hw/ppc/ppc405_boards: Initialize g_autofree pointer
[qemu.git] / migration / multifd.c
blob76b57a7177871ee1082151e651c6d6519338e82c
1 /*
2 * Multifd common code
4 * Copyright (c) 2019-2020 Red Hat Inc
6 * Authors:
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"
14 #include "qemu/rcu.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"
20 #include "ram.h"
21 #include "migration.h"
22 #include "socket.h"
23 #include "tls.h"
24 #include "qemu-file.h"
25 #include "trace.h"
26 #include "multifd.h"
28 #include "qemu/yank.h"
29 #include "io/channel-socket.h"
30 #include "yank_functions.h"
32 /* Multiple fd's */
34 #define MULTIFD_MAGIC 0x11223344U
35 #define MULTIFD_VERSION 1
37 typedef struct {
38 uint32_t magic;
39 uint32_t version;
40 unsigned char uuid[16]; /* QemuUUID */
41 uint8_t id;
42 uint8_t unused1[7]; /* Reserved for future use */
43 uint64_t unused2[4]; /* Reserved for future use */
44 } __attribute__((packed)) MultiFDInit_t;
46 /* Multifd without compression */
48 /**
49 * nocomp_send_setup: setup send side
51 * For no compression this function does nothing.
53 * Returns 0 for success or -1 for error
55 * @p: Params for the channel that we are using
56 * @errp: pointer to an error
58 static int nocomp_send_setup(MultiFDSendParams *p, Error **errp)
60 return 0;
63 /**
64 * nocomp_send_cleanup: cleanup send side
66 * For no compression this function does nothing.
68 * @p: Params for the channel that we are using
69 * @errp: pointer to an error
71 static void nocomp_send_cleanup(MultiFDSendParams *p, Error **errp)
73 return;
76 /**
77 * nocomp_send_prepare: prepare date to be able to send
79 * For no compression we just have to calculate the size of the
80 * packet.
82 * Returns 0 for success or -1 for error
84 * @p: Params for the channel that we are using
85 * @errp: pointer to an error
87 static int nocomp_send_prepare(MultiFDSendParams *p, Error **errp)
89 MultiFDPages_t *pages = p->pages;
90 size_t page_size = qemu_target_page_size();
92 for (int i = 0; i < p->normal_num; i++) {
93 p->iov[p->iovs_num].iov_base = pages->block->host + p->normal[i];
94 p->iov[p->iovs_num].iov_len = page_size;
95 p->iovs_num++;
98 p->next_packet_size = p->normal_num * page_size;
99 p->flags |= MULTIFD_FLAG_NOCOMP;
100 return 0;
104 * nocomp_recv_setup: setup receive side
106 * For no compression this function does nothing.
108 * Returns 0 for success or -1 for error
110 * @p: Params for the channel that we are using
111 * @errp: pointer to an error
113 static int nocomp_recv_setup(MultiFDRecvParams *p, Error **errp)
115 return 0;
119 * nocomp_recv_cleanup: setup receive side
121 * For no compression this function does nothing.
123 * @p: Params for the channel that we are using
125 static void nocomp_recv_cleanup(MultiFDRecvParams *p)
130 * nocomp_recv_pages: read the data from the channel into actual pages
132 * For no compression we just need to read things into the correct place.
134 * Returns 0 for success or -1 for error
136 * @p: Params for the channel that we are using
137 * @errp: pointer to an error
139 static int nocomp_recv_pages(MultiFDRecvParams *p, Error **errp)
141 uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK;
142 size_t page_size = qemu_target_page_size();
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);
147 return -1;
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 = 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 = {};
178 int ret;
180 msg.magic = cpu_to_be32(MULTIFD_MAGIC);
181 msg.version = cpu_to_be32(MULTIFD_VERSION);
182 msg.id = p->id;
183 memcpy(msg.uuid, &qemu_uuid.data, sizeof(msg.uuid));
185 ret = qio_channel_write_all(p->c, (char *)&msg, sizeof(msg), errp);
186 if (ret != 0) {
187 return -1;
189 return 0;
192 static int multifd_recv_initial_packet(QIOChannel *c, Error **errp)
194 MultiFDInit_t msg;
195 int ret;
197 ret = qio_channel_read_all(c, (char *)&msg, sizeof(msg), errp);
198 if (ret != 0) {
199 return -1;
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);
208 return -1;
211 if (msg.version != MULTIFD_VERSION) {
212 error_setg(errp, "multifd: received packet version %u "
213 "expected %u", msg.version, MULTIFD_VERSION);
214 return -1;
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);
223 g_free(uuid);
224 g_free(msg_uuid);
225 return -1;
228 if (msg.id > migrate_multifd_channels()) {
229 error_setg(errp, "multifd: received channel version %u "
230 "expected %u", msg.version, MULTIFD_VERSION);
231 return -1;
234 return msg.id;
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);
244 return pages;
247 static void multifd_pages_clear(MultiFDPages_t *pages)
249 pages->num = 0;
250 pages->allocated = 0;
251 pages->packet_num = 0;
252 pages->block = NULL;
253 g_free(pages->offset);
254 pages->offset = NULL;
255 g_free(pages);
258 static void multifd_send_fill_packet(MultiFDSendParams *p)
260 MultiFDPacket_t *packet = p->packet;
261 int i;
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;
284 size_t page_size = qemu_target_page_size();
285 uint32_t page_count = MULTIFD_PACKET_SIZE / page_size;
286 RAMBlock *block;
287 int i;
289 packet->magic = be32_to_cpu(packet->magic);
290 if (packet->magic != MULTIFD_MAGIC) {
291 error_setg(errp, "multifd: received packet "
292 "magic %x and expected magic %x",
293 packet->magic, MULTIFD_MAGIC);
294 return -1;
297 packet->version = be32_to_cpu(packet->version);
298 if (packet->version != MULTIFD_VERSION) {
299 error_setg(errp, "multifd: received packet "
300 "version %u and expected version %u",
301 packet->version, MULTIFD_VERSION);
302 return -1;
305 p->flags = be32_to_cpu(packet->flags);
307 packet->pages_alloc = be32_to_cpu(packet->pages_alloc);
309 * If we received a packet that is 100 times bigger than expected
310 * just stop migration. It is a magic number.
312 if (packet->pages_alloc > page_count) {
313 error_setg(errp, "multifd: received packet "
314 "with size %u and expected a size of %u",
315 packet->pages_alloc, page_count) ;
316 return -1;
319 p->normal_num = be32_to_cpu(packet->normal_pages);
320 if (p->normal_num > packet->pages_alloc) {
321 error_setg(errp, "multifd: received packet "
322 "with %u pages and expected maximum pages are %u",
323 p->normal_num, packet->pages_alloc) ;
324 return -1;
327 p->next_packet_size = be32_to_cpu(packet->next_packet_size);
328 p->packet_num = be64_to_cpu(packet->packet_num);
330 if (p->normal_num == 0) {
331 return 0;
334 /* make sure that ramblock is 0 terminated */
335 packet->ramblock[255] = 0;
336 block = qemu_ram_block_by_name(packet->ramblock);
337 if (!block) {
338 error_setg(errp, "multifd: unknown ram block %s",
339 packet->ramblock);
340 return -1;
343 p->host = block->host;
344 for (i = 0; i < p->normal_num; i++) {
345 uint64_t offset = be64_to_cpu(packet->offset[i]);
347 if (offset > (block->used_length - page_size)) {
348 error_setg(errp, "multifd: offset too long %" PRIu64
349 " (max " RAM_ADDR_FMT ")",
350 offset, block->used_length);
351 return -1;
353 p->normal[i] = offset;
356 return 0;
359 struct {
360 MultiFDSendParams *params;
361 /* array of pages to sent */
362 MultiFDPages_t *pages;
363 /* global number of generated multifd packets */
364 uint64_t packet_num;
365 /* send channels ready */
366 QemuSemaphore channels_ready;
368 * Have we already run terminate threads. There is a race when it
369 * happens that we got one error while we are exiting.
370 * We will use atomic operations. Only valid values are 0 and 1.
372 int exiting;
373 /* multifd ops */
374 MultiFDMethods *ops;
375 } *multifd_send_state;
378 * How we use multifd_send_state->pages and channel->pages?
380 * We create a pages for each channel, and a main one. Each time that
381 * we need to send a batch of pages we interchange the ones between
382 * multifd_send_state and the channel that is sending it. There are
383 * two reasons for that:
384 * - to not have to do so many mallocs during migration
385 * - to make easier to know what to free at the end of migration
387 * This way we always know who is the owner of each "pages" struct,
388 * and we don't need any locking. It belongs to the migration thread
389 * or to the channel thread. Switching is safe because the migration
390 * thread is using the channel mutex when changing it, and the channel
391 * have to had finish with its own, otherwise pending_job can't be
392 * false.
395 static int multifd_send_pages(QEMUFile *f)
397 int i;
398 static int next_channel;
399 MultiFDSendParams *p = NULL; /* make happy gcc */
400 MultiFDPages_t *pages = multifd_send_state->pages;
401 uint64_t transferred;
403 if (qatomic_read(&multifd_send_state->exiting)) {
404 return -1;
407 qemu_sem_wait(&multifd_send_state->channels_ready);
409 * next_channel can remain from a previous migration that was
410 * using more channels, so ensure it doesn't overflow if the
411 * limit is lower now.
413 next_channel %= migrate_multifd_channels();
414 for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) {
415 p = &multifd_send_state->params[i];
417 qemu_mutex_lock(&p->mutex);
418 if (p->quit) {
419 error_report("%s: channel %d has already quit!", __func__, i);
420 qemu_mutex_unlock(&p->mutex);
421 return -1;
423 if (!p->pending_job) {
424 p->pending_job++;
425 next_channel = (i + 1) % migrate_multifd_channels();
426 break;
428 qemu_mutex_unlock(&p->mutex);
430 assert(!p->pages->num);
431 assert(!p->pages->block);
433 p->packet_num = multifd_send_state->packet_num++;
434 multifd_send_state->pages = p->pages;
435 p->pages = pages;
436 transferred = ((uint64_t) pages->num) * qemu_target_page_size()
437 + p->packet_len;
438 qemu_file_update_transfer(f, transferred);
439 ram_counters.multifd_bytes += transferred;
440 ram_counters.transferred += transferred;
441 qemu_mutex_unlock(&p->mutex);
442 qemu_sem_post(&p->sem);
444 return 1;
447 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
449 MultiFDPages_t *pages = multifd_send_state->pages;
451 if (!pages->block) {
452 pages->block = block;
455 if (pages->block == block) {
456 pages->offset[pages->num] = offset;
457 pages->num++;
459 if (pages->num < pages->allocated) {
460 return 1;
464 if (multifd_send_pages(f) < 0) {
465 return -1;
468 if (pages->block != block) {
469 return multifd_queue_page(f, block, offset);
472 return 1;
475 static void multifd_send_terminate_threads(Error *err)
477 int i;
479 trace_multifd_send_terminate_threads(err != NULL);
481 if (err) {
482 MigrationState *s = migrate_get_current();
483 migrate_set_error(s, err);
484 if (s->state == MIGRATION_STATUS_SETUP ||
485 s->state == MIGRATION_STATUS_PRE_SWITCHOVER ||
486 s->state == MIGRATION_STATUS_DEVICE ||
487 s->state == MIGRATION_STATUS_ACTIVE) {
488 migrate_set_state(&s->state, s->state,
489 MIGRATION_STATUS_FAILED);
494 * We don't want to exit each threads twice. Depending on where
495 * we get the error, or if there are two independent errors in two
496 * threads at the same time, we can end calling this function
497 * twice.
499 if (qatomic_xchg(&multifd_send_state->exiting, 1)) {
500 return;
503 for (i = 0; i < migrate_multifd_channels(); i++) {
504 MultiFDSendParams *p = &multifd_send_state->params[i];
506 qemu_mutex_lock(&p->mutex);
507 p->quit = true;
508 qemu_sem_post(&p->sem);
509 if (p->c) {
510 qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
512 qemu_mutex_unlock(&p->mutex);
516 void multifd_save_cleanup(void)
518 int i;
520 if (!migrate_use_multifd() || !migrate_multifd_is_allowed()) {
521 return;
523 multifd_send_terminate_threads(NULL);
524 for (i = 0; i < migrate_multifd_channels(); i++) {
525 MultiFDSendParams *p = &multifd_send_state->params[i];
527 if (p->running) {
528 qemu_thread_join(&p->thread);
531 for (i = 0; i < migrate_multifd_channels(); i++) {
532 MultiFDSendParams *p = &multifd_send_state->params[i];
533 Error *local_err = NULL;
535 if (p->registered_yank) {
536 migration_ioc_unregister_yank(p->c);
538 socket_send_channel_destroy(p->c);
539 p->c = NULL;
540 qemu_mutex_destroy(&p->mutex);
541 qemu_sem_destroy(&p->sem);
542 qemu_sem_destroy(&p->sem_sync);
543 g_free(p->name);
544 p->name = NULL;
545 g_free(p->tls_hostname);
546 p->tls_hostname = NULL;
547 multifd_pages_clear(p->pages);
548 p->pages = NULL;
549 p->packet_len = 0;
550 g_free(p->packet);
551 p->packet = NULL;
552 g_free(p->iov);
553 p->iov = NULL;
554 g_free(p->normal);
555 p->normal = NULL;
556 multifd_send_state->ops->send_cleanup(p, &local_err);
557 if (local_err) {
558 migrate_set_error(migrate_get_current(), local_err);
559 error_free(local_err);
562 qemu_sem_destroy(&multifd_send_state->channels_ready);
563 g_free(multifd_send_state->params);
564 multifd_send_state->params = NULL;
565 multifd_pages_clear(multifd_send_state->pages);
566 multifd_send_state->pages = NULL;
567 g_free(multifd_send_state);
568 multifd_send_state = NULL;
571 void multifd_send_sync_main(QEMUFile *f)
573 int i;
575 if (!migrate_use_multifd()) {
576 return;
578 if (multifd_send_state->pages->num) {
579 if (multifd_send_pages(f) < 0) {
580 error_report("%s: multifd_send_pages fail", __func__);
581 return;
584 for (i = 0; i < migrate_multifd_channels(); i++) {
585 MultiFDSendParams *p = &multifd_send_state->params[i];
587 trace_multifd_send_sync_main_signal(p->id);
589 qemu_mutex_lock(&p->mutex);
591 if (p->quit) {
592 error_report("%s: channel %d has already quit", __func__, i);
593 qemu_mutex_unlock(&p->mutex);
594 return;
597 p->packet_num = multifd_send_state->packet_num++;
598 p->flags |= MULTIFD_FLAG_SYNC;
599 p->pending_job++;
600 qemu_file_update_transfer(f, p->packet_len);
601 ram_counters.multifd_bytes += p->packet_len;
602 ram_counters.transferred += p->packet_len;
603 qemu_mutex_unlock(&p->mutex);
604 qemu_sem_post(&p->sem);
606 for (i = 0; i < migrate_multifd_channels(); i++) {
607 MultiFDSendParams *p = &multifd_send_state->params[i];
609 trace_multifd_send_sync_main_wait(p->id);
610 qemu_sem_wait(&p->sem_sync);
612 trace_multifd_send_sync_main(multifd_send_state->packet_num);
615 static void *multifd_send_thread(void *opaque)
617 MultiFDSendParams *p = opaque;
618 Error *local_err = NULL;
619 int ret = 0;
621 trace_multifd_send_thread_start(p->id);
622 rcu_register_thread();
624 if (multifd_send_initial_packet(p, &local_err) < 0) {
625 ret = -1;
626 goto out;
628 /* initial packet */
629 p->num_packets = 1;
631 while (true) {
632 qemu_sem_wait(&p->sem);
634 if (qatomic_read(&multifd_send_state->exiting)) {
635 break;
637 qemu_mutex_lock(&p->mutex);
639 if (p->pending_job) {
640 uint64_t packet_num = p->packet_num;
641 uint32_t flags = p->flags;
642 p->iovs_num = 1;
643 p->normal_num = 0;
645 for (int i = 0; i < p->pages->num; i++) {
646 p->normal[p->normal_num] = p->pages->offset[i];
647 p->normal_num++;
650 if (p->normal_num) {
651 ret = multifd_send_state->ops->send_prepare(p, &local_err);
652 if (ret != 0) {
653 qemu_mutex_unlock(&p->mutex);
654 break;
657 multifd_send_fill_packet(p);
658 p->flags = 0;
659 p->num_packets++;
660 p->total_normal_pages += p->normal_num;
661 p->pages->num = 0;
662 p->pages->block = NULL;
663 qemu_mutex_unlock(&p->mutex);
665 trace_multifd_send(p->id, packet_num, p->normal_num, flags,
666 p->next_packet_size);
668 p->iov[0].iov_len = p->packet_len;
669 p->iov[0].iov_base = p->packet;
671 ret = qio_channel_writev_all(p->c, p->iov, p->iovs_num,
672 &local_err);
673 if (ret != 0) {
674 break;
677 qemu_mutex_lock(&p->mutex);
678 p->pending_job--;
679 qemu_mutex_unlock(&p->mutex);
681 if (flags & MULTIFD_FLAG_SYNC) {
682 qemu_sem_post(&p->sem_sync);
684 qemu_sem_post(&multifd_send_state->channels_ready);
685 } else if (p->quit) {
686 qemu_mutex_unlock(&p->mutex);
687 break;
688 } else {
689 qemu_mutex_unlock(&p->mutex);
690 /* sometimes there are spurious wakeups */
694 out:
695 if (local_err) {
696 trace_multifd_send_error(p->id);
697 multifd_send_terminate_threads(local_err);
698 error_free(local_err);
702 * Error happen, I will exit, but I can't just leave, tell
703 * who pay attention to me.
705 if (ret != 0) {
706 qemu_sem_post(&p->sem_sync);
707 qemu_sem_post(&multifd_send_state->channels_ready);
710 qemu_mutex_lock(&p->mutex);
711 p->running = false;
712 qemu_mutex_unlock(&p->mutex);
714 rcu_unregister_thread();
715 trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages);
717 return NULL;
720 static bool multifd_channel_connect(MultiFDSendParams *p,
721 QIOChannel *ioc,
722 Error *error);
724 static void multifd_tls_outgoing_handshake(QIOTask *task,
725 gpointer opaque)
727 MultiFDSendParams *p = opaque;
728 QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
729 Error *err = NULL;
731 if (qio_task_propagate_error(task, &err)) {
732 trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err));
733 } else {
734 trace_multifd_tls_outgoing_handshake_complete(ioc);
737 if (!multifd_channel_connect(p, ioc, err)) {
739 * Error happen, mark multifd_send_thread status as 'quit' although it
740 * is not created, and then tell who pay attention to me.
742 p->quit = true;
743 qemu_sem_post(&multifd_send_state->channels_ready);
744 qemu_sem_post(&p->sem_sync);
748 static void *multifd_tls_handshake_thread(void *opaque)
750 MultiFDSendParams *p = opaque;
751 QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
753 qio_channel_tls_handshake(tioc,
754 multifd_tls_outgoing_handshake,
756 NULL,
757 NULL);
758 return NULL;
761 static void multifd_tls_channel_connect(MultiFDSendParams *p,
762 QIOChannel *ioc,
763 Error **errp)
765 MigrationState *s = migrate_get_current();
766 const char *hostname = p->tls_hostname;
767 QIOChannelTLS *tioc;
769 tioc = migration_tls_client_create(s, ioc, hostname, errp);
770 if (!tioc) {
771 return;
774 object_unref(OBJECT(ioc));
775 trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
776 qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
777 p->c = QIO_CHANNEL(tioc);
778 qemu_thread_create(&p->thread, "multifd-tls-handshake-worker",
779 multifd_tls_handshake_thread, p,
780 QEMU_THREAD_JOINABLE);
783 static bool multifd_channel_connect(MultiFDSendParams *p,
784 QIOChannel *ioc,
785 Error *error)
787 MigrationState *s = migrate_get_current();
789 trace_multifd_set_outgoing_channel(
790 ioc, object_get_typename(OBJECT(ioc)), p->tls_hostname, error);
792 if (!error) {
793 if (s->parameters.tls_creds &&
794 *s->parameters.tls_creds &&
795 !object_dynamic_cast(OBJECT(ioc),
796 TYPE_QIO_CHANNEL_TLS)) {
797 multifd_tls_channel_connect(p, ioc, &error);
798 if (!error) {
800 * tls_channel_connect will call back to this
801 * function after the TLS handshake,
802 * so we mustn't call multifd_send_thread until then
804 return true;
805 } else {
806 return false;
808 } else {
809 migration_ioc_register_yank(ioc);
810 p->registered_yank = true;
811 p->c = ioc;
812 qemu_thread_create(&p->thread, p->name, multifd_send_thread, p,
813 QEMU_THREAD_JOINABLE);
815 return true;
818 return false;
821 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p,
822 QIOChannel *ioc, Error *err)
824 migrate_set_error(migrate_get_current(), err);
825 /* Error happen, we need to tell who pay attention to me */
826 qemu_sem_post(&multifd_send_state->channels_ready);
827 qemu_sem_post(&p->sem_sync);
829 * Although multifd_send_thread is not created, but main migration
830 * thread neet to judge whether it is running, so we need to mark
831 * its status.
833 p->quit = true;
834 object_unref(OBJECT(ioc));
835 error_free(err);
838 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque)
840 MultiFDSendParams *p = opaque;
841 QIOChannel *sioc = QIO_CHANNEL(qio_task_get_source(task));
842 Error *local_err = NULL;
844 trace_multifd_new_send_channel_async(p->id);
845 if (qio_task_propagate_error(task, &local_err)) {
846 goto cleanup;
847 } else {
848 p->c = QIO_CHANNEL(sioc);
849 qio_channel_set_delay(p->c, false);
850 p->running = true;
851 if (!multifd_channel_connect(p, sioc, local_err)) {
852 goto cleanup;
854 return;
857 cleanup:
858 multifd_new_send_channel_cleanup(p, sioc, local_err);
861 static bool migrate_allow_multifd = true;
862 void migrate_protocol_allow_multifd(bool allow)
864 migrate_allow_multifd = allow;
867 bool migrate_multifd_is_allowed(void)
869 return migrate_allow_multifd;
872 int multifd_save_setup(Error **errp)
874 int thread_count;
875 uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
876 uint8_t i;
877 MigrationState *s;
879 if (!migrate_use_multifd()) {
880 return 0;
882 if (!migrate_multifd_is_allowed()) {
883 error_setg(errp, "multifd is not supported by current protocol");
884 return -1;
887 s = migrate_get_current();
888 thread_count = migrate_multifd_channels();
889 multifd_send_state = g_malloc0(sizeof(*multifd_send_state));
890 multifd_send_state->params = g_new0(MultiFDSendParams, thread_count);
891 multifd_send_state->pages = multifd_pages_init(page_count);
892 qemu_sem_init(&multifd_send_state->channels_ready, 0);
893 qatomic_set(&multifd_send_state->exiting, 0);
894 multifd_send_state->ops = multifd_ops[migrate_multifd_compression()];
896 for (i = 0; i < thread_count; i++) {
897 MultiFDSendParams *p = &multifd_send_state->params[i];
899 qemu_mutex_init(&p->mutex);
900 qemu_sem_init(&p->sem, 0);
901 qemu_sem_init(&p->sem_sync, 0);
902 p->quit = false;
903 p->pending_job = 0;
904 p->id = i;
905 p->pages = multifd_pages_init(page_count);
906 p->packet_len = sizeof(MultiFDPacket_t)
907 + sizeof(uint64_t) * page_count;
908 p->packet = g_malloc0(p->packet_len);
909 p->packet->magic = cpu_to_be32(MULTIFD_MAGIC);
910 p->packet->version = cpu_to_be32(MULTIFD_VERSION);
911 p->name = g_strdup_printf("multifdsend_%d", i);
912 p->tls_hostname = g_strdup(s->hostname);
913 /* We need one extra place for the packet header */
914 p->iov = g_new0(struct iovec, page_count + 1);
915 p->normal = g_new0(ram_addr_t, page_count);
916 socket_send_channel_create(multifd_new_send_channel_async, p);
919 for (i = 0; i < thread_count; i++) {
920 MultiFDSendParams *p = &multifd_send_state->params[i];
921 Error *local_err = NULL;
922 int ret;
924 ret = multifd_send_state->ops->send_setup(p, &local_err);
925 if (ret) {
926 error_propagate(errp, local_err);
927 return ret;
930 return 0;
933 struct {
934 MultiFDRecvParams *params;
935 /* number of created threads */
936 int count;
937 /* syncs main thread and channels */
938 QemuSemaphore sem_sync;
939 /* global number of generated multifd packets */
940 uint64_t packet_num;
941 /* multifd ops */
942 MultiFDMethods *ops;
943 } *multifd_recv_state;
945 static void multifd_recv_terminate_threads(Error *err)
947 int i;
949 trace_multifd_recv_terminate_threads(err != NULL);
951 if (err) {
952 MigrationState *s = migrate_get_current();
953 migrate_set_error(s, err);
954 if (s->state == MIGRATION_STATUS_SETUP ||
955 s->state == MIGRATION_STATUS_ACTIVE) {
956 migrate_set_state(&s->state, s->state,
957 MIGRATION_STATUS_FAILED);
961 for (i = 0; i < migrate_multifd_channels(); i++) {
962 MultiFDRecvParams *p = &multifd_recv_state->params[i];
964 qemu_mutex_lock(&p->mutex);
965 p->quit = true;
967 * We could arrive here for two reasons:
968 * - normal quit, i.e. everything went fine, just finished
969 * - error quit: We close the channels so the channel threads
970 * finish the qio_channel_read_all_eof()
972 if (p->c) {
973 qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
975 qemu_mutex_unlock(&p->mutex);
979 int multifd_load_cleanup(Error **errp)
981 int i;
983 if (!migrate_use_multifd() || !migrate_multifd_is_allowed()) {
984 return 0;
986 multifd_recv_terminate_threads(NULL);
987 for (i = 0; i < migrate_multifd_channels(); i++) {
988 MultiFDRecvParams *p = &multifd_recv_state->params[i];
990 if (p->running) {
991 p->quit = true;
993 * multifd_recv_thread may hung at MULTIFD_FLAG_SYNC handle code,
994 * however try to wakeup it without harm in cleanup phase.
996 qemu_sem_post(&p->sem_sync);
997 qemu_thread_join(&p->thread);
1000 for (i = 0; i < migrate_multifd_channels(); i++) {
1001 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1003 migration_ioc_unregister_yank(p->c);
1004 object_unref(OBJECT(p->c));
1005 p->c = NULL;
1006 qemu_mutex_destroy(&p->mutex);
1007 qemu_sem_destroy(&p->sem_sync);
1008 g_free(p->name);
1009 p->name = NULL;
1010 p->packet_len = 0;
1011 g_free(p->packet);
1012 p->packet = NULL;
1013 g_free(p->iov);
1014 p->iov = NULL;
1015 g_free(p->normal);
1016 p->normal = NULL;
1017 multifd_recv_state->ops->recv_cleanup(p);
1019 qemu_sem_destroy(&multifd_recv_state->sem_sync);
1020 g_free(multifd_recv_state->params);
1021 multifd_recv_state->params = NULL;
1022 g_free(multifd_recv_state);
1023 multifd_recv_state = NULL;
1025 return 0;
1028 void multifd_recv_sync_main(void)
1030 int i;
1032 if (!migrate_use_multifd()) {
1033 return;
1035 for (i = 0; i < migrate_multifd_channels(); i++) {
1036 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1038 trace_multifd_recv_sync_main_wait(p->id);
1039 qemu_sem_wait(&multifd_recv_state->sem_sync);
1041 for (i = 0; i < migrate_multifd_channels(); i++) {
1042 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1044 WITH_QEMU_LOCK_GUARD(&p->mutex) {
1045 if (multifd_recv_state->packet_num < p->packet_num) {
1046 multifd_recv_state->packet_num = p->packet_num;
1049 trace_multifd_recv_sync_main_signal(p->id);
1050 qemu_sem_post(&p->sem_sync);
1052 trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
1055 static void *multifd_recv_thread(void *opaque)
1057 MultiFDRecvParams *p = opaque;
1058 Error *local_err = NULL;
1059 int ret;
1061 trace_multifd_recv_thread_start(p->id);
1062 rcu_register_thread();
1064 while (true) {
1065 uint32_t flags;
1067 if (p->quit) {
1068 break;
1071 ret = qio_channel_read_all_eof(p->c, (void *)p->packet,
1072 p->packet_len, &local_err);
1073 if (ret == 0) { /* EOF */
1074 break;
1076 if (ret == -1) { /* Error */
1077 break;
1080 qemu_mutex_lock(&p->mutex);
1081 ret = multifd_recv_unfill_packet(p, &local_err);
1082 if (ret) {
1083 qemu_mutex_unlock(&p->mutex);
1084 break;
1087 flags = p->flags;
1088 /* recv methods don't know how to handle the SYNC flag */
1089 p->flags &= ~MULTIFD_FLAG_SYNC;
1090 trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags,
1091 p->next_packet_size);
1092 p->num_packets++;
1093 p->total_normal_pages += p->normal_num;
1094 qemu_mutex_unlock(&p->mutex);
1096 if (p->normal_num) {
1097 ret = multifd_recv_state->ops->recv_pages(p, &local_err);
1098 if (ret != 0) {
1099 break;
1103 if (flags & MULTIFD_FLAG_SYNC) {
1104 qemu_sem_post(&multifd_recv_state->sem_sync);
1105 qemu_sem_wait(&p->sem_sync);
1109 if (local_err) {
1110 multifd_recv_terminate_threads(local_err);
1111 error_free(local_err);
1113 qemu_mutex_lock(&p->mutex);
1114 p->running = false;
1115 qemu_mutex_unlock(&p->mutex);
1117 rcu_unregister_thread();
1118 trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages);
1120 return NULL;
1123 int multifd_load_setup(Error **errp)
1125 int thread_count;
1126 uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
1127 uint8_t i;
1129 if (!migrate_use_multifd()) {
1130 return 0;
1132 if (!migrate_multifd_is_allowed()) {
1133 error_setg(errp, "multifd is not supported by current protocol");
1134 return -1;
1136 thread_count = migrate_multifd_channels();
1137 multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state));
1138 multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count);
1139 qatomic_set(&multifd_recv_state->count, 0);
1140 qemu_sem_init(&multifd_recv_state->sem_sync, 0);
1141 multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()];
1143 for (i = 0; i < thread_count; i++) {
1144 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1146 qemu_mutex_init(&p->mutex);
1147 qemu_sem_init(&p->sem_sync, 0);
1148 p->quit = false;
1149 p->id = i;
1150 p->packet_len = sizeof(MultiFDPacket_t)
1151 + sizeof(uint64_t) * page_count;
1152 p->packet = g_malloc0(p->packet_len);
1153 p->name = g_strdup_printf("multifdrecv_%d", i);
1154 p->iov = g_new0(struct iovec, page_count);
1155 p->normal = g_new0(ram_addr_t, page_count);
1158 for (i = 0; i < thread_count; i++) {
1159 MultiFDRecvParams *p = &multifd_recv_state->params[i];
1160 Error *local_err = NULL;
1161 int ret;
1163 ret = multifd_recv_state->ops->recv_setup(p, &local_err);
1164 if (ret) {
1165 error_propagate(errp, local_err);
1166 return ret;
1169 return 0;
1172 bool multifd_recv_all_channels_created(void)
1174 int thread_count = migrate_multifd_channels();
1176 if (!migrate_use_multifd()) {
1177 return true;
1180 if (!multifd_recv_state) {
1181 /* Called before any connections created */
1182 return false;
1185 return thread_count == qatomic_read(&multifd_recv_state->count);
1189 * Try to receive all multifd channels to get ready for the migration.
1190 * - Return true and do not set @errp when correctly receiving all channels;
1191 * - Return false and do not set @errp when correctly receiving the current one;
1192 * - Return false and set @errp when failing to receive the current channel.
1194 bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp)
1196 MultiFDRecvParams *p;
1197 Error *local_err = NULL;
1198 int id;
1200 id = multifd_recv_initial_packet(ioc, &local_err);
1201 if (id < 0) {
1202 multifd_recv_terminate_threads(local_err);
1203 error_propagate_prepend(errp, local_err,
1204 "failed to receive packet"
1205 " via multifd channel %d: ",
1206 qatomic_read(&multifd_recv_state->count));
1207 return false;
1209 trace_multifd_recv_new_channel(id);
1211 p = &multifd_recv_state->params[id];
1212 if (p->c != NULL) {
1213 error_setg(&local_err, "multifd: received id '%d' already setup'",
1214 id);
1215 multifd_recv_terminate_threads(local_err);
1216 error_propagate(errp, local_err);
1217 return false;
1219 p->c = ioc;
1220 object_ref(OBJECT(ioc));
1221 /* initial packet */
1222 p->num_packets = 1;
1224 p->running = true;
1225 qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p,
1226 QEMU_THREAD_JOINABLE);
1227 qatomic_inc(&multifd_recv_state->count);
1228 return qatomic_read(&multifd_recv_state->count) ==
1229 migrate_multifd_channels();