arch_init.c

   1 /*
   2  * QEMU System Emulator
   3  *
   4  * Copyright (c) 2003-2008 Fabrice Bellard
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to deal
   8  * in the Software without restriction, including without limitation the rights
   9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  10  * copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  22  * THE SOFTWARE.
  23  */
  24 #include <stdint.h>
  25 #include <stdarg.h>
  26 #include <stdlib.h>
  27 #include <zlib.h>
  28 #ifndef _WIN32
  29 #include <sys/types.h>
  30 #include <sys/mman.h>
  31 #endif
  32 #include "config.h"
  33 #include "monitor/monitor.h"
  34 #include "sysemu/sysemu.h"
  35 #include "qemu/bitops.h"
  36 #include "qemu/bitmap.h"
  37 #include "sysemu/arch_init.h"
  38 #include "audio/audio.h"
  39 #include "hw/i386/pc.h"
  40 #include "hw/pci/pci.h"
  41 #include "hw/audio/audio.h"
  42 #include "sysemu/kvm.h"
  43 #include "migration/migration.h"
  44 #include "hw/i386/smbios.h"
  45 #include "exec/address-spaces.h"
  46 #include "hw/audio/pcspk.h"
  47 #include "migration/page_cache.h"
  48 #include "qemu/config-file.h"
  49 #include "qemu/error-report.h"
  50 #include "qmp-commands.h"
  51 #include "trace.h"
  52 #include "exec/cpu-all.h"
  53 #include "exec/ram_addr.h"
  54 #include "hw/acpi/acpi.h"
  55 #include "qemu/host-utils.h"
  56 #include "qemu/rcu_queue.h"
  57
  58 #ifdef DEBUG_ARCH_INIT
  59 #define DPRINTF(fmt, ...) \
  60     do { fprintf(stdout, "arch_init: " fmt, ## __VA_ARGS__); } while (0)
  61 #else
  62 #define DPRINTF(fmt, ...) \
  63     do { } while (0)
  64 #endif
  65
  66 #ifdef TARGET_SPARC
  67 int graphic_width = 1024;
  68 int graphic_height = 768;
  69 int graphic_depth = 8;
  70 #else
  71 int graphic_width = 800;
  72 int graphic_height = 600;
  73 int graphic_depth = 32;
  74 #endif
  75
  76
  77 #if defined(TARGET_ALPHA)
  78 #define QEMU_ARCH QEMU_ARCH_ALPHA
  79 #elif defined(TARGET_ARM)
  80 #define QEMU_ARCH QEMU_ARCH_ARM
  81 #elif defined(TARGET_CRIS)
  82 #define QEMU_ARCH QEMU_ARCH_CRIS
  83 #elif defined(TARGET_I386)
  84 #define QEMU_ARCH QEMU_ARCH_I386
  85 #elif defined(TARGET_M68K)
  86 #define QEMU_ARCH QEMU_ARCH_M68K
  87 #elif defined(TARGET_LM32)
  88 #define QEMU_ARCH QEMU_ARCH_LM32
  89 #elif defined(TARGET_MICROBLAZE)
  90 #define QEMU_ARCH QEMU_ARCH_MICROBLAZE
  91 #elif defined(TARGET_MIPS)
  92 #define QEMU_ARCH QEMU_ARCH_MIPS
  93 #elif defined(TARGET_MOXIE)
  94 #define QEMU_ARCH QEMU_ARCH_MOXIE
  95 #elif defined(TARGET_OPENRISC)
  96 #define QEMU_ARCH QEMU_ARCH_OPENRISC
  97 #elif defined(TARGET_PPC)
  98 #define QEMU_ARCH QEMU_ARCH_PPC
  99 #elif defined(TARGET_S390X)
 100 #define QEMU_ARCH QEMU_ARCH_S390X
 101 #elif defined(TARGET_SH4)
 102 #define QEMU_ARCH QEMU_ARCH_SH4
 103 #elif defined(TARGET_SPARC)
 104 #define QEMU_ARCH QEMU_ARCH_SPARC
 105 #elif defined(TARGET_XTENSA)
 106 #define QEMU_ARCH QEMU_ARCH_XTENSA
 107 #elif defined(TARGET_UNICORE32)
 108 #define QEMU_ARCH QEMU_ARCH_UNICORE32
 109 #elif defined(TARGET_TRICORE)
 110 #define QEMU_ARCH QEMU_ARCH_TRICORE
 111 #endif
 112
 113 const uint32_t arch_type = QEMU_ARCH;
 114 static bool mig_throttle_on;
 115 static int dirty_rate_high_cnt;
 116 static void check_guest_throttling(void);
 117
 118 static uint64_t bitmap_sync_count;
 119
 120 /***********************************************************/
 121 /* ram save/restore */
 122
 123 #define RAM_SAVE_FLAG_FULL     0x01 /* Obsolete, not used anymore */
 124 #define RAM_SAVE_FLAG_COMPRESS 0x02
 125 #define RAM_SAVE_FLAG_MEM_SIZE 0x04
 126 #define RAM_SAVE_FLAG_PAGE     0x08
 127 #define RAM_SAVE_FLAG_EOS      0x10
 128 #define RAM_SAVE_FLAG_CONTINUE 0x20
 129 #define RAM_SAVE_FLAG_XBZRLE   0x40
 130 /* 0x80 is reserved in migration.h start with 0x100 next */
 131 #define RAM_SAVE_FLAG_COMPRESS_PAGE    0x100
 132
 133 static struct defconfig_file {
 134     const char *filename;
 135     /* Indicates it is an user config file (disabled by -no-user-config) */
 136     bool userconfig;
 137 } default_config_files[] = {
 138     { CONFIG_QEMU_CONFDIR "/qemu.conf",                   true },
 139     { CONFIG_QEMU_CONFDIR "/target-" TARGET_NAME ".conf", true },
 140     { NULL }, /* end of list */
 141 };
 142
 143 static const uint8_t ZERO_TARGET_PAGE[TARGET_PAGE_SIZE];
 144
 145 int qemu_read_default_config_files(bool userconfig)
 146 {
 147     int ret;
 148     struct defconfig_file *f;
 149
 150     for (f = default_config_files; f->filename; f++) {
 151         if (!userconfig && f->userconfig) {
 152             continue;
 153         }
 154         ret = qemu_read_config_file(f->filename);
 155         if (ret < 0 && ret != -ENOENT) {
 156             return ret;
 157         }
 158     }
 159
 160     return 0;
 161 }
 162
 163 static inline bool is_zero_range(uint8_t *p, uint64_t size)
 164 {
 165     return buffer_find_nonzero_offset(p, size) == size;
 166 }
 167
 168 /* struct contains XBZRLE cache and a static page
 169    used by the compression */
 170 static struct {
 171     /* buffer used for XBZRLE encoding */
 172     uint8_t *encoded_buf;
 173     /* buffer for storing page content */
 174     uint8_t *current_buf;
 175     /* Cache for XBZRLE, Protected by lock. */
 176     PageCache *cache;
 177     QemuMutex lock;
 178 } XBZRLE;
 179
 180 /* buffer used for XBZRLE decoding */
 181 static uint8_t *xbzrle_decoded_buf;
 182
 183 static void XBZRLE_cache_lock(void)
 184 {
 185     if (migrate_use_xbzrle())
 186         qemu_mutex_lock(&XBZRLE.lock);
 187 }
 188
 189 static void XBZRLE_cache_unlock(void)
 190 {
 191     if (migrate_use_xbzrle())
 192         qemu_mutex_unlock(&XBZRLE.lock);
 193 }
 194
 195 /*
 196  * called from qmp_migrate_set_cache_size in main thread, possibly while
 197  * a migration is in progress.
 198  * A running migration maybe using the cache and might finish during this
 199  * call, hence changes to the cache are protected by XBZRLE.lock().
 200  */
 201 int64_t xbzrle_cache_resize(int64_t new_size)
 202 {
 203     PageCache *new_cache;
 204     int64_t ret;
 205
 206     if (new_size < TARGET_PAGE_SIZE) {
 207         return -1;
 208     }
 209
 210     XBZRLE_cache_lock();
 211
 212     if (XBZRLE.cache != NULL) {
 213         if (pow2floor(new_size) == migrate_xbzrle_cache_size()) {
 214             goto out_new_size;
 215         }
 216         new_cache = cache_init(new_size / TARGET_PAGE_SIZE,
 217                                         TARGET_PAGE_SIZE);
 218         if (!new_cache) {
 219             error_report("Error creating cache");
 220             ret = -1;
 221             goto out;
 222         }
 223
 224         cache_fini(XBZRLE.cache);
 225         XBZRLE.cache = new_cache;
 226     }
 227
 228 out_new_size:
 229     ret = pow2floor(new_size);
 230 out:
 231     XBZRLE_cache_unlock();
 232     return ret;
 233 }
 234
 235 /* accounting for migration statistics */
 236 typedef struct AccountingInfo {
 237     uint64_t dup_pages;
 238     uint64_t skipped_pages;
 239     uint64_t norm_pages;
 240     uint64_t iterations;
 241     uint64_t xbzrle_bytes;
 242     uint64_t xbzrle_pages;
 243     uint64_t xbzrle_cache_miss;
 244     double xbzrle_cache_miss_rate;
 245     uint64_t xbzrle_overflows;
 246 } AccountingInfo;
 247
 248 static AccountingInfo acct_info;
 249
 250 static void acct_clear(void)
 251 {
 252     memset(&acct_info, 0, sizeof(acct_info));
 253 }
 254
 255 uint64_t dup_mig_bytes_transferred(void)
 256 {
 257     return acct_info.dup_pages * TARGET_PAGE_SIZE;
 258 }
 259
 260 uint64_t dup_mig_pages_transferred(void)
 261 {
 262     return acct_info.dup_pages;
 263 }
 264
 265 uint64_t skipped_mig_bytes_transferred(void)
 266 {
 267     return acct_info.skipped_pages * TARGET_PAGE_SIZE;
 268 }
 269
 270 uint64_t skipped_mig_pages_transferred(void)
 271 {
 272     return acct_info.skipped_pages;
 273 }
 274
 275 uint64_t norm_mig_bytes_transferred(void)
 276 {
 277     return acct_info.norm_pages * TARGET_PAGE_SIZE;
 278 }
 279
 280 uint64_t norm_mig_pages_transferred(void)
 281 {
 282     return acct_info.norm_pages;
 283 }
 284
 285 uint64_t xbzrle_mig_bytes_transferred(void)
 286 {
 287     return acct_info.xbzrle_bytes;
 288 }
 289
 290 uint64_t xbzrle_mig_pages_transferred(void)
 291 {
 292     return acct_info.xbzrle_pages;
 293 }
 294
 295 uint64_t xbzrle_mig_pages_cache_miss(void)
 296 {
 297     return acct_info.xbzrle_cache_miss;
 298 }
 299
 300 double xbzrle_mig_cache_miss_rate(void)
 301 {
 302     return acct_info.xbzrle_cache_miss_rate;
 303 }
 304
 305 uint64_t xbzrle_mig_pages_overflow(void)
 306 {
 307     return acct_info.xbzrle_overflows;
 308 }
 309
 310 /* This is the last block that we have visited serching for dirty pages
 311  */
 312 static RAMBlock *last_seen_block;
 313 /* This is the last block from where we have sent data */
 314 static RAMBlock *last_sent_block;
 315 static ram_addr_t last_offset;
 316 static unsigned long *migration_bitmap;
 317 static uint64_t migration_dirty_pages;
 318 static uint32_t last_version;
 319 static bool ram_bulk_stage;
 320
 321 struct CompressParam {
 322     bool start;
 323     bool done;
 324     QEMUFile *file;
 325     QemuMutex mutex;
 326     QemuCond cond;
 327     RAMBlock *block;
 328     ram_addr_t offset;
 329 };
 330 typedef struct CompressParam CompressParam;
 331
 332 struct DecompressParam {
 333     bool start;
 334     QemuMutex mutex;
 335     QemuCond cond;
 336     void *des;
 337     uint8 *compbuf;
 338     int len;
 339 };
 340 typedef struct DecompressParam DecompressParam;
 341
 342 static CompressParam *comp_param;
 343 static QemuThread *compress_threads;
 344 /* comp_done_cond is used to wake up the migration thread when
 345  * one of the compression threads has finished the compression.
 346  * comp_done_lock is used to co-work with comp_done_cond.
 347  */
 348 static QemuMutex *comp_done_lock;
 349 static QemuCond *comp_done_cond;
 350 /* The empty QEMUFileOps will be used by file in CompressParam */
 351 static const QEMUFileOps empty_ops = { };
 352
 353 static bool compression_switch;
 354 static bool quit_comp_thread;
 355 static bool quit_decomp_thread;
 356 static DecompressParam *decomp_param;
 357 static QemuThread *decompress_threads;
 358 static uint8_t *compressed_data_buf;
 359
 360 static int do_compress_ram_page(CompressParam *param);
 361
 362 static void *do_data_compress(void *opaque)
 363 {
 364     CompressParam *param = opaque;
 365
 366     while (!quit_comp_thread) {
 367         qemu_mutex_lock(&param->mutex);
 368         /* Re-check the quit_comp_thread in case of
 369          * terminate_compression_threads is called just before
 370          * qemu_mutex_lock(&param->mutex) and after
 371          * while(!quit_comp_thread), re-check it here can make
 372          * sure the compression thread terminate as expected.
 373          */
 374         while (!param->start && !quit_comp_thread) {
 375             qemu_cond_wait(&param->cond, &param->mutex);
 376         }
 377         if (!quit_comp_thread) {
 378             do_compress_ram_page(param);
 379         }
 380         param->start = false;
 381         qemu_mutex_unlock(&param->mutex);
 382
 383         qemu_mutex_lock(comp_done_lock);
 384         param->done = true;
 385         qemu_cond_signal(comp_done_cond);
 386         qemu_mutex_unlock(comp_done_lock);
 387     }
 388
 389     return NULL;
 390 }
 391
 392 static inline void terminate_compression_threads(void)
 393 {
 394     int idx, thread_count;
 395
 396     thread_count = migrate_compress_threads();
 397     quit_comp_thread = true;
 398     for (idx = 0; idx < thread_count; idx++) {
 399         qemu_mutex_lock(&comp_param[idx].mutex);
 400         qemu_cond_signal(&comp_param[idx].cond);
 401         qemu_mutex_unlock(&comp_param[idx].mutex);
 402     }
 403 }
 404
 405 void migrate_compress_threads_join(void)
 406 {
 407     int i, thread_count;
 408
 409     if (!migrate_use_compression()) {
 410         return;
 411     }
 412     terminate_compression_threads();
 413     thread_count = migrate_compress_threads();
 414     for (i = 0; i < thread_count; i++) {
 415         qemu_thread_join(compress_threads + i);
 416         qemu_fclose(comp_param[i].file);
 417         qemu_mutex_destroy(&comp_param[i].mutex);
 418         qemu_cond_destroy(&comp_param[i].cond);
 419     }
 420     qemu_mutex_destroy(comp_done_lock);
 421     qemu_cond_destroy(comp_done_cond);
 422     g_free(compress_threads);
 423     g_free(comp_param);
 424     g_free(comp_done_cond);
 425     g_free(comp_done_lock);
 426     compress_threads = NULL;
 427     comp_param = NULL;
 428     comp_done_cond = NULL;
 429     comp_done_lock = NULL;
 430 }
 431
 432 void migrate_compress_threads_create(void)
 433 {
 434     int i, thread_count;
 435
 436     if (!migrate_use_compression()) {
 437         return;
 438     }
 439     quit_comp_thread = false;
 440     compression_switch = true;
 441     thread_count = migrate_compress_threads();
 442     compress_threads = g_new0(QemuThread, thread_count);
 443     comp_param = g_new0(CompressParam, thread_count);
 444     comp_done_cond = g_new0(QemuCond, 1);
 445     comp_done_lock = g_new0(QemuMutex, 1);
 446     qemu_cond_init(comp_done_cond);
 447     qemu_mutex_init(comp_done_lock);
 448     for (i = 0; i < thread_count; i++) {
 449         /* com_param[i].file is just used as a dummy buffer to save data, set
 450          * it's ops to empty.
 451          */
 452         comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops);
 453         comp_param[i].done = true;
 454         qemu_mutex_init(&comp_param[i].mutex);
 455         qemu_cond_init(&comp_param[i].cond);
 456         qemu_thread_create(compress_threads + i, "compress",
 457                            do_data_compress, comp_param + i,
 458                            QEMU_THREAD_JOINABLE);
 459     }
 460 }
 461
 462 /**
 463  * save_page_header: Write page header to wire
 464  *
 465  * If this is the 1st block, it also writes the block identification
 466  *
 467  * Returns: Number of bytes written
 468  *
 469  * @f: QEMUFile where to send the data
 470  * @block: block that contains the page we want to send
 471  * @offset: offset inside the block for the page
 472  *          in the lower bits, it contains flags
 473  */
 474 static size_t save_page_header(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
 475 {
 476     size_t size;
 477
 478     qemu_put_be64(f, offset);
 479     size = 8;
 480
 481     if (!(offset & RAM_SAVE_FLAG_CONTINUE)) {
 482         qemu_put_byte(f, strlen(block->idstr));
 483         qemu_put_buffer(f, (uint8_t *)block->idstr,
 484                         strlen(block->idstr));
 485         size += 1 + strlen(block->idstr);
 486     }
 487     return size;
 488 }
 489
 490 /* Update the xbzrle cache to reflect a page that's been sent as all 0.
 491  * The important thing is that a stale (not-yet-0'd) page be replaced
 492  * by the new data.
 493  * As a bonus, if the page wasn't in the cache it gets added so that
 494  * when a small write is made into the 0'd page it gets XBZRLE sent
 495  */
 496 static void xbzrle_cache_zero_page(ram_addr_t current_addr)
 497 {
 498     if (ram_bulk_stage || !migrate_use_xbzrle()) {
 499         return;
 500     }
 501
 502     /* We don't care if this fails to allocate a new cache page
 503      * as long as it updated an old one */
 504     cache_insert(XBZRLE.cache, current_addr, ZERO_TARGET_PAGE,
 505                  bitmap_sync_count);
 506 }
 507
 508 #define ENCODING_FLAG_XBZRLE 0x1
 509
 510 /**
 511  * save_xbzrle_page: compress and send current page
 512  *
 513  * Returns: 1 means that we wrote the page
 514  *          0 means that page is identical to the one already sent
 515  *          -1 means that xbzrle would be longer than normal
 516  *
 517  * @f: QEMUFile where to send the data
 518  * @current_data:
 519  * @current_addr:
 520  * @block: block that contains the page we want to send
 521  * @offset: offset inside the block for the page
 522  * @last_stage: if we are at the completion stage
 523  * @bytes_transferred: increase it with the number of transferred bytes
 524  */
 525 static int save_xbzrle_page(QEMUFile *f, uint8_t **current_data,
 526                             ram_addr_t current_addr, RAMBlock *block,
 527                             ram_addr_t offset, bool last_stage,
 528                             uint64_t *bytes_transferred)
 529 {
 530     int encoded_len = 0, bytes_xbzrle;
 531     uint8_t *prev_cached_page;
 532
 533     if (!cache_is_cached(XBZRLE.cache, current_addr, bitmap_sync_count)) {
 534         acct_info.xbzrle_cache_miss++;
 535         if (!last_stage) {
 536             if (cache_insert(XBZRLE.cache, current_addr, *current_data,
 537                              bitmap_sync_count) == -1) {
 538                 return -1;
 539             } else {
 540                 /* update *current_data when the page has been
 541                    inserted into cache */
 542                 *current_data = get_cached_data(XBZRLE.cache, current_addr);
 543             }
 544         }
 545         return -1;
 546     }
 547
 548     prev_cached_page = get_cached_data(XBZRLE.cache, current_addr);
 549
 550     /* save current buffer into memory */
 551     memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE);
 552
 553     /* XBZRLE encoding (if there is no overflow) */
 554     encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf,
 555                                        TARGET_PAGE_SIZE, XBZRLE.encoded_buf,
 556                                        TARGET_PAGE_SIZE);
 557     if (encoded_len == 0) {
 558         DPRINTF("Skipping unmodified page\n");
 559         return 0;
 560     } else if (encoded_len == -1) {
 561         DPRINTF("Overflow\n");
 562         acct_info.xbzrle_overflows++;
 563         /* update data in the cache */
 564         if (!last_stage) {
 565             memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE);
 566             *current_data = prev_cached_page;
 567         }
 568         return -1;
 569     }
 570
 571     /* we need to update the data in the cache, in order to get the same data */
 572     if (!last_stage) {
 573         memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE);
 574     }
 575
 576     /* Send XBZRLE based compressed page */
 577     bytes_xbzrle = save_page_header(f, block, offset | RAM_SAVE_FLAG_XBZRLE);
 578     qemu_put_byte(f, ENCODING_FLAG_XBZRLE);
 579     qemu_put_be16(f, encoded_len);
 580     qemu_put_buffer(f, XBZRLE.encoded_buf, encoded_len);
 581     bytes_xbzrle += encoded_len + 1 + 2;
 582     acct_info.xbzrle_pages++;
 583     acct_info.xbzrle_bytes += bytes_xbzrle;
 584     *bytes_transferred += bytes_xbzrle;
 585
 586     return 1;
 587 }
 588
 589 static inline
 590 ram_addr_t migration_bitmap_find_and_reset_dirty(MemoryRegion *mr,
 591                                                  ram_addr_t start)
 592 {
 593     unsigned long base = mr->ram_addr >> TARGET_PAGE_BITS;
 594     unsigned long nr = base + (start >> TARGET_PAGE_BITS);
 595     uint64_t mr_size = TARGET_PAGE_ALIGN(memory_region_size(mr));
 596     unsigned long size = base + (mr_size >> TARGET_PAGE_BITS);
 597
 598     unsigned long next;
 599
 600     if (ram_bulk_stage && nr > base) {
 601         next = nr + 1;
 602     } else {
 603         next = find_next_bit(migration_bitmap, size, nr);
 604     }
 605
 606     if (next < size) {
 607         clear_bit(next, migration_bitmap);
 608         migration_dirty_pages--;
 609     }
 610     return (next - base) << TARGET_PAGE_BITS;
 611 }
 612
 613 static inline bool migration_bitmap_set_dirty(ram_addr_t addr)
 614 {
 615     bool ret;
 616     int nr = addr >> TARGET_PAGE_BITS;
 617
 618     ret = test_and_set_bit(nr, migration_bitmap);
 619
 620     if (!ret) {
 621         migration_dirty_pages++;
 622     }
 623     return ret;
 624 }
 625
 626 static void migration_bitmap_sync_range(ram_addr_t start, ram_addr_t length)
 627 {
 628     ram_addr_t addr;
 629     unsigned long page = BIT_WORD(start >> TARGET_PAGE_BITS);
 630
 631     /* start address is aligned at the start of a word? */
 632     if (((page * BITS_PER_LONG) << TARGET_PAGE_BITS) == start) {
 633         int k;
 634         int nr = BITS_TO_LONGS(length >> TARGET_PAGE_BITS);
 635         unsigned long *src = ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION];
 636
 637         for (k = page; k < page + nr; k++) {
 638             if (src[k]) {
 639                 unsigned long new_dirty;
 640                 new_dirty = ~migration_bitmap[k];
 641                 migration_bitmap[k] |= src[k];
 642                 new_dirty &= src[k];
 643                 migration_dirty_pages += ctpopl(new_dirty);
 644                 src[k] = 0;
 645             }
 646         }
 647     } else {
 648         for (addr = 0; addr < length; addr += TARGET_PAGE_SIZE) {
 649             if (cpu_physical_memory_get_dirty(start + addr,
 650                                               TARGET_PAGE_SIZE,
 651                                               DIRTY_MEMORY_MIGRATION)) {
 652                 cpu_physical_memory_reset_dirty(start + addr,
 653                                                 TARGET_PAGE_SIZE,
 654                                                 DIRTY_MEMORY_MIGRATION);
 655                 migration_bitmap_set_dirty(start + addr);
 656             }
 657         }
 658     }
 659 }
 660
 661
 662 /* Fix me: there are too many global variables used in migration process. */
 663 static int64_t start_time;
 664 static int64_t bytes_xfer_prev;
 665 static int64_t num_dirty_pages_period;
 666 static uint64_t xbzrle_cache_miss_prev;
 667 static uint64_t iterations_prev;
 668
 669 static void migration_bitmap_sync_init(void)
 670 {
 671     start_time = 0;
 672     bytes_xfer_prev = 0;
 673     num_dirty_pages_period = 0;
 674     xbzrle_cache_miss_prev = 0;
 675     iterations_prev = 0;
 676 }
 677
 678 /* Called with iothread lock held, to protect ram_list.dirty_memory[] */
 679 static void migration_bitmap_sync(void)
 680 {
 681     RAMBlock *block;
 682     uint64_t num_dirty_pages_init = migration_dirty_pages;
 683     MigrationState *s = migrate_get_current();
 684     int64_t end_time;
 685     int64_t bytes_xfer_now;
 686
 687     bitmap_sync_count++;
 688
 689     if (!bytes_xfer_prev) {
 690         bytes_xfer_prev = ram_bytes_transferred();
 691     }
 692
 693     if (!start_time) {
 694         start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 695     }
 696
 697     trace_migration_bitmap_sync_start();
 698     address_space_sync_dirty_bitmap(&address_space_memory);
 699
 700     rcu_read_lock();
 701     QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
 702         migration_bitmap_sync_range(block->mr->ram_addr, block->used_length);
 703     }
 704     rcu_read_unlock();
 705
 706     trace_migration_bitmap_sync_end(migration_dirty_pages
 707                                     - num_dirty_pages_init);
 708     num_dirty_pages_period += migration_dirty_pages - num_dirty_pages_init;
 709     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 710
 711     /* more than 1 second = 1000 millisecons */
 712     if (end_time > start_time + 1000) {
 713         if (migrate_auto_converge()) {
 714             /* The following detection logic can be refined later. For now:
 715                Check to see if the dirtied bytes is 50% more than the approx.
 716                amount of bytes that just got transferred since the last time we
 717                were in this routine. If that happens >N times (for now N==4)
 718                we turn on the throttle down logic */
 719             bytes_xfer_now = ram_bytes_transferred();
 720             if (s->dirty_pages_rate &&
 721                (num_dirty_pages_period * TARGET_PAGE_SIZE >
 722                    (bytes_xfer_now - bytes_xfer_prev)/2) &&
 723                (dirty_rate_high_cnt++ > 4)) {
 724                     trace_migration_throttle();
 725                     mig_throttle_on = true;
 726                     dirty_rate_high_cnt = 0;
 727              }
 728              bytes_xfer_prev = bytes_xfer_now;
 729         } else {
 730              mig_throttle_on = false;
 731         }
 732         if (migrate_use_xbzrle()) {
 733             if (iterations_prev != acct_info.iterations) {
 734                 acct_info.xbzrle_cache_miss_rate =
 735                    (double)(acct_info.xbzrle_cache_miss -
 736                             xbzrle_cache_miss_prev) /
 737                    (acct_info.iterations - iterations_prev);
 738             }
 739             iterations_prev = acct_info.iterations;
 740             xbzrle_cache_miss_prev = acct_info.xbzrle_cache_miss;
 741         }
 742         s->dirty_pages_rate = num_dirty_pages_period * 1000
 743             / (end_time - start_time);
 744         s->dirty_bytes_rate = s->dirty_pages_rate * TARGET_PAGE_SIZE;
 745         start_time = end_time;
 746         num_dirty_pages_period = 0;
 747     }
 748     s->dirty_sync_count = bitmap_sync_count;
 749 }
 750
 751 /**
 752  * save_zero_page: Send the zero page to the stream
 753  *
 754  * Returns: Number of pages written.
 755  *
 756  * @f: QEMUFile where to send the data
 757  * @block: block that contains the page we want to send
 758  * @offset: offset inside the block for the page
 759  * @p: pointer to the page
 760  * @bytes_transferred: increase it with the number of transferred bytes
 761  */
 762 static int save_zero_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset,
 763                           uint8_t *p, uint64_t *bytes_transferred)
 764 {
 765     int pages = -1;
 766
 767     if (is_zero_range(p, TARGET_PAGE_SIZE)) {
 768         acct_info.dup_pages++;
 769         *bytes_transferred += save_page_header(f, block,
 770                                                offset | RAM_SAVE_FLAG_COMPRESS);
 771         qemu_put_byte(f, 0);
 772         *bytes_transferred += 1;
 773         pages = 1;
 774     }
 775
 776     return pages;
 777 }
 778
 779 /**
 780  * ram_save_page: Send the given page to the stream
 781  *
 782  * Returns: Number of pages written.
 783  *
 784  * @f: QEMUFile where to send the data
 785  * @block: block that contains the page we want to send
 786  * @offset: offset inside the block for the page
 787  * @last_stage: if we are at the completion stage
 788  * @bytes_transferred: increase it with the number of transferred bytes
 789  */
 790 static int ram_save_page(QEMUFile *f, RAMBlock* block, ram_addr_t offset,
 791                          bool last_stage, uint64_t *bytes_transferred)
 792 {
 793     int pages = -1;
 794     uint64_t bytes_xmit;
 795     ram_addr_t current_addr;
 796     MemoryRegion *mr = block->mr;
 797     uint8_t *p;
 798     int ret;
 799     bool send_async = true;
 800
 801     p = memory_region_get_ram_ptr(mr) + offset;
 802
 803     /* In doubt sent page as normal */
 804     bytes_xmit = 0;
 805     ret = ram_control_save_page(f, block->offset,
 806                            offset, TARGET_PAGE_SIZE, &bytes_xmit);
 807     if (bytes_xmit) {
 808         *bytes_transferred += bytes_xmit;
 809         pages = 1;
 810     }
 811
 812     XBZRLE_cache_lock();
 813
 814     current_addr = block->offset + offset;
 815
 816     if (block == last_sent_block) {
 817         offset |= RAM_SAVE_FLAG_CONTINUE;
 818     }
 819     if (ret != RAM_SAVE_CONTROL_NOT_SUPP) {
 820         if (ret != RAM_SAVE_CONTROL_DELAYED) {
 821             if (bytes_xmit > 0) {
 822                 acct_info.norm_pages++;
 823             } else if (bytes_xmit == 0) {
 824                 acct_info.dup_pages++;
 825             }
 826         }
 827     } else {
 828         pages = save_zero_page(f, block, offset, p, bytes_transferred);
 829         if (pages > 0) {
 830             /* Must let xbzrle know, otherwise a previous (now 0'd) cached
 831              * page would be stale
 832              */
 833             xbzrle_cache_zero_page(current_addr);
 834         } else if (!ram_bulk_stage && migrate_use_xbzrle()) {
 835             pages = save_xbzrle_page(f, &p, current_addr, block,
 836                                      offset, last_stage, bytes_transferred);
 837             if (!last_stage) {
 838                 /* Can't send this cached data async, since the cache page
 839                  * might get updated before it gets to the wire
 840                  */
 841                 send_async = false;
 842             }
 843         }
 844     }
 845
 846     /* XBZRLE overflow or normal page */
 847     if (pages == -1) {
 848         *bytes_transferred += save_page_header(f, block,
 849                                                offset | RAM_SAVE_FLAG_PAGE);
 850         if (send_async) {
 851             qemu_put_buffer_async(f, p, TARGET_PAGE_SIZE);
 852         } else {
 853             qemu_put_buffer(f, p, TARGET_PAGE_SIZE);
 854         }
 855         *bytes_transferred += TARGET_PAGE_SIZE;
 856         pages = 1;
 857         acct_info.norm_pages++;
 858     }
 859
 860     XBZRLE_cache_unlock();
 861
 862     return pages;
 863 }
 864
 865 static int do_compress_ram_page(CompressParam *param)
 866 {
 867     int bytes_sent, blen;
 868     uint8_t *p;
 869     RAMBlock *block = param->block;
 870     ram_addr_t offset = param->offset;
 871
 872     p = memory_region_get_ram_ptr(block->mr) + (offset & TARGET_PAGE_MASK);
 873
 874     bytes_sent = save_page_header(param->file, block, offset |
 875                                   RAM_SAVE_FLAG_COMPRESS_PAGE);
 876     blen = qemu_put_compression_data(param->file, p, TARGET_PAGE_SIZE,
 877                                      migrate_compress_level());
 878     bytes_sent += blen;
 879
 880     return bytes_sent;
 881 }
 882
 883 static inline void start_compression(CompressParam *param)
 884 {
 885     param->done = false;
 886     qemu_mutex_lock(&param->mutex);
 887     param->start = true;
 888     qemu_cond_signal(&param->cond);
 889     qemu_mutex_unlock(&param->mutex);
 890 }
 891
 892 static inline void start_decompression(DecompressParam *param)
 893 {
 894     qemu_mutex_lock(&param->mutex);
 895     param->start = true;
 896     qemu_cond_signal(&param->cond);
 897     qemu_mutex_unlock(&param->mutex);
 898 }
 899
 900 static uint64_t bytes_transferred;
 901
 902 static void flush_compressed_data(QEMUFile *f)
 903 {
 904     int idx, len, thread_count;
 905
 906     if (!migrate_use_compression()) {
 907         return;
 908     }
 909     thread_count = migrate_compress_threads();
 910     for (idx = 0; idx < thread_count; idx++) {
 911         if (!comp_param[idx].done) {
 912             qemu_mutex_lock(comp_done_lock);
 913             while (!comp_param[idx].done && !quit_comp_thread) {
 914                 qemu_cond_wait(comp_done_cond, comp_done_lock);
 915             }
 916             qemu_mutex_unlock(comp_done_lock);
 917         }
 918         if (!quit_comp_thread) {
 919             len = qemu_put_qemu_file(f, comp_param[idx].file);
 920             bytes_transferred += len;
 921         }
 922     }
 923 }
 924
 925 static inline void set_compress_params(CompressParam *param, RAMBlock *block,
 926                                        ram_addr_t offset)
 927 {
 928     param->block = block;
 929     param->offset = offset;
 930 }
 931
 932 static int compress_page_with_multi_thread(QEMUFile *f, RAMBlock *block,
 933                                            ram_addr_t offset,
 934                                            uint64_t *bytes_transferred)
 935 {
 936     int idx, thread_count, bytes_xmit = -1, pages = -1;
 937
 938     thread_count = migrate_compress_threads();
 939     qemu_mutex_lock(comp_done_lock);
 940     while (true) {
 941         for (idx = 0; idx < thread_count; idx++) {
 942             if (comp_param[idx].done) {
 943                 bytes_xmit = qemu_put_qemu_file(f, comp_param[idx].file);
 944                 set_compress_params(&comp_param[idx], block, offset);
 945                 start_compression(&comp_param[idx]);
 946                 pages = 1;
 947                 acct_info.norm_pages++;
 948                 *bytes_transferred += bytes_xmit;
 949                 break;
 950             }
 951         }
 952         if (pages > 0) {
 953             break;
 954         } else {
 955             qemu_cond_wait(comp_done_cond, comp_done_lock);
 956         }
 957     }
 958     qemu_mutex_unlock(comp_done_lock);
 959
 960     return pages;
 961 }
 962
 963 /**
 964  * ram_save_compressed_page: compress the given page and send it to the stream
 965  *
 966  * Returns: Number of pages written.
 967  *
 968  * @f: QEMUFile where to send the data
 969  * @block: block that contains the page we want to send
 970  * @offset: offset inside the block for the page
 971  * @last_stage: if we are at the completion stage
 972  * @bytes_transferred: increase it with the number of transferred bytes
 973  */
 974 static int ram_save_compressed_page(QEMUFile *f, RAMBlock *block,
 975                                     ram_addr_t offset, bool last_stage,
 976                                     uint64_t *bytes_transferred)
 977 {
 978     int pages = -1;
 979     uint64_t bytes_xmit;
 980     MemoryRegion *mr = block->mr;
 981     uint8_t *p;
 982     int ret;
 983
 984     p = memory_region_get_ram_ptr(mr) + offset;
 985
 986     bytes_xmit = 0;
 987     ret = ram_control_save_page(f, block->offset,
 988                                 offset, TARGET_PAGE_SIZE, &bytes_xmit);
 989     if (bytes_xmit) {
 990         *bytes_transferred += bytes_xmit;
 991         pages = 1;
 992     }
 993     if (block == last_sent_block) {
 994         offset |= RAM_SAVE_FLAG_CONTINUE;
 995     }
 996     if (ret != RAM_SAVE_CONTROL_NOT_SUPP) {
 997         if (ret != RAM_SAVE_CONTROL_DELAYED) {
 998             if (bytes_xmit > 0) {
 999                 acct_info.norm_pages++;
1000             } else if (bytes_xmit == 0) {
1001                 acct_info.dup_pages++;
1002             }
1003         }
1004     } else {
1005         /* When starting the process of a new block, the first page of
1006          * the block should be sent out before other pages in the same
1007          * block, and all the pages in last block should have been sent
1008          * out, keeping this order is important, because the 'cont' flag
1009          * is used to avoid resending the block name.
1010          */
1011         if (block != last_sent_block) {
1012             flush_compressed_data(f);
1013             pages = save_zero_page(f, block, offset, p, bytes_transferred);
1014             if (pages == -1) {
1015                 set_compress_params(&comp_param[0], block, offset);
1016                 /* Use the qemu thread to compress the data to make sure the
1017                  * first page is sent out before other pages
1018                  */
1019                 bytes_xmit = do_compress_ram_page(&comp_param[0]);
1020                 acct_info.norm_pages++;
1021                 qemu_put_qemu_file(f, comp_param[0].file);
1022                 *bytes_transferred += bytes_xmit;
1023                 pages = 1;
1024             }
1025         } else {
1026             pages = save_zero_page(f, block, offset, p, bytes_transferred);
1027             if (pages == -1) {
1028                 pages = compress_page_with_multi_thread(f, block, offset,
1029                                                         bytes_transferred);
1030             }
1031         }
1032     }
1033
1034     return pages;
1035 }
1036
1037 /**
1038  * ram_find_and_save_block: Finds a dirty page and sends it to f
1039  *
1040  * Called within an RCU critical section.
1041  *
1042  * Returns:  The number of pages written
1043  *           0 means no dirty pages
1044  *
1045  * @f: QEMUFile where to send the data
1046  * @last_stage: if we are at the completion stage
1047  * @bytes_transferred: increase it with the number of transferred bytes
1048  */
1049
1050 static int ram_find_and_save_block(QEMUFile *f, bool last_stage,
1051                                    uint64_t *bytes_transferred)
1052 {
1053     RAMBlock *block = last_seen_block;
1054     ram_addr_t offset = last_offset;
1055     bool complete_round = false;
1056     int pages = 0;
1057     MemoryRegion *mr;
1058
1059     if (!block)
1060         block = QLIST_FIRST_RCU(&ram_list.blocks);
1061
1062     while (true) {
1063         mr = block->mr;
1064         offset = migration_bitmap_find_and_reset_dirty(mr, offset);
1065         if (complete_round && block == last_seen_block &&
1066             offset >= last_offset) {
1067             break;
1068         }
1069         if (offset >= block->used_length) {
1070             offset = 0;
1071             block = QLIST_NEXT_RCU(block, next);
1072             if (!block) {
1073                 block = QLIST_FIRST_RCU(&ram_list.blocks);
1074                 complete_round = true;
1075                 ram_bulk_stage = false;
1076                 if (migrate_use_xbzrle()) {
1077                     /* If xbzrle is on, stop using the data compression at this
1078                      * point. In theory, xbzrle can do better than compression.
1079                      */
1080                     flush_compressed_data(f);
1081                     compression_switch = false;
1082                 }
1083             }
1084         } else {
1085             if (compression_switch && migrate_use_compression()) {
1086                 pages = ram_save_compressed_page(f, block, offset, last_stage,
1087                                                  bytes_transferred);
1088             } else {
1089                 pages = ram_save_page(f, block, offset, last_stage,
1090                                       bytes_transferred);
1091             }
1092
1093             /* if page is unmodified, continue to the next */
1094             if (pages > 0) {
1095                 last_sent_block = block;
1096                 break;
1097             }
1098         }
1099     }
1100
1101     last_seen_block = block;
1102     last_offset = offset;
1103
1104     return pages;
1105 }
1106
1107 void acct_update_position(QEMUFile *f, size_t size, bool zero)
1108 {
1109     uint64_t pages = size / TARGET_PAGE_SIZE;
1110     if (zero) {
1111         acct_info.dup_pages += pages;
1112     } else {
1113         acct_info.norm_pages += pages;
1114         bytes_transferred += size;
1115         qemu_update_position(f, size);
1116     }
1117 }
1118
1119 static ram_addr_t ram_save_remaining(void)
1120 {
1121     return migration_dirty_pages;
1122 }
1123
1124 uint64_t ram_bytes_remaining(void)
1125 {
1126     return ram_save_remaining() * TARGET_PAGE_SIZE;
1127 }
1128
1129 uint64_t ram_bytes_transferred(void)
1130 {
1131     return bytes_transferred;
1132 }
1133
1134 uint64_t ram_bytes_total(void)
1135 {
1136     RAMBlock *block;
1137     uint64_t total = 0;
1138
1139     rcu_read_lock();
1140     QLIST_FOREACH_RCU(block, &ram_list.blocks, next)
1141         total += block->used_length;
1142     rcu_read_unlock();
1143     return total;
1144 }
1145
1146 void free_xbzrle_decoded_buf(void)
1147 {
1148     g_free(xbzrle_decoded_buf);
1149     xbzrle_decoded_buf = NULL;
1150 }
1151
1152 static void migration_end(void)
1153 {
1154     if (migration_bitmap) {
1155         memory_global_dirty_log_stop();
1156         g_free(migration_bitmap);
1157         migration_bitmap = NULL;
1158     }
1159
1160     XBZRLE_cache_lock();
1161     if (XBZRLE.cache) {
1162         cache_fini(XBZRLE.cache);
1163         g_free(XBZRLE.encoded_buf);
1164         g_free(XBZRLE.current_buf);
1165         XBZRLE.cache = NULL;
1166         XBZRLE.encoded_buf = NULL;
1167         XBZRLE.current_buf = NULL;
1168     }
1169     XBZRLE_cache_unlock();
1170 }
1171
1172 static void ram_migration_cancel(void *opaque)
1173 {
1174     migration_end();
1175 }
1176
1177 static void reset_ram_globals(void)
1178 {
1179     last_seen_block = NULL;
1180     last_sent_block = NULL;
1181     last_offset = 0;
1182     last_version = ram_list.version;
1183     ram_bulk_stage = true;
1184 }
1185
1186 #define MAX_WAIT 50 /* ms, half buffered_file limit */
1187
1188
1189 /* Each of ram_save_setup, ram_save_iterate and ram_save_complete has
1190  * long-running RCU critical section.  When rcu-reclaims in the code
1191  * start to become numerous it will be necessary to reduce the
1192  * granularity of these critical sections.
1193  */
1194
1195 static int ram_save_setup(QEMUFile *f, void *opaque)
1196 {
1197     RAMBlock *block;
1198     int64_t ram_bitmap_pages; /* Size of bitmap in pages, including gaps */
1199
1200     mig_throttle_on = false;
1201     dirty_rate_high_cnt = 0;
1202     bitmap_sync_count = 0;
1203     migration_bitmap_sync_init();
1204
1205     if (migrate_use_xbzrle()) {
1206         XBZRLE_cache_lock();
1207         XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() /
1208                                   TARGET_PAGE_SIZE,
1209                                   TARGET_PAGE_SIZE);
1210         if (!XBZRLE.cache) {
1211             XBZRLE_cache_unlock();
1212             error_report("Error creating cache");
1213             return -1;
1214         }
1215         XBZRLE_cache_unlock();
1216
1217         /* We prefer not to abort if there is no memory */
1218         XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE);
1219         if (!XBZRLE.encoded_buf) {
1220             error_report("Error allocating encoded_buf");
1221             return -1;
1222         }
1223
1224         XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE);
1225         if (!XBZRLE.current_buf) {
1226             error_report("Error allocating current_buf");
1227             g_free(XBZRLE.encoded_buf);
1228             XBZRLE.encoded_buf = NULL;
1229             return -1;
1230         }
1231
1232         acct_clear();
1233     }
1234
1235     /* iothread lock needed for ram_list.dirty_memory[] */
1236     qemu_mutex_lock_iothread();
1237     qemu_mutex_lock_ramlist();
1238     rcu_read_lock();
1239     bytes_transferred = 0;
1240     reset_ram_globals();
1241
1242     ram_bitmap_pages = last_ram_offset() >> TARGET_PAGE_BITS;
1243     migration_bitmap = bitmap_new(ram_bitmap_pages);
1244     bitmap_set(migration_bitmap, 0, ram_bitmap_pages);
1245
1246     /*
1247      * Count the total number of pages used by ram blocks not including any
1248      * gaps due to alignment or unplugs.
1249      */
1250     migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS;
1251
1252     memory_global_dirty_log_start();
1253     migration_bitmap_sync();
1254     qemu_mutex_unlock_ramlist();
1255     qemu_mutex_unlock_iothread();
1256
1257     qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE);
1258
1259     QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
1260         qemu_put_byte(f, strlen(block->idstr));
1261         qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr));
1262         qemu_put_be64(f, block->used_length);
1263     }
1264
1265     rcu_read_unlock();
1266
1267     ram_control_before_iterate(f, RAM_CONTROL_SETUP);
1268     ram_control_after_iterate(f, RAM_CONTROL_SETUP);
1269
1270     qemu_put_be64(f, RAM_SAVE_FLAG_EOS);
1271
1272     return 0;
1273 }
1274
1275 static int ram_save_iterate(QEMUFile *f, void *opaque)
1276 {
1277     int ret;
1278     int i;
1279     int64_t t0;
1280     int pages_sent = 0;
1281
1282     rcu_read_lock();
1283     if (ram_list.version != last_version) {
1284         reset_ram_globals();
1285     }
1286
1287     /* Read version before ram_list.blocks */
1288     smp_rmb();
1289
1290     ram_control_before_iterate(f, RAM_CONTROL_ROUND);
1291
1292     t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1293     i = 0;
1294     while ((ret = qemu_file_rate_limit(f)) == 0) {
1295         int pages;
1296
1297         pages = ram_find_and_save_block(f, false, &bytes_transferred);
1298         /* no more pages to sent */
1299         if (pages == 0) {
1300             break;
1301         }
1302         pages_sent += pages;
1303         acct_info.iterations++;
1304         check_guest_throttling();
1305         /* we want to check in the 1st loop, just in case it was the 1st time
1306            and we had to sync the dirty bitmap.
1307            qemu_get_clock_ns() is a bit expensive, so we only check each some
1308            iterations
1309         */
1310         if ((i & 63) == 0) {
1311             uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000;
1312             if (t1 > MAX_WAIT) {
1313                 DPRINTF("big wait: %" PRIu64 " milliseconds, %d iterations\n",
1314                         t1, i);
1315                 break;
1316             }
1317         }
1318         i++;
1319     }
1320     flush_compressed_data(f);
1321     rcu_read_unlock();
1322
1323     /*
1324      * Must occur before EOS (or any QEMUFile operation)
1325      * because of RDMA protocol.
1326      */
1327     ram_control_after_iterate(f, RAM_CONTROL_ROUND);
1328
1329     qemu_put_be64(f, RAM_SAVE_FLAG_EOS);
1330     bytes_transferred += 8;
1331
1332     ret = qemu_file_get_error(f);
1333     if (ret < 0) {
1334         return ret;
1335     }
1336
1337     return pages_sent;
1338 }
1339
1340 /* Called with iothread lock */
1341 static int ram_save_complete(QEMUFile *f, void *opaque)
1342 {
1343     rcu_read_lock();
1344
1345     migration_bitmap_sync();
1346
1347     ram_control_before_iterate(f, RAM_CONTROL_FINISH);
1348
1349     /* try transferring iterative blocks of memory */
1350
1351     /* flush all remaining blocks regardless of rate limiting */
1352     while (true) {
1353         int pages;
1354
1355         pages = ram_find_and_save_block(f, true, &bytes_transferred);
1356         /* no more blocks to sent */
1357         if (pages == 0) {
1358             break;
1359         }
1360     }
1361
1362     flush_compressed_data(f);
1363     ram_control_after_iterate(f, RAM_CONTROL_FINISH);
1364     migration_end();
1365
1366     rcu_read_unlock();
1367     qemu_put_be64(f, RAM_SAVE_FLAG_EOS);
1368
1369     return 0;
1370 }
1371
1372 static uint64_t ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size)
1373 {
1374     uint64_t remaining_size;
1375
1376     remaining_size = ram_save_remaining() * TARGET_PAGE_SIZE;
1377
1378     if (remaining_size < max_size) {
1379         qemu_mutex_lock_iothread();
1380         rcu_read_lock();
1381         migration_bitmap_sync();
1382         rcu_read_unlock();
1383         qemu_mutex_unlock_iothread();
1384         remaining_size = ram_save_remaining() * TARGET_PAGE_SIZE;
1385     }
1386     return remaining_size;
1387 }
1388
1389 static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host)
1390 {
1391     unsigned int xh_len;
1392     int xh_flags;
1393
1394     if (!xbzrle_decoded_buf) {
1395         xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE);
1396     }
1397
1398     /* extract RLE header */
1399     xh_flags = qemu_get_byte(f);
1400     xh_len = qemu_get_be16(f);
1401
1402     if (xh_flags != ENCODING_FLAG_XBZRLE) {
1403         error_report("Failed to load XBZRLE page - wrong compression!");
1404         return -1;
1405     }
1406
1407     if (xh_len > TARGET_PAGE_SIZE) {
1408         error_report("Failed to load XBZRLE page - len overflow!");
1409         return -1;
1410     }
1411     /* load data and decode */
1412     qemu_get_buffer(f, xbzrle_decoded_buf, xh_len);
1413
1414     /* decode RLE */
1415     if (xbzrle_decode_buffer(xbzrle_decoded_buf, xh_len, host,
1416                              TARGET_PAGE_SIZE) == -1) {
1417         error_report("Failed to load XBZRLE page - decode error!");
1418         return -1;
1419     }
1420
1421     return 0;
1422 }
1423
1424 /* Must be called from within a rcu critical section.
1425  * Returns a pointer from within the RCU-protected ram_list.
1426  */
1427 static inline void *host_from_stream_offset(QEMUFile *f,
1428                                             ram_addr_t offset,
1429                                             int flags)
1430 {
1431     static RAMBlock *block = NULL;
1432     char id[256];
1433     uint8_t len;
1434
1435     if (flags & RAM_SAVE_FLAG_CONTINUE) {
1436         if (!block || block->max_length <= offset) {
1437             error_report("Ack, bad migration stream!");
1438             return NULL;
1439         }
1440
1441         return memory_region_get_ram_ptr(block->mr) + offset;
1442     }
1443
1444     len = qemu_get_byte(f);
1445     qemu_get_buffer(f, (uint8_t *)id, len);
1446     id[len] = 0;
1447
1448     QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
1449         if (!strncmp(id, block->idstr, sizeof(id)) &&
1450             block->max_length > offset) {
1451             return memory_region_get_ram_ptr(block->mr) + offset;
1452         }
1453     }
1454
1455     error_report("Can't find block %s!", id);
1456     return NULL;
1457 }
1458
1459 /*
1460  * If a page (or a whole RDMA chunk) has been
1461  * determined to be zero, then zap it.
1462  */
1463 void ram_handle_compressed(void *host, uint8_t ch, uint64_t size)
1464 {
1465     if (ch != 0 || !is_zero_range(host, size)) {
1466         memset(host, ch, size);
1467     }
1468 }
1469
1470 static void *do_data_decompress(void *opaque)
1471 {
1472     DecompressParam *param = opaque;
1473     unsigned long pagesize;
1474
1475     while (!quit_decomp_thread) {
1476         qemu_mutex_lock(&param->mutex);
1477         while (!param->start && !quit_decomp_thread) {
1478             qemu_cond_wait(&param->cond, &param->mutex);
1479             pagesize = TARGET_PAGE_SIZE;
1480             if (!quit_decomp_thread) {
1481                 /* uncompress() will return failed in some case, especially
1482                  * when the page is dirted when doing the compression, it's
1483                  * not a problem because the dirty page will be retransferred
1484                  * and uncompress() won't break the data in other pages.
1485                  */
1486                 uncompress((Bytef *)param->des, &pagesize,
1487                            (const Bytef *)param->compbuf, param->len);
1488             }
1489             param->start = false;
1490         }
1491         qemu_mutex_unlock(&param->mutex);
1492     }
1493
1494     return NULL;
1495 }
1496
1497 void migrate_decompress_threads_create(void)
1498 {
1499     int i, thread_count;
1500
1501     thread_count = migrate_decompress_threads();
1502     decompress_threads = g_new0(QemuThread, thread_count);
1503     decomp_param = g_new0(DecompressParam, thread_count);
1504     compressed_data_buf = g_malloc0(compressBound(TARGET_PAGE_SIZE));
1505     quit_decomp_thread = false;
1506     for (i = 0; i < thread_count; i++) {
1507         qemu_mutex_init(&decomp_param[i].mutex);
1508         qemu_cond_init(&decomp_param[i].cond);
1509         decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE));
1510         qemu_thread_create(decompress_threads + i, "decompress",
1511                            do_data_decompress, decomp_param + i,
1512                            QEMU_THREAD_JOINABLE);
1513     }
1514 }
1515
1516 void migrate_decompress_threads_join(void)
1517 {
1518     int i, thread_count;
1519
1520     quit_decomp_thread = true;
1521     thread_count = migrate_decompress_threads();
1522     for (i = 0; i < thread_count; i++) {
1523         qemu_mutex_lock(&decomp_param[i].mutex);
1524         qemu_cond_signal(&decomp_param[i].cond);
1525         qemu_mutex_unlock(&decomp_param[i].mutex);
1526     }
1527     for (i = 0; i < thread_count; i++) {
1528         qemu_thread_join(decompress_threads + i);
1529         qemu_mutex_destroy(&decomp_param[i].mutex);
1530         qemu_cond_destroy(&decomp_param[i].cond);
1531         g_free(decomp_param[i].compbuf);
1532     }
1533     g_free(decompress_threads);
1534     g_free(decomp_param);
1535     g_free(compressed_data_buf);
1536     decompress_threads = NULL;
1537     decomp_param = NULL;
1538     compressed_data_buf = NULL;
1539 }
1540
1541 static void decompress_data_with_multi_threads(uint8_t *compbuf,
1542                                                void *host, int len)
1543 {
1544     int idx, thread_count;
1545
1546     thread_count = migrate_decompress_threads();
1547     while (true) {
1548         for (idx = 0; idx < thread_count; idx++) {
1549             if (!decomp_param[idx].start) {
1550                 memcpy(decomp_param[idx].compbuf, compbuf, len);
1551                 decomp_param[idx].des = host;
1552                 decomp_param[idx].len = len;
1553                 start_decompression(&decomp_param[idx]);
1554                 break;
1555             }
1556         }
1557         if (idx < thread_count) {
1558             break;
1559         }
1560     }
1561 }
1562
1563 static int ram_load(QEMUFile *f, void *opaque, int version_id)
1564 {
1565     int flags = 0, ret = 0;
1566     static uint64_t seq_iter;
1567     int len = 0;
1568
1569     seq_iter++;
1570
1571     if (version_id != 4) {
1572         ret = -EINVAL;
1573     }
1574
1575     /* This RCU critical section can be very long running.
1576      * When RCU reclaims in the code start to become numerous,
1577      * it will be necessary to reduce the granularity of this
1578      * critical section.
1579      */
1580     rcu_read_lock();
1581     while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) {
1582         ram_addr_t addr, total_ram_bytes;
1583         void *host;
1584         uint8_t ch;
1585
1586         addr = qemu_get_be64(f);
1587         flags = addr & ~TARGET_PAGE_MASK;
1588         addr &= TARGET_PAGE_MASK;
1589
1590         switch (flags & ~RAM_SAVE_FLAG_CONTINUE) {
1591         case RAM_SAVE_FLAG_MEM_SIZE:
1592             /* Synchronize RAM block list */
1593             total_ram_bytes = addr;
1594             while (!ret && total_ram_bytes) {
1595                 RAMBlock *block;
1596                 uint8_t len;
1597                 char id[256];
1598                 ram_addr_t length;
1599
1600                 len = qemu_get_byte(f);
1601                 qemu_get_buffer(f, (uint8_t *)id, len);
1602                 id[len] = 0;
1603                 length = qemu_get_be64(f);
1604
1605                 QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
1606                     if (!strncmp(id, block->idstr, sizeof(id))) {
1607                         if (length != block->used_length) {
1608                             Error *local_err = NULL;
1609
1610                             ret = qemu_ram_resize(block->offset, length, &local_err);
1611                             if (local_err) {
1612                                 error_report_err(local_err);
1613                             }
1614                         }
1615                         break;
1616                     }
1617                 }
1618
1619                 if (!block) {
1620                     error_report("Unknown ramblock \"%s\", cannot "
1621                                  "accept migration", id);
1622                     ret = -EINVAL;
1623                 }
1624
1625                 total_ram_bytes -= length;
1626             }
1627             break;
1628         case RAM_SAVE_FLAG_COMPRESS:
1629             host = host_from_stream_offset(f, addr, flags);
1630             if (!host) {
1631                 error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
1632                 ret = -EINVAL;
1633                 break;
1634             }
1635             ch = qemu_get_byte(f);
1636             ram_handle_compressed(host, ch, TARGET_PAGE_SIZE);
1637             break;
1638         case RAM_SAVE_FLAG_PAGE:
1639             host = host_from_stream_offset(f, addr, flags);
1640             if (!host) {
1641                 error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
1642                 ret = -EINVAL;
1643                 break;
1644             }
1645             qemu_get_buffer(f, host, TARGET_PAGE_SIZE);
1646             break;
1647         case RAM_SAVE_FLAG_COMPRESS_PAGE:
1648             host = host_from_stream_offset(f, addr, flags);
1649             if (!host) {
1650                 error_report("Invalid RAM offset " RAM_ADDR_FMT, addr);
1651                 ret = -EINVAL;
1652                 break;
1653             }
1654
1655             len = qemu_get_be32(f);
1656             if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) {
1657                 error_report("Invalid compressed data length: %d", len);
1658                 ret = -EINVAL;
1659                 break;
1660             }
1661             qemu_get_buffer(f, compressed_data_buf, len);
1662             decompress_data_with_multi_threads(compressed_data_buf, host, len);
1663             break;
1664         case RAM_SAVE_FLAG_XBZRLE:
1665             host = host_from_stream_offset(f, addr, flags);
1666             if (!host) {
1667                 error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
1668                 ret = -EINVAL;
1669                 break;
1670             }
1671             if (load_xbzrle(f, addr, host) < 0) {
1672                 error_report("Failed to decompress XBZRLE page at "
1673                              RAM_ADDR_FMT, addr);
1674                 ret = -EINVAL;
1675                 break;
1676             }
1677             break;
1678         case RAM_SAVE_FLAG_EOS:
1679             /* normal exit */
1680             break;
1681         default:
1682             if (flags & RAM_SAVE_FLAG_HOOK) {
1683                 ram_control_load_hook(f, flags);
1684             } else {
1685                 error_report("Unknown combination of migration flags: %#x",
1686                              flags);
1687                 ret = -EINVAL;
1688             }
1689         }
1690         if (!ret) {
1691             ret = qemu_file_get_error(f);
1692         }
1693     }
1694
1695     rcu_read_unlock();
1696     DPRINTF("Completed load of VM with exit code %d seq iteration "
1697             "%" PRIu64 "\n", ret, seq_iter);
1698     return ret;
1699 }
1700
1701 static SaveVMHandlers savevm_ram_handlers = {
1702     .save_live_setup = ram_save_setup,
1703     .save_live_iterate = ram_save_iterate,
1704     .save_live_complete = ram_save_complete,
1705     .save_live_pending = ram_save_pending,
1706     .load_state = ram_load,
1707     .cancel = ram_migration_cancel,
1708 };
1709
1710 void ram_mig_init(void)
1711 {
1712     qemu_mutex_init(&XBZRLE.lock);
1713     register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, NULL);
1714 }
1715
1716 struct soundhw {
1717     const char *name;
1718     const char *descr;
1719     int enabled;
1720     int isa;
1721     union {
1722         int (*init_isa) (ISABus *bus);
1723         int (*init_pci) (PCIBus *bus);
1724     } init;
1725 };
1726
1727 static struct soundhw soundhw[9];
1728 static int soundhw_count;
1729
1730 void isa_register_soundhw(const char *name, const char *descr,
1731                           int (*init_isa)(ISABus *bus))
1732 {
1733     assert(soundhw_count < ARRAY_SIZE(soundhw) - 1);
1734     soundhw[soundhw_count].name = name;
1735     soundhw[soundhw_count].descr = descr;
1736     soundhw[soundhw_count].isa = 1;
1737     soundhw[soundhw_count].init.init_isa = init_isa;
1738     soundhw_count++;
1739 }
1740
1741 void pci_register_soundhw(const char *name, const char *descr,
1742                           int (*init_pci)(PCIBus *bus))
1743 {
1744     assert(soundhw_count < ARRAY_SIZE(soundhw) - 1);
1745     soundhw[soundhw_count].name = name;
1746     soundhw[soundhw_count].descr = descr;
1747     soundhw[soundhw_count].isa = 0;
1748     soundhw[soundhw_count].init.init_pci = init_pci;
1749     soundhw_count++;
1750 }
1751
1752 void select_soundhw(const char *optarg)
1753 {
1754     struct soundhw *c;
1755
1756     if (is_help_option(optarg)) {
1757     show_valid_cards:
1758
1759         if (soundhw_count) {
1760              printf("Valid sound card names (comma separated):\n");
1761              for (c = soundhw; c->name; ++c) {
1762                  printf ("%-11s %s\n", c->name, c->descr);
1763              }
1764              printf("\n-soundhw all will enable all of the above\n");
1765         } else {
1766              printf("Machine has no user-selectable audio hardware "
1767                     "(it may or may not have always-present audio hardware).\n");
1768         }
1769         exit(!is_help_option(optarg));
1770     }
1771     else {
1772         size_t l;
1773         const char *p;
1774         char *e;
1775         int bad_card = 0;
1776
1777         if (!strcmp(optarg, "all")) {
1778             for (c = soundhw; c->name; ++c) {
1779                 c->enabled = 1;
1780             }
1781             return;
1782         }
1783
1784         p = optarg;
1785         while (*p) {
1786             e = strchr(p, ',');
1787             l = !e ? strlen(p) : (size_t) (e - p);
1788
1789             for (c = soundhw; c->name; ++c) {
1790                 if (!strncmp(c->name, p, l) && !c->name[l]) {
1791                     c->enabled = 1;
1792                     break;
1793                 }
1794             }
1795
1796             if (!c->name) {
1797                 if (l > 80) {
1798                     error_report("Unknown sound card name (too big to show)");
1799                 }
1800                 else {
1801                     error_report("Unknown sound card name `%.*s'",
1802                                  (int) l, p);
1803                 }
1804                 bad_card = 1;
1805             }
1806             p += l + (e != NULL);
1807         }
1808
1809         if (bad_card) {
1810             goto show_valid_cards;
1811         }
1812     }
1813 }
1814
1815 void audio_init(void)
1816 {
1817     struct soundhw *c;
1818     ISABus *isa_bus = (ISABus *) object_resolve_path_type("", TYPE_ISA_BUS, NULL);
1819     PCIBus *pci_bus = (PCIBus *) object_resolve_path_type("", TYPE_PCI_BUS, NULL);
1820
1821     for (c = soundhw; c->name; ++c) {
1822         if (c->enabled) {
1823             if (c->isa) {
1824                 if (!isa_bus) {
1825                     error_report("ISA bus not available for %s", c->name);
1826                     exit(1);
1827                 }
1828                 c->init.init_isa(isa_bus);
1829             } else {
1830                 if (!pci_bus) {
1831                     error_report("PCI bus not available for %s", c->name);
1832                     exit(1);
1833                 }
1834                 c->init.init_pci(pci_bus);
1835             }
1836         }
1837     }
1838 }
1839
1840 int qemu_uuid_parse(const char *str, uint8_t *uuid)
1841 {
1842     int ret;
1843
1844     if (strlen(str) != 36) {
1845         return -1;
1846     }
1847
1848     ret = sscanf(str, UUID_FMT, &uuid[0], &uuid[1], &uuid[2], &uuid[3],
1849                  &uuid[4], &uuid[5], &uuid[6], &uuid[7], &uuid[8], &uuid[9],
1850                  &uuid[10], &uuid[11], &uuid[12], &uuid[13], &uuid[14],
1851                  &uuid[15]);
1852
1853     if (ret != 16) {
1854         return -1;
1855     }
1856     return 0;
1857 }
1858
1859 void do_acpitable_option(const QemuOpts *opts)
1860 {
1861 #ifdef TARGET_I386
1862     Error *err = NULL;
1863
1864     acpi_table_add(opts, &err);
1865     if (err) {
1866         error_report("Wrong acpi table provided: %s",
1867                      error_get_pretty(err));
1868         error_free(err);
1869         exit(1);
1870     }
1871 #endif
1872 }
1873
1874 void do_smbios_option(QemuOpts *opts)
1875 {
1876 #ifdef TARGET_I386
1877     smbios_entry_add(opts);
1878 #endif
1879 }
1880
1881 void cpudef_init(void)
1882 {
1883 #if defined(cpudef_setup)
1884     cpudef_setup(); /* parse cpu definitions in target config file */
1885 #endif
1886 }
1887
1888 int kvm_available(void)
1889 {
1890 #ifdef CONFIG_KVM
1891     return 1;
1892 #else
1893     return 0;
1894 #endif
1895 }
1896
1897 int xen_available(void)
1898 {
1899 #ifdef CONFIG_XEN
1900     return 1;
1901 #else
1902     return 0;
1903 #endif
1904 }
1905
1906
1907 TargetInfo *qmp_query_target(Error **errp)
1908 {
1909     TargetInfo *info = g_malloc0(sizeof(*info));
1910
1911     info->arch = g_strdup(TARGET_NAME);
1912
1913     return info;
1914 }
1915
1916 /* Stub function that's gets run on the vcpu when its brought out of the
1917    VM to run inside qemu via async_run_on_cpu()*/
1918 static void mig_sleep_cpu(void *opq)
1919 {
1920     qemu_mutex_unlock_iothread();
1921     g_usleep(30*1000);
1922     qemu_mutex_lock_iothread();
1923 }
1924
1925 /* To reduce the dirty rate explicitly disallow the VCPUs from spending
1926    much time in the VM. The migration thread will try to catchup.
1927    Workload will experience a performance drop.
1928 */
1929 static void mig_throttle_guest_down(void)
1930 {
1931     CPUState *cpu;
1932
1933     qemu_mutex_lock_iothread();
1934     CPU_FOREACH(cpu) {
1935         async_run_on_cpu(cpu, mig_sleep_cpu, NULL);
1936     }
1937     qemu_mutex_unlock_iothread();
1938 }
1939
1940 static void check_guest_throttling(void)
1941 {
1942     static int64_t t0;
1943     int64_t        t1;
1944
1945     if (!mig_throttle_on) {
1946         return;
1947     }
1948
1949     if (!t0)  {
1950         t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1951         return;
1952     }
1953
1954     t1 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1955
1956     /* If it has been more than 40 ms since the last time the guest
1957      * was throttled then do it again.
1958      */
1959     if (40 < (t1-t0)/1000000) {
1960         mig_throttle_guest_down();
1961         t0 = t1;
1962     }
1963 }