4 * Copyright (c) 2003-2008 Fabrice Bellard
5 * Copyright (c) 2011-2015 Red Hat Inc
8 * Juan Quintela <quintela@redhat.com>
10 * Permission is hereby granted, free of charge, to any person obtaining a copy
11 * of this software and associated documentation files (the "Software"), to deal
12 * in the Software without restriction, including without limitation the rights
13 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
14 * copies of the Software, and to permit persons to whom the Software is
15 * furnished to do so, subject to the following conditions:
17 * The above copyright notice and this permission notice shall be included in
18 * all copies or substantial portions of the Software.
20 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
23 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
24 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
25 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
29 #include "qemu/osdep.h"
30 #include "qemu/cutils.h"
31 #include "qemu/bitops.h"
32 #include "qemu/bitmap.h"
33 #include "qemu/madvise.h"
34 #include "qemu/main-loop.h"
36 #include "ram-compress.h"
38 #include "migration.h"
39 #include "migration-stats.h"
40 #include "migration/register.h"
41 #include "migration/misc.h"
42 #include "qemu-file.h"
43 #include "postcopy-ram.h"
44 #include "page_cache.h"
45 #include "qemu/error-report.h"
46 #include "qapi/error.h"
47 #include "qapi/qapi-types-migration.h"
48 #include "qapi/qapi-events-migration.h"
49 #include "qapi/qapi-commands-migration.h"
50 #include "qapi/qmp/qerror.h"
52 #include "exec/ram_addr.h"
53 #include "exec/target_page.h"
54 #include "qemu/rcu_queue.h"
55 #include "migration/colo.h"
57 #include "sysemu/cpu-throttle.h"
61 #include "sysemu/runstate.h"
64 #include "sysemu/dirtylimit.h"
65 #include "sysemu/kvm.h"
67 #include "hw/boards.h" /* for machine_dump_guest_core() */
69 #if defined(__linux__)
70 #include "qemu/userfaultfd.h"
71 #endif /* defined(__linux__) */
73 /***********************************************************/
74 /* ram save/restore */
77 * RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it
78 * worked for pages that were filled with the same char. We switched
79 * it to only search for the zero value. And to avoid confusion with
80 * RAM_SAVE_FLAG_COMPRESS_PAGE just rename it.
83 * RAM_SAVE_FLAG_FULL was obsoleted in 2009, it can be reused now
85 #define RAM_SAVE_FLAG_FULL 0x01
86 #define RAM_SAVE_FLAG_ZERO 0x02
87 #define RAM_SAVE_FLAG_MEM_SIZE 0x04
88 #define RAM_SAVE_FLAG_PAGE 0x08
89 #define RAM_SAVE_FLAG_EOS 0x10
90 #define RAM_SAVE_FLAG_CONTINUE 0x20
91 #define RAM_SAVE_FLAG_XBZRLE 0x40
92 /* 0x80 is reserved in rdma.h for RAM_SAVE_FLAG_HOOK */
93 #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100
94 #define RAM_SAVE_FLAG_MULTIFD_FLUSH 0x200
95 /* We can't use any flag that is bigger than 0x200 */
97 XBZRLECacheStats xbzrle_counters
;
99 /* used by the search for pages to send */
100 struct PageSearchStatus
{
101 /* The migration channel used for a specific host page */
102 QEMUFile
*pss_channel
;
103 /* Last block from where we have sent data */
104 RAMBlock
*last_sent_block
;
105 /* Current block being searched */
107 /* Current page to search from */
109 /* Set once we wrap around */
111 /* Whether we're sending a host page */
112 bool host_page_sending
;
113 /* The start/end of current host page. Invalid if host_page_sending==false */
114 unsigned long host_page_start
;
115 unsigned long host_page_end
;
117 typedef struct PageSearchStatus PageSearchStatus
;
119 /* struct contains XBZRLE cache and a static page
120 used by the compression */
122 /* buffer used for XBZRLE encoding */
123 uint8_t *encoded_buf
;
124 /* buffer for storing page content */
125 uint8_t *current_buf
;
126 /* Cache for XBZRLE, Protected by lock. */
129 /* it will store a page full of zeros */
130 uint8_t *zero_target_page
;
131 /* buffer used for XBZRLE decoding */
132 uint8_t *decoded_buf
;
135 static void XBZRLE_cache_lock(void)
137 if (migrate_xbzrle()) {
138 qemu_mutex_lock(&XBZRLE
.lock
);
142 static void XBZRLE_cache_unlock(void)
144 if (migrate_xbzrle()) {
145 qemu_mutex_unlock(&XBZRLE
.lock
);
150 * xbzrle_cache_resize: resize the xbzrle cache
152 * This function is called from migrate_params_apply in main
153 * thread, possibly while a migration is in progress. A running
154 * migration may be using the cache and might finish during this call,
155 * hence changes to the cache are protected by XBZRLE.lock().
157 * Returns 0 for success or -1 for error
159 * @new_size: new cache size
160 * @errp: set *errp if the check failed, with reason
162 int xbzrle_cache_resize(uint64_t new_size
, Error
**errp
)
164 PageCache
*new_cache
;
167 /* Check for truncation */
168 if (new_size
!= (size_t)new_size
) {
169 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
170 "exceeding address space");
174 if (new_size
== migrate_xbzrle_cache_size()) {
181 if (XBZRLE
.cache
!= NULL
) {
182 new_cache
= cache_init(new_size
, TARGET_PAGE_SIZE
, errp
);
188 cache_fini(XBZRLE
.cache
);
189 XBZRLE
.cache
= new_cache
;
192 XBZRLE_cache_unlock();
196 static bool postcopy_preempt_active(void)
198 return migrate_postcopy_preempt() && migration_in_postcopy();
201 bool migrate_ram_is_ignored(RAMBlock
*block
)
203 return !qemu_ram_is_migratable(block
) ||
204 (migrate_ignore_shared() && qemu_ram_is_shared(block
)
205 && qemu_ram_is_named_file(block
));
208 #undef RAMBLOCK_FOREACH
210 int foreach_not_ignored_block(RAMBlockIterFunc func
, void *opaque
)
215 RCU_READ_LOCK_GUARD();
217 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
218 ret
= func(block
, opaque
);
226 static void ramblock_recv_map_init(void)
230 RAMBLOCK_FOREACH_NOT_IGNORED(rb
) {
231 assert(!rb
->receivedmap
);
232 rb
->receivedmap
= bitmap_new(rb
->max_length
>> qemu_target_page_bits());
236 int ramblock_recv_bitmap_test(RAMBlock
*rb
, void *host_addr
)
238 return test_bit(ramblock_recv_bitmap_offset(host_addr
, rb
),
242 bool ramblock_recv_bitmap_test_byte_offset(RAMBlock
*rb
, uint64_t byte_offset
)
244 return test_bit(byte_offset
>> TARGET_PAGE_BITS
, rb
->receivedmap
);
247 void ramblock_recv_bitmap_set(RAMBlock
*rb
, void *host_addr
)
249 set_bit_atomic(ramblock_recv_bitmap_offset(host_addr
, rb
), rb
->receivedmap
);
252 void ramblock_recv_bitmap_set_range(RAMBlock
*rb
, void *host_addr
,
255 bitmap_set_atomic(rb
->receivedmap
,
256 ramblock_recv_bitmap_offset(host_addr
, rb
),
260 #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL)
263 * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes).
265 * Returns >0 if success with sent bytes, or <0 if error.
267 int64_t ramblock_recv_bitmap_send(QEMUFile
*file
,
268 const char *block_name
)
270 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
271 unsigned long *le_bitmap
, nbits
;
275 error_report("%s: invalid block name: %s", __func__
, block_name
);
279 nbits
= block
->postcopy_length
>> TARGET_PAGE_BITS
;
282 * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit
283 * machines we may need 4 more bytes for padding (see below
284 * comment). So extend it a bit before hand.
286 le_bitmap
= bitmap_new(nbits
+ BITS_PER_LONG
);
289 * Always use little endian when sending the bitmap. This is
290 * required that when source and destination VMs are not using the
291 * same endianness. (Note: big endian won't work.)
293 bitmap_to_le(le_bitmap
, block
->receivedmap
, nbits
);
295 /* Size of the bitmap, in bytes */
296 size
= DIV_ROUND_UP(nbits
, 8);
299 * size is always aligned to 8 bytes for 64bit machines, but it
300 * may not be true for 32bit machines. We need this padding to
301 * make sure the migration can survive even between 32bit and
304 size
= ROUND_UP(size
, 8);
306 qemu_put_be64(file
, size
);
307 qemu_put_buffer(file
, (const uint8_t *)le_bitmap
, size
);
309 * Mark as an end, in case the middle part is screwed up due to
310 * some "mysterious" reason.
312 qemu_put_be64(file
, RAMBLOCK_RECV_BITMAP_ENDING
);
317 if (qemu_file_get_error(file
)) {
318 return qemu_file_get_error(file
);
321 return size
+ sizeof(size
);
325 * An outstanding page request, on the source, having been received
328 struct RAMSrcPageRequest
{
333 QSIMPLEQ_ENTRY(RAMSrcPageRequest
) next_req
;
336 /* State of RAM for migration */
339 * PageSearchStatus structures for the channels when send pages.
340 * Protected by the bitmap_mutex.
342 PageSearchStatus pss
[RAM_CHANNEL_MAX
];
343 /* UFFD file descriptor, used in 'write-tracking' migration */
345 /* total ram size in bytes */
346 uint64_t ram_bytes_total
;
347 /* Last block that we have visited searching for dirty pages */
348 RAMBlock
*last_seen_block
;
349 /* Last dirty target page we have sent */
350 ram_addr_t last_page
;
351 /* last ram version we have seen */
352 uint32_t last_version
;
353 /* How many times we have dirty too many pages */
354 int dirty_rate_high_cnt
;
355 /* these variables are used for bitmap sync */
356 /* last time we did a full bitmap_sync */
357 int64_t time_last_bitmap_sync
;
358 /* bytes transferred at start_time */
359 uint64_t bytes_xfer_prev
;
360 /* number of dirty pages since start_time */
361 uint64_t num_dirty_pages_period
;
362 /* xbzrle misses since the beginning of the period */
363 uint64_t xbzrle_cache_miss_prev
;
364 /* Amount of xbzrle pages since the beginning of the period */
365 uint64_t xbzrle_pages_prev
;
366 /* Amount of xbzrle encoded bytes since the beginning of the period */
367 uint64_t xbzrle_bytes_prev
;
368 /* Are we really using XBZRLE (e.g., after the first round). */
370 /* Are we on the last stage of migration */
372 /* compression statistics since the beginning of the period */
373 /* amount of count that no free thread to compress data */
374 uint64_t compress_thread_busy_prev
;
375 /* amount bytes after compression */
376 uint64_t compressed_size_prev
;
377 /* amount of compressed pages */
378 uint64_t compress_pages_prev
;
380 /* total handled target pages at the beginning of period */
381 uint64_t target_page_count_prev
;
382 /* total handled target pages since start */
383 uint64_t target_page_count
;
384 /* number of dirty bits in the bitmap */
385 uint64_t migration_dirty_pages
;
388 * - dirty/clear bitmap
389 * - migration_dirty_pages
392 QemuMutex bitmap_mutex
;
393 /* The RAMBlock used in the last src_page_requests */
394 RAMBlock
*last_req_rb
;
395 /* Queue of outstanding page requests from the destination */
396 QemuMutex src_page_req_mutex
;
397 QSIMPLEQ_HEAD(, RAMSrcPageRequest
) src_page_requests
;
400 * This is only used when postcopy is in recovery phase, to communicate
401 * between the migration thread and the return path thread on dirty
402 * bitmap synchronizations. This field is unused in other stages of
405 unsigned int postcopy_bmap_sync_requested
;
407 typedef struct RAMState RAMState
;
409 static RAMState
*ram_state
;
411 static NotifierWithReturnList precopy_notifier_list
;
413 /* Whether postcopy has queued requests? */
414 static bool postcopy_has_request(RAMState
*rs
)
416 return !QSIMPLEQ_EMPTY_ATOMIC(&rs
->src_page_requests
);
419 void precopy_infrastructure_init(void)
421 notifier_with_return_list_init(&precopy_notifier_list
);
424 void precopy_add_notifier(NotifierWithReturn
*n
)
426 notifier_with_return_list_add(&precopy_notifier_list
, n
);
429 void precopy_remove_notifier(NotifierWithReturn
*n
)
431 notifier_with_return_remove(n
);
434 int precopy_notify(PrecopyNotifyReason reason
, Error
**errp
)
436 PrecopyNotifyData pnd
;
440 return notifier_with_return_list_notify(&precopy_notifier_list
, &pnd
);
443 uint64_t ram_bytes_remaining(void)
445 return ram_state
? (ram_state
->migration_dirty_pages
* TARGET_PAGE_SIZE
) :
449 void ram_transferred_add(uint64_t bytes
)
451 if (runstate_is_running()) {
452 stat64_add(&mig_stats
.precopy_bytes
, bytes
);
453 } else if (migration_in_postcopy()) {
454 stat64_add(&mig_stats
.postcopy_bytes
, bytes
);
456 stat64_add(&mig_stats
.downtime_bytes
, bytes
);
458 stat64_add(&mig_stats
.transferred
, bytes
);
461 struct MigrationOps
{
462 int (*ram_save_target_page
)(RAMState
*rs
, PageSearchStatus
*pss
);
464 typedef struct MigrationOps MigrationOps
;
466 MigrationOps
*migration_ops
;
468 static int ram_save_host_page_urgent(PageSearchStatus
*pss
);
470 /* NOTE: page is the PFN not real ram_addr_t. */
471 static void pss_init(PageSearchStatus
*pss
, RAMBlock
*rb
, ram_addr_t page
)
475 pss
->complete_round
= false;
479 * Check whether two PSSs are actively sending the same page. Return true
480 * if it is, false otherwise.
482 static bool pss_overlap(PageSearchStatus
*pss1
, PageSearchStatus
*pss2
)
484 return pss1
->host_page_sending
&& pss2
->host_page_sending
&&
485 (pss1
->host_page_start
== pss2
->host_page_start
);
489 * save_page_header: write page header to wire
491 * If this is the 1st block, it also writes the block identification
493 * Returns the number of bytes written
495 * @pss: current PSS channel status
496 * @block: block that contains the page we want to send
497 * @offset: offset inside the block for the page
498 * in the lower bits, it contains flags
500 static size_t save_page_header(PageSearchStatus
*pss
, QEMUFile
*f
,
501 RAMBlock
*block
, ram_addr_t offset
)
504 bool same_block
= (block
== pss
->last_sent_block
);
507 offset
|= RAM_SAVE_FLAG_CONTINUE
;
509 qemu_put_be64(f
, offset
);
513 len
= strlen(block
->idstr
);
514 qemu_put_byte(f
, len
);
515 qemu_put_buffer(f
, (uint8_t *)block
->idstr
, len
);
517 pss
->last_sent_block
= block
;
523 * mig_throttle_guest_down: throttle down the guest
525 * Reduce amount of guest cpu execution to hopefully slow down memory
526 * writes. If guest dirty memory rate is reduced below the rate at
527 * which we can transfer pages to the destination then we should be
528 * able to complete migration. Some workloads dirty memory way too
529 * fast and will not effectively converge, even with auto-converge.
531 static void mig_throttle_guest_down(uint64_t bytes_dirty_period
,
532 uint64_t bytes_dirty_threshold
)
534 uint64_t pct_initial
= migrate_cpu_throttle_initial();
535 uint64_t pct_increment
= migrate_cpu_throttle_increment();
536 bool pct_tailslow
= migrate_cpu_throttle_tailslow();
537 int pct_max
= migrate_max_cpu_throttle();
539 uint64_t throttle_now
= cpu_throttle_get_percentage();
540 uint64_t cpu_now
, cpu_ideal
, throttle_inc
;
542 /* We have not started throttling yet. Let's start it. */
543 if (!cpu_throttle_active()) {
544 cpu_throttle_set(pct_initial
);
546 /* Throttling already on, just increase the rate */
548 throttle_inc
= pct_increment
;
550 /* Compute the ideal CPU percentage used by Guest, which may
551 * make the dirty rate match the dirty rate threshold. */
552 cpu_now
= 100 - throttle_now
;
553 cpu_ideal
= cpu_now
* (bytes_dirty_threshold
* 1.0 /
555 throttle_inc
= MIN(cpu_now
- cpu_ideal
, pct_increment
);
557 cpu_throttle_set(MIN(throttle_now
+ throttle_inc
, pct_max
));
561 void mig_throttle_counter_reset(void)
563 RAMState
*rs
= ram_state
;
565 rs
->time_last_bitmap_sync
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
566 rs
->num_dirty_pages_period
= 0;
567 rs
->bytes_xfer_prev
= stat64_get(&mig_stats
.transferred
);
571 * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache
573 * @current_addr: address for the zero page
575 * Update the xbzrle cache to reflect a page that's been sent as all 0.
576 * The important thing is that a stale (not-yet-0'd) page be replaced
578 * As a bonus, if the page wasn't in the cache it gets added so that
579 * when a small write is made into the 0'd page it gets XBZRLE sent.
581 static void xbzrle_cache_zero_page(ram_addr_t current_addr
)
583 /* We don't care if this fails to allocate a new cache page
584 * as long as it updated an old one */
585 cache_insert(XBZRLE
.cache
, current_addr
, XBZRLE
.zero_target_page
,
586 stat64_get(&mig_stats
.dirty_sync_count
));
589 #define ENCODING_FLAG_XBZRLE 0x1
592 * save_xbzrle_page: compress and send current page
594 * Returns: 1 means that we wrote the page
595 * 0 means that page is identical to the one already sent
596 * -1 means that xbzrle would be longer than normal
598 * @rs: current RAM state
599 * @pss: current PSS channel
600 * @current_data: pointer to the address of the page contents
601 * @current_addr: addr of the page
602 * @block: block that contains the page we want to send
603 * @offset: offset inside the block for the page
605 static int save_xbzrle_page(RAMState
*rs
, PageSearchStatus
*pss
,
606 uint8_t **current_data
, ram_addr_t current_addr
,
607 RAMBlock
*block
, ram_addr_t offset
)
609 int encoded_len
= 0, bytes_xbzrle
;
610 uint8_t *prev_cached_page
;
611 QEMUFile
*file
= pss
->pss_channel
;
612 uint64_t generation
= stat64_get(&mig_stats
.dirty_sync_count
);
614 if (!cache_is_cached(XBZRLE
.cache
, current_addr
, generation
)) {
615 xbzrle_counters
.cache_miss
++;
616 if (!rs
->last_stage
) {
617 if (cache_insert(XBZRLE
.cache
, current_addr
, *current_data
,
621 /* update *current_data when the page has been
622 inserted into cache */
623 *current_data
= get_cached_data(XBZRLE
.cache
, current_addr
);
630 * Reaching here means the page has hit the xbzrle cache, no matter what
631 * encoding result it is (normal encoding, overflow or skipping the page),
632 * count the page as encoded. This is used to calculate the encoding rate.
634 * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB,
635 * 2nd page turns out to be skipped (i.e. no new bytes written to the
636 * page), the overall encoding rate will be 8KB / 2KB = 4, which has the
637 * skipped page included. In this way, the encoding rate can tell if the
638 * guest page is good for xbzrle encoding.
640 xbzrle_counters
.pages
++;
641 prev_cached_page
= get_cached_data(XBZRLE
.cache
, current_addr
);
643 /* save current buffer into memory */
644 memcpy(XBZRLE
.current_buf
, *current_data
, TARGET_PAGE_SIZE
);
646 /* XBZRLE encoding (if there is no overflow) */
647 encoded_len
= xbzrle_encode_buffer(prev_cached_page
, XBZRLE
.current_buf
,
648 TARGET_PAGE_SIZE
, XBZRLE
.encoded_buf
,
652 * Update the cache contents, so that it corresponds to the data
653 * sent, in all cases except where we skip the page.
655 if (!rs
->last_stage
&& encoded_len
!= 0) {
656 memcpy(prev_cached_page
, XBZRLE
.current_buf
, TARGET_PAGE_SIZE
);
658 * In the case where we couldn't compress, ensure that the caller
659 * sends the data from the cache, since the guest might have
660 * changed the RAM since we copied it.
662 *current_data
= prev_cached_page
;
665 if (encoded_len
== 0) {
666 trace_save_xbzrle_page_skipping();
668 } else if (encoded_len
== -1) {
669 trace_save_xbzrle_page_overflow();
670 xbzrle_counters
.overflow
++;
671 xbzrle_counters
.bytes
+= TARGET_PAGE_SIZE
;
675 /* Send XBZRLE based compressed page */
676 bytes_xbzrle
= save_page_header(pss
, pss
->pss_channel
, block
,
677 offset
| RAM_SAVE_FLAG_XBZRLE
);
678 qemu_put_byte(file
, ENCODING_FLAG_XBZRLE
);
679 qemu_put_be16(file
, encoded_len
);
680 qemu_put_buffer(file
, XBZRLE
.encoded_buf
, encoded_len
);
681 bytes_xbzrle
+= encoded_len
+ 1 + 2;
683 * Like compressed_size (please see update_compress_thread_counts),
684 * the xbzrle encoded bytes don't count the 8 byte header with
685 * RAM_SAVE_FLAG_CONTINUE.
687 xbzrle_counters
.bytes
+= bytes_xbzrle
- 8;
688 ram_transferred_add(bytes_xbzrle
);
694 * pss_find_next_dirty: find the next dirty page of current ramblock
696 * This function updates pss->page to point to the next dirty page index
697 * within the ramblock to migrate, or the end of ramblock when nothing
698 * found. Note that when pss->host_page_sending==true it means we're
699 * during sending a host page, so we won't look for dirty page that is
700 * outside the host page boundary.
702 * @pss: the current page search status
704 static void pss_find_next_dirty(PageSearchStatus
*pss
)
706 RAMBlock
*rb
= pss
->block
;
707 unsigned long size
= rb
->used_length
>> TARGET_PAGE_BITS
;
708 unsigned long *bitmap
= rb
->bmap
;
710 if (migrate_ram_is_ignored(rb
)) {
711 /* Points directly to the end, so we know no dirty page */
717 * If during sending a host page, only look for dirty pages within the
718 * current host page being send.
720 if (pss
->host_page_sending
) {
721 assert(pss
->host_page_end
);
722 size
= MIN(size
, pss
->host_page_end
);
725 pss
->page
= find_next_bit(bitmap
, size
, pss
->page
);
728 static void migration_clear_memory_region_dirty_bitmap(RAMBlock
*rb
,
734 if (!rb
->clear_bmap
|| !clear_bmap_test_and_clear(rb
, page
)) {
738 shift
= rb
->clear_bmap_shift
;
740 * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this
741 * can make things easier sometimes since then start address
742 * of the small chunk will always be 64 pages aligned so the
743 * bitmap will always be aligned to unsigned long. We should
744 * even be able to remove this restriction but I'm simply
749 size
= 1ULL << (TARGET_PAGE_BITS
+ shift
);
750 start
= QEMU_ALIGN_DOWN((ram_addr_t
)page
<< TARGET_PAGE_BITS
, size
);
751 trace_migration_bitmap_clear_dirty(rb
->idstr
, start
, size
, page
);
752 memory_region_clear_dirty_bitmap(rb
->mr
, start
, size
);
756 migration_clear_memory_region_dirty_bitmap_range(RAMBlock
*rb
,
758 unsigned long npages
)
760 unsigned long i
, chunk_pages
= 1UL << rb
->clear_bmap_shift
;
761 unsigned long chunk_start
= QEMU_ALIGN_DOWN(start
, chunk_pages
);
762 unsigned long chunk_end
= QEMU_ALIGN_UP(start
+ npages
, chunk_pages
);
765 * Clear pages from start to start + npages - 1, so the end boundary is
768 for (i
= chunk_start
; i
< chunk_end
; i
+= chunk_pages
) {
769 migration_clear_memory_region_dirty_bitmap(rb
, i
);
774 * colo_bitmap_find_diry:find contiguous dirty pages from start
776 * Returns the page offset within memory region of the start of the contiguout
779 * @rs: current RAM state
780 * @rb: RAMBlock where to search for dirty pages
781 * @start: page where we start the search
782 * @num: the number of contiguous dirty pages
785 unsigned long colo_bitmap_find_dirty(RAMState
*rs
, RAMBlock
*rb
,
786 unsigned long start
, unsigned long *num
)
788 unsigned long size
= rb
->used_length
>> TARGET_PAGE_BITS
;
789 unsigned long *bitmap
= rb
->bmap
;
790 unsigned long first
, next
;
794 if (migrate_ram_is_ignored(rb
)) {
798 first
= find_next_bit(bitmap
, size
, start
);
802 next
= find_next_zero_bit(bitmap
, size
, first
+ 1);
803 assert(next
>= first
);
808 static inline bool migration_bitmap_clear_dirty(RAMState
*rs
,
815 * Clear dirty bitmap if needed. This _must_ be called before we
816 * send any of the page in the chunk because we need to make sure
817 * we can capture further page content changes when we sync dirty
818 * log the next time. So as long as we are going to send any of
819 * the page in the chunk we clear the remote dirty bitmap for all.
820 * Clearing it earlier won't be a problem, but too late will.
822 migration_clear_memory_region_dirty_bitmap(rb
, page
);
824 ret
= test_and_clear_bit(page
, rb
->bmap
);
826 rs
->migration_dirty_pages
--;
832 static void dirty_bitmap_clear_section(MemoryRegionSection
*section
,
835 const hwaddr offset
= section
->offset_within_region
;
836 const hwaddr size
= int128_get64(section
->size
);
837 const unsigned long start
= offset
>> TARGET_PAGE_BITS
;
838 const unsigned long npages
= size
>> TARGET_PAGE_BITS
;
839 RAMBlock
*rb
= section
->mr
->ram_block
;
840 uint64_t *cleared_bits
= opaque
;
843 * We don't grab ram_state->bitmap_mutex because we expect to run
844 * only when starting migration or during postcopy recovery where
845 * we don't have concurrent access.
847 if (!migration_in_postcopy() && !migrate_background_snapshot()) {
848 migration_clear_memory_region_dirty_bitmap_range(rb
, start
, npages
);
850 *cleared_bits
+= bitmap_count_one_with_offset(rb
->bmap
, start
, npages
);
851 bitmap_clear(rb
->bmap
, start
, npages
);
855 * Exclude all dirty pages from migration that fall into a discarded range as
856 * managed by a RamDiscardManager responsible for the mapped memory region of
857 * the RAMBlock. Clear the corresponding bits in the dirty bitmaps.
859 * Discarded pages ("logically unplugged") have undefined content and must
860 * not get migrated, because even reading these pages for migration might
861 * result in undesired behavior.
863 * Returns the number of cleared bits in the RAMBlock dirty bitmap.
865 * Note: The result is only stable while migrating (precopy/postcopy).
867 static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock
*rb
)
869 uint64_t cleared_bits
= 0;
871 if (rb
->mr
&& rb
->bmap
&& memory_region_has_ram_discard_manager(rb
->mr
)) {
872 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(rb
->mr
);
873 MemoryRegionSection section
= {
875 .offset_within_region
= 0,
876 .size
= int128_make64(qemu_ram_get_used_length(rb
)),
879 ram_discard_manager_replay_discarded(rdm
, §ion
,
880 dirty_bitmap_clear_section
,
887 * Check if a host-page aligned page falls into a discarded range as managed by
888 * a RamDiscardManager responsible for the mapped memory region of the RAMBlock.
890 * Note: The result is only stable while migrating (precopy/postcopy).
892 bool ramblock_page_is_discarded(RAMBlock
*rb
, ram_addr_t start
)
894 if (rb
->mr
&& memory_region_has_ram_discard_manager(rb
->mr
)) {
895 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(rb
->mr
);
896 MemoryRegionSection section
= {
898 .offset_within_region
= start
,
899 .size
= int128_make64(qemu_ram_pagesize(rb
)),
902 return !ram_discard_manager_is_populated(rdm
, §ion
);
907 /* Called with RCU critical section */
908 static void ramblock_sync_dirty_bitmap(RAMState
*rs
, RAMBlock
*rb
)
910 uint64_t new_dirty_pages
=
911 cpu_physical_memory_sync_dirty_bitmap(rb
, 0, rb
->used_length
);
913 rs
->migration_dirty_pages
+= new_dirty_pages
;
914 rs
->num_dirty_pages_period
+= new_dirty_pages
;
918 * ram_pagesize_summary: calculate all the pagesizes of a VM
920 * Returns a summary bitmap of the page sizes of all RAMBlocks
922 * For VMs with just normal pages this is equivalent to the host page
923 * size. If it's got some huge pages then it's the OR of all the
924 * different page sizes.
926 uint64_t ram_pagesize_summary(void)
929 uint64_t summary
= 0;
931 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
932 summary
|= block
->page_size
;
938 uint64_t ram_get_total_transferred_pages(void)
940 return stat64_get(&mig_stats
.normal_pages
) +
941 stat64_get(&mig_stats
.zero_pages
) +
942 ram_compressed_pages() + xbzrle_counters
.pages
;
945 static void migration_update_rates(RAMState
*rs
, int64_t end_time
)
947 uint64_t page_count
= rs
->target_page_count
- rs
->target_page_count_prev
;
948 double compressed_size
;
950 /* calculate period counters */
951 stat64_set(&mig_stats
.dirty_pages_rate
,
952 rs
->num_dirty_pages_period
* 1000 /
953 (end_time
- rs
->time_last_bitmap_sync
));
959 if (migrate_xbzrle()) {
960 double encoded_size
, unencoded_size
;
962 xbzrle_counters
.cache_miss_rate
= (double)(xbzrle_counters
.cache_miss
-
963 rs
->xbzrle_cache_miss_prev
) / page_count
;
964 rs
->xbzrle_cache_miss_prev
= xbzrle_counters
.cache_miss
;
965 unencoded_size
= (xbzrle_counters
.pages
- rs
->xbzrle_pages_prev
) *
967 encoded_size
= xbzrle_counters
.bytes
- rs
->xbzrle_bytes_prev
;
968 if (xbzrle_counters
.pages
== rs
->xbzrle_pages_prev
|| !encoded_size
) {
969 xbzrle_counters
.encoding_rate
= 0;
971 xbzrle_counters
.encoding_rate
= unencoded_size
/ encoded_size
;
973 rs
->xbzrle_pages_prev
= xbzrle_counters
.pages
;
974 rs
->xbzrle_bytes_prev
= xbzrle_counters
.bytes
;
977 if (migrate_compress()) {
978 compression_counters
.busy_rate
= (double)(compression_counters
.busy
-
979 rs
->compress_thread_busy_prev
) / page_count
;
980 rs
->compress_thread_busy_prev
= compression_counters
.busy
;
982 compressed_size
= compression_counters
.compressed_size
-
983 rs
->compressed_size_prev
;
984 if (compressed_size
) {
985 double uncompressed_size
= (compression_counters
.pages
-
986 rs
->compress_pages_prev
) * TARGET_PAGE_SIZE
;
988 /* Compression-Ratio = Uncompressed-size / Compressed-size */
989 compression_counters
.compression_rate
=
990 uncompressed_size
/ compressed_size
;
992 rs
->compress_pages_prev
= compression_counters
.pages
;
993 rs
->compressed_size_prev
= compression_counters
.compressed_size
;
999 * Enable dirty-limit to throttle down the guest
1001 static void migration_dirty_limit_guest(void)
1004 * dirty page rate quota for all vCPUs fetched from
1005 * migration parameter 'vcpu_dirty_limit'
1007 static int64_t quota_dirtyrate
;
1008 MigrationState
*s
= migrate_get_current();
1011 * If dirty limit already enabled and migration parameter
1012 * vcpu-dirty-limit untouched.
1014 if (dirtylimit_in_service() &&
1015 quota_dirtyrate
== s
->parameters
.vcpu_dirty_limit
) {
1019 quota_dirtyrate
= s
->parameters
.vcpu_dirty_limit
;
1022 * Set all vCPU a quota dirtyrate, note that the second
1023 * parameter will be ignored if setting all vCPU for the vm
1025 qmp_set_vcpu_dirty_limit(false, -1, quota_dirtyrate
, NULL
);
1026 trace_migration_dirty_limit_guest(quota_dirtyrate
);
1029 static void migration_trigger_throttle(RAMState
*rs
)
1031 uint64_t threshold
= migrate_throttle_trigger_threshold();
1032 uint64_t bytes_xfer_period
=
1033 stat64_get(&mig_stats
.transferred
) - rs
->bytes_xfer_prev
;
1034 uint64_t bytes_dirty_period
= rs
->num_dirty_pages_period
* TARGET_PAGE_SIZE
;
1035 uint64_t bytes_dirty_threshold
= bytes_xfer_period
* threshold
/ 100;
1037 /* During block migration the auto-converge logic incorrectly detects
1038 * that ram migration makes no progress. Avoid this by disabling the
1039 * throttling logic during the bulk phase of block migration. */
1040 if (blk_mig_bulk_active()) {
1045 * The following detection logic can be refined later. For now:
1046 * Check to see if the ratio between dirtied bytes and the approx.
1047 * amount of bytes that just got transferred since the last time
1048 * we were in this routine reaches the threshold. If that happens
1049 * twice, start or increase throttling.
1051 if ((bytes_dirty_period
> bytes_dirty_threshold
) &&
1052 (++rs
->dirty_rate_high_cnt
>= 2)) {
1053 rs
->dirty_rate_high_cnt
= 0;
1054 if (migrate_auto_converge()) {
1055 trace_migration_throttle();
1056 mig_throttle_guest_down(bytes_dirty_period
,
1057 bytes_dirty_threshold
);
1058 } else if (migrate_dirty_limit()) {
1059 migration_dirty_limit_guest();
1064 static void migration_bitmap_sync(RAMState
*rs
, bool last_stage
)
1069 stat64_add(&mig_stats
.dirty_sync_count
, 1);
1071 if (!rs
->time_last_bitmap_sync
) {
1072 rs
->time_last_bitmap_sync
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1075 trace_migration_bitmap_sync_start();
1076 memory_global_dirty_log_sync(last_stage
);
1078 qemu_mutex_lock(&rs
->bitmap_mutex
);
1079 WITH_RCU_READ_LOCK_GUARD() {
1080 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
1081 ramblock_sync_dirty_bitmap(rs
, block
);
1083 stat64_set(&mig_stats
.dirty_bytes_last_sync
, ram_bytes_remaining());
1085 qemu_mutex_unlock(&rs
->bitmap_mutex
);
1087 memory_global_after_dirty_log_sync();
1088 trace_migration_bitmap_sync_end(rs
->num_dirty_pages_period
);
1090 end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1092 /* more than 1 second = 1000 millisecons */
1093 if (end_time
> rs
->time_last_bitmap_sync
+ 1000) {
1094 migration_trigger_throttle(rs
);
1096 migration_update_rates(rs
, end_time
);
1098 rs
->target_page_count_prev
= rs
->target_page_count
;
1100 /* reset period counters */
1101 rs
->time_last_bitmap_sync
= end_time
;
1102 rs
->num_dirty_pages_period
= 0;
1103 rs
->bytes_xfer_prev
= stat64_get(&mig_stats
.transferred
);
1105 if (migrate_events()) {
1106 uint64_t generation
= stat64_get(&mig_stats
.dirty_sync_count
);
1107 qapi_event_send_migration_pass(generation
);
1111 static void migration_bitmap_sync_precopy(RAMState
*rs
, bool last_stage
)
1113 Error
*local_err
= NULL
;
1116 * The current notifier usage is just an optimization to migration, so we
1117 * don't stop the normal migration process in the error case.
1119 if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC
, &local_err
)) {
1120 error_report_err(local_err
);
1124 migration_bitmap_sync(rs
, last_stage
);
1126 if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC
, &local_err
)) {
1127 error_report_err(local_err
);
1131 void ram_release_page(const char *rbname
, uint64_t offset
)
1133 if (!migrate_release_ram() || !migration_in_postcopy()) {
1137 ram_discard_range(rbname
, offset
, TARGET_PAGE_SIZE
);
1141 * save_zero_page: send the zero page to the stream
1143 * Returns the number of pages written.
1145 * @rs: current RAM state
1146 * @pss: current PSS channel
1147 * @offset: offset inside the block for the page
1149 static int save_zero_page(RAMState
*rs
, PageSearchStatus
*pss
,
1152 uint8_t *p
= pss
->block
->host
+ offset
;
1153 QEMUFile
*file
= pss
->pss_channel
;
1156 if (!buffer_is_zero(p
, TARGET_PAGE_SIZE
)) {
1160 len
+= save_page_header(pss
, file
, pss
->block
, offset
| RAM_SAVE_FLAG_ZERO
);
1161 qemu_put_byte(file
, 0);
1163 ram_release_page(pss
->block
->idstr
, offset
);
1165 stat64_add(&mig_stats
.zero_pages
, 1);
1166 ram_transferred_add(len
);
1169 * Must let xbzrle know, otherwise a previous (now 0'd) cached
1170 * page would be stale.
1172 if (rs
->xbzrle_started
) {
1173 XBZRLE_cache_lock();
1174 xbzrle_cache_zero_page(pss
->block
->offset
+ offset
);
1175 XBZRLE_cache_unlock();
1182 * @pages: the number of pages written by the control path,
1184 * > 0 - number of pages written
1186 * Return true if the pages has been saved, otherwise false is returned.
1188 static bool control_save_page(PageSearchStatus
*pss
,
1189 ram_addr_t offset
, int *pages
)
1193 ret
= rdma_control_save_page(pss
->pss_channel
, pss
->block
->offset
, offset
,
1195 if (ret
== RAM_SAVE_CONTROL_NOT_SUPP
) {
1199 if (ret
== RAM_SAVE_CONTROL_DELAYED
) {
1208 * directly send the page to the stream
1210 * Returns the number of pages written.
1212 * @pss: current PSS channel
1213 * @block: block that contains the page we want to send
1214 * @offset: offset inside the block for the page
1215 * @buf: the page to be sent
1216 * @async: send to page asyncly
1218 static int save_normal_page(PageSearchStatus
*pss
, RAMBlock
*block
,
1219 ram_addr_t offset
, uint8_t *buf
, bool async
)
1221 QEMUFile
*file
= pss
->pss_channel
;
1223 ram_transferred_add(save_page_header(pss
, pss
->pss_channel
, block
,
1224 offset
| RAM_SAVE_FLAG_PAGE
));
1226 qemu_put_buffer_async(file
, buf
, TARGET_PAGE_SIZE
,
1227 migrate_release_ram() &&
1228 migration_in_postcopy());
1230 qemu_put_buffer(file
, buf
, TARGET_PAGE_SIZE
);
1232 ram_transferred_add(TARGET_PAGE_SIZE
);
1233 stat64_add(&mig_stats
.normal_pages
, 1);
1238 * ram_save_page: send the given page to the stream
1240 * Returns the number of pages written.
1242 * >=0 - Number of pages written - this might legally be 0
1243 * if xbzrle noticed the page was the same.
1245 * @rs: current RAM state
1246 * @block: block that contains the page we want to send
1247 * @offset: offset inside the block for the page
1249 static int ram_save_page(RAMState
*rs
, PageSearchStatus
*pss
)
1253 bool send_async
= true;
1254 RAMBlock
*block
= pss
->block
;
1255 ram_addr_t offset
= ((ram_addr_t
)pss
->page
) << TARGET_PAGE_BITS
;
1256 ram_addr_t current_addr
= block
->offset
+ offset
;
1258 p
= block
->host
+ offset
;
1259 trace_ram_save_page(block
->idstr
, (uint64_t)offset
, p
);
1261 XBZRLE_cache_lock();
1262 if (rs
->xbzrle_started
&& !migration_in_postcopy()) {
1263 pages
= save_xbzrle_page(rs
, pss
, &p
, current_addr
,
1265 if (!rs
->last_stage
) {
1266 /* Can't send this cached data async, since the cache page
1267 * might get updated before it gets to the wire
1273 /* XBZRLE overflow or normal page */
1275 pages
= save_normal_page(pss
, block
, offset
, p
, send_async
);
1278 XBZRLE_cache_unlock();
1283 static int ram_save_multifd_page(QEMUFile
*file
, RAMBlock
*block
,
1286 if (multifd_queue_page(file
, block
, offset
) < 0) {
1289 stat64_add(&mig_stats
.normal_pages
, 1);
1294 static int send_queued_data(CompressParam
*param
)
1296 PageSearchStatus
*pss
= &ram_state
->pss
[RAM_CHANNEL_PRECOPY
];
1297 MigrationState
*ms
= migrate_get_current();
1298 QEMUFile
*file
= ms
->to_dst_file
;
1301 RAMBlock
*block
= param
->block
;
1302 ram_addr_t offset
= param
->offset
;
1304 if (param
->result
== RES_NONE
) {
1308 assert(block
== pss
->last_sent_block
);
1310 if (param
->result
== RES_ZEROPAGE
) {
1311 assert(qemu_file_buffer_empty(param
->file
));
1312 len
+= save_page_header(pss
, file
, block
, offset
| RAM_SAVE_FLAG_ZERO
);
1313 qemu_put_byte(file
, 0);
1315 ram_release_page(block
->idstr
, offset
);
1316 } else if (param
->result
== RES_COMPRESS
) {
1317 assert(!qemu_file_buffer_empty(param
->file
));
1318 len
+= save_page_header(pss
, file
, block
,
1319 offset
| RAM_SAVE_FLAG_COMPRESS_PAGE
);
1320 len
+= qemu_put_qemu_file(file
, param
->file
);
1325 update_compress_thread_counts(param
, len
);
1330 static void ram_flush_compressed_data(void)
1332 if (!migrate_compress()) {
1336 flush_compressed_data(send_queued_data
);
1339 #define PAGE_ALL_CLEAN 0
1340 #define PAGE_TRY_AGAIN 1
1341 #define PAGE_DIRTY_FOUND 2
1343 * find_dirty_block: find the next dirty page and update any state
1344 * associated with the search process.
1347 * <0: An error happened
1348 * PAGE_ALL_CLEAN: no dirty page found, give up
1349 * PAGE_TRY_AGAIN: no dirty page found, retry for next block
1350 * PAGE_DIRTY_FOUND: dirty page found
1352 * @rs: current RAM state
1353 * @pss: data about the state of the current dirty page scan
1354 * @again: set to false if the search has scanned the whole of RAM
1356 static int find_dirty_block(RAMState
*rs
, PageSearchStatus
*pss
)
1358 /* Update pss->page for the next dirty bit in ramblock */
1359 pss_find_next_dirty(pss
);
1361 if (pss
->complete_round
&& pss
->block
== rs
->last_seen_block
&&
1362 pss
->page
>= rs
->last_page
) {
1364 * We've been once around the RAM and haven't found anything.
1367 return PAGE_ALL_CLEAN
;
1369 if (!offset_in_ramblock(pss
->block
,
1370 ((ram_addr_t
)pss
->page
) << TARGET_PAGE_BITS
)) {
1371 /* Didn't find anything in this RAM Block */
1373 pss
->block
= QLIST_NEXT_RCU(pss
->block
, next
);
1375 if (migrate_multifd() &&
1376 !migrate_multifd_flush_after_each_section()) {
1377 QEMUFile
*f
= rs
->pss
[RAM_CHANNEL_PRECOPY
].pss_channel
;
1378 int ret
= multifd_send_sync_main(f
);
1382 qemu_put_be64(f
, RAM_SAVE_FLAG_MULTIFD_FLUSH
);
1386 * If memory migration starts over, we will meet a dirtied page
1387 * which may still exists in compression threads's ring, so we
1388 * should flush the compressed data to make sure the new page
1389 * is not overwritten by the old one in the destination.
1391 * Also If xbzrle is on, stop using the data compression at this
1392 * point. In theory, xbzrle can do better than compression.
1394 ram_flush_compressed_data();
1396 /* Hit the end of the list */
1397 pss
->block
= QLIST_FIRST_RCU(&ram_list
.blocks
);
1398 /* Flag that we've looped */
1399 pss
->complete_round
= true;
1400 /* After the first round, enable XBZRLE. */
1401 if (migrate_xbzrle()) {
1402 rs
->xbzrle_started
= true;
1405 /* Didn't find anything this time, but try again on the new block */
1406 return PAGE_TRY_AGAIN
;
1408 /* We've found something */
1409 return PAGE_DIRTY_FOUND
;
1414 * unqueue_page: gets a page of the queue
1416 * Helper for 'get_queued_page' - gets a page off the queue
1418 * Returns the block of the page (or NULL if none available)
1420 * @rs: current RAM state
1421 * @offset: used to return the offset within the RAMBlock
1423 static RAMBlock
*unqueue_page(RAMState
*rs
, ram_addr_t
*offset
)
1425 struct RAMSrcPageRequest
*entry
;
1426 RAMBlock
*block
= NULL
;
1428 if (!postcopy_has_request(rs
)) {
1432 QEMU_LOCK_GUARD(&rs
->src_page_req_mutex
);
1435 * This should _never_ change even after we take the lock, because no one
1436 * should be taking anything off the request list other than us.
1438 assert(postcopy_has_request(rs
));
1440 entry
= QSIMPLEQ_FIRST(&rs
->src_page_requests
);
1442 *offset
= entry
->offset
;
1444 if (entry
->len
> TARGET_PAGE_SIZE
) {
1445 entry
->len
-= TARGET_PAGE_SIZE
;
1446 entry
->offset
+= TARGET_PAGE_SIZE
;
1448 memory_region_unref(block
->mr
);
1449 QSIMPLEQ_REMOVE_HEAD(&rs
->src_page_requests
, next_req
);
1451 migration_consume_urgent_request();
1457 #if defined(__linux__)
1459 * poll_fault_page: try to get next UFFD write fault page and, if pending fault
1460 * is found, return RAM block pointer and page offset
1462 * Returns pointer to the RAMBlock containing faulting page,
1463 * NULL if no write faults are pending
1465 * @rs: current RAM state
1466 * @offset: page offset from the beginning of the block
1468 static RAMBlock
*poll_fault_page(RAMState
*rs
, ram_addr_t
*offset
)
1470 struct uffd_msg uffd_msg
;
1475 if (!migrate_background_snapshot()) {
1479 res
= uffd_read_events(rs
->uffdio_fd
, &uffd_msg
, 1);
1484 page_address
= (void *)(uintptr_t) uffd_msg
.arg
.pagefault
.address
;
1485 block
= qemu_ram_block_from_host(page_address
, false, offset
);
1486 assert(block
&& (block
->flags
& RAM_UF_WRITEPROTECT
) != 0);
1491 * ram_save_release_protection: release UFFD write protection after
1492 * a range of pages has been saved
1494 * @rs: current RAM state
1495 * @pss: page-search-status structure
1496 * @start_page: index of the first page in the range relative to pss->block
1498 * Returns 0 on success, negative value in case of an error
1500 static int ram_save_release_protection(RAMState
*rs
, PageSearchStatus
*pss
,
1501 unsigned long start_page
)
1505 /* Check if page is from UFFD-managed region. */
1506 if (pss
->block
->flags
& RAM_UF_WRITEPROTECT
) {
1507 void *page_address
= pss
->block
->host
+ (start_page
<< TARGET_PAGE_BITS
);
1508 uint64_t run_length
= (pss
->page
- start_page
) << TARGET_PAGE_BITS
;
1510 /* Flush async buffers before un-protect. */
1511 qemu_fflush(pss
->pss_channel
);
1512 /* Un-protect memory range. */
1513 res
= uffd_change_protection(rs
->uffdio_fd
, page_address
, run_length
,
1520 /* ram_write_tracking_available: check if kernel supports required UFFD features
1522 * Returns true if supports, false otherwise
1524 bool ram_write_tracking_available(void)
1526 uint64_t uffd_features
;
1529 res
= uffd_query_features(&uffd_features
);
1531 (uffd_features
& UFFD_FEATURE_PAGEFAULT_FLAG_WP
) != 0);
1534 /* ram_write_tracking_compatible: check if guest configuration is
1535 * compatible with 'write-tracking'
1537 * Returns true if compatible, false otherwise
1539 bool ram_write_tracking_compatible(void)
1541 const uint64_t uffd_ioctls_mask
= BIT(_UFFDIO_WRITEPROTECT
);
1546 /* Open UFFD file descriptor */
1547 uffd_fd
= uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP
, false);
1552 RCU_READ_LOCK_GUARD();
1554 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
1555 uint64_t uffd_ioctls
;
1557 /* Nothing to do with read-only and MMIO-writable regions */
1558 if (block
->mr
->readonly
|| block
->mr
->rom_device
) {
1561 /* Try to register block memory via UFFD-IO to track writes */
1562 if (uffd_register_memory(uffd_fd
, block
->host
, block
->max_length
,
1563 UFFDIO_REGISTER_MODE_WP
, &uffd_ioctls
)) {
1566 if ((uffd_ioctls
& uffd_ioctls_mask
) != uffd_ioctls_mask
) {
1573 uffd_close_fd(uffd_fd
);
1577 static inline void populate_read_range(RAMBlock
*block
, ram_addr_t offset
,
1580 const ram_addr_t end
= offset
+ size
;
1583 * We read one byte of each page; this will preallocate page tables if
1584 * required and populate the shared zeropage on MAP_PRIVATE anonymous memory
1585 * where no page was populated yet. This might require adaption when
1586 * supporting other mappings, like shmem.
1588 for (; offset
< end
; offset
+= block
->page_size
) {
1589 char tmp
= *((char *)block
->host
+ offset
);
1591 /* Don't optimize the read out */
1592 asm volatile("" : "+r" (tmp
));
1596 static inline int populate_read_section(MemoryRegionSection
*section
,
1599 const hwaddr size
= int128_get64(section
->size
);
1600 hwaddr offset
= section
->offset_within_region
;
1601 RAMBlock
*block
= section
->mr
->ram_block
;
1603 populate_read_range(block
, offset
, size
);
1608 * ram_block_populate_read: preallocate page tables and populate pages in the
1609 * RAM block by reading a byte of each page.
1611 * Since it's solely used for userfault_fd WP feature, here we just
1612 * hardcode page size to qemu_real_host_page_size.
1614 * @block: RAM block to populate
1616 static void ram_block_populate_read(RAMBlock
*rb
)
1619 * Skip populating all pages that fall into a discarded range as managed by
1620 * a RamDiscardManager responsible for the mapped memory region of the
1621 * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock
1622 * must not get populated automatically. We don't have to track
1623 * modifications via userfaultfd WP reliably, because these pages will
1624 * not be part of the migration stream either way -- see
1625 * ramblock_dirty_bitmap_exclude_discarded_pages().
1627 * Note: The result is only stable while migrating (precopy/postcopy).
1629 if (rb
->mr
&& memory_region_has_ram_discard_manager(rb
->mr
)) {
1630 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(rb
->mr
);
1631 MemoryRegionSection section
= {
1633 .offset_within_region
= 0,
1634 .size
= rb
->mr
->size
,
1637 ram_discard_manager_replay_populated(rdm
, §ion
,
1638 populate_read_section
, NULL
);
1640 populate_read_range(rb
, 0, rb
->used_length
);
1645 * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking
1647 void ram_write_tracking_prepare(void)
1651 RCU_READ_LOCK_GUARD();
1653 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
1654 /* Nothing to do with read-only and MMIO-writable regions */
1655 if (block
->mr
->readonly
|| block
->mr
->rom_device
) {
1660 * Populate pages of the RAM block before enabling userfault_fd
1663 * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with
1664 * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip
1665 * pages with pte_none() entries in page table.
1667 ram_block_populate_read(block
);
1671 static inline int uffd_protect_section(MemoryRegionSection
*section
,
1674 const hwaddr size
= int128_get64(section
->size
);
1675 const hwaddr offset
= section
->offset_within_region
;
1676 RAMBlock
*rb
= section
->mr
->ram_block
;
1677 int uffd_fd
= (uintptr_t)opaque
;
1679 return uffd_change_protection(uffd_fd
, rb
->host
+ offset
, size
, true,
1683 static int ram_block_uffd_protect(RAMBlock
*rb
, int uffd_fd
)
1685 assert(rb
->flags
& RAM_UF_WRITEPROTECT
);
1687 /* See ram_block_populate_read() */
1688 if (rb
->mr
&& memory_region_has_ram_discard_manager(rb
->mr
)) {
1689 RamDiscardManager
*rdm
= memory_region_get_ram_discard_manager(rb
->mr
);
1690 MemoryRegionSection section
= {
1692 .offset_within_region
= 0,
1693 .size
= rb
->mr
->size
,
1696 return ram_discard_manager_replay_populated(rdm
, §ion
,
1697 uffd_protect_section
,
1698 (void *)(uintptr_t)uffd_fd
);
1700 return uffd_change_protection(uffd_fd
, rb
->host
,
1701 rb
->used_length
, true, false);
1705 * ram_write_tracking_start: start UFFD-WP memory tracking
1707 * Returns 0 for success or negative value in case of error
1709 int ram_write_tracking_start(void)
1712 RAMState
*rs
= ram_state
;
1715 /* Open UFFD file descriptor */
1716 uffd_fd
= uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP
, true);
1720 rs
->uffdio_fd
= uffd_fd
;
1722 RCU_READ_LOCK_GUARD();
1724 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
1725 /* Nothing to do with read-only and MMIO-writable regions */
1726 if (block
->mr
->readonly
|| block
->mr
->rom_device
) {
1730 /* Register block memory with UFFD to track writes */
1731 if (uffd_register_memory(rs
->uffdio_fd
, block
->host
,
1732 block
->max_length
, UFFDIO_REGISTER_MODE_WP
, NULL
)) {
1735 block
->flags
|= RAM_UF_WRITEPROTECT
;
1736 memory_region_ref(block
->mr
);
1738 /* Apply UFFD write protection to the block memory range */
1739 if (ram_block_uffd_protect(block
, uffd_fd
)) {
1743 trace_ram_write_tracking_ramblock_start(block
->idstr
, block
->page_size
,
1744 block
->host
, block
->max_length
);
1750 error_report("ram_write_tracking_start() failed: restoring initial memory state");
1752 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
1753 if ((block
->flags
& RAM_UF_WRITEPROTECT
) == 0) {
1756 uffd_unregister_memory(rs
->uffdio_fd
, block
->host
, block
->max_length
);
1757 /* Cleanup flags and remove reference */
1758 block
->flags
&= ~RAM_UF_WRITEPROTECT
;
1759 memory_region_unref(block
->mr
);
1762 uffd_close_fd(uffd_fd
);
1768 * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection
1770 void ram_write_tracking_stop(void)
1772 RAMState
*rs
= ram_state
;
1775 RCU_READ_LOCK_GUARD();
1777 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
1778 if ((block
->flags
& RAM_UF_WRITEPROTECT
) == 0) {
1781 uffd_unregister_memory(rs
->uffdio_fd
, block
->host
, block
->max_length
);
1783 trace_ram_write_tracking_ramblock_stop(block
->idstr
, block
->page_size
,
1784 block
->host
, block
->max_length
);
1786 /* Cleanup flags and remove reference */
1787 block
->flags
&= ~RAM_UF_WRITEPROTECT
;
1788 memory_region_unref(block
->mr
);
1791 /* Finally close UFFD file descriptor */
1792 uffd_close_fd(rs
->uffdio_fd
);
1797 /* No target OS support, stubs just fail or ignore */
1799 static RAMBlock
*poll_fault_page(RAMState
*rs
, ram_addr_t
*offset
)
1807 static int ram_save_release_protection(RAMState
*rs
, PageSearchStatus
*pss
,
1808 unsigned long start_page
)
1817 bool ram_write_tracking_available(void)
1822 bool ram_write_tracking_compatible(void)
1828 int ram_write_tracking_start(void)
1834 void ram_write_tracking_stop(void)
1838 #endif /* defined(__linux__) */
1841 * get_queued_page: unqueue a page from the postcopy requests
1843 * Skips pages that are already sent (!dirty)
1845 * Returns true if a queued page is found
1847 * @rs: current RAM state
1848 * @pss: data about the state of the current dirty page scan
1850 static bool get_queued_page(RAMState
*rs
, PageSearchStatus
*pss
)
1857 block
= unqueue_page(rs
, &offset
);
1859 * We're sending this page, and since it's postcopy nothing else
1860 * will dirty it, and we must make sure it doesn't get sent again
1861 * even if this queue request was received after the background
1862 * search already sent it.
1867 page
= offset
>> TARGET_PAGE_BITS
;
1868 dirty
= test_bit(page
, block
->bmap
);
1870 trace_get_queued_page_not_dirty(block
->idstr
, (uint64_t)offset
,
1873 trace_get_queued_page(block
->idstr
, (uint64_t)offset
, page
);
1877 } while (block
&& !dirty
);
1881 * Poll write faults too if background snapshot is enabled; that's
1882 * when we have vcpus got blocked by the write protected pages.
1884 block
= poll_fault_page(rs
, &offset
);
1889 * We want the background search to continue from the queued page
1890 * since the guest is likely to want other pages near to the page
1891 * it just requested.
1894 pss
->page
= offset
>> TARGET_PAGE_BITS
;
1897 * This unqueued page would break the "one round" check, even is
1900 pss
->complete_round
= false;
1907 * migration_page_queue_free: drop any remaining pages in the ram
1910 * It should be empty at the end anyway, but in error cases there may
1911 * be some left. in case that there is any page left, we drop it.
1914 static void migration_page_queue_free(RAMState
*rs
)
1916 struct RAMSrcPageRequest
*mspr
, *next_mspr
;
1917 /* This queue generally should be empty - but in the case of a failed
1918 * migration might have some droppings in.
1920 RCU_READ_LOCK_GUARD();
1921 QSIMPLEQ_FOREACH_SAFE(mspr
, &rs
->src_page_requests
, next_req
, next_mspr
) {
1922 memory_region_unref(mspr
->rb
->mr
);
1923 QSIMPLEQ_REMOVE_HEAD(&rs
->src_page_requests
, next_req
);
1929 * ram_save_queue_pages: queue the page for transmission
1931 * A request from postcopy destination for example.
1933 * Returns zero on success or negative on error
1935 * @rbname: Name of the RAMBLock of the request. NULL means the
1936 * same that last one.
1937 * @start: starting address from the start of the RAMBlock
1938 * @len: length (in bytes) to send
1940 int ram_save_queue_pages(const char *rbname
, ram_addr_t start
, ram_addr_t len
)
1943 RAMState
*rs
= ram_state
;
1945 stat64_add(&mig_stats
.postcopy_requests
, 1);
1946 RCU_READ_LOCK_GUARD();
1949 /* Reuse last RAMBlock */
1950 ramblock
= rs
->last_req_rb
;
1954 * Shouldn't happen, we can't reuse the last RAMBlock if
1955 * it's the 1st request.
1957 error_report("ram_save_queue_pages no previous block");
1961 ramblock
= qemu_ram_block_by_name(rbname
);
1964 /* We shouldn't be asked for a non-existent RAMBlock */
1965 error_report("ram_save_queue_pages no block '%s'", rbname
);
1968 rs
->last_req_rb
= ramblock
;
1970 trace_ram_save_queue_pages(ramblock
->idstr
, start
, len
);
1971 if (!offset_in_ramblock(ramblock
, start
+ len
- 1)) {
1972 error_report("%s request overrun start=" RAM_ADDR_FMT
" len="
1973 RAM_ADDR_FMT
" blocklen=" RAM_ADDR_FMT
,
1974 __func__
, start
, len
, ramblock
->used_length
);
1979 * When with postcopy preempt, we send back the page directly in the
1982 if (postcopy_preempt_active()) {
1983 ram_addr_t page_start
= start
>> TARGET_PAGE_BITS
;
1984 size_t page_size
= qemu_ram_pagesize(ramblock
);
1985 PageSearchStatus
*pss
= &ram_state
->pss
[RAM_CHANNEL_POSTCOPY
];
1988 qemu_mutex_lock(&rs
->bitmap_mutex
);
1990 pss_init(pss
, ramblock
, page_start
);
1992 * Always use the preempt channel, and make sure it's there. It's
1993 * safe to access without lock, because when rp-thread is running
1994 * we should be the only one who operates on the qemufile
1996 pss
->pss_channel
= migrate_get_current()->postcopy_qemufile_src
;
1997 assert(pss
->pss_channel
);
2000 * It must be either one or multiple of host page size. Just
2001 * assert; if something wrong we're mostly split brain anyway.
2003 assert(len
% page_size
== 0);
2005 if (ram_save_host_page_urgent(pss
)) {
2006 error_report("%s: ram_save_host_page_urgent() failed: "
2007 "ramblock=%s, start_addr=0x"RAM_ADDR_FMT
,
2008 __func__
, ramblock
->idstr
, start
);
2013 * NOTE: after ram_save_host_page_urgent() succeeded, pss->page
2014 * will automatically be moved and point to the next host page
2015 * we're going to send, so no need to update here.
2017 * Normally QEMU never sends >1 host page in requests, so
2018 * logically we don't even need that as the loop should only
2019 * run once, but just to be consistent.
2023 qemu_mutex_unlock(&rs
->bitmap_mutex
);
2028 struct RAMSrcPageRequest
*new_entry
=
2029 g_new0(struct RAMSrcPageRequest
, 1);
2030 new_entry
->rb
= ramblock
;
2031 new_entry
->offset
= start
;
2032 new_entry
->len
= len
;
2034 memory_region_ref(ramblock
->mr
);
2035 qemu_mutex_lock(&rs
->src_page_req_mutex
);
2036 QSIMPLEQ_INSERT_TAIL(&rs
->src_page_requests
, new_entry
, next_req
);
2037 migration_make_urgent_request();
2038 qemu_mutex_unlock(&rs
->src_page_req_mutex
);
2044 * try to compress the page before posting it out, return true if the page
2045 * has been properly handled by compression, otherwise needs other
2046 * paths to handle it
2048 static bool save_compress_page(RAMState
*rs
, PageSearchStatus
*pss
,
2051 if (!migrate_compress()) {
2056 * When starting the process of a new block, the first page of
2057 * the block should be sent out before other pages in the same
2058 * block, and all the pages in last block should have been sent
2059 * out, keeping this order is important, because the 'cont' flag
2060 * is used to avoid resending the block name.
2062 * We post the fist page as normal page as compression will take
2063 * much CPU resource.
2065 if (pss
->block
!= pss
->last_sent_block
) {
2066 ram_flush_compressed_data();
2070 return compress_page_with_multi_thread(pss
->block
, offset
,
2075 * ram_save_target_page_legacy: save one target page
2077 * Returns the number of pages written
2079 * @rs: current RAM state
2080 * @pss: data about the page we want to send
2082 static int ram_save_target_page_legacy(RAMState
*rs
, PageSearchStatus
*pss
)
2084 RAMBlock
*block
= pss
->block
;
2085 ram_addr_t offset
= ((ram_addr_t
)pss
->page
) << TARGET_PAGE_BITS
;
2088 if (control_save_page(pss
, offset
, &res
)) {
2092 if (save_compress_page(rs
, pss
, offset
)) {
2096 if (save_zero_page(rs
, pss
, offset
)) {
2101 * Do not use multifd in postcopy as one whole host page should be
2102 * placed. Meanwhile postcopy requires atomic update of pages, so even
2103 * if host page size == guest page size the dest guest during run may
2104 * still see partially copied pages which is data corruption.
2106 if (migrate_multifd() && !migration_in_postcopy()) {
2107 return ram_save_multifd_page(pss
->pss_channel
, block
, offset
);
2110 return ram_save_page(rs
, pss
);
2113 /* Should be called before sending a host page */
2114 static void pss_host_page_prepare(PageSearchStatus
*pss
)
2116 /* How many guest pages are there in one host page? */
2117 size_t guest_pfns
= qemu_ram_pagesize(pss
->block
) >> TARGET_PAGE_BITS
;
2119 pss
->host_page_sending
= true;
2120 if (guest_pfns
<= 1) {
2122 * This covers both when guest psize == host psize, or when guest
2123 * has larger psize than the host (guest_pfns==0).
2125 * For the latter, we always send one whole guest page per
2126 * iteration of the host page (example: an Alpha VM on x86 host
2127 * will have guest psize 8K while host psize 4K).
2129 pss
->host_page_start
= pss
->page
;
2130 pss
->host_page_end
= pss
->page
+ 1;
2133 * The host page spans over multiple guest pages, we send them
2134 * within the same host page iteration.
2136 pss
->host_page_start
= ROUND_DOWN(pss
->page
, guest_pfns
);
2137 pss
->host_page_end
= ROUND_UP(pss
->page
+ 1, guest_pfns
);
2142 * Whether the page pointed by PSS is within the host page being sent.
2143 * Must be called after a previous pss_host_page_prepare().
2145 static bool pss_within_range(PageSearchStatus
*pss
)
2147 ram_addr_t ram_addr
;
2149 assert(pss
->host_page_sending
);
2151 /* Over host-page boundary? */
2152 if (pss
->page
>= pss
->host_page_end
) {
2156 ram_addr
= ((ram_addr_t
)pss
->page
) << TARGET_PAGE_BITS
;
2158 return offset_in_ramblock(pss
->block
, ram_addr
);
2161 static void pss_host_page_finish(PageSearchStatus
*pss
)
2163 pss
->host_page_sending
= false;
2164 /* This is not needed, but just to reset it */
2165 pss
->host_page_start
= pss
->host_page_end
= 0;
2169 * Send an urgent host page specified by `pss'. Need to be called with
2170 * bitmap_mutex held.
2172 * Returns 0 if save host page succeeded, false otherwise.
2174 static int ram_save_host_page_urgent(PageSearchStatus
*pss
)
2176 bool page_dirty
, sent
= false;
2177 RAMState
*rs
= ram_state
;
2180 trace_postcopy_preempt_send_host_page(pss
->block
->idstr
, pss
->page
);
2181 pss_host_page_prepare(pss
);
2184 * If precopy is sending the same page, let it be done in precopy, or
2185 * we could send the same page in two channels and none of them will
2186 * receive the whole page.
2188 if (pss_overlap(pss
, &ram_state
->pss
[RAM_CHANNEL_PRECOPY
])) {
2189 trace_postcopy_preempt_hit(pss
->block
->idstr
,
2190 pss
->page
<< TARGET_PAGE_BITS
);
2195 page_dirty
= migration_bitmap_clear_dirty(rs
, pss
->block
, pss
->page
);
2198 /* Be strict to return code; it must be 1, or what else? */
2199 if (migration_ops
->ram_save_target_page(rs
, pss
) != 1) {
2200 error_report_once("%s: ram_save_target_page failed", __func__
);
2206 pss_find_next_dirty(pss
);
2207 } while (pss_within_range(pss
));
2209 pss_host_page_finish(pss
);
2210 /* For urgent requests, flush immediately if sent */
2212 qemu_fflush(pss
->pss_channel
);
2218 * ram_save_host_page: save a whole host page
2220 * Starting at *offset send pages up to the end of the current host
2221 * page. It's valid for the initial offset to point into the middle of
2222 * a host page in which case the remainder of the hostpage is sent.
2223 * Only dirty target pages are sent. Note that the host page size may
2224 * be a huge page for this block.
2226 * The saving stops at the boundary of the used_length of the block
2227 * if the RAMBlock isn't a multiple of the host page size.
2229 * The caller must be with ram_state.bitmap_mutex held to call this
2230 * function. Note that this function can temporarily release the lock, but
2231 * when the function is returned it'll make sure the lock is still held.
2233 * Returns the number of pages written or negative on error
2235 * @rs: current RAM state
2236 * @pss: data about the page we want to send
2238 static int ram_save_host_page(RAMState
*rs
, PageSearchStatus
*pss
)
2240 bool page_dirty
, preempt_active
= postcopy_preempt_active();
2241 int tmppages
, pages
= 0;
2242 size_t pagesize_bits
=
2243 qemu_ram_pagesize(pss
->block
) >> TARGET_PAGE_BITS
;
2244 unsigned long start_page
= pss
->page
;
2247 if (migrate_ram_is_ignored(pss
->block
)) {
2248 error_report("block %s should not be migrated !", pss
->block
->idstr
);
2252 /* Update host page boundary information */
2253 pss_host_page_prepare(pss
);
2256 page_dirty
= migration_bitmap_clear_dirty(rs
, pss
->block
, pss
->page
);
2258 /* Check the pages is dirty and if it is send it */
2261 * Properly yield the lock only in postcopy preempt mode
2262 * because both migration thread and rp-return thread can
2263 * operate on the bitmaps.
2265 if (preempt_active
) {
2266 qemu_mutex_unlock(&rs
->bitmap_mutex
);
2268 tmppages
= migration_ops
->ram_save_target_page(rs
, pss
);
2269 if (tmppages
>= 0) {
2272 * Allow rate limiting to happen in the middle of huge pages if
2273 * something is sent in the current iteration.
2275 if (pagesize_bits
> 1 && tmppages
> 0) {
2276 migration_rate_limit();
2279 if (preempt_active
) {
2280 qemu_mutex_lock(&rs
->bitmap_mutex
);
2287 pss_host_page_finish(pss
);
2291 pss_find_next_dirty(pss
);
2292 } while (pss_within_range(pss
));
2294 pss_host_page_finish(pss
);
2296 res
= ram_save_release_protection(rs
, pss
, start_page
);
2297 return (res
< 0 ? res
: pages
);
2301 * ram_find_and_save_block: finds a dirty page and sends it to f
2303 * Called within an RCU critical section.
2305 * Returns the number of pages written where zero means no dirty pages,
2306 * or negative on error
2308 * @rs: current RAM state
2310 * On systems where host-page-size > target-page-size it will send all the
2311 * pages in a host page that are dirty.
2313 static int ram_find_and_save_block(RAMState
*rs
)
2315 PageSearchStatus
*pss
= &rs
->pss
[RAM_CHANNEL_PRECOPY
];
2318 /* No dirty page as there is zero RAM */
2319 if (!rs
->ram_bytes_total
) {
2324 * Always keep last_seen_block/last_page valid during this procedure,
2325 * because find_dirty_block() relies on these values (e.g., we compare
2326 * last_seen_block with pss.block to see whether we searched all the
2327 * ramblocks) to detect the completion of migration. Having NULL value
2328 * of last_seen_block can conditionally cause below loop to run forever.
2330 if (!rs
->last_seen_block
) {
2331 rs
->last_seen_block
= QLIST_FIRST_RCU(&ram_list
.blocks
);
2335 pss_init(pss
, rs
->last_seen_block
, rs
->last_page
);
2338 if (!get_queued_page(rs
, pss
)) {
2339 /* priority queue empty, so just search for something dirty */
2340 int res
= find_dirty_block(rs
, pss
);
2341 if (res
!= PAGE_DIRTY_FOUND
) {
2342 if (res
== PAGE_ALL_CLEAN
) {
2344 } else if (res
== PAGE_TRY_AGAIN
) {
2346 } else if (res
< 0) {
2352 pages
= ram_save_host_page(rs
, pss
);
2358 rs
->last_seen_block
= pss
->block
;
2359 rs
->last_page
= pss
->page
;
2364 static uint64_t ram_bytes_total_with_ignored(void)
2369 RCU_READ_LOCK_GUARD();
2371 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2372 total
+= block
->used_length
;
2377 uint64_t ram_bytes_total(void)
2382 RCU_READ_LOCK_GUARD();
2384 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
2385 total
+= block
->used_length
;
2390 static void xbzrle_load_setup(void)
2392 XBZRLE
.decoded_buf
= g_malloc(TARGET_PAGE_SIZE
);
2395 static void xbzrle_load_cleanup(void)
2397 g_free(XBZRLE
.decoded_buf
);
2398 XBZRLE
.decoded_buf
= NULL
;
2401 static void ram_state_cleanup(RAMState
**rsp
)
2404 migration_page_queue_free(*rsp
);
2405 qemu_mutex_destroy(&(*rsp
)->bitmap_mutex
);
2406 qemu_mutex_destroy(&(*rsp
)->src_page_req_mutex
);
2412 static void xbzrle_cleanup(void)
2414 XBZRLE_cache_lock();
2416 cache_fini(XBZRLE
.cache
);
2417 g_free(XBZRLE
.encoded_buf
);
2418 g_free(XBZRLE
.current_buf
);
2419 g_free(XBZRLE
.zero_target_page
);
2420 XBZRLE
.cache
= NULL
;
2421 XBZRLE
.encoded_buf
= NULL
;
2422 XBZRLE
.current_buf
= NULL
;
2423 XBZRLE
.zero_target_page
= NULL
;
2425 XBZRLE_cache_unlock();
2428 static void ram_save_cleanup(void *opaque
)
2430 RAMState
**rsp
= opaque
;
2433 /* We don't use dirty log with background snapshots */
2434 if (!migrate_background_snapshot()) {
2435 /* caller have hold iothread lock or is in a bh, so there is
2436 * no writing race against the migration bitmap
2438 if (global_dirty_tracking
& GLOBAL_DIRTY_MIGRATION
) {
2440 * do not stop dirty log without starting it, since
2441 * memory_global_dirty_log_stop will assert that
2442 * memory_global_dirty_log_start/stop used in pairs
2444 memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION
);
2448 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
2449 g_free(block
->clear_bmap
);
2450 block
->clear_bmap
= NULL
;
2451 g_free(block
->bmap
);
2456 compress_threads_save_cleanup();
2457 ram_state_cleanup(rsp
);
2458 g_free(migration_ops
);
2459 migration_ops
= NULL
;
2462 static void ram_state_reset(RAMState
*rs
)
2466 for (i
= 0; i
< RAM_CHANNEL_MAX
; i
++) {
2467 rs
->pss
[i
].last_sent_block
= NULL
;
2470 rs
->last_seen_block
= NULL
;
2472 rs
->last_version
= ram_list
.version
;
2473 rs
->xbzrle_started
= false;
2476 #define MAX_WAIT 50 /* ms, half buffered_file limit */
2478 /* **** functions for postcopy ***** */
2480 void ram_postcopy_migrated_memory_release(MigrationState
*ms
)
2482 struct RAMBlock
*block
;
2484 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
2485 unsigned long *bitmap
= block
->bmap
;
2486 unsigned long range
= block
->used_length
>> TARGET_PAGE_BITS
;
2487 unsigned long run_start
= find_next_zero_bit(bitmap
, range
, 0);
2489 while (run_start
< range
) {
2490 unsigned long run_end
= find_next_bit(bitmap
, range
, run_start
+ 1);
2491 ram_discard_range(block
->idstr
,
2492 ((ram_addr_t
)run_start
) << TARGET_PAGE_BITS
,
2493 ((ram_addr_t
)(run_end
- run_start
))
2494 << TARGET_PAGE_BITS
);
2495 run_start
= find_next_zero_bit(bitmap
, range
, run_end
+ 1);
2501 * postcopy_send_discard_bm_ram: discard a RAMBlock
2503 * Callback from postcopy_each_ram_send_discard for each RAMBlock
2505 * @ms: current migration state
2506 * @block: RAMBlock to discard
2508 static void postcopy_send_discard_bm_ram(MigrationState
*ms
, RAMBlock
*block
)
2510 unsigned long end
= block
->used_length
>> TARGET_PAGE_BITS
;
2511 unsigned long current
;
2512 unsigned long *bitmap
= block
->bmap
;
2514 for (current
= 0; current
< end
; ) {
2515 unsigned long one
= find_next_bit(bitmap
, end
, current
);
2516 unsigned long zero
, discard_length
;
2522 zero
= find_next_zero_bit(bitmap
, end
, one
+ 1);
2525 discard_length
= end
- one
;
2527 discard_length
= zero
- one
;
2529 postcopy_discard_send_range(ms
, one
, discard_length
);
2530 current
= one
+ discard_length
;
2534 static void postcopy_chunk_hostpages_pass(MigrationState
*ms
, RAMBlock
*block
);
2537 * postcopy_each_ram_send_discard: discard all RAMBlocks
2539 * Utility for the outgoing postcopy code.
2540 * Calls postcopy_send_discard_bm_ram for each RAMBlock
2541 * passing it bitmap indexes and name.
2542 * (qemu_ram_foreach_block ends up passing unscaled lengths
2543 * which would mean postcopy code would have to deal with target page)
2545 * @ms: current migration state
2547 static void postcopy_each_ram_send_discard(MigrationState
*ms
)
2549 struct RAMBlock
*block
;
2551 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
2552 postcopy_discard_send_init(ms
, block
->idstr
);
2555 * Deal with TPS != HPS and huge pages. It discard any partially sent
2556 * host-page size chunks, mark any partially dirty host-page size
2557 * chunks as all dirty. In this case the host-page is the host-page
2558 * for the particular RAMBlock, i.e. it might be a huge page.
2560 postcopy_chunk_hostpages_pass(ms
, block
);
2563 * Postcopy sends chunks of bitmap over the wire, but it
2564 * just needs indexes at this point, avoids it having
2565 * target page specific code.
2567 postcopy_send_discard_bm_ram(ms
, block
);
2568 postcopy_discard_send_finish(ms
);
2573 * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages
2575 * Helper for postcopy_chunk_hostpages; it's called twice to
2576 * canonicalize the two bitmaps, that are similar, but one is
2579 * Postcopy requires that all target pages in a hostpage are dirty or
2580 * clean, not a mix. This function canonicalizes the bitmaps.
2582 * @ms: current migration state
2583 * @block: block that contains the page we want to canonicalize
2585 static void postcopy_chunk_hostpages_pass(MigrationState
*ms
, RAMBlock
*block
)
2587 RAMState
*rs
= ram_state
;
2588 unsigned long *bitmap
= block
->bmap
;
2589 unsigned int host_ratio
= block
->page_size
/ TARGET_PAGE_SIZE
;
2590 unsigned long pages
= block
->used_length
>> TARGET_PAGE_BITS
;
2591 unsigned long run_start
;
2593 if (block
->page_size
== TARGET_PAGE_SIZE
) {
2594 /* Easy case - TPS==HPS for a non-huge page RAMBlock */
2598 /* Find a dirty page */
2599 run_start
= find_next_bit(bitmap
, pages
, 0);
2601 while (run_start
< pages
) {
2604 * If the start of this run of pages is in the middle of a host
2605 * page, then we need to fixup this host page.
2607 if (QEMU_IS_ALIGNED(run_start
, host_ratio
)) {
2608 /* Find the end of this run */
2609 run_start
= find_next_zero_bit(bitmap
, pages
, run_start
+ 1);
2611 * If the end isn't at the start of a host page, then the
2612 * run doesn't finish at the end of a host page
2613 * and we need to discard.
2617 if (!QEMU_IS_ALIGNED(run_start
, host_ratio
)) {
2619 unsigned long fixup_start_addr
= QEMU_ALIGN_DOWN(run_start
,
2621 run_start
= QEMU_ALIGN_UP(run_start
, host_ratio
);
2623 /* Clean up the bitmap */
2624 for (page
= fixup_start_addr
;
2625 page
< fixup_start_addr
+ host_ratio
; page
++) {
2627 * Remark them as dirty, updating the count for any pages
2628 * that weren't previously dirty.
2630 rs
->migration_dirty_pages
+= !test_and_set_bit(page
, bitmap
);
2634 /* Find the next dirty page for the next iteration */
2635 run_start
= find_next_bit(bitmap
, pages
, run_start
);
2640 * ram_postcopy_send_discard_bitmap: transmit the discard bitmap
2642 * Transmit the set of pages to be discarded after precopy to the target
2643 * these are pages that:
2644 * a) Have been previously transmitted but are now dirty again
2645 * b) Pages that have never been transmitted, this ensures that
2646 * any pages on the destination that have been mapped by background
2647 * tasks get discarded (transparent huge pages is the specific concern)
2648 * Hopefully this is pretty sparse
2650 * @ms: current migration state
2652 void ram_postcopy_send_discard_bitmap(MigrationState
*ms
)
2654 RAMState
*rs
= ram_state
;
2656 RCU_READ_LOCK_GUARD();
2658 /* This should be our last sync, the src is now paused */
2659 migration_bitmap_sync(rs
, false);
2661 /* Easiest way to make sure we don't resume in the middle of a host-page */
2662 rs
->pss
[RAM_CHANNEL_PRECOPY
].last_sent_block
= NULL
;
2663 rs
->last_seen_block
= NULL
;
2666 postcopy_each_ram_send_discard(ms
);
2668 trace_ram_postcopy_send_discard_bitmap();
2672 * ram_discard_range: discard dirtied pages at the beginning of postcopy
2674 * Returns zero on success
2676 * @rbname: name of the RAMBlock of the request. NULL means the
2677 * same that last one.
2678 * @start: RAMBlock starting page
2679 * @length: RAMBlock size
2681 int ram_discard_range(const char *rbname
, uint64_t start
, size_t length
)
2683 trace_ram_discard_range(rbname
, start
, length
);
2685 RCU_READ_LOCK_GUARD();
2686 RAMBlock
*rb
= qemu_ram_block_by_name(rbname
);
2689 error_report("ram_discard_range: Failed to find block '%s'", rbname
);
2694 * On source VM, we don't need to update the received bitmap since
2695 * we don't even have one.
2697 if (rb
->receivedmap
) {
2698 bitmap_clear(rb
->receivedmap
, start
>> qemu_target_page_bits(),
2699 length
>> qemu_target_page_bits());
2702 return ram_block_discard_range(rb
, start
, length
);
2706 * For every allocation, we will try not to crash the VM if the
2707 * allocation failed.
2709 static int xbzrle_init(void)
2711 Error
*local_err
= NULL
;
2713 if (!migrate_xbzrle()) {
2717 XBZRLE_cache_lock();
2719 XBZRLE
.zero_target_page
= g_try_malloc0(TARGET_PAGE_SIZE
);
2720 if (!XBZRLE
.zero_target_page
) {
2721 error_report("%s: Error allocating zero page", __func__
);
2725 XBZRLE
.cache
= cache_init(migrate_xbzrle_cache_size(),
2726 TARGET_PAGE_SIZE
, &local_err
);
2727 if (!XBZRLE
.cache
) {
2728 error_report_err(local_err
);
2729 goto free_zero_page
;
2732 XBZRLE
.encoded_buf
= g_try_malloc0(TARGET_PAGE_SIZE
);
2733 if (!XBZRLE
.encoded_buf
) {
2734 error_report("%s: Error allocating encoded_buf", __func__
);
2738 XBZRLE
.current_buf
= g_try_malloc(TARGET_PAGE_SIZE
);
2739 if (!XBZRLE
.current_buf
) {
2740 error_report("%s: Error allocating current_buf", __func__
);
2741 goto free_encoded_buf
;
2744 /* We are all good */
2745 XBZRLE_cache_unlock();
2749 g_free(XBZRLE
.encoded_buf
);
2750 XBZRLE
.encoded_buf
= NULL
;
2752 cache_fini(XBZRLE
.cache
);
2753 XBZRLE
.cache
= NULL
;
2755 g_free(XBZRLE
.zero_target_page
);
2756 XBZRLE
.zero_target_page
= NULL
;
2758 XBZRLE_cache_unlock();
2762 static int ram_state_init(RAMState
**rsp
)
2764 *rsp
= g_try_new0(RAMState
, 1);
2767 error_report("%s: Init ramstate fail", __func__
);
2771 qemu_mutex_init(&(*rsp
)->bitmap_mutex
);
2772 qemu_mutex_init(&(*rsp
)->src_page_req_mutex
);
2773 QSIMPLEQ_INIT(&(*rsp
)->src_page_requests
);
2774 (*rsp
)->ram_bytes_total
= ram_bytes_total();
2777 * Count the total number of pages used by ram blocks not including any
2778 * gaps due to alignment or unplugs.
2779 * This must match with the initial values of dirty bitmap.
2781 (*rsp
)->migration_dirty_pages
= (*rsp
)->ram_bytes_total
>> TARGET_PAGE_BITS
;
2782 ram_state_reset(*rsp
);
2787 static void ram_list_init_bitmaps(void)
2789 MigrationState
*ms
= migrate_get_current();
2791 unsigned long pages
;
2794 /* Skip setting bitmap if there is no RAM */
2795 if (ram_bytes_total()) {
2796 shift
= ms
->clear_bitmap_shift
;
2797 if (shift
> CLEAR_BITMAP_SHIFT_MAX
) {
2798 error_report("clear_bitmap_shift (%u) too big, using "
2799 "max value (%u)", shift
, CLEAR_BITMAP_SHIFT_MAX
);
2800 shift
= CLEAR_BITMAP_SHIFT_MAX
;
2801 } else if (shift
< CLEAR_BITMAP_SHIFT_MIN
) {
2802 error_report("clear_bitmap_shift (%u) too small, using "
2803 "min value (%u)", shift
, CLEAR_BITMAP_SHIFT_MIN
);
2804 shift
= CLEAR_BITMAP_SHIFT_MIN
;
2807 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
2808 pages
= block
->max_length
>> TARGET_PAGE_BITS
;
2810 * The initial dirty bitmap for migration must be set with all
2811 * ones to make sure we'll migrate every guest RAM page to
2813 * Here we set RAMBlock.bmap all to 1 because when rebegin a
2814 * new migration after a failed migration, ram_list.
2815 * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole
2818 block
->bmap
= bitmap_new(pages
);
2819 bitmap_set(block
->bmap
, 0, pages
);
2820 block
->clear_bmap_shift
= shift
;
2821 block
->clear_bmap
= bitmap_new(clear_bmap_size(pages
, shift
));
2826 static void migration_bitmap_clear_discarded_pages(RAMState
*rs
)
2828 unsigned long pages
;
2831 RCU_READ_LOCK_GUARD();
2833 RAMBLOCK_FOREACH_NOT_IGNORED(rb
) {
2834 pages
= ramblock_dirty_bitmap_clear_discarded_pages(rb
);
2835 rs
->migration_dirty_pages
-= pages
;
2839 static void ram_init_bitmaps(RAMState
*rs
)
2841 qemu_mutex_lock_ramlist();
2843 WITH_RCU_READ_LOCK_GUARD() {
2844 ram_list_init_bitmaps();
2845 /* We don't use dirty log with background snapshots */
2846 if (!migrate_background_snapshot()) {
2847 memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION
);
2848 migration_bitmap_sync_precopy(rs
, false);
2851 qemu_mutex_unlock_ramlist();
2854 * After an eventual first bitmap sync, fixup the initial bitmap
2855 * containing all 1s to exclude any discarded pages from migration.
2857 migration_bitmap_clear_discarded_pages(rs
);
2860 static int ram_init_all(RAMState
**rsp
)
2862 if (ram_state_init(rsp
)) {
2866 if (xbzrle_init()) {
2867 ram_state_cleanup(rsp
);
2871 ram_init_bitmaps(*rsp
);
2876 static void ram_state_resume_prepare(RAMState
*rs
, QEMUFile
*out
)
2882 * Postcopy is not using xbzrle/compression, so no need for that.
2883 * Also, since source are already halted, we don't need to care
2884 * about dirty page logging as well.
2887 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
2888 pages
+= bitmap_count_one(block
->bmap
,
2889 block
->used_length
>> TARGET_PAGE_BITS
);
2892 /* This may not be aligned with current bitmaps. Recalculate. */
2893 rs
->migration_dirty_pages
= pages
;
2895 ram_state_reset(rs
);
2897 /* Update RAMState cache of output QEMUFile */
2898 rs
->pss
[RAM_CHANNEL_PRECOPY
].pss_channel
= out
;
2900 trace_ram_state_resume_prepare(pages
);
2904 * This function clears bits of the free pages reported by the caller from the
2905 * migration dirty bitmap. @addr is the host address corresponding to the
2906 * start of the continuous guest free pages, and @len is the total bytes of
2909 void qemu_guest_free_page_hint(void *addr
, size_t len
)
2913 size_t used_len
, start
, npages
;
2914 MigrationState
*s
= migrate_get_current();
2916 /* This function is currently expected to be used during live migration */
2917 if (!migration_is_setup_or_active(s
->state
)) {
2921 for (; len
> 0; len
-= used_len
, addr
+= used_len
) {
2922 block
= qemu_ram_block_from_host(addr
, false, &offset
);
2923 if (unlikely(!block
|| offset
>= block
->used_length
)) {
2925 * The implementation might not support RAMBlock resize during
2926 * live migration, but it could happen in theory with future
2927 * updates. So we add a check here to capture that case.
2929 error_report_once("%s unexpected error", __func__
);
2933 if (len
<= block
->used_length
- offset
) {
2936 used_len
= block
->used_length
- offset
;
2939 start
= offset
>> TARGET_PAGE_BITS
;
2940 npages
= used_len
>> TARGET_PAGE_BITS
;
2942 qemu_mutex_lock(&ram_state
->bitmap_mutex
);
2944 * The skipped free pages are equavalent to be sent from clear_bmap's
2945 * perspective, so clear the bits from the memory region bitmap which
2946 * are initially set. Otherwise those skipped pages will be sent in
2947 * the next round after syncing from the memory region bitmap.
2949 migration_clear_memory_region_dirty_bitmap_range(block
, start
, npages
);
2950 ram_state
->migration_dirty_pages
-=
2951 bitmap_count_one_with_offset(block
->bmap
, start
, npages
);
2952 bitmap_clear(block
->bmap
, start
, npages
);
2953 qemu_mutex_unlock(&ram_state
->bitmap_mutex
);
2958 * Each of ram_save_setup, ram_save_iterate and ram_save_complete has
2959 * long-running RCU critical section. When rcu-reclaims in the code
2960 * start to become numerous it will be necessary to reduce the
2961 * granularity of these critical sections.
2965 * ram_save_setup: Setup RAM for migration
2967 * Returns zero to indicate success and negative for error
2969 * @f: QEMUFile where to send the data
2970 * @opaque: RAMState pointer
2972 static int ram_save_setup(QEMUFile
*f
, void *opaque
)
2974 RAMState
**rsp
= opaque
;
2978 if (compress_threads_save_setup()) {
2982 /* migration has already setup the bitmap, reuse it. */
2983 if (!migration_in_colo_state()) {
2984 if (ram_init_all(rsp
) != 0) {
2985 compress_threads_save_cleanup();
2989 (*rsp
)->pss
[RAM_CHANNEL_PRECOPY
].pss_channel
= f
;
2991 WITH_RCU_READ_LOCK_GUARD() {
2992 qemu_put_be64(f
, ram_bytes_total_with_ignored()
2993 | RAM_SAVE_FLAG_MEM_SIZE
);
2995 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2996 qemu_put_byte(f
, strlen(block
->idstr
));
2997 qemu_put_buffer(f
, (uint8_t *)block
->idstr
, strlen(block
->idstr
));
2998 qemu_put_be64(f
, block
->used_length
);
2999 if (migrate_postcopy_ram() && block
->page_size
!=
3000 qemu_host_page_size
) {
3001 qemu_put_be64(f
, block
->page_size
);
3003 if (migrate_ignore_shared()) {
3004 qemu_put_be64(f
, block
->mr
->addr
);
3009 ret
= rdma_registration_start(f
, RAM_CONTROL_SETUP
);
3011 qemu_file_set_error(f
, ret
);
3014 ret
= rdma_registration_stop(f
, RAM_CONTROL_SETUP
);
3016 qemu_file_set_error(f
, ret
);
3019 migration_ops
= g_malloc0(sizeof(MigrationOps
));
3020 migration_ops
->ram_save_target_page
= ram_save_target_page_legacy
;
3022 qemu_mutex_unlock_iothread();
3023 ret
= multifd_send_sync_main(f
);
3024 qemu_mutex_lock_iothread();
3029 if (migrate_multifd() && !migrate_multifd_flush_after_each_section()) {
3030 qemu_put_be64(f
, RAM_SAVE_FLAG_MULTIFD_FLUSH
);
3033 qemu_put_be64(f
, RAM_SAVE_FLAG_EOS
);
3040 * ram_save_iterate: iterative stage for migration
3042 * Returns zero to indicate success and negative for error
3044 * @f: QEMUFile where to send the data
3045 * @opaque: RAMState pointer
3047 static int ram_save_iterate(QEMUFile
*f
, void *opaque
)
3049 RAMState
**temp
= opaque
;
3050 RAMState
*rs
= *temp
;
3056 if (blk_mig_bulk_active()) {
3057 /* Avoid transferring ram during bulk phase of block migration as
3058 * the bulk phase will usually take a long time and transferring
3059 * ram updates during that time is pointless. */
3064 * We'll take this lock a little bit long, but it's okay for two reasons.
3065 * Firstly, the only possible other thread to take it is who calls
3066 * qemu_guest_free_page_hint(), which should be rare; secondly, see
3067 * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which
3068 * guarantees that we'll at least released it in a regular basis.
3070 qemu_mutex_lock(&rs
->bitmap_mutex
);
3071 WITH_RCU_READ_LOCK_GUARD() {
3072 if (ram_list
.version
!= rs
->last_version
) {
3073 ram_state_reset(rs
);
3076 /* Read version before ram_list.blocks */
3079 ret
= rdma_registration_start(f
, RAM_CONTROL_ROUND
);
3081 qemu_file_set_error(f
, ret
);
3084 t0
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
3086 while ((ret
= migration_rate_exceeded(f
)) == 0 ||
3087 postcopy_has_request(rs
)) {
3090 if (qemu_file_get_error(f
)) {
3094 pages
= ram_find_and_save_block(rs
);
3095 /* no more pages to sent */
3102 qemu_file_set_error(f
, pages
);
3106 rs
->target_page_count
+= pages
;
3109 * During postcopy, it is necessary to make sure one whole host
3110 * page is sent in one chunk.
3112 if (migrate_postcopy_ram()) {
3113 ram_flush_compressed_data();
3117 * we want to check in the 1st loop, just in case it was the 1st
3118 * time and we had to sync the dirty bitmap.
3119 * qemu_clock_get_ns() is a bit expensive, so we only check each
3122 if ((i
& 63) == 0) {
3123 uint64_t t1
= (qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - t0
) /
3125 if (t1
> MAX_WAIT
) {
3126 trace_ram_save_iterate_big_wait(t1
, i
);
3133 qemu_mutex_unlock(&rs
->bitmap_mutex
);
3136 * Must occur before EOS (or any QEMUFile operation)
3137 * because of RDMA protocol.
3139 ret
= rdma_registration_stop(f
, RAM_CONTROL_ROUND
);
3141 qemu_file_set_error(f
, ret
);
3146 && migration_is_setup_or_active(migrate_get_current()->state
)) {
3147 if (migrate_multifd() && migrate_multifd_flush_after_each_section()) {
3148 ret
= multifd_send_sync_main(rs
->pss
[RAM_CHANNEL_PRECOPY
].pss_channel
);
3154 qemu_put_be64(f
, RAM_SAVE_FLAG_EOS
);
3156 ram_transferred_add(8);
3158 ret
= qemu_file_get_error(f
);
3168 * ram_save_complete: function called to send the remaining amount of ram
3170 * Returns zero to indicate success or negative on error
3172 * Called with iothread lock
3174 * @f: QEMUFile where to send the data
3175 * @opaque: RAMState pointer
3177 static int ram_save_complete(QEMUFile
*f
, void *opaque
)
3179 RAMState
**temp
= opaque
;
3180 RAMState
*rs
= *temp
;
3183 rs
->last_stage
= !migration_in_colo_state();
3185 WITH_RCU_READ_LOCK_GUARD() {
3186 if (!migration_in_postcopy()) {
3187 migration_bitmap_sync_precopy(rs
, true);
3190 ret
= rdma_registration_start(f
, RAM_CONTROL_FINISH
);
3192 qemu_file_set_error(f
, ret
);
3195 /* try transferring iterative blocks of memory */
3197 /* flush all remaining blocks regardless of rate limiting */
3198 qemu_mutex_lock(&rs
->bitmap_mutex
);
3202 pages
= ram_find_and_save_block(rs
);
3203 /* no more blocks to sent */
3212 qemu_mutex_unlock(&rs
->bitmap_mutex
);
3214 ram_flush_compressed_data();
3216 int ret
= rdma_registration_stop(f
, RAM_CONTROL_FINISH
);
3218 qemu_file_set_error(f
, ret
);
3226 ret
= multifd_send_sync_main(rs
->pss
[RAM_CHANNEL_PRECOPY
].pss_channel
);
3231 if (migrate_multifd() && !migrate_multifd_flush_after_each_section()) {
3232 qemu_put_be64(f
, RAM_SAVE_FLAG_MULTIFD_FLUSH
);
3234 qemu_put_be64(f
, RAM_SAVE_FLAG_EOS
);
3240 static void ram_state_pending_estimate(void *opaque
, uint64_t *must_precopy
,
3241 uint64_t *can_postcopy
)
3243 RAMState
**temp
= opaque
;
3244 RAMState
*rs
= *temp
;
3246 uint64_t remaining_size
= rs
->migration_dirty_pages
* TARGET_PAGE_SIZE
;
3248 if (migrate_postcopy_ram()) {
3249 /* We can do postcopy, and all the data is postcopiable */
3250 *can_postcopy
+= remaining_size
;
3252 *must_precopy
+= remaining_size
;
3256 static void ram_state_pending_exact(void *opaque
, uint64_t *must_precopy
,
3257 uint64_t *can_postcopy
)
3259 MigrationState
*s
= migrate_get_current();
3260 RAMState
**temp
= opaque
;
3261 RAMState
*rs
= *temp
;
3263 uint64_t remaining_size
= rs
->migration_dirty_pages
* TARGET_PAGE_SIZE
;
3265 if (!migration_in_postcopy() && remaining_size
< s
->threshold_size
) {
3266 qemu_mutex_lock_iothread();
3267 WITH_RCU_READ_LOCK_GUARD() {
3268 migration_bitmap_sync_precopy(rs
, false);
3270 qemu_mutex_unlock_iothread();
3271 remaining_size
= rs
->migration_dirty_pages
* TARGET_PAGE_SIZE
;
3274 if (migrate_postcopy_ram()) {
3275 /* We can do postcopy, and all the data is postcopiable */
3276 *can_postcopy
+= remaining_size
;
3278 *must_precopy
+= remaining_size
;
3282 static int load_xbzrle(QEMUFile
*f
, ram_addr_t addr
, void *host
)
3284 unsigned int xh_len
;
3286 uint8_t *loaded_data
;
3288 /* extract RLE header */
3289 xh_flags
= qemu_get_byte(f
);
3290 xh_len
= qemu_get_be16(f
);
3292 if (xh_flags
!= ENCODING_FLAG_XBZRLE
) {
3293 error_report("Failed to load XBZRLE page - wrong compression!");
3297 if (xh_len
> TARGET_PAGE_SIZE
) {
3298 error_report("Failed to load XBZRLE page - len overflow!");
3301 loaded_data
= XBZRLE
.decoded_buf
;
3302 /* load data and decode */
3303 /* it can change loaded_data to point to an internal buffer */
3304 qemu_get_buffer_in_place(f
, &loaded_data
, xh_len
);
3307 if (xbzrle_decode_buffer(loaded_data
, xh_len
, host
,
3308 TARGET_PAGE_SIZE
) == -1) {
3309 error_report("Failed to load XBZRLE page - decode error!");
3317 * ram_block_from_stream: read a RAMBlock id from the migration stream
3319 * Must be called from within a rcu critical section.
3321 * Returns a pointer from within the RCU-protected ram_list.
3323 * @mis: the migration incoming state pointer
3324 * @f: QEMUFile where to read the data from
3325 * @flags: Page flags (mostly to see if it's a continuation of previous block)
3326 * @channel: the channel we're using
3328 static inline RAMBlock
*ram_block_from_stream(MigrationIncomingState
*mis
,
3329 QEMUFile
*f
, int flags
,
3332 RAMBlock
*block
= mis
->last_recv_block
[channel
];
3336 if (flags
& RAM_SAVE_FLAG_CONTINUE
) {
3338 error_report("Ack, bad migration stream!");
3344 len
= qemu_get_byte(f
);
3345 qemu_get_buffer(f
, (uint8_t *)id
, len
);
3348 block
= qemu_ram_block_by_name(id
);
3350 error_report("Can't find block %s", id
);
3354 if (migrate_ram_is_ignored(block
)) {
3355 error_report("block %s should not be migrated !", id
);
3359 mis
->last_recv_block
[channel
] = block
;
3364 static inline void *host_from_ram_block_offset(RAMBlock
*block
,
3367 if (!offset_in_ramblock(block
, offset
)) {
3371 return block
->host
+ offset
;
3374 static void *host_page_from_ram_block_offset(RAMBlock
*block
,
3377 /* Note: Explicitly no check against offset_in_ramblock(). */
3378 return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block
->host
+ offset
),
3382 static ram_addr_t
host_page_offset_from_ram_block_offset(RAMBlock
*block
,
3385 return ((uintptr_t)block
->host
+ offset
) & (block
->page_size
- 1);
3388 void colo_record_bitmap(RAMBlock
*block
, ram_addr_t
*normal
, uint32_t pages
)
3390 qemu_mutex_lock(&ram_state
->bitmap_mutex
);
3391 for (int i
= 0; i
< pages
; i
++) {
3392 ram_addr_t offset
= normal
[i
];
3393 ram_state
->migration_dirty_pages
+= !test_and_set_bit(
3394 offset
>> TARGET_PAGE_BITS
,
3397 qemu_mutex_unlock(&ram_state
->bitmap_mutex
);
3400 static inline void *colo_cache_from_block_offset(RAMBlock
*block
,
3401 ram_addr_t offset
, bool record_bitmap
)
3403 if (!offset_in_ramblock(block
, offset
)) {
3406 if (!block
->colo_cache
) {
3407 error_report("%s: colo_cache is NULL in block :%s",
3408 __func__
, block
->idstr
);
3413 * During colo checkpoint, we need bitmap of these migrated pages.
3414 * It help us to decide which pages in ram cache should be flushed
3415 * into VM's RAM later.
3417 if (record_bitmap
) {
3418 colo_record_bitmap(block
, &offset
, 1);
3420 return block
->colo_cache
+ offset
;
3424 * ram_handle_zero: handle the zero page case
3426 * If a page (or a whole RDMA chunk) has been
3427 * determined to be zero, then zap it.
3429 * @host: host address for the zero page
3430 * @ch: what the page is filled from. We only support zero
3431 * @size: size of the zero page
3433 void ram_handle_zero(void *host
, uint64_t size
)
3435 if (!buffer_is_zero(host
, size
)) {
3436 memset(host
, 0, size
);
3440 static void colo_init_ram_state(void)
3442 ram_state_init(&ram_state
);
3446 * colo cache: this is for secondary VM, we cache the whole
3447 * memory of the secondary VM, it is need to hold the global lock
3448 * to call this helper.
3450 int colo_init_ram_cache(void)
3454 WITH_RCU_READ_LOCK_GUARD() {
3455 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3456 block
->colo_cache
= qemu_anon_ram_alloc(block
->used_length
,
3457 NULL
, false, false);
3458 if (!block
->colo_cache
) {
3459 error_report("%s: Can't alloc memory for COLO cache of block %s,"
3460 "size 0x" RAM_ADDR_FMT
, __func__
, block
->idstr
,
3461 block
->used_length
);
3462 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3463 if (block
->colo_cache
) {
3464 qemu_anon_ram_free(block
->colo_cache
, block
->used_length
);
3465 block
->colo_cache
= NULL
;
3470 if (!machine_dump_guest_core(current_machine
)) {
3471 qemu_madvise(block
->colo_cache
, block
->used_length
,
3472 QEMU_MADV_DONTDUMP
);
3478 * Record the dirty pages that sent by PVM, we use this dirty bitmap together
3479 * with to decide which page in cache should be flushed into SVM's RAM. Here
3480 * we use the same name 'ram_bitmap' as for migration.
3482 if (ram_bytes_total()) {
3483 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3484 unsigned long pages
= block
->max_length
>> TARGET_PAGE_BITS
;
3485 block
->bmap
= bitmap_new(pages
);
3489 colo_init_ram_state();
3493 /* TODO: duplicated with ram_init_bitmaps */
3494 void colo_incoming_start_dirty_log(void)
3496 RAMBlock
*block
= NULL
;
3497 /* For memory_global_dirty_log_start below. */
3498 qemu_mutex_lock_iothread();
3499 qemu_mutex_lock_ramlist();
3501 memory_global_dirty_log_sync(false);
3502 WITH_RCU_READ_LOCK_GUARD() {
3503 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3504 ramblock_sync_dirty_bitmap(ram_state
, block
);
3505 /* Discard this dirty bitmap record */
3506 bitmap_zero(block
->bmap
, block
->max_length
>> TARGET_PAGE_BITS
);
3508 memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION
);
3510 ram_state
->migration_dirty_pages
= 0;
3511 qemu_mutex_unlock_ramlist();
3512 qemu_mutex_unlock_iothread();
3515 /* It is need to hold the global lock to call this helper */
3516 void colo_release_ram_cache(void)
3520 memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION
);
3521 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3522 g_free(block
->bmap
);
3526 WITH_RCU_READ_LOCK_GUARD() {
3527 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3528 if (block
->colo_cache
) {
3529 qemu_anon_ram_free(block
->colo_cache
, block
->used_length
);
3530 block
->colo_cache
= NULL
;
3534 ram_state_cleanup(&ram_state
);
3538 * ram_load_setup: Setup RAM for migration incoming side
3540 * Returns zero to indicate success and negative for error
3542 * @f: QEMUFile where to receive the data
3543 * @opaque: RAMState pointer
3545 static int ram_load_setup(QEMUFile
*f
, void *opaque
)
3547 xbzrle_load_setup();
3548 ramblock_recv_map_init();
3553 static int ram_load_cleanup(void *opaque
)
3557 RAMBLOCK_FOREACH_NOT_IGNORED(rb
) {
3558 qemu_ram_block_writeback(rb
);
3561 xbzrle_load_cleanup();
3563 RAMBLOCK_FOREACH_NOT_IGNORED(rb
) {
3564 g_free(rb
->receivedmap
);
3565 rb
->receivedmap
= NULL
;
3572 * ram_postcopy_incoming_init: allocate postcopy data structures
3574 * Returns 0 for success and negative if there was one error
3576 * @mis: current migration incoming state
3578 * Allocate data structures etc needed by incoming migration with
3579 * postcopy-ram. postcopy-ram's similarly names
3580 * postcopy_ram_incoming_init does the work.
3582 int ram_postcopy_incoming_init(MigrationIncomingState
*mis
)
3584 return postcopy_ram_incoming_init(mis
);
3588 * ram_load_postcopy: load a page in postcopy case
3590 * Returns 0 for success or -errno in case of error
3592 * Called in postcopy mode by ram_load().
3593 * rcu_read_lock is taken prior to this being called.
3595 * @f: QEMUFile where to send the data
3596 * @channel: the channel to use for loading
3598 int ram_load_postcopy(QEMUFile
*f
, int channel
)
3600 int flags
= 0, ret
= 0;
3601 bool place_needed
= false;
3602 bool matches_target_page_size
= false;
3603 MigrationIncomingState
*mis
= migration_incoming_get_current();
3604 PostcopyTmpPage
*tmp_page
= &mis
->postcopy_tmp_pages
[channel
];
3606 while (!ret
&& !(flags
& RAM_SAVE_FLAG_EOS
)) {
3608 void *page_buffer
= NULL
;
3609 void *place_source
= NULL
;
3610 RAMBlock
*block
= NULL
;
3614 addr
= qemu_get_be64(f
);
3617 * If qemu file error, we should stop here, and then "addr"
3620 ret
= qemu_file_get_error(f
);
3625 flags
= addr
& ~TARGET_PAGE_MASK
;
3626 addr
&= TARGET_PAGE_MASK
;
3628 trace_ram_load_postcopy_loop(channel
, (uint64_t)addr
, flags
);
3629 if (flags
& (RAM_SAVE_FLAG_ZERO
| RAM_SAVE_FLAG_PAGE
|
3630 RAM_SAVE_FLAG_COMPRESS_PAGE
)) {
3631 block
= ram_block_from_stream(mis
, f
, flags
, channel
);
3638 * Relying on used_length is racy and can result in false positives.
3639 * We might place pages beyond used_length in case RAM was shrunk
3640 * while in postcopy, which is fine - trying to place via
3641 * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault.
3643 if (!block
->host
|| addr
>= block
->postcopy_length
) {
3644 error_report("Illegal RAM offset " RAM_ADDR_FMT
, addr
);
3648 tmp_page
->target_pages
++;
3649 matches_target_page_size
= block
->page_size
== TARGET_PAGE_SIZE
;
3651 * Postcopy requires that we place whole host pages atomically;
3652 * these may be huge pages for RAMBlocks that are backed by
3654 * To make it atomic, the data is read into a temporary page
3655 * that's moved into place later.
3656 * The migration protocol uses, possibly smaller, target-pages
3657 * however the source ensures it always sends all the components
3658 * of a host page in one chunk.
3660 page_buffer
= tmp_page
->tmp_huge_page
+
3661 host_page_offset_from_ram_block_offset(block
, addr
);
3662 /* If all TP are zero then we can optimise the place */
3663 if (tmp_page
->target_pages
== 1) {
3664 tmp_page
->host_addr
=
3665 host_page_from_ram_block_offset(block
, addr
);
3666 } else if (tmp_page
->host_addr
!=
3667 host_page_from_ram_block_offset(block
, addr
)) {
3668 /* not the 1st TP within the HP */
3669 error_report("Non-same host page detected on channel %d: "
3670 "Target host page %p, received host page %p "
3671 "(rb %s offset 0x"RAM_ADDR_FMT
" target_pages %d)",
3672 channel
, tmp_page
->host_addr
,
3673 host_page_from_ram_block_offset(block
, addr
),
3674 block
->idstr
, addr
, tmp_page
->target_pages
);
3680 * If it's the last part of a host page then we place the host
3683 if (tmp_page
->target_pages
==
3684 (block
->page_size
/ TARGET_PAGE_SIZE
)) {
3685 place_needed
= true;
3687 place_source
= tmp_page
->tmp_huge_page
;
3690 switch (flags
& ~RAM_SAVE_FLAG_CONTINUE
) {
3691 case RAM_SAVE_FLAG_ZERO
:
3692 ch
= qemu_get_byte(f
);
3694 error_report("Found a zero page with value %d", ch
);
3699 * Can skip to set page_buffer when
3700 * this is a zero page and (block->page_size == TARGET_PAGE_SIZE).
3702 if (!matches_target_page_size
) {
3703 memset(page_buffer
, ch
, TARGET_PAGE_SIZE
);
3707 case RAM_SAVE_FLAG_PAGE
:
3708 tmp_page
->all_zero
= false;
3709 if (!matches_target_page_size
) {
3710 /* For huge pages, we always use temporary buffer */
3711 qemu_get_buffer(f
, page_buffer
, TARGET_PAGE_SIZE
);
3714 * For small pages that matches target page size, we
3715 * avoid the qemu_file copy. Instead we directly use
3716 * the buffer of QEMUFile to place the page. Note: we
3717 * cannot do any QEMUFile operation before using that
3718 * buffer to make sure the buffer is valid when
3721 qemu_get_buffer_in_place(f
, (uint8_t **)&place_source
,
3725 case RAM_SAVE_FLAG_COMPRESS_PAGE
:
3726 tmp_page
->all_zero
= false;
3727 len
= qemu_get_be32(f
);
3728 if (len
< 0 || len
> compressBound(TARGET_PAGE_SIZE
)) {
3729 error_report("Invalid compressed data length: %d", len
);
3733 decompress_data_with_multi_threads(f
, page_buffer
, len
);
3735 case RAM_SAVE_FLAG_MULTIFD_FLUSH
:
3736 multifd_recv_sync_main();
3738 case RAM_SAVE_FLAG_EOS
:
3740 if (migrate_multifd() &&
3741 migrate_multifd_flush_after_each_section()) {
3742 multifd_recv_sync_main();
3746 error_report("Unknown combination of migration flags: 0x%x"
3747 " (postcopy mode)", flags
);
3752 /* Got the whole host page, wait for decompress before placing. */
3754 ret
|= wait_for_decompress_done();
3757 /* Detect for any possible file errors */
3758 if (!ret
&& qemu_file_get_error(f
)) {
3759 ret
= qemu_file_get_error(f
);
3762 if (!ret
&& place_needed
) {
3763 if (tmp_page
->all_zero
) {
3764 ret
= postcopy_place_page_zero(mis
, tmp_page
->host_addr
, block
);
3766 ret
= postcopy_place_page(mis
, tmp_page
->host_addr
,
3767 place_source
, block
);
3769 place_needed
= false;
3770 postcopy_temp_page_reset(tmp_page
);
3777 static bool postcopy_is_running(void)
3779 PostcopyState ps
= postcopy_state_get();
3780 return ps
>= POSTCOPY_INCOMING_LISTENING
&& ps
< POSTCOPY_INCOMING_END
;
3784 * Flush content of RAM cache into SVM's memory.
3785 * Only flush the pages that be dirtied by PVM or SVM or both.
3787 void colo_flush_ram_cache(void)
3789 RAMBlock
*block
= NULL
;
3792 unsigned long offset
= 0;
3794 memory_global_dirty_log_sync(false);
3795 qemu_mutex_lock(&ram_state
->bitmap_mutex
);
3796 WITH_RCU_READ_LOCK_GUARD() {
3797 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
3798 ramblock_sync_dirty_bitmap(ram_state
, block
);
3802 trace_colo_flush_ram_cache_begin(ram_state
->migration_dirty_pages
);
3803 WITH_RCU_READ_LOCK_GUARD() {
3804 block
= QLIST_FIRST_RCU(&ram_list
.blocks
);
3807 unsigned long num
= 0;
3809 offset
= colo_bitmap_find_dirty(ram_state
, block
, offset
, &num
);
3810 if (!offset_in_ramblock(block
,
3811 ((ram_addr_t
)offset
) << TARGET_PAGE_BITS
)) {
3814 block
= QLIST_NEXT_RCU(block
, next
);
3816 unsigned long i
= 0;
3818 for (i
= 0; i
< num
; i
++) {
3819 migration_bitmap_clear_dirty(ram_state
, block
, offset
+ i
);
3821 dst_host
= block
->host
3822 + (((ram_addr_t
)offset
) << TARGET_PAGE_BITS
);
3823 src_host
= block
->colo_cache
3824 + (((ram_addr_t
)offset
) << TARGET_PAGE_BITS
);
3825 memcpy(dst_host
, src_host
, TARGET_PAGE_SIZE
* num
);
3830 qemu_mutex_unlock(&ram_state
->bitmap_mutex
);
3831 trace_colo_flush_ram_cache_end();
3834 static int parse_ramblock(QEMUFile
*f
, RAMBlock
*block
, ram_addr_t length
)
3837 /* ADVISE is earlier, it shows the source has the postcopy capability on */
3838 bool postcopy_advised
= migration_incoming_postcopy_advised();
3842 if (!qemu_ram_is_migratable(block
)) {
3843 error_report("block %s should not be migrated !", block
->idstr
);
3847 if (length
!= block
->used_length
) {
3848 Error
*local_err
= NULL
;
3850 ret
= qemu_ram_resize(block
, length
, &local_err
);
3852 error_report_err(local_err
);
3856 /* For postcopy we need to check hugepage sizes match */
3857 if (postcopy_advised
&& migrate_postcopy_ram() &&
3858 block
->page_size
!= qemu_host_page_size
) {
3859 uint64_t remote_page_size
= qemu_get_be64(f
);
3860 if (remote_page_size
!= block
->page_size
) {
3861 error_report("Mismatched RAM page size %s "
3862 "(local) %zd != %" PRId64
, block
->idstr
,
3863 block
->page_size
, remote_page_size
);
3867 if (migrate_ignore_shared()) {
3868 hwaddr addr
= qemu_get_be64(f
);
3869 if (migrate_ram_is_ignored(block
) &&
3870 block
->mr
->addr
!= addr
) {
3871 error_report("Mismatched GPAs for block %s "
3872 "%" PRId64
"!= %" PRId64
, block
->idstr
,
3873 (uint64_t)addr
, (uint64_t)block
->mr
->addr
);
3877 ret
= rdma_block_notification_handle(f
, block
->idstr
);
3879 qemu_file_set_error(f
, ret
);
3885 static int parse_ramblocks(QEMUFile
*f
, ram_addr_t total_ram_bytes
)
3889 /* Synchronize RAM block list */
3890 while (!ret
&& total_ram_bytes
) {
3894 int len
= qemu_get_byte(f
);
3896 qemu_get_buffer(f
, (uint8_t *)id
, len
);
3898 length
= qemu_get_be64(f
);
3900 block
= qemu_ram_block_by_name(id
);
3902 ret
= parse_ramblock(f
, block
, length
);
3904 error_report("Unknown ramblock \"%s\", cannot accept "
3908 total_ram_bytes
-= length
;
3915 * ram_load_precopy: load pages in precopy case
3917 * Returns 0 for success or -errno in case of error
3919 * Called in precopy mode by ram_load().
3920 * rcu_read_lock is taken prior to this being called.
3922 * @f: QEMUFile where to send the data
3924 static int ram_load_precopy(QEMUFile
*f
)
3926 MigrationIncomingState
*mis
= migration_incoming_get_current();
3927 int flags
= 0, ret
= 0, invalid_flags
= 0, len
= 0, i
= 0;
3929 if (!migrate_compress()) {
3930 invalid_flags
|= RAM_SAVE_FLAG_COMPRESS_PAGE
;
3933 while (!ret
&& !(flags
& RAM_SAVE_FLAG_EOS
)) {
3935 void *host
= NULL
, *host_bak
= NULL
;
3939 * Yield periodically to let main loop run, but an iteration of
3940 * the main loop is expensive, so do it each some iterations
3942 if ((i
& 32767) == 0 && qemu_in_coroutine()) {
3943 aio_co_schedule(qemu_get_current_aio_context(),
3944 qemu_coroutine_self());
3945 qemu_coroutine_yield();
3949 addr
= qemu_get_be64(f
);
3950 flags
= addr
& ~TARGET_PAGE_MASK
;
3951 addr
&= TARGET_PAGE_MASK
;
3953 if (flags
& invalid_flags
) {
3954 if (flags
& invalid_flags
& RAM_SAVE_FLAG_COMPRESS_PAGE
) {
3955 error_report("Received an unexpected compressed page");
3962 if (flags
& (RAM_SAVE_FLAG_ZERO
| RAM_SAVE_FLAG_PAGE
|
3963 RAM_SAVE_FLAG_COMPRESS_PAGE
| RAM_SAVE_FLAG_XBZRLE
)) {
3964 RAMBlock
*block
= ram_block_from_stream(mis
, f
, flags
,
3965 RAM_CHANNEL_PRECOPY
);
3967 host
= host_from_ram_block_offset(block
, addr
);
3969 * After going into COLO stage, we should not load the page
3970 * into SVM's memory directly, we put them into colo_cache firstly.
3971 * NOTE: We need to keep a copy of SVM's ram in colo_cache.
3972 * Previously, we copied all these memory in preparing stage of COLO
3973 * while we need to stop VM, which is a time-consuming process.
3974 * Here we optimize it by a trick, back-up every page while in
3975 * migration process while COLO is enabled, though it affects the
3976 * speed of the migration, but it obviously reduce the downtime of
3977 * back-up all SVM'S memory in COLO preparing stage.
3979 if (migration_incoming_colo_enabled()) {
3980 if (migration_incoming_in_colo_state()) {
3981 /* In COLO stage, put all pages into cache temporarily */
3982 host
= colo_cache_from_block_offset(block
, addr
, true);
3985 * In migration stage but before COLO stage,
3986 * Put all pages into both cache and SVM's memory.
3988 host_bak
= colo_cache_from_block_offset(block
, addr
, false);
3992 error_report("Illegal RAM offset " RAM_ADDR_FMT
, addr
);
3996 if (!migration_incoming_in_colo_state()) {
3997 ramblock_recv_bitmap_set(block
, host
);
4000 trace_ram_load_loop(block
->idstr
, (uint64_t)addr
, flags
, host
);
4003 switch (flags
& ~RAM_SAVE_FLAG_CONTINUE
) {
4004 case RAM_SAVE_FLAG_MEM_SIZE
:
4005 ret
= parse_ramblocks(f
, addr
);
4008 case RAM_SAVE_FLAG_ZERO
:
4009 ch
= qemu_get_byte(f
);
4011 error_report("Found a zero page with value %d", ch
);
4015 ram_handle_zero(host
, TARGET_PAGE_SIZE
);
4018 case RAM_SAVE_FLAG_PAGE
:
4019 qemu_get_buffer(f
, host
, TARGET_PAGE_SIZE
);
4022 case RAM_SAVE_FLAG_COMPRESS_PAGE
:
4023 len
= qemu_get_be32(f
);
4024 if (len
< 0 || len
> compressBound(TARGET_PAGE_SIZE
)) {
4025 error_report("Invalid compressed data length: %d", len
);
4029 decompress_data_with_multi_threads(f
, host
, len
);
4032 case RAM_SAVE_FLAG_XBZRLE
:
4033 if (load_xbzrle(f
, addr
, host
) < 0) {
4034 error_report("Failed to decompress XBZRLE page at "
4035 RAM_ADDR_FMT
, addr
);
4040 case RAM_SAVE_FLAG_MULTIFD_FLUSH
:
4041 multifd_recv_sync_main();
4043 case RAM_SAVE_FLAG_EOS
:
4045 if (migrate_multifd() &&
4046 migrate_multifd_flush_after_each_section()) {
4047 multifd_recv_sync_main();
4050 case RAM_SAVE_FLAG_HOOK
:
4051 ret
= rdma_registration_handle(f
);
4053 qemu_file_set_error(f
, ret
);
4057 error_report("Unknown combination of migration flags: 0x%x", flags
);
4061 ret
= qemu_file_get_error(f
);
4063 if (!ret
&& host_bak
) {
4064 memcpy(host_bak
, host
, TARGET_PAGE_SIZE
);
4068 ret
|= wait_for_decompress_done();
4072 static int ram_load(QEMUFile
*f
, void *opaque
, int version_id
)
4075 static uint64_t seq_iter
;
4077 * If system is running in postcopy mode, page inserts to host memory must
4080 bool postcopy_running
= postcopy_is_running();
4084 if (version_id
!= 4) {
4089 * This RCU critical section can be very long running.
4090 * When RCU reclaims in the code start to become numerous,
4091 * it will be necessary to reduce the granularity of this
4094 WITH_RCU_READ_LOCK_GUARD() {
4095 if (postcopy_running
) {
4097 * Note! Here RAM_CHANNEL_PRECOPY is the precopy channel of
4098 * postcopy migration, we have another RAM_CHANNEL_POSTCOPY to
4099 * service fast page faults.
4101 ret
= ram_load_postcopy(f
, RAM_CHANNEL_PRECOPY
);
4103 ret
= ram_load_precopy(f
);
4106 trace_ram_load_complete(ret
, seq_iter
);
4111 static bool ram_has_postcopy(void *opaque
)
4114 RAMBLOCK_FOREACH_NOT_IGNORED(rb
) {
4115 if (ramblock_is_pmem(rb
)) {
4116 info_report("Block: %s, host: %p is a nvdimm memory, postcopy"
4117 "is not supported now!", rb
->idstr
, rb
->host
);
4122 return migrate_postcopy_ram();
4125 /* Sync all the dirty bitmap with destination VM. */
4126 static int ram_dirty_bitmap_sync_all(MigrationState
*s
, RAMState
*rs
)
4129 QEMUFile
*file
= s
->to_dst_file
;
4131 trace_ram_dirty_bitmap_sync_start();
4133 qatomic_set(&rs
->postcopy_bmap_sync_requested
, 0);
4134 RAMBLOCK_FOREACH_NOT_IGNORED(block
) {
4135 qemu_savevm_send_recv_bitmap(file
, block
->idstr
);
4136 trace_ram_dirty_bitmap_request(block
->idstr
);
4137 qatomic_inc(&rs
->postcopy_bmap_sync_requested
);
4140 trace_ram_dirty_bitmap_sync_wait();
4142 /* Wait until all the ramblocks' dirty bitmap synced */
4143 while (qatomic_read(&rs
->postcopy_bmap_sync_requested
)) {
4144 migration_rp_wait(s
);
4147 trace_ram_dirty_bitmap_sync_complete();
4153 * Read the received bitmap, revert it as the initial dirty bitmap.
4154 * This is only used when the postcopy migration is paused but wants
4155 * to resume from a middle point.
4157 int ram_dirty_bitmap_reload(MigrationState
*s
, RAMBlock
*block
)
4160 /* from_dst_file is always valid because we're within rp_thread */
4161 QEMUFile
*file
= s
->rp_state
.from_dst_file
;
4162 g_autofree
unsigned long *le_bitmap
= NULL
;
4163 unsigned long nbits
= block
->used_length
>> TARGET_PAGE_BITS
;
4164 uint64_t local_size
= DIV_ROUND_UP(nbits
, 8);
4165 uint64_t size
, end_mark
;
4166 RAMState
*rs
= ram_state
;
4168 trace_ram_dirty_bitmap_reload_begin(block
->idstr
);
4170 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
4171 error_report("%s: incorrect state %s", __func__
,
4172 MigrationStatus_str(s
->state
));
4177 * Note: see comments in ramblock_recv_bitmap_send() on why we
4178 * need the endianness conversion, and the paddings.
4180 local_size
= ROUND_UP(local_size
, 8);
4183 le_bitmap
= bitmap_new(nbits
+ BITS_PER_LONG
);
4185 size
= qemu_get_be64(file
);
4187 /* The size of the bitmap should match with our ramblock */
4188 if (size
!= local_size
) {
4189 error_report("%s: ramblock '%s' bitmap size mismatch "
4190 "(0x%"PRIx64
" != 0x%"PRIx64
")", __func__
,
4191 block
->idstr
, size
, local_size
);
4195 size
= qemu_get_buffer(file
, (uint8_t *)le_bitmap
, local_size
);
4196 end_mark
= qemu_get_be64(file
);
4198 ret
= qemu_file_get_error(file
);
4199 if (ret
|| size
!= local_size
) {
4200 error_report("%s: read bitmap failed for ramblock '%s': %d"
4201 " (size 0x%"PRIx64
", got: 0x%"PRIx64
")",
4202 __func__
, block
->idstr
, ret
, local_size
, size
);
4206 if (end_mark
!= RAMBLOCK_RECV_BITMAP_ENDING
) {
4207 error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64
,
4208 __func__
, block
->idstr
, end_mark
);
4213 * Endianness conversion. We are during postcopy (though paused).
4214 * The dirty bitmap won't change. We can directly modify it.
4216 bitmap_from_le(block
->bmap
, le_bitmap
, nbits
);
4219 * What we received is "received bitmap". Revert it as the initial
4220 * dirty bitmap for this ramblock.
4222 bitmap_complement(block
->bmap
, block
->bmap
, nbits
);
4224 /* Clear dirty bits of discarded ranges that we don't want to migrate. */
4225 ramblock_dirty_bitmap_clear_discarded_pages(block
);
4227 /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */
4228 trace_ram_dirty_bitmap_reload_complete(block
->idstr
);
4230 qatomic_dec(&rs
->postcopy_bmap_sync_requested
);
4233 * We succeeded to sync bitmap for current ramblock. Always kick the
4234 * migration thread to check whether all requested bitmaps are
4235 * reloaded. NOTE: it's racy to only kick when requested==0, because
4236 * we don't know whether the migration thread may still be increasing
4239 migration_rp_kick(s
);
4244 static int ram_resume_prepare(MigrationState
*s
, void *opaque
)
4246 RAMState
*rs
= *(RAMState
**)opaque
;
4249 ret
= ram_dirty_bitmap_sync_all(s
, rs
);
4254 ram_state_resume_prepare(rs
, s
->to_dst_file
);
4259 void postcopy_preempt_shutdown_file(MigrationState
*s
)
4261 qemu_put_be64(s
->postcopy_qemufile_src
, RAM_SAVE_FLAG_EOS
);
4262 qemu_fflush(s
->postcopy_qemufile_src
);
4265 static SaveVMHandlers savevm_ram_handlers
= {
4266 .save_setup
= ram_save_setup
,
4267 .save_live_iterate
= ram_save_iterate
,
4268 .save_live_complete_postcopy
= ram_save_complete
,
4269 .save_live_complete_precopy
= ram_save_complete
,
4270 .has_postcopy
= ram_has_postcopy
,
4271 .state_pending_exact
= ram_state_pending_exact
,
4272 .state_pending_estimate
= ram_state_pending_estimate
,
4273 .load_state
= ram_load
,
4274 .save_cleanup
= ram_save_cleanup
,
4275 .load_setup
= ram_load_setup
,
4276 .load_cleanup
= ram_load_cleanup
,
4277 .resume_prepare
= ram_resume_prepare
,
4280 static void ram_mig_ram_block_resized(RAMBlockNotifier
*n
, void *host
,
4281 size_t old_size
, size_t new_size
)
4283 PostcopyState ps
= postcopy_state_get();
4285 RAMBlock
*rb
= qemu_ram_block_from_host(host
, false, &offset
);
4289 error_report("RAM block not found");
4293 if (migrate_ram_is_ignored(rb
)) {
4297 if (!migration_is_idle()) {
4299 * Precopy code on the source cannot deal with the size of RAM blocks
4300 * changing at random points in time - especially after sending the
4301 * RAM block sizes in the migration stream, they must no longer change.
4302 * Abort and indicate a proper reason.
4304 error_setg(&err
, "RAM block '%s' resized during precopy.", rb
->idstr
);
4305 migration_cancel(err
);
4310 case POSTCOPY_INCOMING_ADVISE
:
4312 * Update what ram_postcopy_incoming_init()->init_range() does at the
4313 * time postcopy was advised. Syncing RAM blocks with the source will
4314 * result in RAM resizes.
4316 if (old_size
< new_size
) {
4317 if (ram_discard_range(rb
->idstr
, old_size
, new_size
- old_size
)) {
4318 error_report("RAM block '%s' discard of resized RAM failed",
4322 rb
->postcopy_length
= new_size
;
4324 case POSTCOPY_INCOMING_NONE
:
4325 case POSTCOPY_INCOMING_RUNNING
:
4326 case POSTCOPY_INCOMING_END
:
4328 * Once our guest is running, postcopy does no longer care about
4329 * resizes. When growing, the new memory was not available on the
4330 * source, no handler needed.
4334 error_report("RAM block '%s' resized during postcopy state: %d",
4340 static RAMBlockNotifier ram_mig_ram_notifier
= {
4341 .ram_block_resized
= ram_mig_ram_block_resized
,
4344 void ram_mig_init(void)
4346 qemu_mutex_init(&XBZRLE
.lock
);
4347 register_savevm_live("ram", 0, 4, &savevm_ram_handlers
, &ram_state
);
4348 ram_block_notifier_add(&ram_mig_ram_notifier
);