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"
32 #include "qemu/cutils.h"
33 #include "qemu/bitops.h"
34 #include "qemu/bitmap.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/pmem.h"
39 #include "migration.h"
41 #include "migration/register.h"
42 #include "migration/misc.h"
43 #include "qemu-file.h"
44 #include "postcopy-ram.h"
45 #include "page_cache.h"
46 #include "qemu/error-report.h"
47 #include "qapi/error.h"
48 #include "qapi/qapi-events-migration.h"
49 #include "qapi/qmp/qerror.h"
51 #include "exec/ram_addr.h"
52 #include "exec/target_page.h"
53 #include "qemu/rcu_queue.h"
54 #include "migration/colo.h"
56 #include "sysemu/sysemu.h"
57 #include "qemu/uuid.h"
61 /***********************************************************/
62 /* ram save/restore */
64 /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it
65 * worked for pages that where filled with the same char. We switched
66 * it to only search for the zero value. And to avoid confusion with
67 * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it.
70 #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */
71 #define RAM_SAVE_FLAG_ZERO 0x02
72 #define RAM_SAVE_FLAG_MEM_SIZE 0x04
73 #define RAM_SAVE_FLAG_PAGE 0x08
74 #define RAM_SAVE_FLAG_EOS 0x10
75 #define RAM_SAVE_FLAG_CONTINUE 0x20
76 #define RAM_SAVE_FLAG_XBZRLE 0x40
77 /* 0x80 is reserved in migration.h start with 0x100 next */
78 #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100
80 static inline bool is_zero_range(uint8_t *p
, uint64_t size
)
82 return buffer_is_zero(p
, size
);
85 XBZRLECacheStats xbzrle_counters
;
87 /* struct contains XBZRLE cache and a static page
88 used by the compression */
90 /* buffer used for XBZRLE encoding */
92 /* buffer for storing page content */
94 /* Cache for XBZRLE, Protected by lock. */
97 /* it will store a page full of zeros */
98 uint8_t *zero_target_page
;
99 /* buffer used for XBZRLE decoding */
100 uint8_t *decoded_buf
;
103 static void XBZRLE_cache_lock(void)
105 if (migrate_use_xbzrle())
106 qemu_mutex_lock(&XBZRLE
.lock
);
109 static void XBZRLE_cache_unlock(void)
111 if (migrate_use_xbzrle())
112 qemu_mutex_unlock(&XBZRLE
.lock
);
116 * xbzrle_cache_resize: resize the xbzrle cache
118 * This function is called from qmp_migrate_set_cache_size in main
119 * thread, possibly while a migration is in progress. A running
120 * migration may be using the cache and might finish during this call,
121 * hence changes to the cache are protected by XBZRLE.lock().
123 * Returns 0 for success or -1 for error
125 * @new_size: new cache size
126 * @errp: set *errp if the check failed, with reason
128 int xbzrle_cache_resize(int64_t new_size
, Error
**errp
)
130 PageCache
*new_cache
;
133 /* Check for truncation */
134 if (new_size
!= (size_t)new_size
) {
135 error_setg(errp
, QERR_INVALID_PARAMETER_VALUE
, "cache size",
136 "exceeding address space");
140 if (new_size
== migrate_xbzrle_cache_size()) {
147 if (XBZRLE
.cache
!= NULL
) {
148 new_cache
= cache_init(new_size
, TARGET_PAGE_SIZE
, errp
);
154 cache_fini(XBZRLE
.cache
);
155 XBZRLE
.cache
= new_cache
;
158 XBZRLE_cache_unlock();
162 /* Should be holding either ram_list.mutex, or the RCU lock. */
163 #define RAMBLOCK_FOREACH_MIGRATABLE(block) \
164 INTERNAL_RAMBLOCK_FOREACH(block) \
165 if (!qemu_ram_is_migratable(block)) {} else
167 #undef RAMBLOCK_FOREACH
169 static void ramblock_recv_map_init(void)
173 RAMBLOCK_FOREACH_MIGRATABLE(rb
) {
174 assert(!rb
->receivedmap
);
175 rb
->receivedmap
= bitmap_new(rb
->max_length
>> qemu_target_page_bits());
179 int ramblock_recv_bitmap_test(RAMBlock
*rb
, void *host_addr
)
181 return test_bit(ramblock_recv_bitmap_offset(host_addr
, rb
),
185 bool ramblock_recv_bitmap_test_byte_offset(RAMBlock
*rb
, uint64_t byte_offset
)
187 return test_bit(byte_offset
>> TARGET_PAGE_BITS
, rb
->receivedmap
);
190 void ramblock_recv_bitmap_set(RAMBlock
*rb
, void *host_addr
)
192 set_bit_atomic(ramblock_recv_bitmap_offset(host_addr
, rb
), rb
->receivedmap
);
195 void ramblock_recv_bitmap_set_range(RAMBlock
*rb
, void *host_addr
,
198 bitmap_set_atomic(rb
->receivedmap
,
199 ramblock_recv_bitmap_offset(host_addr
, rb
),
203 #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL)
206 * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes).
208 * Returns >0 if success with sent bytes, or <0 if error.
210 int64_t ramblock_recv_bitmap_send(QEMUFile
*file
,
211 const char *block_name
)
213 RAMBlock
*block
= qemu_ram_block_by_name(block_name
);
214 unsigned long *le_bitmap
, nbits
;
218 error_report("%s: invalid block name: %s", __func__
, block_name
);
222 nbits
= block
->used_length
>> TARGET_PAGE_BITS
;
225 * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit
226 * machines we may need 4 more bytes for padding (see below
227 * comment). So extend it a bit before hand.
229 le_bitmap
= bitmap_new(nbits
+ BITS_PER_LONG
);
232 * Always use little endian when sending the bitmap. This is
233 * required that when source and destination VMs are not using the
234 * same endianess. (Note: big endian won't work.)
236 bitmap_to_le(le_bitmap
, block
->receivedmap
, nbits
);
238 /* Size of the bitmap, in bytes */
239 size
= DIV_ROUND_UP(nbits
, 8);
242 * size is always aligned to 8 bytes for 64bit machines, but it
243 * may not be true for 32bit machines. We need this padding to
244 * make sure the migration can survive even between 32bit and
247 size
= ROUND_UP(size
, 8);
249 qemu_put_be64(file
, size
);
250 qemu_put_buffer(file
, (const uint8_t *)le_bitmap
, size
);
252 * Mark as an end, in case the middle part is screwed up due to
253 * some "misterious" reason.
255 qemu_put_be64(file
, RAMBLOCK_RECV_BITMAP_ENDING
);
260 if (qemu_file_get_error(file
)) {
261 return qemu_file_get_error(file
);
264 return size
+ sizeof(size
);
268 * An outstanding page request, on the source, having been received
271 struct RAMSrcPageRequest
{
276 QSIMPLEQ_ENTRY(RAMSrcPageRequest
) next_req
;
279 /* State of RAM for migration */
281 /* QEMUFile used for this migration */
283 /* Last block that we have visited searching for dirty pages */
284 RAMBlock
*last_seen_block
;
285 /* Last block from where we have sent data */
286 RAMBlock
*last_sent_block
;
287 /* Last dirty target page we have sent */
288 ram_addr_t last_page
;
289 /* last ram version we have seen */
290 uint32_t last_version
;
291 /* We are in the first round */
293 /* How many times we have dirty too many pages */
294 int dirty_rate_high_cnt
;
295 /* these variables are used for bitmap sync */
296 /* last time we did a full bitmap_sync */
297 int64_t time_last_bitmap_sync
;
298 /* bytes transferred at start_time */
299 uint64_t bytes_xfer_prev
;
300 /* number of dirty pages since start_time */
301 uint64_t num_dirty_pages_period
;
302 /* xbzrle misses since the beginning of the period */
303 uint64_t xbzrle_cache_miss_prev
;
305 /* compression statistics since the beginning of the period */
306 /* amount of count that no free thread to compress data */
307 uint64_t compress_thread_busy_prev
;
308 /* amount bytes after compression */
309 uint64_t compressed_size_prev
;
310 /* amount of compressed pages */
311 uint64_t compress_pages_prev
;
313 /* total handled target pages at the beginning of period */
314 uint64_t target_page_count_prev
;
315 /* total handled target pages since start */
316 uint64_t target_page_count
;
317 /* number of dirty bits in the bitmap */
318 uint64_t migration_dirty_pages
;
319 /* protects modification of the bitmap */
320 QemuMutex bitmap_mutex
;
321 /* The RAMBlock used in the last src_page_requests */
322 RAMBlock
*last_req_rb
;
323 /* Queue of outstanding page requests from the destination */
324 QemuMutex src_page_req_mutex
;
325 QSIMPLEQ_HEAD(src_page_requests
, RAMSrcPageRequest
) src_page_requests
;
327 typedef struct RAMState RAMState
;
329 static RAMState
*ram_state
;
331 uint64_t ram_bytes_remaining(void)
333 return ram_state
? (ram_state
->migration_dirty_pages
* TARGET_PAGE_SIZE
) :
337 MigrationStats ram_counters
;
339 /* used by the search for pages to send */
340 struct PageSearchStatus
{
341 /* Current block being searched */
343 /* Current page to search from */
345 /* Set once we wrap around */
348 typedef struct PageSearchStatus PageSearchStatus
;
350 CompressionStats compression_counters
;
352 struct CompressParam
{
362 /* internally used fields */
366 typedef struct CompressParam CompressParam
;
368 struct DecompressParam
{
378 typedef struct DecompressParam DecompressParam
;
380 static CompressParam
*comp_param
;
381 static QemuThread
*compress_threads
;
382 /* comp_done_cond is used to wake up the migration thread when
383 * one of the compression threads has finished the compression.
384 * comp_done_lock is used to co-work with comp_done_cond.
386 static QemuMutex comp_done_lock
;
387 static QemuCond comp_done_cond
;
388 /* The empty QEMUFileOps will be used by file in CompressParam */
389 static const QEMUFileOps empty_ops
= { };
391 static QEMUFile
*decomp_file
;
392 static DecompressParam
*decomp_param
;
393 static QemuThread
*decompress_threads
;
394 static QemuMutex decomp_done_lock
;
395 static QemuCond decomp_done_cond
;
397 static bool do_compress_ram_page(QEMUFile
*f
, z_stream
*stream
, RAMBlock
*block
,
398 ram_addr_t offset
, uint8_t *source_buf
);
400 static void *do_data_compress(void *opaque
)
402 CompressParam
*param
= opaque
;
407 qemu_mutex_lock(¶m
->mutex
);
408 while (!param
->quit
) {
410 block
= param
->block
;
411 offset
= param
->offset
;
413 qemu_mutex_unlock(¶m
->mutex
);
415 zero_page
= do_compress_ram_page(param
->file
, ¶m
->stream
,
416 block
, offset
, param
->originbuf
);
418 qemu_mutex_lock(&comp_done_lock
);
420 param
->zero_page
= zero_page
;
421 qemu_cond_signal(&comp_done_cond
);
422 qemu_mutex_unlock(&comp_done_lock
);
424 qemu_mutex_lock(¶m
->mutex
);
426 qemu_cond_wait(¶m
->cond
, ¶m
->mutex
);
429 qemu_mutex_unlock(¶m
->mutex
);
434 static void compress_threads_save_cleanup(void)
438 if (!migrate_use_compression() || !comp_param
) {
442 thread_count
= migrate_compress_threads();
443 for (i
= 0; i
< thread_count
; i
++) {
445 * we use it as a indicator which shows if the thread is
446 * properly init'd or not
448 if (!comp_param
[i
].file
) {
452 qemu_mutex_lock(&comp_param
[i
].mutex
);
453 comp_param
[i
].quit
= true;
454 qemu_cond_signal(&comp_param
[i
].cond
);
455 qemu_mutex_unlock(&comp_param
[i
].mutex
);
457 qemu_thread_join(compress_threads
+ i
);
458 qemu_mutex_destroy(&comp_param
[i
].mutex
);
459 qemu_cond_destroy(&comp_param
[i
].cond
);
460 deflateEnd(&comp_param
[i
].stream
);
461 g_free(comp_param
[i
].originbuf
);
462 qemu_fclose(comp_param
[i
].file
);
463 comp_param
[i
].file
= NULL
;
465 qemu_mutex_destroy(&comp_done_lock
);
466 qemu_cond_destroy(&comp_done_cond
);
467 g_free(compress_threads
);
469 compress_threads
= NULL
;
473 static int compress_threads_save_setup(void)
477 if (!migrate_use_compression()) {
480 thread_count
= migrate_compress_threads();
481 compress_threads
= g_new0(QemuThread
, thread_count
);
482 comp_param
= g_new0(CompressParam
, thread_count
);
483 qemu_cond_init(&comp_done_cond
);
484 qemu_mutex_init(&comp_done_lock
);
485 for (i
= 0; i
< thread_count
; i
++) {
486 comp_param
[i
].originbuf
= g_try_malloc(TARGET_PAGE_SIZE
);
487 if (!comp_param
[i
].originbuf
) {
491 if (deflateInit(&comp_param
[i
].stream
,
492 migrate_compress_level()) != Z_OK
) {
493 g_free(comp_param
[i
].originbuf
);
497 /* comp_param[i].file is just used as a dummy buffer to save data,
498 * set its ops to empty.
500 comp_param
[i
].file
= qemu_fopen_ops(NULL
, &empty_ops
);
501 comp_param
[i
].done
= true;
502 comp_param
[i
].quit
= false;
503 qemu_mutex_init(&comp_param
[i
].mutex
);
504 qemu_cond_init(&comp_param
[i
].cond
);
505 qemu_thread_create(compress_threads
+ i
, "compress",
506 do_data_compress
, comp_param
+ i
,
507 QEMU_THREAD_JOINABLE
);
512 compress_threads_save_cleanup();
518 #define MULTIFD_MAGIC 0x11223344U
519 #define MULTIFD_VERSION 1
521 #define MULTIFD_FLAG_SYNC (1 << 0)
526 unsigned char uuid
[16]; /* QemuUUID */
528 } __attribute__((packed
)) MultiFDInit_t
;
539 } __attribute__((packed
)) MultiFDPacket_t
;
542 /* number of used pages */
544 /* number of allocated pages */
546 /* global number of generated multifd packets */
548 /* offset of each page */
550 /* pointer to each page */
556 /* this fields are not changed once the thread is created */
559 /* channel thread name */
561 /* channel thread id */
563 /* communication channel */
565 /* sem where to wait for more work */
567 /* this mutex protects the following parameters */
569 /* is this channel thread running */
571 /* should this thread finish */
573 /* thread has work to do */
575 /* array of pages to sent */
576 MultiFDPages_t
*pages
;
577 /* packet allocated len */
579 /* pointer to the packet */
580 MultiFDPacket_t
*packet
;
581 /* multifd flags for each packet */
583 /* global number of generated multifd packets */
585 /* thread local variables */
586 /* packets sent through this channel */
587 uint64_t num_packets
;
588 /* pages sent through this channel */
590 /* syncs main thread and channels */
591 QemuSemaphore sem_sync
;
595 /* this fields are not changed once the thread is created */
598 /* channel thread name */
600 /* channel thread id */
602 /* communication channel */
604 /* this mutex protects the following parameters */
606 /* is this channel thread running */
608 /* array of pages to receive */
609 MultiFDPages_t
*pages
;
610 /* packet allocated len */
612 /* pointer to the packet */
613 MultiFDPacket_t
*packet
;
614 /* multifd flags for each packet */
616 /* global number of generated multifd packets */
618 /* thread local variables */
619 /* packets sent through this channel */
620 uint64_t num_packets
;
621 /* pages sent through this channel */
623 /* syncs main thread and channels */
624 QemuSemaphore sem_sync
;
627 static int multifd_send_initial_packet(MultiFDSendParams
*p
, Error
**errp
)
632 msg
.magic
= cpu_to_be32(MULTIFD_MAGIC
);
633 msg
.version
= cpu_to_be32(MULTIFD_VERSION
);
635 memcpy(msg
.uuid
, &qemu_uuid
.data
, sizeof(msg
.uuid
));
637 ret
= qio_channel_write_all(p
->c
, (char *)&msg
, sizeof(msg
), errp
);
644 static int multifd_recv_initial_packet(QIOChannel
*c
, Error
**errp
)
649 ret
= qio_channel_read_all(c
, (char *)&msg
, sizeof(msg
), errp
);
654 msg
.magic
= be32_to_cpu(msg
.magic
);
655 msg
.version
= be32_to_cpu(msg
.version
);
657 if (msg
.magic
!= MULTIFD_MAGIC
) {
658 error_setg(errp
, "multifd: received packet magic %x "
659 "expected %x", msg
.magic
, MULTIFD_MAGIC
);
663 if (msg
.version
!= MULTIFD_VERSION
) {
664 error_setg(errp
, "multifd: received packet version %d "
665 "expected %d", msg
.version
, MULTIFD_VERSION
);
669 if (memcmp(msg
.uuid
, &qemu_uuid
, sizeof(qemu_uuid
))) {
670 char *uuid
= qemu_uuid_unparse_strdup(&qemu_uuid
);
671 char *msg_uuid
= qemu_uuid_unparse_strdup((const QemuUUID
*)msg
.uuid
);
673 error_setg(errp
, "multifd: received uuid '%s' and expected "
674 "uuid '%s' for channel %hhd", msg_uuid
, uuid
, msg
.id
);
680 if (msg
.id
> migrate_multifd_channels()) {
681 error_setg(errp
, "multifd: received channel version %d "
682 "expected %d", msg
.version
, MULTIFD_VERSION
);
689 static MultiFDPages_t
*multifd_pages_init(size_t size
)
691 MultiFDPages_t
*pages
= g_new0(MultiFDPages_t
, 1);
693 pages
->allocated
= size
;
694 pages
->iov
= g_new0(struct iovec
, size
);
695 pages
->offset
= g_new0(ram_addr_t
, size
);
700 static void multifd_pages_clear(MultiFDPages_t
*pages
)
703 pages
->allocated
= 0;
704 pages
->packet_num
= 0;
708 g_free(pages
->offset
);
709 pages
->offset
= NULL
;
713 static void multifd_send_fill_packet(MultiFDSendParams
*p
)
715 MultiFDPacket_t
*packet
= p
->packet
;
718 packet
->magic
= cpu_to_be32(MULTIFD_MAGIC
);
719 packet
->version
= cpu_to_be32(MULTIFD_VERSION
);
720 packet
->flags
= cpu_to_be32(p
->flags
);
721 packet
->size
= cpu_to_be32(migrate_multifd_page_count());
722 packet
->used
= cpu_to_be32(p
->pages
->used
);
723 packet
->packet_num
= cpu_to_be64(p
->packet_num
);
725 if (p
->pages
->block
) {
726 strncpy(packet
->ramblock
, p
->pages
->block
->idstr
, 256);
729 for (i
= 0; i
< p
->pages
->used
; i
++) {
730 packet
->offset
[i
] = cpu_to_be64(p
->pages
->offset
[i
]);
734 static int multifd_recv_unfill_packet(MultiFDRecvParams
*p
, Error
**errp
)
736 MultiFDPacket_t
*packet
= p
->packet
;
740 packet
->magic
= be32_to_cpu(packet
->magic
);
741 if (packet
->magic
!= MULTIFD_MAGIC
) {
742 error_setg(errp
, "multifd: received packet "
743 "magic %x and expected magic %x",
744 packet
->magic
, MULTIFD_MAGIC
);
748 packet
->version
= be32_to_cpu(packet
->version
);
749 if (packet
->version
!= MULTIFD_VERSION
) {
750 error_setg(errp
, "multifd: received packet "
751 "version %d and expected version %d",
752 packet
->version
, MULTIFD_VERSION
);
756 p
->flags
= be32_to_cpu(packet
->flags
);
758 packet
->size
= be32_to_cpu(packet
->size
);
759 if (packet
->size
> migrate_multifd_page_count()) {
760 error_setg(errp
, "multifd: received packet "
761 "with size %d and expected maximum size %d",
762 packet
->size
, migrate_multifd_page_count()) ;
766 p
->pages
->used
= be32_to_cpu(packet
->used
);
767 if (p
->pages
->used
> packet
->size
) {
768 error_setg(errp
, "multifd: received packet "
769 "with size %d and expected maximum size %d",
770 p
->pages
->used
, packet
->size
) ;
774 p
->packet_num
= be64_to_cpu(packet
->packet_num
);
776 if (p
->pages
->used
) {
777 /* make sure that ramblock is 0 terminated */
778 packet
->ramblock
[255] = 0;
779 block
= qemu_ram_block_by_name(packet
->ramblock
);
781 error_setg(errp
, "multifd: unknown ram block %s",
787 for (i
= 0; i
< p
->pages
->used
; i
++) {
788 ram_addr_t offset
= be64_to_cpu(packet
->offset
[i
]);
790 if (offset
> (block
->used_length
- TARGET_PAGE_SIZE
)) {
791 error_setg(errp
, "multifd: offset too long " RAM_ADDR_FMT
792 " (max " RAM_ADDR_FMT
")",
793 offset
, block
->max_length
);
796 p
->pages
->iov
[i
].iov_base
= block
->host
+ offset
;
797 p
->pages
->iov
[i
].iov_len
= TARGET_PAGE_SIZE
;
804 MultiFDSendParams
*params
;
805 /* number of created threads */
807 /* array of pages to sent */
808 MultiFDPages_t
*pages
;
809 /* syncs main thread and channels */
810 QemuSemaphore sem_sync
;
811 /* global number of generated multifd packets */
813 /* send channels ready */
814 QemuSemaphore channels_ready
;
815 } *multifd_send_state
;
818 * How we use multifd_send_state->pages and channel->pages?
820 * We create a pages for each channel, and a main one. Each time that
821 * we need to send a batch of pages we interchange the ones between
822 * multifd_send_state and the channel that is sending it. There are
823 * two reasons for that:
824 * - to not have to do so many mallocs during migration
825 * - to make easier to know what to free at the end of migration
827 * This way we always know who is the owner of each "pages" struct,
828 * and we don't need any loocking. It belongs to the migration thread
829 * or to the channel thread. Switching is safe because the migration
830 * thread is using the channel mutex when changing it, and the channel
831 * have to had finish with its own, otherwise pending_job can't be
835 static void multifd_send_pages(void)
838 static int next_channel
;
839 MultiFDSendParams
*p
= NULL
; /* make happy gcc */
840 MultiFDPages_t
*pages
= multifd_send_state
->pages
;
841 uint64_t transferred
;
843 qemu_sem_wait(&multifd_send_state
->channels_ready
);
844 for (i
= next_channel
;; i
= (i
+ 1) % migrate_multifd_channels()) {
845 p
= &multifd_send_state
->params
[i
];
847 qemu_mutex_lock(&p
->mutex
);
848 if (!p
->pending_job
) {
850 next_channel
= (i
+ 1) % migrate_multifd_channels();
853 qemu_mutex_unlock(&p
->mutex
);
857 p
->packet_num
= multifd_send_state
->packet_num
++;
858 p
->pages
->block
= NULL
;
859 multifd_send_state
->pages
= p
->pages
;
861 transferred
= ((uint64_t) pages
->used
) * TARGET_PAGE_SIZE
+ p
->packet_len
;
862 ram_counters
.multifd_bytes
+= transferred
;
863 ram_counters
.transferred
+= transferred
;;
864 qemu_mutex_unlock(&p
->mutex
);
865 qemu_sem_post(&p
->sem
);
868 static void multifd_queue_page(RAMBlock
*block
, ram_addr_t offset
)
870 MultiFDPages_t
*pages
= multifd_send_state
->pages
;
873 pages
->block
= block
;
876 if (pages
->block
== block
) {
877 pages
->offset
[pages
->used
] = offset
;
878 pages
->iov
[pages
->used
].iov_base
= block
->host
+ offset
;
879 pages
->iov
[pages
->used
].iov_len
= TARGET_PAGE_SIZE
;
882 if (pages
->used
< pages
->allocated
) {
887 multifd_send_pages();
889 if (pages
->block
!= block
) {
890 multifd_queue_page(block
, offset
);
894 static void multifd_send_terminate_threads(Error
*err
)
899 MigrationState
*s
= migrate_get_current();
900 migrate_set_error(s
, err
);
901 if (s
->state
== MIGRATION_STATUS_SETUP
||
902 s
->state
== MIGRATION_STATUS_PRE_SWITCHOVER
||
903 s
->state
== MIGRATION_STATUS_DEVICE
||
904 s
->state
== MIGRATION_STATUS_ACTIVE
) {
905 migrate_set_state(&s
->state
, s
->state
,
906 MIGRATION_STATUS_FAILED
);
910 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
911 MultiFDSendParams
*p
= &multifd_send_state
->params
[i
];
913 qemu_mutex_lock(&p
->mutex
);
915 qemu_sem_post(&p
->sem
);
916 qemu_mutex_unlock(&p
->mutex
);
920 int multifd_save_cleanup(Error
**errp
)
925 if (!migrate_use_multifd()) {
928 multifd_send_terminate_threads(NULL
);
929 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
930 MultiFDSendParams
*p
= &multifd_send_state
->params
[i
];
933 qemu_thread_join(&p
->thread
);
935 socket_send_channel_destroy(p
->c
);
937 qemu_mutex_destroy(&p
->mutex
);
938 qemu_sem_destroy(&p
->sem
);
939 qemu_sem_destroy(&p
->sem_sync
);
942 multifd_pages_clear(p
->pages
);
948 qemu_sem_destroy(&multifd_send_state
->channels_ready
);
949 qemu_sem_destroy(&multifd_send_state
->sem_sync
);
950 g_free(multifd_send_state
->params
);
951 multifd_send_state
->params
= NULL
;
952 multifd_pages_clear(multifd_send_state
->pages
);
953 multifd_send_state
->pages
= NULL
;
954 g_free(multifd_send_state
);
955 multifd_send_state
= NULL
;
959 static void multifd_send_sync_main(void)
963 if (!migrate_use_multifd()) {
966 if (multifd_send_state
->pages
->used
) {
967 multifd_send_pages();
969 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
970 MultiFDSendParams
*p
= &multifd_send_state
->params
[i
];
972 trace_multifd_send_sync_main_signal(p
->id
);
974 qemu_mutex_lock(&p
->mutex
);
976 p
->packet_num
= multifd_send_state
->packet_num
++;
977 p
->flags
|= MULTIFD_FLAG_SYNC
;
979 qemu_mutex_unlock(&p
->mutex
);
980 qemu_sem_post(&p
->sem
);
982 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
983 MultiFDSendParams
*p
= &multifd_send_state
->params
[i
];
985 trace_multifd_send_sync_main_wait(p
->id
);
986 qemu_sem_wait(&multifd_send_state
->sem_sync
);
988 trace_multifd_send_sync_main(multifd_send_state
->packet_num
);
991 static void *multifd_send_thread(void *opaque
)
993 MultiFDSendParams
*p
= opaque
;
994 Error
*local_err
= NULL
;
997 trace_multifd_send_thread_start(p
->id
);
998 rcu_register_thread();
1000 if (multifd_send_initial_packet(p
, &local_err
) < 0) {
1003 /* initial packet */
1007 qemu_sem_wait(&p
->sem
);
1008 qemu_mutex_lock(&p
->mutex
);
1010 if (p
->pending_job
) {
1011 uint32_t used
= p
->pages
->used
;
1012 uint64_t packet_num
= p
->packet_num
;
1013 uint32_t flags
= p
->flags
;
1015 multifd_send_fill_packet(p
);
1018 p
->num_pages
+= used
;
1020 qemu_mutex_unlock(&p
->mutex
);
1022 trace_multifd_send(p
->id
, packet_num
, used
, flags
);
1024 ret
= qio_channel_write_all(p
->c
, (void *)p
->packet
,
1025 p
->packet_len
, &local_err
);
1030 ret
= qio_channel_writev_all(p
->c
, p
->pages
->iov
, used
, &local_err
);
1035 qemu_mutex_lock(&p
->mutex
);
1037 qemu_mutex_unlock(&p
->mutex
);
1039 if (flags
& MULTIFD_FLAG_SYNC
) {
1040 qemu_sem_post(&multifd_send_state
->sem_sync
);
1042 qemu_sem_post(&multifd_send_state
->channels_ready
);
1043 } else if (p
->quit
) {
1044 qemu_mutex_unlock(&p
->mutex
);
1047 qemu_mutex_unlock(&p
->mutex
);
1048 /* sometimes there are spurious wakeups */
1054 multifd_send_terminate_threads(local_err
);
1057 qemu_mutex_lock(&p
->mutex
);
1059 qemu_mutex_unlock(&p
->mutex
);
1061 rcu_unregister_thread();
1062 trace_multifd_send_thread_end(p
->id
, p
->num_packets
, p
->num_pages
);
1067 static void multifd_new_send_channel_async(QIOTask
*task
, gpointer opaque
)
1069 MultiFDSendParams
*p
= opaque
;
1070 QIOChannel
*sioc
= QIO_CHANNEL(qio_task_get_source(task
));
1071 Error
*local_err
= NULL
;
1073 if (qio_task_propagate_error(task
, &local_err
)) {
1074 if (multifd_save_cleanup(&local_err
) != 0) {
1075 migrate_set_error(migrate_get_current(), local_err
);
1078 p
->c
= QIO_CHANNEL(sioc
);
1079 qio_channel_set_delay(p
->c
, false);
1081 qemu_thread_create(&p
->thread
, p
->name
, multifd_send_thread
, p
,
1082 QEMU_THREAD_JOINABLE
);
1084 atomic_inc(&multifd_send_state
->count
);
1088 int multifd_save_setup(void)
1091 uint32_t page_count
= migrate_multifd_page_count();
1094 if (!migrate_use_multifd()) {
1097 thread_count
= migrate_multifd_channels();
1098 multifd_send_state
= g_malloc0(sizeof(*multifd_send_state
));
1099 multifd_send_state
->params
= g_new0(MultiFDSendParams
, thread_count
);
1100 atomic_set(&multifd_send_state
->count
, 0);
1101 multifd_send_state
->pages
= multifd_pages_init(page_count
);
1102 qemu_sem_init(&multifd_send_state
->sem_sync
, 0);
1103 qemu_sem_init(&multifd_send_state
->channels_ready
, 0);
1105 for (i
= 0; i
< thread_count
; i
++) {
1106 MultiFDSendParams
*p
= &multifd_send_state
->params
[i
];
1108 qemu_mutex_init(&p
->mutex
);
1109 qemu_sem_init(&p
->sem
, 0);
1110 qemu_sem_init(&p
->sem_sync
, 0);
1114 p
->pages
= multifd_pages_init(page_count
);
1115 p
->packet_len
= sizeof(MultiFDPacket_t
)
1116 + sizeof(ram_addr_t
) * page_count
;
1117 p
->packet
= g_malloc0(p
->packet_len
);
1118 p
->name
= g_strdup_printf("multifdsend_%d", i
);
1119 socket_send_channel_create(multifd_new_send_channel_async
, p
);
1125 MultiFDRecvParams
*params
;
1126 /* number of created threads */
1128 /* syncs main thread and channels */
1129 QemuSemaphore sem_sync
;
1130 /* global number of generated multifd packets */
1131 uint64_t packet_num
;
1132 } *multifd_recv_state
;
1134 static void multifd_recv_terminate_threads(Error
*err
)
1139 MigrationState
*s
= migrate_get_current();
1140 migrate_set_error(s
, err
);
1141 if (s
->state
== MIGRATION_STATUS_SETUP
||
1142 s
->state
== MIGRATION_STATUS_ACTIVE
) {
1143 migrate_set_state(&s
->state
, s
->state
,
1144 MIGRATION_STATUS_FAILED
);
1148 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
1149 MultiFDRecvParams
*p
= &multifd_recv_state
->params
[i
];
1151 qemu_mutex_lock(&p
->mutex
);
1152 /* We could arrive here for two reasons:
1153 - normal quit, i.e. everything went fine, just finished
1154 - error quit: We close the channels so the channel threads
1155 finish the qio_channel_read_all_eof() */
1156 qio_channel_shutdown(p
->c
, QIO_CHANNEL_SHUTDOWN_BOTH
, NULL
);
1157 qemu_mutex_unlock(&p
->mutex
);
1161 int multifd_load_cleanup(Error
**errp
)
1166 if (!migrate_use_multifd()) {
1169 multifd_recv_terminate_threads(NULL
);
1170 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
1171 MultiFDRecvParams
*p
= &multifd_recv_state
->params
[i
];
1174 qemu_thread_join(&p
->thread
);
1176 object_unref(OBJECT(p
->c
));
1178 qemu_mutex_destroy(&p
->mutex
);
1179 qemu_sem_destroy(&p
->sem_sync
);
1182 multifd_pages_clear(p
->pages
);
1188 qemu_sem_destroy(&multifd_recv_state
->sem_sync
);
1189 g_free(multifd_recv_state
->params
);
1190 multifd_recv_state
->params
= NULL
;
1191 g_free(multifd_recv_state
);
1192 multifd_recv_state
= NULL
;
1197 static void multifd_recv_sync_main(void)
1201 if (!migrate_use_multifd()) {
1204 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
1205 MultiFDRecvParams
*p
= &multifd_recv_state
->params
[i
];
1207 trace_multifd_recv_sync_main_wait(p
->id
);
1208 qemu_sem_wait(&multifd_recv_state
->sem_sync
);
1209 qemu_mutex_lock(&p
->mutex
);
1210 if (multifd_recv_state
->packet_num
< p
->packet_num
) {
1211 multifd_recv_state
->packet_num
= p
->packet_num
;
1213 qemu_mutex_unlock(&p
->mutex
);
1215 for (i
= 0; i
< migrate_multifd_channels(); i
++) {
1216 MultiFDRecvParams
*p
= &multifd_recv_state
->params
[i
];
1218 trace_multifd_recv_sync_main_signal(p
->id
);
1219 qemu_sem_post(&p
->sem_sync
);
1221 trace_multifd_recv_sync_main(multifd_recv_state
->packet_num
);
1224 static void *multifd_recv_thread(void *opaque
)
1226 MultiFDRecvParams
*p
= opaque
;
1227 Error
*local_err
= NULL
;
1230 trace_multifd_recv_thread_start(p
->id
);
1231 rcu_register_thread();
1237 ret
= qio_channel_read_all_eof(p
->c
, (void *)p
->packet
,
1238 p
->packet_len
, &local_err
);
1239 if (ret
== 0) { /* EOF */
1242 if (ret
== -1) { /* Error */
1246 qemu_mutex_lock(&p
->mutex
);
1247 ret
= multifd_recv_unfill_packet(p
, &local_err
);
1249 qemu_mutex_unlock(&p
->mutex
);
1253 used
= p
->pages
->used
;
1255 trace_multifd_recv(p
->id
, p
->packet_num
, used
, flags
);
1257 p
->num_pages
+= used
;
1258 qemu_mutex_unlock(&p
->mutex
);
1260 ret
= qio_channel_readv_all(p
->c
, p
->pages
->iov
, used
, &local_err
);
1265 if (flags
& MULTIFD_FLAG_SYNC
) {
1266 qemu_sem_post(&multifd_recv_state
->sem_sync
);
1267 qemu_sem_wait(&p
->sem_sync
);
1272 multifd_recv_terminate_threads(local_err
);
1274 qemu_mutex_lock(&p
->mutex
);
1276 qemu_mutex_unlock(&p
->mutex
);
1278 rcu_unregister_thread();
1279 trace_multifd_recv_thread_end(p
->id
, p
->num_packets
, p
->num_pages
);
1284 int multifd_load_setup(void)
1287 uint32_t page_count
= migrate_multifd_page_count();
1290 if (!migrate_use_multifd()) {
1293 thread_count
= migrate_multifd_channels();
1294 multifd_recv_state
= g_malloc0(sizeof(*multifd_recv_state
));
1295 multifd_recv_state
->params
= g_new0(MultiFDRecvParams
, thread_count
);
1296 atomic_set(&multifd_recv_state
->count
, 0);
1297 qemu_sem_init(&multifd_recv_state
->sem_sync
, 0);
1299 for (i
= 0; i
< thread_count
; i
++) {
1300 MultiFDRecvParams
*p
= &multifd_recv_state
->params
[i
];
1302 qemu_mutex_init(&p
->mutex
);
1303 qemu_sem_init(&p
->sem_sync
, 0);
1305 p
->pages
= multifd_pages_init(page_count
);
1306 p
->packet_len
= sizeof(MultiFDPacket_t
)
1307 + sizeof(ram_addr_t
) * page_count
;
1308 p
->packet
= g_malloc0(p
->packet_len
);
1309 p
->name
= g_strdup_printf("multifdrecv_%d", i
);
1314 bool multifd_recv_all_channels_created(void)
1316 int thread_count
= migrate_multifd_channels();
1318 if (!migrate_use_multifd()) {
1322 return thread_count
== atomic_read(&multifd_recv_state
->count
);
1325 /* Return true if multifd is ready for the migration, otherwise false */
1326 bool multifd_recv_new_channel(QIOChannel
*ioc
)
1328 MultiFDRecvParams
*p
;
1329 Error
*local_err
= NULL
;
1332 id
= multifd_recv_initial_packet(ioc
, &local_err
);
1334 multifd_recv_terminate_threads(local_err
);
1338 p
= &multifd_recv_state
->params
[id
];
1340 error_setg(&local_err
, "multifd: received id '%d' already setup'",
1342 multifd_recv_terminate_threads(local_err
);
1346 object_ref(OBJECT(ioc
));
1347 /* initial packet */
1351 qemu_thread_create(&p
->thread
, p
->name
, multifd_recv_thread
, p
,
1352 QEMU_THREAD_JOINABLE
);
1353 atomic_inc(&multifd_recv_state
->count
);
1354 return multifd_recv_state
->count
== migrate_multifd_channels();
1358 * save_page_header: write page header to wire
1360 * If this is the 1st block, it also writes the block identification
1362 * Returns the number of bytes written
1364 * @f: QEMUFile where to send the data
1365 * @block: block that contains the page we want to send
1366 * @offset: offset inside the block for the page
1367 * in the lower bits, it contains flags
1369 static size_t save_page_header(RAMState
*rs
, QEMUFile
*f
, RAMBlock
*block
,
1374 if (block
== rs
->last_sent_block
) {
1375 offset
|= RAM_SAVE_FLAG_CONTINUE
;
1377 qemu_put_be64(f
, offset
);
1380 if (!(offset
& RAM_SAVE_FLAG_CONTINUE
)) {
1381 len
= strlen(block
->idstr
);
1382 qemu_put_byte(f
, len
);
1383 qemu_put_buffer(f
, (uint8_t *)block
->idstr
, len
);
1385 rs
->last_sent_block
= block
;
1391 * mig_throttle_guest_down: throotle down the guest
1393 * Reduce amount of guest cpu execution to hopefully slow down memory
1394 * writes. If guest dirty memory rate is reduced below the rate at
1395 * which we can transfer pages to the destination then we should be
1396 * able to complete migration. Some workloads dirty memory way too
1397 * fast and will not effectively converge, even with auto-converge.
1399 static void mig_throttle_guest_down(void)
1401 MigrationState
*s
= migrate_get_current();
1402 uint64_t pct_initial
= s
->parameters
.cpu_throttle_initial
;
1403 uint64_t pct_icrement
= s
->parameters
.cpu_throttle_increment
;
1404 int pct_max
= s
->parameters
.max_cpu_throttle
;
1406 /* We have not started throttling yet. Let's start it. */
1407 if (!cpu_throttle_active()) {
1408 cpu_throttle_set(pct_initial
);
1410 /* Throttling already on, just increase the rate */
1411 cpu_throttle_set(MIN(cpu_throttle_get_percentage() + pct_icrement
,
1417 * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache
1419 * @rs: current RAM state
1420 * @current_addr: address for the zero page
1422 * Update the xbzrle cache to reflect a page that's been sent as all 0.
1423 * The important thing is that a stale (not-yet-0'd) page be replaced
1425 * As a bonus, if the page wasn't in the cache it gets added so that
1426 * when a small write is made into the 0'd page it gets XBZRLE sent.
1428 static void xbzrle_cache_zero_page(RAMState
*rs
, ram_addr_t current_addr
)
1430 if (rs
->ram_bulk_stage
|| !migrate_use_xbzrle()) {
1434 /* We don't care if this fails to allocate a new cache page
1435 * as long as it updated an old one */
1436 cache_insert(XBZRLE
.cache
, current_addr
, XBZRLE
.zero_target_page
,
1437 ram_counters
.dirty_sync_count
);
1440 #define ENCODING_FLAG_XBZRLE 0x1
1443 * save_xbzrle_page: compress and send current page
1445 * Returns: 1 means that we wrote the page
1446 * 0 means that page is identical to the one already sent
1447 * -1 means that xbzrle would be longer than normal
1449 * @rs: current RAM state
1450 * @current_data: pointer to the address of the page contents
1451 * @current_addr: addr of the page
1452 * @block: block that contains the page we want to send
1453 * @offset: offset inside the block for the page
1454 * @last_stage: if we are at the completion stage
1456 static int save_xbzrle_page(RAMState
*rs
, uint8_t **current_data
,
1457 ram_addr_t current_addr
, RAMBlock
*block
,
1458 ram_addr_t offset
, bool last_stage
)
1460 int encoded_len
= 0, bytes_xbzrle
;
1461 uint8_t *prev_cached_page
;
1463 if (!cache_is_cached(XBZRLE
.cache
, current_addr
,
1464 ram_counters
.dirty_sync_count
)) {
1465 xbzrle_counters
.cache_miss
++;
1467 if (cache_insert(XBZRLE
.cache
, current_addr
, *current_data
,
1468 ram_counters
.dirty_sync_count
) == -1) {
1471 /* update *current_data when the page has been
1472 inserted into cache */
1473 *current_data
= get_cached_data(XBZRLE
.cache
, current_addr
);
1479 prev_cached_page
= get_cached_data(XBZRLE
.cache
, current_addr
);
1481 /* save current buffer into memory */
1482 memcpy(XBZRLE
.current_buf
, *current_data
, TARGET_PAGE_SIZE
);
1484 /* XBZRLE encoding (if there is no overflow) */
1485 encoded_len
= xbzrle_encode_buffer(prev_cached_page
, XBZRLE
.current_buf
,
1486 TARGET_PAGE_SIZE
, XBZRLE
.encoded_buf
,
1488 if (encoded_len
== 0) {
1489 trace_save_xbzrle_page_skipping();
1491 } else if (encoded_len
== -1) {
1492 trace_save_xbzrle_page_overflow();
1493 xbzrle_counters
.overflow
++;
1494 /* update data in the cache */
1496 memcpy(prev_cached_page
, *current_data
, TARGET_PAGE_SIZE
);
1497 *current_data
= prev_cached_page
;
1502 /* we need to update the data in the cache, in order to get the same data */
1504 memcpy(prev_cached_page
, XBZRLE
.current_buf
, TARGET_PAGE_SIZE
);
1507 /* Send XBZRLE based compressed page */
1508 bytes_xbzrle
= save_page_header(rs
, rs
->f
, block
,
1509 offset
| RAM_SAVE_FLAG_XBZRLE
);
1510 qemu_put_byte(rs
->f
, ENCODING_FLAG_XBZRLE
);
1511 qemu_put_be16(rs
->f
, encoded_len
);
1512 qemu_put_buffer(rs
->f
, XBZRLE
.encoded_buf
, encoded_len
);
1513 bytes_xbzrle
+= encoded_len
+ 1 + 2;
1514 xbzrle_counters
.pages
++;
1515 xbzrle_counters
.bytes
+= bytes_xbzrle
;
1516 ram_counters
.transferred
+= bytes_xbzrle
;
1522 * migration_bitmap_find_dirty: find the next dirty page from start
1524 * Called with rcu_read_lock() to protect migration_bitmap
1526 * Returns the byte offset within memory region of the start of a dirty page
1528 * @rs: current RAM state
1529 * @rb: RAMBlock where to search for dirty pages
1530 * @start: page where we start the search
1533 unsigned long migration_bitmap_find_dirty(RAMState
*rs
, RAMBlock
*rb
,
1534 unsigned long start
)
1536 unsigned long size
= rb
->used_length
>> TARGET_PAGE_BITS
;
1537 unsigned long *bitmap
= rb
->bmap
;
1540 if (!qemu_ram_is_migratable(rb
)) {
1544 if (rs
->ram_bulk_stage
&& start
> 0) {
1547 next
= find_next_bit(bitmap
, size
, start
);
1553 static inline bool migration_bitmap_clear_dirty(RAMState
*rs
,
1559 ret
= test_and_clear_bit(page
, rb
->bmap
);
1562 rs
->migration_dirty_pages
--;
1567 static void migration_bitmap_sync_range(RAMState
*rs
, RAMBlock
*rb
,
1568 ram_addr_t start
, ram_addr_t length
)
1570 rs
->migration_dirty_pages
+=
1571 cpu_physical_memory_sync_dirty_bitmap(rb
, start
, length
,
1572 &rs
->num_dirty_pages_period
);
1576 * ram_pagesize_summary: calculate all the pagesizes of a VM
1578 * Returns a summary bitmap of the page sizes of all RAMBlocks
1580 * For VMs with just normal pages this is equivalent to the host page
1581 * size. If it's got some huge pages then it's the OR of all the
1582 * different page sizes.
1584 uint64_t ram_pagesize_summary(void)
1587 uint64_t summary
= 0;
1589 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
1590 summary
|= block
->page_size
;
1596 static void migration_update_rates(RAMState
*rs
, int64_t end_time
)
1598 uint64_t page_count
= rs
->target_page_count
- rs
->target_page_count_prev
;
1599 double compressed_size
;
1601 /* calculate period counters */
1602 ram_counters
.dirty_pages_rate
= rs
->num_dirty_pages_period
* 1000
1603 / (end_time
- rs
->time_last_bitmap_sync
);
1609 if (migrate_use_xbzrle()) {
1610 xbzrle_counters
.cache_miss_rate
= (double)(xbzrle_counters
.cache_miss
-
1611 rs
->xbzrle_cache_miss_prev
) / page_count
;
1612 rs
->xbzrle_cache_miss_prev
= xbzrle_counters
.cache_miss
;
1615 if (migrate_use_compression()) {
1616 compression_counters
.busy_rate
= (double)(compression_counters
.busy
-
1617 rs
->compress_thread_busy_prev
) / page_count
;
1618 rs
->compress_thread_busy_prev
= compression_counters
.busy
;
1620 compressed_size
= compression_counters
.compressed_size
-
1621 rs
->compressed_size_prev
;
1622 if (compressed_size
) {
1623 double uncompressed_size
= (compression_counters
.pages
-
1624 rs
->compress_pages_prev
) * TARGET_PAGE_SIZE
;
1626 /* Compression-Ratio = Uncompressed-size / Compressed-size */
1627 compression_counters
.compression_rate
=
1628 uncompressed_size
/ compressed_size
;
1630 rs
->compress_pages_prev
= compression_counters
.pages
;
1631 rs
->compressed_size_prev
= compression_counters
.compressed_size
;
1636 static void migration_bitmap_sync(RAMState
*rs
)
1640 uint64_t bytes_xfer_now
;
1642 ram_counters
.dirty_sync_count
++;
1644 if (!rs
->time_last_bitmap_sync
) {
1645 rs
->time_last_bitmap_sync
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1648 trace_migration_bitmap_sync_start();
1649 memory_global_dirty_log_sync();
1651 qemu_mutex_lock(&rs
->bitmap_mutex
);
1653 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
1654 migration_bitmap_sync_range(rs
, block
, 0, block
->used_length
);
1656 ram_counters
.remaining
= ram_bytes_remaining();
1658 qemu_mutex_unlock(&rs
->bitmap_mutex
);
1660 trace_migration_bitmap_sync_end(rs
->num_dirty_pages_period
);
1662 end_time
= qemu_clock_get_ms(QEMU_CLOCK_REALTIME
);
1664 /* more than 1 second = 1000 millisecons */
1665 if (end_time
> rs
->time_last_bitmap_sync
+ 1000) {
1666 bytes_xfer_now
= ram_counters
.transferred
;
1668 /* During block migration the auto-converge logic incorrectly detects
1669 * that ram migration makes no progress. Avoid this by disabling the
1670 * throttling logic during the bulk phase of block migration. */
1671 if (migrate_auto_converge() && !blk_mig_bulk_active()) {
1672 /* The following detection logic can be refined later. For now:
1673 Check to see if the dirtied bytes is 50% more than the approx.
1674 amount of bytes that just got transferred since the last time we
1675 were in this routine. If that happens twice, start or increase
1678 if ((rs
->num_dirty_pages_period
* TARGET_PAGE_SIZE
>
1679 (bytes_xfer_now
- rs
->bytes_xfer_prev
) / 2) &&
1680 (++rs
->dirty_rate_high_cnt
>= 2)) {
1681 trace_migration_throttle();
1682 rs
->dirty_rate_high_cnt
= 0;
1683 mig_throttle_guest_down();
1687 migration_update_rates(rs
, end_time
);
1689 rs
->target_page_count_prev
= rs
->target_page_count
;
1691 /* reset period counters */
1692 rs
->time_last_bitmap_sync
= end_time
;
1693 rs
->num_dirty_pages_period
= 0;
1694 rs
->bytes_xfer_prev
= bytes_xfer_now
;
1696 if (migrate_use_events()) {
1697 qapi_event_send_migration_pass(ram_counters
.dirty_sync_count
);
1702 * save_zero_page_to_file: send the zero page to the file
1704 * Returns the size of data written to the file, 0 means the page is not
1707 * @rs: current RAM state
1708 * @file: the file where the data is saved
1709 * @block: block that contains the page we want to send
1710 * @offset: offset inside the block for the page
1712 static int save_zero_page_to_file(RAMState
*rs
, QEMUFile
*file
,
1713 RAMBlock
*block
, ram_addr_t offset
)
1715 uint8_t *p
= block
->host
+ offset
;
1718 if (is_zero_range(p
, TARGET_PAGE_SIZE
)) {
1719 len
+= save_page_header(rs
, file
, block
, offset
| RAM_SAVE_FLAG_ZERO
);
1720 qemu_put_byte(file
, 0);
1727 * save_zero_page: send the zero page to the stream
1729 * Returns the number of pages written.
1731 * @rs: current RAM state
1732 * @block: block that contains the page we want to send
1733 * @offset: offset inside the block for the page
1735 static int save_zero_page(RAMState
*rs
, RAMBlock
*block
, ram_addr_t offset
)
1737 int len
= save_zero_page_to_file(rs
, rs
->f
, block
, offset
);
1740 ram_counters
.duplicate
++;
1741 ram_counters
.transferred
+= len
;
1747 static void ram_release_pages(const char *rbname
, uint64_t offset
, int pages
)
1749 if (!migrate_release_ram() || !migration_in_postcopy()) {
1753 ram_discard_range(rbname
, offset
, pages
<< TARGET_PAGE_BITS
);
1757 * @pages: the number of pages written by the control path,
1759 * > 0 - number of pages written
1761 * Return true if the pages has been saved, otherwise false is returned.
1763 static bool control_save_page(RAMState
*rs
, RAMBlock
*block
, ram_addr_t offset
,
1766 uint64_t bytes_xmit
= 0;
1770 ret
= ram_control_save_page(rs
->f
, block
->offset
, offset
, TARGET_PAGE_SIZE
,
1772 if (ret
== RAM_SAVE_CONTROL_NOT_SUPP
) {
1777 ram_counters
.transferred
+= bytes_xmit
;
1781 if (ret
== RAM_SAVE_CONTROL_DELAYED
) {
1785 if (bytes_xmit
> 0) {
1786 ram_counters
.normal
++;
1787 } else if (bytes_xmit
== 0) {
1788 ram_counters
.duplicate
++;
1795 * directly send the page to the stream
1797 * Returns the number of pages written.
1799 * @rs: current RAM state
1800 * @block: block that contains the page we want to send
1801 * @offset: offset inside the block for the page
1802 * @buf: the page to be sent
1803 * @async: send to page asyncly
1805 static int save_normal_page(RAMState
*rs
, RAMBlock
*block
, ram_addr_t offset
,
1806 uint8_t *buf
, bool async
)
1808 ram_counters
.transferred
+= save_page_header(rs
, rs
->f
, block
,
1809 offset
| RAM_SAVE_FLAG_PAGE
);
1811 qemu_put_buffer_async(rs
->f
, buf
, TARGET_PAGE_SIZE
,
1812 migrate_release_ram() &
1813 migration_in_postcopy());
1815 qemu_put_buffer(rs
->f
, buf
, TARGET_PAGE_SIZE
);
1817 ram_counters
.transferred
+= TARGET_PAGE_SIZE
;
1818 ram_counters
.normal
++;
1823 * ram_save_page: send the given page to the stream
1825 * Returns the number of pages written.
1827 * >=0 - Number of pages written - this might legally be 0
1828 * if xbzrle noticed the page was the same.
1830 * @rs: current RAM state
1831 * @block: block that contains the page we want to send
1832 * @offset: offset inside the block for the page
1833 * @last_stage: if we are at the completion stage
1835 static int ram_save_page(RAMState
*rs
, PageSearchStatus
*pss
, bool last_stage
)
1839 bool send_async
= true;
1840 RAMBlock
*block
= pss
->block
;
1841 ram_addr_t offset
= pss
->page
<< TARGET_PAGE_BITS
;
1842 ram_addr_t current_addr
= block
->offset
+ offset
;
1844 p
= block
->host
+ offset
;
1845 trace_ram_save_page(block
->idstr
, (uint64_t)offset
, p
);
1847 XBZRLE_cache_lock();
1848 if (!rs
->ram_bulk_stage
&& !migration_in_postcopy() &&
1849 migrate_use_xbzrle()) {
1850 pages
= save_xbzrle_page(rs
, &p
, current_addr
, block
,
1851 offset
, last_stage
);
1853 /* Can't send this cached data async, since the cache page
1854 * might get updated before it gets to the wire
1860 /* XBZRLE overflow or normal page */
1862 pages
= save_normal_page(rs
, block
, offset
, p
, send_async
);
1865 XBZRLE_cache_unlock();
1870 static int ram_save_multifd_page(RAMState
*rs
, RAMBlock
*block
,
1873 multifd_queue_page(block
, offset
);
1874 ram_counters
.normal
++;
1879 static bool do_compress_ram_page(QEMUFile
*f
, z_stream
*stream
, RAMBlock
*block
,
1880 ram_addr_t offset
, uint8_t *source_buf
)
1882 RAMState
*rs
= ram_state
;
1883 uint8_t *p
= block
->host
+ (offset
& TARGET_PAGE_MASK
);
1884 bool zero_page
= false;
1887 if (save_zero_page_to_file(rs
, f
, block
, offset
)) {
1892 save_page_header(rs
, f
, block
, offset
| RAM_SAVE_FLAG_COMPRESS_PAGE
);
1895 * copy it to a internal buffer to avoid it being modified by VM
1896 * so that we can catch up the error during compression and
1899 memcpy(source_buf
, p
, TARGET_PAGE_SIZE
);
1900 ret
= qemu_put_compression_data(f
, stream
, source_buf
, TARGET_PAGE_SIZE
);
1902 qemu_file_set_error(migrate_get_current()->to_dst_file
, ret
);
1903 error_report("compressed data failed!");
1908 ram_release_pages(block
->idstr
, offset
& TARGET_PAGE_MASK
, 1);
1913 update_compress_thread_counts(const CompressParam
*param
, int bytes_xmit
)
1915 ram_counters
.transferred
+= bytes_xmit
;
1917 if (param
->zero_page
) {
1918 ram_counters
.duplicate
++;
1922 /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */
1923 compression_counters
.compressed_size
+= bytes_xmit
- 8;
1924 compression_counters
.pages
++;
1927 static bool save_page_use_compression(RAMState
*rs
);
1929 static void flush_compressed_data(RAMState
*rs
)
1931 int idx
, len
, thread_count
;
1933 if (!save_page_use_compression(rs
)) {
1936 thread_count
= migrate_compress_threads();
1938 qemu_mutex_lock(&comp_done_lock
);
1939 for (idx
= 0; idx
< thread_count
; idx
++) {
1940 while (!comp_param
[idx
].done
) {
1941 qemu_cond_wait(&comp_done_cond
, &comp_done_lock
);
1944 qemu_mutex_unlock(&comp_done_lock
);
1946 for (idx
= 0; idx
< thread_count
; idx
++) {
1947 qemu_mutex_lock(&comp_param
[idx
].mutex
);
1948 if (!comp_param
[idx
].quit
) {
1949 len
= qemu_put_qemu_file(rs
->f
, comp_param
[idx
].file
);
1951 * it's safe to fetch zero_page without holding comp_done_lock
1952 * as there is no further request submitted to the thread,
1953 * i.e, the thread should be waiting for a request at this point.
1955 update_compress_thread_counts(&comp_param
[idx
], len
);
1957 qemu_mutex_unlock(&comp_param
[idx
].mutex
);
1961 static inline void set_compress_params(CompressParam
*param
, RAMBlock
*block
,
1964 param
->block
= block
;
1965 param
->offset
= offset
;
1968 static int compress_page_with_multi_thread(RAMState
*rs
, RAMBlock
*block
,
1971 int idx
, thread_count
, bytes_xmit
= -1, pages
= -1;
1972 bool wait
= migrate_compress_wait_thread();
1974 thread_count
= migrate_compress_threads();
1975 qemu_mutex_lock(&comp_done_lock
);
1977 for (idx
= 0; idx
< thread_count
; idx
++) {
1978 if (comp_param
[idx
].done
) {
1979 comp_param
[idx
].done
= false;
1980 bytes_xmit
= qemu_put_qemu_file(rs
->f
, comp_param
[idx
].file
);
1981 qemu_mutex_lock(&comp_param
[idx
].mutex
);
1982 set_compress_params(&comp_param
[idx
], block
, offset
);
1983 qemu_cond_signal(&comp_param
[idx
].cond
);
1984 qemu_mutex_unlock(&comp_param
[idx
].mutex
);
1986 update_compress_thread_counts(&comp_param
[idx
], bytes_xmit
);
1992 * wait for the free thread if the user specifies 'compress-wait-thread',
1993 * otherwise we will post the page out in the main thread as normal page.
1995 if (pages
< 0 && wait
) {
1996 qemu_cond_wait(&comp_done_cond
, &comp_done_lock
);
1999 qemu_mutex_unlock(&comp_done_lock
);
2005 * find_dirty_block: find the next dirty page and update any state
2006 * associated with the search process.
2008 * Returns if a page is found
2010 * @rs: current RAM state
2011 * @pss: data about the state of the current dirty page scan
2012 * @again: set to false if the search has scanned the whole of RAM
2014 static bool find_dirty_block(RAMState
*rs
, PageSearchStatus
*pss
, bool *again
)
2016 pss
->page
= migration_bitmap_find_dirty(rs
, pss
->block
, pss
->page
);
2017 if (pss
->complete_round
&& pss
->block
== rs
->last_seen_block
&&
2018 pss
->page
>= rs
->last_page
) {
2020 * We've been once around the RAM and haven't found anything.
2026 if ((pss
->page
<< TARGET_PAGE_BITS
) >= pss
->block
->used_length
) {
2027 /* Didn't find anything in this RAM Block */
2029 pss
->block
= QLIST_NEXT_RCU(pss
->block
, next
);
2032 * If memory migration starts over, we will meet a dirtied page
2033 * which may still exists in compression threads's ring, so we
2034 * should flush the compressed data to make sure the new page
2035 * is not overwritten by the old one in the destination.
2037 * Also If xbzrle is on, stop using the data compression at this
2038 * point. In theory, xbzrle can do better than compression.
2040 flush_compressed_data(rs
);
2042 /* Hit the end of the list */
2043 pss
->block
= QLIST_FIRST_RCU(&ram_list
.blocks
);
2044 /* Flag that we've looped */
2045 pss
->complete_round
= true;
2046 rs
->ram_bulk_stage
= false;
2048 /* Didn't find anything this time, but try again on the new block */
2052 /* Can go around again, but... */
2054 /* We've found something so probably don't need to */
2060 * unqueue_page: gets a page of the queue
2062 * Helper for 'get_queued_page' - gets a page off the queue
2064 * Returns the block of the page (or NULL if none available)
2066 * @rs: current RAM state
2067 * @offset: used to return the offset within the RAMBlock
2069 static RAMBlock
*unqueue_page(RAMState
*rs
, ram_addr_t
*offset
)
2071 RAMBlock
*block
= NULL
;
2073 if (QSIMPLEQ_EMPTY_ATOMIC(&rs
->src_page_requests
)) {
2077 qemu_mutex_lock(&rs
->src_page_req_mutex
);
2078 if (!QSIMPLEQ_EMPTY(&rs
->src_page_requests
)) {
2079 struct RAMSrcPageRequest
*entry
=
2080 QSIMPLEQ_FIRST(&rs
->src_page_requests
);
2082 *offset
= entry
->offset
;
2084 if (entry
->len
> TARGET_PAGE_SIZE
) {
2085 entry
->len
-= TARGET_PAGE_SIZE
;
2086 entry
->offset
+= TARGET_PAGE_SIZE
;
2088 memory_region_unref(block
->mr
);
2089 QSIMPLEQ_REMOVE_HEAD(&rs
->src_page_requests
, next_req
);
2091 migration_consume_urgent_request();
2094 qemu_mutex_unlock(&rs
->src_page_req_mutex
);
2100 * get_queued_page: unqueue a page from the postocpy requests
2102 * Skips pages that are already sent (!dirty)
2104 * Returns if a queued page is found
2106 * @rs: current RAM state
2107 * @pss: data about the state of the current dirty page scan
2109 static bool get_queued_page(RAMState
*rs
, PageSearchStatus
*pss
)
2116 block
= unqueue_page(rs
, &offset
);
2118 * We're sending this page, and since it's postcopy nothing else
2119 * will dirty it, and we must make sure it doesn't get sent again
2120 * even if this queue request was received after the background
2121 * search already sent it.
2126 page
= offset
>> TARGET_PAGE_BITS
;
2127 dirty
= test_bit(page
, block
->bmap
);
2129 trace_get_queued_page_not_dirty(block
->idstr
, (uint64_t)offset
,
2130 page
, test_bit(page
, block
->unsentmap
));
2132 trace_get_queued_page(block
->idstr
, (uint64_t)offset
, page
);
2136 } while (block
&& !dirty
);
2140 * As soon as we start servicing pages out of order, then we have
2141 * to kill the bulk stage, since the bulk stage assumes
2142 * in (migration_bitmap_find_and_reset_dirty) that every page is
2143 * dirty, that's no longer true.
2145 rs
->ram_bulk_stage
= false;
2148 * We want the background search to continue from the queued page
2149 * since the guest is likely to want other pages near to the page
2150 * it just requested.
2153 pss
->page
= offset
>> TARGET_PAGE_BITS
;
2160 * migration_page_queue_free: drop any remaining pages in the ram
2163 * It should be empty at the end anyway, but in error cases there may
2164 * be some left. in case that there is any page left, we drop it.
2167 static void migration_page_queue_free(RAMState
*rs
)
2169 struct RAMSrcPageRequest
*mspr
, *next_mspr
;
2170 /* This queue generally should be empty - but in the case of a failed
2171 * migration might have some droppings in.
2174 QSIMPLEQ_FOREACH_SAFE(mspr
, &rs
->src_page_requests
, next_req
, next_mspr
) {
2175 memory_region_unref(mspr
->rb
->mr
);
2176 QSIMPLEQ_REMOVE_HEAD(&rs
->src_page_requests
, next_req
);
2183 * ram_save_queue_pages: queue the page for transmission
2185 * A request from postcopy destination for example.
2187 * Returns zero on success or negative on error
2189 * @rbname: Name of the RAMBLock of the request. NULL means the
2190 * same that last one.
2191 * @start: starting address from the start of the RAMBlock
2192 * @len: length (in bytes) to send
2194 int ram_save_queue_pages(const char *rbname
, ram_addr_t start
, ram_addr_t len
)
2197 RAMState
*rs
= ram_state
;
2199 ram_counters
.postcopy_requests
++;
2202 /* Reuse last RAMBlock */
2203 ramblock
= rs
->last_req_rb
;
2207 * Shouldn't happen, we can't reuse the last RAMBlock if
2208 * it's the 1st request.
2210 error_report("ram_save_queue_pages no previous block");
2214 ramblock
= qemu_ram_block_by_name(rbname
);
2217 /* We shouldn't be asked for a non-existent RAMBlock */
2218 error_report("ram_save_queue_pages no block '%s'", rbname
);
2221 rs
->last_req_rb
= ramblock
;
2223 trace_ram_save_queue_pages(ramblock
->idstr
, start
, len
);
2224 if (start
+len
> ramblock
->used_length
) {
2225 error_report("%s request overrun start=" RAM_ADDR_FMT
" len="
2226 RAM_ADDR_FMT
" blocklen=" RAM_ADDR_FMT
,
2227 __func__
, start
, len
, ramblock
->used_length
);
2231 struct RAMSrcPageRequest
*new_entry
=
2232 g_malloc0(sizeof(struct RAMSrcPageRequest
));
2233 new_entry
->rb
= ramblock
;
2234 new_entry
->offset
= start
;
2235 new_entry
->len
= len
;
2237 memory_region_ref(ramblock
->mr
);
2238 qemu_mutex_lock(&rs
->src_page_req_mutex
);
2239 QSIMPLEQ_INSERT_TAIL(&rs
->src_page_requests
, new_entry
, next_req
);
2240 migration_make_urgent_request();
2241 qemu_mutex_unlock(&rs
->src_page_req_mutex
);
2251 static bool save_page_use_compression(RAMState
*rs
)
2253 if (!migrate_use_compression()) {
2258 * If xbzrle is on, stop using the data compression after first
2259 * round of migration even if compression is enabled. In theory,
2260 * xbzrle can do better than compression.
2262 if (rs
->ram_bulk_stage
|| !migrate_use_xbzrle()) {
2270 * try to compress the page before posting it out, return true if the page
2271 * has been properly handled by compression, otherwise needs other
2272 * paths to handle it
2274 static bool save_compress_page(RAMState
*rs
, RAMBlock
*block
, ram_addr_t offset
)
2276 if (!save_page_use_compression(rs
)) {
2281 * When starting the process of a new block, the first page of
2282 * the block should be sent out before other pages in the same
2283 * block, and all the pages in last block should have been sent
2284 * out, keeping this order is important, because the 'cont' flag
2285 * is used to avoid resending the block name.
2287 * We post the fist page as normal page as compression will take
2288 * much CPU resource.
2290 if (block
!= rs
->last_sent_block
) {
2291 flush_compressed_data(rs
);
2295 if (compress_page_with_multi_thread(rs
, block
, offset
) > 0) {
2299 compression_counters
.busy
++;
2304 * ram_save_target_page: save one target page
2306 * Returns the number of pages written
2308 * @rs: current RAM state
2309 * @pss: data about the page we want to send
2310 * @last_stage: if we are at the completion stage
2312 static int ram_save_target_page(RAMState
*rs
, PageSearchStatus
*pss
,
2315 RAMBlock
*block
= pss
->block
;
2316 ram_addr_t offset
= pss
->page
<< TARGET_PAGE_BITS
;
2319 if (control_save_page(rs
, block
, offset
, &res
)) {
2323 if (save_compress_page(rs
, block
, offset
)) {
2327 res
= save_zero_page(rs
, block
, offset
);
2329 /* Must let xbzrle know, otherwise a previous (now 0'd) cached
2330 * page would be stale
2332 if (!save_page_use_compression(rs
)) {
2333 XBZRLE_cache_lock();
2334 xbzrle_cache_zero_page(rs
, block
->offset
+ offset
);
2335 XBZRLE_cache_unlock();
2337 ram_release_pages(block
->idstr
, offset
, res
);
2342 * do not use multifd for compression as the first page in the new
2343 * block should be posted out before sending the compressed page
2345 if (!save_page_use_compression(rs
) && migrate_use_multifd()) {
2346 return ram_save_multifd_page(rs
, block
, offset
);
2349 return ram_save_page(rs
, pss
, last_stage
);
2353 * ram_save_host_page: save a whole host page
2355 * Starting at *offset send pages up to the end of the current host
2356 * page. It's valid for the initial offset to point into the middle of
2357 * a host page in which case the remainder of the hostpage is sent.
2358 * Only dirty target pages are sent. Note that the host page size may
2359 * be a huge page for this block.
2360 * The saving stops at the boundary of the used_length of the block
2361 * if the RAMBlock isn't a multiple of the host page size.
2363 * Returns the number of pages written or negative on error
2365 * @rs: current RAM state
2366 * @ms: current migration state
2367 * @pss: data about the page we want to send
2368 * @last_stage: if we are at the completion stage
2370 static int ram_save_host_page(RAMState
*rs
, PageSearchStatus
*pss
,
2373 int tmppages
, pages
= 0;
2374 size_t pagesize_bits
=
2375 qemu_ram_pagesize(pss
->block
) >> TARGET_PAGE_BITS
;
2377 if (!qemu_ram_is_migratable(pss
->block
)) {
2378 error_report("block %s should not be migrated !", pss
->block
->idstr
);
2383 /* Check the pages is dirty and if it is send it */
2384 if (!migration_bitmap_clear_dirty(rs
, pss
->block
, pss
->page
)) {
2389 tmppages
= ram_save_target_page(rs
, pss
, last_stage
);
2395 if (pss
->block
->unsentmap
) {
2396 clear_bit(pss
->page
, pss
->block
->unsentmap
);
2400 } while ((pss
->page
& (pagesize_bits
- 1)) &&
2401 offset_in_ramblock(pss
->block
, pss
->page
<< TARGET_PAGE_BITS
));
2403 /* The offset we leave with is the last one we looked at */
2409 * ram_find_and_save_block: finds a dirty page and sends it to f
2411 * Called within an RCU critical section.
2413 * Returns the number of pages written where zero means no dirty pages,
2414 * or negative on error
2416 * @rs: current RAM state
2417 * @last_stage: if we are at the completion stage
2419 * On systems where host-page-size > target-page-size it will send all the
2420 * pages in a host page that are dirty.
2423 static int ram_find_and_save_block(RAMState
*rs
, bool last_stage
)
2425 PageSearchStatus pss
;
2429 /* No dirty page as there is zero RAM */
2430 if (!ram_bytes_total()) {
2434 pss
.block
= rs
->last_seen_block
;
2435 pss
.page
= rs
->last_page
;
2436 pss
.complete_round
= false;
2439 pss
.block
= QLIST_FIRST_RCU(&ram_list
.blocks
);
2444 found
= get_queued_page(rs
, &pss
);
2447 /* priority queue empty, so just search for something dirty */
2448 found
= find_dirty_block(rs
, &pss
, &again
);
2452 pages
= ram_save_host_page(rs
, &pss
, last_stage
);
2454 } while (!pages
&& again
);
2456 rs
->last_seen_block
= pss
.block
;
2457 rs
->last_page
= pss
.page
;
2462 void acct_update_position(QEMUFile
*f
, size_t size
, bool zero
)
2464 uint64_t pages
= size
/ TARGET_PAGE_SIZE
;
2467 ram_counters
.duplicate
+= pages
;
2469 ram_counters
.normal
+= pages
;
2470 ram_counters
.transferred
+= size
;
2471 qemu_update_position(f
, size
);
2475 uint64_t ram_bytes_total(void)
2481 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2482 total
+= block
->used_length
;
2488 static void xbzrle_load_setup(void)
2490 XBZRLE
.decoded_buf
= g_malloc(TARGET_PAGE_SIZE
);
2493 static void xbzrle_load_cleanup(void)
2495 g_free(XBZRLE
.decoded_buf
);
2496 XBZRLE
.decoded_buf
= NULL
;
2499 static void ram_state_cleanup(RAMState
**rsp
)
2502 migration_page_queue_free(*rsp
);
2503 qemu_mutex_destroy(&(*rsp
)->bitmap_mutex
);
2504 qemu_mutex_destroy(&(*rsp
)->src_page_req_mutex
);
2510 static void xbzrle_cleanup(void)
2512 XBZRLE_cache_lock();
2514 cache_fini(XBZRLE
.cache
);
2515 g_free(XBZRLE
.encoded_buf
);
2516 g_free(XBZRLE
.current_buf
);
2517 g_free(XBZRLE
.zero_target_page
);
2518 XBZRLE
.cache
= NULL
;
2519 XBZRLE
.encoded_buf
= NULL
;
2520 XBZRLE
.current_buf
= NULL
;
2521 XBZRLE
.zero_target_page
= NULL
;
2523 XBZRLE_cache_unlock();
2526 static void ram_save_cleanup(void *opaque
)
2528 RAMState
**rsp
= opaque
;
2531 /* caller have hold iothread lock or is in a bh, so there is
2532 * no writing race against this migration_bitmap
2534 memory_global_dirty_log_stop();
2536 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2537 g_free(block
->bmap
);
2539 g_free(block
->unsentmap
);
2540 block
->unsentmap
= NULL
;
2544 compress_threads_save_cleanup();
2545 ram_state_cleanup(rsp
);
2548 static void ram_state_reset(RAMState
*rs
)
2550 rs
->last_seen_block
= NULL
;
2551 rs
->last_sent_block
= NULL
;
2553 rs
->last_version
= ram_list
.version
;
2554 rs
->ram_bulk_stage
= true;
2557 #define MAX_WAIT 50 /* ms, half buffered_file limit */
2560 * 'expected' is the value you expect the bitmap mostly to be full
2561 * of; it won't bother printing lines that are all this value.
2562 * If 'todump' is null the migration bitmap is dumped.
2564 void ram_debug_dump_bitmap(unsigned long *todump
, bool expected
,
2565 unsigned long pages
)
2568 int64_t linelen
= 128;
2571 for (cur
= 0; cur
< pages
; cur
+= linelen
) {
2575 * Last line; catch the case where the line length
2576 * is longer than remaining ram
2578 if (cur
+ linelen
> pages
) {
2579 linelen
= pages
- cur
;
2581 for (curb
= 0; curb
< linelen
; curb
++) {
2582 bool thisbit
= test_bit(cur
+ curb
, todump
);
2583 linebuf
[curb
] = thisbit
? '1' : '.';
2584 found
= found
|| (thisbit
!= expected
);
2587 linebuf
[curb
] = '\0';
2588 fprintf(stderr
, "0x%08" PRIx64
" : %s\n", cur
, linebuf
);
2593 /* **** functions for postcopy ***** */
2595 void ram_postcopy_migrated_memory_release(MigrationState
*ms
)
2597 struct RAMBlock
*block
;
2599 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2600 unsigned long *bitmap
= block
->bmap
;
2601 unsigned long range
= block
->used_length
>> TARGET_PAGE_BITS
;
2602 unsigned long run_start
= find_next_zero_bit(bitmap
, range
, 0);
2604 while (run_start
< range
) {
2605 unsigned long run_end
= find_next_bit(bitmap
, range
, run_start
+ 1);
2606 ram_discard_range(block
->idstr
, run_start
<< TARGET_PAGE_BITS
,
2607 (run_end
- run_start
) << TARGET_PAGE_BITS
);
2608 run_start
= find_next_zero_bit(bitmap
, range
, run_end
+ 1);
2614 * postcopy_send_discard_bm_ram: discard a RAMBlock
2616 * Returns zero on success
2618 * Callback from postcopy_each_ram_send_discard for each RAMBlock
2619 * Note: At this point the 'unsentmap' is the processed bitmap combined
2620 * with the dirtymap; so a '1' means it's either dirty or unsent.
2622 * @ms: current migration state
2623 * @pds: state for postcopy
2624 * @start: RAMBlock starting page
2625 * @length: RAMBlock size
2627 static int postcopy_send_discard_bm_ram(MigrationState
*ms
,
2628 PostcopyDiscardState
*pds
,
2631 unsigned long end
= block
->used_length
>> TARGET_PAGE_BITS
;
2632 unsigned long current
;
2633 unsigned long *unsentmap
= block
->unsentmap
;
2635 for (current
= 0; current
< end
; ) {
2636 unsigned long one
= find_next_bit(unsentmap
, end
, current
);
2639 unsigned long zero
= find_next_zero_bit(unsentmap
, end
, one
+ 1);
2640 unsigned long discard_length
;
2643 discard_length
= end
- one
;
2645 discard_length
= zero
- one
;
2647 if (discard_length
) {
2648 postcopy_discard_send_range(ms
, pds
, one
, discard_length
);
2650 current
= one
+ discard_length
;
2660 * postcopy_each_ram_send_discard: discard all RAMBlocks
2662 * Returns 0 for success or negative for error
2664 * Utility for the outgoing postcopy code.
2665 * Calls postcopy_send_discard_bm_ram for each RAMBlock
2666 * passing it bitmap indexes and name.
2667 * (qemu_ram_foreach_block ends up passing unscaled lengths
2668 * which would mean postcopy code would have to deal with target page)
2670 * @ms: current migration state
2672 static int postcopy_each_ram_send_discard(MigrationState
*ms
)
2674 struct RAMBlock
*block
;
2677 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2678 PostcopyDiscardState
*pds
=
2679 postcopy_discard_send_init(ms
, block
->idstr
);
2682 * Postcopy sends chunks of bitmap over the wire, but it
2683 * just needs indexes at this point, avoids it having
2684 * target page specific code.
2686 ret
= postcopy_send_discard_bm_ram(ms
, pds
, block
);
2687 postcopy_discard_send_finish(ms
, pds
);
2697 * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages
2699 * Helper for postcopy_chunk_hostpages; it's called twice to
2700 * canonicalize the two bitmaps, that are similar, but one is
2703 * Postcopy requires that all target pages in a hostpage are dirty or
2704 * clean, not a mix. This function canonicalizes the bitmaps.
2706 * @ms: current migration state
2707 * @unsent_pass: if true we need to canonicalize partially unsent host pages
2708 * otherwise we need to canonicalize partially dirty host pages
2709 * @block: block that contains the page we want to canonicalize
2710 * @pds: state for postcopy
2712 static void postcopy_chunk_hostpages_pass(MigrationState
*ms
, bool unsent_pass
,
2714 PostcopyDiscardState
*pds
)
2716 RAMState
*rs
= ram_state
;
2717 unsigned long *bitmap
= block
->bmap
;
2718 unsigned long *unsentmap
= block
->unsentmap
;
2719 unsigned int host_ratio
= block
->page_size
/ TARGET_PAGE_SIZE
;
2720 unsigned long pages
= block
->used_length
>> TARGET_PAGE_BITS
;
2721 unsigned long run_start
;
2723 if (block
->page_size
== TARGET_PAGE_SIZE
) {
2724 /* Easy case - TPS==HPS for a non-huge page RAMBlock */
2729 /* Find a sent page */
2730 run_start
= find_next_zero_bit(unsentmap
, pages
, 0);
2732 /* Find a dirty page */
2733 run_start
= find_next_bit(bitmap
, pages
, 0);
2736 while (run_start
< pages
) {
2737 bool do_fixup
= false;
2738 unsigned long fixup_start_addr
;
2739 unsigned long host_offset
;
2742 * If the start of this run of pages is in the middle of a host
2743 * page, then we need to fixup this host page.
2745 host_offset
= run_start
% host_ratio
;
2748 run_start
-= host_offset
;
2749 fixup_start_addr
= run_start
;
2750 /* For the next pass */
2751 run_start
= run_start
+ host_ratio
;
2753 /* Find the end of this run */
2754 unsigned long run_end
;
2756 run_end
= find_next_bit(unsentmap
, pages
, run_start
+ 1);
2758 run_end
= find_next_zero_bit(bitmap
, pages
, run_start
+ 1);
2761 * If the end isn't at the start of a host page, then the
2762 * run doesn't finish at the end of a host page
2763 * and we need to discard.
2765 host_offset
= run_end
% host_ratio
;
2768 fixup_start_addr
= run_end
- host_offset
;
2770 * This host page has gone, the next loop iteration starts
2771 * from after the fixup
2773 run_start
= fixup_start_addr
+ host_ratio
;
2776 * No discards on this iteration, next loop starts from
2777 * next sent/dirty page
2779 run_start
= run_end
+ 1;
2786 /* Tell the destination to discard this page */
2787 if (unsent_pass
|| !test_bit(fixup_start_addr
, unsentmap
)) {
2788 /* For the unsent_pass we:
2789 * discard partially sent pages
2790 * For the !unsent_pass (dirty) we:
2791 * discard partially dirty pages that were sent
2792 * (any partially sent pages were already discarded
2793 * by the previous unsent_pass)
2795 postcopy_discard_send_range(ms
, pds
, fixup_start_addr
,
2799 /* Clean up the bitmap */
2800 for (page
= fixup_start_addr
;
2801 page
< fixup_start_addr
+ host_ratio
; page
++) {
2802 /* All pages in this host page are now not sent */
2803 set_bit(page
, unsentmap
);
2806 * Remark them as dirty, updating the count for any pages
2807 * that weren't previously dirty.
2809 rs
->migration_dirty_pages
+= !test_and_set_bit(page
, bitmap
);
2814 /* Find the next sent page for the next iteration */
2815 run_start
= find_next_zero_bit(unsentmap
, pages
, run_start
);
2817 /* Find the next dirty page for the next iteration */
2818 run_start
= find_next_bit(bitmap
, pages
, run_start
);
2824 * postcopy_chuck_hostpages: discrad any partially sent host page
2826 * Utility for the outgoing postcopy code.
2828 * Discard any partially sent host-page size chunks, mark any partially
2829 * dirty host-page size chunks as all dirty. In this case the host-page
2830 * is the host-page for the particular RAMBlock, i.e. it might be a huge page
2832 * Returns zero on success
2834 * @ms: current migration state
2835 * @block: block we want to work with
2837 static int postcopy_chunk_hostpages(MigrationState
*ms
, RAMBlock
*block
)
2839 PostcopyDiscardState
*pds
=
2840 postcopy_discard_send_init(ms
, block
->idstr
);
2842 /* First pass: Discard all partially sent host pages */
2843 postcopy_chunk_hostpages_pass(ms
, true, block
, pds
);
2845 * Second pass: Ensure that all partially dirty host pages are made
2848 postcopy_chunk_hostpages_pass(ms
, false, block
, pds
);
2850 postcopy_discard_send_finish(ms
, pds
);
2855 * ram_postcopy_send_discard_bitmap: transmit the discard bitmap
2857 * Returns zero on success
2859 * Transmit the set of pages to be discarded after precopy to the target
2860 * these are pages that:
2861 * a) Have been previously transmitted but are now dirty again
2862 * b) Pages that have never been transmitted, this ensures that
2863 * any pages on the destination that have been mapped by background
2864 * tasks get discarded (transparent huge pages is the specific concern)
2865 * Hopefully this is pretty sparse
2867 * @ms: current migration state
2869 int ram_postcopy_send_discard_bitmap(MigrationState
*ms
)
2871 RAMState
*rs
= ram_state
;
2877 /* This should be our last sync, the src is now paused */
2878 migration_bitmap_sync(rs
);
2880 /* Easiest way to make sure we don't resume in the middle of a host-page */
2881 rs
->last_seen_block
= NULL
;
2882 rs
->last_sent_block
= NULL
;
2885 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
2886 unsigned long pages
= block
->used_length
>> TARGET_PAGE_BITS
;
2887 unsigned long *bitmap
= block
->bmap
;
2888 unsigned long *unsentmap
= block
->unsentmap
;
2891 /* We don't have a safe way to resize the sentmap, so
2892 * if the bitmap was resized it will be NULL at this
2895 error_report("migration ram resized during precopy phase");
2899 /* Deal with TPS != HPS and huge pages */
2900 ret
= postcopy_chunk_hostpages(ms
, block
);
2907 * Update the unsentmap to be unsentmap = unsentmap | dirty
2909 bitmap_or(unsentmap
, unsentmap
, bitmap
, pages
);
2910 #ifdef DEBUG_POSTCOPY
2911 ram_debug_dump_bitmap(unsentmap
, true, pages
);
2914 trace_ram_postcopy_send_discard_bitmap();
2916 ret
= postcopy_each_ram_send_discard(ms
);
2923 * ram_discard_range: discard dirtied pages at the beginning of postcopy
2925 * Returns zero on success
2927 * @rbname: name of the RAMBlock of the request. NULL means the
2928 * same that last one.
2929 * @start: RAMBlock starting page
2930 * @length: RAMBlock size
2932 int ram_discard_range(const char *rbname
, uint64_t start
, size_t length
)
2936 trace_ram_discard_range(rbname
, start
, length
);
2939 RAMBlock
*rb
= qemu_ram_block_by_name(rbname
);
2942 error_report("ram_discard_range: Failed to find block '%s'", rbname
);
2947 * On source VM, we don't need to update the received bitmap since
2948 * we don't even have one.
2950 if (rb
->receivedmap
) {
2951 bitmap_clear(rb
->receivedmap
, start
>> qemu_target_page_bits(),
2952 length
>> qemu_target_page_bits());
2955 ret
= ram_block_discard_range(rb
, start
, length
);
2964 * For every allocation, we will try not to crash the VM if the
2965 * allocation failed.
2967 static int xbzrle_init(void)
2969 Error
*local_err
= NULL
;
2971 if (!migrate_use_xbzrle()) {
2975 XBZRLE_cache_lock();
2977 XBZRLE
.zero_target_page
= g_try_malloc0(TARGET_PAGE_SIZE
);
2978 if (!XBZRLE
.zero_target_page
) {
2979 error_report("%s: Error allocating zero page", __func__
);
2983 XBZRLE
.cache
= cache_init(migrate_xbzrle_cache_size(),
2984 TARGET_PAGE_SIZE
, &local_err
);
2985 if (!XBZRLE
.cache
) {
2986 error_report_err(local_err
);
2987 goto free_zero_page
;
2990 XBZRLE
.encoded_buf
= g_try_malloc0(TARGET_PAGE_SIZE
);
2991 if (!XBZRLE
.encoded_buf
) {
2992 error_report("%s: Error allocating encoded_buf", __func__
);
2996 XBZRLE
.current_buf
= g_try_malloc(TARGET_PAGE_SIZE
);
2997 if (!XBZRLE
.current_buf
) {
2998 error_report("%s: Error allocating current_buf", __func__
);
2999 goto free_encoded_buf
;
3002 /* We are all good */
3003 XBZRLE_cache_unlock();
3007 g_free(XBZRLE
.encoded_buf
);
3008 XBZRLE
.encoded_buf
= NULL
;
3010 cache_fini(XBZRLE
.cache
);
3011 XBZRLE
.cache
= NULL
;
3013 g_free(XBZRLE
.zero_target_page
);
3014 XBZRLE
.zero_target_page
= NULL
;
3016 XBZRLE_cache_unlock();
3020 static int ram_state_init(RAMState
**rsp
)
3022 *rsp
= g_try_new0(RAMState
, 1);
3025 error_report("%s: Init ramstate fail", __func__
);
3029 qemu_mutex_init(&(*rsp
)->bitmap_mutex
);
3030 qemu_mutex_init(&(*rsp
)->src_page_req_mutex
);
3031 QSIMPLEQ_INIT(&(*rsp
)->src_page_requests
);
3034 * Count the total number of pages used by ram blocks not including any
3035 * gaps due to alignment or unplugs.
3037 (*rsp
)->migration_dirty_pages
= ram_bytes_total() >> TARGET_PAGE_BITS
;
3039 ram_state_reset(*rsp
);
3044 static void ram_list_init_bitmaps(void)
3047 unsigned long pages
;
3049 /* Skip setting bitmap if there is no RAM */
3050 if (ram_bytes_total()) {
3051 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3052 pages
= block
->max_length
>> TARGET_PAGE_BITS
;
3053 block
->bmap
= bitmap_new(pages
);
3054 bitmap_set(block
->bmap
, 0, pages
);
3055 if (migrate_postcopy_ram()) {
3056 block
->unsentmap
= bitmap_new(pages
);
3057 bitmap_set(block
->unsentmap
, 0, pages
);
3063 static void ram_init_bitmaps(RAMState
*rs
)
3065 /* For memory_global_dirty_log_start below. */
3066 qemu_mutex_lock_iothread();
3067 qemu_mutex_lock_ramlist();
3070 ram_list_init_bitmaps();
3071 memory_global_dirty_log_start();
3072 migration_bitmap_sync(rs
);
3075 qemu_mutex_unlock_ramlist();
3076 qemu_mutex_unlock_iothread();
3079 static int ram_init_all(RAMState
**rsp
)
3081 if (ram_state_init(rsp
)) {
3085 if (xbzrle_init()) {
3086 ram_state_cleanup(rsp
);
3090 ram_init_bitmaps(*rsp
);
3095 static void ram_state_resume_prepare(RAMState
*rs
, QEMUFile
*out
)
3101 * Postcopy is not using xbzrle/compression, so no need for that.
3102 * Also, since source are already halted, we don't need to care
3103 * about dirty page logging as well.
3106 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3107 pages
+= bitmap_count_one(block
->bmap
,
3108 block
->used_length
>> TARGET_PAGE_BITS
);
3111 /* This may not be aligned with current bitmaps. Recalculate. */
3112 rs
->migration_dirty_pages
= pages
;
3114 rs
->last_seen_block
= NULL
;
3115 rs
->last_sent_block
= NULL
;
3117 rs
->last_version
= ram_list
.version
;
3119 * Disable the bulk stage, otherwise we'll resend the whole RAM no
3120 * matter what we have sent.
3122 rs
->ram_bulk_stage
= false;
3124 /* Update RAMState cache of output QEMUFile */
3127 trace_ram_state_resume_prepare(pages
);
3131 * Each of ram_save_setup, ram_save_iterate and ram_save_complete has
3132 * long-running RCU critical section. When rcu-reclaims in the code
3133 * start to become numerous it will be necessary to reduce the
3134 * granularity of these critical sections.
3138 * ram_save_setup: Setup RAM for migration
3140 * Returns zero to indicate success and negative for error
3142 * @f: QEMUFile where to send the data
3143 * @opaque: RAMState pointer
3145 static int ram_save_setup(QEMUFile
*f
, void *opaque
)
3147 RAMState
**rsp
= opaque
;
3150 if (compress_threads_save_setup()) {
3154 /* migration has already setup the bitmap, reuse it. */
3155 if (!migration_in_colo_state()) {
3156 if (ram_init_all(rsp
) != 0) {
3157 compress_threads_save_cleanup();
3165 qemu_put_be64(f
, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE
);
3167 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3168 qemu_put_byte(f
, strlen(block
->idstr
));
3169 qemu_put_buffer(f
, (uint8_t *)block
->idstr
, strlen(block
->idstr
));
3170 qemu_put_be64(f
, block
->used_length
);
3171 if (migrate_postcopy_ram() && block
->page_size
!= qemu_host_page_size
) {
3172 qemu_put_be64(f
, block
->page_size
);
3178 ram_control_before_iterate(f
, RAM_CONTROL_SETUP
);
3179 ram_control_after_iterate(f
, RAM_CONTROL_SETUP
);
3181 multifd_send_sync_main();
3182 qemu_put_be64(f
, RAM_SAVE_FLAG_EOS
);
3189 * ram_save_iterate: iterative stage for migration
3191 * Returns zero to indicate success and negative for error
3193 * @f: QEMUFile where to send the data
3194 * @opaque: RAMState pointer
3196 static int ram_save_iterate(QEMUFile
*f
, void *opaque
)
3198 RAMState
**temp
= opaque
;
3199 RAMState
*rs
= *temp
;
3205 if (blk_mig_bulk_active()) {
3206 /* Avoid transferring ram during bulk phase of block migration as
3207 * the bulk phase will usually take a long time and transferring
3208 * ram updates during that time is pointless. */
3213 if (ram_list
.version
!= rs
->last_version
) {
3214 ram_state_reset(rs
);
3217 /* Read version before ram_list.blocks */
3220 ram_control_before_iterate(f
, RAM_CONTROL_ROUND
);
3222 t0
= qemu_clock_get_ns(QEMU_CLOCK_REALTIME
);
3224 while ((ret
= qemu_file_rate_limit(f
)) == 0 ||
3225 !QSIMPLEQ_EMPTY(&rs
->src_page_requests
)) {
3228 if (qemu_file_get_error(f
)) {
3232 pages
= ram_find_and_save_block(rs
, false);
3233 /* no more pages to sent */
3240 qemu_file_set_error(f
, pages
);
3244 rs
->target_page_count
+= pages
;
3246 /* we want to check in the 1st loop, just in case it was the 1st time
3247 and we had to sync the dirty bitmap.
3248 qemu_get_clock_ns() is a bit expensive, so we only check each some
3251 if ((i
& 63) == 0) {
3252 uint64_t t1
= (qemu_clock_get_ns(QEMU_CLOCK_REALTIME
) - t0
) / 1000000;
3253 if (t1
> MAX_WAIT
) {
3254 trace_ram_save_iterate_big_wait(t1
, i
);
3263 * Must occur before EOS (or any QEMUFile operation)
3264 * because of RDMA protocol.
3266 ram_control_after_iterate(f
, RAM_CONTROL_ROUND
);
3268 multifd_send_sync_main();
3270 qemu_put_be64(f
, RAM_SAVE_FLAG_EOS
);
3272 ram_counters
.transferred
+= 8;
3274 ret
= qemu_file_get_error(f
);
3283 * ram_save_complete: function called to send the remaining amount of ram
3285 * Returns zero to indicate success or negative on error
3287 * Called with iothread lock
3289 * @f: QEMUFile where to send the data
3290 * @opaque: RAMState pointer
3292 static int ram_save_complete(QEMUFile
*f
, void *opaque
)
3294 RAMState
**temp
= opaque
;
3295 RAMState
*rs
= *temp
;
3300 if (!migration_in_postcopy()) {
3301 migration_bitmap_sync(rs
);
3304 ram_control_before_iterate(f
, RAM_CONTROL_FINISH
);
3306 /* try transferring iterative blocks of memory */
3308 /* flush all remaining blocks regardless of rate limiting */
3312 pages
= ram_find_and_save_block(rs
, !migration_in_colo_state());
3313 /* no more blocks to sent */
3323 flush_compressed_data(rs
);
3324 ram_control_after_iterate(f
, RAM_CONTROL_FINISH
);
3328 multifd_send_sync_main();
3329 qemu_put_be64(f
, RAM_SAVE_FLAG_EOS
);
3335 static void ram_save_pending(QEMUFile
*f
, void *opaque
, uint64_t max_size
,
3336 uint64_t *res_precopy_only
,
3337 uint64_t *res_compatible
,
3338 uint64_t *res_postcopy_only
)
3340 RAMState
**temp
= opaque
;
3341 RAMState
*rs
= *temp
;
3342 uint64_t remaining_size
;
3344 remaining_size
= rs
->migration_dirty_pages
* TARGET_PAGE_SIZE
;
3346 if (!migration_in_postcopy() &&
3347 remaining_size
< max_size
) {
3348 qemu_mutex_lock_iothread();
3350 migration_bitmap_sync(rs
);
3352 qemu_mutex_unlock_iothread();
3353 remaining_size
= rs
->migration_dirty_pages
* TARGET_PAGE_SIZE
;
3356 if (migrate_postcopy_ram()) {
3357 /* We can do postcopy, and all the data is postcopiable */
3358 *res_compatible
+= remaining_size
;
3360 *res_precopy_only
+= remaining_size
;
3364 static int load_xbzrle(QEMUFile
*f
, ram_addr_t addr
, void *host
)
3366 unsigned int xh_len
;
3368 uint8_t *loaded_data
;
3370 /* extract RLE header */
3371 xh_flags
= qemu_get_byte(f
);
3372 xh_len
= qemu_get_be16(f
);
3374 if (xh_flags
!= ENCODING_FLAG_XBZRLE
) {
3375 error_report("Failed to load XBZRLE page - wrong compression!");
3379 if (xh_len
> TARGET_PAGE_SIZE
) {
3380 error_report("Failed to load XBZRLE page - len overflow!");
3383 loaded_data
= XBZRLE
.decoded_buf
;
3384 /* load data and decode */
3385 /* it can change loaded_data to point to an internal buffer */
3386 qemu_get_buffer_in_place(f
, &loaded_data
, xh_len
);
3389 if (xbzrle_decode_buffer(loaded_data
, xh_len
, host
,
3390 TARGET_PAGE_SIZE
) == -1) {
3391 error_report("Failed to load XBZRLE page - decode error!");
3399 * ram_block_from_stream: read a RAMBlock id from the migration stream
3401 * Must be called from within a rcu critical section.
3403 * Returns a pointer from within the RCU-protected ram_list.
3405 * @f: QEMUFile where to read the data from
3406 * @flags: Page flags (mostly to see if it's a continuation of previous block)
3408 static inline RAMBlock
*ram_block_from_stream(QEMUFile
*f
, int flags
)
3410 static RAMBlock
*block
= NULL
;
3414 if (flags
& RAM_SAVE_FLAG_CONTINUE
) {
3416 error_report("Ack, bad migration stream!");
3422 len
= qemu_get_byte(f
);
3423 qemu_get_buffer(f
, (uint8_t *)id
, len
);
3426 block
= qemu_ram_block_by_name(id
);
3428 error_report("Can't find block %s", id
);
3432 if (!qemu_ram_is_migratable(block
)) {
3433 error_report("block %s should not be migrated !", id
);
3440 static inline void *host_from_ram_block_offset(RAMBlock
*block
,
3443 if (!offset_in_ramblock(block
, offset
)) {
3447 return block
->host
+ offset
;
3450 static inline void *colo_cache_from_block_offset(RAMBlock
*block
,
3453 if (!offset_in_ramblock(block
, offset
)) {
3456 if (!block
->colo_cache
) {
3457 error_report("%s: colo_cache is NULL in block :%s",
3458 __func__
, block
->idstr
);
3463 * During colo checkpoint, we need bitmap of these migrated pages.
3464 * It help us to decide which pages in ram cache should be flushed
3465 * into VM's RAM later.
3467 if (!test_and_set_bit(offset
>> TARGET_PAGE_BITS
, block
->bmap
)) {
3468 ram_state
->migration_dirty_pages
++;
3470 return block
->colo_cache
+ offset
;
3474 * ram_handle_compressed: handle the zero page case
3476 * If a page (or a whole RDMA chunk) has been
3477 * determined to be zero, then zap it.
3479 * @host: host address for the zero page
3480 * @ch: what the page is filled from. We only support zero
3481 * @size: size of the zero page
3483 void ram_handle_compressed(void *host
, uint8_t ch
, uint64_t size
)
3485 if (ch
!= 0 || !is_zero_range(host
, size
)) {
3486 memset(host
, ch
, size
);
3490 /* return the size after decompression, or negative value on error */
3492 qemu_uncompress_data(z_stream
*stream
, uint8_t *dest
, size_t dest_len
,
3493 const uint8_t *source
, size_t source_len
)
3497 err
= inflateReset(stream
);
3502 stream
->avail_in
= source_len
;
3503 stream
->next_in
= (uint8_t *)source
;
3504 stream
->avail_out
= dest_len
;
3505 stream
->next_out
= dest
;
3507 err
= inflate(stream
, Z_NO_FLUSH
);
3508 if (err
!= Z_STREAM_END
) {
3512 return stream
->total_out
;
3515 static void *do_data_decompress(void *opaque
)
3517 DecompressParam
*param
= opaque
;
3518 unsigned long pagesize
;
3522 qemu_mutex_lock(¶m
->mutex
);
3523 while (!param
->quit
) {
3528 qemu_mutex_unlock(¶m
->mutex
);
3530 pagesize
= TARGET_PAGE_SIZE
;
3532 ret
= qemu_uncompress_data(¶m
->stream
, des
, pagesize
,
3533 param
->compbuf
, len
);
3534 if (ret
< 0 && migrate_get_current()->decompress_error_check
) {
3535 error_report("decompress data failed");
3536 qemu_file_set_error(decomp_file
, ret
);
3539 qemu_mutex_lock(&decomp_done_lock
);
3541 qemu_cond_signal(&decomp_done_cond
);
3542 qemu_mutex_unlock(&decomp_done_lock
);
3544 qemu_mutex_lock(¶m
->mutex
);
3546 qemu_cond_wait(¶m
->cond
, ¶m
->mutex
);
3549 qemu_mutex_unlock(¶m
->mutex
);
3554 static int wait_for_decompress_done(void)
3556 int idx
, thread_count
;
3558 if (!migrate_use_compression()) {
3562 thread_count
= migrate_decompress_threads();
3563 qemu_mutex_lock(&decomp_done_lock
);
3564 for (idx
= 0; idx
< thread_count
; idx
++) {
3565 while (!decomp_param
[idx
].done
) {
3566 qemu_cond_wait(&decomp_done_cond
, &decomp_done_lock
);
3569 qemu_mutex_unlock(&decomp_done_lock
);
3570 return qemu_file_get_error(decomp_file
);
3573 static void compress_threads_load_cleanup(void)
3575 int i
, thread_count
;
3577 if (!migrate_use_compression()) {
3580 thread_count
= migrate_decompress_threads();
3581 for (i
= 0; i
< thread_count
; i
++) {
3583 * we use it as a indicator which shows if the thread is
3584 * properly init'd or not
3586 if (!decomp_param
[i
].compbuf
) {
3590 qemu_mutex_lock(&decomp_param
[i
].mutex
);
3591 decomp_param
[i
].quit
= true;
3592 qemu_cond_signal(&decomp_param
[i
].cond
);
3593 qemu_mutex_unlock(&decomp_param
[i
].mutex
);
3595 for (i
= 0; i
< thread_count
; i
++) {
3596 if (!decomp_param
[i
].compbuf
) {
3600 qemu_thread_join(decompress_threads
+ i
);
3601 qemu_mutex_destroy(&decomp_param
[i
].mutex
);
3602 qemu_cond_destroy(&decomp_param
[i
].cond
);
3603 inflateEnd(&decomp_param
[i
].stream
);
3604 g_free(decomp_param
[i
].compbuf
);
3605 decomp_param
[i
].compbuf
= NULL
;
3607 g_free(decompress_threads
);
3608 g_free(decomp_param
);
3609 decompress_threads
= NULL
;
3610 decomp_param
= NULL
;
3614 static int compress_threads_load_setup(QEMUFile
*f
)
3616 int i
, thread_count
;
3618 if (!migrate_use_compression()) {
3622 thread_count
= migrate_decompress_threads();
3623 decompress_threads
= g_new0(QemuThread
, thread_count
);
3624 decomp_param
= g_new0(DecompressParam
, thread_count
);
3625 qemu_mutex_init(&decomp_done_lock
);
3626 qemu_cond_init(&decomp_done_cond
);
3628 for (i
= 0; i
< thread_count
; i
++) {
3629 if (inflateInit(&decomp_param
[i
].stream
) != Z_OK
) {
3633 decomp_param
[i
].compbuf
= g_malloc0(compressBound(TARGET_PAGE_SIZE
));
3634 qemu_mutex_init(&decomp_param
[i
].mutex
);
3635 qemu_cond_init(&decomp_param
[i
].cond
);
3636 decomp_param
[i
].done
= true;
3637 decomp_param
[i
].quit
= false;
3638 qemu_thread_create(decompress_threads
+ i
, "decompress",
3639 do_data_decompress
, decomp_param
+ i
,
3640 QEMU_THREAD_JOINABLE
);
3644 compress_threads_load_cleanup();
3648 static void decompress_data_with_multi_threads(QEMUFile
*f
,
3649 void *host
, int len
)
3651 int idx
, thread_count
;
3653 thread_count
= migrate_decompress_threads();
3654 qemu_mutex_lock(&decomp_done_lock
);
3656 for (idx
= 0; idx
< thread_count
; idx
++) {
3657 if (decomp_param
[idx
].done
) {
3658 decomp_param
[idx
].done
= false;
3659 qemu_mutex_lock(&decomp_param
[idx
].mutex
);
3660 qemu_get_buffer(f
, decomp_param
[idx
].compbuf
, len
);
3661 decomp_param
[idx
].des
= host
;
3662 decomp_param
[idx
].len
= len
;
3663 qemu_cond_signal(&decomp_param
[idx
].cond
);
3664 qemu_mutex_unlock(&decomp_param
[idx
].mutex
);
3668 if (idx
< thread_count
) {
3671 qemu_cond_wait(&decomp_done_cond
, &decomp_done_lock
);
3674 qemu_mutex_unlock(&decomp_done_lock
);
3678 * colo cache: this is for secondary VM, we cache the whole
3679 * memory of the secondary VM, it is need to hold the global lock
3680 * to call this helper.
3682 int colo_init_ram_cache(void)
3687 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3688 block
->colo_cache
= qemu_anon_ram_alloc(block
->used_length
,
3691 if (!block
->colo_cache
) {
3692 error_report("%s: Can't alloc memory for COLO cache of block %s,"
3693 "size 0x" RAM_ADDR_FMT
, __func__
, block
->idstr
,
3694 block
->used_length
);
3697 memcpy(block
->colo_cache
, block
->host
, block
->used_length
);
3701 * Record the dirty pages that sent by PVM, we use this dirty bitmap together
3702 * with to decide which page in cache should be flushed into SVM's RAM. Here
3703 * we use the same name 'ram_bitmap' as for migration.
3705 if (ram_bytes_total()) {
3708 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3709 unsigned long pages
= block
->max_length
>> TARGET_PAGE_BITS
;
3711 block
->bmap
= bitmap_new(pages
);
3712 bitmap_set(block
->bmap
, 0, pages
);
3715 ram_state
= g_new0(RAMState
, 1);
3716 ram_state
->migration_dirty_pages
= 0;
3722 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3723 if (block
->colo_cache
) {
3724 qemu_anon_ram_free(block
->colo_cache
, block
->used_length
);
3725 block
->colo_cache
= NULL
;
3733 /* It is need to hold the global lock to call this helper */
3734 void colo_release_ram_cache(void)
3738 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3739 g_free(block
->bmap
);
3745 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
3746 if (block
->colo_cache
) {
3747 qemu_anon_ram_free(block
->colo_cache
, block
->used_length
);
3748 block
->colo_cache
= NULL
;
3758 * ram_load_setup: Setup RAM for migration incoming side
3760 * Returns zero to indicate success and negative for error
3762 * @f: QEMUFile where to receive the data
3763 * @opaque: RAMState pointer
3765 static int ram_load_setup(QEMUFile
*f
, void *opaque
)
3767 if (compress_threads_load_setup(f
)) {
3771 xbzrle_load_setup();
3772 ramblock_recv_map_init();
3777 static int ram_load_cleanup(void *opaque
)
3781 RAMBLOCK_FOREACH_MIGRATABLE(rb
) {
3782 if (ramblock_is_pmem(rb
)) {
3783 pmem_persist(rb
->host
, rb
->used_length
);
3787 xbzrle_load_cleanup();
3788 compress_threads_load_cleanup();
3790 RAMBLOCK_FOREACH_MIGRATABLE(rb
) {
3791 g_free(rb
->receivedmap
);
3792 rb
->receivedmap
= NULL
;
3799 * ram_postcopy_incoming_init: allocate postcopy data structures
3801 * Returns 0 for success and negative if there was one error
3803 * @mis: current migration incoming state
3805 * Allocate data structures etc needed by incoming migration with
3806 * postcopy-ram. postcopy-ram's similarly names
3807 * postcopy_ram_incoming_init does the work.
3809 int ram_postcopy_incoming_init(MigrationIncomingState
*mis
)
3811 return postcopy_ram_incoming_init(mis
);
3815 * ram_load_postcopy: load a page in postcopy case
3817 * Returns 0 for success or -errno in case of error
3819 * Called in postcopy mode by ram_load().
3820 * rcu_read_lock is taken prior to this being called.
3822 * @f: QEMUFile where to send the data
3824 static int ram_load_postcopy(QEMUFile
*f
)
3826 int flags
= 0, ret
= 0;
3827 bool place_needed
= false;
3828 bool matches_target_page_size
= false;
3829 MigrationIncomingState
*mis
= migration_incoming_get_current();
3830 /* Temporary page that is later 'placed' */
3831 void *postcopy_host_page
= postcopy_get_tmp_page(mis
);
3832 void *last_host
= NULL
;
3833 bool all_zero
= false;
3835 while (!ret
&& !(flags
& RAM_SAVE_FLAG_EOS
)) {
3838 void *page_buffer
= NULL
;
3839 void *place_source
= NULL
;
3840 RAMBlock
*block
= NULL
;
3843 addr
= qemu_get_be64(f
);
3846 * If qemu file error, we should stop here, and then "addr"
3849 ret
= qemu_file_get_error(f
);
3854 flags
= addr
& ~TARGET_PAGE_MASK
;
3855 addr
&= TARGET_PAGE_MASK
;
3857 trace_ram_load_postcopy_loop((uint64_t)addr
, flags
);
3858 place_needed
= false;
3859 if (flags
& (RAM_SAVE_FLAG_ZERO
| RAM_SAVE_FLAG_PAGE
)) {
3860 block
= ram_block_from_stream(f
, flags
);
3862 host
= host_from_ram_block_offset(block
, addr
);
3864 error_report("Illegal RAM offset " RAM_ADDR_FMT
, addr
);
3868 matches_target_page_size
= block
->page_size
== TARGET_PAGE_SIZE
;
3870 * Postcopy requires that we place whole host pages atomically;
3871 * these may be huge pages for RAMBlocks that are backed by
3873 * To make it atomic, the data is read into a temporary page
3874 * that's moved into place later.
3875 * The migration protocol uses, possibly smaller, target-pages
3876 * however the source ensures it always sends all the components
3877 * of a host page in order.
3879 page_buffer
= postcopy_host_page
+
3880 ((uintptr_t)host
& (block
->page_size
- 1));
3881 /* If all TP are zero then we can optimise the place */
3882 if (!((uintptr_t)host
& (block
->page_size
- 1))) {
3885 /* not the 1st TP within the HP */
3886 if (host
!= (last_host
+ TARGET_PAGE_SIZE
)) {
3887 error_report("Non-sequential target page %p/%p",
3896 * If it's the last part of a host page then we place the host
3899 place_needed
= (((uintptr_t)host
+ TARGET_PAGE_SIZE
) &
3900 (block
->page_size
- 1)) == 0;
3901 place_source
= postcopy_host_page
;
3905 switch (flags
& ~RAM_SAVE_FLAG_CONTINUE
) {
3906 case RAM_SAVE_FLAG_ZERO
:
3907 ch
= qemu_get_byte(f
);
3908 memset(page_buffer
, ch
, TARGET_PAGE_SIZE
);
3914 case RAM_SAVE_FLAG_PAGE
:
3916 if (!matches_target_page_size
) {
3917 /* For huge pages, we always use temporary buffer */
3918 qemu_get_buffer(f
, page_buffer
, TARGET_PAGE_SIZE
);
3921 * For small pages that matches target page size, we
3922 * avoid the qemu_file copy. Instead we directly use
3923 * the buffer of QEMUFile to place the page. Note: we
3924 * cannot do any QEMUFile operation before using that
3925 * buffer to make sure the buffer is valid when
3928 qemu_get_buffer_in_place(f
, (uint8_t **)&place_source
,
3932 case RAM_SAVE_FLAG_EOS
:
3934 multifd_recv_sync_main();
3937 error_report("Unknown combination of migration flags: %#x"
3938 " (postcopy mode)", flags
);
3943 /* Detect for any possible file errors */
3944 if (!ret
&& qemu_file_get_error(f
)) {
3945 ret
= qemu_file_get_error(f
);
3948 if (!ret
&& place_needed
) {
3949 /* This gets called at the last target page in the host page */
3950 void *place_dest
= host
+ TARGET_PAGE_SIZE
- block
->page_size
;
3953 ret
= postcopy_place_page_zero(mis
, place_dest
,
3956 ret
= postcopy_place_page(mis
, place_dest
,
3957 place_source
, block
);
3965 static bool postcopy_is_advised(void)
3967 PostcopyState ps
= postcopy_state_get();
3968 return ps
>= POSTCOPY_INCOMING_ADVISE
&& ps
< POSTCOPY_INCOMING_END
;
3971 static bool postcopy_is_running(void)
3973 PostcopyState ps
= postcopy_state_get();
3974 return ps
>= POSTCOPY_INCOMING_LISTENING
&& ps
< POSTCOPY_INCOMING_END
;
3977 static int ram_load(QEMUFile
*f
, void *opaque
, int version_id
)
3979 int flags
= 0, ret
= 0, invalid_flags
= 0;
3980 static uint64_t seq_iter
;
3983 * If system is running in postcopy mode, page inserts to host memory must
3986 bool postcopy_running
= postcopy_is_running();
3987 /* ADVISE is earlier, it shows the source has the postcopy capability on */
3988 bool postcopy_advised
= postcopy_is_advised();
3992 if (version_id
!= 4) {
3996 if (!migrate_use_compression()) {
3997 invalid_flags
|= RAM_SAVE_FLAG_COMPRESS_PAGE
;
3999 /* This RCU critical section can be very long running.
4000 * When RCU reclaims in the code start to become numerous,
4001 * it will be necessary to reduce the granularity of this
4006 if (postcopy_running
) {
4007 ret
= ram_load_postcopy(f
);
4010 while (!postcopy_running
&& !ret
&& !(flags
& RAM_SAVE_FLAG_EOS
)) {
4011 ram_addr_t addr
, total_ram_bytes
;
4015 addr
= qemu_get_be64(f
);
4016 flags
= addr
& ~TARGET_PAGE_MASK
;
4017 addr
&= TARGET_PAGE_MASK
;
4019 if (flags
& invalid_flags
) {
4020 if (flags
& invalid_flags
& RAM_SAVE_FLAG_COMPRESS_PAGE
) {
4021 error_report("Received an unexpected compressed page");
4028 if (flags
& (RAM_SAVE_FLAG_ZERO
| RAM_SAVE_FLAG_PAGE
|
4029 RAM_SAVE_FLAG_COMPRESS_PAGE
| RAM_SAVE_FLAG_XBZRLE
)) {
4030 RAMBlock
*block
= ram_block_from_stream(f
, flags
);
4033 * After going into COLO, we should load the Page into colo_cache.
4035 if (migration_incoming_in_colo_state()) {
4036 host
= colo_cache_from_block_offset(block
, addr
);
4038 host
= host_from_ram_block_offset(block
, addr
);
4041 error_report("Illegal RAM offset " RAM_ADDR_FMT
, addr
);
4046 if (!migration_incoming_in_colo_state()) {
4047 ramblock_recv_bitmap_set(block
, host
);
4050 trace_ram_load_loop(block
->idstr
, (uint64_t)addr
, flags
, host
);
4053 switch (flags
& ~RAM_SAVE_FLAG_CONTINUE
) {
4054 case RAM_SAVE_FLAG_MEM_SIZE
:
4055 /* Synchronize RAM block list */
4056 total_ram_bytes
= addr
;
4057 while (!ret
&& total_ram_bytes
) {
4062 len
= qemu_get_byte(f
);
4063 qemu_get_buffer(f
, (uint8_t *)id
, len
);
4065 length
= qemu_get_be64(f
);
4067 block
= qemu_ram_block_by_name(id
);
4068 if (block
&& !qemu_ram_is_migratable(block
)) {
4069 error_report("block %s should not be migrated !", id
);
4072 if (length
!= block
->used_length
) {
4073 Error
*local_err
= NULL
;
4075 ret
= qemu_ram_resize(block
, length
,
4078 error_report_err(local_err
);
4081 /* For postcopy we need to check hugepage sizes match */
4082 if (postcopy_advised
&&
4083 block
->page_size
!= qemu_host_page_size
) {
4084 uint64_t remote_page_size
= qemu_get_be64(f
);
4085 if (remote_page_size
!= block
->page_size
) {
4086 error_report("Mismatched RAM page size %s "
4087 "(local) %zd != %" PRId64
,
4088 id
, block
->page_size
,
4093 ram_control_load_hook(f
, RAM_CONTROL_BLOCK_REG
,
4096 error_report("Unknown ramblock \"%s\", cannot "
4097 "accept migration", id
);
4101 total_ram_bytes
-= length
;
4105 case RAM_SAVE_FLAG_ZERO
:
4106 ch
= qemu_get_byte(f
);
4107 ram_handle_compressed(host
, ch
, TARGET_PAGE_SIZE
);
4110 case RAM_SAVE_FLAG_PAGE
:
4111 qemu_get_buffer(f
, host
, TARGET_PAGE_SIZE
);
4114 case RAM_SAVE_FLAG_COMPRESS_PAGE
:
4115 len
= qemu_get_be32(f
);
4116 if (len
< 0 || len
> compressBound(TARGET_PAGE_SIZE
)) {
4117 error_report("Invalid compressed data length: %d", len
);
4121 decompress_data_with_multi_threads(f
, host
, len
);
4124 case RAM_SAVE_FLAG_XBZRLE
:
4125 if (load_xbzrle(f
, addr
, host
) < 0) {
4126 error_report("Failed to decompress XBZRLE page at "
4127 RAM_ADDR_FMT
, addr
);
4132 case RAM_SAVE_FLAG_EOS
:
4134 multifd_recv_sync_main();
4137 if (flags
& RAM_SAVE_FLAG_HOOK
) {
4138 ram_control_load_hook(f
, RAM_CONTROL_HOOK
, NULL
);
4140 error_report("Unknown combination of migration flags: %#x",
4146 ret
= qemu_file_get_error(f
);
4150 ret
|= wait_for_decompress_done();
4152 trace_ram_load_complete(ret
, seq_iter
);
4156 static bool ram_has_postcopy(void *opaque
)
4159 RAMBLOCK_FOREACH_MIGRATABLE(rb
) {
4160 if (ramblock_is_pmem(rb
)) {
4161 info_report("Block: %s, host: %p is a nvdimm memory, postcopy"
4162 "is not supported now!", rb
->idstr
, rb
->host
);
4167 return migrate_postcopy_ram();
4170 /* Sync all the dirty bitmap with destination VM. */
4171 static int ram_dirty_bitmap_sync_all(MigrationState
*s
, RAMState
*rs
)
4174 QEMUFile
*file
= s
->to_dst_file
;
4175 int ramblock_count
= 0;
4177 trace_ram_dirty_bitmap_sync_start();
4179 RAMBLOCK_FOREACH_MIGRATABLE(block
) {
4180 qemu_savevm_send_recv_bitmap(file
, block
->idstr
);
4181 trace_ram_dirty_bitmap_request(block
->idstr
);
4185 trace_ram_dirty_bitmap_sync_wait();
4187 /* Wait until all the ramblocks' dirty bitmap synced */
4188 while (ramblock_count
--) {
4189 qemu_sem_wait(&s
->rp_state
.rp_sem
);
4192 trace_ram_dirty_bitmap_sync_complete();
4197 static void ram_dirty_bitmap_reload_notify(MigrationState
*s
)
4199 qemu_sem_post(&s
->rp_state
.rp_sem
);
4203 * Read the received bitmap, revert it as the initial dirty bitmap.
4204 * This is only used when the postcopy migration is paused but wants
4205 * to resume from a middle point.
4207 int ram_dirty_bitmap_reload(MigrationState
*s
, RAMBlock
*block
)
4210 QEMUFile
*file
= s
->rp_state
.from_dst_file
;
4211 unsigned long *le_bitmap
, nbits
= block
->used_length
>> TARGET_PAGE_BITS
;
4212 uint64_t local_size
= DIV_ROUND_UP(nbits
, 8);
4213 uint64_t size
, end_mark
;
4215 trace_ram_dirty_bitmap_reload_begin(block
->idstr
);
4217 if (s
->state
!= MIGRATION_STATUS_POSTCOPY_RECOVER
) {
4218 error_report("%s: incorrect state %s", __func__
,
4219 MigrationStatus_str(s
->state
));
4224 * Note: see comments in ramblock_recv_bitmap_send() on why we
4225 * need the endianess convertion, and the paddings.
4227 local_size
= ROUND_UP(local_size
, 8);
4230 le_bitmap
= bitmap_new(nbits
+ BITS_PER_LONG
);
4232 size
= qemu_get_be64(file
);
4234 /* The size of the bitmap should match with our ramblock */
4235 if (size
!= local_size
) {
4236 error_report("%s: ramblock '%s' bitmap size mismatch "
4237 "(0x%"PRIx64
" != 0x%"PRIx64
")", __func__
,
4238 block
->idstr
, size
, local_size
);
4243 size
= qemu_get_buffer(file
, (uint8_t *)le_bitmap
, local_size
);
4244 end_mark
= qemu_get_be64(file
);
4246 ret
= qemu_file_get_error(file
);
4247 if (ret
|| size
!= local_size
) {
4248 error_report("%s: read bitmap failed for ramblock '%s': %d"
4249 " (size 0x%"PRIx64
", got: 0x%"PRIx64
")",
4250 __func__
, block
->idstr
, ret
, local_size
, size
);
4255 if (end_mark
!= RAMBLOCK_RECV_BITMAP_ENDING
) {
4256 error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIu64
,
4257 __func__
, block
->idstr
, end_mark
);
4263 * Endianess convertion. We are during postcopy (though paused).
4264 * The dirty bitmap won't change. We can directly modify it.
4266 bitmap_from_le(block
->bmap
, le_bitmap
, nbits
);
4269 * What we received is "received bitmap". Revert it as the initial
4270 * dirty bitmap for this ramblock.
4272 bitmap_complement(block
->bmap
, block
->bmap
, nbits
);
4274 trace_ram_dirty_bitmap_reload_complete(block
->idstr
);
4277 * We succeeded to sync bitmap for current ramblock. If this is
4278 * the last one to sync, we need to notify the main send thread.
4280 ram_dirty_bitmap_reload_notify(s
);
4288 static int ram_resume_prepare(MigrationState
*s
, void *opaque
)
4290 RAMState
*rs
= *(RAMState
**)opaque
;
4293 ret
= ram_dirty_bitmap_sync_all(s
, rs
);
4298 ram_state_resume_prepare(rs
, s
->to_dst_file
);
4303 static SaveVMHandlers savevm_ram_handlers
= {
4304 .save_setup
= ram_save_setup
,
4305 .save_live_iterate
= ram_save_iterate
,
4306 .save_live_complete_postcopy
= ram_save_complete
,
4307 .save_live_complete_precopy
= ram_save_complete
,
4308 .has_postcopy
= ram_has_postcopy
,
4309 .save_live_pending
= ram_save_pending
,
4310 .load_state
= ram_load
,
4311 .save_cleanup
= ram_save_cleanup
,
4312 .load_setup
= ram_load_setup
,
4313 .load_cleanup
= ram_load_cleanup
,
4314 .resume_prepare
= ram_resume_prepare
,
4317 void ram_mig_init(void)
4319 qemu_mutex_init(&XBZRLE
.lock
);
4320 register_savevm_live(NULL
, "ram", 0, 4, &savevm_ram_handlers
, &ram_state
);