xref: /qemu/migration/multifd.c (revision e4418354)
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 
91     for (int i = 0; i < p->normal_num; i++) {
92         p->iov[p->iovs_num].iov_base = pages->block->host + p->normal[i];
93         p->iov[p->iovs_num].iov_len = p->page_size;
94         p->iovs_num++;
95     }
96 
97     p->next_packet_size = p->normal_num * p->page_size;
98     p->flags |= MULTIFD_FLAG_NOCOMP;
99     return 0;
100 }
101 
102 /**
103  * nocomp_recv_setup: setup receive side
104  *
105  * For no compression this function does nothing.
106  *
107  * Returns 0 for success or -1 for error
108  *
109  * @p: Params for the channel that we are using
110  * @errp: pointer to an error
111  */
112 static int nocomp_recv_setup(MultiFDRecvParams *p, Error **errp)
113 {
114     return 0;
115 }
116 
117 /**
118  * nocomp_recv_cleanup: setup receive side
119  *
120  * For no compression this function does nothing.
121  *
122  * @p: Params for the channel that we are using
123  */
124 static void nocomp_recv_cleanup(MultiFDRecvParams *p)
125 {
126 }
127 
128 /**
129  * nocomp_recv_pages: read the data from the channel into actual pages
130  *
131  * For no compression we just need to read things into the correct place.
132  *
133  * Returns 0 for success or -1 for error
134  *
135  * @p: Params for the channel that we are using
136  * @errp: pointer to an error
137  */
138 static int nocomp_recv_pages(MultiFDRecvParams *p, Error **errp)
139 {
140     uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK;
141 
142     if (flags != MULTIFD_FLAG_NOCOMP) {
143         error_setg(errp, "multifd %u: flags received %x flags expected %x",
144                    p->id, flags, MULTIFD_FLAG_NOCOMP);
145         return -1;
146     }
147     for (int i = 0; i < p->normal_num; i++) {
148         p->iov[i].iov_base = p->host + p->normal[i];
149         p->iov[i].iov_len = p->page_size;
150     }
151     return qio_channel_readv_all(p->c, p->iov, p->normal_num, errp);
152 }
153 
154 static MultiFDMethods multifd_nocomp_ops = {
155     .send_setup = nocomp_send_setup,
156     .send_cleanup = nocomp_send_cleanup,
157     .send_prepare = nocomp_send_prepare,
158     .recv_setup = nocomp_recv_setup,
159     .recv_cleanup = nocomp_recv_cleanup,
160     .recv_pages = nocomp_recv_pages
161 };
162 
163 static MultiFDMethods *multifd_ops[MULTIFD_COMPRESSION__MAX] = {
164     [MULTIFD_COMPRESSION_NONE] = &multifd_nocomp_ops,
165 };
166 
167 void multifd_register_ops(int method, MultiFDMethods *ops)
168 {
169     assert(0 < method && method < MULTIFD_COMPRESSION__MAX);
170     multifd_ops[method] = ops;
171 }
172 
173 static int multifd_send_initial_packet(MultiFDSendParams *p, Error **errp)
174 {
175     MultiFDInit_t msg = {};
176     int ret;
177 
178     msg.magic = cpu_to_be32(MULTIFD_MAGIC);
179     msg.version = cpu_to_be32(MULTIFD_VERSION);
180     msg.id = p->id;
181     memcpy(msg.uuid, &qemu_uuid.data, sizeof(msg.uuid));
182 
183     ret = qio_channel_write_all(p->c, (char *)&msg, sizeof(msg), errp);
184     if (ret != 0) {
185         return -1;
186     }
187     return 0;
188 }
189 
190 static int multifd_recv_initial_packet(QIOChannel *c, Error **errp)
191 {
192     MultiFDInit_t msg;
193     int ret;
194 
195     ret = qio_channel_read_all(c, (char *)&msg, sizeof(msg), errp);
196     if (ret != 0) {
197         return -1;
198     }
199 
200     msg.magic = be32_to_cpu(msg.magic);
201     msg.version = be32_to_cpu(msg.version);
202 
203     if (msg.magic != MULTIFD_MAGIC) {
204         error_setg(errp, "multifd: received packet magic %x "
205                    "expected %x", msg.magic, MULTIFD_MAGIC);
206         return -1;
207     }
208 
209     if (msg.version != MULTIFD_VERSION) {
210         error_setg(errp, "multifd: received packet version %u "
211                    "expected %u", msg.version, MULTIFD_VERSION);
212         return -1;
213     }
214 
215     if (memcmp(msg.uuid, &qemu_uuid, sizeof(qemu_uuid))) {
216         char *uuid = qemu_uuid_unparse_strdup(&qemu_uuid);
217         char *msg_uuid = qemu_uuid_unparse_strdup((const QemuUUID *)msg.uuid);
218 
219         error_setg(errp, "multifd: received uuid '%s' and expected "
220                    "uuid '%s' for channel %hhd", msg_uuid, uuid, msg.id);
221         g_free(uuid);
222         g_free(msg_uuid);
223         return -1;
224     }
225 
226     if (msg.id > migrate_multifd_channels()) {
227         error_setg(errp, "multifd: received channel version %u "
228                    "expected %u", msg.version, MULTIFD_VERSION);
229         return -1;
230     }
231 
232     return msg.id;
233 }
234 
235 static MultiFDPages_t *multifd_pages_init(size_t size)
236 {
237     MultiFDPages_t *pages = g_new0(MultiFDPages_t, 1);
238 
239     pages->allocated = size;
240     pages->offset = g_new0(ram_addr_t, size);
241 
242     return pages;
243 }
244 
245 static void multifd_pages_clear(MultiFDPages_t *pages)
246 {
247     pages->num = 0;
248     pages->allocated = 0;
249     pages->packet_num = 0;
250     pages->block = NULL;
251     g_free(pages->offset);
252     pages->offset = NULL;
253     g_free(pages);
254 }
255 
256 static void multifd_send_fill_packet(MultiFDSendParams *p)
257 {
258     MultiFDPacket_t *packet = p->packet;
259     int i;
260 
261     packet->flags = cpu_to_be32(p->flags);
262     packet->pages_alloc = cpu_to_be32(p->pages->allocated);
263     packet->normal_pages = cpu_to_be32(p->normal_num);
264     packet->next_packet_size = cpu_to_be32(p->next_packet_size);
265     packet->packet_num = cpu_to_be64(p->packet_num);
266 
267     if (p->pages->block) {
268         strncpy(packet->ramblock, p->pages->block->idstr, 256);
269     }
270 
271     for (i = 0; i < p->normal_num; i++) {
272         /* there are architectures where ram_addr_t is 32 bit */
273         uint64_t temp = p->normal[i];
274 
275         packet->offset[i] = cpu_to_be64(temp);
276     }
277 }
278 
279 static int multifd_recv_unfill_packet(MultiFDRecvParams *p, Error **errp)
280 {
281     MultiFDPacket_t *packet = p->packet;
282     RAMBlock *block;
283     int i;
284 
285     packet->magic = be32_to_cpu(packet->magic);
286     if (packet->magic != MULTIFD_MAGIC) {
287         error_setg(errp, "multifd: received packet "
288                    "magic %x and expected magic %x",
289                    packet->magic, MULTIFD_MAGIC);
290         return -1;
291     }
292 
293     packet->version = be32_to_cpu(packet->version);
294     if (packet->version != MULTIFD_VERSION) {
295         error_setg(errp, "multifd: received packet "
296                    "version %u and expected version %u",
297                    packet->version, MULTIFD_VERSION);
298         return -1;
299     }
300 
301     p->flags = be32_to_cpu(packet->flags);
302 
303     packet->pages_alloc = be32_to_cpu(packet->pages_alloc);
304     /*
305      * If we received a packet that is 100 times bigger than expected
306      * just stop migration.  It is a magic number.
307      */
308     if (packet->pages_alloc > p->page_count) {
309         error_setg(errp, "multifd: received packet "
310                    "with size %u and expected a size of %u",
311                    packet->pages_alloc, p->page_count) ;
312         return -1;
313     }
314 
315     p->normal_num = be32_to_cpu(packet->normal_pages);
316     if (p->normal_num > packet->pages_alloc) {
317         error_setg(errp, "multifd: received packet "
318                    "with %u pages and expected maximum pages are %u",
319                    p->normal_num, packet->pages_alloc) ;
320         return -1;
321     }
322 
323     p->next_packet_size = be32_to_cpu(packet->next_packet_size);
324     p->packet_num = be64_to_cpu(packet->packet_num);
325 
326     if (p->normal_num == 0) {
327         return 0;
328     }
329 
330     /* make sure that ramblock is 0 terminated */
331     packet->ramblock[255] = 0;
332     block = qemu_ram_block_by_name(packet->ramblock);
333     if (!block) {
334         error_setg(errp, "multifd: unknown ram block %s",
335                    packet->ramblock);
336         return -1;
337     }
338 
339     p->host = block->host;
340     for (i = 0; i < p->normal_num; i++) {
341         uint64_t offset = be64_to_cpu(packet->offset[i]);
342 
343         if (offset > (block->used_length - p->page_size)) {
344             error_setg(errp, "multifd: offset too long %" PRIu64
345                        " (max " RAM_ADDR_FMT ")",
346                        offset, block->used_length);
347             return -1;
348         }
349         p->normal[i] = offset;
350     }
351 
352     return 0;
353 }
354 
355 struct {
356     MultiFDSendParams *params;
357     /* array of pages to sent */
358     MultiFDPages_t *pages;
359     /* global number of generated multifd packets */
360     uint64_t packet_num;
361     /* send channels ready */
362     QemuSemaphore channels_ready;
363     /*
364      * Have we already run terminate threads.  There is a race when it
365      * happens that we got one error while we are exiting.
366      * We will use atomic operations.  Only valid values are 0 and 1.
367      */
368     int exiting;
369     /* multifd ops */
370     MultiFDMethods *ops;
371 } *multifd_send_state;
372 
373 /*
374  * How we use multifd_send_state->pages and channel->pages?
375  *
376  * We create a pages for each channel, and a main one.  Each time that
377  * we need to send a batch of pages we interchange the ones between
378  * multifd_send_state and the channel that is sending it.  There are
379  * two reasons for that:
380  *    - to not have to do so many mallocs during migration
381  *    - to make easier to know what to free at the end of migration
382  *
383  * This way we always know who is the owner of each "pages" struct,
384  * and we don't need any locking.  It belongs to the migration thread
385  * or to the channel thread.  Switching is safe because the migration
386  * thread is using the channel mutex when changing it, and the channel
387  * have to had finish with its own, otherwise pending_job can't be
388  * false.
389  */
390 
391 static int multifd_send_pages(QEMUFile *f)
392 {
393     int i;
394     static int next_channel;
395     MultiFDSendParams *p = NULL; /* make happy gcc */
396     MultiFDPages_t *pages = multifd_send_state->pages;
397     uint64_t transferred;
398 
399     if (qatomic_read(&multifd_send_state->exiting)) {
400         return -1;
401     }
402 
403     qemu_sem_wait(&multifd_send_state->channels_ready);
404     /*
405      * next_channel can remain from a previous migration that was
406      * using more channels, so ensure it doesn't overflow if the
407      * limit is lower now.
408      */
409     next_channel %= migrate_multifd_channels();
410     for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) {
411         p = &multifd_send_state->params[i];
412 
413         qemu_mutex_lock(&p->mutex);
414         if (p->quit) {
415             error_report("%s: channel %d has already quit!", __func__, i);
416             qemu_mutex_unlock(&p->mutex);
417             return -1;
418         }
419         if (!p->pending_job) {
420             p->pending_job++;
421             next_channel = (i + 1) % migrate_multifd_channels();
422             break;
423         }
424         qemu_mutex_unlock(&p->mutex);
425     }
426     assert(!p->pages->num);
427     assert(!p->pages->block);
428 
429     p->packet_num = multifd_send_state->packet_num++;
430     multifd_send_state->pages = p->pages;
431     p->pages = pages;
432     transferred = ((uint64_t) pages->num) * p->page_size + p->packet_len;
433     qemu_file_acct_rate_limit(f, transferred);
434     ram_counters.multifd_bytes += transferred;
435     stat64_add(&ram_atomic_counters.transferred, transferred);
436     qemu_mutex_unlock(&p->mutex);
437     qemu_sem_post(&p->sem);
438 
439     return 1;
440 }
441 
442 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
443 {
444     MultiFDPages_t *pages = multifd_send_state->pages;
445 
446     if (!pages->block) {
447         pages->block = block;
448     }
449 
450     if (pages->block == block) {
451         pages->offset[pages->num] = offset;
452         pages->num++;
453 
454         if (pages->num < pages->allocated) {
455             return 1;
456         }
457     }
458 
459     if (multifd_send_pages(f) < 0) {
460         return -1;
461     }
462 
463     if (pages->block != block) {
464         return  multifd_queue_page(f, block, offset);
465     }
466 
467     return 1;
468 }
469 
470 static void multifd_send_terminate_threads(Error *err)
471 {
472     int i;
473 
474     trace_multifd_send_terminate_threads(err != NULL);
475 
476     if (err) {
477         MigrationState *s = migrate_get_current();
478         migrate_set_error(s, err);
479         if (s->state == MIGRATION_STATUS_SETUP ||
480             s->state == MIGRATION_STATUS_PRE_SWITCHOVER ||
481             s->state == MIGRATION_STATUS_DEVICE ||
482             s->state == MIGRATION_STATUS_ACTIVE) {
483             migrate_set_state(&s->state, s->state,
484                               MIGRATION_STATUS_FAILED);
485         }
486     }
487 
488     /*
489      * We don't want to exit each threads twice.  Depending on where
490      * we get the error, or if there are two independent errors in two
491      * threads at the same time, we can end calling this function
492      * twice.
493      */
494     if (qatomic_xchg(&multifd_send_state->exiting, 1)) {
495         return;
496     }
497 
498     for (i = 0; i < migrate_multifd_channels(); i++) {
499         MultiFDSendParams *p = &multifd_send_state->params[i];
500 
501         qemu_mutex_lock(&p->mutex);
502         p->quit = true;
503         qemu_sem_post(&p->sem);
504         if (p->c) {
505             qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
506         }
507         qemu_mutex_unlock(&p->mutex);
508     }
509 }
510 
511 void multifd_save_cleanup(void)
512 {
513     int i;
514 
515     if (!migrate_use_multifd() || !migrate_multi_channels_is_allowed()) {
516         return;
517     }
518     multifd_send_terminate_threads(NULL);
519     for (i = 0; i < migrate_multifd_channels(); i++) {
520         MultiFDSendParams *p = &multifd_send_state->params[i];
521 
522         if (p->running) {
523             qemu_thread_join(&p->thread);
524         }
525     }
526     for (i = 0; i < migrate_multifd_channels(); i++) {
527         MultiFDSendParams *p = &multifd_send_state->params[i];
528         Error *local_err = NULL;
529 
530         if (p->registered_yank) {
531             migration_ioc_unregister_yank(p->c);
532         }
533         socket_send_channel_destroy(p->c);
534         p->c = NULL;
535         qemu_mutex_destroy(&p->mutex);
536         qemu_sem_destroy(&p->sem);
537         qemu_sem_destroy(&p->sem_sync);
538         g_free(p->name);
539         p->name = NULL;
540         multifd_pages_clear(p->pages);
541         p->pages = NULL;
542         p->packet_len = 0;
543         g_free(p->packet);
544         p->packet = NULL;
545         g_free(p->iov);
546         p->iov = NULL;
547         g_free(p->normal);
548         p->normal = NULL;
549         multifd_send_state->ops->send_cleanup(p, &local_err);
550         if (local_err) {
551             migrate_set_error(migrate_get_current(), local_err);
552             error_free(local_err);
553         }
554     }
555     qemu_sem_destroy(&multifd_send_state->channels_ready);
556     g_free(multifd_send_state->params);
557     multifd_send_state->params = NULL;
558     multifd_pages_clear(multifd_send_state->pages);
559     multifd_send_state->pages = NULL;
560     g_free(multifd_send_state);
561     multifd_send_state = NULL;
562 }
563 
564 static int multifd_zero_copy_flush(QIOChannel *c)
565 {
566     int ret;
567     Error *err = NULL;
568 
569     ret = qio_channel_flush(c, &err);
570     if (ret < 0) {
571         error_report_err(err);
572         return -1;
573     }
574     if (ret == 1) {
575         dirty_sync_missed_zero_copy();
576     }
577 
578     return ret;
579 }
580 
581 int multifd_send_sync_main(QEMUFile *f)
582 {
583     int i;
584     bool flush_zero_copy;
585 
586     if (!migrate_use_multifd()) {
587         return 0;
588     }
589     if (multifd_send_state->pages->num) {
590         if (multifd_send_pages(f) < 0) {
591             error_report("%s: multifd_send_pages fail", __func__);
592             return -1;
593         }
594     }
595 
596     /*
597      * When using zero-copy, it's necessary to flush the pages before any of
598      * the pages can be sent again, so we'll make sure the new version of the
599      * pages will always arrive _later_ than the old pages.
600      *
601      * Currently we achieve this by flushing the zero-page requested writes
602      * per ram iteration, but in the future we could potentially optimize it
603      * to be less frequent, e.g. only after we finished one whole scanning of
604      * all the dirty bitmaps.
605      */
606 
607     flush_zero_copy = migrate_use_zero_copy_send();
608 
609     for (i = 0; i < migrate_multifd_channels(); i++) {
610         MultiFDSendParams *p = &multifd_send_state->params[i];
611 
612         trace_multifd_send_sync_main_signal(p->id);
613 
614         qemu_mutex_lock(&p->mutex);
615 
616         if (p->quit) {
617             error_report("%s: channel %d has already quit", __func__, i);
618             qemu_mutex_unlock(&p->mutex);
619             return -1;
620         }
621 
622         p->packet_num = multifd_send_state->packet_num++;
623         p->flags |= MULTIFD_FLAG_SYNC;
624         p->pending_job++;
625         qemu_file_acct_rate_limit(f, p->packet_len);
626         ram_counters.multifd_bytes += p->packet_len;
627         stat64_add(&ram_atomic_counters.transferred, p->packet_len);
628         qemu_mutex_unlock(&p->mutex);
629         qemu_sem_post(&p->sem);
630 
631         if (flush_zero_copy && p->c && (multifd_zero_copy_flush(p->c) < 0)) {
632             return -1;
633         }
634     }
635     for (i = 0; i < migrate_multifd_channels(); i++) {
636         MultiFDSendParams *p = &multifd_send_state->params[i];
637 
638         trace_multifd_send_sync_main_wait(p->id);
639         qemu_sem_wait(&p->sem_sync);
640     }
641     trace_multifd_send_sync_main(multifd_send_state->packet_num);
642 
643     return 0;
644 }
645 
646 static void *multifd_send_thread(void *opaque)
647 {
648     MultiFDSendParams *p = opaque;
649     Error *local_err = NULL;
650     int ret = 0;
651     bool use_zero_copy_send = migrate_use_zero_copy_send();
652 
653     trace_multifd_send_thread_start(p->id);
654     rcu_register_thread();
655 
656     if (multifd_send_initial_packet(p, &local_err) < 0) {
657         ret = -1;
658         goto out;
659     }
660     /* initial packet */
661     p->num_packets = 1;
662 
663     while (true) {
664         qemu_sem_wait(&p->sem);
665 
666         if (qatomic_read(&multifd_send_state->exiting)) {
667             break;
668         }
669         qemu_mutex_lock(&p->mutex);
670 
671         if (p->pending_job) {
672             uint64_t packet_num = p->packet_num;
673             uint32_t flags = p->flags;
674             p->normal_num = 0;
675 
676             if (use_zero_copy_send) {
677                 p->iovs_num = 0;
678             } else {
679                 p->iovs_num = 1;
680             }
681 
682             for (int i = 0; i < p->pages->num; i++) {
683                 p->normal[p->normal_num] = p->pages->offset[i];
684                 p->normal_num++;
685             }
686 
687             if (p->normal_num) {
688                 ret = multifd_send_state->ops->send_prepare(p, &local_err);
689                 if (ret != 0) {
690                     qemu_mutex_unlock(&p->mutex);
691                     break;
692                 }
693             }
694             multifd_send_fill_packet(p);
695             p->flags = 0;
696             p->num_packets++;
697             p->total_normal_pages += p->normal_num;
698             p->pages->num = 0;
699             p->pages->block = NULL;
700             qemu_mutex_unlock(&p->mutex);
701 
702             trace_multifd_send(p->id, packet_num, p->normal_num, flags,
703                                p->next_packet_size);
704 
705             if (use_zero_copy_send) {
706                 /* Send header first, without zerocopy */
707                 ret = qio_channel_write_all(p->c, (void *)p->packet,
708                                             p->packet_len, &local_err);
709                 if (ret != 0) {
710                     break;
711                 }
712             } else {
713                 /* Send header using the same writev call */
714                 p->iov[0].iov_len = p->packet_len;
715                 p->iov[0].iov_base = p->packet;
716             }
717 
718             ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
719                                               0, p->write_flags, &local_err);
720             if (ret != 0) {
721                 break;
722             }
723 
724             qemu_mutex_lock(&p->mutex);
725             p->pending_job--;
726             qemu_mutex_unlock(&p->mutex);
727 
728             if (flags & MULTIFD_FLAG_SYNC) {
729                 qemu_sem_post(&p->sem_sync);
730             }
731             qemu_sem_post(&multifd_send_state->channels_ready);
732         } else if (p->quit) {
733             qemu_mutex_unlock(&p->mutex);
734             break;
735         } else {
736             qemu_mutex_unlock(&p->mutex);
737             /* sometimes there are spurious wakeups */
738         }
739     }
740 
741 out:
742     if (local_err) {
743         trace_multifd_send_error(p->id);
744         multifd_send_terminate_threads(local_err);
745         error_free(local_err);
746     }
747 
748     /*
749      * Error happen, I will exit, but I can't just leave, tell
750      * who pay attention to me.
751      */
752     if (ret != 0) {
753         qemu_sem_post(&p->sem_sync);
754         qemu_sem_post(&multifd_send_state->channels_ready);
755     }
756 
757     qemu_mutex_lock(&p->mutex);
758     p->running = false;
759     qemu_mutex_unlock(&p->mutex);
760 
761     rcu_unregister_thread();
762     trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages);
763 
764     return NULL;
765 }
766 
767 static bool multifd_channel_connect(MultiFDSendParams *p,
768                                     QIOChannel *ioc,
769                                     Error *error);
770 
771 static void multifd_tls_outgoing_handshake(QIOTask *task,
772                                            gpointer opaque)
773 {
774     MultiFDSendParams *p = opaque;
775     QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
776     Error *err = NULL;
777 
778     if (qio_task_propagate_error(task, &err)) {
779         trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err));
780     } else {
781         trace_multifd_tls_outgoing_handshake_complete(ioc);
782     }
783 
784     if (!multifd_channel_connect(p, ioc, err)) {
785         /*
786          * Error happen, mark multifd_send_thread status as 'quit' although it
787          * is not created, and then tell who pay attention to me.
788          */
789         p->quit = true;
790         qemu_sem_post(&multifd_send_state->channels_ready);
791         qemu_sem_post(&p->sem_sync);
792     }
793 }
794 
795 static void *multifd_tls_handshake_thread(void *opaque)
796 {
797     MultiFDSendParams *p = opaque;
798     QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
799 
800     qio_channel_tls_handshake(tioc,
801                               multifd_tls_outgoing_handshake,
802                               p,
803                               NULL,
804                               NULL);
805     return NULL;
806 }
807 
808 static void multifd_tls_channel_connect(MultiFDSendParams *p,
809                                         QIOChannel *ioc,
810                                         Error **errp)
811 {
812     MigrationState *s = migrate_get_current();
813     const char *hostname = s->hostname;
814     QIOChannelTLS *tioc;
815 
816     tioc = migration_tls_client_create(s, ioc, hostname, errp);
817     if (!tioc) {
818         return;
819     }
820 
821     object_unref(OBJECT(ioc));
822     trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
823     qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
824     p->c = QIO_CHANNEL(tioc);
825     qemu_thread_create(&p->thread, "multifd-tls-handshake-worker",
826                        multifd_tls_handshake_thread, p,
827                        QEMU_THREAD_JOINABLE);
828 }
829 
830 static bool multifd_channel_connect(MultiFDSendParams *p,
831                                     QIOChannel *ioc,
832                                     Error *error)
833 {
834     trace_multifd_set_outgoing_channel(
835         ioc, object_get_typename(OBJECT(ioc)),
836         migrate_get_current()->hostname, error);
837 
838     if (!error) {
839         if (migrate_channel_requires_tls_upgrade(ioc)) {
840             multifd_tls_channel_connect(p, ioc, &error);
841             if (!error) {
842                 /*
843                  * tls_channel_connect will call back to this
844                  * function after the TLS handshake,
845                  * so we mustn't call multifd_send_thread until then
846                  */
847                 return true;
848             } else {
849                 return false;
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     return false;
862 }
863 
864 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p,
865                                              QIOChannel *ioc, Error *err)
866 {
867      migrate_set_error(migrate_get_current(), err);
868      /* Error happen, we need to tell who pay attention to me */
869      qemu_sem_post(&multifd_send_state->channels_ready);
870      qemu_sem_post(&p->sem_sync);
871      /*
872       * Although multifd_send_thread is not created, but main migration
873       * thread neet to judge whether it is running, so we need to mark
874       * its status.
875       */
876      p->quit = true;
877      object_unref(OBJECT(ioc));
878      error_free(err);
879 }
880 
881 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque)
882 {
883     MultiFDSendParams *p = opaque;
884     QIOChannel *sioc = QIO_CHANNEL(qio_task_get_source(task));
885     Error *local_err = NULL;
886 
887     trace_multifd_new_send_channel_async(p->id);
888     if (qio_task_propagate_error(task, &local_err)) {
889         goto cleanup;
890     } else {
891         p->c = QIO_CHANNEL(sioc);
892         qio_channel_set_delay(p->c, false);
893         p->running = true;
894         if (!multifd_channel_connect(p, sioc, local_err)) {
895             goto cleanup;
896         }
897         return;
898     }
899 
900 cleanup:
901     multifd_new_send_channel_cleanup(p, sioc, local_err);
902 }
903 
904 int multifd_save_setup(Error **errp)
905 {
906     int thread_count;
907     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
908     uint8_t i;
909 
910     if (!migrate_use_multifd()) {
911         return 0;
912     }
913     if (!migrate_multi_channels_is_allowed()) {
914         error_setg(errp, "multifd is not supported by current protocol");
915         return -1;
916     }
917 
918     thread_count = migrate_multifd_channels();
919     multifd_send_state = g_malloc0(sizeof(*multifd_send_state));
920     multifd_send_state->params = g_new0(MultiFDSendParams, thread_count);
921     multifd_send_state->pages = multifd_pages_init(page_count);
922     qemu_sem_init(&multifd_send_state->channels_ready, 0);
923     qatomic_set(&multifd_send_state->exiting, 0);
924     multifd_send_state->ops = multifd_ops[migrate_multifd_compression()];
925 
926     for (i = 0; i < thread_count; i++) {
927         MultiFDSendParams *p = &multifd_send_state->params[i];
928 
929         qemu_mutex_init(&p->mutex);
930         qemu_sem_init(&p->sem, 0);
931         qemu_sem_init(&p->sem_sync, 0);
932         p->quit = false;
933         p->pending_job = 0;
934         p->id = i;
935         p->pages = multifd_pages_init(page_count);
936         p->packet_len = sizeof(MultiFDPacket_t)
937                       + sizeof(uint64_t) * page_count;
938         p->packet = g_malloc0(p->packet_len);
939         p->packet->magic = cpu_to_be32(MULTIFD_MAGIC);
940         p->packet->version = cpu_to_be32(MULTIFD_VERSION);
941         p->name = g_strdup_printf("multifdsend_%d", i);
942         /* We need one extra place for the packet header */
943         p->iov = g_new0(struct iovec, page_count + 1);
944         p->normal = g_new0(ram_addr_t, page_count);
945         p->page_size = qemu_target_page_size();
946         p->page_count = page_count;
947 
948         if (migrate_use_zero_copy_send()) {
949             p->write_flags = QIO_CHANNEL_WRITE_FLAG_ZERO_COPY;
950         } else {
951             p->write_flags = 0;
952         }
953 
954         socket_send_channel_create(multifd_new_send_channel_async, p);
955     }
956 
957     for (i = 0; i < thread_count; i++) {
958         MultiFDSendParams *p = &multifd_send_state->params[i];
959         Error *local_err = NULL;
960         int ret;
961 
962         ret = multifd_send_state->ops->send_setup(p, &local_err);
963         if (ret) {
964             error_propagate(errp, local_err);
965             return ret;
966         }
967     }
968     return 0;
969 }
970 
971 struct {
972     MultiFDRecvParams *params;
973     /* number of created threads */
974     int count;
975     /* syncs main thread and channels */
976     QemuSemaphore sem_sync;
977     /* global number of generated multifd packets */
978     uint64_t packet_num;
979     /* multifd ops */
980     MultiFDMethods *ops;
981 } *multifd_recv_state;
982 
983 static void multifd_recv_terminate_threads(Error *err)
984 {
985     int i;
986 
987     trace_multifd_recv_terminate_threads(err != NULL);
988 
989     if (err) {
990         MigrationState *s = migrate_get_current();
991         migrate_set_error(s, err);
992         if (s->state == MIGRATION_STATUS_SETUP ||
993             s->state == MIGRATION_STATUS_ACTIVE) {
994             migrate_set_state(&s->state, s->state,
995                               MIGRATION_STATUS_FAILED);
996         }
997     }
998 
999     for (i = 0; i < migrate_multifd_channels(); i++) {
1000         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1001 
1002         qemu_mutex_lock(&p->mutex);
1003         p->quit = true;
1004         /*
1005          * We could arrive here for two reasons:
1006          *  - normal quit, i.e. everything went fine, just finished
1007          *  - error quit: We close the channels so the channel threads
1008          *    finish the qio_channel_read_all_eof()
1009          */
1010         if (p->c) {
1011             qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1012         }
1013         qemu_mutex_unlock(&p->mutex);
1014     }
1015 }
1016 
1017 int multifd_load_cleanup(Error **errp)
1018 {
1019     int i;
1020 
1021     if (!migrate_use_multifd() || !migrate_multi_channels_is_allowed()) {
1022         return 0;
1023     }
1024     multifd_recv_terminate_threads(NULL);
1025     for (i = 0; i < migrate_multifd_channels(); i++) {
1026         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1027 
1028         if (p->running) {
1029             p->quit = true;
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             qemu_thread_join(&p->thread);
1036         }
1037     }
1038     for (i = 0; i < migrate_multifd_channels(); i++) {
1039         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1040 
1041         migration_ioc_unregister_yank(p->c);
1042         object_unref(OBJECT(p->c));
1043         p->c = NULL;
1044         qemu_mutex_destroy(&p->mutex);
1045         qemu_sem_destroy(&p->sem_sync);
1046         g_free(p->name);
1047         p->name = NULL;
1048         p->packet_len = 0;
1049         g_free(p->packet);
1050         p->packet = NULL;
1051         g_free(p->iov);
1052         p->iov = NULL;
1053         g_free(p->normal);
1054         p->normal = NULL;
1055         multifd_recv_state->ops->recv_cleanup(p);
1056     }
1057     qemu_sem_destroy(&multifd_recv_state->sem_sync);
1058     g_free(multifd_recv_state->params);
1059     multifd_recv_state->params = NULL;
1060     g_free(multifd_recv_state);
1061     multifd_recv_state = NULL;
1062 
1063     return 0;
1064 }
1065 
1066 void multifd_recv_sync_main(void)
1067 {
1068     int i;
1069 
1070     if (!migrate_use_multifd()) {
1071         return;
1072     }
1073     for (i = 0; i < migrate_multifd_channels(); i++) {
1074         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1075 
1076         trace_multifd_recv_sync_main_wait(p->id);
1077         qemu_sem_wait(&multifd_recv_state->sem_sync);
1078     }
1079     for (i = 0; i < migrate_multifd_channels(); i++) {
1080         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1081 
1082         WITH_QEMU_LOCK_GUARD(&p->mutex) {
1083             if (multifd_recv_state->packet_num < p->packet_num) {
1084                 multifd_recv_state->packet_num = p->packet_num;
1085             }
1086         }
1087         trace_multifd_recv_sync_main_signal(p->id);
1088         qemu_sem_post(&p->sem_sync);
1089     }
1090     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
1091 }
1092 
1093 static void *multifd_recv_thread(void *opaque)
1094 {
1095     MultiFDRecvParams *p = opaque;
1096     Error *local_err = NULL;
1097     int ret;
1098 
1099     trace_multifd_recv_thread_start(p->id);
1100     rcu_register_thread();
1101 
1102     while (true) {
1103         uint32_t flags;
1104 
1105         if (p->quit) {
1106             break;
1107         }
1108 
1109         ret = qio_channel_read_all_eof(p->c, (void *)p->packet,
1110                                        p->packet_len, &local_err);
1111         if (ret == 0) {   /* EOF */
1112             break;
1113         }
1114         if (ret == -1) {   /* Error */
1115             break;
1116         }
1117 
1118         qemu_mutex_lock(&p->mutex);
1119         ret = multifd_recv_unfill_packet(p, &local_err);
1120         if (ret) {
1121             qemu_mutex_unlock(&p->mutex);
1122             break;
1123         }
1124 
1125         flags = p->flags;
1126         /* recv methods don't know how to handle the SYNC flag */
1127         p->flags &= ~MULTIFD_FLAG_SYNC;
1128         trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags,
1129                            p->next_packet_size);
1130         p->num_packets++;
1131         p->total_normal_pages += p->normal_num;
1132         qemu_mutex_unlock(&p->mutex);
1133 
1134         if (p->normal_num) {
1135             ret = multifd_recv_state->ops->recv_pages(p, &local_err);
1136             if (ret != 0) {
1137                 break;
1138             }
1139         }
1140 
1141         if (flags & MULTIFD_FLAG_SYNC) {
1142             qemu_sem_post(&multifd_recv_state->sem_sync);
1143             qemu_sem_wait(&p->sem_sync);
1144         }
1145     }
1146 
1147     if (local_err) {
1148         multifd_recv_terminate_threads(local_err);
1149         error_free(local_err);
1150     }
1151     qemu_mutex_lock(&p->mutex);
1152     p->running = false;
1153     qemu_mutex_unlock(&p->mutex);
1154 
1155     rcu_unregister_thread();
1156     trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages);
1157 
1158     return NULL;
1159 }
1160 
1161 int multifd_load_setup(Error **errp)
1162 {
1163     int thread_count;
1164     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
1165     uint8_t i;
1166 
1167     if (!migrate_use_multifd()) {
1168         return 0;
1169     }
1170     if (!migrate_multi_channels_is_allowed()) {
1171         error_setg(errp, "multifd is not supported by current protocol");
1172         return -1;
1173     }
1174     thread_count = migrate_multifd_channels();
1175     multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state));
1176     multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count);
1177     qatomic_set(&multifd_recv_state->count, 0);
1178     qemu_sem_init(&multifd_recv_state->sem_sync, 0);
1179     multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()];
1180 
1181     for (i = 0; i < thread_count; i++) {
1182         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1183 
1184         qemu_mutex_init(&p->mutex);
1185         qemu_sem_init(&p->sem_sync, 0);
1186         p->quit = false;
1187         p->id = i;
1188         p->packet_len = sizeof(MultiFDPacket_t)
1189                       + sizeof(uint64_t) * page_count;
1190         p->packet = g_malloc0(p->packet_len);
1191         p->name = g_strdup_printf("multifdrecv_%d", i);
1192         p->iov = g_new0(struct iovec, page_count);
1193         p->normal = g_new0(ram_addr_t, page_count);
1194         p->page_count = page_count;
1195         p->page_size = qemu_target_page_size();
1196     }
1197 
1198     for (i = 0; i < thread_count; i++) {
1199         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1200         Error *local_err = NULL;
1201         int ret;
1202 
1203         ret = multifd_recv_state->ops->recv_setup(p, &local_err);
1204         if (ret) {
1205             error_propagate(errp, local_err);
1206             return ret;
1207         }
1208     }
1209     return 0;
1210 }
1211 
1212 bool multifd_recv_all_channels_created(void)
1213 {
1214     int thread_count = migrate_multifd_channels();
1215 
1216     if (!migrate_use_multifd()) {
1217         return true;
1218     }
1219 
1220     if (!multifd_recv_state) {
1221         /* Called before any connections created */
1222         return false;
1223     }
1224 
1225     return thread_count == qatomic_read(&multifd_recv_state->count);
1226 }
1227 
1228 /*
1229  * Try to receive all multifd channels to get ready for the migration.
1230  * - Return true and do not set @errp when correctly receiving all channels;
1231  * - Return false and do not set @errp when correctly receiving the current one;
1232  * - Return false and set @errp when failing to receive the current channel.
1233  */
1234 bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp)
1235 {
1236     MultiFDRecvParams *p;
1237     Error *local_err = NULL;
1238     int id;
1239 
1240     id = multifd_recv_initial_packet(ioc, &local_err);
1241     if (id < 0) {
1242         multifd_recv_terminate_threads(local_err);
1243         error_propagate_prepend(errp, local_err,
1244                                 "failed to receive packet"
1245                                 " via multifd channel %d: ",
1246                                 qatomic_read(&multifd_recv_state->count));
1247         return false;
1248     }
1249     trace_multifd_recv_new_channel(id);
1250 
1251     p = &multifd_recv_state->params[id];
1252     if (p->c != NULL) {
1253         error_setg(&local_err, "multifd: received id '%d' already setup'",
1254                    id);
1255         multifd_recv_terminate_threads(local_err);
1256         error_propagate(errp, local_err);
1257         return false;
1258     }
1259     p->c = ioc;
1260     object_ref(OBJECT(ioc));
1261     /* initial packet */
1262     p->num_packets = 1;
1263 
1264     p->running = true;
1265     qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p,
1266                        QEMU_THREAD_JOINABLE);
1267     qatomic_inc(&multifd_recv_state->count);
1268     return qatomic_read(&multifd_recv_state->count) ==
1269            migrate_multifd_channels();
1270 }
1271