]>
Commit | Line | Data |
---|---|---|
56e93d26 JQ |
1 | /* |
2 | * QEMU System Emulator | |
3 | * | |
4 | * Copyright (c) 2003-2008 Fabrice Bellard | |
76cc7b58 JQ |
5 | * Copyright (c) 2011-2015 Red Hat Inc |
6 | * | |
7 | * Authors: | |
8 | * Juan Quintela <quintela@redhat.com> | |
56e93d26 JQ |
9 | * |
10 | * Permission is hereby granted, free of charge, to any person obtaining a copy | |
11 | * of this software and associated documentation files (the "Software"), to deal | |
12 | * in the Software without restriction, including without limitation the rights | |
13 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell | |
14 | * copies of the Software, and to permit persons to whom the Software is | |
15 | * furnished to do so, subject to the following conditions: | |
16 | * | |
17 | * The above copyright notice and this permission notice shall be included in | |
18 | * all copies or substantial portions of the Software. | |
19 | * | |
20 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
21 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
22 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
23 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | |
24 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | |
25 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | |
26 | * THE SOFTWARE. | |
27 | */ | |
1393a485 | 28 | #include "qemu/osdep.h" |
33c11879 PB |
29 | #include "qemu-common.h" |
30 | #include "cpu.h" | |
56e93d26 | 31 | #include <zlib.h> |
4addcd4f | 32 | #include "qapi-event.h" |
f348b6d1 | 33 | #include "qemu/cutils.h" |
56e93d26 JQ |
34 | #include "qemu/bitops.h" |
35 | #include "qemu/bitmap.h" | |
7205c9ec JQ |
36 | #include "qemu/timer.h" |
37 | #include "qemu/main-loop.h" | |
56e93d26 | 38 | #include "migration/migration.h" |
e0b266f0 | 39 | #include "migration/postcopy-ram.h" |
56e93d26 JQ |
40 | #include "exec/address-spaces.h" |
41 | #include "migration/page_cache.h" | |
56e93d26 | 42 | #include "qemu/error-report.h" |
56e93d26 | 43 | #include "trace.h" |
56e93d26 | 44 | #include "exec/ram_addr.h" |
56e93d26 | 45 | #include "qemu/rcu_queue.h" |
a91246c9 | 46 | #include "migration/colo.h" |
56e93d26 | 47 | |
56e93d26 JQ |
48 | /***********************************************************/ |
49 | /* ram save/restore */ | |
50 | ||
51 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ | |
52 | #define RAM_SAVE_FLAG_COMPRESS 0x02 | |
53 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 | |
54 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
55 | #define RAM_SAVE_FLAG_EOS 0x10 | |
56 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
57 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
58 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
59 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
60 | ||
adb65dec | 61 | static uint8_t *ZERO_TARGET_PAGE; |
56e93d26 JQ |
62 | |
63 | static inline bool is_zero_range(uint8_t *p, uint64_t size) | |
64 | { | |
a1febc49 | 65 | return buffer_is_zero(p, size); |
56e93d26 JQ |
66 | } |
67 | ||
68 | /* struct contains XBZRLE cache and a static page | |
69 | used by the compression */ | |
70 | static struct { | |
71 | /* buffer used for XBZRLE encoding */ | |
72 | uint8_t *encoded_buf; | |
73 | /* buffer for storing page content */ | |
74 | uint8_t *current_buf; | |
75 | /* Cache for XBZRLE, Protected by lock. */ | |
76 | PageCache *cache; | |
77 | QemuMutex lock; | |
78 | } XBZRLE; | |
79 | ||
80 | /* buffer used for XBZRLE decoding */ | |
81 | static uint8_t *xbzrle_decoded_buf; | |
82 | ||
83 | static void XBZRLE_cache_lock(void) | |
84 | { | |
85 | if (migrate_use_xbzrle()) | |
86 | qemu_mutex_lock(&XBZRLE.lock); | |
87 | } | |
88 | ||
89 | static void XBZRLE_cache_unlock(void) | |
90 | { | |
91 | if (migrate_use_xbzrle()) | |
92 | qemu_mutex_unlock(&XBZRLE.lock); | |
93 | } | |
94 | ||
3d0684b2 JQ |
95 | /** |
96 | * xbzrle_cache_resize: resize the xbzrle cache | |
97 | * | |
98 | * This function is called from qmp_migrate_set_cache_size in main | |
99 | * thread, possibly while a migration is in progress. A running | |
100 | * migration may be using the cache and might finish during this call, | |
101 | * hence changes to the cache are protected by XBZRLE.lock(). | |
102 | * | |
103 | * Returns the new_size or negative in case of error. | |
104 | * | |
105 | * @new_size: new cache size | |
56e93d26 JQ |
106 | */ |
107 | int64_t xbzrle_cache_resize(int64_t new_size) | |
108 | { | |
109 | PageCache *new_cache; | |
110 | int64_t ret; | |
111 | ||
112 | if (new_size < TARGET_PAGE_SIZE) { | |
113 | return -1; | |
114 | } | |
115 | ||
116 | XBZRLE_cache_lock(); | |
117 | ||
118 | if (XBZRLE.cache != NULL) { | |
119 | if (pow2floor(new_size) == migrate_xbzrle_cache_size()) { | |
120 | goto out_new_size; | |
121 | } | |
122 | new_cache = cache_init(new_size / TARGET_PAGE_SIZE, | |
123 | TARGET_PAGE_SIZE); | |
124 | if (!new_cache) { | |
125 | error_report("Error creating cache"); | |
126 | ret = -1; | |
127 | goto out; | |
128 | } | |
129 | ||
130 | cache_fini(XBZRLE.cache); | |
131 | XBZRLE.cache = new_cache; | |
132 | } | |
133 | ||
134 | out_new_size: | |
135 | ret = pow2floor(new_size); | |
136 | out: | |
137 | XBZRLE_cache_unlock(); | |
138 | return ret; | |
139 | } | |
140 | ||
eb859c53 JQ |
141 | struct RAMBitmap { |
142 | struct rcu_head rcu; | |
143 | /* Main migration bitmap */ | |
144 | unsigned long *bmap; | |
145 | /* bitmap of pages that haven't been sent even once | |
146 | * only maintained and used in postcopy at the moment | |
147 | * where it's used to send the dirtymap at the start | |
148 | * of the postcopy phase | |
149 | */ | |
150 | unsigned long *unsentmap; | |
151 | }; | |
152 | typedef struct RAMBitmap RAMBitmap; | |
153 | ||
6f37bb8b JQ |
154 | /* State of RAM for migration */ |
155 | struct RAMState { | |
156 | /* Last block that we have visited searching for dirty pages */ | |
157 | RAMBlock *last_seen_block; | |
158 | /* Last block from where we have sent data */ | |
159 | RAMBlock *last_sent_block; | |
160 | /* Last offset we have sent data from */ | |
161 | ram_addr_t last_offset; | |
162 | /* last ram version we have seen */ | |
163 | uint32_t last_version; | |
164 | /* We are in the first round */ | |
165 | bool ram_bulk_stage; | |
8d820d6f JQ |
166 | /* How many times we have dirty too many pages */ |
167 | int dirty_rate_high_cnt; | |
5a987738 JQ |
168 | /* How many times we have synchronized the bitmap */ |
169 | uint64_t bitmap_sync_count; | |
f664da80 JQ |
170 | /* these variables are used for bitmap sync */ |
171 | /* last time we did a full bitmap_sync */ | |
172 | int64_t time_last_bitmap_sync; | |
eac74159 | 173 | /* bytes transferred at start_time */ |
c4bdf0cf | 174 | uint64_t bytes_xfer_prev; |
a66cd90c | 175 | /* number of dirty pages since start_time */ |
68908ed6 | 176 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
177 | /* xbzrle misses since the beginning of the period */ |
178 | uint64_t xbzrle_cache_miss_prev; | |
36040d9c JQ |
179 | /* number of iterations at the beginning of period */ |
180 | uint64_t iterations_prev; | |
f7ccd61b JQ |
181 | /* Accounting fields */ |
182 | /* number of zero pages. It used to be pages filled by the same char. */ | |
183 | uint64_t zero_pages; | |
b4d1c6e7 JQ |
184 | /* number of normal transferred pages */ |
185 | uint64_t norm_pages; | |
23b28c3c JQ |
186 | /* Iterations since start */ |
187 | uint64_t iterations; | |
f36ada95 JQ |
188 | /* xbzrle transmitted bytes. Notice that this is with |
189 | * compression, they can't be calculated from the pages */ | |
07ed50a2 | 190 | uint64_t xbzrle_bytes; |
f36ada95 JQ |
191 | /* xbzrle transmmited pages */ |
192 | uint64_t xbzrle_pages; | |
544c36f1 JQ |
193 | /* xbzrle number of cache miss */ |
194 | uint64_t xbzrle_cache_miss; | |
b07016b6 JQ |
195 | /* xbzrle miss rate */ |
196 | double xbzrle_cache_miss_rate; | |
180f61f7 JQ |
197 | /* xbzrle number of overflows */ |
198 | uint64_t xbzrle_overflows; | |
0d8ec885 JQ |
199 | /* number of dirty bits in the bitmap */ |
200 | uint64_t migration_dirty_pages; | |
2f4fde93 JQ |
201 | /* total number of bytes transferred */ |
202 | uint64_t bytes_transferred; | |
108cfae0 JQ |
203 | /* protects modification of the bitmap */ |
204 | QemuMutex bitmap_mutex; | |
eb859c53 JQ |
205 | /* Ram Bitmap protected by RCU */ |
206 | RAMBitmap *ram_bitmap; | |
68a098f3 JQ |
207 | /* The RAMBlock used in the last src_page_requests */ |
208 | RAMBlock *last_req_rb; | |
6f37bb8b JQ |
209 | }; |
210 | typedef struct RAMState RAMState; | |
211 | ||
212 | static RAMState ram_state; | |
213 | ||
56e93d26 JQ |
214 | uint64_t dup_mig_pages_transferred(void) |
215 | { | |
f7ccd61b | 216 | return ram_state.zero_pages; |
56e93d26 JQ |
217 | } |
218 | ||
56e93d26 JQ |
219 | uint64_t norm_mig_pages_transferred(void) |
220 | { | |
b4d1c6e7 | 221 | return ram_state.norm_pages; |
56e93d26 JQ |
222 | } |
223 | ||
224 | uint64_t xbzrle_mig_bytes_transferred(void) | |
225 | { | |
07ed50a2 | 226 | return ram_state.xbzrle_bytes; |
56e93d26 JQ |
227 | } |
228 | ||
229 | uint64_t xbzrle_mig_pages_transferred(void) | |
230 | { | |
f36ada95 | 231 | return ram_state.xbzrle_pages; |
56e93d26 JQ |
232 | } |
233 | ||
234 | uint64_t xbzrle_mig_pages_cache_miss(void) | |
235 | { | |
544c36f1 | 236 | return ram_state.xbzrle_cache_miss; |
56e93d26 JQ |
237 | } |
238 | ||
239 | double xbzrle_mig_cache_miss_rate(void) | |
240 | { | |
b07016b6 | 241 | return ram_state.xbzrle_cache_miss_rate; |
56e93d26 JQ |
242 | } |
243 | ||
244 | uint64_t xbzrle_mig_pages_overflow(void) | |
245 | { | |
180f61f7 | 246 | return ram_state.xbzrle_overflows; |
56e93d26 JQ |
247 | } |
248 | ||
9edabd4d | 249 | uint64_t ram_bytes_transferred(void) |
0d8ec885 | 250 | { |
9edabd4d | 251 | return ram_state.bytes_transferred; |
0d8ec885 JQ |
252 | } |
253 | ||
9edabd4d | 254 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 255 | { |
9edabd4d | 256 | return ram_state.migration_dirty_pages * TARGET_PAGE_SIZE; |
2f4fde93 JQ |
257 | } |
258 | ||
b8fb8cb7 DDAG |
259 | /* used by the search for pages to send */ |
260 | struct PageSearchStatus { | |
261 | /* Current block being searched */ | |
262 | RAMBlock *block; | |
263 | /* Current offset to search from */ | |
264 | ram_addr_t offset; | |
265 | /* Set once we wrap around */ | |
266 | bool complete_round; | |
267 | }; | |
268 | typedef struct PageSearchStatus PageSearchStatus; | |
269 | ||
56e93d26 | 270 | struct CompressParam { |
56e93d26 | 271 | bool done; |
90e56fb4 | 272 | bool quit; |
56e93d26 JQ |
273 | QEMUFile *file; |
274 | QemuMutex mutex; | |
275 | QemuCond cond; | |
276 | RAMBlock *block; | |
277 | ram_addr_t offset; | |
278 | }; | |
279 | typedef struct CompressParam CompressParam; | |
280 | ||
281 | struct DecompressParam { | |
73a8912b | 282 | bool done; |
90e56fb4 | 283 | bool quit; |
56e93d26 JQ |
284 | QemuMutex mutex; |
285 | QemuCond cond; | |
286 | void *des; | |
d341d9f3 | 287 | uint8_t *compbuf; |
56e93d26 JQ |
288 | int len; |
289 | }; | |
290 | typedef struct DecompressParam DecompressParam; | |
291 | ||
292 | static CompressParam *comp_param; | |
293 | static QemuThread *compress_threads; | |
294 | /* comp_done_cond is used to wake up the migration thread when | |
295 | * one of the compression threads has finished the compression. | |
296 | * comp_done_lock is used to co-work with comp_done_cond. | |
297 | */ | |
0d9f9a5c LL |
298 | static QemuMutex comp_done_lock; |
299 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
300 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
301 | static const QEMUFileOps empty_ops = { }; | |
302 | ||
303 | static bool compression_switch; | |
56e93d26 JQ |
304 | static DecompressParam *decomp_param; |
305 | static QemuThread *decompress_threads; | |
73a8912b LL |
306 | static QemuMutex decomp_done_lock; |
307 | static QemuCond decomp_done_cond; | |
56e93d26 | 308 | |
a7a9a88f LL |
309 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
310 | ram_addr_t offset); | |
56e93d26 JQ |
311 | |
312 | static void *do_data_compress(void *opaque) | |
313 | { | |
314 | CompressParam *param = opaque; | |
a7a9a88f LL |
315 | RAMBlock *block; |
316 | ram_addr_t offset; | |
56e93d26 | 317 | |
a7a9a88f | 318 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 319 | while (!param->quit) { |
a7a9a88f LL |
320 | if (param->block) { |
321 | block = param->block; | |
322 | offset = param->offset; | |
323 | param->block = NULL; | |
324 | qemu_mutex_unlock(¶m->mutex); | |
325 | ||
326 | do_compress_ram_page(param->file, block, offset); | |
327 | ||
0d9f9a5c | 328 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 329 | param->done = true; |
0d9f9a5c LL |
330 | qemu_cond_signal(&comp_done_cond); |
331 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
332 | |
333 | qemu_mutex_lock(¶m->mutex); | |
334 | } else { | |
56e93d26 JQ |
335 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
336 | } | |
56e93d26 | 337 | } |
a7a9a88f | 338 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
339 | |
340 | return NULL; | |
341 | } | |
342 | ||
343 | static inline void terminate_compression_threads(void) | |
344 | { | |
345 | int idx, thread_count; | |
346 | ||
347 | thread_count = migrate_compress_threads(); | |
3d0684b2 | 348 | |
56e93d26 JQ |
349 | for (idx = 0; idx < thread_count; idx++) { |
350 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 351 | comp_param[idx].quit = true; |
56e93d26 JQ |
352 | qemu_cond_signal(&comp_param[idx].cond); |
353 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
354 | } | |
355 | } | |
356 | ||
357 | void migrate_compress_threads_join(void) | |
358 | { | |
359 | int i, thread_count; | |
360 | ||
361 | if (!migrate_use_compression()) { | |
362 | return; | |
363 | } | |
364 | terminate_compression_threads(); | |
365 | thread_count = migrate_compress_threads(); | |
366 | for (i = 0; i < thread_count; i++) { | |
367 | qemu_thread_join(compress_threads + i); | |
368 | qemu_fclose(comp_param[i].file); | |
369 | qemu_mutex_destroy(&comp_param[i].mutex); | |
370 | qemu_cond_destroy(&comp_param[i].cond); | |
371 | } | |
0d9f9a5c LL |
372 | qemu_mutex_destroy(&comp_done_lock); |
373 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
374 | g_free(compress_threads); |
375 | g_free(comp_param); | |
56e93d26 JQ |
376 | compress_threads = NULL; |
377 | comp_param = NULL; | |
56e93d26 JQ |
378 | } |
379 | ||
380 | void migrate_compress_threads_create(void) | |
381 | { | |
382 | int i, thread_count; | |
383 | ||
384 | if (!migrate_use_compression()) { | |
385 | return; | |
386 | } | |
56e93d26 JQ |
387 | compression_switch = true; |
388 | thread_count = migrate_compress_threads(); | |
389 | compress_threads = g_new0(QemuThread, thread_count); | |
390 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
391 | qemu_cond_init(&comp_done_cond); |
392 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 393 | for (i = 0; i < thread_count; i++) { |
e110aa91 C |
394 | /* comp_param[i].file is just used as a dummy buffer to save data, |
395 | * set its ops to empty. | |
56e93d26 JQ |
396 | */ |
397 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
398 | comp_param[i].done = true; | |
90e56fb4 | 399 | comp_param[i].quit = false; |
56e93d26 JQ |
400 | qemu_mutex_init(&comp_param[i].mutex); |
401 | qemu_cond_init(&comp_param[i].cond); | |
402 | qemu_thread_create(compress_threads + i, "compress", | |
403 | do_data_compress, comp_param + i, | |
404 | QEMU_THREAD_JOINABLE); | |
405 | } | |
406 | } | |
407 | ||
408 | /** | |
3d0684b2 | 409 | * save_page_header: write page header to wire |
56e93d26 JQ |
410 | * |
411 | * If this is the 1st block, it also writes the block identification | |
412 | * | |
3d0684b2 | 413 | * Returns the number of bytes written |
56e93d26 JQ |
414 | * |
415 | * @f: QEMUFile where to send the data | |
416 | * @block: block that contains the page we want to send | |
417 | * @offset: offset inside the block for the page | |
418 | * in the lower bits, it contains flags | |
419 | */ | |
420 | static size_t save_page_header(QEMUFile *f, RAMBlock *block, ram_addr_t offset) | |
421 | { | |
9f5f380b | 422 | size_t size, len; |
56e93d26 JQ |
423 | |
424 | qemu_put_be64(f, offset); | |
425 | size = 8; | |
426 | ||
427 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b LL |
428 | len = strlen(block->idstr); |
429 | qemu_put_byte(f, len); | |
430 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
431 | size += 1 + len; | |
56e93d26 JQ |
432 | } |
433 | return size; | |
434 | } | |
435 | ||
3d0684b2 JQ |
436 | /** |
437 | * mig_throttle_guest_down: throotle down the guest | |
438 | * | |
439 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
440 | * writes. If guest dirty memory rate is reduced below the rate at | |
441 | * which we can transfer pages to the destination then we should be | |
442 | * able to complete migration. Some workloads dirty memory way too | |
443 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
444 | */ |
445 | static void mig_throttle_guest_down(void) | |
446 | { | |
447 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
448 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
449 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
450 | |
451 | /* We have not started throttling yet. Let's start it. */ | |
452 | if (!cpu_throttle_active()) { | |
453 | cpu_throttle_set(pct_initial); | |
454 | } else { | |
455 | /* Throttling already on, just increase the rate */ | |
456 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
457 | } | |
458 | } | |
459 | ||
3d0684b2 JQ |
460 | /** |
461 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
462 | * | |
6f37bb8b | 463 | * @rs: current RAM state |
3d0684b2 JQ |
464 | * @current_addr: address for the zero page |
465 | * | |
466 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
467 | * The important thing is that a stale (not-yet-0'd) page be replaced |
468 | * by the new data. | |
469 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 470 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 471 | */ |
6f37bb8b | 472 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 473 | { |
6f37bb8b | 474 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
475 | return; |
476 | } | |
477 | ||
478 | /* We don't care if this fails to allocate a new cache page | |
479 | * as long as it updated an old one */ | |
480 | cache_insert(XBZRLE.cache, current_addr, ZERO_TARGET_PAGE, | |
5a987738 | 481 | rs->bitmap_sync_count); |
56e93d26 JQ |
482 | } |
483 | ||
484 | #define ENCODING_FLAG_XBZRLE 0x1 | |
485 | ||
486 | /** | |
487 | * save_xbzrle_page: compress and send current page | |
488 | * | |
489 | * Returns: 1 means that we wrote the page | |
490 | * 0 means that page is identical to the one already sent | |
491 | * -1 means that xbzrle would be longer than normal | |
492 | * | |
5a987738 | 493 | * @rs: current RAM state |
56e93d26 | 494 | * @f: QEMUFile where to send the data |
3d0684b2 JQ |
495 | * @current_data: pointer to the address of the page contents |
496 | * @current_addr: addr of the page | |
56e93d26 JQ |
497 | * @block: block that contains the page we want to send |
498 | * @offset: offset inside the block for the page | |
499 | * @last_stage: if we are at the completion stage | |
56e93d26 | 500 | */ |
5a987738 | 501 | static int save_xbzrle_page(RAMState *rs, QEMUFile *f, uint8_t **current_data, |
56e93d26 | 502 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 503 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
504 | { |
505 | int encoded_len = 0, bytes_xbzrle; | |
506 | uint8_t *prev_cached_page; | |
507 | ||
5a987738 | 508 | if (!cache_is_cached(XBZRLE.cache, current_addr, rs->bitmap_sync_count)) { |
544c36f1 | 509 | rs->xbzrle_cache_miss++; |
56e93d26 JQ |
510 | if (!last_stage) { |
511 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
5a987738 | 512 | rs->bitmap_sync_count) == -1) { |
56e93d26 JQ |
513 | return -1; |
514 | } else { | |
515 | /* update *current_data when the page has been | |
516 | inserted into cache */ | |
517 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
518 | } | |
519 | } | |
520 | return -1; | |
521 | } | |
522 | ||
523 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
524 | ||
525 | /* save current buffer into memory */ | |
526 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
527 | ||
528 | /* XBZRLE encoding (if there is no overflow) */ | |
529 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
530 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
531 | TARGET_PAGE_SIZE); | |
532 | if (encoded_len == 0) { | |
55c4446b | 533 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
534 | return 0; |
535 | } else if (encoded_len == -1) { | |
55c4446b | 536 | trace_save_xbzrle_page_overflow(); |
180f61f7 | 537 | rs->xbzrle_overflows++; |
56e93d26 JQ |
538 | /* update data in the cache */ |
539 | if (!last_stage) { | |
540 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
541 | *current_data = prev_cached_page; | |
542 | } | |
543 | return -1; | |
544 | } | |
545 | ||
546 | /* we need to update the data in the cache, in order to get the same data */ | |
547 | if (!last_stage) { | |
548 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
549 | } | |
550 | ||
551 | /* Send XBZRLE based compressed page */ | |
552 | bytes_xbzrle = save_page_header(f, block, offset | RAM_SAVE_FLAG_XBZRLE); | |
553 | qemu_put_byte(f, ENCODING_FLAG_XBZRLE); | |
554 | qemu_put_be16(f, encoded_len); | |
555 | qemu_put_buffer(f, XBZRLE.encoded_buf, encoded_len); | |
556 | bytes_xbzrle += encoded_len + 1 + 2; | |
f36ada95 | 557 | rs->xbzrle_pages++; |
07ed50a2 | 558 | rs->xbzrle_bytes += bytes_xbzrle; |
072c2511 | 559 | rs->bytes_transferred += bytes_xbzrle; |
56e93d26 JQ |
560 | |
561 | return 1; | |
562 | } | |
563 | ||
3d0684b2 JQ |
564 | /** |
565 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 566 | * |
3d0684b2 JQ |
567 | * Called with rcu_read_lock() to protect migration_bitmap |
568 | * | |
569 | * Returns the byte offset within memory region of the start of a dirty page | |
570 | * | |
6f37bb8b | 571 | * @rs: current RAM state |
3d0684b2 JQ |
572 | * @rb: RAMBlock where to search for dirty pages |
573 | * @start: starting address (typically so we can continue from previous page) | |
574 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
575 | * within the global ram_addr space | |
f3f491fc | 576 | */ |
56e93d26 | 577 | static inline |
6f37bb8b | 578 | ram_addr_t migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
a82d593b DDAG |
579 | ram_addr_t start, |
580 | ram_addr_t *ram_addr_abs) | |
56e93d26 | 581 | { |
2f68e399 | 582 | unsigned long base = rb->offset >> TARGET_PAGE_BITS; |
56e93d26 | 583 | unsigned long nr = base + (start >> TARGET_PAGE_BITS); |
2f68e399 DDAG |
584 | uint64_t rb_size = rb->used_length; |
585 | unsigned long size = base + (rb_size >> TARGET_PAGE_BITS); | |
2ff64038 | 586 | unsigned long *bitmap; |
56e93d26 JQ |
587 | |
588 | unsigned long next; | |
589 | ||
eb859c53 | 590 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
6f37bb8b | 591 | if (rs->ram_bulk_stage && nr > base) { |
56e93d26 JQ |
592 | next = nr + 1; |
593 | } else { | |
2ff64038 | 594 | next = find_next_bit(bitmap, size, nr); |
56e93d26 JQ |
595 | } |
596 | ||
f3f491fc | 597 | *ram_addr_abs = next << TARGET_PAGE_BITS; |
56e93d26 JQ |
598 | return (next - base) << TARGET_PAGE_BITS; |
599 | } | |
600 | ||
0d8ec885 | 601 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, ram_addr_t addr) |
a82d593b DDAG |
602 | { |
603 | bool ret; | |
604 | int nr = addr >> TARGET_PAGE_BITS; | |
eb859c53 | 605 | unsigned long *bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
a82d593b DDAG |
606 | |
607 | ret = test_and_clear_bit(nr, bitmap); | |
608 | ||
609 | if (ret) { | |
0d8ec885 | 610 | rs->migration_dirty_pages--; |
a82d593b DDAG |
611 | } |
612 | return ret; | |
613 | } | |
614 | ||
a66cd90c JQ |
615 | static void migration_bitmap_sync_range(RAMState *rs, ram_addr_t start, |
616 | ram_addr_t length) | |
56e93d26 | 617 | { |
2ff64038 | 618 | unsigned long *bitmap; |
eb859c53 | 619 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
0d8ec885 JQ |
620 | rs->migration_dirty_pages += |
621 | cpu_physical_memory_sync_dirty_bitmap(bitmap, start, length, | |
622 | &rs->num_dirty_pages_period); | |
56e93d26 JQ |
623 | } |
624 | ||
3d0684b2 JQ |
625 | /** |
626 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
627 | * | |
628 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
629 | * | |
630 | * For VMs with just normal pages this is equivalent to the host page | |
631 | * size. If it's got some huge pages then it's the OR of all the | |
632 | * different page sizes. | |
e8ca1db2 DDAG |
633 | */ |
634 | uint64_t ram_pagesize_summary(void) | |
635 | { | |
636 | RAMBlock *block; | |
637 | uint64_t summary = 0; | |
638 | ||
639 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
640 | summary |= block->page_size; | |
641 | } | |
642 | ||
643 | return summary; | |
644 | } | |
645 | ||
8d820d6f | 646 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
647 | { |
648 | RAMBlock *block; | |
56e93d26 JQ |
649 | MigrationState *s = migrate_get_current(); |
650 | int64_t end_time; | |
c4bdf0cf | 651 | uint64_t bytes_xfer_now; |
56e93d26 | 652 | |
5a987738 | 653 | rs->bitmap_sync_count++; |
56e93d26 | 654 | |
eac74159 JQ |
655 | if (!rs->bytes_xfer_prev) { |
656 | rs->bytes_xfer_prev = ram_bytes_transferred(); | |
56e93d26 JQ |
657 | } |
658 | ||
f664da80 JQ |
659 | if (!rs->time_last_bitmap_sync) { |
660 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
661 | } |
662 | ||
663 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 664 | memory_global_dirty_log_sync(); |
56e93d26 | 665 | |
108cfae0 | 666 | qemu_mutex_lock(&rs->bitmap_mutex); |
56e93d26 JQ |
667 | rcu_read_lock(); |
668 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
a66cd90c | 669 | migration_bitmap_sync_range(rs, block->offset, block->used_length); |
56e93d26 JQ |
670 | } |
671 | rcu_read_unlock(); | |
108cfae0 | 672 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 673 | |
a66cd90c | 674 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 675 | |
56e93d26 JQ |
676 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
677 | ||
678 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 679 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
56e93d26 JQ |
680 | if (migrate_auto_converge()) { |
681 | /* The following detection logic can be refined later. For now: | |
682 | Check to see if the dirtied bytes is 50% more than the approx. | |
683 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
684 | were in this routine. If that happens twice, start or increase |
685 | throttling */ | |
56e93d26 | 686 | bytes_xfer_now = ram_bytes_transferred(); |
070afca2 | 687 | |
56e93d26 | 688 | if (s->dirty_pages_rate && |
a66cd90c | 689 | (rs->num_dirty_pages_period * TARGET_PAGE_SIZE > |
eac74159 | 690 | (bytes_xfer_now - rs->bytes_xfer_prev) / 2) && |
8d820d6f | 691 | (rs->dirty_rate_high_cnt++ >= 2)) { |
56e93d26 | 692 | trace_migration_throttle(); |
8d820d6f | 693 | rs->dirty_rate_high_cnt = 0; |
070afca2 | 694 | mig_throttle_guest_down(); |
56e93d26 | 695 | } |
eac74159 | 696 | rs->bytes_xfer_prev = bytes_xfer_now; |
56e93d26 | 697 | } |
070afca2 | 698 | |
56e93d26 | 699 | if (migrate_use_xbzrle()) { |
23b28c3c | 700 | if (rs->iterations_prev != rs->iterations) { |
b07016b6 | 701 | rs->xbzrle_cache_miss_rate = |
544c36f1 | 702 | (double)(rs->xbzrle_cache_miss - |
b5833fde | 703 | rs->xbzrle_cache_miss_prev) / |
23b28c3c | 704 | (rs->iterations - rs->iterations_prev); |
56e93d26 | 705 | } |
23b28c3c | 706 | rs->iterations_prev = rs->iterations; |
544c36f1 | 707 | rs->xbzrle_cache_miss_prev = rs->xbzrle_cache_miss; |
56e93d26 | 708 | } |
a66cd90c | 709 | s->dirty_pages_rate = rs->num_dirty_pages_period * 1000 |
f664da80 | 710 | / (end_time - rs->time_last_bitmap_sync); |
56e93d26 | 711 | s->dirty_bytes_rate = s->dirty_pages_rate * TARGET_PAGE_SIZE; |
f664da80 | 712 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 713 | rs->num_dirty_pages_period = 0; |
56e93d26 | 714 | } |
5a987738 | 715 | s->dirty_sync_count = rs->bitmap_sync_count; |
4addcd4f | 716 | if (migrate_use_events()) { |
5a987738 | 717 | qapi_event_send_migration_pass(rs->bitmap_sync_count, NULL); |
4addcd4f | 718 | } |
56e93d26 JQ |
719 | } |
720 | ||
721 | /** | |
3d0684b2 | 722 | * save_zero_page: send the zero page to the stream |
56e93d26 | 723 | * |
3d0684b2 | 724 | * Returns the number of pages written. |
56e93d26 | 725 | * |
f7ccd61b | 726 | * @rs: current RAM state |
56e93d26 JQ |
727 | * @f: QEMUFile where to send the data |
728 | * @block: block that contains the page we want to send | |
729 | * @offset: offset inside the block for the page | |
730 | * @p: pointer to the page | |
56e93d26 | 731 | */ |
f7ccd61b | 732 | static int save_zero_page(RAMState *rs, QEMUFile *f, RAMBlock *block, |
072c2511 | 733 | ram_addr_t offset, uint8_t *p) |
56e93d26 JQ |
734 | { |
735 | int pages = -1; | |
736 | ||
737 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
f7ccd61b | 738 | rs->zero_pages++; |
072c2511 JQ |
739 | rs->bytes_transferred += |
740 | save_page_header(f, block, offset | RAM_SAVE_FLAG_COMPRESS); | |
56e93d26 | 741 | qemu_put_byte(f, 0); |
072c2511 | 742 | rs->bytes_transferred += 1; |
56e93d26 JQ |
743 | pages = 1; |
744 | } | |
745 | ||
746 | return pages; | |
747 | } | |
748 | ||
36449157 | 749 | static void ram_release_pages(MigrationState *ms, const char *rbname, |
53f09a10 PB |
750 | uint64_t offset, int pages) |
751 | { | |
752 | if (!migrate_release_ram() || !migration_in_postcopy(ms)) { | |
753 | return; | |
754 | } | |
755 | ||
36449157 | 756 | ram_discard_range(NULL, rbname, offset, pages << TARGET_PAGE_BITS); |
53f09a10 PB |
757 | } |
758 | ||
56e93d26 | 759 | /** |
3d0684b2 | 760 | * ram_save_page: send the given page to the stream |
56e93d26 | 761 | * |
3d0684b2 | 762 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
763 | * < 0 - error |
764 | * >=0 - Number of pages written - this might legally be 0 | |
765 | * if xbzrle noticed the page was the same. | |
56e93d26 | 766 | * |
6f37bb8b | 767 | * @rs: current RAM state |
3d0684b2 | 768 | * @ms: current migration state |
56e93d26 JQ |
769 | * @f: QEMUFile where to send the data |
770 | * @block: block that contains the page we want to send | |
771 | * @offset: offset inside the block for the page | |
772 | * @last_stage: if we are at the completion stage | |
56e93d26 | 773 | */ |
6f37bb8b | 774 | static int ram_save_page(RAMState *rs, MigrationState *ms, QEMUFile *f, |
072c2511 | 775 | PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
776 | { |
777 | int pages = -1; | |
778 | uint64_t bytes_xmit; | |
779 | ram_addr_t current_addr; | |
56e93d26 JQ |
780 | uint8_t *p; |
781 | int ret; | |
782 | bool send_async = true; | |
a08f6890 HZ |
783 | RAMBlock *block = pss->block; |
784 | ram_addr_t offset = pss->offset; | |
56e93d26 | 785 | |
2f68e399 | 786 | p = block->host + offset; |
56e93d26 JQ |
787 | |
788 | /* In doubt sent page as normal */ | |
789 | bytes_xmit = 0; | |
790 | ret = ram_control_save_page(f, block->offset, | |
791 | offset, TARGET_PAGE_SIZE, &bytes_xmit); | |
792 | if (bytes_xmit) { | |
072c2511 | 793 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
794 | pages = 1; |
795 | } | |
796 | ||
797 | XBZRLE_cache_lock(); | |
798 | ||
799 | current_addr = block->offset + offset; | |
800 | ||
6f37bb8b | 801 | if (block == rs->last_sent_block) { |
56e93d26 JQ |
802 | offset |= RAM_SAVE_FLAG_CONTINUE; |
803 | } | |
804 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { | |
805 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
806 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 807 | rs->norm_pages++; |
56e93d26 | 808 | } else if (bytes_xmit == 0) { |
f7ccd61b | 809 | rs->zero_pages++; |
56e93d26 JQ |
810 | } |
811 | } | |
812 | } else { | |
072c2511 | 813 | pages = save_zero_page(rs, f, block, offset, p); |
56e93d26 JQ |
814 | if (pages > 0) { |
815 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
816 | * page would be stale | |
817 | */ | |
6f37bb8b | 818 | xbzrle_cache_zero_page(rs, current_addr); |
53f09a10 | 819 | ram_release_pages(ms, block->idstr, pss->offset, pages); |
6f37bb8b | 820 | } else if (!rs->ram_bulk_stage && |
9eb14766 | 821 | !migration_in_postcopy(ms) && migrate_use_xbzrle()) { |
5a987738 | 822 | pages = save_xbzrle_page(rs, f, &p, current_addr, block, |
072c2511 | 823 | offset, last_stage); |
56e93d26 JQ |
824 | if (!last_stage) { |
825 | /* Can't send this cached data async, since the cache page | |
826 | * might get updated before it gets to the wire | |
827 | */ | |
828 | send_async = false; | |
829 | } | |
830 | } | |
831 | } | |
832 | ||
833 | /* XBZRLE overflow or normal page */ | |
834 | if (pages == -1) { | |
072c2511 | 835 | rs->bytes_transferred += save_page_header(f, block, |
56e93d26 JQ |
836 | offset | RAM_SAVE_FLAG_PAGE); |
837 | if (send_async) { | |
53f09a10 PB |
838 | qemu_put_buffer_async(f, p, TARGET_PAGE_SIZE, |
839 | migrate_release_ram() & | |
840 | migration_in_postcopy(ms)); | |
56e93d26 JQ |
841 | } else { |
842 | qemu_put_buffer(f, p, TARGET_PAGE_SIZE); | |
843 | } | |
072c2511 | 844 | rs->bytes_transferred += TARGET_PAGE_SIZE; |
56e93d26 | 845 | pages = 1; |
b4d1c6e7 | 846 | rs->norm_pages++; |
56e93d26 JQ |
847 | } |
848 | ||
849 | XBZRLE_cache_unlock(); | |
850 | ||
851 | return pages; | |
852 | } | |
853 | ||
a7a9a88f LL |
854 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
855 | ram_addr_t offset) | |
56e93d26 JQ |
856 | { |
857 | int bytes_sent, blen; | |
a7a9a88f | 858 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 | 859 | |
a7a9a88f | 860 | bytes_sent = save_page_header(f, block, offset | |
56e93d26 | 861 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
a7a9a88f | 862 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, |
56e93d26 | 863 | migrate_compress_level()); |
b3be2896 LL |
864 | if (blen < 0) { |
865 | bytes_sent = 0; | |
866 | qemu_file_set_error(migrate_get_current()->to_dst_file, blen); | |
867 | error_report("compressed data failed!"); | |
868 | } else { | |
869 | bytes_sent += blen; | |
53f09a10 PB |
870 | ram_release_pages(migrate_get_current(), block->idstr, |
871 | offset & TARGET_PAGE_MASK, 1); | |
b3be2896 | 872 | } |
56e93d26 JQ |
873 | |
874 | return bytes_sent; | |
875 | } | |
876 | ||
2f4fde93 | 877 | static void flush_compressed_data(RAMState *rs, QEMUFile *f) |
56e93d26 JQ |
878 | { |
879 | int idx, len, thread_count; | |
880 | ||
881 | if (!migrate_use_compression()) { | |
882 | return; | |
883 | } | |
884 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 885 | |
0d9f9a5c | 886 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 887 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 888 | while (!comp_param[idx].done) { |
0d9f9a5c | 889 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 890 | } |
a7a9a88f | 891 | } |
0d9f9a5c | 892 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
893 | |
894 | for (idx = 0; idx < thread_count; idx++) { | |
895 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 896 | if (!comp_param[idx].quit) { |
56e93d26 | 897 | len = qemu_put_qemu_file(f, comp_param[idx].file); |
2f4fde93 | 898 | rs->bytes_transferred += len; |
56e93d26 | 899 | } |
a7a9a88f | 900 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
901 | } |
902 | } | |
903 | ||
904 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
905 | ram_addr_t offset) | |
906 | { | |
907 | param->block = block; | |
908 | param->offset = offset; | |
909 | } | |
910 | ||
b4d1c6e7 | 911 | static int compress_page_with_multi_thread(RAMState *rs, QEMUFile *f, |
072c2511 | 912 | RAMBlock *block, ram_addr_t offset) |
56e93d26 JQ |
913 | { |
914 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
915 | ||
916 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 917 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 JQ |
918 | while (true) { |
919 | for (idx = 0; idx < thread_count; idx++) { | |
920 | if (comp_param[idx].done) { | |
a7a9a88f | 921 | comp_param[idx].done = false; |
56e93d26 | 922 | bytes_xmit = qemu_put_qemu_file(f, comp_param[idx].file); |
a7a9a88f | 923 | qemu_mutex_lock(&comp_param[idx].mutex); |
56e93d26 | 924 | set_compress_params(&comp_param[idx], block, offset); |
a7a9a88f LL |
925 | qemu_cond_signal(&comp_param[idx].cond); |
926 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
56e93d26 | 927 | pages = 1; |
b4d1c6e7 | 928 | rs->norm_pages++; |
072c2511 | 929 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
930 | break; |
931 | } | |
932 | } | |
933 | if (pages > 0) { | |
934 | break; | |
935 | } else { | |
0d9f9a5c | 936 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 JQ |
937 | } |
938 | } | |
0d9f9a5c | 939 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
940 | |
941 | return pages; | |
942 | } | |
943 | ||
944 | /** | |
945 | * ram_save_compressed_page: compress the given page and send it to the stream | |
946 | * | |
3d0684b2 | 947 | * Returns the number of pages written. |
56e93d26 | 948 | * |
6f37bb8b | 949 | * @rs: current RAM state |
3d0684b2 | 950 | * @ms: current migration state |
56e93d26 JQ |
951 | * @f: QEMUFile where to send the data |
952 | * @block: block that contains the page we want to send | |
953 | * @offset: offset inside the block for the page | |
954 | * @last_stage: if we are at the completion stage | |
56e93d26 | 955 | */ |
6f37bb8b JQ |
956 | static int ram_save_compressed_page(RAMState *rs, MigrationState *ms, |
957 | QEMUFile *f, | |
072c2511 | 958 | PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
959 | { |
960 | int pages = -1; | |
fc50438e | 961 | uint64_t bytes_xmit = 0; |
56e93d26 | 962 | uint8_t *p; |
fc50438e | 963 | int ret, blen; |
a08f6890 HZ |
964 | RAMBlock *block = pss->block; |
965 | ram_addr_t offset = pss->offset; | |
56e93d26 | 966 | |
2f68e399 | 967 | p = block->host + offset; |
56e93d26 | 968 | |
56e93d26 JQ |
969 | ret = ram_control_save_page(f, block->offset, |
970 | offset, TARGET_PAGE_SIZE, &bytes_xmit); | |
971 | if (bytes_xmit) { | |
072c2511 | 972 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
973 | pages = 1; |
974 | } | |
56e93d26 JQ |
975 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
976 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
977 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 978 | rs->norm_pages++; |
56e93d26 | 979 | } else if (bytes_xmit == 0) { |
f7ccd61b | 980 | rs->zero_pages++; |
56e93d26 JQ |
981 | } |
982 | } | |
983 | } else { | |
984 | /* When starting the process of a new block, the first page of | |
985 | * the block should be sent out before other pages in the same | |
986 | * block, and all the pages in last block should have been sent | |
987 | * out, keeping this order is important, because the 'cont' flag | |
988 | * is used to avoid resending the block name. | |
989 | */ | |
6f37bb8b | 990 | if (block != rs->last_sent_block) { |
2f4fde93 | 991 | flush_compressed_data(rs, f); |
072c2511 | 992 | pages = save_zero_page(rs, f, block, offset, p); |
56e93d26 | 993 | if (pages == -1) { |
fc50438e LL |
994 | /* Make sure the first page is sent out before other pages */ |
995 | bytes_xmit = save_page_header(f, block, offset | | |
996 | RAM_SAVE_FLAG_COMPRESS_PAGE); | |
997 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, | |
998 | migrate_compress_level()); | |
999 | if (blen > 0) { | |
072c2511 | 1000 | rs->bytes_transferred += bytes_xmit + blen; |
b4d1c6e7 | 1001 | rs->norm_pages++; |
b3be2896 | 1002 | pages = 1; |
fc50438e LL |
1003 | } else { |
1004 | qemu_file_set_error(f, blen); | |
1005 | error_report("compressed data failed!"); | |
b3be2896 | 1006 | } |
56e93d26 | 1007 | } |
53f09a10 PB |
1008 | if (pages > 0) { |
1009 | ram_release_pages(ms, block->idstr, pss->offset, pages); | |
1010 | } | |
56e93d26 | 1011 | } else { |
fc50438e | 1012 | offset |= RAM_SAVE_FLAG_CONTINUE; |
072c2511 | 1013 | pages = save_zero_page(rs, f, block, offset, p); |
56e93d26 | 1014 | if (pages == -1) { |
072c2511 | 1015 | pages = compress_page_with_multi_thread(rs, f, block, offset); |
53f09a10 PB |
1016 | } else { |
1017 | ram_release_pages(ms, block->idstr, pss->offset, pages); | |
56e93d26 JQ |
1018 | } |
1019 | } | |
1020 | } | |
1021 | ||
1022 | return pages; | |
1023 | } | |
1024 | ||
3d0684b2 JQ |
1025 | /** |
1026 | * find_dirty_block: find the next dirty page and update any state | |
1027 | * associated with the search process. | |
b9e60928 | 1028 | * |
3d0684b2 | 1029 | * Returns if a page is found |
b9e60928 | 1030 | * |
6f37bb8b | 1031 | * @rs: current RAM state |
3d0684b2 JQ |
1032 | * @f: QEMUFile where to send the data |
1033 | * @pss: data about the state of the current dirty page scan | |
1034 | * @again: set to false if the search has scanned the whole of RAM | |
1035 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
1036 | * within the global ram_addr space | |
b9e60928 | 1037 | */ |
6f37bb8b | 1038 | static bool find_dirty_block(RAMState *rs, QEMUFile *f, PageSearchStatus *pss, |
f3f491fc | 1039 | bool *again, ram_addr_t *ram_addr_abs) |
b9e60928 | 1040 | { |
6f37bb8b | 1041 | pss->offset = migration_bitmap_find_dirty(rs, pss->block, pss->offset, |
a82d593b | 1042 | ram_addr_abs); |
6f37bb8b JQ |
1043 | if (pss->complete_round && pss->block == rs->last_seen_block && |
1044 | pss->offset >= rs->last_offset) { | |
b9e60928 DDAG |
1045 | /* |
1046 | * We've been once around the RAM and haven't found anything. | |
1047 | * Give up. | |
1048 | */ | |
1049 | *again = false; | |
1050 | return false; | |
1051 | } | |
1052 | if (pss->offset >= pss->block->used_length) { | |
1053 | /* Didn't find anything in this RAM Block */ | |
1054 | pss->offset = 0; | |
1055 | pss->block = QLIST_NEXT_RCU(pss->block, next); | |
1056 | if (!pss->block) { | |
1057 | /* Hit the end of the list */ | |
1058 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1059 | /* Flag that we've looped */ | |
1060 | pss->complete_round = true; | |
6f37bb8b | 1061 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1062 | if (migrate_use_xbzrle()) { |
1063 | /* If xbzrle is on, stop using the data compression at this | |
1064 | * point. In theory, xbzrle can do better than compression. | |
1065 | */ | |
2f4fde93 | 1066 | flush_compressed_data(rs, f); |
b9e60928 DDAG |
1067 | compression_switch = false; |
1068 | } | |
1069 | } | |
1070 | /* Didn't find anything this time, but try again on the new block */ | |
1071 | *again = true; | |
1072 | return false; | |
1073 | } else { | |
1074 | /* Can go around again, but... */ | |
1075 | *again = true; | |
1076 | /* We've found something so probably don't need to */ | |
1077 | return true; | |
1078 | } | |
1079 | } | |
1080 | ||
3d0684b2 JQ |
1081 | /** |
1082 | * unqueue_page: gets a page of the queue | |
1083 | * | |
a82d593b | 1084 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1085 | * |
3d0684b2 JQ |
1086 | * Returns the block of the page (or NULL if none available) |
1087 | * | |
1088 | * @ms: current migration state | |
1089 | * @offset: used to return the offset within the RAMBlock | |
1090 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
1091 | * within the global ram_addr space | |
a82d593b DDAG |
1092 | */ |
1093 | static RAMBlock *unqueue_page(MigrationState *ms, ram_addr_t *offset, | |
1094 | ram_addr_t *ram_addr_abs) | |
1095 | { | |
1096 | RAMBlock *block = NULL; | |
1097 | ||
1098 | qemu_mutex_lock(&ms->src_page_req_mutex); | |
1099 | if (!QSIMPLEQ_EMPTY(&ms->src_page_requests)) { | |
1100 | struct MigrationSrcPageRequest *entry = | |
1101 | QSIMPLEQ_FIRST(&ms->src_page_requests); | |
1102 | block = entry->rb; | |
1103 | *offset = entry->offset; | |
1104 | *ram_addr_abs = (entry->offset + entry->rb->offset) & | |
1105 | TARGET_PAGE_MASK; | |
1106 | ||
1107 | if (entry->len > TARGET_PAGE_SIZE) { | |
1108 | entry->len -= TARGET_PAGE_SIZE; | |
1109 | entry->offset += TARGET_PAGE_SIZE; | |
1110 | } else { | |
1111 | memory_region_unref(block->mr); | |
1112 | QSIMPLEQ_REMOVE_HEAD(&ms->src_page_requests, next_req); | |
1113 | g_free(entry); | |
1114 | } | |
1115 | } | |
1116 | qemu_mutex_unlock(&ms->src_page_req_mutex); | |
1117 | ||
1118 | return block; | |
1119 | } | |
1120 | ||
3d0684b2 JQ |
1121 | /** |
1122 | * get_queued_page: unqueue a page from the postocpy requests | |
1123 | * | |
1124 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1125 | * |
3d0684b2 | 1126 | * Returns if a queued page is found |
a82d593b | 1127 | * |
6f37bb8b | 1128 | * @rs: current RAM state |
3d0684b2 JQ |
1129 | * @ms: current migration state |
1130 | * @pss: data about the state of the current dirty page scan | |
1131 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
1132 | * within the global ram_addr space | |
a82d593b | 1133 | */ |
6f37bb8b JQ |
1134 | static bool get_queued_page(RAMState *rs, MigrationState *ms, |
1135 | PageSearchStatus *pss, | |
a82d593b DDAG |
1136 | ram_addr_t *ram_addr_abs) |
1137 | { | |
1138 | RAMBlock *block; | |
1139 | ram_addr_t offset; | |
1140 | bool dirty; | |
1141 | ||
1142 | do { | |
1143 | block = unqueue_page(ms, &offset, ram_addr_abs); | |
1144 | /* | |
1145 | * We're sending this page, and since it's postcopy nothing else | |
1146 | * will dirty it, and we must make sure it doesn't get sent again | |
1147 | * even if this queue request was received after the background | |
1148 | * search already sent it. | |
1149 | */ | |
1150 | if (block) { | |
1151 | unsigned long *bitmap; | |
eb859c53 | 1152 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
a82d593b DDAG |
1153 | dirty = test_bit(*ram_addr_abs >> TARGET_PAGE_BITS, bitmap); |
1154 | if (!dirty) { | |
1155 | trace_get_queued_page_not_dirty( | |
1156 | block->idstr, (uint64_t)offset, | |
1157 | (uint64_t)*ram_addr_abs, | |
1158 | test_bit(*ram_addr_abs >> TARGET_PAGE_BITS, | |
eb859c53 | 1159 | atomic_rcu_read(&rs->ram_bitmap)->unsentmap)); |
a82d593b DDAG |
1160 | } else { |
1161 | trace_get_queued_page(block->idstr, | |
1162 | (uint64_t)offset, | |
1163 | (uint64_t)*ram_addr_abs); | |
1164 | } | |
1165 | } | |
1166 | ||
1167 | } while (block && !dirty); | |
1168 | ||
1169 | if (block) { | |
1170 | /* | |
1171 | * As soon as we start servicing pages out of order, then we have | |
1172 | * to kill the bulk stage, since the bulk stage assumes | |
1173 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1174 | * dirty, that's no longer true. | |
1175 | */ | |
6f37bb8b | 1176 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1177 | |
1178 | /* | |
1179 | * We want the background search to continue from the queued page | |
1180 | * since the guest is likely to want other pages near to the page | |
1181 | * it just requested. | |
1182 | */ | |
1183 | pss->block = block; | |
1184 | pss->offset = offset; | |
1185 | } | |
1186 | ||
1187 | return !!block; | |
1188 | } | |
1189 | ||
6c595cde | 1190 | /** |
5e58f968 JQ |
1191 | * migration_page_queue_free: drop any remaining pages in the ram |
1192 | * request queue | |
6c595cde | 1193 | * |
3d0684b2 JQ |
1194 | * It should be empty at the end anyway, but in error cases there may |
1195 | * be some left. in case that there is any page left, we drop it. | |
1196 | * | |
1197 | * @ms: current migration state | |
6c595cde | 1198 | */ |
5e58f968 | 1199 | void migration_page_queue_free(MigrationState *ms) |
6c595cde DDAG |
1200 | { |
1201 | struct MigrationSrcPageRequest *mspr, *next_mspr; | |
1202 | /* This queue generally should be empty - but in the case of a failed | |
1203 | * migration might have some droppings in. | |
1204 | */ | |
1205 | rcu_read_lock(); | |
1206 | QSIMPLEQ_FOREACH_SAFE(mspr, &ms->src_page_requests, next_req, next_mspr) { | |
1207 | memory_region_unref(mspr->rb->mr); | |
1208 | QSIMPLEQ_REMOVE_HEAD(&ms->src_page_requests, next_req); | |
1209 | g_free(mspr); | |
1210 | } | |
1211 | rcu_read_unlock(); | |
1212 | } | |
1213 | ||
1214 | /** | |
3d0684b2 JQ |
1215 | * ram_save_queue_pages: queue the page for transmission |
1216 | * | |
1217 | * A request from postcopy destination for example. | |
1218 | * | |
1219 | * Returns zero on success or negative on error | |
1220 | * | |
1221 | * @ms: current migration state | |
1222 | * @rbname: Name of the RAMBLock of the request. NULL means the | |
1223 | * same that last one. | |
1224 | * @start: starting address from the start of the RAMBlock | |
1225 | * @len: length (in bytes) to send | |
6c595cde DDAG |
1226 | */ |
1227 | int ram_save_queue_pages(MigrationState *ms, const char *rbname, | |
1228 | ram_addr_t start, ram_addr_t len) | |
1229 | { | |
1230 | RAMBlock *ramblock; | |
68a098f3 | 1231 | RAMState *rs = &ram_state; |
6c595cde | 1232 | |
d3bf5418 | 1233 | ms->postcopy_requests++; |
6c595cde DDAG |
1234 | rcu_read_lock(); |
1235 | if (!rbname) { | |
1236 | /* Reuse last RAMBlock */ | |
68a098f3 | 1237 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1238 | |
1239 | if (!ramblock) { | |
1240 | /* | |
1241 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1242 | * it's the 1st request. | |
1243 | */ | |
1244 | error_report("ram_save_queue_pages no previous block"); | |
1245 | goto err; | |
1246 | } | |
1247 | } else { | |
1248 | ramblock = qemu_ram_block_by_name(rbname); | |
1249 | ||
1250 | if (!ramblock) { | |
1251 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1252 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1253 | goto err; | |
1254 | } | |
68a098f3 | 1255 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1256 | } |
1257 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1258 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1259 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1260 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1261 | __func__, start, len, ramblock->used_length); |
1262 | goto err; | |
1263 | } | |
1264 | ||
1265 | struct MigrationSrcPageRequest *new_entry = | |
1266 | g_malloc0(sizeof(struct MigrationSrcPageRequest)); | |
1267 | new_entry->rb = ramblock; | |
1268 | new_entry->offset = start; | |
1269 | new_entry->len = len; | |
1270 | ||
1271 | memory_region_ref(ramblock->mr); | |
1272 | qemu_mutex_lock(&ms->src_page_req_mutex); | |
1273 | QSIMPLEQ_INSERT_TAIL(&ms->src_page_requests, new_entry, next_req); | |
1274 | qemu_mutex_unlock(&ms->src_page_req_mutex); | |
1275 | rcu_read_unlock(); | |
1276 | ||
1277 | return 0; | |
1278 | ||
1279 | err: | |
1280 | rcu_read_unlock(); | |
1281 | return -1; | |
1282 | } | |
1283 | ||
a82d593b | 1284 | /** |
3d0684b2 | 1285 | * ram_save_target_page: save one target page |
a82d593b | 1286 | * |
3d0684b2 | 1287 | * Returns the number of pages written |
a82d593b | 1288 | * |
6f37bb8b | 1289 | * @rs: current RAM state |
3d0684b2 | 1290 | * @ms: current migration state |
a82d593b | 1291 | * @f: QEMUFile where to send the data |
3d0684b2 | 1292 | * @pss: data about the page we want to send |
a82d593b | 1293 | * @last_stage: if we are at the completion stage |
3d0684b2 | 1294 | * @dirty_ram_abs: address of the start of the dirty page in ram_addr_t space |
a82d593b | 1295 | */ |
6f37bb8b | 1296 | static int ram_save_target_page(RAMState *rs, MigrationState *ms, QEMUFile *f, |
a08f6890 | 1297 | PageSearchStatus *pss, |
a82d593b | 1298 | bool last_stage, |
a82d593b DDAG |
1299 | ram_addr_t dirty_ram_abs) |
1300 | { | |
1301 | int res = 0; | |
1302 | ||
1303 | /* Check the pages is dirty and if it is send it */ | |
0d8ec885 | 1304 | if (migration_bitmap_clear_dirty(rs, dirty_ram_abs)) { |
a82d593b DDAG |
1305 | unsigned long *unsentmap; |
1306 | if (compression_switch && migrate_use_compression()) { | |
072c2511 | 1307 | res = ram_save_compressed_page(rs, ms, f, pss, last_stage); |
a82d593b | 1308 | } else { |
072c2511 | 1309 | res = ram_save_page(rs, ms, f, pss, last_stage); |
a82d593b DDAG |
1310 | } |
1311 | ||
1312 | if (res < 0) { | |
1313 | return res; | |
1314 | } | |
eb859c53 | 1315 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
a82d593b DDAG |
1316 | if (unsentmap) { |
1317 | clear_bit(dirty_ram_abs >> TARGET_PAGE_BITS, unsentmap); | |
1318 | } | |
3fd3c4b3 DDAG |
1319 | /* Only update last_sent_block if a block was actually sent; xbzrle |
1320 | * might have decided the page was identical so didn't bother writing | |
1321 | * to the stream. | |
1322 | */ | |
1323 | if (res > 0) { | |
6f37bb8b | 1324 | rs->last_sent_block = pss->block; |
3fd3c4b3 | 1325 | } |
a82d593b DDAG |
1326 | } |
1327 | ||
1328 | return res; | |
1329 | } | |
1330 | ||
1331 | /** | |
3d0684b2 | 1332 | * ram_save_host_page: save a whole host page |
a82d593b | 1333 | * |
3d0684b2 JQ |
1334 | * Starting at *offset send pages up to the end of the current host |
1335 | * page. It's valid for the initial offset to point into the middle of | |
1336 | * a host page in which case the remainder of the hostpage is sent. | |
1337 | * Only dirty target pages are sent. Note that the host page size may | |
1338 | * be a huge page for this block. | |
a82d593b | 1339 | * |
3d0684b2 JQ |
1340 | * Returns the number of pages written or negative on error |
1341 | * | |
6f37bb8b | 1342 | * @rs: current RAM state |
3d0684b2 | 1343 | * @ms: current migration state |
a82d593b | 1344 | * @f: QEMUFile where to send the data |
3d0684b2 | 1345 | * @pss: data about the page we want to send |
a82d593b | 1346 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1347 | * @dirty_ram_abs: Address of the start of the dirty page in ram_addr_t space |
1348 | */ | |
6f37bb8b | 1349 | static int ram_save_host_page(RAMState *rs, MigrationState *ms, QEMUFile *f, |
a08f6890 HZ |
1350 | PageSearchStatus *pss, |
1351 | bool last_stage, | |
a82d593b DDAG |
1352 | ram_addr_t dirty_ram_abs) |
1353 | { | |
1354 | int tmppages, pages = 0; | |
4c011c37 DDAG |
1355 | size_t pagesize = qemu_ram_pagesize(pss->block); |
1356 | ||
a82d593b | 1357 | do { |
6f37bb8b | 1358 | tmppages = ram_save_target_page(rs, ms, f, pss, last_stage, |
072c2511 | 1359 | dirty_ram_abs); |
a82d593b DDAG |
1360 | if (tmppages < 0) { |
1361 | return tmppages; | |
1362 | } | |
1363 | ||
1364 | pages += tmppages; | |
a08f6890 | 1365 | pss->offset += TARGET_PAGE_SIZE; |
a82d593b | 1366 | dirty_ram_abs += TARGET_PAGE_SIZE; |
4c011c37 | 1367 | } while (pss->offset & (pagesize - 1)); |
a82d593b DDAG |
1368 | |
1369 | /* The offset we leave with is the last one we looked at */ | |
a08f6890 | 1370 | pss->offset -= TARGET_PAGE_SIZE; |
a82d593b DDAG |
1371 | return pages; |
1372 | } | |
6c595cde | 1373 | |
56e93d26 | 1374 | /** |
3d0684b2 | 1375 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1376 | * |
1377 | * Called within an RCU critical section. | |
1378 | * | |
3d0684b2 | 1379 | * Returns the number of pages written where zero means no dirty pages |
56e93d26 | 1380 | * |
6f37bb8b | 1381 | * @rs: current RAM state |
56e93d26 JQ |
1382 | * @f: QEMUFile where to send the data |
1383 | * @last_stage: if we are at the completion stage | |
a82d593b DDAG |
1384 | * |
1385 | * On systems where host-page-size > target-page-size it will send all the | |
1386 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1387 | */ |
1388 | ||
072c2511 | 1389 | static int ram_find_and_save_block(RAMState *rs, QEMUFile *f, bool last_stage) |
56e93d26 | 1390 | { |
b8fb8cb7 | 1391 | PageSearchStatus pss; |
a82d593b | 1392 | MigrationState *ms = migrate_get_current(); |
56e93d26 | 1393 | int pages = 0; |
b9e60928 | 1394 | bool again, found; |
f3f491fc DDAG |
1395 | ram_addr_t dirty_ram_abs; /* Address of the start of the dirty page in |
1396 | ram_addr_t space */ | |
56e93d26 | 1397 | |
0827b9e9 AA |
1398 | /* No dirty page as there is zero RAM */ |
1399 | if (!ram_bytes_total()) { | |
1400 | return pages; | |
1401 | } | |
1402 | ||
6f37bb8b JQ |
1403 | pss.block = rs->last_seen_block; |
1404 | pss.offset = rs->last_offset; | |
b8fb8cb7 DDAG |
1405 | pss.complete_round = false; |
1406 | ||
1407 | if (!pss.block) { | |
1408 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1409 | } | |
56e93d26 | 1410 | |
b9e60928 | 1411 | do { |
a82d593b | 1412 | again = true; |
6f37bb8b | 1413 | found = get_queued_page(rs, ms, &pss, &dirty_ram_abs); |
b9e60928 | 1414 | |
a82d593b DDAG |
1415 | if (!found) { |
1416 | /* priority queue empty, so just search for something dirty */ | |
6f37bb8b | 1417 | found = find_dirty_block(rs, f, &pss, &again, &dirty_ram_abs); |
a82d593b | 1418 | } |
f3f491fc | 1419 | |
a82d593b | 1420 | if (found) { |
072c2511 | 1421 | pages = ram_save_host_page(rs, ms, f, &pss, last_stage, |
a82d593b | 1422 | dirty_ram_abs); |
56e93d26 | 1423 | } |
b9e60928 | 1424 | } while (!pages && again); |
56e93d26 | 1425 | |
6f37bb8b JQ |
1426 | rs->last_seen_block = pss.block; |
1427 | rs->last_offset = pss.offset; | |
56e93d26 JQ |
1428 | |
1429 | return pages; | |
1430 | } | |
1431 | ||
1432 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1433 | { | |
1434 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b JQ |
1435 | RAMState *rs = &ram_state; |
1436 | ||
56e93d26 | 1437 | if (zero) { |
f7ccd61b | 1438 | rs->zero_pages += pages; |
56e93d26 | 1439 | } else { |
b4d1c6e7 | 1440 | rs->norm_pages += pages; |
2f4fde93 | 1441 | rs->bytes_transferred += size; |
56e93d26 JQ |
1442 | qemu_update_position(f, size); |
1443 | } | |
1444 | } | |
1445 | ||
56e93d26 JQ |
1446 | uint64_t ram_bytes_total(void) |
1447 | { | |
1448 | RAMBlock *block; | |
1449 | uint64_t total = 0; | |
1450 | ||
1451 | rcu_read_lock(); | |
1452 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) | |
1453 | total += block->used_length; | |
1454 | rcu_read_unlock(); | |
1455 | return total; | |
1456 | } | |
1457 | ||
1458 | void free_xbzrle_decoded_buf(void) | |
1459 | { | |
1460 | g_free(xbzrle_decoded_buf); | |
1461 | xbzrle_decoded_buf = NULL; | |
1462 | } | |
1463 | ||
eb859c53 | 1464 | static void migration_bitmap_free(struct RAMBitmap *bmap) |
60be6340 DL |
1465 | { |
1466 | g_free(bmap->bmap); | |
f3f491fc | 1467 | g_free(bmap->unsentmap); |
60be6340 DL |
1468 | g_free(bmap); |
1469 | } | |
1470 | ||
6ad2a215 | 1471 | static void ram_migration_cleanup(void *opaque) |
56e93d26 | 1472 | { |
eb859c53 JQ |
1473 | RAMState *rs = opaque; |
1474 | ||
2ff64038 LZ |
1475 | /* caller have hold iothread lock or is in a bh, so there is |
1476 | * no writing race against this migration_bitmap | |
1477 | */ | |
eb859c53 JQ |
1478 | struct RAMBitmap *bitmap = rs->ram_bitmap; |
1479 | atomic_rcu_set(&rs->ram_bitmap, NULL); | |
2ff64038 | 1480 | if (bitmap) { |
56e93d26 | 1481 | memory_global_dirty_log_stop(); |
60be6340 | 1482 | call_rcu(bitmap, migration_bitmap_free, rcu); |
56e93d26 JQ |
1483 | } |
1484 | ||
1485 | XBZRLE_cache_lock(); | |
1486 | if (XBZRLE.cache) { | |
1487 | cache_fini(XBZRLE.cache); | |
1488 | g_free(XBZRLE.encoded_buf); | |
1489 | g_free(XBZRLE.current_buf); | |
adb65dec | 1490 | g_free(ZERO_TARGET_PAGE); |
56e93d26 JQ |
1491 | XBZRLE.cache = NULL; |
1492 | XBZRLE.encoded_buf = NULL; | |
1493 | XBZRLE.current_buf = NULL; | |
1494 | } | |
1495 | XBZRLE_cache_unlock(); | |
1496 | } | |
1497 | ||
6f37bb8b | 1498 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1499 | { |
6f37bb8b JQ |
1500 | rs->last_seen_block = NULL; |
1501 | rs->last_sent_block = NULL; | |
1502 | rs->last_offset = 0; | |
1503 | rs->last_version = ram_list.version; | |
1504 | rs->ram_bulk_stage = true; | |
56e93d26 JQ |
1505 | } |
1506 | ||
1507 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1508 | ||
dd631697 LZ |
1509 | void migration_bitmap_extend(ram_addr_t old, ram_addr_t new) |
1510 | { | |
0d8ec885 | 1511 | RAMState *rs = &ram_state; |
108cfae0 | 1512 | |
dd631697 LZ |
1513 | /* called in qemu main thread, so there is |
1514 | * no writing race against this migration_bitmap | |
1515 | */ | |
eb859c53 JQ |
1516 | if (rs->ram_bitmap) { |
1517 | struct RAMBitmap *old_bitmap = rs->ram_bitmap, *bitmap; | |
1518 | bitmap = g_new(struct RAMBitmap, 1); | |
60be6340 | 1519 | bitmap->bmap = bitmap_new(new); |
dd631697 LZ |
1520 | |
1521 | /* prevent migration_bitmap content from being set bit | |
1522 | * by migration_bitmap_sync_range() at the same time. | |
1523 | * it is safe to migration if migration_bitmap is cleared bit | |
1524 | * at the same time. | |
1525 | */ | |
108cfae0 | 1526 | qemu_mutex_lock(&rs->bitmap_mutex); |
60be6340 DL |
1527 | bitmap_copy(bitmap->bmap, old_bitmap->bmap, old); |
1528 | bitmap_set(bitmap->bmap, old, new - old); | |
f3f491fc DDAG |
1529 | |
1530 | /* We don't have a way to safely extend the sentmap | |
1531 | * with RCU; so mark it as missing, entry to postcopy | |
1532 | * will fail. | |
1533 | */ | |
1534 | bitmap->unsentmap = NULL; | |
1535 | ||
eb859c53 | 1536 | atomic_rcu_set(&rs->ram_bitmap, bitmap); |
108cfae0 | 1537 | qemu_mutex_unlock(&rs->bitmap_mutex); |
0d8ec885 | 1538 | rs->migration_dirty_pages += new - old; |
60be6340 | 1539 | call_rcu(old_bitmap, migration_bitmap_free, rcu); |
dd631697 LZ |
1540 | } |
1541 | } | |
56e93d26 | 1542 | |
4f2e4252 DDAG |
1543 | /* |
1544 | * 'expected' is the value you expect the bitmap mostly to be full | |
1545 | * of; it won't bother printing lines that are all this value. | |
1546 | * If 'todump' is null the migration bitmap is dumped. | |
1547 | */ | |
1548 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected) | |
1549 | { | |
1550 | int64_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
eb859c53 | 1551 | RAMState *rs = &ram_state; |
4f2e4252 DDAG |
1552 | int64_t cur; |
1553 | int64_t linelen = 128; | |
1554 | char linebuf[129]; | |
1555 | ||
1556 | if (!todump) { | |
eb859c53 | 1557 | todump = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
4f2e4252 DDAG |
1558 | } |
1559 | ||
1560 | for (cur = 0; cur < ram_pages; cur += linelen) { | |
1561 | int64_t curb; | |
1562 | bool found = false; | |
1563 | /* | |
1564 | * Last line; catch the case where the line length | |
1565 | * is longer than remaining ram | |
1566 | */ | |
1567 | if (cur + linelen > ram_pages) { | |
1568 | linelen = ram_pages - cur; | |
1569 | } | |
1570 | for (curb = 0; curb < linelen; curb++) { | |
1571 | bool thisbit = test_bit(cur + curb, todump); | |
1572 | linebuf[curb] = thisbit ? '1' : '.'; | |
1573 | found = found || (thisbit != expected); | |
1574 | } | |
1575 | if (found) { | |
1576 | linebuf[curb] = '\0'; | |
1577 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1578 | } | |
1579 | } | |
1580 | } | |
1581 | ||
e0b266f0 DDAG |
1582 | /* **** functions for postcopy ***** */ |
1583 | ||
ced1c616 PB |
1584 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1585 | { | |
eb859c53 | 1586 | RAMState *rs = &ram_state; |
ced1c616 | 1587 | struct RAMBlock *block; |
eb859c53 | 1588 | unsigned long *bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
ced1c616 PB |
1589 | |
1590 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1591 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1592 | unsigned long range = first + (block->used_length >> TARGET_PAGE_BITS); | |
1593 | unsigned long run_start = find_next_zero_bit(bitmap, range, first); | |
1594 | ||
1595 | while (run_start < range) { | |
1596 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
1597 | ram_discard_range(NULL, block->idstr, run_start << TARGET_PAGE_BITS, | |
1598 | (run_end - run_start) << TARGET_PAGE_BITS); | |
1599 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); | |
1600 | } | |
1601 | } | |
1602 | } | |
1603 | ||
3d0684b2 JQ |
1604 | /** |
1605 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1606 | * | |
1607 | * Returns zero on success | |
1608 | * | |
e0b266f0 DDAG |
1609 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
1610 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1611 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
3d0684b2 JQ |
1612 | * |
1613 | * @ms: current migration state | |
1614 | * @pds: state for postcopy | |
1615 | * @start: RAMBlock starting page | |
1616 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1617 | */ |
1618 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1619 | PostcopyDiscardState *pds, | |
1620 | unsigned long start, | |
1621 | unsigned long length) | |
1622 | { | |
eb859c53 | 1623 | RAMState *rs = &ram_state; |
e0b266f0 DDAG |
1624 | unsigned long end = start + length; /* one after the end */ |
1625 | unsigned long current; | |
1626 | unsigned long *unsentmap; | |
1627 | ||
eb859c53 | 1628 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
e0b266f0 DDAG |
1629 | for (current = start; current < end; ) { |
1630 | unsigned long one = find_next_bit(unsentmap, end, current); | |
1631 | ||
1632 | if (one <= end) { | |
1633 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1634 | unsigned long discard_length; | |
1635 | ||
1636 | if (zero >= end) { | |
1637 | discard_length = end - one; | |
1638 | } else { | |
1639 | discard_length = zero - one; | |
1640 | } | |
d688c62d DDAG |
1641 | if (discard_length) { |
1642 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1643 | } | |
e0b266f0 DDAG |
1644 | current = one + discard_length; |
1645 | } else { | |
1646 | current = one; | |
1647 | } | |
1648 | } | |
1649 | ||
1650 | return 0; | |
1651 | } | |
1652 | ||
3d0684b2 JQ |
1653 | /** |
1654 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1655 | * | |
1656 | * Returns 0 for success or negative for error | |
1657 | * | |
e0b266f0 DDAG |
1658 | * Utility for the outgoing postcopy code. |
1659 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1660 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1661 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1662 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1663 | * |
1664 | * @ms: current migration state | |
e0b266f0 DDAG |
1665 | */ |
1666 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1667 | { | |
1668 | struct RAMBlock *block; | |
1669 | int ret; | |
1670 | ||
1671 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1672 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1673 | PostcopyDiscardState *pds = postcopy_discard_send_init(ms, | |
1674 | first, | |
1675 | block->idstr); | |
1676 | ||
1677 | /* | |
1678 | * Postcopy sends chunks of bitmap over the wire, but it | |
1679 | * just needs indexes at this point, avoids it having | |
1680 | * target page specific code. | |
1681 | */ | |
1682 | ret = postcopy_send_discard_bm_ram(ms, pds, first, | |
1683 | block->used_length >> TARGET_PAGE_BITS); | |
1684 | postcopy_discard_send_finish(ms, pds); | |
1685 | if (ret) { | |
1686 | return ret; | |
1687 | } | |
1688 | } | |
1689 | ||
1690 | return 0; | |
1691 | } | |
1692 | ||
3d0684b2 JQ |
1693 | /** |
1694 | * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages | |
1695 | * | |
1696 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
1697 | * canonicalize the two bitmaps, that are similar, but one is | |
1698 | * inverted. | |
99e314eb | 1699 | * |
3d0684b2 JQ |
1700 | * Postcopy requires that all target pages in a hostpage are dirty or |
1701 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 1702 | * |
3d0684b2 JQ |
1703 | * @ms: current migration state |
1704 | * @unsent_pass: if true we need to canonicalize partially unsent host pages | |
1705 | * otherwise we need to canonicalize partially dirty host pages | |
1706 | * @block: block that contains the page we want to canonicalize | |
1707 | * @pds: state for postcopy | |
99e314eb DDAG |
1708 | */ |
1709 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1710 | RAMBlock *block, | |
1711 | PostcopyDiscardState *pds) | |
1712 | { | |
0d8ec885 | 1713 | RAMState *rs = &ram_state; |
99e314eb DDAG |
1714 | unsigned long *bitmap; |
1715 | unsigned long *unsentmap; | |
29c59172 | 1716 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
99e314eb DDAG |
1717 | unsigned long first = block->offset >> TARGET_PAGE_BITS; |
1718 | unsigned long len = block->used_length >> TARGET_PAGE_BITS; | |
1719 | unsigned long last = first + (len - 1); | |
1720 | unsigned long run_start; | |
1721 | ||
29c59172 DDAG |
1722 | if (block->page_size == TARGET_PAGE_SIZE) { |
1723 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
1724 | return; | |
1725 | } | |
1726 | ||
eb859c53 JQ |
1727 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
1728 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; | |
99e314eb DDAG |
1729 | |
1730 | if (unsent_pass) { | |
1731 | /* Find a sent page */ | |
1732 | run_start = find_next_zero_bit(unsentmap, last + 1, first); | |
1733 | } else { | |
1734 | /* Find a dirty page */ | |
1735 | run_start = find_next_bit(bitmap, last + 1, first); | |
1736 | } | |
1737 | ||
1738 | while (run_start <= last) { | |
1739 | bool do_fixup = false; | |
1740 | unsigned long fixup_start_addr; | |
1741 | unsigned long host_offset; | |
1742 | ||
1743 | /* | |
1744 | * If the start of this run of pages is in the middle of a host | |
1745 | * page, then we need to fixup this host page. | |
1746 | */ | |
1747 | host_offset = run_start % host_ratio; | |
1748 | if (host_offset) { | |
1749 | do_fixup = true; | |
1750 | run_start -= host_offset; | |
1751 | fixup_start_addr = run_start; | |
1752 | /* For the next pass */ | |
1753 | run_start = run_start + host_ratio; | |
1754 | } else { | |
1755 | /* Find the end of this run */ | |
1756 | unsigned long run_end; | |
1757 | if (unsent_pass) { | |
1758 | run_end = find_next_bit(unsentmap, last + 1, run_start + 1); | |
1759 | } else { | |
1760 | run_end = find_next_zero_bit(bitmap, last + 1, run_start + 1); | |
1761 | } | |
1762 | /* | |
1763 | * If the end isn't at the start of a host page, then the | |
1764 | * run doesn't finish at the end of a host page | |
1765 | * and we need to discard. | |
1766 | */ | |
1767 | host_offset = run_end % host_ratio; | |
1768 | if (host_offset) { | |
1769 | do_fixup = true; | |
1770 | fixup_start_addr = run_end - host_offset; | |
1771 | /* | |
1772 | * This host page has gone, the next loop iteration starts | |
1773 | * from after the fixup | |
1774 | */ | |
1775 | run_start = fixup_start_addr + host_ratio; | |
1776 | } else { | |
1777 | /* | |
1778 | * No discards on this iteration, next loop starts from | |
1779 | * next sent/dirty page | |
1780 | */ | |
1781 | run_start = run_end + 1; | |
1782 | } | |
1783 | } | |
1784 | ||
1785 | if (do_fixup) { | |
1786 | unsigned long page; | |
1787 | ||
1788 | /* Tell the destination to discard this page */ | |
1789 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1790 | /* For the unsent_pass we: | |
1791 | * discard partially sent pages | |
1792 | * For the !unsent_pass (dirty) we: | |
1793 | * discard partially dirty pages that were sent | |
1794 | * (any partially sent pages were already discarded | |
1795 | * by the previous unsent_pass) | |
1796 | */ | |
1797 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1798 | host_ratio); | |
1799 | } | |
1800 | ||
1801 | /* Clean up the bitmap */ | |
1802 | for (page = fixup_start_addr; | |
1803 | page < fixup_start_addr + host_ratio; page++) { | |
1804 | /* All pages in this host page are now not sent */ | |
1805 | set_bit(page, unsentmap); | |
1806 | ||
1807 | /* | |
1808 | * Remark them as dirty, updating the count for any pages | |
1809 | * that weren't previously dirty. | |
1810 | */ | |
0d8ec885 | 1811 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
1812 | } |
1813 | } | |
1814 | ||
1815 | if (unsent_pass) { | |
1816 | /* Find the next sent page for the next iteration */ | |
1817 | run_start = find_next_zero_bit(unsentmap, last + 1, | |
1818 | run_start); | |
1819 | } else { | |
1820 | /* Find the next dirty page for the next iteration */ | |
1821 | run_start = find_next_bit(bitmap, last + 1, run_start); | |
1822 | } | |
1823 | } | |
1824 | } | |
1825 | ||
3d0684b2 JQ |
1826 | /** |
1827 | * postcopy_chuck_hostpages: discrad any partially sent host page | |
1828 | * | |
99e314eb DDAG |
1829 | * Utility for the outgoing postcopy code. |
1830 | * | |
1831 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
1832 | * dirty host-page size chunks as all dirty. In this case the host-page |
1833 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 1834 | * |
3d0684b2 JQ |
1835 | * Returns zero on success |
1836 | * | |
1837 | * @ms: current migration state | |
99e314eb DDAG |
1838 | */ |
1839 | static int postcopy_chunk_hostpages(MigrationState *ms) | |
1840 | { | |
6f37bb8b | 1841 | RAMState *rs = &ram_state; |
99e314eb DDAG |
1842 | struct RAMBlock *block; |
1843 | ||
99e314eb | 1844 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
6f37bb8b JQ |
1845 | rs->last_seen_block = NULL; |
1846 | rs->last_sent_block = NULL; | |
1847 | rs->last_offset = 0; | |
99e314eb DDAG |
1848 | |
1849 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1850 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1851 | ||
1852 | PostcopyDiscardState *pds = | |
1853 | postcopy_discard_send_init(ms, first, block->idstr); | |
1854 | ||
1855 | /* First pass: Discard all partially sent host pages */ | |
1856 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1857 | /* | |
1858 | * Second pass: Ensure that all partially dirty host pages are made | |
1859 | * fully dirty. | |
1860 | */ | |
1861 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
1862 | ||
1863 | postcopy_discard_send_finish(ms, pds); | |
1864 | } /* ram_list loop */ | |
1865 | ||
1866 | return 0; | |
1867 | } | |
1868 | ||
3d0684b2 JQ |
1869 | /** |
1870 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
1871 | * | |
1872 | * Returns zero on success | |
1873 | * | |
e0b266f0 DDAG |
1874 | * Transmit the set of pages to be discarded after precopy to the target |
1875 | * these are pages that: | |
1876 | * a) Have been previously transmitted but are now dirty again | |
1877 | * b) Pages that have never been transmitted, this ensures that | |
1878 | * any pages on the destination that have been mapped by background | |
1879 | * tasks get discarded (transparent huge pages is the specific concern) | |
1880 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
1881 | * |
1882 | * @ms: current migration state | |
e0b266f0 DDAG |
1883 | */ |
1884 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1885 | { | |
eb859c53 | 1886 | RAMState *rs = &ram_state; |
e0b266f0 DDAG |
1887 | int ret; |
1888 | unsigned long *bitmap, *unsentmap; | |
1889 | ||
1890 | rcu_read_lock(); | |
1891 | ||
1892 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 1893 | migration_bitmap_sync(rs); |
e0b266f0 | 1894 | |
eb859c53 | 1895 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
e0b266f0 DDAG |
1896 | if (!unsentmap) { |
1897 | /* We don't have a safe way to resize the sentmap, so | |
1898 | * if the bitmap was resized it will be NULL at this | |
1899 | * point. | |
1900 | */ | |
1901 | error_report("migration ram resized during precopy phase"); | |
1902 | rcu_read_unlock(); | |
1903 | return -EINVAL; | |
1904 | } | |
1905 | ||
29c59172 | 1906 | /* Deal with TPS != HPS and huge pages */ |
99e314eb DDAG |
1907 | ret = postcopy_chunk_hostpages(ms); |
1908 | if (ret) { | |
1909 | rcu_read_unlock(); | |
1910 | return ret; | |
1911 | } | |
1912 | ||
e0b266f0 DDAG |
1913 | /* |
1914 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1915 | */ | |
eb859c53 | 1916 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
e0b266f0 DDAG |
1917 | bitmap_or(unsentmap, unsentmap, bitmap, |
1918 | last_ram_offset() >> TARGET_PAGE_BITS); | |
1919 | ||
1920 | ||
1921 | trace_ram_postcopy_send_discard_bitmap(); | |
1922 | #ifdef DEBUG_POSTCOPY | |
1923 | ram_debug_dump_bitmap(unsentmap, true); | |
1924 | #endif | |
1925 | ||
1926 | ret = postcopy_each_ram_send_discard(ms); | |
1927 | rcu_read_unlock(); | |
1928 | ||
1929 | return ret; | |
1930 | } | |
1931 | ||
3d0684b2 JQ |
1932 | /** |
1933 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 1934 | * |
3d0684b2 | 1935 | * Returns zero on success |
e0b266f0 | 1936 | * |
3d0684b2 | 1937 | * @mis: current migration incoming state |
36449157 JQ |
1938 | * @rbname: name of the RAMBlock of the request. NULL means the |
1939 | * same that last one. | |
3d0684b2 JQ |
1940 | * @start: RAMBlock starting page |
1941 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1942 | */ |
1943 | int ram_discard_range(MigrationIncomingState *mis, | |
36449157 | 1944 | const char *rbname, |
e0b266f0 DDAG |
1945 | uint64_t start, size_t length) |
1946 | { | |
1947 | int ret = -1; | |
1948 | ||
36449157 | 1949 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 1950 | |
e0b266f0 | 1951 | rcu_read_lock(); |
36449157 | 1952 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
1953 | |
1954 | if (!rb) { | |
36449157 | 1955 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
e0b266f0 DDAG |
1956 | goto err; |
1957 | } | |
1958 | ||
d3a5038c | 1959 | ret = ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
1960 | |
1961 | err: | |
1962 | rcu_read_unlock(); | |
1963 | ||
1964 | return ret; | |
1965 | } | |
1966 | ||
ceb4d168 | 1967 | static int ram_state_init(RAMState *rs) |
56e93d26 | 1968 | { |
56e93d26 JQ |
1969 | int64_t ram_bitmap_pages; /* Size of bitmap in pages, including gaps */ |
1970 | ||
ceb4d168 | 1971 | memset(rs, 0, sizeof(*rs)); |
108cfae0 | 1972 | qemu_mutex_init(&rs->bitmap_mutex); |
56e93d26 JQ |
1973 | |
1974 | if (migrate_use_xbzrle()) { | |
1975 | XBZRLE_cache_lock(); | |
adb65dec | 1976 | ZERO_TARGET_PAGE = g_malloc0(TARGET_PAGE_SIZE); |
56e93d26 JQ |
1977 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() / |
1978 | TARGET_PAGE_SIZE, | |
1979 | TARGET_PAGE_SIZE); | |
1980 | if (!XBZRLE.cache) { | |
1981 | XBZRLE_cache_unlock(); | |
1982 | error_report("Error creating cache"); | |
1983 | return -1; | |
1984 | } | |
1985 | XBZRLE_cache_unlock(); | |
1986 | ||
1987 | /* We prefer not to abort if there is no memory */ | |
1988 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
1989 | if (!XBZRLE.encoded_buf) { | |
1990 | error_report("Error allocating encoded_buf"); | |
1991 | return -1; | |
1992 | } | |
1993 | ||
1994 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
1995 | if (!XBZRLE.current_buf) { | |
1996 | error_report("Error allocating current_buf"); | |
1997 | g_free(XBZRLE.encoded_buf); | |
1998 | XBZRLE.encoded_buf = NULL; | |
1999 | return -1; | |
2000 | } | |
56e93d26 JQ |
2001 | } |
2002 | ||
49877834 PB |
2003 | /* For memory_global_dirty_log_start below. */ |
2004 | qemu_mutex_lock_iothread(); | |
2005 | ||
56e93d26 JQ |
2006 | qemu_mutex_lock_ramlist(); |
2007 | rcu_read_lock(); | |
6f37bb8b | 2008 | ram_state_reset(rs); |
56e93d26 | 2009 | |
eb859c53 | 2010 | rs->ram_bitmap = g_new0(struct RAMBitmap, 1); |
0827b9e9 AA |
2011 | /* Skip setting bitmap if there is no RAM */ |
2012 | if (ram_bytes_total()) { | |
2013 | ram_bitmap_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
eb859c53 JQ |
2014 | rs->ram_bitmap->bmap = bitmap_new(ram_bitmap_pages); |
2015 | bitmap_set(rs->ram_bitmap->bmap, 0, ram_bitmap_pages); | |
0827b9e9 AA |
2016 | |
2017 | if (migrate_postcopy_ram()) { | |
eb859c53 JQ |
2018 | rs->ram_bitmap->unsentmap = bitmap_new(ram_bitmap_pages); |
2019 | bitmap_set(rs->ram_bitmap->unsentmap, 0, ram_bitmap_pages); | |
0827b9e9 | 2020 | } |
f3f491fc DDAG |
2021 | } |
2022 | ||
56e93d26 JQ |
2023 | /* |
2024 | * Count the total number of pages used by ram blocks not including any | |
2025 | * gaps due to alignment or unplugs. | |
2026 | */ | |
0d8ec885 | 2027 | rs->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
56e93d26 JQ |
2028 | |
2029 | memory_global_dirty_log_start(); | |
8d820d6f | 2030 | migration_bitmap_sync(rs); |
56e93d26 | 2031 | qemu_mutex_unlock_ramlist(); |
49877834 | 2032 | qemu_mutex_unlock_iothread(); |
a91246c9 HZ |
2033 | rcu_read_unlock(); |
2034 | ||
2035 | return 0; | |
2036 | } | |
2037 | ||
3d0684b2 JQ |
2038 | /* |
2039 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2040 | * long-running RCU critical section. When rcu-reclaims in the code |
2041 | * start to become numerous it will be necessary to reduce the | |
2042 | * granularity of these critical sections. | |
2043 | */ | |
2044 | ||
3d0684b2 JQ |
2045 | /** |
2046 | * ram_save_setup: Setup RAM for migration | |
2047 | * | |
2048 | * Returns zero to indicate success and negative for error | |
2049 | * | |
2050 | * @f: QEMUFile where to send the data | |
2051 | * @opaque: RAMState pointer | |
2052 | */ | |
a91246c9 HZ |
2053 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2054 | { | |
6f37bb8b | 2055 | RAMState *rs = opaque; |
a91246c9 HZ |
2056 | RAMBlock *block; |
2057 | ||
2058 | /* migration has already setup the bitmap, reuse it. */ | |
2059 | if (!migration_in_colo_state()) { | |
ceb4d168 | 2060 | if (ram_state_init(rs) < 0) { |
a91246c9 HZ |
2061 | return -1; |
2062 | } | |
2063 | } | |
2064 | ||
2065 | rcu_read_lock(); | |
56e93d26 JQ |
2066 | |
2067 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
2068 | ||
2069 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
2070 | qemu_put_byte(f, strlen(block->idstr)); | |
2071 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2072 | qemu_put_be64(f, block->used_length); | |
ef08fb38 DDAG |
2073 | if (migrate_postcopy_ram() && block->page_size != qemu_host_page_size) { |
2074 | qemu_put_be64(f, block->page_size); | |
2075 | } | |
56e93d26 JQ |
2076 | } |
2077 | ||
2078 | rcu_read_unlock(); | |
2079 | ||
2080 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
2081 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2082 | ||
2083 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2084 | ||
2085 | return 0; | |
2086 | } | |
2087 | ||
3d0684b2 JQ |
2088 | /** |
2089 | * ram_save_iterate: iterative stage for migration | |
2090 | * | |
2091 | * Returns zero to indicate success and negative for error | |
2092 | * | |
2093 | * @f: QEMUFile where to send the data | |
2094 | * @opaque: RAMState pointer | |
2095 | */ | |
56e93d26 JQ |
2096 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2097 | { | |
6f37bb8b | 2098 | RAMState *rs = opaque; |
56e93d26 JQ |
2099 | int ret; |
2100 | int i; | |
2101 | int64_t t0; | |
5c90308f | 2102 | int done = 0; |
56e93d26 JQ |
2103 | |
2104 | rcu_read_lock(); | |
6f37bb8b JQ |
2105 | if (ram_list.version != rs->last_version) { |
2106 | ram_state_reset(rs); | |
56e93d26 JQ |
2107 | } |
2108 | ||
2109 | /* Read version before ram_list.blocks */ | |
2110 | smp_rmb(); | |
2111 | ||
2112 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
2113 | ||
2114 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
2115 | i = 0; | |
2116 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
2117 | int pages; | |
2118 | ||
072c2511 | 2119 | pages = ram_find_and_save_block(rs, f, false); |
56e93d26 JQ |
2120 | /* no more pages to sent */ |
2121 | if (pages == 0) { | |
5c90308f | 2122 | done = 1; |
56e93d26 JQ |
2123 | break; |
2124 | } | |
23b28c3c | 2125 | rs->iterations++; |
070afca2 | 2126 | |
56e93d26 JQ |
2127 | /* we want to check in the 1st loop, just in case it was the 1st time |
2128 | and we had to sync the dirty bitmap. | |
2129 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2130 | iterations | |
2131 | */ | |
2132 | if ((i & 63) == 0) { | |
2133 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2134 | if (t1 > MAX_WAIT) { | |
55c4446b | 2135 | trace_ram_save_iterate_big_wait(t1, i); |
56e93d26 JQ |
2136 | break; |
2137 | } | |
2138 | } | |
2139 | i++; | |
2140 | } | |
2f4fde93 | 2141 | flush_compressed_data(rs, f); |
56e93d26 JQ |
2142 | rcu_read_unlock(); |
2143 | ||
2144 | /* | |
2145 | * Must occur before EOS (or any QEMUFile operation) | |
2146 | * because of RDMA protocol. | |
2147 | */ | |
2148 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2149 | ||
2150 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2f4fde93 | 2151 | rs->bytes_transferred += 8; |
56e93d26 JQ |
2152 | |
2153 | ret = qemu_file_get_error(f); | |
2154 | if (ret < 0) { | |
2155 | return ret; | |
2156 | } | |
2157 | ||
5c90308f | 2158 | return done; |
56e93d26 JQ |
2159 | } |
2160 | ||
3d0684b2 JQ |
2161 | /** |
2162 | * ram_save_complete: function called to send the remaining amount of ram | |
2163 | * | |
2164 | * Returns zero to indicate success | |
2165 | * | |
2166 | * Called with iothread lock | |
2167 | * | |
2168 | * @f: QEMUFile where to send the data | |
2169 | * @opaque: RAMState pointer | |
2170 | */ | |
56e93d26 JQ |
2171 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2172 | { | |
6f37bb8b JQ |
2173 | RAMState *rs = opaque; |
2174 | ||
56e93d26 JQ |
2175 | rcu_read_lock(); |
2176 | ||
663e6c1d | 2177 | if (!migration_in_postcopy(migrate_get_current())) { |
8d820d6f | 2178 | migration_bitmap_sync(rs); |
663e6c1d | 2179 | } |
56e93d26 JQ |
2180 | |
2181 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2182 | ||
2183 | /* try transferring iterative blocks of memory */ | |
2184 | ||
2185 | /* flush all remaining blocks regardless of rate limiting */ | |
2186 | while (true) { | |
2187 | int pages; | |
2188 | ||
072c2511 | 2189 | pages = ram_find_and_save_block(rs, f, !migration_in_colo_state()); |
56e93d26 JQ |
2190 | /* no more blocks to sent */ |
2191 | if (pages == 0) { | |
2192 | break; | |
2193 | } | |
2194 | } | |
2195 | ||
2f4fde93 | 2196 | flush_compressed_data(rs, f); |
56e93d26 | 2197 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 JQ |
2198 | |
2199 | rcu_read_unlock(); | |
d09a6fde | 2200 | |
56e93d26 JQ |
2201 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2202 | ||
2203 | return 0; | |
2204 | } | |
2205 | ||
c31b098f DDAG |
2206 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2207 | uint64_t *non_postcopiable_pending, | |
2208 | uint64_t *postcopiable_pending) | |
56e93d26 | 2209 | { |
8d820d6f | 2210 | RAMState *rs = opaque; |
56e93d26 JQ |
2211 | uint64_t remaining_size; |
2212 | ||
9edabd4d | 2213 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2214 | |
663e6c1d DDAG |
2215 | if (!migration_in_postcopy(migrate_get_current()) && |
2216 | remaining_size < max_size) { | |
56e93d26 JQ |
2217 | qemu_mutex_lock_iothread(); |
2218 | rcu_read_lock(); | |
8d820d6f | 2219 | migration_bitmap_sync(rs); |
56e93d26 JQ |
2220 | rcu_read_unlock(); |
2221 | qemu_mutex_unlock_iothread(); | |
9edabd4d | 2222 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2223 | } |
c31b098f DDAG |
2224 | |
2225 | /* We can do postcopy, and all the data is postcopiable */ | |
2226 | *postcopiable_pending += remaining_size; | |
56e93d26 JQ |
2227 | } |
2228 | ||
2229 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2230 | { | |
2231 | unsigned int xh_len; | |
2232 | int xh_flags; | |
063e760a | 2233 | uint8_t *loaded_data; |
56e93d26 JQ |
2234 | |
2235 | if (!xbzrle_decoded_buf) { | |
2236 | xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE); | |
2237 | } | |
063e760a | 2238 | loaded_data = xbzrle_decoded_buf; |
56e93d26 JQ |
2239 | |
2240 | /* extract RLE header */ | |
2241 | xh_flags = qemu_get_byte(f); | |
2242 | xh_len = qemu_get_be16(f); | |
2243 | ||
2244 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2245 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2246 | return -1; | |
2247 | } | |
2248 | ||
2249 | if (xh_len > TARGET_PAGE_SIZE) { | |
2250 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2251 | return -1; | |
2252 | } | |
2253 | /* load data and decode */ | |
063e760a | 2254 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2255 | |
2256 | /* decode RLE */ | |
063e760a | 2257 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2258 | TARGET_PAGE_SIZE) == -1) { |
2259 | error_report("Failed to load XBZRLE page - decode error!"); | |
2260 | return -1; | |
2261 | } | |
2262 | ||
2263 | return 0; | |
2264 | } | |
2265 | ||
3d0684b2 JQ |
2266 | /** |
2267 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2268 | * | |
2269 | * Must be called from within a rcu critical section. | |
2270 | * | |
56e93d26 | 2271 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2272 | * |
3d0684b2 JQ |
2273 | * @f: QEMUFile where to read the data from |
2274 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2275 | */ |
3d0684b2 | 2276 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2277 | { |
2278 | static RAMBlock *block = NULL; | |
2279 | char id[256]; | |
2280 | uint8_t len; | |
2281 | ||
2282 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2283 | if (!block) { |
56e93d26 JQ |
2284 | error_report("Ack, bad migration stream!"); |
2285 | return NULL; | |
2286 | } | |
4c4bad48 | 2287 | return block; |
56e93d26 JQ |
2288 | } |
2289 | ||
2290 | len = qemu_get_byte(f); | |
2291 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2292 | id[len] = 0; | |
2293 | ||
e3dd7493 | 2294 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2295 | if (!block) { |
2296 | error_report("Can't find block %s", id); | |
2297 | return NULL; | |
56e93d26 JQ |
2298 | } |
2299 | ||
4c4bad48 HZ |
2300 | return block; |
2301 | } | |
2302 | ||
2303 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2304 | ram_addr_t offset) | |
2305 | { | |
2306 | if (!offset_in_ramblock(block, offset)) { | |
2307 | return NULL; | |
2308 | } | |
2309 | ||
2310 | return block->host + offset; | |
56e93d26 JQ |
2311 | } |
2312 | ||
3d0684b2 JQ |
2313 | /** |
2314 | * ram_handle_compressed: handle the zero page case | |
2315 | * | |
56e93d26 JQ |
2316 | * If a page (or a whole RDMA chunk) has been |
2317 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2318 | * |
2319 | * @host: host address for the zero page | |
2320 | * @ch: what the page is filled from. We only support zero | |
2321 | * @size: size of the zero page | |
56e93d26 JQ |
2322 | */ |
2323 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2324 | { | |
2325 | if (ch != 0 || !is_zero_range(host, size)) { | |
2326 | memset(host, ch, size); | |
2327 | } | |
2328 | } | |
2329 | ||
2330 | static void *do_data_decompress(void *opaque) | |
2331 | { | |
2332 | DecompressParam *param = opaque; | |
2333 | unsigned long pagesize; | |
33d151f4 LL |
2334 | uint8_t *des; |
2335 | int len; | |
56e93d26 | 2336 | |
33d151f4 | 2337 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2338 | while (!param->quit) { |
33d151f4 LL |
2339 | if (param->des) { |
2340 | des = param->des; | |
2341 | len = param->len; | |
2342 | param->des = 0; | |
2343 | qemu_mutex_unlock(¶m->mutex); | |
2344 | ||
56e93d26 | 2345 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2346 | /* uncompress() will return failed in some case, especially |
2347 | * when the page is dirted when doing the compression, it's | |
2348 | * not a problem because the dirty page will be retransferred | |
2349 | * and uncompress() won't break the data in other pages. | |
2350 | */ | |
33d151f4 LL |
2351 | uncompress((Bytef *)des, &pagesize, |
2352 | (const Bytef *)param->compbuf, len); | |
73a8912b | 2353 | |
33d151f4 LL |
2354 | qemu_mutex_lock(&decomp_done_lock); |
2355 | param->done = true; | |
2356 | qemu_cond_signal(&decomp_done_cond); | |
2357 | qemu_mutex_unlock(&decomp_done_lock); | |
2358 | ||
2359 | qemu_mutex_lock(¶m->mutex); | |
2360 | } else { | |
2361 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2362 | } | |
56e93d26 | 2363 | } |
33d151f4 | 2364 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2365 | |
2366 | return NULL; | |
2367 | } | |
2368 | ||
5533b2e9 LL |
2369 | static void wait_for_decompress_done(void) |
2370 | { | |
2371 | int idx, thread_count; | |
2372 | ||
2373 | if (!migrate_use_compression()) { | |
2374 | return; | |
2375 | } | |
2376 | ||
2377 | thread_count = migrate_decompress_threads(); | |
2378 | qemu_mutex_lock(&decomp_done_lock); | |
2379 | for (idx = 0; idx < thread_count; idx++) { | |
2380 | while (!decomp_param[idx].done) { | |
2381 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2382 | } | |
2383 | } | |
2384 | qemu_mutex_unlock(&decomp_done_lock); | |
2385 | } | |
2386 | ||
56e93d26 JQ |
2387 | void migrate_decompress_threads_create(void) |
2388 | { | |
2389 | int i, thread_count; | |
2390 | ||
2391 | thread_count = migrate_decompress_threads(); | |
2392 | decompress_threads = g_new0(QemuThread, thread_count); | |
2393 | decomp_param = g_new0(DecompressParam, thread_count); | |
73a8912b LL |
2394 | qemu_mutex_init(&decomp_done_lock); |
2395 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2396 | for (i = 0; i < thread_count; i++) { |
2397 | qemu_mutex_init(&decomp_param[i].mutex); | |
2398 | qemu_cond_init(&decomp_param[i].cond); | |
2399 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2400 | decomp_param[i].done = true; |
90e56fb4 | 2401 | decomp_param[i].quit = false; |
56e93d26 JQ |
2402 | qemu_thread_create(decompress_threads + i, "decompress", |
2403 | do_data_decompress, decomp_param + i, | |
2404 | QEMU_THREAD_JOINABLE); | |
2405 | } | |
2406 | } | |
2407 | ||
2408 | void migrate_decompress_threads_join(void) | |
2409 | { | |
2410 | int i, thread_count; | |
2411 | ||
56e93d26 JQ |
2412 | thread_count = migrate_decompress_threads(); |
2413 | for (i = 0; i < thread_count; i++) { | |
2414 | qemu_mutex_lock(&decomp_param[i].mutex); | |
90e56fb4 | 2415 | decomp_param[i].quit = true; |
56e93d26 JQ |
2416 | qemu_cond_signal(&decomp_param[i].cond); |
2417 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2418 | } | |
2419 | for (i = 0; i < thread_count; i++) { | |
2420 | qemu_thread_join(decompress_threads + i); | |
2421 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2422 | qemu_cond_destroy(&decomp_param[i].cond); | |
2423 | g_free(decomp_param[i].compbuf); | |
2424 | } | |
2425 | g_free(decompress_threads); | |
2426 | g_free(decomp_param); | |
56e93d26 JQ |
2427 | decompress_threads = NULL; |
2428 | decomp_param = NULL; | |
56e93d26 JQ |
2429 | } |
2430 | ||
c1bc6626 | 2431 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2432 | void *host, int len) |
2433 | { | |
2434 | int idx, thread_count; | |
2435 | ||
2436 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2437 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2438 | while (true) { |
2439 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2440 | if (decomp_param[idx].done) { |
33d151f4 LL |
2441 | decomp_param[idx].done = false; |
2442 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2443 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2444 | decomp_param[idx].des = host; |
2445 | decomp_param[idx].len = len; | |
33d151f4 LL |
2446 | qemu_cond_signal(&decomp_param[idx].cond); |
2447 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2448 | break; |
2449 | } | |
2450 | } | |
2451 | if (idx < thread_count) { | |
2452 | break; | |
73a8912b LL |
2453 | } else { |
2454 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2455 | } |
2456 | } | |
73a8912b | 2457 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2458 | } |
2459 | ||
3d0684b2 JQ |
2460 | /** |
2461 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
2462 | * | |
2463 | * Returns 0 for success and negative if there was one error | |
2464 | * | |
2465 | * @mis: current migration incoming state | |
2466 | * | |
2467 | * Allocate data structures etc needed by incoming migration with | |
2468 | * postcopy-ram. postcopy-ram's similarly names | |
2469 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
2470 | */ |
2471 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2472 | { | |
2473 | size_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
2474 | ||
2475 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2476 | } | |
2477 | ||
3d0684b2 JQ |
2478 | /** |
2479 | * ram_load_postcopy: load a page in postcopy case | |
2480 | * | |
2481 | * Returns 0 for success or -errno in case of error | |
2482 | * | |
a7180877 DDAG |
2483 | * Called in postcopy mode by ram_load(). |
2484 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
2485 | * |
2486 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
2487 | */ |
2488 | static int ram_load_postcopy(QEMUFile *f) | |
2489 | { | |
2490 | int flags = 0, ret = 0; | |
2491 | bool place_needed = false; | |
28abd200 | 2492 | bool matching_page_sizes = false; |
a7180877 DDAG |
2493 | MigrationIncomingState *mis = migration_incoming_get_current(); |
2494 | /* Temporary page that is later 'placed' */ | |
2495 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2496 | void *last_host = NULL; |
a3b6ff6d | 2497 | bool all_zero = false; |
a7180877 DDAG |
2498 | |
2499 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2500 | ram_addr_t addr; | |
2501 | void *host = NULL; | |
2502 | void *page_buffer = NULL; | |
2503 | void *place_source = NULL; | |
df9ff5e1 | 2504 | RAMBlock *block = NULL; |
a7180877 | 2505 | uint8_t ch; |
a7180877 DDAG |
2506 | |
2507 | addr = qemu_get_be64(f); | |
2508 | flags = addr & ~TARGET_PAGE_MASK; | |
2509 | addr &= TARGET_PAGE_MASK; | |
2510 | ||
2511 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2512 | place_needed = false; | |
2513 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE)) { | |
df9ff5e1 | 2514 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
2515 | |
2516 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2517 | if (!host) { |
2518 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2519 | ret = -EINVAL; | |
2520 | break; | |
2521 | } | |
28abd200 | 2522 | matching_page_sizes = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 2523 | /* |
28abd200 DDAG |
2524 | * Postcopy requires that we place whole host pages atomically; |
2525 | * these may be huge pages for RAMBlocks that are backed by | |
2526 | * hugetlbfs. | |
a7180877 DDAG |
2527 | * To make it atomic, the data is read into a temporary page |
2528 | * that's moved into place later. | |
2529 | * The migration protocol uses, possibly smaller, target-pages | |
2530 | * however the source ensures it always sends all the components | |
2531 | * of a host page in order. | |
2532 | */ | |
2533 | page_buffer = postcopy_host_page + | |
28abd200 | 2534 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 2535 | /* If all TP are zero then we can optimise the place */ |
28abd200 | 2536 | if (!((uintptr_t)host & (block->page_size - 1))) { |
a7180877 | 2537 | all_zero = true; |
c53b7ddc DDAG |
2538 | } else { |
2539 | /* not the 1st TP within the HP */ | |
2540 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2541 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2542 | host, last_host); |
2543 | ret = -EINVAL; | |
2544 | break; | |
2545 | } | |
a7180877 DDAG |
2546 | } |
2547 | ||
c53b7ddc | 2548 | |
a7180877 DDAG |
2549 | /* |
2550 | * If it's the last part of a host page then we place the host | |
2551 | * page | |
2552 | */ | |
2553 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
28abd200 | 2554 | (block->page_size - 1)) == 0; |
a7180877 DDAG |
2555 | place_source = postcopy_host_page; |
2556 | } | |
c53b7ddc | 2557 | last_host = host; |
a7180877 DDAG |
2558 | |
2559 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
2560 | case RAM_SAVE_FLAG_COMPRESS: | |
2561 | ch = qemu_get_byte(f); | |
2562 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2563 | if (ch) { | |
2564 | all_zero = false; | |
2565 | } | |
2566 | break; | |
2567 | ||
2568 | case RAM_SAVE_FLAG_PAGE: | |
2569 | all_zero = false; | |
2570 | if (!place_needed || !matching_page_sizes) { | |
2571 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2572 | } else { | |
2573 | /* Avoids the qemu_file copy during postcopy, which is | |
2574 | * going to do a copy later; can only do it when we | |
2575 | * do this read in one go (matching page sizes) | |
2576 | */ | |
2577 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2578 | TARGET_PAGE_SIZE); | |
2579 | } | |
2580 | break; | |
2581 | case RAM_SAVE_FLAG_EOS: | |
2582 | /* normal exit */ | |
2583 | break; | |
2584 | default: | |
2585 | error_report("Unknown combination of migration flags: %#x" | |
2586 | " (postcopy mode)", flags); | |
2587 | ret = -EINVAL; | |
2588 | } | |
2589 | ||
2590 | if (place_needed) { | |
2591 | /* This gets called at the last target page in the host page */ | |
df9ff5e1 DDAG |
2592 | void *place_dest = host + TARGET_PAGE_SIZE - block->page_size; |
2593 | ||
a7180877 | 2594 | if (all_zero) { |
df9ff5e1 DDAG |
2595 | ret = postcopy_place_page_zero(mis, place_dest, |
2596 | block->page_size); | |
a7180877 | 2597 | } else { |
df9ff5e1 DDAG |
2598 | ret = postcopy_place_page(mis, place_dest, |
2599 | place_source, block->page_size); | |
a7180877 DDAG |
2600 | } |
2601 | } | |
2602 | if (!ret) { | |
2603 | ret = qemu_file_get_error(f); | |
2604 | } | |
2605 | } | |
2606 | ||
2607 | return ret; | |
2608 | } | |
2609 | ||
56e93d26 JQ |
2610 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2611 | { | |
2612 | int flags = 0, ret = 0; | |
2613 | static uint64_t seq_iter; | |
2614 | int len = 0; | |
a7180877 DDAG |
2615 | /* |
2616 | * If system is running in postcopy mode, page inserts to host memory must | |
2617 | * be atomic | |
2618 | */ | |
2619 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
ef08fb38 DDAG |
2620 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
2621 | bool postcopy_advised = postcopy_state_get() >= POSTCOPY_INCOMING_ADVISE; | |
56e93d26 JQ |
2622 | |
2623 | seq_iter++; | |
2624 | ||
2625 | if (version_id != 4) { | |
2626 | ret = -EINVAL; | |
2627 | } | |
2628 | ||
2629 | /* This RCU critical section can be very long running. | |
2630 | * When RCU reclaims in the code start to become numerous, | |
2631 | * it will be necessary to reduce the granularity of this | |
2632 | * critical section. | |
2633 | */ | |
2634 | rcu_read_lock(); | |
a7180877 DDAG |
2635 | |
2636 | if (postcopy_running) { | |
2637 | ret = ram_load_postcopy(f); | |
2638 | } | |
2639 | ||
2640 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2641 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2642 | void *host = NULL; |
56e93d26 JQ |
2643 | uint8_t ch; |
2644 | ||
2645 | addr = qemu_get_be64(f); | |
2646 | flags = addr & ~TARGET_PAGE_MASK; | |
2647 | addr &= TARGET_PAGE_MASK; | |
2648 | ||
a776aa15 DDAG |
2649 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE | |
2650 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { | |
4c4bad48 HZ |
2651 | RAMBlock *block = ram_block_from_stream(f, flags); |
2652 | ||
2653 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2654 | if (!host) { |
2655 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2656 | ret = -EINVAL; | |
2657 | break; | |
2658 | } | |
2659 | } | |
2660 | ||
56e93d26 JQ |
2661 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2662 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2663 | /* Synchronize RAM block list */ | |
2664 | total_ram_bytes = addr; | |
2665 | while (!ret && total_ram_bytes) { | |
2666 | RAMBlock *block; | |
56e93d26 JQ |
2667 | char id[256]; |
2668 | ram_addr_t length; | |
2669 | ||
2670 | len = qemu_get_byte(f); | |
2671 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2672 | id[len] = 0; | |
2673 | length = qemu_get_be64(f); | |
2674 | ||
e3dd7493 DDAG |
2675 | block = qemu_ram_block_by_name(id); |
2676 | if (block) { | |
2677 | if (length != block->used_length) { | |
2678 | Error *local_err = NULL; | |
56e93d26 | 2679 | |
fa53a0e5 | 2680 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2681 | &local_err); |
2682 | if (local_err) { | |
2683 | error_report_err(local_err); | |
56e93d26 | 2684 | } |
56e93d26 | 2685 | } |
ef08fb38 DDAG |
2686 | /* For postcopy we need to check hugepage sizes match */ |
2687 | if (postcopy_advised && | |
2688 | block->page_size != qemu_host_page_size) { | |
2689 | uint64_t remote_page_size = qemu_get_be64(f); | |
2690 | if (remote_page_size != block->page_size) { | |
2691 | error_report("Mismatched RAM page size %s " | |
2692 | "(local) %zd != %" PRId64, | |
2693 | id, block->page_size, | |
2694 | remote_page_size); | |
2695 | ret = -EINVAL; | |
2696 | } | |
2697 | } | |
e3dd7493 DDAG |
2698 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2699 | block->idstr); | |
2700 | } else { | |
56e93d26 JQ |
2701 | error_report("Unknown ramblock \"%s\", cannot " |
2702 | "accept migration", id); | |
2703 | ret = -EINVAL; | |
2704 | } | |
2705 | ||
2706 | total_ram_bytes -= length; | |
2707 | } | |
2708 | break; | |
a776aa15 | 2709 | |
56e93d26 | 2710 | case RAM_SAVE_FLAG_COMPRESS: |
56e93d26 JQ |
2711 | ch = qemu_get_byte(f); |
2712 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2713 | break; | |
a776aa15 | 2714 | |
56e93d26 | 2715 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2716 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2717 | break; | |
56e93d26 | 2718 | |
a776aa15 | 2719 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2720 | len = qemu_get_be32(f); |
2721 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2722 | error_report("Invalid compressed data length: %d", len); | |
2723 | ret = -EINVAL; | |
2724 | break; | |
2725 | } | |
c1bc6626 | 2726 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2727 | break; |
a776aa15 | 2728 | |
56e93d26 | 2729 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2730 | if (load_xbzrle(f, addr, host) < 0) { |
2731 | error_report("Failed to decompress XBZRLE page at " | |
2732 | RAM_ADDR_FMT, addr); | |
2733 | ret = -EINVAL; | |
2734 | break; | |
2735 | } | |
2736 | break; | |
2737 | case RAM_SAVE_FLAG_EOS: | |
2738 | /* normal exit */ | |
2739 | break; | |
2740 | default: | |
2741 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2742 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2743 | } else { |
2744 | error_report("Unknown combination of migration flags: %#x", | |
2745 | flags); | |
2746 | ret = -EINVAL; | |
2747 | } | |
2748 | } | |
2749 | if (!ret) { | |
2750 | ret = qemu_file_get_error(f); | |
2751 | } | |
2752 | } | |
2753 | ||
5533b2e9 | 2754 | wait_for_decompress_done(); |
56e93d26 | 2755 | rcu_read_unlock(); |
55c4446b | 2756 | trace_ram_load_complete(ret, seq_iter); |
56e93d26 JQ |
2757 | return ret; |
2758 | } | |
2759 | ||
2760 | static SaveVMHandlers savevm_ram_handlers = { | |
2761 | .save_live_setup = ram_save_setup, | |
2762 | .save_live_iterate = ram_save_iterate, | |
763c906b | 2763 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2764 | .save_live_complete_precopy = ram_save_complete, |
56e93d26 JQ |
2765 | .save_live_pending = ram_save_pending, |
2766 | .load_state = ram_load, | |
6ad2a215 | 2767 | .cleanup = ram_migration_cleanup, |
56e93d26 JQ |
2768 | }; |
2769 | ||
2770 | void ram_mig_init(void) | |
2771 | { | |
2772 | qemu_mutex_init(&XBZRLE.lock); | |
6f37bb8b | 2773 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 2774 | } |