migration/multifd.c

   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_update_transfer(f, transferred);
 439     ram_counters.multifd_bytes += transferred;
 440     ram_counters.transferred += transferred;
 441     qemu_mutex_unlock(&p->mutex);
 442     qemu_sem_post(&p->sem);
 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_multifd_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 void multifd_send_sync_main(QEMUFile *f)
 570 {
 571     int i;
 572
 573     if (!migrate_use_multifd()) {
 574         return;
 575     }
 576     if (multifd_send_state->pages->num) {
 577         if (multifd_send_pages(f) < 0) {
 578             error_report("%s: multifd_send_pages fail", __func__);
 579             return;
 580         }
 581     }
 582     for (i = 0; i < migrate_multifd_channels(); i++) {
 583         MultiFDSendParams *p = &multifd_send_state->params[i];
 584
 585         trace_multifd_send_sync_main_signal(p->id);
 586
 587         qemu_mutex_lock(&p->mutex);
 588
 589         if (p->quit) {
 590             error_report("%s: channel %d has already quit", __func__, i);
 591             qemu_mutex_unlock(&p->mutex);
 592             return;
 593         }
 594
 595         p->packet_num = multifd_send_state->packet_num++;
 596         p->flags |= MULTIFD_FLAG_SYNC;
 597         p->pending_job++;
 598         qemu_file_update_transfer(f, p->packet_len);
 599         ram_counters.multifd_bytes += p->packet_len;
 600         ram_counters.transferred += p->packet_len;
 601         qemu_mutex_unlock(&p->mutex);
 602         qemu_sem_post(&p->sem);
 603     }
 604     for (i = 0; i < migrate_multifd_channels(); i++) {
 605         MultiFDSendParams *p = &multifd_send_state->params[i];
 606
 607         trace_multifd_send_sync_main_wait(p->id);
 608         qemu_sem_wait(&p->sem_sync);
 609     }
 610     trace_multifd_send_sync_main(multifd_send_state->packet_num);
 611 }
 612
 613 static void *multifd_send_thread(void *opaque)
 614 {
 615     MultiFDSendParams *p = opaque;
 616     Error *local_err = NULL;
 617     int ret = 0;
 618
 619     trace_multifd_send_thread_start(p->id);
 620     rcu_register_thread();
 621
 622     if (multifd_send_initial_packet(p, &local_err) < 0) {
 623         ret = -1;
 624         goto out;
 625     }
 626     /* initial packet */
 627     p->num_packets = 1;
 628
 629     while (true) {
 630         qemu_sem_wait(&p->sem);
 631
 632         if (qatomic_read(&multifd_send_state->exiting)) {
 633             break;
 634         }
 635         qemu_mutex_lock(&p->mutex);
 636
 637         if (p->pending_job) {
 638             uint64_t packet_num = p->packet_num;
 639             uint32_t flags = p->flags;
 640             p->iovs_num = 1;
 641             p->normal_num = 0;
 642
 643             for (int i = 0; i < p->pages->num; i++) {
 644                 p->normal[p->normal_num] = p->pages->offset[i];
 645                 p->normal_num++;
 646             }
 647
 648             if (p->normal_num) {
 649                 ret = multifd_send_state->ops->send_prepare(p, &local_err);
 650                 if (ret != 0) {
 651                     qemu_mutex_unlock(&p->mutex);
 652                     break;
 653                 }
 654             }
 655             multifd_send_fill_packet(p);
 656             p->flags = 0;
 657             p->num_packets++;
 658             p->total_normal_pages += p->normal_num;
 659             p->pages->num = 0;
 660             p->pages->block = NULL;
 661             qemu_mutex_unlock(&p->mutex);
 662
 663             trace_multifd_send(p->id, packet_num, p->normal_num, flags,
 664                                p->next_packet_size);
 665
 666             p->iov[0].iov_len = p->packet_len;
 667             p->iov[0].iov_base = p->packet;
 668
 669             ret = qio_channel_writev_all(p->c, p->iov, p->iovs_num,
 670                                          &local_err);
 671             if (ret != 0) {
 672                 break;
 673             }
 674
 675             qemu_mutex_lock(&p->mutex);
 676             p->pending_job--;
 677             qemu_mutex_unlock(&p->mutex);
 678
 679             if (flags & MULTIFD_FLAG_SYNC) {
 680                 qemu_sem_post(&p->sem_sync);
 681             }
 682             qemu_sem_post(&multifd_send_state->channels_ready);
 683         } else if (p->quit) {
 684             qemu_mutex_unlock(&p->mutex);
 685             break;
 686         } else {
 687             qemu_mutex_unlock(&p->mutex);
 688             /* sometimes there are spurious wakeups */
 689         }
 690     }
 691
 692 out:
 693     if (local_err) {
 694         trace_multifd_send_error(p->id);
 695         multifd_send_terminate_threads(local_err);
 696         error_free(local_err);
 697     }
 698
 699     /*
 700      * Error happen, I will exit, but I can't just leave, tell
 701      * who pay attention to me.
 702      */
 703     if (ret != 0) {
 704         qemu_sem_post(&p->sem_sync);
 705         qemu_sem_post(&multifd_send_state->channels_ready);
 706     }
 707
 708     qemu_mutex_lock(&p->mutex);
 709     p->running = false;
 710     qemu_mutex_unlock(&p->mutex);
 711
 712     rcu_unregister_thread();
 713     trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages);
 714
 715     return NULL;
 716 }
 717
 718 static bool multifd_channel_connect(MultiFDSendParams *p,
 719                                     QIOChannel *ioc,
 720                                     Error *error);
 721
 722 static void multifd_tls_outgoing_handshake(QIOTask *task,
 723                                            gpointer opaque)
 724 {
 725     MultiFDSendParams *p = opaque;
 726     QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task));
 727     Error *err = NULL;
 728
 729     if (qio_task_propagate_error(task, &err)) {
 730         trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err));
 731     } else {
 732         trace_multifd_tls_outgoing_handshake_complete(ioc);
 733     }
 734
 735     if (!multifd_channel_connect(p, ioc, err)) {
 736         /*
 737          * Error happen, mark multifd_send_thread status as 'quit' although it
 738          * is not created, and then tell who pay attention to me.
 739          */
 740         p->quit = true;
 741         qemu_sem_post(&multifd_send_state->channels_ready);
 742         qemu_sem_post(&p->sem_sync);
 743     }
 744 }
 745
 746 static void *multifd_tls_handshake_thread(void *opaque)
 747 {
 748     MultiFDSendParams *p = opaque;
 749     QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
 750
 751     qio_channel_tls_handshake(tioc,
 752                               multifd_tls_outgoing_handshake,
 753                               p,
 754                               NULL,
 755                               NULL);
 756     return NULL;
 757 }
 758
 759 static void multifd_tls_channel_connect(MultiFDSendParams *p,
 760                                         QIOChannel *ioc,
 761                                         Error **errp)
 762 {
 763     MigrationState *s = migrate_get_current();
 764     const char *hostname = s->hostname;
 765     QIOChannelTLS *tioc;
 766
 767     tioc = migration_tls_client_create(s, ioc, hostname, errp);
 768     if (!tioc) {
 769         return;
 770     }
 771
 772     object_unref(OBJECT(ioc));
 773     trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
 774     qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
 775     p->c = QIO_CHANNEL(tioc);
 776     qemu_thread_create(&p->thread, "multifd-tls-handshake-worker",
 777                        multifd_tls_handshake_thread, p,
 778                        QEMU_THREAD_JOINABLE);
 779 }
 780
 781 static bool multifd_channel_connect(MultiFDSendParams *p,
 782                                     QIOChannel *ioc,
 783                                     Error *error)
 784 {
 785     MigrationState *s = migrate_get_current();
 786
 787     trace_multifd_set_outgoing_channel(
 788         ioc, object_get_typename(OBJECT(ioc)),
 789         migrate_get_current()->hostname, error);
 790
 791     if (!error) {
 792         if (s->parameters.tls_creds &&
 793             *s->parameters.tls_creds &&
 794             !object_dynamic_cast(OBJECT(ioc),
 795                                  TYPE_QIO_CHANNEL_TLS)) {
 796             multifd_tls_channel_connect(p, ioc, &error);
 797             if (!error) {
 798                 /*
 799                  * tls_channel_connect will call back to this
 800                  * function after the TLS handshake,
 801                  * so we mustn't call multifd_send_thread until then
 802                  */
 803                 return true;
 804             } else {
 805                 return false;
 806             }
 807         } else {
 808             migration_ioc_register_yank(ioc);
 809             p->registered_yank = true;
 810             p->c = ioc;
 811             qemu_thread_create(&p->thread, p->name, multifd_send_thread, p,
 812                                    QEMU_THREAD_JOINABLE);
 813        }
 814        return true;
 815     }
 816
 817     return false;
 818 }
 819
 820 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p,
 821                                              QIOChannel *ioc, Error *err)
 822 {
 823      migrate_set_error(migrate_get_current(), err);
 824      /* Error happen, we need to tell who pay attention to me */
 825      qemu_sem_post(&multifd_send_state->channels_ready);
 826      qemu_sem_post(&p->sem_sync);
 827      /*
 828       * Although multifd_send_thread is not created, but main migration
 829       * thread neet to judge whether it is running, so we need to mark
 830       * its status.
 831       */
 832      p->quit = true;
 833      object_unref(OBJECT(ioc));
 834      error_free(err);
 835 }
 836
 837 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque)
 838 {
 839     MultiFDSendParams *p = opaque;
 840     QIOChannel *sioc = QIO_CHANNEL(qio_task_get_source(task));
 841     Error *local_err = NULL;
 842
 843     trace_multifd_new_send_channel_async(p->id);
 844     if (qio_task_propagate_error(task, &local_err)) {
 845         goto cleanup;
 846     } else {
 847         p->c = QIO_CHANNEL(sioc);
 848         qio_channel_set_delay(p->c, false);
 849         p->running = true;
 850         if (!multifd_channel_connect(p, sioc, local_err)) {
 851             goto cleanup;
 852         }
 853         return;
 854     }
 855
 856 cleanup:
 857     multifd_new_send_channel_cleanup(p, sioc, local_err);
 858 }
 859
 860 static bool migrate_allow_multifd = true;
 861 void migrate_protocol_allow_multifd(bool allow)
 862 {
 863     migrate_allow_multifd = allow;
 864 }
 865
 866 bool migrate_multifd_is_allowed(void)
 867 {
 868     return migrate_allow_multifd;
 869 }
 870
 871 int multifd_save_setup(Error **errp)
 872 {
 873     int thread_count;
 874     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
 875     uint8_t i;
 876
 877     if (!migrate_use_multifd()) {
 878         return 0;
 879     }
 880     if (!migrate_multifd_is_allowed()) {
 881         error_setg(errp, "multifd is not supported by current protocol");
 882         return -1;
 883     }
 884
 885     thread_count = migrate_multifd_channels();
 886     multifd_send_state = g_malloc0(sizeof(*multifd_send_state));
 887     multifd_send_state->params = g_new0(MultiFDSendParams, thread_count);
 888     multifd_send_state->pages = multifd_pages_init(page_count);
 889     qemu_sem_init(&multifd_send_state->channels_ready, 0);
 890     qatomic_set(&multifd_send_state->exiting, 0);
 891     multifd_send_state->ops = multifd_ops[migrate_multifd_compression()];
 892
 893     for (i = 0; i < thread_count; i++) {
 894         MultiFDSendParams *p = &multifd_send_state->params[i];
 895
 896         qemu_mutex_init(&p->mutex);
 897         qemu_sem_init(&p->sem, 0);
 898         qemu_sem_init(&p->sem_sync, 0);
 899         p->quit = false;
 900         p->pending_job = 0;
 901         p->id = i;
 902         p->pages = multifd_pages_init(page_count);
 903         p->packet_len = sizeof(MultiFDPacket_t)
 904                       + sizeof(uint64_t) * page_count;
 905         p->packet = g_malloc0(p->packet_len);
 906         p->packet->magic = cpu_to_be32(MULTIFD_MAGIC);
 907         p->packet->version = cpu_to_be32(MULTIFD_VERSION);
 908         p->name = g_strdup_printf("multifdsend_%d", i);
 909         /* We need one extra place for the packet header */
 910         p->iov = g_new0(struct iovec, page_count + 1);
 911         p->normal = g_new0(ram_addr_t, page_count);
 912         socket_send_channel_create(multifd_new_send_channel_async, p);
 913     }
 914
 915     for (i = 0; i < thread_count; i++) {
 916         MultiFDSendParams *p = &multifd_send_state->params[i];
 917         Error *local_err = NULL;
 918         int ret;
 919
 920         ret = multifd_send_state->ops->send_setup(p, &local_err);
 921         if (ret) {
 922             error_propagate(errp, local_err);
 923             return ret;
 924         }
 925     }
 926     return 0;
 927 }
 928
 929 struct {
 930     MultiFDRecvParams *params;
 931     /* number of created threads */
 932     int count;
 933     /* syncs main thread and channels */
 934     QemuSemaphore sem_sync;
 935     /* global number of generated multifd packets */
 936     uint64_t packet_num;
 937     /* multifd ops */
 938     MultiFDMethods *ops;
 939 } *multifd_recv_state;
 940
 941 static void multifd_recv_terminate_threads(Error *err)
 942 {
 943     int i;
 944
 945     trace_multifd_recv_terminate_threads(err != NULL);
 946
 947     if (err) {
 948         MigrationState *s = migrate_get_current();
 949         migrate_set_error(s, err);
 950         if (s->state == MIGRATION_STATUS_SETUP ||
 951             s->state == MIGRATION_STATUS_ACTIVE) {
 952             migrate_set_state(&s->state, s->state,
 953                               MIGRATION_STATUS_FAILED);
 954         }
 955     }
 956
 957     for (i = 0; i < migrate_multifd_channels(); i++) {
 958         MultiFDRecvParams *p = &multifd_recv_state->params[i];
 959
 960         qemu_mutex_lock(&p->mutex);
 961         p->quit = true;
 962         /*
 963          * We could arrive here for two reasons:
 964          *  - normal quit, i.e. everything went fine, just finished
 965          *  - error quit: We close the channels so the channel threads
 966          *    finish the qio_channel_read_all_eof()
 967          */
 968         if (p->c) {
 969             qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
 970         }
 971         qemu_mutex_unlock(&p->mutex);
 972     }
 973 }
 974
 975 int multifd_load_cleanup(Error **errp)
 976 {
 977     int i;
 978
 979     if (!migrate_use_multifd() || !migrate_multifd_is_allowed()) {
 980         return 0;
 981     }
 982     multifd_recv_terminate_threads(NULL);
 983     for (i = 0; i < migrate_multifd_channels(); i++) {
 984         MultiFDRecvParams *p = &multifd_recv_state->params[i];
 985
 986         if (p->running) {
 987             p->quit = true;
 988             /*
 989              * multifd_recv_thread may hung at MULTIFD_FLAG_SYNC handle code,
 990              * however try to wakeup it without harm in cleanup phase.
 991              */
 992             qemu_sem_post(&p->sem_sync);
 993             qemu_thread_join(&p->thread);
 994         }
 995     }
 996     for (i = 0; i < migrate_multifd_channels(); i++) {
 997         MultiFDRecvParams *p = &multifd_recv_state->params[i];
 998
 999         migration_ioc_unregister_yank(p->c);
1000         object_unref(OBJECT(p->c));
1001         p->c = NULL;
1002         qemu_mutex_destroy(&p->mutex);
1003         qemu_sem_destroy(&p->sem_sync);
1004         g_free(p->name);
1005         p->name = NULL;
1006         p->packet_len = 0;
1007         g_free(p->packet);
1008         p->packet = NULL;
1009         g_free(p->iov);
1010         p->iov = NULL;
1011         g_free(p->normal);
1012         p->normal = NULL;
1013         multifd_recv_state->ops->recv_cleanup(p);
1014     }
1015     qemu_sem_destroy(&multifd_recv_state->sem_sync);
1016     g_free(multifd_recv_state->params);
1017     multifd_recv_state->params = NULL;
1018     g_free(multifd_recv_state);
1019     multifd_recv_state = NULL;
1020
1021     return 0;
1022 }
1023
1024 void multifd_recv_sync_main(void)
1025 {
1026     int i;
1027
1028     if (!migrate_use_multifd()) {
1029         return;
1030     }
1031     for (i = 0; i < migrate_multifd_channels(); i++) {
1032         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1033
1034         trace_multifd_recv_sync_main_wait(p->id);
1035         qemu_sem_wait(&multifd_recv_state->sem_sync);
1036     }
1037     for (i = 0; i < migrate_multifd_channels(); i++) {
1038         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1039
1040         WITH_QEMU_LOCK_GUARD(&p->mutex) {
1041             if (multifd_recv_state->packet_num < p->packet_num) {
1042                 multifd_recv_state->packet_num = p->packet_num;
1043             }
1044         }
1045         trace_multifd_recv_sync_main_signal(p->id);
1046         qemu_sem_post(&p->sem_sync);
1047     }
1048     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
1049 }
1050
1051 static void *multifd_recv_thread(void *opaque)
1052 {
1053     MultiFDRecvParams *p = opaque;
1054     Error *local_err = NULL;
1055     int ret;
1056
1057     trace_multifd_recv_thread_start(p->id);
1058     rcu_register_thread();
1059
1060     while (true) {
1061         uint32_t flags;
1062
1063         if (p->quit) {
1064             break;
1065         }
1066
1067         ret = qio_channel_read_all_eof(p->c, (void *)p->packet,
1068                                        p->packet_len, &local_err);
1069         if (ret == 0) {   /* EOF */
1070             break;
1071         }
1072         if (ret == -1) {   /* Error */
1073             break;
1074         }
1075
1076         qemu_mutex_lock(&p->mutex);
1077         ret = multifd_recv_unfill_packet(p, &local_err);
1078         if (ret) {
1079             qemu_mutex_unlock(&p->mutex);
1080             break;
1081         }
1082
1083         flags = p->flags;
1084         /* recv methods don't know how to handle the SYNC flag */
1085         p->flags &= ~MULTIFD_FLAG_SYNC;
1086         trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags,
1087                            p->next_packet_size);
1088         p->num_packets++;
1089         p->total_normal_pages += p->normal_num;
1090         qemu_mutex_unlock(&p->mutex);
1091
1092         if (p->normal_num) {
1093             ret = multifd_recv_state->ops->recv_pages(p, &local_err);
1094             if (ret != 0) {
1095                 break;
1096             }
1097         }
1098
1099         if (flags & MULTIFD_FLAG_SYNC) {
1100             qemu_sem_post(&multifd_recv_state->sem_sync);
1101             qemu_sem_wait(&p->sem_sync);
1102         }
1103     }
1104
1105     if (local_err) {
1106         multifd_recv_terminate_threads(local_err);
1107         error_free(local_err);
1108     }
1109     qemu_mutex_lock(&p->mutex);
1110     p->running = false;
1111     qemu_mutex_unlock(&p->mutex);
1112
1113     rcu_unregister_thread();
1114     trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages);
1115
1116     return NULL;
1117 }
1118
1119 int multifd_load_setup(Error **errp)
1120 {
1121     int thread_count;
1122     uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size();
1123     uint8_t i;
1124
1125     if (!migrate_use_multifd()) {
1126         return 0;
1127     }
1128     if (!migrate_multifd_is_allowed()) {
1129         error_setg(errp, "multifd is not supported by current protocol");
1130         return -1;
1131     }
1132     thread_count = migrate_multifd_channels();
1133     multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state));
1134     multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count);
1135     qatomic_set(&multifd_recv_state->count, 0);
1136     qemu_sem_init(&multifd_recv_state->sem_sync, 0);
1137     multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()];
1138
1139     for (i = 0; i < thread_count; i++) {
1140         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1141
1142         qemu_mutex_init(&p->mutex);
1143         qemu_sem_init(&p->sem_sync, 0);
1144         p->quit = false;
1145         p->id = i;
1146         p->packet_len = sizeof(MultiFDPacket_t)
1147                       + sizeof(uint64_t) * page_count;
1148         p->packet = g_malloc0(p->packet_len);
1149         p->name = g_strdup_printf("multifdrecv_%d", i);
1150         p->iov = g_new0(struct iovec, page_count);
1151         p->normal = g_new0(ram_addr_t, page_count);
1152     }
1153
1154     for (i = 0; i < thread_count; i++) {
1155         MultiFDRecvParams *p = &multifd_recv_state->params[i];
1156         Error *local_err = NULL;
1157         int ret;
1158
1159         ret = multifd_recv_state->ops->recv_setup(p, &local_err);
1160         if (ret) {
1161             error_propagate(errp, local_err);
1162             return ret;
1163         }
1164     }
1165     return 0;
1166 }
1167
1168 bool multifd_recv_all_channels_created(void)
1169 {
1170     int thread_count = migrate_multifd_channels();
1171
1172     if (!migrate_use_multifd()) {
1173         return true;
1174     }
1175
1176     if (!multifd_recv_state) {
1177         /* Called before any connections created */
1178         return false;
1179     }
1180
1181     return thread_count == qatomic_read(&multifd_recv_state->count);
1182 }
1183
1184 /*
1185  * Try to receive all multifd channels to get ready for the migration.
1186  * - Return true and do not set @errp when correctly receiving all channels;
1187  * - Return false and do not set @errp when correctly receiving the current one;
1188  * - Return false and set @errp when failing to receive the current channel.
1189  */
1190 bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp)
1191 {
1192     MultiFDRecvParams *p;
1193     Error *local_err = NULL;
1194     int id;
1195
1196     id = multifd_recv_initial_packet(ioc, &local_err);
1197     if (id < 0) {
1198         multifd_recv_terminate_threads(local_err);
1199         error_propagate_prepend(errp, local_err,
1200                                 "failed to receive packet"
1201                                 " via multifd channel %d: ",
1202                                 qatomic_read(&multifd_recv_state->count));
1203         return false;
1204     }
1205     trace_multifd_recv_new_channel(id);
1206
1207     p = &multifd_recv_state->params[id];
1208     if (p->c != NULL) {
1209         error_setg(&local_err, "multifd: received id '%d' already setup'",
1210                    id);
1211         multifd_recv_terminate_threads(local_err);
1212         error_propagate(errp, local_err);
1213         return false;
1214     }
1215     p->c = ioc;
1216     object_ref(OBJECT(ioc));
1217     /* initial packet */
1218     p->num_packets = 1;
1219
1220     p->running = true;
1221     qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p,
1222                        QEMU_THREAD_JOINABLE);
1223     qatomic_inc(&multifd_recv_state->count);
1224     return qatomic_read(&multifd_recv_state->count) ==
1225            migrate_multifd_channels();
1226 }