]>
Commit | Line | Data |
---|---|---|
c163b5ca | 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 | */ | |
13 | ||
14 | #include "qemu-common.h" | |
15 | #include "block_int.h" | |
16 | #include "hw/hw.h" | |
5e5328be | 17 | #include "qemu-queue.h" |
889ae39c | 18 | #include "qemu-timer.h" |
7184049e | 19 | #include "monitor.h" |
c163b5ca | 20 | #include "block-migration.h" |
889ae39c | 21 | #include "migration.h" |
f48905d4 | 22 | #include "blockdev.h" |
c163b5ca | 23 | #include <assert.h> |
c163b5ca | 24 | |
6ea44308 | 25 | #define BLOCK_SIZE (BDRV_SECTORS_PER_DIRTY_CHUNK << BDRV_SECTOR_BITS) |
c163b5ca | 26 | |
27 | #define BLK_MIG_FLAG_DEVICE_BLOCK 0x01 | |
28 | #define BLK_MIG_FLAG_EOS 0x02 | |
01e61e2d | 29 | #define BLK_MIG_FLAG_PROGRESS 0x04 |
c163b5ca | 30 | |
31 | #define MAX_IS_ALLOCATED_SEARCH 65536 | |
c163b5ca | 32 | |
33 | //#define DEBUG_BLK_MIGRATION | |
34 | ||
35 | #ifdef DEBUG_BLK_MIGRATION | |
d0f2c4c6 | 36 | #define DPRINTF(fmt, ...) \ |
c163b5ca | 37 | do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0) |
38 | #else | |
d0f2c4c6 | 39 | #define DPRINTF(fmt, ...) \ |
c163b5ca | 40 | do { } while (0) |
41 | #endif | |
42 | ||
a55eb92c JK |
43 | typedef struct BlkMigDevState { |
44 | BlockDriverState *bs; | |
45 | int bulk_completed; | |
46 | int shared_base; | |
a55eb92c | 47 | int64_t cur_sector; |
d76cac7d | 48 | int64_t cur_dirty; |
82801d8f | 49 | int64_t completed_sectors; |
a55eb92c JK |
50 | int64_t total_sectors; |
51 | int64_t dirty; | |
5e5328be | 52 | QSIMPLEQ_ENTRY(BlkMigDevState) entry; |
33656af7 | 53 | unsigned long *aio_bitmap; |
a55eb92c JK |
54 | } BlkMigDevState; |
55 | ||
c163b5ca | 56 | typedef struct BlkMigBlock { |
57 | uint8_t *buf; | |
58 | BlkMigDevState *bmds; | |
59 | int64_t sector; | |
33656af7 | 60 | int nr_sectors; |
c163b5ca | 61 | struct iovec iov; |
62 | QEMUIOVector qiov; | |
63 | BlockDriverAIOCB *aiocb; | |
64 | int ret; | |
889ae39c | 65 | int64_t time; |
5e5328be | 66 | QSIMPLEQ_ENTRY(BlkMigBlock) entry; |
c163b5ca | 67 | } BlkMigBlock; |
68 | ||
69 | typedef struct BlkMigState { | |
c163b5ca | 70 | int blk_enable; |
71 | int shared_base; | |
5e5328be JK |
72 | QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list; |
73 | QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list; | |
c163b5ca | 74 | int submitted; |
75 | int read_done; | |
76 | int transferred; | |
82801d8f | 77 | int64_t total_sector_sum; |
01e61e2d | 78 | int prev_progress; |
e970ec0b | 79 | int bulk_completed; |
889ae39c LS |
80 | long double total_time; |
81 | int reads; | |
c163b5ca | 82 | } BlkMigState; |
83 | ||
d11ecd3d | 84 | static BlkMigState block_mig_state; |
c163b5ca | 85 | |
13f0b67f JK |
86 | static void blk_send(QEMUFile *f, BlkMigBlock * blk) |
87 | { | |
88 | int len; | |
89 | ||
90 | /* sector number and flags */ | |
91 | qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS) | |
92 | | BLK_MIG_FLAG_DEVICE_BLOCK); | |
93 | ||
94 | /* device name */ | |
95 | len = strlen(blk->bmds->bs->device_name); | |
96 | qemu_put_byte(f, len); | |
97 | qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len); | |
98 | ||
99 | qemu_put_buffer(f, blk->buf, BLOCK_SIZE); | |
100 | } | |
101 | ||
25f23643 JK |
102 | int blk_mig_active(void) |
103 | { | |
104 | return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list); | |
105 | } | |
106 | ||
107 | uint64_t blk_mig_bytes_transferred(void) | |
108 | { | |
109 | BlkMigDevState *bmds; | |
110 | uint64_t sum = 0; | |
111 | ||
112 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { | |
113 | sum += bmds->completed_sectors; | |
114 | } | |
115 | return sum << BDRV_SECTOR_BITS; | |
116 | } | |
117 | ||
118 | uint64_t blk_mig_bytes_remaining(void) | |
119 | { | |
120 | return blk_mig_bytes_total() - blk_mig_bytes_transferred(); | |
121 | } | |
122 | ||
123 | uint64_t blk_mig_bytes_total(void) | |
124 | { | |
125 | BlkMigDevState *bmds; | |
126 | uint64_t sum = 0; | |
127 | ||
128 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { | |
129 | sum += bmds->total_sectors; | |
130 | } | |
131 | return sum << BDRV_SECTOR_BITS; | |
132 | } | |
133 | ||
889ae39c LS |
134 | static inline void add_avg_read_time(int64_t time) |
135 | { | |
136 | block_mig_state.reads++; | |
137 | block_mig_state.total_time += time; | |
138 | } | |
139 | ||
140 | static inline long double compute_read_bwidth(void) | |
141 | { | |
142 | assert(block_mig_state.total_time != 0); | |
143 | return (block_mig_state.reads * BLOCK_SIZE)/ block_mig_state.total_time; | |
144 | } | |
145 | ||
33656af7 MT |
146 | static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector) |
147 | { | |
148 | int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK; | |
149 | ||
62155e2b | 150 | if ((sector << BDRV_SECTOR_BITS) < bdrv_getlength(bmds->bs)) { |
33656af7 MT |
151 | return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] & |
152 | (1UL << (chunk % (sizeof(unsigned long) * 8)))); | |
153 | } else { | |
154 | return 0; | |
155 | } | |
156 | } | |
157 | ||
158 | static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num, | |
159 | int nb_sectors, int set) | |
160 | { | |
161 | int64_t start, end; | |
162 | unsigned long val, idx, bit; | |
163 | ||
164 | start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK; | |
165 | end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK; | |
166 | ||
167 | for (; start <= end; start++) { | |
168 | idx = start / (sizeof(unsigned long) * 8); | |
169 | bit = start % (sizeof(unsigned long) * 8); | |
170 | val = bmds->aio_bitmap[idx]; | |
171 | if (set) { | |
62155e2b | 172 | val |= 1UL << bit; |
33656af7 | 173 | } else { |
62155e2b | 174 | val &= ~(1UL << bit); |
33656af7 MT |
175 | } |
176 | bmds->aio_bitmap[idx] = val; | |
177 | } | |
178 | } | |
179 | ||
180 | static void alloc_aio_bitmap(BlkMigDevState *bmds) | |
181 | { | |
182 | BlockDriverState *bs = bmds->bs; | |
183 | int64_t bitmap_size; | |
184 | ||
185 | bitmap_size = (bdrv_getlength(bs) >> BDRV_SECTOR_BITS) + | |
186 | BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1; | |
187 | bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8; | |
188 | ||
189 | bmds->aio_bitmap = qemu_mallocz(bitmap_size); | |
190 | } | |
191 | ||
c163b5ca | 192 | static void blk_mig_read_cb(void *opaque, int ret) |
193 | { | |
194 | BlkMigBlock *blk = opaque; | |
a55eb92c | 195 | |
c163b5ca | 196 | blk->ret = ret; |
a55eb92c | 197 | |
889ae39c LS |
198 | blk->time = qemu_get_clock_ns(rt_clock) - blk->time; |
199 | ||
200 | add_avg_read_time(blk->time); | |
201 | ||
5e5328be | 202 | QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry); |
33656af7 | 203 | bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0); |
a55eb92c | 204 | |
d11ecd3d JK |
205 | block_mig_state.submitted--; |
206 | block_mig_state.read_done++; | |
207 | assert(block_mig_state.submitted >= 0); | |
c163b5ca | 208 | } |
209 | ||
7184049e | 210 | static int mig_save_device_bulk(Monitor *mon, QEMUFile *f, |
e970ec0b | 211 | BlkMigDevState *bmds) |
a55eb92c | 212 | { |
57cce12d JK |
213 | int64_t total_sectors = bmds->total_sectors; |
214 | int64_t cur_sector = bmds->cur_sector; | |
215 | BlockDriverState *bs = bmds->bs; | |
c163b5ca | 216 | BlkMigBlock *blk; |
13f0b67f | 217 | int nr_sectors; |
a55eb92c | 218 | |
57cce12d | 219 | if (bmds->shared_base) { |
b1d10856 | 220 | while (cur_sector < total_sectors && |
57cce12d JK |
221 | !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH, |
222 | &nr_sectors)) { | |
c163b5ca | 223 | cur_sector += nr_sectors; |
224 | } | |
225 | } | |
a55eb92c JK |
226 | |
227 | if (cur_sector >= total_sectors) { | |
82801d8f | 228 | bmds->cur_sector = bmds->completed_sectors = total_sectors; |
c163b5ca | 229 | return 1; |
230 | } | |
a55eb92c | 231 | |
82801d8f | 232 | bmds->completed_sectors = cur_sector; |
a55eb92c | 233 | |
57cce12d JK |
234 | cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1); |
235 | ||
6ea44308 JK |
236 | /* we are going to transfer a full block even if it is not allocated */ |
237 | nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK; | |
c163b5ca | 238 | |
6ea44308 | 239 | if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) { |
57cce12d | 240 | nr_sectors = total_sectors - cur_sector; |
c163b5ca | 241 | } |
a55eb92c | 242 | |
13f0b67f JK |
243 | blk = qemu_malloc(sizeof(BlkMigBlock)); |
244 | blk->buf = qemu_malloc(BLOCK_SIZE); | |
245 | blk->bmds = bmds; | |
246 | blk->sector = cur_sector; | |
33656af7 | 247 | blk->nr_sectors = nr_sectors; |
a55eb92c | 248 | |
e970ec0b LS |
249 | blk->iov.iov_base = blk->buf; |
250 | blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE; | |
251 | qemu_iovec_init_external(&blk->qiov, &blk->iov, 1); | |
a55eb92c | 252 | |
889ae39c LS |
253 | blk->time = qemu_get_clock_ns(rt_clock); |
254 | ||
e970ec0b LS |
255 | blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov, |
256 | nr_sectors, blk_mig_read_cb, blk); | |
257 | if (!blk->aiocb) { | |
258 | goto error; | |
c163b5ca | 259 | } |
e970ec0b | 260 | block_mig_state.submitted++; |
d76cac7d | 261 | |
13f0b67f JK |
262 | bdrv_reset_dirty(bs, cur_sector, nr_sectors); |
263 | bmds->cur_sector = cur_sector + nr_sectors; | |
a55eb92c | 264 | |
13f0b67f | 265 | return (bmds->cur_sector >= total_sectors); |
4b640365 JK |
266 | |
267 | error: | |
7184049e | 268 | monitor_printf(mon, "Error reading sector %" PRId64 "\n", cur_sector); |
4b640365 JK |
269 | qemu_file_set_error(f); |
270 | qemu_free(blk->buf); | |
271 | qemu_free(blk); | |
272 | return 0; | |
c163b5ca | 273 | } |
274 | ||
c163b5ca | 275 | static void set_dirty_tracking(int enable) |
276 | { | |
277 | BlkMigDevState *bmds; | |
5e5328be JK |
278 | |
279 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { | |
a55eb92c | 280 | bdrv_set_dirty_tracking(bmds->bs, enable); |
c163b5ca | 281 | } |
c163b5ca | 282 | } |
283 | ||
b66460e4 | 284 | static void init_blk_migration_it(void *opaque, BlockDriverState *bs) |
c163b5ca | 285 | { |
b66460e4 | 286 | Monitor *mon = opaque; |
5e5328be | 287 | BlkMigDevState *bmds; |
792773b2 | 288 | int64_t sectors; |
a55eb92c | 289 | |
d246673d | 290 | if (!bdrv_is_read_only(bs)) { |
b66460e4 | 291 | sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS; |
31f54f24 | 292 | if (sectors <= 0) { |
b66460e4 SH |
293 | return; |
294 | } | |
295 | ||
296 | bmds = qemu_mallocz(sizeof(BlkMigDevState)); | |
297 | bmds->bs = bs; | |
298 | bmds->bulk_completed = 0; | |
299 | bmds->total_sectors = sectors; | |
300 | bmds->completed_sectors = 0; | |
301 | bmds->shared_base = block_mig_state.shared_base; | |
33656af7 | 302 | alloc_aio_bitmap(bmds); |
f48905d4 | 303 | drive_get_ref(drive_get_by_blockdev(bs)); |
b66460e4 SH |
304 | |
305 | block_mig_state.total_sector_sum += sectors; | |
306 | ||
307 | if (bmds->shared_base) { | |
308 | monitor_printf(mon, "Start migration for %s with shared base " | |
309 | "image\n", | |
310 | bs->device_name); | |
311 | } else { | |
312 | monitor_printf(mon, "Start full migration for %s\n", | |
313 | bs->device_name); | |
314 | } | |
315 | ||
316 | QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry); | |
317 | } | |
318 | } | |
319 | ||
320 | static void init_blk_migration(Monitor *mon, QEMUFile *f) | |
321 | { | |
69d63a97 JK |
322 | block_mig_state.submitted = 0; |
323 | block_mig_state.read_done = 0; | |
324 | block_mig_state.transferred = 0; | |
82801d8f | 325 | block_mig_state.total_sector_sum = 0; |
01e61e2d | 326 | block_mig_state.prev_progress = -1; |
e970ec0b | 327 | block_mig_state.bulk_completed = 0; |
889ae39c LS |
328 | block_mig_state.total_time = 0; |
329 | block_mig_state.reads = 0; | |
69d63a97 | 330 | |
b66460e4 | 331 | bdrv_iterate(init_blk_migration_it, mon); |
c163b5ca | 332 | } |
333 | ||
e970ec0b | 334 | static int blk_mig_save_bulked_block(Monitor *mon, QEMUFile *f) |
c163b5ca | 335 | { |
82801d8f | 336 | int64_t completed_sector_sum = 0; |
c163b5ca | 337 | BlkMigDevState *bmds; |
01e61e2d | 338 | int progress; |
82801d8f | 339 | int ret = 0; |
c163b5ca | 340 | |
5e5328be | 341 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { |
a55eb92c | 342 | if (bmds->bulk_completed == 0) { |
e970ec0b | 343 | if (mig_save_device_bulk(mon, f, bmds) == 1) { |
57cce12d JK |
344 | /* completed bulk section for this device */ |
345 | bmds->bulk_completed = 1; | |
c163b5ca | 346 | } |
82801d8f JK |
347 | completed_sector_sum += bmds->completed_sectors; |
348 | ret = 1; | |
349 | break; | |
350 | } else { | |
351 | completed_sector_sum += bmds->completed_sectors; | |
c163b5ca | 352 | } |
353 | } | |
a55eb92c | 354 | |
8b6b2afc PR |
355 | if (block_mig_state.total_sector_sum != 0) { |
356 | progress = completed_sector_sum * 100 / | |
357 | block_mig_state.total_sector_sum; | |
358 | } else { | |
359 | progress = 100; | |
360 | } | |
01e61e2d JK |
361 | if (progress != block_mig_state.prev_progress) { |
362 | block_mig_state.prev_progress = progress; | |
363 | qemu_put_be64(f, (progress << BDRV_SECTOR_BITS) | |
364 | | BLK_MIG_FLAG_PROGRESS); | |
365 | monitor_printf(mon, "Completed %d %%\r", progress); | |
7184049e | 366 | monitor_flush(mon); |
82801d8f JK |
367 | } |
368 | ||
369 | return ret; | |
c163b5ca | 370 | } |
371 | ||
d76cac7d | 372 | static void blk_mig_reset_dirty_cursor(void) |
c163b5ca | 373 | { |
374 | BlkMigDevState *bmds; | |
d76cac7d LS |
375 | |
376 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { | |
377 | bmds->cur_dirty = 0; | |
378 | } | |
379 | } | |
380 | ||
381 | static int mig_save_device_dirty(Monitor *mon, QEMUFile *f, | |
382 | BlkMigDevState *bmds, int is_async) | |
383 | { | |
384 | BlkMigBlock *blk; | |
385 | int64_t total_sectors = bmds->total_sectors; | |
c163b5ca | 386 | int64_t sector; |
d76cac7d | 387 | int nr_sectors; |
a55eb92c | 388 | |
d76cac7d | 389 | for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) { |
62155e2b | 390 | if (bmds_aio_inflight(bmds, sector)) { |
33656af7 | 391 | qemu_aio_flush(); |
62155e2b | 392 | } |
d76cac7d | 393 | if (bdrv_get_dirty(bmds->bs, sector)) { |
575a58d7 | 394 | |
d76cac7d LS |
395 | if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) { |
396 | nr_sectors = total_sectors - sector; | |
397 | } else { | |
398 | nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK; | |
399 | } | |
400 | blk = qemu_malloc(sizeof(BlkMigBlock)); | |
401 | blk->buf = qemu_malloc(BLOCK_SIZE); | |
402 | blk->bmds = bmds; | |
403 | blk->sector = sector; | |
33656af7 | 404 | blk->nr_sectors = nr_sectors; |
d76cac7d | 405 | |
889ae39c | 406 | if (is_async) { |
d76cac7d LS |
407 | blk->iov.iov_base = blk->buf; |
408 | blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE; | |
409 | qemu_iovec_init_external(&blk->qiov, &blk->iov, 1); | |
410 | ||
bd0858bb | 411 | blk->time = qemu_get_clock_ns(rt_clock); |
889ae39c | 412 | |
d76cac7d LS |
413 | blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov, |
414 | nr_sectors, blk_mig_read_cb, blk); | |
415 | if (!blk->aiocb) { | |
416 | goto error; | |
417 | } | |
418 | block_mig_state.submitted++; | |
33656af7 | 419 | bmds_set_aio_inflight(bmds, sector, nr_sectors, 1); |
d76cac7d LS |
420 | } else { |
421 | if (bdrv_read(bmds->bs, sector, blk->buf, | |
422 | nr_sectors) < 0) { | |
423 | goto error; | |
c163b5ca | 424 | } |
d76cac7d | 425 | blk_send(f, blk); |
a55eb92c | 426 | |
d76cac7d LS |
427 | qemu_free(blk->buf); |
428 | qemu_free(blk); | |
a55eb92c | 429 | } |
d76cac7d LS |
430 | |
431 | bdrv_reset_dirty(bmds->bs, sector, nr_sectors); | |
432 | break; | |
c163b5ca | 433 | } |
d76cac7d LS |
434 | sector += BDRV_SECTORS_PER_DIRTY_CHUNK; |
435 | bmds->cur_dirty = sector; | |
c163b5ca | 436 | } |
575a58d7 | 437 | |
d76cac7d LS |
438 | return (bmds->cur_dirty >= bmds->total_sectors); |
439 | ||
889ae39c | 440 | error: |
d76cac7d LS |
441 | monitor_printf(mon, "Error reading sector %" PRId64 "\n", sector); |
442 | qemu_file_set_error(f); | |
443 | qemu_free(blk->buf); | |
444 | qemu_free(blk); | |
445 | return 0; | |
446 | } | |
447 | ||
448 | static int blk_mig_save_dirty_block(Monitor *mon, QEMUFile *f, int is_async) | |
449 | { | |
450 | BlkMigDevState *bmds; | |
451 | int ret = 0; | |
452 | ||
453 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { | |
889ae39c | 454 | if (mig_save_device_dirty(mon, f, bmds, is_async) == 0) { |
d76cac7d LS |
455 | ret = 1; |
456 | break; | |
457 | } | |
458 | } | |
459 | ||
460 | return ret; | |
c163b5ca | 461 | } |
462 | ||
463 | static void flush_blks(QEMUFile* f) | |
464 | { | |
5e5328be | 465 | BlkMigBlock *blk; |
a55eb92c | 466 | |
d0f2c4c6 | 467 | DPRINTF("%s Enter submitted %d read_done %d transferred %d\n", |
d11ecd3d JK |
468 | __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done, |
469 | block_mig_state.transferred); | |
a55eb92c | 470 | |
5e5328be JK |
471 | while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) { |
472 | if (qemu_file_rate_limit(f)) { | |
473 | break; | |
474 | } | |
4b640365 JK |
475 | if (blk->ret < 0) { |
476 | qemu_file_set_error(f); | |
477 | break; | |
478 | } | |
13f0b67f | 479 | blk_send(f, blk); |
a55eb92c | 480 | |
5e5328be | 481 | QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry); |
c163b5ca | 482 | qemu_free(blk->buf); |
483 | qemu_free(blk); | |
a55eb92c | 484 | |
d11ecd3d JK |
485 | block_mig_state.read_done--; |
486 | block_mig_state.transferred++; | |
487 | assert(block_mig_state.read_done >= 0); | |
c163b5ca | 488 | } |
c163b5ca | 489 | |
d0f2c4c6 | 490 | DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__, |
d11ecd3d JK |
491 | block_mig_state.submitted, block_mig_state.read_done, |
492 | block_mig_state.transferred); | |
c163b5ca | 493 | } |
494 | ||
889ae39c LS |
495 | static int64_t get_remaining_dirty(void) |
496 | { | |
497 | BlkMigDevState *bmds; | |
498 | int64_t dirty = 0; | |
499 | ||
500 | QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) { | |
501 | dirty += bdrv_get_dirty_count(bmds->bs); | |
502 | } | |
503 | ||
504 | return dirty * BLOCK_SIZE; | |
505 | } | |
506 | ||
c163b5ca | 507 | static int is_stage2_completed(void) |
508 | { | |
889ae39c LS |
509 | int64_t remaining_dirty; |
510 | long double bwidth; | |
511 | ||
512 | if (block_mig_state.bulk_completed == 1) { | |
513 | ||
514 | remaining_dirty = get_remaining_dirty(); | |
bd0858bb YT |
515 | if (remaining_dirty == 0) { |
516 | return 1; | |
517 | } | |
889ae39c | 518 | |
bd0858bb | 519 | bwidth = compute_read_bwidth(); |
889ae39c | 520 | |
bd0858bb | 521 | if ((remaining_dirty / bwidth) <= |
889ae39c LS |
522 | migrate_max_downtime()) { |
523 | /* finish stage2 because we think that we can finish remaing work | |
524 | below max_downtime */ | |
525 | ||
526 | return 1; | |
527 | } | |
528 | } | |
529 | ||
530 | return 0; | |
c163b5ca | 531 | } |
532 | ||
7184049e | 533 | static void blk_mig_cleanup(Monitor *mon) |
4ec7fcc7 | 534 | { |
82801d8f JK |
535 | BlkMigDevState *bmds; |
536 | BlkMigBlock *blk; | |
4ec7fcc7 | 537 | |
8f794c55 MT |
538 | set_dirty_tracking(0); |
539 | ||
82801d8f JK |
540 | while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) { |
541 | QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry); | |
f48905d4 | 542 | drive_put_ref(drive_get_by_blockdev(bmds->bs)); |
33656af7 | 543 | qemu_free(bmds->aio_bitmap); |
4ec7fcc7 JK |
544 | qemu_free(bmds); |
545 | } | |
546 | ||
82801d8f JK |
547 | while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) { |
548 | QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry); | |
4ec7fcc7 JK |
549 | qemu_free(blk->buf); |
550 | qemu_free(blk); | |
551 | } | |
552 | ||
7184049e | 553 | monitor_printf(mon, "\n"); |
4ec7fcc7 JK |
554 | } |
555 | ||
f327aa0c | 556 | static int block_save_live(Monitor *mon, QEMUFile *f, int stage, void *opaque) |
c163b5ca | 557 | { |
d0f2c4c6 | 558 | DPRINTF("Enter save live stage %d submitted %d transferred %d\n", |
d11ecd3d | 559 | stage, block_mig_state.submitted, block_mig_state.transferred); |
a55eb92c | 560 | |
4ec7fcc7 | 561 | if (stage < 0) { |
7184049e | 562 | blk_mig_cleanup(mon); |
4ec7fcc7 JK |
563 | return 0; |
564 | } | |
565 | ||
d11ecd3d | 566 | if (block_mig_state.blk_enable != 1) { |
c163b5ca | 567 | /* no need to migrate storage */ |
a55eb92c | 568 | qemu_put_be64(f, BLK_MIG_FLAG_EOS); |
c163b5ca | 569 | return 1; |
570 | } | |
a55eb92c JK |
571 | |
572 | if (stage == 1) { | |
7184049e | 573 | init_blk_migration(mon, f); |
a55eb92c | 574 | |
c163b5ca | 575 | /* start track dirty blocks */ |
576 | set_dirty_tracking(1); | |
c163b5ca | 577 | } |
578 | ||
579 | flush_blks(f); | |
a55eb92c | 580 | |
4b640365 | 581 | if (qemu_file_has_error(f)) { |
7184049e | 582 | blk_mig_cleanup(mon); |
4b640365 JK |
583 | return 0; |
584 | } | |
585 | ||
d76cac7d LS |
586 | blk_mig_reset_dirty_cursor(); |
587 | ||
889ae39c | 588 | if (stage == 2) { |
d76cac7d LS |
589 | /* control the rate of transfer */ |
590 | while ((block_mig_state.submitted + | |
591 | block_mig_state.read_done) * BLOCK_SIZE < | |
592 | qemu_file_get_rate_limit(f)) { | |
593 | if (block_mig_state.bulk_completed == 0) { | |
594 | /* first finish the bulk phase */ | |
595 | if (blk_mig_save_bulked_block(mon, f) == 0) { | |
889ae39c | 596 | /* finished saving bulk on all devices */ |
d76cac7d LS |
597 | block_mig_state.bulk_completed = 1; |
598 | } | |
599 | } else { | |
600 | if (blk_mig_save_dirty_block(mon, f, 1) == 0) { | |
601 | /* no more dirty blocks */ | |
602 | break; | |
603 | } | |
604 | } | |
a55eb92c | 605 | } |
a55eb92c | 606 | |
d76cac7d | 607 | flush_blks(f); |
a55eb92c | 608 | |
d76cac7d LS |
609 | if (qemu_file_has_error(f)) { |
610 | blk_mig_cleanup(mon); | |
611 | return 0; | |
612 | } | |
4b640365 JK |
613 | } |
614 | ||
a55eb92c | 615 | if (stage == 3) { |
889ae39c LS |
616 | /* we know for sure that save bulk is completed and |
617 | all async read completed */ | |
618 | assert(block_mig_state.submitted == 0); | |
a55eb92c | 619 | |
889ae39c | 620 | while (blk_mig_save_dirty_block(mon, f, 0) != 0); |
7184049e | 621 | blk_mig_cleanup(mon); |
a55eb92c | 622 | |
01e61e2d JK |
623 | /* report completion */ |
624 | qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS); | |
625 | ||
4b640365 JK |
626 | if (qemu_file_has_error(f)) { |
627 | return 0; | |
628 | } | |
629 | ||
7184049e | 630 | monitor_printf(mon, "Block migration completed\n"); |
c163b5ca | 631 | } |
a55eb92c JK |
632 | |
633 | qemu_put_be64(f, BLK_MIG_FLAG_EOS); | |
634 | ||
c163b5ca | 635 | return ((stage == 2) && is_stage2_completed()); |
636 | } | |
637 | ||
638 | static int block_load(QEMUFile *f, void *opaque, int version_id) | |
639 | { | |
01e61e2d | 640 | static int banner_printed; |
c163b5ca | 641 | int len, flags; |
642 | char device_name[256]; | |
643 | int64_t addr; | |
77358b59 | 644 | BlockDriverState *bs, *bs_prev = NULL; |
c163b5ca | 645 | uint8_t *buf; |
77358b59 PR |
646 | int64_t total_sectors = 0; |
647 | int nr_sectors; | |
a55eb92c | 648 | |
c163b5ca | 649 | do { |
c163b5ca | 650 | addr = qemu_get_be64(f); |
a55eb92c | 651 | |
6ea44308 JK |
652 | flags = addr & ~BDRV_SECTOR_MASK; |
653 | addr >>= BDRV_SECTOR_BITS; | |
a55eb92c JK |
654 | |
655 | if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) { | |
b02bea3a | 656 | int ret; |
c163b5ca | 657 | /* get device name */ |
658 | len = qemu_get_byte(f); | |
c163b5ca | 659 | qemu_get_buffer(f, (uint8_t *)device_name, len); |
660 | device_name[len] = '\0'; | |
a55eb92c | 661 | |
c163b5ca | 662 | bs = bdrv_find(device_name); |
4b640365 JK |
663 | if (!bs) { |
664 | fprintf(stderr, "Error unknown block device %s\n", | |
665 | device_name); | |
666 | return -EINVAL; | |
667 | } | |
a55eb92c | 668 | |
77358b59 PR |
669 | if (bs != bs_prev) { |
670 | bs_prev = bs; | |
671 | total_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS; | |
672 | if (total_sectors <= 0) { | |
673 | error_report("Error getting length of block device %s\n", | |
674 | device_name); | |
675 | return -EINVAL; | |
676 | } | |
677 | } | |
678 | ||
679 | if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) { | |
680 | nr_sectors = total_sectors - addr; | |
681 | } else { | |
682 | nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK; | |
683 | } | |
684 | ||
575a58d7 JK |
685 | buf = qemu_malloc(BLOCK_SIZE); |
686 | ||
a55eb92c | 687 | qemu_get_buffer(f, buf, BLOCK_SIZE); |
77358b59 | 688 | ret = bdrv_write(bs, addr, buf, nr_sectors); |
575a58d7 JK |
689 | |
690 | qemu_free(buf); | |
b02bea3a YT |
691 | if (ret < 0) { |
692 | return ret; | |
693 | } | |
01e61e2d JK |
694 | } else if (flags & BLK_MIG_FLAG_PROGRESS) { |
695 | if (!banner_printed) { | |
696 | printf("Receiving block device images\n"); | |
697 | banner_printed = 1; | |
698 | } | |
699 | printf("Completed %d %%%c", (int)addr, | |
700 | (addr == 100) ? '\n' : '\r'); | |
701 | fflush(stdout); | |
a55eb92c | 702 | } else if (!(flags & BLK_MIG_FLAG_EOS)) { |
4b640365 JK |
703 | fprintf(stderr, "Unknown flags\n"); |
704 | return -EINVAL; | |
705 | } | |
706 | if (qemu_file_has_error(f)) { | |
707 | return -EIO; | |
c163b5ca | 708 | } |
a55eb92c JK |
709 | } while (!(flags & BLK_MIG_FLAG_EOS)); |
710 | ||
c163b5ca | 711 | return 0; |
712 | } | |
713 | ||
714 | static void block_set_params(int blk_enable, int shared_base, void *opaque) | |
715 | { | |
d11ecd3d JK |
716 | block_mig_state.blk_enable = blk_enable; |
717 | block_mig_state.shared_base = shared_base; | |
a55eb92c | 718 | |
c163b5ca | 719 | /* shared base means that blk_enable = 1 */ |
d11ecd3d | 720 | block_mig_state.blk_enable |= shared_base; |
c163b5ca | 721 | } |
722 | ||
c163b5ca | 723 | void blk_mig_init(void) |
a55eb92c | 724 | { |
5e5328be JK |
725 | QSIMPLEQ_INIT(&block_mig_state.bmds_list); |
726 | QSIMPLEQ_INIT(&block_mig_state.blk_list); | |
727 | ||
0be71e32 AW |
728 | register_savevm_live(NULL, "block", 0, 1, block_set_params, |
729 | block_save_live, NULL, block_load, &block_mig_state); | |
c163b5ca | 730 | } |