4 * Copyright IBM, Corp. 2008
7 * Anthony Liguori <aliguori@us.ibm.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
12 * Contributions after 2012-01-13 are licensed under the terms of the
13 * GNU GPL, version 2 or (at your option) any later version.
16 #include "qemu/osdep.h"
17 #include "qemu/cutils.h"
18 #include "qemu/error-report.h"
19 #include "qemu/main-loop.h"
20 #include "migration/blocker.h"
25 #include "sysemu/runstate.h"
26 #include "sysemu/sysemu.h"
27 #include "sysemu/cpu-throttle.h"
30 #include "ram-compress.h"
31 #include "migration/global_state.h"
32 #include "migration/misc.h"
33 #include "migration.h"
34 #include "migration-stats.h"
36 #include "qemu-file.h"
38 #include "migration/vmstate.h"
39 #include "block/block.h"
40 #include "qapi/error.h"
41 #include "qapi/clone-visitor.h"
42 #include "qapi/qapi-visit-migration.h"
43 #include "qapi/qapi-visit-sockets.h"
44 #include "qapi/qapi-commands-migration.h"
45 #include "qapi/qapi-events-migration.h"
46 #include "qapi/qmp/qerror.h"
47 #include "qapi/qmp/qnull.h"
50 #include "postcopy-ram.h"
51 #include "qemu/thread.h"
53 #include "exec/target_page.h"
54 #include "io/channel-buffer.h"
55 #include "io/channel-tls.h"
56 #include "migration/colo.h"
57 #include "hw/boards.h"
58 #include "monitor/monitor.h"
59 #include "net/announce.h"
60 #include "qemu/queue.h"
62 #include "threadinfo.h"
63 #include "qemu/yank.h"
64 #include "sysemu/cpus.h"
65 #include "yank_functions.h"
66 #include "sysemu/qtest.h"
68 #include "sysemu/dirtylimit.h"
69 #include "qemu/sockets.h"
70 #include "sysemu/kvm.h"
72 #define NOTIFIER_ELEM_INIT(array, elem) \
73 [elem] = NOTIFIER_WITH_RETURN_LIST_INITIALIZER((array)[elem])
75 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 transport_supports_seeking(MigrationAddress
*addr
)
160 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
168 migration_channels_and_transport_compatible(MigrationAddress
*addr
,
171 if (migration_needs_seekable_channel() &&
172 !transport_supports_seeking(addr
)) {
173 error_setg(errp
, "Migration requires seekable transport (e.g. file)");
177 if (migration_needs_multiple_sockets() &&
178 !transport_supports_multi_channels(addr
)) {
179 error_setg(errp
, "Migration requires multi-channel URIs (e.g. tcp)");
186 static gint
page_request_addr_cmp(gconstpointer ap
, gconstpointer bp
)
188 uintptr_t a
= (uintptr_t) ap
, b
= (uintptr_t) bp
;
190 return (a
> b
) - (a
< b
);
193 static int migration_stop_vm(MigrationState
*s
, RunState state
)
197 migration_downtime_start(s
);
199 s
->vm_old_state
= runstate_get();
200 global_state_store();
202 ret
= vm_stop_force_state(state
);
204 trace_vmstate_downtime_checkpoint("src-vm-stopped");
205 trace_migration_completion_vm_stop(ret
);
210 void migration_object_init(void)
212 /* This can only be called once. */
213 assert(!current_migration
);
214 current_migration
= MIGRATION_OBJ(object_new(TYPE_MIGRATION
));
217 * Init the migrate incoming object as well no matter whether
218 * we'll use it or not.
220 assert(!current_incoming
);
221 current_incoming
= g_new0(MigrationIncomingState
, 1);
222 current_incoming
->state
= MIGRATION_STATUS_NONE
;
223 current_incoming
->postcopy_remote_fds
=
224 g_array_new(FALSE
, TRUE
, sizeof(struct PostCopyFD
));
225 qemu_mutex_init(¤t_incoming
->rp_mutex
);
226 qemu_mutex_init(¤t_incoming
->postcopy_prio_thread_mutex
);
227 qemu_event_init(¤t_incoming
->main_thread_load_event
, false);
228 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_dst
, 0);
229 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fault
, 0);
230 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fast_load
, 0);
231 qemu_sem_init(¤t_incoming
->postcopy_qemufile_dst_done
, 0);
233 qemu_mutex_init(¤t_incoming
->page_request_mutex
);
234 qemu_cond_init(¤t_incoming
->page_request_cond
);
235 current_incoming
->page_requested
= g_tree_new(page_request_addr_cmp
);
237 migration_object_check(current_migration
, &error_fatal
);
241 dirty_bitmap_mig_init();
250 static void migration_bh_dispatch_bh(void *opaque
)
252 MigrationState
*s
= migrate_get_current();
253 MigrationBH
*migbh
= opaque
;
255 /* cleanup this BH */
256 qemu_bh_delete(migbh
->bh
);
259 /* dispatch the other one */
260 migbh
->cb(migbh
->opaque
);
261 object_unref(OBJECT(s
));
266 void migration_bh_schedule(QEMUBHFunc
*cb
, void *opaque
)
268 MigrationState
*s
= migrate_get_current();
269 MigrationBH
*migbh
= g_new0(MigrationBH
, 1);
270 QEMUBH
*bh
= qemu_bh_new(migration_bh_dispatch_bh
, migbh
);
272 /* Store these to dispatch when the BH runs */
275 migbh
->opaque
= opaque
;
278 * Ref the state for bh, because it may be called when
279 * there're already no other refs
281 object_ref(OBJECT(s
));
282 qemu_bh_schedule(bh
);
285 void migration_cancel(const Error
*error
)
288 migrate_set_error(current_migration
, error
);
290 if (migrate_dirty_limit()) {
291 qmp_cancel_vcpu_dirty_limit(false, -1, NULL
);
293 migrate_fd_cancel(current_migration
);
296 void migration_shutdown(void)
299 * When the QEMU main thread exit, the COLO thread
300 * may wait a semaphore. So, we should wakeup the
301 * COLO thread before migration shutdown.
305 * Cancel the current migration - that will (eventually)
306 * stop the migration using this structure
308 migration_cancel(NULL
);
309 object_unref(OBJECT(current_migration
));
312 * Cancel outgoing migration of dirty bitmaps. It should
313 * at least unref used block nodes.
315 dirty_bitmap_mig_cancel_outgoing();
318 * Cancel incoming migration of dirty bitmaps. Dirty bitmaps
319 * are non-critical data, and their loss never considered as
322 dirty_bitmap_mig_cancel_incoming();
326 MigrationState
*migrate_get_current(void)
328 /* This can only be called after the object created. */
329 assert(current_migration
);
330 return current_migration
;
333 MigrationIncomingState
*migration_incoming_get_current(void)
335 assert(current_incoming
);
336 return current_incoming
;
339 void migration_incoming_transport_cleanup(MigrationIncomingState
*mis
)
341 if (mis
->socket_address_list
) {
342 qapi_free_SocketAddressList(mis
->socket_address_list
);
343 mis
->socket_address_list
= NULL
;
346 if (mis
->transport_cleanup
) {
347 mis
->transport_cleanup(mis
->transport_data
);
348 mis
->transport_data
= mis
->transport_cleanup
= NULL
;
352 void migration_incoming_state_destroy(void)
354 struct MigrationIncomingState
*mis
= migration_incoming_get_current();
356 multifd_recv_cleanup();
357 compress_threads_load_cleanup();
359 if (mis
->to_src_file
) {
360 /* Tell source that we are done */
361 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
362 qemu_fclose(mis
->to_src_file
);
363 mis
->to_src_file
= NULL
;
366 if (mis
->from_src_file
) {
367 migration_ioc_unregister_yank_from_file(mis
->from_src_file
);
368 qemu_fclose(mis
->from_src_file
);
369 mis
->from_src_file
= NULL
;
371 if (mis
->postcopy_remote_fds
) {
372 g_array_free(mis
->postcopy_remote_fds
, TRUE
);
373 mis
->postcopy_remote_fds
= NULL
;
376 migration_incoming_transport_cleanup(mis
);
377 qemu_event_reset(&mis
->main_thread_load_event
);
379 if (mis
->page_requested
) {
380 g_tree_destroy(mis
->page_requested
);
381 mis
->page_requested
= NULL
;
384 if (mis
->postcopy_qemufile_dst
) {
385 migration_ioc_unregister_yank_from_file(mis
->postcopy_qemufile_dst
);
386 qemu_fclose(mis
->postcopy_qemufile_dst
);
387 mis
->postcopy_qemufile_dst
= NULL
;
390 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
393 static void migrate_generate_event(int new_state
)
395 if (migrate_events()) {
396 qapi_event_send_migration(new_state
);
401 * Send a message on the return channel back to the source
404 static int migrate_send_rp_message(MigrationIncomingState
*mis
,
405 enum mig_rp_message_type message_type
,
406 uint16_t len
, void *data
)
410 trace_migrate_send_rp_message((int)message_type
, len
);
411 QEMU_LOCK_GUARD(&mis
->rp_mutex
);
414 * It's possible that the file handle got lost due to network
417 if (!mis
->to_src_file
) {
422 qemu_put_be16(mis
->to_src_file
, (unsigned int)message_type
);
423 qemu_put_be16(mis
->to_src_file
, len
);
424 qemu_put_buffer(mis
->to_src_file
, data
, len
);
425 return qemu_fflush(mis
->to_src_file
);
428 /* Request one page from the source VM at the given start address.
429 * rb: the RAMBlock to request the page in
430 * Start: Address offset within the RB
431 * Len: Length in bytes required - must be a multiple of pagesize
433 int migrate_send_rp_message_req_pages(MigrationIncomingState
*mis
,
434 RAMBlock
*rb
, ram_addr_t start
)
436 uint8_t bufc
[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
437 size_t msglen
= 12; /* start + len */
438 size_t len
= qemu_ram_pagesize(rb
);
439 enum mig_rp_message_type msg_type
;
443 *(uint64_t *)bufc
= cpu_to_be64((uint64_t)start
);
444 *(uint32_t *)(bufc
+ 8) = cpu_to_be32((uint32_t)len
);
447 * We maintain the last ramblock that we requested for page. Note that we
448 * don't need locking because this function will only be called within the
449 * postcopy ram fault thread.
451 if (rb
!= mis
->last_rb
) {
454 rbname
= qemu_ram_get_idstr(rb
);
455 rbname_len
= strlen(rbname
);
457 assert(rbname_len
< 256);
459 bufc
[msglen
++] = rbname_len
;
460 memcpy(bufc
+ msglen
, rbname
, rbname_len
);
461 msglen
+= rbname_len
;
462 msg_type
= MIG_RP_MSG_REQ_PAGES_ID
;
464 msg_type
= MIG_RP_MSG_REQ_PAGES
;
467 return migrate_send_rp_message(mis
, msg_type
, msglen
, bufc
);
470 int migrate_send_rp_req_pages(MigrationIncomingState
*mis
,
471 RAMBlock
*rb
, ram_addr_t start
, uint64_t haddr
)
473 void *aligned
= (void *)(uintptr_t)ROUND_DOWN(haddr
, qemu_ram_pagesize(rb
));
474 bool received
= false;
476 WITH_QEMU_LOCK_GUARD(&mis
->page_request_mutex
) {
477 received
= ramblock_recv_bitmap_test_byte_offset(rb
, start
);
478 if (!received
&& !g_tree_lookup(mis
->page_requested
, aligned
)) {
480 * The page has not been received, and it's not yet in the page
481 * request list. Queue it. Set the value of element to 1, so that
482 * things like g_tree_lookup() will return TRUE (1) when found.
484 g_tree_insert(mis
->page_requested
, aligned
, (gpointer
)1);
485 qatomic_inc(&mis
->page_requested_count
);
486 trace_postcopy_page_req_add(aligned
, mis
->page_requested_count
);
491 * If the page is there, skip sending the message. We don't even need the
492 * lock because as long as the page arrived, it'll be there forever.
498 return migrate_send_rp_message_req_pages(mis
, rb
, start
);
501 static bool migration_colo_enabled
;
502 bool migration_incoming_colo_enabled(void)
504 return migration_colo_enabled
;
507 void migration_incoming_disable_colo(void)
509 ram_block_discard_disable(false);
510 migration_colo_enabled
= false;
513 int migration_incoming_enable_colo(void)
515 #ifndef CONFIG_REPLICATION
516 error_report("ENABLE_COLO command come in migration stream, but COLO "
517 "module is not built in");
521 if (!migrate_colo()) {
522 error_report("ENABLE_COLO command come in migration stream, but c-colo "
523 "capability is not set");
527 if (ram_block_discard_disable(true)) {
528 error_report("COLO: cannot disable RAM discard");
531 migration_colo_enabled
= true;
535 void migrate_add_address(SocketAddress
*address
)
537 MigrationIncomingState
*mis
= migration_incoming_get_current();
539 QAPI_LIST_PREPEND(mis
->socket_address_list
,
540 QAPI_CLONE(SocketAddress
, address
));
543 bool migrate_uri_parse(const char *uri
, MigrationChannel
**channel
,
546 g_autoptr(MigrationChannel
) val
= g_new0(MigrationChannel
, 1);
547 g_autoptr(MigrationAddress
) addr
= g_new0(MigrationAddress
, 1);
548 InetSocketAddress
*isock
= &addr
->u
.rdma
;
549 strList
**tail
= &addr
->u
.exec
.args
;
551 if (strstart(uri
, "exec:", NULL
)) {
552 addr
->transport
= MIGRATION_ADDRESS_TYPE_EXEC
;
554 QAPI_LIST_APPEND(tail
, g_strdup(exec_get_cmd_path()));
555 QAPI_LIST_APPEND(tail
, g_strdup("/c"));
557 QAPI_LIST_APPEND(tail
, g_strdup("/bin/sh"));
558 QAPI_LIST_APPEND(tail
, g_strdup("-c"));
560 QAPI_LIST_APPEND(tail
, g_strdup(uri
+ strlen("exec:")));
561 } else if (strstart(uri
, "rdma:", NULL
)) {
562 if (inet_parse(isock
, uri
+ strlen("rdma:"), errp
)) {
563 qapi_free_InetSocketAddress(isock
);
566 addr
->transport
= MIGRATION_ADDRESS_TYPE_RDMA
;
567 } else if (strstart(uri
, "tcp:", NULL
) ||
568 strstart(uri
, "unix:", NULL
) ||
569 strstart(uri
, "vsock:", NULL
) ||
570 strstart(uri
, "fd:", NULL
)) {
571 addr
->transport
= MIGRATION_ADDRESS_TYPE_SOCKET
;
572 SocketAddress
*saddr
= socket_parse(uri
, errp
);
576 addr
->u
.socket
.type
= saddr
->type
;
577 addr
->u
.socket
.u
= saddr
->u
;
578 /* Don't free the objects inside; their ownership moved to "addr" */
580 } else if (strstart(uri
, "file:", NULL
)) {
581 addr
->transport
= MIGRATION_ADDRESS_TYPE_FILE
;
582 addr
->u
.file
.filename
= g_strdup(uri
+ strlen("file:"));
583 if (file_parse_offset(addr
->u
.file
.filename
, &addr
->u
.file
.offset
,
588 error_setg(errp
, "unknown migration protocol: %s", uri
);
592 val
->channel_type
= MIGRATION_CHANNEL_TYPE_MAIN
;
593 val
->addr
= g_steal_pointer(&addr
);
594 *channel
= g_steal_pointer(&val
);
598 static void qemu_start_incoming_migration(const char *uri
, bool has_channels
,
599 MigrationChannelList
*channels
,
602 g_autoptr(MigrationChannel
) channel
= NULL
;
603 MigrationAddress
*addr
= NULL
;
604 MigrationIncomingState
*mis
= migration_incoming_get_current();
607 * Having preliminary checks for uri and channel
609 if (!uri
== !channels
) {
610 error_setg(errp
, "need either 'uri' or 'channels' argument");
615 /* To verify that Migrate channel list has only item */
616 if (channels
->next
) {
617 error_setg(errp
, "Channel list has more than one entries");
620 addr
= channels
->value
->addr
;
624 /* caller uses the old URI syntax */
625 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
628 addr
= channel
->addr
;
631 /* transport mechanism not suitable for migration? */
632 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
636 migrate_set_state(&mis
->state
, MIGRATION_STATUS_NONE
,
637 MIGRATION_STATUS_SETUP
);
639 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
640 SocketAddress
*saddr
= &addr
->u
.socket
;
641 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
642 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
643 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
644 socket_start_incoming_migration(saddr
, errp
);
645 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
646 fd_start_incoming_migration(saddr
->u
.fd
.str
, errp
);
649 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
650 if (migrate_compress()) {
651 error_setg(errp
, "RDMA and compression can't be used together");
654 if (migrate_xbzrle()) {
655 error_setg(errp
, "RDMA and XBZRLE can't be used together");
658 if (migrate_multifd()) {
659 error_setg(errp
, "RDMA and multifd can't be used together");
662 rdma_start_incoming_migration(&addr
->u
.rdma
, errp
);
664 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
665 exec_start_incoming_migration(addr
->u
.exec
.args
, errp
);
666 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
667 file_start_incoming_migration(&addr
->u
.file
, errp
);
669 error_setg(errp
, "unknown migration protocol: %s", uri
);
673 static void process_incoming_migration_bh(void *opaque
)
675 Error
*local_err
= NULL
;
676 MigrationIncomingState
*mis
= opaque
;
678 trace_vmstate_downtime_checkpoint("dst-precopy-bh-enter");
680 /* If capability late_block_activate is set:
681 * Only fire up the block code now if we're going to restart the
682 * VM, else 'cont' will do it.
683 * This causes file locking to happen; so we don't want it to happen
684 * unless we really are starting the VM.
686 if (!migrate_late_block_activate() ||
687 (autostart
&& (!global_state_received() ||
688 runstate_is_live(global_state_get_runstate())))) {
689 /* Make sure all file formats throw away their mutable metadata.
690 * If we get an error here, just don't restart the VM yet. */
691 bdrv_activate_all(&local_err
);
693 error_report_err(local_err
);
700 * This must happen after all error conditions are dealt with and
701 * we're sure the VM is going to be running on this host.
703 qemu_announce_self(&mis
->announce_timer
, migrate_announce_params());
705 trace_vmstate_downtime_checkpoint("dst-precopy-bh-announced");
707 multifd_recv_shutdown();
709 dirty_bitmap_mig_before_vm_start();
711 if (!global_state_received() ||
712 runstate_is_live(global_state_get_runstate())) {
716 runstate_set(RUN_STATE_PAUSED
);
718 } else if (migration_incoming_colo_enabled()) {
719 migration_incoming_disable_colo();
722 runstate_set(global_state_get_runstate());
724 trace_vmstate_downtime_checkpoint("dst-precopy-bh-vm-started");
726 * This must happen after any state changes since as soon as an external
727 * observer sees this event they might start to prod at the VM assuming
730 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
731 MIGRATION_STATUS_COMPLETED
);
732 migration_incoming_state_destroy();
735 static void coroutine_fn
736 process_incoming_migration_co(void *opaque
)
738 MigrationIncomingState
*mis
= migration_incoming_get_current();
742 assert(mis
->from_src_file
);
744 if (compress_threads_load_setup(mis
->from_src_file
)) {
745 error_report("Failed to setup decompress threads");
749 mis
->largest_page_size
= qemu_ram_pagesize_largest();
750 postcopy_state_set(POSTCOPY_INCOMING_NONE
);
751 migrate_set_state(&mis
->state
, MIGRATION_STATUS_SETUP
,
752 MIGRATION_STATUS_ACTIVE
);
754 mis
->loadvm_co
= qemu_coroutine_self();
755 ret
= qemu_loadvm_state(mis
->from_src_file
);
756 mis
->loadvm_co
= NULL
;
758 trace_vmstate_downtime_checkpoint("dst-precopy-loadvm-completed");
760 ps
= postcopy_state_get();
761 trace_process_incoming_migration_co_end(ret
, ps
);
762 if (ps
!= POSTCOPY_INCOMING_NONE
) {
763 if (ps
== POSTCOPY_INCOMING_ADVISE
) {
765 * Where a migration had postcopy enabled (and thus went to advise)
766 * but managed to complete within the precopy period, we can use
769 postcopy_ram_incoming_cleanup(mis
);
770 } else if (ret
>= 0) {
772 * Postcopy was started, cleanup should happen at the end of the
775 trace_process_incoming_migration_co_postcopy_end_main();
778 /* Else if something went wrong then just fall out of the normal exit */
782 MigrationState
*s
= migrate_get_current();
784 if (migrate_has_error(s
)) {
785 WITH_QEMU_LOCK_GUARD(&s
->error_mutex
) {
786 error_report_err(s
->error
);
789 error_report("load of migration failed: %s", strerror(-ret
));
793 if (colo_incoming_co() < 0) {
797 migration_bh_schedule(process_incoming_migration_bh
, mis
);
800 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
801 MIGRATION_STATUS_FAILED
);
802 qemu_fclose(mis
->from_src_file
);
804 multifd_recv_cleanup();
805 compress_threads_load_cleanup();
811 * migration_incoming_setup: Setup incoming migration
812 * @f: file for main migration channel
814 static void migration_incoming_setup(QEMUFile
*f
)
816 MigrationIncomingState
*mis
= migration_incoming_get_current();
818 if (!mis
->from_src_file
) {
819 mis
->from_src_file
= f
;
821 qemu_file_set_blocking(f
, false);
824 void migration_incoming_process(void)
826 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
, NULL
);
827 qemu_coroutine_enter(co
);
830 /* Returns true if recovered from a paused migration, otherwise false */
831 static bool postcopy_try_recover(void)
833 MigrationIncomingState
*mis
= migration_incoming_get_current();
835 if (mis
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
836 /* Resumed from a paused postcopy migration */
838 /* This should be set already in migration_incoming_setup() */
839 assert(mis
->from_src_file
);
840 /* Postcopy has standalone thread to do vm load */
841 qemu_file_set_blocking(mis
->from_src_file
, true);
843 /* Re-configure the return path */
844 mis
->to_src_file
= qemu_file_get_return_path(mis
->from_src_file
);
846 migrate_set_state(&mis
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
847 MIGRATION_STATUS_POSTCOPY_RECOVER
);
850 * Here, we only wake up the main loading thread (while the
851 * rest threads will still be waiting), so that we can receive
852 * commands from source now, and answer it if needed. The
853 * rest threads will be woken up afterwards until we are sure
854 * that source is ready to reply to page requests.
856 qemu_sem_post(&mis
->postcopy_pause_sem_dst
);
863 void migration_fd_process_incoming(QEMUFile
*f
)
865 migration_incoming_setup(f
);
866 if (postcopy_try_recover()) {
869 migration_incoming_process();
873 * Returns true when we want to start a new incoming migration process,
876 static bool migration_should_start_incoming(bool main_channel
)
878 /* Multifd doesn't start unless all channels are established */
879 if (migrate_multifd()) {
880 return migration_has_all_channels();
883 /* Preempt channel only starts when the main channel is created */
884 if (migrate_postcopy_preempt()) {
889 * For all the rest types of migration, we should only reach here when
890 * it's the main channel that's being created, and we should always
891 * proceed with this channel.
893 assert(main_channel
);
897 void migration_ioc_process_incoming(QIOChannel
*ioc
, Error
**errp
)
899 MigrationIncomingState
*mis
= migration_incoming_get_current();
900 Error
*local_err
= NULL
;
902 bool default_channel
= true;
903 uint32_t channel_magic
= 0;
906 if (migrate_multifd() && !migrate_mapped_ram() &&
907 !migrate_postcopy_ram() &&
908 qio_channel_has_feature(ioc
, QIO_CHANNEL_FEATURE_READ_MSG_PEEK
)) {
910 * With multiple channels, it is possible that we receive channels
911 * out of order on destination side, causing incorrect mapping of
912 * source channels on destination side. Check channel MAGIC to
913 * decide type of channel. Please note this is best effort, postcopy
914 * preempt channel does not send any magic number so avoid it for
915 * postcopy live migration. Also tls live migration already does
916 * tls handshake while initializing main channel so with tls this
917 * issue is not possible.
919 ret
= migration_channel_read_peek(ioc
, (void *)&channel_magic
,
920 sizeof(channel_magic
), errp
);
926 default_channel
= (channel_magic
== cpu_to_be32(QEMU_VM_FILE_MAGIC
));
928 default_channel
= !mis
->from_src_file
;
931 if (multifd_recv_setup(errp
) != 0) {
935 if (default_channel
) {
936 f
= qemu_file_new_input(ioc
);
937 migration_incoming_setup(f
);
939 /* Multiple connections */
940 assert(migration_needs_multiple_sockets());
941 if (migrate_multifd()) {
942 multifd_recv_new_channel(ioc
, &local_err
);
944 assert(migrate_postcopy_preempt());
945 f
= qemu_file_new_input(ioc
);
946 postcopy_preempt_new_channel(mis
, f
);
949 error_propagate(errp
, local_err
);
954 if (migration_should_start_incoming(default_channel
)) {
955 /* If it's a recovery, we're done */
956 if (postcopy_try_recover()) {
959 migration_incoming_process();
964 * @migration_has_all_channels: We have received all channels that we need
966 * Returns true when we have got connections to all the channels that
967 * we need for migration.
969 bool migration_has_all_channels(void)
971 MigrationIncomingState
*mis
= migration_incoming_get_current();
973 if (!mis
->from_src_file
) {
977 if (migrate_multifd()) {
978 return multifd_recv_all_channels_created();
981 if (migrate_postcopy_preempt()) {
982 return mis
->postcopy_qemufile_dst
!= NULL
;
988 int migrate_send_rp_switchover_ack(MigrationIncomingState
*mis
)
990 return migrate_send_rp_message(mis
, MIG_RP_MSG_SWITCHOVER_ACK
, 0, NULL
);
994 * Send a 'SHUT' message on the return channel with the given value
995 * to indicate that we've finished with the RP. Non-0 value indicates
998 void migrate_send_rp_shut(MigrationIncomingState
*mis
,
1003 buf
= cpu_to_be32(value
);
1004 migrate_send_rp_message(mis
, MIG_RP_MSG_SHUT
, sizeof(buf
), &buf
);
1008 * Send a 'PONG' message on the return channel with the given value
1009 * (normally in response to a 'PING')
1011 void migrate_send_rp_pong(MigrationIncomingState
*mis
,
1016 buf
= cpu_to_be32(value
);
1017 migrate_send_rp_message(mis
, MIG_RP_MSG_PONG
, sizeof(buf
), &buf
);
1020 void migrate_send_rp_recv_bitmap(MigrationIncomingState
*mis
,
1028 * First, we send the header part. It contains only the len of
1029 * idstr, and the idstr itself.
1031 len
= strlen(block_name
);
1033 memcpy(buf
+ 1, block_name
, len
);
1035 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
1036 error_report("%s: MSG_RP_RECV_BITMAP only used for recovery",
1041 migrate_send_rp_message(mis
, MIG_RP_MSG_RECV_BITMAP
, len
+ 1, buf
);
1044 * Next, we dump the received bitmap to the stream.
1046 * TODO: currently we are safe since we are the only one that is
1047 * using the to_src_file handle (fault thread is still paused),
1048 * and it's ok even not taking the mutex. However the best way is
1049 * to take the lock before sending the message header, and release
1050 * the lock after sending the bitmap.
1052 qemu_mutex_lock(&mis
->rp_mutex
);
1053 res
= ramblock_recv_bitmap_send(mis
->to_src_file
, block_name
);
1054 qemu_mutex_unlock(&mis
->rp_mutex
);
1056 trace_migrate_send_rp_recv_bitmap(block_name
, res
);
1059 void migrate_send_rp_resume_ack(MigrationIncomingState
*mis
, uint32_t value
)
1063 buf
= cpu_to_be32(value
);
1064 migrate_send_rp_message(mis
, MIG_RP_MSG_RESUME_ACK
, sizeof(buf
), &buf
);
1068 * Return true if we're already in the middle of a migration
1069 * (i.e. any of the active or setup states)
1071 bool migration_is_setup_or_active(void)
1073 MigrationState
*s
= current_migration
;
1076 case MIGRATION_STATUS_ACTIVE
:
1077 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1078 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1079 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1080 case MIGRATION_STATUS_SETUP
:
1081 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1082 case MIGRATION_STATUS_DEVICE
:
1083 case MIGRATION_STATUS_WAIT_UNPLUG
:
1084 case MIGRATION_STATUS_COLO
:
1093 bool migration_is_running(void)
1095 MigrationState
*s
= current_migration
;
1098 case MIGRATION_STATUS_ACTIVE
:
1099 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1100 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1101 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1102 case MIGRATION_STATUS_SETUP
:
1103 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1104 case MIGRATION_STATUS_DEVICE
:
1105 case MIGRATION_STATUS_WAIT_UNPLUG
:
1106 case MIGRATION_STATUS_CANCELLING
:
1115 static bool migrate_show_downtime(MigrationState
*s
)
1117 return (s
->state
== MIGRATION_STATUS_COMPLETED
) || migration_in_postcopy();
1120 static void populate_time_info(MigrationInfo
*info
, MigrationState
*s
)
1122 info
->has_status
= true;
1123 info
->has_setup_time
= true;
1124 info
->setup_time
= s
->setup_time
;
1126 if (s
->state
== MIGRATION_STATUS_COMPLETED
) {
1127 info
->has_total_time
= true;
1128 info
->total_time
= s
->total_time
;
1130 info
->has_total_time
= true;
1131 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) -
1135 if (migrate_show_downtime(s
)) {
1136 info
->has_downtime
= true;
1137 info
->downtime
= s
->downtime
;
1139 info
->has_expected_downtime
= true;
1140 info
->expected_downtime
= s
->expected_downtime
;
1144 static void populate_ram_info(MigrationInfo
*info
, MigrationState
*s
)
1146 size_t page_size
= qemu_target_page_size();
1148 info
->ram
= g_malloc0(sizeof(*info
->ram
));
1149 info
->ram
->transferred
= migration_transferred_bytes();
1150 info
->ram
->total
= ram_bytes_total();
1151 info
->ram
->duplicate
= stat64_get(&mig_stats
.zero_pages
);
1152 /* legacy value. It is not used anymore */
1153 info
->ram
->skipped
= 0;
1154 info
->ram
->normal
= stat64_get(&mig_stats
.normal_pages
);
1155 info
->ram
->normal_bytes
= info
->ram
->normal
* page_size
;
1156 info
->ram
->mbps
= s
->mbps
;
1157 info
->ram
->dirty_sync_count
=
1158 stat64_get(&mig_stats
.dirty_sync_count
);
1159 info
->ram
->dirty_sync_missed_zero_copy
=
1160 stat64_get(&mig_stats
.dirty_sync_missed_zero_copy
);
1161 info
->ram
->postcopy_requests
=
1162 stat64_get(&mig_stats
.postcopy_requests
);
1163 info
->ram
->page_size
= page_size
;
1164 info
->ram
->multifd_bytes
= stat64_get(&mig_stats
.multifd_bytes
);
1165 info
->ram
->pages_per_second
= s
->pages_per_second
;
1166 info
->ram
->precopy_bytes
= stat64_get(&mig_stats
.precopy_bytes
);
1167 info
->ram
->downtime_bytes
= stat64_get(&mig_stats
.downtime_bytes
);
1168 info
->ram
->postcopy_bytes
= stat64_get(&mig_stats
.postcopy_bytes
);
1170 if (migrate_xbzrle()) {
1171 info
->xbzrle_cache
= g_malloc0(sizeof(*info
->xbzrle_cache
));
1172 info
->xbzrle_cache
->cache_size
= migrate_xbzrle_cache_size();
1173 info
->xbzrle_cache
->bytes
= xbzrle_counters
.bytes
;
1174 info
->xbzrle_cache
->pages
= xbzrle_counters
.pages
;
1175 info
->xbzrle_cache
->cache_miss
= xbzrle_counters
.cache_miss
;
1176 info
->xbzrle_cache
->cache_miss_rate
= xbzrle_counters
.cache_miss_rate
;
1177 info
->xbzrle_cache
->encoding_rate
= xbzrle_counters
.encoding_rate
;
1178 info
->xbzrle_cache
->overflow
= xbzrle_counters
.overflow
;
1181 populate_compress(info
);
1183 if (cpu_throttle_active()) {
1184 info
->has_cpu_throttle_percentage
= true;
1185 info
->cpu_throttle_percentage
= cpu_throttle_get_percentage();
1188 if (s
->state
!= MIGRATION_STATUS_COMPLETED
) {
1189 info
->ram
->remaining
= ram_bytes_remaining();
1190 info
->ram
->dirty_pages_rate
=
1191 stat64_get(&mig_stats
.dirty_pages_rate
);
1194 if (migrate_dirty_limit() && dirtylimit_in_service()) {
1195 info
->has_dirty_limit_throttle_time_per_round
= true;
1196 info
->dirty_limit_throttle_time_per_round
=
1197 dirtylimit_throttle_time_per_round();
1199 info
->has_dirty_limit_ring_full_time
= true;
1200 info
->dirty_limit_ring_full_time
= dirtylimit_ring_full_time();
1204 static void populate_disk_info(MigrationInfo
*info
)
1206 if (blk_mig_active()) {
1207 info
->disk
= g_malloc0(sizeof(*info
->disk
));
1208 info
->disk
->transferred
= blk_mig_bytes_transferred();
1209 info
->disk
->remaining
= blk_mig_bytes_remaining();
1210 info
->disk
->total
= blk_mig_bytes_total();
1214 static void fill_source_migration_info(MigrationInfo
*info
)
1216 MigrationState
*s
= migrate_get_current();
1217 int state
= qatomic_read(&s
->state
);
1218 GSList
*cur_blocker
= migration_blockers
[migrate_mode()];
1220 info
->blocked_reasons
= NULL
;
1223 * There are two types of reasons a migration might be blocked;
1224 * a) devices marked in VMState as non-migratable, and
1225 * b) Explicit migration blockers
1226 * We need to add both of them here.
1228 qemu_savevm_non_migratable_list(&info
->blocked_reasons
);
1230 while (cur_blocker
) {
1231 QAPI_LIST_PREPEND(info
->blocked_reasons
,
1232 g_strdup(error_get_pretty(cur_blocker
->data
)));
1233 cur_blocker
= g_slist_next(cur_blocker
);
1235 info
->has_blocked_reasons
= info
->blocked_reasons
!= NULL
;
1238 case MIGRATION_STATUS_NONE
:
1239 /* no migration has happened ever */
1240 /* do not overwrite destination migration status */
1242 case MIGRATION_STATUS_SETUP
:
1243 info
->has_status
= true;
1244 info
->has_total_time
= false;
1246 case MIGRATION_STATUS_ACTIVE
:
1247 case MIGRATION_STATUS_CANCELLING
:
1248 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1249 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1250 case MIGRATION_STATUS_DEVICE
:
1251 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1252 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1253 /* TODO add some postcopy stats */
1254 populate_time_info(info
, s
);
1255 populate_ram_info(info
, s
);
1256 populate_disk_info(info
);
1257 migration_populate_vfio_info(info
);
1259 case MIGRATION_STATUS_COLO
:
1260 info
->has_status
= true;
1261 /* TODO: display COLO specific information (checkpoint info etc.) */
1263 case MIGRATION_STATUS_COMPLETED
:
1264 populate_time_info(info
, s
);
1265 populate_ram_info(info
, s
);
1266 migration_populate_vfio_info(info
);
1268 case MIGRATION_STATUS_FAILED
:
1269 info
->has_status
= true;
1271 case MIGRATION_STATUS_CANCELLED
:
1272 info
->has_status
= true;
1274 case MIGRATION_STATUS_WAIT_UNPLUG
:
1275 info
->has_status
= true;
1278 info
->status
= state
;
1280 QEMU_LOCK_GUARD(&s
->error_mutex
);
1282 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1286 static void fill_destination_migration_info(MigrationInfo
*info
)
1288 MigrationIncomingState
*mis
= migration_incoming_get_current();
1290 if (mis
->socket_address_list
) {
1291 info
->has_socket_address
= true;
1292 info
->socket_address
=
1293 QAPI_CLONE(SocketAddressList
, mis
->socket_address_list
);
1296 switch (mis
->state
) {
1297 case MIGRATION_STATUS_NONE
:
1299 case MIGRATION_STATUS_SETUP
:
1300 case MIGRATION_STATUS_CANCELLING
:
1301 case MIGRATION_STATUS_CANCELLED
:
1302 case MIGRATION_STATUS_ACTIVE
:
1303 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1304 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1305 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1306 case MIGRATION_STATUS_FAILED
:
1307 case MIGRATION_STATUS_COLO
:
1308 info
->has_status
= true;
1310 case MIGRATION_STATUS_COMPLETED
:
1311 info
->has_status
= true;
1312 fill_destination_postcopy_migration_info(info
);
1315 info
->status
= mis
->state
;
1318 MigrationInfo
*qmp_query_migrate(Error
**errp
)
1320 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
1322 fill_destination_migration_info(info
);
1323 fill_source_migration_info(info
);
1328 void qmp_migrate_start_postcopy(Error
**errp
)
1330 MigrationState
*s
= migrate_get_current();
1332 if (!migrate_postcopy()) {
1333 error_setg(errp
, "Enable postcopy with migrate_set_capability before"
1334 " the start of migration");
1338 if (s
->state
== MIGRATION_STATUS_NONE
) {
1339 error_setg(errp
, "Postcopy must be started after migration has been"
1344 * we don't error if migration has finished since that would be racy
1345 * with issuing this command.
1347 qatomic_set(&s
->start_postcopy
, true);
1350 /* shared migration helpers */
1352 void migrate_set_state(int *state
, int old_state
, int new_state
)
1354 assert(new_state
< MIGRATION_STATUS__MAX
);
1355 if (qatomic_cmpxchg(state
, old_state
, new_state
) == old_state
) {
1356 trace_migrate_set_state(MigrationStatus_str(new_state
));
1357 migrate_generate_event(new_state
);
1361 static void migrate_fd_cleanup(MigrationState
*s
)
1363 MigrationEventType type
;
1365 g_free(s
->hostname
);
1367 json_writer_free(s
->vmdesc
);
1370 qemu_savevm_state_cleanup();
1372 close_return_path_on_source(s
);
1374 if (s
->to_dst_file
) {
1377 trace_migrate_fd_cleanup();
1379 if (s
->migration_thread_running
) {
1380 qemu_thread_join(&s
->thread
);
1381 s
->migration_thread_running
= false;
1385 multifd_send_shutdown();
1386 qemu_mutex_lock(&s
->qemu_file_lock
);
1387 tmp
= s
->to_dst_file
;
1388 s
->to_dst_file
= NULL
;
1389 qemu_mutex_unlock(&s
->qemu_file_lock
);
1391 * Close the file handle without the lock to make sure the
1392 * critical section won't block for long.
1394 migration_ioc_unregister_yank_from_file(tmp
);
1398 assert(!migration_is_active());
1400 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1401 migrate_set_state(&s
->state
, MIGRATION_STATUS_CANCELLING
,
1402 MIGRATION_STATUS_CANCELLED
);
1406 /* It is used on info migrate. We can't free it */
1407 error_report_err(error_copy(s
->error
));
1409 type
= migration_has_failed(s
) ? MIG_EVENT_PRECOPY_FAILED
:
1410 MIG_EVENT_PRECOPY_DONE
;
1411 migration_call_notifiers(s
, type
, NULL
);
1412 block_cleanup_parameters();
1413 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1416 static void migrate_fd_cleanup_bh(void *opaque
)
1418 migrate_fd_cleanup(opaque
);
1421 void migrate_set_error(MigrationState
*s
, const Error
*error
)
1423 QEMU_LOCK_GUARD(&s
->error_mutex
);
1425 trace_migrate_error(error_get_pretty(error
));
1428 s
->error
= error_copy(error
);
1432 bool migrate_has_error(MigrationState
*s
)
1434 /* The lock is not helpful here, but still follow the rule */
1435 QEMU_LOCK_GUARD(&s
->error_mutex
);
1436 return qatomic_read(&s
->error
);
1439 static void migrate_error_free(MigrationState
*s
)
1441 QEMU_LOCK_GUARD(&s
->error_mutex
);
1443 error_free(s
->error
);
1448 static void migrate_fd_error(MigrationState
*s
, const Error
*error
)
1450 assert(s
->to_dst_file
== NULL
);
1451 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1452 MIGRATION_STATUS_FAILED
);
1453 migrate_set_error(s
, error
);
1456 static void migrate_fd_cancel(MigrationState
*s
)
1460 trace_migrate_fd_cancel();
1462 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1463 if (s
->rp_state
.from_dst_file
) {
1464 /* shutdown the rp socket, so causing the rp thread to shutdown */
1465 qemu_file_shutdown(s
->rp_state
.from_dst_file
);
1470 old_state
= s
->state
;
1471 if (!migration_is_running()) {
1474 /* If the migration is paused, kick it out of the pause */
1475 if (old_state
== MIGRATION_STATUS_PRE_SWITCHOVER
) {
1476 qemu_sem_post(&s
->pause_sem
);
1478 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_CANCELLING
);
1479 } while (s
->state
!= MIGRATION_STATUS_CANCELLING
);
1482 * If we're unlucky the migration code might be stuck somewhere in a
1483 * send/write while the network has failed and is waiting to timeout;
1484 * if we've got shutdown(2) available then we can force it to quit.
1486 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1487 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1488 if (s
->to_dst_file
) {
1489 qemu_file_shutdown(s
->to_dst_file
);
1493 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& s
->block_inactive
) {
1494 Error
*local_err
= NULL
;
1496 bdrv_activate_all(&local_err
);
1498 error_report_err(local_err
);
1500 s
->block_inactive
= false;
1505 void migration_add_notifier_mode(NotifierWithReturn
*notify
,
1506 MigrationNotifyFunc func
, MigMode mode
)
1508 notify
->notify
= (NotifierWithReturnFunc
)func
;
1509 notifier_with_return_list_add(&migration_state_notifiers
[mode
], notify
);
1512 void migration_add_notifier(NotifierWithReturn
*notify
,
1513 MigrationNotifyFunc func
)
1515 migration_add_notifier_mode(notify
, func
, MIG_MODE_NORMAL
);
1518 void migration_remove_notifier(NotifierWithReturn
*notify
)
1520 if (notify
->notify
) {
1521 notifier_with_return_remove(notify
);
1522 notify
->notify
= NULL
;
1526 int migration_call_notifiers(MigrationState
*s
, MigrationEventType type
,
1529 MigMode mode
= s
->parameters
.mode
;
1534 ret
= notifier_with_return_list_notify(&migration_state_notifiers
[mode
],
1536 assert(!ret
|| type
== MIG_EVENT_PRECOPY_SETUP
);
1540 bool migration_has_failed(MigrationState
*s
)
1542 return (s
->state
== MIGRATION_STATUS_CANCELLED
||
1543 s
->state
== MIGRATION_STATUS_FAILED
);
1546 bool migration_in_postcopy(void)
1548 MigrationState
*s
= migrate_get_current();
1551 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1552 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1553 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1560 bool migration_postcopy_is_alive(int state
)
1563 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1564 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1571 bool migration_in_incoming_postcopy(void)
1573 PostcopyState ps
= postcopy_state_get();
1575 return ps
>= POSTCOPY_INCOMING_DISCARD
&& ps
< POSTCOPY_INCOMING_END
;
1578 bool migration_incoming_postcopy_advised(void)
1580 PostcopyState ps
= postcopy_state_get();
1582 return ps
>= POSTCOPY_INCOMING_ADVISE
&& ps
< POSTCOPY_INCOMING_END
;
1585 bool migration_in_bg_snapshot(void)
1587 return migrate_background_snapshot() &&
1588 migration_is_setup_or_active();
1591 bool migration_is_idle(void)
1593 MigrationState
*s
= current_migration
;
1600 case MIGRATION_STATUS_NONE
:
1601 case MIGRATION_STATUS_CANCELLED
:
1602 case MIGRATION_STATUS_COMPLETED
:
1603 case MIGRATION_STATUS_FAILED
:
1605 case MIGRATION_STATUS_SETUP
:
1606 case MIGRATION_STATUS_CANCELLING
:
1607 case MIGRATION_STATUS_ACTIVE
:
1608 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1609 case MIGRATION_STATUS_COLO
:
1610 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1611 case MIGRATION_STATUS_DEVICE
:
1612 case MIGRATION_STATUS_WAIT_UNPLUG
:
1614 case MIGRATION_STATUS__MAX
:
1615 g_assert_not_reached();
1621 bool migration_is_active(void)
1623 MigrationState
*s
= current_migration
;
1625 return (s
->state
== MIGRATION_STATUS_ACTIVE
||
1626 s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1629 bool migration_is_device(void)
1631 MigrationState
*s
= current_migration
;
1633 return s
->state
== MIGRATION_STATUS_DEVICE
;
1636 bool migration_thread_is_self(void)
1638 MigrationState
*s
= current_migration
;
1640 return qemu_thread_is_self(&s
->thread
);
1643 bool migrate_mode_is_cpr(MigrationState
*s
)
1645 return s
->parameters
.mode
== MIG_MODE_CPR_REBOOT
;
1648 int migrate_init(MigrationState
*s
, Error
**errp
)
1652 ret
= qemu_savevm_state_prepare(errp
);
1658 * Reinitialise all migration state, except
1659 * parameters/capabilities that the user set, and
1662 s
->to_dst_file
= NULL
;
1663 s
->state
= MIGRATION_STATUS_NONE
;
1664 s
->rp_state
.from_dst_file
= NULL
;
1666 s
->pages_per_second
= 0.0;
1668 s
->expected_downtime
= 0;
1670 s
->start_postcopy
= false;
1671 s
->migration_thread_running
= false;
1672 error_free(s
->error
);
1676 migrate_set_state(&s
->state
, MIGRATION_STATUS_NONE
, MIGRATION_STATUS_SETUP
);
1678 s
->start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1680 s
->vm_old_state
= -1;
1681 s
->iteration_initial_bytes
= 0;
1682 s
->threshold_size
= 0;
1683 s
->switchover_acked
= false;
1684 s
->rdma_migration
= false;
1686 * set mig_stats memory to zero for a new migration
1688 memset(&mig_stats
, 0, sizeof(mig_stats
));
1689 migration_reset_vfio_bytes_transferred();
1694 static bool is_busy(Error
**reasonp
, Error
**errp
)
1698 /* Snapshots are similar to migrations, so check RUN_STATE_SAVE_VM too. */
1699 if (runstate_check(RUN_STATE_SAVE_VM
) || !migration_is_idle()) {
1700 error_propagate_prepend(errp
, *reasonp
,
1701 "disallowing migration blocker "
1702 "(migration/snapshot in progress) for: ");
1709 static bool is_only_migratable(Error
**reasonp
, Error
**errp
, int modes
)
1713 if (only_migratable
&& (modes
& BIT(MIG_MODE_NORMAL
))) {
1714 error_propagate_prepend(errp
, *reasonp
,
1715 "disallowing migration blocker "
1716 "(--only-migratable) for: ");
1723 static int get_modes(MigMode mode
, va_list ap
)
1727 while (mode
!= -1 && mode
!= MIG_MODE_ALL
) {
1728 assert(mode
>= MIG_MODE_NORMAL
&& mode
< MIG_MODE__MAX
);
1730 mode
= va_arg(ap
, MigMode
);
1732 if (mode
== MIG_MODE_ALL
) {
1733 modes
= BIT(MIG_MODE__MAX
) - 1;
1738 static int add_blockers(Error
**reasonp
, Error
**errp
, int modes
)
1740 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1741 if (modes
& BIT(mode
)) {
1742 migration_blockers
[mode
] = g_slist_prepend(migration_blockers
[mode
],
1749 int migrate_add_blocker(Error
**reasonp
, Error
**errp
)
1751 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_ALL
);
1754 int migrate_add_blocker_normal(Error
**reasonp
, Error
**errp
)
1756 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_NORMAL
, -1);
1759 int migrate_add_blocker_modes(Error
**reasonp
, Error
**errp
, MigMode mode
, ...)
1765 modes
= get_modes(mode
, ap
);
1768 if (is_only_migratable(reasonp
, errp
, modes
)) {
1770 } else if (is_busy(reasonp
, errp
)) {
1773 return add_blockers(reasonp
, errp
, modes
);
1776 int migrate_add_blocker_internal(Error
**reasonp
, Error
**errp
)
1778 int modes
= BIT(MIG_MODE__MAX
) - 1;
1780 if (is_busy(reasonp
, errp
)) {
1783 return add_blockers(reasonp
, errp
, modes
);
1786 void migrate_del_blocker(Error
**reasonp
)
1789 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1790 migration_blockers
[mode
] = g_slist_remove(migration_blockers
[mode
],
1793 error_free(*reasonp
);
1798 void qmp_migrate_incoming(const char *uri
, bool has_channels
,
1799 MigrationChannelList
*channels
, Error
**errp
)
1801 Error
*local_err
= NULL
;
1802 static bool once
= true;
1805 error_setg(errp
, "The incoming migration has already been started");
1808 if (!runstate_check(RUN_STATE_INMIGRATE
)) {
1809 error_setg(errp
, "'-incoming' was not specified on the command line");
1813 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1817 qemu_start_incoming_migration(uri
, has_channels
, channels
, &local_err
);
1820 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1821 error_propagate(errp
, local_err
);
1828 void qmp_migrate_recover(const char *uri
, Error
**errp
)
1830 MigrationIncomingState
*mis
= migration_incoming_get_current();
1833 * Don't even bother to use ERRP_GUARD() as it _must_ always be set by
1834 * callers (no one should ignore a recover failure); if there is, it's a
1835 * programming error.
1839 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1840 error_setg(errp
, "Migrate recover can only be run "
1841 "when postcopy is paused.");
1845 /* If there's an existing transport, release it */
1846 migration_incoming_transport_cleanup(mis
);
1849 * Note that this call will never start a real migration; it will
1850 * only re-setup the migration stream and poke existing migration
1851 * to continue using that newly established channel.
1853 qemu_start_incoming_migration(uri
, false, NULL
, errp
);
1856 void qmp_migrate_pause(Error
**errp
)
1858 MigrationState
*ms
= migrate_get_current();
1859 MigrationIncomingState
*mis
= migration_incoming_get_current();
1862 if (migration_postcopy_is_alive(ms
->state
)) {
1863 /* Source side, during postcopy */
1864 Error
*error
= NULL
;
1866 /* Tell the core migration that we're pausing */
1867 error_setg(&error
, "Postcopy migration is paused by the user");
1868 migrate_set_error(ms
, error
);
1871 qemu_mutex_lock(&ms
->qemu_file_lock
);
1872 if (ms
->to_dst_file
) {
1873 ret
= qemu_file_shutdown(ms
->to_dst_file
);
1875 qemu_mutex_unlock(&ms
->qemu_file_lock
);
1877 error_setg(errp
, "Failed to pause source migration");
1881 * Kick the migration thread out of any waiting windows (on behalf
1882 * of the rp thread).
1884 migration_rp_kick(ms
);
1889 if (migration_postcopy_is_alive(mis
->state
)) {
1890 ret
= qemu_file_shutdown(mis
->from_src_file
);
1892 error_setg(errp
, "Failed to pause destination migration");
1897 error_setg(errp
, "migrate-pause is currently only supported "
1898 "during postcopy-active or postcopy-recover state");
1901 bool migration_is_blocked(Error
**errp
)
1903 GSList
*blockers
= migration_blockers
[migrate_mode()];
1905 if (qemu_savevm_state_blocked(errp
)) {
1910 error_propagate(errp
, error_copy(blockers
->data
));
1917 /* Returns true if continue to migrate, or false if error detected */
1918 static bool migrate_prepare(MigrationState
*s
, bool blk
, bool blk_inc
,
1919 bool resume
, Error
**errp
)
1922 warn_report("parameter 'inc' is deprecated;"
1923 " use blockdev-mirror with NBD instead");
1927 warn_report("parameter 'blk' is deprecated;"
1928 " use blockdev-mirror with NBD instead");
1932 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1933 error_setg(errp
, "Cannot resume if there is no "
1934 "paused migration");
1939 * Postcopy recovery won't work well with release-ram
1940 * capability since release-ram will drop the page buffer as
1941 * long as the page is put into the send buffer. So if there
1942 * is a network failure happened, any page buffers that have
1943 * not yet reached the destination VM but have already been
1944 * sent from the source VM will be lost forever. Let's refuse
1945 * the client from resuming such a postcopy migration.
1946 * Luckily release-ram was designed to only be used when src
1947 * and destination VMs are on the same host, so it should be
1950 if (migrate_release_ram()) {
1951 error_setg(errp
, "Postcopy recovery cannot work "
1952 "when release-ram capability is set");
1956 /* This is a resume, skip init status */
1960 if (migration_is_running()) {
1961 error_setg(errp
, "There's a migration process in progress");
1965 if (runstate_check(RUN_STATE_INMIGRATE
)) {
1966 error_setg(errp
, "Guest is waiting for an incoming migration");
1970 if (runstate_check(RUN_STATE_POSTMIGRATE
)) {
1971 error_setg(errp
, "Can't migrate the vm that was paused due to "
1972 "previous migration");
1976 if (kvm_hwpoisoned_mem()) {
1977 error_setg(errp
, "Can't migrate this vm with hardware poisoned memory, "
1978 "please reboot the vm and try again");
1982 if (migration_is_blocked(errp
)) {
1986 if (migrate_mapped_ram()) {
1987 if (migrate_tls()) {
1988 error_setg(errp
, "Cannot use TLS with mapped-ram");
1992 if (migrate_multifd_compression()) {
1993 error_setg(errp
, "Cannot use compression with mapped-ram");
1998 if (migrate_mode_is_cpr(s
)) {
1999 const char *conflict
= NULL
;
2001 if (migrate_postcopy()) {
2002 conflict
= "postcopy";
2003 } else if (migrate_background_snapshot()) {
2004 conflict
= "background snapshot";
2005 } else if (migrate_colo()) {
2010 error_setg(errp
, "Cannot use %s with CPR", conflict
);
2015 if (blk
|| blk_inc
) {
2016 if (migrate_colo()) {
2017 error_setg(errp
, "No disk migration is required in COLO mode");
2020 if (migrate_block() || migrate_block_incremental()) {
2021 error_setg(errp
, "Command options are incompatible with "
2022 "current migration capabilities");
2025 if (!migrate_cap_set(MIGRATION_CAPABILITY_BLOCK
, true, errp
)) {
2028 s
->must_remove_block_options
= true;
2032 migrate_set_block_incremental(true);
2035 if (migrate_init(s
, errp
)) {
2042 void qmp_migrate(const char *uri
, bool has_channels
,
2043 MigrationChannelList
*channels
, bool has_blk
, bool blk
,
2044 bool has_inc
, bool inc
, bool has_detach
, bool detach
,
2045 bool has_resume
, bool resume
, Error
**errp
)
2047 bool resume_requested
;
2048 Error
*local_err
= NULL
;
2049 MigrationState
*s
= migrate_get_current();
2050 g_autoptr(MigrationChannel
) channel
= NULL
;
2051 MigrationAddress
*addr
= NULL
;
2054 * Having preliminary checks for uri and channel
2056 if (!uri
== !channels
) {
2057 error_setg(errp
, "need either 'uri' or 'channels' argument");
2062 /* To verify that Migrate channel list has only item */
2063 if (channels
->next
) {
2064 error_setg(errp
, "Channel list has more than one entries");
2067 addr
= channels
->value
->addr
;
2071 /* caller uses the old URI syntax */
2072 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
2075 addr
= channel
->addr
;
2078 /* transport mechanism not suitable for migration? */
2079 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
2083 resume_requested
= has_resume
&& resume
;
2084 if (!migrate_prepare(s
, has_blk
&& blk
, has_inc
&& inc
,
2085 resume_requested
, errp
)) {
2086 /* Error detected, put into errp */
2090 if (!resume_requested
) {
2091 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
2096 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
2097 SocketAddress
*saddr
= &addr
->u
.socket
;
2098 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
2099 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
2100 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
2101 socket_start_outgoing_migration(s
, saddr
, &local_err
);
2102 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
2103 fd_start_outgoing_migration(s
, saddr
->u
.fd
.str
, &local_err
);
2106 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
2107 rdma_start_outgoing_migration(s
, &addr
->u
.rdma
, &local_err
);
2109 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
2110 exec_start_outgoing_migration(s
, addr
->u
.exec
.args
, &local_err
);
2111 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
2112 file_start_outgoing_migration(s
, &addr
->u
.file
, &local_err
);
2114 error_setg(&local_err
, QERR_INVALID_PARAMETER_VALUE
, "uri",
2115 "a valid migration protocol");
2116 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
2117 MIGRATION_STATUS_FAILED
);
2118 block_cleanup_parameters();
2122 if (!resume_requested
) {
2123 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
2125 migrate_fd_error(s
, local_err
);
2126 error_propagate(errp
, local_err
);
2131 void qmp_migrate_cancel(Error
**errp
)
2133 migration_cancel(NULL
);
2136 void qmp_migrate_continue(MigrationStatus state
, Error
**errp
)
2138 MigrationState
*s
= migrate_get_current();
2139 if (s
->state
!= state
) {
2140 error_setg(errp
, "Migration not in expected state: %s",
2141 MigrationStatus_str(s
->state
));
2144 qemu_sem_post(&s
->pause_sem
);
2147 int migration_rp_wait(MigrationState
*s
)
2149 /* If migration has failure already, ignore the wait */
2150 if (migrate_has_error(s
)) {
2154 qemu_sem_wait(&s
->rp_state
.rp_sem
);
2156 /* After wait, double check that there's no failure */
2157 if (migrate_has_error(s
)) {
2164 void migration_rp_kick(MigrationState
*s
)
2166 qemu_sem_post(&s
->rp_state
.rp_sem
);
2169 static struct rp_cmd_args
{
2170 ssize_t len
; /* -1 = variable */
2173 [MIG_RP_MSG_INVALID
] = { .len
= -1, .name
= "INVALID" },
2174 [MIG_RP_MSG_SHUT
] = { .len
= 4, .name
= "SHUT" },
2175 [MIG_RP_MSG_PONG
] = { .len
= 4, .name
= "PONG" },
2176 [MIG_RP_MSG_REQ_PAGES
] = { .len
= 12, .name
= "REQ_PAGES" },
2177 [MIG_RP_MSG_REQ_PAGES_ID
] = { .len
= -1, .name
= "REQ_PAGES_ID" },
2178 [MIG_RP_MSG_RECV_BITMAP
] = { .len
= -1, .name
= "RECV_BITMAP" },
2179 [MIG_RP_MSG_RESUME_ACK
] = { .len
= 4, .name
= "RESUME_ACK" },
2180 [MIG_RP_MSG_SWITCHOVER_ACK
] = { .len
= 0, .name
= "SWITCHOVER_ACK" },
2181 [MIG_RP_MSG_MAX
] = { .len
= -1, .name
= "MAX" },
2185 * Process a request for pages received on the return path,
2186 * We're allowed to send more than requested (e.g. to round to our page size)
2187 * and we don't need to send pages that have already been sent.
2190 migrate_handle_rp_req_pages(MigrationState
*ms
, const char* rbname
,
2191 ram_addr_t start
, size_t len
, Error
**errp
)
2193 long our_host_ps
= qemu_real_host_page_size();
2195 trace_migrate_handle_rp_req_pages(rbname
, start
, len
);
2198 * Since we currently insist on matching page sizes, just sanity check
2199 * we're being asked for whole host pages.
2201 if (!QEMU_IS_ALIGNED(start
, our_host_ps
) ||
2202 !QEMU_IS_ALIGNED(len
, our_host_ps
)) {
2203 error_setg(errp
, "MIG_RP_MSG_REQ_PAGES: Misaligned page request, start:"
2204 RAM_ADDR_FMT
" len: %zd", start
, len
);
2208 ram_save_queue_pages(rbname
, start
, len
, errp
);
2211 static bool migrate_handle_rp_recv_bitmap(MigrationState
*s
, char *block_name
,
2214 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
2217 error_setg(errp
, "MIG_RP_MSG_RECV_BITMAP has invalid block name '%s'",
2222 /* Fetch the received bitmap and refresh the dirty bitmap */
2223 return ram_dirty_bitmap_reload(s
, block
, errp
);
2226 static bool migrate_handle_rp_resume_ack(MigrationState
*s
,
2227 uint32_t value
, Error
**errp
)
2229 trace_source_return_path_thread_resume_ack(value
);
2231 if (value
!= MIGRATION_RESUME_ACK_VALUE
) {
2232 error_setg(errp
, "illegal resume_ack value %"PRIu32
, value
);
2236 /* Now both sides are active. */
2237 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_RECOVER
,
2238 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2240 /* Notify send thread that time to continue send pages */
2241 migration_rp_kick(s
);
2247 * Release ms->rp_state.from_dst_file (and postcopy_qemufile_src if
2248 * existed) in a safe way.
2250 static void migration_release_dst_files(MigrationState
*ms
)
2254 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2256 * Reset the from_dst_file pointer first before releasing it, as we
2257 * can't block within lock section
2259 file
= ms
->rp_state
.from_dst_file
;
2260 ms
->rp_state
.from_dst_file
= NULL
;
2264 * Do the same to postcopy fast path socket too if there is. No
2265 * locking needed because this qemufile should only be managed by
2266 * return path thread.
2268 if (ms
->postcopy_qemufile_src
) {
2269 migration_ioc_unregister_yank_from_file(ms
->postcopy_qemufile_src
);
2270 qemu_file_shutdown(ms
->postcopy_qemufile_src
);
2271 qemu_fclose(ms
->postcopy_qemufile_src
);
2272 ms
->postcopy_qemufile_src
= NULL
;
2279 * Handles messages sent on the return path towards the source VM
2282 static void *source_return_path_thread(void *opaque
)
2284 MigrationState
*ms
= opaque
;
2285 QEMUFile
*rp
= ms
->rp_state
.from_dst_file
;
2286 uint16_t header_len
, header_type
;
2288 uint32_t tmp32
, sibling_error
;
2289 ram_addr_t start
= 0; /* =0 to silence warning */
2290 size_t len
= 0, expected_len
;
2294 trace_source_return_path_thread_entry();
2295 rcu_register_thread();
2297 while (migration_is_setup_or_active()) {
2298 trace_source_return_path_thread_loop_top();
2300 header_type
= qemu_get_be16(rp
);
2301 header_len
= qemu_get_be16(rp
);
2303 if (qemu_file_get_error(rp
)) {
2304 qemu_file_get_error_obj(rp
, &err
);
2308 if (header_type
>= MIG_RP_MSG_MAX
||
2309 header_type
== MIG_RP_MSG_INVALID
) {
2310 error_setg(&err
, "Received invalid message 0x%04x length 0x%04x",
2311 header_type
, header_len
);
2315 if ((rp_cmd_args
[header_type
].len
!= -1 &&
2316 header_len
!= rp_cmd_args
[header_type
].len
) ||
2317 header_len
> sizeof(buf
)) {
2318 error_setg(&err
, "Received '%s' message (0x%04x) with"
2319 "incorrect length %d expecting %zu",
2320 rp_cmd_args
[header_type
].name
, header_type
, header_len
,
2321 (size_t)rp_cmd_args
[header_type
].len
);
2325 /* We know we've got a valid header by this point */
2326 res
= qemu_get_buffer(rp
, buf
, header_len
);
2327 if (res
!= header_len
) {
2328 error_setg(&err
, "Failed reading data for message 0x%04x"
2329 " read %d expected %d",
2330 header_type
, res
, header_len
);
2334 /* OK, we have the message and the data */
2335 switch (header_type
) {
2336 case MIG_RP_MSG_SHUT
:
2337 sibling_error
= ldl_be_p(buf
);
2338 trace_source_return_path_thread_shut(sibling_error
);
2339 if (sibling_error
) {
2340 error_setg(&err
, "Sibling indicated error %d", sibling_error
);
2343 * We'll let the main thread deal with closing the RP
2344 * we could do a shutdown(2) on it, but we're the only user
2345 * anyway, so there's nothing gained.
2349 case MIG_RP_MSG_PONG
:
2350 tmp32
= ldl_be_p(buf
);
2351 trace_source_return_path_thread_pong(tmp32
);
2352 qemu_sem_post(&ms
->rp_state
.rp_pong_acks
);
2355 case MIG_RP_MSG_REQ_PAGES
:
2356 start
= ldq_be_p(buf
);
2357 len
= ldl_be_p(buf
+ 8);
2358 migrate_handle_rp_req_pages(ms
, NULL
, start
, len
, &err
);
2364 case MIG_RP_MSG_REQ_PAGES_ID
:
2365 expected_len
= 12 + 1; /* header + termination */
2367 if (header_len
>= expected_len
) {
2368 start
= ldq_be_p(buf
);
2369 len
= ldl_be_p(buf
+ 8);
2370 /* Now we expect an idstr */
2371 tmp32
= buf
[12]; /* Length of the following idstr */
2372 buf
[13 + tmp32
] = '\0';
2373 expected_len
+= tmp32
;
2375 if (header_len
!= expected_len
) {
2376 error_setg(&err
, "Req_Page_id with length %d expecting %zd",
2377 header_len
, expected_len
);
2380 migrate_handle_rp_req_pages(ms
, (char *)&buf
[13], start
, len
,
2387 case MIG_RP_MSG_RECV_BITMAP
:
2388 if (header_len
< 1) {
2389 error_setg(&err
, "MIG_RP_MSG_RECV_BITMAP missing block name");
2392 /* Format: len (1B) + idstr (<255B). This ends the idstr. */
2393 buf
[buf
[0] + 1] = '\0';
2394 if (!migrate_handle_rp_recv_bitmap(ms
, (char *)(buf
+ 1), &err
)) {
2399 case MIG_RP_MSG_RESUME_ACK
:
2400 tmp32
= ldl_be_p(buf
);
2401 if (!migrate_handle_rp_resume_ack(ms
, tmp32
, &err
)) {
2406 case MIG_RP_MSG_SWITCHOVER_ACK
:
2407 ms
->switchover_acked
= true;
2408 trace_source_return_path_thread_switchover_acked();
2418 migrate_set_error(ms
, err
);
2420 trace_source_return_path_thread_bad_end();
2423 if (ms
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2425 * this will be extremely unlikely: that we got yet another network
2426 * issue during recovering of the 1st network failure.. during this
2427 * period the main migration thread can be waiting on rp_sem for
2428 * this thread to sync with the other side.
2430 * When this happens, explicitly kick the migration thread out of
2431 * RECOVER stage and back to PAUSED, so the admin can try
2434 migration_rp_kick(ms
);
2437 trace_source_return_path_thread_end();
2438 rcu_unregister_thread();
2443 static int open_return_path_on_source(MigrationState
*ms
)
2445 ms
->rp_state
.from_dst_file
= qemu_file_get_return_path(ms
->to_dst_file
);
2446 if (!ms
->rp_state
.from_dst_file
) {
2450 trace_open_return_path_on_source();
2452 qemu_thread_create(&ms
->rp_state
.rp_thread
, "return path",
2453 source_return_path_thread
, ms
, QEMU_THREAD_JOINABLE
);
2454 ms
->rp_state
.rp_thread_created
= true;
2456 trace_open_return_path_on_source_continue();
2461 /* Return true if error detected, or false otherwise */
2462 static bool close_return_path_on_source(MigrationState
*ms
)
2464 if (!ms
->rp_state
.rp_thread_created
) {
2468 trace_migration_return_path_end_before();
2471 * If this is a normal exit then the destination will send a SHUT
2472 * and the rp_thread will exit, however if there's an error we
2473 * need to cause it to exit. shutdown(2), if we have it, will
2474 * cause it to unblock if it's stuck waiting for the destination.
2476 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2477 if (migrate_has_error(ms
) && ms
->rp_state
.from_dst_file
) {
2478 qemu_file_shutdown(ms
->rp_state
.from_dst_file
);
2482 qemu_thread_join(&ms
->rp_state
.rp_thread
);
2483 ms
->rp_state
.rp_thread_created
= false;
2484 migration_release_dst_files(ms
);
2485 trace_migration_return_path_end_after();
2487 /* Return path will persist the error in MigrationState when quit */
2488 return migrate_has_error(ms
);
2492 migration_wait_main_channel(MigrationState
*ms
)
2494 /* Wait until one PONG message received */
2495 qemu_sem_wait(&ms
->rp_state
.rp_pong_acks
);
2499 * Switch from normal iteration to postcopy
2500 * Returns non-0 on error
2502 static int postcopy_start(MigrationState
*ms
, Error
**errp
)
2505 QIOChannelBuffer
*bioc
;
2507 uint64_t bandwidth
= migrate_max_postcopy_bandwidth();
2508 bool restart_block
= false;
2509 int cur_state
= MIGRATION_STATUS_ACTIVE
;
2511 if (migrate_postcopy_preempt()) {
2512 migration_wait_main_channel(ms
);
2513 if (postcopy_preempt_establish_channel(ms
)) {
2514 migrate_set_state(&ms
->state
, ms
->state
, MIGRATION_STATUS_FAILED
);
2515 error_setg(errp
, "%s: Failed to establish preempt channel",
2521 if (!migrate_pause_before_switchover()) {
2522 migrate_set_state(&ms
->state
, MIGRATION_STATUS_ACTIVE
,
2523 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2526 trace_postcopy_start();
2528 trace_postcopy_start_set_run();
2530 ret
= migration_stop_vm(ms
, RUN_STATE_FINISH_MIGRATE
);
2532 error_setg_errno(errp
, -ret
, "%s: Failed to stop the VM", __func__
);
2536 ret
= migration_maybe_pause(ms
, &cur_state
,
2537 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2539 error_setg_errno(errp
, -ret
, "%s: Failed in migration_maybe_pause()",
2544 ret
= bdrv_inactivate_all();
2546 error_setg_errno(errp
, -ret
, "%s: Failed in bdrv_inactivate_all()",
2550 restart_block
= true;
2553 * Cause any non-postcopiable, but iterative devices to
2554 * send out their final data.
2556 qemu_savevm_state_complete_precopy(ms
->to_dst_file
, true, false);
2559 * in Finish migrate and with the io-lock held everything should
2560 * be quiet, but we've potentially still got dirty pages and we
2561 * need to tell the destination to throw any pages it's already received
2564 if (migrate_postcopy_ram()) {
2565 ram_postcopy_send_discard_bitmap(ms
);
2569 * send rest of state - note things that are doing postcopy
2570 * will notice we're in POSTCOPY_ACTIVE and not actually
2571 * wrap their state up here
2573 migration_rate_set(bandwidth
);
2574 if (migrate_postcopy_ram()) {
2575 /* Ping just for debugging, helps line traces up */
2576 qemu_savevm_send_ping(ms
->to_dst_file
, 2);
2580 * While loading the device state we may trigger page transfer
2581 * requests and the fd must be free to process those, and thus
2582 * the destination must read the whole device state off the fd before
2583 * it starts processing it. Unfortunately the ad-hoc migration format
2584 * doesn't allow the destination to know the size to read without fully
2585 * parsing it through each devices load-state code (especially the open
2586 * coded devices that use get/put).
2587 * So we wrap the device state up in a package with a length at the start;
2588 * to do this we use a qemu_buf to hold the whole of the device state.
2590 bioc
= qio_channel_buffer_new(4096);
2591 qio_channel_set_name(QIO_CHANNEL(bioc
), "migration-postcopy-buffer");
2592 fb
= qemu_file_new_output(QIO_CHANNEL(bioc
));
2593 object_unref(OBJECT(bioc
));
2596 * Make sure the receiver can get incoming pages before we send the rest
2599 qemu_savevm_send_postcopy_listen(fb
);
2601 qemu_savevm_state_complete_precopy(fb
, false, false);
2602 if (migrate_postcopy_ram()) {
2603 qemu_savevm_send_ping(fb
, 3);
2606 qemu_savevm_send_postcopy_run(fb
);
2608 /* <><> end of stuff going into the package */
2610 /* Last point of recovery; as soon as we send the package the destination
2611 * can open devices and potentially start running.
2612 * Lets just check again we've not got any errors.
2614 ret
= qemu_file_get_error(ms
->to_dst_file
);
2616 error_setg(errp
, "postcopy_start: Migration stream errored (pre package)");
2620 restart_block
= false;
2622 /* Now send that blob */
2623 if (qemu_savevm_send_packaged(ms
->to_dst_file
, bioc
->data
, bioc
->usage
)) {
2624 error_setg(errp
, "%s: Failed to send packaged data", __func__
);
2629 /* Send a notify to give a chance for anything that needs to happen
2630 * at the transition to postcopy and after the device state; in particular
2631 * spice needs to trigger a transition now
2633 migration_call_notifiers(ms
, MIG_EVENT_PRECOPY_DONE
, NULL
);
2635 migration_downtime_end(ms
);
2639 if (migrate_postcopy_ram()) {
2641 * Although this ping is just for debug, it could potentially be
2642 * used for getting a better measurement of downtime at the source.
2644 qemu_savevm_send_ping(ms
->to_dst_file
, 4);
2647 if (migrate_release_ram()) {
2648 ram_postcopy_migrated_memory_release(ms
);
2651 ret
= qemu_file_get_error(ms
->to_dst_file
);
2653 error_setg_errno(errp
, -ret
, "postcopy_start: Migration stream error");
2657 trace_postcopy_preempt_enabled(migrate_postcopy_preempt());
2664 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2665 MIGRATION_STATUS_FAILED
);
2666 if (restart_block
) {
2667 /* A failure happened early enough that we know the destination hasn't
2668 * accessed block devices, so we're safe to recover.
2670 Error
*local_err
= NULL
;
2672 bdrv_activate_all(&local_err
);
2674 error_report_err(local_err
);
2677 migration_call_notifiers(ms
, MIG_EVENT_PRECOPY_FAILED
, NULL
);
2683 * migration_maybe_pause: Pause if required to by
2684 * migrate_pause_before_switchover called with the BQL locked
2685 * Returns: 0 on success
2687 static int migration_maybe_pause(MigrationState
*s
,
2688 int *current_active_state
,
2691 if (!migrate_pause_before_switchover()) {
2695 /* Since leaving this state is not atomic with posting the semaphore
2696 * it's possible that someone could have issued multiple migrate_continue
2697 * and the semaphore is incorrectly positive at this point;
2698 * the docs say it's undefined to reinit a semaphore that's already
2699 * init'd, so use timedwait to eat up any existing posts.
2701 while (qemu_sem_timedwait(&s
->pause_sem
, 1) == 0) {
2702 /* This block intentionally left blank */
2706 * If the migration is cancelled when it is in the completion phase,
2707 * the migration state is set to MIGRATION_STATUS_CANCELLING.
2708 * So we don't need to wait a semaphore, otherwise we would always
2709 * wait for the 'pause_sem' semaphore.
2711 if (s
->state
!= MIGRATION_STATUS_CANCELLING
) {
2713 migrate_set_state(&s
->state
, *current_active_state
,
2714 MIGRATION_STATUS_PRE_SWITCHOVER
);
2715 qemu_sem_wait(&s
->pause_sem
);
2716 migrate_set_state(&s
->state
, MIGRATION_STATUS_PRE_SWITCHOVER
,
2718 *current_active_state
= new_state
;
2722 return s
->state
== new_state
? 0 : -EINVAL
;
2725 static int migration_completion_precopy(MigrationState
*s
,
2726 int *current_active_state
)
2732 if (!migrate_mode_is_cpr(s
)) {
2733 ret
= migration_stop_vm(s
, RUN_STATE_FINISH_MIGRATE
);
2739 ret
= migration_maybe_pause(s
, current_active_state
,
2740 MIGRATION_STATUS_DEVICE
);
2746 * Inactivate disks except in COLO, and track that we have done so in order
2747 * to remember to reactivate them if migration fails or is cancelled.
2749 s
->block_inactive
= !migrate_colo();
2750 migration_rate_set(RATE_LIMIT_DISABLED
);
2751 ret
= qemu_savevm_state_complete_precopy(s
->to_dst_file
, false,
2758 static void migration_completion_postcopy(MigrationState
*s
)
2760 trace_migration_completion_postcopy_end();
2763 qemu_savevm_state_complete_postcopy(s
->to_dst_file
);
2767 * Shutdown the postcopy fast path thread. This is only needed when dest
2768 * QEMU binary is old (7.1/7.2). QEMU 8.0+ doesn't need this.
2770 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
2771 postcopy_preempt_shutdown_file(s
);
2774 trace_migration_completion_postcopy_end_after_complete();
2777 static void migration_completion_failed(MigrationState
*s
,
2778 int current_active_state
)
2780 if (s
->block_inactive
&& (s
->state
== MIGRATION_STATUS_ACTIVE
||
2781 s
->state
== MIGRATION_STATUS_DEVICE
)) {
2783 * If not doing postcopy, vm_start() will be called: let's
2784 * regain control on images.
2786 Error
*local_err
= NULL
;
2789 bdrv_activate_all(&local_err
);
2791 error_report_err(local_err
);
2793 s
->block_inactive
= false;
2798 migrate_set_state(&s
->state
, current_active_state
,
2799 MIGRATION_STATUS_FAILED
);
2803 * migration_completion: Used by migration_thread when there's not much left.
2804 * The caller 'breaks' the loop when this returns.
2806 * @s: Current migration state
2808 static void migration_completion(MigrationState
*s
)
2811 int current_active_state
= s
->state
;
2812 Error
*local_err
= NULL
;
2814 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2815 ret
= migration_completion_precopy(s
, ¤t_active_state
);
2816 } else if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2817 migration_completion_postcopy(s
);
2826 if (close_return_path_on_source(s
)) {
2830 if (qemu_file_get_error(s
->to_dst_file
)) {
2831 trace_migration_completion_file_err();
2835 if (migrate_colo() && s
->state
== MIGRATION_STATUS_ACTIVE
) {
2836 /* COLO does not support postcopy */
2837 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
2838 MIGRATION_STATUS_COLO
);
2840 migration_completion_end(s
);
2846 if (qemu_file_get_error_obj(s
->to_dst_file
, &local_err
)) {
2847 migrate_set_error(s
, local_err
);
2848 error_free(local_err
);
2850 error_setg_errno(&local_err
, -ret
, "Error in migration completion");
2851 migrate_set_error(s
, local_err
);
2852 error_free(local_err
);
2855 migration_completion_failed(s
, current_active_state
);
2859 * bg_migration_completion: Used by bg_migration_thread when after all the
2860 * RAM has been saved. The caller 'breaks' the loop when this returns.
2862 * @s: Current migration state
2864 static void bg_migration_completion(MigrationState
*s
)
2866 int current_active_state
= s
->state
;
2868 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2870 * By this moment we have RAM content saved into the migration stream.
2871 * The next step is to flush the non-RAM content (device state)
2872 * right after the ram content. The device state has been stored into
2873 * the temporary buffer before RAM saving started.
2875 qemu_put_buffer(s
->to_dst_file
, s
->bioc
->data
, s
->bioc
->usage
);
2876 qemu_fflush(s
->to_dst_file
);
2877 } else if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
2881 if (qemu_file_get_error(s
->to_dst_file
)) {
2882 trace_migration_completion_file_err();
2886 migration_completion_end(s
);
2890 migrate_set_state(&s
->state
, current_active_state
,
2891 MIGRATION_STATUS_FAILED
);
2894 typedef enum MigThrError
{
2895 /* No error detected */
2896 MIG_THR_ERR_NONE
= 0,
2897 /* Detected error, but resumed successfully */
2898 MIG_THR_ERR_RECOVERED
= 1,
2899 /* Detected fatal error, need to exit */
2900 MIG_THR_ERR_FATAL
= 2,
2903 static int postcopy_resume_handshake(MigrationState
*s
)
2905 qemu_savevm_send_postcopy_resume(s
->to_dst_file
);
2907 while (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2908 if (migration_rp_wait(s
)) {
2913 if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2920 /* Return zero if success, or <0 for error */
2921 static int postcopy_do_resume(MigrationState
*s
)
2926 * Call all the resume_prepare() hooks, so that modules can be
2927 * ready for the migration resume.
2929 ret
= qemu_savevm_state_resume_prepare(s
);
2931 error_report("%s: resume_prepare() failure detected: %d",
2937 * If preempt is enabled, re-establish the preempt channel. Note that
2938 * we do it after resume prepare to make sure the main channel will be
2939 * created before the preempt channel. E.g. with weak network, the
2940 * dest QEMU may get messed up with the preempt and main channels on
2941 * the order of connection setup. This guarantees the correct order.
2943 ret
= postcopy_preempt_establish_channel(s
);
2945 error_report("%s: postcopy_preempt_establish_channel(): %d",
2951 * Last handshake with destination on the resume (destination will
2952 * switch to postcopy-active afterwards)
2954 ret
= postcopy_resume_handshake(s
);
2956 error_report("%s: handshake failed: %d", __func__
, ret
);
2964 * We don't return until we are in a safe state to continue current
2965 * postcopy migration. Returns MIG_THR_ERR_RECOVERED if recovered, or
2966 * MIG_THR_ERR_FATAL if unrecovery failure happened.
2968 static MigThrError
postcopy_pause(MigrationState
*s
)
2970 assert(s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2976 * We're already pausing, so ignore any errors on the return
2977 * path and just wait for the thread to finish. It will be
2978 * re-created when we resume.
2980 close_return_path_on_source(s
);
2983 * Current channel is possibly broken. Release it. Note that this is
2984 * guaranteed even without lock because to_dst_file should only be
2985 * modified by the migration thread. That also guarantees that the
2986 * unregister of yank is safe too without the lock. It should be safe
2987 * even to be within the qemu_file_lock, but we didn't do that to avoid
2988 * taking more mutex (yank_lock) within qemu_file_lock. TL;DR: we make
2989 * the qemu_file_lock critical section as small as possible.
2991 assert(s
->to_dst_file
);
2992 migration_ioc_unregister_yank_from_file(s
->to_dst_file
);
2993 qemu_mutex_lock(&s
->qemu_file_lock
);
2994 file
= s
->to_dst_file
;
2995 s
->to_dst_file
= NULL
;
2996 qemu_mutex_unlock(&s
->qemu_file_lock
);
2998 qemu_file_shutdown(file
);
3001 migrate_set_state(&s
->state
, s
->state
,
3002 MIGRATION_STATUS_POSTCOPY_PAUSED
);
3004 error_report("Detected IO failure for postcopy. "
3005 "Migration paused.");
3008 * We wait until things fixed up. Then someone will setup the
3009 * status back for us.
3011 while (s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
3012 qemu_sem_wait(&s
->postcopy_pause_sem
);
3015 if (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
3016 /* Woken up by a recover procedure. Give it a shot */
3018 /* Do the resume logic */
3019 if (postcopy_do_resume(s
) == 0) {
3020 /* Let's continue! */
3021 trace_postcopy_pause_continued();
3022 return MIG_THR_ERR_RECOVERED
;
3025 * Something wrong happened during the recovery, let's
3026 * pause again. Pause is always better than throwing
3032 /* This is not right... Time to quit. */
3033 return MIG_THR_ERR_FATAL
;
3038 void migration_file_set_error(int err
)
3040 MigrationState
*s
= current_migration
;
3042 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
3043 if (s
->to_dst_file
) {
3044 qemu_file_set_error(s
->to_dst_file
, err
);
3049 static MigThrError
migration_detect_error(MigrationState
*s
)
3052 int state
= s
->state
;
3053 Error
*local_error
= NULL
;
3055 if (state
== MIGRATION_STATUS_CANCELLING
||
3056 state
== MIGRATION_STATUS_CANCELLED
) {
3057 /* End the migration, but don't set the state to failed */
3058 return MIG_THR_ERR_FATAL
;
3062 * Try to detect any file errors. Note that postcopy_qemufile_src will
3063 * be NULL when postcopy preempt is not enabled.
3065 ret
= qemu_file_get_error_obj_any(s
->to_dst_file
,
3066 s
->postcopy_qemufile_src
,
3069 /* Everything is fine */
3070 assert(!local_error
);
3071 return MIG_THR_ERR_NONE
;
3075 migrate_set_error(s
, local_error
);
3076 error_free(local_error
);
3079 if (state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
&& ret
) {
3081 * For postcopy, we allow the network to be down for a
3082 * while. After that, it can be continued by a
3085 return postcopy_pause(s
);
3088 * For precopy (or postcopy with error outside IO), we fail
3091 migrate_set_state(&s
->state
, state
, MIGRATION_STATUS_FAILED
);
3092 trace_migration_thread_file_err();
3094 /* Time to stop the migration, now. */
3095 return MIG_THR_ERR_FATAL
;
3099 static void migration_completion_end(MigrationState
*s
)
3101 uint64_t bytes
= migration_transferred_bytes();
3102 int64_t end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3103 int64_t transfer_time
;
3106 * Take the BQL here so that query-migrate on the QMP thread sees:
3107 * - atomic update of s->total_time and s->mbps;
3108 * - correct ordering of s->mbps update vs. s->state;
3111 migration_downtime_end(s
);
3112 s
->total_time
= end_time
- s
->start_time
;
3113 transfer_time
= s
->total_time
- s
->setup_time
;
3114 if (transfer_time
) {
3115 s
->mbps
= ((double) bytes
* 8.0) / transfer_time
/ 1000;
3118 migrate_set_state(&s
->state
, s
->state
,
3119 MIGRATION_STATUS_COMPLETED
);
3123 static void update_iteration_initial_status(MigrationState
*s
)
3126 * Update these three fields at the same time to avoid mismatch info lead
3127 * wrong speed calculation.
3129 s
->iteration_start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3130 s
->iteration_initial_bytes
= migration_transferred_bytes();
3131 s
->iteration_initial_pages
= ram_get_total_transferred_pages();
3134 static void migration_update_counters(MigrationState
*s
,
3135 int64_t current_time
)
3137 uint64_t transferred
, transferred_pages
, time_spent
;
3138 uint64_t current_bytes
; /* bytes transferred since the beginning */
3139 uint64_t switchover_bw
;
3140 /* Expected bandwidth when switching over to destination QEMU */
3141 double expected_bw_per_ms
;
3144 if (current_time
< s
->iteration_start_time
+ BUFFER_DELAY
) {
3148 switchover_bw
= migrate_avail_switchover_bandwidth();
3149 current_bytes
= migration_transferred_bytes();
3150 transferred
= current_bytes
- s
->iteration_initial_bytes
;
3151 time_spent
= current_time
- s
->iteration_start_time
;
3152 bandwidth
= (double)transferred
/ time_spent
;
3154 if (switchover_bw
) {
3156 * If the user specified a switchover bandwidth, let's trust the
3157 * user so that can be more accurate than what we estimated.
3159 expected_bw_per_ms
= switchover_bw
/ 1000;
3161 /* If the user doesn't specify bandwidth, we use the estimated */
3162 expected_bw_per_ms
= bandwidth
;
3165 s
->threshold_size
= expected_bw_per_ms
* migrate_downtime_limit();
3167 s
->mbps
= (((double) transferred
* 8.0) /
3168 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0;
3170 transferred_pages
= ram_get_total_transferred_pages() -
3171 s
->iteration_initial_pages
;
3172 s
->pages_per_second
= (double) transferred_pages
/
3173 (((double) time_spent
/ 1000.0));
3176 * if we haven't sent anything, we don't want to
3177 * recalculate. 10000 is a small enough number for our purposes
3179 if (stat64_get(&mig_stats
.dirty_pages_rate
) &&
3180 transferred
> 10000) {
3181 s
->expected_downtime
=
3182 stat64_get(&mig_stats
.dirty_bytes_last_sync
) / expected_bw_per_ms
;
3185 migration_rate_reset();
3187 update_iteration_initial_status(s
);
3189 trace_migrate_transferred(transferred
, time_spent
,
3190 /* Both in unit bytes/ms */
3191 bandwidth
, switchover_bw
/ 1000,
3195 static bool migration_can_switchover(MigrationState
*s
)
3197 if (!migrate_switchover_ack()) {
3201 /* No reason to wait for switchover ACK if VM is stopped */
3202 if (!runstate_is_running()) {
3206 return s
->switchover_acked
;
3209 /* Migration thread iteration status */
3211 MIG_ITERATE_RESUME
, /* Resume current iteration */
3212 MIG_ITERATE_SKIP
, /* Skip current iteration */
3213 MIG_ITERATE_BREAK
, /* Break the loop */
3217 * Return true if continue to the next iteration directly, false
3220 static MigIterateState
migration_iteration_run(MigrationState
*s
)
3222 uint64_t must_precopy
, can_postcopy
, pending_size
;
3223 Error
*local_err
= NULL
;
3224 bool in_postcopy
= s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
;
3225 bool can_switchover
= migration_can_switchover(s
);
3227 qemu_savevm_state_pending_estimate(&must_precopy
, &can_postcopy
);
3228 pending_size
= must_precopy
+ can_postcopy
;
3229 trace_migrate_pending_estimate(pending_size
, must_precopy
, can_postcopy
);
3231 if (pending_size
< s
->threshold_size
) {
3232 qemu_savevm_state_pending_exact(&must_precopy
, &can_postcopy
);
3233 pending_size
= must_precopy
+ can_postcopy
;
3234 trace_migrate_pending_exact(pending_size
, must_precopy
, can_postcopy
);
3237 if ((!pending_size
|| pending_size
< s
->threshold_size
) && can_switchover
) {
3238 trace_migration_thread_low_pending(pending_size
);
3239 migration_completion(s
);
3240 return MIG_ITERATE_BREAK
;
3243 /* Still a significant amount to transfer */
3244 if (!in_postcopy
&& must_precopy
<= s
->threshold_size
&& can_switchover
&&
3245 qatomic_read(&s
->start_postcopy
)) {
3246 if (postcopy_start(s
, &local_err
)) {
3247 migrate_set_error(s
, local_err
);
3248 error_report_err(local_err
);
3250 return MIG_ITERATE_SKIP
;
3253 /* Just another iteration step */
3254 qemu_savevm_state_iterate(s
->to_dst_file
, in_postcopy
);
3255 return MIG_ITERATE_RESUME
;
3258 static void migration_iteration_finish(MigrationState
*s
)
3260 /* If we enabled cpu throttling for auto-converge, turn it off. */
3261 cpu_throttle_stop();
3265 case MIGRATION_STATUS_COMPLETED
:
3266 runstate_set(RUN_STATE_POSTMIGRATE
);
3268 case MIGRATION_STATUS_COLO
:
3269 assert(migrate_colo());
3270 migrate_start_colo_process(s
);
3271 s
->vm_old_state
= RUN_STATE_RUNNING
;
3273 case MIGRATION_STATUS_FAILED
:
3274 case MIGRATION_STATUS_CANCELLED
:
3275 case MIGRATION_STATUS_CANCELLING
:
3276 if (runstate_is_live(s
->vm_old_state
)) {
3277 if (!runstate_check(RUN_STATE_SHUTDOWN
)) {
3281 if (runstate_check(RUN_STATE_FINISH_MIGRATE
)) {
3282 runstate_set(s
->vm_old_state
);
3288 /* Should not reach here, but if so, forgive the VM. */
3289 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3293 migration_bh_schedule(migrate_fd_cleanup_bh
, s
);
3297 static void bg_migration_iteration_finish(MigrationState
*s
)
3300 * Stop tracking RAM writes - un-protect memory, un-register UFFD
3301 * memory ranges, flush kernel wait queues and wake up threads
3302 * waiting for write fault to be resolved.
3304 ram_write_tracking_stop();
3308 case MIGRATION_STATUS_COMPLETED
:
3309 case MIGRATION_STATUS_ACTIVE
:
3310 case MIGRATION_STATUS_FAILED
:
3311 case MIGRATION_STATUS_CANCELLED
:
3312 case MIGRATION_STATUS_CANCELLING
:
3316 /* Should not reach here, but if so, forgive the VM. */
3317 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3321 migration_bh_schedule(migrate_fd_cleanup_bh
, s
);
3326 * Return true if continue to the next iteration directly, false
3329 static MigIterateState
bg_migration_iteration_run(MigrationState
*s
)
3333 res
= qemu_savevm_state_iterate(s
->to_dst_file
, false);
3335 bg_migration_completion(s
);
3336 return MIG_ITERATE_BREAK
;
3339 return MIG_ITERATE_RESUME
;
3342 void migration_make_urgent_request(void)
3344 qemu_sem_post(&migrate_get_current()->rate_limit_sem
);
3347 void migration_consume_urgent_request(void)
3349 qemu_sem_wait(&migrate_get_current()->rate_limit_sem
);
3352 /* Returns true if the rate limiting was broken by an urgent request */
3353 bool migration_rate_limit(void)
3355 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3356 MigrationState
*s
= migrate_get_current();
3358 bool urgent
= false;
3359 migration_update_counters(s
, now
);
3360 if (migration_rate_exceeded(s
->to_dst_file
)) {
3362 if (qemu_file_get_error(s
->to_dst_file
)) {
3366 * Wait for a delay to do rate limiting OR
3367 * something urgent to post the semaphore.
3369 int ms
= s
->iteration_start_time
+ BUFFER_DELAY
- now
;
3370 trace_migration_rate_limit_pre(ms
);
3371 if (qemu_sem_timedwait(&s
->rate_limit_sem
, ms
) == 0) {
3373 * We were woken by one or more urgent things but
3374 * the timedwait will have consumed one of them.
3375 * The service routine for the urgent wake will dec
3376 * the semaphore itself for each item it consumes,
3377 * so add this one we just eat back.
3379 qemu_sem_post(&s
->rate_limit_sem
);
3382 trace_migration_rate_limit_post(urgent
);
3388 * if failover devices are present, wait they are completely
3392 static void qemu_savevm_wait_unplug(MigrationState
*s
, int old_state
,
3395 if (qemu_savevm_state_guest_unplug_pending()) {
3396 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_WAIT_UNPLUG
);
3398 while (s
->state
== MIGRATION_STATUS_WAIT_UNPLUG
&&
3399 qemu_savevm_state_guest_unplug_pending()) {
3400 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3402 if (s
->state
!= MIGRATION_STATUS_WAIT_UNPLUG
) {
3403 int timeout
= 120; /* 30 seconds */
3405 * migration has been canceled
3406 * but as we have started an unplug we must wait the end
3407 * to be able to plug back the card
3409 while (timeout
-- && qemu_savevm_state_guest_unplug_pending()) {
3410 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3412 if (qemu_savevm_state_guest_unplug_pending() &&
3414 warn_report("migration: partially unplugged device on "
3419 migrate_set_state(&s
->state
, MIGRATION_STATUS_WAIT_UNPLUG
, new_state
);
3421 migrate_set_state(&s
->state
, old_state
, new_state
);
3426 * Master migration thread on the source VM.
3427 * It drives the migration and pumps the data down the outgoing channel.
3429 static void *migration_thread(void *opaque
)
3431 MigrationState
*s
= opaque
;
3432 MigrationThread
*thread
= NULL
;
3433 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3434 MigThrError thr_error
;
3435 bool urgent
= false;
3436 Error
*local_err
= NULL
;
3439 thread
= migration_threads_add("live_migration", qemu_get_thread_id());
3441 rcu_register_thread();
3443 object_ref(OBJECT(s
));
3444 update_iteration_initial_status(s
);
3446 if (!multifd_send_setup()) {
3451 qemu_savevm_state_header(s
->to_dst_file
);
3455 * If we opened the return path, we need to make sure dst has it
3458 if (s
->rp_state
.rp_thread_created
) {
3459 /* Now tell the dest that it should open its end so it can reply */
3460 qemu_savevm_send_open_return_path(s
->to_dst_file
);
3462 /* And do a ping that will make stuff easier to debug */
3463 qemu_savevm_send_ping(s
->to_dst_file
, 1);
3466 if (migrate_postcopy()) {
3468 * Tell the destination that we *might* want to do postcopy later;
3469 * if the other end can't do postcopy it should fail now, nice and
3472 qemu_savevm_send_postcopy_advise(s
->to_dst_file
);
3475 if (migrate_colo()) {
3476 /* Notify migration destination that we enable COLO */
3477 qemu_savevm_send_colo_enable(s
->to_dst_file
);
3481 ret
= qemu_savevm_state_setup(s
->to_dst_file
, &local_err
);
3484 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3485 MIGRATION_STATUS_ACTIVE
);
3488 * Handle SETUP failures after waiting for virtio-net-failover
3489 * devices to unplug. This to preserve migration state transitions.
3492 migrate_set_error(s
, local_err
);
3493 error_free(local_err
);
3494 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3495 MIGRATION_STATUS_FAILED
);
3499 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3501 trace_migration_thread_setup_complete();
3503 while (migration_is_active()) {
3504 if (urgent
|| !migration_rate_exceeded(s
->to_dst_file
)) {
3505 MigIterateState iter_state
= migration_iteration_run(s
);
3506 if (iter_state
== MIG_ITERATE_SKIP
) {
3508 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3514 * Try to detect any kind of failures, and see whether we
3515 * should stop the migration now.
3517 thr_error
= migration_detect_error(s
);
3518 if (thr_error
== MIG_THR_ERR_FATAL
) {
3519 /* Stop migration */
3521 } else if (thr_error
== MIG_THR_ERR_RECOVERED
) {
3523 * Just recovered from a e.g. network failure, reset all
3524 * the local variables. This is important to avoid
3525 * breaking transferred_bytes and bandwidth calculation
3527 update_iteration_initial_status(s
);
3530 urgent
= migration_rate_limit();
3534 trace_migration_thread_after_loop();
3535 migration_iteration_finish(s
);
3536 object_unref(OBJECT(s
));
3537 rcu_unregister_thread();
3538 migration_threads_remove(thread
);
3542 static void bg_migration_vm_start_bh(void *opaque
)
3544 MigrationState
*s
= opaque
;
3546 vm_resume(s
->vm_old_state
);
3547 migration_downtime_end(s
);
3551 * Background snapshot thread, based on live migration code.
3552 * This is an alternative implementation of live migration mechanism
3553 * introduced specifically to support background snapshots.
3555 * It takes advantage of userfault_fd write protection mechanism introduced
3556 * in v5.7 kernel. Compared to existing dirty page logging migration much
3557 * lesser stream traffic is produced resulting in smaller snapshot images,
3558 * simply cause of no page duplicates can get into the stream.
3560 * Another key point is that generated vmstate stream reflects machine state
3561 * 'frozen' at the beginning of snapshot creation compared to dirty page logging
3562 * mechanism, which effectively results in that saved snapshot is the state of VM
3563 * at the end of the process.
3565 static void *bg_migration_thread(void *opaque
)
3567 MigrationState
*s
= opaque
;
3568 int64_t setup_start
;
3569 MigThrError thr_error
;
3571 bool early_fail
= true;
3572 Error
*local_err
= NULL
;
3575 rcu_register_thread();
3576 object_ref(OBJECT(s
));
3578 migration_rate_set(RATE_LIMIT_DISABLED
);
3580 setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3582 * We want to save vmstate for the moment when migration has been
3583 * initiated but also we want to save RAM content while VM is running.
3584 * The RAM content should appear first in the vmstate. So, we first
3585 * stash the non-RAM part of the vmstate to the temporary buffer,
3586 * then write RAM part of the vmstate to the migration stream
3587 * with vCPUs running and, finally, write stashed non-RAM part of
3588 * the vmstate from the buffer to the migration stream.
3590 s
->bioc
= qio_channel_buffer_new(512 * 1024);
3591 qio_channel_set_name(QIO_CHANNEL(s
->bioc
), "vmstate-buffer");
3592 fb
= qemu_file_new_output(QIO_CHANNEL(s
->bioc
));
3593 object_unref(OBJECT(s
->bioc
));
3595 update_iteration_initial_status(s
);
3598 * Prepare for tracking memory writes with UFFD-WP - populate
3599 * RAM pages before protecting.
3602 ram_write_tracking_prepare();
3606 qemu_savevm_state_header(s
->to_dst_file
);
3607 ret
= qemu_savevm_state_setup(s
->to_dst_file
, &local_err
);
3610 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3611 MIGRATION_STATUS_ACTIVE
);
3614 * Handle SETUP failures after waiting for virtio-net-failover
3615 * devices to unplug. This to preserve migration state transitions.
3618 migrate_set_error(s
, local_err
);
3619 error_free(local_err
);
3620 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3621 MIGRATION_STATUS_FAILED
);
3625 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3627 trace_migration_thread_setup_complete();
3631 if (migration_stop_vm(s
, RUN_STATE_PAUSED
)) {
3635 * Put vCPUs in sync with shadow context structures, then
3636 * save their state to channel-buffer along with devices.
3638 cpu_synchronize_all_states();
3639 if (qemu_savevm_state_complete_precopy_non_iterable(fb
, false, false)) {
3643 * Since we are going to get non-iterable state data directly
3644 * from s->bioc->data, explicit flush is needed here.
3648 /* Now initialize UFFD context and start tracking RAM writes */
3649 if (ram_write_tracking_start()) {
3655 * Start VM from BH handler to avoid write-fault lock here.
3656 * UFFD-WP protection for the whole RAM is already enabled so
3657 * calling VM state change notifiers from vm_start() would initiate
3658 * writes to virtio VQs memory which is in write-protected region.
3660 migration_bh_schedule(bg_migration_vm_start_bh
, s
);
3663 while (migration_is_active()) {
3664 MigIterateState iter_state
= bg_migration_iteration_run(s
);
3665 if (iter_state
== MIG_ITERATE_SKIP
) {
3667 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3672 * Try to detect any kind of failures, and see whether we
3673 * should stop the migration now.
3675 thr_error
= migration_detect_error(s
);
3676 if (thr_error
== MIG_THR_ERR_FATAL
) {
3677 /* Stop migration */
3681 migration_update_counters(s
, qemu_clock_get_ms(QEMU_CLOCK_REALTIME
));
3684 trace_migration_thread_after_loop();
3688 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3689 MIGRATION_STATUS_FAILED
);
3694 bg_migration_iteration_finish(s
);
3697 object_unref(OBJECT(s
));
3698 rcu_unregister_thread();
3703 void migrate_fd_connect(MigrationState
*s
, Error
*error_in
)
3705 Error
*local_err
= NULL
;
3706 uint64_t rate_limit
;
3707 bool resume
= s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
;
3711 * If there's a previous error, free it and prepare for another one.
3712 * Meanwhile if migration completes successfully, there won't have an error
3713 * dumped when calling migrate_fd_cleanup().
3715 migrate_error_free(s
);
3717 s
->expected_downtime
= migrate_downtime_limit();
3719 migrate_fd_error(s
, error_in
);
3722 * Don't do cleanup for resume if channel is invalid, but only dump
3723 * the error. We wait for another channel connect from the user.
3724 * The error_report still gives HMP user a hint on what failed.
3725 * It's normally done in migrate_fd_cleanup(), but call it here
3728 error_report_err(error_copy(s
->error
));
3730 migrate_fd_cleanup(s
);
3736 /* This is a resumed migration */
3737 rate_limit
= migrate_max_postcopy_bandwidth();
3739 /* This is a fresh new migration */
3740 rate_limit
= migrate_max_bandwidth();
3742 /* Notify before starting migration thread */
3743 if (migration_call_notifiers(s
, MIG_EVENT_PRECOPY_SETUP
, &local_err
)) {
3748 migration_rate_set(rate_limit
);
3749 qemu_file_set_blocking(s
->to_dst_file
, true);
3752 * Open the return path. For postcopy, it is used exclusively. For
3753 * precopy, only if user specified "return-path" capability would
3754 * QEMU uses the return path.
3756 if (migrate_postcopy_ram() || migrate_return_path()) {
3757 if (open_return_path_on_source(s
)) {
3758 error_setg(&local_err
, "Unable to open return-path for postcopy");
3764 * This needs to be done before resuming a postcopy. Note: for newer
3765 * QEMUs we will delay the channel creation until postcopy_start(), to
3766 * avoid disorder of channel creations.
3768 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
3769 postcopy_preempt_setup(s
);
3773 /* Wakeup the main migration thread to do the recovery */
3774 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
3775 MIGRATION_STATUS_POSTCOPY_RECOVER
);
3776 qemu_sem_post(&s
->postcopy_pause_sem
);
3780 if (migrate_mode_is_cpr(s
)) {
3781 ret
= migration_stop_vm(s
, RUN_STATE_FINISH_MIGRATE
);
3783 error_setg(&local_err
, "migration_stop_vm failed, error %d", -ret
);
3788 if (migrate_background_snapshot()) {
3789 qemu_thread_create(&s
->thread
, "bg_snapshot",
3790 bg_migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3792 qemu_thread_create(&s
->thread
, "live_migration",
3793 migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3795 s
->migration_thread_running
= true;
3799 migrate_set_error(s
, local_err
);
3800 migrate_set_state(&s
->state
, s
->state
, MIGRATION_STATUS_FAILED
);
3801 error_report_err(local_err
);
3802 migrate_fd_cleanup(s
);
3805 static void migration_class_init(ObjectClass
*klass
, void *data
)
3807 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3809 dc
->user_creatable
= false;
3810 device_class_set_props(dc
, migration_properties
);
3813 static void migration_instance_finalize(Object
*obj
)
3815 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3817 qemu_mutex_destroy(&ms
->error_mutex
);
3818 qemu_mutex_destroy(&ms
->qemu_file_lock
);
3819 qemu_sem_destroy(&ms
->wait_unplug_sem
);
3820 qemu_sem_destroy(&ms
->rate_limit_sem
);
3821 qemu_sem_destroy(&ms
->pause_sem
);
3822 qemu_sem_destroy(&ms
->postcopy_pause_sem
);
3823 qemu_sem_destroy(&ms
->rp_state
.rp_sem
);
3824 qemu_sem_destroy(&ms
->rp_state
.rp_pong_acks
);
3825 qemu_sem_destroy(&ms
->postcopy_qemufile_src_sem
);
3826 error_free(ms
->error
);
3829 static void migration_instance_init(Object
*obj
)
3831 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3833 ms
->state
= MIGRATION_STATUS_NONE
;
3835 ms
->pages_per_second
= -1;
3836 qemu_sem_init(&ms
->pause_sem
, 0);
3837 qemu_mutex_init(&ms
->error_mutex
);
3839 migrate_params_init(&ms
->parameters
);
3841 qemu_sem_init(&ms
->postcopy_pause_sem
, 0);
3842 qemu_sem_init(&ms
->rp_state
.rp_sem
, 0);
3843 qemu_sem_init(&ms
->rp_state
.rp_pong_acks
, 0);
3844 qemu_sem_init(&ms
->rate_limit_sem
, 0);
3845 qemu_sem_init(&ms
->wait_unplug_sem
, 0);
3846 qemu_sem_init(&ms
->postcopy_qemufile_src_sem
, 0);
3847 qemu_mutex_init(&ms
->qemu_file_lock
);
3851 * Return true if check pass, false otherwise. Error will be put
3852 * inside errp if provided.
3854 static bool migration_object_check(MigrationState
*ms
, Error
**errp
)
3856 /* Assuming all off */
3857 bool old_caps
[MIGRATION_CAPABILITY__MAX
] = { 0 };
3859 if (!migrate_params_check(&ms
->parameters
, errp
)) {
3863 return migrate_caps_check(old_caps
, ms
->capabilities
, errp
);
3866 static const TypeInfo migration_type
= {
3867 .name
= TYPE_MIGRATION
,
3869 * NOTE: TYPE_MIGRATION is not really a device, as the object is
3870 * not created using qdev_new(), it is not attached to the qdev
3871 * device tree, and it is never realized.
3873 * TODO: Make this TYPE_OBJECT once QOM provides something like
3874 * TYPE_DEVICE's "-global" properties.
3876 .parent
= TYPE_DEVICE
,
3877 .class_init
= migration_class_init
,
3878 .class_size
= sizeof(MigrationClass
),
3879 .instance_size
= sizeof(MigrationState
),
3880 .instance_init
= migration_instance_init
,
3881 .instance_finalize
= migration_instance_finalize
,
3884 static void register_migration_types(void)
3886 type_register_static(&migration_type
);
3889 type_init(register_migration_types
);