block.c 28.4 KB
Newer Older
L
lirans@il.ibm.com 已提交
1 2 3 4 5 6 7 8 9 10 11
/*
 * QEMU live block migration
 *
 * Copyright IBM, Corp. 2009
 *
 * Authors:
 *  Liran Schour   <lirans@il.ibm.com>
 *
 * This work is licensed under the terms of the GNU GPL, version 2.  See
 * the COPYING file in the top-level directory.
 *
12 13
 * Contributions after 2012-01-13 are licensed under the terms of the
 * GNU GPL, version 2 or (at your option) any later version.
L
lirans@il.ibm.com 已提交
14 15
 */

P
Peter Maydell 已提交
16
#include "qemu/osdep.h"
17
#include "qapi/error.h"
L
lirans@il.ibm.com 已提交
18
#include "qemu-common.h"
19 20 21
#include "block/block.h"
#include "qemu/error-report.h"
#include "qemu/main-loop.h"
L
lirans@il.ibm.com 已提交
22
#include "hw/hw.h"
23
#include "qemu/cutils.h"
24 25
#include "qemu/queue.h"
#include "qemu/timer.h"
26 27
#include "migration/block.h"
#include "migration/migration.h"
28
#include "sysemu/blockdev.h"
29
#include "migration/qemu-file.h"
30
#include "sysemu/block-backend.h"
L
lirans@il.ibm.com 已提交
31

32 33
#define BLOCK_SIZE                       (1 << 20)
#define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
L
lirans@il.ibm.com 已提交
34 35 36

#define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
#define BLK_MIG_FLAG_EOS                0x02
37
#define BLK_MIG_FLAG_PROGRESS           0x04
38
#define BLK_MIG_FLAG_ZERO_BLOCK         0x08
L
lirans@il.ibm.com 已提交
39 40 41

#define MAX_IS_ALLOCATED_SEARCH 65536

42 43
#define MAX_INFLIGHT_IO 512

L
lirans@il.ibm.com 已提交
44 45 46
//#define DEBUG_BLK_MIGRATION

#ifdef DEBUG_BLK_MIGRATION
M
malc 已提交
47
#define DPRINTF(fmt, ...) \
L
lirans@il.ibm.com 已提交
48 49
    do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
#else
M
malc 已提交
50
#define DPRINTF(fmt, ...) \
L
lirans@il.ibm.com 已提交
51 52 53
    do { } while (0)
#endif

54
typedef struct BlkMigDevState {
55
    /* Written during setup phase.  Can be read without a lock.  */
56 57
    BlockBackend *blk;
    char *blk_name;
58 59
    int shared_base;
    int64_t total_sectors;
60
    QSIMPLEQ_ENTRY(BlkMigDevState) entry;
61
    Error *blocker;
62 63 64 65 66 67

    /* Only used by migration thread.  Does not need a lock.  */
    int bulk_completed;
    int64_t cur_sector;
    int64_t cur_dirty;

68 69 70
    /* Data in the aio_bitmap is protected by block migration lock.
     * Allocation and free happen during setup and cleanup respectively.
     */
71
    unsigned long *aio_bitmap;
72 73

    /* Protected by block migration lock.  */
74
    int64_t completed_sectors;
75 76 77 78

    /* During migration this is protected by iothread lock / AioContext.
     * Allocation and free happen during setup and cleanup respectively.
     */
F
Fam Zheng 已提交
79
    BdrvDirtyBitmap *dirty_bitmap;
80 81
} BlkMigDevState;

L
lirans@il.ibm.com 已提交
82
typedef struct BlkMigBlock {
83
    /* Only used by migration thread.  */
L
lirans@il.ibm.com 已提交
84 85 86
    uint8_t *buf;
    BlkMigDevState *bmds;
    int64_t sector;
87
    int nr_sectors;
L
lirans@il.ibm.com 已提交
88 89
    struct iovec iov;
    QEMUIOVector qiov;
90
    BlockAIOCB *aiocb;
91

P
Paolo Bonzini 已提交
92
    /* Protected by block migration lock.  */
L
lirans@il.ibm.com 已提交
93
    int ret;
94
    QSIMPLEQ_ENTRY(BlkMigBlock) entry;
L
lirans@il.ibm.com 已提交
95 96 97
} BlkMigBlock;

typedef struct BlkMigState {
98
    QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
99
    int64_t total_sector_sum;
100
    bool zero_blocks;
101

P
Paolo Bonzini 已提交
102
    /* Protected by lock.  */
103
    QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
L
lirans@il.ibm.com 已提交
104 105
    int submitted;
    int read_done;
106 107

    /* Only used by migration thread.  Does not need a lock.  */
L
lirans@il.ibm.com 已提交
108
    int transferred;
109
    int prev_progress;
L
Liran Schour 已提交
110
    int bulk_completed;
P
Paolo Bonzini 已提交
111

112
    /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
P
Paolo Bonzini 已提交
113
    QemuMutex lock;
L
lirans@il.ibm.com 已提交
114 115
} BlkMigState;

116
static BlkMigState block_mig_state;
L
lirans@il.ibm.com 已提交
117

P
Paolo Bonzini 已提交
118 119 120 121 122 123 124 125 126 127
static void blk_mig_lock(void)
{
    qemu_mutex_lock(&block_mig_state.lock);
}

static void blk_mig_unlock(void)
{
    qemu_mutex_unlock(&block_mig_state.lock);
}

128 129 130 131
/* Must run outside of the iothread lock during the bulk phase,
 * or the VM will stall.
 */

132 133 134
static void blk_send(QEMUFile *f, BlkMigBlock * blk)
{
    int len;
135 136 137 138 139 140
    uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;

    if (block_mig_state.zero_blocks &&
        buffer_is_zero(blk->buf, BLOCK_SIZE)) {
        flags |= BLK_MIG_FLAG_ZERO_BLOCK;
    }
141 142 143

    /* sector number and flags */
    qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
144
                     | flags);
145 146

    /* device name */
147
    len = strlen(blk->bmds->blk_name);
148
    qemu_put_byte(f, len);
149
    qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
150

151 152 153 154 155 156 157 158
    /* if a block is zero we need to flush here since the network
     * bandwidth is now a lot higher than the storage device bandwidth.
     * thus if we queue zero blocks we slow down the migration */
    if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
        qemu_fflush(f);
        return;
    }

159 160 161
    qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
}

162 163 164 165 166 167 168 169 170 171
int blk_mig_active(void)
{
    return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
}

uint64_t blk_mig_bytes_transferred(void)
{
    BlkMigDevState *bmds;
    uint64_t sum = 0;

P
Paolo Bonzini 已提交
172
    blk_mig_lock();
173 174 175
    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
        sum += bmds->completed_sectors;
    }
P
Paolo Bonzini 已提交
176
    blk_mig_unlock();
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
    return sum << BDRV_SECTOR_BITS;
}

uint64_t blk_mig_bytes_remaining(void)
{
    return blk_mig_bytes_total() - blk_mig_bytes_transferred();
}

uint64_t blk_mig_bytes_total(void)
{
    BlkMigDevState *bmds;
    uint64_t sum = 0;

    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
        sum += bmds->total_sectors;
    }
    return sum << BDRV_SECTOR_BITS;
}

P
Paolo Bonzini 已提交
196 197 198

/* Called with migration lock held.  */

199 200 201 202
static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
{
    int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;

203
    if (sector < blk_nb_sectors(bmds->blk)) {
204 205 206 207 208 209 210
        return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
            (1UL << (chunk % (sizeof(unsigned long) * 8))));
    } else {
        return 0;
    }
}

P
Paolo Bonzini 已提交
211 212
/* Called with migration lock held.  */

213 214 215 216 217 218 219 220 221 222 223 224 225 226
static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
                             int nb_sectors, int set)
{
    int64_t start, end;
    unsigned long val, idx, bit;

    start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
    end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;

    for (; start <= end; start++) {
        idx = start / (sizeof(unsigned long) * 8);
        bit = start % (sizeof(unsigned long) * 8);
        val = bmds->aio_bitmap[idx];
        if (set) {
227
            val |= 1UL << bit;
228
        } else {
229
            val &= ~(1UL << bit);
230 231 232 233 234 235 236
        }
        bmds->aio_bitmap[idx] = val;
    }
}

static void alloc_aio_bitmap(BlkMigDevState *bmds)
{
237
    BlockBackend *bb = bmds->blk;
238 239
    int64_t bitmap_size;

240
    bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
241 242
    bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;

243
    bmds->aio_bitmap = g_malloc0(bitmap_size);
244 245
}

P
Paolo Bonzini 已提交
246 247
/* Never hold migration lock when yielding to the main loop!  */

L
lirans@il.ibm.com 已提交
248 249 250
static void blk_mig_read_cb(void *opaque, int ret)
{
    BlkMigBlock *blk = opaque;
251

P
Paolo Bonzini 已提交
252
    blk_mig_lock();
L
lirans@il.ibm.com 已提交
253
    blk->ret = ret;
254

255
    QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
256
    bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
257

258 259 260
    block_mig_state.submitted--;
    block_mig_state.read_done++;
    assert(block_mig_state.submitted >= 0);
P
Paolo Bonzini 已提交
261
    blk_mig_unlock();
L
lirans@il.ibm.com 已提交
262 263
}

264 265
/* Called with no lock taken.  */

266
static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
267
{
268 269
    int64_t total_sectors = bmds->total_sectors;
    int64_t cur_sector = bmds->cur_sector;
270
    BlockBackend *bb = bmds->blk;
L
lirans@il.ibm.com 已提交
271
    BlkMigBlock *blk;
272
    int nr_sectors;
273

274
    if (bmds->shared_base) {
275
        qemu_mutex_lock_iothread();
276
        aio_context_acquire(blk_get_aio_context(bb));
277 278
        /* Skip unallocated sectors; intentionally treats failure as
         * an allocated sector */
279
        while (cur_sector < total_sectors &&
280 281
               !bdrv_is_allocated(blk_bs(bb), cur_sector,
                                  MAX_IS_ALLOCATED_SEARCH, &nr_sectors)) {
L
lirans@il.ibm.com 已提交
282 283
            cur_sector += nr_sectors;
        }
284
        aio_context_release(blk_get_aio_context(bb));
285
        qemu_mutex_unlock_iothread();
L
lirans@il.ibm.com 已提交
286
    }
287 288

    if (cur_sector >= total_sectors) {
289
        bmds->cur_sector = bmds->completed_sectors = total_sectors;
L
lirans@il.ibm.com 已提交
290 291
        return 1;
    }
292

293
    bmds->completed_sectors = cur_sector;
294

295 296
    cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);

J
Jan Kiszka 已提交
297 298
    /* we are going to transfer a full block even if it is not allocated */
    nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
L
lirans@il.ibm.com 已提交
299

J
Jan Kiszka 已提交
300
    if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
301
        nr_sectors = total_sectors - cur_sector;
L
lirans@il.ibm.com 已提交
302
    }
303

304
    blk = g_new(BlkMigBlock, 1);
305
    blk->buf = g_malloc(BLOCK_SIZE);
306 307
    blk->bmds = bmds;
    blk->sector = cur_sector;
308
    blk->nr_sectors = nr_sectors;
309

L
Liran Schour 已提交
310 311 312
    blk->iov.iov_base = blk->buf;
    blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
    qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
313

P
Paolo Bonzini 已提交
314
    blk_mig_lock();
315
    block_mig_state.submitted++;
P
Paolo Bonzini 已提交
316
    blk_mig_unlock();
317

318 319 320 321 322 323 324 325
    /* We do not know if bs is under the main thread (and thus does
     * not acquire the AioContext when doing AIO) or rather under
     * dataplane.  Thus acquire both the iothread mutex and the
     * AioContext.
     *
     * This is ugly and will disappear when we make bdrv_* thread-safe,
     * without the need to acquire the AioContext.
     */
326
    qemu_mutex_lock_iothread();
327 328 329
    aio_context_acquire(blk_get_aio_context(bmds->blk));
    blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
                                0, blk_mig_read_cb, blk);
330

331
    bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector, nr_sectors);
332
    aio_context_release(blk_get_aio_context(bmds->blk));
333
    qemu_mutex_unlock_iothread();
334

335
    bmds->cur_sector = cur_sector + nr_sectors;
336
    return (bmds->cur_sector >= total_sectors);
L
lirans@il.ibm.com 已提交
337 338
}

339 340
/* Called with iothread lock taken.  */

341
static int set_dirty_tracking(void)
L
lirans@il.ibm.com 已提交
342 343
{
    BlkMigDevState *bmds;
344 345 346
    int ret;

    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
347 348 349 350
        aio_context_acquire(blk_get_aio_context(bmds->blk));
        bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
                                                      BLOCK_SIZE, NULL, NULL);
        aio_context_release(blk_get_aio_context(bmds->blk));
351 352 353 354 355 356
        if (!bmds->dirty_bitmap) {
            ret = -errno;
            goto fail;
        }
    }
    return 0;
357

358
fail:
359
    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
360
        if (bmds->dirty_bitmap) {
361 362 363
            aio_context_acquire(blk_get_aio_context(bmds->blk));
            bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
            aio_context_release(blk_get_aio_context(bmds->blk));
364
        }
F
Fam Zheng 已提交
365
    }
366
    return ret;
F
Fam Zheng 已提交
367 368
}

369 370
/* Called with iothread lock taken.  */

F
Fam Zheng 已提交
371 372 373 374 375
static void unset_dirty_tracking(void)
{
    BlkMigDevState *bmds;

    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
376 377 378
        aio_context_acquire(blk_get_aio_context(bmds->blk));
        bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
        aio_context_release(blk_get_aio_context(bmds->blk));
L
lirans@il.ibm.com 已提交
379 380 381
    }
}

382
static int init_blk_migration(QEMUFile *f)
L
lirans@il.ibm.com 已提交
383
{
384
    BlockDriverState *bs;
385
    BlkMigDevState *bmds;
386
    int64_t sectors;
K
Kevin Wolf 已提交
387
    BdrvNextIterator it;
388 389 390 391 392
    int i, num_bs = 0;
    struct {
        BlkMigDevState *bmds;
        BlockDriverState *bs;
    } *bmds_bs;
393 394
    Error *local_err = NULL;
    int ret;
395

396 397 398 399 400 401 402 403
    block_mig_state.submitted = 0;
    block_mig_state.read_done = 0;
    block_mig_state.transferred = 0;
    block_mig_state.total_sector_sum = 0;
    block_mig_state.prev_progress = -1;
    block_mig_state.bulk_completed = 0;
    block_mig_state.zero_blocks = migrate_zero_blocks();

K
Kevin Wolf 已提交
404
    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
405 406 407 408 409
        num_bs++;
    }
    bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));

    for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
410 411 412 413
        if (bdrv_is_read_only(bs)) {
            continue;
        }

414
        sectors = bdrv_nb_sectors(bs);
415
        if (sectors <= 0) {
416
            ret = sectors;
417
            goto out;
418 419
        }

420
        bmds = g_new0(BlkMigDevState, 1);
421
        bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
422
        bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
423 424 425
        bmds->bulk_completed = 0;
        bmds->total_sectors = sectors;
        bmds->completed_sectors = 0;
426
        bmds->shared_base = migrate_use_block_incremental();
427 428 429 430

        assert(i < num_bs);
        bmds_bs[i].bmds = bmds;
        bmds_bs[i].bs = bs;
431 432 433 434

        block_mig_state.total_sector_sum += sectors;

        if (bmds->shared_base) {
435
            DPRINTF("Start migration for %s with shared base image\n",
436
                    bdrv_get_device_name(bs));
437
        } else {
438
            DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
439 440 441 442
        }

        QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
    }
443 444 445 446 447 448 449 450

    /* Can only insert new BDSes now because doing so while iterating block
     * devices may end up in a deadlock (iterating the new BDSes, too). */
    for (i = 0; i < num_bs; i++) {
        BlkMigDevState *bmds = bmds_bs[i].bmds;
        BlockDriverState *bs = bmds_bs[i].bs;

        if (bmds) {
451 452 453 454 455
            ret = blk_insert_bs(bmds->blk, bs, &local_err);
            if (ret < 0) {
                error_report_err(local_err);
                goto out;
            }
456 457 458 459 460 461 462

            alloc_aio_bitmap(bmds);
            error_setg(&bmds->blocker, "block device is in use by migration");
            bdrv_op_block_all(bs, bmds->blocker);
        }
    }

463
    ret = 0;
464 465
out:
    g_free(bmds_bs);
466
    return ret;
467 468
}

469 470
/* Called with no lock taken.  */

471
static int blk_mig_save_bulked_block(QEMUFile *f)
L
lirans@il.ibm.com 已提交
472
{
473
    int64_t completed_sector_sum = 0;
L
lirans@il.ibm.com 已提交
474
    BlkMigDevState *bmds;
475
    int progress;
476
    int ret = 0;
L
lirans@il.ibm.com 已提交
477

478
    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
479
        if (bmds->bulk_completed == 0) {
480
            if (mig_save_device_bulk(f, bmds) == 1) {
481 482
                /* completed bulk section for this device */
                bmds->bulk_completed = 1;
L
lirans@il.ibm.com 已提交
483
            }
484 485 486 487 488
            completed_sector_sum += bmds->completed_sectors;
            ret = 1;
            break;
        } else {
            completed_sector_sum += bmds->completed_sectors;
L
lirans@il.ibm.com 已提交
489 490
        }
    }
491

492 493 494 495 496 497
    if (block_mig_state.total_sector_sum != 0) {
        progress = completed_sector_sum * 100 /
                   block_mig_state.total_sector_sum;
    } else {
        progress = 100;
    }
498 499 500 501
    if (progress != block_mig_state.prev_progress) {
        block_mig_state.prev_progress = progress;
        qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
                         | BLK_MIG_FLAG_PROGRESS);
502
        DPRINTF("Completed %d %%\r", progress);
503 504 505
    }

    return ret;
L
lirans@il.ibm.com 已提交
506 507
}

508
static void blk_mig_reset_dirty_cursor(void)
L
lirans@il.ibm.com 已提交
509 510
{
    BlkMigDevState *bmds;
511 512 513 514 515 516

    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
        bmds->cur_dirty = 0;
    }
}

517
/* Called with iothread lock and AioContext taken.  */
518

519 520
static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
                                 int is_async)
521 522
{
    BlkMigBlock *blk;
523
    BlockDriverState *bs = blk_bs(bmds->blk);
524
    int64_t total_sectors = bmds->total_sectors;
L
lirans@il.ibm.com 已提交
525
    int64_t sector;
526
    int nr_sectors;
527
    int ret = -EIO;
528

529
    for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
P
Paolo Bonzini 已提交
530
        blk_mig_lock();
531
        if (bmds_aio_inflight(bmds, sector)) {
P
Paolo Bonzini 已提交
532
            blk_mig_unlock();
533
            blk_drain(bmds->blk);
P
Paolo Bonzini 已提交
534 535
        } else {
            blk_mig_unlock();
536
        }
537
        if (bdrv_get_dirty(bs, bmds->dirty_bitmap, sector)) {
538

539 540 541 542 543
            if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
                nr_sectors = total_sectors - sector;
            } else {
                nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
            }
544
            blk = g_new(BlkMigBlock, 1);
545
            blk->buf = g_malloc(BLOCK_SIZE);
546 547
            blk->bmds = bmds;
            blk->sector = sector;
548
            blk->nr_sectors = nr_sectors;
549

550
            if (is_async) {
551 552 553 554
                blk->iov.iov_base = blk->buf;
                blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
                qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);

555 556 557 558
                blk->aiocb = blk_aio_preadv(bmds->blk,
                                            sector * BDRV_SECTOR_SIZE,
                                            &blk->qiov, 0, blk_mig_read_cb,
                                            blk);
P
Paolo Bonzini 已提交
559 560

                blk_mig_lock();
561
                block_mig_state.submitted++;
562
                bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
P
Paolo Bonzini 已提交
563
                blk_mig_unlock();
564
            } else {
565 566
                ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
                                nr_sectors * BDRV_SECTOR_SIZE);
567
                if (ret < 0) {
568
                    goto error;
L
lirans@il.ibm.com 已提交
569
                }
570
                blk_send(f, blk);
571

572 573
                g_free(blk->buf);
                g_free(blk);
574
            }
575

576
            bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, sector, nr_sectors);
577 578 579
            sector += nr_sectors;
            bmds->cur_dirty = sector;

580
            break;
L
lirans@il.ibm.com 已提交
581
        }
582 583
        sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
        bmds->cur_dirty = sector;
L
lirans@il.ibm.com 已提交
584
    }
585

586 587
    return (bmds->cur_dirty >= bmds->total_sectors);

588
error:
589
    DPRINTF("Error reading sector %" PRId64 "\n", sector);
590 591
    g_free(blk->buf);
    g_free(blk);
592
    return ret;
593 594
}

595 596 597
/* Called with iothread lock taken.
 *
 * return value:
598 599 600
 * 0: too much data for max_downtime
 * 1: few enough data for max_downtime
*/
601
static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
602 603
{
    BlkMigDevState *bmds;
604
    int ret = 1;
605 606

    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
607
        aio_context_acquire(blk_get_aio_context(bmds->blk));
608
        ret = mig_save_device_dirty(f, bmds, is_async);
609
        aio_context_release(blk_get_aio_context(bmds->blk));
610
        if (ret <= 0) {
611 612 613 614 615
            break;
        }
    }

    return ret;
L
lirans@il.ibm.com 已提交
616 617
}

618 619
/* Called with no locks taken.  */

620
static int flush_blks(QEMUFile *f)
L
lirans@il.ibm.com 已提交
621
{
622
    BlkMigBlock *blk;
623
    int ret = 0;
624

M
malc 已提交
625
    DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
626 627
            __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
            block_mig_state.transferred);
628

P
Paolo Bonzini 已提交
629
    blk_mig_lock();
630 631 632 633
    while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
        if (qemu_file_rate_limit(f)) {
            break;
        }
634
        if (blk->ret < 0) {
635
            ret = blk->ret;
636 637
            break;
        }
638

639
        QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
P
Paolo Bonzini 已提交
640
        blk_mig_unlock();
641
        blk_send(f, blk);
P
Paolo Bonzini 已提交
642
        blk_mig_lock();
643

644 645
        g_free(blk->buf);
        g_free(blk);
646

647 648 649
        block_mig_state.read_done--;
        block_mig_state.transferred++;
        assert(block_mig_state.read_done >= 0);
L
lirans@il.ibm.com 已提交
650
    }
P
Paolo Bonzini 已提交
651
    blk_mig_unlock();
L
lirans@il.ibm.com 已提交
652

M
malc 已提交
653
    DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
654 655
            block_mig_state.submitted, block_mig_state.read_done,
            block_mig_state.transferred);
656
    return ret;
L
lirans@il.ibm.com 已提交
657 658
}

659 660
/* Called with iothread lock taken.  */

661 662 663 664 665 666
static int64_t get_remaining_dirty(void)
{
    BlkMigDevState *bmds;
    int64_t dirty = 0;

    QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
667
        aio_context_acquire(blk_get_aio_context(bmds->blk));
668
        dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
669
        aio_context_release(blk_get_aio_context(bmds->blk));
670 671
    }

672
    return dirty << BDRV_SECTOR_BITS;
673 674
}

675 676
/* Called with iothread lock taken.  */

L
Liang Li 已提交
677
static void block_migration_cleanup(void *opaque)
678
{
679 680
    BlkMigDevState *bmds;
    BlkMigBlock *blk;
681
    AioContext *ctx;
682

683 684
    bdrv_drain_all();

F
Fam Zheng 已提交
685
    unset_dirty_tracking();
686

687 688
    while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
        QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
689
        bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
690
        error_free(bmds->blocker);
691

692 693
        /* Save ctx, because bmds->blk can disappear during blk_unref.  */
        ctx = blk_get_aio_context(bmds->blk);
694
        aio_context_acquire(ctx);
695
        blk_unref(bmds->blk);
696 697
        aio_context_release(ctx);

698
        g_free(bmds->blk_name);
699 700
        g_free(bmds->aio_bitmap);
        g_free(bmds);
701 702
    }

703
    blk_mig_lock();
704 705
    while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
        QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
706 707
        g_free(blk->buf);
        g_free(blk);
708
    }
P
Paolo Bonzini 已提交
709
    blk_mig_unlock();
710 711
}

712
static int block_save_setup(QEMUFile *f, void *opaque)
L
lirans@il.ibm.com 已提交
713
{
714 715
    int ret;

716 717
    DPRINTF("Enter save live setup submitted %d transferred %d\n",
            block_mig_state.submitted, block_mig_state.transferred);
718

719
    qemu_mutex_lock_iothread();
720 721 722 723 724
    ret = init_blk_migration(f);
    if (ret < 0) {
        qemu_mutex_unlock_iothread();
        return ret;
    }
725 726

    /* start track dirty blocks */
727 728
    ret = set_dirty_tracking();

729 730
    qemu_mutex_unlock_iothread();

731 732 733 734
    if (ret) {
        return ret;
    }

735
    ret = flush_blks(f);
736 737 738
    blk_mig_reset_dirty_cursor();
    qemu_put_be64(f, BLK_MIG_FLAG_EOS);

739
    return ret;
740 741
}

742
static int block_save_iterate(QEMUFile *f, void *opaque)
743 744
{
    int ret;
745
    int64_t last_ftell = qemu_ftell(f);
G
Gary R Hook 已提交
746
    int64_t delta_ftell;
747

748 749
    DPRINTF("Enter save live iterate submitted %d transferred %d\n",
            block_mig_state.submitted, block_mig_state.transferred);
750

751
    ret = flush_blks(f);
752 753
    if (ret) {
        return ret;
754 755
    }

756 757
    blk_mig_reset_dirty_cursor();

758
    /* control the rate of transfer */
P
Paolo Bonzini 已提交
759
    blk_mig_lock();
760 761
    while ((block_mig_state.submitted +
            block_mig_state.read_done) * BLOCK_SIZE <
762 763 764 765
           qemu_file_get_rate_limit(f) &&
           (block_mig_state.submitted +
            block_mig_state.read_done) <
           MAX_INFLIGHT_IO) {
P
Paolo Bonzini 已提交
766
        blk_mig_unlock();
767 768 769 770 771 772
        if (block_mig_state.bulk_completed == 0) {
            /* first finish the bulk phase */
            if (blk_mig_save_bulked_block(f) == 0) {
                /* finished saving bulk on all devices */
                block_mig_state.bulk_completed = 1;
            }
773
            ret = 0;
774
        } else {
775 776 777 778
            /* Always called with iothread lock taken for
             * simplicity, block_save_complete also calls it.
             */
            qemu_mutex_lock_iothread();
779
            ret = blk_mig_save_dirty_block(f, 1);
780
            qemu_mutex_unlock_iothread();
781 782 783 784
        }
        if (ret < 0) {
            return ret;
        }
P
Paolo Bonzini 已提交
785
        blk_mig_lock();
786 787 788
        if (ret != 0) {
            /* no more dirty blocks */
            break;
789
        }
790
    }
P
Paolo Bonzini 已提交
791
    blk_mig_unlock();
792

793
    ret = flush_blks(f);
794 795
    if (ret) {
        return ret;
796 797
    }

798
    qemu_put_be64(f, BLK_MIG_FLAG_EOS);
G
Gary R Hook 已提交
799 800 801 802 803 804 805 806
    delta_ftell = qemu_ftell(f) - last_ftell;
    if (delta_ftell > 0) {
        return 1;
    } else if (delta_ftell < 0) {
        return -1;
    } else {
        return 0;
    }
807 808
}

809 810
/* Called with iothread lock taken.  */

811 812 813 814 815 816 817
static int block_save_complete(QEMUFile *f, void *opaque)
{
    int ret;

    DPRINTF("Enter save live complete submitted %d transferred %d\n",
            block_mig_state.submitted, block_mig_state.transferred);

818
    ret = flush_blks(f);
819 820 821
    if (ret) {
        return ret;
    }
822

823
    blk_mig_reset_dirty_cursor();
824

825 826
    /* we know for sure that save bulk is completed and
       all async read completed */
P
Paolo Bonzini 已提交
827
    blk_mig_lock();
828
    assert(block_mig_state.submitted == 0);
P
Paolo Bonzini 已提交
829
    blk_mig_unlock();
830

831 832
    do {
        ret = blk_mig_save_dirty_block(f, 0);
833 834 835
        if (ret < 0) {
            return ret;
        }
836
    } while (ret == 0);
837

838 839
    /* report completion */
    qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
840

841 842
    DPRINTF("Block migration completed\n");

843 844
    qemu_put_be64(f, BLK_MIG_FLAG_EOS);

845
    return 0;
L
lirans@il.ibm.com 已提交
846 847
}

848 849 850
static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
                               uint64_t *non_postcopiable_pending,
                               uint64_t *postcopiable_pending)
851
{
852
    /* Estimate pending number of bytes to send */
853 854
    uint64_t pending;

855
    qemu_mutex_lock_iothread();
856 857 858
    pending = get_remaining_dirty();
    qemu_mutex_unlock_iothread();

P
Paolo Bonzini 已提交
859
    blk_mig_lock();
860 861 862
    pending += block_mig_state.submitted * BLOCK_SIZE +
               block_mig_state.read_done * BLOCK_SIZE;
    blk_mig_unlock();
863 864

    /* Report at least one block pending during bulk phase */
865 866
    if (pending <= max_size && !block_mig_state.bulk_completed) {
        pending = max_size + BLOCK_SIZE;
867
    }
868

869
    DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
870 871
    /* We don't do postcopy */
    *non_postcopiable_pending += pending;
872 873
}

L
lirans@il.ibm.com 已提交
874 875
static int block_load(QEMUFile *f, void *opaque, int version_id)
{
876
    static int banner_printed;
L
lirans@il.ibm.com 已提交
877 878 879
    int len, flags;
    char device_name[256];
    int64_t addr;
880
    BlockBackend *blk, *blk_prev = NULL;;
881
    Error *local_err = NULL;
L
lirans@il.ibm.com 已提交
882
    uint8_t *buf;
883 884
    int64_t total_sectors = 0;
    int nr_sectors;
885
    int ret;
886 887
    BlockDriverInfo bdi;
    int cluster_size = BLOCK_SIZE;
888

L
lirans@il.ibm.com 已提交
889 890
    do {
        addr = qemu_get_be64(f);
891

J
Jan Kiszka 已提交
892 893
        flags = addr & ~BDRV_SECTOR_MASK;
        addr >>= BDRV_SECTOR_BITS;
894 895

        if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
L
lirans@il.ibm.com 已提交
896 897 898 899
            /* get device name */
            len = qemu_get_byte(f);
            qemu_get_buffer(f, (uint8_t *)device_name, len);
            device_name[len] = '\0';
900

901 902
            blk = blk_by_name(device_name);
            if (!blk) {
903 904 905 906
                fprintf(stderr, "Error unknown block device %s\n",
                        device_name);
                return -EINVAL;
            }
907

908 909 910
            if (blk != blk_prev) {
                blk_prev = blk;
                total_sectors = blk_nb_sectors(blk);
911
                if (total_sectors <= 0) {
912
                    error_report("Error getting length of block device %s",
913 914 915
                                 device_name);
                    return -EINVAL;
                }
916

917
                blk_invalidate_cache(blk, &local_err);
918 919 920 921
                if (local_err) {
                    error_report_err(local_err);
                    return -EINVAL;
                }
922 923 924 925 926 927 928 929 930

                ret = bdrv_get_info(blk_bs(blk), &bdi);
                if (ret == 0 && bdi.cluster_size > 0 &&
                    bdi.cluster_size <= BLOCK_SIZE &&
                    BLOCK_SIZE % bdi.cluster_size == 0) {
                    cluster_size = bdi.cluster_size;
                } else {
                    cluster_size = BLOCK_SIZE;
                }
931 932 933 934 935 936 937 938
            }

            if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
                nr_sectors = total_sectors - addr;
            } else {
                nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
            }

939
            if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
940 941 942
                ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
                                        nr_sectors * BDRV_SECTOR_SIZE,
                                        BDRV_REQ_MAY_UNMAP);
943
            } else {
944 945 946 947
                int i;
                int64_t cur_addr;
                uint8_t *cur_buf;

948 949
                buf = g_malloc(BLOCK_SIZE);
                qemu_get_buffer(f, buf, BLOCK_SIZE);
950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967
                for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
                    cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
                    cur_buf = buf + i * cluster_size;

                    if ((!block_mig_state.zero_blocks ||
                        cluster_size < BLOCK_SIZE) &&
                        buffer_is_zero(cur_buf, cluster_size)) {
                        ret = blk_pwrite_zeroes(blk, cur_addr,
                                                cluster_size,
                                                BDRV_REQ_MAY_UNMAP);
                    } else {
                        ret = blk_pwrite(blk, cur_addr, cur_buf,
                                         cluster_size, 0);
                    }
                    if (ret < 0) {
                        break;
                    }
                }
968 969
                g_free(buf);
            }
970

971 972 973
            if (ret < 0) {
                return ret;
            }
974 975 976 977 978 979 980 981
        } else if (flags & BLK_MIG_FLAG_PROGRESS) {
            if (!banner_printed) {
                printf("Receiving block device images\n");
                banner_printed = 1;
            }
            printf("Completed %d %%%c", (int)addr,
                   (addr == 100) ? '\n' : '\r');
            fflush(stdout);
982
        } else if (!(flags & BLK_MIG_FLAG_EOS)) {
983
            fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
984 985
            return -EINVAL;
        }
986 987 988
        ret = qemu_file_get_error(f);
        if (ret != 0) {
            return ret;
L
lirans@il.ibm.com 已提交
989
        }
990 991
    } while (!(flags & BLK_MIG_FLAG_EOS));

L
lirans@il.ibm.com 已提交
992 993 994
    return 0;
}

995 996
static bool block_is_active(void *opaque)
{
997
    return migrate_use_block();
998 999
}

1000
static SaveVMHandlers savevm_block_handlers = {
1001
    .save_live_setup = block_save_setup,
1002
    .save_live_iterate = block_save_iterate,
1003
    .save_live_complete_precopy = block_save_complete,
1004
    .save_live_pending = block_save_pending,
1005
    .load_state = block_load,
L
Liang Li 已提交
1006
    .cleanup = block_migration_cleanup,
1007
    .is_active = block_is_active,
1008 1009
};

L
lirans@il.ibm.com 已提交
1010
void blk_mig_init(void)
1011
{
1012 1013
    QSIMPLEQ_INIT(&block_mig_state.bmds_list);
    QSIMPLEQ_INIT(&block_mig_state.blk_list);
P
Paolo Bonzini 已提交
1014
    qemu_mutex_init(&block_mig_state.lock);
1015

1016 1017
    register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
                         &block_mig_state);
L
lirans@il.ibm.com 已提交
1018
}