]>
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 | */ | |
e688df6b | 28 | |
1393a485 | 29 | #include "qemu/osdep.h" |
33c11879 | 30 | #include "cpu.h" |
f348b6d1 | 31 | #include "qemu/cutils.h" |
56e93d26 JQ |
32 | #include "qemu/bitops.h" |
33 | #include "qemu/bitmap.h" | |
7205c9ec | 34 | #include "qemu/main-loop.h" |
709e3fe8 | 35 | #include "xbzrle.h" |
7b1e1a22 | 36 | #include "ram.h" |
6666c96a | 37 | #include "migration.h" |
f2a8f0a6 | 38 | #include "migration/register.h" |
7b1e1a22 | 39 | #include "migration/misc.h" |
08a0aee1 | 40 | #include "qemu-file.h" |
be07b0ac | 41 | #include "postcopy-ram.h" |
53d37d36 | 42 | #include "page_cache.h" |
56e93d26 | 43 | #include "qemu/error-report.h" |
e688df6b | 44 | #include "qapi/error.h" |
ab7cbb0b | 45 | #include "qapi/qapi-types-migration.h" |
9af23989 | 46 | #include "qapi/qapi-events-migration.h" |
8acabf69 | 47 | #include "qapi/qmp/qerror.h" |
56e93d26 | 48 | #include "trace.h" |
56e93d26 | 49 | #include "exec/ram_addr.h" |
f9494614 | 50 | #include "exec/target_page.h" |
56e93d26 | 51 | #include "qemu/rcu_queue.h" |
a91246c9 | 52 | #include "migration/colo.h" |
53d37d36 | 53 | #include "block.h" |
af8b7d2b | 54 | #include "sysemu/sysemu.h" |
edd090c7 | 55 | #include "savevm.h" |
b9ee2f7d | 56 | #include "qemu/iov.h" |
d32ca5ad | 57 | #include "multifd.h" |
56e93d26 | 58 | |
56e93d26 JQ |
59 | /***********************************************************/ |
60 | /* ram save/restore */ | |
61 | ||
bb890ed5 JQ |
62 | /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it |
63 | * worked for pages that where filled with the same char. We switched | |
64 | * it to only search for the zero value. And to avoid confusion with | |
65 | * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it. | |
66 | */ | |
67 | ||
56e93d26 | 68 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ |
bb890ed5 | 69 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
70 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
71 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
72 | #define RAM_SAVE_FLAG_EOS 0x10 | |
73 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
74 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
75 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
76 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
77 | ||
56e93d26 JQ |
78 | static inline bool is_zero_range(uint8_t *p, uint64_t size) |
79 | { | |
a1febc49 | 80 | return buffer_is_zero(p, size); |
56e93d26 JQ |
81 | } |
82 | ||
9360447d JQ |
83 | XBZRLECacheStats xbzrle_counters; |
84 | ||
56e93d26 JQ |
85 | /* struct contains XBZRLE cache and a static page |
86 | used by the compression */ | |
87 | static struct { | |
88 | /* buffer used for XBZRLE encoding */ | |
89 | uint8_t *encoded_buf; | |
90 | /* buffer for storing page content */ | |
91 | uint8_t *current_buf; | |
92 | /* Cache for XBZRLE, Protected by lock. */ | |
93 | PageCache *cache; | |
94 | QemuMutex lock; | |
c00e0928 JQ |
95 | /* it will store a page full of zeros */ |
96 | uint8_t *zero_target_page; | |
f265e0e4 JQ |
97 | /* buffer used for XBZRLE decoding */ |
98 | uint8_t *decoded_buf; | |
56e93d26 JQ |
99 | } XBZRLE; |
100 | ||
56e93d26 JQ |
101 | static void XBZRLE_cache_lock(void) |
102 | { | |
103 | if (migrate_use_xbzrle()) | |
104 | qemu_mutex_lock(&XBZRLE.lock); | |
105 | } | |
106 | ||
107 | static void XBZRLE_cache_unlock(void) | |
108 | { | |
109 | if (migrate_use_xbzrle()) | |
110 | qemu_mutex_unlock(&XBZRLE.lock); | |
111 | } | |
112 | ||
3d0684b2 JQ |
113 | /** |
114 | * xbzrle_cache_resize: resize the xbzrle cache | |
115 | * | |
116 | * This function is called from qmp_migrate_set_cache_size in main | |
117 | * thread, possibly while a migration is in progress. A running | |
118 | * migration may be using the cache and might finish during this call, | |
119 | * hence changes to the cache are protected by XBZRLE.lock(). | |
120 | * | |
c9dede2d | 121 | * Returns 0 for success or -1 for error |
3d0684b2 JQ |
122 | * |
123 | * @new_size: new cache size | |
8acabf69 | 124 | * @errp: set *errp if the check failed, with reason |
56e93d26 | 125 | */ |
c9dede2d | 126 | int xbzrle_cache_resize(int64_t new_size, Error **errp) |
56e93d26 JQ |
127 | { |
128 | PageCache *new_cache; | |
c9dede2d | 129 | int64_t ret = 0; |
56e93d26 | 130 | |
8acabf69 JQ |
131 | /* Check for truncation */ |
132 | if (new_size != (size_t)new_size) { | |
133 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
134 | "exceeding address space"); | |
135 | return -1; | |
136 | } | |
137 | ||
2a313e5c JQ |
138 | if (new_size == migrate_xbzrle_cache_size()) { |
139 | /* nothing to do */ | |
c9dede2d | 140 | return 0; |
2a313e5c JQ |
141 | } |
142 | ||
56e93d26 JQ |
143 | XBZRLE_cache_lock(); |
144 | ||
145 | if (XBZRLE.cache != NULL) { | |
80f8dfde | 146 | new_cache = cache_init(new_size, TARGET_PAGE_SIZE, errp); |
56e93d26 | 147 | if (!new_cache) { |
56e93d26 JQ |
148 | ret = -1; |
149 | goto out; | |
150 | } | |
151 | ||
152 | cache_fini(XBZRLE.cache); | |
153 | XBZRLE.cache = new_cache; | |
154 | } | |
56e93d26 JQ |
155 | out: |
156 | XBZRLE_cache_unlock(); | |
157 | return ret; | |
158 | } | |
159 | ||
fbd162e6 YK |
160 | static bool ramblock_is_ignored(RAMBlock *block) |
161 | { | |
162 | return !qemu_ram_is_migratable(block) || | |
163 | (migrate_ignore_shared() && qemu_ram_is_shared(block)); | |
164 | } | |
165 | ||
b895de50 | 166 | /* Should be holding either ram_list.mutex, or the RCU lock. */ |
fbd162e6 YK |
167 | #define RAMBLOCK_FOREACH_NOT_IGNORED(block) \ |
168 | INTERNAL_RAMBLOCK_FOREACH(block) \ | |
169 | if (ramblock_is_ignored(block)) {} else | |
170 | ||
b895de50 | 171 | #define RAMBLOCK_FOREACH_MIGRATABLE(block) \ |
343f632c | 172 | INTERNAL_RAMBLOCK_FOREACH(block) \ |
b895de50 CLG |
173 | if (!qemu_ram_is_migratable(block)) {} else |
174 | ||
343f632c DDAG |
175 | #undef RAMBLOCK_FOREACH |
176 | ||
fbd162e6 YK |
177 | int foreach_not_ignored_block(RAMBlockIterFunc func, void *opaque) |
178 | { | |
179 | RAMBlock *block; | |
180 | int ret = 0; | |
181 | ||
89ac5a1d DDAG |
182 | RCU_READ_LOCK_GUARD(); |
183 | ||
fbd162e6 YK |
184 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
185 | ret = func(block, opaque); | |
186 | if (ret) { | |
187 | break; | |
188 | } | |
189 | } | |
fbd162e6 YK |
190 | return ret; |
191 | } | |
192 | ||
f9494614 AP |
193 | static void ramblock_recv_map_init(void) |
194 | { | |
195 | RAMBlock *rb; | |
196 | ||
fbd162e6 | 197 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
198 | assert(!rb->receivedmap); |
199 | rb->receivedmap = bitmap_new(rb->max_length >> qemu_target_page_bits()); | |
200 | } | |
201 | } | |
202 | ||
203 | int ramblock_recv_bitmap_test(RAMBlock *rb, void *host_addr) | |
204 | { | |
205 | return test_bit(ramblock_recv_bitmap_offset(host_addr, rb), | |
206 | rb->receivedmap); | |
207 | } | |
208 | ||
1cba9f6e DDAG |
209 | bool ramblock_recv_bitmap_test_byte_offset(RAMBlock *rb, uint64_t byte_offset) |
210 | { | |
211 | return test_bit(byte_offset >> TARGET_PAGE_BITS, rb->receivedmap); | |
212 | } | |
213 | ||
f9494614 AP |
214 | void ramblock_recv_bitmap_set(RAMBlock *rb, void *host_addr) |
215 | { | |
216 | set_bit_atomic(ramblock_recv_bitmap_offset(host_addr, rb), rb->receivedmap); | |
217 | } | |
218 | ||
219 | void ramblock_recv_bitmap_set_range(RAMBlock *rb, void *host_addr, | |
220 | size_t nr) | |
221 | { | |
222 | bitmap_set_atomic(rb->receivedmap, | |
223 | ramblock_recv_bitmap_offset(host_addr, rb), | |
224 | nr); | |
225 | } | |
226 | ||
a335debb PX |
227 | #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL) |
228 | ||
229 | /* | |
230 | * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes). | |
231 | * | |
232 | * Returns >0 if success with sent bytes, or <0 if error. | |
233 | */ | |
234 | int64_t ramblock_recv_bitmap_send(QEMUFile *file, | |
235 | const char *block_name) | |
236 | { | |
237 | RAMBlock *block = qemu_ram_block_by_name(block_name); | |
238 | unsigned long *le_bitmap, nbits; | |
239 | uint64_t size; | |
240 | ||
241 | if (!block) { | |
242 | error_report("%s: invalid block name: %s", __func__, block_name); | |
243 | return -1; | |
244 | } | |
245 | ||
246 | nbits = block->used_length >> TARGET_PAGE_BITS; | |
247 | ||
248 | /* | |
249 | * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit | |
250 | * machines we may need 4 more bytes for padding (see below | |
251 | * comment). So extend it a bit before hand. | |
252 | */ | |
253 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
254 | ||
255 | /* | |
256 | * Always use little endian when sending the bitmap. This is | |
257 | * required that when source and destination VMs are not using the | |
258 | * same endianess. (Note: big endian won't work.) | |
259 | */ | |
260 | bitmap_to_le(le_bitmap, block->receivedmap, nbits); | |
261 | ||
262 | /* Size of the bitmap, in bytes */ | |
a725ef9f | 263 | size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
264 | |
265 | /* | |
266 | * size is always aligned to 8 bytes for 64bit machines, but it | |
267 | * may not be true for 32bit machines. We need this padding to | |
268 | * make sure the migration can survive even between 32bit and | |
269 | * 64bit machines. | |
270 | */ | |
271 | size = ROUND_UP(size, 8); | |
272 | ||
273 | qemu_put_be64(file, size); | |
274 | qemu_put_buffer(file, (const uint8_t *)le_bitmap, size); | |
275 | /* | |
276 | * Mark as an end, in case the middle part is screwed up due to | |
277 | * some "misterious" reason. | |
278 | */ | |
279 | qemu_put_be64(file, RAMBLOCK_RECV_BITMAP_ENDING); | |
280 | qemu_fflush(file); | |
281 | ||
bf269906 | 282 | g_free(le_bitmap); |
a335debb PX |
283 | |
284 | if (qemu_file_get_error(file)) { | |
285 | return qemu_file_get_error(file); | |
286 | } | |
287 | ||
288 | return size + sizeof(size); | |
289 | } | |
290 | ||
ec481c6c JQ |
291 | /* |
292 | * An outstanding page request, on the source, having been received | |
293 | * and queued | |
294 | */ | |
295 | struct RAMSrcPageRequest { | |
296 | RAMBlock *rb; | |
297 | hwaddr offset; | |
298 | hwaddr len; | |
299 | ||
300 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
301 | }; | |
302 | ||
6f37bb8b JQ |
303 | /* State of RAM for migration */ |
304 | struct RAMState { | |
204b88b8 JQ |
305 | /* QEMUFile used for this migration */ |
306 | QEMUFile *f; | |
6f37bb8b JQ |
307 | /* Last block that we have visited searching for dirty pages */ |
308 | RAMBlock *last_seen_block; | |
309 | /* Last block from where we have sent data */ | |
310 | RAMBlock *last_sent_block; | |
269ace29 JQ |
311 | /* Last dirty target page we have sent */ |
312 | ram_addr_t last_page; | |
6f37bb8b JQ |
313 | /* last ram version we have seen */ |
314 | uint32_t last_version; | |
315 | /* We are in the first round */ | |
316 | bool ram_bulk_stage; | |
6eeb63f7 WW |
317 | /* The free page optimization is enabled */ |
318 | bool fpo_enabled; | |
8d820d6f JQ |
319 | /* How many times we have dirty too many pages */ |
320 | int dirty_rate_high_cnt; | |
f664da80 JQ |
321 | /* these variables are used for bitmap sync */ |
322 | /* last time we did a full bitmap_sync */ | |
323 | int64_t time_last_bitmap_sync; | |
eac74159 | 324 | /* bytes transferred at start_time */ |
c4bdf0cf | 325 | uint64_t bytes_xfer_prev; |
a66cd90c | 326 | /* number of dirty pages since start_time */ |
68908ed6 | 327 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
328 | /* xbzrle misses since the beginning of the period */ |
329 | uint64_t xbzrle_cache_miss_prev; | |
76e03000 XG |
330 | |
331 | /* compression statistics since the beginning of the period */ | |
332 | /* amount of count that no free thread to compress data */ | |
333 | uint64_t compress_thread_busy_prev; | |
334 | /* amount bytes after compression */ | |
335 | uint64_t compressed_size_prev; | |
336 | /* amount of compressed pages */ | |
337 | uint64_t compress_pages_prev; | |
338 | ||
be8b02ed XG |
339 | /* total handled target pages at the beginning of period */ |
340 | uint64_t target_page_count_prev; | |
341 | /* total handled target pages since start */ | |
342 | uint64_t target_page_count; | |
9360447d | 343 | /* number of dirty bits in the bitmap */ |
2dfaf12e | 344 | uint64_t migration_dirty_pages; |
386a907b | 345 | /* Protects modification of the bitmap and migration dirty pages */ |
108cfae0 | 346 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
347 | /* The RAMBlock used in the last src_page_requests */ |
348 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
349 | /* Queue of outstanding page requests from the destination */ |
350 | QemuMutex src_page_req_mutex; | |
b58deb34 | 351 | QSIMPLEQ_HEAD(, RAMSrcPageRequest) src_page_requests; |
6f37bb8b JQ |
352 | }; |
353 | typedef struct RAMState RAMState; | |
354 | ||
53518d94 | 355 | static RAMState *ram_state; |
6f37bb8b | 356 | |
bd227060 WW |
357 | static NotifierWithReturnList precopy_notifier_list; |
358 | ||
359 | void precopy_infrastructure_init(void) | |
360 | { | |
361 | notifier_with_return_list_init(&precopy_notifier_list); | |
362 | } | |
363 | ||
364 | void precopy_add_notifier(NotifierWithReturn *n) | |
365 | { | |
366 | notifier_with_return_list_add(&precopy_notifier_list, n); | |
367 | } | |
368 | ||
369 | void precopy_remove_notifier(NotifierWithReturn *n) | |
370 | { | |
371 | notifier_with_return_remove(n); | |
372 | } | |
373 | ||
374 | int precopy_notify(PrecopyNotifyReason reason, Error **errp) | |
375 | { | |
376 | PrecopyNotifyData pnd; | |
377 | pnd.reason = reason; | |
378 | pnd.errp = errp; | |
379 | ||
380 | return notifier_with_return_list_notify(&precopy_notifier_list, &pnd); | |
381 | } | |
382 | ||
6eeb63f7 WW |
383 | void precopy_enable_free_page_optimization(void) |
384 | { | |
385 | if (!ram_state) { | |
386 | return; | |
387 | } | |
388 | ||
389 | ram_state->fpo_enabled = true; | |
390 | } | |
391 | ||
9edabd4d | 392 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 393 | { |
bae416e5 DDAG |
394 | return ram_state ? (ram_state->migration_dirty_pages * TARGET_PAGE_SIZE) : |
395 | 0; | |
2f4fde93 JQ |
396 | } |
397 | ||
9360447d | 398 | MigrationStats ram_counters; |
96506894 | 399 | |
b8fb8cb7 DDAG |
400 | /* used by the search for pages to send */ |
401 | struct PageSearchStatus { | |
402 | /* Current block being searched */ | |
403 | RAMBlock *block; | |
a935e30f JQ |
404 | /* Current page to search from */ |
405 | unsigned long page; | |
b8fb8cb7 DDAG |
406 | /* Set once we wrap around */ |
407 | bool complete_round; | |
408 | }; | |
409 | typedef struct PageSearchStatus PageSearchStatus; | |
410 | ||
76e03000 XG |
411 | CompressionStats compression_counters; |
412 | ||
56e93d26 | 413 | struct CompressParam { |
56e93d26 | 414 | bool done; |
90e56fb4 | 415 | bool quit; |
5e5fdcff | 416 | bool zero_page; |
56e93d26 JQ |
417 | QEMUFile *file; |
418 | QemuMutex mutex; | |
419 | QemuCond cond; | |
420 | RAMBlock *block; | |
421 | ram_addr_t offset; | |
34ab9e97 XG |
422 | |
423 | /* internally used fields */ | |
dcaf446e | 424 | z_stream stream; |
34ab9e97 | 425 | uint8_t *originbuf; |
56e93d26 JQ |
426 | }; |
427 | typedef struct CompressParam CompressParam; | |
428 | ||
429 | struct DecompressParam { | |
73a8912b | 430 | bool done; |
90e56fb4 | 431 | bool quit; |
56e93d26 JQ |
432 | QemuMutex mutex; |
433 | QemuCond cond; | |
434 | void *des; | |
d341d9f3 | 435 | uint8_t *compbuf; |
56e93d26 | 436 | int len; |
797ca154 | 437 | z_stream stream; |
56e93d26 JQ |
438 | }; |
439 | typedef struct DecompressParam DecompressParam; | |
440 | ||
441 | static CompressParam *comp_param; | |
442 | static QemuThread *compress_threads; | |
443 | /* comp_done_cond is used to wake up the migration thread when | |
444 | * one of the compression threads has finished the compression. | |
445 | * comp_done_lock is used to co-work with comp_done_cond. | |
446 | */ | |
0d9f9a5c LL |
447 | static QemuMutex comp_done_lock; |
448 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
449 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
450 | static const QEMUFileOps empty_ops = { }; | |
451 | ||
34ab9e97 | 452 | static QEMUFile *decomp_file; |
56e93d26 JQ |
453 | static DecompressParam *decomp_param; |
454 | static QemuThread *decompress_threads; | |
73a8912b LL |
455 | static QemuMutex decomp_done_lock; |
456 | static QemuCond decomp_done_cond; | |
56e93d26 | 457 | |
5e5fdcff | 458 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 459 | ram_addr_t offset, uint8_t *source_buf); |
56e93d26 JQ |
460 | |
461 | static void *do_data_compress(void *opaque) | |
462 | { | |
463 | CompressParam *param = opaque; | |
a7a9a88f LL |
464 | RAMBlock *block; |
465 | ram_addr_t offset; | |
5e5fdcff | 466 | bool zero_page; |
56e93d26 | 467 | |
a7a9a88f | 468 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 469 | while (!param->quit) { |
a7a9a88f LL |
470 | if (param->block) { |
471 | block = param->block; | |
472 | offset = param->offset; | |
473 | param->block = NULL; | |
474 | qemu_mutex_unlock(¶m->mutex); | |
475 | ||
5e5fdcff XG |
476 | zero_page = do_compress_ram_page(param->file, ¶m->stream, |
477 | block, offset, param->originbuf); | |
a7a9a88f | 478 | |
0d9f9a5c | 479 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 480 | param->done = true; |
5e5fdcff | 481 | param->zero_page = zero_page; |
0d9f9a5c LL |
482 | qemu_cond_signal(&comp_done_cond); |
483 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
484 | |
485 | qemu_mutex_lock(¶m->mutex); | |
486 | } else { | |
56e93d26 JQ |
487 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
488 | } | |
56e93d26 | 489 | } |
a7a9a88f | 490 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
491 | |
492 | return NULL; | |
493 | } | |
494 | ||
f0afa331 | 495 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
496 | { |
497 | int i, thread_count; | |
498 | ||
05306935 | 499 | if (!migrate_use_compression() || !comp_param) { |
56e93d26 JQ |
500 | return; |
501 | } | |
05306935 | 502 | |
56e93d26 JQ |
503 | thread_count = migrate_compress_threads(); |
504 | for (i = 0; i < thread_count; i++) { | |
dcaf446e XG |
505 | /* |
506 | * we use it as a indicator which shows if the thread is | |
507 | * properly init'd or not | |
508 | */ | |
509 | if (!comp_param[i].file) { | |
510 | break; | |
511 | } | |
05306935 FL |
512 | |
513 | qemu_mutex_lock(&comp_param[i].mutex); | |
514 | comp_param[i].quit = true; | |
515 | qemu_cond_signal(&comp_param[i].cond); | |
516 | qemu_mutex_unlock(&comp_param[i].mutex); | |
517 | ||
56e93d26 | 518 | qemu_thread_join(compress_threads + i); |
56e93d26 JQ |
519 | qemu_mutex_destroy(&comp_param[i].mutex); |
520 | qemu_cond_destroy(&comp_param[i].cond); | |
dcaf446e | 521 | deflateEnd(&comp_param[i].stream); |
34ab9e97 | 522 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
523 | qemu_fclose(comp_param[i].file); |
524 | comp_param[i].file = NULL; | |
56e93d26 | 525 | } |
0d9f9a5c LL |
526 | qemu_mutex_destroy(&comp_done_lock); |
527 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
528 | g_free(compress_threads); |
529 | g_free(comp_param); | |
56e93d26 JQ |
530 | compress_threads = NULL; |
531 | comp_param = NULL; | |
56e93d26 JQ |
532 | } |
533 | ||
dcaf446e | 534 | static int compress_threads_save_setup(void) |
56e93d26 JQ |
535 | { |
536 | int i, thread_count; | |
537 | ||
538 | if (!migrate_use_compression()) { | |
dcaf446e | 539 | return 0; |
56e93d26 | 540 | } |
56e93d26 JQ |
541 | thread_count = migrate_compress_threads(); |
542 | compress_threads = g_new0(QemuThread, thread_count); | |
543 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
544 | qemu_cond_init(&comp_done_cond); |
545 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 546 | for (i = 0; i < thread_count; i++) { |
34ab9e97 XG |
547 | comp_param[i].originbuf = g_try_malloc(TARGET_PAGE_SIZE); |
548 | if (!comp_param[i].originbuf) { | |
549 | goto exit; | |
550 | } | |
551 | ||
dcaf446e XG |
552 | if (deflateInit(&comp_param[i].stream, |
553 | migrate_compress_level()) != Z_OK) { | |
34ab9e97 | 554 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
555 | goto exit; |
556 | } | |
557 | ||
e110aa91 C |
558 | /* comp_param[i].file is just used as a dummy buffer to save data, |
559 | * set its ops to empty. | |
56e93d26 JQ |
560 | */ |
561 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
562 | comp_param[i].done = true; | |
90e56fb4 | 563 | comp_param[i].quit = false; |
56e93d26 JQ |
564 | qemu_mutex_init(&comp_param[i].mutex); |
565 | qemu_cond_init(&comp_param[i].cond); | |
566 | qemu_thread_create(compress_threads + i, "compress", | |
567 | do_data_compress, comp_param + i, | |
568 | QEMU_THREAD_JOINABLE); | |
569 | } | |
dcaf446e XG |
570 | return 0; |
571 | ||
572 | exit: | |
573 | compress_threads_save_cleanup(); | |
574 | return -1; | |
56e93d26 JQ |
575 | } |
576 | ||
577 | /** | |
3d0684b2 | 578 | * save_page_header: write page header to wire |
56e93d26 JQ |
579 | * |
580 | * If this is the 1st block, it also writes the block identification | |
581 | * | |
3d0684b2 | 582 | * Returns the number of bytes written |
56e93d26 JQ |
583 | * |
584 | * @f: QEMUFile where to send the data | |
585 | * @block: block that contains the page we want to send | |
586 | * @offset: offset inside the block for the page | |
587 | * in the lower bits, it contains flags | |
588 | */ | |
2bf3aa85 JQ |
589 | static size_t save_page_header(RAMState *rs, QEMUFile *f, RAMBlock *block, |
590 | ram_addr_t offset) | |
56e93d26 | 591 | { |
9f5f380b | 592 | size_t size, len; |
56e93d26 | 593 | |
24795694 JQ |
594 | if (block == rs->last_sent_block) { |
595 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
596 | } | |
2bf3aa85 | 597 | qemu_put_be64(f, offset); |
56e93d26 JQ |
598 | size = 8; |
599 | ||
600 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 601 | len = strlen(block->idstr); |
2bf3aa85 JQ |
602 | qemu_put_byte(f, len); |
603 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 604 | size += 1 + len; |
24795694 | 605 | rs->last_sent_block = block; |
56e93d26 JQ |
606 | } |
607 | return size; | |
608 | } | |
609 | ||
3d0684b2 JQ |
610 | /** |
611 | * mig_throttle_guest_down: throotle down the guest | |
612 | * | |
613 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
614 | * writes. If guest dirty memory rate is reduced below the rate at | |
615 | * which we can transfer pages to the destination then we should be | |
616 | * able to complete migration. Some workloads dirty memory way too | |
617 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
618 | */ |
619 | static void mig_throttle_guest_down(void) | |
620 | { | |
621 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
622 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
623 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
4cbc9c7f | 624 | int pct_max = s->parameters.max_cpu_throttle; |
070afca2 JH |
625 | |
626 | /* We have not started throttling yet. Let's start it. */ | |
627 | if (!cpu_throttle_active()) { | |
628 | cpu_throttle_set(pct_initial); | |
629 | } else { | |
630 | /* Throttling already on, just increase the rate */ | |
4cbc9c7f LQ |
631 | cpu_throttle_set(MIN(cpu_throttle_get_percentage() + pct_icrement, |
632 | pct_max)); | |
070afca2 JH |
633 | } |
634 | } | |
635 | ||
3d0684b2 JQ |
636 | /** |
637 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
638 | * | |
6f37bb8b | 639 | * @rs: current RAM state |
3d0684b2 JQ |
640 | * @current_addr: address for the zero page |
641 | * | |
642 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
643 | * The important thing is that a stale (not-yet-0'd) page be replaced |
644 | * by the new data. | |
645 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 646 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 647 | */ |
6f37bb8b | 648 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 649 | { |
6f37bb8b | 650 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
651 | return; |
652 | } | |
653 | ||
654 | /* We don't care if this fails to allocate a new cache page | |
655 | * as long as it updated an old one */ | |
c00e0928 | 656 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 657 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
658 | } |
659 | ||
660 | #define ENCODING_FLAG_XBZRLE 0x1 | |
661 | ||
662 | /** | |
663 | * save_xbzrle_page: compress and send current page | |
664 | * | |
665 | * Returns: 1 means that we wrote the page | |
666 | * 0 means that page is identical to the one already sent | |
667 | * -1 means that xbzrle would be longer than normal | |
668 | * | |
5a987738 | 669 | * @rs: current RAM state |
3d0684b2 JQ |
670 | * @current_data: pointer to the address of the page contents |
671 | * @current_addr: addr of the page | |
56e93d26 JQ |
672 | * @block: block that contains the page we want to send |
673 | * @offset: offset inside the block for the page | |
674 | * @last_stage: if we are at the completion stage | |
56e93d26 | 675 | */ |
204b88b8 | 676 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 677 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 678 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
679 | { |
680 | int encoded_len = 0, bytes_xbzrle; | |
681 | uint8_t *prev_cached_page; | |
682 | ||
9360447d JQ |
683 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
684 | ram_counters.dirty_sync_count)) { | |
685 | xbzrle_counters.cache_miss++; | |
56e93d26 JQ |
686 | if (!last_stage) { |
687 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
9360447d | 688 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
689 | return -1; |
690 | } else { | |
691 | /* update *current_data when the page has been | |
692 | inserted into cache */ | |
693 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
694 | } | |
695 | } | |
696 | return -1; | |
697 | } | |
698 | ||
699 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
700 | ||
701 | /* save current buffer into memory */ | |
702 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
703 | ||
704 | /* XBZRLE encoding (if there is no overflow) */ | |
705 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
706 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
707 | TARGET_PAGE_SIZE); | |
ca353803 WY |
708 | |
709 | /* | |
710 | * Update the cache contents, so that it corresponds to the data | |
711 | * sent, in all cases except where we skip the page. | |
712 | */ | |
713 | if (!last_stage && encoded_len != 0) { | |
714 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
715 | /* | |
716 | * In the case where we couldn't compress, ensure that the caller | |
717 | * sends the data from the cache, since the guest might have | |
718 | * changed the RAM since we copied it. | |
719 | */ | |
720 | *current_data = prev_cached_page; | |
721 | } | |
722 | ||
56e93d26 | 723 | if (encoded_len == 0) { |
55c4446b | 724 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
725 | return 0; |
726 | } else if (encoded_len == -1) { | |
55c4446b | 727 | trace_save_xbzrle_page_overflow(); |
9360447d | 728 | xbzrle_counters.overflow++; |
56e93d26 JQ |
729 | return -1; |
730 | } | |
731 | ||
56e93d26 | 732 | /* Send XBZRLE based compressed page */ |
2bf3aa85 | 733 | bytes_xbzrle = save_page_header(rs, rs->f, block, |
204b88b8 JQ |
734 | offset | RAM_SAVE_FLAG_XBZRLE); |
735 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
736 | qemu_put_be16(rs->f, encoded_len); | |
737 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 738 | bytes_xbzrle += encoded_len + 1 + 2; |
9360447d JQ |
739 | xbzrle_counters.pages++; |
740 | xbzrle_counters.bytes += bytes_xbzrle; | |
741 | ram_counters.transferred += bytes_xbzrle; | |
56e93d26 JQ |
742 | |
743 | return 1; | |
744 | } | |
745 | ||
3d0684b2 JQ |
746 | /** |
747 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 748 | * |
a5f7b1a6 | 749 | * Returns the page offset within memory region of the start of a dirty page |
3d0684b2 | 750 | * |
6f37bb8b | 751 | * @rs: current RAM state |
3d0684b2 | 752 | * @rb: RAMBlock where to search for dirty pages |
a935e30f | 753 | * @start: page where we start the search |
f3f491fc | 754 | */ |
56e93d26 | 755 | static inline |
a935e30f | 756 | unsigned long migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
f20e2865 | 757 | unsigned long start) |
56e93d26 | 758 | { |
6b6712ef JQ |
759 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
760 | unsigned long *bitmap = rb->bmap; | |
56e93d26 JQ |
761 | unsigned long next; |
762 | ||
fbd162e6 | 763 | if (ramblock_is_ignored(rb)) { |
b895de50 CLG |
764 | return size; |
765 | } | |
766 | ||
6eeb63f7 WW |
767 | /* |
768 | * When the free page optimization is enabled, we need to check the bitmap | |
769 | * to send the non-free pages rather than all the pages in the bulk stage. | |
770 | */ | |
771 | if (!rs->fpo_enabled && rs->ram_bulk_stage && start > 0) { | |
6b6712ef | 772 | next = start + 1; |
56e93d26 | 773 | } else { |
6b6712ef | 774 | next = find_next_bit(bitmap, size, start); |
56e93d26 JQ |
775 | } |
776 | ||
6b6712ef | 777 | return next; |
56e93d26 JQ |
778 | } |
779 | ||
06b10688 | 780 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
781 | RAMBlock *rb, |
782 | unsigned long page) | |
a82d593b DDAG |
783 | { |
784 | bool ret; | |
a82d593b | 785 | |
386a907b | 786 | qemu_mutex_lock(&rs->bitmap_mutex); |
002cad6b PX |
787 | |
788 | /* | |
789 | * Clear dirty bitmap if needed. This _must_ be called before we | |
790 | * send any of the page in the chunk because we need to make sure | |
791 | * we can capture further page content changes when we sync dirty | |
792 | * log the next time. So as long as we are going to send any of | |
793 | * the page in the chunk we clear the remote dirty bitmap for all. | |
794 | * Clearing it earlier won't be a problem, but too late will. | |
795 | */ | |
796 | if (rb->clear_bmap && clear_bmap_test_and_clear(rb, page)) { | |
797 | uint8_t shift = rb->clear_bmap_shift; | |
798 | hwaddr size = 1ULL << (TARGET_PAGE_BITS + shift); | |
8bba004c | 799 | hwaddr start = (((ram_addr_t)page) << TARGET_PAGE_BITS) & (-size); |
002cad6b PX |
800 | |
801 | /* | |
802 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
803 | * can make things easier sometimes since then start address | |
804 | * of the small chunk will always be 64 pages aligned so the | |
805 | * bitmap will always be aligned to unsigned long. We should | |
806 | * even be able to remove this restriction but I'm simply | |
807 | * keeping it. | |
808 | */ | |
809 | assert(shift >= 6); | |
810 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); | |
811 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
812 | } | |
813 | ||
6b6712ef | 814 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b DDAG |
815 | |
816 | if (ret) { | |
0d8ec885 | 817 | rs->migration_dirty_pages--; |
a82d593b | 818 | } |
386a907b WW |
819 | qemu_mutex_unlock(&rs->bitmap_mutex); |
820 | ||
a82d593b DDAG |
821 | return ret; |
822 | } | |
823 | ||
267691b6 | 824 | /* Called with RCU critical section */ |
7a3e9571 | 825 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 826 | { |
0d8ec885 | 827 | rs->migration_dirty_pages += |
5d0980a4 | 828 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length, |
0d8ec885 | 829 | &rs->num_dirty_pages_period); |
56e93d26 JQ |
830 | } |
831 | ||
3d0684b2 JQ |
832 | /** |
833 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
834 | * | |
835 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
836 | * | |
837 | * For VMs with just normal pages this is equivalent to the host page | |
838 | * size. If it's got some huge pages then it's the OR of all the | |
839 | * different page sizes. | |
e8ca1db2 DDAG |
840 | */ |
841 | uint64_t ram_pagesize_summary(void) | |
842 | { | |
843 | RAMBlock *block; | |
844 | uint64_t summary = 0; | |
845 | ||
fbd162e6 | 846 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
847 | summary |= block->page_size; |
848 | } | |
849 | ||
850 | return summary; | |
851 | } | |
852 | ||
aecbfe9c XG |
853 | uint64_t ram_get_total_transferred_pages(void) |
854 | { | |
855 | return ram_counters.normal + ram_counters.duplicate + | |
856 | compression_counters.pages + xbzrle_counters.pages; | |
857 | } | |
858 | ||
b734035b XG |
859 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
860 | { | |
be8b02ed | 861 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 862 | double compressed_size; |
b734035b XG |
863 | |
864 | /* calculate period counters */ | |
865 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 | |
866 | / (end_time - rs->time_last_bitmap_sync); | |
867 | ||
be8b02ed | 868 | if (!page_count) { |
b734035b XG |
869 | return; |
870 | } | |
871 | ||
872 | if (migrate_use_xbzrle()) { | |
873 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - | |
be8b02ed | 874 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b XG |
875 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
876 | } | |
76e03000 XG |
877 | |
878 | if (migrate_use_compression()) { | |
879 | compression_counters.busy_rate = (double)(compression_counters.busy - | |
880 | rs->compress_thread_busy_prev) / page_count; | |
881 | rs->compress_thread_busy_prev = compression_counters.busy; | |
882 | ||
883 | compressed_size = compression_counters.compressed_size - | |
884 | rs->compressed_size_prev; | |
885 | if (compressed_size) { | |
886 | double uncompressed_size = (compression_counters.pages - | |
887 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
888 | ||
889 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
890 | compression_counters.compression_rate = | |
891 | uncompressed_size / compressed_size; | |
892 | ||
893 | rs->compress_pages_prev = compression_counters.pages; | |
894 | rs->compressed_size_prev = compression_counters.compressed_size; | |
895 | } | |
896 | } | |
b734035b XG |
897 | } |
898 | ||
dc14a470 KZ |
899 | static void migration_trigger_throttle(RAMState *rs) |
900 | { | |
901 | MigrationState *s = migrate_get_current(); | |
902 | uint64_t threshold = s->parameters.throttle_trigger_threshold; | |
903 | ||
904 | uint64_t bytes_xfer_period = ram_counters.transferred - rs->bytes_xfer_prev; | |
905 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; | |
906 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
907 | ||
908 | /* During block migration the auto-converge logic incorrectly detects | |
909 | * that ram migration makes no progress. Avoid this by disabling the | |
910 | * throttling logic during the bulk phase of block migration. */ | |
911 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
912 | /* The following detection logic can be refined later. For now: | |
913 | Check to see if the ratio between dirtied bytes and the approx. | |
914 | amount of bytes that just got transferred since the last time | |
915 | we were in this routine reaches the threshold. If that happens | |
916 | twice, start or increase throttling. */ | |
917 | ||
918 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
919 | (++rs->dirty_rate_high_cnt >= 2)) { | |
920 | trace_migration_throttle(); | |
921 | rs->dirty_rate_high_cnt = 0; | |
922 | mig_throttle_guest_down(); | |
923 | } | |
924 | } | |
925 | } | |
926 | ||
8d820d6f | 927 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
928 | { |
929 | RAMBlock *block; | |
56e93d26 | 930 | int64_t end_time; |
56e93d26 | 931 | |
9360447d | 932 | ram_counters.dirty_sync_count++; |
56e93d26 | 933 | |
f664da80 JQ |
934 | if (!rs->time_last_bitmap_sync) { |
935 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
936 | } |
937 | ||
938 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 939 | memory_global_dirty_log_sync(); |
56e93d26 | 940 | |
108cfae0 | 941 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
942 | WITH_RCU_READ_LOCK_GUARD() { |
943 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
944 | ramblock_sync_dirty_bitmap(rs, block); | |
945 | } | |
946 | ram_counters.remaining = ram_bytes_remaining(); | |
56e93d26 | 947 | } |
108cfae0 | 948 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 949 | |
9458a9a1 | 950 | memory_global_after_dirty_log_sync(); |
a66cd90c | 951 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 952 | |
56e93d26 JQ |
953 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
954 | ||
955 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 956 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 957 | migration_trigger_throttle(rs); |
070afca2 | 958 | |
b734035b XG |
959 | migration_update_rates(rs, end_time); |
960 | ||
be8b02ed | 961 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
962 | |
963 | /* reset period counters */ | |
f664da80 | 964 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 965 | rs->num_dirty_pages_period = 0; |
dc14a470 | 966 | rs->bytes_xfer_prev = ram_counters.transferred; |
56e93d26 | 967 | } |
4addcd4f | 968 | if (migrate_use_events()) { |
3ab72385 | 969 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count); |
4addcd4f | 970 | } |
56e93d26 JQ |
971 | } |
972 | ||
bd227060 WW |
973 | static void migration_bitmap_sync_precopy(RAMState *rs) |
974 | { | |
975 | Error *local_err = NULL; | |
976 | ||
977 | /* | |
978 | * The current notifier usage is just an optimization to migration, so we | |
979 | * don't stop the normal migration process in the error case. | |
980 | */ | |
981 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
982 | error_report_err(local_err); | |
b4a1733c | 983 | local_err = NULL; |
bd227060 WW |
984 | } |
985 | ||
986 | migration_bitmap_sync(rs); | |
987 | ||
988 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
989 | error_report_err(local_err); | |
990 | } | |
991 | } | |
992 | ||
6c97ec5f XG |
993 | /** |
994 | * save_zero_page_to_file: send the zero page to the file | |
995 | * | |
996 | * Returns the size of data written to the file, 0 means the page is not | |
997 | * a zero page | |
998 | * | |
999 | * @rs: current RAM state | |
1000 | * @file: the file where the data is saved | |
1001 | * @block: block that contains the page we want to send | |
1002 | * @offset: offset inside the block for the page | |
1003 | */ | |
1004 | static int save_zero_page_to_file(RAMState *rs, QEMUFile *file, | |
1005 | RAMBlock *block, ram_addr_t offset) | |
1006 | { | |
1007 | uint8_t *p = block->host + offset; | |
1008 | int len = 0; | |
1009 | ||
1010 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
1011 | len += save_page_header(rs, file, block, offset | RAM_SAVE_FLAG_ZERO); | |
1012 | qemu_put_byte(file, 0); | |
1013 | len += 1; | |
1014 | } | |
1015 | return len; | |
1016 | } | |
1017 | ||
56e93d26 | 1018 | /** |
3d0684b2 | 1019 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1020 | * |
3d0684b2 | 1021 | * Returns the number of pages written. |
56e93d26 | 1022 | * |
f7ccd61b | 1023 | * @rs: current RAM state |
56e93d26 JQ |
1024 | * @block: block that contains the page we want to send |
1025 | * @offset: offset inside the block for the page | |
56e93d26 | 1026 | */ |
7faccdc3 | 1027 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset) |
56e93d26 | 1028 | { |
6c97ec5f | 1029 | int len = save_zero_page_to_file(rs, rs->f, block, offset); |
56e93d26 | 1030 | |
6c97ec5f | 1031 | if (len) { |
9360447d | 1032 | ram_counters.duplicate++; |
6c97ec5f XG |
1033 | ram_counters.transferred += len; |
1034 | return 1; | |
56e93d26 | 1035 | } |
6c97ec5f | 1036 | return -1; |
56e93d26 JQ |
1037 | } |
1038 | ||
5727309d | 1039 | static void ram_release_pages(const char *rbname, uint64_t offset, int pages) |
53f09a10 | 1040 | { |
5727309d | 1041 | if (!migrate_release_ram() || !migration_in_postcopy()) { |
53f09a10 PB |
1042 | return; |
1043 | } | |
1044 | ||
8bba004c | 1045 | ram_discard_range(rbname, offset, ((ram_addr_t)pages) << TARGET_PAGE_BITS); |
53f09a10 PB |
1046 | } |
1047 | ||
059ff0fb XG |
1048 | /* |
1049 | * @pages: the number of pages written by the control path, | |
1050 | * < 0 - error | |
1051 | * > 0 - number of pages written | |
1052 | * | |
1053 | * Return true if the pages has been saved, otherwise false is returned. | |
1054 | */ | |
1055 | static bool control_save_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, | |
1056 | int *pages) | |
1057 | { | |
1058 | uint64_t bytes_xmit = 0; | |
1059 | int ret; | |
1060 | ||
1061 | *pages = -1; | |
1062 | ret = ram_control_save_page(rs->f, block->offset, offset, TARGET_PAGE_SIZE, | |
1063 | &bytes_xmit); | |
1064 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { | |
1065 | return false; | |
1066 | } | |
1067 | ||
1068 | if (bytes_xmit) { | |
1069 | ram_counters.transferred += bytes_xmit; | |
1070 | *pages = 1; | |
1071 | } | |
1072 | ||
1073 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1074 | return true; | |
1075 | } | |
1076 | ||
1077 | if (bytes_xmit > 0) { | |
1078 | ram_counters.normal++; | |
1079 | } else if (bytes_xmit == 0) { | |
1080 | ram_counters.duplicate++; | |
1081 | } | |
1082 | ||
1083 | return true; | |
1084 | } | |
1085 | ||
65dacaa0 XG |
1086 | /* |
1087 | * directly send the page to the stream | |
1088 | * | |
1089 | * Returns the number of pages written. | |
1090 | * | |
1091 | * @rs: current RAM state | |
1092 | * @block: block that contains the page we want to send | |
1093 | * @offset: offset inside the block for the page | |
1094 | * @buf: the page to be sent | |
1095 | * @async: send to page asyncly | |
1096 | */ | |
1097 | static int save_normal_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, | |
1098 | uint8_t *buf, bool async) | |
1099 | { | |
1100 | ram_counters.transferred += save_page_header(rs, rs->f, block, | |
1101 | offset | RAM_SAVE_FLAG_PAGE); | |
1102 | if (async) { | |
1103 | qemu_put_buffer_async(rs->f, buf, TARGET_PAGE_SIZE, | |
1104 | migrate_release_ram() & | |
1105 | migration_in_postcopy()); | |
1106 | } else { | |
1107 | qemu_put_buffer(rs->f, buf, TARGET_PAGE_SIZE); | |
1108 | } | |
1109 | ram_counters.transferred += TARGET_PAGE_SIZE; | |
1110 | ram_counters.normal++; | |
1111 | return 1; | |
1112 | } | |
1113 | ||
56e93d26 | 1114 | /** |
3d0684b2 | 1115 | * ram_save_page: send the given page to the stream |
56e93d26 | 1116 | * |
3d0684b2 | 1117 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1118 | * < 0 - error |
1119 | * >=0 - Number of pages written - this might legally be 0 | |
1120 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1121 | * |
6f37bb8b | 1122 | * @rs: current RAM state |
56e93d26 JQ |
1123 | * @block: block that contains the page we want to send |
1124 | * @offset: offset inside the block for the page | |
1125 | * @last_stage: if we are at the completion stage | |
56e93d26 | 1126 | */ |
a0a8aa14 | 1127 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
1128 | { |
1129 | int pages = -1; | |
56e93d26 | 1130 | uint8_t *p; |
56e93d26 | 1131 | bool send_async = true; |
a08f6890 | 1132 | RAMBlock *block = pss->block; |
8bba004c | 1133 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1134 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1135 | |
2f68e399 | 1136 | p = block->host + offset; |
1db9d8e5 | 1137 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1138 | |
56e93d26 | 1139 | XBZRLE_cache_lock(); |
d7400a34 XG |
1140 | if (!rs->ram_bulk_stage && !migration_in_postcopy() && |
1141 | migrate_use_xbzrle()) { | |
059ff0fb XG |
1142 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
1143 | offset, last_stage); | |
1144 | if (!last_stage) { | |
1145 | /* Can't send this cached data async, since the cache page | |
1146 | * might get updated before it gets to the wire | |
56e93d26 | 1147 | */ |
059ff0fb | 1148 | send_async = false; |
56e93d26 JQ |
1149 | } |
1150 | } | |
1151 | ||
1152 | /* XBZRLE overflow or normal page */ | |
1153 | if (pages == -1) { | |
65dacaa0 | 1154 | pages = save_normal_page(rs, block, offset, p, send_async); |
56e93d26 JQ |
1155 | } |
1156 | ||
1157 | XBZRLE_cache_unlock(); | |
1158 | ||
1159 | return pages; | |
1160 | } | |
1161 | ||
b9ee2f7d JQ |
1162 | static int ram_save_multifd_page(RAMState *rs, RAMBlock *block, |
1163 | ram_addr_t offset) | |
1164 | { | |
67a4c891 | 1165 | if (multifd_queue_page(rs->f, block, offset) < 0) { |
713f762a IR |
1166 | return -1; |
1167 | } | |
b9ee2f7d JQ |
1168 | ram_counters.normal++; |
1169 | ||
1170 | return 1; | |
1171 | } | |
1172 | ||
5e5fdcff | 1173 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 1174 | ram_addr_t offset, uint8_t *source_buf) |
56e93d26 | 1175 | { |
53518d94 | 1176 | RAMState *rs = ram_state; |
a7a9a88f | 1177 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
5e5fdcff | 1178 | bool zero_page = false; |
6ef3771c | 1179 | int ret; |
56e93d26 | 1180 | |
5e5fdcff XG |
1181 | if (save_zero_page_to_file(rs, f, block, offset)) { |
1182 | zero_page = true; | |
1183 | goto exit; | |
1184 | } | |
1185 | ||
6ef3771c | 1186 | save_page_header(rs, f, block, offset | RAM_SAVE_FLAG_COMPRESS_PAGE); |
34ab9e97 XG |
1187 | |
1188 | /* | |
1189 | * copy it to a internal buffer to avoid it being modified by VM | |
1190 | * so that we can catch up the error during compression and | |
1191 | * decompression | |
1192 | */ | |
1193 | memcpy(source_buf, p, TARGET_PAGE_SIZE); | |
6ef3771c XG |
1194 | ret = qemu_put_compression_data(f, stream, source_buf, TARGET_PAGE_SIZE); |
1195 | if (ret < 0) { | |
1196 | qemu_file_set_error(migrate_get_current()->to_dst_file, ret); | |
b3be2896 | 1197 | error_report("compressed data failed!"); |
5e5fdcff | 1198 | return false; |
b3be2896 | 1199 | } |
56e93d26 | 1200 | |
5e5fdcff | 1201 | exit: |
6ef3771c | 1202 | ram_release_pages(block->idstr, offset & TARGET_PAGE_MASK, 1); |
5e5fdcff XG |
1203 | return zero_page; |
1204 | } | |
1205 | ||
1206 | static void | |
1207 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1208 | { | |
76e03000 XG |
1209 | ram_counters.transferred += bytes_xmit; |
1210 | ||
5e5fdcff XG |
1211 | if (param->zero_page) { |
1212 | ram_counters.duplicate++; | |
76e03000 | 1213 | return; |
5e5fdcff | 1214 | } |
76e03000 XG |
1215 | |
1216 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1217 | compression_counters.compressed_size += bytes_xmit - 8; | |
1218 | compression_counters.pages++; | |
56e93d26 JQ |
1219 | } |
1220 | ||
32b05495 XG |
1221 | static bool save_page_use_compression(RAMState *rs); |
1222 | ||
ce25d337 | 1223 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
1224 | { |
1225 | int idx, len, thread_count; | |
1226 | ||
32b05495 | 1227 | if (!save_page_use_compression(rs)) { |
56e93d26 JQ |
1228 | return; |
1229 | } | |
1230 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1231 | |
0d9f9a5c | 1232 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1233 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1234 | while (!comp_param[idx].done) { |
0d9f9a5c | 1235 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1236 | } |
a7a9a88f | 1237 | } |
0d9f9a5c | 1238 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1239 | |
1240 | for (idx = 0; idx < thread_count; idx++) { | |
1241 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1242 | if (!comp_param[idx].quit) { |
ce25d337 | 1243 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
5e5fdcff XG |
1244 | /* |
1245 | * it's safe to fetch zero_page without holding comp_done_lock | |
1246 | * as there is no further request submitted to the thread, | |
1247 | * i.e, the thread should be waiting for a request at this point. | |
1248 | */ | |
1249 | update_compress_thread_counts(&comp_param[idx], len); | |
56e93d26 | 1250 | } |
a7a9a88f | 1251 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1252 | } |
1253 | } | |
1254 | ||
1255 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1256 | ram_addr_t offset) | |
1257 | { | |
1258 | param->block = block; | |
1259 | param->offset = offset; | |
1260 | } | |
1261 | ||
ce25d337 JQ |
1262 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
1263 | ram_addr_t offset) | |
56e93d26 JQ |
1264 | { |
1265 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1d58872a | 1266 | bool wait = migrate_compress_wait_thread(); |
56e93d26 JQ |
1267 | |
1268 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1269 | qemu_mutex_lock(&comp_done_lock); |
1d58872a XG |
1270 | retry: |
1271 | for (idx = 0; idx < thread_count; idx++) { | |
1272 | if (comp_param[idx].done) { | |
1273 | comp_param[idx].done = false; | |
1274 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); | |
1275 | qemu_mutex_lock(&comp_param[idx].mutex); | |
1276 | set_compress_params(&comp_param[idx], block, offset); | |
1277 | qemu_cond_signal(&comp_param[idx].cond); | |
1278 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
1279 | pages = 1; | |
5e5fdcff | 1280 | update_compress_thread_counts(&comp_param[idx], bytes_xmit); |
56e93d26 | 1281 | break; |
56e93d26 JQ |
1282 | } |
1283 | } | |
1d58872a XG |
1284 | |
1285 | /* | |
1286 | * wait for the free thread if the user specifies 'compress-wait-thread', | |
1287 | * otherwise we will post the page out in the main thread as normal page. | |
1288 | */ | |
1289 | if (pages < 0 && wait) { | |
1290 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); | |
1291 | goto retry; | |
1292 | } | |
0d9f9a5c | 1293 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1294 | |
1295 | return pages; | |
1296 | } | |
1297 | ||
3d0684b2 JQ |
1298 | /** |
1299 | * find_dirty_block: find the next dirty page and update any state | |
1300 | * associated with the search process. | |
b9e60928 | 1301 | * |
a5f7b1a6 | 1302 | * Returns true if a page is found |
b9e60928 | 1303 | * |
6f37bb8b | 1304 | * @rs: current RAM state |
3d0684b2 JQ |
1305 | * @pss: data about the state of the current dirty page scan |
1306 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1307 | */ |
f20e2865 | 1308 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1309 | { |
f20e2865 | 1310 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
6f37bb8b | 1311 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1312 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1313 | /* |
1314 | * We've been once around the RAM and haven't found anything. | |
1315 | * Give up. | |
1316 | */ | |
1317 | *again = false; | |
1318 | return false; | |
1319 | } | |
8bba004c AR |
1320 | if ((((ram_addr_t)pss->page) << TARGET_PAGE_BITS) |
1321 | >= pss->block->used_length) { | |
b9e60928 | 1322 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1323 | pss->page = 0; |
b9e60928 DDAG |
1324 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1325 | if (!pss->block) { | |
48df9d80 XG |
1326 | /* |
1327 | * If memory migration starts over, we will meet a dirtied page | |
1328 | * which may still exists in compression threads's ring, so we | |
1329 | * should flush the compressed data to make sure the new page | |
1330 | * is not overwritten by the old one in the destination. | |
1331 | * | |
1332 | * Also If xbzrle is on, stop using the data compression at this | |
1333 | * point. In theory, xbzrle can do better than compression. | |
1334 | */ | |
1335 | flush_compressed_data(rs); | |
1336 | ||
b9e60928 DDAG |
1337 | /* Hit the end of the list */ |
1338 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1339 | /* Flag that we've looped */ | |
1340 | pss->complete_round = true; | |
6f37bb8b | 1341 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1342 | } |
1343 | /* Didn't find anything this time, but try again on the new block */ | |
1344 | *again = true; | |
1345 | return false; | |
1346 | } else { | |
1347 | /* Can go around again, but... */ | |
1348 | *again = true; | |
1349 | /* We've found something so probably don't need to */ | |
1350 | return true; | |
1351 | } | |
1352 | } | |
1353 | ||
3d0684b2 JQ |
1354 | /** |
1355 | * unqueue_page: gets a page of the queue | |
1356 | * | |
a82d593b | 1357 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1358 | * |
3d0684b2 JQ |
1359 | * Returns the block of the page (or NULL if none available) |
1360 | * | |
ec481c6c | 1361 | * @rs: current RAM state |
3d0684b2 | 1362 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1363 | */ |
f20e2865 | 1364 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b DDAG |
1365 | { |
1366 | RAMBlock *block = NULL; | |
1367 | ||
ae526e32 XG |
1368 | if (QSIMPLEQ_EMPTY_ATOMIC(&rs->src_page_requests)) { |
1369 | return NULL; | |
1370 | } | |
1371 | ||
ec481c6c JQ |
1372 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1373 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
1374 | struct RAMSrcPageRequest *entry = | |
1375 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1376 | block = entry->rb; |
1377 | *offset = entry->offset; | |
a82d593b DDAG |
1378 | |
1379 | if (entry->len > TARGET_PAGE_SIZE) { | |
1380 | entry->len -= TARGET_PAGE_SIZE; | |
1381 | entry->offset += TARGET_PAGE_SIZE; | |
1382 | } else { | |
1383 | memory_region_unref(block->mr); | |
ec481c6c | 1384 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b | 1385 | g_free(entry); |
e03a34f8 | 1386 | migration_consume_urgent_request(); |
a82d593b DDAG |
1387 | } |
1388 | } | |
ec481c6c | 1389 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
a82d593b DDAG |
1390 | |
1391 | return block; | |
1392 | } | |
1393 | ||
3d0684b2 | 1394 | /** |
ff1543af | 1395 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
1396 | * |
1397 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1398 | * |
a5f7b1a6 | 1399 | * Returns true if a queued page is found |
a82d593b | 1400 | * |
6f37bb8b | 1401 | * @rs: current RAM state |
3d0684b2 | 1402 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1403 | */ |
f20e2865 | 1404 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1405 | { |
1406 | RAMBlock *block; | |
1407 | ram_addr_t offset; | |
1408 | bool dirty; | |
1409 | ||
1410 | do { | |
f20e2865 | 1411 | block = unqueue_page(rs, &offset); |
a82d593b DDAG |
1412 | /* |
1413 | * We're sending this page, and since it's postcopy nothing else | |
1414 | * will dirty it, and we must make sure it doesn't get sent again | |
1415 | * even if this queue request was received after the background | |
1416 | * search already sent it. | |
1417 | */ | |
1418 | if (block) { | |
f20e2865 JQ |
1419 | unsigned long page; |
1420 | ||
6b6712ef JQ |
1421 | page = offset >> TARGET_PAGE_BITS; |
1422 | dirty = test_bit(page, block->bmap); | |
a82d593b | 1423 | if (!dirty) { |
06b10688 | 1424 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, |
64737606 | 1425 | page); |
a82d593b | 1426 | } else { |
f20e2865 | 1427 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); |
a82d593b DDAG |
1428 | } |
1429 | } | |
1430 | ||
1431 | } while (block && !dirty); | |
1432 | ||
1433 | if (block) { | |
1434 | /* | |
1435 | * As soon as we start servicing pages out of order, then we have | |
1436 | * to kill the bulk stage, since the bulk stage assumes | |
1437 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1438 | * dirty, that's no longer true. | |
1439 | */ | |
6f37bb8b | 1440 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1441 | |
1442 | /* | |
1443 | * We want the background search to continue from the queued page | |
1444 | * since the guest is likely to want other pages near to the page | |
1445 | * it just requested. | |
1446 | */ | |
1447 | pss->block = block; | |
a935e30f | 1448 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
1449 | |
1450 | /* | |
1451 | * This unqueued page would break the "one round" check, even is | |
1452 | * really rare. | |
1453 | */ | |
1454 | pss->complete_round = false; | |
a82d593b DDAG |
1455 | } |
1456 | ||
1457 | return !!block; | |
1458 | } | |
1459 | ||
6c595cde | 1460 | /** |
5e58f968 JQ |
1461 | * migration_page_queue_free: drop any remaining pages in the ram |
1462 | * request queue | |
6c595cde | 1463 | * |
3d0684b2 JQ |
1464 | * It should be empty at the end anyway, but in error cases there may |
1465 | * be some left. in case that there is any page left, we drop it. | |
1466 | * | |
6c595cde | 1467 | */ |
83c13382 | 1468 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 1469 | { |
ec481c6c | 1470 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
1471 | /* This queue generally should be empty - but in the case of a failed |
1472 | * migration might have some droppings in. | |
1473 | */ | |
89ac5a1d | 1474 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 1475 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1476 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1477 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1478 | g_free(mspr); |
1479 | } | |
6c595cde DDAG |
1480 | } |
1481 | ||
1482 | /** | |
3d0684b2 JQ |
1483 | * ram_save_queue_pages: queue the page for transmission |
1484 | * | |
1485 | * A request from postcopy destination for example. | |
1486 | * | |
1487 | * Returns zero on success or negative on error | |
1488 | * | |
3d0684b2 JQ |
1489 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1490 | * same that last one. | |
1491 | * @start: starting address from the start of the RAMBlock | |
1492 | * @len: length (in bytes) to send | |
6c595cde | 1493 | */ |
96506894 | 1494 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1495 | { |
1496 | RAMBlock *ramblock; | |
53518d94 | 1497 | RAMState *rs = ram_state; |
6c595cde | 1498 | |
9360447d | 1499 | ram_counters.postcopy_requests++; |
89ac5a1d DDAG |
1500 | RCU_READ_LOCK_GUARD(); |
1501 | ||
6c595cde DDAG |
1502 | if (!rbname) { |
1503 | /* Reuse last RAMBlock */ | |
68a098f3 | 1504 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1505 | |
1506 | if (!ramblock) { | |
1507 | /* | |
1508 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1509 | * it's the 1st request. | |
1510 | */ | |
1511 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 1512 | return -1; |
6c595cde DDAG |
1513 | } |
1514 | } else { | |
1515 | ramblock = qemu_ram_block_by_name(rbname); | |
1516 | ||
1517 | if (!ramblock) { | |
1518 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1519 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 1520 | return -1; |
6c595cde | 1521 | } |
68a098f3 | 1522 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1523 | } |
1524 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1525 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1526 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1527 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 1528 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 1529 | return -1; |
6c595cde DDAG |
1530 | } |
1531 | ||
ec481c6c JQ |
1532 | struct RAMSrcPageRequest *new_entry = |
1533 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
1534 | new_entry->rb = ramblock; |
1535 | new_entry->offset = start; | |
1536 | new_entry->len = len; | |
1537 | ||
1538 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
1539 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1540 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 1541 | migration_make_urgent_request(); |
ec481c6c | 1542 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
1543 | |
1544 | return 0; | |
6c595cde DDAG |
1545 | } |
1546 | ||
d7400a34 XG |
1547 | static bool save_page_use_compression(RAMState *rs) |
1548 | { | |
1549 | if (!migrate_use_compression()) { | |
1550 | return false; | |
1551 | } | |
1552 | ||
1553 | /* | |
1554 | * If xbzrle is on, stop using the data compression after first | |
1555 | * round of migration even if compression is enabled. In theory, | |
1556 | * xbzrle can do better than compression. | |
1557 | */ | |
1558 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { | |
1559 | return true; | |
1560 | } | |
1561 | ||
1562 | return false; | |
1563 | } | |
1564 | ||
5e5fdcff XG |
1565 | /* |
1566 | * try to compress the page before posting it out, return true if the page | |
1567 | * has been properly handled by compression, otherwise needs other | |
1568 | * paths to handle it | |
1569 | */ | |
1570 | static bool save_compress_page(RAMState *rs, RAMBlock *block, ram_addr_t offset) | |
1571 | { | |
1572 | if (!save_page_use_compression(rs)) { | |
1573 | return false; | |
1574 | } | |
1575 | ||
1576 | /* | |
1577 | * When starting the process of a new block, the first page of | |
1578 | * the block should be sent out before other pages in the same | |
1579 | * block, and all the pages in last block should have been sent | |
1580 | * out, keeping this order is important, because the 'cont' flag | |
1581 | * is used to avoid resending the block name. | |
1582 | * | |
1583 | * We post the fist page as normal page as compression will take | |
1584 | * much CPU resource. | |
1585 | */ | |
1586 | if (block != rs->last_sent_block) { | |
1587 | flush_compressed_data(rs); | |
1588 | return false; | |
1589 | } | |
1590 | ||
1591 | if (compress_page_with_multi_thread(rs, block, offset) > 0) { | |
1592 | return true; | |
1593 | } | |
1594 | ||
76e03000 | 1595 | compression_counters.busy++; |
5e5fdcff XG |
1596 | return false; |
1597 | } | |
1598 | ||
a82d593b | 1599 | /** |
3d0684b2 | 1600 | * ram_save_target_page: save one target page |
a82d593b | 1601 | * |
3d0684b2 | 1602 | * Returns the number of pages written |
a82d593b | 1603 | * |
6f37bb8b | 1604 | * @rs: current RAM state |
3d0684b2 | 1605 | * @pss: data about the page we want to send |
a82d593b | 1606 | * @last_stage: if we are at the completion stage |
a82d593b | 1607 | */ |
a0a8aa14 | 1608 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1609 | bool last_stage) |
a82d593b | 1610 | { |
a8ec91f9 | 1611 | RAMBlock *block = pss->block; |
8bba004c | 1612 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
1613 | int res; |
1614 | ||
1615 | if (control_save_page(rs, block, offset, &res)) { | |
1616 | return res; | |
1617 | } | |
1618 | ||
5e5fdcff XG |
1619 | if (save_compress_page(rs, block, offset)) { |
1620 | return 1; | |
d7400a34 XG |
1621 | } |
1622 | ||
1623 | res = save_zero_page(rs, block, offset); | |
1624 | if (res > 0) { | |
1625 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
1626 | * page would be stale | |
1627 | */ | |
1628 | if (!save_page_use_compression(rs)) { | |
1629 | XBZRLE_cache_lock(); | |
1630 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
1631 | XBZRLE_cache_unlock(); | |
1632 | } | |
1633 | ram_release_pages(block->idstr, offset, res); | |
1634 | return res; | |
1635 | } | |
1636 | ||
da3f56cb | 1637 | /* |
c6b3a2e0 WY |
1638 | * Do not use multifd for: |
1639 | * 1. Compression as the first page in the new block should be posted out | |
1640 | * before sending the compressed page | |
1641 | * 2. In postcopy as one whole host page should be placed | |
da3f56cb | 1642 | */ |
c6b3a2e0 WY |
1643 | if (!save_page_use_compression(rs) && migrate_use_multifd() |
1644 | && !migration_in_postcopy()) { | |
b9ee2f7d | 1645 | return ram_save_multifd_page(rs, block, offset); |
a82d593b DDAG |
1646 | } |
1647 | ||
1faa5665 | 1648 | return ram_save_page(rs, pss, last_stage); |
a82d593b DDAG |
1649 | } |
1650 | ||
1651 | /** | |
3d0684b2 | 1652 | * ram_save_host_page: save a whole host page |
a82d593b | 1653 | * |
3d0684b2 JQ |
1654 | * Starting at *offset send pages up to the end of the current host |
1655 | * page. It's valid for the initial offset to point into the middle of | |
1656 | * a host page in which case the remainder of the hostpage is sent. | |
1657 | * Only dirty target pages are sent. Note that the host page size may | |
1658 | * be a huge page for this block. | |
1eb3fc0a DDAG |
1659 | * The saving stops at the boundary of the used_length of the block |
1660 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 1661 | * |
3d0684b2 JQ |
1662 | * Returns the number of pages written or negative on error |
1663 | * | |
6f37bb8b | 1664 | * @rs: current RAM state |
3d0684b2 | 1665 | * @ms: current migration state |
3d0684b2 | 1666 | * @pss: data about the page we want to send |
a82d593b | 1667 | * @last_stage: if we are at the completion stage |
a82d593b | 1668 | */ |
a0a8aa14 | 1669 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1670 | bool last_stage) |
a82d593b DDAG |
1671 | { |
1672 | int tmppages, pages = 0; | |
a935e30f JQ |
1673 | size_t pagesize_bits = |
1674 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
4c011c37 | 1675 | |
fbd162e6 | 1676 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
1677 | error_report("block %s should not be migrated !", pss->block->idstr); |
1678 | return 0; | |
1679 | } | |
1680 | ||
a82d593b | 1681 | do { |
1faa5665 XG |
1682 | /* Check the pages is dirty and if it is send it */ |
1683 | if (!migration_bitmap_clear_dirty(rs, pss->block, pss->page)) { | |
1684 | pss->page++; | |
1685 | continue; | |
1686 | } | |
1687 | ||
f20e2865 | 1688 | tmppages = ram_save_target_page(rs, pss, last_stage); |
a82d593b DDAG |
1689 | if (tmppages < 0) { |
1690 | return tmppages; | |
1691 | } | |
1692 | ||
1693 | pages += tmppages; | |
a935e30f | 1694 | pss->page++; |
97e1e067 DDAG |
1695 | /* Allow rate limiting to happen in the middle of huge pages */ |
1696 | migration_rate_limit(); | |
1eb3fc0a | 1697 | } while ((pss->page & (pagesize_bits - 1)) && |
8bba004c AR |
1698 | offset_in_ramblock(pss->block, |
1699 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)); | |
a82d593b DDAG |
1700 | |
1701 | /* The offset we leave with is the last one we looked at */ | |
a935e30f | 1702 | pss->page--; |
a82d593b DDAG |
1703 | return pages; |
1704 | } | |
6c595cde | 1705 | |
56e93d26 | 1706 | /** |
3d0684b2 | 1707 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1708 | * |
1709 | * Called within an RCU critical section. | |
1710 | * | |
e8f3735f XG |
1711 | * Returns the number of pages written where zero means no dirty pages, |
1712 | * or negative on error | |
56e93d26 | 1713 | * |
6f37bb8b | 1714 | * @rs: current RAM state |
56e93d26 | 1715 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1716 | * |
1717 | * On systems where host-page-size > target-page-size it will send all the | |
1718 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1719 | */ |
1720 | ||
ce25d337 | 1721 | static int ram_find_and_save_block(RAMState *rs, bool last_stage) |
56e93d26 | 1722 | { |
b8fb8cb7 | 1723 | PageSearchStatus pss; |
56e93d26 | 1724 | int pages = 0; |
b9e60928 | 1725 | bool again, found; |
56e93d26 | 1726 | |
0827b9e9 AA |
1727 | /* No dirty page as there is zero RAM */ |
1728 | if (!ram_bytes_total()) { | |
1729 | return pages; | |
1730 | } | |
1731 | ||
6f37bb8b | 1732 | pss.block = rs->last_seen_block; |
a935e30f | 1733 | pss.page = rs->last_page; |
b8fb8cb7 DDAG |
1734 | pss.complete_round = false; |
1735 | ||
1736 | if (!pss.block) { | |
1737 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1738 | } | |
56e93d26 | 1739 | |
b9e60928 | 1740 | do { |
a82d593b | 1741 | again = true; |
f20e2865 | 1742 | found = get_queued_page(rs, &pss); |
b9e60928 | 1743 | |
a82d593b DDAG |
1744 | if (!found) { |
1745 | /* priority queue empty, so just search for something dirty */ | |
f20e2865 | 1746 | found = find_dirty_block(rs, &pss, &again); |
a82d593b | 1747 | } |
f3f491fc | 1748 | |
a82d593b | 1749 | if (found) { |
f20e2865 | 1750 | pages = ram_save_host_page(rs, &pss, last_stage); |
56e93d26 | 1751 | } |
b9e60928 | 1752 | } while (!pages && again); |
56e93d26 | 1753 | |
6f37bb8b | 1754 | rs->last_seen_block = pss.block; |
a935e30f | 1755 | rs->last_page = pss.page; |
56e93d26 JQ |
1756 | |
1757 | return pages; | |
1758 | } | |
1759 | ||
1760 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1761 | { | |
1762 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 1763 | |
56e93d26 | 1764 | if (zero) { |
9360447d | 1765 | ram_counters.duplicate += pages; |
56e93d26 | 1766 | } else { |
9360447d JQ |
1767 | ram_counters.normal += pages; |
1768 | ram_counters.transferred += size; | |
56e93d26 JQ |
1769 | qemu_update_position(f, size); |
1770 | } | |
1771 | } | |
1772 | ||
fbd162e6 | 1773 | static uint64_t ram_bytes_total_common(bool count_ignored) |
56e93d26 JQ |
1774 | { |
1775 | RAMBlock *block; | |
1776 | uint64_t total = 0; | |
1777 | ||
89ac5a1d DDAG |
1778 | RCU_READ_LOCK_GUARD(); |
1779 | ||
fbd162e6 YK |
1780 | if (count_ignored) { |
1781 | RAMBLOCK_FOREACH_MIGRATABLE(block) { | |
1782 | total += block->used_length; | |
1783 | } | |
1784 | } else { | |
1785 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1786 | total += block->used_length; | |
1787 | } | |
99e15582 | 1788 | } |
56e93d26 JQ |
1789 | return total; |
1790 | } | |
1791 | ||
fbd162e6 YK |
1792 | uint64_t ram_bytes_total(void) |
1793 | { | |
1794 | return ram_bytes_total_common(false); | |
1795 | } | |
1796 | ||
f265e0e4 | 1797 | static void xbzrle_load_setup(void) |
56e93d26 | 1798 | { |
f265e0e4 | 1799 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
1800 | } |
1801 | ||
f265e0e4 JQ |
1802 | static void xbzrle_load_cleanup(void) |
1803 | { | |
1804 | g_free(XBZRLE.decoded_buf); | |
1805 | XBZRLE.decoded_buf = NULL; | |
1806 | } | |
1807 | ||
7d7c96be PX |
1808 | static void ram_state_cleanup(RAMState **rsp) |
1809 | { | |
b9ccaf6d DDAG |
1810 | if (*rsp) { |
1811 | migration_page_queue_free(*rsp); | |
1812 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
1813 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
1814 | g_free(*rsp); | |
1815 | *rsp = NULL; | |
1816 | } | |
7d7c96be PX |
1817 | } |
1818 | ||
84593a08 PX |
1819 | static void xbzrle_cleanup(void) |
1820 | { | |
1821 | XBZRLE_cache_lock(); | |
1822 | if (XBZRLE.cache) { | |
1823 | cache_fini(XBZRLE.cache); | |
1824 | g_free(XBZRLE.encoded_buf); | |
1825 | g_free(XBZRLE.current_buf); | |
1826 | g_free(XBZRLE.zero_target_page); | |
1827 | XBZRLE.cache = NULL; | |
1828 | XBZRLE.encoded_buf = NULL; | |
1829 | XBZRLE.current_buf = NULL; | |
1830 | XBZRLE.zero_target_page = NULL; | |
1831 | } | |
1832 | XBZRLE_cache_unlock(); | |
1833 | } | |
1834 | ||
f265e0e4 | 1835 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 1836 | { |
53518d94 | 1837 | RAMState **rsp = opaque; |
6b6712ef | 1838 | RAMBlock *block; |
eb859c53 | 1839 | |
2ff64038 | 1840 | /* caller have hold iothread lock or is in a bh, so there is |
4633456c | 1841 | * no writing race against the migration bitmap |
2ff64038 | 1842 | */ |
6b6712ef JQ |
1843 | memory_global_dirty_log_stop(); |
1844 | ||
fbd162e6 | 1845 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
1846 | g_free(block->clear_bmap); |
1847 | block->clear_bmap = NULL; | |
6b6712ef JQ |
1848 | g_free(block->bmap); |
1849 | block->bmap = NULL; | |
56e93d26 JQ |
1850 | } |
1851 | ||
84593a08 | 1852 | xbzrle_cleanup(); |
f0afa331 | 1853 | compress_threads_save_cleanup(); |
7d7c96be | 1854 | ram_state_cleanup(rsp); |
56e93d26 JQ |
1855 | } |
1856 | ||
6f37bb8b | 1857 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1858 | { |
6f37bb8b JQ |
1859 | rs->last_seen_block = NULL; |
1860 | rs->last_sent_block = NULL; | |
269ace29 | 1861 | rs->last_page = 0; |
6f37bb8b JQ |
1862 | rs->last_version = ram_list.version; |
1863 | rs->ram_bulk_stage = true; | |
6eeb63f7 | 1864 | rs->fpo_enabled = false; |
56e93d26 JQ |
1865 | } |
1866 | ||
1867 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1868 | ||
4f2e4252 DDAG |
1869 | /* |
1870 | * 'expected' is the value you expect the bitmap mostly to be full | |
1871 | * of; it won't bother printing lines that are all this value. | |
1872 | * If 'todump' is null the migration bitmap is dumped. | |
1873 | */ | |
6b6712ef JQ |
1874 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected, |
1875 | unsigned long pages) | |
4f2e4252 | 1876 | { |
4f2e4252 DDAG |
1877 | int64_t cur; |
1878 | int64_t linelen = 128; | |
1879 | char linebuf[129]; | |
1880 | ||
6b6712ef | 1881 | for (cur = 0; cur < pages; cur += linelen) { |
4f2e4252 DDAG |
1882 | int64_t curb; |
1883 | bool found = false; | |
1884 | /* | |
1885 | * Last line; catch the case where the line length | |
1886 | * is longer than remaining ram | |
1887 | */ | |
6b6712ef JQ |
1888 | if (cur + linelen > pages) { |
1889 | linelen = pages - cur; | |
4f2e4252 DDAG |
1890 | } |
1891 | for (curb = 0; curb < linelen; curb++) { | |
1892 | bool thisbit = test_bit(cur + curb, todump); | |
1893 | linebuf[curb] = thisbit ? '1' : '.'; | |
1894 | found = found || (thisbit != expected); | |
1895 | } | |
1896 | if (found) { | |
1897 | linebuf[curb] = '\0'; | |
1898 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1899 | } | |
1900 | } | |
1901 | } | |
1902 | ||
e0b266f0 DDAG |
1903 | /* **** functions for postcopy ***** */ |
1904 | ||
ced1c616 PB |
1905 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1906 | { | |
1907 | struct RAMBlock *block; | |
ced1c616 | 1908 | |
fbd162e6 | 1909 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
1910 | unsigned long *bitmap = block->bmap; |
1911 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
1912 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
1913 | |
1914 | while (run_start < range) { | |
1915 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
1916 | ram_discard_range(block->idstr, |
1917 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
1918 | ((ram_addr_t)(run_end - run_start)) | |
1919 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
1920 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
1921 | } | |
1922 | } | |
1923 | } | |
1924 | ||
3d0684b2 JQ |
1925 | /** |
1926 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1927 | * | |
1928 | * Returns zero on success | |
1929 | * | |
e0b266f0 | 1930 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
1931 | * |
1932 | * @ms: current migration state | |
89dab31b | 1933 | * @block: RAMBlock to discard |
e0b266f0 | 1934 | */ |
810cf2bb | 1935 | static int postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 1936 | { |
6b6712ef | 1937 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 1938 | unsigned long current; |
1e7cf8c3 | 1939 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 1940 | |
6b6712ef | 1941 | for (current = 0; current < end; ) { |
1e7cf8c3 | 1942 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 1943 | unsigned long zero, discard_length; |
e0b266f0 | 1944 | |
33a5cb62 WY |
1945 | if (one >= end) { |
1946 | break; | |
1947 | } | |
e0b266f0 | 1948 | |
1e7cf8c3 | 1949 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
1950 | |
1951 | if (zero >= end) { | |
1952 | discard_length = end - one; | |
e0b266f0 | 1953 | } else { |
33a5cb62 WY |
1954 | discard_length = zero - one; |
1955 | } | |
810cf2bb | 1956 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 1957 | current = one + discard_length; |
e0b266f0 DDAG |
1958 | } |
1959 | ||
1960 | return 0; | |
1961 | } | |
1962 | ||
3d0684b2 JQ |
1963 | /** |
1964 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1965 | * | |
1966 | * Returns 0 for success or negative for error | |
1967 | * | |
e0b266f0 DDAG |
1968 | * Utility for the outgoing postcopy code. |
1969 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1970 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1971 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1972 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1973 | * |
1974 | * @ms: current migration state | |
e0b266f0 DDAG |
1975 | */ |
1976 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1977 | { | |
1978 | struct RAMBlock *block; | |
1979 | int ret; | |
1980 | ||
fbd162e6 | 1981 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 1982 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 DDAG |
1983 | |
1984 | /* | |
1985 | * Postcopy sends chunks of bitmap over the wire, but it | |
1986 | * just needs indexes at this point, avoids it having | |
1987 | * target page specific code. | |
1988 | */ | |
810cf2bb WY |
1989 | ret = postcopy_send_discard_bm_ram(ms, block); |
1990 | postcopy_discard_send_finish(ms); | |
e0b266f0 DDAG |
1991 | if (ret) { |
1992 | return ret; | |
1993 | } | |
1994 | } | |
1995 | ||
1996 | return 0; | |
1997 | } | |
1998 | ||
3d0684b2 | 1999 | /** |
8324ef86 | 2000 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2001 | * |
2002 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2003 | * canonicalize the two bitmaps, that are similar, but one is | |
2004 | * inverted. | |
99e314eb | 2005 | * |
3d0684b2 JQ |
2006 | * Postcopy requires that all target pages in a hostpage are dirty or |
2007 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2008 | * |
3d0684b2 | 2009 | * @ms: current migration state |
3d0684b2 | 2010 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2011 | */ |
1e7cf8c3 | 2012 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2013 | { |
53518d94 | 2014 | RAMState *rs = ram_state; |
6b6712ef | 2015 | unsigned long *bitmap = block->bmap; |
29c59172 | 2016 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2017 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2018 | unsigned long run_start; |
2019 | ||
29c59172 DDAG |
2020 | if (block->page_size == TARGET_PAGE_SIZE) { |
2021 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2022 | return; | |
2023 | } | |
2024 | ||
1e7cf8c3 WY |
2025 | /* Find a dirty page */ |
2026 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2027 | |
6b6712ef | 2028 | while (run_start < pages) { |
99e314eb DDAG |
2029 | |
2030 | /* | |
2031 | * If the start of this run of pages is in the middle of a host | |
2032 | * page, then we need to fixup this host page. | |
2033 | */ | |
9dec3cc3 | 2034 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2035 | /* Find the end of this run */ |
1e7cf8c3 | 2036 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2037 | /* |
2038 | * If the end isn't at the start of a host page, then the | |
2039 | * run doesn't finish at the end of a host page | |
2040 | * and we need to discard. | |
2041 | */ | |
99e314eb DDAG |
2042 | } |
2043 | ||
9dec3cc3 | 2044 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2045 | unsigned long page; |
dad45ab2 WY |
2046 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2047 | host_ratio); | |
2048 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2049 | |
99e314eb DDAG |
2050 | /* Clean up the bitmap */ |
2051 | for (page = fixup_start_addr; | |
2052 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2053 | /* |
2054 | * Remark them as dirty, updating the count for any pages | |
2055 | * that weren't previously dirty. | |
2056 | */ | |
0d8ec885 | 2057 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2058 | } |
2059 | } | |
2060 | ||
1e7cf8c3 WY |
2061 | /* Find the next dirty page for the next iteration */ |
2062 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2063 | } |
2064 | } | |
2065 | ||
3d0684b2 | 2066 | /** |
89dab31b | 2067 | * postcopy_chunk_hostpages: discard any partially sent host page |
3d0684b2 | 2068 | * |
99e314eb DDAG |
2069 | * Utility for the outgoing postcopy code. |
2070 | * | |
2071 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
2072 | * dirty host-page size chunks as all dirty. In this case the host-page |
2073 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 2074 | * |
3d0684b2 JQ |
2075 | * Returns zero on success |
2076 | * | |
2077 | * @ms: current migration state | |
6b6712ef | 2078 | * @block: block we want to work with |
99e314eb | 2079 | */ |
6b6712ef | 2080 | static int postcopy_chunk_hostpages(MigrationState *ms, RAMBlock *block) |
99e314eb | 2081 | { |
810cf2bb | 2082 | postcopy_discard_send_init(ms, block->idstr); |
99e314eb | 2083 | |
6b6712ef | 2084 | /* |
1e7cf8c3 | 2085 | * Ensure that all partially dirty host pages are made fully dirty. |
6b6712ef | 2086 | */ |
1e7cf8c3 | 2087 | postcopy_chunk_hostpages_pass(ms, block); |
99e314eb | 2088 | |
810cf2bb | 2089 | postcopy_discard_send_finish(ms); |
99e314eb DDAG |
2090 | return 0; |
2091 | } | |
2092 | ||
3d0684b2 JQ |
2093 | /** |
2094 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2095 | * | |
2096 | * Returns zero on success | |
2097 | * | |
e0b266f0 DDAG |
2098 | * Transmit the set of pages to be discarded after precopy to the target |
2099 | * these are pages that: | |
2100 | * a) Have been previously transmitted but are now dirty again | |
2101 | * b) Pages that have never been transmitted, this ensures that | |
2102 | * any pages on the destination that have been mapped by background | |
2103 | * tasks get discarded (transparent huge pages is the specific concern) | |
2104 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2105 | * |
2106 | * @ms: current migration state | |
e0b266f0 DDAG |
2107 | */ |
2108 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
2109 | { | |
53518d94 | 2110 | RAMState *rs = ram_state; |
6b6712ef | 2111 | RAMBlock *block; |
e0b266f0 | 2112 | int ret; |
e0b266f0 | 2113 | |
89ac5a1d | 2114 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2115 | |
2116 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 2117 | migration_bitmap_sync(rs); |
e0b266f0 | 2118 | |
6b6712ef JQ |
2119 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
2120 | rs->last_seen_block = NULL; | |
2121 | rs->last_sent_block = NULL; | |
2122 | rs->last_page = 0; | |
e0b266f0 | 2123 | |
fbd162e6 | 2124 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2125 | /* Deal with TPS != HPS and huge pages */ |
2126 | ret = postcopy_chunk_hostpages(ms, block); | |
2127 | if (ret) { | |
6b6712ef JQ |
2128 | return ret; |
2129 | } | |
e0b266f0 | 2130 | |
e0b266f0 | 2131 | #ifdef DEBUG_POSTCOPY |
1e7cf8c3 WY |
2132 | ram_debug_dump_bitmap(block->bmap, true, |
2133 | block->used_length >> TARGET_PAGE_BITS); | |
e0b266f0 | 2134 | #endif |
6b6712ef JQ |
2135 | } |
2136 | trace_ram_postcopy_send_discard_bitmap(); | |
e0b266f0 DDAG |
2137 | |
2138 | ret = postcopy_each_ram_send_discard(ms); | |
e0b266f0 DDAG |
2139 | |
2140 | return ret; | |
2141 | } | |
2142 | ||
3d0684b2 JQ |
2143 | /** |
2144 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2145 | * |
3d0684b2 | 2146 | * Returns zero on success |
e0b266f0 | 2147 | * |
36449157 JQ |
2148 | * @rbname: name of the RAMBlock of the request. NULL means the |
2149 | * same that last one. | |
3d0684b2 JQ |
2150 | * @start: RAMBlock starting page |
2151 | * @length: RAMBlock size | |
e0b266f0 | 2152 | */ |
aaa2064c | 2153 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 2154 | { |
36449157 | 2155 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2156 | |
89ac5a1d | 2157 | RCU_READ_LOCK_GUARD(); |
36449157 | 2158 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2159 | |
2160 | if (!rb) { | |
36449157 | 2161 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 2162 | return -1; |
e0b266f0 DDAG |
2163 | } |
2164 | ||
814bb08f PX |
2165 | /* |
2166 | * On source VM, we don't need to update the received bitmap since | |
2167 | * we don't even have one. | |
2168 | */ | |
2169 | if (rb->receivedmap) { | |
2170 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
2171 | length >> qemu_target_page_bits()); | |
2172 | } | |
2173 | ||
03acb4e9 | 2174 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2175 | } |
2176 | ||
84593a08 PX |
2177 | /* |
2178 | * For every allocation, we will try not to crash the VM if the | |
2179 | * allocation failed. | |
2180 | */ | |
2181 | static int xbzrle_init(void) | |
2182 | { | |
2183 | Error *local_err = NULL; | |
2184 | ||
2185 | if (!migrate_use_xbzrle()) { | |
2186 | return 0; | |
2187 | } | |
2188 | ||
2189 | XBZRLE_cache_lock(); | |
2190 | ||
2191 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
2192 | if (!XBZRLE.zero_target_page) { | |
2193 | error_report("%s: Error allocating zero page", __func__); | |
2194 | goto err_out; | |
2195 | } | |
2196 | ||
2197 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
2198 | TARGET_PAGE_SIZE, &local_err); | |
2199 | if (!XBZRLE.cache) { | |
2200 | error_report_err(local_err); | |
2201 | goto free_zero_page; | |
2202 | } | |
2203 | ||
2204 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2205 | if (!XBZRLE.encoded_buf) { | |
2206 | error_report("%s: Error allocating encoded_buf", __func__); | |
2207 | goto free_cache; | |
2208 | } | |
2209 | ||
2210 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2211 | if (!XBZRLE.current_buf) { | |
2212 | error_report("%s: Error allocating current_buf", __func__); | |
2213 | goto free_encoded_buf; | |
2214 | } | |
2215 | ||
2216 | /* We are all good */ | |
2217 | XBZRLE_cache_unlock(); | |
2218 | return 0; | |
2219 | ||
2220 | free_encoded_buf: | |
2221 | g_free(XBZRLE.encoded_buf); | |
2222 | XBZRLE.encoded_buf = NULL; | |
2223 | free_cache: | |
2224 | cache_fini(XBZRLE.cache); | |
2225 | XBZRLE.cache = NULL; | |
2226 | free_zero_page: | |
2227 | g_free(XBZRLE.zero_target_page); | |
2228 | XBZRLE.zero_target_page = NULL; | |
2229 | err_out: | |
2230 | XBZRLE_cache_unlock(); | |
2231 | return -ENOMEM; | |
2232 | } | |
2233 | ||
53518d94 | 2234 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2235 | { |
7d00ee6a PX |
2236 | *rsp = g_try_new0(RAMState, 1); |
2237 | ||
2238 | if (!*rsp) { | |
2239 | error_report("%s: Init ramstate fail", __func__); | |
2240 | return -1; | |
2241 | } | |
53518d94 JQ |
2242 | |
2243 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
2244 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
2245 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
56e93d26 | 2246 | |
7d00ee6a | 2247 | /* |
40c4d4a8 IR |
2248 | * Count the total number of pages used by ram blocks not including any |
2249 | * gaps due to alignment or unplugs. | |
03158519 | 2250 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 2251 | */ |
40c4d4a8 | 2252 | (*rsp)->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
7d00ee6a PX |
2253 | ram_state_reset(*rsp); |
2254 | ||
2255 | return 0; | |
2256 | } | |
2257 | ||
d6eff5d7 | 2258 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 2259 | { |
002cad6b | 2260 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
2261 | RAMBlock *block; |
2262 | unsigned long pages; | |
002cad6b | 2263 | uint8_t shift; |
56e93d26 | 2264 | |
0827b9e9 AA |
2265 | /* Skip setting bitmap if there is no RAM */ |
2266 | if (ram_bytes_total()) { | |
002cad6b PX |
2267 | shift = ms->clear_bitmap_shift; |
2268 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
2269 | error_report("clear_bitmap_shift (%u) too big, using " | |
2270 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
2271 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
2272 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
2273 | error_report("clear_bitmap_shift (%u) too small, using " | |
2274 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
2275 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
2276 | } | |
2277 | ||
fbd162e6 | 2278 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 2279 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
2280 | /* |
2281 | * The initial dirty bitmap for migration must be set with all | |
2282 | * ones to make sure we'll migrate every guest RAM page to | |
2283 | * destination. | |
40c4d4a8 IR |
2284 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
2285 | * new migration after a failed migration, ram_list. | |
2286 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
2287 | * guest memory. | |
03158519 | 2288 | */ |
6b6712ef | 2289 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 2290 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
2291 | block->clear_bmap_shift = shift; |
2292 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 2293 | } |
f3f491fc | 2294 | } |
d6eff5d7 PX |
2295 | } |
2296 | ||
2297 | static void ram_init_bitmaps(RAMState *rs) | |
2298 | { | |
2299 | /* For memory_global_dirty_log_start below. */ | |
2300 | qemu_mutex_lock_iothread(); | |
2301 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 2302 | |
89ac5a1d DDAG |
2303 | WITH_RCU_READ_LOCK_GUARD() { |
2304 | ram_list_init_bitmaps(); | |
2305 | memory_global_dirty_log_start(); | |
2306 | migration_bitmap_sync_precopy(rs); | |
2307 | } | |
56e93d26 | 2308 | qemu_mutex_unlock_ramlist(); |
49877834 | 2309 | qemu_mutex_unlock_iothread(); |
d6eff5d7 PX |
2310 | } |
2311 | ||
2312 | static int ram_init_all(RAMState **rsp) | |
2313 | { | |
2314 | if (ram_state_init(rsp)) { | |
2315 | return -1; | |
2316 | } | |
2317 | ||
2318 | if (xbzrle_init()) { | |
2319 | ram_state_cleanup(rsp); | |
2320 | return -1; | |
2321 | } | |
2322 | ||
2323 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
2324 | |
2325 | return 0; | |
2326 | } | |
2327 | ||
08614f34 PX |
2328 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
2329 | { | |
2330 | RAMBlock *block; | |
2331 | uint64_t pages = 0; | |
2332 | ||
2333 | /* | |
2334 | * Postcopy is not using xbzrle/compression, so no need for that. | |
2335 | * Also, since source are already halted, we don't need to care | |
2336 | * about dirty page logging as well. | |
2337 | */ | |
2338 | ||
fbd162e6 | 2339 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
2340 | pages += bitmap_count_one(block->bmap, |
2341 | block->used_length >> TARGET_PAGE_BITS); | |
2342 | } | |
2343 | ||
2344 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
2345 | rs->migration_dirty_pages = pages; | |
2346 | ||
2347 | rs->last_seen_block = NULL; | |
2348 | rs->last_sent_block = NULL; | |
2349 | rs->last_page = 0; | |
2350 | rs->last_version = ram_list.version; | |
2351 | /* | |
2352 | * Disable the bulk stage, otherwise we'll resend the whole RAM no | |
2353 | * matter what we have sent. | |
2354 | */ | |
2355 | rs->ram_bulk_stage = false; | |
2356 | ||
2357 | /* Update RAMState cache of output QEMUFile */ | |
2358 | rs->f = out; | |
2359 | ||
2360 | trace_ram_state_resume_prepare(pages); | |
2361 | } | |
2362 | ||
6bcb05fc WW |
2363 | /* |
2364 | * This function clears bits of the free pages reported by the caller from the | |
2365 | * migration dirty bitmap. @addr is the host address corresponding to the | |
2366 | * start of the continuous guest free pages, and @len is the total bytes of | |
2367 | * those pages. | |
2368 | */ | |
2369 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
2370 | { | |
2371 | RAMBlock *block; | |
2372 | ram_addr_t offset; | |
2373 | size_t used_len, start, npages; | |
2374 | MigrationState *s = migrate_get_current(); | |
2375 | ||
2376 | /* This function is currently expected to be used during live migration */ | |
2377 | if (!migration_is_setup_or_active(s->state)) { | |
2378 | return; | |
2379 | } | |
2380 | ||
2381 | for (; len > 0; len -= used_len, addr += used_len) { | |
2382 | block = qemu_ram_block_from_host(addr, false, &offset); | |
2383 | if (unlikely(!block || offset >= block->used_length)) { | |
2384 | /* | |
2385 | * The implementation might not support RAMBlock resize during | |
2386 | * live migration, but it could happen in theory with future | |
2387 | * updates. So we add a check here to capture that case. | |
2388 | */ | |
2389 | error_report_once("%s unexpected error", __func__); | |
2390 | return; | |
2391 | } | |
2392 | ||
2393 | if (len <= block->used_length - offset) { | |
2394 | used_len = len; | |
2395 | } else { | |
2396 | used_len = block->used_length - offset; | |
2397 | } | |
2398 | ||
2399 | start = offset >> TARGET_PAGE_BITS; | |
2400 | npages = used_len >> TARGET_PAGE_BITS; | |
2401 | ||
2402 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
2403 | ram_state->migration_dirty_pages -= | |
2404 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
2405 | bitmap_clear(block->bmap, start, npages); | |
2406 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
2407 | } | |
2408 | } | |
2409 | ||
3d0684b2 JQ |
2410 | /* |
2411 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2412 | * long-running RCU critical section. When rcu-reclaims in the code |
2413 | * start to become numerous it will be necessary to reduce the | |
2414 | * granularity of these critical sections. | |
2415 | */ | |
2416 | ||
3d0684b2 JQ |
2417 | /** |
2418 | * ram_save_setup: Setup RAM for migration | |
2419 | * | |
2420 | * Returns zero to indicate success and negative for error | |
2421 | * | |
2422 | * @f: QEMUFile where to send the data | |
2423 | * @opaque: RAMState pointer | |
2424 | */ | |
a91246c9 HZ |
2425 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2426 | { | |
53518d94 | 2427 | RAMState **rsp = opaque; |
a91246c9 HZ |
2428 | RAMBlock *block; |
2429 | ||
dcaf446e XG |
2430 | if (compress_threads_save_setup()) { |
2431 | return -1; | |
2432 | } | |
2433 | ||
a91246c9 HZ |
2434 | /* migration has already setup the bitmap, reuse it. */ |
2435 | if (!migration_in_colo_state()) { | |
7d00ee6a | 2436 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 2437 | compress_threads_save_cleanup(); |
a91246c9 | 2438 | return -1; |
53518d94 | 2439 | } |
a91246c9 | 2440 | } |
53518d94 | 2441 | (*rsp)->f = f; |
a91246c9 | 2442 | |
0e6ebd48 DDAG |
2443 | WITH_RCU_READ_LOCK_GUARD() { |
2444 | qemu_put_be64(f, ram_bytes_total_common(true) | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 2445 | |
0e6ebd48 DDAG |
2446 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
2447 | qemu_put_byte(f, strlen(block->idstr)); | |
2448 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2449 | qemu_put_be64(f, block->used_length); | |
2450 | if (migrate_postcopy_ram() && block->page_size != | |
2451 | qemu_host_page_size) { | |
2452 | qemu_put_be64(f, block->page_size); | |
2453 | } | |
2454 | if (migrate_ignore_shared()) { | |
2455 | qemu_put_be64(f, block->mr->addr); | |
2456 | } | |
fbd162e6 | 2457 | } |
56e93d26 JQ |
2458 | } |
2459 | ||
56e93d26 JQ |
2460 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
2461 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2462 | ||
99f2c6fb | 2463 | multifd_send_sync_main(f); |
56e93d26 | 2464 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 2465 | qemu_fflush(f); |
56e93d26 JQ |
2466 | |
2467 | return 0; | |
2468 | } | |
2469 | ||
3d0684b2 JQ |
2470 | /** |
2471 | * ram_save_iterate: iterative stage for migration | |
2472 | * | |
2473 | * Returns zero to indicate success and negative for error | |
2474 | * | |
2475 | * @f: QEMUFile where to send the data | |
2476 | * @opaque: RAMState pointer | |
2477 | */ | |
56e93d26 JQ |
2478 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2479 | { | |
53518d94 JQ |
2480 | RAMState **temp = opaque; |
2481 | RAMState *rs = *temp; | |
3d4095b2 | 2482 | int ret = 0; |
56e93d26 JQ |
2483 | int i; |
2484 | int64_t t0; | |
5c90308f | 2485 | int done = 0; |
56e93d26 | 2486 | |
b2557345 PL |
2487 | if (blk_mig_bulk_active()) { |
2488 | /* Avoid transferring ram during bulk phase of block migration as | |
2489 | * the bulk phase will usually take a long time and transferring | |
2490 | * ram updates during that time is pointless. */ | |
2491 | goto out; | |
2492 | } | |
2493 | ||
89ac5a1d DDAG |
2494 | WITH_RCU_READ_LOCK_GUARD() { |
2495 | if (ram_list.version != rs->last_version) { | |
2496 | ram_state_reset(rs); | |
2497 | } | |
56e93d26 | 2498 | |
89ac5a1d DDAG |
2499 | /* Read version before ram_list.blocks */ |
2500 | smp_rmb(); | |
56e93d26 | 2501 | |
89ac5a1d | 2502 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 2503 | |
89ac5a1d DDAG |
2504 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
2505 | i = 0; | |
2506 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
2507 | !QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
2508 | int pages; | |
e03a34f8 | 2509 | |
89ac5a1d DDAG |
2510 | if (qemu_file_get_error(f)) { |
2511 | break; | |
2512 | } | |
e8f3735f | 2513 | |
89ac5a1d DDAG |
2514 | pages = ram_find_and_save_block(rs, false); |
2515 | /* no more pages to sent */ | |
2516 | if (pages == 0) { | |
2517 | done = 1; | |
2518 | break; | |
2519 | } | |
e8f3735f | 2520 | |
89ac5a1d DDAG |
2521 | if (pages < 0) { |
2522 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
2523 | break; |
2524 | } | |
89ac5a1d DDAG |
2525 | |
2526 | rs->target_page_count += pages; | |
2527 | ||
644acf99 WY |
2528 | /* |
2529 | * During postcopy, it is necessary to make sure one whole host | |
2530 | * page is sent in one chunk. | |
2531 | */ | |
2532 | if (migrate_postcopy_ram()) { | |
2533 | flush_compressed_data(rs); | |
2534 | } | |
2535 | ||
89ac5a1d DDAG |
2536 | /* |
2537 | * we want to check in the 1st loop, just in case it was the 1st | |
2538 | * time and we had to sync the dirty bitmap. | |
2539 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
2540 | * some iterations | |
2541 | */ | |
2542 | if ((i & 63) == 0) { | |
2543 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
2544 | 1000000; | |
2545 | if (t1 > MAX_WAIT) { | |
2546 | trace_ram_save_iterate_big_wait(t1, i); | |
2547 | break; | |
2548 | } | |
2549 | } | |
2550 | i++; | |
56e93d26 | 2551 | } |
56e93d26 | 2552 | } |
56e93d26 JQ |
2553 | |
2554 | /* | |
2555 | * Must occur before EOS (or any QEMUFile operation) | |
2556 | * because of RDMA protocol. | |
2557 | */ | |
2558 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2559 | ||
b2557345 | 2560 | out: |
b69a0227 JQ |
2561 | if (ret >= 0 |
2562 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
99f2c6fb | 2563 | multifd_send_sync_main(rs->f); |
3d4095b2 JQ |
2564 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2565 | qemu_fflush(f); | |
2566 | ram_counters.transferred += 8; | |
56e93d26 | 2567 | |
3d4095b2 JQ |
2568 | ret = qemu_file_get_error(f); |
2569 | } | |
56e93d26 JQ |
2570 | if (ret < 0) { |
2571 | return ret; | |
2572 | } | |
2573 | ||
5c90308f | 2574 | return done; |
56e93d26 JQ |
2575 | } |
2576 | ||
3d0684b2 JQ |
2577 | /** |
2578 | * ram_save_complete: function called to send the remaining amount of ram | |
2579 | * | |
e8f3735f | 2580 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
2581 | * |
2582 | * Called with iothread lock | |
2583 | * | |
2584 | * @f: QEMUFile where to send the data | |
2585 | * @opaque: RAMState pointer | |
2586 | */ | |
56e93d26 JQ |
2587 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2588 | { | |
53518d94 JQ |
2589 | RAMState **temp = opaque; |
2590 | RAMState *rs = *temp; | |
e8f3735f | 2591 | int ret = 0; |
6f37bb8b | 2592 | |
89ac5a1d DDAG |
2593 | WITH_RCU_READ_LOCK_GUARD() { |
2594 | if (!migration_in_postcopy()) { | |
2595 | migration_bitmap_sync_precopy(rs); | |
2596 | } | |
56e93d26 | 2597 | |
89ac5a1d | 2598 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 2599 | |
89ac5a1d | 2600 | /* try transferring iterative blocks of memory */ |
56e93d26 | 2601 | |
89ac5a1d DDAG |
2602 | /* flush all remaining blocks regardless of rate limiting */ |
2603 | while (true) { | |
2604 | int pages; | |
56e93d26 | 2605 | |
89ac5a1d DDAG |
2606 | pages = ram_find_and_save_block(rs, !migration_in_colo_state()); |
2607 | /* no more blocks to sent */ | |
2608 | if (pages == 0) { | |
2609 | break; | |
2610 | } | |
2611 | if (pages < 0) { | |
2612 | ret = pages; | |
2613 | break; | |
2614 | } | |
e8f3735f | 2615 | } |
56e93d26 | 2616 | |
89ac5a1d DDAG |
2617 | flush_compressed_data(rs); |
2618 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); | |
2619 | } | |
d09a6fde | 2620 | |
3d4095b2 | 2621 | if (ret >= 0) { |
99f2c6fb | 2622 | multifd_send_sync_main(rs->f); |
3d4095b2 JQ |
2623 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2624 | qemu_fflush(f); | |
2625 | } | |
56e93d26 | 2626 | |
e8f3735f | 2627 | return ret; |
56e93d26 JQ |
2628 | } |
2629 | ||
c31b098f | 2630 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
47995026 VSO |
2631 | uint64_t *res_precopy_only, |
2632 | uint64_t *res_compatible, | |
2633 | uint64_t *res_postcopy_only) | |
56e93d26 | 2634 | { |
53518d94 JQ |
2635 | RAMState **temp = opaque; |
2636 | RAMState *rs = *temp; | |
56e93d26 JQ |
2637 | uint64_t remaining_size; |
2638 | ||
9edabd4d | 2639 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2640 | |
5727309d | 2641 | if (!migration_in_postcopy() && |
663e6c1d | 2642 | remaining_size < max_size) { |
56e93d26 | 2643 | qemu_mutex_lock_iothread(); |
89ac5a1d DDAG |
2644 | WITH_RCU_READ_LOCK_GUARD() { |
2645 | migration_bitmap_sync_precopy(rs); | |
2646 | } | |
56e93d26 | 2647 | qemu_mutex_unlock_iothread(); |
9edabd4d | 2648 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2649 | } |
c31b098f | 2650 | |
86e1167e VSO |
2651 | if (migrate_postcopy_ram()) { |
2652 | /* We can do postcopy, and all the data is postcopiable */ | |
47995026 | 2653 | *res_compatible += remaining_size; |
86e1167e | 2654 | } else { |
47995026 | 2655 | *res_precopy_only += remaining_size; |
86e1167e | 2656 | } |
56e93d26 JQ |
2657 | } |
2658 | ||
2659 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2660 | { | |
2661 | unsigned int xh_len; | |
2662 | int xh_flags; | |
063e760a | 2663 | uint8_t *loaded_data; |
56e93d26 | 2664 | |
56e93d26 JQ |
2665 | /* extract RLE header */ |
2666 | xh_flags = qemu_get_byte(f); | |
2667 | xh_len = qemu_get_be16(f); | |
2668 | ||
2669 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2670 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2671 | return -1; | |
2672 | } | |
2673 | ||
2674 | if (xh_len > TARGET_PAGE_SIZE) { | |
2675 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2676 | return -1; | |
2677 | } | |
f265e0e4 | 2678 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 2679 | /* load data and decode */ |
f265e0e4 | 2680 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 2681 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2682 | |
2683 | /* decode RLE */ | |
063e760a | 2684 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2685 | TARGET_PAGE_SIZE) == -1) { |
2686 | error_report("Failed to load XBZRLE page - decode error!"); | |
2687 | return -1; | |
2688 | } | |
2689 | ||
2690 | return 0; | |
2691 | } | |
2692 | ||
3d0684b2 JQ |
2693 | /** |
2694 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2695 | * | |
2696 | * Must be called from within a rcu critical section. | |
2697 | * | |
56e93d26 | 2698 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2699 | * |
3d0684b2 JQ |
2700 | * @f: QEMUFile where to read the data from |
2701 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2702 | */ |
3d0684b2 | 2703 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2704 | { |
2705 | static RAMBlock *block = NULL; | |
2706 | char id[256]; | |
2707 | uint8_t len; | |
2708 | ||
2709 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2710 | if (!block) { |
56e93d26 JQ |
2711 | error_report("Ack, bad migration stream!"); |
2712 | return NULL; | |
2713 | } | |
4c4bad48 | 2714 | return block; |
56e93d26 JQ |
2715 | } |
2716 | ||
2717 | len = qemu_get_byte(f); | |
2718 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2719 | id[len] = 0; | |
2720 | ||
e3dd7493 | 2721 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2722 | if (!block) { |
2723 | error_report("Can't find block %s", id); | |
2724 | return NULL; | |
56e93d26 JQ |
2725 | } |
2726 | ||
fbd162e6 | 2727 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
2728 | error_report("block %s should not be migrated !", id); |
2729 | return NULL; | |
2730 | } | |
2731 | ||
4c4bad48 HZ |
2732 | return block; |
2733 | } | |
2734 | ||
2735 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2736 | ram_addr_t offset) | |
2737 | { | |
2738 | if (!offset_in_ramblock(block, offset)) { | |
2739 | return NULL; | |
2740 | } | |
2741 | ||
2742 | return block->host + offset; | |
56e93d26 JQ |
2743 | } |
2744 | ||
13af18f2 | 2745 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 2746 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
2747 | { |
2748 | if (!offset_in_ramblock(block, offset)) { | |
2749 | return NULL; | |
2750 | } | |
2751 | if (!block->colo_cache) { | |
2752 | error_report("%s: colo_cache is NULL in block :%s", | |
2753 | __func__, block->idstr); | |
2754 | return NULL; | |
2755 | } | |
7d9acafa ZC |
2756 | |
2757 | /* | |
2758 | * During colo checkpoint, we need bitmap of these migrated pages. | |
2759 | * It help us to decide which pages in ram cache should be flushed | |
2760 | * into VM's RAM later. | |
2761 | */ | |
8af66371 HZ |
2762 | if (record_bitmap && |
2763 | !test_and_set_bit(offset >> TARGET_PAGE_BITS, block->bmap)) { | |
7d9acafa ZC |
2764 | ram_state->migration_dirty_pages++; |
2765 | } | |
13af18f2 ZC |
2766 | return block->colo_cache + offset; |
2767 | } | |
2768 | ||
3d0684b2 JQ |
2769 | /** |
2770 | * ram_handle_compressed: handle the zero page case | |
2771 | * | |
56e93d26 JQ |
2772 | * If a page (or a whole RDMA chunk) has been |
2773 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2774 | * |
2775 | * @host: host address for the zero page | |
2776 | * @ch: what the page is filled from. We only support zero | |
2777 | * @size: size of the zero page | |
56e93d26 JQ |
2778 | */ |
2779 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2780 | { | |
2781 | if (ch != 0 || !is_zero_range(host, size)) { | |
2782 | memset(host, ch, size); | |
2783 | } | |
2784 | } | |
2785 | ||
797ca154 XG |
2786 | /* return the size after decompression, or negative value on error */ |
2787 | static int | |
2788 | qemu_uncompress_data(z_stream *stream, uint8_t *dest, size_t dest_len, | |
2789 | const uint8_t *source, size_t source_len) | |
2790 | { | |
2791 | int err; | |
2792 | ||
2793 | err = inflateReset(stream); | |
2794 | if (err != Z_OK) { | |
2795 | return -1; | |
2796 | } | |
2797 | ||
2798 | stream->avail_in = source_len; | |
2799 | stream->next_in = (uint8_t *)source; | |
2800 | stream->avail_out = dest_len; | |
2801 | stream->next_out = dest; | |
2802 | ||
2803 | err = inflate(stream, Z_NO_FLUSH); | |
2804 | if (err != Z_STREAM_END) { | |
2805 | return -1; | |
2806 | } | |
2807 | ||
2808 | return stream->total_out; | |
2809 | } | |
2810 | ||
56e93d26 JQ |
2811 | static void *do_data_decompress(void *opaque) |
2812 | { | |
2813 | DecompressParam *param = opaque; | |
2814 | unsigned long pagesize; | |
33d151f4 | 2815 | uint8_t *des; |
34ab9e97 | 2816 | int len, ret; |
56e93d26 | 2817 | |
33d151f4 | 2818 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2819 | while (!param->quit) { |
33d151f4 LL |
2820 | if (param->des) { |
2821 | des = param->des; | |
2822 | len = param->len; | |
2823 | param->des = 0; | |
2824 | qemu_mutex_unlock(¶m->mutex); | |
2825 | ||
56e93d26 | 2826 | pagesize = TARGET_PAGE_SIZE; |
34ab9e97 XG |
2827 | |
2828 | ret = qemu_uncompress_data(¶m->stream, des, pagesize, | |
2829 | param->compbuf, len); | |
f548222c | 2830 | if (ret < 0 && migrate_get_current()->decompress_error_check) { |
34ab9e97 XG |
2831 | error_report("decompress data failed"); |
2832 | qemu_file_set_error(decomp_file, ret); | |
2833 | } | |
73a8912b | 2834 | |
33d151f4 LL |
2835 | qemu_mutex_lock(&decomp_done_lock); |
2836 | param->done = true; | |
2837 | qemu_cond_signal(&decomp_done_cond); | |
2838 | qemu_mutex_unlock(&decomp_done_lock); | |
2839 | ||
2840 | qemu_mutex_lock(¶m->mutex); | |
2841 | } else { | |
2842 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2843 | } | |
56e93d26 | 2844 | } |
33d151f4 | 2845 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2846 | |
2847 | return NULL; | |
2848 | } | |
2849 | ||
34ab9e97 | 2850 | static int wait_for_decompress_done(void) |
5533b2e9 LL |
2851 | { |
2852 | int idx, thread_count; | |
2853 | ||
2854 | if (!migrate_use_compression()) { | |
34ab9e97 | 2855 | return 0; |
5533b2e9 LL |
2856 | } |
2857 | ||
2858 | thread_count = migrate_decompress_threads(); | |
2859 | qemu_mutex_lock(&decomp_done_lock); | |
2860 | for (idx = 0; idx < thread_count; idx++) { | |
2861 | while (!decomp_param[idx].done) { | |
2862 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2863 | } | |
2864 | } | |
2865 | qemu_mutex_unlock(&decomp_done_lock); | |
34ab9e97 | 2866 | return qemu_file_get_error(decomp_file); |
5533b2e9 LL |
2867 | } |
2868 | ||
f0afa331 | 2869 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
2870 | { |
2871 | int i, thread_count; | |
2872 | ||
3416ab5b JQ |
2873 | if (!migrate_use_compression()) { |
2874 | return; | |
2875 | } | |
56e93d26 JQ |
2876 | thread_count = migrate_decompress_threads(); |
2877 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
2878 | /* |
2879 | * we use it as a indicator which shows if the thread is | |
2880 | * properly init'd or not | |
2881 | */ | |
2882 | if (!decomp_param[i].compbuf) { | |
2883 | break; | |
2884 | } | |
2885 | ||
56e93d26 | 2886 | qemu_mutex_lock(&decomp_param[i].mutex); |
90e56fb4 | 2887 | decomp_param[i].quit = true; |
56e93d26 JQ |
2888 | qemu_cond_signal(&decomp_param[i].cond); |
2889 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2890 | } | |
2891 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
2892 | if (!decomp_param[i].compbuf) { |
2893 | break; | |
2894 | } | |
2895 | ||
56e93d26 JQ |
2896 | qemu_thread_join(decompress_threads + i); |
2897 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2898 | qemu_cond_destroy(&decomp_param[i].cond); | |
797ca154 | 2899 | inflateEnd(&decomp_param[i].stream); |
56e93d26 | 2900 | g_free(decomp_param[i].compbuf); |
797ca154 | 2901 | decomp_param[i].compbuf = NULL; |
56e93d26 JQ |
2902 | } |
2903 | g_free(decompress_threads); | |
2904 | g_free(decomp_param); | |
56e93d26 JQ |
2905 | decompress_threads = NULL; |
2906 | decomp_param = NULL; | |
34ab9e97 | 2907 | decomp_file = NULL; |
56e93d26 JQ |
2908 | } |
2909 | ||
34ab9e97 | 2910 | static int compress_threads_load_setup(QEMUFile *f) |
797ca154 XG |
2911 | { |
2912 | int i, thread_count; | |
2913 | ||
2914 | if (!migrate_use_compression()) { | |
2915 | return 0; | |
2916 | } | |
2917 | ||
2918 | thread_count = migrate_decompress_threads(); | |
2919 | decompress_threads = g_new0(QemuThread, thread_count); | |
2920 | decomp_param = g_new0(DecompressParam, thread_count); | |
2921 | qemu_mutex_init(&decomp_done_lock); | |
2922 | qemu_cond_init(&decomp_done_cond); | |
34ab9e97 | 2923 | decomp_file = f; |
797ca154 XG |
2924 | for (i = 0; i < thread_count; i++) { |
2925 | if (inflateInit(&decomp_param[i].stream) != Z_OK) { | |
2926 | goto exit; | |
2927 | } | |
2928 | ||
2929 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
2930 | qemu_mutex_init(&decomp_param[i].mutex); | |
2931 | qemu_cond_init(&decomp_param[i].cond); | |
2932 | decomp_param[i].done = true; | |
2933 | decomp_param[i].quit = false; | |
2934 | qemu_thread_create(decompress_threads + i, "decompress", | |
2935 | do_data_decompress, decomp_param + i, | |
2936 | QEMU_THREAD_JOINABLE); | |
2937 | } | |
2938 | return 0; | |
2939 | exit: | |
2940 | compress_threads_load_cleanup(); | |
2941 | return -1; | |
2942 | } | |
2943 | ||
c1bc6626 | 2944 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2945 | void *host, int len) |
2946 | { | |
2947 | int idx, thread_count; | |
2948 | ||
2949 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2950 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2951 | while (true) { |
2952 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2953 | if (decomp_param[idx].done) { |
33d151f4 LL |
2954 | decomp_param[idx].done = false; |
2955 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2956 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2957 | decomp_param[idx].des = host; |
2958 | decomp_param[idx].len = len; | |
33d151f4 LL |
2959 | qemu_cond_signal(&decomp_param[idx].cond); |
2960 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2961 | break; |
2962 | } | |
2963 | } | |
2964 | if (idx < thread_count) { | |
2965 | break; | |
73a8912b LL |
2966 | } else { |
2967 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2968 | } |
2969 | } | |
73a8912b | 2970 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2971 | } |
2972 | ||
13af18f2 ZC |
2973 | /* |
2974 | * colo cache: this is for secondary VM, we cache the whole | |
2975 | * memory of the secondary VM, it is need to hold the global lock | |
2976 | * to call this helper. | |
2977 | */ | |
2978 | int colo_init_ram_cache(void) | |
2979 | { | |
2980 | RAMBlock *block; | |
2981 | ||
44901b5a PB |
2982 | WITH_RCU_READ_LOCK_GUARD() { |
2983 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2984 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
2985 | NULL, | |
2986 | false); | |
2987 | if (!block->colo_cache) { | |
2988 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
2989 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
2990 | block->used_length); | |
2991 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2992 | if (block->colo_cache) { | |
2993 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
2994 | block->colo_cache = NULL; | |
2995 | } | |
89ac5a1d | 2996 | } |
44901b5a | 2997 | return -errno; |
89ac5a1d | 2998 | } |
13af18f2 | 2999 | } |
13af18f2 | 3000 | } |
44901b5a | 3001 | |
7d9acafa ZC |
3002 | /* |
3003 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3004 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3005 | * we use the same name 'ram_bitmap' as for migration. | |
3006 | */ | |
3007 | if (ram_bytes_total()) { | |
3008 | RAMBlock *block; | |
3009 | ||
fbd162e6 | 3010 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3011 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3012 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3013 | } |
3014 | } | |
7d9acafa | 3015 | |
0393031a | 3016 | ram_state_init(&ram_state); |
13af18f2 | 3017 | return 0; |
13af18f2 ZC |
3018 | } |
3019 | ||
0393031a HZ |
3020 | /* TODO: duplicated with ram_init_bitmaps */ |
3021 | void colo_incoming_start_dirty_log(void) | |
3022 | { | |
3023 | RAMBlock *block = NULL; | |
3024 | /* For memory_global_dirty_log_start below. */ | |
3025 | qemu_mutex_lock_iothread(); | |
3026 | qemu_mutex_lock_ramlist(); | |
3027 | ||
3028 | memory_global_dirty_log_sync(); | |
3029 | WITH_RCU_READ_LOCK_GUARD() { | |
3030 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3031 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3032 | /* Discard this dirty bitmap record */ | |
3033 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3034 | } | |
3035 | memory_global_dirty_log_start(); | |
3036 | } | |
3037 | ram_state->migration_dirty_pages = 0; | |
3038 | qemu_mutex_unlock_ramlist(); | |
3039 | qemu_mutex_unlock_iothread(); | |
3040 | } | |
3041 | ||
13af18f2 ZC |
3042 | /* It is need to hold the global lock to call this helper */ |
3043 | void colo_release_ram_cache(void) | |
3044 | { | |
3045 | RAMBlock *block; | |
3046 | ||
d1955d22 | 3047 | memory_global_dirty_log_stop(); |
fbd162e6 | 3048 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3049 | g_free(block->bmap); |
3050 | block->bmap = NULL; | |
3051 | } | |
3052 | ||
89ac5a1d DDAG |
3053 | WITH_RCU_READ_LOCK_GUARD() { |
3054 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3055 | if (block->colo_cache) { | |
3056 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3057 | block->colo_cache = NULL; | |
3058 | } | |
13af18f2 ZC |
3059 | } |
3060 | } | |
0393031a | 3061 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
3062 | } |
3063 | ||
f265e0e4 JQ |
3064 | /** |
3065 | * ram_load_setup: Setup RAM for migration incoming side | |
3066 | * | |
3067 | * Returns zero to indicate success and negative for error | |
3068 | * | |
3069 | * @f: QEMUFile where to receive the data | |
3070 | * @opaque: RAMState pointer | |
3071 | */ | |
3072 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
3073 | { | |
34ab9e97 | 3074 | if (compress_threads_load_setup(f)) { |
797ca154 XG |
3075 | return -1; |
3076 | } | |
3077 | ||
f265e0e4 | 3078 | xbzrle_load_setup(); |
f9494614 | 3079 | ramblock_recv_map_init(); |
13af18f2 | 3080 | |
f265e0e4 JQ |
3081 | return 0; |
3082 | } | |
3083 | ||
3084 | static int ram_load_cleanup(void *opaque) | |
3085 | { | |
f9494614 | 3086 | RAMBlock *rb; |
56eb90af | 3087 | |
fbd162e6 | 3088 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 3089 | qemu_ram_block_writeback(rb); |
56eb90af JH |
3090 | } |
3091 | ||
f265e0e4 | 3092 | xbzrle_load_cleanup(); |
f0afa331 | 3093 | compress_threads_load_cleanup(); |
f9494614 | 3094 | |
fbd162e6 | 3095 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
3096 | g_free(rb->receivedmap); |
3097 | rb->receivedmap = NULL; | |
3098 | } | |
13af18f2 | 3099 | |
f265e0e4 JQ |
3100 | return 0; |
3101 | } | |
3102 | ||
3d0684b2 JQ |
3103 | /** |
3104 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
3105 | * | |
3106 | * Returns 0 for success and negative if there was one error | |
3107 | * | |
3108 | * @mis: current migration incoming state | |
3109 | * | |
3110 | * Allocate data structures etc needed by incoming migration with | |
3111 | * postcopy-ram. postcopy-ram's similarly names | |
3112 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
3113 | */ |
3114 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
3115 | { | |
c136180c | 3116 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
3117 | } |
3118 | ||
3d0684b2 JQ |
3119 | /** |
3120 | * ram_load_postcopy: load a page in postcopy case | |
3121 | * | |
3122 | * Returns 0 for success or -errno in case of error | |
3123 | * | |
a7180877 DDAG |
3124 | * Called in postcopy mode by ram_load(). |
3125 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
3126 | * |
3127 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
3128 | */ |
3129 | static int ram_load_postcopy(QEMUFile *f) | |
3130 | { | |
3131 | int flags = 0, ret = 0; | |
3132 | bool place_needed = false; | |
1aa83678 | 3133 | bool matches_target_page_size = false; |
a7180877 DDAG |
3134 | MigrationIncomingState *mis = migration_incoming_get_current(); |
3135 | /* Temporary page that is later 'placed' */ | |
3414322a | 3136 | void *postcopy_host_page = mis->postcopy_tmp_page; |
91ba442f | 3137 | void *this_host = NULL; |
a3b6ff6d | 3138 | bool all_zero = false; |
4cbb3c63 | 3139 | int target_pages = 0; |
a7180877 DDAG |
3140 | |
3141 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
3142 | ram_addr_t addr; | |
3143 | void *host = NULL; | |
3144 | void *page_buffer = NULL; | |
3145 | void *place_source = NULL; | |
df9ff5e1 | 3146 | RAMBlock *block = NULL; |
a7180877 | 3147 | uint8_t ch; |
644acf99 | 3148 | int len; |
a7180877 DDAG |
3149 | |
3150 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
3151 | |
3152 | /* | |
3153 | * If qemu file error, we should stop here, and then "addr" | |
3154 | * may be invalid | |
3155 | */ | |
3156 | ret = qemu_file_get_error(f); | |
3157 | if (ret) { | |
3158 | break; | |
3159 | } | |
3160 | ||
a7180877 DDAG |
3161 | flags = addr & ~TARGET_PAGE_MASK; |
3162 | addr &= TARGET_PAGE_MASK; | |
3163 | ||
3164 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
3165 | place_needed = false; | |
644acf99 WY |
3166 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
3167 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
df9ff5e1 | 3168 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
3169 | |
3170 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
3171 | if (!host) { |
3172 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
3173 | ret = -EINVAL; | |
3174 | break; | |
3175 | } | |
4cbb3c63 | 3176 | target_pages++; |
1aa83678 | 3177 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 3178 | /* |
28abd200 DDAG |
3179 | * Postcopy requires that we place whole host pages atomically; |
3180 | * these may be huge pages for RAMBlocks that are backed by | |
3181 | * hugetlbfs. | |
a7180877 DDAG |
3182 | * To make it atomic, the data is read into a temporary page |
3183 | * that's moved into place later. | |
3184 | * The migration protocol uses, possibly smaller, target-pages | |
3185 | * however the source ensures it always sends all the components | |
91ba442f | 3186 | * of a host page in one chunk. |
a7180877 DDAG |
3187 | */ |
3188 | page_buffer = postcopy_host_page + | |
28abd200 | 3189 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 3190 | /* If all TP are zero then we can optimise the place */ |
e5e73b0f | 3191 | if (target_pages == 1) { |
a7180877 | 3192 | all_zero = true; |
91ba442f WY |
3193 | this_host = (void *)QEMU_ALIGN_DOWN((uintptr_t)host, |
3194 | block->page_size); | |
c53b7ddc DDAG |
3195 | } else { |
3196 | /* not the 1st TP within the HP */ | |
91ba442f WY |
3197 | if (QEMU_ALIGN_DOWN((uintptr_t)host, block->page_size) != |
3198 | (uintptr_t)this_host) { | |
3199 | error_report("Non-same host page %p/%p", | |
3200 | host, this_host); | |
c53b7ddc DDAG |
3201 | ret = -EINVAL; |
3202 | break; | |
3203 | } | |
a7180877 DDAG |
3204 | } |
3205 | ||
3206 | /* | |
3207 | * If it's the last part of a host page then we place the host | |
3208 | * page | |
3209 | */ | |
4cbb3c63 WY |
3210 | if (target_pages == (block->page_size / TARGET_PAGE_SIZE)) { |
3211 | place_needed = true; | |
3212 | target_pages = 0; | |
3213 | } | |
a7180877 DDAG |
3214 | place_source = postcopy_host_page; |
3215 | } | |
3216 | ||
3217 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 3218 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 3219 | ch = qemu_get_byte(f); |
2e36bc1b WY |
3220 | /* |
3221 | * Can skip to set page_buffer when | |
3222 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
3223 | */ | |
3224 | if (ch || !matches_target_page_size) { | |
3225 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
3226 | } | |
a7180877 DDAG |
3227 | if (ch) { |
3228 | all_zero = false; | |
3229 | } | |
3230 | break; | |
3231 | ||
3232 | case RAM_SAVE_FLAG_PAGE: | |
3233 | all_zero = false; | |
1aa83678 PX |
3234 | if (!matches_target_page_size) { |
3235 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
3236 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
3237 | } else { | |
1aa83678 PX |
3238 | /* |
3239 | * For small pages that matches target page size, we | |
3240 | * avoid the qemu_file copy. Instead we directly use | |
3241 | * the buffer of QEMUFile to place the page. Note: we | |
3242 | * cannot do any QEMUFile operation before using that | |
3243 | * buffer to make sure the buffer is valid when | |
3244 | * placing the page. | |
a7180877 DDAG |
3245 | */ |
3246 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
3247 | TARGET_PAGE_SIZE); | |
3248 | } | |
3249 | break; | |
644acf99 WY |
3250 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
3251 | all_zero = false; | |
3252 | len = qemu_get_be32(f); | |
3253 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
3254 | error_report("Invalid compressed data length: %d", len); | |
3255 | ret = -EINVAL; | |
3256 | break; | |
3257 | } | |
3258 | decompress_data_with_multi_threads(f, page_buffer, len); | |
3259 | break; | |
3260 | ||
a7180877 DDAG |
3261 | case RAM_SAVE_FLAG_EOS: |
3262 | /* normal exit */ | |
6df264ac | 3263 | multifd_recv_sync_main(); |
a7180877 DDAG |
3264 | break; |
3265 | default: | |
3266 | error_report("Unknown combination of migration flags: %#x" | |
3267 | " (postcopy mode)", flags); | |
3268 | ret = -EINVAL; | |
7a9ddfbf PX |
3269 | break; |
3270 | } | |
3271 | ||
644acf99 WY |
3272 | /* Got the whole host page, wait for decompress before placing. */ |
3273 | if (place_needed) { | |
3274 | ret |= wait_for_decompress_done(); | |
3275 | } | |
3276 | ||
7a9ddfbf PX |
3277 | /* Detect for any possible file errors */ |
3278 | if (!ret && qemu_file_get_error(f)) { | |
3279 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
3280 | } |
3281 | ||
7a9ddfbf | 3282 | if (!ret && place_needed) { |
a7180877 | 3283 | /* This gets called at the last target page in the host page */ |
91ba442f WY |
3284 | void *place_dest = (void *)QEMU_ALIGN_DOWN((uintptr_t)host, |
3285 | block->page_size); | |
df9ff5e1 | 3286 | |
a7180877 | 3287 | if (all_zero) { |
df9ff5e1 | 3288 | ret = postcopy_place_page_zero(mis, place_dest, |
8be4620b | 3289 | block); |
a7180877 | 3290 | } else { |
df9ff5e1 | 3291 | ret = postcopy_place_page(mis, place_dest, |
8be4620b | 3292 | place_source, block); |
a7180877 DDAG |
3293 | } |
3294 | } | |
a7180877 DDAG |
3295 | } |
3296 | ||
3297 | return ret; | |
3298 | } | |
3299 | ||
acab30b8 DHB |
3300 | static bool postcopy_is_advised(void) |
3301 | { | |
3302 | PostcopyState ps = postcopy_state_get(); | |
3303 | return ps >= POSTCOPY_INCOMING_ADVISE && ps < POSTCOPY_INCOMING_END; | |
3304 | } | |
3305 | ||
3306 | static bool postcopy_is_running(void) | |
3307 | { | |
3308 | PostcopyState ps = postcopy_state_get(); | |
3309 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
3310 | } | |
3311 | ||
e6f4aa18 ZC |
3312 | /* |
3313 | * Flush content of RAM cache into SVM's memory. | |
3314 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
3315 | */ | |
3316 | static void colo_flush_ram_cache(void) | |
3317 | { | |
3318 | RAMBlock *block = NULL; | |
3319 | void *dst_host; | |
3320 | void *src_host; | |
3321 | unsigned long offset = 0; | |
3322 | ||
d1955d22 | 3323 | memory_global_dirty_log_sync(); |
89ac5a1d DDAG |
3324 | WITH_RCU_READ_LOCK_GUARD() { |
3325 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3326 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3327 | } | |
d1955d22 | 3328 | } |
d1955d22 | 3329 | |
e6f4aa18 | 3330 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
3331 | WITH_RCU_READ_LOCK_GUARD() { |
3332 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 3333 | |
89ac5a1d DDAG |
3334 | while (block) { |
3335 | offset = migration_bitmap_find_dirty(ram_state, block, offset); | |
e6f4aa18 | 3336 | |
8bba004c AR |
3337 | if (((ram_addr_t)offset) << TARGET_PAGE_BITS |
3338 | >= block->used_length) { | |
89ac5a1d DDAG |
3339 | offset = 0; |
3340 | block = QLIST_NEXT_RCU(block, next); | |
3341 | } else { | |
3342 | migration_bitmap_clear_dirty(ram_state, block, offset); | |
8bba004c AR |
3343 | dst_host = block->host |
3344 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
3345 | src_host = block->colo_cache | |
3346 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
89ac5a1d DDAG |
3347 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE); |
3348 | } | |
e6f4aa18 ZC |
3349 | } |
3350 | } | |
e6f4aa18 ZC |
3351 | trace_colo_flush_ram_cache_end(); |
3352 | } | |
3353 | ||
10da4a36 WY |
3354 | /** |
3355 | * ram_load_precopy: load pages in precopy case | |
3356 | * | |
3357 | * Returns 0 for success or -errno in case of error | |
3358 | * | |
3359 | * Called in precopy mode by ram_load(). | |
3360 | * rcu_read_lock is taken prior to this being called. | |
3361 | * | |
3362 | * @f: QEMUFile where to send the data | |
3363 | */ | |
3364 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 3365 | { |
e65cec5e | 3366 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 3367 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
acab30b8 | 3368 | bool postcopy_advised = postcopy_is_advised(); |
edc60127 JQ |
3369 | if (!migrate_use_compression()) { |
3370 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
3371 | } | |
a7180877 | 3372 | |
10da4a36 | 3373 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 3374 | ram_addr_t addr, total_ram_bytes; |
0393031a | 3375 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
3376 | uint8_t ch; |
3377 | ||
e65cec5e YK |
3378 | /* |
3379 | * Yield periodically to let main loop run, but an iteration of | |
3380 | * the main loop is expensive, so do it each some iterations | |
3381 | */ | |
3382 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
3383 | aio_co_schedule(qemu_get_current_aio_context(), | |
3384 | qemu_coroutine_self()); | |
3385 | qemu_coroutine_yield(); | |
3386 | } | |
3387 | i++; | |
3388 | ||
56e93d26 JQ |
3389 | addr = qemu_get_be64(f); |
3390 | flags = addr & ~TARGET_PAGE_MASK; | |
3391 | addr &= TARGET_PAGE_MASK; | |
3392 | ||
edc60127 JQ |
3393 | if (flags & invalid_flags) { |
3394 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
3395 | error_report("Received an unexpected compressed page"); | |
3396 | } | |
3397 | ||
3398 | ret = -EINVAL; | |
3399 | break; | |
3400 | } | |
3401 | ||
bb890ed5 | 3402 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 3403 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
4c4bad48 HZ |
3404 | RAMBlock *block = ram_block_from_stream(f, flags); |
3405 | ||
0393031a | 3406 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 3407 | /* |
0393031a HZ |
3408 | * After going into COLO stage, we should not load the page |
3409 | * into SVM's memory directly, we put them into colo_cache firstly. | |
3410 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
3411 | * Previously, we copied all these memory in preparing stage of COLO | |
3412 | * while we need to stop VM, which is a time-consuming process. | |
3413 | * Here we optimize it by a trick, back-up every page while in | |
3414 | * migration process while COLO is enabled, though it affects the | |
3415 | * speed of the migration, but it obviously reduce the downtime of | |
3416 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 3417 | */ |
0393031a HZ |
3418 | if (migration_incoming_colo_enabled()) { |
3419 | if (migration_incoming_in_colo_state()) { | |
3420 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 3421 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
3422 | } else { |
3423 | /* | |
3424 | * In migration stage but before COLO stage, | |
3425 | * Put all pages into both cache and SVM's memory. | |
3426 | */ | |
8af66371 | 3427 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 3428 | } |
13af18f2 | 3429 | } |
a776aa15 DDAG |
3430 | if (!host) { |
3431 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
3432 | ret = -EINVAL; | |
3433 | break; | |
3434 | } | |
13af18f2 ZC |
3435 | if (!migration_incoming_in_colo_state()) { |
3436 | ramblock_recv_bitmap_set(block, host); | |
3437 | } | |
3438 | ||
1db9d8e5 | 3439 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
3440 | } |
3441 | ||
56e93d26 JQ |
3442 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
3443 | case RAM_SAVE_FLAG_MEM_SIZE: | |
3444 | /* Synchronize RAM block list */ | |
3445 | total_ram_bytes = addr; | |
3446 | while (!ret && total_ram_bytes) { | |
3447 | RAMBlock *block; | |
56e93d26 JQ |
3448 | char id[256]; |
3449 | ram_addr_t length; | |
3450 | ||
3451 | len = qemu_get_byte(f); | |
3452 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3453 | id[len] = 0; | |
3454 | length = qemu_get_be64(f); | |
3455 | ||
e3dd7493 | 3456 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
3457 | if (block && !qemu_ram_is_migratable(block)) { |
3458 | error_report("block %s should not be migrated !", id); | |
3459 | ret = -EINVAL; | |
3460 | } else if (block) { | |
e3dd7493 DDAG |
3461 | if (length != block->used_length) { |
3462 | Error *local_err = NULL; | |
56e93d26 | 3463 | |
fa53a0e5 | 3464 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
3465 | &local_err); |
3466 | if (local_err) { | |
3467 | error_report_err(local_err); | |
56e93d26 | 3468 | } |
56e93d26 | 3469 | } |
ef08fb38 DDAG |
3470 | /* For postcopy we need to check hugepage sizes match */ |
3471 | if (postcopy_advised && | |
3472 | block->page_size != qemu_host_page_size) { | |
3473 | uint64_t remote_page_size = qemu_get_be64(f); | |
3474 | if (remote_page_size != block->page_size) { | |
3475 | error_report("Mismatched RAM page size %s " | |
3476 | "(local) %zd != %" PRId64, | |
3477 | id, block->page_size, | |
3478 | remote_page_size); | |
3479 | ret = -EINVAL; | |
3480 | } | |
3481 | } | |
fbd162e6 YK |
3482 | if (migrate_ignore_shared()) { |
3483 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
3484 | if (ramblock_is_ignored(block) && |
3485 | block->mr->addr != addr) { | |
3486 | error_report("Mismatched GPAs for block %s " | |
3487 | "%" PRId64 "!= %" PRId64, | |
3488 | id, (uint64_t)addr, | |
3489 | (uint64_t)block->mr->addr); | |
3490 | ret = -EINVAL; | |
3491 | } | |
3492 | } | |
e3dd7493 DDAG |
3493 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
3494 | block->idstr); | |
3495 | } else { | |
56e93d26 JQ |
3496 | error_report("Unknown ramblock \"%s\", cannot " |
3497 | "accept migration", id); | |
3498 | ret = -EINVAL; | |
3499 | } | |
3500 | ||
3501 | total_ram_bytes -= length; | |
3502 | } | |
3503 | break; | |
a776aa15 | 3504 | |
bb890ed5 | 3505 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
3506 | ch = qemu_get_byte(f); |
3507 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
3508 | break; | |
a776aa15 | 3509 | |
56e93d26 | 3510 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
3511 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
3512 | break; | |
56e93d26 | 3513 | |
a776aa15 | 3514 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
3515 | len = qemu_get_be32(f); |
3516 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
3517 | error_report("Invalid compressed data length: %d", len); | |
3518 | ret = -EINVAL; | |
3519 | break; | |
3520 | } | |
c1bc6626 | 3521 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 3522 | break; |
a776aa15 | 3523 | |
56e93d26 | 3524 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
3525 | if (load_xbzrle(f, addr, host) < 0) { |
3526 | error_report("Failed to decompress XBZRLE page at " | |
3527 | RAM_ADDR_FMT, addr); | |
3528 | ret = -EINVAL; | |
3529 | break; | |
3530 | } | |
3531 | break; | |
3532 | case RAM_SAVE_FLAG_EOS: | |
3533 | /* normal exit */ | |
6df264ac | 3534 | multifd_recv_sync_main(); |
56e93d26 JQ |
3535 | break; |
3536 | default: | |
3537 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 3538 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
3539 | } else { |
3540 | error_report("Unknown combination of migration flags: %#x", | |
3541 | flags); | |
3542 | ret = -EINVAL; | |
3543 | } | |
3544 | } | |
3545 | if (!ret) { | |
3546 | ret = qemu_file_get_error(f); | |
3547 | } | |
0393031a HZ |
3548 | if (!ret && host_bak) { |
3549 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
3550 | } | |
56e93d26 JQ |
3551 | } |
3552 | ||
ca1a6b70 | 3553 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
3554 | return ret; |
3555 | } | |
3556 | ||
3557 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
3558 | { | |
3559 | int ret = 0; | |
3560 | static uint64_t seq_iter; | |
3561 | /* | |
3562 | * If system is running in postcopy mode, page inserts to host memory must | |
3563 | * be atomic | |
3564 | */ | |
3565 | bool postcopy_running = postcopy_is_running(); | |
3566 | ||
3567 | seq_iter++; | |
3568 | ||
3569 | if (version_id != 4) { | |
3570 | return -EINVAL; | |
3571 | } | |
3572 | ||
3573 | /* | |
3574 | * This RCU critical section can be very long running. | |
3575 | * When RCU reclaims in the code start to become numerous, | |
3576 | * it will be necessary to reduce the granularity of this | |
3577 | * critical section. | |
3578 | */ | |
89ac5a1d DDAG |
3579 | WITH_RCU_READ_LOCK_GUARD() { |
3580 | if (postcopy_running) { | |
3581 | ret = ram_load_postcopy(f); | |
3582 | } else { | |
3583 | ret = ram_load_precopy(f); | |
3584 | } | |
10da4a36 | 3585 | } |
55c4446b | 3586 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 ZC |
3587 | |
3588 | if (!ret && migration_incoming_in_colo_state()) { | |
3589 | colo_flush_ram_cache(); | |
3590 | } | |
56e93d26 JQ |
3591 | return ret; |
3592 | } | |
3593 | ||
c6467627 VSO |
3594 | static bool ram_has_postcopy(void *opaque) |
3595 | { | |
469dd51b | 3596 | RAMBlock *rb; |
fbd162e6 | 3597 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
3598 | if (ramblock_is_pmem(rb)) { |
3599 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
3600 | "is not supported now!", rb->idstr, rb->host); | |
3601 | return false; | |
3602 | } | |
3603 | } | |
3604 | ||
c6467627 VSO |
3605 | return migrate_postcopy_ram(); |
3606 | } | |
3607 | ||
edd090c7 PX |
3608 | /* Sync all the dirty bitmap with destination VM. */ |
3609 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
3610 | { | |
3611 | RAMBlock *block; | |
3612 | QEMUFile *file = s->to_dst_file; | |
3613 | int ramblock_count = 0; | |
3614 | ||
3615 | trace_ram_dirty_bitmap_sync_start(); | |
3616 | ||
fbd162e6 | 3617 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
3618 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
3619 | trace_ram_dirty_bitmap_request(block->idstr); | |
3620 | ramblock_count++; | |
3621 | } | |
3622 | ||
3623 | trace_ram_dirty_bitmap_sync_wait(); | |
3624 | ||
3625 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
3626 | while (ramblock_count--) { | |
3627 | qemu_sem_wait(&s->rp_state.rp_sem); | |
3628 | } | |
3629 | ||
3630 | trace_ram_dirty_bitmap_sync_complete(); | |
3631 | ||
3632 | return 0; | |
3633 | } | |
3634 | ||
3635 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
3636 | { | |
3637 | qemu_sem_post(&s->rp_state.rp_sem); | |
3638 | } | |
3639 | ||
a335debb PX |
3640 | /* |
3641 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
3642 | * This is only used when the postcopy migration is paused but wants | |
3643 | * to resume from a middle point. | |
3644 | */ | |
3645 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
3646 | { | |
3647 | int ret = -EINVAL; | |
3648 | QEMUFile *file = s->rp_state.from_dst_file; | |
3649 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 3650 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
3651 | uint64_t size, end_mark; |
3652 | ||
3653 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
3654 | ||
3655 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
3656 | error_report("%s: incorrect state %s", __func__, | |
3657 | MigrationStatus_str(s->state)); | |
3658 | return -EINVAL; | |
3659 | } | |
3660 | ||
3661 | /* | |
3662 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3663 | * need the endianess convertion, and the paddings. | |
3664 | */ | |
3665 | local_size = ROUND_UP(local_size, 8); | |
3666 | ||
3667 | /* Add paddings */ | |
3668 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
3669 | ||
3670 | size = qemu_get_be64(file); | |
3671 | ||
3672 | /* The size of the bitmap should match with our ramblock */ | |
3673 | if (size != local_size) { | |
3674 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
3675 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
3676 | block->idstr, size, local_size); | |
3677 | ret = -EINVAL; | |
3678 | goto out; | |
3679 | } | |
3680 | ||
3681 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
3682 | end_mark = qemu_get_be64(file); | |
3683 | ||
3684 | ret = qemu_file_get_error(file); | |
3685 | if (ret || size != local_size) { | |
3686 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
3687 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
3688 | __func__, block->idstr, ret, local_size, size); | |
3689 | ret = -EIO; | |
3690 | goto out; | |
3691 | } | |
3692 | ||
3693 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
3694 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIu64, | |
3695 | __func__, block->idstr, end_mark); | |
3696 | ret = -EINVAL; | |
3697 | goto out; | |
3698 | } | |
3699 | ||
3700 | /* | |
3701 | * Endianess convertion. We are during postcopy (though paused). | |
3702 | * The dirty bitmap won't change. We can directly modify it. | |
3703 | */ | |
3704 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
3705 | ||
3706 | /* | |
3707 | * What we received is "received bitmap". Revert it as the initial | |
3708 | * dirty bitmap for this ramblock. | |
3709 | */ | |
3710 | bitmap_complement(block->bmap, block->bmap, nbits); | |
3711 | ||
3712 | trace_ram_dirty_bitmap_reload_complete(block->idstr); | |
3713 | ||
edd090c7 PX |
3714 | /* |
3715 | * We succeeded to sync bitmap for current ramblock. If this is | |
3716 | * the last one to sync, we need to notify the main send thread. | |
3717 | */ | |
3718 | ram_dirty_bitmap_reload_notify(s); | |
3719 | ||
a335debb PX |
3720 | ret = 0; |
3721 | out: | |
bf269906 | 3722 | g_free(le_bitmap); |
a335debb PX |
3723 | return ret; |
3724 | } | |
3725 | ||
edd090c7 PX |
3726 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
3727 | { | |
3728 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 3729 | int ret; |
edd090c7 | 3730 | |
08614f34 PX |
3731 | ret = ram_dirty_bitmap_sync_all(s, rs); |
3732 | if (ret) { | |
3733 | return ret; | |
3734 | } | |
3735 | ||
3736 | ram_state_resume_prepare(rs, s->to_dst_file); | |
3737 | ||
3738 | return 0; | |
edd090c7 PX |
3739 | } |
3740 | ||
56e93d26 | 3741 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 3742 | .save_setup = ram_save_setup, |
56e93d26 | 3743 | .save_live_iterate = ram_save_iterate, |
763c906b | 3744 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 3745 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 3746 | .has_postcopy = ram_has_postcopy, |
56e93d26 JQ |
3747 | .save_live_pending = ram_save_pending, |
3748 | .load_state = ram_load, | |
f265e0e4 JQ |
3749 | .save_cleanup = ram_save_cleanup, |
3750 | .load_setup = ram_load_setup, | |
3751 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 3752 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
3753 | }; |
3754 | ||
3755 | void ram_mig_init(void) | |
3756 | { | |
3757 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 3758 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 3759 | } |