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              |  33 +++
  35  include/migration/snapshot.h |   2 +
  36  include/monitor/hmp.h        |   5 +
  37  migration/meson.build        |   1 +
  38  migration/savevm-async.c     | 548 +++++++++++++++++++++++++++++++++++
  39  monitor/hmp-cmds.c           |  58 ++++
  40  qapi/migration.json          |  34 +++
  41  qapi/misc.json               |  32 ++
  42  qemu-options.hx              |  12 +
  43  softmmu/vl.c                 |  10 +
  44  11 files changed, 748 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..b66d7fc4ab 100644
  73 --- a/hmp-commands.hx
  74 +++ b/hmp-commands.hx
  75 @@ -1846,3 +1846,36 @@ 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       = "snapshot-drive",
  90 +        .args_type  = "device:s,name:s",
  91 +        .params     = "device name",
  92 +        .help       = "Create internal snapshot.",
  93 +        .cmd = hmp_snapshot_drive,
  94 +    },
  95 +
  96 +    {
  97 +        .name       = "delete-drive-snapshot",
  98 +        .args_type  = "device:s,name:s",
  99 +        .params     = "device name",
 100 +        .help       = "Delete internal snapshot.",
 101 +        .cmd = hmp_delete_drive_snapshot,
 102 +    },
 103 +
 104 +    {
 105 +        .name       = "savevm-end",
 106 +        .args_type  = "",
 107 +        .params     = "",
 108 +        .help       = "Resume VM after snaphot.",
 109 +        .cmd        = hmp_savevm_end,
 110 +        .coroutine  = true,
 111 +    },
 112 diff --git a/include/migration/snapshot.h b/include/migration/snapshot.h
 113 index e72083b117..c846d37806 100644
 114 --- a/include/migration/snapshot.h
 115 +++ b/include/migration/snapshot.h
 116 @@ -61,4 +61,6 @@ bool delete_snapshot(const char *name,
 117                      bool has_devices, strList *devices,
 118                      Error **errp);
 119
 120 +int load_snapshot_from_blockdev(const char *filename, Error **errp);
 121 +
 122  #endif
 123 diff --git a/include/monitor/hmp.h b/include/monitor/hmp.h
 124 index fdb69b7f9c..c012bad741 100644
 125 --- a/include/monitor/hmp.h
 126 +++ b/include/monitor/hmp.h
 127 @@ -28,6 +28,7 @@ void hmp_info_status(Monitor *mon, const QDict *qdict);
 128  void hmp_info_uuid(Monitor *mon, const QDict *qdict);
 129  void hmp_info_chardev(Monitor *mon, const QDict *qdict);
 130  void hmp_info_mice(Monitor *mon, const QDict *qdict);
 131 +void hmp_info_savevm(Monitor *mon, const QDict *qdict);
 132  void hmp_info_migrate(Monitor *mon, const QDict *qdict);
 133  void hmp_info_migrate_capabilities(Monitor *mon, const QDict *qdict);
 134  void hmp_info_migrate_parameters(Monitor *mon, const QDict *qdict);
 135 @@ -94,6 +95,10 @@ void hmp_closefd(Monitor *mon, const QDict *qdict);
 136  void hmp_mouse_move(Monitor *mon, const QDict *qdict);
 137  void hmp_mouse_button(Monitor *mon, const QDict *qdict);
 138  void hmp_mouse_set(Monitor *mon, const QDict *qdict);
 139 +void hmp_savevm_start(Monitor *mon, const QDict *qdict);
 140 +void hmp_snapshot_drive(Monitor *mon, const QDict *qdict);
 141 +void hmp_delete_drive_snapshot(Monitor *mon, const QDict *qdict);
 142 +void hmp_savevm_end(Monitor *mon, const QDict *qdict);
 143  void hmp_sendkey(Monitor *mon, const QDict *qdict);
 144  void coroutine_fn hmp_screendump(Monitor *mon, const QDict *qdict);
 145  void hmp_chardev_add(Monitor *mon, const QDict *qdict);
 146 diff --git a/migration/meson.build b/migration/meson.build
 147 index 8a142fc7a9..a7824b5266 100644
 148 --- a/migration/meson.build
 149 +++ b/migration/meson.build
 150 @@ -25,6 +25,7 @@ softmmu_ss.add(files(
 151    'multifd-zlib.c',
 152    'postcopy-ram.c',
 153    'savevm.c',
 154 +  'savevm-async.c',
 155    'socket.c',
 156    'tls.c',
 157    'threadinfo.c',
 158 diff --git a/migration/savevm-async.c b/migration/savevm-async.c
 159 new file mode 100644
 160 index 0000000000..c5db9e9c1e
 161 --- /dev/null
 162 +++ b/migration/savevm-async.c
 163 @@ -0,0 +1,548 @@
 164 +#include "qemu/osdep.h"
 165 +#include "migration/channel-savevm-async.h"
 166 +#include "migration/migration.h"
 167 +#include "migration/savevm.h"
 168 +#include "migration/snapshot.h"
 169 +#include "migration/global_state.h"
 170 +#include "migration/ram.h"
 171 +#include "migration/qemu-file.h"
 172 +#include "sysemu/sysemu.h"
 173 +#include "sysemu/runstate.h"
 174 +#include "block/block.h"
 175 +#include "sysemu/block-backend.h"
 176 +#include "qapi/error.h"
 177 +#include "qapi/qmp/qerror.h"
 178 +#include "qapi/qmp/qdict.h"
 179 +#include "qapi/qapi-commands-migration.h"
 180 +#include "qapi/qapi-commands-misc.h"
 181 +#include "qapi/qapi-commands-block.h"
 182 +#include "qemu/cutils.h"
 183 +#include "qemu/timer.h"
 184 +#include "qemu/main-loop.h"
 185 +#include "qemu/rcu.h"
 186 +#include "qemu/yank.h"
 187 +
 188 +/* #define DEBUG_SAVEVM_STATE */
 189 +
 190 +#ifdef DEBUG_SAVEVM_STATE
 191 +#define DPRINTF(fmt, ...) \
 192 +    do { printf("savevm-async: " fmt, ## __VA_ARGS__); } while (0)
 193 +#else
 194 +#define DPRINTF(fmt, ...) \
 195 +    do { } while (0)
 196 +#endif
 197 +
 198 +enum {
 199 +    SAVE_STATE_DONE,
 200 +    SAVE_STATE_ERROR,
 201 +    SAVE_STATE_ACTIVE,
 202 +    SAVE_STATE_COMPLETED,
 203 +    SAVE_STATE_CANCELLED
 204 +};
 205 +
 206 +
 207 +static struct SnapshotState {
 208 +    BlockBackend *target;
 209 +    size_t bs_pos;
 210 +    int state;
 211 +    Error *error;
 212 +    Error *blocker;
 213 +    int saved_vm_running;
 214 +    QEMUFile *file;
 215 +    int64_t total_time;
 216 +    QEMUBH *finalize_bh;
 217 +    Coroutine *co;
 218 +    QemuCoSleep target_close_wait;
 219 +} snap_state;
 220 +
 221 +static bool savevm_aborted(void)
 222 +{
 223 +    return snap_state.state == SAVE_STATE_CANCELLED ||
 224 +        snap_state.state == SAVE_STATE_ERROR;
 225 +}
 226 +
 227 +SaveVMInfo *qmp_query_savevm(Error **errp)
 228 +{
 229 +    SaveVMInfo *info = g_malloc0(sizeof(*info));
 230 +    struct SnapshotState *s = &snap_state;
 231 +
 232 +    if (s->state != SAVE_STATE_DONE) {
 233 +        info->has_bytes = true;
 234 +        info->bytes = s->bs_pos;
 235 +        switch (s->state) {
 236 +        case SAVE_STATE_ERROR:
 237 +            info->status = g_strdup("failed");
 238 +            info->has_total_time = true;
 239 +            info->total_time = s->total_time;
 240 +            if (s->error) {
 241 +                info->error = g_strdup(error_get_pretty(s->error));
 242 +            }
 243 +            break;
 244 +        case SAVE_STATE_ACTIVE:
 245 +            info->status = g_strdup("active");
 246 +            info->has_total_time = true;
 247 +            info->total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME)
 248 +                - s->total_time;
 249 +            break;
 250 +        case SAVE_STATE_COMPLETED:
 251 +            info->status = g_strdup("completed");
 252 +            info->has_total_time = true;
 253 +            info->total_time = s->total_time;
 254 +            break;
 255 +        }
 256 +    }
 257 +
 258 +    return info;
 259 +}
 260 +
 261 +static int save_snapshot_cleanup(void)
 262 +{
 263 +    int ret = 0;
 264 +
 265 +    DPRINTF("save_snapshot_cleanup\n");
 266 +
 267 +    snap_state.total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
 268 +        snap_state.total_time;
 269 +
 270 +    if (snap_state.file) {
 271 +        ret = qemu_fclose(snap_state.file);
 272 +        snap_state.file = NULL;
 273 +    }
 274 +
 275 +    if (snap_state.target) {
 276 +        if (!savevm_aborted()) {
 277 +            /* try to truncate, but ignore errors (will fail on block devices).
 278 +            * note1: bdrv_read() need whole blocks, so we need to round up
 279 +            * note2: PVE requires 1024 (BDRV_SECTOR_SIZE*2) alignment
 280 +            */
 281 +            size_t size = QEMU_ALIGN_UP(snap_state.bs_pos, BDRV_SECTOR_SIZE*2);
 282 +            blk_truncate(snap_state.target, size, false, PREALLOC_MODE_OFF, 0, NULL);
 283 +        }
 284 +        blk_op_unblock_all(snap_state.target, snap_state.blocker);
 285 +        error_free(snap_state.blocker);
 286 +        snap_state.blocker = NULL;
 287 +        blk_unref(snap_state.target);
 288 +        snap_state.target = NULL;
 289 +
 290 +        qemu_co_sleep_wake(&snap_state.target_close_wait);
 291 +    }
 292 +
 293 +    return ret;
 294 +}
 295 +
 296 +static void save_snapshot_error(const char *fmt, ...)
 297 +{
 298 +    va_list ap;
 299 +    char *msg;
 300 +
 301 +    va_start(ap, fmt);
 302 +    msg = g_strdup_vprintf(fmt, ap);
 303 +    va_end(ap);
 304 +
 305 +    DPRINTF("save_snapshot_error: %s\n", msg);
 306 +
 307 +    if (!snap_state.error) {
 308 +        error_set(&snap_state.error, ERROR_CLASS_GENERIC_ERROR, "%s", msg);
 309 +    }
 310 +
 311 +    g_free (msg);
 312 +
 313 +    snap_state.state = SAVE_STATE_ERROR;
 314 +}
 315 +
 316 +static void process_savevm_finalize(void *opaque)
 317 +{
 318 +    int ret;
 319 +    AioContext *iohandler_ctx = iohandler_get_aio_context();
 320 +    MigrationState *ms = migrate_get_current();
 321 +
 322 +    bool aborted = savevm_aborted();
 323 +
 324 +#ifdef DEBUG_SAVEVM_STATE
 325 +    int64_t start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 326 +#endif
 327 +
 328 +    qemu_bh_delete(snap_state.finalize_bh);
 329 +    snap_state.finalize_bh = NULL;
 330 +    snap_state.co = NULL;
 331 +
 332 +    /* We need to own the target bdrv's context for the following functions,
 333 +     * so move it back. It can stay in the main context and live out its live
 334 +     * there, since we're done with it after this method ends anyway.
 335 +     */
 336 +    aio_context_acquire(iohandler_ctx);
 337 +    blk_set_aio_context(snap_state.target, qemu_get_aio_context(), NULL);
 338 +    aio_context_release(iohandler_ctx);
 339 +
 340 +    ret = vm_stop_force_state(RUN_STATE_FINISH_MIGRATE);
 341 +    if (ret < 0) {
 342 +        save_snapshot_error("vm_stop_force_state error %d", ret);
 343 +    }
 344 +
 345 +    if (!aborted) {
 346 +        /* skip state saving if we aborted, snapshot will be invalid anyway */
 347 +        (void)qemu_savevm_state_complete_precopy(snap_state.file, false, false);
 348 +        ret = qemu_file_get_error(snap_state.file);
 349 +        if (ret < 0) {
 350 +            save_snapshot_error("qemu_savevm_state_complete_precopy error %d", ret);
 351 +        }
 352 +    }
 353 +
 354 +    DPRINTF("state saving complete\n");
 355 +    DPRINTF("timing: process_savevm_finalize (state saving) took %ld ms\n",
 356 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time);
 357 +
 358 +    /* clear migration state */
 359 +    migrate_set_state(&ms->state, MIGRATION_STATUS_SETUP,
 360 +        ret || aborted ? MIGRATION_STATUS_FAILED : MIGRATION_STATUS_COMPLETED);
 361 +    ms->to_dst_file = NULL;
 362 +
 363 +    qemu_savevm_state_cleanup();
 364 +
 365 +    ret = save_snapshot_cleanup();
 366 +    if (ret < 0) {
 367 +        save_snapshot_error("save_snapshot_cleanup error %d", ret);
 368 +    } else if (snap_state.state == SAVE_STATE_ACTIVE) {
 369 +        snap_state.state = SAVE_STATE_COMPLETED;
 370 +    } else if (aborted) {
 371 +        /*
 372 +         * If there was an error, there's no need to set a new one here.
 373 +         * If the snapshot was canceled, leave setting the state to
 374 +         * qmp_savevm_end(), which is waked by save_snapshot_cleanup().
 375 +         */
 376 +    } else {
 377 +        save_snapshot_error("process_savevm_cleanup: invalid state: %d",
 378 +                            snap_state.state);
 379 +    }
 380 +    if (snap_state.saved_vm_running) {
 381 +        vm_start();
 382 +        snap_state.saved_vm_running = false;
 383 +    }
 384 +
 385 +    DPRINTF("timing: process_savevm_finalize (full) took %ld ms\n",
 386 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time);
 387 +}
 388 +
 389 +static void coroutine_fn process_savevm_co(void *opaque)
 390 +{
 391 +    int ret;
 392 +    int64_t maxlen;
 393 +    BdrvNextIterator it;
 394 +    BlockDriverState *bs = NULL;
 395 +
 396 +#ifdef DEBUG_SAVEVM_STATE
 397 +    int64_t start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 398 +#endif
 399 +
 400 +    ret = qemu_file_get_error(snap_state.file);
 401 +    if (ret < 0) {
 402 +        save_snapshot_error("qemu_savevm_state_setup failed");
 403 +        return;
 404 +    }
 405 +
 406 +    while (snap_state.state == SAVE_STATE_ACTIVE) {
 407 +        uint64_t pending_size, pend_precopy, pend_postcopy;
 408 +        uint64_t threshold = 400 * 1000;
 409 +
 410 +        /*
 411 +         * pending_{estimate,exact} are expected to be called without iothread
 412 +         * lock. Similar to what is done in migration.c, call the exact variant
 413 +         * only once pend_precopy in the estimate is below the threshold.
 414 +         */
 415 +        qemu_mutex_unlock_iothread();
 416 +        qemu_savevm_state_pending_estimate(&pend_precopy, &pend_postcopy);
 417 +        if (pend_precopy <= threshold) {
 418 +            qemu_savevm_state_pending_exact(&pend_precopy, &pend_postcopy);
 419 +        }
 420 +        qemu_mutex_lock_iothread();
 421 +        pending_size = pend_precopy + pend_postcopy;
 422 +
 423 +        /*
 424 +         * A guest reaching this cutoff is dirtying lots of RAM. It should be
 425 +         * large enough so that the guest can't dirty this much between the
 426 +         * check and the guest actually being stopped, but it should be small
 427 +         * enough to avoid long downtimes for non-hibernation snapshots.
 428 +         */
 429 +        maxlen = blk_getlength(snap_state.target) - 100*1024*1024;
 430 +
 431 +        /* Note that there is no progress for pend_postcopy when iterating */
 432 +        if (pend_precopy > threshold && snap_state.bs_pos + pending_size < maxlen) {
 433 +            ret = qemu_savevm_state_iterate(snap_state.file, false);
 434 +            if (ret < 0) {
 435 +                save_snapshot_error("qemu_savevm_state_iterate error %d", ret);
 436 +                break;
 437 +            }
 438 +            DPRINTF("savevm iterate pending size %lu ret %d\n", pending_size, ret);
 439 +        } else {
 440 +            qemu_system_wakeup_request(QEMU_WAKEUP_REASON_OTHER, NULL);
 441 +            ret = global_state_store();
 442 +            if (ret) {
 443 +                save_snapshot_error("global_state_store error %d", ret);
 444 +                break;
 445 +            }
 446 +
 447 +            DPRINTF("savevm iterate complete\n");
 448 +            break;
 449 +        }
 450 +    }
 451 +
 452 +    DPRINTF("timing: process_savevm_co took %ld ms\n",
 453 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time);
 454 +
 455 +#ifdef DEBUG_SAVEVM_STATE
 456 +    int64_t start_time_flush = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 457 +#endif
 458 +    /* If a drive runs in an IOThread we can flush it async, and only
 459 +     * need to sync-flush whatever IO happens between now and
 460 +     * vm_stop_force_state. bdrv_next can only be called from main AioContext,
 461 +     * so move there now and after every flush.
 462 +     */
 463 +    aio_co_reschedule_self(qemu_get_aio_context());
 464 +    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 465 +        /* target has BDRV_O_NO_FLUSH, no sense calling bdrv_flush on it */
 466 +        if (bs == blk_bs(snap_state.target)) {
 467 +            continue;
 468 +        }
 469 +
 470 +        AioContext *bs_ctx = bdrv_get_aio_context(bs);
 471 +        if (bs_ctx != qemu_get_aio_context()) {
 472 +            DPRINTF("savevm: async flushing drive %s\n", bs->filename);
 473 +            aio_co_reschedule_self(bs_ctx);
 474 +            bdrv_graph_co_rdlock();
 475 +            bdrv_flush(bs);
 476 +            bdrv_graph_co_rdunlock();
 477 +            aio_co_reschedule_self(qemu_get_aio_context());
 478 +        }
 479 +    }
 480 +
 481 +    DPRINTF("timing: async flushing took %ld ms\n",
 482 +        qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - start_time_flush);
 483 +
 484 +    qemu_bh_schedule(snap_state.finalize_bh);
 485 +}
 486 +
 487 +void qmp_savevm_start(const char *statefile, Error **errp)
 488 +{
 489 +    Error *local_err = NULL;
 490 +    MigrationState *ms = migrate_get_current();
 491 +    AioContext *iohandler_ctx = iohandler_get_aio_context();
 492 +
 493 +    int bdrv_oflags = BDRV_O_RDWR | BDRV_O_RESIZE | BDRV_O_NO_FLUSH;
 494 +
 495 +    if (snap_state.state != SAVE_STATE_DONE) {
 496 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR,
 497 +                  "VM snapshot already started\n");
 498 +        return;
 499 +    }
 500 +
 501 +    if (migration_is_running(ms->state)) {
 502 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR, QERR_MIGRATION_ACTIVE);
 503 +        return;
 504 +    }
 505 +
 506 +    if (migrate_use_block()) {
 507 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR,
 508 +                  "Block migration and snapshots are incompatible");
 509 +        return;
 510 +    }
 511 +
 512 +    /* initialize snapshot info */
 513 +    snap_state.saved_vm_running = runstate_is_running();
 514 +    snap_state.bs_pos = 0;
 515 +    snap_state.total_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
 516 +    snap_state.blocker = NULL;
 517 +    snap_state.target_close_wait = (QemuCoSleep){ .to_wake = NULL };
 518 +
 519 +    if (snap_state.error) {
 520 +        error_free(snap_state.error);
 521 +        snap_state.error = NULL;
 522 +    }
 523 +
 524 +    if (!statefile) {
 525 +        vm_stop(RUN_STATE_SAVE_VM);
 526 +        snap_state.state = SAVE_STATE_COMPLETED;
 527 +        return;
 528 +    }
 529 +
 530 +    if (qemu_savevm_state_blocked(errp)) {
 531 +        return;
 532 +    }
 533 +
 534 +    /* Open the image */
 535 +    QDict *options = NULL;
 536 +    options = qdict_new();
 537 +    qdict_put_str(options, "driver", "raw");
 538 +    snap_state.target = blk_new_open(statefile, NULL, options, bdrv_oflags, &local_err);
 539 +    if (!snap_state.target) {
 540 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR, "failed to open '%s'", statefile);
 541 +        goto restart;
 542 +    }
 543 +
 544 +    QIOChannel *ioc = QIO_CHANNEL(qio_channel_savevm_async_new(snap_state.target,
 545 +                                                               &snap_state.bs_pos));
 546 +    snap_state.file = qemu_file_new_output(ioc);
 547 +
 548 +    if (!snap_state.file) {
 549 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR, "failed to open '%s'", statefile);
 550 +        goto restart;
 551 +    }
 552 +
 553 +    /*
 554 +     * qemu_savevm_* paths use migration code and expect a migration state.
 555 +     * State is cleared in process_savevm_co, but has to be initialized
 556 +     * here (blocking main thread, from QMP) to avoid race conditions.
 557 +     */
 558 +    migrate_init(ms);
 559 +    memset(&ram_counters, 0, sizeof(ram_counters));
 560 +    memset(&compression_counters, 0, sizeof(compression_counters));
 561 +    ms->to_dst_file = snap_state.file;
 562 +
 563 +    error_setg(&snap_state.blocker, "block device is in use by savevm");
 564 +    blk_op_block_all(snap_state.target, snap_state.blocker);
 565 +
 566 +    snap_state.state = SAVE_STATE_ACTIVE;
 567 +    snap_state.finalize_bh = qemu_bh_new(process_savevm_finalize, &snap_state);
 568 +    snap_state.co = qemu_coroutine_create(&process_savevm_co, NULL);
 569 +    qemu_mutex_unlock_iothread();
 570 +    qemu_savevm_state_header(snap_state.file);
 571 +    qemu_savevm_state_setup(snap_state.file);
 572 +    qemu_mutex_lock_iothread();
 573 +
 574 +    /* Async processing from here on out happens in iohandler context, so let
 575 +     * the target bdrv have its home there.
 576 +     */
 577 +    blk_set_aio_context(snap_state.target, iohandler_ctx, &local_err);
 578 +
 579 +    aio_co_schedule(iohandler_ctx, snap_state.co);
 580 +
 581 +    return;
 582 +
 583 +restart:
 584 +
 585 +    save_snapshot_error("setup failed");
 586 +
 587 +    if (snap_state.saved_vm_running) {
 588 +        vm_start();
 589 +        snap_state.saved_vm_running = false;
 590 +    }
 591 +}
 592 +
 593 +void coroutine_fn qmp_savevm_end(Error **errp)
 594 +{
 595 +    int64_t timeout;
 596 +
 597 +    if (snap_state.state == SAVE_STATE_DONE) {
 598 +        error_set(errp, ERROR_CLASS_GENERIC_ERROR,
 599 +                  "VM snapshot not started\n");
 600 +        return;
 601 +    }
 602 +
 603 +    if (snap_state.state == SAVE_STATE_ACTIVE) {
 604 +        snap_state.state = SAVE_STATE_CANCELLED;
 605 +        goto wait_for_close;
 606 +    }
 607 +
 608 +    if (snap_state.saved_vm_running) {
 609 +        vm_start();
 610 +        snap_state.saved_vm_running = false;
 611 +    }
 612 +
 613 +    snap_state.state = SAVE_STATE_DONE;
 614 +
 615 +wait_for_close:
 616 +    if (!snap_state.target) {
 617 +        DPRINTF("savevm-end: no target file open\n");
 618 +        return;
 619 +    }
 620 +
 621 +    /* wait until cleanup is done before returning, this ensures that after this
 622 +     * call exits the statefile will be closed and can be removed immediately */
 623 +    DPRINTF("savevm-end: waiting for cleanup\n");
 624 +    timeout = 30L * 1000 * 1000 * 1000;
 625 +    qemu_co_sleep_ns_wakeable(&snap_state.target_close_wait,
 626 +                              QEMU_CLOCK_REALTIME, timeout);
 627 +    if (snap_state.target) {
 628 +        save_snapshot_error("timeout waiting for target file close in "
 629 +                            "qmp_savevm_end");
 630 +        /* we cannot assume the snapshot finished in this case, so leave the
 631 +         * state alone - caller has to figure something out */
 632 +        return;
 633 +    }
 634 +
 635 +    // File closed and no other error, so ensure next snapshot can be started.
 636 +    if (snap_state.state != SAVE_STATE_ERROR) {
 637 +        snap_state.state = SAVE_STATE_DONE;
 638 +    }
 639 +
 640 +    DPRINTF("savevm-end: cleanup done\n");
 641 +}
 642 +
 643 +// FIXME: Deprecated
 644 +void qmp_snapshot_drive(const char *device, const char *name, Error **errp)
 645 +{
 646 +    // Compatibility to older qemu-server.
 647 +    qmp_blockdev_snapshot_internal_sync(device, name, errp);
 648 +}
 649 +
 650 +// FIXME: Deprecated
 651 +void qmp_delete_drive_snapshot(const char *device, const char *name,
 652 +                               Error **errp)
 653 +{
 654 +    // Compatibility to older qemu-server.
 655 +    (void)qmp_blockdev_snapshot_delete_internal_sync(device, NULL, name, errp);
 656 +}
 657 +
 658 +int load_snapshot_from_blockdev(const char *filename, Error **errp)
 659 +{
 660 +    BlockBackend *be;
 661 +    Error *local_err = NULL;
 662 +    Error *blocker = NULL;
 663 +
 664 +    QEMUFile *f;
 665 +    size_t bs_pos = 0;
 666 +    int ret = -EINVAL;
 667 +
 668 +    be = blk_new_open(filename, NULL, NULL, 0, &local_err);
 669 +
 670 +    if (!be) {
 671 +        error_setg(errp, "Could not open VM state file");
 672 +        goto the_end;
 673 +    }
 674 +
 675 +    error_setg(&blocker, "block device is in use by load state");
 676 +    blk_op_block_all(be, blocker);
 677 +
 678 +    /* restore the VM state */
 679 +    f = qemu_file_new_input(QIO_CHANNEL(qio_channel_savevm_async_new(be, &bs_pos)));
 680 +    if (!f) {
 681 +        error_setg(errp, "Could not open VM state file");
 682 +        goto the_end;
 683 +    }
 684 +
 685 +    qemu_system_reset(SHUTDOWN_CAUSE_NONE);
 686 +    ret = qemu_loadvm_state(f);
 687 +
 688 +    /* dirty bitmap migration has a special case we need to trigger manually */
 689 +    dirty_bitmap_mig_before_vm_start();
 690 +
 691 +    qemu_fclose(f);
 692 +
 693 +    /* state_destroy assumes a real migration which would have added a yank */
 694 +    yank_register_instance(MIGRATION_YANK_INSTANCE, &error_abort);
 695 +
 696 +    migration_incoming_state_destroy();
 697 +    if (ret < 0) {
 698 +        error_setg_errno(errp, -ret, "Error while loading VM state");
 699 +        goto the_end;
 700 +    }
 701 +
 702 +    ret = 0;
 703 +
 704 + the_end:
 705 +    if (be) {
 706 +        blk_op_unblock_all(be, blocker);
 707 +        error_free(blocker);
 708 +        blk_unref(be);
 709 +    }
 710 +    return ret;
 711 +}
 712 diff --git a/monitor/hmp-cmds.c b/monitor/hmp-cmds.c
 713 index 6c559b48c8..435f9334f9 100644
 714 --- a/monitor/hmp-cmds.c
 715 +++ b/monitor/hmp-cmds.c
 716 @@ -22,6 +22,7 @@
 717  #include "monitor/monitor-internal.h"
 718  #include "qapi/error.h"
 719  #include "qapi/qapi-commands-control.h"
 720 +#include "qapi/qapi-commands-migration.h"
 721  #include "qapi/qapi-commands-misc.h"
 722  #include "qapi/qmp/qdict.h"
 723  #include "qapi/qmp/qerror.h"
 724 @@ -443,3 +444,60 @@ void hmp_info_mtree(Monitor *mon, const QDict *qdict)
 725
 726      mtree_info(flatview, dispatch_tree, owner, disabled);
 727  }
 728 +
 729 +void hmp_savevm_start(Monitor *mon, const QDict *qdict)
 730 +{
 731 +    Error *errp = NULL;
 732 +    const char *statefile = qdict_get_try_str(qdict, "statefile");
 733 +
 734 +    qmp_savevm_start(statefile, &errp);
 735 +    hmp_handle_error(mon, errp);
 736 +}
 737 +
 738 +void hmp_snapshot_drive(Monitor *mon, const QDict *qdict)
 739 +{
 740 +    Error *errp = NULL;
 741 +    const char *name = qdict_get_str(qdict, "name");
 742 +    const char *device = qdict_get_str(qdict, "device");
 743 +
 744 +    qmp_snapshot_drive(device, name, &errp);
 745 +    hmp_handle_error(mon, errp);
 746 +}
 747 +
 748 +void hmp_delete_drive_snapshot(Monitor *mon, const QDict *qdict)
 749 +{
 750 +    Error *errp = NULL;
 751 +    const char *name = qdict_get_str(qdict, "name");
 752 +    const char *device = qdict_get_str(qdict, "device");
 753 +
 754 +    qmp_delete_drive_snapshot(device, name, &errp);
 755 +    hmp_handle_error(mon, errp);
 756 +}
 757 +
 758 +void coroutine_fn hmp_savevm_end(Monitor *mon, const QDict *qdict)
 759 +{
 760 +    Error *errp = NULL;
 761 +
 762 +    qmp_savevm_end(&errp);
 763 +    hmp_handle_error(mon, errp);
 764 +}
 765 +
 766 +void hmp_info_savevm(Monitor *mon, const QDict *qdict)
 767 +{
 768 +    SaveVMInfo *info;
 769 +    info = qmp_query_savevm(NULL);
 770 +
 771 +    if (info->status) {
 772 +        monitor_printf(mon, "savevm status: %s\n", info->status);
 773 +        monitor_printf(mon, "total time: %" PRIu64 " milliseconds\n",
 774 +                       info->total_time);
 775 +    } else {
 776 +        monitor_printf(mon, "savevm status: not running\n");
 777 +    }
 778 +    if (info->has_bytes) {
 779 +        monitor_printf(mon, "Bytes saved: %"PRIu64"\n", info->bytes);
 780 +    }
 781 +    if (info->error) {
 782 +        monitor_printf(mon, "Error: %s\n", info->error);
 783 +    }
 784 +}
 785 diff --git a/qapi/migration.json b/qapi/migration.json
 786 index c84fa10e86..1702b92553 100644
 787 --- a/qapi/migration.json
 788 +++ b/qapi/migration.json
 789 @@ -261,6 +261,40 @@
 790             '*compression': 'CompressionStats',
 791             '*socket-address': ['SocketAddress'] } }
 792
 793 +##
 794 +# @SaveVMInfo:
 795 +#
 796 +# Information about current migration process.
 797 +#
 798 +# @status: string describing the current savevm status.
 799 +#          This can be 'active', 'completed', 'failed'.
 800 +#          If this field is not returned, no savevm process
 801 +#          has been initiated
 802 +#
 803 +# @error: string containing error message is status is failed.
 804 +#
 805 +# @total-time: total amount of milliseconds since savevm started.
 806 +#              If savevm has ended, it returns the total save time
 807 +#
 808 +# @bytes: total amount of data transfered
 809 +#
 810 +# Since: 1.3
 811 +##
 812 +{ 'struct': 'SaveVMInfo',
 813 +  'data': {'*status': 'str', '*error': 'str',
 814 +           '*total-time': 'int', '*bytes': 'int'} }
 815 +
 816 +##
 817 +# @query-savevm:
 818 +#
 819 +# Returns information about current savevm process.
 820 +#
 821 +# Returns: @SaveVMInfo
 822 +#
 823 +# Since: 1.3
 824 +##
 825 +{ 'command': 'query-savevm', 'returns': 'SaveVMInfo' }
 826 +
 827  ##
 828  # @query-migrate:
 829  #
 830 diff --git a/qapi/misc.json b/qapi/misc.json
 831 index 6ddd16ea28..098c9bbe93 100644
 832 --- a/qapi/misc.json
 833 +++ b/qapi/misc.json
 834 @@ -469,6 +469,38 @@
 835  ##
 836  { 'command': 'query-fdsets', 'returns': ['FdsetInfo'] }
 837
 838 +##
 839 +# @savevm-start:
 840 +#
 841 +# Prepare for snapshot and halt VM. Save VM state to statefile.
 842 +#
 843 +##
 844 +{ 'command': 'savevm-start', 'data': { '*statefile': 'str' } }
 845 +
 846 +##
 847 +# @snapshot-drive:
 848 +#
 849 +# Create an internal drive snapshot.
 850 +#
 851 +##
 852 +{ 'command': 'snapshot-drive', 'data': { 'device': 'str', 'name': 'str' } }
 853 +
 854 +##
 855 +# @delete-drive-snapshot:
 856 +#
 857 +# Delete a drive snapshot.
 858 +#
 859 +##
 860 +{ 'command': 'delete-drive-snapshot', 'data': { 'device': 'str', 'name': 'str' } }
 861 +
 862 +##
 863 +# @savevm-end:
 864 +#
 865 +# Resume VM after a snapshot.
 866 +#
 867 +##
 868 +{ 'command': 'savevm-end', 'coroutine': true }
 869 +
 870  ##
 871  # @CommandLineParameterType:
 872  #
 873 diff --git a/qemu-options.hx b/qemu-options.hx
 874 index 59bdf67a2c..fc6cb23dd9 100644
 875 --- a/qemu-options.hx
 876 +++ b/qemu-options.hx
 877 @@ -4378,6 +4378,18 @@ SRST
 878      Start right away with a saved state (``loadvm`` in monitor)
 879  ERST
 880
 881 +DEF("loadstate", HAS_ARG, QEMU_OPTION_loadstate, \
 882 +    "-loadstate file\n" \
 883 +    "                start right away with a saved state\n",
 884 +    QEMU_ARCH_ALL)
 885 +SRST
 886 +``-loadstate file``
 887 +  Start right away with a saved state. This option does not rollback
 888 +  disk state like @code{loadvm}, so user must make sure that disk
 889 +  have correct state. @var{file} can be any valid device URL. See the section
 890 +  for "Device URL Syntax" for more information.
 891 +ERST
 892 +
 893  #ifndef _WIN32
 894  DEF("daemonize", 0, QEMU_OPTION_daemonize, \
 895      "-daemonize      daemonize QEMU after initializing\n", QEMU_ARCH_ALL)
 896 diff --git a/softmmu/vl.c b/softmmu/vl.c
 897 index ea20b23e4c..0eabc71b68 100644
 898 --- a/softmmu/vl.c
 899 +++ b/softmmu/vl.c
 900 @@ -164,6 +164,7 @@ static const char *accelerators;
 901  static bool have_custom_ram_size;
 902  static const char *ram_memdev_id;
 903  static QDict *machine_opts_dict;
 904 +static const char *loadstate;
 905  static QTAILQ_HEAD(, ObjectOption) object_opts = QTAILQ_HEAD_INITIALIZER(object_opts);
 906  static QTAILQ_HEAD(, DeviceOption) device_opts = QTAILQ_HEAD_INITIALIZER(device_opts);
 907  static int display_remote;
 908 @@ -2612,6 +2613,12 @@ void qmp_x_exit_preconfig(Error **errp)
 909
 910      if (loadvm) {
 911          load_snapshot(loadvm, NULL, false, NULL, &error_fatal);
 912 +    } else if (loadstate) {
 913 +        Error *local_err = NULL;
 914 +        if (load_snapshot_from_blockdev(loadstate, &local_err) < 0) {
 915 +            error_report_err(local_err);
 916 +            autostart = 0;
 917 +        }
 918      }
 919      if (replay_mode != REPLAY_MODE_NONE) {
 920          replay_vmstate_init();
 921 @@ -3159,6 +3166,9 @@ void qemu_init(int argc, char **argv)
 922              case QEMU_OPTION_loadvm:
 923                  loadvm = optarg;
 924                  break;
 925 +            case QEMU_OPTION_loadstate:
 926 +                loadstate = optarg;
 927 +                break;
 928              case QEMU_OPTION_full_screen:
 929                  dpy.has_full_screen = true;
 930                  dpy.full_screen = true;