1 /*
2  * QEMU live block migration
3  *
4  * Copyright IBM, Corp. 2009
5  *
6  * Authors:
7  *  Liran Schour   <lirans@il.ibm.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15 
16 #include "qemu/osdep.h"
17 #include "qapi/error.h"
18 #include "qemu/error-report.h"
19 #include "qemu/cutils.h"
20 #include "qemu/queue.h"
21 #include "block.h"
22 #include "migration/misc.h"
23 #include "migration.h"
24 #include "migration/register.h"
25 #include "qemu-file.h"
26 #include "migration/vmstate.h"
27 #include "sysemu/block-backend.h"
28 
29 #define BLOCK_SIZE                       (1 << 20)
30 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
31 
32 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
33 #define BLK_MIG_FLAG_EOS                0x02
34 #define BLK_MIG_FLAG_PROGRESS           0x04
35 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
36 
37 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
38 
39 #define MAX_IO_BUFFERS 512
40 #define MAX_PARALLEL_IO 16
41 
42 //#define DEBUG_BLK_MIGRATION
43 
44 #ifdef DEBUG_BLK_MIGRATION
45 #define DPRINTF(fmt, ...) \
46     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
47 #else
48 #define DPRINTF(fmt, ...) \
49     do { } while (0)
50 #endif
51 
52 typedef struct BlkMigDevState {
53     /* Written during setup phase.  Can be read without a lock.  */
54     BlockBackend *blk;
55     char *blk_name;
56     int shared_base;
57     int64_t total_sectors;
58     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
59     Error *blocker;
60 
61     /* Only used by migration thread.  Does not need a lock.  */
62     int bulk_completed;
63     int64_t cur_sector;
64     int64_t cur_dirty;
65 
66     /* Data in the aio_bitmap is protected by block migration lock.
67      * Allocation and free happen during setup and cleanup respectively.
68      */
69     unsigned long *aio_bitmap;
70 
71     /* Protected by block migration lock.  */
72     int64_t completed_sectors;
73 
74     /* During migration this is protected by iothread lock / AioContext.
75      * Allocation and free happen during setup and cleanup respectively.
76      */
77     BdrvDirtyBitmap *dirty_bitmap;
78 } BlkMigDevState;
79 
80 typedef struct BlkMigBlock {
81     /* Only used by migration thread.  */
82     uint8_t *buf;
83     BlkMigDevState *bmds;
84     int64_t sector;
85     int nr_sectors;
86     struct iovec iov;
87     QEMUIOVector qiov;
88     BlockAIOCB *aiocb;
89 
90     /* Protected by block migration lock.  */
91     int ret;
92     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
93 } BlkMigBlock;
94 
95 typedef struct BlkMigState {
96     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
97     int64_t total_sector_sum;
98     bool zero_blocks;
99 
100     /* Protected by lock.  */
101     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
102     int submitted;
103     int read_done;
104 
105     /* Only used by migration thread.  Does not need a lock.  */
106     int transferred;
107     int prev_progress;
108     int bulk_completed;
109 
110     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
111     QemuMutex lock;
112 } BlkMigState;
113 
114 static BlkMigState block_mig_state;
115 
blk_mig_lock(void)116 static void blk_mig_lock(void)
117 {
118     qemu_mutex_lock(&block_mig_state.lock);
119 }
120 
blk_mig_unlock(void)121 static void blk_mig_unlock(void)
122 {
123     qemu_mutex_unlock(&block_mig_state.lock);
124 }
125 
126 /* Must run outside of the iothread lock during the bulk phase,
127  * or the VM will stall.
128  */
129 
blk_send(QEMUFile * f,BlkMigBlock * blk)130 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
131 {
132     int len;
133     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
134 
135     if (block_mig_state.zero_blocks &&
136         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
137         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
138     }
139 
140     /* sector number and flags */
141     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
142                      | flags);
143 
144     /* device name */
145     len = strlen(blk->bmds->blk_name);
146     qemu_put_byte(f, len);
147     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
148 
149     /* if a block is zero we need to flush here since the network
150      * bandwidth is now a lot higher than the storage device bandwidth.
151      * thus if we queue zero blocks we slow down the migration */
152     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
153         qemu_fflush(f);
154         return;
155     }
156 
157     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
158 }
159 
blk_mig_active(void)160 int blk_mig_active(void)
161 {
162     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
163 }
164 
blk_mig_bulk_active(void)165 int blk_mig_bulk_active(void)
166 {
167     return blk_mig_active() && !block_mig_state.bulk_completed;
168 }
169 
blk_mig_bytes_transferred(void)170 uint64_t blk_mig_bytes_transferred(void)
171 {
172     BlkMigDevState *bmds;
173     uint64_t sum = 0;
174 
175     blk_mig_lock();
176     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
177         sum += bmds->completed_sectors;
178     }
179     blk_mig_unlock();
180     return sum << BDRV_SECTOR_BITS;
181 }
182 
blk_mig_bytes_remaining(void)183 uint64_t blk_mig_bytes_remaining(void)
184 {
185     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
186 }
187 
blk_mig_bytes_total(void)188 uint64_t blk_mig_bytes_total(void)
189 {
190     BlkMigDevState *bmds;
191     uint64_t sum = 0;
192 
193     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
194         sum += bmds->total_sectors;
195     }
196     return sum << BDRV_SECTOR_BITS;
197 }
198 
199 
200 /* Called with migration lock held.  */
201 
bmds_aio_inflight(BlkMigDevState * bmds,int64_t sector)202 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
203 {
204     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
205 
206     if (sector < blk_nb_sectors(bmds->blk)) {
207         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
208             (1UL << (chunk % (sizeof(unsigned long) * 8))));
209     } else {
210         return 0;
211     }
212 }
213 
214 /* Called with migration lock held.  */
215 
bmds_set_aio_inflight(BlkMigDevState * bmds,int64_t sector_num,int nb_sectors,int set)216 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
217                              int nb_sectors, int set)
218 {
219     int64_t start, end;
220     unsigned long val, idx, bit;
221 
222     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
223     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
224 
225     for (; start <= end; start++) {
226         idx = start / (sizeof(unsigned long) * 8);
227         bit = start % (sizeof(unsigned long) * 8);
228         val = bmds->aio_bitmap[idx];
229         if (set) {
230             val |= 1UL << bit;
231         } else {
232             val &= ~(1UL << bit);
233         }
234         bmds->aio_bitmap[idx] = val;
235     }
236 }
237 
alloc_aio_bitmap(BlkMigDevState * bmds)238 static void alloc_aio_bitmap(BlkMigDevState *bmds)
239 {
240     BlockBackend *bb = bmds->blk;
241     int64_t bitmap_size;
242 
243     bitmap_size = blk_nb_sectors(bb) + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
244     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
245 
246     bmds->aio_bitmap = g_malloc0(bitmap_size);
247 }
248 
249 /* Never hold migration lock when yielding to the main loop!  */
250 
blk_mig_read_cb(void * opaque,int ret)251 static void blk_mig_read_cb(void *opaque, int ret)
252 {
253     BlkMigBlock *blk = opaque;
254 
255     blk_mig_lock();
256     blk->ret = ret;
257 
258     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
259     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
260 
261     block_mig_state.submitted--;
262     block_mig_state.read_done++;
263     assert(block_mig_state.submitted >= 0);
264     blk_mig_unlock();
265 }
266 
267 /* Called with no lock taken.  */
268 
mig_save_device_bulk(QEMUFile * f,BlkMigDevState * bmds)269 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
270 {
271     int64_t total_sectors = bmds->total_sectors;
272     int64_t cur_sector = bmds->cur_sector;
273     BlockBackend *bb = bmds->blk;
274     BlkMigBlock *blk;
275     int nr_sectors;
276     int64_t count;
277 
278     if (bmds->shared_base) {
279         qemu_mutex_lock_iothread();
280         aio_context_acquire(blk_get_aio_context(bb));
281         /* Skip unallocated sectors; intentionally treats failure or
282          * partial sector as an allocated sector */
283         while (cur_sector < total_sectors &&
284                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
285                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
286             if (count < BDRV_SECTOR_SIZE) {
287                 break;
288             }
289             cur_sector += count >> BDRV_SECTOR_BITS;
290         }
291         aio_context_release(blk_get_aio_context(bb));
292         qemu_mutex_unlock_iothread();
293     }
294 
295     if (cur_sector >= total_sectors) {
296         bmds->cur_sector = bmds->completed_sectors = total_sectors;
297         return 1;
298     }
299 
300     bmds->completed_sectors = cur_sector;
301 
302     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
303 
304     /* we are going to transfer a full block even if it is not allocated */
305     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
306 
307     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
308         nr_sectors = total_sectors - cur_sector;
309     }
310 
311     blk = g_new(BlkMigBlock, 1);
312     blk->buf = g_malloc(BLOCK_SIZE);
313     blk->bmds = bmds;
314     blk->sector = cur_sector;
315     blk->nr_sectors = nr_sectors;
316 
317     blk->iov.iov_base = blk->buf;
318     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
319     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
320 
321     blk_mig_lock();
322     block_mig_state.submitted++;
323     blk_mig_unlock();
324 
325     /* We do not know if bs is under the main thread (and thus does
326      * not acquire the AioContext when doing AIO) or rather under
327      * dataplane.  Thus acquire both the iothread mutex and the
328      * AioContext.
329      *
330      * This is ugly and will disappear when we make bdrv_* thread-safe,
331      * without the need to acquire the AioContext.
332      */
333     qemu_mutex_lock_iothread();
334     aio_context_acquire(blk_get_aio_context(bmds->blk));
335     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
336                             nr_sectors * BDRV_SECTOR_SIZE);
337     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
338                                 0, blk_mig_read_cb, blk);
339     aio_context_release(blk_get_aio_context(bmds->blk));
340     qemu_mutex_unlock_iothread();
341 
342     bmds->cur_sector = cur_sector + nr_sectors;
343     return (bmds->cur_sector >= total_sectors);
344 }
345 
346 /* Called with iothread lock taken.  */
347 
set_dirty_tracking(void)348 static int set_dirty_tracking(void)
349 {
350     BlkMigDevState *bmds;
351     int ret;
352 
353     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
354         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
355                                                       BLOCK_SIZE, NULL, NULL);
356         if (!bmds->dirty_bitmap) {
357             ret = -errno;
358             goto fail;
359         }
360     }
361     return 0;
362 
363 fail:
364     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
365         if (bmds->dirty_bitmap) {
366             bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
367         }
368     }
369     return ret;
370 }
371 
372 /* Called with iothread lock taken.  */
373 
unset_dirty_tracking(void)374 static void unset_dirty_tracking(void)
375 {
376     BlkMigDevState *bmds;
377 
378     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
379         bdrv_release_dirty_bitmap(blk_bs(bmds->blk), bmds->dirty_bitmap);
380     }
381 }
382 
init_blk_migration(QEMUFile * f)383 static int init_blk_migration(QEMUFile *f)
384 {
385     BlockDriverState *bs;
386     BlkMigDevState *bmds;
387     int64_t sectors;
388     BdrvNextIterator it;
389     int i, num_bs = 0;
390     struct {
391         BlkMigDevState *bmds;
392         BlockDriverState *bs;
393     } *bmds_bs;
394     Error *local_err = NULL;
395     int ret;
396 
397     block_mig_state.submitted = 0;
398     block_mig_state.read_done = 0;
399     block_mig_state.transferred = 0;
400     block_mig_state.total_sector_sum = 0;
401     block_mig_state.prev_progress = -1;
402     block_mig_state.bulk_completed = 0;
403     block_mig_state.zero_blocks = migrate_zero_blocks();
404 
405     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
406         num_bs++;
407     }
408     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
409 
410     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
411         if (bdrv_is_read_only(bs)) {
412             continue;
413         }
414 
415         sectors = bdrv_nb_sectors(bs);
416         if (sectors <= 0) {
417             ret = sectors;
418             bdrv_next_cleanup(&it);
419             goto out;
420         }
421 
422         bmds = g_new0(BlkMigDevState, 1);
423         bmds->blk = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
424         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
425         bmds->bulk_completed = 0;
426         bmds->total_sectors = sectors;
427         bmds->completed_sectors = 0;
428         bmds->shared_base = migrate_use_block_incremental();
429 
430         assert(i < num_bs);
431         bmds_bs[i].bmds = bmds;
432         bmds_bs[i].bs = bs;
433 
434         block_mig_state.total_sector_sum += sectors;
435 
436         if (bmds->shared_base) {
437             DPRINTF("Start migration for %s with shared base image\n",
438                     bdrv_get_device_name(bs));
439         } else {
440             DPRINTF("Start full migration for %s\n", bdrv_get_device_name(bs));
441         }
442 
443         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
444     }
445 
446     /* Can only insert new BDSes now because doing so while iterating block
447      * devices may end up in a deadlock (iterating the new BDSes, too). */
448     for (i = 0; i < num_bs; i++) {
449         BlkMigDevState *bmds = bmds_bs[i].bmds;
450         BlockDriverState *bs = bmds_bs[i].bs;
451 
452         if (bmds) {
453             ret = blk_insert_bs(bmds->blk, bs, &local_err);
454             if (ret < 0) {
455                 error_report_err(local_err);
456                 goto out;
457             }
458 
459             alloc_aio_bitmap(bmds);
460             error_setg(&bmds->blocker, "block device is in use by migration");
461             bdrv_op_block_all(bs, bmds->blocker);
462         }
463     }
464 
465     ret = 0;
466 out:
467     g_free(bmds_bs);
468     return ret;
469 }
470 
471 /* Called with no lock taken.  */
472 
blk_mig_save_bulked_block(QEMUFile * f)473 static int blk_mig_save_bulked_block(QEMUFile *f)
474 {
475     int64_t completed_sector_sum = 0;
476     BlkMigDevState *bmds;
477     int progress;
478     int ret = 0;
479 
480     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
481         if (bmds->bulk_completed == 0) {
482             if (mig_save_device_bulk(f, bmds) == 1) {
483                 /* completed bulk section for this device */
484                 bmds->bulk_completed = 1;
485             }
486             completed_sector_sum += bmds->completed_sectors;
487             ret = 1;
488             break;
489         } else {
490             completed_sector_sum += bmds->completed_sectors;
491         }
492     }
493 
494     if (block_mig_state.total_sector_sum != 0) {
495         progress = completed_sector_sum * 100 /
496                    block_mig_state.total_sector_sum;
497     } else {
498         progress = 100;
499     }
500     if (progress != block_mig_state.prev_progress) {
501         block_mig_state.prev_progress = progress;
502         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
503                          | BLK_MIG_FLAG_PROGRESS);
504         DPRINTF("Completed %d %%\r", progress);
505     }
506 
507     return ret;
508 }
509 
blk_mig_reset_dirty_cursor(void)510 static void blk_mig_reset_dirty_cursor(void)
511 {
512     BlkMigDevState *bmds;
513 
514     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
515         bmds->cur_dirty = 0;
516     }
517 }
518 
519 /* Called with iothread lock and AioContext taken.  */
520 
mig_save_device_dirty(QEMUFile * f,BlkMigDevState * bmds,int is_async)521 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
522                                  int is_async)
523 {
524     BlkMigBlock *blk;
525     BlockDriverState *bs = blk_bs(bmds->blk);
526     int64_t total_sectors = bmds->total_sectors;
527     int64_t sector;
528     int nr_sectors;
529     int ret = -EIO;
530 
531     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
532         blk_mig_lock();
533         if (bmds_aio_inflight(bmds, sector)) {
534             blk_mig_unlock();
535             blk_drain(bmds->blk);
536         } else {
537             blk_mig_unlock();
538         }
539         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
540         if (bdrv_get_dirty_locked(bs, bmds->dirty_bitmap,
541                                   sector * BDRV_SECTOR_SIZE)) {
542             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
543                 nr_sectors = total_sectors - sector;
544             } else {
545                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
546             }
547             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
548                                            sector * BDRV_SECTOR_SIZE,
549                                            nr_sectors * BDRV_SECTOR_SIZE);
550             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
551 
552             blk = g_new(BlkMigBlock, 1);
553             blk->buf = g_malloc(BLOCK_SIZE);
554             blk->bmds = bmds;
555             blk->sector = sector;
556             blk->nr_sectors = nr_sectors;
557 
558             if (is_async) {
559                 blk->iov.iov_base = blk->buf;
560                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
561                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
562 
563                 blk->aiocb = blk_aio_preadv(bmds->blk,
564                                             sector * BDRV_SECTOR_SIZE,
565                                             &blk->qiov, 0, blk_mig_read_cb,
566                                             blk);
567 
568                 blk_mig_lock();
569                 block_mig_state.submitted++;
570                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
571                 blk_mig_unlock();
572             } else {
573                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE, blk->buf,
574                                 nr_sectors * BDRV_SECTOR_SIZE);
575                 if (ret < 0) {
576                     goto error;
577                 }
578                 blk_send(f, blk);
579 
580                 g_free(blk->buf);
581                 g_free(blk);
582             }
583 
584             sector += nr_sectors;
585             bmds->cur_dirty = sector;
586             break;
587         }
588 
589         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
590         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
591         bmds->cur_dirty = sector;
592     }
593 
594     return (bmds->cur_dirty >= bmds->total_sectors);
595 
596 error:
597     DPRINTF("Error reading sector %" PRId64 "\n", sector);
598     g_free(blk->buf);
599     g_free(blk);
600     return ret;
601 }
602 
603 /* Called with iothread lock taken.
604  *
605  * return value:
606  * 0: too much data for max_downtime
607  * 1: few enough data for max_downtime
608 */
blk_mig_save_dirty_block(QEMUFile * f,int is_async)609 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
610 {
611     BlkMigDevState *bmds;
612     int ret = 1;
613 
614     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
615         aio_context_acquire(blk_get_aio_context(bmds->blk));
616         ret = mig_save_device_dirty(f, bmds, is_async);
617         aio_context_release(blk_get_aio_context(bmds->blk));
618         if (ret <= 0) {
619             break;
620         }
621     }
622 
623     return ret;
624 }
625 
626 /* Called with no locks taken.  */
627 
flush_blks(QEMUFile * f)628 static int flush_blks(QEMUFile *f)
629 {
630     BlkMigBlock *blk;
631     int ret = 0;
632 
633     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
634             __func__, block_mig_state.submitted, block_mig_state.read_done,
635             block_mig_state.transferred);
636 
637     blk_mig_lock();
638     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
639         if (qemu_file_rate_limit(f)) {
640             break;
641         }
642         if (blk->ret < 0) {
643             ret = blk->ret;
644             break;
645         }
646 
647         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
648         blk_mig_unlock();
649         blk_send(f, blk);
650         blk_mig_lock();
651 
652         g_free(blk->buf);
653         g_free(blk);
654 
655         block_mig_state.read_done--;
656         block_mig_state.transferred++;
657         assert(block_mig_state.read_done >= 0);
658     }
659     blk_mig_unlock();
660 
661     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __func__,
662             block_mig_state.submitted, block_mig_state.read_done,
663             block_mig_state.transferred);
664     return ret;
665 }
666 
667 /* Called with iothread lock taken.  */
668 
get_remaining_dirty(void)669 static int64_t get_remaining_dirty(void)
670 {
671     BlkMigDevState *bmds;
672     int64_t dirty = 0;
673 
674     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
675         aio_context_acquire(blk_get_aio_context(bmds->blk));
676         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
677         aio_context_release(blk_get_aio_context(bmds->blk));
678     }
679 
680     return dirty;
681 }
682 
683 
684 
685 /* Called with iothread lock taken.  */
block_migration_cleanup_bmds(void)686 static void block_migration_cleanup_bmds(void)
687 {
688     BlkMigDevState *bmds;
689     AioContext *ctx;
690 
691     unset_dirty_tracking();
692 
693     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
694         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
695         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
696         error_free(bmds->blocker);
697 
698         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
699         ctx = blk_get_aio_context(bmds->blk);
700         aio_context_acquire(ctx);
701         blk_unref(bmds->blk);
702         aio_context_release(ctx);
703 
704         g_free(bmds->blk_name);
705         g_free(bmds->aio_bitmap);
706         g_free(bmds);
707     }
708 }
709 
710 /* Called with iothread lock taken.  */
block_migration_cleanup(void * opaque)711 static void block_migration_cleanup(void *opaque)
712 {
713     BlkMigBlock *blk;
714 
715     bdrv_drain_all();
716 
717     block_migration_cleanup_bmds();
718 
719     blk_mig_lock();
720     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
721         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
722         g_free(blk->buf);
723         g_free(blk);
724     }
725     blk_mig_unlock();
726 }
727 
block_save_setup(QEMUFile * f,void * opaque)728 static int block_save_setup(QEMUFile *f, void *opaque)
729 {
730     int ret;
731 
732     DPRINTF("Enter save live setup submitted %d transferred %d\n",
733             block_mig_state.submitted, block_mig_state.transferred);
734 
735     qemu_mutex_lock_iothread();
736     ret = init_blk_migration(f);
737     if (ret < 0) {
738         qemu_mutex_unlock_iothread();
739         return ret;
740     }
741 
742     /* start track dirty blocks */
743     ret = set_dirty_tracking();
744 
745     qemu_mutex_unlock_iothread();
746 
747     if (ret) {
748         return ret;
749     }
750 
751     ret = flush_blks(f);
752     blk_mig_reset_dirty_cursor();
753     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
754 
755     return ret;
756 }
757 
block_save_iterate(QEMUFile * f,void * opaque)758 static int block_save_iterate(QEMUFile *f, void *opaque)
759 {
760     int ret;
761     int64_t last_ftell = qemu_ftell(f);
762     int64_t delta_ftell;
763 
764     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
765             block_mig_state.submitted, block_mig_state.transferred);
766 
767     ret = flush_blks(f);
768     if (ret) {
769         return ret;
770     }
771 
772     blk_mig_reset_dirty_cursor();
773 
774     /* control the rate of transfer */
775     blk_mig_lock();
776     while (block_mig_state.read_done * BLOCK_SIZE <
777            qemu_file_get_rate_limit(f) &&
778            block_mig_state.submitted < MAX_PARALLEL_IO &&
779            (block_mig_state.submitted + block_mig_state.read_done) <
780            MAX_IO_BUFFERS) {
781         blk_mig_unlock();
782         if (block_mig_state.bulk_completed == 0) {
783             /* first finish the bulk phase */
784             if (blk_mig_save_bulked_block(f) == 0) {
785                 /* finished saving bulk on all devices */
786                 block_mig_state.bulk_completed = 1;
787             }
788             ret = 0;
789         } else {
790             /* Always called with iothread lock taken for
791              * simplicity, block_save_complete also calls it.
792              */
793             qemu_mutex_lock_iothread();
794             ret = blk_mig_save_dirty_block(f, 1);
795             qemu_mutex_unlock_iothread();
796         }
797         if (ret < 0) {
798             return ret;
799         }
800         blk_mig_lock();
801         if (ret != 0) {
802             /* no more dirty blocks */
803             break;
804         }
805     }
806     blk_mig_unlock();
807 
808     ret = flush_blks(f);
809     if (ret) {
810         return ret;
811     }
812 
813     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
814     delta_ftell = qemu_ftell(f) - last_ftell;
815     if (delta_ftell > 0) {
816         return 1;
817     } else if (delta_ftell < 0) {
818         return -1;
819     } else {
820         return 0;
821     }
822 }
823 
824 /* Called with iothread lock taken.  */
825 
block_save_complete(QEMUFile * f,void * opaque)826 static int block_save_complete(QEMUFile *f, void *opaque)
827 {
828     int ret;
829 
830     DPRINTF("Enter save live complete submitted %d transferred %d\n",
831             block_mig_state.submitted, block_mig_state.transferred);
832 
833     ret = flush_blks(f);
834     if (ret) {
835         return ret;
836     }
837 
838     blk_mig_reset_dirty_cursor();
839 
840     /* we know for sure that save bulk is completed and
841        all async read completed */
842     blk_mig_lock();
843     assert(block_mig_state.submitted == 0);
844     blk_mig_unlock();
845 
846     do {
847         ret = blk_mig_save_dirty_block(f, 0);
848         if (ret < 0) {
849             return ret;
850         }
851     } while (ret == 0);
852 
853     /* report completion */
854     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
855 
856     DPRINTF("Block migration completed\n");
857 
858     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
859 
860     /* Make sure that our BlockBackends are gone, so that the block driver
861      * nodes can be inactivated. */
862     block_migration_cleanup_bmds();
863 
864     return 0;
865 }
866 
block_save_pending(QEMUFile * f,void * opaque,uint64_t max_size,uint64_t * res_precopy_only,uint64_t * res_compatible,uint64_t * res_postcopy_only)867 static void block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size,
868                                uint64_t *res_precopy_only,
869                                uint64_t *res_compatible,
870                                uint64_t *res_postcopy_only)
871 {
872     /* Estimate pending number of bytes to send */
873     uint64_t pending;
874 
875     qemu_mutex_lock_iothread();
876     pending = get_remaining_dirty();
877     qemu_mutex_unlock_iothread();
878 
879     blk_mig_lock();
880     pending += block_mig_state.submitted * BLOCK_SIZE +
881                block_mig_state.read_done * BLOCK_SIZE;
882     blk_mig_unlock();
883 
884     /* Report at least one block pending during bulk phase */
885     if (pending <= max_size && !block_mig_state.bulk_completed) {
886         pending = max_size + BLOCK_SIZE;
887     }
888 
889     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
890     /* We don't do postcopy */
891     *res_precopy_only += pending;
892 }
893 
block_load(QEMUFile * f,void * opaque,int version_id)894 static int block_load(QEMUFile *f, void *opaque, int version_id)
895 {
896     static int banner_printed;
897     int len, flags;
898     char device_name[256];
899     int64_t addr;
900     BlockBackend *blk, *blk_prev = NULL;
901     Error *local_err = NULL;
902     uint8_t *buf;
903     int64_t total_sectors = 0;
904     int nr_sectors;
905     int ret;
906     BlockDriverInfo bdi;
907     int cluster_size = BLOCK_SIZE;
908 
909     do {
910         addr = qemu_get_be64(f);
911 
912         flags = addr & ~BDRV_SECTOR_MASK;
913         addr >>= BDRV_SECTOR_BITS;
914 
915         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
916             /* get device name */
917             len = qemu_get_byte(f);
918             qemu_get_buffer(f, (uint8_t *)device_name, len);
919             device_name[len] = '\0';
920 
921             blk = blk_by_name(device_name);
922             if (!blk) {
923                 fprintf(stderr, "Error unknown block device %s\n",
924                         device_name);
925                 return -EINVAL;
926             }
927 
928             if (blk != blk_prev) {
929                 blk_prev = blk;
930                 total_sectors = blk_nb_sectors(blk);
931                 if (total_sectors <= 0) {
932                     error_report("Error getting length of block device %s",
933                                  device_name);
934                     return -EINVAL;
935                 }
936 
937                 blk_invalidate_cache(blk, &local_err);
938                 if (local_err) {
939                     error_report_err(local_err);
940                     return -EINVAL;
941                 }
942 
943                 ret = bdrv_get_info(blk_bs(blk), &bdi);
944                 if (ret == 0 && bdi.cluster_size > 0 &&
945                     bdi.cluster_size <= BLOCK_SIZE &&
946                     BLOCK_SIZE % bdi.cluster_size == 0) {
947                     cluster_size = bdi.cluster_size;
948                 } else {
949                     cluster_size = BLOCK_SIZE;
950                 }
951             }
952 
953             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
954                 nr_sectors = total_sectors - addr;
955             } else {
956                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
957             }
958 
959             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
960                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
961                                         nr_sectors * BDRV_SECTOR_SIZE,
962                                         BDRV_REQ_MAY_UNMAP);
963             } else {
964                 int i;
965                 int64_t cur_addr;
966                 uint8_t *cur_buf;
967 
968                 buf = g_malloc(BLOCK_SIZE);
969                 qemu_get_buffer(f, buf, BLOCK_SIZE);
970                 for (i = 0; i < BLOCK_SIZE / cluster_size; i++) {
971                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
972                     cur_buf = buf + i * cluster_size;
973 
974                     if ((!block_mig_state.zero_blocks ||
975                         cluster_size < BLOCK_SIZE) &&
976                         buffer_is_zero(cur_buf, cluster_size)) {
977                         ret = blk_pwrite_zeroes(blk, cur_addr,
978                                                 cluster_size,
979                                                 BDRV_REQ_MAY_UNMAP);
980                     } else {
981                         ret = blk_pwrite(blk, cur_addr, cur_buf,
982                                          cluster_size, 0);
983                     }
984                     if (ret < 0) {
985                         break;
986                     }
987                 }
988                 g_free(buf);
989             }
990 
991             if (ret < 0) {
992                 return ret;
993             }
994         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
995             if (!banner_printed) {
996                 printf("Receiving block device images\n");
997                 banner_printed = 1;
998             }
999             printf("Completed %d %%%c", (int)addr,
1000                    (addr == 100) ? '\n' : '\r');
1001             fflush(stdout);
1002         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
1003             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
1004             return -EINVAL;
1005         }
1006         ret = qemu_file_get_error(f);
1007         if (ret != 0) {
1008             return ret;
1009         }
1010     } while (!(flags & BLK_MIG_FLAG_EOS));
1011 
1012     return 0;
1013 }
1014 
block_is_active(void * opaque)1015 static bool block_is_active(void *opaque)
1016 {
1017     return migrate_use_block();
1018 }
1019 
1020 static SaveVMHandlers savevm_block_handlers = {
1021     .save_setup = block_save_setup,
1022     .save_live_iterate = block_save_iterate,
1023     .save_live_complete_precopy = block_save_complete,
1024     .save_live_pending = block_save_pending,
1025     .load_state = block_load,
1026     .save_cleanup = block_migration_cleanup,
1027     .is_active = block_is_active,
1028 };
1029 
blk_mig_init(void)1030 void blk_mig_init(void)
1031 {
1032     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1033     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1034     qemu_mutex_init(&block_mig_state.lock);
1035 
1036     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
1037                          &block_mig_state);
1038 }
1039