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 static bool quit_comp_thread;
 353 static bool quit_decomp_thread;
 354 static DecompressParam *decomp_param;
 355 static QemuThread *decompress_threads;
 356 static uint8_t *compressed_data_buf;
 357
 358 static int do_compress_ram_page(CompressParam *param);
 359
 360 static void *do_data_compress(void *opaque)
 361 {
 362     CompressParam *param = opaque;
 363
 364     while (!quit_comp_thread) {
 365         qemu_mutex_lock(&param->mutex);
 366         /* Re-check the quit_comp_thread in case of
 367          * terminate_compression_threads is called just before
 368          * qemu_mutex_lock(&param->mutex) and after
 369          * while(!quit_comp_thread), re-check it here can make
 370          * sure the compression thread terminate as expected.
 371          */
 372         while (!param->start && !quit_comp_thread) {
 373             qemu_cond_wait(&param->cond, &param->mutex);
 374         }
 375         if (!quit_comp_thread) {
 376             do_compress_ram_page(param);
 377         }
 378         param->start = false;
 379         qemu_mutex_unlock(&param->mutex);
 380
 381         qemu_mutex_lock(comp_done_lock);
 382         param->done = true;
 383         qemu_cond_signal(comp_done_cond);
 384         qemu_mutex_unlock(comp_done_lock);
 385     }
 386
 387     return NULL;
 388 }
 389
 390 static inline void terminate_compression_threads(void)
 391 {
 392     int idx, thread_count;
 393
 394     thread_count = migrate_compress_threads();
 395     quit_comp_thread = true;
 396     for (idx = 0; idx < thread_count; idx++) {
 397         qemu_mutex_lock(&comp_param[idx].mutex);
 398         qemu_cond_signal(&comp_param[idx].cond);
 399         qemu_mutex_unlock(&comp_param[idx].mutex);
 400     }
 401 }
 402
 403 void migrate_compress_threads_join(void)
 404 {
 405     int i, thread_count;
 406
 407     if (!migrate_use_compression()) {
 408         return;
 409     }
 410     terminate_compression_threads();
 411     thread_count = migrate_compress_threads();
 412     for (i = 0; i < thread_count; i++) {
 413         qemu_thread_join(compress_threads + i);
 414         qemu_fclose(comp_param[i].file);
 415         qemu_mutex_destroy(&comp_param[i].mutex);
 416         qemu_cond_destroy(&comp_param[i].cond);
 417     }
 418     qemu_mutex_destroy(comp_done_lock);
 419     qemu_cond_destroy(comp_done_cond);
 420     g_free(compress_threads);
 421     g_free(comp_param);
 422     g_free(comp_done_cond);
 423     g_free(comp_done_lock);
 424     compress_threads = NULL;
 425     comp_param = NULL;
 426     comp_done_cond = NULL;
 427     comp_done_lock = NULL;
 428 }
 429
 430 void migrate_compress_threads_create(void)
 431 {
 432     int i, thread_count;
 433
 434     if (!migrate_use_compression()) {
 435         return;
 436     }
 437     quit_comp_thread = false;
 438     thread_count = migrate_compress_threads();
 439     compress_threads = g_new0(QemuThread, thread_count);
 440     comp_param = g_new0(CompressParam, thread_count);
 441     comp_done_cond = g_new0(QemuCond, 1);
 442     comp_done_lock = g_new0(QemuMutex, 1);
 443     qemu_cond_init(comp_done_cond);
 444     qemu_mutex_init(comp_done_lock);
 445     for (i = 0; i < thread_count; i++) {
 446         /* com_param[i].file is just used as a dummy buffer to save data, set
 447          * it's ops to empty.
 448          */
 449         comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops);
 450         comp_param[i].done = true;
 451         qemu_mutex_init(&comp_param[i].mutex);
 452         qemu_cond_init(&comp_param[i].cond);
 453         qemu_thread_create(compress_threads + i, "compress",
 454                            do_data_compress, comp_param + i,
 455                            QEMU_THREAD_JOINABLE);
 456     }
 457 }
 458
 459 /**
 460  * save_page_header: Write page header to wire
 461  *
 462  * If this is the 1st block, it also writes the block identification
 463  *
 464  * Returns: Number of bytes written
 465  *
 466  * @f: QEMUFile where to send the data
 467  * @block: block that contains the page we want to send
 468  * @offset: offset inside the block for the page
 469  *          in the lower bits, it contains flags
 470  */
 471 static size_t save_page_header(QEMUFile *f, RAMBlock *block, ram_addr_t offset)
 472 {
 473     size_t size;
 474
 475     qemu_put_be64(f, offset);
 476     size = 8;
 477
 478     if (!(offset & RAM_SAVE_FLAG_CONTINUE)) {
 479         qemu_put_byte(f, strlen(block->idstr));
 480         qemu_put_buffer(f, (uint8_t *)block->idstr,
 481                         strlen(block->idstr));
 482         size += 1 + strlen(block->idstr);
 483     }
 484     return size;
 485 }
 486
 487 /* Update the xbzrle cache to reflect a page that's been sent as all 0.
 488  * The important thing is that a stale (not-yet-0'd) page be replaced
 489  * by the new data.
 490  * As a bonus, if the page wasn't in the cache it gets added so that
 491  * when a small write is made into the 0'd page it gets XBZRLE sent
 492  */
 493 static void xbzrle_cache_zero_page(ram_addr_t current_addr)
 494 {
 495     if (ram_bulk_stage || !migrate_use_xbzrle()) {
 496         return;
 497     }
 498
 499     /* We don't care if this fails to allocate a new cache page
 500      * as long as it updated an old one */
 501     cache_insert(XBZRLE.cache, current_addr, ZERO_TARGET_PAGE,
 502                  bitmap_sync_count);
 503 }
 504
 505 #define ENCODING_FLAG_XBZRLE 0x1
 506
 507 /**
 508  * save_xbzrle_page: compress and send current page
 509  *
 510  * Returns: 1 means that we wrote the page
 511  *          0 means that page is identical to the one already sent
 512  *          -1 means that xbzrle would be longer than normal
 513  *
 514  * @f: QEMUFile where to send the data
 515  * @current_data:
 516  * @current_addr:
 517  * @block: block that contains the page we want to send
 518  * @offset: offset inside the block for the page
 519  * @last_stage: if we are at the completion stage
 520  * @bytes_transferred: increase it with the number of transferred bytes
 521  */
 522 static int save_xbzrle_page(QEMUFile *f, uint8_t **current_data,
 523                             ram_addr_t current_addr, RAMBlock *block,
 524                             ram_addr_t offset, bool last_stage,
 525                             uint64_t *bytes_transferred)
 526 {
 527     int encoded_len = 0, bytes_xbzrle;
 528     uint8_t *prev_cached_page;
 529
 530     if (!cache_is_cached(XBZRLE.cache, current_addr, bitmap_sync_count)) {
 531         acct_info.xbzrle_cache_miss++;
 532         if (!last_stage) {
 533             if (cache_insert(XBZRLE.cache, current_addr, *current_data,
 534                              bitmap_sync_count) == -1) {
 535                 return -1;
 536             } else {
 537                 /* update *current_data when the page has been
 538                    inserted into cache */
 539                 *current_data = get_cached_data(XBZRLE.cache, current_addr);
 540             }
 541         }
 542         return -1;
 543     }
 544
 545     prev_cached_page = get_cached_data(XBZRLE.cache, current_addr);
 546
 547     /* save current buffer into memory */
 548     memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE);
 549
 550     /* XBZRLE encoding (if there is no overflow) */
 551     encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf,
 552                                        TARGET_PAGE_SIZE, XBZRLE.encoded_buf,
 553                                        TARGET_PAGE_SIZE);
 554     if (encoded_len == 0) {
 555         DPRINTF("Skipping unmodified page\n");
 556         return 0;
 557     } else if (encoded_len == -1) {
 558         DPRINTF("Overflow\n");
 559         acct_info.xbzrle_overflows++;
 560         /* update data in the cache */
 561         if (!last_stage) {
 562             memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE);
 563             *current_data = prev_cached_page;
 564         }
 565         return -1;
 566     }
 567
 568     /* we need to update the data in the cache, in order to get the same data */
 569     if (!last_stage) {
 570         memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE);
 571     }
 572
 573     /* Send XBZRLE based compressed page */
 574     bytes_xbzrle = save_page_header(f, block, offset | RAM_SAVE_FLAG_XBZRLE);
 575     qemu_put_byte(f, ENCODING_FLAG_XBZRLE);
 576     qemu_put_be16(f, encoded_len);
 577     qemu_put_buffer(f, XBZRLE.encoded_buf, encoded_len);
 578     bytes_xbzrle += encoded_len + 1 + 2;
 579     acct_info.xbzrle_pages++;
 580     acct_info.xbzrle_bytes += bytes_xbzrle;
 581     *bytes_transferred += bytes_xbzrle;
 582
 583     return 1;
 584 }
 585
 586 static inline
 587 ram_addr_t migration_bitmap_find_and_reset_dirty(MemoryRegion *mr,
 588                                                  ram_addr_t start)
 589 {
 590     unsigned long base = mr->ram_addr >> TARGET_PAGE_BITS;
 591     unsigned long nr = base + (start >> TARGET_PAGE_BITS);
 592     uint64_t mr_size = TARGET_PAGE_ALIGN(memory_region_size(mr));
 593     unsigned long size = base + (mr_size >> TARGET_PAGE_BITS);
 594
 595     unsigned long next;
 596
 597     if (ram_bulk_stage && nr > base) {
 598         next = nr + 1;
 599     } else {
 600         next = find_next_bit(migration_bitmap, size, nr);
 601     }
 602
 603     if (next < size) {
 604         clear_bit(next, migration_bitmap);
 605         migration_dirty_pages--;
 606     }
 607     return (next - base) << TARGET_PAGE_BITS;
 608 }
 609
 610 static inline bool migration_bitmap_set_dirty(ram_addr_t addr)
 611 {
 612     bool ret;
 613     int nr = addr >> TARGET_PAGE_BITS;
 614
 615     ret = test_and_set_bit(nr, migration_bitmap);
 616
 617     if (!ret) {
 618         migration_dirty_pages++;
 619     }
 620     return ret;
 621 }
 622
 623 static void migration_bitmap_sync_range(ram_addr_t start, ram_addr_t length)
 624 {
 625     ram_addr_t addr;
 626     unsigned long page = BIT_WORD(start >> TARGET_PAGE_BITS);
 627
 628     /* start address is aligned at the start of a word? */
 629     if (((page * BITS_PER_LONG) << TARGET_PAGE_BITS) == start) {
 630         int k;
 631         int nr = BITS_TO_LONGS(length >> TARGET_PAGE_BITS);
 632         unsigned long *src = ram_list.dirty_memory[DIRTY_MEMORY_MIGRATION];
 633
 634         for (k = page; k < page + nr; k++) {
 635             if (src[k]) {
 636                 unsigned long new_dirty;
 637                 new_dirty = ~migration_bitmap[k];
 638                 migration_bitmap[k] |= src[k];
 639                 new_dirty &= src[k];
 640                 migration_dirty_pages += ctpopl(new_dirty);
 641                 src[k] = 0;
 642             }
 643         }
 644     } else {
 645         for (addr = 0; addr < length; addr += TARGET_PAGE_SIZE) {
 646             if (cpu_physical_memory_get_dirty(start + addr,
 647                                               TARGET_PAGE_SIZE,
 648                                               DIRTY_MEMORY_MIGRATION)) {
 649                 cpu_physical_memory_reset_dirty(start + addr,
 650                                                 TARGET_PAGE_SIZE,
 651                                                 DIRTY_MEMORY_MIGRATION);
 652                 migration_bitmap_set_dirty(start + addr);
 653             }
 654         }
 655     }
 656 }
 657
 658
 659 /* Fix me: there are too many global variables used in migration process. */
 660 static int64_t start_time;
 661 static int64_t bytes_xfer_prev;
 662 static int64_t num_dirty_pages_period;
 663
 664 static void migration_bitmap_sync_init(void)
 665 {
 666     start_time = 0;
 667     bytes_xfer_prev = 0;
 668     num_dirty_pages_period = 0;
 669 }
 670
 671 /* Called with iothread lock held, to protect ram_list.dirty_memory[] */
 672 static void migration_bitmap_sync(void)
 673 {
 674     RAMBlock *block;
 675     uint64_t num_dirty_pages_init = migration_dirty_pages;
 676     MigrationState *s = migrate_get_current();
 677     int64_t end_time;
 678     int64_t bytes_xfer_now;
 679     static uint64_t xbzrle_cache_miss_prev;
 680     static uint64_t iterations_prev;
 681
 682     bitmap_sync_count++;
 683
 684     if (!bytes_xfer_prev) {
 685         bytes_xfer_prev = ram_bytes_transferred();
 686     }
 687
 688     if (!start_time) {
 689         start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 690     }
 691
 692     trace_migration_bitmap_sync_start();
 693     address_space_sync_dirty_bitmap(&address_space_memory);
 694
 695     rcu_read_lock();
 696     QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
 697         migration_bitmap_sync_range(block->mr->ram_addr, block->used_length);
 698     }
 699     rcu_read_unlock();
 700
 701     trace_migration_bitmap_sync_end(migration_dirty_pages
 702                                     - num_dirty_pages_init);
 703     num_dirty_pages_period += migration_dirty_pages - num_dirty_pages_init;
 704     end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 705
 706     /* more than 1 second = 1000 millisecons */
 707     if (end_time > start_time + 1000) {
 708         if (migrate_auto_converge()) {
 709             /* The following detection logic can be refined later. For now:
 710                Check to see if the dirtied bytes is 50% more than the approx.
 711                amount of bytes that just got transferred since the last time we
 712                were in this routine. If that happens >N times (for now N==4)
 713                we turn on the throttle down logic */
 714             bytes_xfer_now = ram_bytes_transferred();
 715             if (s->dirty_pages_rate &&
 716                (num_dirty_pages_period * TARGET_PAGE_SIZE >
 717                    (bytes_xfer_now - bytes_xfer_prev)/2) &&
 718                (dirty_rate_high_cnt++ > 4)) {
 719                     trace_migration_throttle();
 720                     mig_throttle_on = true;
 721                     dirty_rate_high_cnt = 0;
 722              }
 723              bytes_xfer_prev = bytes_xfer_now;
 724         } else {
 725              mig_throttle_on = false;
 726         }
 727         if (migrate_use_xbzrle()) {
 728             if (iterations_prev != 0) {
 729                 acct_info.xbzrle_cache_miss_rate =
 730                    (double)(acct_info.xbzrle_cache_miss -
 731                             xbzrle_cache_miss_prev) /
 732                    (acct_info.iterations - iterations_prev);
 733             }
 734             iterations_prev = acct_info.iterations;
 735             xbzrle_cache_miss_prev = acct_info.xbzrle_cache_miss;
 736         }
 737         s->dirty_pages_rate = num_dirty_pages_period * 1000
 738             / (end_time - start_time);
 739         s->dirty_bytes_rate = s->dirty_pages_rate * TARGET_PAGE_SIZE;
 740         start_time = end_time;
 741         num_dirty_pages_period = 0;
 742         s->dirty_sync_count = bitmap_sync_count;
 743     }
 744 }
 745
 746 /**
 747  * save_zero_page: Send the zero page to the stream
 748  *
 749  * Returns: Number of pages written.
 750  *
 751  * @f: QEMUFile where to send the data
 752  * @block: block that contains the page we want to send
 753  * @offset: offset inside the block for the page
 754  * @p: pointer to the page
 755  * @bytes_transferred: increase it with the number of transferred bytes
 756  */
 757 static int save_zero_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset,
 758                           uint8_t *p, uint64_t *bytes_transferred)
 759 {
 760     int pages = -1;
 761
 762     if (is_zero_range(p, TARGET_PAGE_SIZE)) {
 763         acct_info.dup_pages++;
 764         *bytes_transferred += save_page_header(f, block,
 765                                                offset | RAM_SAVE_FLAG_COMPRESS);
 766         qemu_put_byte(f, 0);
 767         *bytes_transferred += 1;
 768         pages = 1;
 769     }
 770
 771     return pages;
 772 }
 773
 774 /**
 775  * ram_save_page: Send the given page to the stream
 776  *
 777  * Returns: Number of pages written.
 778  *
 779  * @f: QEMUFile where to send the data
 780  * @block: block that contains the page we want to send
 781  * @offset: offset inside the block for the page
 782  * @last_stage: if we are at the completion stage
 783  * @bytes_transferred: increase it with the number of transferred bytes
 784  */
 785 static int ram_save_page(QEMUFile *f, RAMBlock* block, ram_addr_t offset,
 786                          bool last_stage, uint64_t *bytes_transferred)
 787 {
 788     int pages = -1;
 789     uint64_t bytes_xmit;
 790     ram_addr_t current_addr;
 791     MemoryRegion *mr = block->mr;
 792     uint8_t *p;
 793     int ret;
 794     bool send_async = true;
 795
 796     p = memory_region_get_ram_ptr(mr) + offset;
 797
 798     /* In doubt sent page as normal */
 799     bytes_xmit = 0;
 800     ret = ram_control_save_page(f, block->offset,
 801                            offset, TARGET_PAGE_SIZE, &bytes_xmit);
 802     if (bytes_xmit) {
 803         *bytes_transferred += bytes_xmit;
 804         pages = 1;
 805     }
 806
 807     XBZRLE_cache_lock();
 808
 809     current_addr = block->offset + offset;
 810
 811     if (block == last_sent_block) {
 812         offset |= RAM_SAVE_FLAG_CONTINUE;
 813     }
 814     if (ret != RAM_SAVE_CONTROL_NOT_SUPP) {
 815         if (ret != RAM_SAVE_CONTROL_DELAYED) {
 816             if (bytes_xmit > 0) {
 817                 acct_info.norm_pages++;
 818             } else if (bytes_xmit == 0) {
 819                 acct_info.dup_pages++;
 820             }
 821         }
 822     } else {
 823         pages = save_zero_page(f, block, offset, p, bytes_transferred);
 824         if (pages > 0) {
 825             /* Must let xbzrle know, otherwise a previous (now 0'd) cached
 826              * page would be stale
 827              */
 828             xbzrle_cache_zero_page(current_addr);
 829         } else if (!ram_bulk_stage && migrate_use_xbzrle()) {
 830             pages = save_xbzrle_page(f, &p, current_addr, block,
 831                                      offset, last_stage, bytes_transferred);
 832             if (!last_stage) {
 833                 /* Can't send this cached data async, since the cache page
 834                  * might get updated before it gets to the wire
 835                  */
 836                 send_async = false;
 837             }
 838         }
 839     }
 840
 841     /* XBZRLE overflow or normal page */
 842     if (pages == -1) {
 843         *bytes_transferred += save_page_header(f, block,
 844                                                offset | RAM_SAVE_FLAG_PAGE);
 845         if (send_async) {
 846             qemu_put_buffer_async(f, p, TARGET_PAGE_SIZE);
 847         } else {
 848             qemu_put_buffer(f, p, TARGET_PAGE_SIZE);
 849         }
 850         *bytes_transferred += TARGET_PAGE_SIZE;
 851         pages = 1;
 852         acct_info.norm_pages++;
 853     }
 854
 855     XBZRLE_cache_unlock();
 856
 857     return pages;
 858 }
 859
 860 static int do_compress_ram_page(CompressParam *param)
 861 {
 862     int bytes_sent, blen;
 863     uint8_t *p;
 864     RAMBlock *block = param->block;
 865     ram_addr_t offset = param->offset;
 866
 867     p = memory_region_get_ram_ptr(block->mr) + (offset & TARGET_PAGE_MASK);
 868
 869     bytes_sent = save_page_header(param->file, block, offset |
 870                                   RAM_SAVE_FLAG_COMPRESS_PAGE);
 871     blen = qemu_put_compression_data(param->file, p, TARGET_PAGE_SIZE,
 872                                      migrate_compress_level());
 873     bytes_sent += blen;
 874
 875     return bytes_sent;
 876 }
 877
 878 static inline void start_compression(CompressParam *param)
 879 {
 880     param->done = false;
 881     qemu_mutex_lock(&param->mutex);
 882     param->start = true;
 883     qemu_cond_signal(&param->cond);
 884     qemu_mutex_unlock(&param->mutex);
 885 }
 886
 887
 888 static uint64_t bytes_transferred;
 889
 890 static void flush_compressed_data(QEMUFile *f)
 891 {
 892     int idx, len, thread_count;
 893
 894     if (!migrate_use_compression()) {
 895         return;
 896     }
 897     thread_count = migrate_compress_threads();
 898     for (idx = 0; idx < thread_count; idx++) {
 899         if (!comp_param[idx].done) {
 900             qemu_mutex_lock(comp_done_lock);
 901             while (!comp_param[idx].done && !quit_comp_thread) {
 902                 qemu_cond_wait(comp_done_cond, comp_done_lock);
 903             }
 904             qemu_mutex_unlock(comp_done_lock);
 905         }
 906         if (!quit_comp_thread) {
 907             len = qemu_put_qemu_file(f, comp_param[idx].file);
 908             bytes_transferred += len;
 909         }
 910     }
 911 }
 912
 913 static inline void set_compress_params(CompressParam *param, RAMBlock *block,
 914                                        ram_addr_t offset)
 915 {
 916     param->block = block;
 917     param->offset = offset;
 918 }
 919
 920 static int compress_page_with_multi_thread(QEMUFile *f, RAMBlock *block,
 921                                            ram_addr_t offset,
 922                                            uint64_t *bytes_transferred)
 923 {
 924     int idx, thread_count, bytes_xmit = -1, pages = -1;
 925
 926     thread_count = migrate_compress_threads();
 927     qemu_mutex_lock(comp_done_lock);
 928     while (true) {
 929         for (idx = 0; idx < thread_count; idx++) {
 930             if (comp_param[idx].done) {
 931                 bytes_xmit = qemu_put_qemu_file(f, comp_param[idx].file);
 932                 set_compress_params(&comp_param[idx], block, offset);
 933                 start_compression(&comp_param[idx]);
 934                 pages = 1;
 935                 acct_info.norm_pages++;
 936                 *bytes_transferred += bytes_xmit;
 937                 break;
 938             }
 939         }
 940         if (pages > 0) {
 941             break;
 942         } else {
 943             qemu_cond_wait(comp_done_cond, comp_done_lock);
 944         }
 945     }
 946     qemu_mutex_unlock(comp_done_lock);
 947
 948     return pages;
 949 }
 950
 951 /**
 952  * ram_save_compressed_page: compress the given page and send it to the stream
 953  *
 954  * Returns: Number of pages written.
 955  *
 956  * @f: QEMUFile where to send the data
 957  * @block: block that contains the page we want to send
 958  * @offset: offset inside the block for the page
 959  * @last_stage: if we are at the completion stage
 960  * @bytes_transferred: increase it with the number of transferred bytes
 961  */
 962 static int ram_save_compressed_page(QEMUFile *f, RAMBlock *block,
 963                                     ram_addr_t offset, bool last_stage,
 964                                     uint64_t *bytes_transferred)
 965 {
 966     int pages = -1;
 967     uint64_t bytes_xmit;
 968     MemoryRegion *mr = block->mr;
 969     uint8_t *p;
 970     int ret;
 971
 972     p = memory_region_get_ram_ptr(mr) + offset;
 973
 974     bytes_xmit = 0;
 975     ret = ram_control_save_page(f, block->offset,
 976                                 offset, TARGET_PAGE_SIZE, &bytes_xmit);
 977     if (bytes_xmit) {
 978         *bytes_transferred += bytes_xmit;
 979         pages = 1;
 980     }
 981     if (block == last_sent_block) {
 982         offset |= RAM_SAVE_FLAG_CONTINUE;
 983     }
 984     if (ret != RAM_SAVE_CONTROL_NOT_SUPP) {
 985         if (ret != RAM_SAVE_CONTROL_DELAYED) {
 986             if (bytes_xmit > 0) {
 987                 acct_info.norm_pages++;
 988             } else if (bytes_xmit == 0) {
 989                 acct_info.dup_pages++;
 990             }
 991         }
 992     } else {
 993         /* When starting the process of a new block, the first page of
 994          * the block should be sent out before other pages in the same
 995          * block, and all the pages in last block should have been sent
 996          * out, keeping this order is important, because the 'cont' flag
 997          * is used to avoid resending the block name.
 998          */
 999         if (block != last_sent_block) {
1000             flush_compressed_data(f);
1001             pages = save_zero_page(f, block, offset, p, bytes_transferred);
1002             if (pages == -1) {
1003                 set_compress_params(&comp_param[0], block, offset);
1004                 /* Use the qemu thread to compress the data to make sure the
1005                  * first page is sent out before other pages
1006                  */
1007                 bytes_xmit = do_compress_ram_page(&comp_param[0]);
1008                 acct_info.norm_pages++;
1009                 qemu_put_qemu_file(f, comp_param[0].file);
1010                 *bytes_transferred += bytes_xmit;
1011                 pages = 1;
1012             }
1013         } else {
1014             pages = save_zero_page(f, block, offset, p, bytes_transferred);
1015             if (pages == -1) {
1016                 pages = compress_page_with_multi_thread(f, block, offset,
1017                                                         bytes_transferred);
1018             }
1019         }
1020     }
1021
1022     return pages;
1023 }
1024
1025 /**
1026  * ram_find_and_save_block: Finds a dirty page and sends it to f
1027  *
1028  * Called within an RCU critical section.
1029  *
1030  * Returns:  The number of pages written
1031  *           0 means no dirty pages
1032  *
1033  * @f: QEMUFile where to send the data
1034  * @last_stage: if we are at the completion stage
1035  * @bytes_transferred: increase it with the number of transferred bytes
1036  */
1037
1038 static int ram_find_and_save_block(QEMUFile *f, bool last_stage,
1039                                    uint64_t *bytes_transferred)
1040 {
1041     RAMBlock *block = last_seen_block;
1042     ram_addr_t offset = last_offset;
1043     bool complete_round = false;
1044     int pages = 0;
1045     MemoryRegion *mr;
1046
1047     if (!block)
1048         block = QLIST_FIRST_RCU(&ram_list.blocks);
1049
1050     while (true) {
1051         mr = block->mr;
1052         offset = migration_bitmap_find_and_reset_dirty(mr, offset);
1053         if (complete_round && block == last_seen_block &&
1054             offset >= last_offset) {
1055             break;
1056         }
1057         if (offset >= block->used_length) {
1058             offset = 0;
1059             block = QLIST_NEXT_RCU(block, next);
1060             if (!block) {
1061                 block = QLIST_FIRST_RCU(&ram_list.blocks);
1062                 complete_round = true;
1063                 ram_bulk_stage = false;
1064             }
1065         } else {
1066             if (migrate_use_compression()) {
1067                 pages = ram_save_compressed_page(f, block, offset, last_stage,
1068                                                  bytes_transferred);
1069             } else {
1070                 pages = ram_save_page(f, block, offset, last_stage,
1071                                       bytes_transferred);
1072             }
1073
1074             /* if page is unmodified, continue to the next */
1075             if (pages > 0) {
1076                 last_sent_block = block;
1077                 break;
1078             }
1079         }
1080     }
1081
1082     last_seen_block = block;
1083     last_offset = offset;
1084
1085     return pages;
1086 }
1087
1088 void acct_update_position(QEMUFile *f, size_t size, bool zero)
1089 {
1090     uint64_t pages = size / TARGET_PAGE_SIZE;
1091     if (zero) {
1092         acct_info.dup_pages += pages;
1093     } else {
1094         acct_info.norm_pages += pages;
1095         bytes_transferred += size;
1096         qemu_update_position(f, size);
1097     }
1098 }
1099
1100 static ram_addr_t ram_save_remaining(void)
1101 {
1102     return migration_dirty_pages;
1103 }
1104
1105 uint64_t ram_bytes_remaining(void)
1106 {
1107     return ram_save_remaining() * TARGET_PAGE_SIZE;
1108 }
1109
1110 uint64_t ram_bytes_transferred(void)
1111 {
1112     return bytes_transferred;
1113 }
1114
1115 uint64_t ram_bytes_total(void)
1116 {
1117     RAMBlock *block;
1118     uint64_t total = 0;
1119
1120     rcu_read_lock();
1121     QLIST_FOREACH_RCU(block, &ram_list.blocks, next)
1122         total += block->used_length;
1123     rcu_read_unlock();
1124     return total;
1125 }
1126
1127 void free_xbzrle_decoded_buf(void)
1128 {
1129     g_free(xbzrle_decoded_buf);
1130     xbzrle_decoded_buf = NULL;
1131 }
1132
1133 static void migration_end(void)
1134 {
1135     if (migration_bitmap) {
1136         memory_global_dirty_log_stop();
1137         g_free(migration_bitmap);
1138         migration_bitmap = NULL;
1139     }
1140
1141     XBZRLE_cache_lock();
1142     if (XBZRLE.cache) {
1143         cache_fini(XBZRLE.cache);
1144         g_free(XBZRLE.encoded_buf);
1145         g_free(XBZRLE.current_buf);
1146         XBZRLE.cache = NULL;
1147         XBZRLE.encoded_buf = NULL;
1148         XBZRLE.current_buf = NULL;
1149     }
1150     XBZRLE_cache_unlock();
1151 }
1152
1153 static void ram_migration_cancel(void *opaque)
1154 {
1155     migration_end();
1156 }
1157
1158 static void reset_ram_globals(void)
1159 {
1160     last_seen_block = NULL;
1161     last_sent_block = NULL;
1162     last_offset = 0;
1163     last_version = ram_list.version;
1164     ram_bulk_stage = true;
1165 }
1166
1167 #define MAX_WAIT 50 /* ms, half buffered_file limit */
1168
1169
1170 /* Each of ram_save_setup, ram_save_iterate and ram_save_complete has
1171  * long-running RCU critical section.  When rcu-reclaims in the code
1172  * start to become numerous it will be necessary to reduce the
1173  * granularity of these critical sections.
1174  */
1175
1176 static int ram_save_setup(QEMUFile *f, void *opaque)
1177 {
1178     RAMBlock *block;
1179     int64_t ram_bitmap_pages; /* Size of bitmap in pages, including gaps */
1180
1181     mig_throttle_on = false;
1182     dirty_rate_high_cnt = 0;
1183     bitmap_sync_count = 0;
1184     migration_bitmap_sync_init();
1185
1186     if (migrate_use_xbzrle()) {
1187         XBZRLE_cache_lock();
1188         XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() /
1189                                   TARGET_PAGE_SIZE,
1190                                   TARGET_PAGE_SIZE);
1191         if (!XBZRLE.cache) {
1192             XBZRLE_cache_unlock();
1193             error_report("Error creating cache");
1194             return -1;
1195         }
1196         XBZRLE_cache_unlock();
1197
1198         /* We prefer not to abort if there is no memory */
1199         XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE);
1200         if (!XBZRLE.encoded_buf) {
1201             error_report("Error allocating encoded_buf");
1202             return -1;
1203         }
1204
1205         XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE);
1206         if (!XBZRLE.current_buf) {
1207             error_report("Error allocating current_buf");
1208             g_free(XBZRLE.encoded_buf);
1209             XBZRLE.encoded_buf = NULL;
1210             return -1;
1211         }
1212
1213         acct_clear();
1214     }
1215
1216     /* iothread lock needed for ram_list.dirty_memory[] */
1217     qemu_mutex_lock_iothread();
1218     qemu_mutex_lock_ramlist();
1219     rcu_read_lock();
1220     bytes_transferred = 0;
1221     reset_ram_globals();
1222
1223     ram_bitmap_pages = last_ram_offset() >> TARGET_PAGE_BITS;
1224     migration_bitmap = bitmap_new(ram_bitmap_pages);
1225     bitmap_set(migration_bitmap, 0, ram_bitmap_pages);
1226
1227     /*
1228      * Count the total number of pages used by ram blocks not including any
1229      * gaps due to alignment or unplugs.
1230      */
1231     migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS;
1232
1233     memory_global_dirty_log_start();
1234     migration_bitmap_sync();
1235     qemu_mutex_unlock_ramlist();
1236     qemu_mutex_unlock_iothread();
1237
1238     qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE);
1239
1240     QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
1241         qemu_put_byte(f, strlen(block->idstr));
1242         qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr));
1243         qemu_put_be64(f, block->used_length);
1244     }
1245
1246     rcu_read_unlock();
1247
1248     ram_control_before_iterate(f, RAM_CONTROL_SETUP);
1249     ram_control_after_iterate(f, RAM_CONTROL_SETUP);
1250
1251     qemu_put_be64(f, RAM_SAVE_FLAG_EOS);
1252
1253     return 0;
1254 }
1255
1256 static int ram_save_iterate(QEMUFile *f, void *opaque)
1257 {
1258     int ret;
1259     int i;
1260     int64_t t0;
1261     int pages_sent = 0;
1262
1263     rcu_read_lock();
1264     if (ram_list.version != last_version) {
1265         reset_ram_globals();
1266     }
1267
1268     /* Read version before ram_list.blocks */
1269     smp_rmb();
1270
1271     ram_control_before_iterate(f, RAM_CONTROL_ROUND);
1272
1273     t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1274     i = 0;
1275     while ((ret = qemu_file_rate_limit(f)) == 0) {
1276         int pages;
1277
1278         pages = ram_find_and_save_block(f, false, &bytes_transferred);
1279         /* no more pages to sent */
1280         if (pages == 0) {
1281             break;
1282         }
1283         pages_sent += pages;
1284         acct_info.iterations++;
1285         check_guest_throttling();
1286         /* we want to check in the 1st loop, just in case it was the 1st time
1287            and we had to sync the dirty bitmap.
1288            qemu_get_clock_ns() is a bit expensive, so we only check each some
1289            iterations
1290         */
1291         if ((i & 63) == 0) {
1292             uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000;
1293             if (t1 > MAX_WAIT) {
1294                 DPRINTF("big wait: %" PRIu64 " milliseconds, %d iterations\n",
1295                         t1, i);
1296                 break;
1297             }
1298         }
1299         i++;
1300     }
1301     flush_compressed_data(f);
1302     rcu_read_unlock();
1303
1304     /*
1305      * Must occur before EOS (or any QEMUFile operation)
1306      * because of RDMA protocol.
1307      */
1308     ram_control_after_iterate(f, RAM_CONTROL_ROUND);
1309
1310     qemu_put_be64(f, RAM_SAVE_FLAG_EOS);
1311     bytes_transferred += 8;
1312
1313     ret = qemu_file_get_error(f);
1314     if (ret < 0) {
1315         return ret;
1316     }
1317
1318     return pages_sent;
1319 }
1320
1321 /* Called with iothread lock */
1322 static int ram_save_complete(QEMUFile *f, void *opaque)
1323 {
1324     rcu_read_lock();
1325
1326     migration_bitmap_sync();
1327
1328     ram_control_before_iterate(f, RAM_CONTROL_FINISH);
1329
1330     /* try transferring iterative blocks of memory */
1331
1332     /* flush all remaining blocks regardless of rate limiting */
1333     while (true) {
1334         int pages;
1335
1336         pages = ram_find_and_save_block(f, true, &bytes_transferred);
1337         /* no more blocks to sent */
1338         if (pages == 0) {
1339             break;
1340         }
1341     }
1342
1343     flush_compressed_data(f);
1344     ram_control_after_iterate(f, RAM_CONTROL_FINISH);
1345     migration_end();
1346
1347     rcu_read_unlock();
1348     qemu_put_be64(f, RAM_SAVE_FLAG_EOS);
1349
1350     return 0;
1351 }
1352
1353 static uint64_t ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size)
1354 {
1355     uint64_t remaining_size;
1356
1357     remaining_size = ram_save_remaining() * TARGET_PAGE_SIZE;
1358
1359     if (remaining_size < max_size) {
1360         qemu_mutex_lock_iothread();
1361         rcu_read_lock();
1362         migration_bitmap_sync();
1363         rcu_read_unlock();
1364         qemu_mutex_unlock_iothread();
1365         remaining_size = ram_save_remaining() * TARGET_PAGE_SIZE;
1366     }
1367     return remaining_size;
1368 }
1369
1370 static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host)
1371 {
1372     unsigned int xh_len;
1373     int xh_flags;
1374
1375     if (!xbzrle_decoded_buf) {
1376         xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE);
1377     }
1378
1379     /* extract RLE header */
1380     xh_flags = qemu_get_byte(f);
1381     xh_len = qemu_get_be16(f);
1382
1383     if (xh_flags != ENCODING_FLAG_XBZRLE) {
1384         error_report("Failed to load XBZRLE page - wrong compression!");
1385         return -1;
1386     }
1387
1388     if (xh_len > TARGET_PAGE_SIZE) {
1389         error_report("Failed to load XBZRLE page - len overflow!");
1390         return -1;
1391     }
1392     /* load data and decode */
1393     qemu_get_buffer(f, xbzrle_decoded_buf, xh_len);
1394
1395     /* decode RLE */
1396     if (xbzrle_decode_buffer(xbzrle_decoded_buf, xh_len, host,
1397                              TARGET_PAGE_SIZE) == -1) {
1398         error_report("Failed to load XBZRLE page - decode error!");
1399         return -1;
1400     }
1401
1402     return 0;
1403 }
1404
1405 /* Must be called from within a rcu critical section.
1406  * Returns a pointer from within the RCU-protected ram_list.
1407  */
1408 static inline void *host_from_stream_offset(QEMUFile *f,
1409                                             ram_addr_t offset,
1410                                             int flags)
1411 {
1412     static RAMBlock *block = NULL;
1413     char id[256];
1414     uint8_t len;
1415
1416     if (flags & RAM_SAVE_FLAG_CONTINUE) {
1417         if (!block || block->max_length <= offset) {
1418             error_report("Ack, bad migration stream!");
1419             return NULL;
1420         }
1421
1422         return memory_region_get_ram_ptr(block->mr) + offset;
1423     }
1424
1425     len = qemu_get_byte(f);
1426     qemu_get_buffer(f, (uint8_t *)id, len);
1427     id[len] = 0;
1428
1429     QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
1430         if (!strncmp(id, block->idstr, sizeof(id)) &&
1431             block->max_length > offset) {
1432             return memory_region_get_ram_ptr(block->mr) + offset;
1433         }
1434     }
1435
1436     error_report("Can't find block %s!", id);
1437     return NULL;
1438 }
1439
1440 /*
1441  * If a page (or a whole RDMA chunk) has been
1442  * determined to be zero, then zap it.
1443  */
1444 void ram_handle_compressed(void *host, uint8_t ch, uint64_t size)
1445 {
1446     if (ch != 0 || !is_zero_range(host, size)) {
1447         memset(host, ch, size);
1448     }
1449 }
1450
1451 static void *do_data_decompress(void *opaque)
1452 {
1453     while (!quit_decomp_thread) {
1454         /* To be done */
1455     }
1456
1457     return NULL;
1458 }
1459
1460 void migrate_decompress_threads_create(void)
1461 {
1462     int i, thread_count;
1463
1464     thread_count = migrate_decompress_threads();
1465     decompress_threads = g_new0(QemuThread, thread_count);
1466     decomp_param = g_new0(DecompressParam, thread_count);
1467     compressed_data_buf = g_malloc0(compressBound(TARGET_PAGE_SIZE));
1468     quit_decomp_thread = false;
1469     for (i = 0; i < thread_count; i++) {
1470         qemu_mutex_init(&decomp_param[i].mutex);
1471         qemu_cond_init(&decomp_param[i].cond);
1472         decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE));
1473         qemu_thread_create(decompress_threads + i, "decompress",
1474                            do_data_decompress, decomp_param + i,
1475                            QEMU_THREAD_JOINABLE);
1476     }
1477 }
1478
1479 void migrate_decompress_threads_join(void)
1480 {
1481     int i, thread_count;
1482
1483     quit_decomp_thread = true;
1484     thread_count = migrate_decompress_threads();
1485     for (i = 0; i < thread_count; i++) {
1486         qemu_thread_join(decompress_threads + i);
1487         qemu_mutex_destroy(&decomp_param[i].mutex);
1488         qemu_cond_destroy(&decomp_param[i].cond);
1489         g_free(decomp_param[i].compbuf);
1490     }
1491     g_free(decompress_threads);
1492     g_free(decomp_param);
1493     g_free(compressed_data_buf);
1494     decompress_threads = NULL;
1495     decomp_param = NULL;
1496     compressed_data_buf = NULL;
1497 }
1498
1499 static void decompress_data_with_multi_threads(uint8_t *compbuf,
1500                                                void *host, int len)
1501 {
1502     /* To be done */
1503 }
1504
1505 static int ram_load(QEMUFile *f, void *opaque, int version_id)
1506 {
1507     int flags = 0, ret = 0;
1508     static uint64_t seq_iter;
1509     int len = 0;
1510
1511     seq_iter++;
1512
1513     if (version_id != 4) {
1514         ret = -EINVAL;
1515     }
1516
1517     /* This RCU critical section can be very long running.
1518      * When RCU reclaims in the code start to become numerous,
1519      * it will be necessary to reduce the granularity of this
1520      * critical section.
1521      */
1522     rcu_read_lock();
1523     while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) {
1524         ram_addr_t addr, total_ram_bytes;
1525         void *host;
1526         uint8_t ch;
1527
1528         addr = qemu_get_be64(f);
1529         flags = addr & ~TARGET_PAGE_MASK;
1530         addr &= TARGET_PAGE_MASK;
1531
1532         switch (flags & ~RAM_SAVE_FLAG_CONTINUE) {
1533         case RAM_SAVE_FLAG_MEM_SIZE:
1534             /* Synchronize RAM block list */
1535             total_ram_bytes = addr;
1536             while (!ret && total_ram_bytes) {
1537                 RAMBlock *block;
1538                 uint8_t len;
1539                 char id[256];
1540                 ram_addr_t length;
1541
1542                 len = qemu_get_byte(f);
1543                 qemu_get_buffer(f, (uint8_t *)id, len);
1544                 id[len] = 0;
1545                 length = qemu_get_be64(f);
1546
1547                 QLIST_FOREACH_RCU(block, &ram_list.blocks, next) {
1548                     if (!strncmp(id, block->idstr, sizeof(id))) {
1549                         if (length != block->used_length) {
1550                             Error *local_err = NULL;
1551
1552                             ret = qemu_ram_resize(block->offset, length, &local_err);
1553                             if (local_err) {
1554                                 error_report_err(local_err);
1555                             }
1556                         }
1557                         break;
1558                     }
1559                 }
1560
1561                 if (!block) {
1562                     error_report("Unknown ramblock \"%s\", cannot "
1563                                  "accept migration", id);
1564                     ret = -EINVAL;
1565                 }
1566
1567                 total_ram_bytes -= length;
1568             }
1569             break;
1570         case RAM_SAVE_FLAG_COMPRESS:
1571             host = host_from_stream_offset(f, addr, flags);
1572             if (!host) {
1573                 error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
1574                 ret = -EINVAL;
1575                 break;
1576             }
1577             ch = qemu_get_byte(f);
1578             ram_handle_compressed(host, ch, TARGET_PAGE_SIZE);
1579             break;
1580         case RAM_SAVE_FLAG_PAGE:
1581             host = host_from_stream_offset(f, addr, flags);
1582             if (!host) {
1583                 error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
1584                 ret = -EINVAL;
1585                 break;
1586             }
1587             qemu_get_buffer(f, host, TARGET_PAGE_SIZE);
1588             break;
1589         case RAM_SAVE_FLAG_COMPRESS_PAGE:
1590             host = host_from_stream_offset(f, addr, flags);
1591             if (!host) {
1592                 error_report("Invalid RAM offset " RAM_ADDR_FMT, addr);
1593                 ret = -EINVAL;
1594                 break;
1595             }
1596
1597             len = qemu_get_be32(f);
1598             if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) {
1599                 error_report("Invalid compressed data length: %d", len);
1600                 ret = -EINVAL;
1601                 break;
1602             }
1603             qemu_get_buffer(f, compressed_data_buf, len);
1604             decompress_data_with_multi_threads(compressed_data_buf, host, len);
1605             break;
1606         case RAM_SAVE_FLAG_XBZRLE:
1607             host = host_from_stream_offset(f, addr, flags);
1608             if (!host) {
1609                 error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
1610                 ret = -EINVAL;
1611                 break;
1612             }
1613             if (load_xbzrle(f, addr, host) < 0) {
1614                 error_report("Failed to decompress XBZRLE page at "
1615                              RAM_ADDR_FMT, addr);
1616                 ret = -EINVAL;
1617                 break;
1618             }
1619             break;
1620         case RAM_SAVE_FLAG_EOS:
1621             /* normal exit */
1622             break;
1623         default:
1624             if (flags & RAM_SAVE_FLAG_HOOK) {
1625                 ram_control_load_hook(f, flags);
1626             } else {
1627                 error_report("Unknown combination of migration flags: %#x",
1628                              flags);
1629                 ret = -EINVAL;
1630             }
1631         }
1632         if (!ret) {
1633             ret = qemu_file_get_error(f);
1634         }
1635     }
1636
1637     rcu_read_unlock();
1638     DPRINTF("Completed load of VM with exit code %d seq iteration "
1639             "%" PRIu64 "\n", ret, seq_iter);
1640     return ret;
1641 }
1642
1643 static SaveVMHandlers savevm_ram_handlers = {
1644     .save_live_setup = ram_save_setup,
1645     .save_live_iterate = ram_save_iterate,
1646     .save_live_complete = ram_save_complete,
1647     .save_live_pending = ram_save_pending,
1648     .load_state = ram_load,
1649     .cancel = ram_migration_cancel,
1650 };
1651
1652 void ram_mig_init(void)
1653 {
1654     qemu_mutex_init(&XBZRLE.lock);
1655     register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, NULL);
1656 }
1657
1658 struct soundhw {
1659     const char *name;
1660     const char *descr;
1661     int enabled;
1662     int isa;
1663     union {
1664         int (*init_isa) (ISABus *bus);
1665         int (*init_pci) (PCIBus *bus);
1666     } init;
1667 };
1668
1669 static struct soundhw soundhw[9];
1670 static int soundhw_count;
1671
1672 void isa_register_soundhw(const char *name, const char *descr,
1673                           int (*init_isa)(ISABus *bus))
1674 {
1675     assert(soundhw_count < ARRAY_SIZE(soundhw) - 1);
1676     soundhw[soundhw_count].name = name;
1677     soundhw[soundhw_count].descr = descr;
1678     soundhw[soundhw_count].isa = 1;
1679     soundhw[soundhw_count].init.init_isa = init_isa;
1680     soundhw_count++;
1681 }
1682
1683 void pci_register_soundhw(const char *name, const char *descr,
1684                           int (*init_pci)(PCIBus *bus))
1685 {
1686     assert(soundhw_count < ARRAY_SIZE(soundhw) - 1);
1687     soundhw[soundhw_count].name = name;
1688     soundhw[soundhw_count].descr = descr;
1689     soundhw[soundhw_count].isa = 0;
1690     soundhw[soundhw_count].init.init_pci = init_pci;
1691     soundhw_count++;
1692 }
1693
1694 void select_soundhw(const char *optarg)
1695 {
1696     struct soundhw *c;
1697
1698     if (is_help_option(optarg)) {
1699     show_valid_cards:
1700
1701         if (soundhw_count) {
1702              printf("Valid sound card names (comma separated):\n");
1703              for (c = soundhw; c->name; ++c) {
1704                  printf ("%-11s %s\n", c->name, c->descr);
1705              }
1706              printf("\n-soundhw all will enable all of the above\n");
1707         } else {
1708              printf("Machine has no user-selectable audio hardware "
1709                     "(it may or may not have always-present audio hardware).\n");
1710         }
1711         exit(!is_help_option(optarg));
1712     }
1713     else {
1714         size_t l;
1715         const char *p;
1716         char *e;
1717         int bad_card = 0;
1718
1719         if (!strcmp(optarg, "all")) {
1720             for (c = soundhw; c->name; ++c) {
1721                 c->enabled = 1;
1722             }
1723             return;
1724         }
1725
1726         p = optarg;
1727         while (*p) {
1728             e = strchr(p, ',');
1729             l = !e ? strlen(p) : (size_t) (e - p);
1730
1731             for (c = soundhw; c->name; ++c) {
1732                 if (!strncmp(c->name, p, l) && !c->name[l]) {
1733                     c->enabled = 1;
1734                     break;
1735                 }
1736             }
1737
1738             if (!c->name) {
1739                 if (l > 80) {
1740                     error_report("Unknown sound card name (too big to show)");
1741                 }
1742                 else {
1743                     error_report("Unknown sound card name `%.*s'",
1744                                  (int) l, p);
1745                 }
1746                 bad_card = 1;
1747             }
1748             p += l + (e != NULL);
1749         }
1750
1751         if (bad_card) {
1752             goto show_valid_cards;
1753         }
1754     }
1755 }
1756
1757 void audio_init(void)
1758 {
1759     struct soundhw *c;
1760     ISABus *isa_bus = (ISABus *) object_resolve_path_type("", TYPE_ISA_BUS, NULL);
1761     PCIBus *pci_bus = (PCIBus *) object_resolve_path_type("", TYPE_PCI_BUS, NULL);
1762
1763     for (c = soundhw; c->name; ++c) {
1764         if (c->enabled) {
1765             if (c->isa) {
1766                 if (!isa_bus) {
1767                     error_report("ISA bus not available for %s", c->name);
1768                     exit(1);
1769                 }
1770                 c->init.init_isa(isa_bus);
1771             } else {
1772                 if (!pci_bus) {
1773                     error_report("PCI bus not available for %s", c->name);
1774                     exit(1);
1775                 }
1776                 c->init.init_pci(pci_bus);
1777             }
1778         }
1779     }
1780 }
1781
1782 int qemu_uuid_parse(const char *str, uint8_t *uuid)
1783 {
1784     int ret;
1785
1786     if (strlen(str) != 36) {
1787         return -1;
1788     }
1789
1790     ret = sscanf(str, UUID_FMT, &uuid[0], &uuid[1], &uuid[2], &uuid[3],
1791                  &uuid[4], &uuid[5], &uuid[6], &uuid[7], &uuid[8], &uuid[9],
1792                  &uuid[10], &uuid[11], &uuid[12], &uuid[13], &uuid[14],
1793                  &uuid[15]);
1794
1795     if (ret != 16) {
1796         return -1;
1797     }
1798     return 0;
1799 }
1800
1801 void do_acpitable_option(const QemuOpts *opts)
1802 {
1803 #ifdef TARGET_I386
1804     Error *err = NULL;
1805
1806     acpi_table_add(opts, &err);
1807     if (err) {
1808         error_report("Wrong acpi table provided: %s",
1809                      error_get_pretty(err));
1810         error_free(err);
1811         exit(1);
1812     }
1813 #endif
1814 }
1815
1816 void do_smbios_option(QemuOpts *opts)
1817 {
1818 #ifdef TARGET_I386
1819     smbios_entry_add(opts);
1820 #endif
1821 }
1822
1823 void cpudef_init(void)
1824 {
1825 #if defined(cpudef_setup)
1826     cpudef_setup(); /* parse cpu definitions in target config file */
1827 #endif
1828 }
1829
1830 int kvm_available(void)
1831 {
1832 #ifdef CONFIG_KVM
1833     return 1;
1834 #else
1835     return 0;
1836 #endif
1837 }
1838
1839 int xen_available(void)
1840 {
1841 #ifdef CONFIG_XEN
1842     return 1;
1843 #else
1844     return 0;
1845 #endif
1846 }
1847
1848
1849 TargetInfo *qmp_query_target(Error **errp)
1850 {
1851     TargetInfo *info = g_malloc0(sizeof(*info));
1852
1853     info->arch = g_strdup(TARGET_NAME);
1854
1855     return info;
1856 }
1857
1858 /* Stub function that's gets run on the vcpu when its brought out of the
1859    VM to run inside qemu via async_run_on_cpu()*/
1860 static void mig_sleep_cpu(void *opq)
1861 {
1862     qemu_mutex_unlock_iothread();
1863     g_usleep(30*1000);
1864     qemu_mutex_lock_iothread();
1865 }
1866
1867 /* To reduce the dirty rate explicitly disallow the VCPUs from spending
1868    much time in the VM. The migration thread will try to catchup.
1869    Workload will experience a performance drop.
1870 */
1871 static void mig_throttle_guest_down(void)
1872 {
1873     CPUState *cpu;
1874
1875     qemu_mutex_lock_iothread();
1876     CPU_FOREACH(cpu) {
1877         async_run_on_cpu(cpu, mig_sleep_cpu, NULL);
1878     }
1879     qemu_mutex_unlock_iothread();
1880 }
1881
1882 static void check_guest_throttling(void)
1883 {
1884     static int64_t t0;
1885     int64_t        t1;
1886
1887     if (!mig_throttle_on) {
1888         return;
1889     }
1890
1891     if (!t0)  {
1892         t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1893         return;
1894     }
1895
1896     t1 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
1897
1898     /* If it has been more than 40 ms since the last time the guest
1899      * was throttled then do it again.
1900      */
1901     if (40 < (t1-t0)/1000000) {
1902         mig_throttle_guest_down();
1903         t0 = t1;
1904     }
1905 }