1 /*
2  * Block dirty bitmap postcopy migration
3  *
4  * Copyright IBM, Corp. 2009
5  * Copyright (c) 2016-2017 Virtuozzo International GmbH. All rights reserved.
6  *
7  * Authors:
8  *  Liran Schour   <lirans@il.ibm.com>
9  *  Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
10  *
11  * This work is licensed under the terms of the GNU GPL, version 2.  See
12  * the COPYING file in the top-level directory.
13  * This file is derived from migration/block.c, so it's author and IBM copyright
14  * are here, although content is quite different.
15  *
16  * Contributions after 2012-01-13 are licensed under the terms of the
17  * GNU GPL, version 2 or (at your option) any later version.
18  *
19  *                                ***
20  *
21  * Here postcopy migration of dirty bitmaps is realized. Only QMP-addressable
22  * bitmaps are migrated.
23  *
24  * Bitmap migration implies creating bitmap with the same name and granularity
25  * in destination QEMU. If the bitmap with the same name (for the same node)
26  * already exists on destination an error will be generated.
27  *
28  * format of migration:
29  *
30  * # Header (shared for different chunk types)
31  * 1, 2 or 4 bytes: flags (see qemu_{put,put}_flags)
32  * [ 1 byte: node alias size ] \  flags & DEVICE_NAME
33  * [ n bytes: node alias     ] /
34  * [ 1 byte: bitmap alias size ] \  flags & BITMAP_NAME
35  * [ n bytes: bitmap alias     ] /
36  *
37  * # Start of bitmap migration (flags & START)
38  * header
39  * be64: granularity
40  * 1 byte: bitmap flags (corresponds to BdrvDirtyBitmap)
41  *   bit 0    -  bitmap is enabled
42  *   bit 1    -  bitmap is persistent
43  *   bit 2    -  bitmap is autoloading
44  *   bits 3-7 - reserved, must be zero
45  *
46  * # Complete of bitmap migration (flags & COMPLETE)
47  * header
48  *
49  * # Data chunk of bitmap migration
50  * header
51  * be64: start sector
52  * be32: number of sectors
53  * [ be64: buffer size  ] \ ! (flags & ZEROES)
54  * [ n bytes: buffer    ] /
55  *
56  * The last chunk in stream should contain flags & EOS. The chunk may skip
57  * device and/or bitmap names, assuming them to be the same with the previous
58  * chunk.
59  */
60 
61 #include "qemu/osdep.h"
62 #include "block/block.h"
63 #include "block/block_int.h"
64 #include "block/dirty-bitmap.h"
65 #include "sysemu/block-backend.h"
66 #include "sysemu/runstate.h"
67 #include "qemu/main-loop.h"
68 #include "qemu/error-report.h"
69 #include "migration/misc.h"
70 #include "migration/migration.h"
71 #include "qemu-file.h"
72 #include "migration/vmstate.h"
73 #include "migration/register.h"
74 #include "qemu/hbitmap.h"
75 #include "qemu/cutils.h"
76 #include "qemu/id.h"
77 #include "qapi/error.h"
78 #include "qapi/qapi-commands-migration.h"
79 #include "qapi/qapi-visit-migration.h"
80 #include "qapi/clone-visitor.h"
81 #include "trace.h"
82 #include "options.h"
83 
84 #define CHUNK_SIZE     (1 << 10)
85 
86 /* Flags occupy one, two or four bytes (Big Endian). The size is determined as
87  * follows:
88  * in first (most significant) byte bit 8 is clear  -->  one byte
89  * in first byte bit 8 is set    -->  two or four bytes, depending on second
90  *                                    byte:
91  *    | in second byte bit 8 is clear  -->  two bytes
92  *    | in second byte bit 8 is set    -->  four bytes
93  */
94 #define DIRTY_BITMAP_MIG_FLAG_EOS           0x01
95 #define DIRTY_BITMAP_MIG_FLAG_ZEROES        0x02
96 #define DIRTY_BITMAP_MIG_FLAG_BITMAP_NAME   0x04
97 #define DIRTY_BITMAP_MIG_FLAG_DEVICE_NAME   0x08
98 #define DIRTY_BITMAP_MIG_FLAG_START         0x10
99 #define DIRTY_BITMAP_MIG_FLAG_COMPLETE      0x20
100 #define DIRTY_BITMAP_MIG_FLAG_BITS          0x40
101 
102 #define DIRTY_BITMAP_MIG_EXTRA_FLAGS        0x80
103 
104 #define DIRTY_BITMAP_MIG_START_FLAG_ENABLED          0x01
105 #define DIRTY_BITMAP_MIG_START_FLAG_PERSISTENT       0x02
106 /* 0x04 was "AUTOLOAD" flags on older versions, now it is ignored */
107 #define DIRTY_BITMAP_MIG_START_FLAG_RESERVED_MASK    0xf8
108 
109 /* State of one bitmap during save process */
110 typedef struct SaveBitmapState {
111     /* Written during setup phase. */
112     BlockDriverState *bs;
113     char *node_alias;
114     char *bitmap_alias;
115     BdrvDirtyBitmap *bitmap;
116     uint64_t total_sectors;
117     uint64_t sectors_per_chunk;
118     QSIMPLEQ_ENTRY(SaveBitmapState) entry;
119     uint8_t flags;
120 
121     /* For bulk phase. */
122     bool bulk_completed;
123     uint64_t cur_sector;
124 } SaveBitmapState;
125 
126 /* State of the dirty bitmap migration (DBM) during save process */
127 typedef struct DBMSaveState {
128     QSIMPLEQ_HEAD(, SaveBitmapState) dbms_list;
129 
130     bool bulk_completed;
131     bool no_bitmaps;
132 
133     /* for send_bitmap_bits() */
134     BlockDriverState *prev_bs;
135     BdrvDirtyBitmap *prev_bitmap;
136 } DBMSaveState;
137 
138 typedef struct LoadBitmapState {
139     BlockDriverState *bs;
140     BdrvDirtyBitmap *bitmap;
141     bool migrated;
142     bool enabled;
143 } LoadBitmapState;
144 
145 /* State of the dirty bitmap migration (DBM) during load process */
146 typedef struct DBMLoadState {
147     uint32_t flags;
148     char node_alias[256];
149     char bitmap_alias[256];
150     char bitmap_name[BDRV_BITMAP_MAX_NAME_SIZE + 1];
151     BlockDriverState *bs;
152     BdrvDirtyBitmap *bitmap;
153 
154     bool before_vm_start_handled; /* set in dirty_bitmap_mig_before_vm_start */
155     BitmapMigrationBitmapAlias *bmap_inner;
156 
157     /*
158      * cancelled
159      * Incoming migration is cancelled for some reason. That means that we
160      * still should read our chunks from migration stream, to not affect other
161      * migration objects (like RAM), but just ignore them and do not touch any
162      * bitmaps or nodes.
163      */
164     bool cancelled;
165 
166     GSList *bitmaps;
167     QemuMutex lock; /* protect bitmaps */
168 } DBMLoadState;
169 
170 typedef struct DBMState {
171     DBMSaveState save;
172     DBMLoadState load;
173 } DBMState;
174 
175 static DBMState dbm_state;
176 
177 /* For hash tables that map node/bitmap names to aliases */
178 typedef struct AliasMapInnerNode {
179     char *string;
180     GHashTable *subtree;
181 } AliasMapInnerNode;
182 
183 static void free_alias_map_inner_node(void *amin_ptr)
184 {
185     AliasMapInnerNode *amin = amin_ptr;
186 
187     g_free(amin->string);
188     g_hash_table_unref(amin->subtree);
189     g_free(amin);
190 }
191 
192 /**
193  * Construct an alias map based on the given QMP structure.
194  *
195  * (Note that we cannot store such maps in the MigrationParameters
196  * object, because that struct is defined by the QAPI schema, which
197  * makes it basically impossible to have dicts with arbitrary keys.
198  * Therefore, we instead have to construct these maps when migration
199  * starts.)
200  *
201  * @bbm is the block_bitmap_mapping from the migration parameters.
202  *
203  * If @name_to_alias is true, the returned hash table will map node
204  * and bitmap names to their respective aliases (for outgoing
205  * migration).
206  *
207  * If @name_to_alias is false, the returned hash table will map node
208  * and bitmap aliases to their respective names (for incoming
209  * migration).
210  *
211  * The hash table maps node names/aliases to AliasMapInnerNode
212  * objects, whose .string is the respective node alias/name, and whose
213  * .subtree table maps bitmap names/aliases to the respective bitmap
214  * alias/name.
215  */
216 static GHashTable *construct_alias_map(const BitmapMigrationNodeAliasList *bbm,
217                                        bool name_to_alias,
218                                        Error **errp)
219 {
220     GHashTable *alias_map;
221     size_t max_node_name_len = sizeof_field(BlockDriverState, node_name) - 1;
222 
223     alias_map = g_hash_table_new_full(g_str_hash, g_str_equal,
224                                       g_free, free_alias_map_inner_node);
225 
226     for (; bbm; bbm = bbm->next) {
227         const BitmapMigrationNodeAlias *bmna = bbm->value;
228         const BitmapMigrationBitmapAliasList *bmbal;
229         AliasMapInnerNode *amin;
230         GHashTable *bitmaps_map;
231         const char *node_map_from, *node_map_to;
232         GDestroyNotify gdn;
233 
234         if (!id_wellformed(bmna->alias)) {
235             error_setg(errp, "The node alias '%s' is not well-formed",
236                        bmna->alias);
237             goto fail;
238         }
239 
240         if (strlen(bmna->alias) > UINT8_MAX) {
241             error_setg(errp, "The node alias '%s' is longer than %u bytes",
242                        bmna->alias, UINT8_MAX);
243             goto fail;
244         }
245 
246         if (strlen(bmna->node_name) > max_node_name_len) {
247             error_setg(errp, "The node name '%s' is longer than %zu bytes",
248                        bmna->node_name, max_node_name_len);
249             goto fail;
250         }
251 
252         if (name_to_alias) {
253             if (g_hash_table_contains(alias_map, bmna->node_name)) {
254                 error_setg(errp, "The node name '%s' is mapped twice",
255                            bmna->node_name);
256                 goto fail;
257             }
258 
259             node_map_from = bmna->node_name;
260             node_map_to = bmna->alias;
261         } else {
262             if (g_hash_table_contains(alias_map, bmna->alias)) {
263                 error_setg(errp, "The node alias '%s' is used twice",
264                            bmna->alias);
265                 goto fail;
266             }
267 
268             node_map_from = bmna->alias;
269             node_map_to = bmna->node_name;
270         }
271 
272         gdn = (GDestroyNotify) qapi_free_BitmapMigrationBitmapAlias;
273         bitmaps_map = g_hash_table_new_full(g_str_hash, g_str_equal, g_free,
274                                             gdn);
275 
276         amin = g_new(AliasMapInnerNode, 1);
277         *amin = (AliasMapInnerNode){
278             .string = g_strdup(node_map_to),
279             .subtree = bitmaps_map,
280         };
281 
282         g_hash_table_insert(alias_map, g_strdup(node_map_from), amin);
283 
284         for (bmbal = bmna->bitmaps; bmbal; bmbal = bmbal->next) {
285             const BitmapMigrationBitmapAlias *bmba = bmbal->value;
286             const char *bmap_map_from;
287 
288             if (strlen(bmba->alias) > UINT8_MAX) {
289                 error_setg(errp,
290                            "The bitmap alias '%s' is longer than %u bytes",
291                            bmba->alias, UINT8_MAX);
292                 goto fail;
293             }
294 
295             if (strlen(bmba->name) > BDRV_BITMAP_MAX_NAME_SIZE) {
296                 error_setg(errp, "The bitmap name '%s' is longer than %d bytes",
297                            bmba->name, BDRV_BITMAP_MAX_NAME_SIZE);
298                 goto fail;
299             }
300 
301             if (name_to_alias) {
302                 bmap_map_from = bmba->name;
303 
304                 if (g_hash_table_contains(bitmaps_map, bmba->name)) {
305                     error_setg(errp, "The bitmap '%s'/'%s' is mapped twice",
306                                bmna->node_name, bmba->name);
307                     goto fail;
308                 }
309             } else {
310                 bmap_map_from = bmba->alias;
311 
312                 if (g_hash_table_contains(bitmaps_map, bmba->alias)) {
313                     error_setg(errp, "The bitmap alias '%s'/'%s' is used twice",
314                                bmna->alias, bmba->alias);
315                     goto fail;
316                 }
317             }
318 
319             g_hash_table_insert(bitmaps_map, g_strdup(bmap_map_from),
320                                 QAPI_CLONE(BitmapMigrationBitmapAlias, bmba));
321         }
322     }
323 
324     return alias_map;
325 
326 fail:
327     g_hash_table_destroy(alias_map);
328     return NULL;
329 }
330 
331 /**
332  * Run construct_alias_map() in both directions to check whether @bbm
333  * is valid.
334  * (This function is to be used by migration/migration.c to validate
335  * the user-specified block-bitmap-mapping migration parameter.)
336  *
337  * Returns true if and only if the mapping is valid.
338  */
339 bool check_dirty_bitmap_mig_alias_map(const BitmapMigrationNodeAliasList *bbm,
340                                       Error **errp)
341 {
342     GHashTable *alias_map;
343 
344     alias_map = construct_alias_map(bbm, true, errp);
345     if (!alias_map) {
346         return false;
347     }
348     g_hash_table_destroy(alias_map);
349 
350     alias_map = construct_alias_map(bbm, false, errp);
351     if (!alias_map) {
352         return false;
353     }
354     g_hash_table_destroy(alias_map);
355 
356     return true;
357 }
358 
359 static uint32_t qemu_get_bitmap_flags(QEMUFile *f)
360 {
361     uint8_t flags = qemu_get_byte(f);
362     if (flags & DIRTY_BITMAP_MIG_EXTRA_FLAGS) {
363         flags = flags << 8 | qemu_get_byte(f);
364         if (flags & DIRTY_BITMAP_MIG_EXTRA_FLAGS) {
365             flags = flags << 16 | qemu_get_be16(f);
366         }
367     }
368 
369     return flags;
370 }
371 
372 static void qemu_put_bitmap_flags(QEMUFile *f, uint32_t flags)
373 {
374     /* The code currently does not send flags as more than one byte */
375     assert(!(flags & (0xffffff00 | DIRTY_BITMAP_MIG_EXTRA_FLAGS)));
376 
377     qemu_put_byte(f, flags);
378 }
379 
380 static void send_bitmap_header(QEMUFile *f, DBMSaveState *s,
381                                SaveBitmapState *dbms, uint32_t additional_flags)
382 {
383     BlockDriverState *bs = dbms->bs;
384     BdrvDirtyBitmap *bitmap = dbms->bitmap;
385     uint32_t flags = additional_flags;
386     trace_send_bitmap_header_enter();
387 
388     if (bs != s->prev_bs) {
389         s->prev_bs = bs;
390         flags |= DIRTY_BITMAP_MIG_FLAG_DEVICE_NAME;
391     }
392 
393     if (bitmap != s->prev_bitmap) {
394         s->prev_bitmap = bitmap;
395         flags |= DIRTY_BITMAP_MIG_FLAG_BITMAP_NAME;
396     }
397 
398     qemu_put_bitmap_flags(f, flags);
399 
400     if (flags & DIRTY_BITMAP_MIG_FLAG_DEVICE_NAME) {
401         qemu_put_counted_string(f, dbms->node_alias);
402     }
403 
404     if (flags & DIRTY_BITMAP_MIG_FLAG_BITMAP_NAME) {
405         qemu_put_counted_string(f, dbms->bitmap_alias);
406     }
407 }
408 
409 static void send_bitmap_start(QEMUFile *f, DBMSaveState *s,
410                               SaveBitmapState *dbms)
411 {
412     send_bitmap_header(f, s, dbms, DIRTY_BITMAP_MIG_FLAG_START);
413     qemu_put_be32(f, bdrv_dirty_bitmap_granularity(dbms->bitmap));
414     qemu_put_byte(f, dbms->flags);
415 }
416 
417 static void send_bitmap_complete(QEMUFile *f, DBMSaveState *s,
418                                  SaveBitmapState *dbms)
419 {
420     send_bitmap_header(f, s, dbms, DIRTY_BITMAP_MIG_FLAG_COMPLETE);
421 }
422 
423 static void send_bitmap_bits(QEMUFile *f, DBMSaveState *s,
424                              SaveBitmapState *dbms,
425                              uint64_t start_sector, uint32_t nr_sectors)
426 {
427     /* align for buffer_is_zero() */
428     uint64_t align = 4 * sizeof(long);
429     uint64_t unaligned_size =
430         bdrv_dirty_bitmap_serialization_size(
431             dbms->bitmap, start_sector << BDRV_SECTOR_BITS,
432             (uint64_t)nr_sectors << BDRV_SECTOR_BITS);
433     uint64_t buf_size = QEMU_ALIGN_UP(unaligned_size, align);
434     uint8_t *buf = g_malloc0(buf_size);
435     uint32_t flags = DIRTY_BITMAP_MIG_FLAG_BITS;
436 
437     bdrv_dirty_bitmap_serialize_part(
438         dbms->bitmap, buf, start_sector << BDRV_SECTOR_BITS,
439         (uint64_t)nr_sectors << BDRV_SECTOR_BITS);
440 
441     if (buffer_is_zero(buf, buf_size)) {
442         g_free(buf);
443         buf = NULL;
444         flags |= DIRTY_BITMAP_MIG_FLAG_ZEROES;
445     }
446 
447     trace_send_bitmap_bits(flags, start_sector, nr_sectors, buf_size);
448 
449     send_bitmap_header(f, s, dbms, flags);
450 
451     qemu_put_be64(f, start_sector);
452     qemu_put_be32(f, nr_sectors);
453 
454     /* if a block is zero we need to flush here since the network
455      * bandwidth is now a lot higher than the storage device bandwidth.
456      * thus if we queue zero blocks we slow down the migration. */
457     if (flags & DIRTY_BITMAP_MIG_FLAG_ZEROES) {
458         qemu_fflush(f);
459     } else {
460         qemu_put_be64(f, buf_size);
461         qemu_put_buffer(f, buf, buf_size);
462     }
463 
464     g_free(buf);
465 }
466 
467 /* Called with iothread lock taken.  */
468 static void dirty_bitmap_do_save_cleanup(DBMSaveState *s)
469 {
470     SaveBitmapState *dbms;
471 
472     while ((dbms = QSIMPLEQ_FIRST(&s->dbms_list)) != NULL) {
473         QSIMPLEQ_REMOVE_HEAD(&s->dbms_list, entry);
474         bdrv_dirty_bitmap_set_busy(dbms->bitmap, false);
475         bdrv_unref(dbms->bs);
476         g_free(dbms->node_alias);
477         g_free(dbms->bitmap_alias);
478         g_free(dbms);
479     }
480 }
481 
482 /* Called with iothread lock taken. */
483 static int add_bitmaps_to_list(DBMSaveState *s, BlockDriverState *bs,
484                                const char *bs_name, GHashTable *alias_map)
485 {
486     BdrvDirtyBitmap *bitmap;
487     SaveBitmapState *dbms;
488     GHashTable *bitmap_aliases;
489     const char *node_alias, *bitmap_name, *bitmap_alias;
490     Error *local_err = NULL;
491 
492     /* When an alias map is given, @bs_name must be @bs's node name */
493     assert(!alias_map || !strcmp(bs_name, bdrv_get_node_name(bs)));
494 
495     FOR_EACH_DIRTY_BITMAP(bs, bitmap) {
496         if (bdrv_dirty_bitmap_name(bitmap)) {
497             break;
498         }
499     }
500     if (!bitmap) {
501         return 0;
502     }
503 
504     bitmap_name = bdrv_dirty_bitmap_name(bitmap);
505 
506     if (!bs_name || strcmp(bs_name, "") == 0) {
507         error_report("Bitmap '%s' in unnamed node can't be migrated",
508                      bitmap_name);
509         return -1;
510     }
511 
512     if (alias_map) {
513         const AliasMapInnerNode *amin = g_hash_table_lookup(alias_map, bs_name);
514 
515         if (!amin) {
516             /* Skip bitmaps on nodes with no alias */
517             return 0;
518         }
519 
520         node_alias = amin->string;
521         bitmap_aliases = amin->subtree;
522     } else {
523         node_alias = bs_name;
524         bitmap_aliases = NULL;
525     }
526 
527     if (node_alias[0] == '#') {
528         error_report("Bitmap '%s' in a node with auto-generated "
529                      "name '%s' can't be migrated",
530                      bitmap_name, node_alias);
531         return -1;
532     }
533 
534     FOR_EACH_DIRTY_BITMAP(bs, bitmap) {
535         BitmapMigrationBitmapAliasTransform *bitmap_transform = NULL;
536         bitmap_name = bdrv_dirty_bitmap_name(bitmap);
537         if (!bitmap_name) {
538             continue;
539         }
540 
541         if (bdrv_dirty_bitmap_check(bitmap, BDRV_BITMAP_DEFAULT, &local_err)) {
542             error_report_err(local_err);
543             return -1;
544         }
545 
546         if (bitmap_aliases) {
547             BitmapMigrationBitmapAlias *bmap_inner;
548 
549             bmap_inner = g_hash_table_lookup(bitmap_aliases, bitmap_name);
550             if (!bmap_inner) {
551                 /* Skip bitmaps with no alias */
552                 continue;
553             }
554 
555             bitmap_alias = bmap_inner->alias;
556             if (bmap_inner->transform) {
557                 bitmap_transform = bmap_inner->transform;
558             }
559         } else {
560             if (strlen(bitmap_name) > UINT8_MAX) {
561                 error_report("Cannot migrate bitmap '%s' on node '%s': "
562                              "Name is longer than %u bytes",
563                              bitmap_name, bs_name, UINT8_MAX);
564                 return -1;
565             }
566             bitmap_alias = bitmap_name;
567         }
568 
569         bdrv_ref(bs);
570         bdrv_dirty_bitmap_set_busy(bitmap, true);
571 
572         dbms = g_new0(SaveBitmapState, 1);
573         dbms->bs = bs;
574         dbms->node_alias = g_strdup(node_alias);
575         dbms->bitmap_alias = g_strdup(bitmap_alias);
576         dbms->bitmap = bitmap;
577         dbms->total_sectors = bdrv_nb_sectors(bs);
578         dbms->sectors_per_chunk = CHUNK_SIZE * 8LLU *
579             (bdrv_dirty_bitmap_granularity(bitmap) >> BDRV_SECTOR_BITS);
580         assert(dbms->sectors_per_chunk != 0);
581         if (bdrv_dirty_bitmap_enabled(bitmap)) {
582             dbms->flags |= DIRTY_BITMAP_MIG_START_FLAG_ENABLED;
583         }
584         if (bitmap_transform &&
585             bitmap_transform->has_persistent) {
586             if (bitmap_transform->persistent) {
587                 dbms->flags |= DIRTY_BITMAP_MIG_START_FLAG_PERSISTENT;
588             }
589         } else {
590             if (bdrv_dirty_bitmap_get_persistence(bitmap)) {
591                 dbms->flags |= DIRTY_BITMAP_MIG_START_FLAG_PERSISTENT;
592             }
593         }
594 
595         QSIMPLEQ_INSERT_TAIL(&s->dbms_list, dbms, entry);
596     }
597 
598     return 0;
599 }
600 
601 /* Called with iothread lock taken. */
602 static int init_dirty_bitmap_migration(DBMSaveState *s)
603 {
604     BlockDriverState *bs;
605     SaveBitmapState *dbms;
606     GHashTable *handled_by_blk = g_hash_table_new(NULL, NULL);
607     BlockBackend *blk;
608     GHashTable *alias_map = NULL;
609 
610     if (migrate_has_block_bitmap_mapping()) {
611         alias_map = construct_alias_map(migrate_block_bitmap_mapping(), true,
612                                         &error_abort);
613     }
614 
615     s->bulk_completed = false;
616     s->prev_bs = NULL;
617     s->prev_bitmap = NULL;
618     s->no_bitmaps = false;
619 
620     if (!alias_map) {
621         /*
622          * Use blockdevice name for direct (or filtered) children of named block
623          * backends.
624          */
625         for (blk = blk_next(NULL); blk; blk = blk_next(blk)) {
626             const char *name = blk_name(blk);
627 
628             if (!name || strcmp(name, "") == 0) {
629                 continue;
630             }
631 
632             bs = blk_bs(blk);
633 
634             /* Skip filters without bitmaps */
635             while (bs && bs->drv && bs->drv->is_filter &&
636                    !bdrv_has_named_bitmaps(bs))
637             {
638                 bs = bdrv_filter_bs(bs);
639             }
640 
641             if (bs && bs->drv && !bs->drv->is_filter) {
642                 if (add_bitmaps_to_list(s, bs, name, NULL)) {
643                     goto fail;
644                 }
645                 g_hash_table_add(handled_by_blk, bs);
646             }
647         }
648     }
649 
650     for (bs = bdrv_next_all_states(NULL); bs; bs = bdrv_next_all_states(bs)) {
651         if (g_hash_table_contains(handled_by_blk, bs)) {
652             continue;
653         }
654 
655         if (add_bitmaps_to_list(s, bs, bdrv_get_node_name(bs), alias_map)) {
656             goto fail;
657         }
658     }
659 
660     /* unset migration flags here, to not roll back it */
661     QSIMPLEQ_FOREACH(dbms, &s->dbms_list, entry) {
662         bdrv_dirty_bitmap_skip_store(dbms->bitmap, true);
663     }
664 
665     if (QSIMPLEQ_EMPTY(&s->dbms_list)) {
666         s->no_bitmaps = true;
667     }
668 
669     g_hash_table_destroy(handled_by_blk);
670     if (alias_map) {
671         g_hash_table_destroy(alias_map);
672     }
673 
674     return 0;
675 
676 fail:
677     g_hash_table_destroy(handled_by_blk);
678     if (alias_map) {
679         g_hash_table_destroy(alias_map);
680     }
681     dirty_bitmap_do_save_cleanup(s);
682 
683     return -1;
684 }
685 
686 /* Called with no lock taken.  */
687 static void bulk_phase_send_chunk(QEMUFile *f, DBMSaveState *s,
688                                   SaveBitmapState *dbms)
689 {
690     uint32_t nr_sectors = MIN(dbms->total_sectors - dbms->cur_sector,
691                              dbms->sectors_per_chunk);
692 
693     send_bitmap_bits(f, s, dbms, dbms->cur_sector, nr_sectors);
694 
695     dbms->cur_sector += nr_sectors;
696     if (dbms->cur_sector >= dbms->total_sectors) {
697         dbms->bulk_completed = true;
698     }
699 }
700 
701 /* Called with no lock taken.  */
702 static void bulk_phase(QEMUFile *f, DBMSaveState *s, bool limit)
703 {
704     SaveBitmapState *dbms;
705 
706     QSIMPLEQ_FOREACH(dbms, &s->dbms_list, entry) {
707         while (!dbms->bulk_completed) {
708             bulk_phase_send_chunk(f, s, dbms);
709             if (limit && migration_rate_exceeded(f)) {
710                 return;
711             }
712         }
713     }
714 
715     s->bulk_completed = true;
716 }
717 
718 /* for SaveVMHandlers */
719 static void dirty_bitmap_save_cleanup(void *opaque)
720 {
721     DBMSaveState *s = &((DBMState *)opaque)->save;
722 
723     dirty_bitmap_do_save_cleanup(s);
724 }
725 
726 static int dirty_bitmap_save_iterate(QEMUFile *f, void *opaque)
727 {
728     DBMSaveState *s = &((DBMState *)opaque)->save;
729 
730     trace_dirty_bitmap_save_iterate(migration_in_postcopy());
731 
732     if (migration_in_postcopy() && !s->bulk_completed) {
733         bulk_phase(f, s, true);
734     }
735 
736     qemu_put_bitmap_flags(f, DIRTY_BITMAP_MIG_FLAG_EOS);
737 
738     return s->bulk_completed;
739 }
740 
741 /* Called with iothread lock taken.  */
742 
743 static int dirty_bitmap_save_complete(QEMUFile *f, void *opaque)
744 {
745     DBMSaveState *s = &((DBMState *)opaque)->save;
746     SaveBitmapState *dbms;
747     trace_dirty_bitmap_save_complete_enter();
748 
749     if (!s->bulk_completed) {
750         bulk_phase(f, s, false);
751     }
752 
753     QSIMPLEQ_FOREACH(dbms, &s->dbms_list, entry) {
754         send_bitmap_complete(f, s, dbms);
755     }
756 
757     qemu_put_bitmap_flags(f, DIRTY_BITMAP_MIG_FLAG_EOS);
758 
759     trace_dirty_bitmap_save_complete_finish();
760 
761     dirty_bitmap_save_cleanup(opaque);
762     return 0;
763 }
764 
765 static void dirty_bitmap_state_pending(void *opaque,
766                                        uint64_t *must_precopy,
767                                        uint64_t *can_postcopy)
768 {
769     DBMSaveState *s = &((DBMState *)opaque)->save;
770     SaveBitmapState *dbms;
771     uint64_t pending = 0;
772 
773     qemu_mutex_lock_iothread();
774 
775     QSIMPLEQ_FOREACH(dbms, &s->dbms_list, entry) {
776         uint64_t gran = bdrv_dirty_bitmap_granularity(dbms->bitmap);
777         uint64_t sectors = dbms->bulk_completed ? 0 :
778                            dbms->total_sectors - dbms->cur_sector;
779 
780         pending += DIV_ROUND_UP(sectors * BDRV_SECTOR_SIZE, gran);
781     }
782 
783     qemu_mutex_unlock_iothread();
784 
785     trace_dirty_bitmap_state_pending(pending);
786 
787     *can_postcopy += pending;
788 }
789 
790 /* First occurrence of this bitmap. It should be created if doesn't exist */
791 static int dirty_bitmap_load_start(QEMUFile *f, DBMLoadState *s)
792 {
793     Error *local_err = NULL;
794     uint32_t granularity = qemu_get_be32(f);
795     uint8_t flags = qemu_get_byte(f);
796     LoadBitmapState *b;
797     bool persistent;
798 
799     if (s->cancelled) {
800         return 0;
801     }
802 
803     if (s->bitmap) {
804         error_report("Bitmap with the same name ('%s') already exists on "
805                      "destination", bdrv_dirty_bitmap_name(s->bitmap));
806         return -EINVAL;
807     } else {
808         s->bitmap = bdrv_create_dirty_bitmap(s->bs, granularity,
809                                              s->bitmap_name, &local_err);
810         if (!s->bitmap) {
811             error_report_err(local_err);
812             return -EINVAL;
813         }
814     }
815 
816     if (flags & DIRTY_BITMAP_MIG_START_FLAG_RESERVED_MASK) {
817         error_report("Unknown flags in migrated dirty bitmap header: %x",
818                      flags);
819         return -EINVAL;
820     }
821 
822     if (s->bmap_inner &&
823         s->bmap_inner->transform &&
824         s->bmap_inner->transform->has_persistent) {
825         persistent = s->bmap_inner->transform->persistent;
826     } else {
827         persistent = flags & DIRTY_BITMAP_MIG_START_FLAG_PERSISTENT;
828     }
829 
830     if (persistent) {
831         bdrv_dirty_bitmap_set_persistence(s->bitmap, true);
832     }
833 
834     bdrv_disable_dirty_bitmap(s->bitmap);
835     if (flags & DIRTY_BITMAP_MIG_START_FLAG_ENABLED) {
836         bdrv_dirty_bitmap_create_successor(s->bitmap, &local_err);
837         if (local_err) {
838             error_report_err(local_err);
839             return -EINVAL;
840         }
841     } else {
842         bdrv_dirty_bitmap_set_busy(s->bitmap, true);
843     }
844 
845     b = g_new(LoadBitmapState, 1);
846     b->bs = s->bs;
847     b->bitmap = s->bitmap;
848     b->migrated = false;
849     b->enabled = flags & DIRTY_BITMAP_MIG_START_FLAG_ENABLED;
850 
851     s->bitmaps = g_slist_prepend(s->bitmaps, b);
852 
853     return 0;
854 }
855 
856 /*
857  * before_vm_start_handle_item
858  *
859  * g_slist_foreach helper
860  *
861  * item is LoadBitmapState*
862  * opaque is DBMLoadState*
863  */
864 static void before_vm_start_handle_item(void *item, void *opaque)
865 {
866     DBMLoadState *s = opaque;
867     LoadBitmapState *b = item;
868 
869     if (b->enabled) {
870         if (b->migrated) {
871             bdrv_enable_dirty_bitmap(b->bitmap);
872         } else {
873             bdrv_dirty_bitmap_enable_successor(b->bitmap);
874         }
875     }
876 
877     if (b->migrated) {
878         s->bitmaps = g_slist_remove(s->bitmaps, b);
879         g_free(b);
880     }
881 }
882 
883 void dirty_bitmap_mig_before_vm_start(void)
884 {
885     DBMLoadState *s = &dbm_state.load;
886     qemu_mutex_lock(&s->lock);
887 
888     assert(!s->before_vm_start_handled);
889     g_slist_foreach(s->bitmaps, before_vm_start_handle_item, s);
890     s->before_vm_start_handled = true;
891 
892     qemu_mutex_unlock(&s->lock);
893 }
894 
895 static void cancel_incoming_locked(DBMLoadState *s)
896 {
897     GSList *item;
898 
899     if (s->cancelled) {
900         return;
901     }
902 
903     s->cancelled = true;
904     s->bs = NULL;
905     s->bitmap = NULL;
906 
907     /* Drop all unfinished bitmaps */
908     for (item = s->bitmaps; item; item = g_slist_next(item)) {
909         LoadBitmapState *b = item->data;
910 
911         /*
912          * Bitmap must be unfinished, as finished bitmaps should already be
913          * removed from the list.
914          */
915         assert(!s->before_vm_start_handled || !b->migrated);
916         if (bdrv_dirty_bitmap_has_successor(b->bitmap)) {
917             bdrv_reclaim_dirty_bitmap(b->bitmap, &error_abort);
918         } else {
919             bdrv_dirty_bitmap_set_busy(b->bitmap, false);
920         }
921         bdrv_release_dirty_bitmap(b->bitmap);
922     }
923 
924     g_slist_free_full(s->bitmaps, g_free);
925     s->bitmaps = NULL;
926 }
927 
928 void dirty_bitmap_mig_cancel_outgoing(void)
929 {
930     dirty_bitmap_do_save_cleanup(&dbm_state.save);
931 }
932 
933 void dirty_bitmap_mig_cancel_incoming(void)
934 {
935     DBMLoadState *s = &dbm_state.load;
936 
937     qemu_mutex_lock(&s->lock);
938 
939     cancel_incoming_locked(s);
940 
941     qemu_mutex_unlock(&s->lock);
942 }
943 
944 static void dirty_bitmap_load_complete(QEMUFile *f, DBMLoadState *s)
945 {
946     GSList *item;
947     trace_dirty_bitmap_load_complete();
948 
949     if (s->cancelled) {
950         return;
951     }
952 
953     bdrv_dirty_bitmap_deserialize_finish(s->bitmap);
954 
955     if (bdrv_dirty_bitmap_has_successor(s->bitmap)) {
956         bdrv_reclaim_dirty_bitmap(s->bitmap, &error_abort);
957     } else {
958         bdrv_dirty_bitmap_set_busy(s->bitmap, false);
959     }
960 
961     for (item = s->bitmaps; item; item = g_slist_next(item)) {
962         LoadBitmapState *b = item->data;
963 
964         if (b->bitmap == s->bitmap) {
965             b->migrated = true;
966             if (s->before_vm_start_handled) {
967                 s->bitmaps = g_slist_remove(s->bitmaps, b);
968                 g_free(b);
969             }
970             break;
971         }
972     }
973 }
974 
975 static int dirty_bitmap_load_bits(QEMUFile *f, DBMLoadState *s)
976 {
977     uint64_t first_byte = qemu_get_be64(f) << BDRV_SECTOR_BITS;
978     uint64_t nr_bytes = (uint64_t)qemu_get_be32(f) << BDRV_SECTOR_BITS;
979     trace_dirty_bitmap_load_bits_enter(first_byte >> BDRV_SECTOR_BITS,
980                                        nr_bytes >> BDRV_SECTOR_BITS);
981 
982     if (s->flags & DIRTY_BITMAP_MIG_FLAG_ZEROES) {
983         trace_dirty_bitmap_load_bits_zeroes();
984         if (!s->cancelled) {
985             bdrv_dirty_bitmap_deserialize_zeroes(s->bitmap, first_byte,
986                                                  nr_bytes, false);
987         }
988     } else {
989         size_t ret;
990         g_autofree uint8_t *buf = NULL;
991         uint64_t buf_size = qemu_get_be64(f);
992         uint64_t needed_size;
993 
994         /*
995          * The actual check for buf_size is done a bit later. We can't do it in
996          * cancelled mode as we don't have the bitmap to check the constraints
997          * (so, we allocate a buffer and read prior to the check). On the other
998          * hand, we shouldn't blindly g_malloc the number from the stream.
999          * Actually one chunk should not be larger than CHUNK_SIZE. Let's allow
1000          * a bit larger (which means that bitmap migration will fail anyway and
1001          * the whole migration will most probably fail soon due to broken
1002          * stream).
1003          */
1004         if (buf_size > 10 * CHUNK_SIZE) {
1005             error_report("Bitmap migration stream buffer allocation request "
1006                          "is too large");
1007             return -EIO;
1008         }
1009 
1010         buf = g_malloc(buf_size);
1011         ret = qemu_get_buffer(f, buf, buf_size);
1012         if (ret != buf_size) {
1013             error_report("Failed to read bitmap bits");
1014             return -EIO;
1015         }
1016 
1017         if (s->cancelled) {
1018             return 0;
1019         }
1020 
1021         needed_size = bdrv_dirty_bitmap_serialization_size(s->bitmap,
1022                                                            first_byte,
1023                                                            nr_bytes);
1024 
1025         if (needed_size > buf_size ||
1026             buf_size > QEMU_ALIGN_UP(needed_size, 4 * sizeof(long))
1027              /* Here used same alignment as in send_bitmap_bits */
1028         ) {
1029             error_report("Migrated bitmap granularity doesn't "
1030                          "match the destination bitmap '%s' granularity",
1031                          bdrv_dirty_bitmap_name(s->bitmap));
1032             cancel_incoming_locked(s);
1033             return 0;
1034         }
1035 
1036         bdrv_dirty_bitmap_deserialize_part(s->bitmap, buf, first_byte, nr_bytes,
1037                                            false);
1038     }
1039 
1040     return 0;
1041 }
1042 
1043 static int dirty_bitmap_load_header(QEMUFile *f, DBMLoadState *s,
1044                                     GHashTable *alias_map)
1045 {
1046     GHashTable *bitmap_alias_map = NULL;
1047     Error *local_err = NULL;
1048     bool nothing;
1049     s->flags = qemu_get_bitmap_flags(f);
1050     trace_dirty_bitmap_load_header(s->flags);
1051 
1052     nothing = s->flags == (s->flags & DIRTY_BITMAP_MIG_FLAG_EOS);
1053 
1054     if (s->flags & DIRTY_BITMAP_MIG_FLAG_DEVICE_NAME) {
1055         if (!qemu_get_counted_string(f, s->node_alias)) {
1056             error_report("Unable to read node alias string");
1057             return -EINVAL;
1058         }
1059 
1060         if (!s->cancelled) {
1061             if (alias_map) {
1062                 const AliasMapInnerNode *amin;
1063 
1064                 amin = g_hash_table_lookup(alias_map, s->node_alias);
1065                 if (!amin) {
1066                     error_setg(&local_err, "Error: Unknown node alias '%s'",
1067                                s->node_alias);
1068                     s->bs = NULL;
1069                 } else {
1070                     bitmap_alias_map = amin->subtree;
1071                     s->bs = bdrv_lookup_bs(NULL, amin->string, &local_err);
1072                 }
1073             } else {
1074                 s->bs = bdrv_lookup_bs(s->node_alias, s->node_alias,
1075                                        &local_err);
1076             }
1077             if (!s->bs) {
1078                 error_report_err(local_err);
1079                 cancel_incoming_locked(s);
1080             }
1081         }
1082     } else if (s->bs) {
1083         if (alias_map) {
1084             const AliasMapInnerNode *amin;
1085 
1086             /* Must be present in the map, or s->bs would not be set */
1087             amin = g_hash_table_lookup(alias_map, s->node_alias);
1088             assert(amin != NULL);
1089 
1090             bitmap_alias_map = amin->subtree;
1091         }
1092     } else if (!nothing && !s->cancelled) {
1093         error_report("Error: block device name is not set");
1094         cancel_incoming_locked(s);
1095     }
1096 
1097     assert(nothing || s->cancelled || !!alias_map == !!bitmap_alias_map);
1098 
1099     if (s->flags & DIRTY_BITMAP_MIG_FLAG_BITMAP_NAME) {
1100         const char *bitmap_name;
1101 
1102         if (!qemu_get_counted_string(f, s->bitmap_alias)) {
1103             error_report("Unable to read bitmap alias string");
1104             return -EINVAL;
1105         }
1106 
1107         bitmap_name = s->bitmap_alias;
1108         if (!s->cancelled && bitmap_alias_map) {
1109             BitmapMigrationBitmapAlias *bmap_inner;
1110 
1111             bmap_inner = g_hash_table_lookup(bitmap_alias_map, s->bitmap_alias);
1112             if (!bmap_inner) {
1113                 error_report("Error: Unknown bitmap alias '%s' on node "
1114                              "'%s' (alias '%s')", s->bitmap_alias,
1115                              s->bs->node_name, s->node_alias);
1116                 cancel_incoming_locked(s);
1117             } else {
1118                 bitmap_name = bmap_inner->name;
1119             }
1120 
1121             s->bmap_inner = bmap_inner;
1122         }
1123 
1124         if (!s->cancelled) {
1125             g_strlcpy(s->bitmap_name, bitmap_name, sizeof(s->bitmap_name));
1126             s->bitmap = bdrv_find_dirty_bitmap(s->bs, s->bitmap_name);
1127 
1128             /*
1129              * bitmap may be NULL here, it wouldn't be an error if it is the
1130              * first occurrence of the bitmap
1131              */
1132             if (!s->bitmap && !(s->flags & DIRTY_BITMAP_MIG_FLAG_START)) {
1133                 error_report("Error: unknown dirty bitmap "
1134                              "'%s' for block device '%s'",
1135                              s->bitmap_name, s->bs->node_name);
1136                 cancel_incoming_locked(s);
1137             }
1138         }
1139     } else if (!s->bitmap && !nothing && !s->cancelled) {
1140         error_report("Error: block device name is not set");
1141         cancel_incoming_locked(s);
1142     }
1143 
1144     return 0;
1145 }
1146 
1147 /*
1148  * dirty_bitmap_load
1149  *
1150  * Load sequence of dirty bitmap chunks. Return error only on fatal io stream
1151  * violations. On other errors just cancel bitmaps incoming migration and return
1152  * 0.
1153  *
1154  * Note, than when incoming bitmap migration is canceled, we still must read all
1155  * our chunks (and just ignore them), to not affect other migration objects.
1156  */
1157 static int dirty_bitmap_load(QEMUFile *f, void *opaque, int version_id)
1158 {
1159     GHashTable *alias_map = NULL;
1160     DBMLoadState *s = &((DBMState *)opaque)->load;
1161     int ret = 0;
1162 
1163     trace_dirty_bitmap_load_enter();
1164 
1165     if (version_id != 1) {
1166         QEMU_LOCK_GUARD(&s->lock);
1167         cancel_incoming_locked(s);
1168         return -EINVAL;
1169     }
1170 
1171     if (migrate_has_block_bitmap_mapping()) {
1172         alias_map = construct_alias_map(migrate_block_bitmap_mapping(), false,
1173                                         &error_abort);
1174     }
1175 
1176     do {
1177         QEMU_LOCK_GUARD(&s->lock);
1178 
1179         ret = dirty_bitmap_load_header(f, s, alias_map);
1180         if (ret < 0) {
1181             cancel_incoming_locked(s);
1182             goto fail;
1183         }
1184 
1185         if (s->flags & DIRTY_BITMAP_MIG_FLAG_START) {
1186             ret = dirty_bitmap_load_start(f, s);
1187         } else if (s->flags & DIRTY_BITMAP_MIG_FLAG_COMPLETE) {
1188             dirty_bitmap_load_complete(f, s);
1189         } else if (s->flags & DIRTY_BITMAP_MIG_FLAG_BITS) {
1190             ret = dirty_bitmap_load_bits(f, s);
1191         }
1192 
1193         if (!ret) {
1194             ret = qemu_file_get_error(f);
1195         }
1196 
1197         if (ret) {
1198             cancel_incoming_locked(s);
1199             goto fail;
1200         }
1201     } while (!(s->flags & DIRTY_BITMAP_MIG_FLAG_EOS));
1202 
1203     trace_dirty_bitmap_load_success();
1204     ret = 0;
1205 fail:
1206     if (alias_map) {
1207         g_hash_table_destroy(alias_map);
1208     }
1209     return ret;
1210 }
1211 
1212 static int dirty_bitmap_save_setup(QEMUFile *f, void *opaque)
1213 {
1214     DBMSaveState *s = &((DBMState *)opaque)->save;
1215     SaveBitmapState *dbms = NULL;
1216 
1217     if (init_dirty_bitmap_migration(s) < 0) {
1218         return -1;
1219     }
1220 
1221     QSIMPLEQ_FOREACH(dbms, &s->dbms_list, entry) {
1222         send_bitmap_start(f, s, dbms);
1223     }
1224     qemu_put_bitmap_flags(f, DIRTY_BITMAP_MIG_FLAG_EOS);
1225     return 0;
1226 }
1227 
1228 static bool dirty_bitmap_is_active(void *opaque)
1229 {
1230     DBMSaveState *s = &((DBMState *)opaque)->save;
1231 
1232     return migrate_dirty_bitmaps() && !s->no_bitmaps;
1233 }
1234 
1235 static bool dirty_bitmap_is_active_iterate(void *opaque)
1236 {
1237     return dirty_bitmap_is_active(opaque) && !runstate_is_running();
1238 }
1239 
1240 static bool dirty_bitmap_has_postcopy(void *opaque)
1241 {
1242     return true;
1243 }
1244 
1245 static SaveVMHandlers savevm_dirty_bitmap_handlers = {
1246     .save_setup = dirty_bitmap_save_setup,
1247     .save_live_complete_postcopy = dirty_bitmap_save_complete,
1248     .save_live_complete_precopy = dirty_bitmap_save_complete,
1249     .has_postcopy = dirty_bitmap_has_postcopy,
1250     .state_pending_exact = dirty_bitmap_state_pending,
1251     .state_pending_estimate = dirty_bitmap_state_pending,
1252     .save_live_iterate = dirty_bitmap_save_iterate,
1253     .is_active_iterate = dirty_bitmap_is_active_iterate,
1254     .load_state = dirty_bitmap_load,
1255     .save_cleanup = dirty_bitmap_save_cleanup,
1256     .is_active = dirty_bitmap_is_active,
1257 };
1258 
1259 void dirty_bitmap_mig_init(void)
1260 {
1261     QSIMPLEQ_INIT(&dbm_state.save.dbms_list);
1262     qemu_mutex_init(&dbm_state.load.lock);
1263 
1264     register_savevm_live("dirty-bitmap", 0, 1,
1265                          &savevm_dirty_bitmap_handlers,
1266                          &dbm_state);
1267 }
1268