2 * Postcopy migration for RAM
4 * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates
7 * Dave Gilbert <dgilbert@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
15 * Postcopy is a migration technique where the execution flips from the
16 * source to the destination before all the data has been copied.
19 #include "qemu/osdep.h"
21 #include "exec/target_page.h"
22 #include "migration.h"
23 #include "qemu-file.h"
25 #include "postcopy-ram.h"
27 #include "qapi/error.h"
28 #include "qemu/notify.h"
30 #include "sysemu/sysemu.h"
31 #include "qemu/error-report.h"
33 #include "hw/boards.h"
34 #include "exec/ramblock.h"
36 /* Arbitrary limit on size of each discard command,
37 * keeps them around ~200 bytes
39 #define MAX_DISCARDS_PER_COMMAND 12
41 struct PostcopyDiscardState
{
42 const char *ramblock_name
;
45 * Start and length of a discard range (bytes)
47 uint64_t start_list
[MAX_DISCARDS_PER_COMMAND
];
48 uint64_t length_list
[MAX_DISCARDS_PER_COMMAND
];
49 unsigned int nsentwords
;
50 unsigned int nsentcmds
;
53 static NotifierWithReturnList postcopy_notifier_list
;
55 void postcopy_infrastructure_init(void)
57 notifier_with_return_list_init(&postcopy_notifier_list
);
60 void postcopy_add_notifier(NotifierWithReturn
*nn
)
62 notifier_with_return_list_add(&postcopy_notifier_list
, nn
);
65 void postcopy_remove_notifier(NotifierWithReturn
*n
)
67 notifier_with_return_remove(n
);
70 int postcopy_notify(enum PostcopyNotifyReason reason
, Error
**errp
)
72 struct PostcopyNotifyData pnd
;
76 return notifier_with_return_list_notify(&postcopy_notifier_list
,
80 /* Postcopy needs to detect accesses to pages that haven't yet been copied
81 * across, and efficiently map new pages in, the techniques for doing this
82 * are target OS specific.
84 #if defined(__linux__)
87 #include <sys/ioctl.h>
88 #include <sys/syscall.h>
89 #include <asm/types.h> /* for __u64 */
92 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
93 #include <sys/eventfd.h>
94 #include <linux/userfaultfd.h>
96 typedef struct PostcopyBlocktimeContext
{
97 /* time when page fault initiated per vCPU */
98 uint32_t *page_fault_vcpu_time
;
99 /* page address per vCPU */
100 uintptr_t *vcpu_addr
;
101 uint32_t total_blocktime
;
102 /* blocktime per vCPU */
103 uint32_t *vcpu_blocktime
;
104 /* point in time when last page fault was initiated */
106 /* number of vCPU are suspended */
111 * Handler for exit event, necessary for
112 * releasing whole blocktime_ctx
114 Notifier exit_notifier
;
115 } PostcopyBlocktimeContext
;
117 static void destroy_blocktime_context(struct PostcopyBlocktimeContext
*ctx
)
119 g_free(ctx
->page_fault_vcpu_time
);
120 g_free(ctx
->vcpu_addr
);
121 g_free(ctx
->vcpu_blocktime
);
125 static void migration_exit_cb(Notifier
*n
, void *data
)
127 PostcopyBlocktimeContext
*ctx
= container_of(n
, PostcopyBlocktimeContext
,
129 destroy_blocktime_context(ctx
);
132 static struct PostcopyBlocktimeContext
*blocktime_context_new(void)
134 MachineState
*ms
= MACHINE(qdev_get_machine());
135 unsigned int smp_cpus
= ms
->smp
.cpus
;
136 PostcopyBlocktimeContext
*ctx
= g_new0(PostcopyBlocktimeContext
, 1);
137 ctx
->page_fault_vcpu_time
= g_new0(uint32_t, smp_cpus
);
138 ctx
->vcpu_addr
= g_new0(uintptr_t, smp_cpus
);
139 ctx
->vcpu_blocktime
= g_new0(uint32_t, smp_cpus
);
141 ctx
->exit_notifier
.notify
= migration_exit_cb
;
142 ctx
->start_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
143 qemu_add_exit_notifier(&ctx
->exit_notifier
);
147 static uint32List
*get_vcpu_blocktime_list(PostcopyBlocktimeContext
*ctx
)
149 MachineState
*ms
= MACHINE(qdev_get_machine());
150 uint32List
*list
= NULL
;
153 for (i
= ms
->smp
.cpus
- 1; i
>= 0; i
--) {
154 QAPI_LIST_PREPEND(list
, ctx
->vcpu_blocktime
[i
]);
161 * This function just populates MigrationInfo from postcopy's
162 * blocktime context. It will not populate MigrationInfo,
163 * unless postcopy-blocktime capability was set.
165 * @info: pointer to MigrationInfo to populate
167 void fill_destination_postcopy_migration_info(MigrationInfo
*info
)
169 MigrationIncomingState
*mis
= migration_incoming_get_current();
170 PostcopyBlocktimeContext
*bc
= mis
->blocktime_ctx
;
176 info
->has_postcopy_blocktime
= true;
177 info
->postcopy_blocktime
= bc
->total_blocktime
;
178 info
->has_postcopy_vcpu_blocktime
= true;
179 info
->postcopy_vcpu_blocktime
= get_vcpu_blocktime_list(bc
);
182 static uint32_t get_postcopy_total_blocktime(void)
184 MigrationIncomingState
*mis
= migration_incoming_get_current();
185 PostcopyBlocktimeContext
*bc
= mis
->blocktime_ctx
;
191 return bc
->total_blocktime
;
195 * receive_ufd_features: check userfault fd features, to request only supported
196 * features in the future.
198 * Returns: true on success
200 * __NR_userfaultfd - should be checked before
201 * @features: out parameter will contain uffdio_api.features provided by kernel
204 static bool receive_ufd_features(uint64_t *features
)
206 struct uffdio_api api_struct
= {0};
210 /* if we are here __NR_userfaultfd should exists */
211 ufd
= syscall(__NR_userfaultfd
, O_CLOEXEC
);
213 error_report("%s: syscall __NR_userfaultfd failed: %s", __func__
,
219 api_struct
.api
= UFFD_API
;
220 api_struct
.features
= 0;
221 if (ioctl(ufd
, UFFDIO_API
, &api_struct
)) {
222 error_report("%s: UFFDIO_API failed: %s", __func__
,
228 *features
= api_struct
.features
;
236 * request_ufd_features: this function should be called only once on a newly
237 * opened ufd, subsequent calls will lead to error.
239 * Returns: true on success
241 * @ufd: fd obtained from userfaultfd syscall
242 * @features: bit mask see UFFD_API_FEATURES
244 static bool request_ufd_features(int ufd
, uint64_t features
)
246 struct uffdio_api api_struct
= {0};
249 api_struct
.api
= UFFD_API
;
250 api_struct
.features
= features
;
251 if (ioctl(ufd
, UFFDIO_API
, &api_struct
)) {
252 error_report("%s failed: UFFDIO_API failed: %s", __func__
,
257 ioctl_mask
= (__u64
)1 << _UFFDIO_REGISTER
|
258 (__u64
)1 << _UFFDIO_UNREGISTER
;
259 if ((api_struct
.ioctls
& ioctl_mask
) != ioctl_mask
) {
260 error_report("Missing userfault features: %" PRIx64
,
261 (uint64_t)(~api_struct
.ioctls
& ioctl_mask
));
268 static bool ufd_check_and_apply(int ufd
, MigrationIncomingState
*mis
)
270 uint64_t asked_features
= 0;
271 static uint64_t supported_features
;
274 * it's not possible to
275 * request UFFD_API twice per one fd
276 * userfault fd features is persistent
278 if (!supported_features
) {
279 if (!receive_ufd_features(&supported_features
)) {
280 error_report("%s failed", __func__
);
285 #ifdef UFFD_FEATURE_THREAD_ID
286 if (migrate_postcopy_blocktime() && mis
&&
287 UFFD_FEATURE_THREAD_ID
& supported_features
) {
288 /* kernel supports that feature */
289 /* don't create blocktime_context if it exists */
290 if (!mis
->blocktime_ctx
) {
291 mis
->blocktime_ctx
= blocktime_context_new();
294 asked_features
|= UFFD_FEATURE_THREAD_ID
;
299 * request features, even if asked_features is 0, due to
300 * kernel expects UFFD_API before UFFDIO_REGISTER, per
301 * userfault file descriptor
303 if (!request_ufd_features(ufd
, asked_features
)) {
304 error_report("%s failed: features %" PRIu64
, __func__
,
309 if (qemu_real_host_page_size
!= ram_pagesize_summary()) {
310 bool have_hp
= false;
311 /* We've got a huge page */
312 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
313 have_hp
= supported_features
& UFFD_FEATURE_MISSING_HUGETLBFS
;
316 error_report("Userfault on this host does not support huge pages");
323 /* Callback from postcopy_ram_supported_by_host block iterator.
325 static int test_ramblock_postcopiable(RAMBlock
*rb
, void *opaque
)
327 const char *block_name
= qemu_ram_get_idstr(rb
);
328 ram_addr_t length
= qemu_ram_get_used_length(rb
);
329 size_t pagesize
= qemu_ram_pagesize(rb
);
331 if (length
% pagesize
) {
332 error_report("Postcopy requires RAM blocks to be a page size multiple,"
333 " block %s is 0x" RAM_ADDR_FMT
" bytes with a "
334 "page size of 0x%zx", block_name
, length
, pagesize
);
341 * Note: This has the side effect of munlock'ing all of RAM, that's
342 * normally fine since if the postcopy succeeds it gets turned back on at the
345 bool postcopy_ram_supported_by_host(MigrationIncomingState
*mis
)
347 long pagesize
= qemu_real_host_page_size
;
349 bool ret
= false; /* Error unless we change it */
350 void *testarea
= NULL
;
351 struct uffdio_register reg_struct
;
352 struct uffdio_range range_struct
;
353 uint64_t feature_mask
;
354 Error
*local_err
= NULL
;
356 if (qemu_target_page_size() > pagesize
) {
357 error_report("Target page size bigger than host page size");
361 ufd
= syscall(__NR_userfaultfd
, O_CLOEXEC
);
363 error_report("%s: userfaultfd not available: %s", __func__
,
368 /* Give devices a chance to object */
369 if (postcopy_notify(POSTCOPY_NOTIFY_PROBE
, &local_err
)) {
370 error_report_err(local_err
);
374 /* Version and features check */
375 if (!ufd_check_and_apply(ufd
, mis
)) {
379 /* We don't support postcopy with shared RAM yet */
380 if (foreach_not_ignored_block(test_ramblock_postcopiable
, NULL
)) {
385 * userfault and mlock don't go together; we'll put it back later if
389 error_report("%s: munlockall: %s", __func__
, strerror(errno
));
394 * We need to check that the ops we need are supported on anon memory
395 * To do that we need to register a chunk and see the flags that
398 testarea
= mmap(NULL
, pagesize
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
399 MAP_ANONYMOUS
, -1, 0);
400 if (testarea
== MAP_FAILED
) {
401 error_report("%s: Failed to map test area: %s", __func__
,
405 g_assert(QEMU_PTR_IS_ALIGNED(testarea
, pagesize
));
407 reg_struct
.range
.start
= (uintptr_t)testarea
;
408 reg_struct
.range
.len
= pagesize
;
409 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
411 if (ioctl(ufd
, UFFDIO_REGISTER
, ®_struct
)) {
412 error_report("%s userfault register: %s", __func__
, strerror(errno
));
416 range_struct
.start
= (uintptr_t)testarea
;
417 range_struct
.len
= pagesize
;
418 if (ioctl(ufd
, UFFDIO_UNREGISTER
, &range_struct
)) {
419 error_report("%s userfault unregister: %s", __func__
, strerror(errno
));
423 feature_mask
= (__u64
)1 << _UFFDIO_WAKE
|
424 (__u64
)1 << _UFFDIO_COPY
|
425 (__u64
)1 << _UFFDIO_ZEROPAGE
;
426 if ((reg_struct
.ioctls
& feature_mask
) != feature_mask
) {
427 error_report("Missing userfault map features: %" PRIx64
,
428 (uint64_t)(~reg_struct
.ioctls
& feature_mask
));
436 munmap(testarea
, pagesize
);
445 * Setup an area of RAM so that it *can* be used for postcopy later; this
446 * must be done right at the start prior to pre-copy.
447 * opaque should be the MIS.
449 static int init_range(RAMBlock
*rb
, void *opaque
)
451 const char *block_name
= qemu_ram_get_idstr(rb
);
452 void *host_addr
= qemu_ram_get_host_addr(rb
);
453 ram_addr_t offset
= qemu_ram_get_offset(rb
);
454 ram_addr_t length
= qemu_ram_get_used_length(rb
);
455 trace_postcopy_init_range(block_name
, host_addr
, offset
, length
);
458 * Save the used_length before running the guest. In case we have to
459 * resize RAM blocks when syncing RAM block sizes from the source during
460 * precopy, we'll update it manually via the ram block notifier.
462 rb
->postcopy_length
= length
;
465 * We need the whole of RAM to be truly empty for postcopy, so things
466 * like ROMs and any data tables built during init must be zero'd
467 * - we're going to get the copy from the source anyway.
468 * (Precopy will just overwrite this data, so doesn't need the discard)
470 if (ram_discard_range(block_name
, 0, length
)) {
478 * At the end of migration, undo the effects of init_range
479 * opaque should be the MIS.
481 static int cleanup_range(RAMBlock
*rb
, void *opaque
)
483 const char *block_name
= qemu_ram_get_idstr(rb
);
484 void *host_addr
= qemu_ram_get_host_addr(rb
);
485 ram_addr_t offset
= qemu_ram_get_offset(rb
);
486 ram_addr_t length
= rb
->postcopy_length
;
487 MigrationIncomingState
*mis
= opaque
;
488 struct uffdio_range range_struct
;
489 trace_postcopy_cleanup_range(block_name
, host_addr
, offset
, length
);
492 * We turned off hugepage for the precopy stage with postcopy enabled
493 * we can turn it back on now.
495 qemu_madvise(host_addr
, length
, QEMU_MADV_HUGEPAGE
);
498 * We can also turn off userfault now since we should have all the
499 * pages. It can be useful to leave it on to debug postcopy
500 * if you're not sure it's always getting every page.
502 range_struct
.start
= (uintptr_t)host_addr
;
503 range_struct
.len
= length
;
505 if (ioctl(mis
->userfault_fd
, UFFDIO_UNREGISTER
, &range_struct
)) {
506 error_report("%s: userfault unregister %s", __func__
, strerror(errno
));
515 * Initialise postcopy-ram, setting the RAM to a state where we can go into
516 * postcopy later; must be called prior to any precopy.
517 * called from arch_init's similarly named ram_postcopy_incoming_init
519 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
)
521 if (foreach_not_ignored_block(init_range
, NULL
)) {
529 * At the end of a migration where postcopy_ram_incoming_init was called.
531 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
533 trace_postcopy_ram_incoming_cleanup_entry();
535 if (mis
->have_fault_thread
) {
536 Error
*local_err
= NULL
;
538 /* Let the fault thread quit */
539 qatomic_set(&mis
->fault_thread_quit
, 1);
540 postcopy_fault_thread_notify(mis
);
541 trace_postcopy_ram_incoming_cleanup_join();
542 qemu_thread_join(&mis
->fault_thread
);
544 if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END
, &local_err
)) {
545 error_report_err(local_err
);
549 if (foreach_not_ignored_block(cleanup_range
, mis
)) {
553 trace_postcopy_ram_incoming_cleanup_closeuf();
554 close(mis
->userfault_fd
);
555 close(mis
->userfault_event_fd
);
556 mis
->have_fault_thread
= false;
560 if (os_mlock() < 0) {
561 error_report("mlock: %s", strerror(errno
));
563 * It doesn't feel right to fail at this point, we have a valid
569 if (mis
->postcopy_tmp_page
) {
570 munmap(mis
->postcopy_tmp_page
, mis
->largest_page_size
);
571 mis
->postcopy_tmp_page
= NULL
;
573 if (mis
->postcopy_tmp_zero_page
) {
574 munmap(mis
->postcopy_tmp_zero_page
, mis
->largest_page_size
);
575 mis
->postcopy_tmp_zero_page
= NULL
;
577 trace_postcopy_ram_incoming_cleanup_blocktime(
578 get_postcopy_total_blocktime());
580 trace_postcopy_ram_incoming_cleanup_exit();
585 * Disable huge pages on an area
587 static int nhp_range(RAMBlock
*rb
, void *opaque
)
589 const char *block_name
= qemu_ram_get_idstr(rb
);
590 void *host_addr
= qemu_ram_get_host_addr(rb
);
591 ram_addr_t offset
= qemu_ram_get_offset(rb
);
592 ram_addr_t length
= rb
->postcopy_length
;
593 trace_postcopy_nhp_range(block_name
, host_addr
, offset
, length
);
596 * Before we do discards we need to ensure those discards really
597 * do delete areas of the page, even if THP thinks a hugepage would
598 * be a good idea, so force hugepages off.
600 qemu_madvise(host_addr
, length
, QEMU_MADV_NOHUGEPAGE
);
606 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
607 * however leaving it until after precopy means that most of the precopy
610 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
612 if (foreach_not_ignored_block(nhp_range
, mis
)) {
616 postcopy_state_set(POSTCOPY_INCOMING_DISCARD
);
622 * Mark the given area of RAM as requiring notification to unwritten areas
623 * Used as a callback on foreach_not_ignored_block.
624 * host_addr: Base of area to mark
625 * offset: Offset in the whole ram arena
626 * length: Length of the section
627 * opaque: MigrationIncomingState pointer
628 * Returns 0 on success
630 static int ram_block_enable_notify(RAMBlock
*rb
, void *opaque
)
632 MigrationIncomingState
*mis
= opaque
;
633 struct uffdio_register reg_struct
;
635 reg_struct
.range
.start
= (uintptr_t)qemu_ram_get_host_addr(rb
);
636 reg_struct
.range
.len
= rb
->postcopy_length
;
637 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
639 /* Now tell our userfault_fd that it's responsible for this area */
640 if (ioctl(mis
->userfault_fd
, UFFDIO_REGISTER
, ®_struct
)) {
641 error_report("%s userfault register: %s", __func__
, strerror(errno
));
644 if (!(reg_struct
.ioctls
& ((__u64
)1 << _UFFDIO_COPY
))) {
645 error_report("%s userfault: Region doesn't support COPY", __func__
);
648 if (reg_struct
.ioctls
& ((__u64
)1 << _UFFDIO_ZEROPAGE
)) {
649 qemu_ram_set_uf_zeroable(rb
);
655 int postcopy_wake_shared(struct PostCopyFD
*pcfd
,
656 uint64_t client_addr
,
659 size_t pagesize
= qemu_ram_pagesize(rb
);
660 struct uffdio_range range
;
662 trace_postcopy_wake_shared(client_addr
, qemu_ram_get_idstr(rb
));
663 range
.start
= ROUND_DOWN(client_addr
, pagesize
);
664 range
.len
= pagesize
;
665 ret
= ioctl(pcfd
->fd
, UFFDIO_WAKE
, &range
);
667 error_report("%s: Failed to wake: %zx in %s (%s)",
668 __func__
, (size_t)client_addr
, qemu_ram_get_idstr(rb
),
674 static int postcopy_request_page(MigrationIncomingState
*mis
, RAMBlock
*rb
,
675 ram_addr_t start
, uint64_t haddr
)
677 void *aligned
= (void *)(uintptr_t)ROUND_DOWN(haddr
, qemu_ram_pagesize(rb
));
680 * Discarded pages (via RamDiscardManager) are never migrated. On unlikely
681 * access, place a zeropage, which will also set the relevant bits in the
682 * recv_bitmap accordingly, so we won't try placing a zeropage twice.
684 * Checking a single bit is sufficient to handle pagesize > TPS as either
685 * all relevant bits are set or not.
687 assert(QEMU_IS_ALIGNED(start
, qemu_ram_pagesize(rb
)));
688 if (ramblock_page_is_discarded(rb
, start
)) {
689 bool received
= ramblock_recv_bitmap_test_byte_offset(rb
, start
);
691 return received
? 0 : postcopy_place_page_zero(mis
, aligned
, rb
);
694 return migrate_send_rp_req_pages(mis
, rb
, start
, haddr
);
698 * Callback from shared fault handlers to ask for a page,
699 * the page must be specified by a RAMBlock and an offset in that rb
700 * Note: Only for use by shared fault handlers (in fault thread)
702 int postcopy_request_shared_page(struct PostCopyFD
*pcfd
, RAMBlock
*rb
,
703 uint64_t client_addr
, uint64_t rb_offset
)
705 uint64_t aligned_rbo
= ROUND_DOWN(rb_offset
, qemu_ram_pagesize(rb
));
706 MigrationIncomingState
*mis
= migration_incoming_get_current();
708 trace_postcopy_request_shared_page(pcfd
->idstr
, qemu_ram_get_idstr(rb
),
710 if (ramblock_recv_bitmap_test_byte_offset(rb
, aligned_rbo
)) {
711 trace_postcopy_request_shared_page_present(pcfd
->idstr
,
712 qemu_ram_get_idstr(rb
), rb_offset
);
713 return postcopy_wake_shared(pcfd
, client_addr
, rb
);
715 postcopy_request_page(mis
, rb
, aligned_rbo
, client_addr
);
719 static int get_mem_fault_cpu_index(uint32_t pid
)
723 CPU_FOREACH(cpu_iter
) {
724 if (cpu_iter
->thread_id
== pid
) {
725 trace_get_mem_fault_cpu_index(cpu_iter
->cpu_index
, pid
);
726 return cpu_iter
->cpu_index
;
729 trace_get_mem_fault_cpu_index(-1, pid
);
733 static uint32_t get_low_time_offset(PostcopyBlocktimeContext
*dc
)
735 int64_t start_time_offset
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
) -
737 return start_time_offset
< 1 ? 1 : start_time_offset
& UINT32_MAX
;
741 * This function is being called when pagefault occurs. It
742 * tracks down vCPU blocking time.
744 * @addr: faulted host virtual address
745 * @ptid: faulted process thread id
746 * @rb: ramblock appropriate to addr
748 static void mark_postcopy_blocktime_begin(uintptr_t addr
, uint32_t ptid
,
751 int cpu
, already_received
;
752 MigrationIncomingState
*mis
= migration_incoming_get_current();
753 PostcopyBlocktimeContext
*dc
= mis
->blocktime_ctx
;
754 uint32_t low_time_offset
;
756 if (!dc
|| ptid
== 0) {
759 cpu
= get_mem_fault_cpu_index(ptid
);
764 low_time_offset
= get_low_time_offset(dc
);
765 if (dc
->vcpu_addr
[cpu
] == 0) {
766 qatomic_inc(&dc
->smp_cpus_down
);
769 qatomic_xchg(&dc
->last_begin
, low_time_offset
);
770 qatomic_xchg(&dc
->page_fault_vcpu_time
[cpu
], low_time_offset
);
771 qatomic_xchg(&dc
->vcpu_addr
[cpu
], addr
);
774 * check it here, not at the beginning of the function,
775 * due to, check could occur early than bitmap_set in
776 * qemu_ufd_copy_ioctl
778 already_received
= ramblock_recv_bitmap_test(rb
, (void *)addr
);
779 if (already_received
) {
780 qatomic_xchg(&dc
->vcpu_addr
[cpu
], 0);
781 qatomic_xchg(&dc
->page_fault_vcpu_time
[cpu
], 0);
782 qatomic_dec(&dc
->smp_cpus_down
);
784 trace_mark_postcopy_blocktime_begin(addr
, dc
, dc
->page_fault_vcpu_time
[cpu
],
785 cpu
, already_received
);
789 * This function just provide calculated blocktime per cpu and trace it.
790 * Total blocktime is calculated in mark_postcopy_blocktime_end.
793 * Assume we have 3 CPU
796 * -----***********------------xxx***************------------------------> CPU1
799 * ------------****************xxx---------------------------------------> CPU2
802 * ------------------------****xxx********-------------------------------> CPU3
804 * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
805 * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
806 * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
807 * it's a part of total blocktime.
808 * S1 - here is last_begin
809 * Legend of the picture is following:
810 * * - means blocktime per vCPU
811 * x - means overlapped blocktime (total blocktime)
813 * @addr: host virtual address
815 static void mark_postcopy_blocktime_end(uintptr_t addr
)
817 MigrationIncomingState
*mis
= migration_incoming_get_current();
818 PostcopyBlocktimeContext
*dc
= mis
->blocktime_ctx
;
819 MachineState
*ms
= MACHINE(qdev_get_machine());
820 unsigned int smp_cpus
= ms
->smp
.cpus
;
821 int i
, affected_cpu
= 0;
822 bool vcpu_total_blocktime
= false;
823 uint32_t read_vcpu_time
, low_time_offset
;
829 low_time_offset
= get_low_time_offset(dc
);
830 /* lookup cpu, to clear it,
831 * that algorithm looks straightforward, but it's not
832 * optimal, more optimal algorithm is keeping tree or hash
833 * where key is address value is a list of */
834 for (i
= 0; i
< smp_cpus
; i
++) {
835 uint32_t vcpu_blocktime
= 0;
837 read_vcpu_time
= qatomic_fetch_add(&dc
->page_fault_vcpu_time
[i
], 0);
838 if (qatomic_fetch_add(&dc
->vcpu_addr
[i
], 0) != addr
||
839 read_vcpu_time
== 0) {
842 qatomic_xchg(&dc
->vcpu_addr
[i
], 0);
843 vcpu_blocktime
= low_time_offset
- read_vcpu_time
;
845 /* we need to know is that mark_postcopy_end was due to
846 * faulted page, another possible case it's prefetched
847 * page and in that case we shouldn't be here */
848 if (!vcpu_total_blocktime
&&
849 qatomic_fetch_add(&dc
->smp_cpus_down
, 0) == smp_cpus
) {
850 vcpu_total_blocktime
= true;
852 /* continue cycle, due to one page could affect several vCPUs */
853 dc
->vcpu_blocktime
[i
] += vcpu_blocktime
;
856 qatomic_sub(&dc
->smp_cpus_down
, affected_cpu
);
857 if (vcpu_total_blocktime
) {
858 dc
->total_blocktime
+= low_time_offset
- qatomic_fetch_add(
861 trace_mark_postcopy_blocktime_end(addr
, dc
, dc
->total_blocktime
,
865 static bool postcopy_pause_fault_thread(MigrationIncomingState
*mis
)
867 trace_postcopy_pause_fault_thread();
869 qemu_sem_wait(&mis
->postcopy_pause_sem_fault
);
871 trace_postcopy_pause_fault_thread_continued();
877 * Handle faults detected by the USERFAULT markings
879 static void *postcopy_ram_fault_thread(void *opaque
)
881 MigrationIncomingState
*mis
= opaque
;
887 trace_postcopy_ram_fault_thread_entry();
888 rcu_register_thread();
889 mis
->last_rb
= NULL
; /* last RAMBlock we sent part of */
890 qemu_sem_post(&mis
->fault_thread_sem
);
893 size_t pfd_len
= 2 + mis
->postcopy_remote_fds
->len
;
895 pfd
= g_new0(struct pollfd
, pfd_len
);
897 pfd
[0].fd
= mis
->userfault_fd
;
898 pfd
[0].events
= POLLIN
;
899 pfd
[1].fd
= mis
->userfault_event_fd
;
900 pfd
[1].events
= POLLIN
; /* Waiting for eventfd to go positive */
901 trace_postcopy_ram_fault_thread_fds_core(pfd
[0].fd
, pfd
[1].fd
);
902 for (index
= 0; index
< mis
->postcopy_remote_fds
->len
; index
++) {
903 struct PostCopyFD
*pcfd
= &g_array_index(mis
->postcopy_remote_fds
,
904 struct PostCopyFD
, index
);
905 pfd
[2 + index
].fd
= pcfd
->fd
;
906 pfd
[2 + index
].events
= POLLIN
;
907 trace_postcopy_ram_fault_thread_fds_extra(2 + index
, pcfd
->idstr
,
912 ram_addr_t rb_offset
;
916 * We're mainly waiting for the kernel to give us a faulting HVA,
917 * however we can be told to quit via userfault_quit_fd which is
921 poll_result
= poll(pfd
, pfd_len
, -1 /* Wait forever */);
922 if (poll_result
== -1) {
923 error_report("%s: userfault poll: %s", __func__
, strerror(errno
));
927 if (!mis
->to_src_file
) {
929 * Possibly someone tells us that the return path is
930 * broken already using the event. We should hold until
931 * the channel is rebuilt.
933 if (postcopy_pause_fault_thread(mis
)) {
934 /* Continue to read the userfaultfd */
936 error_report("%s: paused but don't allow to continue",
942 if (pfd
[1].revents
) {
945 /* Consume the signal */
946 if (read(mis
->userfault_event_fd
, &tmp64
, 8) != 8) {
947 /* Nothing obviously nicer than posting this error. */
948 error_report("%s: read() failed", __func__
);
951 if (qatomic_read(&mis
->fault_thread_quit
)) {
952 trace_postcopy_ram_fault_thread_quit();
957 if (pfd
[0].revents
) {
959 ret
= read(mis
->userfault_fd
, &msg
, sizeof(msg
));
960 if (ret
!= sizeof(msg
)) {
961 if (errno
== EAGAIN
) {
963 * if a wake up happens on the other thread just after
964 * the poll, there is nothing to read.
969 error_report("%s: Failed to read full userfault "
971 __func__
, strerror(errno
));
974 error_report("%s: Read %d bytes from userfaultfd "
976 __func__
, ret
, sizeof(msg
));
977 break; /* Lost alignment, don't know what we'd read next */
980 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
) {
981 error_report("%s: Read unexpected event %ud from userfaultfd",
982 __func__
, msg
.event
);
983 continue; /* It's not a page fault, shouldn't happen */
986 rb
= qemu_ram_block_from_host(
987 (void *)(uintptr_t)msg
.arg
.pagefault
.address
,
990 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
991 PRIx64
, (uint64_t)msg
.arg
.pagefault
.address
);
995 rb_offset
= ROUND_DOWN(rb_offset
, qemu_ram_pagesize(rb
));
996 trace_postcopy_ram_fault_thread_request(msg
.arg
.pagefault
.address
,
997 qemu_ram_get_idstr(rb
),
999 msg
.arg
.pagefault
.feat
.ptid
);
1000 mark_postcopy_blocktime_begin(
1001 (uintptr_t)(msg
.arg
.pagefault
.address
),
1002 msg
.arg
.pagefault
.feat
.ptid
, rb
);
1006 * Send the request to the source - we want to request one
1007 * of our host page sizes (which is >= TPS)
1009 ret
= postcopy_request_page(mis
, rb
, rb_offset
,
1010 msg
.arg
.pagefault
.address
);
1012 /* May be network failure, try to wait for recovery */
1013 if (ret
== -EIO
&& postcopy_pause_fault_thread(mis
)) {
1014 /* We got reconnected somehow, try to continue */
1017 /* This is a unavoidable fault */
1018 error_report("%s: postcopy_request_page() get %d",
1025 /* Now handle any requests from external processes on shared memory */
1026 /* TODO: May need to handle devices deregistering during postcopy */
1027 for (index
= 2; index
< pfd_len
&& poll_result
; index
++) {
1028 if (pfd
[index
].revents
) {
1029 struct PostCopyFD
*pcfd
=
1030 &g_array_index(mis
->postcopy_remote_fds
,
1031 struct PostCopyFD
, index
- 2);
1034 if (pfd
[index
].revents
& POLLERR
) {
1035 error_report("%s: POLLERR on poll %zd fd=%d",
1036 __func__
, index
, pcfd
->fd
);
1037 pfd
[index
].events
= 0;
1041 ret
= read(pcfd
->fd
, &msg
, sizeof(msg
));
1042 if (ret
!= sizeof(msg
)) {
1043 if (errno
== EAGAIN
) {
1045 * if a wake up happens on the other thread just after
1046 * the poll, there is nothing to read.
1051 error_report("%s: Failed to read full userfault "
1052 "message: %s (shared) revents=%d",
1053 __func__
, strerror(errno
),
1054 pfd
[index
].revents
);
1055 /*TODO: Could just disable this sharer */
1058 error_report("%s: Read %d bytes from userfaultfd "
1059 "expected %zd (shared)",
1060 __func__
, ret
, sizeof(msg
));
1061 /*TODO: Could just disable this sharer */
1062 break; /*Lost alignment,don't know what we'd read next*/
1065 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
) {
1066 error_report("%s: Read unexpected event %ud "
1067 "from userfaultfd (shared)",
1068 __func__
, msg
.event
);
1069 continue; /* It's not a page fault, shouldn't happen */
1071 /* Call the device handler registered with us */
1072 ret
= pcfd
->handler(pcfd
, &msg
);
1074 error_report("%s: Failed to resolve shared fault on %zd/%s",
1075 __func__
, index
, pcfd
->idstr
);
1076 /* TODO: Fail? Disable this sharer? */
1081 rcu_unregister_thread();
1082 trace_postcopy_ram_fault_thread_exit();
1087 int postcopy_ram_incoming_setup(MigrationIncomingState
*mis
)
1089 /* Open the fd for the kernel to give us userfaults */
1090 mis
->userfault_fd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
1091 if (mis
->userfault_fd
== -1) {
1092 error_report("%s: Failed to open userfault fd: %s", __func__
,
1098 * Although the host check already tested the API, we need to
1099 * do the check again as an ABI handshake on the new fd.
1101 if (!ufd_check_and_apply(mis
->userfault_fd
, mis
)) {
1105 /* Now an eventfd we use to tell the fault-thread to quit */
1106 mis
->userfault_event_fd
= eventfd(0, EFD_CLOEXEC
);
1107 if (mis
->userfault_event_fd
== -1) {
1108 error_report("%s: Opening userfault_event_fd: %s", __func__
,
1110 close(mis
->userfault_fd
);
1114 qemu_sem_init(&mis
->fault_thread_sem
, 0);
1115 qemu_thread_create(&mis
->fault_thread
, "postcopy/fault",
1116 postcopy_ram_fault_thread
, mis
, QEMU_THREAD_JOINABLE
);
1117 qemu_sem_wait(&mis
->fault_thread_sem
);
1118 qemu_sem_destroy(&mis
->fault_thread_sem
);
1119 mis
->have_fault_thread
= true;
1121 /* Mark so that we get notified of accesses to unwritten areas */
1122 if (foreach_not_ignored_block(ram_block_enable_notify
, mis
)) {
1123 error_report("ram_block_enable_notify failed");
1127 mis
->postcopy_tmp_page
= mmap(NULL
, mis
->largest_page_size
,
1128 PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
1129 MAP_ANONYMOUS
, -1, 0);
1130 if (mis
->postcopy_tmp_page
== MAP_FAILED
) {
1131 mis
->postcopy_tmp_page
= NULL
;
1132 error_report("%s: Failed to map postcopy_tmp_page %s",
1133 __func__
, strerror(errno
));
1138 * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages
1140 mis
->postcopy_tmp_zero_page
= mmap(NULL
, mis
->largest_page_size
,
1141 PROT_READ
| PROT_WRITE
,
1142 MAP_PRIVATE
| MAP_ANONYMOUS
,
1144 if (mis
->postcopy_tmp_zero_page
== MAP_FAILED
) {
1146 mis
->postcopy_tmp_zero_page
= NULL
;
1147 error_report("%s: Failed to map large zero page %s",
1148 __func__
, strerror(e
));
1151 memset(mis
->postcopy_tmp_zero_page
, '\0', mis
->largest_page_size
);
1153 trace_postcopy_ram_enable_notify();
1158 static int qemu_ufd_copy_ioctl(MigrationIncomingState
*mis
, void *host_addr
,
1159 void *from_addr
, uint64_t pagesize
, RAMBlock
*rb
)
1161 int userfault_fd
= mis
->userfault_fd
;
1165 struct uffdio_copy copy_struct
;
1166 copy_struct
.dst
= (uint64_t)(uintptr_t)host_addr
;
1167 copy_struct
.src
= (uint64_t)(uintptr_t)from_addr
;
1168 copy_struct
.len
= pagesize
;
1169 copy_struct
.mode
= 0;
1170 ret
= ioctl(userfault_fd
, UFFDIO_COPY
, ©_struct
);
1172 struct uffdio_zeropage zero_struct
;
1173 zero_struct
.range
.start
= (uint64_t)(uintptr_t)host_addr
;
1174 zero_struct
.range
.len
= pagesize
;
1175 zero_struct
.mode
= 0;
1176 ret
= ioctl(userfault_fd
, UFFDIO_ZEROPAGE
, &zero_struct
);
1179 qemu_mutex_lock(&mis
->page_request_mutex
);
1180 ramblock_recv_bitmap_set_range(rb
, host_addr
,
1181 pagesize
/ qemu_target_page_size());
1183 * If this page resolves a page fault for a previous recorded faulted
1184 * address, take a special note to maintain the requested page list.
1186 if (g_tree_lookup(mis
->page_requested
, host_addr
)) {
1187 g_tree_remove(mis
->page_requested
, host_addr
);
1188 mis
->page_requested_count
--;
1189 trace_postcopy_page_req_del(host_addr
, mis
->page_requested_count
);
1191 qemu_mutex_unlock(&mis
->page_request_mutex
);
1192 mark_postcopy_blocktime_end((uintptr_t)host_addr
);
1197 int postcopy_notify_shared_wake(RAMBlock
*rb
, uint64_t offset
)
1200 MigrationIncomingState
*mis
= migration_incoming_get_current();
1201 GArray
*pcrfds
= mis
->postcopy_remote_fds
;
1203 for (i
= 0; i
< pcrfds
->len
; i
++) {
1204 struct PostCopyFD
*cur
= &g_array_index(pcrfds
, struct PostCopyFD
, i
);
1205 int ret
= cur
->waker(cur
, rb
, offset
);
1214 * Place a host page (from) at (host) atomically
1215 * returns 0 on success
1217 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
1220 size_t pagesize
= qemu_ram_pagesize(rb
);
1222 /* copy also acks to the kernel waking the stalled thread up
1223 * TODO: We can inhibit that ack and only do it if it was requested
1224 * which would be slightly cheaper, but we'd have to be careful
1225 * of the order of updating our page state.
1227 if (qemu_ufd_copy_ioctl(mis
, host
, from
, pagesize
, rb
)) {
1229 error_report("%s: %s copy host: %p from: %p (size: %zd)",
1230 __func__
, strerror(e
), host
, from
, pagesize
);
1235 trace_postcopy_place_page(host
);
1236 return postcopy_notify_shared_wake(rb
,
1237 qemu_ram_block_host_offset(rb
, host
));
1241 * Place a zero page at (host) atomically
1242 * returns 0 on success
1244 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
1247 size_t pagesize
= qemu_ram_pagesize(rb
);
1248 trace_postcopy_place_page_zero(host
);
1250 /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1251 * but it's not available for everything (e.g. hugetlbpages)
1253 if (qemu_ram_is_uf_zeroable(rb
)) {
1254 if (qemu_ufd_copy_ioctl(mis
, host
, NULL
, pagesize
, rb
)) {
1256 error_report("%s: %s zero host: %p",
1257 __func__
, strerror(e
), host
);
1261 return postcopy_notify_shared_wake(rb
,
1262 qemu_ram_block_host_offset(rb
,
1265 return postcopy_place_page(mis
, host
, mis
->postcopy_tmp_zero_page
, rb
);
1270 /* No target OS support, stubs just fail */
1271 void fill_destination_postcopy_migration_info(MigrationInfo
*info
)
1275 bool postcopy_ram_supported_by_host(MigrationIncomingState
*mis
)
1277 error_report("%s: No OS support", __func__
);
1281 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
)
1283 error_report("postcopy_ram_incoming_init: No OS support");
1287 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
1293 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
1299 int postcopy_request_shared_page(struct PostCopyFD
*pcfd
, RAMBlock
*rb
,
1300 uint64_t client_addr
, uint64_t rb_offset
)
1306 int postcopy_ram_incoming_setup(MigrationIncomingState
*mis
)
1312 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
1319 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
1326 int postcopy_wake_shared(struct PostCopyFD
*pcfd
,
1327 uint64_t client_addr
,
1335 /* ------------------------------------------------------------------------- */
1337 void postcopy_fault_thread_notify(MigrationIncomingState
*mis
)
1342 * Wakeup the fault_thread. It's an eventfd that should currently
1343 * be at 0, we're going to increment it to 1
1345 if (write(mis
->userfault_event_fd
, &tmp64
, 8) != 8) {
1346 /* Not much we can do here, but may as well report it */
1347 error_report("%s: incrementing failed: %s", __func__
,
1353 * postcopy_discard_send_init: Called at the start of each RAMBlock before
1354 * asking to discard individual ranges.
1356 * @ms: The current migration state.
1357 * @offset: the bitmap offset of the named RAMBlock in the migration bitmap.
1358 * @name: RAMBlock that discards will operate on.
1360 static PostcopyDiscardState pds
= {0};
1361 void postcopy_discard_send_init(MigrationState
*ms
, const char *name
)
1363 pds
.ramblock_name
= name
;
1370 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1371 * discard. May send a discard message, may just leave it queued to
1374 * @ms: Current migration state.
1375 * @start,@length: a range of pages in the migration bitmap in the
1376 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1378 void postcopy_discard_send_range(MigrationState
*ms
, unsigned long start
,
1379 unsigned long length
)
1381 size_t tp_size
= qemu_target_page_size();
1382 /* Convert to byte offsets within the RAM block */
1383 pds
.start_list
[pds
.cur_entry
] = start
* tp_size
;
1384 pds
.length_list
[pds
.cur_entry
] = length
* tp_size
;
1385 trace_postcopy_discard_send_range(pds
.ramblock_name
, start
, length
);
1389 if (pds
.cur_entry
== MAX_DISCARDS_PER_COMMAND
) {
1390 /* Full set, ship it! */
1391 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
1402 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1403 * bitmap code. Sends any outstanding discard messages, frees the PDS
1405 * @ms: Current migration state.
1407 void postcopy_discard_send_finish(MigrationState
*ms
)
1409 /* Anything unsent? */
1410 if (pds
.cur_entry
) {
1411 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
1419 trace_postcopy_discard_send_finish(pds
.ramblock_name
, pds
.nsentwords
,
1424 * Current state of incoming postcopy; note this is not part of
1425 * MigrationIncomingState since it's state is used during cleanup
1426 * at the end as MIS is being freed.
1428 static PostcopyState incoming_postcopy_state
;
1430 PostcopyState
postcopy_state_get(void)
1432 return qatomic_mb_read(&incoming_postcopy_state
);
1435 /* Set the state and return the old state */
1436 PostcopyState
postcopy_state_set(PostcopyState new_state
)
1438 return qatomic_xchg(&incoming_postcopy_state
, new_state
);
1441 /* Register a handler for external shared memory postcopy
1442 * called on the destination.
1444 void postcopy_register_shared_ufd(struct PostCopyFD
*pcfd
)
1446 MigrationIncomingState
*mis
= migration_incoming_get_current();
1448 mis
->postcopy_remote_fds
= g_array_append_val(mis
->postcopy_remote_fds
,
1452 /* Unregister a handler for external shared memory postcopy
1454 void postcopy_unregister_shared_ufd(struct PostCopyFD
*pcfd
)
1457 MigrationIncomingState
*mis
= migration_incoming_get_current();
1458 GArray
*pcrfds
= mis
->postcopy_remote_fds
;
1461 /* migration has already finished and freed the array */
1464 for (i
= 0; i
< pcrfds
->len
; i
++) {
1465 struct PostCopyFD
*cur
= &g_array_index(pcrfds
, struct PostCopyFD
, i
);
1466 if (cur
->fd
== pcfd
->fd
) {
1467 mis
->postcopy_remote_fds
= g_array_remove_index(pcrfds
, i
);