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"
69 #include "qemu/sockets.h"
70 #include "sysemu/kvm.h"
72 #define NOTIFIER_ELEM_INIT(array, elem) \
73 [elem] = NOTIFIER_WITH_RETURN_LIST_INITIALIZER((array)[elem])
75 #define INMIGRATE_DEFAULT_EXIT_ON_ERROR true
77 static NotifierWithReturnList migration_state_notifiers
[] = {
78 NOTIFIER_ELEM_INIT(migration_state_notifiers
, MIG_MODE_NORMAL
),
79 NOTIFIER_ELEM_INIT(migration_state_notifiers
, MIG_MODE_CPR_REBOOT
),
82 /* Messages sent on the return path from destination to source */
83 enum mig_rp_message_type
{
84 MIG_RP_MSG_INVALID
= 0, /* Must be 0 */
85 MIG_RP_MSG_SHUT
, /* sibling will not send any more RP messages */
86 MIG_RP_MSG_PONG
, /* Response to a PING; data (seq: be32 ) */
88 MIG_RP_MSG_REQ_PAGES_ID
, /* data (start: be64, len: be32, id: string) */
89 MIG_RP_MSG_REQ_PAGES
, /* data (start: be64, len: be32) */
90 MIG_RP_MSG_RECV_BITMAP
, /* send recved_bitmap back to source */
91 MIG_RP_MSG_RESUME_ACK
, /* tell source that we are ready to resume */
92 MIG_RP_MSG_SWITCHOVER_ACK
, /* Tell source it's OK to do switchover */
97 /* When we add fault tolerance, we could have several
98 migrations at once. For now we don't need to add
99 dynamic creation of migration */
101 static MigrationState
*current_migration
;
102 static MigrationIncomingState
*current_incoming
;
104 static GSList
*migration_blockers
[MIG_MODE__MAX
];
106 static bool migration_object_check(MigrationState
*ms
, Error
**errp
);
107 static int migration_maybe_pause(MigrationState
*s
,
108 int *current_active_state
,
110 static void migrate_fd_cancel(MigrationState
*s
);
111 static bool close_return_path_on_source(MigrationState
*s
);
112 static void migration_completion_end(MigrationState
*s
);
114 static void migration_downtime_start(MigrationState
*s
)
116 trace_vmstate_downtime_checkpoint("src-downtime-start");
117 s
->downtime_start
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
120 static void migration_downtime_end(MigrationState
*s
)
122 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
125 * If downtime already set, should mean that postcopy already set it,
126 * then that should be the real downtime already.
129 s
->downtime
= now
- s
->downtime_start
;
132 trace_vmstate_downtime_checkpoint("src-downtime-end");
135 static bool migration_needs_multiple_sockets(void)
137 return migrate_multifd() || migrate_postcopy_preempt();
140 static bool transport_supports_multi_channels(MigrationAddress
*addr
)
142 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
143 SocketAddress
*saddr
= &addr
->u
.socket
;
145 return (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
146 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
147 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
);
148 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
149 return migrate_mapped_ram();
155 static bool migration_needs_seekable_channel(void)
157 return migrate_mapped_ram();
160 static bool transport_supports_seeking(MigrationAddress
*addr
)
162 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
170 migration_channels_and_transport_compatible(MigrationAddress
*addr
,
173 if (migration_needs_seekable_channel() &&
174 !transport_supports_seeking(addr
)) {
175 error_setg(errp
, "Migration requires seekable transport (e.g. file)");
179 if (migration_needs_multiple_sockets() &&
180 !transport_supports_multi_channels(addr
)) {
181 error_setg(errp
, "Migration requires multi-channel URIs (e.g. tcp)");
188 static gint
page_request_addr_cmp(gconstpointer ap
, gconstpointer bp
)
190 uintptr_t a
= (uintptr_t) ap
, b
= (uintptr_t) bp
;
192 return (a
> b
) - (a
< b
);
195 static int migration_stop_vm(MigrationState
*s
, RunState state
)
199 migration_downtime_start(s
);
201 s
->vm_old_state
= runstate_get();
202 global_state_store();
204 ret
= vm_stop_force_state(state
);
206 trace_vmstate_downtime_checkpoint("src-vm-stopped");
207 trace_migration_completion_vm_stop(ret
);
212 void migration_object_init(void)
214 /* This can only be called once. */
215 assert(!current_migration
);
216 current_migration
= MIGRATION_OBJ(object_new(TYPE_MIGRATION
));
219 * Init the migrate incoming object as well no matter whether
220 * we'll use it or not.
222 assert(!current_incoming
);
223 current_incoming
= g_new0(MigrationIncomingState
, 1);
224 current_incoming
->state
= MIGRATION_STATUS_NONE
;
225 current_incoming
->postcopy_remote_fds
=
226 g_array_new(FALSE
, TRUE
, sizeof(struct PostCopyFD
));
227 qemu_mutex_init(¤t_incoming
->rp_mutex
);
228 qemu_mutex_init(¤t_incoming
->postcopy_prio_thread_mutex
);
229 qemu_event_init(¤t_incoming
->main_thread_load_event
, false);
230 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_dst
, 0);
231 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fault
, 0);
232 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fast_load
, 0);
233 qemu_sem_init(¤t_incoming
->postcopy_qemufile_dst_done
, 0);
235 qemu_mutex_init(¤t_incoming
->page_request_mutex
);
236 qemu_cond_init(¤t_incoming
->page_request_cond
);
237 current_incoming
->page_requested
= g_tree_new(page_request_addr_cmp
);
239 current_incoming
->exit_on_error
= INMIGRATE_DEFAULT_EXIT_ON_ERROR
;
241 migration_object_check(current_migration
, &error_fatal
);
245 dirty_bitmap_mig_init();
254 static void migration_bh_dispatch_bh(void *opaque
)
256 MigrationState
*s
= migrate_get_current();
257 MigrationBH
*migbh
= opaque
;
259 /* cleanup this BH */
260 qemu_bh_delete(migbh
->bh
);
263 /* dispatch the other one */
264 migbh
->cb(migbh
->opaque
);
265 object_unref(OBJECT(s
));
270 void migration_bh_schedule(QEMUBHFunc
*cb
, void *opaque
)
272 MigrationState
*s
= migrate_get_current();
273 MigrationBH
*migbh
= g_new0(MigrationBH
, 1);
274 QEMUBH
*bh
= qemu_bh_new(migration_bh_dispatch_bh
, migbh
);
276 /* Store these to dispatch when the BH runs */
279 migbh
->opaque
= opaque
;
282 * Ref the state for bh, because it may be called when
283 * there're already no other refs
285 object_ref(OBJECT(s
));
286 qemu_bh_schedule(bh
);
289 void migration_cancel(const Error
*error
)
292 migrate_set_error(current_migration
, error
);
294 if (migrate_dirty_limit()) {
295 qmp_cancel_vcpu_dirty_limit(false, -1, NULL
);
297 migrate_fd_cancel(current_migration
);
300 void migration_shutdown(void)
303 * When the QEMU main thread exit, the COLO thread
304 * may wait a semaphore. So, we should wakeup the
305 * COLO thread before migration shutdown.
309 * Cancel the current migration - that will (eventually)
310 * stop the migration using this structure
312 migration_cancel(NULL
);
313 object_unref(OBJECT(current_migration
));
316 * Cancel outgoing migration of dirty bitmaps. It should
317 * at least unref used block nodes.
319 dirty_bitmap_mig_cancel_outgoing();
322 * Cancel incoming migration of dirty bitmaps. Dirty bitmaps
323 * are non-critical data, and their loss never considered as
326 dirty_bitmap_mig_cancel_incoming();
330 MigrationState
*migrate_get_current(void)
332 /* This can only be called after the object created. */
333 assert(current_migration
);
334 return current_migration
;
337 MigrationIncomingState
*migration_incoming_get_current(void)
339 assert(current_incoming
);
340 return current_incoming
;
343 void migration_incoming_transport_cleanup(MigrationIncomingState
*mis
)
345 if (mis
->socket_address_list
) {
346 qapi_free_SocketAddressList(mis
->socket_address_list
);
347 mis
->socket_address_list
= NULL
;
350 if (mis
->transport_cleanup
) {
351 mis
->transport_cleanup(mis
->transport_data
);
352 mis
->transport_data
= mis
->transport_cleanup
= NULL
;
356 void migration_incoming_state_destroy(void)
358 struct MigrationIncomingState
*mis
= migration_incoming_get_current();
360 multifd_recv_cleanup();
361 compress_threads_load_cleanup();
363 if (mis
->to_src_file
) {
364 /* Tell source that we are done */
365 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
366 qemu_fclose(mis
->to_src_file
);
367 mis
->to_src_file
= NULL
;
370 if (mis
->from_src_file
) {
371 migration_ioc_unregister_yank_from_file(mis
->from_src_file
);
372 qemu_fclose(mis
->from_src_file
);
373 mis
->from_src_file
= NULL
;
375 if (mis
->postcopy_remote_fds
) {
376 g_array_free(mis
->postcopy_remote_fds
, TRUE
);
377 mis
->postcopy_remote_fds
= NULL
;
380 migration_incoming_transport_cleanup(mis
);
381 qemu_event_reset(&mis
->main_thread_load_event
);
383 if (mis
->page_requested
) {
384 g_tree_destroy(mis
->page_requested
);
385 mis
->page_requested
= NULL
;
388 if (mis
->postcopy_qemufile_dst
) {
389 migration_ioc_unregister_yank_from_file(mis
->postcopy_qemufile_dst
);
390 qemu_fclose(mis
->postcopy_qemufile_dst
);
391 mis
->postcopy_qemufile_dst
= NULL
;
394 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
397 static void migrate_generate_event(int new_state
)
399 if (migrate_events()) {
400 qapi_event_send_migration(new_state
);
405 * Send a message on the return channel back to the source
408 static int migrate_send_rp_message(MigrationIncomingState
*mis
,
409 enum mig_rp_message_type message_type
,
410 uint16_t len
, void *data
)
414 trace_migrate_send_rp_message((int)message_type
, len
);
415 QEMU_LOCK_GUARD(&mis
->rp_mutex
);
418 * It's possible that the file handle got lost due to network
421 if (!mis
->to_src_file
) {
426 qemu_put_be16(mis
->to_src_file
, (unsigned int)message_type
);
427 qemu_put_be16(mis
->to_src_file
, len
);
428 qemu_put_buffer(mis
->to_src_file
, data
, len
);
429 return qemu_fflush(mis
->to_src_file
);
432 /* Request one page from the source VM at the given start address.
433 * rb: the RAMBlock to request the page in
434 * Start: Address offset within the RB
435 * Len: Length in bytes required - must be a multiple of pagesize
437 int migrate_send_rp_message_req_pages(MigrationIncomingState
*mis
,
438 RAMBlock
*rb
, ram_addr_t start
)
440 uint8_t bufc
[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
441 size_t msglen
= 12; /* start + len */
442 size_t len
= qemu_ram_pagesize(rb
);
443 enum mig_rp_message_type msg_type
;
447 *(uint64_t *)bufc
= cpu_to_be64((uint64_t)start
);
448 *(uint32_t *)(bufc
+ 8) = cpu_to_be32((uint32_t)len
);
451 * We maintain the last ramblock that we requested for page. Note that we
452 * don't need locking because this function will only be called within the
453 * postcopy ram fault thread.
455 if (rb
!= mis
->last_rb
) {
458 rbname
= qemu_ram_get_idstr(rb
);
459 rbname_len
= strlen(rbname
);
461 assert(rbname_len
< 256);
463 bufc
[msglen
++] = rbname_len
;
464 memcpy(bufc
+ msglen
, rbname
, rbname_len
);
465 msglen
+= rbname_len
;
466 msg_type
= MIG_RP_MSG_REQ_PAGES_ID
;
468 msg_type
= MIG_RP_MSG_REQ_PAGES
;
471 return migrate_send_rp_message(mis
, msg_type
, msglen
, bufc
);
474 int migrate_send_rp_req_pages(MigrationIncomingState
*mis
,
475 RAMBlock
*rb
, ram_addr_t start
, uint64_t haddr
)
477 void *aligned
= (void *)(uintptr_t)ROUND_DOWN(haddr
, qemu_ram_pagesize(rb
));
478 bool received
= false;
480 WITH_QEMU_LOCK_GUARD(&mis
->page_request_mutex
) {
481 received
= ramblock_recv_bitmap_test_byte_offset(rb
, start
);
482 if (!received
&& !g_tree_lookup(mis
->page_requested
, aligned
)) {
484 * The page has not been received, and it's not yet in the page
485 * request list. Queue it. Set the value of element to 1, so that
486 * things like g_tree_lookup() will return TRUE (1) when found.
488 g_tree_insert(mis
->page_requested
, aligned
, (gpointer
)1);
489 qatomic_inc(&mis
->page_requested_count
);
490 trace_postcopy_page_req_add(aligned
, mis
->page_requested_count
);
495 * If the page is there, skip sending the message. We don't even need the
496 * lock because as long as the page arrived, it'll be there forever.
502 return migrate_send_rp_message_req_pages(mis
, rb
, start
);
505 static bool migration_colo_enabled
;
506 bool migration_incoming_colo_enabled(void)
508 return migration_colo_enabled
;
511 void migration_incoming_disable_colo(void)
513 ram_block_discard_disable(false);
514 migration_colo_enabled
= false;
517 int migration_incoming_enable_colo(void)
519 #ifndef CONFIG_REPLICATION
520 error_report("ENABLE_COLO command come in migration stream, but COLO "
521 "module is not built in");
525 if (!migrate_colo()) {
526 error_report("ENABLE_COLO command come in migration stream, but c-colo "
527 "capability is not set");
531 if (ram_block_discard_disable(true)) {
532 error_report("COLO: cannot disable RAM discard");
535 migration_colo_enabled
= true;
539 void migrate_add_address(SocketAddress
*address
)
541 MigrationIncomingState
*mis
= migration_incoming_get_current();
543 QAPI_LIST_PREPEND(mis
->socket_address_list
,
544 QAPI_CLONE(SocketAddress
, address
));
547 bool migrate_uri_parse(const char *uri
, MigrationChannel
**channel
,
550 g_autoptr(MigrationChannel
) val
= g_new0(MigrationChannel
, 1);
551 g_autoptr(MigrationAddress
) addr
= g_new0(MigrationAddress
, 1);
552 InetSocketAddress
*isock
= &addr
->u
.rdma
;
553 strList
**tail
= &addr
->u
.exec
.args
;
555 if (strstart(uri
, "exec:", NULL
)) {
556 addr
->transport
= MIGRATION_ADDRESS_TYPE_EXEC
;
558 QAPI_LIST_APPEND(tail
, g_strdup(exec_get_cmd_path()));
559 QAPI_LIST_APPEND(tail
, g_strdup("/c"));
561 QAPI_LIST_APPEND(tail
, g_strdup("/bin/sh"));
562 QAPI_LIST_APPEND(tail
, g_strdup("-c"));
564 QAPI_LIST_APPEND(tail
, g_strdup(uri
+ strlen("exec:")));
565 } else if (strstart(uri
, "rdma:", NULL
)) {
566 if (inet_parse(isock
, uri
+ strlen("rdma:"), errp
)) {
567 qapi_free_InetSocketAddress(isock
);
570 addr
->transport
= MIGRATION_ADDRESS_TYPE_RDMA
;
571 } else if (strstart(uri
, "tcp:", NULL
) ||
572 strstart(uri
, "unix:", NULL
) ||
573 strstart(uri
, "vsock:", NULL
) ||
574 strstart(uri
, "fd:", NULL
)) {
575 addr
->transport
= MIGRATION_ADDRESS_TYPE_SOCKET
;
576 SocketAddress
*saddr
= socket_parse(uri
, errp
);
580 addr
->u
.socket
.type
= saddr
->type
;
581 addr
->u
.socket
.u
= saddr
->u
;
582 /* Don't free the objects inside; their ownership moved to "addr" */
584 } else if (strstart(uri
, "file:", NULL
)) {
585 addr
->transport
= MIGRATION_ADDRESS_TYPE_FILE
;
586 addr
->u
.file
.filename
= g_strdup(uri
+ strlen("file:"));
587 if (file_parse_offset(addr
->u
.file
.filename
, &addr
->u
.file
.offset
,
592 error_setg(errp
, "unknown migration protocol: %s", uri
);
596 val
->channel_type
= MIGRATION_CHANNEL_TYPE_MAIN
;
597 val
->addr
= g_steal_pointer(&addr
);
598 *channel
= g_steal_pointer(&val
);
602 static void qemu_start_incoming_migration(const char *uri
, bool has_channels
,
603 MigrationChannelList
*channels
,
606 g_autoptr(MigrationChannel
) channel
= NULL
;
607 MigrationAddress
*addr
= NULL
;
608 MigrationIncomingState
*mis
= migration_incoming_get_current();
611 * Having preliminary checks for uri and channel
613 if (!uri
== !channels
) {
614 error_setg(errp
, "need either 'uri' or 'channels' argument");
619 /* To verify that Migrate channel list has only item */
620 if (channels
->next
) {
621 error_setg(errp
, "Channel list has more than one entries");
624 addr
= channels
->value
->addr
;
628 /* caller uses the old URI syntax */
629 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
632 addr
= channel
->addr
;
635 /* transport mechanism not suitable for migration? */
636 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
640 migrate_set_state(&mis
->state
, MIGRATION_STATUS_NONE
,
641 MIGRATION_STATUS_SETUP
);
643 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
644 SocketAddress
*saddr
= &addr
->u
.socket
;
645 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
646 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
647 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
648 socket_start_incoming_migration(saddr
, errp
);
649 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
650 fd_start_incoming_migration(saddr
->u
.fd
.str
, errp
);
653 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
654 if (migrate_compress()) {
655 error_setg(errp
, "RDMA and compression can't be used together");
658 if (migrate_xbzrle()) {
659 error_setg(errp
, "RDMA and XBZRLE can't be used together");
662 if (migrate_multifd()) {
663 error_setg(errp
, "RDMA and multifd can't be used together");
666 rdma_start_incoming_migration(&addr
->u
.rdma
, errp
);
668 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
669 exec_start_incoming_migration(addr
->u
.exec
.args
, errp
);
670 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
671 file_start_incoming_migration(&addr
->u
.file
, errp
);
673 error_setg(errp
, "unknown migration protocol: %s", uri
);
677 static void process_incoming_migration_bh(void *opaque
)
679 Error
*local_err
= NULL
;
680 MigrationIncomingState
*mis
= opaque
;
682 trace_vmstate_downtime_checkpoint("dst-precopy-bh-enter");
684 /* If capability late_block_activate is set:
685 * Only fire up the block code now if we're going to restart the
686 * VM, else 'cont' will do it.
687 * This causes file locking to happen; so we don't want it to happen
688 * unless we really are starting the VM.
690 if (!migrate_late_block_activate() ||
691 (autostart
&& (!global_state_received() ||
692 runstate_is_live(global_state_get_runstate())))) {
693 /* Make sure all file formats throw away their mutable metadata.
694 * If we get an error here, just don't restart the VM yet. */
695 bdrv_activate_all(&local_err
);
697 error_report_err(local_err
);
704 * This must happen after all error conditions are dealt with and
705 * we're sure the VM is going to be running on this host.
707 qemu_announce_self(&mis
->announce_timer
, migrate_announce_params());
709 trace_vmstate_downtime_checkpoint("dst-precopy-bh-announced");
711 multifd_recv_shutdown();
713 dirty_bitmap_mig_before_vm_start();
715 if (!global_state_received() ||
716 runstate_is_live(global_state_get_runstate())) {
720 runstate_set(RUN_STATE_PAUSED
);
722 } else if (migration_incoming_colo_enabled()) {
723 migration_incoming_disable_colo();
726 runstate_set(global_state_get_runstate());
728 trace_vmstate_downtime_checkpoint("dst-precopy-bh-vm-started");
730 * This must happen after any state changes since as soon as an external
731 * observer sees this event they might start to prod at the VM assuming
734 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
735 MIGRATION_STATUS_COMPLETED
);
736 migration_incoming_state_destroy();
739 static void coroutine_fn
740 process_incoming_migration_co(void *opaque
)
742 MigrationState
*s
= migrate_get_current();
743 MigrationIncomingState
*mis
= migration_incoming_get_current();
746 Error
*local_err
= NULL
;
748 assert(mis
->from_src_file
);
750 if (compress_threads_load_setup(mis
->from_src_file
)) {
751 error_setg(&local_err
, "Failed to setup decompress threads");
755 mis
->largest_page_size
= qemu_ram_pagesize_largest();
756 postcopy_state_set(POSTCOPY_INCOMING_NONE
);
757 migrate_set_state(&mis
->state
, MIGRATION_STATUS_SETUP
,
758 MIGRATION_STATUS_ACTIVE
);
760 mis
->loadvm_co
= qemu_coroutine_self();
761 ret
= qemu_loadvm_state(mis
->from_src_file
);
762 mis
->loadvm_co
= NULL
;
764 trace_vmstate_downtime_checkpoint("dst-precopy-loadvm-completed");
766 ps
= postcopy_state_get();
767 trace_process_incoming_migration_co_end(ret
, ps
);
768 if (ps
!= POSTCOPY_INCOMING_NONE
) {
769 if (ps
== POSTCOPY_INCOMING_ADVISE
) {
771 * Where a migration had postcopy enabled (and thus went to advise)
772 * but managed to complete within the precopy period, we can use
775 postcopy_ram_incoming_cleanup(mis
);
776 } else if (ret
>= 0) {
778 * Postcopy was started, cleanup should happen at the end of the
781 trace_process_incoming_migration_co_postcopy_end_main();
784 /* Else if something went wrong then just fall out of the normal exit */
788 error_setg(&local_err
, "load of migration failed: %s", strerror(-ret
));
792 if (colo_incoming_co() < 0) {
793 error_setg(&local_err
, "colo incoming failed");
797 migration_bh_schedule(process_incoming_migration_bh
, mis
);
800 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
801 MIGRATION_STATUS_FAILED
);
802 migrate_set_error(s
, local_err
);
803 error_free(local_err
);
805 migration_incoming_state_destroy();
807 if (mis
->exit_on_error
) {
808 WITH_QEMU_LOCK_GUARD(&s
->error_mutex
) {
809 error_report_err(s
->error
);
818 * migration_incoming_setup: Setup incoming migration
819 * @f: file for main migration channel
821 static void migration_incoming_setup(QEMUFile
*f
)
823 MigrationIncomingState
*mis
= migration_incoming_get_current();
825 if (!mis
->from_src_file
) {
826 mis
->from_src_file
= f
;
828 qemu_file_set_blocking(f
, false);
831 void migration_incoming_process(void)
833 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
, NULL
);
834 qemu_coroutine_enter(co
);
837 /* Returns true if recovered from a paused migration, otherwise false */
838 static bool postcopy_try_recover(void)
840 MigrationIncomingState
*mis
= migration_incoming_get_current();
842 if (mis
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
843 /* Resumed from a paused postcopy migration */
845 /* This should be set already in migration_incoming_setup() */
846 assert(mis
->from_src_file
);
847 /* Postcopy has standalone thread to do vm load */
848 qemu_file_set_blocking(mis
->from_src_file
, true);
850 /* Re-configure the return path */
851 mis
->to_src_file
= qemu_file_get_return_path(mis
->from_src_file
);
853 migrate_set_state(&mis
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
854 MIGRATION_STATUS_POSTCOPY_RECOVER
);
857 * Here, we only wake up the main loading thread (while the
858 * rest threads will still be waiting), so that we can receive
859 * commands from source now, and answer it if needed. The
860 * rest threads will be woken up afterwards until we are sure
861 * that source is ready to reply to page requests.
863 qemu_sem_post(&mis
->postcopy_pause_sem_dst
);
870 void migration_fd_process_incoming(QEMUFile
*f
)
872 migration_incoming_setup(f
);
873 if (postcopy_try_recover()) {
876 migration_incoming_process();
880 * Returns true when we want to start a new incoming migration process,
883 static bool migration_should_start_incoming(bool main_channel
)
885 /* Multifd doesn't start unless all channels are established */
886 if (migrate_multifd()) {
887 return migration_has_all_channels();
890 /* Preempt channel only starts when the main channel is created */
891 if (migrate_postcopy_preempt()) {
896 * For all the rest types of migration, we should only reach here when
897 * it's the main channel that's being created, and we should always
898 * proceed with this channel.
900 assert(main_channel
);
904 void migration_ioc_process_incoming(QIOChannel
*ioc
, Error
**errp
)
906 MigrationIncomingState
*mis
= migration_incoming_get_current();
907 Error
*local_err
= NULL
;
909 bool default_channel
= true;
910 uint32_t channel_magic
= 0;
913 if (migrate_multifd() && !migrate_mapped_ram() &&
914 !migrate_postcopy_ram() &&
915 qio_channel_has_feature(ioc
, QIO_CHANNEL_FEATURE_READ_MSG_PEEK
)) {
917 * With multiple channels, it is possible that we receive channels
918 * out of order on destination side, causing incorrect mapping of
919 * source channels on destination side. Check channel MAGIC to
920 * decide type of channel. Please note this is best effort, postcopy
921 * preempt channel does not send any magic number so avoid it for
922 * postcopy live migration. Also tls live migration already does
923 * tls handshake while initializing main channel so with tls this
924 * issue is not possible.
926 ret
= migration_channel_read_peek(ioc
, (void *)&channel_magic
,
927 sizeof(channel_magic
), errp
);
933 default_channel
= (channel_magic
== cpu_to_be32(QEMU_VM_FILE_MAGIC
));
935 default_channel
= !mis
->from_src_file
;
938 if (multifd_recv_setup(errp
) != 0) {
942 if (default_channel
) {
943 f
= qemu_file_new_input(ioc
);
944 migration_incoming_setup(f
);
946 /* Multiple connections */
947 assert(migration_needs_multiple_sockets());
948 if (migrate_multifd()) {
949 multifd_recv_new_channel(ioc
, &local_err
);
951 assert(migrate_postcopy_preempt());
952 f
= qemu_file_new_input(ioc
);
953 postcopy_preempt_new_channel(mis
, f
);
956 error_propagate(errp
, local_err
);
961 if (migration_should_start_incoming(default_channel
)) {
962 /* If it's a recovery, we're done */
963 if (postcopy_try_recover()) {
966 migration_incoming_process();
971 * @migration_has_all_channels: We have received all channels that we need
973 * Returns true when we have got connections to all the channels that
974 * we need for migration.
976 bool migration_has_all_channels(void)
978 MigrationIncomingState
*mis
= migration_incoming_get_current();
980 if (!mis
->from_src_file
) {
984 if (migrate_multifd()) {
985 return multifd_recv_all_channels_created();
988 if (migrate_postcopy_preempt()) {
989 return mis
->postcopy_qemufile_dst
!= NULL
;
995 int migrate_send_rp_switchover_ack(MigrationIncomingState
*mis
)
997 return migrate_send_rp_message(mis
, MIG_RP_MSG_SWITCHOVER_ACK
, 0, NULL
);
1001 * Send a 'SHUT' message on the return channel with the given value
1002 * to indicate that we've finished with the RP. Non-0 value indicates
1005 void migrate_send_rp_shut(MigrationIncomingState
*mis
,
1010 buf
= cpu_to_be32(value
);
1011 migrate_send_rp_message(mis
, MIG_RP_MSG_SHUT
, sizeof(buf
), &buf
);
1015 * Send a 'PONG' message on the return channel with the given value
1016 * (normally in response to a 'PING')
1018 void migrate_send_rp_pong(MigrationIncomingState
*mis
,
1023 buf
= cpu_to_be32(value
);
1024 migrate_send_rp_message(mis
, MIG_RP_MSG_PONG
, sizeof(buf
), &buf
);
1027 void migrate_send_rp_recv_bitmap(MigrationIncomingState
*mis
,
1035 * First, we send the header part. It contains only the len of
1036 * idstr, and the idstr itself.
1038 len
= strlen(block_name
);
1040 memcpy(buf
+ 1, block_name
, len
);
1042 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
1043 error_report("%s: MSG_RP_RECV_BITMAP only used for recovery",
1048 migrate_send_rp_message(mis
, MIG_RP_MSG_RECV_BITMAP
, len
+ 1, buf
);
1051 * Next, we dump the received bitmap to the stream.
1053 * TODO: currently we are safe since we are the only one that is
1054 * using the to_src_file handle (fault thread is still paused),
1055 * and it's ok even not taking the mutex. However the best way is
1056 * to take the lock before sending the message header, and release
1057 * the lock after sending the bitmap.
1059 qemu_mutex_lock(&mis
->rp_mutex
);
1060 res
= ramblock_recv_bitmap_send(mis
->to_src_file
, block_name
);
1061 qemu_mutex_unlock(&mis
->rp_mutex
);
1063 trace_migrate_send_rp_recv_bitmap(block_name
, res
);
1066 void migrate_send_rp_resume_ack(MigrationIncomingState
*mis
, uint32_t value
)
1070 buf
= cpu_to_be32(value
);
1071 migrate_send_rp_message(mis
, MIG_RP_MSG_RESUME_ACK
, sizeof(buf
), &buf
);
1075 * Return true if we're already in the middle of a migration
1076 * (i.e. any of the active or setup states)
1078 bool migration_is_setup_or_active(void)
1080 MigrationState
*s
= current_migration
;
1083 case MIGRATION_STATUS_ACTIVE
:
1084 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1085 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1086 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1087 case MIGRATION_STATUS_SETUP
:
1088 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1089 case MIGRATION_STATUS_DEVICE
:
1090 case MIGRATION_STATUS_WAIT_UNPLUG
:
1091 case MIGRATION_STATUS_COLO
:
1100 bool migration_is_running(void)
1102 MigrationState
*s
= current_migration
;
1105 case MIGRATION_STATUS_ACTIVE
:
1106 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1107 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1108 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1109 case MIGRATION_STATUS_SETUP
:
1110 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1111 case MIGRATION_STATUS_DEVICE
:
1112 case MIGRATION_STATUS_WAIT_UNPLUG
:
1113 case MIGRATION_STATUS_CANCELLING
:
1122 static bool migrate_show_downtime(MigrationState
*s
)
1124 return (s
->state
== MIGRATION_STATUS_COMPLETED
) || migration_in_postcopy();
1127 static void populate_time_info(MigrationInfo
*info
, MigrationState
*s
)
1129 info
->has_status
= true;
1130 info
->has_setup_time
= true;
1131 info
->setup_time
= s
->setup_time
;
1133 if (s
->state
== MIGRATION_STATUS_COMPLETED
) {
1134 info
->has_total_time
= true;
1135 info
->total_time
= s
->total_time
;
1137 info
->has_total_time
= true;
1138 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) -
1142 if (migrate_show_downtime(s
)) {
1143 info
->has_downtime
= true;
1144 info
->downtime
= s
->downtime
;
1146 info
->has_expected_downtime
= true;
1147 info
->expected_downtime
= s
->expected_downtime
;
1151 static void populate_ram_info(MigrationInfo
*info
, MigrationState
*s
)
1153 size_t page_size
= qemu_target_page_size();
1155 info
->ram
= g_malloc0(sizeof(*info
->ram
));
1156 info
->ram
->transferred
= migration_transferred_bytes();
1157 info
->ram
->total
= ram_bytes_total();
1158 info
->ram
->duplicate
= stat64_get(&mig_stats
.zero_pages
);
1159 /* legacy value. It is not used anymore */
1160 info
->ram
->skipped
= 0;
1161 info
->ram
->normal
= stat64_get(&mig_stats
.normal_pages
);
1162 info
->ram
->normal_bytes
= info
->ram
->normal
* page_size
;
1163 info
->ram
->mbps
= s
->mbps
;
1164 info
->ram
->dirty_sync_count
=
1165 stat64_get(&mig_stats
.dirty_sync_count
);
1166 info
->ram
->dirty_sync_missed_zero_copy
=
1167 stat64_get(&mig_stats
.dirty_sync_missed_zero_copy
);
1168 info
->ram
->postcopy_requests
=
1169 stat64_get(&mig_stats
.postcopy_requests
);
1170 info
->ram
->page_size
= page_size
;
1171 info
->ram
->multifd_bytes
= stat64_get(&mig_stats
.multifd_bytes
);
1172 info
->ram
->pages_per_second
= s
->pages_per_second
;
1173 info
->ram
->precopy_bytes
= stat64_get(&mig_stats
.precopy_bytes
);
1174 info
->ram
->downtime_bytes
= stat64_get(&mig_stats
.downtime_bytes
);
1175 info
->ram
->postcopy_bytes
= stat64_get(&mig_stats
.postcopy_bytes
);
1177 if (migrate_xbzrle()) {
1178 info
->xbzrle_cache
= g_malloc0(sizeof(*info
->xbzrle_cache
));
1179 info
->xbzrle_cache
->cache_size
= migrate_xbzrle_cache_size();
1180 info
->xbzrle_cache
->bytes
= xbzrle_counters
.bytes
;
1181 info
->xbzrle_cache
->pages
= xbzrle_counters
.pages
;
1182 info
->xbzrle_cache
->cache_miss
= xbzrle_counters
.cache_miss
;
1183 info
->xbzrle_cache
->cache_miss_rate
= xbzrle_counters
.cache_miss_rate
;
1184 info
->xbzrle_cache
->encoding_rate
= xbzrle_counters
.encoding_rate
;
1185 info
->xbzrle_cache
->overflow
= xbzrle_counters
.overflow
;
1188 populate_compress(info
);
1190 if (cpu_throttle_active()) {
1191 info
->has_cpu_throttle_percentage
= true;
1192 info
->cpu_throttle_percentage
= cpu_throttle_get_percentage();
1195 if (s
->state
!= MIGRATION_STATUS_COMPLETED
) {
1196 info
->ram
->remaining
= ram_bytes_remaining();
1197 info
->ram
->dirty_pages_rate
=
1198 stat64_get(&mig_stats
.dirty_pages_rate
);
1201 if (migrate_dirty_limit() && dirtylimit_in_service()) {
1202 info
->has_dirty_limit_throttle_time_per_round
= true;
1203 info
->dirty_limit_throttle_time_per_round
=
1204 dirtylimit_throttle_time_per_round();
1206 info
->has_dirty_limit_ring_full_time
= true;
1207 info
->dirty_limit_ring_full_time
= dirtylimit_ring_full_time();
1211 static void populate_disk_info(MigrationInfo
*info
)
1213 if (blk_mig_active()) {
1214 info
->disk
= g_malloc0(sizeof(*info
->disk
));
1215 info
->disk
->transferred
= blk_mig_bytes_transferred();
1216 info
->disk
->remaining
= blk_mig_bytes_remaining();
1217 info
->disk
->total
= blk_mig_bytes_total();
1221 static void fill_source_migration_info(MigrationInfo
*info
)
1223 MigrationState
*s
= migrate_get_current();
1224 int state
= qatomic_read(&s
->state
);
1225 GSList
*cur_blocker
= migration_blockers
[migrate_mode()];
1227 info
->blocked_reasons
= NULL
;
1230 * There are two types of reasons a migration might be blocked;
1231 * a) devices marked in VMState as non-migratable, and
1232 * b) Explicit migration blockers
1233 * We need to add both of them here.
1235 qemu_savevm_non_migratable_list(&info
->blocked_reasons
);
1237 while (cur_blocker
) {
1238 QAPI_LIST_PREPEND(info
->blocked_reasons
,
1239 g_strdup(error_get_pretty(cur_blocker
->data
)));
1240 cur_blocker
= g_slist_next(cur_blocker
);
1242 info
->has_blocked_reasons
= info
->blocked_reasons
!= NULL
;
1245 case MIGRATION_STATUS_NONE
:
1246 /* no migration has happened ever */
1247 /* do not overwrite destination migration status */
1249 case MIGRATION_STATUS_SETUP
:
1250 info
->has_status
= true;
1251 info
->has_total_time
= false;
1253 case MIGRATION_STATUS_ACTIVE
:
1254 case MIGRATION_STATUS_CANCELLING
:
1255 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1256 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1257 case MIGRATION_STATUS_DEVICE
:
1258 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1259 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1260 /* TODO add some postcopy stats */
1261 populate_time_info(info
, s
);
1262 populate_ram_info(info
, s
);
1263 populate_disk_info(info
);
1264 migration_populate_vfio_info(info
);
1266 case MIGRATION_STATUS_COLO
:
1267 info
->has_status
= true;
1268 /* TODO: display COLO specific information (checkpoint info etc.) */
1270 case MIGRATION_STATUS_COMPLETED
:
1271 populate_time_info(info
, s
);
1272 populate_ram_info(info
, s
);
1273 migration_populate_vfio_info(info
);
1275 case MIGRATION_STATUS_FAILED
:
1276 info
->has_status
= true;
1278 case MIGRATION_STATUS_CANCELLED
:
1279 info
->has_status
= true;
1281 case MIGRATION_STATUS_WAIT_UNPLUG
:
1282 info
->has_status
= true;
1285 info
->status
= state
;
1287 QEMU_LOCK_GUARD(&s
->error_mutex
);
1289 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1293 static void fill_destination_migration_info(MigrationInfo
*info
)
1295 MigrationIncomingState
*mis
= migration_incoming_get_current();
1297 if (mis
->socket_address_list
) {
1298 info
->has_socket_address
= true;
1299 info
->socket_address
=
1300 QAPI_CLONE(SocketAddressList
, mis
->socket_address_list
);
1303 switch (mis
->state
) {
1304 case MIGRATION_STATUS_NONE
:
1306 case MIGRATION_STATUS_SETUP
:
1307 case MIGRATION_STATUS_CANCELLING
:
1308 case MIGRATION_STATUS_CANCELLED
:
1309 case MIGRATION_STATUS_ACTIVE
:
1310 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1311 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1312 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1313 case MIGRATION_STATUS_FAILED
:
1314 case MIGRATION_STATUS_COLO
:
1315 info
->has_status
= true;
1317 case MIGRATION_STATUS_COMPLETED
:
1318 info
->has_status
= true;
1319 fill_destination_postcopy_migration_info(info
);
1322 info
->status
= mis
->state
;
1324 if (!info
->error_desc
) {
1325 MigrationState
*s
= migrate_get_current();
1326 QEMU_LOCK_GUARD(&s
->error_mutex
);
1329 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1334 MigrationInfo
*qmp_query_migrate(Error
**errp
)
1336 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
1338 fill_destination_migration_info(info
);
1339 fill_source_migration_info(info
);
1344 void qmp_migrate_start_postcopy(Error
**errp
)
1346 MigrationState
*s
= migrate_get_current();
1348 if (!migrate_postcopy()) {
1349 error_setg(errp
, "Enable postcopy with migrate_set_capability before"
1350 " the start of migration");
1354 if (s
->state
== MIGRATION_STATUS_NONE
) {
1355 error_setg(errp
, "Postcopy must be started after migration has been"
1360 * we don't error if migration has finished since that would be racy
1361 * with issuing this command.
1363 qatomic_set(&s
->start_postcopy
, true);
1366 /* shared migration helpers */
1368 void migrate_set_state(int *state
, int old_state
, int new_state
)
1370 assert(new_state
< MIGRATION_STATUS__MAX
);
1371 if (qatomic_cmpxchg(state
, old_state
, new_state
) == old_state
) {
1372 trace_migrate_set_state(MigrationStatus_str(new_state
));
1373 migrate_generate_event(new_state
);
1377 static void migrate_fd_cleanup(MigrationState
*s
)
1379 MigrationEventType type
;
1381 g_free(s
->hostname
);
1383 json_writer_free(s
->vmdesc
);
1386 qemu_savevm_state_cleanup();
1388 close_return_path_on_source(s
);
1390 if (s
->to_dst_file
) {
1393 trace_migrate_fd_cleanup();
1395 if (s
->migration_thread_running
) {
1396 qemu_thread_join(&s
->thread
);
1397 s
->migration_thread_running
= false;
1401 multifd_send_shutdown();
1402 qemu_mutex_lock(&s
->qemu_file_lock
);
1403 tmp
= s
->to_dst_file
;
1404 s
->to_dst_file
= NULL
;
1405 qemu_mutex_unlock(&s
->qemu_file_lock
);
1407 * Close the file handle without the lock to make sure the
1408 * critical section won't block for long.
1410 migration_ioc_unregister_yank_from_file(tmp
);
1414 assert(!migration_is_active());
1416 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1417 migrate_set_state(&s
->state
, MIGRATION_STATUS_CANCELLING
,
1418 MIGRATION_STATUS_CANCELLED
);
1422 /* It is used on info migrate. We can't free it */
1423 error_report_err(error_copy(s
->error
));
1425 type
= migration_has_failed(s
) ? MIG_EVENT_PRECOPY_FAILED
:
1426 MIG_EVENT_PRECOPY_DONE
;
1427 migration_call_notifiers(s
, type
, NULL
);
1428 block_cleanup_parameters();
1429 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1432 static void migrate_fd_cleanup_bh(void *opaque
)
1434 migrate_fd_cleanup(opaque
);
1437 void migrate_set_error(MigrationState
*s
, const Error
*error
)
1439 QEMU_LOCK_GUARD(&s
->error_mutex
);
1441 trace_migrate_error(error_get_pretty(error
));
1444 s
->error
= error_copy(error
);
1448 bool migrate_has_error(MigrationState
*s
)
1450 /* The lock is not helpful here, but still follow the rule */
1451 QEMU_LOCK_GUARD(&s
->error_mutex
);
1452 return qatomic_read(&s
->error
);
1455 static void migrate_error_free(MigrationState
*s
)
1457 QEMU_LOCK_GUARD(&s
->error_mutex
);
1459 error_free(s
->error
);
1464 static void migrate_fd_error(MigrationState
*s
, const Error
*error
)
1466 assert(s
->to_dst_file
== NULL
);
1467 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1468 MIGRATION_STATUS_FAILED
);
1469 migrate_set_error(s
, error
);
1472 static void migrate_fd_cancel(MigrationState
*s
)
1476 trace_migrate_fd_cancel();
1478 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1479 if (s
->rp_state
.from_dst_file
) {
1480 /* shutdown the rp socket, so causing the rp thread to shutdown */
1481 qemu_file_shutdown(s
->rp_state
.from_dst_file
);
1486 old_state
= s
->state
;
1487 if (!migration_is_running()) {
1490 /* If the migration is paused, kick it out of the pause */
1491 if (old_state
== MIGRATION_STATUS_PRE_SWITCHOVER
) {
1492 qemu_sem_post(&s
->pause_sem
);
1494 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_CANCELLING
);
1495 } while (s
->state
!= MIGRATION_STATUS_CANCELLING
);
1498 * If we're unlucky the migration code might be stuck somewhere in a
1499 * send/write while the network has failed and is waiting to timeout;
1500 * if we've got shutdown(2) available then we can force it to quit.
1502 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1503 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1504 if (s
->to_dst_file
) {
1505 qemu_file_shutdown(s
->to_dst_file
);
1509 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& s
->block_inactive
) {
1510 Error
*local_err
= NULL
;
1512 bdrv_activate_all(&local_err
);
1514 error_report_err(local_err
);
1516 s
->block_inactive
= false;
1521 void migration_add_notifier_mode(NotifierWithReturn
*notify
,
1522 MigrationNotifyFunc func
, MigMode mode
)
1524 notify
->notify
= (NotifierWithReturnFunc
)func
;
1525 notifier_with_return_list_add(&migration_state_notifiers
[mode
], notify
);
1528 void migration_add_notifier(NotifierWithReturn
*notify
,
1529 MigrationNotifyFunc func
)
1531 migration_add_notifier_mode(notify
, func
, MIG_MODE_NORMAL
);
1534 void migration_remove_notifier(NotifierWithReturn
*notify
)
1536 if (notify
->notify
) {
1537 notifier_with_return_remove(notify
);
1538 notify
->notify
= NULL
;
1542 int migration_call_notifiers(MigrationState
*s
, MigrationEventType type
,
1545 MigMode mode
= s
->parameters
.mode
;
1550 ret
= notifier_with_return_list_notify(&migration_state_notifiers
[mode
],
1552 assert(!ret
|| type
== MIG_EVENT_PRECOPY_SETUP
);
1556 bool migration_has_failed(MigrationState
*s
)
1558 return (s
->state
== MIGRATION_STATUS_CANCELLED
||
1559 s
->state
== MIGRATION_STATUS_FAILED
);
1562 bool migration_in_postcopy(void)
1564 MigrationState
*s
= migrate_get_current();
1567 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1568 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1569 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1576 bool migration_postcopy_is_alive(int state
)
1579 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1580 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1587 bool migration_in_incoming_postcopy(void)
1589 PostcopyState ps
= postcopy_state_get();
1591 return ps
>= POSTCOPY_INCOMING_DISCARD
&& ps
< POSTCOPY_INCOMING_END
;
1594 bool migration_incoming_postcopy_advised(void)
1596 PostcopyState ps
= postcopy_state_get();
1598 return ps
>= POSTCOPY_INCOMING_ADVISE
&& ps
< POSTCOPY_INCOMING_END
;
1601 bool migration_in_bg_snapshot(void)
1603 return migrate_background_snapshot() &&
1604 migration_is_setup_or_active();
1607 bool migration_is_idle(void)
1609 MigrationState
*s
= current_migration
;
1616 case MIGRATION_STATUS_NONE
:
1617 case MIGRATION_STATUS_CANCELLED
:
1618 case MIGRATION_STATUS_COMPLETED
:
1619 case MIGRATION_STATUS_FAILED
:
1621 case MIGRATION_STATUS_SETUP
:
1622 case MIGRATION_STATUS_CANCELLING
:
1623 case MIGRATION_STATUS_ACTIVE
:
1624 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1625 case MIGRATION_STATUS_COLO
:
1626 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1627 case MIGRATION_STATUS_DEVICE
:
1628 case MIGRATION_STATUS_WAIT_UNPLUG
:
1630 case MIGRATION_STATUS__MAX
:
1631 g_assert_not_reached();
1637 bool migration_is_active(void)
1639 MigrationState
*s
= current_migration
;
1641 return (s
->state
== MIGRATION_STATUS_ACTIVE
||
1642 s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1645 bool migration_is_device(void)
1647 MigrationState
*s
= current_migration
;
1649 return s
->state
== MIGRATION_STATUS_DEVICE
;
1652 bool migration_thread_is_self(void)
1654 MigrationState
*s
= current_migration
;
1656 return qemu_thread_is_self(&s
->thread
);
1659 bool migrate_mode_is_cpr(MigrationState
*s
)
1661 return s
->parameters
.mode
== MIG_MODE_CPR_REBOOT
;
1664 int migrate_init(MigrationState
*s
, Error
**errp
)
1668 ret
= qemu_savevm_state_prepare(errp
);
1674 * Reinitialise all migration state, except
1675 * parameters/capabilities that the user set, and
1678 s
->to_dst_file
= NULL
;
1679 s
->state
= MIGRATION_STATUS_NONE
;
1680 s
->rp_state
.from_dst_file
= NULL
;
1682 s
->pages_per_second
= 0.0;
1684 s
->expected_downtime
= 0;
1686 s
->start_postcopy
= false;
1687 s
->migration_thread_running
= false;
1688 error_free(s
->error
);
1692 migrate_set_state(&s
->state
, MIGRATION_STATUS_NONE
, MIGRATION_STATUS_SETUP
);
1694 s
->start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1696 s
->vm_old_state
= -1;
1697 s
->iteration_initial_bytes
= 0;
1698 s
->threshold_size
= 0;
1699 s
->switchover_acked
= false;
1700 s
->rdma_migration
= false;
1702 * set mig_stats memory to zero for a new migration
1704 memset(&mig_stats
, 0, sizeof(mig_stats
));
1705 migration_reset_vfio_bytes_transferred();
1710 static bool is_busy(Error
**reasonp
, Error
**errp
)
1714 /* Snapshots are similar to migrations, so check RUN_STATE_SAVE_VM too. */
1715 if (runstate_check(RUN_STATE_SAVE_VM
) || !migration_is_idle()) {
1716 error_propagate_prepend(errp
, *reasonp
,
1717 "disallowing migration blocker "
1718 "(migration/snapshot in progress) for: ");
1725 static bool is_only_migratable(Error
**reasonp
, Error
**errp
, int modes
)
1729 if (only_migratable
&& (modes
& BIT(MIG_MODE_NORMAL
))) {
1730 error_propagate_prepend(errp
, *reasonp
,
1731 "disallowing migration blocker "
1732 "(--only-migratable) for: ");
1739 static int get_modes(MigMode mode
, va_list ap
)
1743 while (mode
!= -1 && mode
!= MIG_MODE_ALL
) {
1744 assert(mode
>= MIG_MODE_NORMAL
&& mode
< MIG_MODE__MAX
);
1746 mode
= va_arg(ap
, MigMode
);
1748 if (mode
== MIG_MODE_ALL
) {
1749 modes
= BIT(MIG_MODE__MAX
) - 1;
1754 static int add_blockers(Error
**reasonp
, Error
**errp
, int modes
)
1756 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1757 if (modes
& BIT(mode
)) {
1758 migration_blockers
[mode
] = g_slist_prepend(migration_blockers
[mode
],
1765 int migrate_add_blocker(Error
**reasonp
, Error
**errp
)
1767 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_ALL
);
1770 int migrate_add_blocker_normal(Error
**reasonp
, Error
**errp
)
1772 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_NORMAL
, -1);
1775 int migrate_add_blocker_modes(Error
**reasonp
, Error
**errp
, MigMode mode
, ...)
1781 modes
= get_modes(mode
, ap
);
1784 if (is_only_migratable(reasonp
, errp
, modes
)) {
1786 } else if (is_busy(reasonp
, errp
)) {
1789 return add_blockers(reasonp
, errp
, modes
);
1792 int migrate_add_blocker_internal(Error
**reasonp
, Error
**errp
)
1794 int modes
= BIT(MIG_MODE__MAX
) - 1;
1796 if (is_busy(reasonp
, errp
)) {
1799 return add_blockers(reasonp
, errp
, modes
);
1802 void migrate_del_blocker(Error
**reasonp
)
1805 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1806 migration_blockers
[mode
] = g_slist_remove(migration_blockers
[mode
],
1809 error_free(*reasonp
);
1814 void qmp_migrate_incoming(const char *uri
, bool has_channels
,
1815 MigrationChannelList
*channels
,
1816 bool has_exit_on_error
, bool exit_on_error
,
1819 Error
*local_err
= NULL
;
1820 static bool once
= true;
1821 MigrationIncomingState
*mis
= migration_incoming_get_current();
1824 error_setg(errp
, "The incoming migration has already been started");
1827 if (!runstate_check(RUN_STATE_INMIGRATE
)) {
1828 error_setg(errp
, "'-incoming' was not specified on the command line");
1832 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1836 mis
->exit_on_error
=
1837 has_exit_on_error
? exit_on_error
: INMIGRATE_DEFAULT_EXIT_ON_ERROR
;
1839 qemu_start_incoming_migration(uri
, has_channels
, channels
, &local_err
);
1842 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1843 error_propagate(errp
, local_err
);
1850 void qmp_migrate_recover(const char *uri
, Error
**errp
)
1852 MigrationIncomingState
*mis
= migration_incoming_get_current();
1855 * Don't even bother to use ERRP_GUARD() as it _must_ always be set by
1856 * callers (no one should ignore a recover failure); if there is, it's a
1857 * programming error.
1861 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1862 error_setg(errp
, "Migrate recover can only be run "
1863 "when postcopy is paused.");
1867 /* If there's an existing transport, release it */
1868 migration_incoming_transport_cleanup(mis
);
1871 * Note that this call will never start a real migration; it will
1872 * only re-setup the migration stream and poke existing migration
1873 * to continue using that newly established channel.
1875 qemu_start_incoming_migration(uri
, false, NULL
, errp
);
1878 void qmp_migrate_pause(Error
**errp
)
1880 MigrationState
*ms
= migrate_get_current();
1881 MigrationIncomingState
*mis
= migration_incoming_get_current();
1884 if (migration_postcopy_is_alive(ms
->state
)) {
1885 /* Source side, during postcopy */
1886 Error
*error
= NULL
;
1888 /* Tell the core migration that we're pausing */
1889 error_setg(&error
, "Postcopy migration is paused by the user");
1890 migrate_set_error(ms
, error
);
1893 qemu_mutex_lock(&ms
->qemu_file_lock
);
1894 if (ms
->to_dst_file
) {
1895 ret
= qemu_file_shutdown(ms
->to_dst_file
);
1897 qemu_mutex_unlock(&ms
->qemu_file_lock
);
1899 error_setg(errp
, "Failed to pause source migration");
1903 * Kick the migration thread out of any waiting windows (on behalf
1904 * of the rp thread).
1906 migration_rp_kick(ms
);
1911 if (migration_postcopy_is_alive(mis
->state
)) {
1912 ret
= qemu_file_shutdown(mis
->from_src_file
);
1914 error_setg(errp
, "Failed to pause destination migration");
1919 error_setg(errp
, "migrate-pause is currently only supported "
1920 "during postcopy-active or postcopy-recover state");
1923 bool migration_is_blocked(Error
**errp
)
1925 GSList
*blockers
= migration_blockers
[migrate_mode()];
1927 if (qemu_savevm_state_blocked(errp
)) {
1932 error_propagate(errp
, error_copy(blockers
->data
));
1939 /* Returns true if continue to migrate, or false if error detected */
1940 static bool migrate_prepare(MigrationState
*s
, bool blk
, bool blk_inc
,
1941 bool resume
, Error
**errp
)
1944 warn_report("parameter 'inc' is deprecated;"
1945 " use blockdev-mirror with NBD instead");
1949 warn_report("parameter 'blk' is deprecated;"
1950 " use blockdev-mirror with NBD instead");
1954 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1955 error_setg(errp
, "Cannot resume if there is no "
1956 "paused migration");
1961 * Postcopy recovery won't work well with release-ram
1962 * capability since release-ram will drop the page buffer as
1963 * long as the page is put into the send buffer. So if there
1964 * is a network failure happened, any page buffers that have
1965 * not yet reached the destination VM but have already been
1966 * sent from the source VM will be lost forever. Let's refuse
1967 * the client from resuming such a postcopy migration.
1968 * Luckily release-ram was designed to only be used when src
1969 * and destination VMs are on the same host, so it should be
1972 if (migrate_release_ram()) {
1973 error_setg(errp
, "Postcopy recovery cannot work "
1974 "when release-ram capability is set");
1978 /* This is a resume, skip init status */
1982 if (migration_is_running()) {
1983 error_setg(errp
, "There's a migration process in progress");
1987 if (runstate_check(RUN_STATE_INMIGRATE
)) {
1988 error_setg(errp
, "Guest is waiting for an incoming migration");
1992 if (runstate_check(RUN_STATE_POSTMIGRATE
)) {
1993 error_setg(errp
, "Can't migrate the vm that was paused due to "
1994 "previous migration");
1998 if (kvm_hwpoisoned_mem()) {
1999 error_setg(errp
, "Can't migrate this vm with hardware poisoned memory, "
2000 "please reboot the vm and try again");
2004 if (migration_is_blocked(errp
)) {
2008 if (migrate_mapped_ram()) {
2009 if (migrate_tls()) {
2010 error_setg(errp
, "Cannot use TLS with mapped-ram");
2014 if (migrate_multifd_compression()) {
2015 error_setg(errp
, "Cannot use compression with mapped-ram");
2020 if (migrate_mode_is_cpr(s
)) {
2021 const char *conflict
= NULL
;
2023 if (migrate_postcopy()) {
2024 conflict
= "postcopy";
2025 } else if (migrate_background_snapshot()) {
2026 conflict
= "background snapshot";
2027 } else if (migrate_colo()) {
2032 error_setg(errp
, "Cannot use %s with CPR", conflict
);
2037 if (blk
|| blk_inc
) {
2038 if (migrate_colo()) {
2039 error_setg(errp
, "No disk migration is required in COLO mode");
2042 if (migrate_block() || migrate_block_incremental()) {
2043 error_setg(errp
, "Command options are incompatible with "
2044 "current migration capabilities");
2047 if (!migrate_cap_set(MIGRATION_CAPABILITY_BLOCK
, true, errp
)) {
2050 s
->must_remove_block_options
= true;
2054 migrate_set_block_incremental(true);
2057 if (migrate_init(s
, errp
)) {
2064 void qmp_migrate(const char *uri
, bool has_channels
,
2065 MigrationChannelList
*channels
, bool has_blk
, bool blk
,
2066 bool has_inc
, bool inc
, bool has_detach
, bool detach
,
2067 bool has_resume
, bool resume
, Error
**errp
)
2069 bool resume_requested
;
2070 Error
*local_err
= NULL
;
2071 MigrationState
*s
= migrate_get_current();
2072 g_autoptr(MigrationChannel
) channel
= NULL
;
2073 MigrationAddress
*addr
= NULL
;
2076 * Having preliminary checks for uri and channel
2078 if (!uri
== !channels
) {
2079 error_setg(errp
, "need either 'uri' or 'channels' argument");
2084 /* To verify that Migrate channel list has only item */
2085 if (channels
->next
) {
2086 error_setg(errp
, "Channel list has more than one entries");
2089 addr
= channels
->value
->addr
;
2093 /* caller uses the old URI syntax */
2094 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
2097 addr
= channel
->addr
;
2100 /* transport mechanism not suitable for migration? */
2101 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
2105 resume_requested
= has_resume
&& resume
;
2106 if (!migrate_prepare(s
, has_blk
&& blk
, has_inc
&& inc
,
2107 resume_requested
, errp
)) {
2108 /* Error detected, put into errp */
2112 if (!resume_requested
) {
2113 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
2118 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
2119 SocketAddress
*saddr
= &addr
->u
.socket
;
2120 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
2121 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
2122 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
2123 socket_start_outgoing_migration(s
, saddr
, &local_err
);
2124 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
2125 fd_start_outgoing_migration(s
, saddr
->u
.fd
.str
, &local_err
);
2128 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
2129 rdma_start_outgoing_migration(s
, &addr
->u
.rdma
, &local_err
);
2131 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
2132 exec_start_outgoing_migration(s
, addr
->u
.exec
.args
, &local_err
);
2133 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
2134 file_start_outgoing_migration(s
, &addr
->u
.file
, &local_err
);
2136 error_setg(&local_err
, QERR_INVALID_PARAMETER_VALUE
, "uri",
2137 "a valid migration protocol");
2138 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
2139 MIGRATION_STATUS_FAILED
);
2140 block_cleanup_parameters();
2144 if (!resume_requested
) {
2145 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
2147 migrate_fd_error(s
, local_err
);
2148 error_propagate(errp
, local_err
);
2153 void qmp_migrate_cancel(Error
**errp
)
2155 migration_cancel(NULL
);
2158 void qmp_migrate_continue(MigrationStatus state
, Error
**errp
)
2160 MigrationState
*s
= migrate_get_current();
2161 if (s
->state
!= state
) {
2162 error_setg(errp
, "Migration not in expected state: %s",
2163 MigrationStatus_str(s
->state
));
2166 qemu_sem_post(&s
->pause_sem
);
2169 int migration_rp_wait(MigrationState
*s
)
2171 /* If migration has failure already, ignore the wait */
2172 if (migrate_has_error(s
)) {
2176 qemu_sem_wait(&s
->rp_state
.rp_sem
);
2178 /* After wait, double check that there's no failure */
2179 if (migrate_has_error(s
)) {
2186 void migration_rp_kick(MigrationState
*s
)
2188 qemu_sem_post(&s
->rp_state
.rp_sem
);
2191 static struct rp_cmd_args
{
2192 ssize_t len
; /* -1 = variable */
2195 [MIG_RP_MSG_INVALID
] = { .len
= -1, .name
= "INVALID" },
2196 [MIG_RP_MSG_SHUT
] = { .len
= 4, .name
= "SHUT" },
2197 [MIG_RP_MSG_PONG
] = { .len
= 4, .name
= "PONG" },
2198 [MIG_RP_MSG_REQ_PAGES
] = { .len
= 12, .name
= "REQ_PAGES" },
2199 [MIG_RP_MSG_REQ_PAGES_ID
] = { .len
= -1, .name
= "REQ_PAGES_ID" },
2200 [MIG_RP_MSG_RECV_BITMAP
] = { .len
= -1, .name
= "RECV_BITMAP" },
2201 [MIG_RP_MSG_RESUME_ACK
] = { .len
= 4, .name
= "RESUME_ACK" },
2202 [MIG_RP_MSG_SWITCHOVER_ACK
] = { .len
= 0, .name
= "SWITCHOVER_ACK" },
2203 [MIG_RP_MSG_MAX
] = { .len
= -1, .name
= "MAX" },
2207 * Process a request for pages received on the return path,
2208 * We're allowed to send more than requested (e.g. to round to our page size)
2209 * and we don't need to send pages that have already been sent.
2212 migrate_handle_rp_req_pages(MigrationState
*ms
, const char* rbname
,
2213 ram_addr_t start
, size_t len
, Error
**errp
)
2215 long our_host_ps
= qemu_real_host_page_size();
2217 trace_migrate_handle_rp_req_pages(rbname
, start
, len
);
2220 * Since we currently insist on matching page sizes, just sanity check
2221 * we're being asked for whole host pages.
2223 if (!QEMU_IS_ALIGNED(start
, our_host_ps
) ||
2224 !QEMU_IS_ALIGNED(len
, our_host_ps
)) {
2225 error_setg(errp
, "MIG_RP_MSG_REQ_PAGES: Misaligned page request, start:"
2226 RAM_ADDR_FMT
" len: %zd", start
, len
);
2230 ram_save_queue_pages(rbname
, start
, len
, errp
);
2233 static bool migrate_handle_rp_recv_bitmap(MigrationState
*s
, char *block_name
,
2236 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
2239 error_setg(errp
, "MIG_RP_MSG_RECV_BITMAP has invalid block name '%s'",
2244 /* Fetch the received bitmap and refresh the dirty bitmap */
2245 return ram_dirty_bitmap_reload(s
, block
, errp
);
2248 static bool migrate_handle_rp_resume_ack(MigrationState
*s
,
2249 uint32_t value
, Error
**errp
)
2251 trace_source_return_path_thread_resume_ack(value
);
2253 if (value
!= MIGRATION_RESUME_ACK_VALUE
) {
2254 error_setg(errp
, "illegal resume_ack value %"PRIu32
, value
);
2258 /* Now both sides are active. */
2259 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_RECOVER
,
2260 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2262 /* Notify send thread that time to continue send pages */
2263 migration_rp_kick(s
);
2269 * Release ms->rp_state.from_dst_file (and postcopy_qemufile_src if
2270 * existed) in a safe way.
2272 static void migration_release_dst_files(MigrationState
*ms
)
2276 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2278 * Reset the from_dst_file pointer first before releasing it, as we
2279 * can't block within lock section
2281 file
= ms
->rp_state
.from_dst_file
;
2282 ms
->rp_state
.from_dst_file
= NULL
;
2286 * Do the same to postcopy fast path socket too if there is. No
2287 * locking needed because this qemufile should only be managed by
2288 * return path thread.
2290 if (ms
->postcopy_qemufile_src
) {
2291 migration_ioc_unregister_yank_from_file(ms
->postcopy_qemufile_src
);
2292 qemu_file_shutdown(ms
->postcopy_qemufile_src
);
2293 qemu_fclose(ms
->postcopy_qemufile_src
);
2294 ms
->postcopy_qemufile_src
= NULL
;
2301 * Handles messages sent on the return path towards the source VM
2304 static void *source_return_path_thread(void *opaque
)
2306 MigrationState
*ms
= opaque
;
2307 QEMUFile
*rp
= ms
->rp_state
.from_dst_file
;
2308 uint16_t header_len
, header_type
;
2310 uint32_t tmp32
, sibling_error
;
2311 ram_addr_t start
= 0; /* =0 to silence warning */
2312 size_t len
= 0, expected_len
;
2316 trace_source_return_path_thread_entry();
2317 rcu_register_thread();
2319 while (migration_is_setup_or_active()) {
2320 trace_source_return_path_thread_loop_top();
2322 header_type
= qemu_get_be16(rp
);
2323 header_len
= qemu_get_be16(rp
);
2325 if (qemu_file_get_error(rp
)) {
2326 qemu_file_get_error_obj(rp
, &err
);
2330 if (header_type
>= MIG_RP_MSG_MAX
||
2331 header_type
== MIG_RP_MSG_INVALID
) {
2332 error_setg(&err
, "Received invalid message 0x%04x length 0x%04x",
2333 header_type
, header_len
);
2337 if ((rp_cmd_args
[header_type
].len
!= -1 &&
2338 header_len
!= rp_cmd_args
[header_type
].len
) ||
2339 header_len
> sizeof(buf
)) {
2340 error_setg(&err
, "Received '%s' message (0x%04x) with"
2341 "incorrect length %d expecting %zu",
2342 rp_cmd_args
[header_type
].name
, header_type
, header_len
,
2343 (size_t)rp_cmd_args
[header_type
].len
);
2347 /* We know we've got a valid header by this point */
2348 res
= qemu_get_buffer(rp
, buf
, header_len
);
2349 if (res
!= header_len
) {
2350 error_setg(&err
, "Failed reading data for message 0x%04x"
2351 " read %d expected %d",
2352 header_type
, res
, header_len
);
2356 /* OK, we have the message and the data */
2357 switch (header_type
) {
2358 case MIG_RP_MSG_SHUT
:
2359 sibling_error
= ldl_be_p(buf
);
2360 trace_source_return_path_thread_shut(sibling_error
);
2361 if (sibling_error
) {
2362 error_setg(&err
, "Sibling indicated error %d", sibling_error
);
2365 * We'll let the main thread deal with closing the RP
2366 * we could do a shutdown(2) on it, but we're the only user
2367 * anyway, so there's nothing gained.
2371 case MIG_RP_MSG_PONG
:
2372 tmp32
= ldl_be_p(buf
);
2373 trace_source_return_path_thread_pong(tmp32
);
2374 qemu_sem_post(&ms
->rp_state
.rp_pong_acks
);
2377 case MIG_RP_MSG_REQ_PAGES
:
2378 start
= ldq_be_p(buf
);
2379 len
= ldl_be_p(buf
+ 8);
2380 migrate_handle_rp_req_pages(ms
, NULL
, start
, len
, &err
);
2386 case MIG_RP_MSG_REQ_PAGES_ID
:
2387 expected_len
= 12 + 1; /* header + termination */
2389 if (header_len
>= expected_len
) {
2390 start
= ldq_be_p(buf
);
2391 len
= ldl_be_p(buf
+ 8);
2392 /* Now we expect an idstr */
2393 tmp32
= buf
[12]; /* Length of the following idstr */
2394 buf
[13 + tmp32
] = '\0';
2395 expected_len
+= tmp32
;
2397 if (header_len
!= expected_len
) {
2398 error_setg(&err
, "Req_Page_id with length %d expecting %zd",
2399 header_len
, expected_len
);
2402 migrate_handle_rp_req_pages(ms
, (char *)&buf
[13], start
, len
,
2409 case MIG_RP_MSG_RECV_BITMAP
:
2410 if (header_len
< 1) {
2411 error_setg(&err
, "MIG_RP_MSG_RECV_BITMAP missing block name");
2414 /* Format: len (1B) + idstr (<255B). This ends the idstr. */
2415 buf
[buf
[0] + 1] = '\0';
2416 if (!migrate_handle_rp_recv_bitmap(ms
, (char *)(buf
+ 1), &err
)) {
2421 case MIG_RP_MSG_RESUME_ACK
:
2422 tmp32
= ldl_be_p(buf
);
2423 if (!migrate_handle_rp_resume_ack(ms
, tmp32
, &err
)) {
2428 case MIG_RP_MSG_SWITCHOVER_ACK
:
2429 ms
->switchover_acked
= true;
2430 trace_source_return_path_thread_switchover_acked();
2440 migrate_set_error(ms
, err
);
2442 trace_source_return_path_thread_bad_end();
2445 if (ms
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2447 * this will be extremely unlikely: that we got yet another network
2448 * issue during recovering of the 1st network failure.. during this
2449 * period the main migration thread can be waiting on rp_sem for
2450 * this thread to sync with the other side.
2452 * When this happens, explicitly kick the migration thread out of
2453 * RECOVER stage and back to PAUSED, so the admin can try
2456 migration_rp_kick(ms
);
2459 trace_source_return_path_thread_end();
2460 rcu_unregister_thread();
2465 static int open_return_path_on_source(MigrationState
*ms
)
2467 ms
->rp_state
.from_dst_file
= qemu_file_get_return_path(ms
->to_dst_file
);
2468 if (!ms
->rp_state
.from_dst_file
) {
2472 trace_open_return_path_on_source();
2474 qemu_thread_create(&ms
->rp_state
.rp_thread
, "return path",
2475 source_return_path_thread
, ms
, QEMU_THREAD_JOINABLE
);
2476 ms
->rp_state
.rp_thread_created
= true;
2478 trace_open_return_path_on_source_continue();
2483 /* Return true if error detected, or false otherwise */
2484 static bool close_return_path_on_source(MigrationState
*ms
)
2486 if (!ms
->rp_state
.rp_thread_created
) {
2490 trace_migration_return_path_end_before();
2493 * If this is a normal exit then the destination will send a SHUT
2494 * and the rp_thread will exit, however if there's an error we
2495 * need to cause it to exit. shutdown(2), if we have it, will
2496 * cause it to unblock if it's stuck waiting for the destination.
2498 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2499 if (migrate_has_error(ms
) && ms
->rp_state
.from_dst_file
) {
2500 qemu_file_shutdown(ms
->rp_state
.from_dst_file
);
2504 qemu_thread_join(&ms
->rp_state
.rp_thread
);
2505 ms
->rp_state
.rp_thread_created
= false;
2506 migration_release_dst_files(ms
);
2507 trace_migration_return_path_end_after();
2509 /* Return path will persist the error in MigrationState when quit */
2510 return migrate_has_error(ms
);
2514 migration_wait_main_channel(MigrationState
*ms
)
2516 /* Wait until one PONG message received */
2517 qemu_sem_wait(&ms
->rp_state
.rp_pong_acks
);
2521 * Switch from normal iteration to postcopy
2522 * Returns non-0 on error
2524 static int postcopy_start(MigrationState
*ms
, Error
**errp
)
2527 QIOChannelBuffer
*bioc
;
2529 uint64_t bandwidth
= migrate_max_postcopy_bandwidth();
2530 bool restart_block
= false;
2531 int cur_state
= MIGRATION_STATUS_ACTIVE
;
2533 if (migrate_postcopy_preempt()) {
2534 migration_wait_main_channel(ms
);
2535 if (postcopy_preempt_establish_channel(ms
)) {
2536 migrate_set_state(&ms
->state
, ms
->state
, MIGRATION_STATUS_FAILED
);
2537 error_setg(errp
, "%s: Failed to establish preempt channel",
2543 if (!migrate_pause_before_switchover()) {
2544 migrate_set_state(&ms
->state
, MIGRATION_STATUS_ACTIVE
,
2545 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2548 trace_postcopy_start();
2550 trace_postcopy_start_set_run();
2552 ret
= migration_stop_vm(ms
, RUN_STATE_FINISH_MIGRATE
);
2554 error_setg_errno(errp
, -ret
, "%s: Failed to stop the VM", __func__
);
2558 ret
= migration_maybe_pause(ms
, &cur_state
,
2559 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2561 error_setg_errno(errp
, -ret
, "%s: Failed in migration_maybe_pause()",
2566 ret
= bdrv_inactivate_all();
2568 error_setg_errno(errp
, -ret
, "%s: Failed in bdrv_inactivate_all()",
2572 restart_block
= true;
2575 * Cause any non-postcopiable, but iterative devices to
2576 * send out their final data.
2578 qemu_savevm_state_complete_precopy(ms
->to_dst_file
, true, false);
2581 * in Finish migrate and with the io-lock held everything should
2582 * be quiet, but we've potentially still got dirty pages and we
2583 * need to tell the destination to throw any pages it's already received
2586 if (migrate_postcopy_ram()) {
2587 ram_postcopy_send_discard_bitmap(ms
);
2591 * send rest of state - note things that are doing postcopy
2592 * will notice we're in POSTCOPY_ACTIVE and not actually
2593 * wrap their state up here
2595 migration_rate_set(bandwidth
);
2596 if (migrate_postcopy_ram()) {
2597 /* Ping just for debugging, helps line traces up */
2598 qemu_savevm_send_ping(ms
->to_dst_file
, 2);
2602 * While loading the device state we may trigger page transfer
2603 * requests and the fd must be free to process those, and thus
2604 * the destination must read the whole device state off the fd before
2605 * it starts processing it. Unfortunately the ad-hoc migration format
2606 * doesn't allow the destination to know the size to read without fully
2607 * parsing it through each devices load-state code (especially the open
2608 * coded devices that use get/put).
2609 * So we wrap the device state up in a package with a length at the start;
2610 * to do this we use a qemu_buf to hold the whole of the device state.
2612 bioc
= qio_channel_buffer_new(4096);
2613 qio_channel_set_name(QIO_CHANNEL(bioc
), "migration-postcopy-buffer");
2614 fb
= qemu_file_new_output(QIO_CHANNEL(bioc
));
2615 object_unref(OBJECT(bioc
));
2618 * Make sure the receiver can get incoming pages before we send the rest
2621 qemu_savevm_send_postcopy_listen(fb
);
2623 qemu_savevm_state_complete_precopy(fb
, false, false);
2624 if (migrate_postcopy_ram()) {
2625 qemu_savevm_send_ping(fb
, 3);
2628 qemu_savevm_send_postcopy_run(fb
);
2630 /* <><> end of stuff going into the package */
2632 /* Last point of recovery; as soon as we send the package the destination
2633 * can open devices and potentially start running.
2634 * Lets just check again we've not got any errors.
2636 ret
= qemu_file_get_error(ms
->to_dst_file
);
2638 error_setg(errp
, "postcopy_start: Migration stream errored (pre package)");
2642 restart_block
= false;
2644 /* Now send that blob */
2645 if (qemu_savevm_send_packaged(ms
->to_dst_file
, bioc
->data
, bioc
->usage
)) {
2646 error_setg(errp
, "%s: Failed to send packaged data", __func__
);
2651 /* Send a notify to give a chance for anything that needs to happen
2652 * at the transition to postcopy and after the device state; in particular
2653 * spice needs to trigger a transition now
2655 migration_call_notifiers(ms
, MIG_EVENT_PRECOPY_DONE
, NULL
);
2657 migration_downtime_end(ms
);
2661 if (migrate_postcopy_ram()) {
2663 * Although this ping is just for debug, it could potentially be
2664 * used for getting a better measurement of downtime at the source.
2666 qemu_savevm_send_ping(ms
->to_dst_file
, 4);
2669 if (migrate_release_ram()) {
2670 ram_postcopy_migrated_memory_release(ms
);
2673 ret
= qemu_file_get_error(ms
->to_dst_file
);
2675 error_setg_errno(errp
, -ret
, "postcopy_start: Migration stream error");
2679 trace_postcopy_preempt_enabled(migrate_postcopy_preempt());
2686 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2687 MIGRATION_STATUS_FAILED
);
2688 if (restart_block
) {
2689 /* A failure happened early enough that we know the destination hasn't
2690 * accessed block devices, so we're safe to recover.
2692 Error
*local_err
= NULL
;
2694 bdrv_activate_all(&local_err
);
2696 error_report_err(local_err
);
2699 migration_call_notifiers(ms
, MIG_EVENT_PRECOPY_FAILED
, NULL
);
2705 * migration_maybe_pause: Pause if required to by
2706 * migrate_pause_before_switchover called with the BQL locked
2707 * Returns: 0 on success
2709 static int migration_maybe_pause(MigrationState
*s
,
2710 int *current_active_state
,
2713 if (!migrate_pause_before_switchover()) {
2717 /* Since leaving this state is not atomic with posting the semaphore
2718 * it's possible that someone could have issued multiple migrate_continue
2719 * and the semaphore is incorrectly positive at this point;
2720 * the docs say it's undefined to reinit a semaphore that's already
2721 * init'd, so use timedwait to eat up any existing posts.
2723 while (qemu_sem_timedwait(&s
->pause_sem
, 1) == 0) {
2724 /* This block intentionally left blank */
2728 * If the migration is cancelled when it is in the completion phase,
2729 * the migration state is set to MIGRATION_STATUS_CANCELLING.
2730 * So we don't need to wait a semaphore, otherwise we would always
2731 * wait for the 'pause_sem' semaphore.
2733 if (s
->state
!= MIGRATION_STATUS_CANCELLING
) {
2735 migrate_set_state(&s
->state
, *current_active_state
,
2736 MIGRATION_STATUS_PRE_SWITCHOVER
);
2737 qemu_sem_wait(&s
->pause_sem
);
2738 migrate_set_state(&s
->state
, MIGRATION_STATUS_PRE_SWITCHOVER
,
2740 *current_active_state
= new_state
;
2744 return s
->state
== new_state
? 0 : -EINVAL
;
2747 static int migration_completion_precopy(MigrationState
*s
,
2748 int *current_active_state
)
2754 if (!migrate_mode_is_cpr(s
)) {
2755 ret
= migration_stop_vm(s
, RUN_STATE_FINISH_MIGRATE
);
2761 ret
= migration_maybe_pause(s
, current_active_state
,
2762 MIGRATION_STATUS_DEVICE
);
2768 * Inactivate disks except in COLO, and track that we have done so in order
2769 * to remember to reactivate them if migration fails or is cancelled.
2771 s
->block_inactive
= !migrate_colo();
2772 migration_rate_set(RATE_LIMIT_DISABLED
);
2773 ret
= qemu_savevm_state_complete_precopy(s
->to_dst_file
, false,
2780 static void migration_completion_postcopy(MigrationState
*s
)
2782 trace_migration_completion_postcopy_end();
2785 qemu_savevm_state_complete_postcopy(s
->to_dst_file
);
2789 * Shutdown the postcopy fast path thread. This is only needed when dest
2790 * QEMU binary is old (7.1/7.2). QEMU 8.0+ doesn't need this.
2792 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
2793 postcopy_preempt_shutdown_file(s
);
2796 trace_migration_completion_postcopy_end_after_complete();
2799 static void migration_completion_failed(MigrationState
*s
,
2800 int current_active_state
)
2802 if (s
->block_inactive
&& (s
->state
== MIGRATION_STATUS_ACTIVE
||
2803 s
->state
== MIGRATION_STATUS_DEVICE
)) {
2805 * If not doing postcopy, vm_start() will be called: let's
2806 * regain control on images.
2808 Error
*local_err
= NULL
;
2811 bdrv_activate_all(&local_err
);
2813 error_report_err(local_err
);
2815 s
->block_inactive
= false;
2820 migrate_set_state(&s
->state
, current_active_state
,
2821 MIGRATION_STATUS_FAILED
);
2825 * migration_completion: Used by migration_thread when there's not much left.
2826 * The caller 'breaks' the loop when this returns.
2828 * @s: Current migration state
2830 static void migration_completion(MigrationState
*s
)
2833 int current_active_state
= s
->state
;
2834 Error
*local_err
= NULL
;
2836 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2837 ret
= migration_completion_precopy(s
, ¤t_active_state
);
2838 } else if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2839 migration_completion_postcopy(s
);
2848 if (close_return_path_on_source(s
)) {
2852 if (qemu_file_get_error(s
->to_dst_file
)) {
2853 trace_migration_completion_file_err();
2857 if (migrate_colo() && s
->state
== MIGRATION_STATUS_ACTIVE
) {
2858 /* COLO does not support postcopy */
2859 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
2860 MIGRATION_STATUS_COLO
);
2862 migration_completion_end(s
);
2868 if (qemu_file_get_error_obj(s
->to_dst_file
, &local_err
)) {
2869 migrate_set_error(s
, local_err
);
2870 error_free(local_err
);
2872 error_setg_errno(&local_err
, -ret
, "Error in migration completion");
2873 migrate_set_error(s
, local_err
);
2874 error_free(local_err
);
2877 migration_completion_failed(s
, current_active_state
);
2881 * bg_migration_completion: Used by bg_migration_thread when after all the
2882 * RAM has been saved. The caller 'breaks' the loop when this returns.
2884 * @s: Current migration state
2886 static void bg_migration_completion(MigrationState
*s
)
2888 int current_active_state
= s
->state
;
2890 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2892 * By this moment we have RAM content saved into the migration stream.
2893 * The next step is to flush the non-RAM content (device state)
2894 * right after the ram content. The device state has been stored into
2895 * the temporary buffer before RAM saving started.
2897 qemu_put_buffer(s
->to_dst_file
, s
->bioc
->data
, s
->bioc
->usage
);
2898 qemu_fflush(s
->to_dst_file
);
2899 } else if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
2903 if (qemu_file_get_error(s
->to_dst_file
)) {
2904 trace_migration_completion_file_err();
2908 migration_completion_end(s
);
2912 migrate_set_state(&s
->state
, current_active_state
,
2913 MIGRATION_STATUS_FAILED
);
2916 typedef enum MigThrError
{
2917 /* No error detected */
2918 MIG_THR_ERR_NONE
= 0,
2919 /* Detected error, but resumed successfully */
2920 MIG_THR_ERR_RECOVERED
= 1,
2921 /* Detected fatal error, need to exit */
2922 MIG_THR_ERR_FATAL
= 2,
2925 static int postcopy_resume_handshake(MigrationState
*s
)
2927 qemu_savevm_send_postcopy_resume(s
->to_dst_file
);
2929 while (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2930 if (migration_rp_wait(s
)) {
2935 if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2942 /* Return zero if success, or <0 for error */
2943 static int postcopy_do_resume(MigrationState
*s
)
2948 * Call all the resume_prepare() hooks, so that modules can be
2949 * ready for the migration resume.
2951 ret
= qemu_savevm_state_resume_prepare(s
);
2953 error_report("%s: resume_prepare() failure detected: %d",
2959 * If preempt is enabled, re-establish the preempt channel. Note that
2960 * we do it after resume prepare to make sure the main channel will be
2961 * created before the preempt channel. E.g. with weak network, the
2962 * dest QEMU may get messed up with the preempt and main channels on
2963 * the order of connection setup. This guarantees the correct order.
2965 ret
= postcopy_preempt_establish_channel(s
);
2967 error_report("%s: postcopy_preempt_establish_channel(): %d",
2973 * Last handshake with destination on the resume (destination will
2974 * switch to postcopy-active afterwards)
2976 ret
= postcopy_resume_handshake(s
);
2978 error_report("%s: handshake failed: %d", __func__
, ret
);
2986 * We don't return until we are in a safe state to continue current
2987 * postcopy migration. Returns MIG_THR_ERR_RECOVERED if recovered, or
2988 * MIG_THR_ERR_FATAL if unrecovery failure happened.
2990 static MigThrError
postcopy_pause(MigrationState
*s
)
2992 assert(s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2998 * We're already pausing, so ignore any errors on the return
2999 * path and just wait for the thread to finish. It will be
3000 * re-created when we resume.
3002 close_return_path_on_source(s
);
3005 * Current channel is possibly broken. Release it. Note that this is
3006 * guaranteed even without lock because to_dst_file should only be
3007 * modified by the migration thread. That also guarantees that the
3008 * unregister of yank is safe too without the lock. It should be safe
3009 * even to be within the qemu_file_lock, but we didn't do that to avoid
3010 * taking more mutex (yank_lock) within qemu_file_lock. TL;DR: we make
3011 * the qemu_file_lock critical section as small as possible.
3013 assert(s
->to_dst_file
);
3014 migration_ioc_unregister_yank_from_file(s
->to_dst_file
);
3015 qemu_mutex_lock(&s
->qemu_file_lock
);
3016 file
= s
->to_dst_file
;
3017 s
->to_dst_file
= NULL
;
3018 qemu_mutex_unlock(&s
->qemu_file_lock
);
3020 qemu_file_shutdown(file
);
3023 migrate_set_state(&s
->state
, s
->state
,
3024 MIGRATION_STATUS_POSTCOPY_PAUSED
);
3026 error_report("Detected IO failure for postcopy. "
3027 "Migration paused.");
3030 * We wait until things fixed up. Then someone will setup the
3031 * status back for us.
3033 while (s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
3034 qemu_sem_wait(&s
->postcopy_pause_sem
);
3037 if (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
3038 /* Woken up by a recover procedure. Give it a shot */
3040 /* Do the resume logic */
3041 if (postcopy_do_resume(s
) == 0) {
3042 /* Let's continue! */
3043 trace_postcopy_pause_continued();
3044 return MIG_THR_ERR_RECOVERED
;
3047 * Something wrong happened during the recovery, let's
3048 * pause again. Pause is always better than throwing
3054 /* This is not right... Time to quit. */
3055 return MIG_THR_ERR_FATAL
;
3060 void migration_file_set_error(int err
)
3062 MigrationState
*s
= current_migration
;
3064 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
3065 if (s
->to_dst_file
) {
3066 qemu_file_set_error(s
->to_dst_file
, err
);
3071 static MigThrError
migration_detect_error(MigrationState
*s
)
3074 int state
= s
->state
;
3075 Error
*local_error
= NULL
;
3077 if (state
== MIGRATION_STATUS_CANCELLING
||
3078 state
== MIGRATION_STATUS_CANCELLED
) {
3079 /* End the migration, but don't set the state to failed */
3080 return MIG_THR_ERR_FATAL
;
3084 * Try to detect any file errors. Note that postcopy_qemufile_src will
3085 * be NULL when postcopy preempt is not enabled.
3087 ret
= qemu_file_get_error_obj_any(s
->to_dst_file
,
3088 s
->postcopy_qemufile_src
,
3091 /* Everything is fine */
3092 assert(!local_error
);
3093 return MIG_THR_ERR_NONE
;
3097 migrate_set_error(s
, local_error
);
3098 error_free(local_error
);
3101 if (state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
&& ret
) {
3103 * For postcopy, we allow the network to be down for a
3104 * while. After that, it can be continued by a
3107 return postcopy_pause(s
);
3110 * For precopy (or postcopy with error outside IO), we fail
3113 migrate_set_state(&s
->state
, state
, MIGRATION_STATUS_FAILED
);
3114 trace_migration_thread_file_err();
3116 /* Time to stop the migration, now. */
3117 return MIG_THR_ERR_FATAL
;
3121 static void migration_completion_end(MigrationState
*s
)
3123 uint64_t bytes
= migration_transferred_bytes();
3124 int64_t end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3125 int64_t transfer_time
;
3128 * Take the BQL here so that query-migrate on the QMP thread sees:
3129 * - atomic update of s->total_time and s->mbps;
3130 * - correct ordering of s->mbps update vs. s->state;
3133 migration_downtime_end(s
);
3134 s
->total_time
= end_time
- s
->start_time
;
3135 transfer_time
= s
->total_time
- s
->setup_time
;
3136 if (transfer_time
) {
3137 s
->mbps
= ((double) bytes
* 8.0) / transfer_time
/ 1000;
3140 migrate_set_state(&s
->state
, s
->state
,
3141 MIGRATION_STATUS_COMPLETED
);
3145 static void update_iteration_initial_status(MigrationState
*s
)
3148 * Update these three fields at the same time to avoid mismatch info lead
3149 * wrong speed calculation.
3151 s
->iteration_start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3152 s
->iteration_initial_bytes
= migration_transferred_bytes();
3153 s
->iteration_initial_pages
= ram_get_total_transferred_pages();
3156 static void migration_update_counters(MigrationState
*s
,
3157 int64_t current_time
)
3159 uint64_t transferred
, transferred_pages
, time_spent
;
3160 uint64_t current_bytes
; /* bytes transferred since the beginning */
3161 uint64_t switchover_bw
;
3162 /* Expected bandwidth when switching over to destination QEMU */
3163 double expected_bw_per_ms
;
3166 if (current_time
< s
->iteration_start_time
+ BUFFER_DELAY
) {
3170 switchover_bw
= migrate_avail_switchover_bandwidth();
3171 current_bytes
= migration_transferred_bytes();
3172 transferred
= current_bytes
- s
->iteration_initial_bytes
;
3173 time_spent
= current_time
- s
->iteration_start_time
;
3174 bandwidth
= (double)transferred
/ time_spent
;
3176 if (switchover_bw
) {
3178 * If the user specified a switchover bandwidth, let's trust the
3179 * user so that can be more accurate than what we estimated.
3181 expected_bw_per_ms
= switchover_bw
/ 1000;
3183 /* If the user doesn't specify bandwidth, we use the estimated */
3184 expected_bw_per_ms
= bandwidth
;
3187 s
->threshold_size
= expected_bw_per_ms
* migrate_downtime_limit();
3189 s
->mbps
= (((double) transferred
* 8.0) /
3190 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0;
3192 transferred_pages
= ram_get_total_transferred_pages() -
3193 s
->iteration_initial_pages
;
3194 s
->pages_per_second
= (double) transferred_pages
/
3195 (((double) time_spent
/ 1000.0));
3198 * if we haven't sent anything, we don't want to
3199 * recalculate. 10000 is a small enough number for our purposes
3201 if (stat64_get(&mig_stats
.dirty_pages_rate
) &&
3202 transferred
> 10000) {
3203 s
->expected_downtime
=
3204 stat64_get(&mig_stats
.dirty_bytes_last_sync
) / expected_bw_per_ms
;
3207 migration_rate_reset();
3209 update_iteration_initial_status(s
);
3211 trace_migrate_transferred(transferred
, time_spent
,
3212 /* Both in unit bytes/ms */
3213 bandwidth
, switchover_bw
/ 1000,
3217 static bool migration_can_switchover(MigrationState
*s
)
3219 if (!migrate_switchover_ack()) {
3223 /* No reason to wait for switchover ACK if VM is stopped */
3224 if (!runstate_is_running()) {
3228 return s
->switchover_acked
;
3231 /* Migration thread iteration status */
3233 MIG_ITERATE_RESUME
, /* Resume current iteration */
3234 MIG_ITERATE_SKIP
, /* Skip current iteration */
3235 MIG_ITERATE_BREAK
, /* Break the loop */
3239 * Return true if continue to the next iteration directly, false
3242 static MigIterateState
migration_iteration_run(MigrationState
*s
)
3244 uint64_t must_precopy
, can_postcopy
, pending_size
;
3245 Error
*local_err
= NULL
;
3246 bool in_postcopy
= s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
;
3247 bool can_switchover
= migration_can_switchover(s
);
3249 qemu_savevm_state_pending_estimate(&must_precopy
, &can_postcopy
);
3250 pending_size
= must_precopy
+ can_postcopy
;
3251 trace_migrate_pending_estimate(pending_size
, must_precopy
, can_postcopy
);
3253 if (pending_size
< s
->threshold_size
) {
3254 qemu_savevm_state_pending_exact(&must_precopy
, &can_postcopy
);
3255 pending_size
= must_precopy
+ can_postcopy
;
3256 trace_migrate_pending_exact(pending_size
, must_precopy
, can_postcopy
);
3259 if ((!pending_size
|| pending_size
< s
->threshold_size
) && can_switchover
) {
3260 trace_migration_thread_low_pending(pending_size
);
3261 migration_completion(s
);
3262 return MIG_ITERATE_BREAK
;
3265 /* Still a significant amount to transfer */
3266 if (!in_postcopy
&& must_precopy
<= s
->threshold_size
&& can_switchover
&&
3267 qatomic_read(&s
->start_postcopy
)) {
3268 if (postcopy_start(s
, &local_err
)) {
3269 migrate_set_error(s
, local_err
);
3270 error_report_err(local_err
);
3272 return MIG_ITERATE_SKIP
;
3275 /* Just another iteration step */
3276 qemu_savevm_state_iterate(s
->to_dst_file
, in_postcopy
);
3277 return MIG_ITERATE_RESUME
;
3280 static void migration_iteration_finish(MigrationState
*s
)
3282 /* If we enabled cpu throttling for auto-converge, turn it off. */
3283 cpu_throttle_stop();
3287 case MIGRATION_STATUS_COMPLETED
:
3288 runstate_set(RUN_STATE_POSTMIGRATE
);
3290 case MIGRATION_STATUS_COLO
:
3291 assert(migrate_colo());
3292 migrate_start_colo_process(s
);
3293 s
->vm_old_state
= RUN_STATE_RUNNING
;
3295 case MIGRATION_STATUS_FAILED
:
3296 case MIGRATION_STATUS_CANCELLED
:
3297 case MIGRATION_STATUS_CANCELLING
:
3298 if (runstate_is_live(s
->vm_old_state
)) {
3299 if (!runstate_check(RUN_STATE_SHUTDOWN
)) {
3303 if (runstate_check(RUN_STATE_FINISH_MIGRATE
)) {
3304 runstate_set(s
->vm_old_state
);
3310 /* Should not reach here, but if so, forgive the VM. */
3311 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3315 migration_bh_schedule(migrate_fd_cleanup_bh
, s
);
3319 static void bg_migration_iteration_finish(MigrationState
*s
)
3322 * Stop tracking RAM writes - un-protect memory, un-register UFFD
3323 * memory ranges, flush kernel wait queues and wake up threads
3324 * waiting for write fault to be resolved.
3326 ram_write_tracking_stop();
3330 case MIGRATION_STATUS_COMPLETED
:
3331 case MIGRATION_STATUS_ACTIVE
:
3332 case MIGRATION_STATUS_FAILED
:
3333 case MIGRATION_STATUS_CANCELLED
:
3334 case MIGRATION_STATUS_CANCELLING
:
3338 /* Should not reach here, but if so, forgive the VM. */
3339 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3343 migration_bh_schedule(migrate_fd_cleanup_bh
, s
);
3348 * Return true if continue to the next iteration directly, false
3351 static MigIterateState
bg_migration_iteration_run(MigrationState
*s
)
3355 res
= qemu_savevm_state_iterate(s
->to_dst_file
, false);
3357 bg_migration_completion(s
);
3358 return MIG_ITERATE_BREAK
;
3361 return MIG_ITERATE_RESUME
;
3364 void migration_make_urgent_request(void)
3366 qemu_sem_post(&migrate_get_current()->rate_limit_sem
);
3369 void migration_consume_urgent_request(void)
3371 qemu_sem_wait(&migrate_get_current()->rate_limit_sem
);
3374 /* Returns true if the rate limiting was broken by an urgent request */
3375 bool migration_rate_limit(void)
3377 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3378 MigrationState
*s
= migrate_get_current();
3380 bool urgent
= false;
3381 migration_update_counters(s
, now
);
3382 if (migration_rate_exceeded(s
->to_dst_file
)) {
3384 if (qemu_file_get_error(s
->to_dst_file
)) {
3388 * Wait for a delay to do rate limiting OR
3389 * something urgent to post the semaphore.
3391 int ms
= s
->iteration_start_time
+ BUFFER_DELAY
- now
;
3392 trace_migration_rate_limit_pre(ms
);
3393 if (qemu_sem_timedwait(&s
->rate_limit_sem
, ms
) == 0) {
3395 * We were woken by one or more urgent things but
3396 * the timedwait will have consumed one of them.
3397 * The service routine for the urgent wake will dec
3398 * the semaphore itself for each item it consumes,
3399 * so add this one we just eat back.
3401 qemu_sem_post(&s
->rate_limit_sem
);
3404 trace_migration_rate_limit_post(urgent
);
3410 * if failover devices are present, wait they are completely
3414 static void qemu_savevm_wait_unplug(MigrationState
*s
, int old_state
,
3417 if (qemu_savevm_state_guest_unplug_pending()) {
3418 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_WAIT_UNPLUG
);
3420 while (s
->state
== MIGRATION_STATUS_WAIT_UNPLUG
&&
3421 qemu_savevm_state_guest_unplug_pending()) {
3422 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3424 if (s
->state
!= MIGRATION_STATUS_WAIT_UNPLUG
) {
3425 int timeout
= 120; /* 30 seconds */
3427 * migration has been canceled
3428 * but as we have started an unplug we must wait the end
3429 * to be able to plug back the card
3431 while (timeout
-- && qemu_savevm_state_guest_unplug_pending()) {
3432 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3434 if (qemu_savevm_state_guest_unplug_pending() &&
3436 warn_report("migration: partially unplugged device on "
3441 migrate_set_state(&s
->state
, MIGRATION_STATUS_WAIT_UNPLUG
, new_state
);
3443 migrate_set_state(&s
->state
, old_state
, new_state
);
3448 * Master migration thread on the source VM.
3449 * It drives the migration and pumps the data down the outgoing channel.
3451 static void *migration_thread(void *opaque
)
3453 MigrationState
*s
= opaque
;
3454 MigrationThread
*thread
= NULL
;
3455 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3456 MigThrError thr_error
;
3457 bool urgent
= false;
3458 Error
*local_err
= NULL
;
3461 thread
= migration_threads_add("live_migration", qemu_get_thread_id());
3463 rcu_register_thread();
3465 object_ref(OBJECT(s
));
3466 update_iteration_initial_status(s
);
3468 if (!multifd_send_setup()) {
3473 qemu_savevm_state_header(s
->to_dst_file
);
3477 * If we opened the return path, we need to make sure dst has it
3480 if (s
->rp_state
.rp_thread_created
) {
3481 /* Now tell the dest that it should open its end so it can reply */
3482 qemu_savevm_send_open_return_path(s
->to_dst_file
);
3484 /* And do a ping that will make stuff easier to debug */
3485 qemu_savevm_send_ping(s
->to_dst_file
, 1);
3488 if (migrate_postcopy()) {
3490 * Tell the destination that we *might* want to do postcopy later;
3491 * if the other end can't do postcopy it should fail now, nice and
3494 qemu_savevm_send_postcopy_advise(s
->to_dst_file
);
3497 if (migrate_colo()) {
3498 /* Notify migration destination that we enable COLO */
3499 qemu_savevm_send_colo_enable(s
->to_dst_file
);
3503 ret
= qemu_savevm_state_setup(s
->to_dst_file
, &local_err
);
3506 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3507 MIGRATION_STATUS_ACTIVE
);
3510 * Handle SETUP failures after waiting for virtio-net-failover
3511 * devices to unplug. This to preserve migration state transitions.
3514 migrate_set_error(s
, local_err
);
3515 error_free(local_err
);
3516 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3517 MIGRATION_STATUS_FAILED
);
3521 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3523 trace_migration_thread_setup_complete();
3525 while (migration_is_active()) {
3526 if (urgent
|| !migration_rate_exceeded(s
->to_dst_file
)) {
3527 MigIterateState iter_state
= migration_iteration_run(s
);
3528 if (iter_state
== MIG_ITERATE_SKIP
) {
3530 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3536 * Try to detect any kind of failures, and see whether we
3537 * should stop the migration now.
3539 thr_error
= migration_detect_error(s
);
3540 if (thr_error
== MIG_THR_ERR_FATAL
) {
3541 /* Stop migration */
3543 } else if (thr_error
== MIG_THR_ERR_RECOVERED
) {
3545 * Just recovered from a e.g. network failure, reset all
3546 * the local variables. This is important to avoid
3547 * breaking transferred_bytes and bandwidth calculation
3549 update_iteration_initial_status(s
);
3552 urgent
= migration_rate_limit();
3556 trace_migration_thread_after_loop();
3557 migration_iteration_finish(s
);
3558 object_unref(OBJECT(s
));
3559 rcu_unregister_thread();
3560 migration_threads_remove(thread
);
3564 static void bg_migration_vm_start_bh(void *opaque
)
3566 MigrationState
*s
= opaque
;
3568 vm_resume(s
->vm_old_state
);
3569 migration_downtime_end(s
);
3573 * Background snapshot thread, based on live migration code.
3574 * This is an alternative implementation of live migration mechanism
3575 * introduced specifically to support background snapshots.
3577 * It takes advantage of userfault_fd write protection mechanism introduced
3578 * in v5.7 kernel. Compared to existing dirty page logging migration much
3579 * lesser stream traffic is produced resulting in smaller snapshot images,
3580 * simply cause of no page duplicates can get into the stream.
3582 * Another key point is that generated vmstate stream reflects machine state
3583 * 'frozen' at the beginning of snapshot creation compared to dirty page logging
3584 * mechanism, which effectively results in that saved snapshot is the state of VM
3585 * at the end of the process.
3587 static void *bg_migration_thread(void *opaque
)
3589 MigrationState
*s
= opaque
;
3590 int64_t setup_start
;
3591 MigThrError thr_error
;
3593 bool early_fail
= true;
3594 Error
*local_err
= NULL
;
3597 rcu_register_thread();
3598 object_ref(OBJECT(s
));
3600 migration_rate_set(RATE_LIMIT_DISABLED
);
3602 setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3604 * We want to save vmstate for the moment when migration has been
3605 * initiated but also we want to save RAM content while VM is running.
3606 * The RAM content should appear first in the vmstate. So, we first
3607 * stash the non-RAM part of the vmstate to the temporary buffer,
3608 * then write RAM part of the vmstate to the migration stream
3609 * with vCPUs running and, finally, write stashed non-RAM part of
3610 * the vmstate from the buffer to the migration stream.
3612 s
->bioc
= qio_channel_buffer_new(512 * 1024);
3613 qio_channel_set_name(QIO_CHANNEL(s
->bioc
), "vmstate-buffer");
3614 fb
= qemu_file_new_output(QIO_CHANNEL(s
->bioc
));
3615 object_unref(OBJECT(s
->bioc
));
3617 update_iteration_initial_status(s
);
3620 * Prepare for tracking memory writes with UFFD-WP - populate
3621 * RAM pages before protecting.
3624 ram_write_tracking_prepare();
3628 qemu_savevm_state_header(s
->to_dst_file
);
3629 ret
= qemu_savevm_state_setup(s
->to_dst_file
, &local_err
);
3632 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3633 MIGRATION_STATUS_ACTIVE
);
3636 * Handle SETUP failures after waiting for virtio-net-failover
3637 * devices to unplug. This to preserve migration state transitions.
3640 migrate_set_error(s
, local_err
);
3641 error_free(local_err
);
3642 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3643 MIGRATION_STATUS_FAILED
);
3647 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3649 trace_migration_thread_setup_complete();
3653 if (migration_stop_vm(s
, RUN_STATE_PAUSED
)) {
3657 * Put vCPUs in sync with shadow context structures, then
3658 * save their state to channel-buffer along with devices.
3660 cpu_synchronize_all_states();
3661 if (qemu_savevm_state_complete_precopy_non_iterable(fb
, false, false)) {
3665 * Since we are going to get non-iterable state data directly
3666 * from s->bioc->data, explicit flush is needed here.
3670 /* Now initialize UFFD context and start tracking RAM writes */
3671 if (ram_write_tracking_start()) {
3677 * Start VM from BH handler to avoid write-fault lock here.
3678 * UFFD-WP protection for the whole RAM is already enabled so
3679 * calling VM state change notifiers from vm_start() would initiate
3680 * writes to virtio VQs memory which is in write-protected region.
3682 migration_bh_schedule(bg_migration_vm_start_bh
, s
);
3685 while (migration_is_active()) {
3686 MigIterateState iter_state
= bg_migration_iteration_run(s
);
3687 if (iter_state
== MIG_ITERATE_SKIP
) {
3689 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3694 * Try to detect any kind of failures, and see whether we
3695 * should stop the migration now.
3697 thr_error
= migration_detect_error(s
);
3698 if (thr_error
== MIG_THR_ERR_FATAL
) {
3699 /* Stop migration */
3703 migration_update_counters(s
, qemu_clock_get_ms(QEMU_CLOCK_REALTIME
));
3706 trace_migration_thread_after_loop();
3710 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3711 MIGRATION_STATUS_FAILED
);
3716 bg_migration_iteration_finish(s
);
3719 object_unref(OBJECT(s
));
3720 rcu_unregister_thread();
3725 void migrate_fd_connect(MigrationState
*s
, Error
*error_in
)
3727 Error
*local_err
= NULL
;
3728 uint64_t rate_limit
;
3729 bool resume
= s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
;
3733 * If there's a previous error, free it and prepare for another one.
3734 * Meanwhile if migration completes successfully, there won't have an error
3735 * dumped when calling migrate_fd_cleanup().
3737 migrate_error_free(s
);
3739 s
->expected_downtime
= migrate_downtime_limit();
3741 migrate_fd_error(s
, error_in
);
3744 * Don't do cleanup for resume if channel is invalid, but only dump
3745 * the error. We wait for another channel connect from the user.
3746 * The error_report still gives HMP user a hint on what failed.
3747 * It's normally done in migrate_fd_cleanup(), but call it here
3750 error_report_err(error_copy(s
->error
));
3752 migrate_fd_cleanup(s
);
3758 /* This is a resumed migration */
3759 rate_limit
= migrate_max_postcopy_bandwidth();
3761 /* This is a fresh new migration */
3762 rate_limit
= migrate_max_bandwidth();
3764 /* Notify before starting migration thread */
3765 if (migration_call_notifiers(s
, MIG_EVENT_PRECOPY_SETUP
, &local_err
)) {
3770 migration_rate_set(rate_limit
);
3771 qemu_file_set_blocking(s
->to_dst_file
, true);
3774 * Open the return path. For postcopy, it is used exclusively. For
3775 * precopy, only if user specified "return-path" capability would
3776 * QEMU uses the return path.
3778 if (migrate_postcopy_ram() || migrate_return_path()) {
3779 if (open_return_path_on_source(s
)) {
3780 error_setg(&local_err
, "Unable to open return-path for postcopy");
3786 * This needs to be done before resuming a postcopy. Note: for newer
3787 * QEMUs we will delay the channel creation until postcopy_start(), to
3788 * avoid disorder of channel creations.
3790 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
3791 postcopy_preempt_setup(s
);
3795 /* Wakeup the main migration thread to do the recovery */
3796 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
3797 MIGRATION_STATUS_POSTCOPY_RECOVER
);
3798 qemu_sem_post(&s
->postcopy_pause_sem
);
3802 if (migrate_mode_is_cpr(s
)) {
3803 ret
= migration_stop_vm(s
, RUN_STATE_FINISH_MIGRATE
);
3805 error_setg(&local_err
, "migration_stop_vm failed, error %d", -ret
);
3810 if (migrate_background_snapshot()) {
3811 qemu_thread_create(&s
->thread
, "bg_snapshot",
3812 bg_migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3814 qemu_thread_create(&s
->thread
, "live_migration",
3815 migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3817 s
->migration_thread_running
= true;
3821 migrate_set_error(s
, local_err
);
3822 migrate_set_state(&s
->state
, s
->state
, MIGRATION_STATUS_FAILED
);
3823 error_report_err(local_err
);
3824 migrate_fd_cleanup(s
);
3827 static void migration_class_init(ObjectClass
*klass
, void *data
)
3829 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3831 dc
->user_creatable
= false;
3832 device_class_set_props(dc
, migration_properties
);
3835 static void migration_instance_finalize(Object
*obj
)
3837 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3839 qemu_mutex_destroy(&ms
->error_mutex
);
3840 qemu_mutex_destroy(&ms
->qemu_file_lock
);
3841 qemu_sem_destroy(&ms
->wait_unplug_sem
);
3842 qemu_sem_destroy(&ms
->rate_limit_sem
);
3843 qemu_sem_destroy(&ms
->pause_sem
);
3844 qemu_sem_destroy(&ms
->postcopy_pause_sem
);
3845 qemu_sem_destroy(&ms
->rp_state
.rp_sem
);
3846 qemu_sem_destroy(&ms
->rp_state
.rp_pong_acks
);
3847 qemu_sem_destroy(&ms
->postcopy_qemufile_src_sem
);
3848 error_free(ms
->error
);
3851 static void migration_instance_init(Object
*obj
)
3853 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3855 ms
->state
= MIGRATION_STATUS_NONE
;
3857 ms
->pages_per_second
= -1;
3858 qemu_sem_init(&ms
->pause_sem
, 0);
3859 qemu_mutex_init(&ms
->error_mutex
);
3861 migrate_params_init(&ms
->parameters
);
3863 qemu_sem_init(&ms
->postcopy_pause_sem
, 0);
3864 qemu_sem_init(&ms
->rp_state
.rp_sem
, 0);
3865 qemu_sem_init(&ms
->rp_state
.rp_pong_acks
, 0);
3866 qemu_sem_init(&ms
->rate_limit_sem
, 0);
3867 qemu_sem_init(&ms
->wait_unplug_sem
, 0);
3868 qemu_sem_init(&ms
->postcopy_qemufile_src_sem
, 0);
3869 qemu_mutex_init(&ms
->qemu_file_lock
);
3873 * Return true if check pass, false otherwise. Error will be put
3874 * inside errp if provided.
3876 static bool migration_object_check(MigrationState
*ms
, Error
**errp
)
3878 /* Assuming all off */
3879 bool old_caps
[MIGRATION_CAPABILITY__MAX
] = { 0 };
3881 if (!migrate_params_check(&ms
->parameters
, errp
)) {
3885 return migrate_caps_check(old_caps
, ms
->capabilities
, errp
);
3888 static const TypeInfo migration_type
= {
3889 .name
= TYPE_MIGRATION
,
3891 * NOTE: TYPE_MIGRATION is not really a device, as the object is
3892 * not created using qdev_new(), it is not attached to the qdev
3893 * device tree, and it is never realized.
3895 * TODO: Make this TYPE_OBJECT once QOM provides something like
3896 * TYPE_DEVICE's "-global" properties.
3898 .parent
= TYPE_DEVICE
,
3899 .class_init
= migration_class_init
,
3900 .class_size
= sizeof(MigrationClass
),
3901 .instance_size
= sizeof(MigrationState
),
3902 .instance_init
= migration_instance_init
,
3903 .instance_finalize
= migration_instance_finalize
,
3906 static void register_migration_types(void)
3908 type_register_static(&migration_type
);
3911 type_init(register_migration_types
);