debian/patches/pve/0017-PVE-add-savevm-async-for-background-state-snapshots.patch

   1 From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
   2 From: Dietmar Maurer <dietmar@proxmox.com>
   3 Date: Mon, 6 Apr 2020 12:16:46 +0200
   4 Subject: [PATCH] PVE: add savevm-async for background state snapshots
   5
   6 Put qemu_savevm_state_{header,setup} into the main loop and the rest
   7 of the iteration into a coroutine. The former need to lock the
   8 iothread (and we can't unlock it in the coroutine), and the latter
   9 can't deal with being in a separate thread, so a coroutine it must
  10 be.
  11
  12 Truncate output file at 1024 boundary.
  13
  14 Do not block the VM and save the state on aborting a snapshot, as the
  15 snapshot will be invalid anyway.
  16
  17 Also, when aborting, wait for the target file to be closed, otherwise a
  18 client might run into race-conditions when trying to remove the file
  19 still opened by QEMU.
  20
  21 Signed-off-by: Thomas Lamprecht <t.lamprecht@proxmox.com>
  22 Signed-off-by: Dietmar Maurer <dietmar@proxmox.com>
  23 Signed-off-by: Wolfgang Bumiller <w.bumiller@proxmox.com>
  24 [SR: improve aborting
  25      register yank before migration_incoming_state_destroy]
  26 Signed-off-by: Stefan Reiter <s.reiter@proxmox.com>
  27 [FE: further improve aborting
  28      adapt to removal of QEMUFileOps
  29      improve condition for entering final stage
  30      adapt to QAPI and other changes for 8.0]
  31 Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
  32 ---
  33  hmp-commands-info.hx         |  13 +
  34  hmp-commands.hx              |  17 ++
  35  include/migration/snapshot.h |   2 +
  36  include/monitor/hmp.h        |   3 +
  37  migration/meson.build        |   1 +
  38  migration/savevm-async.c     | 533 +++++++++++++++++++++++++++++++++++
  39  monitor/hmp-cmds.c           |  38 +++
  40  qapi/migration.json          |  34 +++
  41  qapi/misc.json               |  16 ++
  42  qemu-options.hx              |  12 +
  43  softmmu/vl.c                 |  10 +
  44  11 files changed, 679 insertions(+)
  45  create mode 100644 migration/savevm-async.c
  46
  47 diff --git a/hmp-commands-info.hx b/hmp-commands-info.hx
  48 index 47d63d26db..a166bff3d5 100644
  49 --- a/hmp-commands-info.hx
  50 +++ b/hmp-commands-info.hx
  51 @@ -540,6 +540,19 @@ SRST
  52      Show current migration parameters.
  53  ERST
  54
  55 +    {
  56 +        .name       = "savevm",
  57 +        .args_type  = "",
  58 +        .params     = "",
  59 +        .help       = "show savevm status",
  60 +        .cmd = hmp_info_savevm,
  61 +    },
  62 +
  63 +SRST
  64 +  ``info savevm``
  65 +    Show savevm status.
  66 +ERST
  67 +
  68      {
  69          .name       = "balloon",
  70          .args_type  = "",
  71 diff --git a/hmp-commands.hx b/hmp-commands.hx
  72 index bb85ee1d26..d9f9f42d11 100644
  73 --- a/hmp-commands.hx
  74 +++ b/hmp-commands.hx
  75 @@ -1846,3 +1846,20 @@ SRST
  76    List event channels in the guest
  77  ERST
  78  #endif
  79 +
  80 +    {
  81 +        .name       = "savevm-start",
  82 +        .args_type  = "statefile:s?",
  83 +        .params     = "[statefile]",
  84 +        .help       = "Prepare for snapshot and halt VM. Save VM state to statefile.",
  85 +        .cmd = hmp_savevm_start,
  86 +    },
  87 +
  88 +    {
  89 +        .name       = "savevm-end",
  90 +        .args_type  = "",
  91 +        .params     = "",
  92 +        .help       = "Resume VM after snaphot.",
  93 +        .cmd        = hmp_savevm_end,
  94 +        .coroutine  = true,
  95 +    },
  96 diff --git a/include/migration/snapshot.h b/include/migration/snapshot.h
  97 index e72083b117..c846d37806 100644
  98 --- a/include/migration/snapshot.h
  99 +++ b/include/migration/snapshot.h
 100 @@ -61,4 +61,6 @@ bool delete_snapshot(const char *name,
 101                      bool has_devices, strList *devices,
 102                      Error **errp);
 103
 104 +int load_snapshot_from_blockdev(const char *filename, Error **errp);
 105 +
 106  #endif
 107 diff --git a/include/monitor/hmp.h b/include/monitor/hmp.h
 108 index fdb69b7f9c..fdf6b45fb8 100644
 109 --- a/include/monitor/hmp.h
 110 +++ b/include/monitor/hmp.h
 111 @@ -28,6 +28,7 @@ void hmp_info_status(Monitor *mon, const QDict *qdict);
 112  void hmp_info_uuid(Monitor *mon, const QDict *qdict);
 113  void hmp_info_chardev(Monitor *mon, const QDict *qdict);
 114  void hmp_info_mice(Monitor *mon, const QDict *qdict);
 115 +void hmp_info_savevm(Monitor *mon, const QDict *qdict);
 116  void hmp_info_migrate(Monitor *mon, const QDict *qdict);
 117  void hmp_info_migrate_capabilities(Monitor *mon, const QDict *qdict);
 118  void hmp_info_migrate_parameters(Monitor *mon, const QDict *qdict);
 119 @@ -94,6 +95,8 @@ void hmp_closefd(Monitor *mon, const QDict *qdict);
 120  void hmp_mouse_move(Monitor *mon, const QDict *qdict);
 121  void hmp_mouse_button(Monitor *mon, const QDict *qdict);
 122  void hmp_mouse_set(Monitor *mon, const QDict *qdict);
 123 +void hmp_savevm_start(Monitor *mon, const QDict *qdict);
 124 +void hmp_savevm_end(Monitor *mon, const QDict *qdict);
 125  void hmp_sendkey(Monitor *mon, const QDict *qdict);
 126  void coroutine_fn hmp_screendump(Monitor *mon, const QDict *qdict);
 127  void hmp_chardev_add(Monitor *mon, const QDict *qdict);
 128 diff --git a/migration/meson.build b/migration/meson.build
 129 index 8a142fc7a9..a7824b5266 100644
 130 --- a/migration/meson.build
 131 +++ b/migration/meson.build
 132 @@ -25,6 +25,7 @@ softmmu_ss.add(files(
 133    'multifd-zlib.c',
 134    'postcopy-ram.c',
 135    'savevm.c',
 136 +  'savevm-async.c',
 137    'socket.c',
 138    'tls.c',
 139    'threadinfo.c',
 140 diff --git a/migration/savevm-async.c b/migration/savevm-async.c
 141 new file mode 100644
 142 index 0000000000..ac1fac6378
 143 --- /dev/null
 144 +++ b/migration/savevm-async.c
 145 @@ -0,0 +1,533 @@
 146 +#include "qemu/osdep.h"
 147 +#include "migration/channel-savevm-async.h"
 148 +#include "migration/migration.h"
 149 +#include "migration/savevm.h"
 150 +#include "migration/snapshot.h"
 151 +#include "migration/global_state.h"
 152 +#include "migration/ram.h"
 153 +#include "migration/qemu-file.h"
 154 +#include "sysemu/sysemu.h"
 155 +#include "sysemu/runstate.h"
 156 +#include "block/block.h"
 157 +#include "sysemu/block-backend.h"
 158 +#include "qapi/error.h"
 159 +#include "qapi/qmp/qerror.h"
 160 +#include "qapi/qmp/qdict.h"
 161 +#include "qapi/qapi-commands-migration.h"
 162 +#include "qapi/qapi-commands-misc.h"
 163 +#include "qapi/qapi-commands-block.h"
 164 +#include "qemu/cutils.h"
 165 +#include "qemu/timer.h"
 166 +#include "qemu/main-loop.h"
 167 +#include "qemu/rcu.h"
 168 +#include "qemu/yank.h"
 169 +
 170 +/* #define DEBUG_SAVEVM_STATE */
 171 +
 172 +#ifdef DEBUG_SAVEVM_STATE
 173 +#define DPRINTF(fmt, ...) \
 174 +    do { printf("savevm-async: " fmt, ## __VA_ARGS__); } while (0)
 175 +#else
 176 +#define DPRINTF(fmt, ...) \
 177 +    do { } while (0)
 178 +#endif
 179 +
 180 +enum {
 181 +    SAVE_STATE_DONE,
 182 +    SAVE_STATE_ERROR,
 183 +    SAVE_STATE_ACTIVE,
 184 +    SAVE_STATE_COMPLETED,
 185 +    SAVE_STATE_CANCELLED
 186 +};
 187 +
 188 +
 189 +static struct SnapshotState {
 190 +    BlockBackend *target;
 191 +    size_t bs_pos;
 192 +    int state;
 193 +    Error *error;
 194 +    Error *blocker;
 195 +    int saved_vm_running;
 196 +    QEMUFile *file;
 197 +    int64_t total_time;
 198 +    QEMUBH *finalize_bh;
 199 +    Coroutine *co;
 200 +    QemuCoSleep target_close_wait;
 201 +} snap_state;
 202 +
 203 +static bool savevm_aborted(void)
 204 +{
 205 +    return snap_state.state == SAVE_STATE_CANCELLED ||
 206 +        snap_state.state == SAVE_STATE_ERROR;
 207 +}
 208 +
 209 +SaveVMInfo *qmp_query_savevm(Error **errp)
 210 +{
 211 +    SaveVMInfo *info = g_malloc0(sizeof(*info));
 212 +    struct SnapshotState *s = &snap_state;
 213 +
 214 +    if (s->state != SAVE_STATE_DONE) {
 215 +        info->has_bytes = true;
 216 +        info->bytes = s->bs_pos;
 217 +        switch (s->state) {
 218 +        case SAVE_STATE_ERROR:
 219 +            info->status = g_strdup("failed");
 220 +            info->has_total_time = true;
 221 +            info->total_time = s->total_time;
 222 +            if (s->error) {
 223 +                info->error = g_strdup(error_get_pretty(s->error));
 224 +            }
 225 +            break;
 226 +        case SAVE_STATE_ACTIVE:
 227 +            info->status = g_strdup("active");
 228 +            info->has_total_time = true;
 229 +            info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 230 +                - s->total_time;
 231 +            break;
 232 +        case SAVE_STATE_COMPLETED:
 233 +            info->status = g_strdup("completed");
 234 +            info->has_total_time = true;
 235 +            info->total_time = s->total_time;
 236 +            break;
 237 +        }
 238 +    }
 239 +
 240 +    return info;
 241 +}
 242 +
 243 +static int save_snapshot_cleanup(void)
 244 +{
 245 +    int ret = 0;
 246 +
 247 +    DPRINTF("save_snapshot_cleanup\n");
 248 +
 249 +    snap_state.total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
 250 +        snap_state.total_time;
 251 +
 252 +    if (snap_state.file) {
 253 +        ret = qemu_fclose(snap_state.file);
 254 +        snap_state.file = NULL;
 255 +    }
 256 +
 257 +    if (snap_state.target) {
 258 +        if (!savevm_aborted()) {
 259 +            /* try to truncate, but ignore errors (will fail on block devices).
 260 +            * note1: bdrv_read() need whole blocks, so we need to round up
 261 +            * note2: PVE requires 1024 (BDRV_SECTOR_SIZE*2) alignment
 262 +            */
 263 +            size_t size = QEMU_ALIGN_UP(snap_state.bs_pos, BDRV_SECTOR_SIZE*2);
 264 +            blk_truncate(snap_state.target, size, false, PREALLOC_MODE_OFF, 0, NULL);
 265 +        }
 266 +        blk_op_unblock_all(snap_state.target, snap_state.blocker);
 267 +        error_free(snap_state.blocker);
 268 +        snap_state.blocker = NULL;
 269 +        blk_unref(snap_state.target);
 270 +        snap_state.target = NULL;
 271 +
 272 +        qemu_co_sleep_wake(&snap_state.target_close_wait);
 273 +    }
 274 +
 275 +    return ret;
 276 +}
 277 +
 278 +static void save_snapshot_error(const char *fmt, ...)
 279 +{
 280 +    va_list ap;
 281 +    char *msg;
 282 +
 283 +    va_start(ap, fmt);
 284 +    msg = g_strdup_vprintf(fmt, ap);
 285 +    va_end(ap);
 286 +
 287 +    DPRINTF("save_snapshot_error: %s\n", msg);
 288 +
 289 +    if (!snap_state.error) {
 290 +        error_set(&snap_state.error, ERROR_CLASS_GENERIC_ERROR, "%s", msg);
 291 +    }
 292 +
 293 +    g_free (msg);
 294 +
 295 +    snap_state.state = SAVE_STATE_ERROR;
 296 +}
 297 +
 298 +static void process_savevm_finalize(void *opaque)
 299 +{
 300 +    int ret;
 301 +    AioContext *iohandler_ctx = iohandler_get_aio_context();
 302 +    MigrationState *ms = migrate_get_current();
 303 +
 304 +    bool aborted = savevm_aborted();
 305 +
 306 +#ifdef DEBUG_SAVEVM_STATE
 307 +    int64_t start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 308 +#endif
 309 +
 310 +    qemu_bh_delete(snap_state.finalize_bh);
 311 +    snap_state.finalize_bh = NULL;
 312 +    snap_state.co = NULL;
 313 +
 314 +    /* We need to own the target bdrv's context for the following functions,
 315 +     * so move it back. It can stay in the main context and live out its live
 316 +     * there, since we're done with it after this method ends anyway.
 317 +     */
 318 +    aio_context_acquire(iohandler_ctx);
 319 +    blk_set_aio_context(snap_state.target, qemu_get_aio_context(), NULL);
 320 +    aio_context_release(iohandler_ctx);
 321 +
 322 +    ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 323 +    if (ret < 0) {
 324 +        save_snapshot_error("vm_stop_force_state error %d", ret);
 325 +    }
 326 +
 327 +    if (!aborted) {
 328 +        /* skip state saving if we aborted, snapshot will be invalid anyway */
 329 +        (void)qemu_savevm_state_complete_precopy(snap_state.file, false, false);
 330 +        ret = qemu_file_get_error(snap_state.file);
 331 +        if (ret < 0) {
 332 +            save_snapshot_error("qemu_savevm_state_complete_precopy error %d", ret);
 333 +        }
 334 +    }
 335 +
 336 +    DPRINTF("state saving complete\n");
 337 +    DPRINTF("timing: process_savevm_finalize (state saving) took %ld ms\n",
 338 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time);
 339 +
 340 +    /* clear migration state */
 341 +    migrate_set_state(&ms->state, MIGRATION_STATUS_SETUP,
 342 +        ret || aborted ? MIGRATION_STATUS_FAILED : MIGRATION_STATUS_COMPLETED);
 343 +    ms->to_dst_file = NULL;
 344 +
 345 +    qemu_savevm_state_cleanup();
 346 +
 347 +    ret = save_snapshot_cleanup();
 348 +    if (ret < 0) {
 349 +        save_snapshot_error("save_snapshot_cleanup error %d", ret);
 350 +    } else if (snap_state.state == SAVE_STATE_ACTIVE) {
 351 +        snap_state.state = SAVE_STATE_COMPLETED;
 352 +    } else if (aborted) {
 353 +        /*
 354 +         * If there was an error, there's no need to set a new one here.
 355 +         * If the snapshot was canceled, leave setting the state to
 356 +         * qmp_savevm_end(), which is waked by save_snapshot_cleanup().
 357 +         */
 358 +    } else {
 359 +        save_snapshot_error("process_savevm_cleanup: invalid state: %d",
 360 +                            snap_state.state);
 361 +    }
 362 +    if (snap_state.saved_vm_running) {
 363 +        vm_start();
 364 +        snap_state.saved_vm_running = false;
 365 +    }
 366 +
 367 +    DPRINTF("timing: process_savevm_finalize (full) took %ld ms\n",
 368 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time);
 369 +}
 370 +
 371 +static void coroutine_fn process_savevm_co(void *opaque)
 372 +{
 373 +    int ret;
 374 +    int64_t maxlen;
 375 +    BdrvNextIterator it;
 376 +    BlockDriverState *bs = NULL;
 377 +
 378 +#ifdef DEBUG_SAVEVM_STATE
 379 +    int64_t start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 380 +#endif
 381 +
 382 +    ret = qemu_file_get_error(snap_state.file);
 383 +    if (ret < 0) {
 384 +        save_snapshot_error("qemu_savevm_state_setup failed");
 385 +        return;
 386 +    }
 387 +
 388 +    while (snap_state.state == SAVE_STATE_ACTIVE) {
 389 +        uint64_t pending_size, pend_precopy, pend_postcopy;
 390 +        uint64_t threshold = 400 * 1000;
 391 +
 392 +        /*
 393 +         * pending_{estimate,exact} are expected to be called without iothread
 394 +         * lock. Similar to what is done in migration.c, call the exact variant
 395 +         * only once pend_precopy in the estimate is below the threshold.
 396 +         */
 397 +        qemu_mutex_unlock_iothread();
 398 +        qemu_savevm_state_pending_estimate(&pend_precopy, &pend_postcopy);
 399 +        if (pend_precopy <= threshold) {
 400 +            qemu_savevm_state_pending_exact(&pend_precopy, &pend_postcopy);
 401 +        }
 402 +        qemu_mutex_lock_iothread();
 403 +        pending_size = pend_precopy + pend_postcopy;
 404 +
 405 +        /*
 406 +         * A guest reaching this cutoff is dirtying lots of RAM. It should be
 407 +         * large enough so that the guest can't dirty this much between the
 408 +         * check and the guest actually being stopped, but it should be small
 409 +         * enough to avoid long downtimes for non-hibernation snapshots.
 410 +         */
 411 +        maxlen = blk_getlength(snap_state.target) - 100*1024*1024;
 412 +
 413 +        /* Note that there is no progress for pend_postcopy when iterating */
 414 +        if (pend_precopy > threshold && snap_state.bs_pos + pending_size < maxlen) {
 415 +            ret = qemu_savevm_state_iterate(snap_state.file, false);
 416 +            if (ret < 0) {
 417 +                save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
 418 +                break;
 419 +            }
 420 +            DPRINTF("savevm iterate pending size %lu ret %d\n", pending_size, ret);
 421 +        } else {
 422 +            qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER, NULL);
 423 +            ret = global_state_store();
 424 +            if (ret) {
 425 +                save_snapshot_error("global_state_store error %d", ret);
 426 +                break;
 427 +            }
 428 +
 429 +            DPRINTF("savevm iterate complete\n");
 430 +            break;
 431 +        }
 432 +    }
 433 +
 434 +    DPRINTF("timing: process_savevm_co took %ld ms\n",
 435 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time);
 436 +
 437 +#ifdef DEBUG_SAVEVM_STATE
 438 +    int64_t start_time_flush = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 439 +#endif
 440 +    /* If a drive runs in an IOThread we can flush it async, and only
 441 +     * need to sync-flush whatever IO happens between now and
 442 +     * vm_stop_force_state. bdrv_next can only be called from main AioContext,
 443 +     * so move there now and after every flush.
 444 +     */
 445 +    aio_co_reschedule_self(qemu_get_aio_context());
 446 +    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 447 +        /* target has BDRV_O_NO_FLUSH, no sense calling bdrv_flush on it */
 448 +        if (bs == blk_bs(snap_state.target)) {
 449 +            continue;
 450 +        }
 451 +
 452 +        AioContext *bs_ctx = bdrv_get_aio_context(bs);
 453 +        if (bs_ctx != qemu_get_aio_context()) {
 454 +            DPRINTF("savevm: async flushing drive %s\n", bs->filename);
 455 +            aio_co_reschedule_self(bs_ctx);
 456 +            bdrv_graph_co_rdlock();
 457 +            bdrv_flush(bs);
 458 +            bdrv_graph_co_rdunlock();
 459 +            aio_co_reschedule_self(qemu_get_aio_context());
 460 +        }
 461 +    }
 462 +
 463 +    DPRINTF("timing: async flushing took %ld ms\n",
 464 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time_flush);
 465 +
 466 +    qemu_bh_schedule(snap_state.finalize_bh);
 467 +}
 468 +
 469 +void qmp_savevm_start(const char *statefile, Error **errp)
 470 +{
 471 +    Error *local_err = NULL;
 472 +    MigrationState *ms = migrate_get_current();
 473 +    AioContext *iohandler_ctx = iohandler_get_aio_context();
 474 +
 475 +    int bdrv_oflags = BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_NO_FLUSH;
 476 +
 477 +    if (snap_state.state != SAVE_STATE_DONE) {
 478 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR,
 479 +                  "VM snapshot already started\n");
 480 +        return;
 481 +    }
 482 +
 483 +    if (migration_is_running(ms->state)) {
 484 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR, QERR_MIGRATION_ACTIVE);
 485 +        return;
 486 +    }
 487 +
 488 +    if (migrate_use_block()) {
 489 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR,
 490 +                  "Block migration and snapshots are incompatible");
 491 +        return;
 492 +    }
 493 +
 494 +    /* initialize snapshot info */
 495 +    snap_state.saved_vm_running = runstate_is_running();
 496 +    snap_state.bs_pos = 0;
 497 +    snap_state.total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 498 +    snap_state.blocker = NULL;
 499 +    snap_state.target_close_wait = (QemuCoSleep){ .to_wake = NULL };
 500 +
 501 +    if (snap_state.error) {
 502 +        error_free(snap_state.error);
 503 +        snap_state.error = NULL;
 504 +    }
 505 +
 506 +    if (!statefile) {
 507 +        vm_stop(RUN_STATE_SAVE_VM);
 508 +        snap_state.state = SAVE_STATE_COMPLETED;
 509 +        return;
 510 +    }
 511 +
 512 +    if (qemu_savevm_state_blocked(errp)) {
 513 +        return;
 514 +    }
 515 +
 516 +    /* Open the image */
 517 +    QDict *options = NULL;
 518 +    options = qdict_new();
 519 +    qdict_put_str(options, "driver", "raw");
 520 +    snap_state.target = blk_new_open(statefile, NULL, options, bdrv_oflags, &local_err);
 521 +    if (!snap_state.target) {
 522 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR, "failed to open '%s'", statefile);
 523 +        goto restart;
 524 +    }
 525 +
 526 +    QIOChannel *ioc = QIO_CHANNEL(qio_channel_savevm_async_new(snap_state.target,
 527 +                                                               &snap_state.bs_pos));
 528 +    snap_state.file = qemu_file_new_output(ioc);
 529 +
 530 +    if (!snap_state.file) {
 531 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR, "failed to open '%s'", statefile);
 532 +        goto restart;
 533 +    }
 534 +
 535 +    /*
 536 +     * qemu_savevm_* paths use migration code and expect a migration state.
 537 +     * State is cleared in process_savevm_co, but has to be initialized
 538 +     * here (blocking main thread, from QMP) to avoid race conditions.
 539 +     */
 540 +    migrate_init(ms);
 541 +    memset(&ram_counters, 0, sizeof(ram_counters));
 542 +    memset(&compression_counters, 0, sizeof(compression_counters));
 543 +    ms->to_dst_file = snap_state.file;
 544 +
 545 +    error_setg(&snap_state.blocker, "block device is in use by savevm");
 546 +    blk_op_block_all(snap_state.target, snap_state.blocker);
 547 +
 548 +    snap_state.state = SAVE_STATE_ACTIVE;
 549 +    snap_state.finalize_bh = qemu_bh_new(process_savevm_finalize, &snap_state);
 550 +    snap_state.co = qemu_coroutine_create(&process_savevm_co, NULL);
 551 +    qemu_mutex_unlock_iothread();
 552 +    qemu_savevm_state_header(snap_state.file);
 553 +    qemu_savevm_state_setup(snap_state.file);
 554 +    qemu_mutex_lock_iothread();
 555 +
 556 +    /* Async processing from here on out happens in iohandler context, so let
 557 +     * the target bdrv have its home there.
 558 +     */
 559 +    blk_set_aio_context(snap_state.target, iohandler_ctx, &local_err);
 560 +
 561 +    aio_co_schedule(iohandler_ctx, snap_state.co);
 562 +
 563 +    return;
 564 +
 565 +restart:
 566 +
 567 +    save_snapshot_error("setup failed");
 568 +
 569 +    if (snap_state.saved_vm_running) {
 570 +        vm_start();
 571 +        snap_state.saved_vm_running = false;
 572 +    }
 573 +}
 574 +
 575 +void coroutine_fn qmp_savevm_end(Error **errp)
 576 +{
 577 +    int64_t timeout;
 578 +
 579 +    if (snap_state.state == SAVE_STATE_DONE) {
 580 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR,
 581 +                  "VM snapshot not started\n");
 582 +        return;
 583 +    }
 584 +
 585 +    if (snap_state.state == SAVE_STATE_ACTIVE) {
 586 +        snap_state.state = SAVE_STATE_CANCELLED;
 587 +        goto wait_for_close;
 588 +    }
 589 +
 590 +    if (snap_state.saved_vm_running) {
 591 +        vm_start();
 592 +        snap_state.saved_vm_running = false;
 593 +    }
 594 +
 595 +    snap_state.state = SAVE_STATE_DONE;
 596 +
 597 +wait_for_close:
 598 +    if (!snap_state.target) {
 599 +        DPRINTF("savevm-end: no target file open\n");
 600 +        return;
 601 +    }
 602 +
 603 +    /* wait until cleanup is done before returning, this ensures that after this
 604 +     * call exits the statefile will be closed and can be removed immediately */
 605 +    DPRINTF("savevm-end: waiting for cleanup\n");
 606 +    timeout = 30L * 1000 * 1000 * 1000;
 607 +    qemu_co_sleep_ns_wakeable(&snap_state.target_close_wait,
 608 +                              QEMU_CLOCK_REALTIME, timeout);
 609 +    if (snap_state.target) {
 610 +        save_snapshot_error("timeout waiting for target file close in "
 611 +                            "qmp_savevm_end");
 612 +        /* we cannot assume the snapshot finished in this case, so leave the
 613 +         * state alone - caller has to figure something out */
 614 +        return;
 615 +    }
 616 +
 617 +    // File closed and no other error, so ensure next snapshot can be started.
 618 +    if (snap_state.state != SAVE_STATE_ERROR) {
 619 +        snap_state.state = SAVE_STATE_DONE;
 620 +    }
 621 +
 622 +    DPRINTF("savevm-end: cleanup done\n");
 623 +}
 624 +
 625 +int load_snapshot_from_blockdev(const char *filename, Error **errp)
 626 +{
 627 +    BlockBackend *be;
 628 +    Error *local_err = NULL;
 629 +    Error *blocker = NULL;
 630 +
 631 +    QEMUFile *f;
 632 +    size_t bs_pos = 0;
 633 +    int ret = -EINVAL;
 634 +
 635 +    be = blk_new_open(filename, NULL, NULL, 0, &local_err);
 636 +
 637 +    if (!be) {
 638 +        error_setg(errp, "Could not open VM state file");
 639 +        goto the_end;
 640 +    }
 641 +
 642 +    error_setg(&blocker, "block device is in use by load state");
 643 +    blk_op_block_all(be, blocker);
 644 +
 645 +    /* restore the VM state */
 646 +    f = qemu_file_new_input(QIO_CHANNEL(qio_channel_savevm_async_new(be, &bs_pos)));
 647 +    if (!f) {
 648 +        error_setg(errp, "Could not open VM state file");
 649 +        goto the_end;
 650 +    }
 651 +
 652 +    qemu_system_reset(SHUTDOWN_CAUSE_NONE);
 653 +    ret = qemu_loadvm_state(f);
 654 +
 655 +    /* dirty bitmap migration has a special case we need to trigger manually */
 656 +    dirty_bitmap_mig_before_vm_start();
 657 +
 658 +    qemu_fclose(f);
 659 +
 660 +    /* state_destroy assumes a real migration which would have added a yank */
 661 +    yank_register_instance(MIGRATION_YANK_INSTANCE, &error_abort);
 662 +
 663 +    migration_incoming_state_destroy();
 664 +    if (ret < 0) {
 665 +        error_setg_errno(errp, -ret, "Error while loading VM state");
 666 +        goto the_end;
 667 +    }
 668 +
 669 +    ret = 0;
 670 +
 671 + the_end:
 672 +    if (be) {
 673 +        blk_op_unblock_all(be, blocker);
 674 +        error_free(blocker);
 675 +        blk_unref(be);
 676 +    }
 677 +    return ret;
 678 +}
 679 diff --git a/monitor/hmp-cmds.c b/monitor/hmp-cmds.c
 680 index 6c559b48c8..91be698308 100644
 681 --- a/monitor/hmp-cmds.c
 682 +++ b/monitor/hmp-cmds.c
 683 @@ -22,6 +22,7 @@
 684  #include "monitor/monitor-internal.h"
 685  #include "qapi/error.h"
 686  #include "qapi/qapi-commands-control.h"
 687 +#include "qapi/qapi-commands-migration.h"
 688  #include "qapi/qapi-commands-misc.h"
 689  #include "qapi/qmp/qdict.h"
 690  #include "qapi/qmp/qerror.h"
 691 @@ -443,3 +444,40 @@ void hmp_info_mtree(Monitor *mon, const QDict *qdict)
 692
 693      mtree_info(flatview, dispatch_tree, owner, disabled);
 694  }
 695 +
 696 +void hmp_savevm_start(Monitor *mon, const QDict *qdict)
 697 +{
 698 +    Error *errp = NULL;
 699 +    const char *statefile = qdict_get_try_str(qdict, "statefile");
 700 +
 701 +    qmp_savevm_start(statefile, &errp);
 702 +    hmp_handle_error(mon, errp);
 703 +}
 704 +
 705 +void coroutine_fn hmp_savevm_end(Monitor *mon, const QDict *qdict)
 706 +{
 707 +    Error *errp = NULL;
 708 +
 709 +    qmp_savevm_end(&errp);
 710 +    hmp_handle_error(mon, errp);
 711 +}
 712 +
 713 +void hmp_info_savevm(Monitor *mon, const QDict *qdict)
 714 +{
 715 +    SaveVMInfo *info;
 716 +    info = qmp_query_savevm(NULL);
 717 +
 718 +    if (info->status) {
 719 +        monitor_printf(mon, "savevm status: %s\n", info->status);
 720 +        monitor_printf(mon, "total time: %" PRIu64 " milliseconds\n",
 721 +                       info->total_time);
 722 +    } else {
 723 +        monitor_printf(mon, "savevm status: not running\n");
 724 +    }
 725 +    if (info->has_bytes) {
 726 +        monitor_printf(mon, "Bytes saved: %"PRIu64"\n", info->bytes);
 727 +    }
 728 +    if (info->error) {
 729 +        monitor_printf(mon, "Error: %s\n", info->error);
 730 +    }
 731 +}
 732 diff --git a/qapi/migration.json b/qapi/migration.json
 733 index c84fa10e86..1702b92553 100644
 734 --- a/qapi/migration.json
 735 +++ b/qapi/migration.json
 736 @@ -261,6 +261,40 @@
 737             '*compression': 'CompressionStats',
 738             '*socket-address': ['SocketAddress'] } }
 739
 740 +##
 741 +# @SaveVMInfo:
 742 +#
 743 +# Information about current migration process.
 744 +#
 745 +# @status: string describing the current savevm status.
 746 +#          This can be 'active', 'completed', 'failed'.
 747 +#          If this field is not returned, no savevm process
 748 +#          has been initiated
 749 +#
 750 +# @error: string containing error message is status is failed.
 751 +#
 752 +# @total-time: total amount of milliseconds since savevm started.
 753 +#              If savevm has ended, it returns the total save time
 754 +#
 755 +# @bytes: total amount of data transfered
 756 +#
 757 +# Since: 1.3
 758 +##
 759 +{ 'struct': 'SaveVMInfo',
 760 +  'data': {'*status': 'str', '*error': 'str',
 761 +           '*total-time': 'int', '*bytes': 'int'} }
 762 +
 763 +##
 764 +# @query-savevm:
 765 +#
 766 +# Returns information about current savevm process.
 767 +#
 768 +# Returns: @SaveVMInfo
 769 +#
 770 +# Since: 1.3
 771 +##
 772 +{ 'command': 'query-savevm', 'returns': 'SaveVMInfo' }
 773 +
 774  ##
 775  # @query-migrate:
 776  #
 777 diff --git a/qapi/misc.json b/qapi/misc.json
 778 index 6ddd16ea28..e5681ae8a2 100644
 779 --- a/qapi/misc.json
 780 +++ b/qapi/misc.json
 781 @@ -469,6 +469,22 @@
 782  ##
 783  { 'command': 'query-fdsets', 'returns': ['FdsetInfo'] }
 784
 785 +##
 786 +# @savevm-start:
 787 +#
 788 +# Prepare for snapshot and halt VM. Save VM state to statefile.
 789 +#
 790 +##
 791 +{ 'command': 'savevm-start', 'data': { '*statefile': 'str' } }
 792 +
 793 +##
 794 +# @savevm-end:
 795 +#
 796 +# Resume VM after a snapshot.
 797 +#
 798 +##
 799 +{ 'command': 'savevm-end', 'coroutine': true }
 800 +
 801  ##
 802  # @CommandLineParameterType:
 803  #
 804 diff --git a/qemu-options.hx b/qemu-options.hx
 805 index fdddfab6ff..fdd551c2bb 100644
 806 --- a/qemu-options.hx
 807 +++ b/qemu-options.hx
 808 @@ -4398,6 +4398,18 @@ SRST
 809      Start right away with a saved state (``loadvm`` in monitor)
 810  ERST
 811
 812 +DEF("loadstate", HAS_ARG, QEMU_OPTION_loadstate, \
 813 +    "-loadstate file\n" \
 814 +    "                start right away with a saved state\n",
 815 +    QEMU_ARCH_ALL)
 816 +SRST
 817 +``-loadstate file``
 818 +  Start right away with a saved state. This option does not rollback
 819 +  disk state like @code{loadvm}, so user must make sure that disk
 820 +  have correct state. @var{file} can be any valid device URL. See the section
 821 +  for "Device URL Syntax" for more information.
 822 +ERST
 823 +
 824  #ifndef _WIN32
 825  DEF("daemonize", 0, QEMU_OPTION_daemonize, \
 826      "-daemonize      daemonize QEMU after initializing\n", QEMU_ARCH_ALL)
 827 diff --git a/softmmu/vl.c b/softmmu/vl.c
 828 index ea20b23e4c..0eabc71b68 100644
 829 --- a/softmmu/vl.c
 830 +++ b/softmmu/vl.c
 831 @@ -164,6 +164,7 @@ static const char *accelerators;
 832  static bool have_custom_ram_size;
 833  static const char *ram_memdev_id;
 834  static QDict *machine_opts_dict;
 835 +static const char *loadstate;
 836  static QTAILQ_HEAD(, ObjectOption) object_opts = QTAILQ_HEAD_INITIALIZER(object_opts);
 837  static QTAILQ_HEAD(, DeviceOption) device_opts = QTAILQ_HEAD_INITIALIZER(device_opts);
 838  static int display_remote;
 839 @@ -2612,6 +2613,12 @@ void qmp_x_exit_preconfig(Error **errp)
 840
 841      if (loadvm) {
 842          load_snapshot(loadvm, NULL, false, NULL, &error_fatal);
 843 +    } else if (loadstate) {
 844 +        Error *local_err = NULL;
 845 +        if (load_snapshot_from_blockdev(loadstate, &local_err) < 0) {
 846 +            error_report_err(local_err);
 847 +            autostart = 0;
 848 +        }
 849      }
 850      if (replay_mode != REPLAY_MODE_NONE) {
 851          replay_vmstate_init();
 852 @@ -3159,6 +3166,9 @@ void qemu_init(int argc, char **argv)
 853              case QEMU_OPTION_loadvm:
 854                  loadvm = optarg;
 855                  break;
 856 +            case QEMU_OPTION_loadstate:
 857 +                loadstate = optarg;
 858 +                break;
 859              case QEMU_OPTION_full_screen:
 860                  dpy.has_full_screen = true;
 861                  dpy.full_screen = true;