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 <zlib.h>
  15 #include "qapi/error.h"
  16 #include "cpu.h"
  17 #include "exec/ramblock.h"
  18 #include "qemu/rcu_queue.h"
  19 #include "qapi/qapi-commands-migration.h"
  20 #include "ram.h"
  21 #include "trace.h"
  22 #include "dirtyrate.h"
  23 #include "monitor/hmp.h"
  24 #include "monitor/monitor.h"
  25 #include "qapi/qmp/qdict.h"
  26
  27 static int CalculatingState = DIRTY_RATE_STATUS_UNSTARTED;
  28 static struct DirtyRateStat DirtyStat;
  29
  30 static int64_t set_sample_page_period(int64_t msec, int64_t initial_time)
  31 {
  32     int64_t current_time;
  33
  34     current_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
  35     if ((current_time - initial_time) >= msec) {
  36         msec = current_time - initial_time;
  37     } else {
  38         g_usleep((msec + initial_time - current_time) * 1000);
  39     }
  40
  41     return msec;
  42 }
  43
  44 static bool is_sample_period_valid(int64_t sec)
  45 {
  46     if (sec < MIN_FETCH_DIRTYRATE_TIME_SEC ||
  47         sec > MAX_FETCH_DIRTYRATE_TIME_SEC) {
  48         return false;
  49     }
  50
  51     return true;
  52 }
  53
  54 static bool is_sample_pages_valid(int64_t pages)
  55 {
  56     return pages >= MIN_SAMPLE_PAGE_COUNT &&
  57            pages <= MAX_SAMPLE_PAGE_COUNT;
  58 }
  59
  60 static int dirtyrate_set_state(int *state, int old_state, int new_state)
  61 {
  62     assert(new_state < DIRTY_RATE_STATUS__MAX);
  63     trace_dirtyrate_set_state(DirtyRateStatus_str(new_state));
  64     if (qatomic_cmpxchg(state, old_state, new_state) == old_state) {
  65         return 0;
  66     } else {
  67         return -1;
  68     }
  69 }
  70
  71 static struct DirtyRateInfo *query_dirty_rate_info(void)
  72 {
  73     int64_t dirty_rate = DirtyStat.dirty_rate;
  74     struct DirtyRateInfo *info = g_malloc0(sizeof(DirtyRateInfo));
  75
  76     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURED) {
  77         info->has_dirty_rate = true;
  78         info->dirty_rate = dirty_rate;
  79     }
  80
  81     info->status = CalculatingState;
  82     info->start_time = DirtyStat.start_time;
  83     info->calc_time = DirtyStat.calc_time;
  84     info->sample_pages = DirtyStat.sample_pages;
  85
  86     trace_query_dirty_rate_info(DirtyRateStatus_str(CalculatingState));
  87
  88     return info;
  89 }
  90
  91 static void init_dirtyrate_stat(int64_t start_time,
  92                                 struct DirtyRateConfig config)
  93 {
  94     DirtyStat.dirty_rate = -1;
  95     DirtyStat.start_time = start_time;
  96     DirtyStat.calc_time = config.sample_period_seconds;
  97     DirtyStat.sample_pages = config.sample_pages_per_gigabytes;
  98
  99     switch (config.mode) {
 100     case DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING:
 101         DirtyStat.page_sampling.total_dirty_samples = 0;
 102         DirtyStat.page_sampling.total_sample_count = 0;
 103         DirtyStat.page_sampling.total_block_mem_MB = 0;
 104         break;
 105     case DIRTY_RATE_MEASURE_MODE_DIRTY_RING:
 106         DirtyStat.dirty_ring.nvcpu = -1;
 107         DirtyStat.dirty_ring.rates = NULL;
 108         break;
 109     default:
 110         break;
 111     }
 112 }
 113
 114 static void update_dirtyrate_stat(struct RamblockDirtyInfo *info)
 115 {
 116     DirtyStat.page_sampling.total_dirty_samples += info->sample_dirty_count;
 117     DirtyStat.page_sampling.total_sample_count += info->sample_pages_count;
 118     /* size of total pages in MB */
 119     DirtyStat.page_sampling.total_block_mem_MB += (info->ramblock_pages *
 120                                                    TARGET_PAGE_SIZE) >> 20;
 121 }
 122
 123 static void update_dirtyrate(uint64_t msec)
 124 {
 125     uint64_t dirtyrate;
 126     uint64_t total_dirty_samples = DirtyStat.page_sampling.total_dirty_samples;
 127     uint64_t total_sample_count = DirtyStat.page_sampling.total_sample_count;
 128     uint64_t total_block_mem_MB = DirtyStat.page_sampling.total_block_mem_MB;
 129
 130     dirtyrate = total_dirty_samples * total_block_mem_MB *
 131                 1000 / (total_sample_count * msec);
 132
 133     DirtyStat.dirty_rate = dirtyrate;
 134 }
 135
 136 /*
 137  * get hash result for the sampled memory with length of TARGET_PAGE_SIZE
 138  * in ramblock, which starts from ramblock base address.
 139  */
 140 static uint32_t get_ramblock_vfn_hash(struct RamblockDirtyInfo *info,
 141                                       uint64_t vfn)
 142 {
 143     uint32_t crc;
 144
 145     crc = crc32(0, (info->ramblock_addr +
 146                 vfn * TARGET_PAGE_SIZE), TARGET_PAGE_SIZE);
 147
 148     trace_get_ramblock_vfn_hash(info->idstr, vfn, crc);
 149     return crc;
 150 }
 151
 152 static bool save_ramblock_hash(struct RamblockDirtyInfo *info)
 153 {
 154     unsigned int sample_pages_count;
 155     int i;
 156     GRand *rand;
 157
 158     sample_pages_count = info->sample_pages_count;
 159
 160     /* ramblock size less than one page, return success to skip this ramblock */
 161     if (unlikely(info->ramblock_pages == 0 || sample_pages_count == 0)) {
 162         return true;
 163     }
 164
 165     info->hash_result = g_try_malloc0_n(sample_pages_count,
 166                                         sizeof(uint32_t));
 167     if (!info->hash_result) {
 168         return false;
 169     }
 170
 171     info->sample_page_vfn = g_try_malloc0_n(sample_pages_count,
 172                                             sizeof(uint64_t));
 173     if (!info->sample_page_vfn) {
 174         g_free(info->hash_result);
 175         return false;
 176     }
 177
 178     rand  = g_rand_new();
 179     for (i = 0; i < sample_pages_count; i++) {
 180         info->sample_page_vfn[i] = g_rand_int_range(rand, 0,
 181                                                     info->ramblock_pages - 1);
 182         info->hash_result[i] = get_ramblock_vfn_hash(info,
 183                                                      info->sample_page_vfn[i]);
 184     }
 185     g_rand_free(rand);
 186
 187     return true;
 188 }
 189
 190 static void get_ramblock_dirty_info(RAMBlock *block,
 191                                     struct RamblockDirtyInfo *info,
 192                                     struct DirtyRateConfig *config)
 193 {
 194     uint64_t sample_pages_per_gigabytes = config->sample_pages_per_gigabytes;
 195
 196     /* Right shift 30 bits to calc ramblock size in GB */
 197     info->sample_pages_count = (qemu_ram_get_used_length(block) *
 198                                 sample_pages_per_gigabytes) >> 30;
 199     /* Right shift TARGET_PAGE_BITS to calc page count */
 200     info->ramblock_pages = qemu_ram_get_used_length(block) >>
 201                            TARGET_PAGE_BITS;
 202     info->ramblock_addr = qemu_ram_get_host_addr(block);
 203     strcpy(info->idstr, qemu_ram_get_idstr(block));
 204 }
 205
 206 static void free_ramblock_dirty_info(struct RamblockDirtyInfo *infos, int count)
 207 {
 208     int i;
 209
 210     if (!infos) {
 211         return;
 212     }
 213
 214     for (i = 0; i < count; i++) {
 215         g_free(infos[i].sample_page_vfn);
 216         g_free(infos[i].hash_result);
 217     }
 218     g_free(infos);
 219 }
 220
 221 static bool skip_sample_ramblock(RAMBlock *block)
 222 {
 223     /*
 224      * Sample only blocks larger than MIN_RAMBLOCK_SIZE.
 225      */
 226     if (qemu_ram_get_used_length(block) < (MIN_RAMBLOCK_SIZE << 10)) {
 227         trace_skip_sample_ramblock(block->idstr,
 228                                    qemu_ram_get_used_length(block));
 229         return true;
 230     }
 231
 232     return false;
 233 }
 234
 235 static bool record_ramblock_hash_info(struct RamblockDirtyInfo **block_dinfo,
 236                                       struct DirtyRateConfig config,
 237                                       int *block_count)
 238 {
 239     struct RamblockDirtyInfo *info = NULL;
 240     struct RamblockDirtyInfo *dinfo = NULL;
 241     RAMBlock *block = NULL;
 242     int total_count = 0;
 243     int index = 0;
 244     bool ret = false;
 245
 246     RAMBLOCK_FOREACH_MIGRATABLE(block) {
 247         if (skip_sample_ramblock(block)) {
 248             continue;
 249         }
 250         total_count++;
 251     }
 252
 253     dinfo = g_try_malloc0_n(total_count, sizeof(struct RamblockDirtyInfo));
 254     if (dinfo == NULL) {
 255         goto out;
 256     }
 257
 258     RAMBLOCK_FOREACH_MIGRATABLE(block) {
 259         if (skip_sample_ramblock(block)) {
 260             continue;
 261         }
 262         if (index >= total_count) {
 263             break;
 264         }
 265         info = &dinfo[index];
 266         get_ramblock_dirty_info(block, info, &config);
 267         if (!save_ramblock_hash(info)) {
 268             goto out;
 269         }
 270         index++;
 271     }
 272     ret = true;
 273
 274 out:
 275     *block_count = index;
 276     *block_dinfo = dinfo;
 277     return ret;
 278 }
 279
 280 static void calc_page_dirty_rate(struct RamblockDirtyInfo *info)
 281 {
 282     uint32_t crc;
 283     int i;
 284
 285     for (i = 0; i < info->sample_pages_count; i++) {
 286         crc = get_ramblock_vfn_hash(info, info->sample_page_vfn[i]);
 287         if (crc != info->hash_result[i]) {
 288             trace_calc_page_dirty_rate(info->idstr, crc, info->hash_result[i]);
 289             info->sample_dirty_count++;
 290         }
 291     }
 292 }
 293
 294 static struct RamblockDirtyInfo *
 295 find_block_matched(RAMBlock *block, int count,
 296                   struct RamblockDirtyInfo *infos)
 297 {
 298     int i;
 299     struct RamblockDirtyInfo *matched;
 300
 301     for (i = 0; i < count; i++) {
 302         if (!strcmp(infos[i].idstr, qemu_ram_get_idstr(block))) {
 303             break;
 304         }
 305     }
 306
 307     if (i == count) {
 308         return NULL;
 309     }
 310
 311     if (infos[i].ramblock_addr != qemu_ram_get_host_addr(block) ||
 312         infos[i].ramblock_pages !=
 313             (qemu_ram_get_used_length(block) >> TARGET_PAGE_BITS)) {
 314         trace_find_page_matched(block->idstr);
 315         return NULL;
 316     }
 317
 318     matched = &infos[i];
 319
 320     return matched;
 321 }
 322
 323 static bool compare_page_hash_info(struct RamblockDirtyInfo *info,
 324                                   int block_count)
 325 {
 326     struct RamblockDirtyInfo *block_dinfo = NULL;
 327     RAMBlock *block = NULL;
 328
 329     RAMBLOCK_FOREACH_MIGRATABLE(block) {
 330         if (skip_sample_ramblock(block)) {
 331             continue;
 332         }
 333         block_dinfo = find_block_matched(block, block_count, info);
 334         if (block_dinfo == NULL) {
 335             continue;
 336         }
 337         calc_page_dirty_rate(block_dinfo);
 338         update_dirtyrate_stat(block_dinfo);
 339     }
 340
 341     if (DirtyStat.page_sampling.total_sample_count == 0) {
 342         return false;
 343     }
 344
 345     return true;
 346 }
 347
 348 static void calculate_dirtyrate(struct DirtyRateConfig config)
 349 {
 350     struct RamblockDirtyInfo *block_dinfo = NULL;
 351     int block_count = 0;
 352     int64_t msec = 0;
 353     int64_t initial_time;
 354
 355     rcu_read_lock();
 356     initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 357     if (!record_ramblock_hash_info(&block_dinfo, config, &block_count)) {
 358         goto out;
 359     }
 360     rcu_read_unlock();
 361
 362     msec = config.sample_period_seconds * 1000;
 363     msec = set_sample_page_period(msec, initial_time);
 364     DirtyStat.start_time = initial_time / 1000;
 365     DirtyStat.calc_time = msec / 1000;
 366
 367     rcu_read_lock();
 368     if (!compare_page_hash_info(block_dinfo, block_count)) {
 369         goto out;
 370     }
 371
 372     update_dirtyrate(msec);
 373
 374 out:
 375     rcu_read_unlock();
 376     free_ramblock_dirty_info(block_dinfo, block_count);
 377 }
 378
 379 void *get_dirtyrate_thread(void *arg)
 380 {
 381     struct DirtyRateConfig config = *(struct DirtyRateConfig *)arg;
 382     int ret;
 383     int64_t start_time;
 384     rcu_register_thread();
 385
 386     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_UNSTARTED,
 387                               DIRTY_RATE_STATUS_MEASURING);
 388     if (ret == -1) {
 389         error_report("change dirtyrate state failed.");
 390         return NULL;
 391     }
 392
 393     start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) / 1000;
 394     init_dirtyrate_stat(start_time, config);
 395
 396     calculate_dirtyrate(config);
 397
 398     ret = dirtyrate_set_state(&CalculatingState, DIRTY_RATE_STATUS_MEASURING,
 399                               DIRTY_RATE_STATUS_MEASURED);
 400     if (ret == -1) {
 401         error_report("change dirtyrate state failed.");
 402     }
 403
 404     rcu_unregister_thread();
 405     return NULL;
 406 }
 407
 408 void qmp_calc_dirty_rate(int64_t calc_time, bool has_sample_pages,
 409                          int64_t sample_pages, Error **errp)
 410 {
 411     static struct DirtyRateConfig config;
 412     QemuThread thread;
 413     int ret;
 414
 415     /*
 416      * If the dirty rate is already being measured, don't attempt to start.
 417      */
 418     if (qatomic_read(&CalculatingState) == DIRTY_RATE_STATUS_MEASURING) {
 419         error_setg(errp, "the dirty rate is already being measured.");
 420         return;
 421     }
 422
 423     if (!is_sample_period_valid(calc_time)) {
 424         error_setg(errp, "calc-time is out of range[%d, %d].",
 425                          MIN_FETCH_DIRTYRATE_TIME_SEC,
 426                          MAX_FETCH_DIRTYRATE_TIME_SEC);
 427         return;
 428     }
 429
 430     if (has_sample_pages) {
 431         if (!is_sample_pages_valid(sample_pages)) {
 432             error_setg(errp, "sample-pages is out of range[%d, %d].",
 433                             MIN_SAMPLE_PAGE_COUNT,
 434                             MAX_SAMPLE_PAGE_COUNT);
 435             return;
 436         }
 437     } else {
 438         sample_pages = DIRTYRATE_DEFAULT_SAMPLE_PAGES;
 439     }
 440
 441     /*
 442      * Init calculation state as unstarted.
 443      */
 444     ret = dirtyrate_set_state(&CalculatingState, CalculatingState,
 445                               DIRTY_RATE_STATUS_UNSTARTED);
 446     if (ret == -1) {
 447         error_setg(errp, "init dirty rate calculation state failed.");
 448         return;
 449     }
 450
 451     config.sample_period_seconds = calc_time;
 452     config.sample_pages_per_gigabytes = sample_pages;
 453     config.mode = DIRTY_RATE_MEASURE_MODE_PAGE_SAMPLING;
 454     qemu_thread_create(&thread, "get_dirtyrate", get_dirtyrate_thread,
 455                        (void *)&config, QEMU_THREAD_DETACHED);
 456 }
 457
 458 struct DirtyRateInfo *qmp_query_dirty_rate(Error **errp)
 459 {
 460     return query_dirty_rate_info();
 461 }
 462
 463 void hmp_info_dirty_rate(Monitor *mon, const QDict *qdict)
 464 {
 465     DirtyRateInfo *info = query_dirty_rate_info();
 466
 467     monitor_printf(mon, "Status: %s\n",
 468                    DirtyRateStatus_str(info->status));
 469     monitor_printf(mon, "Start Time: %"PRIi64" (ms)\n",
 470                    info->start_time);
 471     monitor_printf(mon, "Sample Pages: %"PRIu64" (per GB)\n",
 472                    info->sample_pages);
 473     monitor_printf(mon, "Period: %"PRIi64" (sec)\n",
 474                    info->calc_time);
 475     monitor_printf(mon, "Dirty rate: ");
 476     if (info->has_dirty_rate) {
 477         monitor_printf(mon, "%"PRIi64" (MB/s)\n", info->dirty_rate);
 478     } else {
 479         monitor_printf(mon, "(not ready)\n");
 480     }
 481     g_free(info);
 482 }
 483
 484 void hmp_calc_dirty_rate(Monitor *mon, const QDict *qdict)
 485 {
 486     int64_t sec = qdict_get_try_int(qdict, "second", 0);
 487     int64_t sample_pages = qdict_get_try_int(qdict, "sample_pages_per_GB", -1);
 488     bool has_sample_pages = (sample_pages != -1);
 489     Error *err = NULL;
 490
 491     if (!sec) {
 492         monitor_printf(mon, "Incorrect period length specified!\n");
 493         return;
 494     }
 495
 496     qmp_calc_dirty_rate(sec, has_sample_pages, sample_pages, &err);
 497     if (err) {
 498         hmp_handle_error(mon, err);
 499         return;
 500     }
 501
 502     monitor_printf(mon, "Starting dirty rate measurement with period %"PRIi64
 503                    " seconds\n", sec);
 504     monitor_printf(mon, "[Please use 'info dirty_rate' to check results]\n");
 505 }