migration/dirtyrate.c

   1 /*
   2  * Dirtyrate implement code
   3  *
   4  * Copyright (c) 2020 HUAWEI TECHNOLOGIES CO.,LTD.
   5  *
   6  * Authors:
   7  *  Chuan Zheng <zhengchuan@huawei.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  10  * See the COPYING file in the top-level directory.
  11  */
  12
  13 #include "qemu/osdep.h"
  14 #include "qemu/error-report.h"
  15 #include <zlib.h>
  16 #include "hw/core/cpu.h"
  17 #include "qapi/error.h"
  18 #include "exec/ramblock.h"
  19 #include "exec/target_page.h"
  20 #include "qemu/rcu_queue.h"
  21 #include "qemu/main-loop.h"
  22 #include "qapi/qapi-commands-migration.h"
  23 #include "ram.h"
  24 #include "trace.h"
  25 #include "dirtyrate.h"
  26 #include "monitor/hmp.h"
  27 #include "monitor/monitor.h"
  28 #include "qapi/qmp/qdict.h"
  29 #include "sysemu/kvm.h"
  30 #include "sysemu/runstate.h"
  31 #include "exec/memory.h"
  32 #include "qemu/xxhash.h"
  33
  34 /*
  35  * total_dirty_pages is procted by BQL and is used
  36  * to stat dirty pages during the period of two
  37  * memory_global_dirty_log_sync
  38  */
  39 uint64_t total_dirty_pages;
  40
  41 typedef struct DirtyPageRecord {
  42     uint64_t start_pages;
  43     uint64_t end_pages;
  44 } DirtyPageRecord;
  45
  46 static int CalculatingState = DIRTY_RATE_STATUS_UNSTARTED;
  47 static struct DirtyRateStat DirtyStat;
  48 static DirtyRateMeasureMode dirtyrate_mode =
  49                 DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING;
  50
  51 static int64_t dirty_stat_wait(int64_t msec, int64_t initial_time)
  52 {
  53     int64_t current_time;
  54
  55     current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
  56     if ((current_time - initial_time) >= msec) {
  57         msec = current_time - initial_time;
  58     } else {
  59         g_usleep((msec + initial_time - current_time) * 1000);
  60         /* g_usleep may overshoot */
  61         msec = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - initial_time;
  62     }
  63
  64     return msec;
  65 }
  66
  67 static inline void record_dirtypages(DirtyPageRecord *dirty_pages,
  68                                      CPUState *cpu, bool start)
  69 {
  70     if (start) {
  71         dirty_pages[cpu->cpu_index].start_pages = cpu->dirty_pages;
  72     } else {
  73         dirty_pages[cpu->cpu_index].end_pages = cpu->dirty_pages;
  74     }
  75 }
  76
  77 static int64_t do_calculate_dirtyrate(DirtyPageRecord dirty_pages,
  78                                       int64_t calc_time_ms)
  79 {
  80     uint64_t increased_dirty_pages =
  81         dirty_pages.end_pages - dirty_pages.start_pages;
  82
  83     /*
  84      * multiply by 1000ms/s _before_ converting down to megabytes
  85      * to avoid losing precision
  86      */
  87     return qemu_target_pages_to_MiB(increased_dirty_pages * 1000) /
  88         calc_time_ms;
  89 }
  90
  91 void global_dirty_log_change(unsigned int flag, bool start)
  92 {
  93     bql_lock();
  94     if (start) {
  95         memory_global_dirty_log_start(flag);
  96     } else {
  97         memory_global_dirty_log_stop(flag);
  98     }
  99     bql_unlock();
 100 }
 101
 102 /*
 103  * global_dirty_log_sync
 104  * 1. sync dirty log from kvm
 105  * 2. stop dirty tracking if needed.
 106  */
 107 static void global_dirty_log_sync(unsigned int flag, bool one_shot)
 108 {
 109     bql_lock();
 110     memory_global_dirty_log_sync(false);
 111     if (one_shot) {
 112         memory_global_dirty_log_stop(flag);
 113     }
 114     bql_unlock();
 115 }
 116
 117 static DirtyPageRecord *vcpu_dirty_stat_alloc(VcpuStat *stat)
 118 {
 119     CPUState *cpu;
 120     int nvcpu = 0;
 121
 122     CPU_FOREACH(cpu) {
 123         nvcpu++;
 124     }
 125
 126     stat->nvcpu = nvcpu;
 127     stat->rates = g_new0(DirtyRateVcpu, nvcpu);
 128
 129     return g_new0(DirtyPageRecord, nvcpu);
 130 }
 131
 132 static void vcpu_dirty_stat_collect(DirtyPageRecord *records,
 133                                     bool start)
 134 {
 135     CPUState *cpu;
 136
 137     CPU_FOREACH(cpu) {
 138         record_dirtypages(records, cpu, start);
 139     }
 140 }
 141
 142 int64_t vcpu_calculate_dirtyrate(int64_t calc_time_ms,
 143                                  VcpuStat *stat,
 144                                  unsigned int flag,
 145                                  bool one_shot)
 146 {
 147     DirtyPageRecord *records;
 148     int64_t init_time_ms;
 149     int64_t duration;
 150     int64_t dirtyrate;
 151     int i = 0;
 152     unsigned int gen_id;
 153
 154 retry:
 155     init_time_ms = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 156
 157     WITH_QEMU_LOCK_GUARD(&qemu_cpu_list_lock) {
 158         gen_id = cpu_list_generation_id_get();
 159         records = vcpu_dirty_stat_alloc(stat);
 160         vcpu_dirty_stat_collect(records, true);
 161     }
 162
 163     duration = dirty_stat_wait(calc_time_ms, init_time_ms);
 164
 165     global_dirty_log_sync(flag, one_shot);
 166
 167     WITH_QEMU_LOCK_GUARD(&qemu_cpu_list_lock) {
 168         if (gen_id != cpu_list_generation_id_get()) {
 169             g_free(records);
 170             g_free(stat->rates);
 171             cpu_list_unlock();
 172             goto retry;
 173         }
 174         vcpu_dirty_stat_collect(records, false);
 175     }
 176
 177     for (i = 0; i < stat->nvcpu; i++) {
 178         dirtyrate = do_calculate_dirtyrate(records[i], duration);
 179
 180         stat->rates[i].id = i;
 181         stat->rates[i].dirty_rate = dirtyrate;
 182
 183         trace_dirtyrate_do_calculate_vcpu(i, dirtyrate);
 184     }
 185
 186     g_free(records);
 187
 188     return duration;
 189 }
 190
 191 static bool is_calc_time_valid(int64_t msec)
 192 {
 193     if ((msec < MIN_CALC_TIME_MS) || (msec > MAX_CALC_TIME_MS)) {
 194         return false;
 195     }
 196
 197     return true;
 198 }
 199
 200 static bool is_sample_pages_valid(int64_t pages)
 201 {
 202     return pages >= MIN_SAMPLE_PAGE_COUNT &&
 203            pages <= MAX_SAMPLE_PAGE_COUNT;
 204 }
 205
 206 static int dirtyrate_set_state(int *state, int old_state, int new_state)
 207 {
 208     assert(new_state < DIRTY_RATE_STATUS__MAX);
 209     trace_dirtyrate_set_state(DirtyRateStatus_str(new_state));
 210     if (qatomic_cmpxchg(state, old_state, new_state) == old_state) {
 211         return 0;
 212     } else {
 213         return -1;
 214     }
 215 }
 216
 217 /* Decimal power of given time unit relative to one second */
 218 static int time_unit_to_power(TimeUnit time_unit)
 219 {
 220     switch (time_unit) {
 221     case TIME_UNIT_SECOND:
 222         return 0;
 223     case TIME_UNIT_MILLISECOND:
 224         return -3;
 225     default:
 226         assert(false); /* unreachable */
 227         return 0;
 228     }
 229 }
 230
 231 static int64_t convert_time_unit(int64_t value, TimeUnit unit_from,
 232                                  TimeUnit unit_to)
 233 {
 234     int power = time_unit_to_power(unit_from) -
 235                 time_unit_to_power(unit_to);
 236     while (power < 0) {
 237         value /= 10;
 238         power += 1;
 239     }
 240     while (power > 0) {
 241         value *= 10;
 242         power -= 1;
 243     }
 244     return value;
 245 }
 246
 247
 248 static struct DirtyRateInfo *
 249 query_dirty_rate_info(TimeUnit calc_time_unit)
 250 {
 251     int i;
 252     int64_t dirty_rate = DirtyStat.dirty_rate;
 253     struct DirtyRateInfo *info = g_new0(DirtyRateInfo, 1);
 254     DirtyRateVcpuList *head = NULL, **tail = &head;
 255
 256     info->status = CalculatingState;
 257     info->start_time = DirtyStat.start_time;
 258     info->calc_time = convert_time_unit(DirtyStat.calc_time_ms,
 259                                         TIME_UNIT_MILLISECOND,
 260                                         calc_time_unit);
 261     info->calc_time_unit = calc_time_unit;
 262     info->sample_pages = DirtyStat.sample_pages;
 263     info->mode = dirtyrate_mode;
 264
 265     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURED) {
 266         info->has_dirty_rate = true;
 267         info->dirty_rate = dirty_rate;
 268
 269         if (dirtyrate_mode == DIRTY_RATE_MEASURE_MODE_DIRTY_RING) {
 270             /*
 271              * set sample_pages with 0 to indicate page sampling
 272              * isn't enabled
 273              **/
 274             info->sample_pages = 0;
 275             info->has_vcpu_dirty_rate = true;
 276             for (i = 0; i < DirtyStat.dirty_ring.nvcpu; i++) {
 277                 DirtyRateVcpu *rate = g_new0(DirtyRateVcpu, 1);
 278                 rate->id = DirtyStat.dirty_ring.rates[i].id;
 279                 rate->dirty_rate = DirtyStat.dirty_ring.rates[i].dirty_rate;
 280                 QAPI_LIST_APPEND(tail, rate);
 281             }
 282             info->vcpu_dirty_rate = head;
 283         }
 284
 285         if (dirtyrate_mode == DIRTY_RATE_MEASURE_MODE_DIRTY_BITMAP) {
 286             info->sample_pages = 0;
 287         }
 288     }
 289
 290     trace_query_dirty_rate_info(DirtyRateStatus_str(CalculatingState));
 291
 292     return info;
 293 }
 294
 295 static void init_dirtyrate_stat(struct DirtyRateConfig config)
 296 {
 297     DirtyStat.dirty_rate = -1;
 298     DirtyStat.start_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) / 1000;
 299     DirtyStat.calc_time_ms = config.calc_time_ms;
 300     DirtyStat.sample_pages = config.sample_pages_per_gigabytes;
 301
 302     switch (config.mode) {
 303     case DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING:
 304         DirtyStat.page_sampling.total_dirty_samples = 0;
 305         DirtyStat.page_sampling.total_sample_count = 0;
 306         DirtyStat.page_sampling.total_block_mem_MB = 0;
 307         break;
 308     case DIRTY_RATE_MEASURE_MODE_DIRTY_RING:
 309         DirtyStat.dirty_ring.nvcpu = -1;
 310         DirtyStat.dirty_ring.rates = NULL;
 311         break;
 312     default:
 313         break;
 314     }
 315 }
 316
 317 static void cleanup_dirtyrate_stat(struct DirtyRateConfig config)
 318 {
 319     /* last calc-dirty-rate qmp use dirty ring mode */
 320     if (dirtyrate_mode == DIRTY_RATE_MEASURE_MODE_DIRTY_RING) {
 321         free(DirtyStat.dirty_ring.rates);
 322         DirtyStat.dirty_ring.rates = NULL;
 323     }
 324 }
 325
 326 static void update_dirtyrate_stat(struct RamblockDirtyInfo *info)
 327 {
 328     DirtyStat.page_sampling.total_dirty_samples += info->sample_dirty_count;
 329     DirtyStat.page_sampling.total_sample_count += info->sample_pages_count;
 330     /* size of total pages in MB */
 331     DirtyStat.page_sampling.total_block_mem_MB +=
 332         qemu_target_pages_to_MiB(info->ramblock_pages);
 333 }
 334
 335 static void update_dirtyrate(uint64_t msec)
 336 {
 337     uint64_t dirtyrate;
 338     uint64_t total_dirty_samples = DirtyStat.page_sampling.total_dirty_samples;
 339     uint64_t total_sample_count = DirtyStat.page_sampling.total_sample_count;
 340     uint64_t total_block_mem_MB = DirtyStat.page_sampling.total_block_mem_MB;
 341
 342     dirtyrate = total_dirty_samples * total_block_mem_MB *
 343                 1000 / (total_sample_count * msec);
 344
 345     DirtyStat.dirty_rate = dirtyrate;
 346 }
 347
 348 /*
 349  * Compute hash of a single page of size TARGET_PAGE_SIZE.
 350  */
 351 static uint32_t compute_page_hash(void *ptr)
 352 {
 353     size_t page_size = qemu_target_page_size();
 354     uint32_t i;
 355     uint64_t v1, v2, v3, v4;
 356     uint64_t res;
 357     const uint64_t *p = ptr;
 358
 359     v1 = QEMU_XXHASH_SEED + XXH_PRIME64_1 + XXH_PRIME64_2;
 360     v2 = QEMU_XXHASH_SEED + XXH_PRIME64_2;
 361     v3 = QEMU_XXHASH_SEED + 0;
 362     v4 = QEMU_XXHASH_SEED - XXH_PRIME64_1;
 363     for (i = 0; i < page_size / 8; i += 4) {
 364         v1 = XXH64_round(v1, p[i + 0]);
 365         v2 = XXH64_round(v2, p[i + 1]);
 366         v3 = XXH64_round(v3, p[i + 2]);
 367         v4 = XXH64_round(v4, p[i + 3]);
 368     }
 369     res = XXH64_mergerounds(v1, v2, v3, v4);
 370     res += page_size;
 371     res = XXH64_avalanche(res);
 372     return (uint32_t)(res & UINT32_MAX);
 373 }
 374
 375
 376 /*
 377  * get hash result for the sampled memory with length of TARGET_PAGE_SIZE
 378  * in ramblock, which starts from ramblock base address.
 379  */
 380 static uint32_t get_ramblock_vfn_hash(struct RamblockDirtyInfo *info,
 381                                       uint64_t vfn)
 382 {
 383     uint32_t hash;
 384
 385     hash = compute_page_hash(info->ramblock_addr +
 386                              vfn * qemu_target_page_size());
 387
 388     trace_get_ramblock_vfn_hash(info->idstr, vfn, hash);
 389     return hash;
 390 }
 391
 392 static bool save_ramblock_hash(struct RamblockDirtyInfo *info)
 393 {
 394     unsigned int sample_pages_count;
 395     int i;
 396     GRand *rand;
 397
 398     sample_pages_count = info->sample_pages_count;
 399
 400     /* ramblock size less than one page, return success to skip this ramblock */
 401     if (unlikely(info->ramblock_pages == 0 || sample_pages_count == 0)) {
 402         return true;
 403     }
 404
 405     info->hash_result = g_try_malloc0_n(sample_pages_count,
 406                                         sizeof(uint32_t));
 407     if (!info->hash_result) {
 408         return false;
 409     }
 410
 411     info->sample_page_vfn = g_try_malloc0_n(sample_pages_count,
 412                                             sizeof(uint64_t));
 413     if (!info->sample_page_vfn) {
 414         g_free(info->hash_result);
 415         return false;
 416     }
 417
 418     rand  = g_rand_new();
 419     for (i = 0; i < sample_pages_count; i++) {
 420         info->sample_page_vfn[i] = g_rand_int_range(rand, 0,
 421                                                     info->ramblock_pages - 1);
 422         info->hash_result[i] = get_ramblock_vfn_hash(info,
 423                                                      info->sample_page_vfn[i]);
 424     }
 425     g_rand_free(rand);
 426
 427     return true;
 428 }
 429
 430 static void get_ramblock_dirty_info(RAMBlock *block,
 431                                     struct RamblockDirtyInfo *info,
 432                                     struct DirtyRateConfig *config)
 433 {
 434     uint64_t sample_pages_per_gigabytes = config->sample_pages_per_gigabytes;
 435
 436     /* Right shift 30 bits to calc ramblock size in GB */
 437     info->sample_pages_count = (qemu_ram_get_used_length(block) *
 438                                 sample_pages_per_gigabytes) >> 30;
 439     /* Right shift TARGET_PAGE_BITS to calc page count */
 440     info->ramblock_pages = qemu_ram_get_used_length(block) >>
 441                            qemu_target_page_bits();
 442     info->ramblock_addr = qemu_ram_get_host_addr(block);
 443     strcpy(info->idstr, qemu_ram_get_idstr(block));
 444 }
 445
 446 static void free_ramblock_dirty_info(struct RamblockDirtyInfo *infos, int count)
 447 {
 448     int i;
 449
 450     if (!infos) {
 451         return;
 452     }
 453
 454     for (i = 0; i < count; i++) {
 455         g_free(infos[i].sample_page_vfn);
 456         g_free(infos[i].hash_result);
 457     }
 458     g_free(infos);
 459 }
 460
 461 static bool skip_sample_ramblock(RAMBlock *block)
 462 {
 463     /*
 464      * Sample only blocks larger than MIN_RAMBLOCK_SIZE.
 465      */
 466     if (qemu_ram_get_used_length(block) < (MIN_RAMBLOCK_SIZE << 10)) {
 467         trace_skip_sample_ramblock(block->idstr,
 468                                    qemu_ram_get_used_length(block));
 469         return true;
 470     }
 471
 472     return false;
 473 }
 474
 475 static bool record_ramblock_hash_info(struct RamblockDirtyInfo **block_dinfo,
 476                                       struct DirtyRateConfig config,
 477                                       int *block_count)
 478 {
 479     struct RamblockDirtyInfo *info = NULL;
 480     struct RamblockDirtyInfo *dinfo = NULL;
 481     RAMBlock *block = NULL;
 482     int total_count = 0;
 483     int index = 0;
 484     bool ret = false;
 485
 486     RAMBLOCK_FOREACH_MIGRATABLE(block) {
 487         if (skip_sample_ramblock(block)) {
 488             continue;
 489         }
 490         total_count++;
 491     }
 492
 493     dinfo = g_try_malloc0_n(total_count, sizeof(struct RamblockDirtyInfo));
 494     if (dinfo == NULL) {
 495         goto out;
 496     }
 497
 498     RAMBLOCK_FOREACH_MIGRATABLE(block) {
 499         if (skip_sample_ramblock(block)) {
 500             continue;
 501         }
 502         if (index >= total_count) {
 503             break;
 504         }
 505         info = &dinfo[index];
 506         get_ramblock_dirty_info(block, info, &config);
 507         if (!save_ramblock_hash(info)) {
 508             goto out;
 509         }
 510         index++;
 511     }
 512     ret = true;
 513
 514 out:
 515     *block_count = index;
 516     *block_dinfo = dinfo;
 517     return ret;
 518 }
 519
 520 static void calc_page_dirty_rate(struct RamblockDirtyInfo *info)
 521 {
 522     uint32_t hash;
 523     int i;
 524
 525     for (i = 0; i < info->sample_pages_count; i++) {
 526         hash = get_ramblock_vfn_hash(info, info->sample_page_vfn[i]);
 527         if (hash != info->hash_result[i]) {
 528             trace_calc_page_dirty_rate(info->idstr, hash, info->hash_result[i]);
 529             info->sample_dirty_count++;
 530         }
 531     }
 532 }
 533
 534 static struct RamblockDirtyInfo *
 535 find_block_matched(RAMBlock *block, int count,
 536                   struct RamblockDirtyInfo *infos)
 537 {
 538     int i;
 539
 540     for (i = 0; i < count; i++) {
 541         if (!strcmp(infos[i].idstr, qemu_ram_get_idstr(block))) {
 542             break;
 543         }
 544     }
 545
 546     if (i == count) {
 547         return NULL;
 548     }
 549
 550     if (infos[i].ramblock_addr != qemu_ram_get_host_addr(block) ||
 551         infos[i].ramblock_pages !=
 552             (qemu_ram_get_used_length(block) >> qemu_target_page_bits())) {
 553         trace_find_page_matched(block->idstr);
 554         return NULL;
 555     }
 556
 557     return &infos[i];
 558 }
 559
 560 static bool compare_page_hash_info(struct RamblockDirtyInfo *info,
 561                                   int block_count)
 562 {
 563     struct RamblockDirtyInfo *block_dinfo = NULL;
 564     RAMBlock *block = NULL;
 565
 566     RAMBLOCK_FOREACH_MIGRATABLE(block) {
 567         if (skip_sample_ramblock(block)) {
 568             continue;
 569         }
 570         block_dinfo = find_block_matched(block, block_count, info);
 571         if (block_dinfo == NULL) {
 572             continue;
 573         }
 574         calc_page_dirty_rate(block_dinfo);
 575         update_dirtyrate_stat(block_dinfo);
 576     }
 577
 578     if (DirtyStat.page_sampling.total_sample_count == 0) {
 579         return false;
 580     }
 581
 582     return true;
 583 }
 584
 585 static inline void record_dirtypages_bitmap(DirtyPageRecord *dirty_pages,
 586                                             bool start)
 587 {
 588     if (start) {
 589         dirty_pages->start_pages = total_dirty_pages;
 590     } else {
 591         dirty_pages->end_pages = total_dirty_pages;
 592     }
 593 }
 594
 595 static inline void dirtyrate_manual_reset_protect(void)
 596 {
 597     RAMBlock *block = NULL;
 598
 599     WITH_RCU_READ_LOCK_GUARD() {
 600         RAMBLOCK_FOREACH_MIGRATABLE(block) {
 601             memory_region_clear_dirty_bitmap(block->mr, 0,
 602                                              block->used_length);
 603         }
 604     }
 605 }
 606
 607 static void calculate_dirtyrate_dirty_bitmap(struct DirtyRateConfig config)
 608 {
 609     int64_t start_time;
 610     DirtyPageRecord dirty_pages;
 611
 612     bql_lock();
 613     memory_global_dirty_log_start(GLOBAL_DIRTY_DIRTY_RATE);
 614
 615     /*
 616      * 1'round of log sync may return all 1 bits with
 617      * KVM_DIRTY_LOG_INITIALLY_SET enable
 618      * skip it unconditionally and start dirty tracking
 619      * from 2'round of log sync
 620      */
 621     memory_global_dirty_log_sync(false);
 622
 623     /*
 624      * reset page protect manually and unconditionally.
 625      * this make sure kvm dirty log be cleared if
 626      * KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE cap is enabled.
 627      */
 628     dirtyrate_manual_reset_protect();
 629     bql_unlock();
 630
 631     record_dirtypages_bitmap(&dirty_pages, true);
 632
 633     start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 634     DirtyStat.start_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) / 1000;
 635
 636     DirtyStat.calc_time_ms = dirty_stat_wait(config.calc_time_ms, start_time);
 637
 638     /*
 639      * do two things.
 640      * 1. fetch dirty bitmap from kvm
 641      * 2. stop dirty tracking
 642      */
 643     global_dirty_log_sync(GLOBAL_DIRTY_DIRTY_RATE, true);
 644
 645     record_dirtypages_bitmap(&dirty_pages, false);
 646
 647     DirtyStat.dirty_rate = do_calculate_dirtyrate(dirty_pages,
 648                                                   DirtyStat.calc_time_ms);
 649 }
 650
 651 static void calculate_dirtyrate_dirty_ring(struct DirtyRateConfig config)
 652 {
 653     uint64_t dirtyrate = 0;
 654     uint64_t dirtyrate_sum = 0;
 655     int i = 0;
 656
 657     /* start log sync */
 658     global_dirty_log_change(GLOBAL_DIRTY_DIRTY_RATE, true);
 659
 660     DirtyStat.start_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) / 1000;
 661
 662     /* calculate vcpu dirtyrate */
 663     DirtyStat.calc_time_ms = vcpu_calculate_dirtyrate(config.calc_time_ms,
 664                                                       &DirtyStat.dirty_ring,
 665                                                       GLOBAL_DIRTY_DIRTY_RATE,
 666                                                       true);
 667
 668     /* calculate vm dirtyrate */
 669     for (i = 0; i < DirtyStat.dirty_ring.nvcpu; i++) {
 670         dirtyrate = DirtyStat.dirty_ring.rates[i].dirty_rate;
 671         DirtyStat.dirty_ring.rates[i].dirty_rate = dirtyrate;
 672         dirtyrate_sum += dirtyrate;
 673     }
 674
 675     DirtyStat.dirty_rate = dirtyrate_sum;
 676 }
 677
 678 static void calculate_dirtyrate_sample_vm(struct DirtyRateConfig config)
 679 {
 680     struct RamblockDirtyInfo *block_dinfo = NULL;
 681     int block_count = 0;
 682     int64_t initial_time;
 683
 684     rcu_read_lock();
 685     initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 686     DirtyStat.start_time = qemu_clock_get_ms(QEMU_CLOCK_HOST) / 1000;
 687     if (!record_ramblock_hash_info(&block_dinfo, config, &block_count)) {
 688         goto out;
 689     }
 690     rcu_read_unlock();
 691
 692     DirtyStat.calc_time_ms = dirty_stat_wait(config.calc_time_ms,
 693                                              initial_time);
 694
 695     rcu_read_lock();
 696     if (!compare_page_hash_info(block_dinfo, block_count)) {
 697         goto out;
 698     }
 699
 700     update_dirtyrate(DirtyStat.calc_time_ms);
 701
 702 out:
 703     rcu_read_unlock();
 704     free_ramblock_dirty_info(block_dinfo, block_count);
 705 }
 706
 707 static void calculate_dirtyrate(struct DirtyRateConfig config)
 708 {
 709     if (config.mode == DIRTY_RATE_MEASURE_MODE_DIRTY_BITMAP) {
 710         calculate_dirtyrate_dirty_bitmap(config);
 711     } else if (config.mode == DIRTY_RATE_MEASURE_MODE_DIRTY_RING) {
 712         calculate_dirtyrate_dirty_ring(config);
 713     } else {
 714         calculate_dirtyrate_sample_vm(config);
 715     }
 716
 717     trace_dirtyrate_calculate(DirtyStat.dirty_rate);
 718 }
 719
 720 void *get_dirtyrate_thread(void *arg)
 721 {
 722     struct DirtyRateConfig config = *(struct DirtyRateConfig *)arg;
 723     int ret;
 724     rcu_register_thread();
 725
 726     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_UNSTARTED,
 727                               DIRTY_RATE_STATUS_MEASURING);
 728     if (ret == -1) {
 729         error_report("change dirtyrate state failed.");
 730         return NULL;
 731     }
 732
 733     calculate_dirtyrate(config);
 734
 735     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_MEASURING,
 736                               DIRTY_RATE_STATUS_MEASURED);
 737     if (ret == -1) {
 738         error_report("change dirtyrate state failed.");
 739     }
 740
 741     rcu_unregister_thread();
 742     return NULL;
 743 }
 744
 745 void qmp_calc_dirty_rate(int64_t calc_time,
 746                          bool has_calc_time_unit,
 747                          TimeUnit calc_time_unit,
 748                          bool has_sample_pages,
 749                          int64_t sample_pages,
 750                          bool has_mode,
 751                          DirtyRateMeasureMode mode,
 752                          Error **errp)
 753 {
 754     static struct DirtyRateConfig config;
 755     QemuThread thread;
 756     int ret;
 757
 758     /*
 759      * If the dirty rate is already being measured, don't attempt to start.
 760      */
 761     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURING) {
 762         error_setg(errp, "the dirty rate is already being measured.");
 763         return;
 764     }
 765
 766     int64_t calc_time_ms = convert_time_unit(
 767         calc_time,
 768         has_calc_time_unit ? calc_time_unit : TIME_UNIT_SECOND,
 769         TIME_UNIT_MILLISECOND
 770     );
 771
 772     if (!is_calc_time_valid(calc_time_ms)) {
 773         error_setg(errp, "Calculation time is out of range [%dms, %dms].",
 774                          MIN_CALC_TIME_MS, MAX_CALC_TIME_MS);
 775         return;
 776     }
 777
 778     if (!has_mode) {
 779         mode =  DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING;
 780     }
 781
 782     if (has_sample_pages && mode != DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING) {
 783         error_setg(errp, "sample-pages is used only in page-sampling mode");
 784         return;
 785     }
 786
 787     if (has_sample_pages) {
 788         if (!is_sample_pages_valid(sample_pages)) {
 789             error_setg(errp, "sample-pages is out of range[%d, %d].",
 790                             MIN_SAMPLE_PAGE_COUNT,
 791                             MAX_SAMPLE_PAGE_COUNT);
 792             return;
 793         }
 794     } else {
 795         sample_pages = DIRTYRATE_DEFAULT_SAMPLE_PAGES;
 796     }
 797
 798     /*
 799      * dirty ring mode only works when kvm dirty ring is enabled.
 800      * on the contrary, dirty bitmap mode is not.
 801      */
 802     if (((mode == DIRTY_RATE_MEASURE_MODE_DIRTY_RING) &&
 803         !kvm_dirty_ring_enabled()) ||
 804         ((mode == DIRTY_RATE_MEASURE_MODE_DIRTY_BITMAP) &&
 805          kvm_dirty_ring_enabled())) {
 806         error_setg(errp, "mode %s is not enabled, use other method instead.",
 807                          DirtyRateMeasureMode_str(mode));
 808          return;
 809     }
 810
 811     /*
 812      * Init calculation state as unstarted.
 813      */
 814     ret = dirtyrate_set_state(&CalculatingState, CalculatingState,
 815                               DIRTY_RATE_STATUS_UNSTARTED);
 816     if (ret == -1) {
 817         error_setg(errp, "init dirty rate calculation state failed.");
 818         return;
 819     }
 820
 821     config.calc_time_ms = calc_time_ms;
 822     config.sample_pages_per_gigabytes = sample_pages;
 823     config.mode = mode;
 824
 825     cleanup_dirtyrate_stat(config);
 826
 827     /*
 828      * update dirty rate mode so that we can figure out what mode has
 829      * been used in last calculation
 830      **/
 831     dirtyrate_mode = mode;
 832
 833     init_dirtyrate_stat(config);
 834
 835     qemu_thread_create(&thread, "get_dirtyrate", get_dirtyrate_thread,
 836                        (void *)&config, QEMU_THREAD_DETACHED);
 837 }
 838
 839
 840 struct DirtyRateInfo *qmp_query_dirty_rate(bool has_calc_time_unit,
 841                                            TimeUnit calc_time_unit,
 842                                            Error **errp)
 843 {
 844     return query_dirty_rate_info(
 845         has_calc_time_unit ? calc_time_unit : TIME_UNIT_SECOND);
 846 }
 847
 848 void hmp_info_dirty_rate(Monitor *mon, const QDict *qdict)
 849 {
 850     DirtyRateInfo *info = query_dirty_rate_info(TIME_UNIT_SECOND);
 851
 852     monitor_printf(mon, "Status: %s\n",
 853                    DirtyRateStatus_str(info->status));
 854     monitor_printf(mon, "Start Time: %"PRIi64" (ms)\n",
 855                    info->start_time);
 856     if (info->mode == DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING) {
 857         monitor_printf(mon, "Sample Pages: %"PRIu64" (per GB)\n",
 858                        info->sample_pages);
 859     }
 860     monitor_printf(mon, "Period: %"PRIi64" (sec)\n",
 861                    info->calc_time);
 862     monitor_printf(mon, "Mode: %s\n",
 863                    DirtyRateMeasureMode_str(info->mode));
 864     monitor_printf(mon, "Dirty rate: ");
 865     if (info->has_dirty_rate) {
 866         monitor_printf(mon, "%"PRIi64" (MB/s)\n", info->dirty_rate);
 867         if (info->has_vcpu_dirty_rate) {
 868             DirtyRateVcpuList *rate, *head = info->vcpu_dirty_rate;
 869             for (rate = head; rate != NULL; rate = rate->next) {
 870                 monitor_printf(mon, "vcpu[%"PRIi64"], Dirty rate: %"PRIi64
 871                                " (MB/s)\n", rate->value->id,
 872                                rate->value->dirty_rate);
 873             }
 874         }
 875     } else {
 876         monitor_printf(mon, "(not ready)\n");
 877     }
 878
 879     qapi_free_DirtyRateVcpuList(info->vcpu_dirty_rate);
 880     g_free(info);
 881 }
 882
 883 void hmp_calc_dirty_rate(Monitor *mon, const QDict *qdict)
 884 {
 885     int64_t sec = qdict_get_try_int(qdict, "second", 0);
 886     int64_t sample_pages = qdict_get_try_int(qdict, "sample_pages_per_GB", -1);
 887     bool has_sample_pages = (sample_pages != -1);
 888     bool dirty_ring = qdict_get_try_bool(qdict, "dirty_ring", false);
 889     bool dirty_bitmap = qdict_get_try_bool(qdict, "dirty_bitmap", false);
 890     DirtyRateMeasureMode mode = DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING;
 891     Error *err = NULL;
 892
 893     if (!sec) {
 894         monitor_printf(mon, "Incorrect period length specified!\n");
 895         return;
 896     }
 897
 898     if (dirty_ring && dirty_bitmap) {
 899         monitor_printf(mon, "Either dirty ring or dirty bitmap "
 900                        "can be specified!\n");
 901         return;
 902     }
 903
 904     if (dirty_bitmap) {
 905         mode = DIRTY_RATE_MEASURE_MODE_DIRTY_BITMAP;
 906     } else if (dirty_ring) {
 907         mode = DIRTY_RATE_MEASURE_MODE_DIRTY_RING;
 908     }
 909
 910     qmp_calc_dirty_rate(sec, /* calc-time */
 911                         false, TIME_UNIT_SECOND, /* calc-time-unit */
 912                         has_sample_pages, sample_pages,
 913                         true, mode,
 914                         &err);
 915     if (err) {
 916         hmp_handle_error(mon, err);
 917         return;
 918     }
 919
 920     monitor_printf(mon, "Starting dirty rate measurement with period %"PRIi64
 921                    " seconds\n", sec);
 922     monitor_printf(mon, "[Please use 'info dirty_rate' to check results]\n");
 923 }