xref: /openbmc/qemu/migration/multifd.c (revision 954a6c4f)
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 "socket.h"
23 #include "tls.h"
24 #include "qemu-file.h"
25 #include "trace.h"
26 #include "multifd.h"
27 
28 #include "qemu/yank.h"
29 #include "io/channel-socket.h"
30 #include "yank_functions.h"
31 
32 /* Multiple fd's */
33 
34 #define MULTIFD_MAGIC 0x11223344U
35 #define MULTIFD_VERSION 1
36 
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;
45 
46 /* Multifd without compression */
47 
48 /**
49  * nocomp_send_setup: setup send side
50  *
51  * For no compression this function does nothing.
52  *
53  * Returns 0 for success or -1 for error
54  *
55  * @p: Params for the channel that we are using
56  * @errp: pointer to an error
57  */
58 static int nocomp_send_setup(MultiFDSendParams *p, Error **errp)
59 {
60     return 0;
61 }
62 
63 /**
64  * nocomp_send_cleanup: cleanup send side
65  *
66  * For no compression this function does nothing.
67  *
68  * @p: Params for the channel that we are using
69  * @errp: pointer to an error
70  */
71 static void nocomp_send_cleanup(MultiFDSendParams *p, Error **errp)
72 {
73     return;
74 }
75 
76 /**
77  * nocomp_send_prepare: prepare date to be able to send
78  *
79  * For no compression we just have to calculate the size of the
80  * packet.
81  *
82  * Returns 0 for success or -1 for error
83  *
84  * @p: Params for the channel that we are using
85  * @errp: pointer to an error
86  */
87 static int nocomp_send_prepare(MultiFDSendParams *p, Error **errp)
88 {
89     MultiFDPages_t *pages = p->pages;
90     size_t page_size = qemu_target_page_size();
91 
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++;
96     }
97 
98     p->next_packet_size = p->normal_num * page_size;
99     p->flags |= MULTIFD_FLAG_NOCOMP;
100     return 0;
101 }
102 
103 /**
104  * nocomp_recv_setup: setup receive side
105  *
106  * For no compression this function does nothing.
107  *
108  * Returns 0 for success or -1 for error
109  *
110  * @p: Params for the channel that we are using
111  * @errp: pointer to an error
112  */
113 static int nocomp_recv_setup(MultiFDRecvParams *p, Error **errp)
114 {
115     return 0;
116 }
117 
118 /**
119  * nocomp_recv_cleanup: setup receive side
120  *
121  * For no compression this function does nothing.
122  *
123  * @p: Params for the channel that we are using
124  */
125 static void nocomp_recv_cleanup(MultiFDRecvParams *p)
126 {
127 }
128 
129 /**
130  * nocomp_recv_pages: read the data from the channel into actual pages
131  *
132  * For no compression we just need to read things into the correct place.
133  *
134  * Returns 0 for success or -1 for error
135  *
136  * @p: Params for the channel that we are using
137  * @errp: pointer to an error
138  */
139 static int nocomp_recv_pages(MultiFDRecvParams *p, Error **errp)
140 {
141     uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK;
142     size_t page_size = qemu_target_page_size();
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 = 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     int ret;
179 
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));
184 
185     ret = qio_channel_write_all(p->c, (char *)&msg, sizeof(msg), errp);
186     if (ret != 0) {
187         return -1;
188     }
189     return 0;
190 }
191 
192 static int multifd_recv_initial_packet(QIOChannel *c, Error **errp)
193 {
194     MultiFDInit_t msg;
195     int ret;
196 
197     ret = qio_channel_read_all(c, (char *)&msg, sizeof(msg), errp);
198     if (ret != 0) {
199         return -1;
200     }
201 
202     msg.magic = be32_to_cpu(msg.magic);
203     msg.version = be32_to_cpu(msg.version);
204 
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;
209     }
210 
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;
215     }
216 
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);
220 
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;
226     }
227 
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;
232     }
233 
234     return msg.id;
235 }
236 
237 static MultiFDPages_t *multifd_pages_init(size_t size)
238 {
239     MultiFDPages_t *pages = g_new0(MultiFDPages_t, 1);
240 
241     pages->allocated = size;
242     pages->offset = g_new0(ram_addr_t, size);
243 
244     return pages;
245 }
246 
247 static void multifd_pages_clear(MultiFDPages_t *pages)
248 {
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);
256 }
257 
258 static void multifd_send_fill_packet(MultiFDSendParams *p)
259 {
260     MultiFDPacket_t *packet = p->packet;
261     int i;
262 
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);
268 
269     if (p->pages->block) {
270         strncpy(packet->ramblock, p->pages->block->idstr, 256);
271     }
272 
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];
276 
277         packet->offset[i] = cpu_to_be64(temp);
278     }
279 }
280 
281 static int multifd_recv_unfill_packet(MultiFDRecvParams *p, Error **errp)
282 {
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;
288 
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;
295     }
296 
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;
303     }
304 
305     p->flags = be32_to_cpu(packet->flags);
306 
307     packet->pages_alloc = be32_to_cpu(packet->pages_alloc);
308     /*
309      * If we received a packet that is 100 times bigger than expected
310      * just stop migration.  It is a magic number.
311      */
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;
317     }
318 
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;
325     }
326 
327     p->next_packet_size = be32_to_cpu(packet->next_packet_size);
328     p->packet_num = be64_to_cpu(packet->packet_num);
329 
330     if (p->normal_num == 0) {
331         return 0;
332     }
333 
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;
341     }
342 
343     p->host = block->host;
344     for (i = 0; i < p->normal_num; i++) {
345         uint64_t offset = be64_to_cpu(packet->offset[i]);
346 
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;
352         }
353         p->normal[i] = offset;
354     }
355 
356     return 0;
357 }
358 
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;
367     /*
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.
371      */
372     int exiting;
373     /* multifd ops */
374     MultiFDMethods *ops;
375 } *multifd_send_state;
376 
377 /*
378  * How we use multifd_send_state->pages and channel->pages?
379  *
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
386  *
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.
393  */
394 
395 static int multifd_send_pages(QEMUFile *f)
396 {
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;
402 
403     if (qatomic_read(&multifd_send_state->exiting)) {
404         return -1;
405     }
406 
407     qemu_sem_wait(&multifd_send_state->channels_ready);
408     /*
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.
412      */
413     next_channel %= migrate_multifd_channels();
414     for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) {
415         p = &multifd_send_state->params[i];
416 
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;
422         }
423         if (!p->pending_job) {
424             p->pending_job++;
425             next_channel = (i + 1) % migrate_multifd_channels();
426             break;
427         }
428         qemu_mutex_unlock(&p->mutex);
429     }
430     assert(!p->pages->num);
431     assert(!p->pages->block);
432 
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_acct_rate_limit(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);
443 
444     return 1;
445 }
446 
447 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
448 {
449     MultiFDPages_t *pages = multifd_send_state->pages;
450 
451     if (!pages->block) {
452         pages->block = block;
453     }
454 
455     if (pages->block == block) {
456         pages->offset[pages->num] = offset;
457         pages->num++;
458 
459         if (pages->num < pages->allocated) {
460             return 1;
461         }
462     }
463 
464     if (multifd_send_pages(f) < 0) {
465         return -1;
466     }
467 
468     if (pages->block != block) {
469         return  multifd_queue_page(f, block, offset);
470     }
471 
472     return 1;
473 }
474 
475 static void multifd_send_terminate_threads(Error *err)
476 {
477     int i;
478 
479     trace_multifd_send_terminate_threads(err != NULL);
480 
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);
490         }
491     }
492 
493     /*
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.
498      */
499     if (qatomic_xchg(&multifd_send_state->exiting, 1)) {
500         return;
501     }
502 
503     for (i = 0; i < migrate_multifd_channels(); i++) {
504         MultiFDSendParams *p = &multifd_send_state->params[i];
505 
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);
511         }
512         qemu_mutex_unlock(&p->mutex);
513     }
514 }
515 
516 void multifd_save_cleanup(void)
517 {
518     int i;
519 
520     if (!migrate_use_multifd() || !migrate_multi_channels_is_allowed()) {
521         return;
522     }
523     multifd_send_terminate_threads(NULL);
524     for (i = 0; i < migrate_multifd_channels(); i++) {
525         MultiFDSendParams *p = &multifd_send_state->params[i];
526 
527         if (p->running) {
528             qemu_thread_join(&p->thread);
529         }
530     }
531     for (i = 0; i < migrate_multifd_channels(); i++) {
532         MultiFDSendParams *p = &multifd_send_state->params[i];
533         Error *local_err = NULL;
534 
535         if (p->registered_yank) {
536             migration_ioc_unregister_yank(p->c);
537         }
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         multifd_pages_clear(p->pages);
546         p->pages = NULL;
547         p->packet_len = 0;
548         g_free(p->packet);
549         p->packet = NULL;
550         g_free(p->iov);
551         p->iov = NULL;
552         g_free(p->normal);
553         p->normal = NULL;
554         multifd_send_state->ops->send_cleanup(p, &local_err);
555         if (local_err) {
556             migrate_set_error(migrate_get_current(), local_err);
557             error_free(local_err);
558         }
559     }
560     qemu_sem_destroy(&multifd_send_state->channels_ready);
561     g_free(multifd_send_state->params);
562     multifd_send_state->params = NULL;
563     multifd_pages_clear(multifd_send_state->pages);
564     multifd_send_state->pages = NULL;
565     g_free(multifd_send_state);
566     multifd_send_state = NULL;
567 }
568 
569 static int multifd_zero_copy_flush(QIOChannel *c)
570 {
571     int ret;
572     Error *err = NULL;
573 
574     ret = qio_channel_flush(c, &err);
575     if (ret < 0) {
576         error_report_err(err);
577         return -1;
578     }
579     if (ret == 1) {
580         dirty_sync_missed_zero_copy();
581     }
582 
583     return ret;
584 }
585 
586 int multifd_send_sync_main(QEMUFile *f)
587 {
588     int i;
589     bool flush_zero_copy;
590 
591     if (!migrate_use_multifd()) {
592         return 0;
593     }
594     if (multifd_send_state->pages->num) {
595         if (multifd_send_pages(f) < 0) {
596             error_report("%s: multifd_send_pages fail", __func__);
597             return -1;
598         }
599     }
600 
601     /*
602      * When using zero-copy, it's necessary to flush the pages before any of
603      * the pages can be sent again, so we'll make sure the new version of the
604      * pages will always arrive _later_ than the old pages.
605      *
606      * Currently we achieve this by flushing the zero-page requested writes
607      * per ram iteration, but in the future we could potentially optimize it
608      * to be less frequent, e.g. only after we finished one whole scanning of
609      * all the dirty bitmaps.
610      */
611 
612     flush_zero_copy = migrate_use_zero_copy_send();
613 
614     for (i = 0; i < migrate_multifd_channels(); i++) {
615         MultiFDSendParams *p = &multifd_send_state->params[i];
616 
617         trace_multifd_send_sync_main_signal(p->id);
618 
619         qemu_mutex_lock(&p->mutex);
620 
621         if (p->quit) {
622             error_report("%s: channel %d has already quit", __func__, i);
623             qemu_mutex_unlock(&p->mutex);
624             return -1;
625         }
626 
627         p->packet_num = multifd_send_state->packet_num++;
628         p->flags |= MULTIFD_FLAG_SYNC;
629         p->pending_job++;
630         qemu_file_acct_rate_limit(f, p->packet_len);
631         ram_counters.multifd_bytes += p->packet_len;
632         ram_counters.transferred += p->packet_len;
633         qemu_mutex_unlock(&p->mutex);
634         qemu_sem_post(&p->sem);
635 
636         if (flush_zero_copy && p->c && (multifd_zero_copy_flush(p->c) < 0)) {
637             return -1;
638         }
639     }
640     for (i = 0; i < migrate_multifd_channels(); i++) {
641         MultiFDSendParams *p = &multifd_send_state->params[i];
642 
643         trace_multifd_send_sync_main_wait(p->id);
644         qemu_sem_wait(&p->sem_sync);
645     }
646     trace_multifd_send_sync_main(multifd_send_state->packet_num);
647 
648     return 0;
649 }
650 
651 static void *multifd_send_thread(void *opaque)
652 {
653     MultiFDSendParams *p = opaque;
654     Error *local_err = NULL;
655     int ret = 0;
656     bool use_zero_copy_send = migrate_use_zero_copy_send();
657 
658     trace_multifd_send_thread_start(p->id);
659     rcu_register_thread();
660 
661     if (multifd_send_initial_packet(p, &local_err) < 0) {
662         ret = -1;
663         goto out;
664     }
665     /* initial packet */
666     p->num_packets = 1;
667 
668     while (true) {
669         qemu_sem_wait(&p->sem);
670 
671         if (qatomic_read(&multifd_send_state->exiting)) {
672             break;
673         }
674         qemu_mutex_lock(&p->mutex);
675 
676         if (p->pending_job) {
677             uint64_t packet_num = p->packet_num;
678             uint32_t flags = p->flags;
679             p->normal_num = 0;
680 
681             if (use_zero_copy_send) {
682                 p->iovs_num = 0;
683             } else {
684                 p->iovs_num = 1;
685             }
686 
687             for (int i = 0; i < p->pages->num; i++) {
688                 p->normal[p->normal_num] = p->pages->offset[i];
689                 p->normal_num++;
690             }
691 
692             if (p->normal_num) {
693                 ret = multifd_send_state->ops->send_prepare(p, &local_err);
694                 if (ret != 0) {
695                     qemu_mutex_unlock(&p->mutex);
696                     break;
697                 }
698             }
699             multifd_send_fill_packet(p);
700             p->flags = 0;
701             p->num_packets++;
702             p->total_normal_pages += p->normal_num;
703             p->pages->num = 0;
704             p->pages->block = NULL;
705             qemu_mutex_unlock(&p->mutex);
706 
707             trace_multifd_send(p->id, packet_num, p->normal_num, flags,
708                                p->next_packet_size);
709 
710             if (use_zero_copy_send) {
711                 /* Send header first, without zerocopy */
712                 ret = qio_channel_write_all(p->c, (void *)p->packet,
713                                             p->packet_len, &local_err);
714                 if (ret != 0) {
715                     break;
716                 }
717             } else {
718                 /* Send header using the same writev call */
719                 p->iov[0].iov_len = p->packet_len;
720                 p->iov[0].iov_base = p->packet;
721             }
722 
723             ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
724                                               0, p->write_flags, &local_err);
725             if (ret != 0) {
726                 break;
727             }
728 
729             qemu_mutex_lock(&p->mutex);
730             p->pending_job--;
731             qemu_mutex_unlock(&p->mutex);
732 
733             if (flags & MULTIFD_FLAG_SYNC) {
734                 qemu_sem_post(&p->sem_sync);
735             }
736             qemu_sem_post(&multifd_send_state->channels_ready);
737         } else if (p->quit) {
738             qemu_mutex_unlock(&p->mutex);
739             break;
740         } else {
741             qemu_mutex_unlock(&p->mutex);
742             /* sometimes there are spurious wakeups */
743         }
744     }
745 
746 out:
747     if (local_err) {
748         trace_multifd_send_error(p->id);
749         multifd_send_terminate_threads(local_err);
750         error_free(local_err);
751     }
752 
753     /*
754      * Error happen, I will exit, but I can't just leave, tell
755      * who pay attention to me.
756      */
757     if (ret != 0) {
758         qemu_sem_post(&p->sem_sync);
759         qemu_sem_post(&multifd_send_state->channels_ready);
760     }
761 
762     qemu_mutex_lock(&p->mutex);
763     p->running = false;
764     qemu_mutex_unlock(&p->mutex);
765 
766     rcu_unregister_thread();
767     trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages);
768 
769     return NULL;
770 }
771 
772 static bool multifd_channel_connect(MultiFDSendParams *p,
773                                     QIOChannel *ioc,
774                                     Error *error);
775 
776 static void multifd_tls_outgoing_handshake(QIOTask *task,
777                                            gpointer opaque)
778 {
779     MultiFDSendParams *p = opaque;
780     QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
781     Error *err = NULL;
782 
783     if (qio_task_propagate_error(task, &err)) {
784         trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err));
785     } else {
786         trace_multifd_tls_outgoing_handshake_complete(ioc);
787     }
788 
789     if (!multifd_channel_connect(p, ioc, err)) {
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 
800 static void *multifd_tls_handshake_thread(void *opaque)
801 {
802     MultiFDSendParams *p = opaque;
803     QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
804 
805     qio_channel_tls_handshake(tioc,
806                               multifd_tls_outgoing_handshake,
807                               p,
808                               NULL,
809                               NULL);
810     return NULL;
811 }
812 
813 static void multifd_tls_channel_connect(MultiFDSendParams *p,
814                                         QIOChannel *ioc,
815                                         Error **errp)
816 {
817     MigrationState *s = migrate_get_current();
818     const char *hostname = s->hostname;
819     QIOChannelTLS *tioc;
820 
821     tioc = migration_tls_client_create(s, ioc, hostname, errp);
822     if (!tioc) {
823         return;
824     }
825 
826     object_unref(OBJECT(ioc));
827     trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
828     qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
829     p->c = QIO_CHANNEL(tioc);
830     qemu_thread_create(&p->thread, "multifd-tls-handshake-worker",
831                        multifd_tls_handshake_thread, p,
832                        QEMU_THREAD_JOINABLE);
833 }
834 
835 static bool multifd_channel_connect(MultiFDSendParams *p,
836                                     QIOChannel *ioc,
837                                     Error *error)
838 {
839     trace_multifd_set_outgoing_channel(
840         ioc, object_get_typename(OBJECT(ioc)),
841         migrate_get_current()->hostname, error);
842 
843     if (!error) {
844         if (migrate_channel_requires_tls_upgrade(ioc)) {
845             multifd_tls_channel_connect(p, ioc, &error);
846             if (!error) {
847                 /*
848                  * tls_channel_connect will call back to this
849                  * function after the TLS handshake,
850                  * so we mustn't call multifd_send_thread until then
851                  */
852                 return true;
853             } else {
854                 return false;
855             }
856         } else {
857             migration_ioc_register_yank(ioc);
858             p->registered_yank = true;
859             p->c = ioc;
860             qemu_thread_create(&p->thread, p->name, multifd_send_thread, p,
861                                    QEMU_THREAD_JOINABLE);
862        }
863        return true;
864     }
865 
866     return false;
867 }
868 
869 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p,
870                                              QIOChannel *ioc, Error *err)
871 {
872      migrate_set_error(migrate_get_current(), err);
873      /* Error happen, we need to tell who pay attention to me */
874      qemu_sem_post(&multifd_send_state->channels_ready);
875      qemu_sem_post(&p->sem_sync);
876      /*
877       * Although multifd_send_thread is not created, but main migration
878       * thread neet to judge whether it is running, so we need to mark
879       * its status.
880       */
881      p->quit = true;
882      object_unref(OBJECT(ioc));
883      error_free(err);
884 }
885 
886 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque)
887 {
888     MultiFDSendParams *p = opaque;
889     QIOChannel *sioc = QIO_CHANNEL(qio_task_get_source(task));
890     Error *local_err = NULL;
891 
892     trace_multifd_new_send_channel_async(p->id);
893     if (qio_task_propagate_error(task, &local_err)) {
894         goto cleanup;
895     } else {
896         p->c = QIO_CHANNEL(sioc);
897         qio_channel_set_delay(p->c, false);
898         p->running = true;
899         if (!multifd_channel_connect(p, sioc, local_err)) {
900             goto cleanup;
901         }
902         return;
903     }
904 
905 cleanup:
906     multifd_new_send_channel_cleanup(p, sioc, local_err);
907 }
908 
909 int multifd_save_setup(Error **errp)
910 {
911     int thread_count;
912     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
913     uint8_t i;
914 
915     if (!migrate_use_multifd()) {
916         return 0;
917     }
918     if (!migrate_multi_channels_is_allowed()) {
919         error_setg(errp, "multifd is not supported by current protocol");
920         return -1;
921     }
922 
923     thread_count = migrate_multifd_channels();
924     multifd_send_state = g_malloc0(sizeof(*multifd_send_state));
925     multifd_send_state->params = g_new0(MultiFDSendParams, thread_count);
926     multifd_send_state->pages = multifd_pages_init(page_count);
927     qemu_sem_init(&multifd_send_state->channels_ready, 0);
928     qatomic_set(&multifd_send_state->exiting, 0);
929     multifd_send_state->ops = multifd_ops[migrate_multifd_compression()];
930 
931     for (i = 0; i < thread_count; i++) {
932         MultiFDSendParams *p = &multifd_send_state->params[i];
933 
934         qemu_mutex_init(&p->mutex);
935         qemu_sem_init(&p->sem, 0);
936         qemu_sem_init(&p->sem_sync, 0);
937         p->quit = false;
938         p->pending_job = 0;
939         p->id = i;
940         p->pages = multifd_pages_init(page_count);
941         p->packet_len = sizeof(MultiFDPacket_t)
942                       + sizeof(uint64_t) * page_count;
943         p->packet = g_malloc0(p->packet_len);
944         p->packet->magic = cpu_to_be32(MULTIFD_MAGIC);
945         p->packet->version = cpu_to_be32(MULTIFD_VERSION);
946         p->name = g_strdup_printf("multifdsend_%d", i);
947         /* We need one extra place for the packet header */
948         p->iov = g_new0(struct iovec, page_count + 1);
949         p->normal = g_new0(ram_addr_t, page_count);
950 
951         if (migrate_use_zero_copy_send()) {
952             p->write_flags = QIO_CHANNEL_WRITE_FLAG_ZERO_COPY;
953         } else {
954             p->write_flags = 0;
955         }
956 
957         socket_send_channel_create(multifd_new_send_channel_async, p);
958     }
959 
960     for (i = 0; i < thread_count; i++) {
961         MultiFDSendParams *p = &multifd_send_state->params[i];
962         Error *local_err = NULL;
963         int ret;
964 
965         ret = multifd_send_state->ops->send_setup(p, &local_err);
966         if (ret) {
967             error_propagate(errp, local_err);
968             return ret;
969         }
970     }
971     return 0;
972 }
973 
974 struct {
975     MultiFDRecvParams *params;
976     /* number of created threads */
977     int count;
978     /* syncs main thread and channels */
979     QemuSemaphore sem_sync;
980     /* global number of generated multifd packets */
981     uint64_t packet_num;
982     /* multifd ops */
983     MultiFDMethods *ops;
984 } *multifd_recv_state;
985 
986 static void multifd_recv_terminate_threads(Error *err)
987 {
988     int i;
989 
990     trace_multifd_recv_terminate_threads(err != NULL);
991 
992     if (err) {
993         MigrationState *s = migrate_get_current();
994         migrate_set_error(s, err);
995         if (s->state == MIGRATION_STATUS_SETUP ||
996             s->state == MIGRATION_STATUS_ACTIVE) {
997             migrate_set_state(&s->state, s->state,
998                               MIGRATION_STATUS_FAILED);
999         }
1000     }
1001 
1002     for (i = 0; i < migrate_multifd_channels(); i++) {
1003         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1004 
1005         qemu_mutex_lock(&p->mutex);
1006         p->quit = true;
1007         /*
1008          * We could arrive here for two reasons:
1009          *  - normal quit, i.e. everything went fine, just finished
1010          *  - error quit: We close the channels so the channel threads
1011          *    finish the qio_channel_read_all_eof()
1012          */
1013         if (p->c) {
1014             qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1015         }
1016         qemu_mutex_unlock(&p->mutex);
1017     }
1018 }
1019 
1020 int multifd_load_cleanup(Error **errp)
1021 {
1022     int i;
1023 
1024     if (!migrate_use_multifd() || !migrate_multi_channels_is_allowed()) {
1025         return 0;
1026     }
1027     multifd_recv_terminate_threads(NULL);
1028     for (i = 0; i < migrate_multifd_channels(); i++) {
1029         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1030 
1031         if (p->running) {
1032             p->quit = true;
1033             /*
1034              * multifd_recv_thread may hung at MULTIFD_FLAG_SYNC handle code,
1035              * however try to wakeup it without harm in cleanup phase.
1036              */
1037             qemu_sem_post(&p->sem_sync);
1038             qemu_thread_join(&p->thread);
1039         }
1040     }
1041     for (i = 0; i < migrate_multifd_channels(); i++) {
1042         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1043 
1044         migration_ioc_unregister_yank(p->c);
1045         object_unref(OBJECT(p->c));
1046         p->c = NULL;
1047         qemu_mutex_destroy(&p->mutex);
1048         qemu_sem_destroy(&p->sem_sync);
1049         g_free(p->name);
1050         p->name = NULL;
1051         p->packet_len = 0;
1052         g_free(p->packet);
1053         p->packet = NULL;
1054         g_free(p->iov);
1055         p->iov = NULL;
1056         g_free(p->normal);
1057         p->normal = NULL;
1058         multifd_recv_state->ops->recv_cleanup(p);
1059     }
1060     qemu_sem_destroy(&multifd_recv_state->sem_sync);
1061     g_free(multifd_recv_state->params);
1062     multifd_recv_state->params = NULL;
1063     g_free(multifd_recv_state);
1064     multifd_recv_state = NULL;
1065 
1066     return 0;
1067 }
1068 
1069 void multifd_recv_sync_main(void)
1070 {
1071     int i;
1072 
1073     if (!migrate_use_multifd()) {
1074         return;
1075     }
1076     for (i = 0; i < migrate_multifd_channels(); i++) {
1077         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1078 
1079         trace_multifd_recv_sync_main_wait(p->id);
1080         qemu_sem_wait(&multifd_recv_state->sem_sync);
1081     }
1082     for (i = 0; i < migrate_multifd_channels(); i++) {
1083         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1084 
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;
1088             }
1089         }
1090         trace_multifd_recv_sync_main_signal(p->id);
1091         qemu_sem_post(&p->sem_sync);
1092     }
1093     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
1094 }
1095 
1096 static void *multifd_recv_thread(void *opaque)
1097 {
1098     MultiFDRecvParams *p = opaque;
1099     Error *local_err = NULL;
1100     int ret;
1101 
1102     trace_multifd_recv_thread_start(p->id);
1103     rcu_register_thread();
1104 
1105     while (true) {
1106         uint32_t flags;
1107 
1108         if (p->quit) {
1109             break;
1110         }
1111 
1112         ret = qio_channel_read_all_eof(p->c, (void *)p->packet,
1113                                        p->packet_len, &local_err);
1114         if (ret == 0) {   /* EOF */
1115             break;
1116         }
1117         if (ret == -1) {   /* Error */
1118             break;
1119         }
1120 
1121         qemu_mutex_lock(&p->mutex);
1122         ret = multifd_recv_unfill_packet(p, &local_err);
1123         if (ret) {
1124             qemu_mutex_unlock(&p->mutex);
1125             break;
1126         }
1127 
1128         flags = p->flags;
1129         /* recv methods don't know how to handle the SYNC flag */
1130         p->flags &= ~MULTIFD_FLAG_SYNC;
1131         trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags,
1132                            p->next_packet_size);
1133         p->num_packets++;
1134         p->total_normal_pages += p->normal_num;
1135         qemu_mutex_unlock(&p->mutex);
1136 
1137         if (p->normal_num) {
1138             ret = multifd_recv_state->ops->recv_pages(p, &local_err);
1139             if (ret != 0) {
1140                 break;
1141             }
1142         }
1143 
1144         if (flags & MULTIFD_FLAG_SYNC) {
1145             qemu_sem_post(&multifd_recv_state->sem_sync);
1146             qemu_sem_wait(&p->sem_sync);
1147         }
1148     }
1149 
1150     if (local_err) {
1151         multifd_recv_terminate_threads(local_err);
1152         error_free(local_err);
1153     }
1154     qemu_mutex_lock(&p->mutex);
1155     p->running = false;
1156     qemu_mutex_unlock(&p->mutex);
1157 
1158     rcu_unregister_thread();
1159     trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages);
1160 
1161     return NULL;
1162 }
1163 
1164 int multifd_load_setup(Error **errp)
1165 {
1166     int thread_count;
1167     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
1168     uint8_t i;
1169 
1170     if (!migrate_use_multifd()) {
1171         return 0;
1172     }
1173     if (!migrate_multi_channels_is_allowed()) {
1174         error_setg(errp, "multifd is not supported by current protocol");
1175         return -1;
1176     }
1177     thread_count = migrate_multifd_channels();
1178     multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state));
1179     multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count);
1180     qatomic_set(&multifd_recv_state->count, 0);
1181     qemu_sem_init(&multifd_recv_state->sem_sync, 0);
1182     multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()];
1183 
1184     for (i = 0; i < thread_count; i++) {
1185         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1186 
1187         qemu_mutex_init(&p->mutex);
1188         qemu_sem_init(&p->sem_sync, 0);
1189         p->quit = false;
1190         p->id = i;
1191         p->packet_len = sizeof(MultiFDPacket_t)
1192                       + sizeof(uint64_t) * page_count;
1193         p->packet = g_malloc0(p->packet_len);
1194         p->name = g_strdup_printf("multifdrecv_%d", i);
1195         p->iov = g_new0(struct iovec, page_count);
1196         p->normal = g_new0(ram_addr_t, page_count);
1197     }
1198 
1199     for (i = 0; i < thread_count; i++) {
1200         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1201         Error *local_err = NULL;
1202         int ret;
1203 
1204         ret = multifd_recv_state->ops->recv_setup(p, &local_err);
1205         if (ret) {
1206             error_propagate(errp, local_err);
1207             return ret;
1208         }
1209     }
1210     return 0;
1211 }
1212 
1213 bool multifd_recv_all_channels_created(void)
1214 {
1215     int thread_count = migrate_multifd_channels();
1216 
1217     if (!migrate_use_multifd()) {
1218         return true;
1219     }
1220 
1221     if (!multifd_recv_state) {
1222         /* Called before any connections created */
1223         return false;
1224     }
1225 
1226     return thread_count == qatomic_read(&multifd_recv_state->count);
1227 }
1228 
1229 /*
1230  * Try to receive all multifd channels to get ready for the migration.
1231  * - Return true and do not set @errp when correctly receiving all channels;
1232  * - Return false and do not set @errp when correctly receiving the current one;
1233  * - Return false and set @errp when failing to receive the current channel.
1234  */
1235 bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp)
1236 {
1237     MultiFDRecvParams *p;
1238     Error *local_err = NULL;
1239     int id;
1240 
1241     id = multifd_recv_initial_packet(ioc, &local_err);
1242     if (id < 0) {
1243         multifd_recv_terminate_threads(local_err);
1244         error_propagate_prepend(errp, local_err,
1245                                 "failed to receive packet"
1246                                 " via multifd channel %d: ",
1247                                 qatomic_read(&multifd_recv_state->count));
1248         return false;
1249     }
1250     trace_multifd_recv_new_channel(id);
1251 
1252     p = &multifd_recv_state->params[id];
1253     if (p->c != NULL) {
1254         error_setg(&local_err, "multifd: received id '%d' already setup'",
1255                    id);
1256         multifd_recv_terminate_threads(local_err);
1257         error_propagate(errp, local_err);
1258         return false;
1259     }
1260     p->c = ioc;
1261     object_ref(OBJECT(ioc));
1262     /* initial packet */
1263     p->num_packets = 1;
1264 
1265     p->running = true;
1266     qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p,
1267                        QEMU_THREAD_JOINABLE);
1268     qatomic_inc(&multifd_recv_state->count);
1269     return qatomic_read(&multifd_recv_state->count) ==
1270            migrate_multifd_channels();
1271 }
1272