E.g. with weak network, the
2540 * dest QEMU may get messed up with the preempt and main channels on
2541 * the order of connection setup. This guarantees the correct order.
2543 ret
= postcopy_preempt_establish_channel(s
);
2545 error_report("%s: postcopy_preempt_establish_channel(): %d",
2551 * Last handshake with destination on the resume (destination will
2552 * switch to postcopy-active afterwards)
2554 ret
= postcopy_resume_handshake(s
);
2556 error_report("%s: handshake failed: %d", __func__
, ret
);
2564 * We don't return until we are in a safe state to continue current
2565 * postcopy migration. Returns MIG_THR_ERR_RECOVERED if recovered, or
2566 * MIG_THR_ERR_FATAL if unrecovery failure happened.
2568 static MigThrError
postcopy_pause(MigrationState
*s
)
2570 assert(s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2576 * Current channel is possibly broken. Release it. Note that this is
2577 * guaranteed even without lock because to_dst_file should only be
2578 * modified by the migration thread. That also guarantees that the
2579 * unregister of yank is safe too without the lock. It should be safe
2580 * even to be within the qemu_file_lock, but we didn't do that to avoid
2581 * taking more mutex (yank_lock) within qemu_file_lock. TL;DR: we make
2582 * the qemu_file_lock critical section as small as possible.
2584 assert(s
->to_dst_file
);
2585 migration_ioc_unregister_yank_from_file(s
->to_dst_file
);
2586 qemu_mutex_lock(&s
->qemu_file_lock
);
2587 file
= s
->to_dst_file
;
2588 s
->to_dst_file
= NULL
;
2589 qemu_mutex_unlock(&s
->qemu_file_lock
);
2591 qemu_file_shutdown(file
);
2595 * We're already pausing, so ignore any errors on the return
2596 * path and just wait for the thread to finish. It will be
2597 * re-created when we resume.
2599 close_return_path_on_source(s
);
2601 migrate_set_state(&s
->state
, s
->state
,
2602 MIGRATION_STATUS_POSTCOPY_PAUSED
);
2604 error_report("Detected IO failure for postcopy. "
2605 "Migration paused.");
2608 * We wait until things fixed up. Then someone will setup the
2609 * status back for us.
2611 while (s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
2612 qemu_sem_wait(&s
->postcopy_pause_sem
);
2615 if (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2616 /* Woken up by a recover procedure. Give it a shot */
2618 /* Do the resume logic */
2619 if (postcopy_do_resume(s
) == 0) {
2620 /* Let's continue! */
2621 trace_postcopy_pause_continued();
2622 return MIG_THR_ERR_RECOVERED
;
2625 * Something wrong happened during the recovery, let's
2626 * pause again. Pause is always better than throwing
2632 /* This is not right... Time to quit. */
2633 return MIG_THR_ERR_FATAL
;
2638 static MigThrError
migration_detect_error(MigrationState
*s
)
2641 int state
= s
->state
;
2642 Error
*local_error
= NULL
;
2644 if (state
== MIGRATION_STATUS_CANCELLING
||
2645 state
== MIGRATION_STATUS_CANCELLED
) {
2646 /* End the migration, but don't set the state to failed */
2647 return MIG_THR_ERR_FATAL
;
2651 * Try to detect any file errors. Note that postcopy_qemufile_src will
2652 * be NULL when postcopy preempt is not enabled.
2654 ret
= qemu_file_get_error_obj_any(s
->to_dst_file
,
2655 s
->postcopy_qemufile_src
,
2658 /* Everything is fine */
2659 assert(!local_error
);
2660 return MIG_THR_ERR_NONE
;
2664 migrate_set_error(s
, local_error
);
2665 error_free(local_error
);
2668 if (state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
&& ret
) {
2670 * For postcopy, we allow the network to be down for a
2671 * while. After that, it can be continued by a
2674 return postcopy_pause(s
);
2677 * For precopy (or postcopy with error outside IO), we fail
2680 migrate_set_state(&s
->state
, state
, MIGRATION_STATUS_FAILED
);
2681 trace_migration_thread_file_err();
2683 /* Time to stop the migration, now. */
2684 return MIG_THR_ERR_FATAL
;
2688 static void migration_calculate_complete(MigrationState
*s
)
2690 uint64_t bytes
= migration_transferred_bytes(s
->to_dst_file
);
2691 int64_t end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2692 int64_t transfer_time
;
2694 s
->total_time
= end_time
- s
->start_time
;
2697 * It's still not set, so we are precopy migration. For
2698 * postcopy, downtime is calculated during postcopy_start().
2700 s
->downtime
= end_time
- s
->downtime_start
;
2703 transfer_time
= s
->total_time
- s
->setup_time
;
2704 if (transfer_time
) {
2705 s
->mbps
= ((double) bytes
* 8.0) / transfer_time
/ 1000;
2709 static void update_iteration_initial_status(MigrationState
*s
)
2712 * Update these three fields at the same time to avoid mismatch info lead
2713 * wrong speed calculation.
2715 s
->iteration_start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2716 s
->iteration_initial_bytes
= migration_transferred_bytes(s
->to_dst_file
);
2717 s
->iteration_initial_pages
= ram_get_total_transferred_pages();
2720 static void migration_update_counters(MigrationState
*s
,
2721 int64_t current_time
)
2723 uint64_t transferred
, transferred_pages
, time_spent
;
2724 uint64_t current_bytes
; /* bytes transferred since the beginning */
2725 uint64_t switchover_bw
;
2726 /* Expected bandwidth when switching over to destination QEMU */
2727 double expected_bw_per_ms
;
2730 if (current_time
< s
->iteration_start_time
+ BUFFER_DELAY
) {
2734 switchover_bw
= migrate_avail_switchover_bandwidth();
2735 current_bytes
= migration_transferred_bytes(s
->to_dst_file
);
2736 transferred
= current_bytes
- s
->iteration_initial_bytes
;
2737 time_spent
= current_time
- s
->iteration_start_time
;
2738 bandwidth
= (double)transferred
/ time_spent
;
2740 if (switchover_bw
) {
2742 * If the user specified a switchover bandwidth, let's trust the
2743 * user so that can be more accurate than what we estimated.
2745 expected_bw_per_ms
= switchover_bw
/ 1000;
2747 /* If the user doesn't specify bandwidth, we use the estimated */
2748 expected_bw_per_ms
= bandwidth
;
2751 s
->threshold_size
= expected_bw_per_ms
* migrate_downtime_limit();
2753 s
->mbps
= (((double) transferred
* 8.0) /
2754 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0;
2756 transferred_pages
= ram_get_total_transferred_pages() -
2757 s
->iteration_initial_pages
;
2758 s
->pages_per_second
= (double) transferred_pages
/
2759 (((double) time_spent
/ 1000.0));
2762 * if we haven't sent anything, we don't want to
2763 * recalculate. 10000 is a small enough number for our purposes
2765 if (stat64_get(&mig_stats
.dirty_pages_rate
) &&
2766 transferred
> 10000) {
2767 s
->expected_downtime
=
2768 stat64_get(&mig_stats
.dirty_bytes_last_sync
) / expected_bw_per_ms
;
2771 migration_rate_reset(s
->to_dst_file
);
2773 update_iteration_initial_status(s
);
2775 trace_migrate_transferred(transferred
, time_spent
,
2776 /* Both in unit bytes/ms */
2777 bandwidth
, switchover_bw
/ 1000,
2781 static bool migration_can_switchover(MigrationState
*s
)
2783 if (!migrate_switchover_ack()) {
2787 /* No reason to wait for switchover ACK if VM is stopped */
2788 if (!runstate_is_running()) {
2792 return s
->switchover_acked
;
2795 /* Migration thread iteration status */
2797 MIG_ITERATE_RESUME
, /* Resume current iteration */
2798 MIG_ITERATE_SKIP
, /* Skip current iteration */
2799 MIG_ITERATE_BREAK
, /* Break the loop */
2803 * Return true if continue to the next iteration directly, false
2806 static MigIterateState
migration_iteration_run(MigrationState
*s
)
2808 uint64_t must_precopy
, can_postcopy
;
2809 Error
*local_err
= NULL
;
2810 bool in_postcopy
= s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
;
2811 bool can_switchover
= migration_can_switchover(s
);
2813 qemu_savevm_state_pending_estimate(&must_precopy
, &can_postcopy
);
2814 uint64_t pending_size
= must_precopy
+ can_postcopy
;
2816 trace_migrate_pending_estimate(pending_size
, must_precopy
, can_postcopy
);
2818 if (must_precopy
<= s
->threshold_size
) {
2819 qemu_savevm_state_pending_exact(&must_precopy
, &can_postcopy
);
2820 pending_size
= must_precopy
+ can_postcopy
;
2821 trace_migrate_pending_exact(pending_size
, must_precopy
, can_postcopy
);
2824 if ((!pending_size
|| pending_size
< s
->threshold_size
) && can_switchover
) {
2825 trace_migration_thread_low_pending(pending_size
);
2826 migration_completion(s
);
2827 return MIG_ITERATE_BREAK
;
2830 /* Still a significant amount to transfer */
2831 if (!in_postcopy
&& must_precopy
<= s
->threshold_size
&& can_switchover
&&
2832 qatomic_read(&s
->start_postcopy
)) {
2833 if (postcopy_start(s
, &local_err
)) {
2834 migrate_set_error(s
, local_err
);
2835 error_report_err(local_err
);
2837 return MIG_ITERATE_SKIP
;
2840 /* Just another iteration step */
2841 qemu_savevm_state_iterate(s
->to_dst_file
, in_postcopy
);
2842 return MIG_ITERATE_RESUME
;
2845 static void migration_iteration_finish(MigrationState
*s
)
2847 /* If we enabled cpu throttling for auto-converge, turn it off. */
2848 cpu_throttle_stop();
2850 qemu_mutex_lock_iothread();
2852 case MIGRATION_STATUS_COMPLETED
:
2853 migration_calculate_complete(s
);
2854 runstate_set(RUN_STATE_POSTMIGRATE
);
2856 case MIGRATION_STATUS_COLO
:
2857 assert(migrate_colo());
2858 migrate_start_colo_process(s
);
2859 s
->vm_old_state
= RUN_STATE_RUNNING
;
2861 case MIGRATION_STATUS_FAILED
:
2862 case MIGRATION_STATUS_CANCELLED
:
2863 case MIGRATION_STATUS_CANCELLING
:
2864 if (s
->vm_old_state
== RUN_STATE_RUNNING
) {
2865 if (!runstate_check(RUN_STATE_SHUTDOWN
)) {
2869 if (runstate_check(RUN_STATE_FINISH_MIGRATE
)) {
2870 runstate_set(s
->vm_old_state
);
2876 /* Should not reach here, but if so, forgive the VM. */
2877 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
2880 migrate_fd_cleanup_schedule(s
);
2881 qemu_mutex_unlock_iothread();
2884 static void bg_migration_iteration_finish(MigrationState
*s
)
2887 * Stop tracking RAM writes - un-protect memory, un-register UFFD
2888 * memory ranges, flush kernel wait queues and wake up threads
2889 * waiting for write fault to be resolved.
2891 ram_write_tracking_stop();
2893 qemu_mutex_lock_iothread();
2895 case MIGRATION_STATUS_COMPLETED
:
2896 migration_calculate_complete(s
);
2899 case MIGRATION_STATUS_ACTIVE
:
2900 case MIGRATION_STATUS_FAILED
:
2901 case MIGRATION_STATUS_CANCELLED
:
2902 case MIGRATION_STATUS_CANCELLING
:
2906 /* Should not reach here, but if so, forgive the VM. */
2907 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
2911 migrate_fd_cleanup_schedule(s
);
2912 qemu_mutex_unlock_iothread();
2916 * Return true if continue to the next iteration directly, false
2919 static MigIterateState
bg_migration_iteration_run(MigrationState
*s
)
2923 res
= qemu_savevm_state_iterate(s
->to_dst_file
, false);
2925 bg_migration_completion(s
);
2926 return MIG_ITERATE_BREAK
;
2929 return MIG_ITERATE_RESUME
;
2932 void migration_make_urgent_request(void)
2934 qemu_sem_post(&migrate_get_current()->rate_limit_sem
);
2937 void migration_consume_urgent_request(void)
2939 qemu_sem_wait(&migrate_get_current()->rate_limit_sem
);
2942 /* Returns true if the rate limiting was broken by an urgent request */
2943 bool migration_rate_limit(void)
2945 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2946 MigrationState
*s
= migrate_get_current();
2948 bool urgent
= false;
2949 migration_update_counters(s
, now
);
2950 if (migration_rate_exceeded(s
->to_dst_file
)) {
2952 if (qemu_file_get_error(s
->to_dst_file
)) {
2956 * Wait for a delay to do rate limiting OR
2957 * something urgent to post the semaphore.
2959 int ms
= s
->iteration_start_time
+ BUFFER_DELAY
- now
;
2960 trace_migration_rate_limit_pre(ms
);
2961 if (qemu_sem_timedwait(&s
->rate_limit_sem
, ms
) == 0) {
2963 * We were woken by one or more urgent things but
2964 * the timedwait will have consumed one of them.
2965 * The service routine for the urgent wake will dec
2966 * the semaphore itself for each item it consumes,
2967 * so add this one we just eat back.
2969 qemu_sem_post(&s
->rate_limit_sem
);
2972 trace_migration_rate_limit_post(urgent
);
2978 * if failover devices are present, wait they are completely
2982 static void qemu_savevm_wait_unplug(MigrationState
*s
, int old_state
,
2985 if (qemu_savevm_state_guest_unplug_pending()) {
2986 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_WAIT_UNPLUG
);
2988 while (s
->state
== MIGRATION_STATUS_WAIT_UNPLUG
&&
2989 qemu_savevm_state_guest_unplug_pending()) {
2990 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
2992 if (s
->state
!= MIGRATION_STATUS_WAIT_UNPLUG
) {
2993 int timeout
= 120; /* 30 seconds */
2995 * migration has been canceled
2996 * but as we have started an unplug we must wait the end
2997 * to be able to plug back the card
2999 while (timeout
-- && qemu_savevm_state_guest_unplug_pending()) {
3000 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3002 if (qemu_savevm_state_guest_unplug_pending() &&
3004 warn_report("migration: partially unplugged device on "
3009 migrate_set_state(&s
->state
, MIGRATION_STATUS_WAIT_UNPLUG
, new_state
);
3011 migrate_set_state(&s
->state
, old_state
, new_state
);
3016 * Master migration thread on the source VM.
3017 * It drives the migration and pumps the data down the outgoing channel.
3019 static void *migration_thread(void *opaque
)
3021 MigrationState
*s
= opaque
;
3022 MigrationThread
*thread
= NULL
;
3023 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3024 MigThrError thr_error
;
3025 bool urgent
= false;
3027 thread
= migration_threads_add("live_migration", qemu_get_thread_id());
3029 rcu_register_thread();
3031 object_ref(OBJECT(s
));
3032 update_iteration_initial_status(s
);
3034 qemu_mutex_lock_iothread();
3035 qemu_savevm_state_header(s
->to_dst_file
);
3036 qemu_mutex_unlock_iothread();
3039 * If we opened the return path, we need to make sure dst has it
3042 if (s
->rp_state
.rp_thread_created
) {
3043 /* Now tell the dest that it should open its end so it can reply */
3044 qemu_savevm_send_open_return_path(s
->to_dst_file
);
3046 /* And do a ping that will make stuff easier to debug */
3047 qemu_savevm_send_ping(s
->to_dst_file
, 1);
3050 if (migrate_postcopy()) {
3052 * Tell the destination that we *might* want to do postcopy later;
3053 * if the other end can't do postcopy it should fail now, nice and
3056 qemu_savevm_send_postcopy_advise(s
->to_dst_file
);
3059 if (migrate_colo()) {
3060 /* Notify migration destination that we enable COLO */
3061 qemu_savevm_send_colo_enable(s
->to_dst_file
);
3064 qemu_mutex_lock_iothread();
3065 qemu_savevm_state_setup(s
->to_dst_file
);
3066 qemu_mutex_unlock_iothread();
3068 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3069 MIGRATION_STATUS_ACTIVE
);
3071 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3073 trace_migration_thread_setup_complete();
3075 while (migration_is_active(s
)) {
3076 if (urgent
|| !migration_rate_exceeded(s
->to_dst_file
)) {
3077 MigIterateState iter_state
= migration_iteration_run(s
);
3078 if (iter_state
== MIG_ITERATE_SKIP
) {
3080 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3086 * Try to detect any kind of failures, and see whether we
3087 * should stop the migration now.
3089 thr_error
= migration_detect_error(s
);
3090 if (thr_error
== MIG_THR_ERR_FATAL
) {
3091 /* Stop migration */
3093 } else if (thr_error
== MIG_THR_ERR_RECOVERED
) {
3095 * Just recovered from a e.g. network failure, reset all
3096 * the local variables. This is important to avoid
3097 * breaking transferred_bytes and bandwidth calculation
3099 update_iteration_initial_status(s
);
3102 urgent
= migration_rate_limit();
3105 trace_migration_thread_after_loop();
3106 migration_iteration_finish(s
);
3107 object_unref(OBJECT(s
));
3108 rcu_unregister_thread();
3109 migration_threads_remove(thread
);
3113 static void bg_migration_vm_start_bh(void *opaque
)
3115 MigrationState
*s
= opaque
;
3117 qemu_bh_delete(s
->vm_start_bh
);
3118 s
->vm_start_bh
= NULL
;
3121 s
->downtime
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) - s
->downtime_start
;
3125 * Background snapshot thread, based on live migration code.
3126 * This is an alternative implementation of live migration mechanism
3127 * introduced specifically to support background snapshots.
3129 * It takes advantage of userfault_fd write protection mechanism introduced
3130 * in v5.7 kernel. Compared to existing dirty page logging migration much
3131 * lesser stream traffic is produced resulting in smaller snapshot images,
3132 * simply cause of no page duplicates can get into the stream.
3134 * Another key point is that generated vmstate stream reflects machine state
3135 * 'frozen' at the beginning of snapshot creation compared to dirty page logging
3136 * mechanism, which effectively results in that saved snapshot is the state of VM
3137 * at the end of the process.
3139 static void *bg_migration_thread(void *opaque
)
3141 MigrationState
*s
= opaque
;
3142 int64_t setup_start
;
3143 MigThrError thr_error
;
3145 bool early_fail
= true;
3147 rcu_register_thread();
3148 object_ref(OBJECT(s
));
3150 migration_rate_set(RATE_LIMIT_DISABLED
);
3152 setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3154 * We want to save vmstate for the moment when migration has been
3155 * initiated but also we want to save RAM content while VM is running.
3156 * The RAM content should appear first in the vmstate. So, we first
3157 * stash the non-RAM part of the vmstate to the temporary buffer,
3158 * then write RAM part of the vmstate to the migration stream
3159 * with vCPUs running and, finally, write stashed non-RAM part of
3160 * the vmstate from the buffer to the migration stream.
3162 s
->bioc
= qio_channel_buffer_new(512 * 1024);
3163 qio_channel_set_name(QIO_CHANNEL(s
->bioc
), "vmstate-buffer");
3164 fb
= qemu_file_new_output(QIO_CHANNEL(s
->bioc
));
3165 object_unref(OBJECT(s
->bioc
));
3167 update_iteration_initial_status(s
);
3170 * Prepare for tracking memory writes with UFFD-WP - populate
3171 * RAM pages before protecting.
3174 ram_write_tracking_prepare();
3177 qemu_mutex_lock_iothread();
3178 qemu_savevm_state_header(s
->to_dst_file
);
3179 qemu_savevm_state_setup(s
->to_dst_file
);
3180 qemu_mutex_unlock_iothread();
3182 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3183 MIGRATION_STATUS_ACTIVE
);
3185 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3187 trace_migration_thread_setup_complete();
3188 s
->downtime_start
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3190 qemu_mutex_lock_iothread();
3193 * If VM is currently in suspended state, then, to make a valid runstate
3194 * transition in vm_stop_force_state() we need to wakeup it up.
3196 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
, NULL
);
3197 s
->vm_old_state
= runstate_get();
3199 global_state_store();
3200 /* Forcibly stop VM before saving state of vCPUs and devices */
3201 if (vm_stop_force_state(RUN_STATE_PAUSED
)) {
3205 * Put vCPUs in sync with shadow context structures, then
3206 * save their state to channel-buffer along with devices.
3208 cpu_synchronize_all_states();
3209 if (qemu_savevm_state_complete_precopy_non_iterable(fb
, false, false)) {
3213 * Since we are going to get non-iterable state data directly
3214 * from s->bioc->data, explicit flush is needed here.
3218 /* Now initialize UFFD context and start tracking RAM writes */
3219 if (ram_write_tracking_start()) {
3225 * Start VM from BH handler to avoid write-fault lock here.
3226 * UFFD-WP protection for the whole RAM is already enabled so
3227 * calling VM state change notifiers from vm_start() would initiate
3228 * writes to virtio VQs memory which is in write-protected region.
3230 s
->vm_start_bh
= qemu_bh_new(bg_migration_vm_start_bh
, s
);
3231 qemu_bh_schedule(s
->vm_start_bh
);
3233 qemu_mutex_unlock_iothread();
3235 while (migration_is_active(s
)) {
3236 MigIterateState iter_state
= bg_migration_iteration_run(s
);
3237 if (iter_state
== MIG_ITERATE_SKIP
) {
3239 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3244 * Try to detect any kind of failures, and see whether we
3245 * should stop the migration now.
3247 thr_error
= migration_detect_error(s
);
3248 if (thr_error
== MIG_THR_ERR_FATAL
) {
3249 /* Stop migration */
3253 migration_update_counters(s
, qemu_clock_get_ms(QEMU_CLOCK_REALTIME
));
3256 trace_migration_thread_after_loop();
3260 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3261 MIGRATION_STATUS_FAILED
);
3262 qemu_mutex_unlock_iothread();
3265 bg_migration_iteration_finish(s
);
3268 object_unref(OBJECT(s
));
3269 rcu_unregister_thread();
3274 void migrate_fd_connect(MigrationState
*s
, Error
*error_in
)
3276 Error
*local_err
= NULL
;
3277 uint64_t rate_limit
;
3278 bool resume
= s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
;
3281 * If there's a previous error, free it and prepare for another one.
3282 * Meanwhile if migration completes successfully, there won't have an error
3283 * dumped when calling migrate_fd_cleanup().
3285 migrate_error_free(s
);
3287 s
->expected_downtime
= migrate_downtime_limit();
3289 assert(s
->cleanup_bh
);
3291 assert(!s
->cleanup_bh
);
3292 s
->cleanup_bh
= qemu_bh_new(migrate_fd_cleanup_bh
, s
);
3295 migrate_fd_error(s
, error_in
);
3298 * Don't do cleanup for resume if channel is invalid, but only dump
3299 * the error. We wait for another channel connect from the user.
3300 * The error_report still gives HMP user a hint on what failed.
3301 * It's normally done in migrate_fd_cleanup(), but call it here
3304 error_report_err(error_copy(s
->error
));
3306 migrate_fd_cleanup(s
);
3312 /* This is a resumed migration */
3313 rate_limit
= migrate_max_postcopy_bandwidth();
3315 /* This is a fresh new migration */
3316 rate_limit
= migrate_max_bandwidth();
3318 /* Notify before starting migration thread */
3319 notifier_list_notify(&migration_state_notifiers
, s
);
3322 migration_rate_set(rate_limit
);
3323 qemu_file_set_blocking(s
->to_dst_file
, true);
3326 * Open the return path. For postcopy, it is used exclusively. For
3327 * precopy, only if user specified "return-path" capability would
3328 * QEMU uses the return path.
3330 if (migrate_postcopy_ram() || migrate_return_path()) {
3331 if (open_return_path_on_source(s
)) {
3332 error_setg(&local_err
, "Unable to open return-path for postcopy");
3333 migrate_set_state(&s
->state
, s
->state
, MIGRATION_STATUS_FAILED
);
3334 migrate_set_error(s
, local_err
);
3335 error_report_err(local_err
);
3336 migrate_fd_cleanup(s
);
3342 * This needs to be done before resuming a postcopy. Note: for newer
3343 * QEMUs we will delay the channel creation until postcopy_start(), to
3344 * avoid disorder of channel creations.
3346 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
3347 postcopy_preempt_setup(s
);
3351 /* Wakeup the main migration thread to do the recovery */
3352 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
3353 MIGRATION_STATUS_POSTCOPY_RECOVER
);
3354 qemu_sem_post(&s
->postcopy_pause_sem
);
3358 if (multifd_save_setup(&local_err
) != 0) {
3359 migrate_set_error(s
, local_err
);
3360 error_report_err(local_err
);
3361 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
3362 MIGRATION_STATUS_FAILED
);
3363 migrate_fd_cleanup(s
);
3367 if (migrate_background_snapshot()) {
3368 qemu_thread_create(&s
->thread
, "bg_snapshot",
3369 bg_migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3371 qemu_thread_create(&s
->thread
, "live_migration",
3372 migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3374 s
->migration_thread_running
= true;
3377 static void migration_class_init(ObjectClass
*klass
, void *data
)
3379 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3381 dc
->user_creatable
= false;
3382 device_class_set_props(dc
, migration_properties
);
3385 static void migration_instance_finalize(Object
*obj
)
3387 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3389 qemu_mutex_destroy(&ms
->error_mutex
);
3390 qemu_mutex_destroy(&ms
->qemu_file_lock
);
3391 qemu_sem_destroy(&ms
->wait_unplug_sem
);
3392 qemu_sem_destroy(&ms
->rate_limit_sem
);
3393 qemu_sem_destroy(&ms
->pause_sem
);
3394 qemu_sem_destroy(&ms
->postcopy_pause_sem
);
3395 qemu_sem_destroy(&ms
->rp_state
.rp_sem
);
3396 qemu_sem_destroy(&ms
->rp_state
.rp_pong_acks
);
3397 qemu_sem_destroy(&ms
->postcopy_qemufile_src_sem
);
3398 error_free(ms
->error
);
3401 static void migration_instance_init(Object
*obj
)
3403 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3405 ms
->state
= MIGRATION_STATUS_NONE
;
3407 ms
->pages_per_second
= -1;
3408 qemu_sem_init(&ms
->pause_sem
, 0);
3409 qemu_mutex_init(&ms
->error_mutex
);
3411 migrate_params_init(&ms
->parameters
);
3413 qemu_sem_init(&ms
->postcopy_pause_sem
, 0);
3414 qemu_sem_init(&ms
->rp_state
.rp_sem
, 0);
3415 qemu_sem_init(&ms
->rp_state
.rp_pong_acks
, 0);
3416 qemu_sem_init(&ms
->rate_limit_sem
, 0);
3417 qemu_sem_init(&ms
->wait_unplug_sem
, 0);
3418 qemu_sem_init(&ms
->postcopy_qemufile_src_sem
, 0);
3419 qemu_mutex_init(&ms
->qemu_file_lock
);
3423 * Return true if check pass, false otherwise. Error will be put
3424 * inside errp if provided.
3426 static bool migration_object_check(MigrationState
*ms
, Error
**errp
)
3428 /* Assuming all off */
3429 bool old_caps
[MIGRATION_CAPABILITY__MAX
] = { 0 };
3431 if (!migrate_params_check(&ms
->parameters
, errp
)) {
3435 return migrate_caps_check(old_caps
, ms
->capabilities
, errp
);
3438 static const TypeInfo migration_type
= {
3439 .name
= TYPE_MIGRATION
,
3441 * NOTE: TYPE_MIGRATION is not really a device, as the object is
3442 * not created using qdev_new(), it is not attached to the qdev
3443 * device tree, and it is never realized.
3445 * TODO: Make this TYPE_OBJECT once QOM provides something like
3446 * TYPE_DEVICE's "-global" properties.
3448 .parent
= TYPE_DEVICE
,
3449 .class_init
= migration_class_init
,
3450 .class_size
= sizeof(MigrationClass
),
3451 .instance_size
= sizeof(MigrationState
),
3452 .instance_init
= migration_instance_init
,
3453 .instance_finalize
= migration_instance_finalize
,
3456 static void register_migration_types(void)
3458 type_register_static(&migration_type
);
3461 type_init(register_migration_types
);