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"
71 static NotifierList migration_state_notifiers
=
72 NOTIFIER_LIST_INITIALIZER(migration_state_notifiers
);
74 /* Messages sent on the return path from destination to source */
75 enum mig_rp_message_type
{
76 MIG_RP_MSG_INVALID
= 0, /* Must be 0 */
77 MIG_RP_MSG_SHUT
, /* sibling will not send any more RP messages */
78 MIG_RP_MSG_PONG
, /* Response to a PING; data (seq: be32 ) */
80 MIG_RP_MSG_REQ_PAGES_ID
, /* data (start: be64, len: be32, id: string) */
81 MIG_RP_MSG_REQ_PAGES
, /* data (start: be64, len: be32) */
82 MIG_RP_MSG_RECV_BITMAP
, /* send recved_bitmap back to source */
83 MIG_RP_MSG_RESUME_ACK
, /* tell source that we are ready to resume */
84 MIG_RP_MSG_SWITCHOVER_ACK
, /* Tell source it's OK to do switchover */
89 /* When we add fault tolerance, we could have several
90 migrations at once. For now we don't need to add
91 dynamic creation of migration */
93 static MigrationState
*current_migration
;
94 static MigrationIncomingState
*current_incoming
;
96 static GSList
*migration_blockers
[MIG_MODE__MAX
];
98 static bool migration_object_check(MigrationState
*ms
, Error
**errp
);
99 static int migration_maybe_pause(MigrationState
*s
,
100 int *current_active_state
,
102 static void migrate_fd_cancel(MigrationState
*s
);
103 static bool close_return_path_on_source(MigrationState
*s
);
105 static void migration_downtime_start(MigrationState
*s
)
107 trace_vmstate_downtime_checkpoint("src-downtime-start");
108 s
->downtime_start
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
111 static void migration_downtime_end(MigrationState
*s
)
113 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
116 * If downtime already set, should mean that postcopy already set it,
117 * then that should be the real downtime already.
120 s
->downtime
= now
- s
->downtime_start
;
123 trace_vmstate_downtime_checkpoint("src-downtime-end");
126 static bool migration_needs_multiple_sockets(void)
128 return migrate_multifd() || migrate_postcopy_preempt();
131 static bool transport_supports_multi_channels(SocketAddress
*saddr
)
133 return saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
134 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
135 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
;
139 migration_channels_and_transport_compatible(MigrationAddress
*addr
,
142 if (migration_needs_multiple_sockets() &&
143 (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) &&
144 !transport_supports_multi_channels(&addr
->u
.socket
)) {
145 error_setg(errp
, "Migration requires multi-channel URIs (e.g. tcp)");
152 static gint
page_request_addr_cmp(gconstpointer ap
, gconstpointer bp
)
154 uintptr_t a
= (uintptr_t) ap
, b
= (uintptr_t) bp
;
156 return (a
> b
) - (a
< b
);
159 int migration_stop_vm(RunState state
)
161 int ret
= vm_stop_force_state(state
);
163 trace_vmstate_downtime_checkpoint("src-vm-stopped");
168 void migration_object_init(void)
170 /* This can only be called once. */
171 assert(!current_migration
);
172 current_migration
= MIGRATION_OBJ(object_new(TYPE_MIGRATION
));
175 * Init the migrate incoming object as well no matter whether
176 * we'll use it or not.
178 assert(!current_incoming
);
179 current_incoming
= g_new0(MigrationIncomingState
, 1);
180 current_incoming
->state
= MIGRATION_STATUS_NONE
;
181 current_incoming
->postcopy_remote_fds
=
182 g_array_new(FALSE
, TRUE
, sizeof(struct PostCopyFD
));
183 qemu_mutex_init(¤t_incoming
->rp_mutex
);
184 qemu_mutex_init(¤t_incoming
->postcopy_prio_thread_mutex
);
185 qemu_event_init(¤t_incoming
->main_thread_load_event
, false);
186 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_dst
, 0);
187 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fault
, 0);
188 qemu_sem_init(¤t_incoming
->postcopy_pause_sem_fast_load
, 0);
189 qemu_sem_init(¤t_incoming
->postcopy_qemufile_dst_done
, 0);
191 qemu_mutex_init(¤t_incoming
->page_request_mutex
);
192 qemu_cond_init(¤t_incoming
->page_request_cond
);
193 current_incoming
->page_requested
= g_tree_new(page_request_addr_cmp
);
195 migration_object_check(current_migration
, &error_fatal
);
199 dirty_bitmap_mig_init();
202 void migration_cancel(const Error
*error
)
205 migrate_set_error(current_migration
, error
);
207 if (migrate_dirty_limit()) {
208 qmp_cancel_vcpu_dirty_limit(false, -1, NULL
);
210 migrate_fd_cancel(current_migration
);
213 void migration_shutdown(void)
216 * When the QEMU main thread exit, the COLO thread
217 * may wait a semaphore. So, we should wakeup the
218 * COLO thread before migration shutdown.
222 * Cancel the current migration - that will (eventually)
223 * stop the migration using this structure
225 migration_cancel(NULL
);
226 object_unref(OBJECT(current_migration
));
229 * Cancel outgoing migration of dirty bitmaps. It should
230 * at least unref used block nodes.
232 dirty_bitmap_mig_cancel_outgoing();
235 * Cancel incoming migration of dirty bitmaps. Dirty bitmaps
236 * are non-critical data, and their loss never considered as
239 dirty_bitmap_mig_cancel_incoming();
243 MigrationState
*migrate_get_current(void)
245 /* This can only be called after the object created. */
246 assert(current_migration
);
247 return current_migration
;
250 MigrationIncomingState
*migration_incoming_get_current(void)
252 assert(current_incoming
);
253 return current_incoming
;
256 void migration_incoming_transport_cleanup(MigrationIncomingState
*mis
)
258 if (mis
->socket_address_list
) {
259 qapi_free_SocketAddressList(mis
->socket_address_list
);
260 mis
->socket_address_list
= NULL
;
263 if (mis
->transport_cleanup
) {
264 mis
->transport_cleanup(mis
->transport_data
);
265 mis
->transport_data
= mis
->transport_cleanup
= NULL
;
269 void migration_incoming_state_destroy(void)
271 struct MigrationIncomingState
*mis
= migration_incoming_get_current();
273 multifd_load_cleanup();
274 compress_threads_load_cleanup();
276 if (mis
->to_src_file
) {
277 /* Tell source that we are done */
278 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
279 qemu_fclose(mis
->to_src_file
);
280 mis
->to_src_file
= NULL
;
283 if (mis
->from_src_file
) {
284 migration_ioc_unregister_yank_from_file(mis
->from_src_file
);
285 qemu_fclose(mis
->from_src_file
);
286 mis
->from_src_file
= NULL
;
288 if (mis
->postcopy_remote_fds
) {
289 g_array_free(mis
->postcopy_remote_fds
, TRUE
);
290 mis
->postcopy_remote_fds
= NULL
;
293 migration_incoming_transport_cleanup(mis
);
294 qemu_event_reset(&mis
->main_thread_load_event
);
296 if (mis
->page_requested
) {
297 g_tree_destroy(mis
->page_requested
);
298 mis
->page_requested
= NULL
;
301 if (mis
->postcopy_qemufile_dst
) {
302 migration_ioc_unregister_yank_from_file(mis
->postcopy_qemufile_dst
);
303 qemu_fclose(mis
->postcopy_qemufile_dst
);
304 mis
->postcopy_qemufile_dst
= NULL
;
307 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
310 static void migrate_generate_event(int new_state
)
312 if (migrate_events()) {
313 qapi_event_send_migration(new_state
);
318 * Send a message on the return channel back to the source
321 static int migrate_send_rp_message(MigrationIncomingState
*mis
,
322 enum mig_rp_message_type message_type
,
323 uint16_t len
, void *data
)
327 trace_migrate_send_rp_message((int)message_type
, len
);
328 QEMU_LOCK_GUARD(&mis
->rp_mutex
);
331 * It's possible that the file handle got lost due to network
334 if (!mis
->to_src_file
) {
339 qemu_put_be16(mis
->to_src_file
, (unsigned int)message_type
);
340 qemu_put_be16(mis
->to_src_file
, len
);
341 qemu_put_buffer(mis
->to_src_file
, data
, len
);
342 return qemu_fflush(mis
->to_src_file
);
345 /* Request one page from the source VM at the given start address.
346 * rb: the RAMBlock to request the page in
347 * Start: Address offset within the RB
348 * Len: Length in bytes required - must be a multiple of pagesize
350 int migrate_send_rp_message_req_pages(MigrationIncomingState
*mis
,
351 RAMBlock
*rb
, ram_addr_t start
)
353 uint8_t bufc
[12 + 1 + 255]; /* start (8), len (4), rbname up to 256 */
354 size_t msglen
= 12; /* start + len */
355 size_t len
= qemu_ram_pagesize(rb
);
356 enum mig_rp_message_type msg_type
;
360 *(uint64_t *)bufc
= cpu_to_be64((uint64_t)start
);
361 *(uint32_t *)(bufc
+ 8) = cpu_to_be32((uint32_t)len
);
364 * We maintain the last ramblock that we requested for page. Note that we
365 * don't need locking because this function will only be called within the
366 * postcopy ram fault thread.
368 if (rb
!= mis
->last_rb
) {
371 rbname
= qemu_ram_get_idstr(rb
);
372 rbname_len
= strlen(rbname
);
374 assert(rbname_len
< 256);
376 bufc
[msglen
++] = rbname_len
;
377 memcpy(bufc
+ msglen
, rbname
, rbname_len
);
378 msglen
+= rbname_len
;
379 msg_type
= MIG_RP_MSG_REQ_PAGES_ID
;
381 msg_type
= MIG_RP_MSG_REQ_PAGES
;
384 return migrate_send_rp_message(mis
, msg_type
, msglen
, bufc
);
387 int migrate_send_rp_req_pages(MigrationIncomingState
*mis
,
388 RAMBlock
*rb
, ram_addr_t start
, uint64_t haddr
)
390 void *aligned
= (void *)(uintptr_t)ROUND_DOWN(haddr
, qemu_ram_pagesize(rb
));
391 bool received
= false;
393 WITH_QEMU_LOCK_GUARD(&mis
->page_request_mutex
) {
394 received
= ramblock_recv_bitmap_test_byte_offset(rb
, start
);
395 if (!received
&& !g_tree_lookup(mis
->page_requested
, aligned
)) {
397 * The page has not been received, and it's not yet in the page
398 * request list. Queue it. Set the value of element to 1, so that
399 * things like g_tree_lookup() will return TRUE (1) when found.
401 g_tree_insert(mis
->page_requested
, aligned
, (gpointer
)1);
402 qatomic_inc(&mis
->page_requested_count
);
403 trace_postcopy_page_req_add(aligned
, mis
->page_requested_count
);
408 * If the page is there, skip sending the message. We don't even need the
409 * lock because as long as the page arrived, it'll be there forever.
415 return migrate_send_rp_message_req_pages(mis
, rb
, start
);
418 static bool migration_colo_enabled
;
419 bool migration_incoming_colo_enabled(void)
421 return migration_colo_enabled
;
424 void migration_incoming_disable_colo(void)
426 ram_block_discard_disable(false);
427 migration_colo_enabled
= false;
430 int migration_incoming_enable_colo(void)
432 #ifndef CONFIG_REPLICATION
433 error_report("ENABLE_COLO command come in migration stream, but COLO "
434 "module is not built in");
438 if (!migrate_colo()) {
439 error_report("ENABLE_COLO command come in migration stream, but c-colo "
440 "capability is not set");
444 if (ram_block_discard_disable(true)) {
445 error_report("COLO: cannot disable RAM discard");
448 migration_colo_enabled
= true;
452 void migrate_add_address(SocketAddress
*address
)
454 MigrationIncomingState
*mis
= migration_incoming_get_current();
456 QAPI_LIST_PREPEND(mis
->socket_address_list
,
457 QAPI_CLONE(SocketAddress
, address
));
460 bool migrate_uri_parse(const char *uri
, MigrationChannel
**channel
,
463 g_autoptr(MigrationChannel
) val
= g_new0(MigrationChannel
, 1);
464 g_autoptr(MigrationAddress
) addr
= g_new0(MigrationAddress
, 1);
465 SocketAddress
*saddr
= NULL
;
466 InetSocketAddress
*isock
= &addr
->u
.rdma
;
467 strList
**tail
= &addr
->u
.exec
.args
;
469 if (strstart(uri
, "exec:", NULL
)) {
470 addr
->transport
= MIGRATION_ADDRESS_TYPE_EXEC
;
472 QAPI_LIST_APPEND(tail
, g_strdup(exec_get_cmd_path()));
473 QAPI_LIST_APPEND(tail
, g_strdup("/c"));
475 QAPI_LIST_APPEND(tail
, g_strdup("/bin/sh"));
476 QAPI_LIST_APPEND(tail
, g_strdup("-c"));
478 QAPI_LIST_APPEND(tail
, g_strdup(uri
+ strlen("exec:")));
479 } else if (strstart(uri
, "rdma:", NULL
)) {
480 if (inet_parse(isock
, uri
+ strlen("rdma:"), errp
)) {
481 qapi_free_InetSocketAddress(isock
);
484 addr
->transport
= MIGRATION_ADDRESS_TYPE_RDMA
;
485 } else if (strstart(uri
, "tcp:", NULL
) ||
486 strstart(uri
, "unix:", NULL
) ||
487 strstart(uri
, "vsock:", NULL
) ||
488 strstart(uri
, "fd:", NULL
)) {
489 addr
->transport
= MIGRATION_ADDRESS_TYPE_SOCKET
;
490 saddr
= socket_parse(uri
, errp
);
494 addr
->u
.socket
.type
= saddr
->type
;
495 addr
->u
.socket
.u
= saddr
->u
;
496 } else if (strstart(uri
, "file:", NULL
)) {
497 addr
->transport
= MIGRATION_ADDRESS_TYPE_FILE
;
498 addr
->u
.file
.filename
= g_strdup(uri
+ strlen("file:"));
499 if (file_parse_offset(addr
->u
.file
.filename
, &addr
->u
.file
.offset
,
504 error_setg(errp
, "unknown migration protocol: %s", uri
);
508 val
->channel_type
= MIGRATION_CHANNEL_TYPE_MAIN
;
509 val
->addr
= g_steal_pointer(&addr
);
510 *channel
= g_steal_pointer(&val
);
514 static void qemu_start_incoming_migration(const char *uri
, bool has_channels
,
515 MigrationChannelList
*channels
,
518 MigrationChannel
*channel
= NULL
;
519 MigrationAddress
*addr
= NULL
;
520 MigrationIncomingState
*mis
= migration_incoming_get_current();
523 * Having preliminary checks for uri and channel
525 if (uri
&& has_channels
) {
526 error_setg(errp
, "'uri' and 'channels' arguments are mutually "
527 "exclusive; exactly one of the two should be present in "
528 "'migrate-incoming' qmp command ");
530 } else if (channels
) {
531 /* To verify that Migrate channel list has only item */
532 if (channels
->next
) {
533 error_setg(errp
, "Channel list has more than one entries");
536 channel
= channels
->value
;
538 /* caller uses the old URI syntax */
539 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
543 error_setg(errp
, "neither 'uri' or 'channels' argument are "
544 "specified in 'migrate-incoming' qmp command ");
547 addr
= channel
->addr
;
549 /* transport mechanism not suitable for migration? */
550 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
554 migrate_set_state(&mis
->state
, MIGRATION_STATUS_NONE
,
555 MIGRATION_STATUS_SETUP
);
557 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
558 SocketAddress
*saddr
= &addr
->u
.socket
;
559 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
560 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
561 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
562 socket_start_incoming_migration(saddr
, errp
);
563 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
564 fd_start_incoming_migration(saddr
->u
.fd
.str
, errp
);
567 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
568 if (migrate_compress()) {
569 error_setg(errp
, "RDMA and compression can't be used together");
572 if (migrate_xbzrle()) {
573 error_setg(errp
, "RDMA and XBZRLE can't be used together");
576 if (migrate_multifd()) {
577 error_setg(errp
, "RDMA and multifd can't be used together");
580 rdma_start_incoming_migration(&addr
->u
.rdma
, errp
);
582 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
583 exec_start_incoming_migration(addr
->u
.exec
.args
, errp
);
584 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
585 file_start_incoming_migration(&addr
->u
.file
, errp
);
587 error_setg(errp
, "unknown migration protocol: %s", uri
);
591 static void process_incoming_migration_bh(void *opaque
)
593 Error
*local_err
= NULL
;
594 MigrationIncomingState
*mis
= opaque
;
596 trace_vmstate_downtime_checkpoint("dst-precopy-bh-enter");
598 /* If capability late_block_activate is set:
599 * Only fire up the block code now if we're going to restart the
600 * VM, else 'cont' will do it.
601 * This causes file locking to happen; so we don't want it to happen
602 * unless we really are starting the VM.
604 if (!migrate_late_block_activate() ||
605 (autostart
&& (!global_state_received() ||
606 global_state_get_runstate() == RUN_STATE_RUNNING
))) {
607 /* Make sure all file formats throw away their mutable metadata.
608 * If we get an error here, just don't restart the VM yet. */
609 bdrv_activate_all(&local_err
);
611 error_report_err(local_err
);
618 * This must happen after all error conditions are dealt with and
619 * we're sure the VM is going to be running on this host.
621 qemu_announce_self(&mis
->announce_timer
, migrate_announce_params());
623 trace_vmstate_downtime_checkpoint("dst-precopy-bh-announced");
625 multifd_load_shutdown();
627 dirty_bitmap_mig_before_vm_start();
629 if (!global_state_received() ||
630 global_state_get_runstate() == RUN_STATE_RUNNING
) {
634 runstate_set(RUN_STATE_PAUSED
);
636 } else if (migration_incoming_colo_enabled()) {
637 migration_incoming_disable_colo();
640 runstate_set(global_state_get_runstate());
642 trace_vmstate_downtime_checkpoint("dst-precopy-bh-vm-started");
644 * This must happen after any state changes since as soon as an external
645 * observer sees this event they might start to prod at the VM assuming
648 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
649 MIGRATION_STATUS_COMPLETED
);
650 qemu_bh_delete(mis
->bh
);
651 migration_incoming_state_destroy();
654 static void coroutine_fn
655 process_incoming_migration_co(void *opaque
)
657 MigrationIncomingState
*mis
= migration_incoming_get_current();
661 assert(mis
->from_src_file
);
663 if (compress_threads_load_setup(mis
->from_src_file
)) {
664 error_report("Failed to setup decompress threads");
668 mis
->largest_page_size
= qemu_ram_pagesize_largest();
669 postcopy_state_set(POSTCOPY_INCOMING_NONE
);
670 migrate_set_state(&mis
->state
, MIGRATION_STATUS_SETUP
,
671 MIGRATION_STATUS_ACTIVE
);
673 mis
->loadvm_co
= qemu_coroutine_self();
674 ret
= qemu_loadvm_state(mis
->from_src_file
);
675 mis
->loadvm_co
= NULL
;
677 trace_vmstate_downtime_checkpoint("dst-precopy-loadvm-completed");
679 ps
= postcopy_state_get();
680 trace_process_incoming_migration_co_end(ret
, ps
);
681 if (ps
!= POSTCOPY_INCOMING_NONE
) {
682 if (ps
== POSTCOPY_INCOMING_ADVISE
) {
684 * Where a migration had postcopy enabled (and thus went to advise)
685 * but managed to complete within the precopy period, we can use
688 postcopy_ram_incoming_cleanup(mis
);
689 } else if (ret
>= 0) {
691 * Postcopy was started, cleanup should happen at the end of the
694 trace_process_incoming_migration_co_postcopy_end_main();
697 /* Else if something went wrong then just fall out of the normal exit */
701 error_report("load of migration failed: %s", strerror(-ret
));
705 if (colo_incoming_co() < 0) {
709 mis
->bh
= qemu_bh_new(process_incoming_migration_bh
, mis
);
710 qemu_bh_schedule(mis
->bh
);
713 migrate_set_state(&mis
->state
, MIGRATION_STATUS_ACTIVE
,
714 MIGRATION_STATUS_FAILED
);
715 qemu_fclose(mis
->from_src_file
);
717 multifd_load_cleanup();
718 compress_threads_load_cleanup();
724 * migration_incoming_setup: Setup incoming migration
725 * @f: file for main migration channel
726 * @errp: where to put errors
728 * Returns: %true on success, %false on error.
730 static bool migration_incoming_setup(QEMUFile
*f
, Error
**errp
)
732 MigrationIncomingState
*mis
= migration_incoming_get_current();
734 if (!mis
->from_src_file
) {
735 mis
->from_src_file
= f
;
737 qemu_file_set_blocking(f
, false);
741 void migration_incoming_process(void)
743 Coroutine
*co
= qemu_coroutine_create(process_incoming_migration_co
, NULL
);
744 qemu_coroutine_enter(co
);
747 /* Returns true if recovered from a paused migration, otherwise false */
748 static bool postcopy_try_recover(void)
750 MigrationIncomingState
*mis
= migration_incoming_get_current();
752 if (mis
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
753 /* Resumed from a paused postcopy migration */
755 /* This should be set already in migration_incoming_setup() */
756 assert(mis
->from_src_file
);
757 /* Postcopy has standalone thread to do vm load */
758 qemu_file_set_blocking(mis
->from_src_file
, true);
760 /* Re-configure the return path */
761 mis
->to_src_file
= qemu_file_get_return_path(mis
->from_src_file
);
763 migrate_set_state(&mis
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
764 MIGRATION_STATUS_POSTCOPY_RECOVER
);
767 * Here, we only wake up the main loading thread (while the
768 * rest threads will still be waiting), so that we can receive
769 * commands from source now, and answer it if needed. The
770 * rest threads will be woken up afterwards until we are sure
771 * that source is ready to reply to page requests.
773 qemu_sem_post(&mis
->postcopy_pause_sem_dst
);
780 void migration_fd_process_incoming(QEMUFile
*f
, Error
**errp
)
782 if (!migration_incoming_setup(f
, errp
)) {
785 if (postcopy_try_recover()) {
788 migration_incoming_process();
792 * Returns true when we want to start a new incoming migration process,
795 static bool migration_should_start_incoming(bool main_channel
)
797 /* Multifd doesn't start unless all channels are established */
798 if (migrate_multifd()) {
799 return migration_has_all_channels();
802 /* Preempt channel only starts when the main channel is created */
803 if (migrate_postcopy_preempt()) {
808 * For all the rest types of migration, we should only reach here when
809 * it's the main channel that's being created, and we should always
810 * proceed with this channel.
812 assert(main_channel
);
816 void migration_ioc_process_incoming(QIOChannel
*ioc
, Error
**errp
)
818 MigrationIncomingState
*mis
= migration_incoming_get_current();
819 Error
*local_err
= NULL
;
821 bool default_channel
= true;
822 uint32_t channel_magic
= 0;
825 if (migrate_multifd() && !migrate_postcopy_ram() &&
826 qio_channel_has_feature(ioc
, QIO_CHANNEL_FEATURE_READ_MSG_PEEK
)) {
828 * With multiple channels, it is possible that we receive channels
829 * out of order on destination side, causing incorrect mapping of
830 * source channels on destination side. Check channel MAGIC to
831 * decide type of channel. Please note this is best effort, postcopy
832 * preempt channel does not send any magic number so avoid it for
833 * postcopy live migration. Also tls live migration already does
834 * tls handshake while initializing main channel so with tls this
835 * issue is not possible.
837 ret
= migration_channel_read_peek(ioc
, (void *)&channel_magic
,
838 sizeof(channel_magic
), &local_err
);
841 error_propagate(errp
, local_err
);
845 default_channel
= (channel_magic
== cpu_to_be32(QEMU_VM_FILE_MAGIC
));
847 default_channel
= !mis
->from_src_file
;
850 if (multifd_load_setup(errp
) != 0) {
851 error_setg(errp
, "Failed to setup multifd channels");
855 if (default_channel
) {
856 f
= qemu_file_new_input(ioc
);
858 if (!migration_incoming_setup(f
, errp
)) {
862 /* Multiple connections */
863 assert(migration_needs_multiple_sockets());
864 if (migrate_multifd()) {
865 multifd_recv_new_channel(ioc
, &local_err
);
867 assert(migrate_postcopy_preempt());
868 f
= qemu_file_new_input(ioc
);
869 postcopy_preempt_new_channel(mis
, f
);
872 error_propagate(errp
, local_err
);
877 if (migration_should_start_incoming(default_channel
)) {
878 /* If it's a recovery, we're done */
879 if (postcopy_try_recover()) {
882 migration_incoming_process();
887 * @migration_has_all_channels: We have received all channels that we need
889 * Returns true when we have got connections to all the channels that
890 * we need for migration.
892 bool migration_has_all_channels(void)
894 MigrationIncomingState
*mis
= migration_incoming_get_current();
896 if (!mis
->from_src_file
) {
900 if (migrate_multifd()) {
901 return multifd_recv_all_channels_created();
904 if (migrate_postcopy_preempt()) {
905 return mis
->postcopy_qemufile_dst
!= NULL
;
911 int migrate_send_rp_switchover_ack(MigrationIncomingState
*mis
)
913 return migrate_send_rp_message(mis
, MIG_RP_MSG_SWITCHOVER_ACK
, 0, NULL
);
917 * Send a 'SHUT' message on the return channel with the given value
918 * to indicate that we've finished with the RP. Non-0 value indicates
921 void migrate_send_rp_shut(MigrationIncomingState
*mis
,
926 buf
= cpu_to_be32(value
);
927 migrate_send_rp_message(mis
, MIG_RP_MSG_SHUT
, sizeof(buf
), &buf
);
931 * Send a 'PONG' message on the return channel with the given value
932 * (normally in response to a 'PING')
934 void migrate_send_rp_pong(MigrationIncomingState
*mis
,
939 buf
= cpu_to_be32(value
);
940 migrate_send_rp_message(mis
, MIG_RP_MSG_PONG
, sizeof(buf
), &buf
);
943 void migrate_send_rp_recv_bitmap(MigrationIncomingState
*mis
,
951 * First, we send the header part. It contains only the len of
952 * idstr, and the idstr itself.
954 len
= strlen(block_name
);
956 memcpy(buf
+ 1, block_name
, len
);
958 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
959 error_report("%s: MSG_RP_RECV_BITMAP only used for recovery",
964 migrate_send_rp_message(mis
, MIG_RP_MSG_RECV_BITMAP
, len
+ 1, buf
);
967 * Next, we dump the received bitmap to the stream.
969 * TODO: currently we are safe since we are the only one that is
970 * using the to_src_file handle (fault thread is still paused),
971 * and it's ok even not taking the mutex. However the best way is
972 * to take the lock before sending the message header, and release
973 * the lock after sending the bitmap.
975 qemu_mutex_lock(&mis
->rp_mutex
);
976 res
= ramblock_recv_bitmap_send(mis
->to_src_file
, block_name
);
977 qemu_mutex_unlock(&mis
->rp_mutex
);
979 trace_migrate_send_rp_recv_bitmap(block_name
, res
);
982 void migrate_send_rp_resume_ack(MigrationIncomingState
*mis
, uint32_t value
)
986 buf
= cpu_to_be32(value
);
987 migrate_send_rp_message(mis
, MIG_RP_MSG_RESUME_ACK
, sizeof(buf
), &buf
);
991 * Return true if we're already in the middle of a migration
992 * (i.e. any of the active or setup states)
994 bool migration_is_setup_or_active(int state
)
997 case MIGRATION_STATUS_ACTIVE
:
998 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
999 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1000 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1001 case MIGRATION_STATUS_SETUP
:
1002 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1003 case MIGRATION_STATUS_DEVICE
:
1004 case MIGRATION_STATUS_WAIT_UNPLUG
:
1005 case MIGRATION_STATUS_COLO
:
1014 bool migration_is_running(int state
)
1017 case MIGRATION_STATUS_ACTIVE
:
1018 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1019 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1020 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1021 case MIGRATION_STATUS_SETUP
:
1022 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1023 case MIGRATION_STATUS_DEVICE
:
1024 case MIGRATION_STATUS_WAIT_UNPLUG
:
1025 case MIGRATION_STATUS_CANCELLING
:
1034 static bool migrate_show_downtime(MigrationState
*s
)
1036 return (s
->state
== MIGRATION_STATUS_COMPLETED
) || migration_in_postcopy();
1039 static void populate_time_info(MigrationInfo
*info
, MigrationState
*s
)
1041 info
->has_status
= true;
1042 info
->has_setup_time
= true;
1043 info
->setup_time
= s
->setup_time
;
1045 if (s
->state
== MIGRATION_STATUS_COMPLETED
) {
1046 info
->has_total_time
= true;
1047 info
->total_time
= s
->total_time
;
1049 info
->has_total_time
= true;
1050 info
->total_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) -
1054 if (migrate_show_downtime(s
)) {
1055 info
->has_downtime
= true;
1056 info
->downtime
= s
->downtime
;
1058 info
->has_expected_downtime
= true;
1059 info
->expected_downtime
= s
->expected_downtime
;
1063 static void populate_ram_info(MigrationInfo
*info
, MigrationState
*s
)
1065 size_t page_size
= qemu_target_page_size();
1067 info
->ram
= g_malloc0(sizeof(*info
->ram
));
1068 info
->ram
->transferred
= migration_transferred_bytes();
1069 info
->ram
->total
= ram_bytes_total();
1070 info
->ram
->duplicate
= stat64_get(&mig_stats
.zero_pages
);
1071 /* legacy value. It is not used anymore */
1072 info
->ram
->skipped
= 0;
1073 info
->ram
->normal
= stat64_get(&mig_stats
.normal_pages
);
1074 info
->ram
->normal_bytes
= info
->ram
->normal
* page_size
;
1075 info
->ram
->mbps
= s
->mbps
;
1076 info
->ram
->dirty_sync_count
=
1077 stat64_get(&mig_stats
.dirty_sync_count
);
1078 info
->ram
->dirty_sync_missed_zero_copy
=
1079 stat64_get(&mig_stats
.dirty_sync_missed_zero_copy
);
1080 info
->ram
->postcopy_requests
=
1081 stat64_get(&mig_stats
.postcopy_requests
);
1082 info
->ram
->page_size
= page_size
;
1083 info
->ram
->multifd_bytes
= stat64_get(&mig_stats
.multifd_bytes
);
1084 info
->ram
->pages_per_second
= s
->pages_per_second
;
1085 info
->ram
->precopy_bytes
= stat64_get(&mig_stats
.precopy_bytes
);
1086 info
->ram
->downtime_bytes
= stat64_get(&mig_stats
.downtime_bytes
);
1087 info
->ram
->postcopy_bytes
= stat64_get(&mig_stats
.postcopy_bytes
);
1089 if (migrate_xbzrle()) {
1090 info
->xbzrle_cache
= g_malloc0(sizeof(*info
->xbzrle_cache
));
1091 info
->xbzrle_cache
->cache_size
= migrate_xbzrle_cache_size();
1092 info
->xbzrle_cache
->bytes
= xbzrle_counters
.bytes
;
1093 info
->xbzrle_cache
->pages
= xbzrle_counters
.pages
;
1094 info
->xbzrle_cache
->cache_miss
= xbzrle_counters
.cache_miss
;
1095 info
->xbzrle_cache
->cache_miss_rate
= xbzrle_counters
.cache_miss_rate
;
1096 info
->xbzrle_cache
->encoding_rate
= xbzrle_counters
.encoding_rate
;
1097 info
->xbzrle_cache
->overflow
= xbzrle_counters
.overflow
;
1100 populate_compress(info
);
1102 if (cpu_throttle_active()) {
1103 info
->has_cpu_throttle_percentage
= true;
1104 info
->cpu_throttle_percentage
= cpu_throttle_get_percentage();
1107 if (s
->state
!= MIGRATION_STATUS_COMPLETED
) {
1108 info
->ram
->remaining
= ram_bytes_remaining();
1109 info
->ram
->dirty_pages_rate
=
1110 stat64_get(&mig_stats
.dirty_pages_rate
);
1113 if (migrate_dirty_limit() && dirtylimit_in_service()) {
1114 info
->has_dirty_limit_throttle_time_per_round
= true;
1115 info
->dirty_limit_throttle_time_per_round
=
1116 dirtylimit_throttle_time_per_round();
1118 info
->has_dirty_limit_ring_full_time
= true;
1119 info
->dirty_limit_ring_full_time
= dirtylimit_ring_full_time();
1123 static void populate_disk_info(MigrationInfo
*info
)
1125 if (blk_mig_active()) {
1126 info
->disk
= g_malloc0(sizeof(*info
->disk
));
1127 info
->disk
->transferred
= blk_mig_bytes_transferred();
1128 info
->disk
->remaining
= blk_mig_bytes_remaining();
1129 info
->disk
->total
= blk_mig_bytes_total();
1133 static void fill_source_migration_info(MigrationInfo
*info
)
1135 MigrationState
*s
= migrate_get_current();
1136 int state
= qatomic_read(&s
->state
);
1137 GSList
*cur_blocker
= migration_blockers
[migrate_mode()];
1139 info
->blocked_reasons
= NULL
;
1142 * There are two types of reasons a migration might be blocked;
1143 * a) devices marked in VMState as non-migratable, and
1144 * b) Explicit migration blockers
1145 * We need to add both of them here.
1147 qemu_savevm_non_migratable_list(&info
->blocked_reasons
);
1149 while (cur_blocker
) {
1150 QAPI_LIST_PREPEND(info
->blocked_reasons
,
1151 g_strdup(error_get_pretty(cur_blocker
->data
)));
1152 cur_blocker
= g_slist_next(cur_blocker
);
1154 info
->has_blocked_reasons
= info
->blocked_reasons
!= NULL
;
1157 case MIGRATION_STATUS_NONE
:
1158 /* no migration has happened ever */
1159 /* do not overwrite destination migration status */
1161 case MIGRATION_STATUS_SETUP
:
1162 info
->has_status
= true;
1163 info
->has_total_time
= false;
1165 case MIGRATION_STATUS_ACTIVE
:
1166 case MIGRATION_STATUS_CANCELLING
:
1167 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1168 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1169 case MIGRATION_STATUS_DEVICE
:
1170 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1171 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1172 /* TODO add some postcopy stats */
1173 populate_time_info(info
, s
);
1174 populate_ram_info(info
, s
);
1175 populate_disk_info(info
);
1176 migration_populate_vfio_info(info
);
1178 case MIGRATION_STATUS_COLO
:
1179 info
->has_status
= true;
1180 /* TODO: display COLO specific information (checkpoint info etc.) */
1182 case MIGRATION_STATUS_COMPLETED
:
1183 populate_time_info(info
, s
);
1184 populate_ram_info(info
, s
);
1185 migration_populate_vfio_info(info
);
1187 case MIGRATION_STATUS_FAILED
:
1188 info
->has_status
= true;
1190 case MIGRATION_STATUS_CANCELLED
:
1191 info
->has_status
= true;
1193 case MIGRATION_STATUS_WAIT_UNPLUG
:
1194 info
->has_status
= true;
1197 info
->status
= state
;
1199 QEMU_LOCK_GUARD(&s
->error_mutex
);
1201 info
->error_desc
= g_strdup(error_get_pretty(s
->error
));
1205 static void fill_destination_migration_info(MigrationInfo
*info
)
1207 MigrationIncomingState
*mis
= migration_incoming_get_current();
1209 if (mis
->socket_address_list
) {
1210 info
->has_socket_address
= true;
1211 info
->socket_address
=
1212 QAPI_CLONE(SocketAddressList
, mis
->socket_address_list
);
1215 switch (mis
->state
) {
1216 case MIGRATION_STATUS_NONE
:
1218 case MIGRATION_STATUS_SETUP
:
1219 case MIGRATION_STATUS_CANCELLING
:
1220 case MIGRATION_STATUS_CANCELLED
:
1221 case MIGRATION_STATUS_ACTIVE
:
1222 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1223 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1224 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1225 case MIGRATION_STATUS_FAILED
:
1226 case MIGRATION_STATUS_COLO
:
1227 info
->has_status
= true;
1229 case MIGRATION_STATUS_COMPLETED
:
1230 info
->has_status
= true;
1231 fill_destination_postcopy_migration_info(info
);
1234 info
->status
= mis
->state
;
1237 MigrationInfo
*qmp_query_migrate(Error
**errp
)
1239 MigrationInfo
*info
= g_malloc0(sizeof(*info
));
1241 fill_destination_migration_info(info
);
1242 fill_source_migration_info(info
);
1247 void qmp_migrate_start_postcopy(Error
**errp
)
1249 MigrationState
*s
= migrate_get_current();
1251 if (!migrate_postcopy()) {
1252 error_setg(errp
, "Enable postcopy with migrate_set_capability before"
1253 " the start of migration");
1257 if (s
->state
== MIGRATION_STATUS_NONE
) {
1258 error_setg(errp
, "Postcopy must be started after migration has been"
1263 * we don't error if migration has finished since that would be racy
1264 * with issuing this command.
1266 qatomic_set(&s
->start_postcopy
, true);
1269 /* shared migration helpers */
1271 void migrate_set_state(int *state
, int old_state
, int new_state
)
1273 assert(new_state
< MIGRATION_STATUS__MAX
);
1274 if (qatomic_cmpxchg(state
, old_state
, new_state
) == old_state
) {
1275 trace_migrate_set_state(MigrationStatus_str(new_state
));
1276 migrate_generate_event(new_state
);
1280 static void migrate_fd_cleanup(MigrationState
*s
)
1282 qemu_bh_delete(s
->cleanup_bh
);
1283 s
->cleanup_bh
= NULL
;
1285 g_free(s
->hostname
);
1287 json_writer_free(s
->vmdesc
);
1290 qemu_savevm_state_cleanup();
1292 if (s
->to_dst_file
) {
1295 trace_migrate_fd_cleanup();
1296 qemu_mutex_unlock_iothread();
1297 if (s
->migration_thread_running
) {
1298 qemu_thread_join(&s
->thread
);
1299 s
->migration_thread_running
= false;
1301 qemu_mutex_lock_iothread();
1303 multifd_save_cleanup();
1304 qemu_mutex_lock(&s
->qemu_file_lock
);
1305 tmp
= s
->to_dst_file
;
1306 s
->to_dst_file
= NULL
;
1307 qemu_mutex_unlock(&s
->qemu_file_lock
);
1309 * Close the file handle without the lock to make sure the
1310 * critical section won't block for long.
1312 migration_ioc_unregister_yank_from_file(tmp
);
1317 * We already cleaned up to_dst_file, so errors from the return
1318 * path might be due to that, ignore them.
1320 close_return_path_on_source(s
);
1322 assert(!migration_is_active(s
));
1324 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1325 migrate_set_state(&s
->state
, MIGRATION_STATUS_CANCELLING
,
1326 MIGRATION_STATUS_CANCELLED
);
1330 /* It is used on info migrate. We can't free it */
1331 error_report_err(error_copy(s
->error
));
1333 migration_call_notifiers(s
);
1334 block_cleanup_parameters();
1335 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1338 static void migrate_fd_cleanup_schedule(MigrationState
*s
)
1341 * Ref the state for bh, because it may be called when
1342 * there're already no other refs
1344 object_ref(OBJECT(s
));
1345 qemu_bh_schedule(s
->cleanup_bh
);
1348 static void migrate_fd_cleanup_bh(void *opaque
)
1350 MigrationState
*s
= opaque
;
1351 migrate_fd_cleanup(s
);
1352 object_unref(OBJECT(s
));
1355 void migrate_set_error(MigrationState
*s
, const Error
*error
)
1357 QEMU_LOCK_GUARD(&s
->error_mutex
);
1359 s
->error
= error_copy(error
);
1363 bool migrate_has_error(MigrationState
*s
)
1365 /* The lock is not helpful here, but still follow the rule */
1366 QEMU_LOCK_GUARD(&s
->error_mutex
);
1367 return qatomic_read(&s
->error
);
1370 static void migrate_error_free(MigrationState
*s
)
1372 QEMU_LOCK_GUARD(&s
->error_mutex
);
1374 error_free(s
->error
);
1379 static void migrate_fd_error(MigrationState
*s
, const Error
*error
)
1381 trace_migrate_fd_error(error_get_pretty(error
));
1382 assert(s
->to_dst_file
== NULL
);
1383 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
1384 MIGRATION_STATUS_FAILED
);
1385 migrate_set_error(s
, error
);
1388 static void migrate_fd_cancel(MigrationState
*s
)
1392 trace_migrate_fd_cancel();
1394 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1395 if (s
->rp_state
.from_dst_file
) {
1396 /* shutdown the rp socket, so causing the rp thread to shutdown */
1397 qemu_file_shutdown(s
->rp_state
.from_dst_file
);
1402 old_state
= s
->state
;
1403 if (!migration_is_running(old_state
)) {
1406 /* If the migration is paused, kick it out of the pause */
1407 if (old_state
== MIGRATION_STATUS_PRE_SWITCHOVER
) {
1408 qemu_sem_post(&s
->pause_sem
);
1410 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_CANCELLING
);
1411 } while (s
->state
!= MIGRATION_STATUS_CANCELLING
);
1414 * If we're unlucky the migration code might be stuck somewhere in a
1415 * send/write while the network has failed and is waiting to timeout;
1416 * if we've got shutdown(2) available then we can force it to quit.
1418 if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
1419 WITH_QEMU_LOCK_GUARD(&s
->qemu_file_lock
) {
1420 if (s
->to_dst_file
) {
1421 qemu_file_shutdown(s
->to_dst_file
);
1425 if (s
->state
== MIGRATION_STATUS_CANCELLING
&& s
->block_inactive
) {
1426 Error
*local_err
= NULL
;
1428 bdrv_activate_all(&local_err
);
1430 error_report_err(local_err
);
1432 s
->block_inactive
= false;
1437 void migration_add_notifier(Notifier
*notify
,
1438 void (*func
)(Notifier
*notifier
, void *data
))
1440 notify
->notify
= func
;
1441 notifier_list_add(&migration_state_notifiers
, notify
);
1444 void migration_remove_notifier(Notifier
*notify
)
1446 if (notify
->notify
) {
1447 notifier_remove(notify
);
1448 notify
->notify
= NULL
;
1452 void migration_call_notifiers(MigrationState
*s
)
1454 notifier_list_notify(&migration_state_notifiers
, s
);
1457 bool migration_in_setup(MigrationState
*s
)
1459 return s
->state
== MIGRATION_STATUS_SETUP
;
1462 bool migration_has_finished(MigrationState
*s
)
1464 return s
->state
== MIGRATION_STATUS_COMPLETED
;
1467 bool migration_has_failed(MigrationState
*s
)
1469 return (s
->state
== MIGRATION_STATUS_CANCELLED
||
1470 s
->state
== MIGRATION_STATUS_FAILED
);
1473 bool migration_in_postcopy(void)
1475 MigrationState
*s
= migrate_get_current();
1478 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1479 case MIGRATION_STATUS_POSTCOPY_PAUSED
:
1480 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1487 bool migration_postcopy_is_alive(int state
)
1490 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1491 case MIGRATION_STATUS_POSTCOPY_RECOVER
:
1498 bool migration_in_postcopy_after_devices(MigrationState
*s
)
1500 return migration_in_postcopy() && s
->postcopy_after_devices
;
1503 bool migration_in_incoming_postcopy(void)
1505 PostcopyState ps
= postcopy_state_get();
1507 return ps
>= POSTCOPY_INCOMING_DISCARD
&& ps
< POSTCOPY_INCOMING_END
;
1510 bool migration_incoming_postcopy_advised(void)
1512 PostcopyState ps
= postcopy_state_get();
1514 return ps
>= POSTCOPY_INCOMING_ADVISE
&& ps
< POSTCOPY_INCOMING_END
;
1517 bool migration_in_bg_snapshot(void)
1519 MigrationState
*s
= migrate_get_current();
1521 return migrate_background_snapshot() &&
1522 migration_is_setup_or_active(s
->state
);
1525 bool migration_is_idle(void)
1527 MigrationState
*s
= current_migration
;
1534 case MIGRATION_STATUS_NONE
:
1535 case MIGRATION_STATUS_CANCELLED
:
1536 case MIGRATION_STATUS_COMPLETED
:
1537 case MIGRATION_STATUS_FAILED
:
1539 case MIGRATION_STATUS_SETUP
:
1540 case MIGRATION_STATUS_CANCELLING
:
1541 case MIGRATION_STATUS_ACTIVE
:
1542 case MIGRATION_STATUS_POSTCOPY_ACTIVE
:
1543 case MIGRATION_STATUS_COLO
:
1544 case MIGRATION_STATUS_PRE_SWITCHOVER
:
1545 case MIGRATION_STATUS_DEVICE
:
1546 case MIGRATION_STATUS_WAIT_UNPLUG
:
1548 case MIGRATION_STATUS__MAX
:
1549 g_assert_not_reached();
1555 bool migration_is_active(MigrationState
*s
)
1557 return (s
->state
== MIGRATION_STATUS_ACTIVE
||
1558 s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
1561 int migrate_init(MigrationState
*s
, Error
**errp
)
1565 ret
= qemu_savevm_state_prepare(errp
);
1571 * Reinitialise all migration state, except
1572 * parameters/capabilities that the user set, and
1577 s
->to_dst_file
= NULL
;
1578 s
->state
= MIGRATION_STATUS_NONE
;
1579 s
->rp_state
.from_dst_file
= NULL
;
1581 s
->pages_per_second
= 0.0;
1583 s
->expected_downtime
= 0;
1585 s
->start_postcopy
= false;
1586 s
->postcopy_after_devices
= false;
1587 s
->migration_thread_running
= false;
1588 error_free(s
->error
);
1593 migrate_set_state(&s
->state
, MIGRATION_STATUS_NONE
, MIGRATION_STATUS_SETUP
);
1595 s
->start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1597 s
->vm_old_state
= -1;
1598 s
->iteration_initial_bytes
= 0;
1599 s
->threshold_size
= 0;
1600 s
->switchover_acked
= false;
1601 s
->rdma_migration
= false;
1603 * set mig_stats memory to zero for a new migration
1605 memset(&mig_stats
, 0, sizeof(mig_stats
));
1606 migration_reset_vfio_bytes_transferred();
1611 static bool is_busy(Error
**reasonp
, Error
**errp
)
1615 /* Snapshots are similar to migrations, so check RUN_STATE_SAVE_VM too. */
1616 if (runstate_check(RUN_STATE_SAVE_VM
) || !migration_is_idle()) {
1617 error_propagate_prepend(errp
, *reasonp
,
1618 "disallowing migration blocker "
1619 "(migration/snapshot in progress) for: ");
1626 static bool is_only_migratable(Error
**reasonp
, Error
**errp
, int modes
)
1630 if (only_migratable
&& (modes
& BIT(MIG_MODE_NORMAL
))) {
1631 error_propagate_prepend(errp
, *reasonp
,
1632 "disallowing migration blocker "
1633 "(--only-migratable) for: ");
1640 static int get_modes(MigMode mode
, va_list ap
)
1644 while (mode
!= -1 && mode
!= MIG_MODE_ALL
) {
1645 assert(mode
>= MIG_MODE_NORMAL
&& mode
< MIG_MODE__MAX
);
1647 mode
= va_arg(ap
, MigMode
);
1649 if (mode
== MIG_MODE_ALL
) {
1650 modes
= BIT(MIG_MODE__MAX
) - 1;
1655 static int add_blockers(Error
**reasonp
, Error
**errp
, int modes
)
1657 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1658 if (modes
& BIT(mode
)) {
1659 migration_blockers
[mode
] = g_slist_prepend(migration_blockers
[mode
],
1666 int migrate_add_blocker(Error
**reasonp
, Error
**errp
)
1668 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_ALL
);
1671 int migrate_add_blocker_normal(Error
**reasonp
, Error
**errp
)
1673 return migrate_add_blocker_modes(reasonp
, errp
, MIG_MODE_NORMAL
, -1);
1676 int migrate_add_blocker_modes(Error
**reasonp
, Error
**errp
, MigMode mode
, ...)
1682 modes
= get_modes(mode
, ap
);
1685 if (is_only_migratable(reasonp
, errp
, modes
)) {
1687 } else if (is_busy(reasonp
, errp
)) {
1690 return add_blockers(reasonp
, errp
, modes
);
1693 int migrate_add_blocker_internal(Error
**reasonp
, Error
**errp
)
1695 int modes
= BIT(MIG_MODE__MAX
) - 1;
1697 if (is_busy(reasonp
, errp
)) {
1700 return add_blockers(reasonp
, errp
, modes
);
1703 void migrate_del_blocker(Error
**reasonp
)
1706 for (MigMode mode
= 0; mode
< MIG_MODE__MAX
; mode
++) {
1707 migration_blockers
[mode
] = g_slist_remove(migration_blockers
[mode
],
1710 error_free(*reasonp
);
1715 void qmp_migrate_incoming(const char *uri
, bool has_channels
,
1716 MigrationChannelList
*channels
, Error
**errp
)
1718 Error
*local_err
= NULL
;
1719 static bool once
= true;
1722 error_setg(errp
, "The incoming migration has already been started");
1725 if (!runstate_check(RUN_STATE_INMIGRATE
)) {
1726 error_setg(errp
, "'-incoming' was not specified on the command line");
1730 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1734 qemu_start_incoming_migration(uri
, has_channels
, channels
, &local_err
);
1737 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
1738 error_propagate(errp
, local_err
);
1745 void qmp_migrate_recover(const char *uri
, Error
**errp
)
1747 MigrationIncomingState
*mis
= migration_incoming_get_current();
1750 * Don't even bother to use ERRP_GUARD() as it _must_ always be set by
1751 * callers (no one should ignore a recover failure); if there is, it's a
1752 * programming error.
1756 if (mis
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1757 error_setg(errp
, "Migrate recover can only be run "
1758 "when postcopy is paused.");
1762 /* If there's an existing transport, release it */
1763 migration_incoming_transport_cleanup(mis
);
1766 * Note that this call will never start a real migration; it will
1767 * only re-setup the migration stream and poke existing migration
1768 * to continue using that newly established channel.
1770 qemu_start_incoming_migration(uri
, false, NULL
, errp
);
1773 void qmp_migrate_pause(Error
**errp
)
1775 MigrationState
*ms
= migrate_get_current();
1776 MigrationIncomingState
*mis
= migration_incoming_get_current();
1779 if (migration_postcopy_is_alive(ms
->state
)) {
1780 /* Source side, during postcopy */
1781 Error
*error
= NULL
;
1783 /* Tell the core migration that we're pausing */
1784 error_setg(&error
, "Postcopy migration is paused by the user");
1785 migrate_set_error(ms
, error
);
1788 qemu_mutex_lock(&ms
->qemu_file_lock
);
1789 if (ms
->to_dst_file
) {
1790 ret
= qemu_file_shutdown(ms
->to_dst_file
);
1792 qemu_mutex_unlock(&ms
->qemu_file_lock
);
1794 error_setg(errp
, "Failed to pause source migration");
1798 * Kick the migration thread out of any waiting windows (on behalf
1799 * of the rp thread).
1801 migration_rp_kick(ms
);
1806 if (migration_postcopy_is_alive(mis
->state
)) {
1807 ret
= qemu_file_shutdown(mis
->from_src_file
);
1809 error_setg(errp
, "Failed to pause destination migration");
1814 error_setg(errp
, "migrate-pause is currently only supported "
1815 "during postcopy-active or postcopy-recover state");
1818 bool migration_is_blocked(Error
**errp
)
1820 GSList
*blockers
= migration_blockers
[migrate_mode()];
1822 if (qemu_savevm_state_blocked(errp
)) {
1827 error_propagate(errp
, error_copy(blockers
->data
));
1834 /* Returns true if continue to migrate, or false if error detected */
1835 static bool migrate_prepare(MigrationState
*s
, bool blk
, bool blk_inc
,
1836 bool resume
, Error
**errp
)
1838 Error
*local_err
= NULL
;
1841 warn_report("parameter 'inc' is deprecated;"
1842 " use blockdev-mirror with NBD instead");
1846 warn_report("parameter 'blk' is deprecated;"
1847 " use blockdev-mirror with NBD instead");
1851 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_PAUSED
) {
1852 error_setg(errp
, "Cannot resume if there is no "
1853 "paused migration");
1858 * Postcopy recovery won't work well with release-ram
1859 * capability since release-ram will drop the page buffer as
1860 * long as the page is put into the send buffer. So if there
1861 * is a network failure happened, any page buffers that have
1862 * not yet reached the destination VM but have already been
1863 * sent from the source VM will be lost forever. Let's refuse
1864 * the client from resuming such a postcopy migration.
1865 * Luckily release-ram was designed to only be used when src
1866 * and destination VMs are on the same host, so it should be
1869 if (migrate_release_ram()) {
1870 error_setg(errp
, "Postcopy recovery cannot work "
1871 "when release-ram capability is set");
1875 /* This is a resume, skip init status */
1879 if (migration_is_running(s
->state
)) {
1880 error_setg(errp
, QERR_MIGRATION_ACTIVE
);
1884 if (runstate_check(RUN_STATE_INMIGRATE
)) {
1885 error_setg(errp
, "Guest is waiting for an incoming migration");
1889 if (runstate_check(RUN_STATE_POSTMIGRATE
)) {
1890 error_setg(errp
, "Can't migrate the vm that was paused due to "
1891 "previous migration");
1895 if (migration_is_blocked(errp
)) {
1899 if (blk
|| blk_inc
) {
1900 if (migrate_colo()) {
1901 error_setg(errp
, "No disk migration is required in COLO mode");
1904 if (migrate_block() || migrate_block_incremental()) {
1905 error_setg(errp
, "Command options are incompatible with "
1906 "current migration capabilities");
1909 if (!migrate_cap_set(MIGRATION_CAPABILITY_BLOCK
, true, &local_err
)) {
1910 error_propagate(errp
, local_err
);
1913 s
->must_remove_block_options
= true;
1917 migrate_set_block_incremental(true);
1920 if (migrate_init(s
, errp
)) {
1927 void qmp_migrate(const char *uri
, bool has_channels
,
1928 MigrationChannelList
*channels
, bool has_blk
, bool blk
,
1929 bool has_inc
, bool inc
, bool has_detach
, bool detach
,
1930 bool has_resume
, bool resume
, Error
**errp
)
1932 bool resume_requested
;
1933 Error
*local_err
= NULL
;
1934 MigrationState
*s
= migrate_get_current();
1935 MigrationChannel
*channel
= NULL
;
1936 MigrationAddress
*addr
= NULL
;
1939 * Having preliminary checks for uri and channel
1941 if (uri
&& has_channels
) {
1942 error_setg(errp
, "'uri' and 'channels' arguments are mutually "
1943 "exclusive; exactly one of the two should be present in "
1944 "'migrate' qmp command ");
1946 } else if (channels
) {
1947 /* To verify that Migrate channel list has only item */
1948 if (channels
->next
) {
1949 error_setg(errp
, "Channel list has more than one entries");
1952 channel
= channels
->value
;
1954 /* caller uses the old URI syntax */
1955 if (!migrate_uri_parse(uri
, &channel
, errp
)) {
1959 error_setg(errp
, "neither 'uri' or 'channels' argument are "
1960 "specified in 'migrate' qmp command ");
1963 addr
= channel
->addr
;
1965 /* transport mechanism not suitable for migration? */
1966 if (!migration_channels_and_transport_compatible(addr
, errp
)) {
1970 resume_requested
= has_resume
&& resume
;
1971 if (!migrate_prepare(s
, has_blk
&& blk
, has_inc
&& inc
,
1972 resume_requested
, errp
)) {
1973 /* Error detected, put into errp */
1977 if (!resume_requested
) {
1978 if (!yank_register_instance(MIGRATION_YANK_INSTANCE
, errp
)) {
1983 if (addr
->transport
== MIGRATION_ADDRESS_TYPE_SOCKET
) {
1984 SocketAddress
*saddr
= &addr
->u
.socket
;
1985 if (saddr
->type
== SOCKET_ADDRESS_TYPE_INET
||
1986 saddr
->type
== SOCKET_ADDRESS_TYPE_UNIX
||
1987 saddr
->type
== SOCKET_ADDRESS_TYPE_VSOCK
) {
1988 socket_start_outgoing_migration(s
, saddr
, &local_err
);
1989 } else if (saddr
->type
== SOCKET_ADDRESS_TYPE_FD
) {
1990 fd_start_outgoing_migration(s
, saddr
->u
.fd
.str
, &local_err
);
1993 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_RDMA
) {
1994 rdma_start_outgoing_migration(s
, &addr
->u
.rdma
, &local_err
);
1996 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_EXEC
) {
1997 exec_start_outgoing_migration(s
, addr
->u
.exec
.args
, &local_err
);
1998 } else if (addr
->transport
== MIGRATION_ADDRESS_TYPE_FILE
) {
1999 file_start_outgoing_migration(s
, &addr
->u
.file
, &local_err
);
2001 error_setg(&local_err
, QERR_INVALID_PARAMETER_VALUE
, "uri",
2002 "a valid migration protocol");
2003 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
2004 MIGRATION_STATUS_FAILED
);
2005 block_cleanup_parameters();
2009 if (!resume_requested
) {
2010 yank_unregister_instance(MIGRATION_YANK_INSTANCE
);
2012 migrate_fd_error(s
, local_err
);
2013 error_propagate(errp
, local_err
);
2018 void qmp_migrate_cancel(Error
**errp
)
2020 migration_cancel(NULL
);
2023 void qmp_migrate_continue(MigrationStatus state
, Error
**errp
)
2025 MigrationState
*s
= migrate_get_current();
2026 if (s
->state
!= state
) {
2027 error_setg(errp
, "Migration not in expected state: %s",
2028 MigrationStatus_str(s
->state
));
2031 qemu_sem_post(&s
->pause_sem
);
2034 int migration_rp_wait(MigrationState
*s
)
2036 /* If migration has failure already, ignore the wait */
2037 if (migrate_has_error(s
)) {
2041 qemu_sem_wait(&s
->rp_state
.rp_sem
);
2043 /* After wait, double check that there's no failure */
2044 if (migrate_has_error(s
)) {
2051 void migration_rp_kick(MigrationState
*s
)
2053 qemu_sem_post(&s
->rp_state
.rp_sem
);
2056 static struct rp_cmd_args
{
2057 ssize_t len
; /* -1 = variable */
2060 [MIG_RP_MSG_INVALID
] = { .len
= -1, .name
= "INVALID" },
2061 [MIG_RP_MSG_SHUT
] = { .len
= 4, .name
= "SHUT" },
2062 [MIG_RP_MSG_PONG
] = { .len
= 4, .name
= "PONG" },
2063 [MIG_RP_MSG_REQ_PAGES
] = { .len
= 12, .name
= "REQ_PAGES" },
2064 [MIG_RP_MSG_REQ_PAGES_ID
] = { .len
= -1, .name
= "REQ_PAGES_ID" },
2065 [MIG_RP_MSG_RECV_BITMAP
] = { .len
= -1, .name
= "RECV_BITMAP" },
2066 [MIG_RP_MSG_RESUME_ACK
] = { .len
= 4, .name
= "RESUME_ACK" },
2067 [MIG_RP_MSG_SWITCHOVER_ACK
] = { .len
= 0, .name
= "SWITCHOVER_ACK" },
2068 [MIG_RP_MSG_MAX
] = { .len
= -1, .name
= "MAX" },
2072 * Process a request for pages received on the return path,
2073 * We're allowed to send more than requested (e.g. to round to our page size)
2074 * and we don't need to send pages that have already been sent.
2077 migrate_handle_rp_req_pages(MigrationState
*ms
, const char* rbname
,
2078 ram_addr_t start
, size_t len
, Error
**errp
)
2080 long our_host_ps
= qemu_real_host_page_size();
2082 trace_migrate_handle_rp_req_pages(rbname
, start
, len
);
2085 * Since we currently insist on matching page sizes, just sanity check
2086 * we're being asked for whole host pages.
2088 if (!QEMU_IS_ALIGNED(start
, our_host_ps
) ||
2089 !QEMU_IS_ALIGNED(len
, our_host_ps
)) {
2090 error_setg(errp
, "MIG_RP_MSG_REQ_PAGES: Misaligned page request, start:"
2091 RAM_ADDR_FMT
" len: %zd", start
, len
);
2095 ram_save_queue_pages(rbname
, start
, len
, errp
);
2098 static bool migrate_handle_rp_recv_bitmap(MigrationState
*s
, char *block_name
,
2101 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
2104 error_setg(errp
, "MIG_RP_MSG_RECV_BITMAP has invalid block name '%s'",
2109 /* Fetch the received bitmap and refresh the dirty bitmap */
2110 return ram_dirty_bitmap_reload(s
, block
, errp
);
2113 static bool migrate_handle_rp_resume_ack(MigrationState
*s
,
2114 uint32_t value
, Error
**errp
)
2116 trace_source_return_path_thread_resume_ack(value
);
2118 if (value
!= MIGRATION_RESUME_ACK_VALUE
) {
2119 error_setg(errp
, "illegal resume_ack value %"PRIu32
, value
);
2123 /* Now both sides are active. */
2124 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_RECOVER
,
2125 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2127 /* Notify send thread that time to continue send pages */
2128 migration_rp_kick(s
);
2134 * Release ms->rp_state.from_dst_file (and postcopy_qemufile_src if
2135 * existed) in a safe way.
2137 static void migration_release_dst_files(MigrationState
*ms
)
2141 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2143 * Reset the from_dst_file pointer first before releasing it, as we
2144 * can't block within lock section
2146 file
= ms
->rp_state
.from_dst_file
;
2147 ms
->rp_state
.from_dst_file
= NULL
;
2151 * Do the same to postcopy fast path socket too if there is. No
2152 * locking needed because this qemufile should only be managed by
2153 * return path thread.
2155 if (ms
->postcopy_qemufile_src
) {
2156 migration_ioc_unregister_yank_from_file(ms
->postcopy_qemufile_src
);
2157 qemu_file_shutdown(ms
->postcopy_qemufile_src
);
2158 qemu_fclose(ms
->postcopy_qemufile_src
);
2159 ms
->postcopy_qemufile_src
= NULL
;
2166 * Handles messages sent on the return path towards the source VM
2169 static void *source_return_path_thread(void *opaque
)
2171 MigrationState
*ms
= opaque
;
2172 QEMUFile
*rp
= ms
->rp_state
.from_dst_file
;
2173 uint16_t header_len
, header_type
;
2175 uint32_t tmp32
, sibling_error
;
2176 ram_addr_t start
= 0; /* =0 to silence warning */
2177 size_t len
= 0, expected_len
;
2181 trace_source_return_path_thread_entry();
2182 rcu_register_thread();
2184 while (migration_is_setup_or_active(ms
->state
)) {
2185 trace_source_return_path_thread_loop_top();
2187 header_type
= qemu_get_be16(rp
);
2188 header_len
= qemu_get_be16(rp
);
2190 if (qemu_file_get_error(rp
)) {
2191 qemu_file_get_error_obj(rp
, &err
);
2195 if (header_type
>= MIG_RP_MSG_MAX
||
2196 header_type
== MIG_RP_MSG_INVALID
) {
2197 error_setg(&err
, "Received invalid message 0x%04x length 0x%04x",
2198 header_type
, header_len
);
2202 if ((rp_cmd_args
[header_type
].len
!= -1 &&
2203 header_len
!= rp_cmd_args
[header_type
].len
) ||
2204 header_len
> sizeof(buf
)) {
2205 error_setg(&err
, "Received '%s' message (0x%04x) with"
2206 "incorrect length %d expecting %zu",
2207 rp_cmd_args
[header_type
].name
, header_type
, header_len
,
2208 (size_t)rp_cmd_args
[header_type
].len
);
2212 /* We know we've got a valid header by this point */
2213 res
= qemu_get_buffer(rp
, buf
, header_len
);
2214 if (res
!= header_len
) {
2215 error_setg(&err
, "Failed reading data for message 0x%04x"
2216 " read %d expected %d",
2217 header_type
, res
, header_len
);
2221 /* OK, we have the message and the data */
2222 switch (header_type
) {
2223 case MIG_RP_MSG_SHUT
:
2224 sibling_error
= ldl_be_p(buf
);
2225 trace_source_return_path_thread_shut(sibling_error
);
2226 if (sibling_error
) {
2227 error_setg(&err
, "Sibling indicated error %d", sibling_error
);
2230 * We'll let the main thread deal with closing the RP
2231 * we could do a shutdown(2) on it, but we're the only user
2232 * anyway, so there's nothing gained.
2236 case MIG_RP_MSG_PONG
:
2237 tmp32
= ldl_be_p(buf
);
2238 trace_source_return_path_thread_pong(tmp32
);
2239 qemu_sem_post(&ms
->rp_state
.rp_pong_acks
);
2242 case MIG_RP_MSG_REQ_PAGES
:
2243 start
= ldq_be_p(buf
);
2244 len
= ldl_be_p(buf
+ 8);
2245 migrate_handle_rp_req_pages(ms
, NULL
, start
, len
, &err
);
2251 case MIG_RP_MSG_REQ_PAGES_ID
:
2252 expected_len
= 12 + 1; /* header + termination */
2254 if (header_len
>= expected_len
) {
2255 start
= ldq_be_p(buf
);
2256 len
= ldl_be_p(buf
+ 8);
2257 /* Now we expect an idstr */
2258 tmp32
= buf
[12]; /* Length of the following idstr */
2259 buf
[13 + tmp32
] = '\0';
2260 expected_len
+= tmp32
;
2262 if (header_len
!= expected_len
) {
2263 error_setg(&err
, "Req_Page_id with length %d expecting %zd",
2264 header_len
, expected_len
);
2267 migrate_handle_rp_req_pages(ms
, (char *)&buf
[13], start
, len
,
2274 case MIG_RP_MSG_RECV_BITMAP
:
2275 if (header_len
< 1) {
2276 error_setg(&err
, "MIG_RP_MSG_RECV_BITMAP missing block name");
2279 /* Format: len (1B) + idstr (<255B). This ends the idstr. */
2280 buf
[buf
[0] + 1] = '\0';
2281 if (!migrate_handle_rp_recv_bitmap(ms
, (char *)(buf
+ 1), &err
)) {
2286 case MIG_RP_MSG_RESUME_ACK
:
2287 tmp32
= ldl_be_p(buf
);
2288 if (!migrate_handle_rp_resume_ack(ms
, tmp32
, &err
)) {
2293 case MIG_RP_MSG_SWITCHOVER_ACK
:
2294 ms
->switchover_acked
= true;
2295 trace_source_return_path_thread_switchover_acked();
2305 migrate_set_error(ms
, err
);
2307 trace_source_return_path_thread_bad_end();
2310 if (ms
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2312 * this will be extremely unlikely: that we got yet another network
2313 * issue during recovering of the 1st network failure.. during this
2314 * period the main migration thread can be waiting on rp_sem for
2315 * this thread to sync with the other side.
2317 * When this happens, explicitly kick the migration thread out of
2318 * RECOVER stage and back to PAUSED, so the admin can try
2321 migration_rp_kick(ms
);
2324 trace_source_return_path_thread_end();
2325 rcu_unregister_thread();
2330 static int open_return_path_on_source(MigrationState
*ms
)
2332 ms
->rp_state
.from_dst_file
= qemu_file_get_return_path(ms
->to_dst_file
);
2333 if (!ms
->rp_state
.from_dst_file
) {
2337 trace_open_return_path_on_source();
2339 qemu_thread_create(&ms
->rp_state
.rp_thread
, "return path",
2340 source_return_path_thread
, ms
, QEMU_THREAD_JOINABLE
);
2341 ms
->rp_state
.rp_thread_created
= true;
2343 trace_open_return_path_on_source_continue();
2348 /* Return true if error detected, or false otherwise */
2349 static bool close_return_path_on_source(MigrationState
*ms
)
2351 if (!ms
->rp_state
.rp_thread_created
) {
2355 trace_migration_return_path_end_before();
2358 * If this is a normal exit then the destination will send a SHUT
2359 * and the rp_thread will exit, however if there's an error we
2360 * need to cause it to exit. shutdown(2), if we have it, will
2361 * cause it to unblock if it's stuck waiting for the destination.
2363 WITH_QEMU_LOCK_GUARD(&ms
->qemu_file_lock
) {
2364 if (ms
->to_dst_file
&& ms
->rp_state
.from_dst_file
&&
2365 qemu_file_get_error(ms
->to_dst_file
)) {
2366 qemu_file_shutdown(ms
->rp_state
.from_dst_file
);
2370 qemu_thread_join(&ms
->rp_state
.rp_thread
);
2371 ms
->rp_state
.rp_thread_created
= false;
2372 migration_release_dst_files(ms
);
2373 trace_migration_return_path_end_after();
2375 /* Return path will persist the error in MigrationState when quit */
2376 return migrate_has_error(ms
);
2380 migration_wait_main_channel(MigrationState
*ms
)
2382 /* Wait until one PONG message received */
2383 qemu_sem_wait(&ms
->rp_state
.rp_pong_acks
);
2387 * Switch from normal iteration to postcopy
2388 * Returns non-0 on error
2390 static int postcopy_start(MigrationState
*ms
, Error
**errp
)
2393 QIOChannelBuffer
*bioc
;
2395 uint64_t bandwidth
= migrate_max_postcopy_bandwidth();
2396 bool restart_block
= false;
2397 int cur_state
= MIGRATION_STATUS_ACTIVE
;
2399 if (migrate_postcopy_preempt()) {
2400 migration_wait_main_channel(ms
);
2401 if (postcopy_preempt_establish_channel(ms
)) {
2402 migrate_set_state(&ms
->state
, ms
->state
, MIGRATION_STATUS_FAILED
);
2407 if (!migrate_pause_before_switchover()) {
2408 migrate_set_state(&ms
->state
, MIGRATION_STATUS_ACTIVE
,
2409 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2412 trace_postcopy_start();
2413 qemu_mutex_lock_iothread();
2414 trace_postcopy_start_set_run();
2416 migration_downtime_start(ms
);
2418 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
, NULL
);
2419 global_state_store();
2420 ret
= migration_stop_vm(RUN_STATE_FINISH_MIGRATE
);
2425 ret
= migration_maybe_pause(ms
, &cur_state
,
2426 MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2431 ret
= bdrv_inactivate_all();
2435 restart_block
= true;
2438 * Cause any non-postcopiable, but iterative devices to
2439 * send out their final data.
2441 qemu_savevm_state_complete_precopy(ms
->to_dst_file
, true, false);
2444 * in Finish migrate and with the io-lock held everything should
2445 * be quiet, but we've potentially still got dirty pages and we
2446 * need to tell the destination to throw any pages it's already received
2449 if (migrate_postcopy_ram()) {
2450 ram_postcopy_send_discard_bitmap(ms
);
2454 * send rest of state - note things that are doing postcopy
2455 * will notice we're in POSTCOPY_ACTIVE and not actually
2456 * wrap their state up here
2458 migration_rate_set(bandwidth
);
2459 if (migrate_postcopy_ram()) {
2460 /* Ping just for debugging, helps line traces up */
2461 qemu_savevm_send_ping(ms
->to_dst_file
, 2);
2465 * While loading the device state we may trigger page transfer
2466 * requests and the fd must be free to process those, and thus
2467 * the destination must read the whole device state off the fd before
2468 * it starts processing it. Unfortunately the ad-hoc migration format
2469 * doesn't allow the destination to know the size to read without fully
2470 * parsing it through each devices load-state code (especially the open
2471 * coded devices that use get/put).
2472 * So we wrap the device state up in a package with a length at the start;
2473 * to do this we use a qemu_buf to hold the whole of the device state.
2475 bioc
= qio_channel_buffer_new(4096);
2476 qio_channel_set_name(QIO_CHANNEL(bioc
), "migration-postcopy-buffer");
2477 fb
= qemu_file_new_output(QIO_CHANNEL(bioc
));
2478 object_unref(OBJECT(bioc
));
2481 * Make sure the receiver can get incoming pages before we send the rest
2484 qemu_savevm_send_postcopy_listen(fb
);
2486 qemu_savevm_state_complete_precopy(fb
, false, false);
2487 if (migrate_postcopy_ram()) {
2488 qemu_savevm_send_ping(fb
, 3);
2491 qemu_savevm_send_postcopy_run(fb
);
2493 /* <><> end of stuff going into the package */
2495 /* Last point of recovery; as soon as we send the package the destination
2496 * can open devices and potentially start running.
2497 * Lets just check again we've not got any errors.
2499 ret
= qemu_file_get_error(ms
->to_dst_file
);
2501 error_setg(errp
, "postcopy_start: Migration stream errored (pre package)");
2505 restart_block
= false;
2507 /* Now send that blob */
2508 if (qemu_savevm_send_packaged(ms
->to_dst_file
, bioc
->data
, bioc
->usage
)) {
2513 /* Send a notify to give a chance for anything that needs to happen
2514 * at the transition to postcopy and after the device state; in particular
2515 * spice needs to trigger a transition now
2517 ms
->postcopy_after_devices
= true;
2518 migration_call_notifiers(ms
);
2520 migration_downtime_end(ms
);
2522 qemu_mutex_unlock_iothread();
2524 if (migrate_postcopy_ram()) {
2526 * Although this ping is just for debug, it could potentially be
2527 * used for getting a better measurement of downtime at the source.
2529 qemu_savevm_send_ping(ms
->to_dst_file
, 4);
2532 if (migrate_release_ram()) {
2533 ram_postcopy_migrated_memory_release(ms
);
2536 ret
= qemu_file_get_error(ms
->to_dst_file
);
2538 error_setg(errp
, "postcopy_start: Migration stream errored");
2539 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2540 MIGRATION_STATUS_FAILED
);
2543 trace_postcopy_preempt_enabled(migrate_postcopy_preempt());
2550 migrate_set_state(&ms
->state
, MIGRATION_STATUS_POSTCOPY_ACTIVE
,
2551 MIGRATION_STATUS_FAILED
);
2552 if (restart_block
) {
2553 /* A failure happened early enough that we know the destination hasn't
2554 * accessed block devices, so we're safe to recover.
2556 Error
*local_err
= NULL
;
2558 bdrv_activate_all(&local_err
);
2560 error_report_err(local_err
);
2563 qemu_mutex_unlock_iothread();
2568 * migration_maybe_pause: Pause if required to by
2569 * migrate_pause_before_switchover called with the iothread locked
2570 * Returns: 0 on success
2572 static int migration_maybe_pause(MigrationState
*s
,
2573 int *current_active_state
,
2576 if (!migrate_pause_before_switchover()) {
2580 /* Since leaving this state is not atomic with posting the semaphore
2581 * it's possible that someone could have issued multiple migrate_continue
2582 * and the semaphore is incorrectly positive at this point;
2583 * the docs say it's undefined to reinit a semaphore that's already
2584 * init'd, so use timedwait to eat up any existing posts.
2586 while (qemu_sem_timedwait(&s
->pause_sem
, 1) == 0) {
2587 /* This block intentionally left blank */
2591 * If the migration is cancelled when it is in the completion phase,
2592 * the migration state is set to MIGRATION_STATUS_CANCELLING.
2593 * So we don't need to wait a semaphore, otherwise we would always
2594 * wait for the 'pause_sem' semaphore.
2596 if (s
->state
!= MIGRATION_STATUS_CANCELLING
) {
2597 qemu_mutex_unlock_iothread();
2598 migrate_set_state(&s
->state
, *current_active_state
,
2599 MIGRATION_STATUS_PRE_SWITCHOVER
);
2600 qemu_sem_wait(&s
->pause_sem
);
2601 migrate_set_state(&s
->state
, MIGRATION_STATUS_PRE_SWITCHOVER
,
2603 *current_active_state
= new_state
;
2604 qemu_mutex_lock_iothread();
2607 return s
->state
== new_state
? 0 : -EINVAL
;
2610 static int migration_completion_precopy(MigrationState
*s
,
2611 int *current_active_state
)
2615 qemu_mutex_lock_iothread();
2616 migration_downtime_start(s
);
2617 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
, NULL
);
2619 s
->vm_old_state
= runstate_get();
2620 global_state_store();
2622 ret
= migration_stop_vm(RUN_STATE_FINISH_MIGRATE
);
2623 trace_migration_completion_vm_stop(ret
);
2628 ret
= migration_maybe_pause(s
, current_active_state
,
2629 MIGRATION_STATUS_DEVICE
);
2635 * Inactivate disks except in COLO, and track that we have done so in order
2636 * to remember to reactivate them if migration fails or is cancelled.
2638 s
->block_inactive
= !migrate_colo();
2639 migration_rate_set(RATE_LIMIT_DISABLED
);
2640 ret
= qemu_savevm_state_complete_precopy(s
->to_dst_file
, false,
2643 qemu_mutex_unlock_iothread();
2647 static void migration_completion_postcopy(MigrationState
*s
)
2649 trace_migration_completion_postcopy_end();
2651 qemu_mutex_lock_iothread();
2652 qemu_savevm_state_complete_postcopy(s
->to_dst_file
);
2653 qemu_mutex_unlock_iothread();
2656 * Shutdown the postcopy fast path thread. This is only needed when dest
2657 * QEMU binary is old (7.1/7.2). QEMU 8.0+ doesn't need this.
2659 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
2660 postcopy_preempt_shutdown_file(s
);
2663 trace_migration_completion_postcopy_end_after_complete();
2666 static void migration_completion_failed(MigrationState
*s
,
2667 int current_active_state
)
2669 if (s
->block_inactive
&& (s
->state
== MIGRATION_STATUS_ACTIVE
||
2670 s
->state
== MIGRATION_STATUS_DEVICE
)) {
2672 * If not doing postcopy, vm_start() will be called: let's
2673 * regain control on images.
2675 Error
*local_err
= NULL
;
2677 qemu_mutex_lock_iothread();
2678 bdrv_activate_all(&local_err
);
2680 error_report_err(local_err
);
2682 s
->block_inactive
= false;
2684 qemu_mutex_unlock_iothread();
2687 migrate_set_state(&s
->state
, current_active_state
,
2688 MIGRATION_STATUS_FAILED
);
2692 * migration_completion: Used by migration_thread when there's not much left.
2693 * The caller 'breaks' the loop when this returns.
2695 * @s: Current migration state
2697 static void migration_completion(MigrationState
*s
)
2700 int current_active_state
= s
->state
;
2702 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2703 ret
= migration_completion_precopy(s
, ¤t_active_state
);
2704 } else if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2705 migration_completion_postcopy(s
);
2714 if (close_return_path_on_source(s
)) {
2718 if (qemu_file_get_error(s
->to_dst_file
)) {
2719 trace_migration_completion_file_err();
2723 if (migrate_colo() && s
->state
== MIGRATION_STATUS_ACTIVE
) {
2724 /* COLO does not support postcopy */
2725 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
2726 MIGRATION_STATUS_COLO
);
2728 migrate_set_state(&s
->state
, current_active_state
,
2729 MIGRATION_STATUS_COMPLETED
);
2735 migration_completion_failed(s
, current_active_state
);
2739 * bg_migration_completion: Used by bg_migration_thread when after all the
2740 * RAM has been saved. The caller 'breaks' the loop when this returns.
2742 * @s: Current migration state
2744 static void bg_migration_completion(MigrationState
*s
)
2746 int current_active_state
= s
->state
;
2748 if (s
->state
== MIGRATION_STATUS_ACTIVE
) {
2750 * By this moment we have RAM content saved into the migration stream.
2751 * The next step is to flush the non-RAM content (device state)
2752 * right after the ram content. The device state has been stored into
2753 * the temporary buffer before RAM saving started.
2755 qemu_put_buffer(s
->to_dst_file
, s
->bioc
->data
, s
->bioc
->usage
);
2756 qemu_fflush(s
->to_dst_file
);
2757 } else if (s
->state
== MIGRATION_STATUS_CANCELLING
) {
2761 if (qemu_file_get_error(s
->to_dst_file
)) {
2762 trace_migration_completion_file_err();
2766 migrate_set_state(&s
->state
, current_active_state
,
2767 MIGRATION_STATUS_COMPLETED
);
2771 migrate_set_state(&s
->state
, current_active_state
,
2772 MIGRATION_STATUS_FAILED
);
2775 typedef enum MigThrError
{
2776 /* No error detected */
2777 MIG_THR_ERR_NONE
= 0,
2778 /* Detected error, but resumed successfully */
2779 MIG_THR_ERR_RECOVERED
= 1,
2780 /* Detected fatal error, need to exit */
2781 MIG_THR_ERR_FATAL
= 2,
2784 static int postcopy_resume_handshake(MigrationState
*s
)
2786 qemu_savevm_send_postcopy_resume(s
->to_dst_file
);
2788 while (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2789 if (migration_rp_wait(s
)) {
2794 if (s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
) {
2801 /* Return zero if success, or <0 for error */
2802 static int postcopy_do_resume(MigrationState
*s
)
2807 * Call all the resume_prepare() hooks, so that modules can be
2808 * ready for the migration resume.
2810 ret
= qemu_savevm_state_resume_prepare(s
);
2812 error_report("%s: resume_prepare() failure detected: %d",
2818 * If preempt is enabled, re-establish the preempt channel. Note that
2819 * we do it after resume prepare to make sure the main channel will be
2820 * created before the preempt channel. E.g. with weak network, the
2821 * dest QEMU may get messed up with the preempt and main channels on
2822 * the order of connection setup. This guarantees the correct order.
2824 ret
= postcopy_preempt_establish_channel(s
);
2826 error_report("%s: postcopy_preempt_establish_channel(): %d",
2832 * Last handshake with destination on the resume (destination will
2833 * switch to postcopy-active afterwards)
2835 ret
= postcopy_resume_handshake(s
);
2837 error_report("%s: handshake failed: %d", __func__
, ret
);
2845 * We don't return until we are in a safe state to continue current
2846 * postcopy migration. Returns MIG_THR_ERR_RECOVERED if recovered, or
2847 * MIG_THR_ERR_FATAL if unrecovery failure happened.
2849 static MigThrError
postcopy_pause(MigrationState
*s
)
2851 assert(s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
);
2857 * Current channel is possibly broken. Release it. Note that this is
2858 * guaranteed even without lock because to_dst_file should only be
2859 * modified by the migration thread. That also guarantees that the
2860 * unregister of yank is safe too without the lock. It should be safe
2861 * even to be within the qemu_file_lock, but we didn't do that to avoid
2862 * taking more mutex (yank_lock) within qemu_file_lock. TL;DR: we make
2863 * the qemu_file_lock critical section as small as possible.
2865 assert(s
->to_dst_file
);
2866 migration_ioc_unregister_yank_from_file(s
->to_dst_file
);
2867 qemu_mutex_lock(&s
->qemu_file_lock
);
2868 file
= s
->to_dst_file
;
2869 s
->to_dst_file
= NULL
;
2870 qemu_mutex_unlock(&s
->qemu_file_lock
);
2872 qemu_file_shutdown(file
);
2876 * We're already pausing, so ignore any errors on the return
2877 * path and just wait for the thread to finish. It will be
2878 * re-created when we resume.
2880 close_return_path_on_source(s
);
2882 migrate_set_state(&s
->state
, s
->state
,
2883 MIGRATION_STATUS_POSTCOPY_PAUSED
);
2885 error_report("Detected IO failure for postcopy. "
2886 "Migration paused.");
2889 * We wait until things fixed up. Then someone will setup the
2890 * status back for us.
2892 while (s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
) {
2893 qemu_sem_wait(&s
->postcopy_pause_sem
);
2896 if (s
->state
== MIGRATION_STATUS_POSTCOPY_RECOVER
) {
2897 /* Woken up by a recover procedure. Give it a shot */
2899 /* Do the resume logic */
2900 if (postcopy_do_resume(s
) == 0) {
2901 /* Let's continue! */
2902 trace_postcopy_pause_continued();
2903 return MIG_THR_ERR_RECOVERED
;
2906 * Something wrong happened during the recovery, let's
2907 * pause again. Pause is always better than throwing
2913 /* This is not right... Time to quit. */
2914 return MIG_THR_ERR_FATAL
;
2919 static MigThrError
migration_detect_error(MigrationState
*s
)
2922 int state
= s
->state
;
2923 Error
*local_error
= NULL
;
2925 if (state
== MIGRATION_STATUS_CANCELLING
||
2926 state
== MIGRATION_STATUS_CANCELLED
) {
2927 /* End the migration, but don't set the state to failed */
2928 return MIG_THR_ERR_FATAL
;
2932 * Try to detect any file errors. Note that postcopy_qemufile_src will
2933 * be NULL when postcopy preempt is not enabled.
2935 ret
= qemu_file_get_error_obj_any(s
->to_dst_file
,
2936 s
->postcopy_qemufile_src
,
2939 /* Everything is fine */
2940 assert(!local_error
);
2941 return MIG_THR_ERR_NONE
;
2945 migrate_set_error(s
, local_error
);
2946 error_free(local_error
);
2949 if (state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
&& ret
) {
2951 * For postcopy, we allow the network to be down for a
2952 * while. After that, it can be continued by a
2955 return postcopy_pause(s
);
2958 * For precopy (or postcopy with error outside IO), we fail
2961 migrate_set_state(&s
->state
, state
, MIGRATION_STATUS_FAILED
);
2962 trace_migration_thread_file_err();
2964 /* Time to stop the migration, now. */
2965 return MIG_THR_ERR_FATAL
;
2969 static void migration_calculate_complete(MigrationState
*s
)
2971 uint64_t bytes
= migration_transferred_bytes();
2972 int64_t end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2973 int64_t transfer_time
;
2975 migration_downtime_end(s
);
2976 s
->total_time
= end_time
- s
->start_time
;
2977 transfer_time
= s
->total_time
- s
->setup_time
;
2978 if (transfer_time
) {
2979 s
->mbps
= ((double) bytes
* 8.0) / transfer_time
/ 1000;
2983 static void update_iteration_initial_status(MigrationState
*s
)
2986 * Update these three fields at the same time to avoid mismatch info lead
2987 * wrong speed calculation.
2989 s
->iteration_start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
2990 s
->iteration_initial_bytes
= migration_transferred_bytes();
2991 s
->iteration_initial_pages
= ram_get_total_transferred_pages();
2994 static void migration_update_counters(MigrationState
*s
,
2995 int64_t current_time
)
2997 uint64_t transferred
, transferred_pages
, time_spent
;
2998 uint64_t current_bytes
; /* bytes transferred since the beginning */
2999 uint64_t switchover_bw
;
3000 /* Expected bandwidth when switching over to destination QEMU */
3001 double expected_bw_per_ms
;
3004 if (current_time
< s
->iteration_start_time
+ BUFFER_DELAY
) {
3008 switchover_bw
= migrate_avail_switchover_bandwidth();
3009 current_bytes
= migration_transferred_bytes();
3010 transferred
= current_bytes
- s
->iteration_initial_bytes
;
3011 time_spent
= current_time
- s
->iteration_start_time
;
3012 bandwidth
= (double)transferred
/ time_spent
;
3014 if (switchover_bw
) {
3016 * If the user specified a switchover bandwidth, let's trust the
3017 * user so that can be more accurate than what we estimated.
3019 expected_bw_per_ms
= switchover_bw
/ 1000;
3021 /* If the user doesn't specify bandwidth, we use the estimated */
3022 expected_bw_per_ms
= bandwidth
;
3025 s
->threshold_size
= expected_bw_per_ms
* migrate_downtime_limit();
3027 s
->mbps
= (((double) transferred
* 8.0) /
3028 ((double) time_spent
/ 1000.0)) / 1000.0 / 1000.0;
3030 transferred_pages
= ram_get_total_transferred_pages() -
3031 s
->iteration_initial_pages
;
3032 s
->pages_per_second
= (double) transferred_pages
/
3033 (((double) time_spent
/ 1000.0));
3036 * if we haven't sent anything, we don't want to
3037 * recalculate. 10000 is a small enough number for our purposes
3039 if (stat64_get(&mig_stats
.dirty_pages_rate
) &&
3040 transferred
> 10000) {
3041 s
->expected_downtime
=
3042 stat64_get(&mig_stats
.dirty_bytes_last_sync
) / expected_bw_per_ms
;
3045 migration_rate_reset();
3047 update_iteration_initial_status(s
);
3049 trace_migrate_transferred(transferred
, time_spent
,
3050 /* Both in unit bytes/ms */
3051 bandwidth
, switchover_bw
/ 1000,
3055 static bool migration_can_switchover(MigrationState
*s
)
3057 if (!migrate_switchover_ack()) {
3061 /* No reason to wait for switchover ACK if VM is stopped */
3062 if (!runstate_is_running()) {
3066 return s
->switchover_acked
;
3069 /* Migration thread iteration status */
3071 MIG_ITERATE_RESUME
, /* Resume current iteration */
3072 MIG_ITERATE_SKIP
, /* Skip current iteration */
3073 MIG_ITERATE_BREAK
, /* Break the loop */
3077 * Return true if continue to the next iteration directly, false
3080 static MigIterateState
migration_iteration_run(MigrationState
*s
)
3082 uint64_t must_precopy
, can_postcopy
;
3083 Error
*local_err
= NULL
;
3084 bool in_postcopy
= s
->state
== MIGRATION_STATUS_POSTCOPY_ACTIVE
;
3085 bool can_switchover
= migration_can_switchover(s
);
3087 qemu_savevm_state_pending_estimate(&must_precopy
, &can_postcopy
);
3088 uint64_t pending_size
= must_precopy
+ can_postcopy
;
3090 trace_migrate_pending_estimate(pending_size
, must_precopy
, can_postcopy
);
3092 if (must_precopy
<= s
->threshold_size
) {
3093 qemu_savevm_state_pending_exact(&must_precopy
, &can_postcopy
);
3094 pending_size
= must_precopy
+ can_postcopy
;
3095 trace_migrate_pending_exact(pending_size
, must_precopy
, can_postcopy
);
3098 if ((!pending_size
|| pending_size
< s
->threshold_size
) && can_switchover
) {
3099 trace_migration_thread_low_pending(pending_size
);
3100 migration_completion(s
);
3101 return MIG_ITERATE_BREAK
;
3104 /* Still a significant amount to transfer */
3105 if (!in_postcopy
&& must_precopy
<= s
->threshold_size
&& can_switchover
&&
3106 qatomic_read(&s
->start_postcopy
)) {
3107 if (postcopy_start(s
, &local_err
)) {
3108 migrate_set_error(s
, local_err
);
3109 error_report_err(local_err
);
3111 return MIG_ITERATE_SKIP
;
3114 /* Just another iteration step */
3115 qemu_savevm_state_iterate(s
->to_dst_file
, in_postcopy
);
3116 return MIG_ITERATE_RESUME
;
3119 static void migration_iteration_finish(MigrationState
*s
)
3121 /* If we enabled cpu throttling for auto-converge, turn it off. */
3122 cpu_throttle_stop();
3124 qemu_mutex_lock_iothread();
3126 case MIGRATION_STATUS_COMPLETED
:
3127 migration_calculate_complete(s
);
3128 runstate_set(RUN_STATE_POSTMIGRATE
);
3130 case MIGRATION_STATUS_COLO
:
3131 assert(migrate_colo());
3132 migrate_start_colo_process(s
);
3133 s
->vm_old_state
= RUN_STATE_RUNNING
;
3135 case MIGRATION_STATUS_FAILED
:
3136 case MIGRATION_STATUS_CANCELLED
:
3137 case MIGRATION_STATUS_CANCELLING
:
3138 if (s
->vm_old_state
== RUN_STATE_RUNNING
) {
3139 if (!runstate_check(RUN_STATE_SHUTDOWN
)) {
3143 if (runstate_check(RUN_STATE_FINISH_MIGRATE
)) {
3144 runstate_set(s
->vm_old_state
);
3150 /* Should not reach here, but if so, forgive the VM. */
3151 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3154 migrate_fd_cleanup_schedule(s
);
3155 qemu_mutex_unlock_iothread();
3158 static void bg_migration_iteration_finish(MigrationState
*s
)
3161 * Stop tracking RAM writes - un-protect memory, un-register UFFD
3162 * memory ranges, flush kernel wait queues and wake up threads
3163 * waiting for write fault to be resolved.
3165 ram_write_tracking_stop();
3167 qemu_mutex_lock_iothread();
3169 case MIGRATION_STATUS_COMPLETED
:
3170 migration_calculate_complete(s
);
3173 case MIGRATION_STATUS_ACTIVE
:
3174 case MIGRATION_STATUS_FAILED
:
3175 case MIGRATION_STATUS_CANCELLED
:
3176 case MIGRATION_STATUS_CANCELLING
:
3180 /* Should not reach here, but if so, forgive the VM. */
3181 error_report("%s: Unknown ending state %d", __func__
, s
->state
);
3185 migrate_fd_cleanup_schedule(s
);
3186 qemu_mutex_unlock_iothread();
3190 * Return true if continue to the next iteration directly, false
3193 static MigIterateState
bg_migration_iteration_run(MigrationState
*s
)
3197 res
= qemu_savevm_state_iterate(s
->to_dst_file
, false);
3199 bg_migration_completion(s
);
3200 return MIG_ITERATE_BREAK
;
3203 return MIG_ITERATE_RESUME
;
3206 void migration_make_urgent_request(void)
3208 qemu_sem_post(&migrate_get_current()->rate_limit_sem
);
3211 void migration_consume_urgent_request(void)
3213 qemu_sem_wait(&migrate_get_current()->rate_limit_sem
);
3216 /* Returns true if the rate limiting was broken by an urgent request */
3217 bool migration_rate_limit(void)
3219 int64_t now
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
3220 MigrationState
*s
= migrate_get_current();
3222 bool urgent
= false;
3223 migration_update_counters(s
, now
);
3224 if (migration_rate_exceeded(s
->to_dst_file
)) {
3226 if (qemu_file_get_error(s
->to_dst_file
)) {
3230 * Wait for a delay to do rate limiting OR
3231 * something urgent to post the semaphore.
3233 int ms
= s
->iteration_start_time
+ BUFFER_DELAY
- now
;
3234 trace_migration_rate_limit_pre(ms
);
3235 if (qemu_sem_timedwait(&s
->rate_limit_sem
, ms
) == 0) {
3237 * We were woken by one or more urgent things but
3238 * the timedwait will have consumed one of them.
3239 * The service routine for the urgent wake will dec
3240 * the semaphore itself for each item it consumes,
3241 * so add this one we just eat back.
3243 qemu_sem_post(&s
->rate_limit_sem
);
3246 trace_migration_rate_limit_post(urgent
);
3252 * if failover devices are present, wait they are completely
3256 static void qemu_savevm_wait_unplug(MigrationState
*s
, int old_state
,
3259 if (qemu_savevm_state_guest_unplug_pending()) {
3260 migrate_set_state(&s
->state
, old_state
, MIGRATION_STATUS_WAIT_UNPLUG
);
3262 while (s
->state
== MIGRATION_STATUS_WAIT_UNPLUG
&&
3263 qemu_savevm_state_guest_unplug_pending()) {
3264 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3266 if (s
->state
!= MIGRATION_STATUS_WAIT_UNPLUG
) {
3267 int timeout
= 120; /* 30 seconds */
3269 * migration has been canceled
3270 * but as we have started an unplug we must wait the end
3271 * to be able to plug back the card
3273 while (timeout
-- && qemu_savevm_state_guest_unplug_pending()) {
3274 qemu_sem_timedwait(&s
->wait_unplug_sem
, 250);
3276 if (qemu_savevm_state_guest_unplug_pending() &&
3278 warn_report("migration: partially unplugged device on "
3283 migrate_set_state(&s
->state
, MIGRATION_STATUS_WAIT_UNPLUG
, new_state
);
3285 migrate_set_state(&s
->state
, old_state
, new_state
);
3290 * Master migration thread on the source VM.
3291 * It drives the migration and pumps the data down the outgoing channel.
3293 static void *migration_thread(void *opaque
)
3295 MigrationState
*s
= opaque
;
3296 MigrationThread
*thread
= NULL
;
3297 int64_t setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3298 MigThrError thr_error
;
3299 bool urgent
= false;
3301 thread
= migration_threads_add("live_migration", qemu_get_thread_id());
3303 rcu_register_thread();
3305 object_ref(OBJECT(s
));
3306 update_iteration_initial_status(s
);
3308 qemu_mutex_lock_iothread();
3309 qemu_savevm_state_header(s
->to_dst_file
);
3310 qemu_mutex_unlock_iothread();
3313 * If we opened the return path, we need to make sure dst has it
3316 if (s
->rp_state
.rp_thread_created
) {
3317 /* Now tell the dest that it should open its end so it can reply */
3318 qemu_savevm_send_open_return_path(s
->to_dst_file
);
3320 /* And do a ping that will make stuff easier to debug */
3321 qemu_savevm_send_ping(s
->to_dst_file
, 1);
3324 if (migrate_postcopy()) {
3326 * Tell the destination that we *might* want to do postcopy later;
3327 * if the other end can't do postcopy it should fail now, nice and
3330 qemu_savevm_send_postcopy_advise(s
->to_dst_file
);
3333 if (migrate_colo()) {
3334 /* Notify migration destination that we enable COLO */
3335 qemu_savevm_send_colo_enable(s
->to_dst_file
);
3338 qemu_mutex_lock_iothread();
3339 qemu_savevm_state_setup(s
->to_dst_file
);
3340 qemu_mutex_unlock_iothread();
3342 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3343 MIGRATION_STATUS_ACTIVE
);
3345 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3347 trace_migration_thread_setup_complete();
3349 while (migration_is_active(s
)) {
3350 if (urgent
|| !migration_rate_exceeded(s
->to_dst_file
)) {
3351 MigIterateState iter_state
= migration_iteration_run(s
);
3352 if (iter_state
== MIG_ITERATE_SKIP
) {
3354 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3360 * Try to detect any kind of failures, and see whether we
3361 * should stop the migration now.
3363 thr_error
= migration_detect_error(s
);
3364 if (thr_error
== MIG_THR_ERR_FATAL
) {
3365 /* Stop migration */
3367 } else if (thr_error
== MIG_THR_ERR_RECOVERED
) {
3369 * Just recovered from a e.g. network failure, reset all
3370 * the local variables. This is important to avoid
3371 * breaking transferred_bytes and bandwidth calculation
3373 update_iteration_initial_status(s
);
3376 urgent
= migration_rate_limit();
3379 trace_migration_thread_after_loop();
3380 migration_iteration_finish(s
);
3381 object_unref(OBJECT(s
));
3382 rcu_unregister_thread();
3383 migration_threads_remove(thread
);
3387 static void bg_migration_vm_start_bh(void *opaque
)
3389 MigrationState
*s
= opaque
;
3391 qemu_bh_delete(s
->vm_start_bh
);
3392 s
->vm_start_bh
= NULL
;
3395 migration_downtime_end(s
);
3399 * Background snapshot thread, based on live migration code.
3400 * This is an alternative implementation of live migration mechanism
3401 * introduced specifically to support background snapshots.
3403 * It takes advantage of userfault_fd write protection mechanism introduced
3404 * in v5.7 kernel. Compared to existing dirty page logging migration much
3405 * lesser stream traffic is produced resulting in smaller snapshot images,
3406 * simply cause of no page duplicates can get into the stream.
3408 * Another key point is that generated vmstate stream reflects machine state
3409 * 'frozen' at the beginning of snapshot creation compared to dirty page logging
3410 * mechanism, which effectively results in that saved snapshot is the state of VM
3411 * at the end of the process.
3413 static void *bg_migration_thread(void *opaque
)
3415 MigrationState
*s
= opaque
;
3416 int64_t setup_start
;
3417 MigThrError thr_error
;
3419 bool early_fail
= true;
3421 rcu_register_thread();
3422 object_ref(OBJECT(s
));
3424 migration_rate_set(RATE_LIMIT_DISABLED
);
3426 setup_start
= qemu_clock_get_ms(QEMU_CLOCK_HOST
);
3428 * We want to save vmstate for the moment when migration has been
3429 * initiated but also we want to save RAM content while VM is running.
3430 * The RAM content should appear first in the vmstate. So, we first
3431 * stash the non-RAM part of the vmstate to the temporary buffer,
3432 * then write RAM part of the vmstate to the migration stream
3433 * with vCPUs running and, finally, write stashed non-RAM part of
3434 * the vmstate from the buffer to the migration stream.
3436 s
->bioc
= qio_channel_buffer_new(512 * 1024);
3437 qio_channel_set_name(QIO_CHANNEL(s
->bioc
), "vmstate-buffer");
3438 fb
= qemu_file_new_output(QIO_CHANNEL(s
->bioc
));
3439 object_unref(OBJECT(s
->bioc
));
3441 update_iteration_initial_status(s
);
3444 * Prepare for tracking memory writes with UFFD-WP - populate
3445 * RAM pages before protecting.
3448 ram_write_tracking_prepare();
3451 qemu_mutex_lock_iothread();
3452 qemu_savevm_state_header(s
->to_dst_file
);
3453 qemu_savevm_state_setup(s
->to_dst_file
);
3454 qemu_mutex_unlock_iothread();
3456 qemu_savevm_wait_unplug(s
, MIGRATION_STATUS_SETUP
,
3457 MIGRATION_STATUS_ACTIVE
);
3459 s
->setup_time
= qemu_clock_get_ms(QEMU_CLOCK_HOST
) - setup_start
;
3461 trace_migration_thread_setup_complete();
3462 migration_downtime_start(s
);
3464 qemu_mutex_lock_iothread();
3467 * If VM is currently in suspended state, then, to make a valid runstate
3468 * transition in vm_stop_force_state() we need to wakeup it up.
3470 qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER
, NULL
);
3471 s
->vm_old_state
= runstate_get();
3473 global_state_store();
3474 /* Forcibly stop VM before saving state of vCPUs and devices */
3475 if (migration_stop_vm(RUN_STATE_PAUSED
)) {
3479 * Put vCPUs in sync with shadow context structures, then
3480 * save their state to channel-buffer along with devices.
3482 cpu_synchronize_all_states();
3483 if (qemu_savevm_state_complete_precopy_non_iterable(fb
, false, false)) {
3487 * Since we are going to get non-iterable state data directly
3488 * from s->bioc->data, explicit flush is needed here.
3492 /* Now initialize UFFD context and start tracking RAM writes */
3493 if (ram_write_tracking_start()) {
3499 * Start VM from BH handler to avoid write-fault lock here.
3500 * UFFD-WP protection for the whole RAM is already enabled so
3501 * calling VM state change notifiers from vm_start() would initiate
3502 * writes to virtio VQs memory which is in write-protected region.
3504 s
->vm_start_bh
= qemu_bh_new(bg_migration_vm_start_bh
, s
);
3505 qemu_bh_schedule(s
->vm_start_bh
);
3507 qemu_mutex_unlock_iothread();
3509 while (migration_is_active(s
)) {
3510 MigIterateState iter_state
= bg_migration_iteration_run(s
);
3511 if (iter_state
== MIG_ITERATE_SKIP
) {
3513 } else if (iter_state
== MIG_ITERATE_BREAK
) {
3518 * Try to detect any kind of failures, and see whether we
3519 * should stop the migration now.
3521 thr_error
= migration_detect_error(s
);
3522 if (thr_error
== MIG_THR_ERR_FATAL
) {
3523 /* Stop migration */
3527 migration_update_counters(s
, qemu_clock_get_ms(QEMU_CLOCK_REALTIME
));
3530 trace_migration_thread_after_loop();
3534 migrate_set_state(&s
->state
, MIGRATION_STATUS_ACTIVE
,
3535 MIGRATION_STATUS_FAILED
);
3536 qemu_mutex_unlock_iothread();
3539 bg_migration_iteration_finish(s
);
3542 object_unref(OBJECT(s
));
3543 rcu_unregister_thread();
3548 void migrate_fd_connect(MigrationState
*s
, Error
*error_in
)
3550 Error
*local_err
= NULL
;
3551 uint64_t rate_limit
;
3552 bool resume
= s
->state
== MIGRATION_STATUS_POSTCOPY_PAUSED
;
3555 * If there's a previous error, free it and prepare for another one.
3556 * Meanwhile if migration completes successfully, there won't have an error
3557 * dumped when calling migrate_fd_cleanup().
3559 migrate_error_free(s
);
3561 s
->expected_downtime
= migrate_downtime_limit();
3563 assert(s
->cleanup_bh
);
3565 assert(!s
->cleanup_bh
);
3566 s
->cleanup_bh
= qemu_bh_new(migrate_fd_cleanup_bh
, s
);
3569 migrate_fd_error(s
, error_in
);
3572 * Don't do cleanup for resume if channel is invalid, but only dump
3573 * the error. We wait for another channel connect from the user.
3574 * The error_report still gives HMP user a hint on what failed.
3575 * It's normally done in migrate_fd_cleanup(), but call it here
3578 error_report_err(error_copy(s
->error
));
3580 migrate_fd_cleanup(s
);
3586 /* This is a resumed migration */
3587 rate_limit
= migrate_max_postcopy_bandwidth();
3589 /* This is a fresh new migration */
3590 rate_limit
= migrate_max_bandwidth();
3592 /* Notify before starting migration thread */
3593 migration_call_notifiers(s
);
3596 migration_rate_set(rate_limit
);
3597 qemu_file_set_blocking(s
->to_dst_file
, true);
3600 * Open the return path. For postcopy, it is used exclusively. For
3601 * precopy, only if user specified "return-path" capability would
3602 * QEMU uses the return path.
3604 if (migrate_postcopy_ram() || migrate_return_path()) {
3605 if (open_return_path_on_source(s
)) {
3606 error_setg(&local_err
, "Unable to open return-path for postcopy");
3607 migrate_set_state(&s
->state
, s
->state
, MIGRATION_STATUS_FAILED
);
3608 migrate_set_error(s
, local_err
);
3609 error_report_err(local_err
);
3610 migrate_fd_cleanup(s
);
3616 * This needs to be done before resuming a postcopy. Note: for newer
3617 * QEMUs we will delay the channel creation until postcopy_start(), to
3618 * avoid disorder of channel creations.
3620 if (migrate_postcopy_preempt() && s
->preempt_pre_7_2
) {
3621 postcopy_preempt_setup(s
);
3625 /* Wakeup the main migration thread to do the recovery */
3626 migrate_set_state(&s
->state
, MIGRATION_STATUS_POSTCOPY_PAUSED
,
3627 MIGRATION_STATUS_POSTCOPY_RECOVER
);
3628 qemu_sem_post(&s
->postcopy_pause_sem
);
3632 if (multifd_save_setup(&local_err
) != 0) {
3633 migrate_set_error(s
, local_err
);
3634 error_report_err(local_err
);
3635 migrate_set_state(&s
->state
, MIGRATION_STATUS_SETUP
,
3636 MIGRATION_STATUS_FAILED
);
3637 migrate_fd_cleanup(s
);
3641 if (migrate_background_snapshot()) {
3642 qemu_thread_create(&s
->thread
, "bg_snapshot",
3643 bg_migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3645 qemu_thread_create(&s
->thread
, "live_migration",
3646 migration_thread
, s
, QEMU_THREAD_JOINABLE
);
3648 s
->migration_thread_running
= true;
3651 static void migration_class_init(ObjectClass
*klass
, void *data
)
3653 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3655 dc
->user_creatable
= false;
3656 device_class_set_props(dc
, migration_properties
);
3659 static void migration_instance_finalize(Object
*obj
)
3661 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3663 qemu_mutex_destroy(&ms
->error_mutex
);
3664 qemu_mutex_destroy(&ms
->qemu_file_lock
);
3665 qemu_sem_destroy(&ms
->wait_unplug_sem
);
3666 qemu_sem_destroy(&ms
->rate_limit_sem
);
3667 qemu_sem_destroy(&ms
->pause_sem
);
3668 qemu_sem_destroy(&ms
->postcopy_pause_sem
);
3669 qemu_sem_destroy(&ms
->rp_state
.rp_sem
);
3670 qemu_sem_destroy(&ms
->rp_state
.rp_pong_acks
);
3671 qemu_sem_destroy(&ms
->postcopy_qemufile_src_sem
);
3672 error_free(ms
->error
);
3675 static void migration_instance_init(Object
*obj
)
3677 MigrationState
*ms
= MIGRATION_OBJ(obj
);
3679 ms
->state
= MIGRATION_STATUS_NONE
;
3681 ms
->pages_per_second
= -1;
3682 qemu_sem_init(&ms
->pause_sem
, 0);
3683 qemu_mutex_init(&ms
->error_mutex
);
3685 migrate_params_init(&ms
->parameters
);
3687 qemu_sem_init(&ms
->postcopy_pause_sem
, 0);
3688 qemu_sem_init(&ms
->rp_state
.rp_sem
, 0);
3689 qemu_sem_init(&ms
->rp_state
.rp_pong_acks
, 0);
3690 qemu_sem_init(&ms
->rate_limit_sem
, 0);
3691 qemu_sem_init(&ms
->wait_unplug_sem
, 0);
3692 qemu_sem_init(&ms
->postcopy_qemufile_src_sem
, 0);
3693 qemu_mutex_init(&ms
->qemu_file_lock
);
3697 * Return true if check pass, false otherwise. Error will be put
3698 * inside errp if provided.
3700 static bool migration_object_check(MigrationState
*ms
, Error
**errp
)
3702 /* Assuming all off */
3703 bool old_caps
[MIGRATION_CAPABILITY__MAX
] = { 0 };
3705 if (!migrate_params_check(&ms
->parameters
, errp
)) {
3709 return migrate_caps_check(old_caps
, ms
->capabilities
, errp
);
3712 static const TypeInfo migration_type
= {
3713 .name
= TYPE_MIGRATION
,
3715 * NOTE: TYPE_MIGRATION is not really a device, as the object is
3716 * not created using qdev_new(), it is not attached to the qdev
3717 * device tree, and it is never realized.
3719 * TODO: Make this TYPE_OBJECT once QOM provides something like
3720 * TYPE_DEVICE's "-global" properties.
3722 .parent
= TYPE_DEVICE
,
3723 .class_init
= migration_class_init
,
3724 .class_size
= sizeof(MigrationClass
),
3725 .instance_size
= sizeof(MigrationState
),
3726 .instance_init
= migration_instance_init
,
3727 .instance_finalize
= migration_instance_finalize
,
3730 static void register_migration_types(void)
3732 type_register_static(&migration_type
);
3735 type_init(register_migration_types
);