]>
Commit | Line | Data |
---|---|---|
56e93d26 JQ |
1 | /* |
2 | * QEMU System Emulator | |
3 | * | |
4 | * Copyright (c) 2003-2008 Fabrice Bellard | |
76cc7b58 JQ |
5 | * Copyright (c) 2011-2015 Red Hat Inc |
6 | * | |
7 | * Authors: | |
8 | * Juan Quintela <quintela@redhat.com> | |
56e93d26 JQ |
9 | * |
10 | * Permission is hereby granted, free of charge, to any person obtaining a copy | |
11 | * of this software and associated documentation files (the "Software"), to deal | |
12 | * in the Software without restriction, including without limitation the rights | |
13 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell | |
14 | * copies of the Software, and to permit persons to whom the Software is | |
15 | * furnished to do so, subject to the following conditions: | |
16 | * | |
17 | * The above copyright notice and this permission notice shall be included in | |
18 | * all copies or substantial portions of the Software. | |
19 | * | |
20 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
21 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
22 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
23 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | |
24 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | |
25 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | |
26 | * THE SOFTWARE. | |
27 | */ | |
1393a485 | 28 | #include "qemu/osdep.h" |
33c11879 PB |
29 | #include "qemu-common.h" |
30 | #include "cpu.h" | |
56e93d26 | 31 | #include <zlib.h> |
4addcd4f | 32 | #include "qapi-event.h" |
f348b6d1 | 33 | #include "qemu/cutils.h" |
56e93d26 JQ |
34 | #include "qemu/bitops.h" |
35 | #include "qemu/bitmap.h" | |
7205c9ec JQ |
36 | #include "qemu/timer.h" |
37 | #include "qemu/main-loop.h" | |
56e93d26 | 38 | #include "migration/migration.h" |
e0b266f0 | 39 | #include "migration/postcopy-ram.h" |
56e93d26 JQ |
40 | #include "exec/address-spaces.h" |
41 | #include "migration/page_cache.h" | |
56e93d26 | 42 | #include "qemu/error-report.h" |
56e93d26 | 43 | #include "trace.h" |
56e93d26 | 44 | #include "exec/ram_addr.h" |
56e93d26 JQ |
45 | #include "qemu/rcu_queue.h" |
46 | ||
47 | #ifdef DEBUG_MIGRATION_RAM | |
48 | #define DPRINTF(fmt, ...) \ | |
49 | do { fprintf(stdout, "migration_ram: " fmt, ## __VA_ARGS__); } while (0) | |
50 | #else | |
51 | #define DPRINTF(fmt, ...) \ | |
52 | do { } while (0) | |
53 | #endif | |
54 | ||
56e93d26 | 55 | static int dirty_rate_high_cnt; |
56e93d26 JQ |
56 | |
57 | static uint64_t bitmap_sync_count; | |
58 | ||
59 | /***********************************************************/ | |
60 | /* ram save/restore */ | |
61 | ||
62 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ | |
63 | #define RAM_SAVE_FLAG_COMPRESS 0x02 | |
64 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 | |
65 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
66 | #define RAM_SAVE_FLAG_EOS 0x10 | |
67 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
68 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
69 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
70 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
71 | ||
72 | static const uint8_t ZERO_TARGET_PAGE[TARGET_PAGE_SIZE]; | |
73 | ||
74 | static inline bool is_zero_range(uint8_t *p, uint64_t size) | |
75 | { | |
76 | return buffer_find_nonzero_offset(p, size) == size; | |
77 | } | |
78 | ||
79 | /* struct contains XBZRLE cache and a static page | |
80 | used by the compression */ | |
81 | static struct { | |
82 | /* buffer used for XBZRLE encoding */ | |
83 | uint8_t *encoded_buf; | |
84 | /* buffer for storing page content */ | |
85 | uint8_t *current_buf; | |
86 | /* Cache for XBZRLE, Protected by lock. */ | |
87 | PageCache *cache; | |
88 | QemuMutex lock; | |
89 | } XBZRLE; | |
90 | ||
91 | /* buffer used for XBZRLE decoding */ | |
92 | static uint8_t *xbzrle_decoded_buf; | |
93 | ||
94 | static void XBZRLE_cache_lock(void) | |
95 | { | |
96 | if (migrate_use_xbzrle()) | |
97 | qemu_mutex_lock(&XBZRLE.lock); | |
98 | } | |
99 | ||
100 | static void XBZRLE_cache_unlock(void) | |
101 | { | |
102 | if (migrate_use_xbzrle()) | |
103 | qemu_mutex_unlock(&XBZRLE.lock); | |
104 | } | |
105 | ||
106 | /* | |
107 | * called from qmp_migrate_set_cache_size in main thread, possibly while | |
108 | * a migration is in progress. | |
109 | * A running migration maybe using the cache and might finish during this | |
110 | * call, hence changes to the cache are protected by XBZRLE.lock(). | |
111 | */ | |
112 | int64_t xbzrle_cache_resize(int64_t new_size) | |
113 | { | |
114 | PageCache *new_cache; | |
115 | int64_t ret; | |
116 | ||
117 | if (new_size < TARGET_PAGE_SIZE) { | |
118 | return -1; | |
119 | } | |
120 | ||
121 | XBZRLE_cache_lock(); | |
122 | ||
123 | if (XBZRLE.cache != NULL) { | |
124 | if (pow2floor(new_size) == migrate_xbzrle_cache_size()) { | |
125 | goto out_new_size; | |
126 | } | |
127 | new_cache = cache_init(new_size / TARGET_PAGE_SIZE, | |
128 | TARGET_PAGE_SIZE); | |
129 | if (!new_cache) { | |
130 | error_report("Error creating cache"); | |
131 | ret = -1; | |
132 | goto out; | |
133 | } | |
134 | ||
135 | cache_fini(XBZRLE.cache); | |
136 | XBZRLE.cache = new_cache; | |
137 | } | |
138 | ||
139 | out_new_size: | |
140 | ret = pow2floor(new_size); | |
141 | out: | |
142 | XBZRLE_cache_unlock(); | |
143 | return ret; | |
144 | } | |
145 | ||
146 | /* accounting for migration statistics */ | |
147 | typedef struct AccountingInfo { | |
148 | uint64_t dup_pages; | |
149 | uint64_t skipped_pages; | |
150 | uint64_t norm_pages; | |
151 | uint64_t iterations; | |
152 | uint64_t xbzrle_bytes; | |
153 | uint64_t xbzrle_pages; | |
154 | uint64_t xbzrle_cache_miss; | |
155 | double xbzrle_cache_miss_rate; | |
156 | uint64_t xbzrle_overflows; | |
157 | } AccountingInfo; | |
158 | ||
159 | static AccountingInfo acct_info; | |
160 | ||
161 | static void acct_clear(void) | |
162 | { | |
163 | memset(&acct_info, 0, sizeof(acct_info)); | |
164 | } | |
165 | ||
166 | uint64_t dup_mig_bytes_transferred(void) | |
167 | { | |
168 | return acct_info.dup_pages * TARGET_PAGE_SIZE; | |
169 | } | |
170 | ||
171 | uint64_t dup_mig_pages_transferred(void) | |
172 | { | |
173 | return acct_info.dup_pages; | |
174 | } | |
175 | ||
176 | uint64_t skipped_mig_bytes_transferred(void) | |
177 | { | |
178 | return acct_info.skipped_pages * TARGET_PAGE_SIZE; | |
179 | } | |
180 | ||
181 | uint64_t skipped_mig_pages_transferred(void) | |
182 | { | |
183 | return acct_info.skipped_pages; | |
184 | } | |
185 | ||
186 | uint64_t norm_mig_bytes_transferred(void) | |
187 | { | |
188 | return acct_info.norm_pages * TARGET_PAGE_SIZE; | |
189 | } | |
190 | ||
191 | uint64_t norm_mig_pages_transferred(void) | |
192 | { | |
193 | return acct_info.norm_pages; | |
194 | } | |
195 | ||
196 | uint64_t xbzrle_mig_bytes_transferred(void) | |
197 | { | |
198 | return acct_info.xbzrle_bytes; | |
199 | } | |
200 | ||
201 | uint64_t xbzrle_mig_pages_transferred(void) | |
202 | { | |
203 | return acct_info.xbzrle_pages; | |
204 | } | |
205 | ||
206 | uint64_t xbzrle_mig_pages_cache_miss(void) | |
207 | { | |
208 | return acct_info.xbzrle_cache_miss; | |
209 | } | |
210 | ||
211 | double xbzrle_mig_cache_miss_rate(void) | |
212 | { | |
213 | return acct_info.xbzrle_cache_miss_rate; | |
214 | } | |
215 | ||
216 | uint64_t xbzrle_mig_pages_overflow(void) | |
217 | { | |
218 | return acct_info.xbzrle_overflows; | |
219 | } | |
220 | ||
221 | /* This is the last block that we have visited serching for dirty pages | |
222 | */ | |
223 | static RAMBlock *last_seen_block; | |
224 | /* This is the last block from where we have sent data */ | |
225 | static RAMBlock *last_sent_block; | |
226 | static ram_addr_t last_offset; | |
dd631697 | 227 | static QemuMutex migration_bitmap_mutex; |
56e93d26 JQ |
228 | static uint64_t migration_dirty_pages; |
229 | static uint32_t last_version; | |
230 | static bool ram_bulk_stage; | |
231 | ||
b8fb8cb7 DDAG |
232 | /* used by the search for pages to send */ |
233 | struct PageSearchStatus { | |
234 | /* Current block being searched */ | |
235 | RAMBlock *block; | |
236 | /* Current offset to search from */ | |
237 | ram_addr_t offset; | |
238 | /* Set once we wrap around */ | |
239 | bool complete_round; | |
240 | }; | |
241 | typedef struct PageSearchStatus PageSearchStatus; | |
242 | ||
60be6340 DL |
243 | static struct BitmapRcu { |
244 | struct rcu_head rcu; | |
f3f491fc | 245 | /* Main migration bitmap */ |
60be6340 | 246 | unsigned long *bmap; |
f3f491fc DDAG |
247 | /* bitmap of pages that haven't been sent even once |
248 | * only maintained and used in postcopy at the moment | |
249 | * where it's used to send the dirtymap at the start | |
250 | * of the postcopy phase | |
251 | */ | |
252 | unsigned long *unsentmap; | |
60be6340 DL |
253 | } *migration_bitmap_rcu; |
254 | ||
56e93d26 JQ |
255 | struct CompressParam { |
256 | bool start; | |
257 | bool done; | |
258 | QEMUFile *file; | |
259 | QemuMutex mutex; | |
260 | QemuCond cond; | |
261 | RAMBlock *block; | |
262 | ram_addr_t offset; | |
263 | }; | |
264 | typedef struct CompressParam CompressParam; | |
265 | ||
266 | struct DecompressParam { | |
267 | bool start; | |
73a8912b | 268 | bool done; |
56e93d26 JQ |
269 | QemuMutex mutex; |
270 | QemuCond cond; | |
271 | void *des; | |
d341d9f3 | 272 | uint8_t *compbuf; |
56e93d26 JQ |
273 | int len; |
274 | }; | |
275 | typedef struct DecompressParam DecompressParam; | |
276 | ||
277 | static CompressParam *comp_param; | |
278 | static QemuThread *compress_threads; | |
279 | /* comp_done_cond is used to wake up the migration thread when | |
280 | * one of the compression threads has finished the compression. | |
281 | * comp_done_lock is used to co-work with comp_done_cond. | |
282 | */ | |
283 | static QemuMutex *comp_done_lock; | |
284 | static QemuCond *comp_done_cond; | |
285 | /* The empty QEMUFileOps will be used by file in CompressParam */ | |
286 | static const QEMUFileOps empty_ops = { }; | |
287 | ||
288 | static bool compression_switch; | |
289 | static bool quit_comp_thread; | |
290 | static bool quit_decomp_thread; | |
291 | static DecompressParam *decomp_param; | |
292 | static QemuThread *decompress_threads; | |
73a8912b LL |
293 | static QemuMutex decomp_done_lock; |
294 | static QemuCond decomp_done_cond; | |
56e93d26 JQ |
295 | |
296 | static int do_compress_ram_page(CompressParam *param); | |
297 | ||
298 | static void *do_data_compress(void *opaque) | |
299 | { | |
300 | CompressParam *param = opaque; | |
301 | ||
302 | while (!quit_comp_thread) { | |
303 | qemu_mutex_lock(¶m->mutex); | |
304 | /* Re-check the quit_comp_thread in case of | |
305 | * terminate_compression_threads is called just before | |
306 | * qemu_mutex_lock(¶m->mutex) and after | |
307 | * while(!quit_comp_thread), re-check it here can make | |
308 | * sure the compression thread terminate as expected. | |
309 | */ | |
310 | while (!param->start && !quit_comp_thread) { | |
311 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
312 | } | |
313 | if (!quit_comp_thread) { | |
314 | do_compress_ram_page(param); | |
315 | } | |
316 | param->start = false; | |
317 | qemu_mutex_unlock(¶m->mutex); | |
318 | ||
319 | qemu_mutex_lock(comp_done_lock); | |
320 | param->done = true; | |
321 | qemu_cond_signal(comp_done_cond); | |
322 | qemu_mutex_unlock(comp_done_lock); | |
323 | } | |
324 | ||
325 | return NULL; | |
326 | } | |
327 | ||
328 | static inline void terminate_compression_threads(void) | |
329 | { | |
330 | int idx, thread_count; | |
331 | ||
332 | thread_count = migrate_compress_threads(); | |
333 | quit_comp_thread = true; | |
334 | for (idx = 0; idx < thread_count; idx++) { | |
335 | qemu_mutex_lock(&comp_param[idx].mutex); | |
336 | qemu_cond_signal(&comp_param[idx].cond); | |
337 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
338 | } | |
339 | } | |
340 | ||
341 | void migrate_compress_threads_join(void) | |
342 | { | |
343 | int i, thread_count; | |
344 | ||
345 | if (!migrate_use_compression()) { | |
346 | return; | |
347 | } | |
348 | terminate_compression_threads(); | |
349 | thread_count = migrate_compress_threads(); | |
350 | for (i = 0; i < thread_count; i++) { | |
351 | qemu_thread_join(compress_threads + i); | |
352 | qemu_fclose(comp_param[i].file); | |
353 | qemu_mutex_destroy(&comp_param[i].mutex); | |
354 | qemu_cond_destroy(&comp_param[i].cond); | |
355 | } | |
356 | qemu_mutex_destroy(comp_done_lock); | |
357 | qemu_cond_destroy(comp_done_cond); | |
358 | g_free(compress_threads); | |
359 | g_free(comp_param); | |
360 | g_free(comp_done_cond); | |
361 | g_free(comp_done_lock); | |
362 | compress_threads = NULL; | |
363 | comp_param = NULL; | |
364 | comp_done_cond = NULL; | |
365 | comp_done_lock = NULL; | |
366 | } | |
367 | ||
368 | void migrate_compress_threads_create(void) | |
369 | { | |
370 | int i, thread_count; | |
371 | ||
372 | if (!migrate_use_compression()) { | |
373 | return; | |
374 | } | |
375 | quit_comp_thread = false; | |
376 | compression_switch = true; | |
377 | thread_count = migrate_compress_threads(); | |
378 | compress_threads = g_new0(QemuThread, thread_count); | |
379 | comp_param = g_new0(CompressParam, thread_count); | |
380 | comp_done_cond = g_new0(QemuCond, 1); | |
381 | comp_done_lock = g_new0(QemuMutex, 1); | |
382 | qemu_cond_init(comp_done_cond); | |
383 | qemu_mutex_init(comp_done_lock); | |
384 | for (i = 0; i < thread_count; i++) { | |
385 | /* com_param[i].file is just used as a dummy buffer to save data, set | |
386 | * it's ops to empty. | |
387 | */ | |
388 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
389 | comp_param[i].done = true; | |
390 | qemu_mutex_init(&comp_param[i].mutex); | |
391 | qemu_cond_init(&comp_param[i].cond); | |
392 | qemu_thread_create(compress_threads + i, "compress", | |
393 | do_data_compress, comp_param + i, | |
394 | QEMU_THREAD_JOINABLE); | |
395 | } | |
396 | } | |
397 | ||
398 | /** | |
399 | * save_page_header: Write page header to wire | |
400 | * | |
401 | * If this is the 1st block, it also writes the block identification | |
402 | * | |
403 | * Returns: Number of bytes written | |
404 | * | |
405 | * @f: QEMUFile where to send the data | |
406 | * @block: block that contains the page we want to send | |
407 | * @offset: offset inside the block for the page | |
408 | * in the lower bits, it contains flags | |
409 | */ | |
410 | static size_t save_page_header(QEMUFile *f, RAMBlock *block, ram_addr_t offset) | |
411 | { | |
9f5f380b | 412 | size_t size, len; |
56e93d26 JQ |
413 | |
414 | qemu_put_be64(f, offset); | |
415 | size = 8; | |
416 | ||
417 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b LL |
418 | len = strlen(block->idstr); |
419 | qemu_put_byte(f, len); | |
420 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
421 | size += 1 + len; | |
56e93d26 JQ |
422 | } |
423 | return size; | |
424 | } | |
425 | ||
070afca2 JH |
426 | /* Reduce amount of guest cpu execution to hopefully slow down memory writes. |
427 | * If guest dirty memory rate is reduced below the rate at which we can | |
428 | * transfer pages to the destination then we should be able to complete | |
429 | * migration. Some workloads dirty memory way too fast and will not effectively | |
430 | * converge, even with auto-converge. | |
431 | */ | |
432 | static void mig_throttle_guest_down(void) | |
433 | { | |
434 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
435 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
436 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
437 | |
438 | /* We have not started throttling yet. Let's start it. */ | |
439 | if (!cpu_throttle_active()) { | |
440 | cpu_throttle_set(pct_initial); | |
441 | } else { | |
442 | /* Throttling already on, just increase the rate */ | |
443 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
444 | } | |
445 | } | |
446 | ||
56e93d26 JQ |
447 | /* Update the xbzrle cache to reflect a page that's been sent as all 0. |
448 | * The important thing is that a stale (not-yet-0'd) page be replaced | |
449 | * by the new data. | |
450 | * As a bonus, if the page wasn't in the cache it gets added so that | |
451 | * when a small write is made into the 0'd page it gets XBZRLE sent | |
452 | */ | |
453 | static void xbzrle_cache_zero_page(ram_addr_t current_addr) | |
454 | { | |
455 | if (ram_bulk_stage || !migrate_use_xbzrle()) { | |
456 | return; | |
457 | } | |
458 | ||
459 | /* We don't care if this fails to allocate a new cache page | |
460 | * as long as it updated an old one */ | |
461 | cache_insert(XBZRLE.cache, current_addr, ZERO_TARGET_PAGE, | |
462 | bitmap_sync_count); | |
463 | } | |
464 | ||
465 | #define ENCODING_FLAG_XBZRLE 0x1 | |
466 | ||
467 | /** | |
468 | * save_xbzrle_page: compress and send current page | |
469 | * | |
470 | * Returns: 1 means that we wrote the page | |
471 | * 0 means that page is identical to the one already sent | |
472 | * -1 means that xbzrle would be longer than normal | |
473 | * | |
474 | * @f: QEMUFile where to send the data | |
475 | * @current_data: | |
476 | * @current_addr: | |
477 | * @block: block that contains the page we want to send | |
478 | * @offset: offset inside the block for the page | |
479 | * @last_stage: if we are at the completion stage | |
480 | * @bytes_transferred: increase it with the number of transferred bytes | |
481 | */ | |
482 | static int save_xbzrle_page(QEMUFile *f, uint8_t **current_data, | |
483 | ram_addr_t current_addr, RAMBlock *block, | |
484 | ram_addr_t offset, bool last_stage, | |
485 | uint64_t *bytes_transferred) | |
486 | { | |
487 | int encoded_len = 0, bytes_xbzrle; | |
488 | uint8_t *prev_cached_page; | |
489 | ||
490 | if (!cache_is_cached(XBZRLE.cache, current_addr, bitmap_sync_count)) { | |
491 | acct_info.xbzrle_cache_miss++; | |
492 | if (!last_stage) { | |
493 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
494 | bitmap_sync_count) == -1) { | |
495 | return -1; | |
496 | } else { | |
497 | /* update *current_data when the page has been | |
498 | inserted into cache */ | |
499 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
500 | } | |
501 | } | |
502 | return -1; | |
503 | } | |
504 | ||
505 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
506 | ||
507 | /* save current buffer into memory */ | |
508 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
509 | ||
510 | /* XBZRLE encoding (if there is no overflow) */ | |
511 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
512 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
513 | TARGET_PAGE_SIZE); | |
514 | if (encoded_len == 0) { | |
515 | DPRINTF("Skipping unmodified page\n"); | |
516 | return 0; | |
517 | } else if (encoded_len == -1) { | |
518 | DPRINTF("Overflow\n"); | |
519 | acct_info.xbzrle_overflows++; | |
520 | /* update data in the cache */ | |
521 | if (!last_stage) { | |
522 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
523 | *current_data = prev_cached_page; | |
524 | } | |
525 | return -1; | |
526 | } | |
527 | ||
528 | /* we need to update the data in the cache, in order to get the same data */ | |
529 | if (!last_stage) { | |
530 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
531 | } | |
532 | ||
533 | /* Send XBZRLE based compressed page */ | |
534 | bytes_xbzrle = save_page_header(f, block, offset | RAM_SAVE_FLAG_XBZRLE); | |
535 | qemu_put_byte(f, ENCODING_FLAG_XBZRLE); | |
536 | qemu_put_be16(f, encoded_len); | |
537 | qemu_put_buffer(f, XBZRLE.encoded_buf, encoded_len); | |
538 | bytes_xbzrle += encoded_len + 1 + 2; | |
539 | acct_info.xbzrle_pages++; | |
540 | acct_info.xbzrle_bytes += bytes_xbzrle; | |
541 | *bytes_transferred += bytes_xbzrle; | |
542 | ||
543 | return 1; | |
544 | } | |
545 | ||
f3f491fc DDAG |
546 | /* Called with rcu_read_lock() to protect migration_bitmap |
547 | * rb: The RAMBlock to search for dirty pages in | |
548 | * start: Start address (typically so we can continue from previous page) | |
549 | * ram_addr_abs: Pointer into which to store the address of the dirty page | |
550 | * within the global ram_addr space | |
551 | * | |
552 | * Returns: byte offset within memory region of the start of a dirty page | |
553 | */ | |
56e93d26 | 554 | static inline |
a82d593b DDAG |
555 | ram_addr_t migration_bitmap_find_dirty(RAMBlock *rb, |
556 | ram_addr_t start, | |
557 | ram_addr_t *ram_addr_abs) | |
56e93d26 | 558 | { |
2f68e399 | 559 | unsigned long base = rb->offset >> TARGET_PAGE_BITS; |
56e93d26 | 560 | unsigned long nr = base + (start >> TARGET_PAGE_BITS); |
2f68e399 DDAG |
561 | uint64_t rb_size = rb->used_length; |
562 | unsigned long size = base + (rb_size >> TARGET_PAGE_BITS); | |
2ff64038 | 563 | unsigned long *bitmap; |
56e93d26 JQ |
564 | |
565 | unsigned long next; | |
566 | ||
60be6340 | 567 | bitmap = atomic_rcu_read(&migration_bitmap_rcu)->bmap; |
56e93d26 JQ |
568 | if (ram_bulk_stage && nr > base) { |
569 | next = nr + 1; | |
570 | } else { | |
2ff64038 | 571 | next = find_next_bit(bitmap, size, nr); |
56e93d26 JQ |
572 | } |
573 | ||
f3f491fc | 574 | *ram_addr_abs = next << TARGET_PAGE_BITS; |
56e93d26 JQ |
575 | return (next - base) << TARGET_PAGE_BITS; |
576 | } | |
577 | ||
a82d593b DDAG |
578 | static inline bool migration_bitmap_clear_dirty(ram_addr_t addr) |
579 | { | |
580 | bool ret; | |
581 | int nr = addr >> TARGET_PAGE_BITS; | |
582 | unsigned long *bitmap = atomic_rcu_read(&migration_bitmap_rcu)->bmap; | |
583 | ||
584 | ret = test_and_clear_bit(nr, bitmap); | |
585 | ||
586 | if (ret) { | |
587 | migration_dirty_pages--; | |
588 | } | |
589 | return ret; | |
590 | } | |
591 | ||
56e93d26 JQ |
592 | static void migration_bitmap_sync_range(ram_addr_t start, ram_addr_t length) |
593 | { | |
2ff64038 | 594 | unsigned long *bitmap; |
60be6340 | 595 | bitmap = atomic_rcu_read(&migration_bitmap_rcu)->bmap; |
56e93d26 | 596 | migration_dirty_pages += |
2ff64038 | 597 | cpu_physical_memory_sync_dirty_bitmap(bitmap, start, length); |
56e93d26 JQ |
598 | } |
599 | ||
56e93d26 JQ |
600 | /* Fix me: there are too many global variables used in migration process. */ |
601 | static int64_t start_time; | |
602 | static int64_t bytes_xfer_prev; | |
603 | static int64_t num_dirty_pages_period; | |
604 | static uint64_t xbzrle_cache_miss_prev; | |
605 | static uint64_t iterations_prev; | |
606 | ||
607 | static void migration_bitmap_sync_init(void) | |
608 | { | |
609 | start_time = 0; | |
610 | bytes_xfer_prev = 0; | |
611 | num_dirty_pages_period = 0; | |
612 | xbzrle_cache_miss_prev = 0; | |
613 | iterations_prev = 0; | |
614 | } | |
615 | ||
56e93d26 JQ |
616 | static void migration_bitmap_sync(void) |
617 | { | |
618 | RAMBlock *block; | |
619 | uint64_t num_dirty_pages_init = migration_dirty_pages; | |
620 | MigrationState *s = migrate_get_current(); | |
621 | int64_t end_time; | |
622 | int64_t bytes_xfer_now; | |
623 | ||
624 | bitmap_sync_count++; | |
625 | ||
626 | if (!bytes_xfer_prev) { | |
627 | bytes_xfer_prev = ram_bytes_transferred(); | |
628 | } | |
629 | ||
630 | if (!start_time) { | |
631 | start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
632 | } | |
633 | ||
634 | trace_migration_bitmap_sync_start(); | |
635 | address_space_sync_dirty_bitmap(&address_space_memory); | |
636 | ||
dd631697 | 637 | qemu_mutex_lock(&migration_bitmap_mutex); |
56e93d26 JQ |
638 | rcu_read_lock(); |
639 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
2f68e399 | 640 | migration_bitmap_sync_range(block->offset, block->used_length); |
56e93d26 JQ |
641 | } |
642 | rcu_read_unlock(); | |
dd631697 | 643 | qemu_mutex_unlock(&migration_bitmap_mutex); |
56e93d26 JQ |
644 | |
645 | trace_migration_bitmap_sync_end(migration_dirty_pages | |
646 | - num_dirty_pages_init); | |
647 | num_dirty_pages_period += migration_dirty_pages - num_dirty_pages_init; | |
648 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
649 | ||
650 | /* more than 1 second = 1000 millisecons */ | |
651 | if (end_time > start_time + 1000) { | |
652 | if (migrate_auto_converge()) { | |
653 | /* The following detection logic can be refined later. For now: | |
654 | Check to see if the dirtied bytes is 50% more than the approx. | |
655 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
656 | were in this routine. If that happens twice, start or increase |
657 | throttling */ | |
56e93d26 | 658 | bytes_xfer_now = ram_bytes_transferred(); |
070afca2 | 659 | |
56e93d26 JQ |
660 | if (s->dirty_pages_rate && |
661 | (num_dirty_pages_period * TARGET_PAGE_SIZE > | |
662 | (bytes_xfer_now - bytes_xfer_prev)/2) && | |
070afca2 | 663 | (dirty_rate_high_cnt++ >= 2)) { |
56e93d26 | 664 | trace_migration_throttle(); |
56e93d26 | 665 | dirty_rate_high_cnt = 0; |
070afca2 | 666 | mig_throttle_guest_down(); |
56e93d26 JQ |
667 | } |
668 | bytes_xfer_prev = bytes_xfer_now; | |
56e93d26 | 669 | } |
070afca2 | 670 | |
56e93d26 JQ |
671 | if (migrate_use_xbzrle()) { |
672 | if (iterations_prev != acct_info.iterations) { | |
673 | acct_info.xbzrle_cache_miss_rate = | |
674 | (double)(acct_info.xbzrle_cache_miss - | |
675 | xbzrle_cache_miss_prev) / | |
676 | (acct_info.iterations - iterations_prev); | |
677 | } | |
678 | iterations_prev = acct_info.iterations; | |
679 | xbzrle_cache_miss_prev = acct_info.xbzrle_cache_miss; | |
680 | } | |
681 | s->dirty_pages_rate = num_dirty_pages_period * 1000 | |
682 | / (end_time - start_time); | |
683 | s->dirty_bytes_rate = s->dirty_pages_rate * TARGET_PAGE_SIZE; | |
684 | start_time = end_time; | |
685 | num_dirty_pages_period = 0; | |
686 | } | |
687 | s->dirty_sync_count = bitmap_sync_count; | |
4addcd4f DDAG |
688 | if (migrate_use_events()) { |
689 | qapi_event_send_migration_pass(bitmap_sync_count, NULL); | |
690 | } | |
56e93d26 JQ |
691 | } |
692 | ||
693 | /** | |
694 | * save_zero_page: Send the zero page to the stream | |
695 | * | |
696 | * Returns: Number of pages written. | |
697 | * | |
698 | * @f: QEMUFile where to send the data | |
699 | * @block: block that contains the page we want to send | |
700 | * @offset: offset inside the block for the page | |
701 | * @p: pointer to the page | |
702 | * @bytes_transferred: increase it with the number of transferred bytes | |
703 | */ | |
704 | static int save_zero_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset, | |
705 | uint8_t *p, uint64_t *bytes_transferred) | |
706 | { | |
707 | int pages = -1; | |
708 | ||
709 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
710 | acct_info.dup_pages++; | |
711 | *bytes_transferred += save_page_header(f, block, | |
712 | offset | RAM_SAVE_FLAG_COMPRESS); | |
713 | qemu_put_byte(f, 0); | |
714 | *bytes_transferred += 1; | |
715 | pages = 1; | |
716 | } | |
717 | ||
718 | return pages; | |
719 | } | |
720 | ||
721 | /** | |
722 | * ram_save_page: Send the given page to the stream | |
723 | * | |
724 | * Returns: Number of pages written. | |
3fd3c4b3 DDAG |
725 | * < 0 - error |
726 | * >=0 - Number of pages written - this might legally be 0 | |
727 | * if xbzrle noticed the page was the same. | |
56e93d26 JQ |
728 | * |
729 | * @f: QEMUFile where to send the data | |
730 | * @block: block that contains the page we want to send | |
731 | * @offset: offset inside the block for the page | |
732 | * @last_stage: if we are at the completion stage | |
733 | * @bytes_transferred: increase it with the number of transferred bytes | |
734 | */ | |
a08f6890 | 735 | static int ram_save_page(QEMUFile *f, PageSearchStatus *pss, |
56e93d26 JQ |
736 | bool last_stage, uint64_t *bytes_transferred) |
737 | { | |
738 | int pages = -1; | |
739 | uint64_t bytes_xmit; | |
740 | ram_addr_t current_addr; | |
56e93d26 JQ |
741 | uint8_t *p; |
742 | int ret; | |
743 | bool send_async = true; | |
a08f6890 HZ |
744 | RAMBlock *block = pss->block; |
745 | ram_addr_t offset = pss->offset; | |
56e93d26 | 746 | |
2f68e399 | 747 | p = block->host + offset; |
56e93d26 JQ |
748 | |
749 | /* In doubt sent page as normal */ | |
750 | bytes_xmit = 0; | |
751 | ret = ram_control_save_page(f, block->offset, | |
752 | offset, TARGET_PAGE_SIZE, &bytes_xmit); | |
753 | if (bytes_xmit) { | |
754 | *bytes_transferred += bytes_xmit; | |
755 | pages = 1; | |
756 | } | |
757 | ||
758 | XBZRLE_cache_lock(); | |
759 | ||
760 | current_addr = block->offset + offset; | |
761 | ||
762 | if (block == last_sent_block) { | |
763 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
764 | } | |
765 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { | |
766 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
767 | if (bytes_xmit > 0) { | |
768 | acct_info.norm_pages++; | |
769 | } else if (bytes_xmit == 0) { | |
770 | acct_info.dup_pages++; | |
771 | } | |
772 | } | |
773 | } else { | |
774 | pages = save_zero_page(f, block, offset, p, bytes_transferred); | |
775 | if (pages > 0) { | |
776 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
777 | * page would be stale | |
778 | */ | |
779 | xbzrle_cache_zero_page(current_addr); | |
780 | } else if (!ram_bulk_stage && migrate_use_xbzrle()) { | |
781 | pages = save_xbzrle_page(f, &p, current_addr, block, | |
782 | offset, last_stage, bytes_transferred); | |
783 | if (!last_stage) { | |
784 | /* Can't send this cached data async, since the cache page | |
785 | * might get updated before it gets to the wire | |
786 | */ | |
787 | send_async = false; | |
788 | } | |
789 | } | |
790 | } | |
791 | ||
792 | /* XBZRLE overflow or normal page */ | |
793 | if (pages == -1) { | |
794 | *bytes_transferred += save_page_header(f, block, | |
795 | offset | RAM_SAVE_FLAG_PAGE); | |
796 | if (send_async) { | |
797 | qemu_put_buffer_async(f, p, TARGET_PAGE_SIZE); | |
798 | } else { | |
799 | qemu_put_buffer(f, p, TARGET_PAGE_SIZE); | |
800 | } | |
801 | *bytes_transferred += TARGET_PAGE_SIZE; | |
802 | pages = 1; | |
803 | acct_info.norm_pages++; | |
804 | } | |
805 | ||
806 | XBZRLE_cache_unlock(); | |
807 | ||
808 | return pages; | |
809 | } | |
810 | ||
811 | static int do_compress_ram_page(CompressParam *param) | |
812 | { | |
813 | int bytes_sent, blen; | |
814 | uint8_t *p; | |
815 | RAMBlock *block = param->block; | |
816 | ram_addr_t offset = param->offset; | |
817 | ||
2f68e399 | 818 | p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 JQ |
819 | |
820 | bytes_sent = save_page_header(param->file, block, offset | | |
821 | RAM_SAVE_FLAG_COMPRESS_PAGE); | |
822 | blen = qemu_put_compression_data(param->file, p, TARGET_PAGE_SIZE, | |
823 | migrate_compress_level()); | |
824 | bytes_sent += blen; | |
825 | ||
826 | return bytes_sent; | |
827 | } | |
828 | ||
829 | static inline void start_compression(CompressParam *param) | |
830 | { | |
831 | param->done = false; | |
832 | qemu_mutex_lock(¶m->mutex); | |
833 | param->start = true; | |
834 | qemu_cond_signal(¶m->cond); | |
835 | qemu_mutex_unlock(¶m->mutex); | |
836 | } | |
837 | ||
838 | static inline void start_decompression(DecompressParam *param) | |
839 | { | |
73a8912b | 840 | param->done = false; |
56e93d26 JQ |
841 | qemu_mutex_lock(¶m->mutex); |
842 | param->start = true; | |
843 | qemu_cond_signal(¶m->cond); | |
844 | qemu_mutex_unlock(¶m->mutex); | |
845 | } | |
846 | ||
847 | static uint64_t bytes_transferred; | |
848 | ||
849 | static void flush_compressed_data(QEMUFile *f) | |
850 | { | |
851 | int idx, len, thread_count; | |
852 | ||
853 | if (!migrate_use_compression()) { | |
854 | return; | |
855 | } | |
856 | thread_count = migrate_compress_threads(); | |
857 | for (idx = 0; idx < thread_count; idx++) { | |
858 | if (!comp_param[idx].done) { | |
859 | qemu_mutex_lock(comp_done_lock); | |
860 | while (!comp_param[idx].done && !quit_comp_thread) { | |
861 | qemu_cond_wait(comp_done_cond, comp_done_lock); | |
862 | } | |
863 | qemu_mutex_unlock(comp_done_lock); | |
864 | } | |
865 | if (!quit_comp_thread) { | |
866 | len = qemu_put_qemu_file(f, comp_param[idx].file); | |
867 | bytes_transferred += len; | |
868 | } | |
869 | } | |
870 | } | |
871 | ||
872 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
873 | ram_addr_t offset) | |
874 | { | |
875 | param->block = block; | |
876 | param->offset = offset; | |
877 | } | |
878 | ||
879 | static int compress_page_with_multi_thread(QEMUFile *f, RAMBlock *block, | |
880 | ram_addr_t offset, | |
881 | uint64_t *bytes_transferred) | |
882 | { | |
883 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
884 | ||
885 | thread_count = migrate_compress_threads(); | |
886 | qemu_mutex_lock(comp_done_lock); | |
887 | while (true) { | |
888 | for (idx = 0; idx < thread_count; idx++) { | |
889 | if (comp_param[idx].done) { | |
890 | bytes_xmit = qemu_put_qemu_file(f, comp_param[idx].file); | |
891 | set_compress_params(&comp_param[idx], block, offset); | |
892 | start_compression(&comp_param[idx]); | |
893 | pages = 1; | |
894 | acct_info.norm_pages++; | |
895 | *bytes_transferred += bytes_xmit; | |
896 | break; | |
897 | } | |
898 | } | |
899 | if (pages > 0) { | |
900 | break; | |
901 | } else { | |
902 | qemu_cond_wait(comp_done_cond, comp_done_lock); | |
903 | } | |
904 | } | |
905 | qemu_mutex_unlock(comp_done_lock); | |
906 | ||
907 | return pages; | |
908 | } | |
909 | ||
910 | /** | |
911 | * ram_save_compressed_page: compress the given page and send it to the stream | |
912 | * | |
913 | * Returns: Number of pages written. | |
914 | * | |
915 | * @f: QEMUFile where to send the data | |
916 | * @block: block that contains the page we want to send | |
917 | * @offset: offset inside the block for the page | |
918 | * @last_stage: if we are at the completion stage | |
919 | * @bytes_transferred: increase it with the number of transferred bytes | |
920 | */ | |
a08f6890 HZ |
921 | static int ram_save_compressed_page(QEMUFile *f, PageSearchStatus *pss, |
922 | bool last_stage, | |
56e93d26 JQ |
923 | uint64_t *bytes_transferred) |
924 | { | |
925 | int pages = -1; | |
926 | uint64_t bytes_xmit; | |
56e93d26 JQ |
927 | uint8_t *p; |
928 | int ret; | |
a08f6890 HZ |
929 | RAMBlock *block = pss->block; |
930 | ram_addr_t offset = pss->offset; | |
56e93d26 | 931 | |
2f68e399 | 932 | p = block->host + offset; |
56e93d26 JQ |
933 | |
934 | bytes_xmit = 0; | |
935 | ret = ram_control_save_page(f, block->offset, | |
936 | offset, TARGET_PAGE_SIZE, &bytes_xmit); | |
937 | if (bytes_xmit) { | |
938 | *bytes_transferred += bytes_xmit; | |
939 | pages = 1; | |
940 | } | |
941 | if (block == last_sent_block) { | |
942 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
943 | } | |
944 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { | |
945 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
946 | if (bytes_xmit > 0) { | |
947 | acct_info.norm_pages++; | |
948 | } else if (bytes_xmit == 0) { | |
949 | acct_info.dup_pages++; | |
950 | } | |
951 | } | |
952 | } else { | |
953 | /* When starting the process of a new block, the first page of | |
954 | * the block should be sent out before other pages in the same | |
955 | * block, and all the pages in last block should have been sent | |
956 | * out, keeping this order is important, because the 'cont' flag | |
957 | * is used to avoid resending the block name. | |
958 | */ | |
959 | if (block != last_sent_block) { | |
960 | flush_compressed_data(f); | |
961 | pages = save_zero_page(f, block, offset, p, bytes_transferred); | |
962 | if (pages == -1) { | |
963 | set_compress_params(&comp_param[0], block, offset); | |
964 | /* Use the qemu thread to compress the data to make sure the | |
965 | * first page is sent out before other pages | |
966 | */ | |
967 | bytes_xmit = do_compress_ram_page(&comp_param[0]); | |
968 | acct_info.norm_pages++; | |
969 | qemu_put_qemu_file(f, comp_param[0].file); | |
970 | *bytes_transferred += bytes_xmit; | |
971 | pages = 1; | |
972 | } | |
973 | } else { | |
974 | pages = save_zero_page(f, block, offset, p, bytes_transferred); | |
975 | if (pages == -1) { | |
976 | pages = compress_page_with_multi_thread(f, block, offset, | |
977 | bytes_transferred); | |
978 | } | |
979 | } | |
980 | } | |
981 | ||
982 | return pages; | |
983 | } | |
984 | ||
b9e60928 DDAG |
985 | /* |
986 | * Find the next dirty page and update any state associated with | |
987 | * the search process. | |
988 | * | |
989 | * Returns: True if a page is found | |
990 | * | |
991 | * @f: Current migration stream. | |
992 | * @pss: Data about the state of the current dirty page scan. | |
993 | * @*again: Set to false if the search has scanned the whole of RAM | |
e0b266f0 DDAG |
994 | * *ram_addr_abs: Pointer into which to store the address of the dirty page |
995 | * within the global ram_addr space | |
b9e60928 DDAG |
996 | */ |
997 | static bool find_dirty_block(QEMUFile *f, PageSearchStatus *pss, | |
f3f491fc | 998 | bool *again, ram_addr_t *ram_addr_abs) |
b9e60928 | 999 | { |
a82d593b DDAG |
1000 | pss->offset = migration_bitmap_find_dirty(pss->block, pss->offset, |
1001 | ram_addr_abs); | |
b9e60928 DDAG |
1002 | if (pss->complete_round && pss->block == last_seen_block && |
1003 | pss->offset >= last_offset) { | |
1004 | /* | |
1005 | * We've been once around the RAM and haven't found anything. | |
1006 | * Give up. | |
1007 | */ | |
1008 | *again = false; | |
1009 | return false; | |
1010 | } | |
1011 | if (pss->offset >= pss->block->used_length) { | |
1012 | /* Didn't find anything in this RAM Block */ | |
1013 | pss->offset = 0; | |
1014 | pss->block = QLIST_NEXT_RCU(pss->block, next); | |
1015 | if (!pss->block) { | |
1016 | /* Hit the end of the list */ | |
1017 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1018 | /* Flag that we've looped */ | |
1019 | pss->complete_round = true; | |
1020 | ram_bulk_stage = false; | |
1021 | if (migrate_use_xbzrle()) { | |
1022 | /* If xbzrle is on, stop using the data compression at this | |
1023 | * point. In theory, xbzrle can do better than compression. | |
1024 | */ | |
1025 | flush_compressed_data(f); | |
1026 | compression_switch = false; | |
1027 | } | |
1028 | } | |
1029 | /* Didn't find anything this time, but try again on the new block */ | |
1030 | *again = true; | |
1031 | return false; | |
1032 | } else { | |
1033 | /* Can go around again, but... */ | |
1034 | *again = true; | |
1035 | /* We've found something so probably don't need to */ | |
1036 | return true; | |
1037 | } | |
1038 | } | |
1039 | ||
a82d593b DDAG |
1040 | /* |
1041 | * Helper for 'get_queued_page' - gets a page off the queue | |
1042 | * ms: MigrationState in | |
1043 | * *offset: Used to return the offset within the RAMBlock | |
1044 | * ram_addr_abs: global offset in the dirty/sent bitmaps | |
1045 | * | |
1046 | * Returns: block (or NULL if none available) | |
1047 | */ | |
1048 | static RAMBlock *unqueue_page(MigrationState *ms, ram_addr_t *offset, | |
1049 | ram_addr_t *ram_addr_abs) | |
1050 | { | |
1051 | RAMBlock *block = NULL; | |
1052 | ||
1053 | qemu_mutex_lock(&ms->src_page_req_mutex); | |
1054 | if (!QSIMPLEQ_EMPTY(&ms->src_page_requests)) { | |
1055 | struct MigrationSrcPageRequest *entry = | |
1056 | QSIMPLEQ_FIRST(&ms->src_page_requests); | |
1057 | block = entry->rb; | |
1058 | *offset = entry->offset; | |
1059 | *ram_addr_abs = (entry->offset + entry->rb->offset) & | |
1060 | TARGET_PAGE_MASK; | |
1061 | ||
1062 | if (entry->len > TARGET_PAGE_SIZE) { | |
1063 | entry->len -= TARGET_PAGE_SIZE; | |
1064 | entry->offset += TARGET_PAGE_SIZE; | |
1065 | } else { | |
1066 | memory_region_unref(block->mr); | |
1067 | QSIMPLEQ_REMOVE_HEAD(&ms->src_page_requests, next_req); | |
1068 | g_free(entry); | |
1069 | } | |
1070 | } | |
1071 | qemu_mutex_unlock(&ms->src_page_req_mutex); | |
1072 | ||
1073 | return block; | |
1074 | } | |
1075 | ||
1076 | /* | |
1077 | * Unqueue a page from the queue fed by postcopy page requests; skips pages | |
1078 | * that are already sent (!dirty) | |
1079 | * | |
1080 | * ms: MigrationState in | |
1081 | * pss: PageSearchStatus structure updated with found block/offset | |
1082 | * ram_addr_abs: global offset in the dirty/sent bitmaps | |
1083 | * | |
1084 | * Returns: true if a queued page is found | |
1085 | */ | |
1086 | static bool get_queued_page(MigrationState *ms, PageSearchStatus *pss, | |
1087 | ram_addr_t *ram_addr_abs) | |
1088 | { | |
1089 | RAMBlock *block; | |
1090 | ram_addr_t offset; | |
1091 | bool dirty; | |
1092 | ||
1093 | do { | |
1094 | block = unqueue_page(ms, &offset, ram_addr_abs); | |
1095 | /* | |
1096 | * We're sending this page, and since it's postcopy nothing else | |
1097 | * will dirty it, and we must make sure it doesn't get sent again | |
1098 | * even if this queue request was received after the background | |
1099 | * search already sent it. | |
1100 | */ | |
1101 | if (block) { | |
1102 | unsigned long *bitmap; | |
1103 | bitmap = atomic_rcu_read(&migration_bitmap_rcu)->bmap; | |
1104 | dirty = test_bit(*ram_addr_abs >> TARGET_PAGE_BITS, bitmap); | |
1105 | if (!dirty) { | |
1106 | trace_get_queued_page_not_dirty( | |
1107 | block->idstr, (uint64_t)offset, | |
1108 | (uint64_t)*ram_addr_abs, | |
1109 | test_bit(*ram_addr_abs >> TARGET_PAGE_BITS, | |
1110 | atomic_rcu_read(&migration_bitmap_rcu)->unsentmap)); | |
1111 | } else { | |
1112 | trace_get_queued_page(block->idstr, | |
1113 | (uint64_t)offset, | |
1114 | (uint64_t)*ram_addr_abs); | |
1115 | } | |
1116 | } | |
1117 | ||
1118 | } while (block && !dirty); | |
1119 | ||
1120 | if (block) { | |
1121 | /* | |
1122 | * As soon as we start servicing pages out of order, then we have | |
1123 | * to kill the bulk stage, since the bulk stage assumes | |
1124 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1125 | * dirty, that's no longer true. | |
1126 | */ | |
1127 | ram_bulk_stage = false; | |
1128 | ||
1129 | /* | |
1130 | * We want the background search to continue from the queued page | |
1131 | * since the guest is likely to want other pages near to the page | |
1132 | * it just requested. | |
1133 | */ | |
1134 | pss->block = block; | |
1135 | pss->offset = offset; | |
1136 | } | |
1137 | ||
1138 | return !!block; | |
1139 | } | |
1140 | ||
6c595cde DDAG |
1141 | /** |
1142 | * flush_page_queue: Flush any remaining pages in the ram request queue | |
1143 | * it should be empty at the end anyway, but in error cases there may be | |
1144 | * some left. | |
1145 | * | |
1146 | * ms: MigrationState | |
1147 | */ | |
1148 | void flush_page_queue(MigrationState *ms) | |
1149 | { | |
1150 | struct MigrationSrcPageRequest *mspr, *next_mspr; | |
1151 | /* This queue generally should be empty - but in the case of a failed | |
1152 | * migration might have some droppings in. | |
1153 | */ | |
1154 | rcu_read_lock(); | |
1155 | QSIMPLEQ_FOREACH_SAFE(mspr, &ms->src_page_requests, next_req, next_mspr) { | |
1156 | memory_region_unref(mspr->rb->mr); | |
1157 | QSIMPLEQ_REMOVE_HEAD(&ms->src_page_requests, next_req); | |
1158 | g_free(mspr); | |
1159 | } | |
1160 | rcu_read_unlock(); | |
1161 | } | |
1162 | ||
1163 | /** | |
1164 | * Queue the pages for transmission, e.g. a request from postcopy destination | |
1165 | * ms: MigrationStatus in which the queue is held | |
1166 | * rbname: The RAMBlock the request is for - may be NULL (to mean reuse last) | |
1167 | * start: Offset from the start of the RAMBlock | |
1168 | * len: Length (in bytes) to send | |
1169 | * Return: 0 on success | |
1170 | */ | |
1171 | int ram_save_queue_pages(MigrationState *ms, const char *rbname, | |
1172 | ram_addr_t start, ram_addr_t len) | |
1173 | { | |
1174 | RAMBlock *ramblock; | |
1175 | ||
d3bf5418 | 1176 | ms->postcopy_requests++; |
6c595cde DDAG |
1177 | rcu_read_lock(); |
1178 | if (!rbname) { | |
1179 | /* Reuse last RAMBlock */ | |
1180 | ramblock = ms->last_req_rb; | |
1181 | ||
1182 | if (!ramblock) { | |
1183 | /* | |
1184 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1185 | * it's the 1st request. | |
1186 | */ | |
1187 | error_report("ram_save_queue_pages no previous block"); | |
1188 | goto err; | |
1189 | } | |
1190 | } else { | |
1191 | ramblock = qemu_ram_block_by_name(rbname); | |
1192 | ||
1193 | if (!ramblock) { | |
1194 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1195 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1196 | goto err; | |
1197 | } | |
1198 | ms->last_req_rb = ramblock; | |
1199 | } | |
1200 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1201 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1202 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1203 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1204 | __func__, start, len, ramblock->used_length); |
1205 | goto err; | |
1206 | } | |
1207 | ||
1208 | struct MigrationSrcPageRequest *new_entry = | |
1209 | g_malloc0(sizeof(struct MigrationSrcPageRequest)); | |
1210 | new_entry->rb = ramblock; | |
1211 | new_entry->offset = start; | |
1212 | new_entry->len = len; | |
1213 | ||
1214 | memory_region_ref(ramblock->mr); | |
1215 | qemu_mutex_lock(&ms->src_page_req_mutex); | |
1216 | QSIMPLEQ_INSERT_TAIL(&ms->src_page_requests, new_entry, next_req); | |
1217 | qemu_mutex_unlock(&ms->src_page_req_mutex); | |
1218 | rcu_read_unlock(); | |
1219 | ||
1220 | return 0; | |
1221 | ||
1222 | err: | |
1223 | rcu_read_unlock(); | |
1224 | return -1; | |
1225 | } | |
1226 | ||
a82d593b DDAG |
1227 | /** |
1228 | * ram_save_target_page: Save one target page | |
1229 | * | |
1230 | * | |
1231 | * @f: QEMUFile where to send the data | |
1232 | * @block: pointer to block that contains the page we want to send | |
1233 | * @offset: offset inside the block for the page; | |
1234 | * @last_stage: if we are at the completion stage | |
1235 | * @bytes_transferred: increase it with the number of transferred bytes | |
1236 | * @dirty_ram_abs: Address of the start of the dirty page in ram_addr_t space | |
1237 | * | |
1238 | * Returns: Number of pages written. | |
1239 | */ | |
1240 | static int ram_save_target_page(MigrationState *ms, QEMUFile *f, | |
a08f6890 | 1241 | PageSearchStatus *pss, |
a82d593b DDAG |
1242 | bool last_stage, |
1243 | uint64_t *bytes_transferred, | |
1244 | ram_addr_t dirty_ram_abs) | |
1245 | { | |
1246 | int res = 0; | |
1247 | ||
1248 | /* Check the pages is dirty and if it is send it */ | |
1249 | if (migration_bitmap_clear_dirty(dirty_ram_abs)) { | |
1250 | unsigned long *unsentmap; | |
1251 | if (compression_switch && migrate_use_compression()) { | |
a08f6890 | 1252 | res = ram_save_compressed_page(f, pss, |
a82d593b DDAG |
1253 | last_stage, |
1254 | bytes_transferred); | |
1255 | } else { | |
a08f6890 | 1256 | res = ram_save_page(f, pss, last_stage, |
a82d593b DDAG |
1257 | bytes_transferred); |
1258 | } | |
1259 | ||
1260 | if (res < 0) { | |
1261 | return res; | |
1262 | } | |
1263 | unsentmap = atomic_rcu_read(&migration_bitmap_rcu)->unsentmap; | |
1264 | if (unsentmap) { | |
1265 | clear_bit(dirty_ram_abs >> TARGET_PAGE_BITS, unsentmap); | |
1266 | } | |
3fd3c4b3 DDAG |
1267 | /* Only update last_sent_block if a block was actually sent; xbzrle |
1268 | * might have decided the page was identical so didn't bother writing | |
1269 | * to the stream. | |
1270 | */ | |
1271 | if (res > 0) { | |
a08f6890 | 1272 | last_sent_block = pss->block; |
3fd3c4b3 | 1273 | } |
a82d593b DDAG |
1274 | } |
1275 | ||
1276 | return res; | |
1277 | } | |
1278 | ||
1279 | /** | |
cb8d4c8f | 1280 | * ram_save_host_page: Starting at *offset send pages up to the end |
a82d593b DDAG |
1281 | * of the current host page. It's valid for the initial |
1282 | * offset to point into the middle of a host page | |
1283 | * in which case the remainder of the hostpage is sent. | |
1284 | * Only dirty target pages are sent. | |
1285 | * | |
1286 | * Returns: Number of pages written. | |
1287 | * | |
1288 | * @f: QEMUFile where to send the data | |
1289 | * @block: pointer to block that contains the page we want to send | |
1290 | * @offset: offset inside the block for the page; updated to last target page | |
1291 | * sent | |
1292 | * @last_stage: if we are at the completion stage | |
1293 | * @bytes_transferred: increase it with the number of transferred bytes | |
1294 | * @dirty_ram_abs: Address of the start of the dirty page in ram_addr_t space | |
1295 | */ | |
a08f6890 HZ |
1296 | static int ram_save_host_page(MigrationState *ms, QEMUFile *f, |
1297 | PageSearchStatus *pss, | |
1298 | bool last_stage, | |
a82d593b DDAG |
1299 | uint64_t *bytes_transferred, |
1300 | ram_addr_t dirty_ram_abs) | |
1301 | { | |
1302 | int tmppages, pages = 0; | |
1303 | do { | |
a08f6890 | 1304 | tmppages = ram_save_target_page(ms, f, pss, last_stage, |
a82d593b DDAG |
1305 | bytes_transferred, dirty_ram_abs); |
1306 | if (tmppages < 0) { | |
1307 | return tmppages; | |
1308 | } | |
1309 | ||
1310 | pages += tmppages; | |
a08f6890 | 1311 | pss->offset += TARGET_PAGE_SIZE; |
a82d593b | 1312 | dirty_ram_abs += TARGET_PAGE_SIZE; |
a08f6890 | 1313 | } while (pss->offset & (qemu_host_page_size - 1)); |
a82d593b DDAG |
1314 | |
1315 | /* The offset we leave with is the last one we looked at */ | |
a08f6890 | 1316 | pss->offset -= TARGET_PAGE_SIZE; |
a82d593b DDAG |
1317 | return pages; |
1318 | } | |
6c595cde | 1319 | |
56e93d26 JQ |
1320 | /** |
1321 | * ram_find_and_save_block: Finds a dirty page and sends it to f | |
1322 | * | |
1323 | * Called within an RCU critical section. | |
1324 | * | |
1325 | * Returns: The number of pages written | |
1326 | * 0 means no dirty pages | |
1327 | * | |
1328 | * @f: QEMUFile where to send the data | |
1329 | * @last_stage: if we are at the completion stage | |
1330 | * @bytes_transferred: increase it with the number of transferred bytes | |
a82d593b DDAG |
1331 | * |
1332 | * On systems where host-page-size > target-page-size it will send all the | |
1333 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1334 | */ |
1335 | ||
1336 | static int ram_find_and_save_block(QEMUFile *f, bool last_stage, | |
1337 | uint64_t *bytes_transferred) | |
1338 | { | |
b8fb8cb7 | 1339 | PageSearchStatus pss; |
a82d593b | 1340 | MigrationState *ms = migrate_get_current(); |
56e93d26 | 1341 | int pages = 0; |
b9e60928 | 1342 | bool again, found; |
f3f491fc DDAG |
1343 | ram_addr_t dirty_ram_abs; /* Address of the start of the dirty page in |
1344 | ram_addr_t space */ | |
56e93d26 | 1345 | |
b8fb8cb7 DDAG |
1346 | pss.block = last_seen_block; |
1347 | pss.offset = last_offset; | |
1348 | pss.complete_round = false; | |
1349 | ||
1350 | if (!pss.block) { | |
1351 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1352 | } | |
56e93d26 | 1353 | |
b9e60928 | 1354 | do { |
a82d593b DDAG |
1355 | again = true; |
1356 | found = get_queued_page(ms, &pss, &dirty_ram_abs); | |
b9e60928 | 1357 | |
a82d593b DDAG |
1358 | if (!found) { |
1359 | /* priority queue empty, so just search for something dirty */ | |
1360 | found = find_dirty_block(f, &pss, &again, &dirty_ram_abs); | |
1361 | } | |
f3f491fc | 1362 | |
a82d593b | 1363 | if (found) { |
a08f6890 | 1364 | pages = ram_save_host_page(ms, f, &pss, |
a82d593b DDAG |
1365 | last_stage, bytes_transferred, |
1366 | dirty_ram_abs); | |
56e93d26 | 1367 | } |
b9e60928 | 1368 | } while (!pages && again); |
56e93d26 | 1369 | |
b8fb8cb7 DDAG |
1370 | last_seen_block = pss.block; |
1371 | last_offset = pss.offset; | |
56e93d26 JQ |
1372 | |
1373 | return pages; | |
1374 | } | |
1375 | ||
1376 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1377 | { | |
1378 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
1379 | if (zero) { | |
1380 | acct_info.dup_pages += pages; | |
1381 | } else { | |
1382 | acct_info.norm_pages += pages; | |
1383 | bytes_transferred += size; | |
1384 | qemu_update_position(f, size); | |
1385 | } | |
1386 | } | |
1387 | ||
1388 | static ram_addr_t ram_save_remaining(void) | |
1389 | { | |
1390 | return migration_dirty_pages; | |
1391 | } | |
1392 | ||
1393 | uint64_t ram_bytes_remaining(void) | |
1394 | { | |
1395 | return ram_save_remaining() * TARGET_PAGE_SIZE; | |
1396 | } | |
1397 | ||
1398 | uint64_t ram_bytes_transferred(void) | |
1399 | { | |
1400 | return bytes_transferred; | |
1401 | } | |
1402 | ||
1403 | uint64_t ram_bytes_total(void) | |
1404 | { | |
1405 | RAMBlock *block; | |
1406 | uint64_t total = 0; | |
1407 | ||
1408 | rcu_read_lock(); | |
1409 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) | |
1410 | total += block->used_length; | |
1411 | rcu_read_unlock(); | |
1412 | return total; | |
1413 | } | |
1414 | ||
1415 | void free_xbzrle_decoded_buf(void) | |
1416 | { | |
1417 | g_free(xbzrle_decoded_buf); | |
1418 | xbzrle_decoded_buf = NULL; | |
1419 | } | |
1420 | ||
60be6340 DL |
1421 | static void migration_bitmap_free(struct BitmapRcu *bmap) |
1422 | { | |
1423 | g_free(bmap->bmap); | |
f3f491fc | 1424 | g_free(bmap->unsentmap); |
60be6340 DL |
1425 | g_free(bmap); |
1426 | } | |
1427 | ||
6ad2a215 | 1428 | static void ram_migration_cleanup(void *opaque) |
56e93d26 | 1429 | { |
2ff64038 LZ |
1430 | /* caller have hold iothread lock or is in a bh, so there is |
1431 | * no writing race against this migration_bitmap | |
1432 | */ | |
60be6340 DL |
1433 | struct BitmapRcu *bitmap = migration_bitmap_rcu; |
1434 | atomic_rcu_set(&migration_bitmap_rcu, NULL); | |
2ff64038 | 1435 | if (bitmap) { |
56e93d26 | 1436 | memory_global_dirty_log_stop(); |
60be6340 | 1437 | call_rcu(bitmap, migration_bitmap_free, rcu); |
56e93d26 JQ |
1438 | } |
1439 | ||
1440 | XBZRLE_cache_lock(); | |
1441 | if (XBZRLE.cache) { | |
1442 | cache_fini(XBZRLE.cache); | |
1443 | g_free(XBZRLE.encoded_buf); | |
1444 | g_free(XBZRLE.current_buf); | |
1445 | XBZRLE.cache = NULL; | |
1446 | XBZRLE.encoded_buf = NULL; | |
1447 | XBZRLE.current_buf = NULL; | |
1448 | } | |
1449 | XBZRLE_cache_unlock(); | |
1450 | } | |
1451 | ||
56e93d26 JQ |
1452 | static void reset_ram_globals(void) |
1453 | { | |
1454 | last_seen_block = NULL; | |
1455 | last_sent_block = NULL; | |
1456 | last_offset = 0; | |
1457 | last_version = ram_list.version; | |
1458 | ram_bulk_stage = true; | |
1459 | } | |
1460 | ||
1461 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1462 | ||
dd631697 LZ |
1463 | void migration_bitmap_extend(ram_addr_t old, ram_addr_t new) |
1464 | { | |
1465 | /* called in qemu main thread, so there is | |
1466 | * no writing race against this migration_bitmap | |
1467 | */ | |
60be6340 DL |
1468 | if (migration_bitmap_rcu) { |
1469 | struct BitmapRcu *old_bitmap = migration_bitmap_rcu, *bitmap; | |
1470 | bitmap = g_new(struct BitmapRcu, 1); | |
1471 | bitmap->bmap = bitmap_new(new); | |
dd631697 LZ |
1472 | |
1473 | /* prevent migration_bitmap content from being set bit | |
1474 | * by migration_bitmap_sync_range() at the same time. | |
1475 | * it is safe to migration if migration_bitmap is cleared bit | |
1476 | * at the same time. | |
1477 | */ | |
1478 | qemu_mutex_lock(&migration_bitmap_mutex); | |
60be6340 DL |
1479 | bitmap_copy(bitmap->bmap, old_bitmap->bmap, old); |
1480 | bitmap_set(bitmap->bmap, old, new - old); | |
f3f491fc DDAG |
1481 | |
1482 | /* We don't have a way to safely extend the sentmap | |
1483 | * with RCU; so mark it as missing, entry to postcopy | |
1484 | * will fail. | |
1485 | */ | |
1486 | bitmap->unsentmap = NULL; | |
1487 | ||
60be6340 | 1488 | atomic_rcu_set(&migration_bitmap_rcu, bitmap); |
dd631697 LZ |
1489 | qemu_mutex_unlock(&migration_bitmap_mutex); |
1490 | migration_dirty_pages += new - old; | |
60be6340 | 1491 | call_rcu(old_bitmap, migration_bitmap_free, rcu); |
dd631697 LZ |
1492 | } |
1493 | } | |
56e93d26 | 1494 | |
4f2e4252 DDAG |
1495 | /* |
1496 | * 'expected' is the value you expect the bitmap mostly to be full | |
1497 | * of; it won't bother printing lines that are all this value. | |
1498 | * If 'todump' is null the migration bitmap is dumped. | |
1499 | */ | |
1500 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected) | |
1501 | { | |
1502 | int64_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
1503 | ||
1504 | int64_t cur; | |
1505 | int64_t linelen = 128; | |
1506 | char linebuf[129]; | |
1507 | ||
1508 | if (!todump) { | |
1509 | todump = atomic_rcu_read(&migration_bitmap_rcu)->bmap; | |
1510 | } | |
1511 | ||
1512 | for (cur = 0; cur < ram_pages; cur += linelen) { | |
1513 | int64_t curb; | |
1514 | bool found = false; | |
1515 | /* | |
1516 | * Last line; catch the case where the line length | |
1517 | * is longer than remaining ram | |
1518 | */ | |
1519 | if (cur + linelen > ram_pages) { | |
1520 | linelen = ram_pages - cur; | |
1521 | } | |
1522 | for (curb = 0; curb < linelen; curb++) { | |
1523 | bool thisbit = test_bit(cur + curb, todump); | |
1524 | linebuf[curb] = thisbit ? '1' : '.'; | |
1525 | found = found || (thisbit != expected); | |
1526 | } | |
1527 | if (found) { | |
1528 | linebuf[curb] = '\0'; | |
1529 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1530 | } | |
1531 | } | |
1532 | } | |
1533 | ||
e0b266f0 DDAG |
1534 | /* **** functions for postcopy ***** */ |
1535 | ||
1536 | /* | |
1537 | * Callback from postcopy_each_ram_send_discard for each RAMBlock | |
1538 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1539 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
1540 | * start,length: Indexes into the bitmap for the first bit | |
1541 | * representing the named block and length in target-pages | |
1542 | */ | |
1543 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1544 | PostcopyDiscardState *pds, | |
1545 | unsigned long start, | |
1546 | unsigned long length) | |
1547 | { | |
1548 | unsigned long end = start + length; /* one after the end */ | |
1549 | unsigned long current; | |
1550 | unsigned long *unsentmap; | |
1551 | ||
1552 | unsentmap = atomic_rcu_read(&migration_bitmap_rcu)->unsentmap; | |
1553 | for (current = start; current < end; ) { | |
1554 | unsigned long one = find_next_bit(unsentmap, end, current); | |
1555 | ||
1556 | if (one <= end) { | |
1557 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1558 | unsigned long discard_length; | |
1559 | ||
1560 | if (zero >= end) { | |
1561 | discard_length = end - one; | |
1562 | } else { | |
1563 | discard_length = zero - one; | |
1564 | } | |
d688c62d DDAG |
1565 | if (discard_length) { |
1566 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1567 | } | |
e0b266f0 DDAG |
1568 | current = one + discard_length; |
1569 | } else { | |
1570 | current = one; | |
1571 | } | |
1572 | } | |
1573 | ||
1574 | return 0; | |
1575 | } | |
1576 | ||
1577 | /* | |
1578 | * Utility for the outgoing postcopy code. | |
1579 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1580 | * passing it bitmap indexes and name. | |
1581 | * Returns: 0 on success | |
1582 | * (qemu_ram_foreach_block ends up passing unscaled lengths | |
1583 | * which would mean postcopy code would have to deal with target page) | |
1584 | */ | |
1585 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1586 | { | |
1587 | struct RAMBlock *block; | |
1588 | int ret; | |
1589 | ||
1590 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1591 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1592 | PostcopyDiscardState *pds = postcopy_discard_send_init(ms, | |
1593 | first, | |
1594 | block->idstr); | |
1595 | ||
1596 | /* | |
1597 | * Postcopy sends chunks of bitmap over the wire, but it | |
1598 | * just needs indexes at this point, avoids it having | |
1599 | * target page specific code. | |
1600 | */ | |
1601 | ret = postcopy_send_discard_bm_ram(ms, pds, first, | |
1602 | block->used_length >> TARGET_PAGE_BITS); | |
1603 | postcopy_discard_send_finish(ms, pds); | |
1604 | if (ret) { | |
1605 | return ret; | |
1606 | } | |
1607 | } | |
1608 | ||
1609 | return 0; | |
1610 | } | |
1611 | ||
99e314eb DDAG |
1612 | /* |
1613 | * Helper for postcopy_chunk_hostpages; it's called twice to cleanup | |
1614 | * the two bitmaps, that are similar, but one is inverted. | |
1615 | * | |
1616 | * We search for runs of target-pages that don't start or end on a | |
1617 | * host page boundary; | |
1618 | * unsent_pass=true: Cleans up partially unsent host pages by searching | |
1619 | * the unsentmap | |
1620 | * unsent_pass=false: Cleans up partially dirty host pages by searching | |
1621 | * the main migration bitmap | |
1622 | * | |
1623 | */ | |
1624 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1625 | RAMBlock *block, | |
1626 | PostcopyDiscardState *pds) | |
1627 | { | |
1628 | unsigned long *bitmap; | |
1629 | unsigned long *unsentmap; | |
1630 | unsigned int host_ratio = qemu_host_page_size / TARGET_PAGE_SIZE; | |
1631 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1632 | unsigned long len = block->used_length >> TARGET_PAGE_BITS; | |
1633 | unsigned long last = first + (len - 1); | |
1634 | unsigned long run_start; | |
1635 | ||
1636 | bitmap = atomic_rcu_read(&migration_bitmap_rcu)->bmap; | |
1637 | unsentmap = atomic_rcu_read(&migration_bitmap_rcu)->unsentmap; | |
1638 | ||
1639 | if (unsent_pass) { | |
1640 | /* Find a sent page */ | |
1641 | run_start = find_next_zero_bit(unsentmap, last + 1, first); | |
1642 | } else { | |
1643 | /* Find a dirty page */ | |
1644 | run_start = find_next_bit(bitmap, last + 1, first); | |
1645 | } | |
1646 | ||
1647 | while (run_start <= last) { | |
1648 | bool do_fixup = false; | |
1649 | unsigned long fixup_start_addr; | |
1650 | unsigned long host_offset; | |
1651 | ||
1652 | /* | |
1653 | * If the start of this run of pages is in the middle of a host | |
1654 | * page, then we need to fixup this host page. | |
1655 | */ | |
1656 | host_offset = run_start % host_ratio; | |
1657 | if (host_offset) { | |
1658 | do_fixup = true; | |
1659 | run_start -= host_offset; | |
1660 | fixup_start_addr = run_start; | |
1661 | /* For the next pass */ | |
1662 | run_start = run_start + host_ratio; | |
1663 | } else { | |
1664 | /* Find the end of this run */ | |
1665 | unsigned long run_end; | |
1666 | if (unsent_pass) { | |
1667 | run_end = find_next_bit(unsentmap, last + 1, run_start + 1); | |
1668 | } else { | |
1669 | run_end = find_next_zero_bit(bitmap, last + 1, run_start + 1); | |
1670 | } | |
1671 | /* | |
1672 | * If the end isn't at the start of a host page, then the | |
1673 | * run doesn't finish at the end of a host page | |
1674 | * and we need to discard. | |
1675 | */ | |
1676 | host_offset = run_end % host_ratio; | |
1677 | if (host_offset) { | |
1678 | do_fixup = true; | |
1679 | fixup_start_addr = run_end - host_offset; | |
1680 | /* | |
1681 | * This host page has gone, the next loop iteration starts | |
1682 | * from after the fixup | |
1683 | */ | |
1684 | run_start = fixup_start_addr + host_ratio; | |
1685 | } else { | |
1686 | /* | |
1687 | * No discards on this iteration, next loop starts from | |
1688 | * next sent/dirty page | |
1689 | */ | |
1690 | run_start = run_end + 1; | |
1691 | } | |
1692 | } | |
1693 | ||
1694 | if (do_fixup) { | |
1695 | unsigned long page; | |
1696 | ||
1697 | /* Tell the destination to discard this page */ | |
1698 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1699 | /* For the unsent_pass we: | |
1700 | * discard partially sent pages | |
1701 | * For the !unsent_pass (dirty) we: | |
1702 | * discard partially dirty pages that were sent | |
1703 | * (any partially sent pages were already discarded | |
1704 | * by the previous unsent_pass) | |
1705 | */ | |
1706 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1707 | host_ratio); | |
1708 | } | |
1709 | ||
1710 | /* Clean up the bitmap */ | |
1711 | for (page = fixup_start_addr; | |
1712 | page < fixup_start_addr + host_ratio; page++) { | |
1713 | /* All pages in this host page are now not sent */ | |
1714 | set_bit(page, unsentmap); | |
1715 | ||
1716 | /* | |
1717 | * Remark them as dirty, updating the count for any pages | |
1718 | * that weren't previously dirty. | |
1719 | */ | |
1720 | migration_dirty_pages += !test_and_set_bit(page, bitmap); | |
1721 | } | |
1722 | } | |
1723 | ||
1724 | if (unsent_pass) { | |
1725 | /* Find the next sent page for the next iteration */ | |
1726 | run_start = find_next_zero_bit(unsentmap, last + 1, | |
1727 | run_start); | |
1728 | } else { | |
1729 | /* Find the next dirty page for the next iteration */ | |
1730 | run_start = find_next_bit(bitmap, last + 1, run_start); | |
1731 | } | |
1732 | } | |
1733 | } | |
1734 | ||
1735 | /* | |
1736 | * Utility for the outgoing postcopy code. | |
1737 | * | |
1738 | * Discard any partially sent host-page size chunks, mark any partially | |
1739 | * dirty host-page size chunks as all dirty. | |
1740 | * | |
1741 | * Returns: 0 on success | |
1742 | */ | |
1743 | static int postcopy_chunk_hostpages(MigrationState *ms) | |
1744 | { | |
1745 | struct RAMBlock *block; | |
1746 | ||
1747 | if (qemu_host_page_size == TARGET_PAGE_SIZE) { | |
1748 | /* Easy case - TPS==HPS - nothing to be done */ | |
1749 | return 0; | |
1750 | } | |
1751 | ||
1752 | /* Easiest way to make sure we don't resume in the middle of a host-page */ | |
1753 | last_seen_block = NULL; | |
1754 | last_sent_block = NULL; | |
1755 | last_offset = 0; | |
1756 | ||
1757 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1758 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1759 | ||
1760 | PostcopyDiscardState *pds = | |
1761 | postcopy_discard_send_init(ms, first, block->idstr); | |
1762 | ||
1763 | /* First pass: Discard all partially sent host pages */ | |
1764 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1765 | /* | |
1766 | * Second pass: Ensure that all partially dirty host pages are made | |
1767 | * fully dirty. | |
1768 | */ | |
1769 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
1770 | ||
1771 | postcopy_discard_send_finish(ms, pds); | |
1772 | } /* ram_list loop */ | |
1773 | ||
1774 | return 0; | |
1775 | } | |
1776 | ||
e0b266f0 DDAG |
1777 | /* |
1778 | * Transmit the set of pages to be discarded after precopy to the target | |
1779 | * these are pages that: | |
1780 | * a) Have been previously transmitted but are now dirty again | |
1781 | * b) Pages that have never been transmitted, this ensures that | |
1782 | * any pages on the destination that have been mapped by background | |
1783 | * tasks get discarded (transparent huge pages is the specific concern) | |
1784 | * Hopefully this is pretty sparse | |
1785 | */ | |
1786 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1787 | { | |
1788 | int ret; | |
1789 | unsigned long *bitmap, *unsentmap; | |
1790 | ||
1791 | rcu_read_lock(); | |
1792 | ||
1793 | /* This should be our last sync, the src is now paused */ | |
1794 | migration_bitmap_sync(); | |
1795 | ||
1796 | unsentmap = atomic_rcu_read(&migration_bitmap_rcu)->unsentmap; | |
1797 | if (!unsentmap) { | |
1798 | /* We don't have a safe way to resize the sentmap, so | |
1799 | * if the bitmap was resized it will be NULL at this | |
1800 | * point. | |
1801 | */ | |
1802 | error_report("migration ram resized during precopy phase"); | |
1803 | rcu_read_unlock(); | |
1804 | return -EINVAL; | |
1805 | } | |
1806 | ||
99e314eb DDAG |
1807 | /* Deal with TPS != HPS */ |
1808 | ret = postcopy_chunk_hostpages(ms); | |
1809 | if (ret) { | |
1810 | rcu_read_unlock(); | |
1811 | return ret; | |
1812 | } | |
1813 | ||
e0b266f0 DDAG |
1814 | /* |
1815 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1816 | */ | |
1817 | bitmap = atomic_rcu_read(&migration_bitmap_rcu)->bmap; | |
1818 | bitmap_or(unsentmap, unsentmap, bitmap, | |
1819 | last_ram_offset() >> TARGET_PAGE_BITS); | |
1820 | ||
1821 | ||
1822 | trace_ram_postcopy_send_discard_bitmap(); | |
1823 | #ifdef DEBUG_POSTCOPY | |
1824 | ram_debug_dump_bitmap(unsentmap, true); | |
1825 | #endif | |
1826 | ||
1827 | ret = postcopy_each_ram_send_discard(ms); | |
1828 | rcu_read_unlock(); | |
1829 | ||
1830 | return ret; | |
1831 | } | |
1832 | ||
1833 | /* | |
1834 | * At the start of the postcopy phase of migration, any now-dirty | |
1835 | * precopied pages are discarded. | |
1836 | * | |
1837 | * start, length describe a byte address range within the RAMBlock | |
1838 | * | |
1839 | * Returns 0 on success. | |
1840 | */ | |
1841 | int ram_discard_range(MigrationIncomingState *mis, | |
1842 | const char *block_name, | |
1843 | uint64_t start, size_t length) | |
1844 | { | |
1845 | int ret = -1; | |
1846 | ||
1847 | rcu_read_lock(); | |
1848 | RAMBlock *rb = qemu_ram_block_by_name(block_name); | |
1849 | ||
1850 | if (!rb) { | |
1851 | error_report("ram_discard_range: Failed to find block '%s'", | |
1852 | block_name); | |
1853 | goto err; | |
1854 | } | |
1855 | ||
1856 | uint8_t *host_startaddr = rb->host + start; | |
1857 | ||
1858 | if ((uintptr_t)host_startaddr & (qemu_host_page_size - 1)) { | |
1859 | error_report("ram_discard_range: Unaligned start address: %p", | |
1860 | host_startaddr); | |
1861 | goto err; | |
1862 | } | |
1863 | ||
1864 | if ((start + length) <= rb->used_length) { | |
1865 | uint8_t *host_endaddr = host_startaddr + length; | |
1866 | if ((uintptr_t)host_endaddr & (qemu_host_page_size - 1)) { | |
1867 | error_report("ram_discard_range: Unaligned end address: %p", | |
1868 | host_endaddr); | |
1869 | goto err; | |
1870 | } | |
1871 | ret = postcopy_ram_discard_range(mis, host_startaddr, length); | |
1872 | } else { | |
1873 | error_report("ram_discard_range: Overrun block '%s' (%" PRIu64 | |
9458ad6b | 1874 | "/%zx/" RAM_ADDR_FMT")", |
e0b266f0 DDAG |
1875 | block_name, start, length, rb->used_length); |
1876 | } | |
1877 | ||
1878 | err: | |
1879 | rcu_read_unlock(); | |
1880 | ||
1881 | return ret; | |
1882 | } | |
1883 | ||
1884 | ||
56e93d26 JQ |
1885 | /* Each of ram_save_setup, ram_save_iterate and ram_save_complete has |
1886 | * long-running RCU critical section. When rcu-reclaims in the code | |
1887 | * start to become numerous it will be necessary to reduce the | |
1888 | * granularity of these critical sections. | |
1889 | */ | |
1890 | ||
1891 | static int ram_save_setup(QEMUFile *f, void *opaque) | |
1892 | { | |
1893 | RAMBlock *block; | |
1894 | int64_t ram_bitmap_pages; /* Size of bitmap in pages, including gaps */ | |
1895 | ||
56e93d26 JQ |
1896 | dirty_rate_high_cnt = 0; |
1897 | bitmap_sync_count = 0; | |
1898 | migration_bitmap_sync_init(); | |
dd631697 | 1899 | qemu_mutex_init(&migration_bitmap_mutex); |
56e93d26 JQ |
1900 | |
1901 | if (migrate_use_xbzrle()) { | |
1902 | XBZRLE_cache_lock(); | |
1903 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() / | |
1904 | TARGET_PAGE_SIZE, | |
1905 | TARGET_PAGE_SIZE); | |
1906 | if (!XBZRLE.cache) { | |
1907 | XBZRLE_cache_unlock(); | |
1908 | error_report("Error creating cache"); | |
1909 | return -1; | |
1910 | } | |
1911 | XBZRLE_cache_unlock(); | |
1912 | ||
1913 | /* We prefer not to abort if there is no memory */ | |
1914 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
1915 | if (!XBZRLE.encoded_buf) { | |
1916 | error_report("Error allocating encoded_buf"); | |
1917 | return -1; | |
1918 | } | |
1919 | ||
1920 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
1921 | if (!XBZRLE.current_buf) { | |
1922 | error_report("Error allocating current_buf"); | |
1923 | g_free(XBZRLE.encoded_buf); | |
1924 | XBZRLE.encoded_buf = NULL; | |
1925 | return -1; | |
1926 | } | |
1927 | ||
1928 | acct_clear(); | |
1929 | } | |
1930 | ||
49877834 PB |
1931 | /* For memory_global_dirty_log_start below. */ |
1932 | qemu_mutex_lock_iothread(); | |
1933 | ||
56e93d26 JQ |
1934 | qemu_mutex_lock_ramlist(); |
1935 | rcu_read_lock(); | |
1936 | bytes_transferred = 0; | |
1937 | reset_ram_globals(); | |
1938 | ||
1939 | ram_bitmap_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
f3f491fc | 1940 | migration_bitmap_rcu = g_new0(struct BitmapRcu, 1); |
60be6340 DL |
1941 | migration_bitmap_rcu->bmap = bitmap_new(ram_bitmap_pages); |
1942 | bitmap_set(migration_bitmap_rcu->bmap, 0, ram_bitmap_pages); | |
56e93d26 | 1943 | |
f3f491fc DDAG |
1944 | if (migrate_postcopy_ram()) { |
1945 | migration_bitmap_rcu->unsentmap = bitmap_new(ram_bitmap_pages); | |
1946 | bitmap_set(migration_bitmap_rcu->unsentmap, 0, ram_bitmap_pages); | |
1947 | } | |
1948 | ||
56e93d26 JQ |
1949 | /* |
1950 | * Count the total number of pages used by ram blocks not including any | |
1951 | * gaps due to alignment or unplugs. | |
1952 | */ | |
1953 | migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; | |
1954 | ||
1955 | memory_global_dirty_log_start(); | |
1956 | migration_bitmap_sync(); | |
1957 | qemu_mutex_unlock_ramlist(); | |
49877834 | 1958 | qemu_mutex_unlock_iothread(); |
56e93d26 JQ |
1959 | |
1960 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
1961 | ||
1962 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1963 | qemu_put_byte(f, strlen(block->idstr)); | |
1964 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
1965 | qemu_put_be64(f, block->used_length); | |
1966 | } | |
1967 | ||
1968 | rcu_read_unlock(); | |
1969 | ||
1970 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
1971 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
1972 | ||
1973 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
1974 | ||
1975 | return 0; | |
1976 | } | |
1977 | ||
1978 | static int ram_save_iterate(QEMUFile *f, void *opaque) | |
1979 | { | |
1980 | int ret; | |
1981 | int i; | |
1982 | int64_t t0; | |
1983 | int pages_sent = 0; | |
1984 | ||
1985 | rcu_read_lock(); | |
1986 | if (ram_list.version != last_version) { | |
1987 | reset_ram_globals(); | |
1988 | } | |
1989 | ||
1990 | /* Read version before ram_list.blocks */ | |
1991 | smp_rmb(); | |
1992 | ||
1993 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
1994 | ||
1995 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
1996 | i = 0; | |
1997 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
1998 | int pages; | |
1999 | ||
2000 | pages = ram_find_and_save_block(f, false, &bytes_transferred); | |
2001 | /* no more pages to sent */ | |
2002 | if (pages == 0) { | |
2003 | break; | |
2004 | } | |
2005 | pages_sent += pages; | |
2006 | acct_info.iterations++; | |
070afca2 | 2007 | |
56e93d26 JQ |
2008 | /* we want to check in the 1st loop, just in case it was the 1st time |
2009 | and we had to sync the dirty bitmap. | |
2010 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2011 | iterations | |
2012 | */ | |
2013 | if ((i & 63) == 0) { | |
2014 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2015 | if (t1 > MAX_WAIT) { | |
2016 | DPRINTF("big wait: %" PRIu64 " milliseconds, %d iterations\n", | |
2017 | t1, i); | |
2018 | break; | |
2019 | } | |
2020 | } | |
2021 | i++; | |
2022 | } | |
2023 | flush_compressed_data(f); | |
2024 | rcu_read_unlock(); | |
2025 | ||
2026 | /* | |
2027 | * Must occur before EOS (or any QEMUFile operation) | |
2028 | * because of RDMA protocol. | |
2029 | */ | |
2030 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2031 | ||
2032 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2033 | bytes_transferred += 8; | |
2034 | ||
2035 | ret = qemu_file_get_error(f); | |
2036 | if (ret < 0) { | |
2037 | return ret; | |
2038 | } | |
2039 | ||
2040 | return pages_sent; | |
2041 | } | |
2042 | ||
2043 | /* Called with iothread lock */ | |
2044 | static int ram_save_complete(QEMUFile *f, void *opaque) | |
2045 | { | |
2046 | rcu_read_lock(); | |
2047 | ||
663e6c1d DDAG |
2048 | if (!migration_in_postcopy(migrate_get_current())) { |
2049 | migration_bitmap_sync(); | |
2050 | } | |
56e93d26 JQ |
2051 | |
2052 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2053 | ||
2054 | /* try transferring iterative blocks of memory */ | |
2055 | ||
2056 | /* flush all remaining blocks regardless of rate limiting */ | |
2057 | while (true) { | |
2058 | int pages; | |
2059 | ||
2060 | pages = ram_find_and_save_block(f, true, &bytes_transferred); | |
2061 | /* no more blocks to sent */ | |
2062 | if (pages == 0) { | |
2063 | break; | |
2064 | } | |
2065 | } | |
2066 | ||
2067 | flush_compressed_data(f); | |
2068 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); | |
56e93d26 JQ |
2069 | |
2070 | rcu_read_unlock(); | |
d09a6fde | 2071 | |
56e93d26 JQ |
2072 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2073 | ||
2074 | return 0; | |
2075 | } | |
2076 | ||
c31b098f DDAG |
2077 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2078 | uint64_t *non_postcopiable_pending, | |
2079 | uint64_t *postcopiable_pending) | |
56e93d26 JQ |
2080 | { |
2081 | uint64_t remaining_size; | |
2082 | ||
2083 | remaining_size = ram_save_remaining() * TARGET_PAGE_SIZE; | |
2084 | ||
663e6c1d DDAG |
2085 | if (!migration_in_postcopy(migrate_get_current()) && |
2086 | remaining_size < max_size) { | |
56e93d26 JQ |
2087 | qemu_mutex_lock_iothread(); |
2088 | rcu_read_lock(); | |
2089 | migration_bitmap_sync(); | |
2090 | rcu_read_unlock(); | |
2091 | qemu_mutex_unlock_iothread(); | |
2092 | remaining_size = ram_save_remaining() * TARGET_PAGE_SIZE; | |
2093 | } | |
c31b098f DDAG |
2094 | |
2095 | /* We can do postcopy, and all the data is postcopiable */ | |
2096 | *postcopiable_pending += remaining_size; | |
56e93d26 JQ |
2097 | } |
2098 | ||
2099 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2100 | { | |
2101 | unsigned int xh_len; | |
2102 | int xh_flags; | |
063e760a | 2103 | uint8_t *loaded_data; |
56e93d26 JQ |
2104 | |
2105 | if (!xbzrle_decoded_buf) { | |
2106 | xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE); | |
2107 | } | |
063e760a | 2108 | loaded_data = xbzrle_decoded_buf; |
56e93d26 JQ |
2109 | |
2110 | /* extract RLE header */ | |
2111 | xh_flags = qemu_get_byte(f); | |
2112 | xh_len = qemu_get_be16(f); | |
2113 | ||
2114 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2115 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2116 | return -1; | |
2117 | } | |
2118 | ||
2119 | if (xh_len > TARGET_PAGE_SIZE) { | |
2120 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2121 | return -1; | |
2122 | } | |
2123 | /* load data and decode */ | |
063e760a | 2124 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2125 | |
2126 | /* decode RLE */ | |
063e760a | 2127 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2128 | TARGET_PAGE_SIZE) == -1) { |
2129 | error_report("Failed to load XBZRLE page - decode error!"); | |
2130 | return -1; | |
2131 | } | |
2132 | ||
2133 | return 0; | |
2134 | } | |
2135 | ||
2136 | /* Must be called from within a rcu critical section. | |
2137 | * Returns a pointer from within the RCU-protected ram_list. | |
2138 | */ | |
a7180877 | 2139 | /* |
4c4bad48 | 2140 | * Read a RAMBlock ID from the stream f. |
a7180877 DDAG |
2141 | * |
2142 | * f: Stream to read from | |
a7180877 DDAG |
2143 | * flags: Page flags (mostly to see if it's a continuation of previous block) |
2144 | */ | |
4c4bad48 HZ |
2145 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, |
2146 | int flags) | |
56e93d26 JQ |
2147 | { |
2148 | static RAMBlock *block = NULL; | |
2149 | char id[256]; | |
2150 | uint8_t len; | |
2151 | ||
2152 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2153 | if (!block) { |
56e93d26 JQ |
2154 | error_report("Ack, bad migration stream!"); |
2155 | return NULL; | |
2156 | } | |
4c4bad48 | 2157 | return block; |
56e93d26 JQ |
2158 | } |
2159 | ||
2160 | len = qemu_get_byte(f); | |
2161 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2162 | id[len] = 0; | |
2163 | ||
e3dd7493 | 2164 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2165 | if (!block) { |
2166 | error_report("Can't find block %s", id); | |
2167 | return NULL; | |
56e93d26 JQ |
2168 | } |
2169 | ||
4c4bad48 HZ |
2170 | return block; |
2171 | } | |
2172 | ||
2173 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2174 | ram_addr_t offset) | |
2175 | { | |
2176 | if (!offset_in_ramblock(block, offset)) { | |
2177 | return NULL; | |
2178 | } | |
2179 | ||
2180 | return block->host + offset; | |
56e93d26 JQ |
2181 | } |
2182 | ||
2183 | /* | |
2184 | * If a page (or a whole RDMA chunk) has been | |
2185 | * determined to be zero, then zap it. | |
2186 | */ | |
2187 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2188 | { | |
2189 | if (ch != 0 || !is_zero_range(host, size)) { | |
2190 | memset(host, ch, size); | |
2191 | } | |
2192 | } | |
2193 | ||
2194 | static void *do_data_decompress(void *opaque) | |
2195 | { | |
2196 | DecompressParam *param = opaque; | |
2197 | unsigned long pagesize; | |
2198 | ||
2199 | while (!quit_decomp_thread) { | |
2200 | qemu_mutex_lock(¶m->mutex); | |
2201 | while (!param->start && !quit_decomp_thread) { | |
2202 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
73a8912b LL |
2203 | } |
2204 | if (!quit_decomp_thread) { | |
56e93d26 | 2205 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2206 | /* uncompress() will return failed in some case, especially |
2207 | * when the page is dirted when doing the compression, it's | |
2208 | * not a problem because the dirty page will be retransferred | |
2209 | * and uncompress() won't break the data in other pages. | |
2210 | */ | |
2211 | uncompress((Bytef *)param->des, &pagesize, | |
2212 | (const Bytef *)param->compbuf, param->len); | |
56e93d26 | 2213 | } |
73a8912b | 2214 | param->start = false; |
56e93d26 | 2215 | qemu_mutex_unlock(¶m->mutex); |
73a8912b LL |
2216 | |
2217 | qemu_mutex_lock(&decomp_done_lock); | |
2218 | param->done = true; | |
2219 | qemu_cond_signal(&decomp_done_cond); | |
2220 | qemu_mutex_unlock(&decomp_done_lock); | |
56e93d26 JQ |
2221 | } |
2222 | ||
2223 | return NULL; | |
2224 | } | |
2225 | ||
5533b2e9 LL |
2226 | static void wait_for_decompress_done(void) |
2227 | { | |
2228 | int idx, thread_count; | |
2229 | ||
2230 | if (!migrate_use_compression()) { | |
2231 | return; | |
2232 | } | |
2233 | ||
2234 | thread_count = migrate_decompress_threads(); | |
2235 | qemu_mutex_lock(&decomp_done_lock); | |
2236 | for (idx = 0; idx < thread_count; idx++) { | |
2237 | while (!decomp_param[idx].done) { | |
2238 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2239 | } | |
2240 | } | |
2241 | qemu_mutex_unlock(&decomp_done_lock); | |
2242 | } | |
2243 | ||
56e93d26 JQ |
2244 | void migrate_decompress_threads_create(void) |
2245 | { | |
2246 | int i, thread_count; | |
2247 | ||
2248 | thread_count = migrate_decompress_threads(); | |
2249 | decompress_threads = g_new0(QemuThread, thread_count); | |
2250 | decomp_param = g_new0(DecompressParam, thread_count); | |
56e93d26 | 2251 | quit_decomp_thread = false; |
73a8912b LL |
2252 | qemu_mutex_init(&decomp_done_lock); |
2253 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2254 | for (i = 0; i < thread_count; i++) { |
2255 | qemu_mutex_init(&decomp_param[i].mutex); | |
2256 | qemu_cond_init(&decomp_param[i].cond); | |
2257 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2258 | decomp_param[i].done = true; |
56e93d26 JQ |
2259 | qemu_thread_create(decompress_threads + i, "decompress", |
2260 | do_data_decompress, decomp_param + i, | |
2261 | QEMU_THREAD_JOINABLE); | |
2262 | } | |
2263 | } | |
2264 | ||
2265 | void migrate_decompress_threads_join(void) | |
2266 | { | |
2267 | int i, thread_count; | |
2268 | ||
2269 | quit_decomp_thread = true; | |
2270 | thread_count = migrate_decompress_threads(); | |
2271 | for (i = 0; i < thread_count; i++) { | |
2272 | qemu_mutex_lock(&decomp_param[i].mutex); | |
2273 | qemu_cond_signal(&decomp_param[i].cond); | |
2274 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2275 | } | |
2276 | for (i = 0; i < thread_count; i++) { | |
2277 | qemu_thread_join(decompress_threads + i); | |
2278 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2279 | qemu_cond_destroy(&decomp_param[i].cond); | |
2280 | g_free(decomp_param[i].compbuf); | |
2281 | } | |
2282 | g_free(decompress_threads); | |
2283 | g_free(decomp_param); | |
56e93d26 JQ |
2284 | decompress_threads = NULL; |
2285 | decomp_param = NULL; | |
56e93d26 JQ |
2286 | } |
2287 | ||
c1bc6626 | 2288 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2289 | void *host, int len) |
2290 | { | |
2291 | int idx, thread_count; | |
2292 | ||
2293 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2294 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2295 | while (true) { |
2296 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2297 | if (decomp_param[idx].done) { |
c1bc6626 | 2298 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2299 | decomp_param[idx].des = host; |
2300 | decomp_param[idx].len = len; | |
2301 | start_decompression(&decomp_param[idx]); | |
2302 | break; | |
2303 | } | |
2304 | } | |
2305 | if (idx < thread_count) { | |
2306 | break; | |
73a8912b LL |
2307 | } else { |
2308 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2309 | } |
2310 | } | |
73a8912b | 2311 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2312 | } |
2313 | ||
1caddf8a DDAG |
2314 | /* |
2315 | * Allocate data structures etc needed by incoming migration with postcopy-ram | |
2316 | * postcopy-ram's similarly names postcopy_ram_incoming_init does the work | |
2317 | */ | |
2318 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2319 | { | |
2320 | size_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
2321 | ||
2322 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2323 | } | |
2324 | ||
a7180877 DDAG |
2325 | /* |
2326 | * Called in postcopy mode by ram_load(). | |
2327 | * rcu_read_lock is taken prior to this being called. | |
2328 | */ | |
2329 | static int ram_load_postcopy(QEMUFile *f) | |
2330 | { | |
2331 | int flags = 0, ret = 0; | |
2332 | bool place_needed = false; | |
2333 | bool matching_page_sizes = qemu_host_page_size == TARGET_PAGE_SIZE; | |
2334 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
2335 | /* Temporary page that is later 'placed' */ | |
2336 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2337 | void *last_host = NULL; |
a3b6ff6d | 2338 | bool all_zero = false; |
a7180877 DDAG |
2339 | |
2340 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2341 | ram_addr_t addr; | |
2342 | void *host = NULL; | |
2343 | void *page_buffer = NULL; | |
2344 | void *place_source = NULL; | |
2345 | uint8_t ch; | |
a7180877 DDAG |
2346 | |
2347 | addr = qemu_get_be64(f); | |
2348 | flags = addr & ~TARGET_PAGE_MASK; | |
2349 | addr &= TARGET_PAGE_MASK; | |
2350 | ||
2351 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2352 | place_needed = false; | |
2353 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE)) { | |
4c4bad48 HZ |
2354 | RAMBlock *block = ram_block_from_stream(f, flags); |
2355 | ||
2356 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2357 | if (!host) { |
2358 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2359 | ret = -EINVAL; | |
2360 | break; | |
2361 | } | |
a7180877 DDAG |
2362 | /* |
2363 | * Postcopy requires that we place whole host pages atomically. | |
2364 | * To make it atomic, the data is read into a temporary page | |
2365 | * that's moved into place later. | |
2366 | * The migration protocol uses, possibly smaller, target-pages | |
2367 | * however the source ensures it always sends all the components | |
2368 | * of a host page in order. | |
2369 | */ | |
2370 | page_buffer = postcopy_host_page + | |
2371 | ((uintptr_t)host & ~qemu_host_page_mask); | |
2372 | /* If all TP are zero then we can optimise the place */ | |
2373 | if (!((uintptr_t)host & ~qemu_host_page_mask)) { | |
2374 | all_zero = true; | |
c53b7ddc DDAG |
2375 | } else { |
2376 | /* not the 1st TP within the HP */ | |
2377 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2378 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2379 | host, last_host); |
2380 | ret = -EINVAL; | |
2381 | break; | |
2382 | } | |
a7180877 DDAG |
2383 | } |
2384 | ||
c53b7ddc | 2385 | |
a7180877 DDAG |
2386 | /* |
2387 | * If it's the last part of a host page then we place the host | |
2388 | * page | |
2389 | */ | |
2390 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
2391 | ~qemu_host_page_mask) == 0; | |
2392 | place_source = postcopy_host_page; | |
2393 | } | |
c53b7ddc | 2394 | last_host = host; |
a7180877 DDAG |
2395 | |
2396 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
2397 | case RAM_SAVE_FLAG_COMPRESS: | |
2398 | ch = qemu_get_byte(f); | |
2399 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2400 | if (ch) { | |
2401 | all_zero = false; | |
2402 | } | |
2403 | break; | |
2404 | ||
2405 | case RAM_SAVE_FLAG_PAGE: | |
2406 | all_zero = false; | |
2407 | if (!place_needed || !matching_page_sizes) { | |
2408 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2409 | } else { | |
2410 | /* Avoids the qemu_file copy during postcopy, which is | |
2411 | * going to do a copy later; can only do it when we | |
2412 | * do this read in one go (matching page sizes) | |
2413 | */ | |
2414 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2415 | TARGET_PAGE_SIZE); | |
2416 | } | |
2417 | break; | |
2418 | case RAM_SAVE_FLAG_EOS: | |
2419 | /* normal exit */ | |
2420 | break; | |
2421 | default: | |
2422 | error_report("Unknown combination of migration flags: %#x" | |
2423 | " (postcopy mode)", flags); | |
2424 | ret = -EINVAL; | |
2425 | } | |
2426 | ||
2427 | if (place_needed) { | |
2428 | /* This gets called at the last target page in the host page */ | |
2429 | if (all_zero) { | |
2430 | ret = postcopy_place_page_zero(mis, | |
2431 | host + TARGET_PAGE_SIZE - | |
2432 | qemu_host_page_size); | |
2433 | } else { | |
2434 | ret = postcopy_place_page(mis, host + TARGET_PAGE_SIZE - | |
2435 | qemu_host_page_size, | |
2436 | place_source); | |
2437 | } | |
2438 | } | |
2439 | if (!ret) { | |
2440 | ret = qemu_file_get_error(f); | |
2441 | } | |
2442 | } | |
2443 | ||
2444 | return ret; | |
2445 | } | |
2446 | ||
56e93d26 JQ |
2447 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2448 | { | |
2449 | int flags = 0, ret = 0; | |
2450 | static uint64_t seq_iter; | |
2451 | int len = 0; | |
a7180877 DDAG |
2452 | /* |
2453 | * If system is running in postcopy mode, page inserts to host memory must | |
2454 | * be atomic | |
2455 | */ | |
2456 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
56e93d26 JQ |
2457 | |
2458 | seq_iter++; | |
2459 | ||
2460 | if (version_id != 4) { | |
2461 | ret = -EINVAL; | |
2462 | } | |
2463 | ||
2464 | /* This RCU critical section can be very long running. | |
2465 | * When RCU reclaims in the code start to become numerous, | |
2466 | * it will be necessary to reduce the granularity of this | |
2467 | * critical section. | |
2468 | */ | |
2469 | rcu_read_lock(); | |
a7180877 DDAG |
2470 | |
2471 | if (postcopy_running) { | |
2472 | ret = ram_load_postcopy(f); | |
2473 | } | |
2474 | ||
2475 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2476 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2477 | void *host = NULL; |
56e93d26 JQ |
2478 | uint8_t ch; |
2479 | ||
2480 | addr = qemu_get_be64(f); | |
2481 | flags = addr & ~TARGET_PAGE_MASK; | |
2482 | addr &= TARGET_PAGE_MASK; | |
2483 | ||
a776aa15 DDAG |
2484 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE | |
2485 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { | |
4c4bad48 HZ |
2486 | RAMBlock *block = ram_block_from_stream(f, flags); |
2487 | ||
2488 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2489 | if (!host) { |
2490 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2491 | ret = -EINVAL; | |
2492 | break; | |
2493 | } | |
2494 | } | |
2495 | ||
56e93d26 JQ |
2496 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2497 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2498 | /* Synchronize RAM block list */ | |
2499 | total_ram_bytes = addr; | |
2500 | while (!ret && total_ram_bytes) { | |
2501 | RAMBlock *block; | |
56e93d26 JQ |
2502 | char id[256]; |
2503 | ram_addr_t length; | |
2504 | ||
2505 | len = qemu_get_byte(f); | |
2506 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2507 | id[len] = 0; | |
2508 | length = qemu_get_be64(f); | |
2509 | ||
e3dd7493 DDAG |
2510 | block = qemu_ram_block_by_name(id); |
2511 | if (block) { | |
2512 | if (length != block->used_length) { | |
2513 | Error *local_err = NULL; | |
56e93d26 | 2514 | |
fa53a0e5 | 2515 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2516 | &local_err); |
2517 | if (local_err) { | |
2518 | error_report_err(local_err); | |
56e93d26 | 2519 | } |
56e93d26 | 2520 | } |
e3dd7493 DDAG |
2521 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2522 | block->idstr); | |
2523 | } else { | |
56e93d26 JQ |
2524 | error_report("Unknown ramblock \"%s\", cannot " |
2525 | "accept migration", id); | |
2526 | ret = -EINVAL; | |
2527 | } | |
2528 | ||
2529 | total_ram_bytes -= length; | |
2530 | } | |
2531 | break; | |
a776aa15 | 2532 | |
56e93d26 | 2533 | case RAM_SAVE_FLAG_COMPRESS: |
56e93d26 JQ |
2534 | ch = qemu_get_byte(f); |
2535 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2536 | break; | |
a776aa15 | 2537 | |
56e93d26 | 2538 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2539 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2540 | break; | |
56e93d26 | 2541 | |
a776aa15 | 2542 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2543 | len = qemu_get_be32(f); |
2544 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2545 | error_report("Invalid compressed data length: %d", len); | |
2546 | ret = -EINVAL; | |
2547 | break; | |
2548 | } | |
c1bc6626 | 2549 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2550 | break; |
a776aa15 | 2551 | |
56e93d26 | 2552 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2553 | if (load_xbzrle(f, addr, host) < 0) { |
2554 | error_report("Failed to decompress XBZRLE page at " | |
2555 | RAM_ADDR_FMT, addr); | |
2556 | ret = -EINVAL; | |
2557 | break; | |
2558 | } | |
2559 | break; | |
2560 | case RAM_SAVE_FLAG_EOS: | |
2561 | /* normal exit */ | |
2562 | break; | |
2563 | default: | |
2564 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2565 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2566 | } else { |
2567 | error_report("Unknown combination of migration flags: %#x", | |
2568 | flags); | |
2569 | ret = -EINVAL; | |
2570 | } | |
2571 | } | |
2572 | if (!ret) { | |
2573 | ret = qemu_file_get_error(f); | |
2574 | } | |
2575 | } | |
2576 | ||
5533b2e9 | 2577 | wait_for_decompress_done(); |
56e93d26 JQ |
2578 | rcu_read_unlock(); |
2579 | DPRINTF("Completed load of VM with exit code %d seq iteration " | |
2580 | "%" PRIu64 "\n", ret, seq_iter); | |
2581 | return ret; | |
2582 | } | |
2583 | ||
2584 | static SaveVMHandlers savevm_ram_handlers = { | |
2585 | .save_live_setup = ram_save_setup, | |
2586 | .save_live_iterate = ram_save_iterate, | |
763c906b | 2587 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2588 | .save_live_complete_precopy = ram_save_complete, |
56e93d26 JQ |
2589 | .save_live_pending = ram_save_pending, |
2590 | .load_state = ram_load, | |
6ad2a215 | 2591 | .cleanup = ram_migration_cleanup, |
56e93d26 JQ |
2592 | }; |
2593 | ||
2594 | void ram_mig_init(void) | |
2595 | { | |
2596 | qemu_mutex_init(&XBZRLE.lock); | |
2597 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, NULL); | |
2598 | } |