4 * Copyright (c) 2019-2020 Red Hat Inc
7 * Juan Quintela <quintela@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
13 #include "qemu/osdep.h"
15 #include "exec/target_page.h"
16 #include "sysemu/sysemu.h"
17 #include "exec/ramblock.h"
18 #include "qemu/error-report.h"
19 #include "qapi/error.h"
21 #include "migration.h"
22 #include "migration-stats.h"
25 #include "qemu-file.h"
28 #include "threadinfo.h"
30 #include "qemu/yank.h"
31 #include "io/channel-socket.h"
32 #include "yank_functions.h"
36 #define MULTIFD_MAGIC 0x11223344U
37 #define MULTIFD_VERSION 1
42 unsigned char uuid[16]; /* QemuUUID */
44 uint8_t unused1[7]; /* Reserved for future use */
45 uint64_t unused2[4]; /* Reserved for future use */
46 } __attribute__((packed)) MultiFDInit_t;
48 /* Multifd without compression */
51 * nocomp_send_setup: setup send side
53 * For no compression this function does nothing.
55 * Returns 0 for success or -1 for error
57 * @p: Params for the channel that we are using
58 * @errp: pointer to an error
60 static int nocomp_send_setup(MultiFDSendParams *p, Error **errp)
66 * nocomp_send_cleanup: cleanup send side
68 * For no compression this function does nothing.
70 * @p: Params for the channel that we are using
71 * @errp: pointer to an error
73 static void nocomp_send_cleanup(MultiFDSendParams *p, Error **errp)
79 * nocomp_send_prepare: prepare date to be able to send
81 * For no compression we just have to calculate the size of the
84 * Returns 0 for success or -1 for error
86 * @p: Params for the channel that we are using
87 * @errp: pointer to an error
89 static int nocomp_send_prepare(MultiFDSendParams *p, Error **errp)
91 MultiFDPages_t *pages = p->pages;
93 for (int i = 0; i < p->normal_num; i++) {
94 p->iov[p->iovs_num].iov_base = pages->block->host + p->normal[i];
95 p->iov[p->iovs_num].iov_len = p->page_size;
99 p->next_packet_size = p->normal_num * p->page_size;
100 p->flags |= MULTIFD_FLAG_NOCOMP;
105 * nocomp_recv_setup: setup receive side
107 * For no compression this function does nothing.
109 * Returns 0 for success or -1 for error
111 * @p: Params for the channel that we are using
112 * @errp: pointer to an error
114 static int nocomp_recv_setup(MultiFDRecvParams *p, Error **errp)
120 * nocomp_recv_cleanup: setup receive side
122 * For no compression this function does nothing.
124 * @p: Params for the channel that we are using
126 static void nocomp_recv_cleanup(MultiFDRecvParams *p)
131 * nocomp_recv_pages: read the data from the channel into actual pages
133 * For no compression we just need to read things into the correct place.
135 * Returns 0 for success or -1 for error
137 * @p: Params for the channel that we are using
138 * @errp: pointer to an error
140 static int nocomp_recv_pages(MultiFDRecvParams *p, Error **errp)
142 uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK;
144 if (flags != MULTIFD_FLAG_NOCOMP) {
145 error_setg(errp, "multifd %u: flags received %x flags expected %x",
146 p->id, flags, MULTIFD_FLAG_NOCOMP);
149 for (int i = 0; i < p->normal_num; i++) {
150 p->iov[i].iov_base = p->host + p->normal[i];
151 p->iov[i].iov_len = p->page_size;
153 return qio_channel_readv_all(p->c, p->iov, p->normal_num, errp);
156 static MultiFDMethods multifd_nocomp_ops = {
157 .send_setup = nocomp_send_setup,
158 .send_cleanup = nocomp_send_cleanup,
159 .send_prepare = nocomp_send_prepare,
160 .recv_setup = nocomp_recv_setup,
161 .recv_cleanup = nocomp_recv_cleanup,
162 .recv_pages = nocomp_recv_pages
165 static MultiFDMethods *multifd_ops[MULTIFD_COMPRESSION__MAX] = {
166 [MULTIFD_COMPRESSION_NONE] = &multifd_nocomp_ops,
169 void multifd_register_ops(int method, MultiFDMethods *ops)
171 assert(0 < method && method < MULTIFD_COMPRESSION__MAX);
172 multifd_ops[method] = ops;
175 static int multifd_send_initial_packet(MultiFDSendParams *p, Error **errp)
177 MultiFDInit_t msg = {};
180 msg.magic = cpu_to_be32(MULTIFD_MAGIC);
181 msg.version = cpu_to_be32(MULTIFD_VERSION);
183 memcpy(msg.uuid, &qemu_uuid.data, sizeof(msg.uuid));
185 ret = qio_channel_write_all(p->c, (char *)&msg, sizeof(msg), errp);
192 static int multifd_recv_initial_packet(QIOChannel *c, Error **errp)
197 ret = qio_channel_read_all(c, (char *)&msg, sizeof(msg), errp);
202 msg.magic = be32_to_cpu(msg.magic);
203 msg.version = be32_to_cpu(msg.version);
205 if (msg.magic != MULTIFD_MAGIC) {
206 error_setg(errp, "multifd: received packet magic %x "
207 "expected %x", msg.magic, MULTIFD_MAGIC);
211 if (msg.version != MULTIFD_VERSION) {
212 error_setg(errp, "multifd: received packet version %u "
213 "expected %u", msg.version, MULTIFD_VERSION);
217 if (memcmp(msg.uuid, &qemu_uuid, sizeof(qemu_uuid))) {
218 char *uuid = qemu_uuid_unparse_strdup(&qemu_uuid);
219 char *msg_uuid = qemu_uuid_unparse_strdup((const QemuUUID *)msg.uuid);
221 error_setg(errp, "multifd: received uuid '%s' and expected "
222 "uuid '%s' for channel %hhd", msg_uuid, uuid, msg.id);
228 if (msg.id > migrate_multifd_channels()) {
229 error_setg(errp, "multifd: received channel version %u "
230 "expected %u", msg.version, MULTIFD_VERSION);
237 static MultiFDPages_t *multifd_pages_init(size_t size)
239 MultiFDPages_t *pages = g_new0(MultiFDPages_t, 1);
241 pages->allocated = size;
242 pages->offset = g_new0(ram_addr_t, size);
247 static void multifd_pages_clear(MultiFDPages_t *pages)
250 pages->allocated = 0;
251 pages->packet_num = 0;
253 g_free(pages->offset);
254 pages->offset = NULL;
258 static void multifd_send_fill_packet(MultiFDSendParams *p)
260 MultiFDPacket_t *packet = p->packet;
263 packet->flags = cpu_to_be32(p->flags);
264 packet->pages_alloc = cpu_to_be32(p->pages->allocated);
265 packet->normal_pages = cpu_to_be32(p->normal_num);
266 packet->next_packet_size = cpu_to_be32(p->next_packet_size);
267 packet->packet_num = cpu_to_be64(p->packet_num);
269 if (p->pages->block) {
270 strncpy(packet->ramblock, p->pages->block->idstr, 256);
273 for (i = 0; i < p->normal_num; i++) {
274 /* there are architectures where ram_addr_t is 32 bit */
275 uint64_t temp = p->normal[i];
277 packet->offset[i] = cpu_to_be64(temp);
281 static int multifd_recv_unfill_packet(MultiFDRecvParams *p, Error **errp)
283 MultiFDPacket_t *packet = p->packet;
286 packet->magic = be32_to_cpu(packet->magic);
287 if (packet->magic != MULTIFD_MAGIC) {
288 error_setg(errp, "multifd: received packet "
289 "magic %x and expected magic %x",
290 packet->magic, MULTIFD_MAGIC);
294 packet->version = be32_to_cpu(packet->version);
295 if (packet->version != MULTIFD_VERSION) {
296 error_setg(errp, "multifd: received packet "
297 "version %u and expected version %u",
298 packet->version, MULTIFD_VERSION);
302 p->flags = be32_to_cpu(packet->flags);
304 packet->pages_alloc = be32_to_cpu(packet->pages_alloc);
306 * If we received a packet that is 100 times bigger than expected
307 * just stop migration. It is a magic number.
309 if (packet->pages_alloc > p->page_count) {
310 error_setg(errp, "multifd: received packet "
311 "with size %u and expected a size of %u",
312 packet->pages_alloc, p->page_count) ;
316 p->normal_num = be32_to_cpu(packet->normal_pages);
317 if (p->normal_num > packet->pages_alloc) {
318 error_setg(errp, "multifd: received packet "
319 "with %u pages and expected maximum pages are %u",
320 p->normal_num, packet->pages_alloc) ;
324 p->next_packet_size = be32_to_cpu(packet->next_packet_size);
325 p->packet_num = be64_to_cpu(packet->packet_num);
327 if (p->normal_num == 0) {
331 /* make sure that ramblock is 0 terminated */
332 packet->ramblock[255] = 0;
333 p->block = qemu_ram_block_by_name(packet->ramblock);
335 error_setg(errp, "multifd: unknown ram block %s",
340 p->host = p->block->host;
341 for (i = 0; i < p->normal_num; i++) {
342 uint64_t offset = be64_to_cpu(packet->offset[i]);
344 if (offset > (p->block->used_length - p->page_size)) {
345 error_setg(errp, "multifd: offset too long %" PRIu64
346 " (max " RAM_ADDR_FMT ")",
347 offset, p->block->used_length);
350 p->normal[i] = offset;
357 MultiFDSendParams *params;
358 /* array of pages to sent */
359 MultiFDPages_t *pages;
360 /* global number of generated multifd packets */
362 /* send channels ready */
363 QemuSemaphore channels_ready;
365 * Have we already run terminate threads. There is a race when it
366 * happens that we got one error while we are exiting.
367 * We will use atomic operations. Only valid values are 0 and 1.
372 } *multifd_send_state;
375 * How we use multifd_send_state->pages and channel->pages?
377 * We create a pages for each channel, and a main one. Each time that
378 * we need to send a batch of pages we interchange the ones between
379 * multifd_send_state and the channel that is sending it. There are
380 * two reasons for that:
381 * - to not have to do so many mallocs during migration
382 * - to make easier to know what to free at the end of migration
384 * This way we always know who is the owner of each "pages" struct,
385 * and we don't need any locking. It belongs to the migration thread
386 * or to the channel thread. Switching is safe because the migration
387 * thread is using the channel mutex when changing it, and the channel
388 * have to had finish with its own, otherwise pending_job can't be
392 static int multifd_send_pages(QEMUFile *f)
395 static int next_channel;
396 MultiFDSendParams *p = NULL; /* make happy gcc */
397 MultiFDPages_t *pages = multifd_send_state->pages;
398 uint64_t transferred;
400 if (qatomic_read(&multifd_send_state->exiting)) {
404 qemu_sem_wait(&multifd_send_state->channels_ready);
406 * next_channel can remain from a previous migration that was
407 * using more channels, so ensure it doesn't overflow if the
408 * limit is lower now.
410 next_channel %= migrate_multifd_channels();
411 for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) {
412 p = &multifd_send_state->params[i];
414 qemu_mutex_lock(&p->mutex);
416 error_report("%s: channel %d has already quit!", __func__, i);
417 qemu_mutex_unlock(&p->mutex);
420 if (!p->pending_job) {
422 next_channel = (i + 1) % migrate_multifd_channels();
425 qemu_mutex_unlock(&p->mutex);
427 assert(!p->pages->num);
428 assert(!p->pages->block);
430 p->packet_num = multifd_send_state->packet_num++;
431 multifd_send_state->pages = p->pages;
433 transferred = ((uint64_t) pages->num) * p->page_size + p->packet_len;
434 qemu_file_acct_rate_limit(f, transferred);
435 qemu_mutex_unlock(&p->mutex);
436 stat64_add(&mig_stats.transferred, transferred);
437 stat64_add(&mig_stats.multifd_bytes, transferred);
438 qemu_sem_post(&p->sem);
443 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
445 MultiFDPages_t *pages = multifd_send_state->pages;
446 bool changed = false;
449 pages->block = block;
452 if (pages->block == block) {
453 pages->offset[pages->num] = offset;
456 if (pages->num < pages->allocated) {
463 if (multifd_send_pages(f) < 0) {
468 return multifd_queue_page(f, block, offset);
474 static void multifd_send_terminate_threads(Error *err)
478 trace_multifd_send_terminate_threads(err != NULL);
481 MigrationState *s = migrate_get_current();
482 migrate_set_error(s, err);
483 if (s->state == MIGRATION_STATUS_SETUP ||
484 s->state == MIGRATION_STATUS_PRE_SWITCHOVER ||
485 s->state == MIGRATION_STATUS_DEVICE ||
486 s->state == MIGRATION_STATUS_ACTIVE) {
487 migrate_set_state(&s->state, s->state,
488 MIGRATION_STATUS_FAILED);
493 * We don't want to exit each threads twice. Depending on where
494 * we get the error, or if there are two independent errors in two
495 * threads at the same time, we can end calling this function
498 if (qatomic_xchg(&multifd_send_state->exiting, 1)) {
502 for (i = 0; i < migrate_multifd_channels(); i++) {
503 MultiFDSendParams *p = &multifd_send_state->params[i];
505 qemu_mutex_lock(&p->mutex);
507 qemu_sem_post(&p->sem);
509 qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
511 qemu_mutex_unlock(&p->mutex);
515 void multifd_save_cleanup(void)
519 if (!migrate_multifd()) {
522 multifd_send_terminate_threads(NULL);
523 for (i = 0; i < migrate_multifd_channels(); i++) {
524 MultiFDSendParams *p = &multifd_send_state->params[i];
527 qemu_thread_join(&p->thread);
530 for (i = 0; i < migrate_multifd_channels(); i++) {
531 MultiFDSendParams *p = &multifd_send_state->params[i];
532 Error *local_err = NULL;
534 if (p->registered_yank) {
535 migration_ioc_unregister_yank(p->c);
537 socket_send_channel_destroy(p->c);
539 qemu_mutex_destroy(&p->mutex);
540 qemu_sem_destroy(&p->sem);
541 qemu_sem_destroy(&p->sem_sync);
544 multifd_pages_clear(p->pages);
553 multifd_send_state->ops->send_cleanup(p, &local_err);
555 migrate_set_error(migrate_get_current(), local_err);
556 error_free(local_err);
559 qemu_sem_destroy(&multifd_send_state->channels_ready);
560 g_free(multifd_send_state->params);
561 multifd_send_state->params = NULL;
562 multifd_pages_clear(multifd_send_state->pages);
563 multifd_send_state->pages = NULL;
564 g_free(multifd_send_state);
565 multifd_send_state = NULL;
568 static int multifd_zero_copy_flush(QIOChannel *c)
573 ret = qio_channel_flush(c, &err);
575 error_report_err(err);
579 stat64_add(&mig_stats.dirty_sync_missed_zero_copy, 1);
585 int multifd_send_sync_main(QEMUFile *f)
588 bool flush_zero_copy;
590 if (!migrate_multifd()) {
593 if (multifd_send_state->pages->num) {
594 if (multifd_send_pages(f) < 0) {
595 error_report("%s: multifd_send_pages fail", __func__);
601 * When using zero-copy, it's necessary to flush the pages before any of
602 * the pages can be sent again, so we'll make sure the new version of the
603 * pages will always arrive _later_ than the old pages.
605 * Currently we achieve this by flushing the zero-page requested writes
606 * per ram iteration, but in the future we could potentially optimize it
607 * to be less frequent, e.g. only after we finished one whole scanning of
608 * all the dirty bitmaps.
611 flush_zero_copy = migrate_zero_copy_send();
613 for (i = 0; i < migrate_multifd_channels(); i++) {
614 MultiFDSendParams *p = &multifd_send_state->params[i];
616 trace_multifd_send_sync_main_signal(p->id);
618 qemu_mutex_lock(&p->mutex);
621 error_report("%s: channel %d has already quit", __func__, i);
622 qemu_mutex_unlock(&p->mutex);
626 p->packet_num = multifd_send_state->packet_num++;
627 p->flags |= MULTIFD_FLAG_SYNC;
629 qemu_mutex_unlock(&p->mutex);
630 qemu_sem_post(&p->sem);
632 for (i = 0; i < migrate_multifd_channels(); i++) {
633 MultiFDSendParams *p = &multifd_send_state->params[i];
635 qemu_sem_wait(&multifd_send_state->channels_ready);
636 trace_multifd_send_sync_main_wait(p->id);
637 qemu_sem_wait(&p->sem_sync);
639 if (flush_zero_copy && p->c && (multifd_zero_copy_flush(p->c) < 0)) {
643 trace_multifd_send_sync_main(multifd_send_state->packet_num);
648 static void *multifd_send_thread(void *opaque)
650 MultiFDSendParams *p = opaque;
651 MigrationThread *thread = NULL;
652 Error *local_err = NULL;
654 bool use_zero_copy_send = migrate_zero_copy_send();
656 thread = MigrationThreadAdd(p->name, qemu_get_thread_id());
658 trace_multifd_send_thread_start(p->id);
659 rcu_register_thread();
661 if (multifd_send_initial_packet(p, &local_err) < 0) {
669 qemu_sem_post(&multifd_send_state->channels_ready);
670 qemu_sem_wait(&p->sem);
672 if (qatomic_read(&multifd_send_state->exiting)) {
675 qemu_mutex_lock(&p->mutex);
677 if (p->pending_job) {
678 uint64_t packet_num = p->packet_num;
682 if (use_zero_copy_send) {
688 for (int i = 0; i < p->pages->num; i++) {
689 p->normal[p->normal_num] = p->pages->offset[i];
694 ret = multifd_send_state->ops->send_prepare(p, &local_err);
696 qemu_mutex_unlock(&p->mutex);
700 multifd_send_fill_packet(p);
704 p->total_normal_pages += p->normal_num;
706 p->pages->block = NULL;
707 qemu_mutex_unlock(&p->mutex);
709 trace_multifd_send(p->id, packet_num, p->normal_num, flags,
710 p->next_packet_size);
712 if (use_zero_copy_send) {
713 /* Send header first, without zerocopy */
714 ret = qio_channel_write_all(p->c, (void *)p->packet,
715 p->packet_len, &local_err);
720 /* Send header using the same writev call */
721 p->iov[0].iov_len = p->packet_len;
722 p->iov[0].iov_base = p->packet;
725 ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
726 0, p->write_flags, &local_err);
731 qemu_mutex_lock(&p->mutex);
733 qemu_mutex_unlock(&p->mutex);
735 if (flags & MULTIFD_FLAG_SYNC) {
736 qemu_sem_post(&p->sem_sync);
738 } else if (p->quit) {
739 qemu_mutex_unlock(&p->mutex);
742 qemu_mutex_unlock(&p->mutex);
743 /* sometimes there are spurious wakeups */
749 trace_multifd_send_error(p->id);
750 multifd_send_terminate_threads(local_err);
751 error_free(local_err);
755 * Error happen, I will exit, but I can't just leave, tell
756 * who pay attention to me.
759 qemu_sem_post(&p->sem_sync);
760 qemu_sem_post(&multifd_send_state->channels_ready);
763 qemu_mutex_lock(&p->mutex);
765 qemu_mutex_unlock(&p->mutex);
767 rcu_unregister_thread();
768 MigrationThreadDel(thread);
769 trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages);
774 static bool multifd_channel_connect(MultiFDSendParams *p,
778 static void multifd_tls_outgoing_handshake(QIOTask *task,
781 MultiFDSendParams *p = opaque;
782 QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
785 if (qio_task_propagate_error(task, &err)) {
786 trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err));
788 trace_multifd_tls_outgoing_handshake_complete(ioc);
791 if (!multifd_channel_connect(p, ioc, err)) {
793 * Error happen, mark multifd_send_thread status as 'quit' although it
794 * is not created, and then tell who pay attention to me.
797 qemu_sem_post(&multifd_send_state->channels_ready);
798 qemu_sem_post(&p->sem_sync);
802 static void *multifd_tls_handshake_thread(void *opaque)
804 MultiFDSendParams *p = opaque;
805 QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
807 qio_channel_tls_handshake(tioc,
808 multifd_tls_outgoing_handshake,
815 static void multifd_tls_channel_connect(MultiFDSendParams *p,
819 MigrationState *s = migrate_get_current();
820 const char *hostname = s->hostname;
823 tioc = migration_tls_client_create(ioc, hostname, errp);
828 object_unref(OBJECT(ioc));
829 trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
830 qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
831 p->c = QIO_CHANNEL(tioc);
832 qemu_thread_create(&p->thread, "multifd-tls-handshake-worker",
833 multifd_tls_handshake_thread, p,
834 QEMU_THREAD_JOINABLE);
837 static bool multifd_channel_connect(MultiFDSendParams *p,
841 trace_multifd_set_outgoing_channel(
842 ioc, object_get_typename(OBJECT(ioc)),
843 migrate_get_current()->hostname, error);
848 if (migrate_channel_requires_tls_upgrade(ioc)) {
849 multifd_tls_channel_connect(p, ioc, &error);
852 * tls_channel_connect will call back to this
853 * function after the TLS handshake,
854 * so we mustn't call multifd_send_thread until then
861 migration_ioc_register_yank(ioc);
862 p->registered_yank = true;
864 qemu_thread_create(&p->thread, p->name, multifd_send_thread, p,
865 QEMU_THREAD_JOINABLE);
870 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p,
871 QIOChannel *ioc, Error *err)
873 migrate_set_error(migrate_get_current(), err);
874 /* Error happen, we need to tell who pay attention to me */
875 qemu_sem_post(&multifd_send_state->channels_ready);
876 qemu_sem_post(&p->sem_sync);
878 * Although multifd_send_thread is not created, but main migration
879 * thread neet to judge whether it is running, so we need to mark
883 object_unref(OBJECT(ioc));
887 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque)
889 MultiFDSendParams *p = opaque;
890 QIOChannel *sioc = QIO_CHANNEL(qio_task_get_source(task));
891 Error *local_err = NULL;
893 trace_multifd_new_send_channel_async(p->id);
894 if (!qio_task_propagate_error(task, &local_err)) {
895 p->c = QIO_CHANNEL(sioc);
896 qio_channel_set_delay(p->c, false);
898 if (multifd_channel_connect(p, sioc, local_err)) {
903 multifd_new_send_channel_cleanup(p, sioc, local_err);
906 int multifd_save_setup(Error **errp)
909 uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
912 if (!migrate_multifd()) {
916 thread_count = migrate_multifd_channels();
917 multifd_send_state = g_malloc0(sizeof(*multifd_send_state));
918 multifd_send_state->params = g_new0(MultiFDSendParams, thread_count);
919 multifd_send_state->pages = multifd_pages_init(page_count);
920 qemu_sem_init(&multifd_send_state->channels_ready, 0);
921 qatomic_set(&multifd_send_state->exiting, 0);
922 multifd_send_state->ops = multifd_ops[migrate_multifd_compression()];
924 for (i = 0; i < thread_count; i++) {
925 MultiFDSendParams *p = &multifd_send_state->params[i];
927 qemu_mutex_init(&p->mutex);
928 qemu_sem_init(&p->sem, 0);
929 qemu_sem_init(&p->sem_sync, 0);
933 p->pages = multifd_pages_init(page_count);
934 p->packet_len = sizeof(MultiFDPacket_t)
935 + sizeof(uint64_t) * page_count;
936 p->packet = g_malloc0(p->packet_len);
937 p->packet->magic = cpu_to_be32(MULTIFD_MAGIC);
938 p->packet->version = cpu_to_be32(MULTIFD_VERSION);
939 p->name = g_strdup_printf("multifdsend_%d", i);
940 /* We need one extra place for the packet header */
941 p->iov = g_new0(struct iovec, page_count + 1);
942 p->normal = g_new0(ram_addr_t, page_count);
943 p->page_size = qemu_target_page_size();
944 p->page_count = page_count;
946 if (migrate_zero_copy_send()) {
947 p->write_flags = QIO_CHANNEL_WRITE_FLAG_ZERO_COPY;
952 socket_send_channel_create(multifd_new_send_channel_async, p);
955 for (i = 0; i < thread_count; i++) {
956 MultiFDSendParams *p = &multifd_send_state->params[i];
957 Error *local_err = NULL;
960 ret = multifd_send_state->ops->send_setup(p, &local_err);
962 error_propagate(errp, local_err);
970 MultiFDRecvParams *params;
971 /* number of created threads */
973 /* syncs main thread and channels */
974 QemuSemaphore sem_sync;
975 /* global number of generated multifd packets */
979 } *multifd_recv_state;
981 static void multifd_recv_terminate_threads(Error *err)
985 trace_multifd_recv_terminate_threads(err != NULL);
988 MigrationState *s = migrate_get_current();
989 migrate_set_error(s, err);
990 if (s->state == MIGRATION_STATUS_SETUP ||
991 s->state == MIGRATION_STATUS_ACTIVE) {
992 migrate_set_state(&s->state, s->state,
993 MIGRATION_STATUS_FAILED);
997 for (i = 0; i < migrate_multifd_channels(); i++) {
998 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1000 qemu_mutex_lock(&p->mutex);
1003 * We could arrive here for two reasons:
1004 * - normal quit, i.e. everything went fine, just finished
1005 * - error quit: We close the channels so the channel threads
1006 * finish the qio_channel_read_all_eof()
1009 qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1011 qemu_mutex_unlock(&p->mutex);
1015 void multifd_load_shutdown(void)
1017 if (migrate_multifd()) {
1018 multifd_recv_terminate_threads(NULL);
1022 void multifd_load_cleanup(void)
1026 if (!migrate_multifd()) {
1029 multifd_recv_terminate_threads(NULL);
1030 for (i = 0; i < migrate_multifd_channels(); i++) {
1031 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1035 * multifd_recv_thread may hung at MULTIFD_FLAG_SYNC handle code,
1036 * however try to wakeup it without harm in cleanup phase.
1038 qemu_sem_post(&p->sem_sync);
1041 qemu_thread_join(&p->thread);
1043 for (i = 0; i < migrate_multifd_channels(); i++) {
1044 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1046 migration_ioc_unregister_yank(p->c);
1047 object_unref(OBJECT(p->c));
1049 qemu_mutex_destroy(&p->mutex);
1050 qemu_sem_destroy(&p->sem_sync);
1060 multifd_recv_state->ops->recv_cleanup(p);
1062 qemu_sem_destroy(&multifd_recv_state->sem_sync);
1063 g_free(multifd_recv_state->params);
1064 multifd_recv_state->params = NULL;
1065 g_free(multifd_recv_state);
1066 multifd_recv_state = NULL;
1069 void multifd_recv_sync_main(void)
1073 if (!migrate_multifd()) {
1076 for (i = 0; i < migrate_multifd_channels(); i++) {
1077 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1079 trace_multifd_recv_sync_main_wait(p->id);
1080 qemu_sem_wait(&multifd_recv_state->sem_sync);
1082 for (i = 0; i < migrate_multifd_channels(); i++) {
1083 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1085 WITH_QEMU_LOCK_GUARD(&p->mutex) {
1086 if (multifd_recv_state->packet_num < p->packet_num) {
1087 multifd_recv_state->packet_num = p->packet_num;
1090 trace_multifd_recv_sync_main_signal(p->id);
1091 qemu_sem_post(&p->sem_sync);
1093 trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
1096 static void *multifd_recv_thread(void *opaque)
1098 MultiFDRecvParams *p = opaque;
1099 Error *local_err = NULL;
1102 trace_multifd_recv_thread_start(p->id);
1103 rcu_register_thread();
1112 ret = qio_channel_read_all_eof(p->c, (void *)p->packet,
1113 p->packet_len, &local_err);
1114 if (ret == 0 || ret == -1) { /* 0: EOF -1: Error */
1118 qemu_mutex_lock(&p->mutex);
1119 ret = multifd_recv_unfill_packet(p, &local_err);
1121 qemu_mutex_unlock(&p->mutex);
1126 /* recv methods don't know how to handle the SYNC flag */
1127 p->flags &= ~MULTIFD_FLAG_SYNC;
1128 trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags,
1129 p->next_packet_size);
1131 p->total_normal_pages += p->normal_num;
1132 qemu_mutex_unlock(&p->mutex);
1134 if (p->normal_num) {
1135 ret = multifd_recv_state->ops->recv_pages(p, &local_err);
1141 if (flags & MULTIFD_FLAG_SYNC) {
1142 qemu_sem_post(&multifd_recv_state->sem_sync);
1143 qemu_sem_wait(&p->sem_sync);
1148 multifd_recv_terminate_threads(local_err);
1149 error_free(local_err);
1151 qemu_mutex_lock(&p->mutex);
1153 qemu_mutex_unlock(&p->mutex);
1155 rcu_unregister_thread();
1156 trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages);
1161 int multifd_load_setup(Error **errp)
1164 uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
1168 * Return successfully if multiFD recv state is already initialised
1169 * or multiFD is not enabled.
1171 if (multifd_recv_state || !migrate_multifd()) {
1175 thread_count = migrate_multifd_channels();
1176 multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state));
1177 multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count);
1178 qatomic_set(&multifd_recv_state->count, 0);
1179 qemu_sem_init(&multifd_recv_state->sem_sync, 0);
1180 multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()];
1182 for (i = 0; i < thread_count; i++) {
1183 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1185 qemu_mutex_init(&p->mutex);
1186 qemu_sem_init(&p->sem_sync, 0);
1189 p->packet_len = sizeof(MultiFDPacket_t)
1190 + sizeof(uint64_t) * page_count;
1191 p->packet = g_malloc0(p->packet_len);
1192 p->name = g_strdup_printf("multifdrecv_%d", i);
1193 p->iov = g_new0(struct iovec, page_count);
1194 p->normal = g_new0(ram_addr_t, page_count);
1195 p->page_count = page_count;
1196 p->page_size = qemu_target_page_size();
1199 for (i = 0; i < thread_count; i++) {
1200 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1201 Error *local_err = NULL;
1204 ret = multifd_recv_state->ops->recv_setup(p, &local_err);
1206 error_propagate(errp, local_err);
1213 bool multifd_recv_all_channels_created(void)
1215 int thread_count = migrate_multifd_channels();
1217 if (!migrate_multifd()) {
1221 if (!multifd_recv_state) {
1222 /* Called before any connections created */
1226 return thread_count == qatomic_read(&multifd_recv_state->count);
1230 * Try to receive all multifd channels to get ready for the migration.
1231 * Sets @errp when failing to receive the current channel.
1233 void multifd_recv_new_channel(QIOChannel *ioc, Error **errp)
1235 MultiFDRecvParams *p;
1236 Error *local_err = NULL;
1239 id = multifd_recv_initial_packet(ioc, &local_err);
1241 multifd_recv_terminate_threads(local_err);
1242 error_propagate_prepend(errp, local_err,
1243 "failed to receive packet"
1244 " via multifd channel %d: ",
1245 qatomic_read(&multifd_recv_state->count));
1248 trace_multifd_recv_new_channel(id);
1250 p = &multifd_recv_state->params[id];
1252 error_setg(&local_err, "multifd: received id '%d' already setup'",
1254 multifd_recv_terminate_threads(local_err);
1255 error_propagate(errp, local_err);
1259 object_ref(OBJECT(ioc));
1260 /* initial packet */
1264 qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p,
1265 QEMU_THREAD_JOINABLE);
1266 qatomic_inc(&multifd_recv_state->count);