xref: /openbmc/qemu/migration/multifd.c (revision 5db05230)
1 /*
2  * Multifd common code
3  *
4  * Copyright (c) 2019-2020 Red Hat Inc
5  *
6  * Authors:
7  *  Juan Quintela <quintela@redhat.com>
8  *
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.
11  */
12 
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 "migration-stats.h"
23 #include "socket.h"
24 #include "tls.h"
25 #include "qemu-file.h"
26 #include "trace.h"
27 #include "multifd.h"
28 #include "threadinfo.h"
29 #include "options.h"
30 #include "qemu/yank.h"
31 #include "io/channel-socket.h"
32 #include "yank_functions.h"
33 
34 /* Multiple fd's */
35 
36 #define MULTIFD_MAGIC 0x11223344U
37 #define MULTIFD_VERSION 1
38 
39 typedef struct {
40     uint32_t magic;
41     uint32_t version;
42     unsigned char uuid[16]; /* QemuUUID */
43     uint8_t id;
44     uint8_t unused1[7];     /* Reserved for future use */
45     uint64_t unused2[4];    /* Reserved for future use */
46 } __attribute__((packed)) MultiFDInit_t;
47 
48 /* Multifd without compression */
49 
50 /**
51  * nocomp_send_setup: setup send side
52  *
53  * For no compression this function does nothing.
54  *
55  * Returns 0 for success or -1 for error
56  *
57  * @p: Params for the channel that we are using
58  * @errp: pointer to an error
59  */
60 static int nocomp_send_setup(MultiFDSendParams *p, Error **errp)
61 {
62     return 0;
63 }
64 
65 /**
66  * nocomp_send_cleanup: cleanup send side
67  *
68  * For no compression this function does nothing.
69  *
70  * @p: Params for the channel that we are using
71  * @errp: pointer to an error
72  */
73 static void nocomp_send_cleanup(MultiFDSendParams *p, Error **errp)
74 {
75     return;
76 }
77 
78 /**
79  * nocomp_send_prepare: prepare date to be able to send
80  *
81  * For no compression we just have to calculate the size of the
82  * packet.
83  *
84  * Returns 0 for success or -1 for error
85  *
86  * @p: Params for the channel that we are using
87  * @errp: pointer to an error
88  */
89 static int nocomp_send_prepare(MultiFDSendParams *p, Error **errp)
90 {
91     MultiFDPages_t *pages = p->pages;
92 
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;
96         p->iovs_num++;
97     }
98 
99     p->next_packet_size = p->normal_num * p->page_size;
100     p->flags |= MULTIFD_FLAG_NOCOMP;
101     return 0;
102 }
103 
104 /**
105  * nocomp_recv_setup: setup receive side
106  *
107  * For no compression this function does nothing.
108  *
109  * Returns 0 for success or -1 for error
110  *
111  * @p: Params for the channel that we are using
112  * @errp: pointer to an error
113  */
114 static int nocomp_recv_setup(MultiFDRecvParams *p, Error **errp)
115 {
116     return 0;
117 }
118 
119 /**
120  * nocomp_recv_cleanup: setup receive side
121  *
122  * For no compression this function does nothing.
123  *
124  * @p: Params for the channel that we are using
125  */
126 static void nocomp_recv_cleanup(MultiFDRecvParams *p)
127 {
128 }
129 
130 /**
131  * nocomp_recv_pages: read the data from the channel into actual pages
132  *
133  * For no compression we just need to read things into the correct place.
134  *
135  * Returns 0 for success or -1 for error
136  *
137  * @p: Params for the channel that we are using
138  * @errp: pointer to an error
139  */
140 static int nocomp_recv_pages(MultiFDRecvParams *p, Error **errp)
141 {
142     uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK;
143 
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;
148     }
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;
152     }
153     return qio_channel_readv_all(p->c, p->iov, p->normal_num, errp);
154 }
155 
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
163 };
164 
165 static MultiFDMethods *multifd_ops[MULTIFD_COMPRESSION__MAX] = {
166     [MULTIFD_COMPRESSION_NONE] = &multifd_nocomp_ops,
167 };
168 
169 void multifd_register_ops(int method, MultiFDMethods *ops)
170 {
171     assert(0 < method && method < MULTIFD_COMPRESSION__MAX);
172     multifd_ops[method] = ops;
173 }
174 
175 static int multifd_send_initial_packet(MultiFDSendParams *p, Error **errp)
176 {
177     MultiFDInit_t msg = {};
178     size_t size = sizeof(msg);
179     int ret;
180 
181     msg.magic = cpu_to_be32(MULTIFD_MAGIC);
182     msg.version = cpu_to_be32(MULTIFD_VERSION);
183     msg.id = p->id;
184     memcpy(msg.uuid, &qemu_uuid.data, sizeof(msg.uuid));
185 
186     ret = qio_channel_write_all(p->c, (char *)&msg, size, errp);
187     if (ret != 0) {
188         return -1;
189     }
190     stat64_add(&mig_stats.multifd_bytes, size);
191     return 0;
192 }
193 
194 static int multifd_recv_initial_packet(QIOChannel *c, Error **errp)
195 {
196     MultiFDInit_t msg;
197     int ret;
198 
199     ret = qio_channel_read_all(c, (char *)&msg, sizeof(msg), errp);
200     if (ret != 0) {
201         return -1;
202     }
203 
204     msg.magic = be32_to_cpu(msg.magic);
205     msg.version = be32_to_cpu(msg.version);
206 
207     if (msg.magic != MULTIFD_MAGIC) {
208         error_setg(errp, "multifd: received packet magic %x "
209                    "expected %x", msg.magic, MULTIFD_MAGIC);
210         return -1;
211     }
212 
213     if (msg.version != MULTIFD_VERSION) {
214         error_setg(errp, "multifd: received packet version %u "
215                    "expected %u", msg.version, MULTIFD_VERSION);
216         return -1;
217     }
218 
219     if (memcmp(msg.uuid, &qemu_uuid, sizeof(qemu_uuid))) {
220         char *uuid = qemu_uuid_unparse_strdup(&qemu_uuid);
221         char *msg_uuid = qemu_uuid_unparse_strdup((const QemuUUID *)msg.uuid);
222 
223         error_setg(errp, "multifd: received uuid '%s' and expected "
224                    "uuid '%s' for channel %hhd", msg_uuid, uuid, msg.id);
225         g_free(uuid);
226         g_free(msg_uuid);
227         return -1;
228     }
229 
230     if (msg.id > migrate_multifd_channels()) {
231         error_setg(errp, "multifd: received channel version %u "
232                    "expected %u", msg.version, MULTIFD_VERSION);
233         return -1;
234     }
235 
236     return msg.id;
237 }
238 
239 static MultiFDPages_t *multifd_pages_init(size_t size)
240 {
241     MultiFDPages_t *pages = g_new0(MultiFDPages_t, 1);
242 
243     pages->allocated = size;
244     pages->offset = g_new0(ram_addr_t, size);
245 
246     return pages;
247 }
248 
249 static void multifd_pages_clear(MultiFDPages_t *pages)
250 {
251     pages->num = 0;
252     pages->allocated = 0;
253     pages->packet_num = 0;
254     pages->block = NULL;
255     g_free(pages->offset);
256     pages->offset = NULL;
257     g_free(pages);
258 }
259 
260 static void multifd_send_fill_packet(MultiFDSendParams *p)
261 {
262     MultiFDPacket_t *packet = p->packet;
263     int i;
264 
265     packet->flags = cpu_to_be32(p->flags);
266     packet->pages_alloc = cpu_to_be32(p->pages->allocated);
267     packet->normal_pages = cpu_to_be32(p->normal_num);
268     packet->next_packet_size = cpu_to_be32(p->next_packet_size);
269     packet->packet_num = cpu_to_be64(p->packet_num);
270 
271     if (p->pages->block) {
272         strncpy(packet->ramblock, p->pages->block->idstr, 256);
273     }
274 
275     for (i = 0; i < p->normal_num; i++) {
276         /* there are architectures where ram_addr_t is 32 bit */
277         uint64_t temp = p->normal[i];
278 
279         packet->offset[i] = cpu_to_be64(temp);
280     }
281 }
282 
283 static int multifd_recv_unfill_packet(MultiFDRecvParams *p, Error **errp)
284 {
285     MultiFDPacket_t *packet = p->packet;
286     int i;
287 
288     packet->magic = be32_to_cpu(packet->magic);
289     if (packet->magic != MULTIFD_MAGIC) {
290         error_setg(errp, "multifd: received packet "
291                    "magic %x and expected magic %x",
292                    packet->magic, MULTIFD_MAGIC);
293         return -1;
294     }
295 
296     packet->version = be32_to_cpu(packet->version);
297     if (packet->version != MULTIFD_VERSION) {
298         error_setg(errp, "multifd: received packet "
299                    "version %u and expected version %u",
300                    packet->version, MULTIFD_VERSION);
301         return -1;
302     }
303 
304     p->flags = be32_to_cpu(packet->flags);
305 
306     packet->pages_alloc = be32_to_cpu(packet->pages_alloc);
307     /*
308      * If we received a packet that is 100 times bigger than expected
309      * just stop migration.  It is a magic number.
310      */
311     if (packet->pages_alloc > p->page_count) {
312         error_setg(errp, "multifd: received packet "
313                    "with size %u and expected a size of %u",
314                    packet->pages_alloc, p->page_count) ;
315         return -1;
316     }
317 
318     p->normal_num = be32_to_cpu(packet->normal_pages);
319     if (p->normal_num > packet->pages_alloc) {
320         error_setg(errp, "multifd: received packet "
321                    "with %u pages and expected maximum pages are %u",
322                    p->normal_num, packet->pages_alloc) ;
323         return -1;
324     }
325 
326     p->next_packet_size = be32_to_cpu(packet->next_packet_size);
327     p->packet_num = be64_to_cpu(packet->packet_num);
328 
329     if (p->normal_num == 0) {
330         return 0;
331     }
332 
333     /* make sure that ramblock is 0 terminated */
334     packet->ramblock[255] = 0;
335     p->block = qemu_ram_block_by_name(packet->ramblock);
336     if (!p->block) {
337         error_setg(errp, "multifd: unknown ram block %s",
338                    packet->ramblock);
339         return -1;
340     }
341 
342     p->host = p->block->host;
343     for (i = 0; i < p->normal_num; i++) {
344         uint64_t offset = be64_to_cpu(packet->offset[i]);
345 
346         if (offset > (p->block->used_length - p->page_size)) {
347             error_setg(errp, "multifd: offset too long %" PRIu64
348                        " (max " RAM_ADDR_FMT ")",
349                        offset, p->block->used_length);
350             return -1;
351         }
352         p->normal[i] = offset;
353     }
354 
355     return 0;
356 }
357 
358 struct {
359     MultiFDSendParams *params;
360     /* array of pages to sent */
361     MultiFDPages_t *pages;
362     /* global number of generated multifd packets */
363     uint64_t packet_num;
364     /* send channels ready */
365     QemuSemaphore channels_ready;
366     /*
367      * Have we already run terminate threads.  There is a race when it
368      * happens that we got one error while we are exiting.
369      * We will use atomic operations.  Only valid values are 0 and 1.
370      */
371     int exiting;
372     /* multifd ops */
373     MultiFDMethods *ops;
374 } *multifd_send_state;
375 
376 /*
377  * How we use multifd_send_state->pages and channel->pages?
378  *
379  * We create a pages for each channel, and a main one.  Each time that
380  * we need to send a batch of pages we interchange the ones between
381  * multifd_send_state and the channel that is sending it.  There are
382  * two reasons for that:
383  *    - to not have to do so many mallocs during migration
384  *    - to make easier to know what to free at the end of migration
385  *
386  * This way we always know who is the owner of each "pages" struct,
387  * and we don't need any locking.  It belongs to the migration thread
388  * or to the channel thread.  Switching is safe because the migration
389  * thread is using the channel mutex when changing it, and the channel
390  * have to had finish with its own, otherwise pending_job can't be
391  * false.
392  */
393 
394 static int multifd_send_pages(QEMUFile *f)
395 {
396     int i;
397     static int next_channel;
398     MultiFDSendParams *p = NULL; /* make happy gcc */
399     MultiFDPages_t *pages = multifd_send_state->pages;
400 
401     if (qatomic_read(&multifd_send_state->exiting)) {
402         return -1;
403     }
404 
405     qemu_sem_wait(&multifd_send_state->channels_ready);
406     /*
407      * next_channel can remain from a previous migration that was
408      * using more channels, so ensure it doesn't overflow if the
409      * limit is lower now.
410      */
411     next_channel %= migrate_multifd_channels();
412     for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) {
413         p = &multifd_send_state->params[i];
414 
415         qemu_mutex_lock(&p->mutex);
416         if (p->quit) {
417             error_report("%s: channel %d has already quit!", __func__, i);
418             qemu_mutex_unlock(&p->mutex);
419             return -1;
420         }
421         if (!p->pending_job) {
422             p->pending_job++;
423             next_channel = (i + 1) % migrate_multifd_channels();
424             break;
425         }
426         qemu_mutex_unlock(&p->mutex);
427     }
428     assert(!p->pages->num);
429     assert(!p->pages->block);
430 
431     p->packet_num = multifd_send_state->packet_num++;
432     multifd_send_state->pages = p->pages;
433     p->pages = pages;
434     qemu_mutex_unlock(&p->mutex);
435     qemu_sem_post(&p->sem);
436 
437     return 1;
438 }
439 
440 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
441 {
442     MultiFDPages_t *pages = multifd_send_state->pages;
443     bool changed = false;
444 
445     if (!pages->block) {
446         pages->block = block;
447     }
448 
449     if (pages->block == block) {
450         pages->offset[pages->num] = offset;
451         pages->num++;
452 
453         if (pages->num < pages->allocated) {
454             return 1;
455         }
456     } else {
457         changed = true;
458     }
459 
460     if (multifd_send_pages(f) < 0) {
461         return -1;
462     }
463 
464     if (changed) {
465         return multifd_queue_page(f, block, offset);
466     }
467 
468     return 1;
469 }
470 
471 static void multifd_send_terminate_threads(Error *err)
472 {
473     int i;
474 
475     trace_multifd_send_terminate_threads(err != NULL);
476 
477     if (err) {
478         MigrationState *s = migrate_get_current();
479         migrate_set_error(s, err);
480         if (s->state == MIGRATION_STATUS_SETUP ||
481             s->state == MIGRATION_STATUS_PRE_SWITCHOVER ||
482             s->state == MIGRATION_STATUS_DEVICE ||
483             s->state == MIGRATION_STATUS_ACTIVE) {
484             migrate_set_state(&s->state, s->state,
485                               MIGRATION_STATUS_FAILED);
486         }
487     }
488 
489     /*
490      * We don't want to exit each threads twice.  Depending on where
491      * we get the error, or if there are two independent errors in two
492      * threads at the same time, we can end calling this function
493      * twice.
494      */
495     if (qatomic_xchg(&multifd_send_state->exiting, 1)) {
496         return;
497     }
498 
499     for (i = 0; i < migrate_multifd_channels(); i++) {
500         MultiFDSendParams *p = &multifd_send_state->params[i];
501 
502         qemu_mutex_lock(&p->mutex);
503         p->quit = true;
504         qemu_sem_post(&p->sem);
505         if (p->c) {
506             qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
507         }
508         qemu_mutex_unlock(&p->mutex);
509     }
510 }
511 
512 static int multifd_send_channel_destroy(QIOChannel *send)
513 {
514     return socket_send_channel_destroy(send);
515 }
516 
517 void multifd_save_cleanup(void)
518 {
519     int i;
520 
521     if (!migrate_multifd()) {
522         return;
523     }
524     multifd_send_terminate_threads(NULL);
525     for (i = 0; i < migrate_multifd_channels(); i++) {
526         MultiFDSendParams *p = &multifd_send_state->params[i];
527 
528         if (p->running) {
529             qemu_thread_join(&p->thread);
530         }
531     }
532     for (i = 0; i < migrate_multifd_channels(); i++) {
533         MultiFDSendParams *p = &multifd_send_state->params[i];
534         Error *local_err = NULL;
535 
536         if (p->registered_yank) {
537             migration_ioc_unregister_yank(p->c);
538         }
539         multifd_send_channel_destroy(p->c);
540         p->c = NULL;
541         qemu_mutex_destroy(&p->mutex);
542         qemu_sem_destroy(&p->sem);
543         qemu_sem_destroy(&p->sem_sync);
544         g_free(p->name);
545         p->name = NULL;
546         multifd_pages_clear(p->pages);
547         p->pages = NULL;
548         p->packet_len = 0;
549         g_free(p->packet);
550         p->packet = NULL;
551         g_free(p->iov);
552         p->iov = NULL;
553         g_free(p->normal);
554         p->normal = NULL;
555         multifd_send_state->ops->send_cleanup(p, &local_err);
556         if (local_err) {
557             migrate_set_error(migrate_get_current(), local_err);
558             error_free(local_err);
559         }
560     }
561     qemu_sem_destroy(&multifd_send_state->channels_ready);
562     g_free(multifd_send_state->params);
563     multifd_send_state->params = NULL;
564     multifd_pages_clear(multifd_send_state->pages);
565     multifd_send_state->pages = NULL;
566     g_free(multifd_send_state);
567     multifd_send_state = NULL;
568 }
569 
570 static int multifd_zero_copy_flush(QIOChannel *c)
571 {
572     int ret;
573     Error *err = NULL;
574 
575     ret = qio_channel_flush(c, &err);
576     if (ret < 0) {
577         error_report_err(err);
578         return -1;
579     }
580     if (ret == 1) {
581         stat64_add(&mig_stats.dirty_sync_missed_zero_copy, 1);
582     }
583 
584     return ret;
585 }
586 
587 int multifd_send_sync_main(QEMUFile *f)
588 {
589     int i;
590     bool flush_zero_copy;
591 
592     if (!migrate_multifd()) {
593         return 0;
594     }
595     if (multifd_send_state->pages->num) {
596         if (multifd_send_pages(f) < 0) {
597             error_report("%s: multifd_send_pages fail", __func__);
598             return -1;
599         }
600     }
601 
602     /*
603      * When using zero-copy, it's necessary to flush the pages before any of
604      * the pages can be sent again, so we'll make sure the new version of the
605      * pages will always arrive _later_ than the old pages.
606      *
607      * Currently we achieve this by flushing the zero-page requested writes
608      * per ram iteration, but in the future we could potentially optimize it
609      * to be less frequent, e.g. only after we finished one whole scanning of
610      * all the dirty bitmaps.
611      */
612 
613     flush_zero_copy = migrate_zero_copy_send();
614 
615     for (i = 0; i < migrate_multifd_channels(); i++) {
616         MultiFDSendParams *p = &multifd_send_state->params[i];
617 
618         trace_multifd_send_sync_main_signal(p->id);
619 
620         qemu_mutex_lock(&p->mutex);
621 
622         if (p->quit) {
623             error_report("%s: channel %d has already quit", __func__, i);
624             qemu_mutex_unlock(&p->mutex);
625             return -1;
626         }
627 
628         p->packet_num = multifd_send_state->packet_num++;
629         p->flags |= MULTIFD_FLAG_SYNC;
630         p->pending_job++;
631         qemu_mutex_unlock(&p->mutex);
632         qemu_sem_post(&p->sem);
633     }
634     for (i = 0; i < migrate_multifd_channels(); i++) {
635         MultiFDSendParams *p = &multifd_send_state->params[i];
636 
637         qemu_sem_wait(&multifd_send_state->channels_ready);
638         trace_multifd_send_sync_main_wait(p->id);
639         qemu_sem_wait(&p->sem_sync);
640 
641         if (flush_zero_copy && p->c && (multifd_zero_copy_flush(p->c) < 0)) {
642             return -1;
643         }
644     }
645     trace_multifd_send_sync_main(multifd_send_state->packet_num);
646 
647     return 0;
648 }
649 
650 static void *multifd_send_thread(void *opaque)
651 {
652     MultiFDSendParams *p = opaque;
653     MigrationThread *thread = NULL;
654     Error *local_err = NULL;
655     int ret = 0;
656     bool use_zero_copy_send = migrate_zero_copy_send();
657 
658     thread = migration_threads_add(p->name, qemu_get_thread_id());
659 
660     trace_multifd_send_thread_start(p->id);
661     rcu_register_thread();
662 
663     if (multifd_send_initial_packet(p, &local_err) < 0) {
664         ret = -1;
665         goto out;
666     }
667     /* initial packet */
668     p->num_packets = 1;
669 
670     while (true) {
671         qemu_sem_post(&multifd_send_state->channels_ready);
672         qemu_sem_wait(&p->sem);
673 
674         if (qatomic_read(&multifd_send_state->exiting)) {
675             break;
676         }
677         qemu_mutex_lock(&p->mutex);
678 
679         if (p->pending_job) {
680             uint64_t packet_num = p->packet_num;
681             uint32_t flags;
682             p->normal_num = 0;
683 
684             if (use_zero_copy_send) {
685                 p->iovs_num = 0;
686             } else {
687                 p->iovs_num = 1;
688             }
689 
690             for (int i = 0; i < p->pages->num; i++) {
691                 p->normal[p->normal_num] = p->pages->offset[i];
692                 p->normal_num++;
693             }
694 
695             if (p->normal_num) {
696                 ret = multifd_send_state->ops->send_prepare(p, &local_err);
697                 if (ret != 0) {
698                     qemu_mutex_unlock(&p->mutex);
699                     break;
700                 }
701             }
702             multifd_send_fill_packet(p);
703             flags = p->flags;
704             p->flags = 0;
705             p->num_packets++;
706             p->total_normal_pages += p->normal_num;
707             p->pages->num = 0;
708             p->pages->block = NULL;
709             qemu_mutex_unlock(&p->mutex);
710 
711             trace_multifd_send(p->id, packet_num, p->normal_num, flags,
712                                p->next_packet_size);
713 
714             if (use_zero_copy_send) {
715                 /* Send header first, without zerocopy */
716                 ret = qio_channel_write_all(p->c, (void *)p->packet,
717                                             p->packet_len, &local_err);
718                 if (ret != 0) {
719                     break;
720                 }
721             } else {
722                 /* Send header using the same writev call */
723                 p->iov[0].iov_len = p->packet_len;
724                 p->iov[0].iov_base = p->packet;
725             }
726 
727             ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
728                                               0, p->write_flags, &local_err);
729             if (ret != 0) {
730                 break;
731             }
732 
733             stat64_add(&mig_stats.multifd_bytes,
734                        p->next_packet_size + p->packet_len);
735             p->next_packet_size = 0;
736             qemu_mutex_lock(&p->mutex);
737             p->pending_job--;
738             qemu_mutex_unlock(&p->mutex);
739 
740             if (flags & MULTIFD_FLAG_SYNC) {
741                 qemu_sem_post(&p->sem_sync);
742             }
743         } else {
744             qemu_mutex_unlock(&p->mutex);
745             /* sometimes there are spurious wakeups */
746         }
747     }
748 
749 out:
750     if (ret) {
751         assert(local_err);
752         trace_multifd_send_error(p->id);
753         multifd_send_terminate_threads(local_err);
754         qemu_sem_post(&p->sem_sync);
755         qemu_sem_post(&multifd_send_state->channels_ready);
756         error_free(local_err);
757     }
758 
759     qemu_mutex_lock(&p->mutex);
760     p->running = false;
761     qemu_mutex_unlock(&p->mutex);
762 
763     rcu_unregister_thread();
764     migration_threads_remove(thread);
765     trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages);
766 
767     return NULL;
768 }
769 
770 static bool multifd_channel_connect(MultiFDSendParams *p,
771                                     QIOChannel *ioc,
772                                     Error **errp);
773 
774 static void multifd_tls_outgoing_handshake(QIOTask *task,
775                                            gpointer opaque)
776 {
777     MultiFDSendParams *p = opaque;
778     QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
779     Error *err = NULL;
780 
781     if (!qio_task_propagate_error(task, &err)) {
782         trace_multifd_tls_outgoing_handshake_complete(ioc);
783         if (multifd_channel_connect(p, ioc, &err)) {
784             return;
785         }
786     }
787 
788     trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err));
789 
790     /*
791      * Error happen, mark multifd_send_thread status as 'quit' although it
792      * is not created, and then tell who pay attention to me.
793      */
794     p->quit = true;
795     qemu_sem_post(&multifd_send_state->channels_ready);
796     qemu_sem_post(&p->sem_sync);
797 }
798 
799 static void *multifd_tls_handshake_thread(void *opaque)
800 {
801     MultiFDSendParams *p = opaque;
802     QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
803 
804     qio_channel_tls_handshake(tioc,
805                               multifd_tls_outgoing_handshake,
806                               p,
807                               NULL,
808                               NULL);
809     return NULL;
810 }
811 
812 static bool multifd_tls_channel_connect(MultiFDSendParams *p,
813                                         QIOChannel *ioc,
814                                         Error **errp)
815 {
816     MigrationState *s = migrate_get_current();
817     const char *hostname = s->hostname;
818     QIOChannelTLS *tioc;
819 
820     tioc = migration_tls_client_create(ioc, hostname, errp);
821     if (!tioc) {
822         return false;
823     }
824 
825     object_unref(OBJECT(ioc));
826     trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
827     qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
828     p->c = QIO_CHANNEL(tioc);
829     qemu_thread_create(&p->thread, "multifd-tls-handshake-worker",
830                        multifd_tls_handshake_thread, p,
831                        QEMU_THREAD_JOINABLE);
832     return true;
833 }
834 
835 static bool multifd_channel_connect(MultiFDSendParams *p,
836                                     QIOChannel *ioc,
837                                     Error **errp)
838 {
839     trace_multifd_set_outgoing_channel(
840         ioc, object_get_typename(OBJECT(ioc)),
841         migrate_get_current()->hostname);
842 
843     if (migrate_channel_requires_tls_upgrade(ioc)) {
844         /*
845          * tls_channel_connect will call back to this
846          * function after the TLS handshake,
847          * so we mustn't call multifd_send_thread until then
848          */
849         return multifd_tls_channel_connect(p, ioc, errp);
850 
851     } else {
852         migration_ioc_register_yank(ioc);
853         p->registered_yank = true;
854         p->c = ioc;
855         qemu_thread_create(&p->thread, p->name, multifd_send_thread, p,
856                            QEMU_THREAD_JOINABLE);
857     }
858     return true;
859 }
860 
861 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p,
862                                              QIOChannel *ioc, Error *err)
863 {
864      migrate_set_error(migrate_get_current(), err);
865      /* Error happen, we need to tell who pay attention to me */
866      qemu_sem_post(&multifd_send_state->channels_ready);
867      qemu_sem_post(&p->sem_sync);
868      /*
869       * Although multifd_send_thread is not created, but main migration
870       * thread need to judge whether it is running, so we need to mark
871       * its status.
872       */
873      p->quit = true;
874      object_unref(OBJECT(ioc));
875      error_free(err);
876 }
877 
878 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque)
879 {
880     MultiFDSendParams *p = opaque;
881     QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
882     Error *local_err = NULL;
883 
884     trace_multifd_new_send_channel_async(p->id);
885     if (!qio_task_propagate_error(task, &local_err)) {
886         qio_channel_set_delay(ioc, false);
887         p->running = true;
888         if (multifd_channel_connect(p, ioc, &local_err)) {
889             return;
890         }
891     }
892 
893     trace_multifd_new_send_channel_async_error(p->id, local_err);
894     multifd_new_send_channel_cleanup(p, ioc, local_err);
895 }
896 
897 static void multifd_new_send_channel_create(gpointer opaque)
898 {
899     socket_send_channel_create(multifd_new_send_channel_async, opaque);
900 }
901 
902 int multifd_save_setup(Error **errp)
903 {
904     int thread_count;
905     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
906     uint8_t i;
907 
908     if (!migrate_multifd()) {
909         return 0;
910     }
911 
912     thread_count = migrate_multifd_channels();
913     multifd_send_state = g_malloc0(sizeof(*multifd_send_state));
914     multifd_send_state->params = g_new0(MultiFDSendParams, thread_count);
915     multifd_send_state->pages = multifd_pages_init(page_count);
916     qemu_sem_init(&multifd_send_state->channels_ready, 0);
917     qatomic_set(&multifd_send_state->exiting, 0);
918     multifd_send_state->ops = multifd_ops[migrate_multifd_compression()];
919 
920     for (i = 0; i < thread_count; i++) {
921         MultiFDSendParams *p = &multifd_send_state->params[i];
922 
923         qemu_mutex_init(&p->mutex);
924         qemu_sem_init(&p->sem, 0);
925         qemu_sem_init(&p->sem_sync, 0);
926         p->quit = false;
927         p->pending_job = 0;
928         p->id = i;
929         p->pages = multifd_pages_init(page_count);
930         p->packet_len = sizeof(MultiFDPacket_t)
931                       + sizeof(uint64_t) * page_count;
932         p->packet = g_malloc0(p->packet_len);
933         p->packet->magic = cpu_to_be32(MULTIFD_MAGIC);
934         p->packet->version = cpu_to_be32(MULTIFD_VERSION);
935         p->name = g_strdup_printf("multifdsend_%d", i);
936         /* We need one extra place for the packet header */
937         p->iov = g_new0(struct iovec, page_count + 1);
938         p->normal = g_new0(ram_addr_t, page_count);
939         p->page_size = qemu_target_page_size();
940         p->page_count = page_count;
941 
942         if (migrate_zero_copy_send()) {
943             p->write_flags = QIO_CHANNEL_WRITE_FLAG_ZERO_COPY;
944         } else {
945             p->write_flags = 0;
946         }
947 
948         multifd_new_send_channel_create(p);
949     }
950 
951     for (i = 0; i < thread_count; i++) {
952         MultiFDSendParams *p = &multifd_send_state->params[i];
953         Error *local_err = NULL;
954         int ret;
955 
956         ret = multifd_send_state->ops->send_setup(p, &local_err);
957         if (ret) {
958             error_propagate(errp, local_err);
959             return ret;
960         }
961     }
962     return 0;
963 }
964 
965 struct {
966     MultiFDRecvParams *params;
967     /* number of created threads */
968     int count;
969     /* syncs main thread and channels */
970     QemuSemaphore sem_sync;
971     /* global number of generated multifd packets */
972     uint64_t packet_num;
973     /* multifd ops */
974     MultiFDMethods *ops;
975 } *multifd_recv_state;
976 
977 static void multifd_recv_terminate_threads(Error *err)
978 {
979     int i;
980 
981     trace_multifd_recv_terminate_threads(err != NULL);
982 
983     if (err) {
984         MigrationState *s = migrate_get_current();
985         migrate_set_error(s, err);
986         if (s->state == MIGRATION_STATUS_SETUP ||
987             s->state == MIGRATION_STATUS_ACTIVE) {
988             migrate_set_state(&s->state, s->state,
989                               MIGRATION_STATUS_FAILED);
990         }
991     }
992 
993     for (i = 0; i < migrate_multifd_channels(); i++) {
994         MultiFDRecvParams *p = &multifd_recv_state->params[i];
995 
996         qemu_mutex_lock(&p->mutex);
997         p->quit = true;
998         /*
999          * We could arrive here for two reasons:
1000          *  - normal quit, i.e. everything went fine, just finished
1001          *  - error quit: We close the channels so the channel threads
1002          *    finish the qio_channel_read_all_eof()
1003          */
1004         if (p->c) {
1005             qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1006         }
1007         qemu_mutex_unlock(&p->mutex);
1008     }
1009 }
1010 
1011 void multifd_load_shutdown(void)
1012 {
1013     if (migrate_multifd()) {
1014         multifd_recv_terminate_threads(NULL);
1015     }
1016 }
1017 
1018 void multifd_load_cleanup(void)
1019 {
1020     int i;
1021 
1022     if (!migrate_multifd()) {
1023         return;
1024     }
1025     multifd_recv_terminate_threads(NULL);
1026     for (i = 0; i < migrate_multifd_channels(); i++) {
1027         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1028 
1029         if (p->running) {
1030             /*
1031              * multifd_recv_thread may hung at MULTIFD_FLAG_SYNC handle code,
1032              * however try to wakeup it without harm in cleanup phase.
1033              */
1034             qemu_sem_post(&p->sem_sync);
1035         }
1036 
1037         qemu_thread_join(&p->thread);
1038     }
1039     for (i = 0; i < migrate_multifd_channels(); i++) {
1040         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1041 
1042         migration_ioc_unregister_yank(p->c);
1043         object_unref(OBJECT(p->c));
1044         p->c = NULL;
1045         qemu_mutex_destroy(&p->mutex);
1046         qemu_sem_destroy(&p->sem_sync);
1047         g_free(p->name);
1048         p->name = NULL;
1049         p->packet_len = 0;
1050         g_free(p->packet);
1051         p->packet = NULL;
1052         g_free(p->iov);
1053         p->iov = NULL;
1054         g_free(p->normal);
1055         p->normal = NULL;
1056         multifd_recv_state->ops->recv_cleanup(p);
1057     }
1058     qemu_sem_destroy(&multifd_recv_state->sem_sync);
1059     g_free(multifd_recv_state->params);
1060     multifd_recv_state->params = NULL;
1061     g_free(multifd_recv_state);
1062     multifd_recv_state = NULL;
1063 }
1064 
1065 void multifd_recv_sync_main(void)
1066 {
1067     int i;
1068 
1069     if (!migrate_multifd()) {
1070         return;
1071     }
1072     for (i = 0; i < migrate_multifd_channels(); i++) {
1073         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1074 
1075         trace_multifd_recv_sync_main_wait(p->id);
1076         qemu_sem_wait(&multifd_recv_state->sem_sync);
1077     }
1078     for (i = 0; i < migrate_multifd_channels(); i++) {
1079         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1080 
1081         WITH_QEMU_LOCK_GUARD(&p->mutex) {
1082             if (multifd_recv_state->packet_num < p->packet_num) {
1083                 multifd_recv_state->packet_num = p->packet_num;
1084             }
1085         }
1086         trace_multifd_recv_sync_main_signal(p->id);
1087         qemu_sem_post(&p->sem_sync);
1088     }
1089     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
1090 }
1091 
1092 static void *multifd_recv_thread(void *opaque)
1093 {
1094     MultiFDRecvParams *p = opaque;
1095     Error *local_err = NULL;
1096     int ret;
1097 
1098     trace_multifd_recv_thread_start(p->id);
1099     rcu_register_thread();
1100 
1101     while (true) {
1102         uint32_t flags;
1103 
1104         if (p->quit) {
1105             break;
1106         }
1107 
1108         ret = qio_channel_read_all_eof(p->c, (void *)p->packet,
1109                                        p->packet_len, &local_err);
1110         if (ret == 0 || ret == -1) {   /* 0: EOF  -1: Error */
1111             break;
1112         }
1113 
1114         qemu_mutex_lock(&p->mutex);
1115         ret = multifd_recv_unfill_packet(p, &local_err);
1116         if (ret) {
1117             qemu_mutex_unlock(&p->mutex);
1118             break;
1119         }
1120 
1121         flags = p->flags;
1122         /* recv methods don't know how to handle the SYNC flag */
1123         p->flags &= ~MULTIFD_FLAG_SYNC;
1124         trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags,
1125                            p->next_packet_size);
1126         p->num_packets++;
1127         p->total_normal_pages += p->normal_num;
1128         qemu_mutex_unlock(&p->mutex);
1129 
1130         if (p->normal_num) {
1131             ret = multifd_recv_state->ops->recv_pages(p, &local_err);
1132             if (ret != 0) {
1133                 break;
1134             }
1135         }
1136 
1137         if (flags & MULTIFD_FLAG_SYNC) {
1138             qemu_sem_post(&multifd_recv_state->sem_sync);
1139             qemu_sem_wait(&p->sem_sync);
1140         }
1141     }
1142 
1143     if (local_err) {
1144         multifd_recv_terminate_threads(local_err);
1145         error_free(local_err);
1146     }
1147     qemu_mutex_lock(&p->mutex);
1148     p->running = false;
1149     qemu_mutex_unlock(&p->mutex);
1150 
1151     rcu_unregister_thread();
1152     trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages);
1153 
1154     return NULL;
1155 }
1156 
1157 int multifd_load_setup(Error **errp)
1158 {
1159     int thread_count;
1160     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
1161     uint8_t i;
1162 
1163     /*
1164      * Return successfully if multiFD recv state is already initialised
1165      * or multiFD is not enabled.
1166      */
1167     if (multifd_recv_state || !migrate_multifd()) {
1168         return 0;
1169     }
1170 
1171     thread_count = migrate_multifd_channels();
1172     multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state));
1173     multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count);
1174     qatomic_set(&multifd_recv_state->count, 0);
1175     qemu_sem_init(&multifd_recv_state->sem_sync, 0);
1176     multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()];
1177 
1178     for (i = 0; i < thread_count; i++) {
1179         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1180 
1181         qemu_mutex_init(&p->mutex);
1182         qemu_sem_init(&p->sem_sync, 0);
1183         p->quit = false;
1184         p->id = i;
1185         p->packet_len = sizeof(MultiFDPacket_t)
1186                       + sizeof(uint64_t) * page_count;
1187         p->packet = g_malloc0(p->packet_len);
1188         p->name = g_strdup_printf("multifdrecv_%d", i);
1189         p->iov = g_new0(struct iovec, page_count);
1190         p->normal = g_new0(ram_addr_t, page_count);
1191         p->page_count = page_count;
1192         p->page_size = qemu_target_page_size();
1193     }
1194 
1195     for (i = 0; i < thread_count; i++) {
1196         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1197         Error *local_err = NULL;
1198         int ret;
1199 
1200         ret = multifd_recv_state->ops->recv_setup(p, &local_err);
1201         if (ret) {
1202             error_propagate(errp, local_err);
1203             return ret;
1204         }
1205     }
1206     return 0;
1207 }
1208 
1209 bool multifd_recv_all_channels_created(void)
1210 {
1211     int thread_count = migrate_multifd_channels();
1212 
1213     if (!migrate_multifd()) {
1214         return true;
1215     }
1216 
1217     if (!multifd_recv_state) {
1218         /* Called before any connections created */
1219         return false;
1220     }
1221 
1222     return thread_count == qatomic_read(&multifd_recv_state->count);
1223 }
1224 
1225 /*
1226  * Try to receive all multifd channels to get ready for the migration.
1227  * Sets @errp when failing to receive the current channel.
1228  */
1229 void multifd_recv_new_channel(QIOChannel *ioc, Error **errp)
1230 {
1231     MultiFDRecvParams *p;
1232     Error *local_err = NULL;
1233     int id;
1234 
1235     id = multifd_recv_initial_packet(ioc, &local_err);
1236     if (id < 0) {
1237         multifd_recv_terminate_threads(local_err);
1238         error_propagate_prepend(errp, local_err,
1239                                 "failed to receive packet"
1240                                 " via multifd channel %d: ",
1241                                 qatomic_read(&multifd_recv_state->count));
1242         return;
1243     }
1244     trace_multifd_recv_new_channel(id);
1245 
1246     p = &multifd_recv_state->params[id];
1247     if (p->c != NULL) {
1248         error_setg(&local_err, "multifd: received id '%d' already setup'",
1249                    id);
1250         multifd_recv_terminate_threads(local_err);
1251         error_propagate(errp, local_err);
1252         return;
1253     }
1254     p->c = ioc;
1255     object_ref(OBJECT(ioc));
1256     /* initial packet */
1257     p->num_packets = 1;
1258 
1259     p->running = true;
1260     qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p,
1261                        QEMU_THREAD_JOINABLE);
1262     qatomic_inc(&multifd_recv_state->count);
1263 }
1264