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 "migration/global_state.h"
31 #include "migration/misc.h"
32 #include "migration.h"
33 #include "migration-stats.h"
35 #include "qemu-file.h"
37 #include "migration/vmstate.h"
38 #include "block/block.h"
39 #include "qapi/error.h"
40 #include "qapi/clone-visitor.h"
41 #include "qapi/qapi-visit-migration.h"
42 #include "qapi/qapi-visit-sockets.h"
43 #include "qapi/qapi-commands-migration.h"
44 #include "qapi/qapi-events-migration.h"
45 #include "qapi/qmp/qerror.h"
46 #include "qapi/qmp/qnull.h"
48 #include "postcopy-ram.h"
49 #include "qemu/thread.h"
51 #include "exec/target_page.h"
52 #include "io/channel-buffer.h"
53 #include "io/channel-tls.h"
54 #include "migration/colo.h"
55 #include "hw/boards.h"
56 #include "monitor/monitor.h"
57 #include "net/announce.h"
58 #include "qemu/queue.h"
60 #include "threadinfo.h"
61 #include "qemu/yank.h"
62 #include "sysemu/cpus.h"
63 #include "yank_functions.h"
64 #include "sysemu/qtest.h"
66 #include "sysemu/dirtylimit.h"
67 #include "qemu/sockets.h"
68 #include "sysemu/kvm.h"
70 #define NOTIFIER_ELEM_INIT(array, elem) \
71 [elem] = NOTIFIER_WITH_RETURN_LIST_INITIALIZER((array)[elem])
73 #define INMIGRATE_DEFAULT_EXIT_ON_ERROR true
75 static NotifierWithReturnList migration_state_notifiers
[] = {
76 NOTIFIER_ELEM_INIT(migration_state_notifiers
, MIG_MODE_NORMAL
),
77 NOTIFIER_ELEM_INIT(migration_state_notifiers
, MIG_MODE_CPR_REBOOT
),
80 /* Messages sent on the return path from destination to source */
81 enum mig_rp_message_type
{
82 MIG_RP_MSG_INVALID
= 0, /* Must be 0 */
83 MIG_RP_MSG_SHUT
, /* sibling will not send any more RP messages */
84 MIG_RP_MSG_PONG
, /* Response to a PING; data (seq: be32 ) */
86 MIG_RP_MSG_REQ_PAGES_ID
, /* data (start: be64, len: be32, id: string) */
87 MIG_RP_MSG_REQ_PAGES
, /* data (start: be64, len: be32) */
88 MIG_RP_MSG_RECV_BITMAP
, /* send recved_bitmap back to source */
89 MIG_RP_MSG_RESUME_ACK
, /* tell source that we are ready to resume */
90 MIG_RP_MSG_SWITCHOVER_ACK
, /* Tell source it's OK to do switchover */
95 /* When we add fault tolerance, we could have several
96 migrations at once. For now we don't need to add
97 dynamic creation of migration */
99 static MigrationState
*current_migration
;
100 static MigrationIncomingState
*current_incoming
;
102 static GSList
*migration_blockers
[MIG_MODE__MAX
];
104 static bool migration_object_check(MigrationState
*ms
, Error
**errp
);
105 static int migration_maybe_pause(MigrationState
*s
,
106 int *current_active_state
,
108 static void migrate_fd_cancel(MigrationState
*s
);
109 static bool close_return_path_on_source(MigrationState
*s
);
110 static void migration_completion_end(MigrationState
*s
);
112 static void migration_downtime_start(MigrationState
*s
)
114 trace_vmstate_downtime_checkpoint("src-downtime-start");
115 s
->downtime_start
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
118 static void migration_downtime_end(MigrationState
*s
)
120 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
123 * If downtime already set, should mean that postcopy already set it,
124 * then that should be the real downtime already.
127 s
->downtime
= now
- s
->downtime_start
;
130 trace_vmstate_downtime_checkpoint("src-downtime-end");
133 static bool migration_needs_multiple_sockets(void)
135 return migrate_multifd() || migrate_postcopy_preempt();
138 static bool transport_supports_multi_channels(MigrationAddress
*addr
)
140 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
141 SocketAddress
*saddr
= &addr
->u
.socket
;
143 return (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
144 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
145 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
);
146 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
147 return migrate_mapped_ram();
153 static bool migration_needs_seekable_channel(void)
155 return migrate_mapped_ram();
158 static bool migration_needs_extra_fds(void)
161 * When doing direct-io, multifd requires two different,
162 * non-duplicated file descriptors so we can use one of them for
165 return migrate_multifd() && migrate_direct_io();
168 static bool transport_supports_seeking(MigrationAddress
*addr
)
170 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
177 static bool transport_supports_extra_fds(MigrationAddress
*addr
)
179 /* file: works because QEMU can open it multiple times */
180 return addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
;
184 migration_channels_and_transport_compatible(MigrationAddress
*addr
,
187 if (migration_needs_seekable_channel() &&
188 !transport_supports_seeking(addr
)) {
189 error_setg(errp
, "Migration requires seekable transport (e.g. file)");
193 if (migration_needs_multiple_sockets() &&
194 !transport_supports_multi_channels(addr
)) {
195 error_setg(errp
, "Migration requires multi-channel URIs (e.g. tcp)");
199 if (migration_needs_extra_fds() &&
200 !transport_supports_extra_fds(addr
)) {
202 "Migration requires a transport that allows for extra fds (e.g. file)");
209 static gint
page_request_addr_cmp(gconstpointer ap
, gconstpointer bp
)
211 uintptr_t a
= (uintptr_t) ap
, b
= (uintptr_t) bp
;
213 return (a
> b
) - (a
< b
);
216 static int migration_stop_vm(MigrationState
*s
, RunState state
)
220 migration_downtime_start(s
);
222 s
->vm_old_state
= runstate_get();
223 global_state_store();
225 ret
= vm_stop_force_state(state
);
227 trace_vmstate_downtime_checkpoint("src-vm-stopped");
228 trace_migration_completion_vm_stop(ret
);
233 void migration_object_init(void)
235 /* This can only be called once. */
236 assert(!current_migration
);
237 current_migration
= MIGRATION_OBJ(object_new(TYPE_MIGRATION
));
240 * Init the migrate incoming object as well no matter whether
241 * we'll use it or not.
243 assert(!current_incoming
);
244 current_incoming
= g_new0(MigrationIncomingState
, 1);
245 current_incoming
->state
= MIGRATION_STATUS_NONE
;
246 current_incoming
->postcopy_remote_fds
=
247 g_array_new(FALSE
, TRUE
, sizeof(struct PostCopyFD
));
248 qemu_mutex_init(¤t_incoming
->rp_mutex
);
249 qemu_mutex_init(¤t_incoming
->postcopy_prio_thread_mutex
);
250 qemu_event_init(¤t_incoming
->main_thread_load_event
, false);
251 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_dst
, 0);
252 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fault
, 0);
253 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fast_load
, 0);
254 qemu_sem_init(¤t_incoming
->postcopy_qemufile_dst_done
, 0);
256 qemu_mutex_init(¤t_incoming
->page_request_mutex
);
257 qemu_cond_init(¤t_incoming
->page_request_cond
);
258 current_incoming
->page_requested
= g_tree_new(page_request_addr_cmp
);
260 current_incoming
->exit_on_error
= INMIGRATE_DEFAULT_EXIT_ON_ERROR
;
262 migration_object_check(current_migration
, &error_fatal
);
265 dirty_bitmap_mig_init();
274 static void migration_bh_dispatch_bh(void *opaque
)
276 MigrationState
*s
= migrate_get_current();
277 MigrationBH
*migbh
= opaque
;
279 /* cleanup this BH */
280 qemu_bh_delete(migbh
->bh
);
283 /* dispatch the other one */
284 migbh
->cb(migbh
->opaque
);
285 object_unref(OBJECT(s
));
290 void migration_bh_schedule(QEMUBHFunc
*cb
, void *opaque
)
292 MigrationState
*s
= migrate_get_current();
293 MigrationBH
*migbh
= g_new0(MigrationBH
, 1);
294 QEMUBH
*bh
= qemu_bh_new(migration_bh_dispatch_bh
, migbh
);
296 /* Store these to dispatch when the BH runs */
299 migbh
->opaque
= opaque
;
302 * Ref the state for bh, because it may be called when
303 * there're already no other refs
305 object_ref(OBJECT(s
));
306 qemu_bh_schedule(bh
);
309 void migration_cancel(const Error
*error
)
312 migrate_set_error(current_migration
, error
);
314 if (migrate_dirty_limit()) {
315 qmp_cancel_vcpu_dirty_limit(false, -1, NULL
);
317 migrate_fd_cancel(current_migration
);
320 void migration_shutdown(void)
323 * When the QEMU main thread exit, the COLO thread
324 * may wait a semaphore. So, we should wakeup the
325 * COLO thread before migration shutdown.
329 * Cancel the current migration - that will (eventually)
330 * stop the migration using this structure
332 migration_cancel(NULL
);
333 object_unref(OBJECT(current_migration
));
336 * Cancel outgoing migration of dirty bitmaps. It should
337 * at least unref used block nodes.
339 dirty_bitmap_mig_cancel_outgoing();
342 * Cancel incoming migration of dirty bitmaps. Dirty bitmaps
343 * are non-critical data, and their loss never considered as
346 dirty_bitmap_mig_cancel_incoming();
350 MigrationState
*migrate_get_current(void)
352 /* This can only be called after the object created. */
353 assert(current_migration
);
354 return current_migration
;
357 MigrationIncomingState
*migration_incoming_get_current(void)
359 assert(current_incoming
);
360 return current_incoming
;
363 void migration_incoming_transport_cleanup(MigrationIncomingState
*mis
)
365 if (mis
->socket_address_list
) {
366 qapi_free_SocketAddressList(mis
->socket_address_list
);
367 mis
->socket_address_list
= NULL
;
370 if (mis
->transport_cleanup
) {
371 mis
->transport_cleanup(mis
->transport_data
);
372 mis
->transport_data
= mis
->transport_cleanup
= NULL
;
376 void migration_incoming_state_destroy(void)
378 struct MigrationIncomingState
*mis
= migration_incoming_get_current();
380 multifd_recv_cleanup();
382 if (mis
->to_src_file
) {
383 /* Tell source that we are done */
384 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
385 qemu_fclose(mis
->to_src_file
);
386 mis
->to_src_file
= NULL
;
389 if (mis
->from_src_file
) {
390 migration_ioc_unregister_yank_from_file(mis
->from_src_file
);
391 qemu_fclose(mis
->from_src_file
);
392 mis
->from_src_file
= NULL
;
394 if (mis
->postcopy_remote_fds
) {
395 g_array_free(mis
->postcopy_remote_fds
, TRUE
);
396 mis
->postcopy_remote_fds
= NULL
;
399 migration_incoming_transport_cleanup(mis
);
400 qemu_event_reset(&mis
->main_thread_load_event
);
402 if (mis
->page_requested
) {
403 g_tree_destroy(mis
->page_requested
);
404 mis
->page_requested
= NULL
;
407 if (mis
->postcopy_qemufile_dst
) {
408 migration_ioc_unregister_yank_from_file(mis
->postcopy_qemufile_dst
);
409 qemu_fclose(mis
->postcopy_qemufile_dst
);
410 mis
->postcopy_qemufile_dst
= NULL
;
413 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
416 static void migrate_generate_event(MigrationStatus new_state
)
418 if (migrate_events()) {
419 qapi_event_send_migration(new_state
);
424 * Send a message on the return channel back to the source
427 static int migrate_send_rp_message(MigrationIncomingState
*mis
,
428 enum mig_rp_message_type message_type
,
429 uint16_t len
, void *data
)
433 trace_migrate_send_rp_message((int)message_type
, len
);
434 QEMU_LOCK_GUARD(&mis
->rp_mutex
);
437 * It's possible that the file handle got lost due to network
440 if (!mis
->to_src_file
) {
445 qemu_put_be16(mis
->to_src_file
, (unsigned int)message_type
);
446 qemu_put_be16(mis
->to_src_file
, len
);
447 qemu_put_buffer(mis
->to_src_file
, data
, len
);
448 return qemu_fflush(mis
->to_src_file
);
451 /* Request one page from the source VM at the given start address.
452 * rb: the RAMBlock to request the page in
453 * Start: Address offset within the RB
454 * Len: Length in bytes required - must be a multiple of pagesize
456 int migrate_send_rp_message_req_pages(MigrationIncomingState
*mis
,
457 RAMBlock
*rb
, ram_addr_t start
)
459 uint8_t bufc
[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
460 size_t msglen
= 12; /* start + len */
461 size_t len
= qemu_ram_pagesize(rb
);
462 enum mig_rp_message_type msg_type
;
466 *(uint64_t *)bufc
= cpu_to_be64((uint64_t)start
);
467 *(uint32_t *)(bufc
+ 8) = cpu_to_be32((uint32_t)len
);
470 * We maintain the last ramblock that we requested for page. Note that we
471 * don't need locking because this function will only be called within the
472 * postcopy ram fault thread.
474 if (rb
!= mis
->last_rb
) {
477 rbname
= qemu_ram_get_idstr(rb
);
478 rbname_len
= strlen(rbname
);
480 assert(rbname_len
< 256);
482 bufc
[msglen
++] = rbname_len
;
483 memcpy(bufc
+ msglen
, rbname
, rbname_len
);
484 msglen
+= rbname_len
;
485 msg_type
= MIG_RP_MSG_REQ_PAGES_ID
;
487 msg_type
= MIG_RP_MSG_REQ_PAGES
;
490 return migrate_send_rp_message(mis
, msg_type
, msglen
, bufc
);
493 int migrate_send_rp_req_pages(MigrationIncomingState
*mis
,
494 RAMBlock
*rb
, ram_addr_t start
, uint64_t haddr
)
496 void *aligned
= (void *)(uintptr_t)ROUND_DOWN(haddr
, qemu_ram_pagesize(rb
));
497 bool received
= false;
499 WITH_QEMU_LOCK_GUARD(&mis
->page_request_mutex
) {
500 received
= ramblock_recv_bitmap_test_byte_offset(rb
, start
);
501 if (!received
&& !g_tree_lookup(mis
->page_requested
, aligned
)) {
503 * The page has not been received, and it's not yet in the page
504 * request list. Queue it. Set the value of element to 1, so that
505 * things like g_tree_lookup() will return TRUE (1) when found.
507 g_tree_insert(mis
->page_requested
, aligned
, (gpointer
)1);
508 qatomic_inc(&mis
->page_requested_count
);
509 trace_postcopy_page_req_add(aligned
, mis
->page_requested_count
);
514 * If the page is there, skip sending the message. We don't even need the
515 * lock because as long as the page arrived, it'll be there forever.
521 return migrate_send_rp_message_req_pages(mis
, rb
, start
);
524 static bool migration_colo_enabled
;
525 bool migration_incoming_colo_enabled(void)
527 return migration_colo_enabled
;
530 void migration_incoming_disable_colo(void)
532 ram_block_discard_disable(false);
533 migration_colo_enabled
= false;
536 int migration_incoming_enable_colo(void)
538 #ifndef CONFIG_REPLICATION
539 error_report("ENABLE_COLO command come in migration stream, but the "
540 "replication module is not built in");
544 if (!migrate_colo()) {
545 error_report("ENABLE_COLO command come in migration stream, but x-colo "
546 "capability is not set");
550 if (ram_block_discard_disable(true)) {
551 error_report("COLO: cannot disable RAM discard");
554 migration_colo_enabled
= true;
558 void migrate_add_address(SocketAddress
*address
)
560 MigrationIncomingState
*mis
= migration_incoming_get_current();
562 QAPI_LIST_PREPEND(mis
->socket_address_list
,
563 QAPI_CLONE(SocketAddress
, address
));
566 bool migrate_uri_parse(const char *uri
, MigrationChannel
**channel
,
569 g_autoptr(MigrationChannel
) val
= g_new0(MigrationChannel
, 1);
570 g_autoptr(MigrationAddress
) addr
= g_new0(MigrationAddress
, 1);
571 InetSocketAddress
*isock
= &addr
->u
.rdma
;
572 strList
**tail
= &addr
->u
.exec
.args
;
574 if (strstart(uri
, "exec:", NULL
)) {
575 addr
->transport
= MIGRATION_ADDRESS_TYPE_EXEC
;
577 QAPI_LIST_APPEND(tail
, g_strdup(exec_get_cmd_path()));
578 QAPI_LIST_APPEND(tail
, g_strdup("/c"));
580 QAPI_LIST_APPEND(tail
, g_strdup("/bin/sh"));
581 QAPI_LIST_APPEND(tail
, g_strdup("-c"));
583 QAPI_LIST_APPEND(tail
, g_strdup(uri
+ strlen("exec:")));
584 } else if (strstart(uri
, "rdma:", NULL
)) {
585 if (inet_parse(isock
, uri
+ strlen("rdma:"), errp
)) {
586 qapi_free_InetSocketAddress(isock
);
589 addr
->transport
= MIGRATION_ADDRESS_TYPE_RDMA
;
590 } else if (strstart(uri
, "tcp:", NULL
) ||
591 strstart(uri
, "unix:", NULL
) ||
592 strstart(uri
, "vsock:", NULL
) ||
593 strstart(uri
, "fd:", NULL
)) {
594 addr
->transport
= MIGRATION_ADDRESS_TYPE_SOCKET
;
595 SocketAddress
*saddr
= socket_parse(uri
, errp
);
599 addr
->u
.socket
.type
= saddr
->type
;
600 addr
->u
.socket
.u
= saddr
->u
;
601 /* Don't free the objects inside; their ownership moved to "addr" */
603 } else if (strstart(uri
, "file:", NULL
)) {
604 addr
->transport
= MIGRATION_ADDRESS_TYPE_FILE
;
605 addr
->u
.file
.filename
= g_strdup(uri
+ strlen("file:"));
606 if (file_parse_offset(addr
->u
.file
.filename
, &addr
->u
.file
.offset
,
611 error_setg(errp
, "unknown migration protocol: %s", uri
);
615 val
->channel_type
= MIGRATION_CHANNEL_TYPE_MAIN
;
616 val
->addr
= g_steal_pointer(&addr
);
617 *channel
= g_steal_pointer(&val
);
622 migration_incoming_state_setup(MigrationIncomingState
*mis
, Error
**errp
)
624 MigrationStatus current
= mis
->state
;
626 if (current
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
628 * Incoming postcopy migration will stay in PAUSED state even if
629 * reconnection happened.
634 if (current
!= MIGRATION_STATUS_NONE
) {
635 error_setg(errp
, "Illegal migration incoming state: %s",
636 MigrationStatus_str(current
));
640 migrate_set_state(&mis
->state
, current
, MIGRATION_STATUS_SETUP
);
644 static void qemu_start_incoming_migration(const char *uri
, bool has_channels
,
645 MigrationChannelList
*channels
,
648 g_autoptr(MigrationChannel
) channel
= NULL
;
649 MigrationAddress
*addr
= NULL
;
650 MigrationIncomingState
*mis
= migration_incoming_get_current();
653 * Having preliminary checks for uri and channel
655 if (!uri
== !channels
) {
656 error_setg(errp
, "need either 'uri' or 'channels' argument");
661 /* To verify that Migrate channel list has only item */
662 if (channels
->next
) {
663 error_setg(errp
, "Channel list has more than one entries");
666 addr
= channels
->value
->addr
;
670 /* caller uses the old URI syntax */
671 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
674 addr
= channel
->addr
;
677 /* transport mechanism not suitable for migration? */
678 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
682 if (!migration_incoming_state_setup(mis
, errp
)) {
686 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
687 SocketAddress
*saddr
= &addr
->u
.socket
;
688 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
689 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
690 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
691 socket_start_incoming_migration(saddr
, errp
);
692 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
693 fd_start_incoming_migration(saddr
->u
.fd
.str
, errp
);
696 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
697 if (migrate_xbzrle()) {
698 error_setg(errp
, "RDMA and XBZRLE can't be used together");
701 if (migrate_multifd()) {
702 error_setg(errp
, "RDMA and multifd can't be used together");
705 rdma_start_incoming_migration(&addr
->u
.rdma
, errp
);
707 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
708 exec_start_incoming_migration(addr
->u
.exec
.args
, errp
);
709 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
710 file_start_incoming_migration(&addr
->u
.file
, errp
);
712 error_setg(errp
, "unknown migration protocol: %s", uri
);
716 static void process_incoming_migration_bh(void *opaque
)
718 Error
*local_err
= NULL
;
719 MigrationIncomingState
*mis
= opaque
;
721 trace_vmstate_downtime_checkpoint("dst-precopy-bh-enter");
723 /* If capability late_block_activate is set:
724 * Only fire up the block code now if we're going to restart the
725 * VM, else 'cont' will do it.
726 * This causes file locking to happen; so we don't want it to happen
727 * unless we really are starting the VM.
729 if (!migrate_late_block_activate() ||
730 (autostart
&& (!global_state_received() ||
731 runstate_is_live(global_state_get_runstate())))) {
732 /* Make sure all file formats throw away their mutable metadata.
733 * If we get an error here, just don't restart the VM yet. */
734 bdrv_activate_all(&local_err
);
736 error_report_err(local_err
);
743 * This must happen after all error conditions are dealt with and
744 * we're sure the VM is going to be running on this host.
746 qemu_announce_self(&mis
->announce_timer
, migrate_announce_params());
748 trace_vmstate_downtime_checkpoint("dst-precopy-bh-announced");
750 multifd_recv_shutdown();
752 dirty_bitmap_mig_before_vm_start();
754 if (!global_state_received() ||
755 runstate_is_live(global_state_get_runstate())) {
759 runstate_set(RUN_STATE_PAUSED
);
761 } else if (migration_incoming_colo_enabled()) {
762 migration_incoming_disable_colo();
765 runstate_set(global_state_get_runstate());
767 trace_vmstate_downtime_checkpoint("dst-precopy-bh-vm-started");
769 * This must happen after any state changes since as soon as an external
770 * observer sees this event they might start to prod at the VM assuming
773 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
774 MIGRATION_STATUS_COMPLETED
);
775 migration_incoming_state_destroy();
778 static void coroutine_fn
779 process_incoming_migration_co(void *opaque
)
781 MigrationState
*s
= migrate_get_current();
782 MigrationIncomingState
*mis
= migration_incoming_get_current();
785 Error
*local_err
= NULL
;
787 assert(mis
->from_src_file
);
789 mis
->largest_page_size
= qemu_ram_pagesize_largest();
790 postcopy_state_set(POSTCOPY_INCOMING_NONE
);
791 migrate_set_state(&mis
->state
, MIGRATION_STATUS_SETUP
,
792 MIGRATION_STATUS_ACTIVE
);
794 mis
->loadvm_co
= qemu_coroutine_self();
795 ret
= qemu_loadvm_state(mis
->from_src_file
);
796 mis
->loadvm_co
= NULL
;
798 trace_vmstate_downtime_checkpoint("dst-precopy-loadvm-completed");
800 ps
= postcopy_state_get();
801 trace_process_incoming_migration_co_end(ret
, ps
);
802 if (ps
!= POSTCOPY_INCOMING_NONE
) {
803 if (ps
== POSTCOPY_INCOMING_ADVISE
) {
805 * Where a migration had postcopy enabled (and thus went to advise)
806 * but managed to complete within the precopy period, we can use
809 postcopy_ram_incoming_cleanup(mis
);
810 } else if (ret
>= 0) {
812 * Postcopy was started, cleanup should happen at the end of the
815 trace_process_incoming_migration_co_postcopy_end_main();
818 /* Else if something went wrong then just fall out of the normal exit */
822 error_setg(&local_err
, "load of migration failed: %s", strerror(-ret
));
826 if (migration_incoming_colo_enabled()) {
827 /* yield until COLO exit */
831 migration_bh_schedule(process_incoming_migration_bh
, mis
);
834 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
835 MIGRATION_STATUS_FAILED
);
836 migrate_set_error(s
, local_err
);
837 error_free(local_err
);
839 migration_incoming_state_destroy();
841 if (mis
->exit_on_error
) {
842 WITH_QEMU_LOCK_GUARD(&s
->error_mutex
) {
843 error_report_err(s
->error
);
852 * migration_incoming_setup: Setup incoming migration
853 * @f: file for main migration channel
855 static void migration_incoming_setup(QEMUFile
*f
)
857 MigrationIncomingState
*mis
= migration_incoming_get_current();
859 if (!mis
->from_src_file
) {
860 mis
->from_src_file
= f
;
862 qemu_file_set_blocking(f
, false);
865 void migration_incoming_process(void)
867 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
, NULL
);
868 qemu_coroutine_enter(co
);
871 /* Returns true if recovered from a paused migration, otherwise false */
872 static bool postcopy_try_recover(void)
874 MigrationIncomingState
*mis
= migration_incoming_get_current();
876 if (mis
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
877 /* Resumed from a paused postcopy migration */
879 /* This should be set already in migration_incoming_setup() */
880 assert(mis
->from_src_file
);
881 /* Postcopy has standalone thread to do vm load */
882 qemu_file_set_blocking(mis
->from_src_file
, true);
884 /* Re-configure the return path */
885 mis
->to_src_file
= qemu_file_get_return_path(mis
->from_src_file
);
887 migrate_set_state(&mis
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
888 MIGRATION_STATUS_POSTCOPY_RECOVER
);
891 * Here, we only wake up the main loading thread (while the
892 * rest threads will still be waiting), so that we can receive
893 * commands from source now, and answer it if needed. The
894 * rest threads will be woken up afterwards until we are sure
895 * that source is ready to reply to page requests.
897 qemu_sem_post(&mis
->postcopy_pause_sem_dst
);
904 void migration_fd_process_incoming(QEMUFile
*f
)
906 migration_incoming_setup(f
);
907 if (postcopy_try_recover()) {
910 migration_incoming_process();
914 * Returns true when we want to start a new incoming migration process,
917 static bool migration_should_start_incoming(bool main_channel
)
919 /* Multifd doesn't start unless all channels are established */
920 if (migrate_multifd()) {
921 return migration_has_all_channels();
924 /* Preempt channel only starts when the main channel is created */
925 if (migrate_postcopy_preempt()) {
930 * For all the rest types of migration, we should only reach here when
931 * it's the main channel that's being created, and we should always
932 * proceed with this channel.
934 assert(main_channel
);
938 void migration_ioc_process_incoming(QIOChannel
*ioc
, Error
**errp
)
940 MigrationIncomingState
*mis
= migration_incoming_get_current();
941 Error
*local_err
= NULL
;
943 bool default_channel
= true;
944 uint32_t channel_magic
= 0;
947 if (migrate_multifd() && !migrate_mapped_ram() &&
948 !migrate_postcopy_ram() &&
949 qio_channel_has_feature(ioc
, QIO_CHANNEL_FEATURE_READ_MSG_PEEK
)) {
951 * With multiple channels, it is possible that we receive channels
952 * out of order on destination side, causing incorrect mapping of
953 * source channels on destination side. Check channel MAGIC to
954 * decide type of channel. Please note this is best effort, postcopy
955 * preempt channel does not send any magic number so avoid it for
956 * postcopy live migration. Also tls live migration already does
957 * tls handshake while initializing main channel so with tls this
958 * issue is not possible.
960 ret
= migration_channel_read_peek(ioc
, (void *)&channel_magic
,
961 sizeof(channel_magic
), errp
);
967 default_channel
= (channel_magic
== cpu_to_be32(QEMU_VM_FILE_MAGIC
));
969 default_channel
= !mis
->from_src_file
;
972 if (multifd_recv_setup(errp
) != 0) {
976 if (default_channel
) {
977 f
= qemu_file_new_input(ioc
);
978 migration_incoming_setup(f
);
980 /* Multiple connections */
981 assert(migration_needs_multiple_sockets());
982 if (migrate_multifd()) {
983 multifd_recv_new_channel(ioc
, &local_err
);
985 assert(migrate_postcopy_preempt());
986 f
= qemu_file_new_input(ioc
);
987 postcopy_preempt_new_channel(mis
, f
);
990 error_propagate(errp
, local_err
);
995 if (migration_should_start_incoming(default_channel
)) {
996 /* If it's a recovery, we're done */
997 if (postcopy_try_recover()) {
1000 migration_incoming_process();
1005 * @migration_has_all_channels: We have received all channels that we need
1007 * Returns true when we have got connections to all the channels that
1008 * we need for migration.
1010 bool migration_has_all_channels(void)
1012 MigrationIncomingState
*mis
= migration_incoming_get_current();
1014 if (!mis
->from_src_file
) {
1018 if (migrate_multifd()) {
1019 return multifd_recv_all_channels_created();
1022 if (migrate_postcopy_preempt()) {
1023 return mis
->postcopy_qemufile_dst
!= NULL
;
1029 int migrate_send_rp_switchover_ack(MigrationIncomingState
*mis
)
1031 return migrate_send_rp_message(mis
, MIG_RP_MSG_SWITCHOVER_ACK
, 0, NULL
);
1035 * Send a 'SHUT' message on the return channel with the given value
1036 * to indicate that we've finished with the RP. Non-0 value indicates
1039 void migrate_send_rp_shut(MigrationIncomingState
*mis
,
1044 buf
= cpu_to_be32(value
);
1045 migrate_send_rp_message(mis
, MIG_RP_MSG_SHUT
, sizeof(buf
), &buf
);
1049 * Send a 'PONG' message on the return channel with the given value
1050 * (normally in response to a 'PING')
1052 void migrate_send_rp_pong(MigrationIncomingState
*mis
,
1057 buf
= cpu_to_be32(value
);
1058 migrate_send_rp_message(mis
, MIG_RP_MSG_PONG
, sizeof(buf
), &buf
);
1061 void migrate_send_rp_recv_bitmap(MigrationIncomingState
*mis
,
1069 * First, we send the header part. It contains only the len of
1070 * idstr, and the idstr itself.
1072 len
= strlen(block_name
);
1074 memcpy(buf
+ 1, block_name
, len
);
1076 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
1077 error_report("%s: MSG_RP_RECV_BITMAP only used for recovery",
1082 migrate_send_rp_message(mis
, MIG_RP_MSG_RECV_BITMAP
, len
+ 1, buf
);
1085 * Next, we dump the received bitmap to the stream.
1087 * TODO: currently we are safe since we are the only one that is
1088 * using the to_src_file handle (fault thread is still paused),
1089 * and it's ok even not taking the mutex. However the best way is
1090 * to take the lock before sending the message header, and release
1091 * the lock after sending the bitmap.
1093 qemu_mutex_lock(&mis
->rp_mutex
);
1094 res
= ramblock_recv_bitmap_send(mis
->to_src_file
, block_name
);
1095 qemu_mutex_unlock(&mis
->rp_mutex
);
1097 trace_migrate_send_rp_recv_bitmap(block_name
, res
);
1100 void migrate_send_rp_resume_ack(MigrationIncomingState
*mis
, uint32_t value
)
1104 buf
= cpu_to_be32(value
);
1105 migrate_send_rp_message(mis
, MIG_RP_MSG_RESUME_ACK
, sizeof(buf
), &buf
);
1109 * Return true if we're already in the middle of a migration
1110 * (i.e. any of the active or setup states)
1112 bool migration_is_setup_or_active(void)
1114 MigrationState
*s
= current_migration
;
1117 case MIGRATION_STATUS_ACTIVE
:
1118 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1119 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1120 case MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
:
1121 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1122 case MIGRATION_STATUS_SETUP
:
1123 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1124 case MIGRATION_STATUS_DEVICE
:
1125 case MIGRATION_STATUS_WAIT_UNPLUG
:
1126 case MIGRATION_STATUS_COLO
:
1135 bool migration_is_running(void)
1137 MigrationState
*s
= current_migration
;
1140 case MIGRATION_STATUS_ACTIVE
:
1141 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1142 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1143 case MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
:
1144 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1145 case MIGRATION_STATUS_SETUP
:
1146 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1147 case MIGRATION_STATUS_DEVICE
:
1148 case MIGRATION_STATUS_WAIT_UNPLUG
:
1149 case MIGRATION_STATUS_CANCELLING
:
1158 static bool migrate_show_downtime(MigrationState
*s
)
1160 return (s
->state
== MIGRATION_STATUS_COMPLETED
) || migration_in_postcopy();
1163 static void populate_time_info(MigrationInfo
*info
, MigrationState
*s
)
1165 info
->has_status
= true;
1166 info
->has_setup_time
= true;
1167 info
->setup_time
= s
->setup_time
;
1169 if (s
->state
== MIGRATION_STATUS_COMPLETED
) {
1170 info
->has_total_time
= true;
1171 info
->total_time
= s
->total_time
;
1173 info
->has_total_time
= true;
1174 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) -
1178 if (migrate_show_downtime(s
)) {
1179 info
->has_downtime
= true;
1180 info
->downtime
= s
->downtime
;
1182 info
->has_expected_downtime
= true;
1183 info
->expected_downtime
= s
->expected_downtime
;
1187 static void populate_ram_info(MigrationInfo
*info
, MigrationState
*s
)
1189 size_t page_size
= qemu_target_page_size();
1191 info
->ram
= g_malloc0(sizeof(*info
->ram
));
1192 info
->ram
->transferred
= migration_transferred_bytes();
1193 info
->ram
->total
= ram_bytes_total();
1194 info
->ram
->duplicate
= stat64_get(&mig_stats
.zero_pages
);
1195 info
->ram
->normal
= stat64_get(&mig_stats
.normal_pages
);
1196 info
->ram
->normal_bytes
= info
->ram
->normal
* page_size
;
1197 info
->ram
->mbps
= s
->mbps
;
1198 info
->ram
->dirty_sync_count
=
1199 stat64_get(&mig_stats
.dirty_sync_count
);
1200 info
->ram
->dirty_sync_missed_zero_copy
=
1201 stat64_get(&mig_stats
.dirty_sync_missed_zero_copy
);
1202 info
->ram
->postcopy_requests
=
1203 stat64_get(&mig_stats
.postcopy_requests
);
1204 info
->ram
->page_size
= page_size
;
1205 info
->ram
->multifd_bytes
= stat64_get(&mig_stats
.multifd_bytes
);
1206 info
->ram
->pages_per_second
= s
->pages_per_second
;
1207 info
->ram
->precopy_bytes
= stat64_get(&mig_stats
.precopy_bytes
);
1208 info
->ram
->downtime_bytes
= stat64_get(&mig_stats
.downtime_bytes
);
1209 info
->ram
->postcopy_bytes
= stat64_get(&mig_stats
.postcopy_bytes
);
1211 if (migrate_xbzrle()) {
1212 info
->xbzrle_cache
= g_malloc0(sizeof(*info
->xbzrle_cache
));
1213 info
->xbzrle_cache
->cache_size
= migrate_xbzrle_cache_size();
1214 info
->xbzrle_cache
->bytes
= xbzrle_counters
.bytes
;
1215 info
->xbzrle_cache
->pages
= xbzrle_counters
.pages
;
1216 info
->xbzrle_cache
->cache_miss
= xbzrle_counters
.cache_miss
;
1217 info
->xbzrle_cache
->cache_miss_rate
= xbzrle_counters
.cache_miss_rate
;
1218 info
->xbzrle_cache
->encoding_rate
= xbzrle_counters
.encoding_rate
;
1219 info
->xbzrle_cache
->overflow
= xbzrle_counters
.overflow
;
1222 if (cpu_throttle_active()) {
1223 info
->has_cpu_throttle_percentage
= true;
1224 info
->cpu_throttle_percentage
= cpu_throttle_get_percentage();
1227 if (s
->state
!= MIGRATION_STATUS_COMPLETED
) {
1228 info
->ram
->remaining
= ram_bytes_remaining();
1229 info
->ram
->dirty_pages_rate
=
1230 stat64_get(&mig_stats
.dirty_pages_rate
);
1233 if (migrate_dirty_limit() && dirtylimit_in_service()) {
1234 info
->has_dirty_limit_throttle_time_per_round
= true;
1235 info
->dirty_limit_throttle_time_per_round
=
1236 dirtylimit_throttle_time_per_round();
1238 info
->has_dirty_limit_ring_full_time
= true;
1239 info
->dirty_limit_ring_full_time
= dirtylimit_ring_full_time();
1243 static void fill_source_migration_info(MigrationInfo
*info
)
1245 MigrationState
*s
= migrate_get_current();
1246 int state
= qatomic_read(&s
->state
);
1247 GSList
*cur_blocker
= migration_blockers
[migrate_mode()];
1249 info
->blocked_reasons
= NULL
;
1252 * There are two types of reasons a migration might be blocked;
1253 * a) devices marked in VMState as non-migratable, and
1254 * b) Explicit migration blockers
1255 * We need to add both of them here.
1257 qemu_savevm_non_migratable_list(&info
->blocked_reasons
);
1259 while (cur_blocker
) {
1260 QAPI_LIST_PREPEND(info
->blocked_reasons
,
1261 g_strdup(error_get_pretty(cur_blocker
->data
)));
1262 cur_blocker
= g_slist_next(cur_blocker
);
1264 info
->has_blocked_reasons
= info
->blocked_reasons
!= NULL
;
1267 case MIGRATION_STATUS_NONE
:
1268 /* no migration has happened ever */
1269 /* do not overwrite destination migration status */
1271 case MIGRATION_STATUS_SETUP
:
1272 info
->has_status
= true;
1273 info
->has_total_time
= false;
1275 case MIGRATION_STATUS_ACTIVE
:
1276 case MIGRATION_STATUS_CANCELLING
:
1277 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1278 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1279 case MIGRATION_STATUS_DEVICE
:
1280 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1281 case MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
:
1282 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1283 /* TODO add some postcopy stats */
1284 populate_time_info(info
, s
);
1285 populate_ram_info(info
, s
);
1286 migration_populate_vfio_info(info
);
1288 case MIGRATION_STATUS_COLO
:
1289 info
->has_status
= true;
1290 /* TODO: display COLO specific information (checkpoint info etc.) */
1292 case MIGRATION_STATUS_COMPLETED
:
1293 populate_time_info(info
, s
);
1294 populate_ram_info(info
, s
);
1295 migration_populate_vfio_info(info
);
1297 case MIGRATION_STATUS_FAILED
:
1298 info
->has_status
= true;
1300 case MIGRATION_STATUS_CANCELLED
:
1301 info
->has_status
= true;
1303 case MIGRATION_STATUS_WAIT_UNPLUG
:
1304 info
->has_status
= true;
1307 info
->status
= state
;
1309 QEMU_LOCK_GUARD(&s
->error_mutex
);
1311 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1315 static void fill_destination_migration_info(MigrationInfo
*info
)
1317 MigrationIncomingState
*mis
= migration_incoming_get_current();
1319 if (mis
->socket_address_list
) {
1320 info
->has_socket_address
= true;
1321 info
->socket_address
=
1322 QAPI_CLONE(SocketAddressList
, mis
->socket_address_list
);
1325 switch (mis
->state
) {
1326 case MIGRATION_STATUS_SETUP
:
1327 case MIGRATION_STATUS_CANCELLING
:
1328 case MIGRATION_STATUS_CANCELLED
:
1329 case MIGRATION_STATUS_ACTIVE
:
1330 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1331 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1332 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1333 case MIGRATION_STATUS_FAILED
:
1334 case MIGRATION_STATUS_COLO
:
1335 info
->has_status
= true;
1337 case MIGRATION_STATUS_COMPLETED
:
1338 info
->has_status
= true;
1339 fill_destination_postcopy_migration_info(info
);
1344 info
->status
= mis
->state
;
1346 if (!info
->error_desc
) {
1347 MigrationState
*s
= migrate_get_current();
1348 QEMU_LOCK_GUARD(&s
->error_mutex
);
1351 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1356 MigrationInfo
*qmp_query_migrate(Error
**errp
)
1358 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
1360 fill_destination_migration_info(info
);
1361 fill_source_migration_info(info
);
1366 void qmp_migrate_start_postcopy(Error
**errp
)
1368 MigrationState
*s
= migrate_get_current();
1370 if (!migrate_postcopy()) {
1371 error_setg(errp
, "Enable postcopy with migrate_set_capability before"
1372 " the start of migration");
1376 if (s
->state
== MIGRATION_STATUS_NONE
) {
1377 error_setg(errp
, "Postcopy must be started after migration has been"
1382 * we don't error if migration has finished since that would be racy
1383 * with issuing this command.
1385 qatomic_set(&s
->start_postcopy
, true);
1388 /* shared migration helpers */
1390 void migrate_set_state(MigrationStatus
*state
, MigrationStatus old_state
,
1391 MigrationStatus new_state
)
1393 assert(new_state
< MIGRATION_STATUS__MAX
);
1394 if (qatomic_cmpxchg(state
, old_state
, new_state
) == old_state
) {
1395 trace_migrate_set_state(MigrationStatus_str(new_state
));
1396 migrate_generate_event(new_state
);
1400 static void migrate_fd_cleanup(MigrationState
*s
)
1402 MigrationEventType type
;
1404 g_free(s
->hostname
);
1406 json_writer_free(s
->vmdesc
);
1409 qemu_savevm_state_cleanup();
1411 close_return_path_on_source(s
);
1413 if (s
->to_dst_file
) {
1416 trace_migrate_fd_cleanup();
1418 if (s
->migration_thread_running
) {
1419 qemu_thread_join(&s
->thread
);
1420 s
->migration_thread_running
= false;
1424 multifd_send_shutdown();
1425 qemu_mutex_lock(&s
->qemu_file_lock
);
1426 tmp
= s
->to_dst_file
;
1427 s
->to_dst_file
= NULL
;
1428 qemu_mutex_unlock(&s
->qemu_file_lock
);
1430 * Close the file handle without the lock to make sure the
1431 * critical section won't block for long.
1433 migration_ioc_unregister_yank_from_file(tmp
);
1437 assert(!migration_is_active());
1439 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1440 migrate_set_state(&s
->state
, MIGRATION_STATUS_CANCELLING
,
1441 MIGRATION_STATUS_CANCELLED
);
1445 /* It is used on info migrate. We can't free it */
1446 error_report_err(error_copy(s
->error
));
1448 type
= migration_has_failed(s
) ? MIG_EVENT_PRECOPY_FAILED
:
1449 MIG_EVENT_PRECOPY_DONE
;
1450 migration_call_notifiers(s
, type
, NULL
);
1451 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1454 static void migrate_fd_cleanup_bh(void *opaque
)
1456 migrate_fd_cleanup(opaque
);
1459 void migrate_set_error(MigrationState
*s
, const Error
*error
)
1461 QEMU_LOCK_GUARD(&s
->error_mutex
);
1463 trace_migrate_error(error_get_pretty(error
));
1466 s
->error
= error_copy(error
);
1470 bool migrate_has_error(MigrationState
*s
)
1472 /* The lock is not helpful here, but still follow the rule */
1473 QEMU_LOCK_GUARD(&s
->error_mutex
);
1474 return qatomic_read(&s
->error
);
1477 static void migrate_error_free(MigrationState
*s
)
1479 QEMU_LOCK_GUARD(&s
->error_mutex
);
1481 error_free(s
->error
);
1486 static void migrate_fd_error(MigrationState
*s
, const Error
*error
)
1488 MigrationStatus current
= s
->state
;
1489 MigrationStatus next
;
1491 assert(s
->to_dst_file
== NULL
);
1494 case MIGRATION_STATUS_SETUP
:
1495 next
= MIGRATION_STATUS_FAILED
;
1497 case MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
:
1498 /* Never fail a postcopy migration; switch back to PAUSED instead */
1499 next
= MIGRATION_STATUS_POSTCOPY_PAUSED
;
1503 * This really shouldn't happen. Just be careful to not crash a VM
1504 * just for this. Instead, dump something.
1506 error_report("%s: Illegal migration status (%s) detected",
1507 __func__
, MigrationStatus_str(current
));
1511 migrate_set_state(&s
->state
, current
, next
);
1512 migrate_set_error(s
, error
);
1515 static void migrate_fd_cancel(MigrationState
*s
)
1519 trace_migrate_fd_cancel();
1521 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1522 if (s
->rp_state
.from_dst_file
) {
1523 /* shutdown the rp socket, so causing the rp thread to shutdown */
1524 qemu_file_shutdown(s
->rp_state
.from_dst_file
);
1529 old_state
= s
->state
;
1530 if (!migration_is_running()) {
1533 /* If the migration is paused, kick it out of the pause */
1534 if (old_state
== MIGRATION_STATUS_PRE_SWITCHOVER
) {
1535 qemu_sem_post(&s
->pause_sem
);
1537 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_CANCELLING
);
1538 } while (s
->state
!= MIGRATION_STATUS_CANCELLING
);
1541 * If we're unlucky the migration code might be stuck somewhere in a
1542 * send/write while the network has failed and is waiting to timeout;
1543 * if we've got shutdown(2) available then we can force it to quit.
1545 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1546 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1547 if (s
->to_dst_file
) {
1548 qemu_file_shutdown(s
->to_dst_file
);
1552 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& s
->block_inactive
) {
1553 Error
*local_err
= NULL
;
1555 bdrv_activate_all(&local_err
);
1557 error_report_err(local_err
);
1559 s
->block_inactive
= false;
1564 void migration_add_notifier_mode(NotifierWithReturn
*notify
,
1565 MigrationNotifyFunc func
, MigMode mode
)
1567 notify
->notify
= (NotifierWithReturnFunc
)func
;
1568 notifier_with_return_list_add(&migration_state_notifiers
[mode
], notify
);
1571 void migration_add_notifier(NotifierWithReturn
*notify
,
1572 MigrationNotifyFunc func
)
1574 migration_add_notifier_mode(notify
, func
, MIG_MODE_NORMAL
);
1577 void migration_remove_notifier(NotifierWithReturn
*notify
)
1579 if (notify
->notify
) {
1580 notifier_with_return_remove(notify
);
1581 notify
->notify
= NULL
;
1585 int migration_call_notifiers(MigrationState
*s
, MigrationEventType type
,
1588 MigMode mode
= s
->parameters
.mode
;
1593 ret
= notifier_with_return_list_notify(&migration_state_notifiers
[mode
],
1595 assert(!ret
|| type
== MIG_EVENT_PRECOPY_SETUP
);
1599 bool migration_has_failed(MigrationState
*s
)
1601 return (s
->state
== MIGRATION_STATUS_CANCELLED
||
1602 s
->state
== MIGRATION_STATUS_FAILED
);
1605 bool migration_in_postcopy(void)
1607 MigrationState
*s
= migrate_get_current();
1610 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1611 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1612 case MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
:
1613 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1620 bool migration_postcopy_is_alive(MigrationStatus state
)
1623 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1624 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1631 bool migration_in_incoming_postcopy(void)
1633 PostcopyState ps
= postcopy_state_get();
1635 return ps
>= POSTCOPY_INCOMING_DISCARD
&& ps
< POSTCOPY_INCOMING_END
;
1638 bool migration_incoming_postcopy_advised(void)
1640 PostcopyState ps
= postcopy_state_get();
1642 return ps
>= POSTCOPY_INCOMING_ADVISE
&& ps
< POSTCOPY_INCOMING_END
;
1645 bool migration_in_bg_snapshot(void)
1647 return migrate_background_snapshot() &&
1648 migration_is_setup_or_active();
1651 bool migration_is_idle(void)
1653 MigrationState
*s
= current_migration
;
1660 case MIGRATION_STATUS_NONE
:
1661 case MIGRATION_STATUS_CANCELLED
:
1662 case MIGRATION_STATUS_COMPLETED
:
1663 case MIGRATION_STATUS_FAILED
:
1670 bool migration_is_active(void)
1672 MigrationState
*s
= current_migration
;
1674 return (s
->state
== MIGRATION_STATUS_ACTIVE
||
1675 s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1678 bool migration_is_device(void)
1680 MigrationState
*s
= current_migration
;
1682 return s
->state
== MIGRATION_STATUS_DEVICE
;
1685 bool migration_thread_is_self(void)
1687 MigrationState
*s
= current_migration
;
1689 return qemu_thread_is_self(&s
->thread
);
1692 bool migrate_mode_is_cpr(MigrationState
*s
)
1694 return s
->parameters
.mode
== MIG_MODE_CPR_REBOOT
;
1697 int migrate_init(MigrationState
*s
, Error
**errp
)
1701 ret
= qemu_savevm_state_prepare(errp
);
1707 * Reinitialise all migration state, except
1708 * parameters/capabilities that the user set, and
1711 s
->to_dst_file
= NULL
;
1712 s
->state
= MIGRATION_STATUS_NONE
;
1713 s
->rp_state
.from_dst_file
= NULL
;
1715 s
->pages_per_second
= 0.0;
1717 s
->expected_downtime
= 0;
1719 s
->start_postcopy
= false;
1720 s
->migration_thread_running
= false;
1721 error_free(s
->error
);
1725 migrate_set_state(&s
->state
, MIGRATION_STATUS_NONE
, MIGRATION_STATUS_SETUP
);
1727 s
->start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1729 s
->vm_old_state
= -1;
1730 s
->iteration_initial_bytes
= 0;
1731 s
->threshold_size
= 0;
1732 s
->switchover_acked
= false;
1733 s
->rdma_migration
= false;
1735 * set mig_stats memory to zero for a new migration
1737 memset(&mig_stats
, 0, sizeof(mig_stats
));
1738 migration_reset_vfio_bytes_transferred();
1743 static bool is_busy(Error
**reasonp
, Error
**errp
)
1747 /* Snapshots are similar to migrations, so check RUN_STATE_SAVE_VM too. */
1748 if (runstate_check(RUN_STATE_SAVE_VM
) || !migration_is_idle()) {
1749 error_propagate_prepend(errp
, *reasonp
,
1750 "disallowing migration blocker "
1751 "(migration/snapshot in progress) for: ");
1758 static bool is_only_migratable(Error
**reasonp
, Error
**errp
, int modes
)
1762 if (only_migratable
&& (modes
& BIT(MIG_MODE_NORMAL
))) {
1763 error_propagate_prepend(errp
, *reasonp
,
1764 "disallowing migration blocker "
1765 "(--only-migratable) for: ");
1772 static int get_modes(MigMode mode
, va_list ap
)
1776 while (mode
!= -1 && mode
!= MIG_MODE_ALL
) {
1777 assert(mode
>= MIG_MODE_NORMAL
&& mode
< MIG_MODE__MAX
);
1779 mode
= va_arg(ap
, MigMode
);
1781 if (mode
== MIG_MODE_ALL
) {
1782 modes
= BIT(MIG_MODE__MAX
) - 1;
1787 static int add_blockers(Error
**reasonp
, Error
**errp
, int modes
)
1789 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1790 if (modes
& BIT(mode
)) {
1791 migration_blockers
[mode
] = g_slist_prepend(migration_blockers
[mode
],
1798 int migrate_add_blocker(Error
**reasonp
, Error
**errp
)
1800 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_ALL
);
1803 int migrate_add_blocker_normal(Error
**reasonp
, Error
**errp
)
1805 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_NORMAL
, -1);
1808 int migrate_add_blocker_modes(Error
**reasonp
, Error
**errp
, MigMode mode
, ...)
1814 modes
= get_modes(mode
, ap
);
1817 if (is_only_migratable(reasonp
, errp
, modes
)) {
1819 } else if (is_busy(reasonp
, errp
)) {
1822 return add_blockers(reasonp
, errp
, modes
);
1825 int migrate_add_blocker_internal(Error
**reasonp
, Error
**errp
)
1827 int modes
= BIT(MIG_MODE__MAX
) - 1;
1829 if (is_busy(reasonp
, errp
)) {
1832 return add_blockers(reasonp
, errp
, modes
);
1835 void migrate_del_blocker(Error
**reasonp
)
1838 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1839 migration_blockers
[mode
] = g_slist_remove(migration_blockers
[mode
],
1842 error_free(*reasonp
);
1847 void qmp_migrate_incoming(const char *uri
, bool has_channels
,
1848 MigrationChannelList
*channels
,
1849 bool has_exit_on_error
, bool exit_on_error
,
1852 Error
*local_err
= NULL
;
1853 static bool once
= true;
1854 MigrationIncomingState
*mis
= migration_incoming_get_current();
1857 error_setg(errp
, "The incoming migration has already been started");
1860 if (!runstate_check(RUN_STATE_INMIGRATE
)) {
1861 error_setg(errp
, "'-incoming' was not specified on the command line");
1865 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1869 mis
->exit_on_error
=
1870 has_exit_on_error
? exit_on_error
: INMIGRATE_DEFAULT_EXIT_ON_ERROR
;
1872 qemu_start_incoming_migration(uri
, has_channels
, channels
, &local_err
);
1875 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1876 error_propagate(errp
, local_err
);
1883 void qmp_migrate_recover(const char *uri
, Error
**errp
)
1885 MigrationIncomingState
*mis
= migration_incoming_get_current();
1888 * Don't even bother to use ERRP_GUARD() as it _must_ always be set by
1889 * callers (no one should ignore a recover failure); if there is, it's a
1890 * programming error.
1894 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1895 error_setg(errp
, "Migrate recover can only be run "
1896 "when postcopy is paused.");
1900 /* If there's an existing transport, release it */
1901 migration_incoming_transport_cleanup(mis
);
1904 * Note that this call will never start a real migration; it will
1905 * only re-setup the migration stream and poke existing migration
1906 * to continue using that newly established channel.
1908 qemu_start_incoming_migration(uri
, false, NULL
, errp
);
1911 void qmp_migrate_pause(Error
**errp
)
1913 MigrationState
*ms
= migrate_get_current();
1914 MigrationIncomingState
*mis
= migration_incoming_get_current();
1917 if (migration_postcopy_is_alive(ms
->state
)) {
1918 /* Source side, during postcopy */
1919 Error
*error
= NULL
;
1921 /* Tell the core migration that we're pausing */
1922 error_setg(&error
, "Postcopy migration is paused by the user");
1923 migrate_set_error(ms
, error
);
1926 qemu_mutex_lock(&ms
->qemu_file_lock
);
1927 if (ms
->to_dst_file
) {
1928 ret
= qemu_file_shutdown(ms
->to_dst_file
);
1930 qemu_mutex_unlock(&ms
->qemu_file_lock
);
1932 error_setg(errp
, "Failed to pause source migration");
1936 * Kick the migration thread out of any waiting windows (on behalf
1937 * of the rp thread).
1939 migration_rp_kick(ms
);
1944 if (migration_postcopy_is_alive(mis
->state
)) {
1945 ret
= qemu_file_shutdown(mis
->from_src_file
);
1947 error_setg(errp
, "Failed to pause destination migration");
1952 error_setg(errp
, "migrate-pause is currently only supported "
1953 "during postcopy-active or postcopy-recover state");
1956 bool migration_is_blocked(Error
**errp
)
1958 GSList
*blockers
= migration_blockers
[migrate_mode()];
1960 if (qemu_savevm_state_blocked(errp
)) {
1965 error_propagate(errp
, error_copy(blockers
->data
));
1972 /* Returns true if continue to migrate, or false if error detected */
1973 static bool migrate_prepare(MigrationState
*s
, bool resume
, Error
**errp
)
1976 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1977 error_setg(errp
, "Cannot resume if there is no "
1978 "paused migration");
1983 * Postcopy recovery won't work well with release-ram
1984 * capability since release-ram will drop the page buffer as
1985 * long as the page is put into the send buffer. So if there
1986 * is a network failure happened, any page buffers that have
1987 * not yet reached the destination VM but have already been
1988 * sent from the source VM will be lost forever. Let's refuse
1989 * the client from resuming such a postcopy migration.
1990 * Luckily release-ram was designed to only be used when src
1991 * and destination VMs are on the same host, so it should be
1994 if (migrate_release_ram()) {
1995 error_setg(errp
, "Postcopy recovery cannot work "
1996 "when release-ram capability is set");
2000 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
2001 MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
);
2003 /* This is a resume, skip init status */
2007 if (migration_is_running()) {
2008 error_setg(errp
, "There's a migration process in progress");
2012 if (runstate_check(RUN_STATE_INMIGRATE
)) {
2013 error_setg(errp
, "Guest is waiting for an incoming migration");
2017 if (runstate_check(RUN_STATE_POSTMIGRATE
)) {
2018 error_setg(errp
, "Can't migrate the vm that was paused due to "
2019 "previous migration");
2023 if (kvm_hwpoisoned_mem()) {
2024 error_setg(errp
, "Can't migrate this vm with hardware poisoned memory, "
2025 "please reboot the vm and try again");
2029 if (migration_is_blocked(errp
)) {
2033 if (migrate_mapped_ram()) {
2034 if (migrate_tls()) {
2035 error_setg(errp
, "Cannot use TLS with mapped-ram");
2039 if (migrate_multifd_compression()) {
2040 error_setg(errp
, "Cannot use compression with mapped-ram");
2045 if (migrate_mode_is_cpr(s
)) {
2046 const char *conflict
= NULL
;
2048 if (migrate_postcopy()) {
2049 conflict
= "postcopy";
2050 } else if (migrate_background_snapshot()) {
2051 conflict
= "background snapshot";
2052 } else if (migrate_colo()) {
2057 error_setg(errp
, "Cannot use %s with CPR", conflict
);
2062 if (migrate_init(s
, errp
)) {
2069 void qmp_migrate(const char *uri
, bool has_channels
,
2070 MigrationChannelList
*channels
, bool has_detach
, bool detach
,
2071 bool has_resume
, bool resume
, Error
**errp
)
2073 bool resume_requested
;
2074 Error
*local_err
= NULL
;
2075 MigrationState
*s
= migrate_get_current();
2076 g_autoptr(MigrationChannel
) channel
= NULL
;
2077 MigrationAddress
*addr
= NULL
;
2080 * Having preliminary checks for uri and channel
2082 if (!uri
== !channels
) {
2083 error_setg(errp
, "need either 'uri' or 'channels' argument");
2088 /* To verify that Migrate channel list has only item */
2089 if (channels
->next
) {
2090 error_setg(errp
, "Channel list has more than one entries");
2093 addr
= channels
->value
->addr
;
2097 /* caller uses the old URI syntax */
2098 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
2101 addr
= channel
->addr
;
2104 /* transport mechanism not suitable for migration? */
2105 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
2109 resume_requested
= has_resume
&& resume
;
2110 if (!migrate_prepare(s
, resume_requested
, errp
)) {
2111 /* Error detected, put into errp */
2115 if (!resume_requested
) {
2116 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
2121 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
2122 SocketAddress
*saddr
= &addr
->u
.socket
;
2123 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
2124 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
2125 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
2126 socket_start_outgoing_migration(s
, saddr
, &local_err
);
2127 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
2128 fd_start_outgoing_migration(s
, saddr
->u
.fd
.str
, &local_err
);
2131 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
2132 rdma_start_outgoing_migration(s
, &addr
->u
.rdma
, &local_err
);
2134 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
2135 exec_start_outgoing_migration(s
, addr
->u
.exec
.args
, &local_err
);
2136 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
2137 file_start_outgoing_migration(s
, &addr
->u
.file
, &local_err
);
2139 error_setg(&local_err
, QERR_INVALID_PARAMETER_VALUE
, "uri",
2140 "a valid migration protocol");
2141 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
2142 MIGRATION_STATUS_FAILED
);
2146 if (!resume_requested
) {
2147 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
2149 migrate_fd_error(s
, local_err
);
2150 error_propagate(errp
, local_err
);
2155 void qmp_migrate_cancel(Error
**errp
)
2157 migration_cancel(NULL
);
2160 void qmp_migrate_continue(MigrationStatus state
, Error
**errp
)
2162 MigrationState
*s
= migrate_get_current();
2163 if (s
->state
!= state
) {
2164 error_setg(errp
, "Migration not in expected state: %s",
2165 MigrationStatus_str(s
->state
));
2168 qemu_sem_post(&s
->pause_sem
);
2171 int migration_rp_wait(MigrationState
*s
)
2173 /* If migration has failure already, ignore the wait */
2174 if (migrate_has_error(s
)) {
2178 qemu_sem_wait(&s
->rp_state
.rp_sem
);
2180 /* After wait, double check that there's no failure */
2181 if (migrate_has_error(s
)) {
2188 void migration_rp_kick(MigrationState
*s
)
2190 qemu_sem_post(&s
->rp_state
.rp_sem
);
2193 static struct rp_cmd_args
{
2194 ssize_t len
; /* -1 = variable */
2197 [MIG_RP_MSG_INVALID
] = { .len
= -1, .name
= "INVALID" },
2198 [MIG_RP_MSG_SHUT
] = { .len
= 4, .name
= "SHUT" },
2199 [MIG_RP_MSG_PONG
] = { .len
= 4, .name
= "PONG" },
2200 [MIG_RP_MSG_REQ_PAGES
] = { .len
= 12, .name
= "REQ_PAGES" },
2201 [MIG_RP_MSG_REQ_PAGES_ID
] = { .len
= -1, .name
= "REQ_PAGES_ID" },
2202 [MIG_RP_MSG_RECV_BITMAP
] = { .len
= -1, .name
= "RECV_BITMAP" },
2203 [MIG_RP_MSG_RESUME_ACK
] = { .len
= 4, .name
= "RESUME_ACK" },
2204 [MIG_RP_MSG_SWITCHOVER_ACK
] = { .len
= 0, .name
= "SWITCHOVER_ACK" },
2205 [MIG_RP_MSG_MAX
] = { .len
= -1, .name
= "MAX" },
2209 * Process a request for pages received on the return path,
2210 * We're allowed to send more than requested (e.g. to round to our page size)
2211 * and we don't need to send pages that have already been sent.
2214 migrate_handle_rp_req_pages(MigrationState
*ms
, const char* rbname
,
2215 ram_addr_t start
, size_t len
, Error
**errp
)
2217 long our_host_ps
= qemu_real_host_page_size();
2219 trace_migrate_handle_rp_req_pages(rbname
, start
, len
);
2222 * Since we currently insist on matching page sizes, just sanity check
2223 * we're being asked for whole host pages.
2225 if (!QEMU_IS_ALIGNED(start
, our_host_ps
) ||
2226 !QEMU_IS_ALIGNED(len
, our_host_ps
)) {
2227 error_setg(errp
, "MIG_RP_MSG_REQ_PAGES: Misaligned page request, start:"
2228 RAM_ADDR_FMT
" len: %zd", start
, len
);
2232 ram_save_queue_pages(rbname
, start
, len
, errp
);
2235 static bool migrate_handle_rp_recv_bitmap(MigrationState
*s
, char *block_name
,
2238 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
2241 error_setg(errp
, "MIG_RP_MSG_RECV_BITMAP has invalid block name '%s'",
2246 /* Fetch the received bitmap and refresh the dirty bitmap */
2247 return ram_dirty_bitmap_reload(s
, block
, errp
);
2250 static bool migrate_handle_rp_resume_ack(MigrationState
*s
,
2251 uint32_t value
, Error
**errp
)
2253 trace_source_return_path_thread_resume_ack(value
);
2255 if (value
!= MIGRATION_RESUME_ACK_VALUE
) {
2256 error_setg(errp
, "illegal resume_ack value %"PRIu32
, value
);
2260 /* Now both sides are active. */
2261 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_RECOVER
,
2262 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2264 /* Notify send thread that time to continue send pages */
2265 migration_rp_kick(s
);
2271 * Release ms->rp_state.from_dst_file (and postcopy_qemufile_src if
2272 * existed) in a safe way.
2274 static void migration_release_dst_files(MigrationState
*ms
)
2278 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2280 * Reset the from_dst_file pointer first before releasing it, as we
2281 * can't block within lock section
2283 file
= ms
->rp_state
.from_dst_file
;
2284 ms
->rp_state
.from_dst_file
= NULL
;
2288 * Do the same to postcopy fast path socket too if there is. No
2289 * locking needed because this qemufile should only be managed by
2290 * return path thread.
2292 if (ms
->postcopy_qemufile_src
) {
2293 migration_ioc_unregister_yank_from_file(ms
->postcopy_qemufile_src
);
2294 qemu_file_shutdown(ms
->postcopy_qemufile_src
);
2295 qemu_fclose(ms
->postcopy_qemufile_src
);
2296 ms
->postcopy_qemufile_src
= NULL
;
2303 * Handles messages sent on the return path towards the source VM
2306 static void *source_return_path_thread(void *opaque
)
2308 MigrationState
*ms
= opaque
;
2309 QEMUFile
*rp
= ms
->rp_state
.from_dst_file
;
2310 uint16_t header_len
, header_type
;
2312 uint32_t tmp32
, sibling_error
;
2313 ram_addr_t start
= 0; /* =0 to silence warning */
2314 size_t len
= 0, expected_len
;
2318 trace_source_return_path_thread_entry();
2319 rcu_register_thread();
2321 while (migration_is_setup_or_active()) {
2322 trace_source_return_path_thread_loop_top();
2324 header_type
= qemu_get_be16(rp
);
2325 header_len
= qemu_get_be16(rp
);
2327 if (qemu_file_get_error(rp
)) {
2328 qemu_file_get_error_obj(rp
, &err
);
2332 if (header_type
>= MIG_RP_MSG_MAX
||
2333 header_type
== MIG_RP_MSG_INVALID
) {
2334 error_setg(&err
, "Received invalid message 0x%04x length 0x%04x",
2335 header_type
, header_len
);
2339 if ((rp_cmd_args
[header_type
].len
!= -1 &&
2340 header_len
!= rp_cmd_args
[header_type
].len
) ||
2341 header_len
> sizeof(buf
)) {
2342 error_setg(&err
, "Received '%s' message (0x%04x) with"
2343 "incorrect length %d expecting %zu",
2344 rp_cmd_args
[header_type
].name
, header_type
, header_len
,
2345 (size_t)rp_cmd_args
[header_type
].len
);
2349 /* We know we've got a valid header by this point */
2350 res
= qemu_get_buffer(rp
, buf
, header_len
);
2351 if (res
!= header_len
) {
2352 error_setg(&err
, "Failed reading data for message 0x%04x"
2353 " read %d expected %d",
2354 header_type
, res
, header_len
);
2358 /* OK, we have the message and the data */
2359 switch (header_type
) {
2360 case MIG_RP_MSG_SHUT
:
2361 sibling_error
= ldl_be_p(buf
);
2362 trace_source_return_path_thread_shut(sibling_error
);
2363 if (sibling_error
) {
2364 error_setg(&err
, "Sibling indicated error %d", sibling_error
);
2367 * We'll let the main thread deal with closing the RP
2368 * we could do a shutdown(2) on it, but we're the only user
2369 * anyway, so there's nothing gained.
2373 case MIG_RP_MSG_PONG
:
2374 tmp32
= ldl_be_p(buf
);
2375 trace_source_return_path_thread_pong(tmp32
);
2376 qemu_sem_post(&ms
->rp_state
.rp_pong_acks
);
2379 case MIG_RP_MSG_REQ_PAGES
:
2380 start
= ldq_be_p(buf
);
2381 len
= ldl_be_p(buf
+ 8);
2382 migrate_handle_rp_req_pages(ms
, NULL
, start
, len
, &err
);
2388 case MIG_RP_MSG_REQ_PAGES_ID
:
2389 expected_len
= 12 + 1; /* header + termination */
2391 if (header_len
>= expected_len
) {
2392 start
= ldq_be_p(buf
);
2393 len
= ldl_be_p(buf
+ 8);
2394 /* Now we expect an idstr */
2395 tmp32
= buf
[12]; /* Length of the following idstr */
2396 buf
[13 + tmp32
] = '\0';
2397 expected_len
+= tmp32
;
2399 if (header_len
!= expected_len
) {
2400 error_setg(&err
, "Req_Page_id with length %d expecting %zd",
2401 header_len
, expected_len
);
2404 migrate_handle_rp_req_pages(ms
, (char *)&buf
[13], start
, len
,
2411 case MIG_RP_MSG_RECV_BITMAP
:
2412 if (header_len
< 1) {
2413 error_setg(&err
, "MIG_RP_MSG_RECV_BITMAP missing block name");
2416 /* Format: len (1B) + idstr (<255B). This ends the idstr. */
2417 buf
[buf
[0] + 1] = '\0';
2418 if (!migrate_handle_rp_recv_bitmap(ms
, (char *)(buf
+ 1), &err
)) {
2423 case MIG_RP_MSG_RESUME_ACK
:
2424 tmp32
= ldl_be_p(buf
);
2425 if (!migrate_handle_rp_resume_ack(ms
, tmp32
, &err
)) {
2430 case MIG_RP_MSG_SWITCHOVER_ACK
:
2431 ms
->switchover_acked
= true;
2432 trace_source_return_path_thread_switchover_acked();
2442 migrate_set_error(ms
, err
);
2444 trace_source_return_path_thread_bad_end();
2447 if (ms
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2449 * this will be extremely unlikely: that we got yet another network
2450 * issue during recovering of the 1st network failure.. during this
2451 * period the main migration thread can be waiting on rp_sem for
2452 * this thread to sync with the other side.
2454 * When this happens, explicitly kick the migration thread out of
2455 * RECOVER stage and back to PAUSED, so the admin can try
2458 migration_rp_kick(ms
);
2461 trace_source_return_path_thread_end();
2462 rcu_unregister_thread();
2467 static int open_return_path_on_source(MigrationState
*ms
)
2469 ms
->rp_state
.from_dst_file
= qemu_file_get_return_path(ms
->to_dst_file
);
2470 if (!ms
->rp_state
.from_dst_file
) {
2474 trace_open_return_path_on_source();
2476 qemu_thread_create(&ms
->rp_state
.rp_thread
, "mig/src/rp-thr",
2477 source_return_path_thread
, ms
, QEMU_THREAD_JOINABLE
);
2478 ms
->rp_state
.rp_thread_created
= true;
2480 trace_open_return_path_on_source_continue();
2485 /* Return true if error detected, or false otherwise */
2486 static bool close_return_path_on_source(MigrationState
*ms
)
2488 if (!ms
->rp_state
.rp_thread_created
) {
2492 trace_migration_return_path_end_before();
2495 * If this is a normal exit then the destination will send a SHUT
2496 * and the rp_thread will exit, however if there's an error we
2497 * need to cause it to exit. shutdown(2), if we have it, will
2498 * cause it to unblock if it's stuck waiting for the destination.
2500 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2501 if (migrate_has_error(ms
) && ms
->rp_state
.from_dst_file
) {
2502 qemu_file_shutdown(ms
->rp_state
.from_dst_file
);
2506 qemu_thread_join(&ms
->rp_state
.rp_thread
);
2507 ms
->rp_state
.rp_thread_created
= false;
2508 migration_release_dst_files(ms
);
2509 trace_migration_return_path_end_after();
2511 /* Return path will persist the error in MigrationState when quit */
2512 return migrate_has_error(ms
);
2516 migration_wait_main_channel(MigrationState
*ms
)
2518 /* Wait until one PONG message received */
2519 qemu_sem_wait(&ms
->rp_state
.rp_pong_acks
);
2523 * Switch from normal iteration to postcopy
2524 * Returns non-0 on error
2526 static int postcopy_start(MigrationState
*ms
, Error
**errp
)
2529 QIOChannelBuffer
*bioc
;
2531 uint64_t bandwidth
= migrate_max_postcopy_bandwidth();
2532 bool restart_block
= false;
2533 int cur_state
= MIGRATION_STATUS_ACTIVE
;
2535 if (migrate_postcopy_preempt()) {
2536 migration_wait_main_channel(ms
);
2537 if (postcopy_preempt_establish_channel(ms
)) {
2538 migrate_set_state(&ms
->state
, ms
->state
, MIGRATION_STATUS_FAILED
);
2539 error_setg(errp
, "%s: Failed to establish preempt channel",
2545 if (!migrate_pause_before_switchover()) {
2546 migrate_set_state(&ms
->state
, MIGRATION_STATUS_ACTIVE
,
2547 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2550 trace_postcopy_start();
2552 trace_postcopy_start_set_run();
2554 ret
= migration_stop_vm(ms
, RUN_STATE_FINISH_MIGRATE
);
2556 error_setg_errno(errp
, -ret
, "%s: Failed to stop the VM", __func__
);
2560 ret
= migration_maybe_pause(ms
, &cur_state
,
2561 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2563 error_setg_errno(errp
, -ret
, "%s: Failed in migration_maybe_pause()",
2568 ret
= bdrv_inactivate_all();
2570 error_setg_errno(errp
, -ret
, "%s: Failed in bdrv_inactivate_all()",
2574 restart_block
= true;
2577 * Cause any non-postcopiable, but iterative devices to
2578 * send out their final data.
2580 qemu_savevm_state_complete_precopy(ms
->to_dst_file
, true, false);
2583 * in Finish migrate and with the io-lock held everything should
2584 * be quiet, but we've potentially still got dirty pages and we
2585 * need to tell the destination to throw any pages it's already received
2588 if (migrate_postcopy_ram()) {
2589 ram_postcopy_send_discard_bitmap(ms
);
2593 * send rest of state - note things that are doing postcopy
2594 * will notice we're in POSTCOPY_ACTIVE and not actually
2595 * wrap their state up here
2597 migration_rate_set(bandwidth
);
2598 if (migrate_postcopy_ram()) {
2599 /* Ping just for debugging, helps line traces up */
2600 qemu_savevm_send_ping(ms
->to_dst_file
, 2);
2604 * While loading the device state we may trigger page transfer
2605 * requests and the fd must be free to process those, and thus
2606 * the destination must read the whole device state off the fd before
2607 * it starts processing it. Unfortunately the ad-hoc migration format
2608 * doesn't allow the destination to know the size to read without fully
2609 * parsing it through each devices load-state code (especially the open
2610 * coded devices that use get/put).
2611 * So we wrap the device state up in a package with a length at the start;
2612 * to do this we use a qemu_buf to hold the whole of the device state.
2614 bioc
= qio_channel_buffer_new(4096);
2615 qio_channel_set_name(QIO_CHANNEL(bioc
), "migration-postcopy-buffer");
2616 fb
= qemu_file_new_output(QIO_CHANNEL(bioc
));
2617 object_unref(OBJECT(bioc
));
2620 * Make sure the receiver can get incoming pages before we send the rest
2623 qemu_savevm_send_postcopy_listen(fb
);
2625 qemu_savevm_state_complete_precopy(fb
, false, false);
2626 if (migrate_postcopy_ram()) {
2627 qemu_savevm_send_ping(fb
, 3);
2630 qemu_savevm_send_postcopy_run(fb
);
2632 /* <><> end of stuff going into the package */
2634 /* Last point of recovery; as soon as we send the package the destination
2635 * can open devices and potentially start running.
2636 * Lets just check again we've not got any errors.
2638 ret
= qemu_file_get_error(ms
->to_dst_file
);
2640 error_setg(errp
, "postcopy_start: Migration stream errored (pre package)");
2644 restart_block
= false;
2646 /* Now send that blob */
2647 if (qemu_savevm_send_packaged(ms
->to_dst_file
, bioc
->data
, bioc
->usage
)) {
2648 error_setg(errp
, "%s: Failed to send packaged data", __func__
);
2653 /* Send a notify to give a chance for anything that needs to happen
2654 * at the transition to postcopy and after the device state; in particular
2655 * spice needs to trigger a transition now
2657 migration_call_notifiers(ms
, MIG_EVENT_PRECOPY_DONE
, NULL
);
2659 migration_downtime_end(ms
);
2663 if (migrate_postcopy_ram()) {
2665 * Although this ping is just for debug, it could potentially be
2666 * used for getting a better measurement of downtime at the source.
2668 qemu_savevm_send_ping(ms
->to_dst_file
, 4);
2671 if (migrate_release_ram()) {
2672 ram_postcopy_migrated_memory_release(ms
);
2675 ret
= qemu_file_get_error(ms
->to_dst_file
);
2677 error_setg_errno(errp
, -ret
, "postcopy_start: Migration stream error");
2681 trace_postcopy_preempt_enabled(migrate_postcopy_preempt());
2688 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2689 MIGRATION_STATUS_FAILED
);
2690 if (restart_block
) {
2691 /* A failure happened early enough that we know the destination hasn't
2692 * accessed block devices, so we're safe to recover.
2694 Error
*local_err
= NULL
;
2696 bdrv_activate_all(&local_err
);
2698 error_report_err(local_err
);
2701 migration_call_notifiers(ms
, MIG_EVENT_PRECOPY_FAILED
, NULL
);
2707 * migration_maybe_pause: Pause if required to by
2708 * migrate_pause_before_switchover called with the BQL locked
2709 * Returns: 0 on success
2711 static int migration_maybe_pause(MigrationState
*s
,
2712 int *current_active_state
,
2715 if (!migrate_pause_before_switchover()) {
2719 /* Since leaving this state is not atomic with posting the semaphore
2720 * it's possible that someone could have issued multiple migrate_continue
2721 * and the semaphore is incorrectly positive at this point;
2722 * the docs say it's undefined to reinit a semaphore that's already
2723 * init'd, so use timedwait to eat up any existing posts.
2725 while (qemu_sem_timedwait(&s
->pause_sem
, 1) == 0) {
2726 /* This block intentionally left blank */
2730 * If the migration is cancelled when it is in the completion phase,
2731 * the migration state is set to MIGRATION_STATUS_CANCELLING.
2732 * So we don't need to wait a semaphore, otherwise we would always
2733 * wait for the 'pause_sem' semaphore.
2735 if (s
->state
!= MIGRATION_STATUS_CANCELLING
) {
2737 migrate_set_state(&s
->state
, *current_active_state
,
2738 MIGRATION_STATUS_PRE_SWITCHOVER
);
2739 qemu_sem_wait(&s
->pause_sem
);
2740 migrate_set_state(&s
->state
, MIGRATION_STATUS_PRE_SWITCHOVER
,
2742 *current_active_state
= new_state
;
2746 return s
->state
== new_state
? 0 : -EINVAL
;
2749 static int migration_completion_precopy(MigrationState
*s
,
2750 int *current_active_state
)
2756 if (!migrate_mode_is_cpr(s
)) {
2757 ret
= migration_stop_vm(s
, RUN_STATE_FINISH_MIGRATE
);
2763 ret
= migration_maybe_pause(s
, current_active_state
,
2764 MIGRATION_STATUS_DEVICE
);
2770 * Inactivate disks except in COLO, and track that we have done so in order
2771 * to remember to reactivate them if migration fails or is cancelled.
2773 s
->block_inactive
= !migrate_colo();
2774 migration_rate_set(RATE_LIMIT_DISABLED
);
2775 ret
= qemu_savevm_state_complete_precopy(s
->to_dst_file
, false,
2782 static void migration_completion_postcopy(MigrationState
*s
)
2784 trace_migration_completion_postcopy_end();
2787 qemu_savevm_state_complete_postcopy(s
->to_dst_file
);
2791 * Shutdown the postcopy fast path thread. This is only needed when dest
2792 * QEMU binary is old (7.1/7.2). QEMU 8.0+ doesn't need this.
2794 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
2795 postcopy_preempt_shutdown_file(s
);
2798 trace_migration_completion_postcopy_end_after_complete();
2801 static void migration_completion_failed(MigrationState
*s
,
2802 int current_active_state
)
2804 if (s
->block_inactive
&& (s
->state
== MIGRATION_STATUS_ACTIVE
||
2805 s
->state
== MIGRATION_STATUS_DEVICE
)) {
2807 * If not doing postcopy, vm_start() will be called: let's
2808 * regain control on images.
2810 Error
*local_err
= NULL
;
2813 bdrv_activate_all(&local_err
);
2815 error_report_err(local_err
);
2817 s
->block_inactive
= false;
2822 migrate_set_state(&s
->state
, current_active_state
,
2823 MIGRATION_STATUS_FAILED
);
2827 * migration_completion: Used by migration_thread when there's not much left.
2828 * The caller 'breaks' the loop when this returns.
2830 * @s: Current migration state
2832 static void migration_completion(MigrationState
*s
)
2835 int current_active_state
= s
->state
;
2836 Error
*local_err
= NULL
;
2838 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2839 ret
= migration_completion_precopy(s
, ¤t_active_state
);
2840 } else if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2841 migration_completion_postcopy(s
);
2850 if (close_return_path_on_source(s
)) {
2854 if (qemu_file_get_error(s
->to_dst_file
)) {
2855 trace_migration_completion_file_err();
2859 if (migrate_colo() && s
->state
== MIGRATION_STATUS_ACTIVE
) {
2860 /* COLO does not support postcopy */
2861 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
2862 MIGRATION_STATUS_COLO
);
2864 migration_completion_end(s
);
2870 if (qemu_file_get_error_obj(s
->to_dst_file
, &local_err
)) {
2871 migrate_set_error(s
, local_err
);
2872 error_free(local_err
);
2874 error_setg_errno(&local_err
, -ret
, "Error in migration completion");
2875 migrate_set_error(s
, local_err
);
2876 error_free(local_err
);
2879 migration_completion_failed(s
, current_active_state
);
2883 * bg_migration_completion: Used by bg_migration_thread when after all the
2884 * RAM has been saved. The caller 'breaks' the loop when this returns.
2886 * @s: Current migration state
2888 static void bg_migration_completion(MigrationState
*s
)
2890 int current_active_state
= s
->state
;
2892 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2894 * By this moment we have RAM content saved into the migration stream.
2895 * The next step is to flush the non-RAM content (device state)
2896 * right after the ram content. The device state has been stored into
2897 * the temporary buffer before RAM saving started.
2899 qemu_put_buffer(s
->to_dst_file
, s
->bioc
->data
, s
->bioc
->usage
);
2900 qemu_fflush(s
->to_dst_file
);
2901 } else if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
2905 if (qemu_file_get_error(s
->to_dst_file
)) {
2906 trace_migration_completion_file_err();
2910 migration_completion_end(s
);
2914 migrate_set_state(&s
->state
, current_active_state
,
2915 MIGRATION_STATUS_FAILED
);
2918 typedef enum MigThrError
{
2919 /* No error detected */
2920 MIG_THR_ERR_NONE
= 0,
2921 /* Detected error, but resumed successfully */
2922 MIG_THR_ERR_RECOVERED
= 1,
2923 /* Detected fatal error, need to exit */
2924 MIG_THR_ERR_FATAL
= 2,
2927 static int postcopy_resume_handshake(MigrationState
*s
)
2929 qemu_savevm_send_postcopy_resume(s
->to_dst_file
);
2931 while (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2932 if (migration_rp_wait(s
)) {
2937 if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2944 /* Return zero if success, or <0 for error */
2945 static int postcopy_do_resume(MigrationState
*s
)
2950 * Call all the resume_prepare() hooks, so that modules can be
2951 * ready for the migration resume.
2953 ret
= qemu_savevm_state_resume_prepare(s
);
2955 error_report("%s: resume_prepare() failure detected: %d",
2961 * If preempt is enabled, re-establish the preempt channel. Note that
2962 * we do it after resume prepare to make sure the main channel will be
2963 * created before the preempt channel. E.g. with weak network, the
2964 * dest QEMU may get messed up with the preempt and main channels on
2965 * the order of connection setup. This guarantees the correct order.
2967 ret
= postcopy_preempt_establish_channel(s
);
2969 error_report("%s: postcopy_preempt_establish_channel(): %d",
2975 * Last handshake with destination on the resume (destination will
2976 * switch to postcopy-active afterwards)
2978 ret
= postcopy_resume_handshake(s
);
2980 error_report("%s: handshake failed: %d", __func__
, ret
);
2988 * We don't return until we are in a safe state to continue current
2989 * postcopy migration. Returns MIG_THR_ERR_RECOVERED if recovered, or
2990 * MIG_THR_ERR_FATAL if unrecovery failure happened.
2992 static MigThrError
postcopy_pause(MigrationState
*s
)
2994 assert(s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
3000 * We're already pausing, so ignore any errors on the return
3001 * path and just wait for the thread to finish. It will be
3002 * re-created when we resume.
3004 close_return_path_on_source(s
);
3007 * Current channel is possibly broken. Release it. Note that this is
3008 * guaranteed even without lock because to_dst_file should only be
3009 * modified by the migration thread. That also guarantees that the
3010 * unregister of yank is safe too without the lock. It should be safe
3011 * even to be within the qemu_file_lock, but we didn't do that to avoid
3012 * taking more mutex (yank_lock) within qemu_file_lock. TL;DR: we make
3013 * the qemu_file_lock critical section as small as possible.
3015 assert(s
->to_dst_file
);
3016 migration_ioc_unregister_yank_from_file(s
->to_dst_file
);
3017 qemu_mutex_lock(&s
->qemu_file_lock
);
3018 file
= s
->to_dst_file
;
3019 s
->to_dst_file
= NULL
;
3020 qemu_mutex_unlock(&s
->qemu_file_lock
);
3022 qemu_file_shutdown(file
);
3025 migrate_set_state(&s
->state
, s
->state
,
3026 MIGRATION_STATUS_POSTCOPY_PAUSED
);
3028 error_report("Detected IO failure for postcopy. "
3029 "Migration paused.");
3032 * We wait until things fixed up. Then someone will setup the
3033 * status back for us.
3036 qemu_sem_wait(&s
->postcopy_pause_sem
);
3037 } while (postcopy_is_paused(s
->state
));
3039 if (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
3040 /* Woken up by a recover procedure. Give it a shot */
3042 /* Do the resume logic */
3043 if (postcopy_do_resume(s
) == 0) {
3044 /* Let's continue! */
3045 trace_postcopy_pause_continued();
3046 return MIG_THR_ERR_RECOVERED
;
3049 * Something wrong happened during the recovery, let's
3050 * pause again. Pause is always better than throwing
3056 /* This is not right... Time to quit. */
3057 return MIG_THR_ERR_FATAL
;
3062 void migration_file_set_error(int ret
, Error
*err
)
3064 MigrationState
*s
= current_migration
;
3066 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
3067 if (s
->to_dst_file
) {
3068 qemu_file_set_error_obj(s
->to_dst_file
, ret
, err
);
3070 error_report_err(err
);
3075 static MigThrError
migration_detect_error(MigrationState
*s
)
3078 int state
= s
->state
;
3079 Error
*local_error
= NULL
;
3081 if (state
== MIGRATION_STATUS_CANCELLING
||
3082 state
== MIGRATION_STATUS_CANCELLED
) {
3083 /* End the migration, but don't set the state to failed */
3084 return MIG_THR_ERR_FATAL
;
3088 * Try to detect any file errors. Note that postcopy_qemufile_src will
3089 * be NULL when postcopy preempt is not enabled.
3091 ret
= qemu_file_get_error_obj_any(s
->to_dst_file
,
3092 s
->postcopy_qemufile_src
,
3095 /* Everything is fine */
3096 assert(!local_error
);
3097 return MIG_THR_ERR_NONE
;
3101 migrate_set_error(s
, local_error
);
3102 error_free(local_error
);
3105 if (state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
&& ret
) {
3107 * For postcopy, we allow the network to be down for a
3108 * while. After that, it can be continued by a
3111 return postcopy_pause(s
);
3114 * For precopy (or postcopy with error outside IO), we fail
3117 migrate_set_state(&s
->state
, state
, MIGRATION_STATUS_FAILED
);
3118 trace_migration_thread_file_err();
3120 /* Time to stop the migration, now. */
3121 return MIG_THR_ERR_FATAL
;
3125 static void migration_completion_end(MigrationState
*s
)
3127 uint64_t bytes
= migration_transferred_bytes();
3128 int64_t end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3129 int64_t transfer_time
;
3132 * Take the BQL here so that query-migrate on the QMP thread sees:
3133 * - atomic update of s->total_time and s->mbps;
3134 * - correct ordering of s->mbps update vs. s->state;
3137 migration_downtime_end(s
);
3138 s
->total_time
= end_time
- s
->start_time
;
3139 transfer_time
= s
->total_time
- s
->setup_time
;
3140 if (transfer_time
) {
3141 s
->mbps
= ((double) bytes
* 8.0) / transfer_time
/ 1000;
3144 migrate_set_state(&s
->state
, s
->state
,
3145 MIGRATION_STATUS_COMPLETED
);
3149 static void update_iteration_initial_status(MigrationState
*s
)
3152 * Update these three fields at the same time to avoid mismatch info lead
3153 * wrong speed calculation.
3155 s
->iteration_start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3156 s
->iteration_initial_bytes
= migration_transferred_bytes();
3157 s
->iteration_initial_pages
= ram_get_total_transferred_pages();
3160 static void migration_update_counters(MigrationState
*s
,
3161 int64_t current_time
)
3163 uint64_t transferred
, transferred_pages
, time_spent
;
3164 uint64_t current_bytes
; /* bytes transferred since the beginning */
3165 uint64_t switchover_bw
;
3166 /* Expected bandwidth when switching over to destination QEMU */
3167 double expected_bw_per_ms
;
3170 if (current_time
< s
->iteration_start_time
+ BUFFER_DELAY
) {
3174 switchover_bw
= migrate_avail_switchover_bandwidth();
3175 current_bytes
= migration_transferred_bytes();
3176 transferred
= current_bytes
- s
->iteration_initial_bytes
;
3177 time_spent
= current_time
- s
->iteration_start_time
;
3178 bandwidth
= (double)transferred
/ time_spent
;
3180 if (switchover_bw
) {
3182 * If the user specified a switchover bandwidth, let's trust the
3183 * user so that can be more accurate than what we estimated.
3185 expected_bw_per_ms
= switchover_bw
/ 1000;
3187 /* If the user doesn't specify bandwidth, we use the estimated */
3188 expected_bw_per_ms
= bandwidth
;
3191 s
->threshold_size
= expected_bw_per_ms
* migrate_downtime_limit();
3193 s
->mbps
= (((double) transferred
* 8.0) /
3194 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0;
3196 transferred_pages
= ram_get_total_transferred_pages() -
3197 s
->iteration_initial_pages
;
3198 s
->pages_per_second
= (double) transferred_pages
/
3199 (((double) time_spent
/ 1000.0));
3202 * if we haven't sent anything, we don't want to
3203 * recalculate. 10000 is a small enough number for our purposes
3205 if (stat64_get(&mig_stats
.dirty_pages_rate
) &&
3206 transferred
> 10000) {
3207 s
->expected_downtime
=
3208 stat64_get(&mig_stats
.dirty_bytes_last_sync
) / expected_bw_per_ms
;
3211 migration_rate_reset();
3213 update_iteration_initial_status(s
);
3215 trace_migrate_transferred(transferred
, time_spent
,
3216 /* Both in unit bytes/ms */
3217 bandwidth
, switchover_bw
/ 1000,
3221 static bool migration_can_switchover(MigrationState
*s
)
3223 if (!migrate_switchover_ack()) {
3227 /* No reason to wait for switchover ACK if VM is stopped */
3228 if (!runstate_is_running()) {
3232 return s
->switchover_acked
;
3235 /* Migration thread iteration status */
3237 MIG_ITERATE_RESUME
, /* Resume current iteration */
3238 MIG_ITERATE_SKIP
, /* Skip current iteration */
3239 MIG_ITERATE_BREAK
, /* Break the loop */
3243 * Return true if continue to the next iteration directly, false
3246 static MigIterateState
migration_iteration_run(MigrationState
*s
)
3248 uint64_t must_precopy
, can_postcopy
, pending_size
;
3249 Error
*local_err
= NULL
;
3250 bool in_postcopy
= s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
;
3251 bool can_switchover
= migration_can_switchover(s
);
3253 qemu_savevm_state_pending_estimate(&must_precopy
, &can_postcopy
);
3254 pending_size
= must_precopy
+ can_postcopy
;
3255 trace_migrate_pending_estimate(pending_size
, must_precopy
, can_postcopy
);
3257 if (pending_size
< s
->threshold_size
) {
3258 qemu_savevm_state_pending_exact(&must_precopy
, &can_postcopy
);
3259 pending_size
= must_precopy
+ can_postcopy
;
3260 trace_migrate_pending_exact(pending_size
, must_precopy
, can_postcopy
);
3263 if ((!pending_size
|| pending_size
< s
->threshold_size
) && can_switchover
) {
3264 trace_migration_thread_low_pending(pending_size
);
3265 migration_completion(s
);
3266 return MIG_ITERATE_BREAK
;
3269 /* Still a significant amount to transfer */
3270 if (!in_postcopy
&& must_precopy
<= s
->threshold_size
&& can_switchover
&&
3271 qatomic_read(&s
->start_postcopy
)) {
3272 if (postcopy_start(s
, &local_err
)) {
3273 migrate_set_error(s
, local_err
);
3274 error_report_err(local_err
);
3276 return MIG_ITERATE_SKIP
;
3279 /* Just another iteration step */
3280 qemu_savevm_state_iterate(s
->to_dst_file
, in_postcopy
);
3281 return MIG_ITERATE_RESUME
;
3284 static void migration_iteration_finish(MigrationState
*s
)
3286 /* If we enabled cpu throttling for auto-converge, turn it off. */
3287 cpu_throttle_stop();
3291 case MIGRATION_STATUS_COMPLETED
:
3292 runstate_set(RUN_STATE_POSTMIGRATE
);
3294 case MIGRATION_STATUS_COLO
:
3295 assert(migrate_colo());
3296 migrate_start_colo_process(s
);
3297 s
->vm_old_state
= RUN_STATE_RUNNING
;
3299 case MIGRATION_STATUS_FAILED
:
3300 case MIGRATION_STATUS_CANCELLED
:
3301 case MIGRATION_STATUS_CANCELLING
:
3302 if (runstate_is_live(s
->vm_old_state
)) {
3303 if (!runstate_check(RUN_STATE_SHUTDOWN
)) {
3307 if (runstate_check(RUN_STATE_FINISH_MIGRATE
)) {
3308 runstate_set(s
->vm_old_state
);
3314 /* Should not reach here, but if so, forgive the VM. */
3315 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3319 migration_bh_schedule(migrate_fd_cleanup_bh
, s
);
3323 static void bg_migration_iteration_finish(MigrationState
*s
)
3326 * Stop tracking RAM writes - un-protect memory, un-register UFFD
3327 * memory ranges, flush kernel wait queues and wake up threads
3328 * waiting for write fault to be resolved.
3330 ram_write_tracking_stop();
3334 case MIGRATION_STATUS_COMPLETED
:
3335 case MIGRATION_STATUS_ACTIVE
:
3336 case MIGRATION_STATUS_FAILED
:
3337 case MIGRATION_STATUS_CANCELLED
:
3338 case MIGRATION_STATUS_CANCELLING
:
3342 /* Should not reach here, but if so, forgive the VM. */
3343 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3347 migration_bh_schedule(migrate_fd_cleanup_bh
, s
);
3352 * Return true if continue to the next iteration directly, false
3355 static MigIterateState
bg_migration_iteration_run(MigrationState
*s
)
3359 res
= qemu_savevm_state_iterate(s
->to_dst_file
, false);
3361 bg_migration_completion(s
);
3362 return MIG_ITERATE_BREAK
;
3365 return MIG_ITERATE_RESUME
;
3368 void migration_make_urgent_request(void)
3370 qemu_sem_post(&migrate_get_current()->rate_limit_sem
);
3373 void migration_consume_urgent_request(void)
3375 qemu_sem_wait(&migrate_get_current()->rate_limit_sem
);
3378 /* Returns true if the rate limiting was broken by an urgent request */
3379 bool migration_rate_limit(void)
3381 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3382 MigrationState
*s
= migrate_get_current();
3384 bool urgent
= false;
3385 migration_update_counters(s
, now
);
3386 if (migration_rate_exceeded(s
->to_dst_file
)) {
3388 if (qemu_file_get_error(s
->to_dst_file
)) {
3392 * Wait for a delay to do rate limiting OR
3393 * something urgent to post the semaphore.
3395 int ms
= s
->iteration_start_time
+ BUFFER_DELAY
- now
;
3396 trace_migration_rate_limit_pre(ms
);
3397 if (qemu_sem_timedwait(&s
->rate_limit_sem
, ms
) == 0) {
3399 * We were woken by one or more urgent things but
3400 * the timedwait will have consumed one of them.
3401 * The service routine for the urgent wake will dec
3402 * the semaphore itself for each item it consumes,
3403 * so add this one we just eat back.
3405 qemu_sem_post(&s
->rate_limit_sem
);
3408 trace_migration_rate_limit_post(urgent
);
3414 * if failover devices are present, wait they are completely
3418 static void qemu_savevm_wait_unplug(MigrationState
*s
, int old_state
,
3421 if (qemu_savevm_state_guest_unplug_pending()) {
3422 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_WAIT_UNPLUG
);
3424 while (s
->state
== MIGRATION_STATUS_WAIT_UNPLUG
&&
3425 qemu_savevm_state_guest_unplug_pending()) {
3426 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3428 if (s
->state
!= MIGRATION_STATUS_WAIT_UNPLUG
) {
3429 int timeout
= 120; /* 30 seconds */
3431 * migration has been canceled
3432 * but as we have started an unplug we must wait the end
3433 * to be able to plug back the card
3435 while (timeout
-- && qemu_savevm_state_guest_unplug_pending()) {
3436 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3438 if (qemu_savevm_state_guest_unplug_pending() &&
3440 warn_report("migration: partially unplugged device on "
3445 migrate_set_state(&s
->state
, MIGRATION_STATUS_WAIT_UNPLUG
, new_state
);
3447 migrate_set_state(&s
->state
, old_state
, new_state
);
3452 * Master migration thread on the source VM.
3453 * It drives the migration and pumps the data down the outgoing channel.
3455 static void *migration_thread(void *opaque
)
3457 MigrationState
*s
= opaque
;
3458 MigrationThread
*thread
= NULL
;
3459 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3460 MigThrError thr_error
;
3461 bool urgent
= false;
3462 Error
*local_err
= NULL
;
3465 thread
= migration_threads_add("live_migration", qemu_get_thread_id());
3467 rcu_register_thread();
3469 object_ref(OBJECT(s
));
3470 update_iteration_initial_status(s
);
3472 if (!multifd_send_setup()) {
3477 qemu_savevm_state_header(s
->to_dst_file
);
3481 * If we opened the return path, we need to make sure dst has it
3484 if (s
->rp_state
.rp_thread_created
) {
3485 /* Now tell the dest that it should open its end so it can reply */
3486 qemu_savevm_send_open_return_path(s
->to_dst_file
);
3488 /* And do a ping that will make stuff easier to debug */
3489 qemu_savevm_send_ping(s
->to_dst_file
, 1);
3492 if (migrate_postcopy()) {
3494 * Tell the destination that we *might* want to do postcopy later;
3495 * if the other end can't do postcopy it should fail now, nice and
3498 qemu_savevm_send_postcopy_advise(s
->to_dst_file
);
3501 if (migrate_colo()) {
3502 /* Notify migration destination that we enable COLO */
3503 qemu_savevm_send_colo_enable(s
->to_dst_file
);
3507 ret
= qemu_savevm_state_setup(s
->to_dst_file
, &local_err
);
3510 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3511 MIGRATION_STATUS_ACTIVE
);
3514 * Handle SETUP failures after waiting for virtio-net-failover
3515 * devices to unplug. This to preserve migration state transitions.
3518 migrate_set_error(s
, local_err
);
3519 error_free(local_err
);
3520 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3521 MIGRATION_STATUS_FAILED
);
3525 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3527 trace_migration_thread_setup_complete();
3529 while (migration_is_active()) {
3530 if (urgent
|| !migration_rate_exceeded(s
->to_dst_file
)) {
3531 MigIterateState iter_state
= migration_iteration_run(s
);
3532 if (iter_state
== MIG_ITERATE_SKIP
) {
3534 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3540 * Try to detect any kind of failures, and see whether we
3541 * should stop the migration now.
3543 thr_error
= migration_detect_error(s
);
3544 if (thr_error
== MIG_THR_ERR_FATAL
) {
3545 /* Stop migration */
3547 } else if (thr_error
== MIG_THR_ERR_RECOVERED
) {
3549 * Just recovered from a e.g. network failure, reset all
3550 * the local variables. This is important to avoid
3551 * breaking transferred_bytes and bandwidth calculation
3553 update_iteration_initial_status(s
);
3556 urgent
= migration_rate_limit();
3560 trace_migration_thread_after_loop();
3561 migration_iteration_finish(s
);
3562 object_unref(OBJECT(s
));
3563 rcu_unregister_thread();
3564 migration_threads_remove(thread
);
3568 static void bg_migration_vm_start_bh(void *opaque
)
3570 MigrationState
*s
= opaque
;
3572 vm_resume(s
->vm_old_state
);
3573 migration_downtime_end(s
);
3577 * Background snapshot thread, based on live migration code.
3578 * This is an alternative implementation of live migration mechanism
3579 * introduced specifically to support background snapshots.
3581 * It takes advantage of userfault_fd write protection mechanism introduced
3582 * in v5.7 kernel. Compared to existing dirty page logging migration much
3583 * lesser stream traffic is produced resulting in smaller snapshot images,
3584 * simply cause of no page duplicates can get into the stream.
3586 * Another key point is that generated vmstate stream reflects machine state
3587 * 'frozen' at the beginning of snapshot creation compared to dirty page logging
3588 * mechanism, which effectively results in that saved snapshot is the state of VM
3589 * at the end of the process.
3591 static void *bg_migration_thread(void *opaque
)
3593 MigrationState
*s
= opaque
;
3594 int64_t setup_start
;
3595 MigThrError thr_error
;
3597 bool early_fail
= true;
3598 Error
*local_err
= NULL
;
3601 rcu_register_thread();
3602 object_ref(OBJECT(s
));
3604 migration_rate_set(RATE_LIMIT_DISABLED
);
3606 setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3608 * We want to save vmstate for the moment when migration has been
3609 * initiated but also we want to save RAM content while VM is running.
3610 * The RAM content should appear first in the vmstate. So, we first
3611 * stash the non-RAM part of the vmstate to the temporary buffer,
3612 * then write RAM part of the vmstate to the migration stream
3613 * with vCPUs running and, finally, write stashed non-RAM part of
3614 * the vmstate from the buffer to the migration stream.
3616 s
->bioc
= qio_channel_buffer_new(512 * 1024);
3617 qio_channel_set_name(QIO_CHANNEL(s
->bioc
), "vmstate-buffer");
3618 fb
= qemu_file_new_output(QIO_CHANNEL(s
->bioc
));
3619 object_unref(OBJECT(s
->bioc
));
3621 update_iteration_initial_status(s
);
3624 * Prepare for tracking memory writes with UFFD-WP - populate
3625 * RAM pages before protecting.
3628 ram_write_tracking_prepare();
3632 qemu_savevm_state_header(s
->to_dst_file
);
3633 ret
= qemu_savevm_state_setup(s
->to_dst_file
, &local_err
);
3636 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3637 MIGRATION_STATUS_ACTIVE
);
3640 * Handle SETUP failures after waiting for virtio-net-failover
3641 * devices to unplug. This to preserve migration state transitions.
3644 migrate_set_error(s
, local_err
);
3645 error_free(local_err
);
3646 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3647 MIGRATION_STATUS_FAILED
);
3651 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3653 trace_migration_thread_setup_complete();
3657 if (migration_stop_vm(s
, RUN_STATE_PAUSED
)) {
3661 * Put vCPUs in sync with shadow context structures, then
3662 * save their state to channel-buffer along with devices.
3664 cpu_synchronize_all_states();
3665 if (qemu_savevm_state_complete_precopy_non_iterable(fb
, false, false)) {
3669 * Since we are going to get non-iterable state data directly
3670 * from s->bioc->data, explicit flush is needed here.
3674 /* Now initialize UFFD context and start tracking RAM writes */
3675 if (ram_write_tracking_start()) {
3681 * Start VM from BH handler to avoid write-fault lock here.
3682 * UFFD-WP protection for the whole RAM is already enabled so
3683 * calling VM state change notifiers from vm_start() would initiate
3684 * writes to virtio VQs memory which is in write-protected region.
3686 migration_bh_schedule(bg_migration_vm_start_bh
, s
);
3689 while (migration_is_active()) {
3690 MigIterateState iter_state
= bg_migration_iteration_run(s
);
3691 if (iter_state
== MIG_ITERATE_SKIP
) {
3693 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3698 * Try to detect any kind of failures, and see whether we
3699 * should stop the migration now.
3701 thr_error
= migration_detect_error(s
);
3702 if (thr_error
== MIG_THR_ERR_FATAL
) {
3703 /* Stop migration */
3707 migration_update_counters(s
, qemu_clock_get_ms(QEMU_CLOCK_REALTIME
));
3710 trace_migration_thread_after_loop();
3714 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3715 MIGRATION_STATUS_FAILED
);
3720 bg_migration_iteration_finish(s
);
3723 object_unref(OBJECT(s
));
3724 rcu_unregister_thread();
3729 void migrate_fd_connect(MigrationState
*s
, Error
*error_in
)
3731 Error
*local_err
= NULL
;
3732 uint64_t rate_limit
;
3733 bool resume
= (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
);
3737 * If there's a previous error, free it and prepare for another one.
3738 * Meanwhile if migration completes successfully, there won't have an error
3739 * dumped when calling migrate_fd_cleanup().
3741 migrate_error_free(s
);
3743 s
->expected_downtime
= migrate_downtime_limit();
3745 migrate_fd_error(s
, error_in
);
3748 * Don't do cleanup for resume if channel is invalid, but only dump
3749 * the error. We wait for another channel connect from the user.
3750 * The error_report still gives HMP user a hint on what failed.
3751 * It's normally done in migrate_fd_cleanup(), but call it here
3754 error_report_err(error_copy(s
->error
));
3756 migrate_fd_cleanup(s
);
3762 /* This is a resumed migration */
3763 rate_limit
= migrate_max_postcopy_bandwidth();
3765 /* This is a fresh new migration */
3766 rate_limit
= migrate_max_bandwidth();
3768 /* Notify before starting migration thread */
3769 if (migration_call_notifiers(s
, MIG_EVENT_PRECOPY_SETUP
, &local_err
)) {
3774 migration_rate_set(rate_limit
);
3775 qemu_file_set_blocking(s
->to_dst_file
, true);
3778 * Open the return path. For postcopy, it is used exclusively. For
3779 * precopy, only if user specified "return-path" capability would
3780 * QEMU uses the return path.
3782 if (migrate_postcopy_ram() || migrate_return_path()) {
3783 if (open_return_path_on_source(s
)) {
3784 error_setg(&local_err
, "Unable to open return-path for postcopy");
3790 * This needs to be done before resuming a postcopy. Note: for newer
3791 * QEMUs we will delay the channel creation until postcopy_start(), to
3792 * avoid disorder of channel creations.
3794 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
3795 postcopy_preempt_setup(s
);
3799 /* Wakeup the main migration thread to do the recovery */
3800 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP
,
3801 MIGRATION_STATUS_POSTCOPY_RECOVER
);
3802 qemu_sem_post(&s
->postcopy_pause_sem
);
3806 if (migrate_mode_is_cpr(s
)) {
3807 ret
= migration_stop_vm(s
, RUN_STATE_FINISH_MIGRATE
);
3809 error_setg(&local_err
, "migration_stop_vm failed, error %d", -ret
);
3814 if (migrate_background_snapshot()) {
3815 qemu_thread_create(&s
->thread
, "mig/snapshot",
3816 bg_migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3818 qemu_thread_create(&s
->thread
, "mig/src/main",
3819 migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3821 s
->migration_thread_running
= true;
3825 migrate_set_error(s
, local_err
);
3826 migrate_set_state(&s
->state
, s
->state
, MIGRATION_STATUS_FAILED
);
3827 error_report_err(local_err
);
3828 migrate_fd_cleanup(s
);
3831 static void migration_class_init(ObjectClass
*klass
, void *data
)
3833 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3835 dc
->user_creatable
= false;
3836 device_class_set_props(dc
, migration_properties
);
3839 static void migration_instance_finalize(Object
*obj
)
3841 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3843 qemu_mutex_destroy(&ms
->error_mutex
);
3844 qemu_mutex_destroy(&ms
->qemu_file_lock
);
3845 qemu_sem_destroy(&ms
->wait_unplug_sem
);
3846 qemu_sem_destroy(&ms
->rate_limit_sem
);
3847 qemu_sem_destroy(&ms
->pause_sem
);
3848 qemu_sem_destroy(&ms
->postcopy_pause_sem
);
3849 qemu_sem_destroy(&ms
->rp_state
.rp_sem
);
3850 qemu_sem_destroy(&ms
->rp_state
.rp_pong_acks
);
3851 qemu_sem_destroy(&ms
->postcopy_qemufile_src_sem
);
3852 error_free(ms
->error
);
3855 static void migration_instance_init(Object
*obj
)
3857 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3859 ms
->state
= MIGRATION_STATUS_NONE
;
3861 ms
->pages_per_second
= -1;
3862 qemu_sem_init(&ms
->pause_sem
, 0);
3863 qemu_mutex_init(&ms
->error_mutex
);
3865 migrate_params_init(&ms
->parameters
);
3867 qemu_sem_init(&ms
->postcopy_pause_sem
, 0);
3868 qemu_sem_init(&ms
->rp_state
.rp_sem
, 0);
3869 qemu_sem_init(&ms
->rp_state
.rp_pong_acks
, 0);
3870 qemu_sem_init(&ms
->rate_limit_sem
, 0);
3871 qemu_sem_init(&ms
->wait_unplug_sem
, 0);
3872 qemu_sem_init(&ms
->postcopy_qemufile_src_sem
, 0);
3873 qemu_mutex_init(&ms
->qemu_file_lock
);
3877 * Return true if check pass, false otherwise. Error will be put
3878 * inside errp if provided.
3880 static bool migration_object_check(MigrationState
*ms
, Error
**errp
)
3882 /* Assuming all off */
3883 bool old_caps
[MIGRATION_CAPABILITY__MAX
] = { 0 };
3885 if (!migrate_params_check(&ms
->parameters
, errp
)) {
3889 return migrate_caps_check(old_caps
, ms
->capabilities
, errp
);
3892 static const TypeInfo migration_type
= {
3893 .name
= TYPE_MIGRATION
,
3895 * NOTE: TYPE_MIGRATION is not really a device, as the object is
3896 * not created using qdev_new(), it is not attached to the qdev
3897 * device tree, and it is never realized.
3899 * TODO: Make this TYPE_OBJECT once QOM provides something like
3900 * TYPE_DEVICE's "-global" properties.
3902 .parent
= TYPE_DEVICE
,
3903 .class_init
= migration_class_init
,
3904 .class_size
= sizeof(MigrationClass
),
3905 .instance_size
= sizeof(MigrationState
),
3906 .instance_init
= migration_instance_init
,
3907 .instance_finalize
= migration_instance_finalize
,
3910 static void register_migration_types(void)
3912 type_register_static(&migration_type
);
3915 type_init(register_migration_types
);