]>
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 | ||
ec481c6c JQ |
154 | /* |
155 | * An outstanding page request, on the source, having been received | |
156 | * and queued | |
157 | */ | |
158 | struct RAMSrcPageRequest { | |
159 | RAMBlock *rb; | |
160 | hwaddr offset; | |
161 | hwaddr len; | |
162 | ||
163 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
164 | }; | |
165 | ||
6f37bb8b JQ |
166 | /* State of RAM for migration */ |
167 | struct RAMState { | |
204b88b8 JQ |
168 | /* QEMUFile used for this migration */ |
169 | QEMUFile *f; | |
6f37bb8b JQ |
170 | /* Last block that we have visited searching for dirty pages */ |
171 | RAMBlock *last_seen_block; | |
172 | /* Last block from where we have sent data */ | |
173 | RAMBlock *last_sent_block; | |
174 | /* Last offset we have sent data from */ | |
175 | ram_addr_t last_offset; | |
176 | /* last ram version we have seen */ | |
177 | uint32_t last_version; | |
178 | /* We are in the first round */ | |
179 | bool ram_bulk_stage; | |
8d820d6f JQ |
180 | /* How many times we have dirty too many pages */ |
181 | int dirty_rate_high_cnt; | |
5a987738 JQ |
182 | /* How many times we have synchronized the bitmap */ |
183 | uint64_t bitmap_sync_count; | |
f664da80 JQ |
184 | /* these variables are used for bitmap sync */ |
185 | /* last time we did a full bitmap_sync */ | |
186 | int64_t time_last_bitmap_sync; | |
eac74159 | 187 | /* bytes transferred at start_time */ |
c4bdf0cf | 188 | uint64_t bytes_xfer_prev; |
a66cd90c | 189 | /* number of dirty pages since start_time */ |
68908ed6 | 190 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
191 | /* xbzrle misses since the beginning of the period */ |
192 | uint64_t xbzrle_cache_miss_prev; | |
36040d9c JQ |
193 | /* number of iterations at the beginning of period */ |
194 | uint64_t iterations_prev; | |
f7ccd61b JQ |
195 | /* Accounting fields */ |
196 | /* number of zero pages. It used to be pages filled by the same char. */ | |
197 | uint64_t zero_pages; | |
b4d1c6e7 JQ |
198 | /* number of normal transferred pages */ |
199 | uint64_t norm_pages; | |
23b28c3c JQ |
200 | /* Iterations since start */ |
201 | uint64_t iterations; | |
f36ada95 JQ |
202 | /* xbzrle transmitted bytes. Notice that this is with |
203 | * compression, they can't be calculated from the pages */ | |
07ed50a2 | 204 | uint64_t xbzrle_bytes; |
f36ada95 JQ |
205 | /* xbzrle transmmited pages */ |
206 | uint64_t xbzrle_pages; | |
544c36f1 JQ |
207 | /* xbzrle number of cache miss */ |
208 | uint64_t xbzrle_cache_miss; | |
b07016b6 JQ |
209 | /* xbzrle miss rate */ |
210 | double xbzrle_cache_miss_rate; | |
180f61f7 JQ |
211 | /* xbzrle number of overflows */ |
212 | uint64_t xbzrle_overflows; | |
0d8ec885 JQ |
213 | /* number of dirty bits in the bitmap */ |
214 | uint64_t migration_dirty_pages; | |
2f4fde93 JQ |
215 | /* total number of bytes transferred */ |
216 | uint64_t bytes_transferred; | |
47ad8619 JQ |
217 | /* number of dirtied pages in the last second */ |
218 | uint64_t dirty_pages_rate; | |
96506894 JQ |
219 | /* Count of requests incoming from destination */ |
220 | uint64_t postcopy_requests; | |
108cfae0 JQ |
221 | /* protects modification of the bitmap */ |
222 | QemuMutex bitmap_mutex; | |
eb859c53 JQ |
223 | /* Ram Bitmap protected by RCU */ |
224 | RAMBitmap *ram_bitmap; | |
68a098f3 JQ |
225 | /* The RAMBlock used in the last src_page_requests */ |
226 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
227 | /* Queue of outstanding page requests from the destination */ |
228 | QemuMutex src_page_req_mutex; | |
229 | QSIMPLEQ_HEAD(src_page_requests, RAMSrcPageRequest) src_page_requests; | |
6f37bb8b JQ |
230 | }; |
231 | typedef struct RAMState RAMState; | |
232 | ||
233 | static RAMState ram_state; | |
234 | ||
56e93d26 JQ |
235 | uint64_t dup_mig_pages_transferred(void) |
236 | { | |
f7ccd61b | 237 | return ram_state.zero_pages; |
56e93d26 JQ |
238 | } |
239 | ||
56e93d26 JQ |
240 | uint64_t norm_mig_pages_transferred(void) |
241 | { | |
b4d1c6e7 | 242 | return ram_state.norm_pages; |
56e93d26 JQ |
243 | } |
244 | ||
245 | uint64_t xbzrle_mig_bytes_transferred(void) | |
246 | { | |
07ed50a2 | 247 | return ram_state.xbzrle_bytes; |
56e93d26 JQ |
248 | } |
249 | ||
250 | uint64_t xbzrle_mig_pages_transferred(void) | |
251 | { | |
f36ada95 | 252 | return ram_state.xbzrle_pages; |
56e93d26 JQ |
253 | } |
254 | ||
255 | uint64_t xbzrle_mig_pages_cache_miss(void) | |
256 | { | |
544c36f1 | 257 | return ram_state.xbzrle_cache_miss; |
56e93d26 JQ |
258 | } |
259 | ||
260 | double xbzrle_mig_cache_miss_rate(void) | |
261 | { | |
b07016b6 | 262 | return ram_state.xbzrle_cache_miss_rate; |
56e93d26 JQ |
263 | } |
264 | ||
265 | uint64_t xbzrle_mig_pages_overflow(void) | |
266 | { | |
180f61f7 | 267 | return ram_state.xbzrle_overflows; |
56e93d26 JQ |
268 | } |
269 | ||
9edabd4d | 270 | uint64_t ram_bytes_transferred(void) |
0d8ec885 | 271 | { |
9edabd4d | 272 | return ram_state.bytes_transferred; |
0d8ec885 JQ |
273 | } |
274 | ||
9edabd4d | 275 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 276 | { |
9edabd4d | 277 | return ram_state.migration_dirty_pages * TARGET_PAGE_SIZE; |
2f4fde93 JQ |
278 | } |
279 | ||
42d219d3 JQ |
280 | uint64_t ram_dirty_sync_count(void) |
281 | { | |
282 | return ram_state.bitmap_sync_count; | |
283 | } | |
284 | ||
47ad8619 JQ |
285 | uint64_t ram_dirty_pages_rate(void) |
286 | { | |
287 | return ram_state.dirty_pages_rate; | |
288 | } | |
289 | ||
96506894 JQ |
290 | uint64_t ram_postcopy_requests(void) |
291 | { | |
292 | return ram_state.postcopy_requests; | |
293 | } | |
294 | ||
b8fb8cb7 DDAG |
295 | /* used by the search for pages to send */ |
296 | struct PageSearchStatus { | |
297 | /* Current block being searched */ | |
298 | RAMBlock *block; | |
299 | /* Current offset to search from */ | |
300 | ram_addr_t offset; | |
301 | /* Set once we wrap around */ | |
302 | bool complete_round; | |
303 | }; | |
304 | typedef struct PageSearchStatus PageSearchStatus; | |
305 | ||
56e93d26 | 306 | struct CompressParam { |
56e93d26 | 307 | bool done; |
90e56fb4 | 308 | bool quit; |
56e93d26 JQ |
309 | QEMUFile *file; |
310 | QemuMutex mutex; | |
311 | QemuCond cond; | |
312 | RAMBlock *block; | |
313 | ram_addr_t offset; | |
314 | }; | |
315 | typedef struct CompressParam CompressParam; | |
316 | ||
317 | struct DecompressParam { | |
73a8912b | 318 | bool done; |
90e56fb4 | 319 | bool quit; |
56e93d26 JQ |
320 | QemuMutex mutex; |
321 | QemuCond cond; | |
322 | void *des; | |
d341d9f3 | 323 | uint8_t *compbuf; |
56e93d26 JQ |
324 | int len; |
325 | }; | |
326 | typedef struct DecompressParam DecompressParam; | |
327 | ||
328 | static CompressParam *comp_param; | |
329 | static QemuThread *compress_threads; | |
330 | /* comp_done_cond is used to wake up the migration thread when | |
331 | * one of the compression threads has finished the compression. | |
332 | * comp_done_lock is used to co-work with comp_done_cond. | |
333 | */ | |
0d9f9a5c LL |
334 | static QemuMutex comp_done_lock; |
335 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
336 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
337 | static const QEMUFileOps empty_ops = { }; | |
338 | ||
339 | static bool compression_switch; | |
56e93d26 JQ |
340 | static DecompressParam *decomp_param; |
341 | static QemuThread *decompress_threads; | |
73a8912b LL |
342 | static QemuMutex decomp_done_lock; |
343 | static QemuCond decomp_done_cond; | |
56e93d26 | 344 | |
a7a9a88f LL |
345 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
346 | ram_addr_t offset); | |
56e93d26 JQ |
347 | |
348 | static void *do_data_compress(void *opaque) | |
349 | { | |
350 | CompressParam *param = opaque; | |
a7a9a88f LL |
351 | RAMBlock *block; |
352 | ram_addr_t offset; | |
56e93d26 | 353 | |
a7a9a88f | 354 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 355 | while (!param->quit) { |
a7a9a88f LL |
356 | if (param->block) { |
357 | block = param->block; | |
358 | offset = param->offset; | |
359 | param->block = NULL; | |
360 | qemu_mutex_unlock(¶m->mutex); | |
361 | ||
362 | do_compress_ram_page(param->file, block, offset); | |
363 | ||
0d9f9a5c | 364 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 365 | param->done = true; |
0d9f9a5c LL |
366 | qemu_cond_signal(&comp_done_cond); |
367 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
368 | |
369 | qemu_mutex_lock(¶m->mutex); | |
370 | } else { | |
56e93d26 JQ |
371 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
372 | } | |
56e93d26 | 373 | } |
a7a9a88f | 374 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
375 | |
376 | return NULL; | |
377 | } | |
378 | ||
379 | static inline void terminate_compression_threads(void) | |
380 | { | |
381 | int idx, thread_count; | |
382 | ||
383 | thread_count = migrate_compress_threads(); | |
3d0684b2 | 384 | |
56e93d26 JQ |
385 | for (idx = 0; idx < thread_count; idx++) { |
386 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 387 | comp_param[idx].quit = true; |
56e93d26 JQ |
388 | qemu_cond_signal(&comp_param[idx].cond); |
389 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
390 | } | |
391 | } | |
392 | ||
393 | void migrate_compress_threads_join(void) | |
394 | { | |
395 | int i, thread_count; | |
396 | ||
397 | if (!migrate_use_compression()) { | |
398 | return; | |
399 | } | |
400 | terminate_compression_threads(); | |
401 | thread_count = migrate_compress_threads(); | |
402 | for (i = 0; i < thread_count; i++) { | |
403 | qemu_thread_join(compress_threads + i); | |
404 | qemu_fclose(comp_param[i].file); | |
405 | qemu_mutex_destroy(&comp_param[i].mutex); | |
406 | qemu_cond_destroy(&comp_param[i].cond); | |
407 | } | |
0d9f9a5c LL |
408 | qemu_mutex_destroy(&comp_done_lock); |
409 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
410 | g_free(compress_threads); |
411 | g_free(comp_param); | |
56e93d26 JQ |
412 | compress_threads = NULL; |
413 | comp_param = NULL; | |
56e93d26 JQ |
414 | } |
415 | ||
416 | void migrate_compress_threads_create(void) | |
417 | { | |
418 | int i, thread_count; | |
419 | ||
420 | if (!migrate_use_compression()) { | |
421 | return; | |
422 | } | |
56e93d26 JQ |
423 | compression_switch = true; |
424 | thread_count = migrate_compress_threads(); | |
425 | compress_threads = g_new0(QemuThread, thread_count); | |
426 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
427 | qemu_cond_init(&comp_done_cond); |
428 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 429 | for (i = 0; i < thread_count; i++) { |
e110aa91 C |
430 | /* comp_param[i].file is just used as a dummy buffer to save data, |
431 | * set its ops to empty. | |
56e93d26 JQ |
432 | */ |
433 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
434 | comp_param[i].done = true; | |
90e56fb4 | 435 | comp_param[i].quit = false; |
56e93d26 JQ |
436 | qemu_mutex_init(&comp_param[i].mutex); |
437 | qemu_cond_init(&comp_param[i].cond); | |
438 | qemu_thread_create(compress_threads + i, "compress", | |
439 | do_data_compress, comp_param + i, | |
440 | QEMU_THREAD_JOINABLE); | |
441 | } | |
442 | } | |
443 | ||
444 | /** | |
3d0684b2 | 445 | * save_page_header: write page header to wire |
56e93d26 JQ |
446 | * |
447 | * If this is the 1st block, it also writes the block identification | |
448 | * | |
3d0684b2 | 449 | * Returns the number of bytes written |
56e93d26 JQ |
450 | * |
451 | * @f: QEMUFile where to send the data | |
452 | * @block: block that contains the page we want to send | |
453 | * @offset: offset inside the block for the page | |
454 | * in the lower bits, it contains flags | |
455 | */ | |
456 | static size_t save_page_header(QEMUFile *f, RAMBlock *block, ram_addr_t offset) | |
457 | { | |
9f5f380b | 458 | size_t size, len; |
56e93d26 JQ |
459 | |
460 | qemu_put_be64(f, offset); | |
461 | size = 8; | |
462 | ||
463 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b LL |
464 | len = strlen(block->idstr); |
465 | qemu_put_byte(f, len); | |
466 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
467 | size += 1 + len; | |
56e93d26 JQ |
468 | } |
469 | return size; | |
470 | } | |
471 | ||
3d0684b2 JQ |
472 | /** |
473 | * mig_throttle_guest_down: throotle down the guest | |
474 | * | |
475 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
476 | * writes. If guest dirty memory rate is reduced below the rate at | |
477 | * which we can transfer pages to the destination then we should be | |
478 | * able to complete migration. Some workloads dirty memory way too | |
479 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
480 | */ |
481 | static void mig_throttle_guest_down(void) | |
482 | { | |
483 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
484 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
485 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
486 | |
487 | /* We have not started throttling yet. Let's start it. */ | |
488 | if (!cpu_throttle_active()) { | |
489 | cpu_throttle_set(pct_initial); | |
490 | } else { | |
491 | /* Throttling already on, just increase the rate */ | |
492 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
493 | } | |
494 | } | |
495 | ||
3d0684b2 JQ |
496 | /** |
497 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
498 | * | |
6f37bb8b | 499 | * @rs: current RAM state |
3d0684b2 JQ |
500 | * @current_addr: address for the zero page |
501 | * | |
502 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
503 | * The important thing is that a stale (not-yet-0'd) page be replaced |
504 | * by the new data. | |
505 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 506 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 507 | */ |
6f37bb8b | 508 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 509 | { |
6f37bb8b | 510 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
511 | return; |
512 | } | |
513 | ||
514 | /* We don't care if this fails to allocate a new cache page | |
515 | * as long as it updated an old one */ | |
516 | cache_insert(XBZRLE.cache, current_addr, ZERO_TARGET_PAGE, | |
5a987738 | 517 | rs->bitmap_sync_count); |
56e93d26 JQ |
518 | } |
519 | ||
520 | #define ENCODING_FLAG_XBZRLE 0x1 | |
521 | ||
522 | /** | |
523 | * save_xbzrle_page: compress and send current page | |
524 | * | |
525 | * Returns: 1 means that we wrote the page | |
526 | * 0 means that page is identical to the one already sent | |
527 | * -1 means that xbzrle would be longer than normal | |
528 | * | |
5a987738 | 529 | * @rs: current RAM state |
3d0684b2 JQ |
530 | * @current_data: pointer to the address of the page contents |
531 | * @current_addr: addr of the page | |
56e93d26 JQ |
532 | * @block: block that contains the page we want to send |
533 | * @offset: offset inside the block for the page | |
534 | * @last_stage: if we are at the completion stage | |
56e93d26 | 535 | */ |
204b88b8 | 536 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 537 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 538 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
539 | { |
540 | int encoded_len = 0, bytes_xbzrle; | |
541 | uint8_t *prev_cached_page; | |
542 | ||
5a987738 | 543 | if (!cache_is_cached(XBZRLE.cache, current_addr, rs->bitmap_sync_count)) { |
544c36f1 | 544 | rs->xbzrle_cache_miss++; |
56e93d26 JQ |
545 | if (!last_stage) { |
546 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
5a987738 | 547 | rs->bitmap_sync_count) == -1) { |
56e93d26 JQ |
548 | return -1; |
549 | } else { | |
550 | /* update *current_data when the page has been | |
551 | inserted into cache */ | |
552 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
553 | } | |
554 | } | |
555 | return -1; | |
556 | } | |
557 | ||
558 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
559 | ||
560 | /* save current buffer into memory */ | |
561 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
562 | ||
563 | /* XBZRLE encoding (if there is no overflow) */ | |
564 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
565 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
566 | TARGET_PAGE_SIZE); | |
567 | if (encoded_len == 0) { | |
55c4446b | 568 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
569 | return 0; |
570 | } else if (encoded_len == -1) { | |
55c4446b | 571 | trace_save_xbzrle_page_overflow(); |
180f61f7 | 572 | rs->xbzrle_overflows++; |
56e93d26 JQ |
573 | /* update data in the cache */ |
574 | if (!last_stage) { | |
575 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
576 | *current_data = prev_cached_page; | |
577 | } | |
578 | return -1; | |
579 | } | |
580 | ||
581 | /* we need to update the data in the cache, in order to get the same data */ | |
582 | if (!last_stage) { | |
583 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
584 | } | |
585 | ||
586 | /* Send XBZRLE based compressed page */ | |
204b88b8 JQ |
587 | bytes_xbzrle = save_page_header(rs->f, block, |
588 | offset | RAM_SAVE_FLAG_XBZRLE); | |
589 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
590 | qemu_put_be16(rs->f, encoded_len); | |
591 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 592 | bytes_xbzrle += encoded_len + 1 + 2; |
f36ada95 | 593 | rs->xbzrle_pages++; |
07ed50a2 | 594 | rs->xbzrle_bytes += bytes_xbzrle; |
072c2511 | 595 | rs->bytes_transferred += bytes_xbzrle; |
56e93d26 JQ |
596 | |
597 | return 1; | |
598 | } | |
599 | ||
3d0684b2 JQ |
600 | /** |
601 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 602 | * |
3d0684b2 JQ |
603 | * Called with rcu_read_lock() to protect migration_bitmap |
604 | * | |
605 | * Returns the byte offset within memory region of the start of a dirty page | |
606 | * | |
6f37bb8b | 607 | * @rs: current RAM state |
3d0684b2 JQ |
608 | * @rb: RAMBlock where to search for dirty pages |
609 | * @start: starting address (typically so we can continue from previous page) | |
610 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
611 | * within the global ram_addr space | |
f3f491fc | 612 | */ |
56e93d26 | 613 | static inline |
6f37bb8b | 614 | ram_addr_t migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
a82d593b DDAG |
615 | ram_addr_t start, |
616 | ram_addr_t *ram_addr_abs) | |
56e93d26 | 617 | { |
2f68e399 | 618 | unsigned long base = rb->offset >> TARGET_PAGE_BITS; |
56e93d26 | 619 | unsigned long nr = base + (start >> TARGET_PAGE_BITS); |
2f68e399 DDAG |
620 | uint64_t rb_size = rb->used_length; |
621 | unsigned long size = base + (rb_size >> TARGET_PAGE_BITS); | |
2ff64038 | 622 | unsigned long *bitmap; |
56e93d26 JQ |
623 | |
624 | unsigned long next; | |
625 | ||
eb859c53 | 626 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
6f37bb8b | 627 | if (rs->ram_bulk_stage && nr > base) { |
56e93d26 JQ |
628 | next = nr + 1; |
629 | } else { | |
2ff64038 | 630 | next = find_next_bit(bitmap, size, nr); |
56e93d26 JQ |
631 | } |
632 | ||
f3f491fc | 633 | *ram_addr_abs = next << TARGET_PAGE_BITS; |
56e93d26 JQ |
634 | return (next - base) << TARGET_PAGE_BITS; |
635 | } | |
636 | ||
0d8ec885 | 637 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, ram_addr_t addr) |
a82d593b DDAG |
638 | { |
639 | bool ret; | |
640 | int nr = addr >> TARGET_PAGE_BITS; | |
eb859c53 | 641 | unsigned long *bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
a82d593b DDAG |
642 | |
643 | ret = test_and_clear_bit(nr, bitmap); | |
644 | ||
645 | if (ret) { | |
0d8ec885 | 646 | rs->migration_dirty_pages--; |
a82d593b DDAG |
647 | } |
648 | return ret; | |
649 | } | |
650 | ||
a66cd90c JQ |
651 | static void migration_bitmap_sync_range(RAMState *rs, ram_addr_t start, |
652 | ram_addr_t length) | |
56e93d26 | 653 | { |
2ff64038 | 654 | unsigned long *bitmap; |
eb859c53 | 655 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
0d8ec885 JQ |
656 | rs->migration_dirty_pages += |
657 | cpu_physical_memory_sync_dirty_bitmap(bitmap, start, length, | |
658 | &rs->num_dirty_pages_period); | |
56e93d26 JQ |
659 | } |
660 | ||
3d0684b2 JQ |
661 | /** |
662 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
663 | * | |
664 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
665 | * | |
666 | * For VMs with just normal pages this is equivalent to the host page | |
667 | * size. If it's got some huge pages then it's the OR of all the | |
668 | * different page sizes. | |
e8ca1db2 DDAG |
669 | */ |
670 | uint64_t ram_pagesize_summary(void) | |
671 | { | |
672 | RAMBlock *block; | |
673 | uint64_t summary = 0; | |
674 | ||
675 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
676 | summary |= block->page_size; | |
677 | } | |
678 | ||
679 | return summary; | |
680 | } | |
681 | ||
8d820d6f | 682 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
683 | { |
684 | RAMBlock *block; | |
56e93d26 | 685 | int64_t end_time; |
c4bdf0cf | 686 | uint64_t bytes_xfer_now; |
56e93d26 | 687 | |
5a987738 | 688 | rs->bitmap_sync_count++; |
56e93d26 | 689 | |
eac74159 JQ |
690 | if (!rs->bytes_xfer_prev) { |
691 | rs->bytes_xfer_prev = ram_bytes_transferred(); | |
56e93d26 JQ |
692 | } |
693 | ||
f664da80 JQ |
694 | if (!rs->time_last_bitmap_sync) { |
695 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
696 | } |
697 | ||
698 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 699 | memory_global_dirty_log_sync(); |
56e93d26 | 700 | |
108cfae0 | 701 | qemu_mutex_lock(&rs->bitmap_mutex); |
56e93d26 JQ |
702 | rcu_read_lock(); |
703 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
a66cd90c | 704 | migration_bitmap_sync_range(rs, block->offset, block->used_length); |
56e93d26 JQ |
705 | } |
706 | rcu_read_unlock(); | |
108cfae0 | 707 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 708 | |
a66cd90c | 709 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 710 | |
56e93d26 JQ |
711 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
712 | ||
713 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 714 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
56e93d26 JQ |
715 | if (migrate_auto_converge()) { |
716 | /* The following detection logic can be refined later. For now: | |
717 | Check to see if the dirtied bytes is 50% more than the approx. | |
718 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
719 | were in this routine. If that happens twice, start or increase |
720 | throttling */ | |
56e93d26 | 721 | bytes_xfer_now = ram_bytes_transferred(); |
070afca2 | 722 | |
47ad8619 | 723 | if (rs->dirty_pages_rate && |
a66cd90c | 724 | (rs->num_dirty_pages_period * TARGET_PAGE_SIZE > |
eac74159 | 725 | (bytes_xfer_now - rs->bytes_xfer_prev) / 2) && |
8d820d6f | 726 | (rs->dirty_rate_high_cnt++ >= 2)) { |
56e93d26 | 727 | trace_migration_throttle(); |
8d820d6f | 728 | rs->dirty_rate_high_cnt = 0; |
070afca2 | 729 | mig_throttle_guest_down(); |
56e93d26 | 730 | } |
eac74159 | 731 | rs->bytes_xfer_prev = bytes_xfer_now; |
56e93d26 | 732 | } |
070afca2 | 733 | |
56e93d26 | 734 | if (migrate_use_xbzrle()) { |
23b28c3c | 735 | if (rs->iterations_prev != rs->iterations) { |
b07016b6 | 736 | rs->xbzrle_cache_miss_rate = |
544c36f1 | 737 | (double)(rs->xbzrle_cache_miss - |
b5833fde | 738 | rs->xbzrle_cache_miss_prev) / |
23b28c3c | 739 | (rs->iterations - rs->iterations_prev); |
56e93d26 | 740 | } |
23b28c3c | 741 | rs->iterations_prev = rs->iterations; |
544c36f1 | 742 | rs->xbzrle_cache_miss_prev = rs->xbzrle_cache_miss; |
56e93d26 | 743 | } |
47ad8619 | 744 | rs->dirty_pages_rate = rs->num_dirty_pages_period * 1000 |
f664da80 | 745 | / (end_time - rs->time_last_bitmap_sync); |
f664da80 | 746 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 747 | rs->num_dirty_pages_period = 0; |
56e93d26 | 748 | } |
4addcd4f | 749 | if (migrate_use_events()) { |
5a987738 | 750 | qapi_event_send_migration_pass(rs->bitmap_sync_count, NULL); |
4addcd4f | 751 | } |
56e93d26 JQ |
752 | } |
753 | ||
754 | /** | |
3d0684b2 | 755 | * save_zero_page: send the zero page to the stream |
56e93d26 | 756 | * |
3d0684b2 | 757 | * Returns the number of pages written. |
56e93d26 | 758 | * |
f7ccd61b | 759 | * @rs: current RAM state |
56e93d26 JQ |
760 | * @f: QEMUFile where to send the data |
761 | * @block: block that contains the page we want to send | |
762 | * @offset: offset inside the block for the page | |
763 | * @p: pointer to the page | |
56e93d26 | 764 | */ |
f7ccd61b | 765 | static int save_zero_page(RAMState *rs, QEMUFile *f, RAMBlock *block, |
072c2511 | 766 | ram_addr_t offset, uint8_t *p) |
56e93d26 JQ |
767 | { |
768 | int pages = -1; | |
769 | ||
770 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
f7ccd61b | 771 | rs->zero_pages++; |
072c2511 JQ |
772 | rs->bytes_transferred += |
773 | save_page_header(f, block, offset | RAM_SAVE_FLAG_COMPRESS); | |
56e93d26 | 774 | qemu_put_byte(f, 0); |
072c2511 | 775 | rs->bytes_transferred += 1; |
56e93d26 JQ |
776 | pages = 1; |
777 | } | |
778 | ||
779 | return pages; | |
780 | } | |
781 | ||
36449157 | 782 | static void ram_release_pages(MigrationState *ms, const char *rbname, |
53f09a10 PB |
783 | uint64_t offset, int pages) |
784 | { | |
785 | if (!migrate_release_ram() || !migration_in_postcopy(ms)) { | |
786 | return; | |
787 | } | |
788 | ||
36449157 | 789 | ram_discard_range(NULL, rbname, offset, pages << TARGET_PAGE_BITS); |
53f09a10 PB |
790 | } |
791 | ||
56e93d26 | 792 | /** |
3d0684b2 | 793 | * ram_save_page: send the given page to the stream |
56e93d26 | 794 | * |
3d0684b2 | 795 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
796 | * < 0 - error |
797 | * >=0 - Number of pages written - this might legally be 0 | |
798 | * if xbzrle noticed the page was the same. | |
56e93d26 | 799 | * |
6f37bb8b | 800 | * @rs: current RAM state |
3d0684b2 | 801 | * @ms: current migration state |
56e93d26 JQ |
802 | * @f: QEMUFile where to send the data |
803 | * @block: block that contains the page we want to send | |
804 | * @offset: offset inside the block for the page | |
805 | * @last_stage: if we are at the completion stage | |
56e93d26 | 806 | */ |
6f37bb8b | 807 | static int ram_save_page(RAMState *rs, MigrationState *ms, QEMUFile *f, |
072c2511 | 808 | PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
809 | { |
810 | int pages = -1; | |
811 | uint64_t bytes_xmit; | |
812 | ram_addr_t current_addr; | |
56e93d26 JQ |
813 | uint8_t *p; |
814 | int ret; | |
815 | bool send_async = true; | |
a08f6890 HZ |
816 | RAMBlock *block = pss->block; |
817 | ram_addr_t offset = pss->offset; | |
56e93d26 | 818 | |
2f68e399 | 819 | p = block->host + offset; |
56e93d26 JQ |
820 | |
821 | /* In doubt sent page as normal */ | |
822 | bytes_xmit = 0; | |
823 | ret = ram_control_save_page(f, block->offset, | |
824 | offset, TARGET_PAGE_SIZE, &bytes_xmit); | |
825 | if (bytes_xmit) { | |
072c2511 | 826 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
827 | pages = 1; |
828 | } | |
829 | ||
830 | XBZRLE_cache_lock(); | |
831 | ||
832 | current_addr = block->offset + offset; | |
833 | ||
6f37bb8b | 834 | if (block == rs->last_sent_block) { |
56e93d26 JQ |
835 | offset |= RAM_SAVE_FLAG_CONTINUE; |
836 | } | |
837 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { | |
838 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
839 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 840 | rs->norm_pages++; |
56e93d26 | 841 | } else if (bytes_xmit == 0) { |
f7ccd61b | 842 | rs->zero_pages++; |
56e93d26 JQ |
843 | } |
844 | } | |
845 | } else { | |
072c2511 | 846 | pages = save_zero_page(rs, f, block, offset, p); |
56e93d26 JQ |
847 | if (pages > 0) { |
848 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
849 | * page would be stale | |
850 | */ | |
6f37bb8b | 851 | xbzrle_cache_zero_page(rs, current_addr); |
53f09a10 | 852 | ram_release_pages(ms, block->idstr, pss->offset, pages); |
6f37bb8b | 853 | } else if (!rs->ram_bulk_stage && |
9eb14766 | 854 | !migration_in_postcopy(ms) && migrate_use_xbzrle()) { |
204b88b8 | 855 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
072c2511 | 856 | offset, last_stage); |
56e93d26 JQ |
857 | if (!last_stage) { |
858 | /* Can't send this cached data async, since the cache page | |
859 | * might get updated before it gets to the wire | |
860 | */ | |
861 | send_async = false; | |
862 | } | |
863 | } | |
864 | } | |
865 | ||
866 | /* XBZRLE overflow or normal page */ | |
867 | if (pages == -1) { | |
072c2511 | 868 | rs->bytes_transferred += save_page_header(f, block, |
56e93d26 JQ |
869 | offset | RAM_SAVE_FLAG_PAGE); |
870 | if (send_async) { | |
53f09a10 PB |
871 | qemu_put_buffer_async(f, p, TARGET_PAGE_SIZE, |
872 | migrate_release_ram() & | |
873 | migration_in_postcopy(ms)); | |
56e93d26 JQ |
874 | } else { |
875 | qemu_put_buffer(f, p, TARGET_PAGE_SIZE); | |
876 | } | |
072c2511 | 877 | rs->bytes_transferred += TARGET_PAGE_SIZE; |
56e93d26 | 878 | pages = 1; |
b4d1c6e7 | 879 | rs->norm_pages++; |
56e93d26 JQ |
880 | } |
881 | ||
882 | XBZRLE_cache_unlock(); | |
883 | ||
884 | return pages; | |
885 | } | |
886 | ||
a7a9a88f LL |
887 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
888 | ram_addr_t offset) | |
56e93d26 JQ |
889 | { |
890 | int bytes_sent, blen; | |
a7a9a88f | 891 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 | 892 | |
a7a9a88f | 893 | bytes_sent = save_page_header(f, block, offset | |
56e93d26 | 894 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
a7a9a88f | 895 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, |
56e93d26 | 896 | migrate_compress_level()); |
b3be2896 LL |
897 | if (blen < 0) { |
898 | bytes_sent = 0; | |
899 | qemu_file_set_error(migrate_get_current()->to_dst_file, blen); | |
900 | error_report("compressed data failed!"); | |
901 | } else { | |
902 | bytes_sent += blen; | |
53f09a10 PB |
903 | ram_release_pages(migrate_get_current(), block->idstr, |
904 | offset & TARGET_PAGE_MASK, 1); | |
b3be2896 | 905 | } |
56e93d26 JQ |
906 | |
907 | return bytes_sent; | |
908 | } | |
909 | ||
2f4fde93 | 910 | static void flush_compressed_data(RAMState *rs, QEMUFile *f) |
56e93d26 JQ |
911 | { |
912 | int idx, len, thread_count; | |
913 | ||
914 | if (!migrate_use_compression()) { | |
915 | return; | |
916 | } | |
917 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 918 | |
0d9f9a5c | 919 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 920 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 921 | while (!comp_param[idx].done) { |
0d9f9a5c | 922 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 923 | } |
a7a9a88f | 924 | } |
0d9f9a5c | 925 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
926 | |
927 | for (idx = 0; idx < thread_count; idx++) { | |
928 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 929 | if (!comp_param[idx].quit) { |
56e93d26 | 930 | len = qemu_put_qemu_file(f, comp_param[idx].file); |
2f4fde93 | 931 | rs->bytes_transferred += len; |
56e93d26 | 932 | } |
a7a9a88f | 933 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
934 | } |
935 | } | |
936 | ||
937 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
938 | ram_addr_t offset) | |
939 | { | |
940 | param->block = block; | |
941 | param->offset = offset; | |
942 | } | |
943 | ||
b4d1c6e7 | 944 | static int compress_page_with_multi_thread(RAMState *rs, QEMUFile *f, |
072c2511 | 945 | RAMBlock *block, ram_addr_t offset) |
56e93d26 JQ |
946 | { |
947 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
948 | ||
949 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 950 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 JQ |
951 | while (true) { |
952 | for (idx = 0; idx < thread_count; idx++) { | |
953 | if (comp_param[idx].done) { | |
a7a9a88f | 954 | comp_param[idx].done = false; |
56e93d26 | 955 | bytes_xmit = qemu_put_qemu_file(f, comp_param[idx].file); |
a7a9a88f | 956 | qemu_mutex_lock(&comp_param[idx].mutex); |
56e93d26 | 957 | set_compress_params(&comp_param[idx], block, offset); |
a7a9a88f LL |
958 | qemu_cond_signal(&comp_param[idx].cond); |
959 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
56e93d26 | 960 | pages = 1; |
b4d1c6e7 | 961 | rs->norm_pages++; |
072c2511 | 962 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
963 | break; |
964 | } | |
965 | } | |
966 | if (pages > 0) { | |
967 | break; | |
968 | } else { | |
0d9f9a5c | 969 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 JQ |
970 | } |
971 | } | |
0d9f9a5c | 972 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
973 | |
974 | return pages; | |
975 | } | |
976 | ||
977 | /** | |
978 | * ram_save_compressed_page: compress the given page and send it to the stream | |
979 | * | |
3d0684b2 | 980 | * Returns the number of pages written. |
56e93d26 | 981 | * |
6f37bb8b | 982 | * @rs: current RAM state |
3d0684b2 | 983 | * @ms: current migration state |
56e93d26 JQ |
984 | * @f: QEMUFile where to send the data |
985 | * @block: block that contains the page we want to send | |
986 | * @offset: offset inside the block for the page | |
987 | * @last_stage: if we are at the completion stage | |
56e93d26 | 988 | */ |
6f37bb8b JQ |
989 | static int ram_save_compressed_page(RAMState *rs, MigrationState *ms, |
990 | QEMUFile *f, | |
072c2511 | 991 | PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
992 | { |
993 | int pages = -1; | |
fc50438e | 994 | uint64_t bytes_xmit = 0; |
56e93d26 | 995 | uint8_t *p; |
fc50438e | 996 | int ret, blen; |
a08f6890 HZ |
997 | RAMBlock *block = pss->block; |
998 | ram_addr_t offset = pss->offset; | |
56e93d26 | 999 | |
2f68e399 | 1000 | p = block->host + offset; |
56e93d26 | 1001 | |
56e93d26 JQ |
1002 | ret = ram_control_save_page(f, block->offset, |
1003 | offset, TARGET_PAGE_SIZE, &bytes_xmit); | |
1004 | if (bytes_xmit) { | |
072c2511 | 1005 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
1006 | pages = 1; |
1007 | } | |
56e93d26 JQ |
1008 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
1009 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
1010 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 1011 | rs->norm_pages++; |
56e93d26 | 1012 | } else if (bytes_xmit == 0) { |
f7ccd61b | 1013 | rs->zero_pages++; |
56e93d26 JQ |
1014 | } |
1015 | } | |
1016 | } else { | |
1017 | /* When starting the process of a new block, the first page of | |
1018 | * the block should be sent out before other pages in the same | |
1019 | * block, and all the pages in last block should have been sent | |
1020 | * out, keeping this order is important, because the 'cont' flag | |
1021 | * is used to avoid resending the block name. | |
1022 | */ | |
6f37bb8b | 1023 | if (block != rs->last_sent_block) { |
2f4fde93 | 1024 | flush_compressed_data(rs, f); |
072c2511 | 1025 | pages = save_zero_page(rs, f, block, offset, p); |
56e93d26 | 1026 | if (pages == -1) { |
fc50438e LL |
1027 | /* Make sure the first page is sent out before other pages */ |
1028 | bytes_xmit = save_page_header(f, block, offset | | |
1029 | RAM_SAVE_FLAG_COMPRESS_PAGE); | |
1030 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, | |
1031 | migrate_compress_level()); | |
1032 | if (blen > 0) { | |
072c2511 | 1033 | rs->bytes_transferred += bytes_xmit + blen; |
b4d1c6e7 | 1034 | rs->norm_pages++; |
b3be2896 | 1035 | pages = 1; |
fc50438e LL |
1036 | } else { |
1037 | qemu_file_set_error(f, blen); | |
1038 | error_report("compressed data failed!"); | |
b3be2896 | 1039 | } |
56e93d26 | 1040 | } |
53f09a10 PB |
1041 | if (pages > 0) { |
1042 | ram_release_pages(ms, block->idstr, pss->offset, pages); | |
1043 | } | |
56e93d26 | 1044 | } else { |
fc50438e | 1045 | offset |= RAM_SAVE_FLAG_CONTINUE; |
072c2511 | 1046 | pages = save_zero_page(rs, f, block, offset, p); |
56e93d26 | 1047 | if (pages == -1) { |
072c2511 | 1048 | pages = compress_page_with_multi_thread(rs, f, block, offset); |
53f09a10 PB |
1049 | } else { |
1050 | ram_release_pages(ms, block->idstr, pss->offset, pages); | |
56e93d26 JQ |
1051 | } |
1052 | } | |
1053 | } | |
1054 | ||
1055 | return pages; | |
1056 | } | |
1057 | ||
3d0684b2 JQ |
1058 | /** |
1059 | * find_dirty_block: find the next dirty page and update any state | |
1060 | * associated with the search process. | |
b9e60928 | 1061 | * |
3d0684b2 | 1062 | * Returns if a page is found |
b9e60928 | 1063 | * |
6f37bb8b | 1064 | * @rs: current RAM state |
3d0684b2 JQ |
1065 | * @f: QEMUFile where to send the data |
1066 | * @pss: data about the state of the current dirty page scan | |
1067 | * @again: set to false if the search has scanned the whole of RAM | |
1068 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
1069 | * within the global ram_addr space | |
b9e60928 | 1070 | */ |
6f37bb8b | 1071 | static bool find_dirty_block(RAMState *rs, QEMUFile *f, PageSearchStatus *pss, |
f3f491fc | 1072 | bool *again, ram_addr_t *ram_addr_abs) |
b9e60928 | 1073 | { |
6f37bb8b | 1074 | pss->offset = migration_bitmap_find_dirty(rs, pss->block, pss->offset, |
a82d593b | 1075 | ram_addr_abs); |
6f37bb8b JQ |
1076 | if (pss->complete_round && pss->block == rs->last_seen_block && |
1077 | pss->offset >= rs->last_offset) { | |
b9e60928 DDAG |
1078 | /* |
1079 | * We've been once around the RAM and haven't found anything. | |
1080 | * Give up. | |
1081 | */ | |
1082 | *again = false; | |
1083 | return false; | |
1084 | } | |
1085 | if (pss->offset >= pss->block->used_length) { | |
1086 | /* Didn't find anything in this RAM Block */ | |
1087 | pss->offset = 0; | |
1088 | pss->block = QLIST_NEXT_RCU(pss->block, next); | |
1089 | if (!pss->block) { | |
1090 | /* Hit the end of the list */ | |
1091 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1092 | /* Flag that we've looped */ | |
1093 | pss->complete_round = true; | |
6f37bb8b | 1094 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1095 | if (migrate_use_xbzrle()) { |
1096 | /* If xbzrle is on, stop using the data compression at this | |
1097 | * point. In theory, xbzrle can do better than compression. | |
1098 | */ | |
2f4fde93 | 1099 | flush_compressed_data(rs, f); |
b9e60928 DDAG |
1100 | compression_switch = false; |
1101 | } | |
1102 | } | |
1103 | /* Didn't find anything this time, but try again on the new block */ | |
1104 | *again = true; | |
1105 | return false; | |
1106 | } else { | |
1107 | /* Can go around again, but... */ | |
1108 | *again = true; | |
1109 | /* We've found something so probably don't need to */ | |
1110 | return true; | |
1111 | } | |
1112 | } | |
1113 | ||
3d0684b2 JQ |
1114 | /** |
1115 | * unqueue_page: gets a page of the queue | |
1116 | * | |
a82d593b | 1117 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1118 | * |
3d0684b2 JQ |
1119 | * Returns the block of the page (or NULL if none available) |
1120 | * | |
ec481c6c | 1121 | * @rs: current RAM state |
3d0684b2 JQ |
1122 | * @offset: used to return the offset within the RAMBlock |
1123 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
1124 | * within the global ram_addr space | |
a82d593b | 1125 | */ |
ec481c6c | 1126 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset, |
a82d593b DDAG |
1127 | ram_addr_t *ram_addr_abs) |
1128 | { | |
1129 | RAMBlock *block = NULL; | |
1130 | ||
ec481c6c JQ |
1131 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1132 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
1133 | struct RAMSrcPageRequest *entry = | |
1134 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1135 | block = entry->rb; |
1136 | *offset = entry->offset; | |
1137 | *ram_addr_abs = (entry->offset + entry->rb->offset) & | |
1138 | TARGET_PAGE_MASK; | |
1139 | ||
1140 | if (entry->len > TARGET_PAGE_SIZE) { | |
1141 | entry->len -= TARGET_PAGE_SIZE; | |
1142 | entry->offset += TARGET_PAGE_SIZE; | |
1143 | } else { | |
1144 | memory_region_unref(block->mr); | |
ec481c6c | 1145 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b DDAG |
1146 | g_free(entry); |
1147 | } | |
1148 | } | |
ec481c6c | 1149 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
a82d593b DDAG |
1150 | |
1151 | return block; | |
1152 | } | |
1153 | ||
3d0684b2 JQ |
1154 | /** |
1155 | * get_queued_page: unqueue a page from the postocpy requests | |
1156 | * | |
1157 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1158 | * |
3d0684b2 | 1159 | * Returns if a queued page is found |
a82d593b | 1160 | * |
6f37bb8b | 1161 | * @rs: current RAM state |
3d0684b2 JQ |
1162 | * @pss: data about the state of the current dirty page scan |
1163 | * @ram_addr_abs: pointer into which to store the address of the dirty page | |
1164 | * within the global ram_addr space | |
a82d593b | 1165 | */ |
ec481c6c | 1166 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss, |
a82d593b DDAG |
1167 | ram_addr_t *ram_addr_abs) |
1168 | { | |
1169 | RAMBlock *block; | |
1170 | ram_addr_t offset; | |
1171 | bool dirty; | |
1172 | ||
1173 | do { | |
ec481c6c | 1174 | block = unqueue_page(rs, &offset, ram_addr_abs); |
a82d593b DDAG |
1175 | /* |
1176 | * We're sending this page, and since it's postcopy nothing else | |
1177 | * will dirty it, and we must make sure it doesn't get sent again | |
1178 | * even if this queue request was received after the background | |
1179 | * search already sent it. | |
1180 | */ | |
1181 | if (block) { | |
1182 | unsigned long *bitmap; | |
eb859c53 | 1183 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
a82d593b DDAG |
1184 | dirty = test_bit(*ram_addr_abs >> TARGET_PAGE_BITS, bitmap); |
1185 | if (!dirty) { | |
1186 | trace_get_queued_page_not_dirty( | |
1187 | block->idstr, (uint64_t)offset, | |
1188 | (uint64_t)*ram_addr_abs, | |
1189 | test_bit(*ram_addr_abs >> TARGET_PAGE_BITS, | |
eb859c53 | 1190 | atomic_rcu_read(&rs->ram_bitmap)->unsentmap)); |
a82d593b DDAG |
1191 | } else { |
1192 | trace_get_queued_page(block->idstr, | |
1193 | (uint64_t)offset, | |
1194 | (uint64_t)*ram_addr_abs); | |
1195 | } | |
1196 | } | |
1197 | ||
1198 | } while (block && !dirty); | |
1199 | ||
1200 | if (block) { | |
1201 | /* | |
1202 | * As soon as we start servicing pages out of order, then we have | |
1203 | * to kill the bulk stage, since the bulk stage assumes | |
1204 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1205 | * dirty, that's no longer true. | |
1206 | */ | |
6f37bb8b | 1207 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1208 | |
1209 | /* | |
1210 | * We want the background search to continue from the queued page | |
1211 | * since the guest is likely to want other pages near to the page | |
1212 | * it just requested. | |
1213 | */ | |
1214 | pss->block = block; | |
1215 | pss->offset = offset; | |
1216 | } | |
1217 | ||
1218 | return !!block; | |
1219 | } | |
1220 | ||
6c595cde | 1221 | /** |
5e58f968 JQ |
1222 | * migration_page_queue_free: drop any remaining pages in the ram |
1223 | * request queue | |
6c595cde | 1224 | * |
3d0684b2 JQ |
1225 | * It should be empty at the end anyway, but in error cases there may |
1226 | * be some left. in case that there is any page left, we drop it. | |
1227 | * | |
6c595cde | 1228 | */ |
ec481c6c | 1229 | void migration_page_queue_free(void) |
6c595cde | 1230 | { |
ec481c6c JQ |
1231 | struct RAMSrcPageRequest *mspr, *next_mspr; |
1232 | RAMState *rs = &ram_state; | |
6c595cde DDAG |
1233 | /* This queue generally should be empty - but in the case of a failed |
1234 | * migration might have some droppings in. | |
1235 | */ | |
1236 | rcu_read_lock(); | |
ec481c6c | 1237 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1238 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1239 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1240 | g_free(mspr); |
1241 | } | |
1242 | rcu_read_unlock(); | |
1243 | } | |
1244 | ||
1245 | /** | |
3d0684b2 JQ |
1246 | * ram_save_queue_pages: queue the page for transmission |
1247 | * | |
1248 | * A request from postcopy destination for example. | |
1249 | * | |
1250 | * Returns zero on success or negative on error | |
1251 | * | |
3d0684b2 JQ |
1252 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1253 | * same that last one. | |
1254 | * @start: starting address from the start of the RAMBlock | |
1255 | * @len: length (in bytes) to send | |
6c595cde | 1256 | */ |
96506894 | 1257 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1258 | { |
1259 | RAMBlock *ramblock; | |
68a098f3 | 1260 | RAMState *rs = &ram_state; |
6c595cde | 1261 | |
96506894 | 1262 | rs->postcopy_requests++; |
6c595cde DDAG |
1263 | rcu_read_lock(); |
1264 | if (!rbname) { | |
1265 | /* Reuse last RAMBlock */ | |
68a098f3 | 1266 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1267 | |
1268 | if (!ramblock) { | |
1269 | /* | |
1270 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1271 | * it's the 1st request. | |
1272 | */ | |
1273 | error_report("ram_save_queue_pages no previous block"); | |
1274 | goto err; | |
1275 | } | |
1276 | } else { | |
1277 | ramblock = qemu_ram_block_by_name(rbname); | |
1278 | ||
1279 | if (!ramblock) { | |
1280 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1281 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1282 | goto err; | |
1283 | } | |
68a098f3 | 1284 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1285 | } |
1286 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1287 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1288 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1289 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1290 | __func__, start, len, ramblock->used_length); |
1291 | goto err; | |
1292 | } | |
1293 | ||
ec481c6c JQ |
1294 | struct RAMSrcPageRequest *new_entry = |
1295 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
1296 | new_entry->rb = ramblock; |
1297 | new_entry->offset = start; | |
1298 | new_entry->len = len; | |
1299 | ||
1300 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
1301 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1302 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
1303 | qemu_mutex_unlock(&rs->src_page_req_mutex); | |
6c595cde DDAG |
1304 | rcu_read_unlock(); |
1305 | ||
1306 | return 0; | |
1307 | ||
1308 | err: | |
1309 | rcu_read_unlock(); | |
1310 | return -1; | |
1311 | } | |
1312 | ||
a82d593b | 1313 | /** |
3d0684b2 | 1314 | * ram_save_target_page: save one target page |
a82d593b | 1315 | * |
3d0684b2 | 1316 | * Returns the number of pages written |
a82d593b | 1317 | * |
6f37bb8b | 1318 | * @rs: current RAM state |
3d0684b2 | 1319 | * @ms: current migration state |
a82d593b | 1320 | * @f: QEMUFile where to send the data |
3d0684b2 | 1321 | * @pss: data about the page we want to send |
a82d593b | 1322 | * @last_stage: if we are at the completion stage |
3d0684b2 | 1323 | * @dirty_ram_abs: address of the start of the dirty page in ram_addr_t space |
a82d593b | 1324 | */ |
6f37bb8b | 1325 | static int ram_save_target_page(RAMState *rs, MigrationState *ms, QEMUFile *f, |
a08f6890 | 1326 | PageSearchStatus *pss, |
a82d593b | 1327 | bool last_stage, |
a82d593b DDAG |
1328 | ram_addr_t dirty_ram_abs) |
1329 | { | |
1330 | int res = 0; | |
1331 | ||
1332 | /* Check the pages is dirty and if it is send it */ | |
0d8ec885 | 1333 | if (migration_bitmap_clear_dirty(rs, dirty_ram_abs)) { |
a82d593b DDAG |
1334 | unsigned long *unsentmap; |
1335 | if (compression_switch && migrate_use_compression()) { | |
072c2511 | 1336 | res = ram_save_compressed_page(rs, ms, f, pss, last_stage); |
a82d593b | 1337 | } else { |
072c2511 | 1338 | res = ram_save_page(rs, ms, f, pss, last_stage); |
a82d593b DDAG |
1339 | } |
1340 | ||
1341 | if (res < 0) { | |
1342 | return res; | |
1343 | } | |
eb859c53 | 1344 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
a82d593b DDAG |
1345 | if (unsentmap) { |
1346 | clear_bit(dirty_ram_abs >> TARGET_PAGE_BITS, unsentmap); | |
1347 | } | |
3fd3c4b3 DDAG |
1348 | /* Only update last_sent_block if a block was actually sent; xbzrle |
1349 | * might have decided the page was identical so didn't bother writing | |
1350 | * to the stream. | |
1351 | */ | |
1352 | if (res > 0) { | |
6f37bb8b | 1353 | rs->last_sent_block = pss->block; |
3fd3c4b3 | 1354 | } |
a82d593b DDAG |
1355 | } |
1356 | ||
1357 | return res; | |
1358 | } | |
1359 | ||
1360 | /** | |
3d0684b2 | 1361 | * ram_save_host_page: save a whole host page |
a82d593b | 1362 | * |
3d0684b2 JQ |
1363 | * Starting at *offset send pages up to the end of the current host |
1364 | * page. It's valid for the initial offset to point into the middle of | |
1365 | * a host page in which case the remainder of the hostpage is sent. | |
1366 | * Only dirty target pages are sent. Note that the host page size may | |
1367 | * be a huge page for this block. | |
a82d593b | 1368 | * |
3d0684b2 JQ |
1369 | * Returns the number of pages written or negative on error |
1370 | * | |
6f37bb8b | 1371 | * @rs: current RAM state |
3d0684b2 | 1372 | * @ms: current migration state |
a82d593b | 1373 | * @f: QEMUFile where to send the data |
3d0684b2 | 1374 | * @pss: data about the page we want to send |
a82d593b | 1375 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1376 | * @dirty_ram_abs: Address of the start of the dirty page in ram_addr_t space |
1377 | */ | |
6f37bb8b | 1378 | static int ram_save_host_page(RAMState *rs, MigrationState *ms, QEMUFile *f, |
a08f6890 HZ |
1379 | PageSearchStatus *pss, |
1380 | bool last_stage, | |
a82d593b DDAG |
1381 | ram_addr_t dirty_ram_abs) |
1382 | { | |
1383 | int tmppages, pages = 0; | |
4c011c37 DDAG |
1384 | size_t pagesize = qemu_ram_pagesize(pss->block); |
1385 | ||
a82d593b | 1386 | do { |
6f37bb8b | 1387 | tmppages = ram_save_target_page(rs, ms, f, pss, last_stage, |
072c2511 | 1388 | dirty_ram_abs); |
a82d593b DDAG |
1389 | if (tmppages < 0) { |
1390 | return tmppages; | |
1391 | } | |
1392 | ||
1393 | pages += tmppages; | |
a08f6890 | 1394 | pss->offset += TARGET_PAGE_SIZE; |
a82d593b | 1395 | dirty_ram_abs += TARGET_PAGE_SIZE; |
4c011c37 | 1396 | } while (pss->offset & (pagesize - 1)); |
a82d593b DDAG |
1397 | |
1398 | /* The offset we leave with is the last one we looked at */ | |
a08f6890 | 1399 | pss->offset -= TARGET_PAGE_SIZE; |
a82d593b DDAG |
1400 | return pages; |
1401 | } | |
6c595cde | 1402 | |
56e93d26 | 1403 | /** |
3d0684b2 | 1404 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1405 | * |
1406 | * Called within an RCU critical section. | |
1407 | * | |
3d0684b2 | 1408 | * Returns the number of pages written where zero means no dirty pages |
56e93d26 | 1409 | * |
6f37bb8b | 1410 | * @rs: current RAM state |
56e93d26 JQ |
1411 | * @f: QEMUFile where to send the data |
1412 | * @last_stage: if we are at the completion stage | |
a82d593b DDAG |
1413 | * |
1414 | * On systems where host-page-size > target-page-size it will send all the | |
1415 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1416 | */ |
1417 | ||
072c2511 | 1418 | static int ram_find_and_save_block(RAMState *rs, QEMUFile *f, bool last_stage) |
56e93d26 | 1419 | { |
b8fb8cb7 | 1420 | PageSearchStatus pss; |
a82d593b | 1421 | MigrationState *ms = migrate_get_current(); |
56e93d26 | 1422 | int pages = 0; |
b9e60928 | 1423 | bool again, found; |
f3f491fc DDAG |
1424 | ram_addr_t dirty_ram_abs; /* Address of the start of the dirty page in |
1425 | ram_addr_t space */ | |
56e93d26 | 1426 | |
0827b9e9 AA |
1427 | /* No dirty page as there is zero RAM */ |
1428 | if (!ram_bytes_total()) { | |
1429 | return pages; | |
1430 | } | |
1431 | ||
6f37bb8b JQ |
1432 | pss.block = rs->last_seen_block; |
1433 | pss.offset = rs->last_offset; | |
b8fb8cb7 DDAG |
1434 | pss.complete_round = false; |
1435 | ||
1436 | if (!pss.block) { | |
1437 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1438 | } | |
56e93d26 | 1439 | |
b9e60928 | 1440 | do { |
a82d593b | 1441 | again = true; |
ec481c6c | 1442 | found = get_queued_page(rs, &pss, &dirty_ram_abs); |
b9e60928 | 1443 | |
a82d593b DDAG |
1444 | if (!found) { |
1445 | /* priority queue empty, so just search for something dirty */ | |
6f37bb8b | 1446 | found = find_dirty_block(rs, f, &pss, &again, &dirty_ram_abs); |
a82d593b | 1447 | } |
f3f491fc | 1448 | |
a82d593b | 1449 | if (found) { |
072c2511 | 1450 | pages = ram_save_host_page(rs, ms, f, &pss, last_stage, |
a82d593b | 1451 | dirty_ram_abs); |
56e93d26 | 1452 | } |
b9e60928 | 1453 | } while (!pages && again); |
56e93d26 | 1454 | |
6f37bb8b JQ |
1455 | rs->last_seen_block = pss.block; |
1456 | rs->last_offset = pss.offset; | |
56e93d26 JQ |
1457 | |
1458 | return pages; | |
1459 | } | |
1460 | ||
1461 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1462 | { | |
1463 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b JQ |
1464 | RAMState *rs = &ram_state; |
1465 | ||
56e93d26 | 1466 | if (zero) { |
f7ccd61b | 1467 | rs->zero_pages += pages; |
56e93d26 | 1468 | } else { |
b4d1c6e7 | 1469 | rs->norm_pages += pages; |
2f4fde93 | 1470 | rs->bytes_transferred += size; |
56e93d26 JQ |
1471 | qemu_update_position(f, size); |
1472 | } | |
1473 | } | |
1474 | ||
56e93d26 JQ |
1475 | uint64_t ram_bytes_total(void) |
1476 | { | |
1477 | RAMBlock *block; | |
1478 | uint64_t total = 0; | |
1479 | ||
1480 | rcu_read_lock(); | |
1481 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) | |
1482 | total += block->used_length; | |
1483 | rcu_read_unlock(); | |
1484 | return total; | |
1485 | } | |
1486 | ||
1487 | void free_xbzrle_decoded_buf(void) | |
1488 | { | |
1489 | g_free(xbzrle_decoded_buf); | |
1490 | xbzrle_decoded_buf = NULL; | |
1491 | } | |
1492 | ||
eb859c53 | 1493 | static void migration_bitmap_free(struct RAMBitmap *bmap) |
60be6340 DL |
1494 | { |
1495 | g_free(bmap->bmap); | |
f3f491fc | 1496 | g_free(bmap->unsentmap); |
60be6340 DL |
1497 | g_free(bmap); |
1498 | } | |
1499 | ||
6ad2a215 | 1500 | static void ram_migration_cleanup(void *opaque) |
56e93d26 | 1501 | { |
eb859c53 JQ |
1502 | RAMState *rs = opaque; |
1503 | ||
2ff64038 LZ |
1504 | /* caller have hold iothread lock or is in a bh, so there is |
1505 | * no writing race against this migration_bitmap | |
1506 | */ | |
eb859c53 JQ |
1507 | struct RAMBitmap *bitmap = rs->ram_bitmap; |
1508 | atomic_rcu_set(&rs->ram_bitmap, NULL); | |
2ff64038 | 1509 | if (bitmap) { |
56e93d26 | 1510 | memory_global_dirty_log_stop(); |
60be6340 | 1511 | call_rcu(bitmap, migration_bitmap_free, rcu); |
56e93d26 JQ |
1512 | } |
1513 | ||
1514 | XBZRLE_cache_lock(); | |
1515 | if (XBZRLE.cache) { | |
1516 | cache_fini(XBZRLE.cache); | |
1517 | g_free(XBZRLE.encoded_buf); | |
1518 | g_free(XBZRLE.current_buf); | |
adb65dec | 1519 | g_free(ZERO_TARGET_PAGE); |
56e93d26 JQ |
1520 | XBZRLE.cache = NULL; |
1521 | XBZRLE.encoded_buf = NULL; | |
1522 | XBZRLE.current_buf = NULL; | |
1523 | } | |
1524 | XBZRLE_cache_unlock(); | |
1525 | } | |
1526 | ||
6f37bb8b | 1527 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1528 | { |
6f37bb8b JQ |
1529 | rs->last_seen_block = NULL; |
1530 | rs->last_sent_block = NULL; | |
1531 | rs->last_offset = 0; | |
1532 | rs->last_version = ram_list.version; | |
1533 | rs->ram_bulk_stage = true; | |
56e93d26 JQ |
1534 | } |
1535 | ||
1536 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1537 | ||
dd631697 LZ |
1538 | void migration_bitmap_extend(ram_addr_t old, ram_addr_t new) |
1539 | { | |
0d8ec885 | 1540 | RAMState *rs = &ram_state; |
108cfae0 | 1541 | |
dd631697 LZ |
1542 | /* called in qemu main thread, so there is |
1543 | * no writing race against this migration_bitmap | |
1544 | */ | |
eb859c53 JQ |
1545 | if (rs->ram_bitmap) { |
1546 | struct RAMBitmap *old_bitmap = rs->ram_bitmap, *bitmap; | |
1547 | bitmap = g_new(struct RAMBitmap, 1); | |
60be6340 | 1548 | bitmap->bmap = bitmap_new(new); |
dd631697 LZ |
1549 | |
1550 | /* prevent migration_bitmap content from being set bit | |
1551 | * by migration_bitmap_sync_range() at the same time. | |
1552 | * it is safe to migration if migration_bitmap is cleared bit | |
1553 | * at the same time. | |
1554 | */ | |
108cfae0 | 1555 | qemu_mutex_lock(&rs->bitmap_mutex); |
60be6340 DL |
1556 | bitmap_copy(bitmap->bmap, old_bitmap->bmap, old); |
1557 | bitmap_set(bitmap->bmap, old, new - old); | |
f3f491fc DDAG |
1558 | |
1559 | /* We don't have a way to safely extend the sentmap | |
1560 | * with RCU; so mark it as missing, entry to postcopy | |
1561 | * will fail. | |
1562 | */ | |
1563 | bitmap->unsentmap = NULL; | |
1564 | ||
eb859c53 | 1565 | atomic_rcu_set(&rs->ram_bitmap, bitmap); |
108cfae0 | 1566 | qemu_mutex_unlock(&rs->bitmap_mutex); |
0d8ec885 | 1567 | rs->migration_dirty_pages += new - old; |
60be6340 | 1568 | call_rcu(old_bitmap, migration_bitmap_free, rcu); |
dd631697 LZ |
1569 | } |
1570 | } | |
56e93d26 | 1571 | |
4f2e4252 DDAG |
1572 | /* |
1573 | * 'expected' is the value you expect the bitmap mostly to be full | |
1574 | * of; it won't bother printing lines that are all this value. | |
1575 | * If 'todump' is null the migration bitmap is dumped. | |
1576 | */ | |
1577 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected) | |
1578 | { | |
1579 | int64_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
eb859c53 | 1580 | RAMState *rs = &ram_state; |
4f2e4252 DDAG |
1581 | int64_t cur; |
1582 | int64_t linelen = 128; | |
1583 | char linebuf[129]; | |
1584 | ||
1585 | if (!todump) { | |
eb859c53 | 1586 | todump = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
4f2e4252 DDAG |
1587 | } |
1588 | ||
1589 | for (cur = 0; cur < ram_pages; cur += linelen) { | |
1590 | int64_t curb; | |
1591 | bool found = false; | |
1592 | /* | |
1593 | * Last line; catch the case where the line length | |
1594 | * is longer than remaining ram | |
1595 | */ | |
1596 | if (cur + linelen > ram_pages) { | |
1597 | linelen = ram_pages - cur; | |
1598 | } | |
1599 | for (curb = 0; curb < linelen; curb++) { | |
1600 | bool thisbit = test_bit(cur + curb, todump); | |
1601 | linebuf[curb] = thisbit ? '1' : '.'; | |
1602 | found = found || (thisbit != expected); | |
1603 | } | |
1604 | if (found) { | |
1605 | linebuf[curb] = '\0'; | |
1606 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1607 | } | |
1608 | } | |
1609 | } | |
1610 | ||
e0b266f0 DDAG |
1611 | /* **** functions for postcopy ***** */ |
1612 | ||
ced1c616 PB |
1613 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1614 | { | |
eb859c53 | 1615 | RAMState *rs = &ram_state; |
ced1c616 | 1616 | struct RAMBlock *block; |
eb859c53 | 1617 | unsigned long *bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
ced1c616 PB |
1618 | |
1619 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1620 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1621 | unsigned long range = first + (block->used_length >> TARGET_PAGE_BITS); | |
1622 | unsigned long run_start = find_next_zero_bit(bitmap, range, first); | |
1623 | ||
1624 | while (run_start < range) { | |
1625 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
1626 | ram_discard_range(NULL, block->idstr, run_start << TARGET_PAGE_BITS, | |
1627 | (run_end - run_start) << TARGET_PAGE_BITS); | |
1628 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); | |
1629 | } | |
1630 | } | |
1631 | } | |
1632 | ||
3d0684b2 JQ |
1633 | /** |
1634 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1635 | * | |
1636 | * Returns zero on success | |
1637 | * | |
e0b266f0 DDAG |
1638 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
1639 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1640 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
3d0684b2 JQ |
1641 | * |
1642 | * @ms: current migration state | |
1643 | * @pds: state for postcopy | |
1644 | * @start: RAMBlock starting page | |
1645 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1646 | */ |
1647 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1648 | PostcopyDiscardState *pds, | |
1649 | unsigned long start, | |
1650 | unsigned long length) | |
1651 | { | |
eb859c53 | 1652 | RAMState *rs = &ram_state; |
e0b266f0 DDAG |
1653 | unsigned long end = start + length; /* one after the end */ |
1654 | unsigned long current; | |
1655 | unsigned long *unsentmap; | |
1656 | ||
eb859c53 | 1657 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
e0b266f0 DDAG |
1658 | for (current = start; current < end; ) { |
1659 | unsigned long one = find_next_bit(unsentmap, end, current); | |
1660 | ||
1661 | if (one <= end) { | |
1662 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1663 | unsigned long discard_length; | |
1664 | ||
1665 | if (zero >= end) { | |
1666 | discard_length = end - one; | |
1667 | } else { | |
1668 | discard_length = zero - one; | |
1669 | } | |
d688c62d DDAG |
1670 | if (discard_length) { |
1671 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1672 | } | |
e0b266f0 DDAG |
1673 | current = one + discard_length; |
1674 | } else { | |
1675 | current = one; | |
1676 | } | |
1677 | } | |
1678 | ||
1679 | return 0; | |
1680 | } | |
1681 | ||
3d0684b2 JQ |
1682 | /** |
1683 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1684 | * | |
1685 | * Returns 0 for success or negative for error | |
1686 | * | |
e0b266f0 DDAG |
1687 | * Utility for the outgoing postcopy code. |
1688 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1689 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1690 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1691 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1692 | * |
1693 | * @ms: current migration state | |
e0b266f0 DDAG |
1694 | */ |
1695 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1696 | { | |
1697 | struct RAMBlock *block; | |
1698 | int ret; | |
1699 | ||
1700 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1701 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1702 | PostcopyDiscardState *pds = postcopy_discard_send_init(ms, | |
1703 | first, | |
1704 | block->idstr); | |
1705 | ||
1706 | /* | |
1707 | * Postcopy sends chunks of bitmap over the wire, but it | |
1708 | * just needs indexes at this point, avoids it having | |
1709 | * target page specific code. | |
1710 | */ | |
1711 | ret = postcopy_send_discard_bm_ram(ms, pds, first, | |
1712 | block->used_length >> TARGET_PAGE_BITS); | |
1713 | postcopy_discard_send_finish(ms, pds); | |
1714 | if (ret) { | |
1715 | return ret; | |
1716 | } | |
1717 | } | |
1718 | ||
1719 | return 0; | |
1720 | } | |
1721 | ||
3d0684b2 JQ |
1722 | /** |
1723 | * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages | |
1724 | * | |
1725 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
1726 | * canonicalize the two bitmaps, that are similar, but one is | |
1727 | * inverted. | |
99e314eb | 1728 | * |
3d0684b2 JQ |
1729 | * Postcopy requires that all target pages in a hostpage are dirty or |
1730 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 1731 | * |
3d0684b2 JQ |
1732 | * @ms: current migration state |
1733 | * @unsent_pass: if true we need to canonicalize partially unsent host pages | |
1734 | * otherwise we need to canonicalize partially dirty host pages | |
1735 | * @block: block that contains the page we want to canonicalize | |
1736 | * @pds: state for postcopy | |
99e314eb DDAG |
1737 | */ |
1738 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1739 | RAMBlock *block, | |
1740 | PostcopyDiscardState *pds) | |
1741 | { | |
0d8ec885 | 1742 | RAMState *rs = &ram_state; |
99e314eb DDAG |
1743 | unsigned long *bitmap; |
1744 | unsigned long *unsentmap; | |
29c59172 | 1745 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
99e314eb DDAG |
1746 | unsigned long first = block->offset >> TARGET_PAGE_BITS; |
1747 | unsigned long len = block->used_length >> TARGET_PAGE_BITS; | |
1748 | unsigned long last = first + (len - 1); | |
1749 | unsigned long run_start; | |
1750 | ||
29c59172 DDAG |
1751 | if (block->page_size == TARGET_PAGE_SIZE) { |
1752 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
1753 | return; | |
1754 | } | |
1755 | ||
eb859c53 JQ |
1756 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
1757 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; | |
99e314eb DDAG |
1758 | |
1759 | if (unsent_pass) { | |
1760 | /* Find a sent page */ | |
1761 | run_start = find_next_zero_bit(unsentmap, last + 1, first); | |
1762 | } else { | |
1763 | /* Find a dirty page */ | |
1764 | run_start = find_next_bit(bitmap, last + 1, first); | |
1765 | } | |
1766 | ||
1767 | while (run_start <= last) { | |
1768 | bool do_fixup = false; | |
1769 | unsigned long fixup_start_addr; | |
1770 | unsigned long host_offset; | |
1771 | ||
1772 | /* | |
1773 | * If the start of this run of pages is in the middle of a host | |
1774 | * page, then we need to fixup this host page. | |
1775 | */ | |
1776 | host_offset = run_start % host_ratio; | |
1777 | if (host_offset) { | |
1778 | do_fixup = true; | |
1779 | run_start -= host_offset; | |
1780 | fixup_start_addr = run_start; | |
1781 | /* For the next pass */ | |
1782 | run_start = run_start + host_ratio; | |
1783 | } else { | |
1784 | /* Find the end of this run */ | |
1785 | unsigned long run_end; | |
1786 | if (unsent_pass) { | |
1787 | run_end = find_next_bit(unsentmap, last + 1, run_start + 1); | |
1788 | } else { | |
1789 | run_end = find_next_zero_bit(bitmap, last + 1, run_start + 1); | |
1790 | } | |
1791 | /* | |
1792 | * If the end isn't at the start of a host page, then the | |
1793 | * run doesn't finish at the end of a host page | |
1794 | * and we need to discard. | |
1795 | */ | |
1796 | host_offset = run_end % host_ratio; | |
1797 | if (host_offset) { | |
1798 | do_fixup = true; | |
1799 | fixup_start_addr = run_end - host_offset; | |
1800 | /* | |
1801 | * This host page has gone, the next loop iteration starts | |
1802 | * from after the fixup | |
1803 | */ | |
1804 | run_start = fixup_start_addr + host_ratio; | |
1805 | } else { | |
1806 | /* | |
1807 | * No discards on this iteration, next loop starts from | |
1808 | * next sent/dirty page | |
1809 | */ | |
1810 | run_start = run_end + 1; | |
1811 | } | |
1812 | } | |
1813 | ||
1814 | if (do_fixup) { | |
1815 | unsigned long page; | |
1816 | ||
1817 | /* Tell the destination to discard this page */ | |
1818 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1819 | /* For the unsent_pass we: | |
1820 | * discard partially sent pages | |
1821 | * For the !unsent_pass (dirty) we: | |
1822 | * discard partially dirty pages that were sent | |
1823 | * (any partially sent pages were already discarded | |
1824 | * by the previous unsent_pass) | |
1825 | */ | |
1826 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1827 | host_ratio); | |
1828 | } | |
1829 | ||
1830 | /* Clean up the bitmap */ | |
1831 | for (page = fixup_start_addr; | |
1832 | page < fixup_start_addr + host_ratio; page++) { | |
1833 | /* All pages in this host page are now not sent */ | |
1834 | set_bit(page, unsentmap); | |
1835 | ||
1836 | /* | |
1837 | * Remark them as dirty, updating the count for any pages | |
1838 | * that weren't previously dirty. | |
1839 | */ | |
0d8ec885 | 1840 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
1841 | } |
1842 | } | |
1843 | ||
1844 | if (unsent_pass) { | |
1845 | /* Find the next sent page for the next iteration */ | |
1846 | run_start = find_next_zero_bit(unsentmap, last + 1, | |
1847 | run_start); | |
1848 | } else { | |
1849 | /* Find the next dirty page for the next iteration */ | |
1850 | run_start = find_next_bit(bitmap, last + 1, run_start); | |
1851 | } | |
1852 | } | |
1853 | } | |
1854 | ||
3d0684b2 JQ |
1855 | /** |
1856 | * postcopy_chuck_hostpages: discrad any partially sent host page | |
1857 | * | |
99e314eb DDAG |
1858 | * Utility for the outgoing postcopy code. |
1859 | * | |
1860 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
1861 | * dirty host-page size chunks as all dirty. In this case the host-page |
1862 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 1863 | * |
3d0684b2 JQ |
1864 | * Returns zero on success |
1865 | * | |
1866 | * @ms: current migration state | |
99e314eb DDAG |
1867 | */ |
1868 | static int postcopy_chunk_hostpages(MigrationState *ms) | |
1869 | { | |
6f37bb8b | 1870 | RAMState *rs = &ram_state; |
99e314eb DDAG |
1871 | struct RAMBlock *block; |
1872 | ||
99e314eb | 1873 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
6f37bb8b JQ |
1874 | rs->last_seen_block = NULL; |
1875 | rs->last_sent_block = NULL; | |
1876 | rs->last_offset = 0; | |
99e314eb DDAG |
1877 | |
1878 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1879 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1880 | ||
1881 | PostcopyDiscardState *pds = | |
1882 | postcopy_discard_send_init(ms, first, block->idstr); | |
1883 | ||
1884 | /* First pass: Discard all partially sent host pages */ | |
1885 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1886 | /* | |
1887 | * Second pass: Ensure that all partially dirty host pages are made | |
1888 | * fully dirty. | |
1889 | */ | |
1890 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
1891 | ||
1892 | postcopy_discard_send_finish(ms, pds); | |
1893 | } /* ram_list loop */ | |
1894 | ||
1895 | return 0; | |
1896 | } | |
1897 | ||
3d0684b2 JQ |
1898 | /** |
1899 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
1900 | * | |
1901 | * Returns zero on success | |
1902 | * | |
e0b266f0 DDAG |
1903 | * Transmit the set of pages to be discarded after precopy to the target |
1904 | * these are pages that: | |
1905 | * a) Have been previously transmitted but are now dirty again | |
1906 | * b) Pages that have never been transmitted, this ensures that | |
1907 | * any pages on the destination that have been mapped by background | |
1908 | * tasks get discarded (transparent huge pages is the specific concern) | |
1909 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
1910 | * |
1911 | * @ms: current migration state | |
e0b266f0 DDAG |
1912 | */ |
1913 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1914 | { | |
eb859c53 | 1915 | RAMState *rs = &ram_state; |
e0b266f0 DDAG |
1916 | int ret; |
1917 | unsigned long *bitmap, *unsentmap; | |
1918 | ||
1919 | rcu_read_lock(); | |
1920 | ||
1921 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 1922 | migration_bitmap_sync(rs); |
e0b266f0 | 1923 | |
eb859c53 | 1924 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
e0b266f0 DDAG |
1925 | if (!unsentmap) { |
1926 | /* We don't have a safe way to resize the sentmap, so | |
1927 | * if the bitmap was resized it will be NULL at this | |
1928 | * point. | |
1929 | */ | |
1930 | error_report("migration ram resized during precopy phase"); | |
1931 | rcu_read_unlock(); | |
1932 | return -EINVAL; | |
1933 | } | |
1934 | ||
29c59172 | 1935 | /* Deal with TPS != HPS and huge pages */ |
99e314eb DDAG |
1936 | ret = postcopy_chunk_hostpages(ms); |
1937 | if (ret) { | |
1938 | rcu_read_unlock(); | |
1939 | return ret; | |
1940 | } | |
1941 | ||
e0b266f0 DDAG |
1942 | /* |
1943 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1944 | */ | |
eb859c53 | 1945 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
e0b266f0 DDAG |
1946 | bitmap_or(unsentmap, unsentmap, bitmap, |
1947 | last_ram_offset() >> TARGET_PAGE_BITS); | |
1948 | ||
1949 | ||
1950 | trace_ram_postcopy_send_discard_bitmap(); | |
1951 | #ifdef DEBUG_POSTCOPY | |
1952 | ram_debug_dump_bitmap(unsentmap, true); | |
1953 | #endif | |
1954 | ||
1955 | ret = postcopy_each_ram_send_discard(ms); | |
1956 | rcu_read_unlock(); | |
1957 | ||
1958 | return ret; | |
1959 | } | |
1960 | ||
3d0684b2 JQ |
1961 | /** |
1962 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 1963 | * |
3d0684b2 | 1964 | * Returns zero on success |
e0b266f0 | 1965 | * |
3d0684b2 | 1966 | * @mis: current migration incoming state |
36449157 JQ |
1967 | * @rbname: name of the RAMBlock of the request. NULL means the |
1968 | * same that last one. | |
3d0684b2 JQ |
1969 | * @start: RAMBlock starting page |
1970 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1971 | */ |
1972 | int ram_discard_range(MigrationIncomingState *mis, | |
36449157 | 1973 | const char *rbname, |
e0b266f0 DDAG |
1974 | uint64_t start, size_t length) |
1975 | { | |
1976 | int ret = -1; | |
1977 | ||
36449157 | 1978 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 1979 | |
e0b266f0 | 1980 | rcu_read_lock(); |
36449157 | 1981 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
1982 | |
1983 | if (!rb) { | |
36449157 | 1984 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
e0b266f0 DDAG |
1985 | goto err; |
1986 | } | |
1987 | ||
d3a5038c | 1988 | ret = ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
1989 | |
1990 | err: | |
1991 | rcu_read_unlock(); | |
1992 | ||
1993 | return ret; | |
1994 | } | |
1995 | ||
ceb4d168 | 1996 | static int ram_state_init(RAMState *rs) |
56e93d26 | 1997 | { |
56e93d26 JQ |
1998 | int64_t ram_bitmap_pages; /* Size of bitmap in pages, including gaps */ |
1999 | ||
ceb4d168 | 2000 | memset(rs, 0, sizeof(*rs)); |
108cfae0 | 2001 | qemu_mutex_init(&rs->bitmap_mutex); |
ec481c6c JQ |
2002 | qemu_mutex_init(&rs->src_page_req_mutex); |
2003 | QSIMPLEQ_INIT(&rs->src_page_requests); | |
56e93d26 JQ |
2004 | |
2005 | if (migrate_use_xbzrle()) { | |
2006 | XBZRLE_cache_lock(); | |
adb65dec | 2007 | ZERO_TARGET_PAGE = g_malloc0(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2008 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() / |
2009 | TARGET_PAGE_SIZE, | |
2010 | TARGET_PAGE_SIZE); | |
2011 | if (!XBZRLE.cache) { | |
2012 | XBZRLE_cache_unlock(); | |
2013 | error_report("Error creating cache"); | |
2014 | return -1; | |
2015 | } | |
2016 | XBZRLE_cache_unlock(); | |
2017 | ||
2018 | /* We prefer not to abort if there is no memory */ | |
2019 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2020 | if (!XBZRLE.encoded_buf) { | |
2021 | error_report("Error allocating encoded_buf"); | |
2022 | return -1; | |
2023 | } | |
2024 | ||
2025 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2026 | if (!XBZRLE.current_buf) { | |
2027 | error_report("Error allocating current_buf"); | |
2028 | g_free(XBZRLE.encoded_buf); | |
2029 | XBZRLE.encoded_buf = NULL; | |
2030 | return -1; | |
2031 | } | |
56e93d26 JQ |
2032 | } |
2033 | ||
49877834 PB |
2034 | /* For memory_global_dirty_log_start below. */ |
2035 | qemu_mutex_lock_iothread(); | |
2036 | ||
56e93d26 JQ |
2037 | qemu_mutex_lock_ramlist(); |
2038 | rcu_read_lock(); | |
6f37bb8b | 2039 | ram_state_reset(rs); |
56e93d26 | 2040 | |
eb859c53 | 2041 | rs->ram_bitmap = g_new0(struct RAMBitmap, 1); |
0827b9e9 AA |
2042 | /* Skip setting bitmap if there is no RAM */ |
2043 | if (ram_bytes_total()) { | |
2044 | ram_bitmap_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
eb859c53 JQ |
2045 | rs->ram_bitmap->bmap = bitmap_new(ram_bitmap_pages); |
2046 | bitmap_set(rs->ram_bitmap->bmap, 0, ram_bitmap_pages); | |
0827b9e9 AA |
2047 | |
2048 | if (migrate_postcopy_ram()) { | |
eb859c53 JQ |
2049 | rs->ram_bitmap->unsentmap = bitmap_new(ram_bitmap_pages); |
2050 | bitmap_set(rs->ram_bitmap->unsentmap, 0, ram_bitmap_pages); | |
0827b9e9 | 2051 | } |
f3f491fc DDAG |
2052 | } |
2053 | ||
56e93d26 JQ |
2054 | /* |
2055 | * Count the total number of pages used by ram blocks not including any | |
2056 | * gaps due to alignment or unplugs. | |
2057 | */ | |
0d8ec885 | 2058 | rs->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
56e93d26 JQ |
2059 | |
2060 | memory_global_dirty_log_start(); | |
8d820d6f | 2061 | migration_bitmap_sync(rs); |
56e93d26 | 2062 | qemu_mutex_unlock_ramlist(); |
49877834 | 2063 | qemu_mutex_unlock_iothread(); |
a91246c9 HZ |
2064 | rcu_read_unlock(); |
2065 | ||
2066 | return 0; | |
2067 | } | |
2068 | ||
3d0684b2 JQ |
2069 | /* |
2070 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2071 | * long-running RCU critical section. When rcu-reclaims in the code |
2072 | * start to become numerous it will be necessary to reduce the | |
2073 | * granularity of these critical sections. | |
2074 | */ | |
2075 | ||
3d0684b2 JQ |
2076 | /** |
2077 | * ram_save_setup: Setup RAM for migration | |
2078 | * | |
2079 | * Returns zero to indicate success and negative for error | |
2080 | * | |
2081 | * @f: QEMUFile where to send the data | |
2082 | * @opaque: RAMState pointer | |
2083 | */ | |
a91246c9 HZ |
2084 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2085 | { | |
6f37bb8b | 2086 | RAMState *rs = opaque; |
a91246c9 HZ |
2087 | RAMBlock *block; |
2088 | ||
2089 | /* migration has already setup the bitmap, reuse it. */ | |
2090 | if (!migration_in_colo_state()) { | |
ceb4d168 | 2091 | if (ram_state_init(rs) < 0) { |
a91246c9 HZ |
2092 | return -1; |
2093 | } | |
2094 | } | |
204b88b8 | 2095 | rs->f = f; |
a91246c9 HZ |
2096 | |
2097 | rcu_read_lock(); | |
56e93d26 JQ |
2098 | |
2099 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
2100 | ||
2101 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
2102 | qemu_put_byte(f, strlen(block->idstr)); | |
2103 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2104 | qemu_put_be64(f, block->used_length); | |
ef08fb38 DDAG |
2105 | if (migrate_postcopy_ram() && block->page_size != qemu_host_page_size) { |
2106 | qemu_put_be64(f, block->page_size); | |
2107 | } | |
56e93d26 JQ |
2108 | } |
2109 | ||
2110 | rcu_read_unlock(); | |
2111 | ||
2112 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
2113 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2114 | ||
2115 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2116 | ||
2117 | return 0; | |
2118 | } | |
2119 | ||
3d0684b2 JQ |
2120 | /** |
2121 | * ram_save_iterate: iterative stage for migration | |
2122 | * | |
2123 | * Returns zero to indicate success and negative for error | |
2124 | * | |
2125 | * @f: QEMUFile where to send the data | |
2126 | * @opaque: RAMState pointer | |
2127 | */ | |
56e93d26 JQ |
2128 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2129 | { | |
6f37bb8b | 2130 | RAMState *rs = opaque; |
56e93d26 JQ |
2131 | int ret; |
2132 | int i; | |
2133 | int64_t t0; | |
5c90308f | 2134 | int done = 0; |
56e93d26 JQ |
2135 | |
2136 | rcu_read_lock(); | |
6f37bb8b JQ |
2137 | if (ram_list.version != rs->last_version) { |
2138 | ram_state_reset(rs); | |
56e93d26 JQ |
2139 | } |
2140 | ||
2141 | /* Read version before ram_list.blocks */ | |
2142 | smp_rmb(); | |
2143 | ||
2144 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
2145 | ||
2146 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
2147 | i = 0; | |
2148 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
2149 | int pages; | |
2150 | ||
072c2511 | 2151 | pages = ram_find_and_save_block(rs, f, false); |
56e93d26 JQ |
2152 | /* no more pages to sent */ |
2153 | if (pages == 0) { | |
5c90308f | 2154 | done = 1; |
56e93d26 JQ |
2155 | break; |
2156 | } | |
23b28c3c | 2157 | rs->iterations++; |
070afca2 | 2158 | |
56e93d26 JQ |
2159 | /* we want to check in the 1st loop, just in case it was the 1st time |
2160 | and we had to sync the dirty bitmap. | |
2161 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2162 | iterations | |
2163 | */ | |
2164 | if ((i & 63) == 0) { | |
2165 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2166 | if (t1 > MAX_WAIT) { | |
55c4446b | 2167 | trace_ram_save_iterate_big_wait(t1, i); |
56e93d26 JQ |
2168 | break; |
2169 | } | |
2170 | } | |
2171 | i++; | |
2172 | } | |
2f4fde93 | 2173 | flush_compressed_data(rs, f); |
56e93d26 JQ |
2174 | rcu_read_unlock(); |
2175 | ||
2176 | /* | |
2177 | * Must occur before EOS (or any QEMUFile operation) | |
2178 | * because of RDMA protocol. | |
2179 | */ | |
2180 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2181 | ||
2182 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2f4fde93 | 2183 | rs->bytes_transferred += 8; |
56e93d26 JQ |
2184 | |
2185 | ret = qemu_file_get_error(f); | |
2186 | if (ret < 0) { | |
2187 | return ret; | |
2188 | } | |
2189 | ||
5c90308f | 2190 | return done; |
56e93d26 JQ |
2191 | } |
2192 | ||
3d0684b2 JQ |
2193 | /** |
2194 | * ram_save_complete: function called to send the remaining amount of ram | |
2195 | * | |
2196 | * Returns zero to indicate success | |
2197 | * | |
2198 | * Called with iothread lock | |
2199 | * | |
2200 | * @f: QEMUFile where to send the data | |
2201 | * @opaque: RAMState pointer | |
2202 | */ | |
56e93d26 JQ |
2203 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2204 | { | |
6f37bb8b JQ |
2205 | RAMState *rs = opaque; |
2206 | ||
56e93d26 JQ |
2207 | rcu_read_lock(); |
2208 | ||
663e6c1d | 2209 | if (!migration_in_postcopy(migrate_get_current())) { |
8d820d6f | 2210 | migration_bitmap_sync(rs); |
663e6c1d | 2211 | } |
56e93d26 JQ |
2212 | |
2213 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2214 | ||
2215 | /* try transferring iterative blocks of memory */ | |
2216 | ||
2217 | /* flush all remaining blocks regardless of rate limiting */ | |
2218 | while (true) { | |
2219 | int pages; | |
2220 | ||
072c2511 | 2221 | pages = ram_find_and_save_block(rs, f, !migration_in_colo_state()); |
56e93d26 JQ |
2222 | /* no more blocks to sent */ |
2223 | if (pages == 0) { | |
2224 | break; | |
2225 | } | |
2226 | } | |
2227 | ||
2f4fde93 | 2228 | flush_compressed_data(rs, f); |
56e93d26 | 2229 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 JQ |
2230 | |
2231 | rcu_read_unlock(); | |
d09a6fde | 2232 | |
56e93d26 JQ |
2233 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2234 | ||
2235 | return 0; | |
2236 | } | |
2237 | ||
c31b098f DDAG |
2238 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2239 | uint64_t *non_postcopiable_pending, | |
2240 | uint64_t *postcopiable_pending) | |
56e93d26 | 2241 | { |
8d820d6f | 2242 | RAMState *rs = opaque; |
56e93d26 JQ |
2243 | uint64_t remaining_size; |
2244 | ||
9edabd4d | 2245 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2246 | |
663e6c1d DDAG |
2247 | if (!migration_in_postcopy(migrate_get_current()) && |
2248 | remaining_size < max_size) { | |
56e93d26 JQ |
2249 | qemu_mutex_lock_iothread(); |
2250 | rcu_read_lock(); | |
8d820d6f | 2251 | migration_bitmap_sync(rs); |
56e93d26 JQ |
2252 | rcu_read_unlock(); |
2253 | qemu_mutex_unlock_iothread(); | |
9edabd4d | 2254 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2255 | } |
c31b098f DDAG |
2256 | |
2257 | /* We can do postcopy, and all the data is postcopiable */ | |
2258 | *postcopiable_pending += remaining_size; | |
56e93d26 JQ |
2259 | } |
2260 | ||
2261 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2262 | { | |
2263 | unsigned int xh_len; | |
2264 | int xh_flags; | |
063e760a | 2265 | uint8_t *loaded_data; |
56e93d26 JQ |
2266 | |
2267 | if (!xbzrle_decoded_buf) { | |
2268 | xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE); | |
2269 | } | |
063e760a | 2270 | loaded_data = xbzrle_decoded_buf; |
56e93d26 JQ |
2271 | |
2272 | /* extract RLE header */ | |
2273 | xh_flags = qemu_get_byte(f); | |
2274 | xh_len = qemu_get_be16(f); | |
2275 | ||
2276 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2277 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2278 | return -1; | |
2279 | } | |
2280 | ||
2281 | if (xh_len > TARGET_PAGE_SIZE) { | |
2282 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2283 | return -1; | |
2284 | } | |
2285 | /* load data and decode */ | |
063e760a | 2286 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2287 | |
2288 | /* decode RLE */ | |
063e760a | 2289 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2290 | TARGET_PAGE_SIZE) == -1) { |
2291 | error_report("Failed to load XBZRLE page - decode error!"); | |
2292 | return -1; | |
2293 | } | |
2294 | ||
2295 | return 0; | |
2296 | } | |
2297 | ||
3d0684b2 JQ |
2298 | /** |
2299 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2300 | * | |
2301 | * Must be called from within a rcu critical section. | |
2302 | * | |
56e93d26 | 2303 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2304 | * |
3d0684b2 JQ |
2305 | * @f: QEMUFile where to read the data from |
2306 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2307 | */ |
3d0684b2 | 2308 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2309 | { |
2310 | static RAMBlock *block = NULL; | |
2311 | char id[256]; | |
2312 | uint8_t len; | |
2313 | ||
2314 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2315 | if (!block) { |
56e93d26 JQ |
2316 | error_report("Ack, bad migration stream!"); |
2317 | return NULL; | |
2318 | } | |
4c4bad48 | 2319 | return block; |
56e93d26 JQ |
2320 | } |
2321 | ||
2322 | len = qemu_get_byte(f); | |
2323 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2324 | id[len] = 0; | |
2325 | ||
e3dd7493 | 2326 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2327 | if (!block) { |
2328 | error_report("Can't find block %s", id); | |
2329 | return NULL; | |
56e93d26 JQ |
2330 | } |
2331 | ||
4c4bad48 HZ |
2332 | return block; |
2333 | } | |
2334 | ||
2335 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2336 | ram_addr_t offset) | |
2337 | { | |
2338 | if (!offset_in_ramblock(block, offset)) { | |
2339 | return NULL; | |
2340 | } | |
2341 | ||
2342 | return block->host + offset; | |
56e93d26 JQ |
2343 | } |
2344 | ||
3d0684b2 JQ |
2345 | /** |
2346 | * ram_handle_compressed: handle the zero page case | |
2347 | * | |
56e93d26 JQ |
2348 | * If a page (or a whole RDMA chunk) has been |
2349 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2350 | * |
2351 | * @host: host address for the zero page | |
2352 | * @ch: what the page is filled from. We only support zero | |
2353 | * @size: size of the zero page | |
56e93d26 JQ |
2354 | */ |
2355 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2356 | { | |
2357 | if (ch != 0 || !is_zero_range(host, size)) { | |
2358 | memset(host, ch, size); | |
2359 | } | |
2360 | } | |
2361 | ||
2362 | static void *do_data_decompress(void *opaque) | |
2363 | { | |
2364 | DecompressParam *param = opaque; | |
2365 | unsigned long pagesize; | |
33d151f4 LL |
2366 | uint8_t *des; |
2367 | int len; | |
56e93d26 | 2368 | |
33d151f4 | 2369 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2370 | while (!param->quit) { |
33d151f4 LL |
2371 | if (param->des) { |
2372 | des = param->des; | |
2373 | len = param->len; | |
2374 | param->des = 0; | |
2375 | qemu_mutex_unlock(¶m->mutex); | |
2376 | ||
56e93d26 | 2377 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2378 | /* uncompress() will return failed in some case, especially |
2379 | * when the page is dirted when doing the compression, it's | |
2380 | * not a problem because the dirty page will be retransferred | |
2381 | * and uncompress() won't break the data in other pages. | |
2382 | */ | |
33d151f4 LL |
2383 | uncompress((Bytef *)des, &pagesize, |
2384 | (const Bytef *)param->compbuf, len); | |
73a8912b | 2385 | |
33d151f4 LL |
2386 | qemu_mutex_lock(&decomp_done_lock); |
2387 | param->done = true; | |
2388 | qemu_cond_signal(&decomp_done_cond); | |
2389 | qemu_mutex_unlock(&decomp_done_lock); | |
2390 | ||
2391 | qemu_mutex_lock(¶m->mutex); | |
2392 | } else { | |
2393 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2394 | } | |
56e93d26 | 2395 | } |
33d151f4 | 2396 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2397 | |
2398 | return NULL; | |
2399 | } | |
2400 | ||
5533b2e9 LL |
2401 | static void wait_for_decompress_done(void) |
2402 | { | |
2403 | int idx, thread_count; | |
2404 | ||
2405 | if (!migrate_use_compression()) { | |
2406 | return; | |
2407 | } | |
2408 | ||
2409 | thread_count = migrate_decompress_threads(); | |
2410 | qemu_mutex_lock(&decomp_done_lock); | |
2411 | for (idx = 0; idx < thread_count; idx++) { | |
2412 | while (!decomp_param[idx].done) { | |
2413 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2414 | } | |
2415 | } | |
2416 | qemu_mutex_unlock(&decomp_done_lock); | |
2417 | } | |
2418 | ||
56e93d26 JQ |
2419 | void migrate_decompress_threads_create(void) |
2420 | { | |
2421 | int i, thread_count; | |
2422 | ||
2423 | thread_count = migrate_decompress_threads(); | |
2424 | decompress_threads = g_new0(QemuThread, thread_count); | |
2425 | decomp_param = g_new0(DecompressParam, thread_count); | |
73a8912b LL |
2426 | qemu_mutex_init(&decomp_done_lock); |
2427 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2428 | for (i = 0; i < thread_count; i++) { |
2429 | qemu_mutex_init(&decomp_param[i].mutex); | |
2430 | qemu_cond_init(&decomp_param[i].cond); | |
2431 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2432 | decomp_param[i].done = true; |
90e56fb4 | 2433 | decomp_param[i].quit = false; |
56e93d26 JQ |
2434 | qemu_thread_create(decompress_threads + i, "decompress", |
2435 | do_data_decompress, decomp_param + i, | |
2436 | QEMU_THREAD_JOINABLE); | |
2437 | } | |
2438 | } | |
2439 | ||
2440 | void migrate_decompress_threads_join(void) | |
2441 | { | |
2442 | int i, thread_count; | |
2443 | ||
56e93d26 JQ |
2444 | thread_count = migrate_decompress_threads(); |
2445 | for (i = 0; i < thread_count; i++) { | |
2446 | qemu_mutex_lock(&decomp_param[i].mutex); | |
90e56fb4 | 2447 | decomp_param[i].quit = true; |
56e93d26 JQ |
2448 | qemu_cond_signal(&decomp_param[i].cond); |
2449 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2450 | } | |
2451 | for (i = 0; i < thread_count; i++) { | |
2452 | qemu_thread_join(decompress_threads + i); | |
2453 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2454 | qemu_cond_destroy(&decomp_param[i].cond); | |
2455 | g_free(decomp_param[i].compbuf); | |
2456 | } | |
2457 | g_free(decompress_threads); | |
2458 | g_free(decomp_param); | |
56e93d26 JQ |
2459 | decompress_threads = NULL; |
2460 | decomp_param = NULL; | |
56e93d26 JQ |
2461 | } |
2462 | ||
c1bc6626 | 2463 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2464 | void *host, int len) |
2465 | { | |
2466 | int idx, thread_count; | |
2467 | ||
2468 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2469 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2470 | while (true) { |
2471 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2472 | if (decomp_param[idx].done) { |
33d151f4 LL |
2473 | decomp_param[idx].done = false; |
2474 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2475 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2476 | decomp_param[idx].des = host; |
2477 | decomp_param[idx].len = len; | |
33d151f4 LL |
2478 | qemu_cond_signal(&decomp_param[idx].cond); |
2479 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2480 | break; |
2481 | } | |
2482 | } | |
2483 | if (idx < thread_count) { | |
2484 | break; | |
73a8912b LL |
2485 | } else { |
2486 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2487 | } |
2488 | } | |
73a8912b | 2489 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2490 | } |
2491 | ||
3d0684b2 JQ |
2492 | /** |
2493 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
2494 | * | |
2495 | * Returns 0 for success and negative if there was one error | |
2496 | * | |
2497 | * @mis: current migration incoming state | |
2498 | * | |
2499 | * Allocate data structures etc needed by incoming migration with | |
2500 | * postcopy-ram. postcopy-ram's similarly names | |
2501 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
2502 | */ |
2503 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2504 | { | |
2505 | size_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
2506 | ||
2507 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2508 | } | |
2509 | ||
3d0684b2 JQ |
2510 | /** |
2511 | * ram_load_postcopy: load a page in postcopy case | |
2512 | * | |
2513 | * Returns 0 for success or -errno in case of error | |
2514 | * | |
a7180877 DDAG |
2515 | * Called in postcopy mode by ram_load(). |
2516 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
2517 | * |
2518 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
2519 | */ |
2520 | static int ram_load_postcopy(QEMUFile *f) | |
2521 | { | |
2522 | int flags = 0, ret = 0; | |
2523 | bool place_needed = false; | |
28abd200 | 2524 | bool matching_page_sizes = false; |
a7180877 DDAG |
2525 | MigrationIncomingState *mis = migration_incoming_get_current(); |
2526 | /* Temporary page that is later 'placed' */ | |
2527 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2528 | void *last_host = NULL; |
a3b6ff6d | 2529 | bool all_zero = false; |
a7180877 DDAG |
2530 | |
2531 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2532 | ram_addr_t addr; | |
2533 | void *host = NULL; | |
2534 | void *page_buffer = NULL; | |
2535 | void *place_source = NULL; | |
df9ff5e1 | 2536 | RAMBlock *block = NULL; |
a7180877 | 2537 | uint8_t ch; |
a7180877 DDAG |
2538 | |
2539 | addr = qemu_get_be64(f); | |
2540 | flags = addr & ~TARGET_PAGE_MASK; | |
2541 | addr &= TARGET_PAGE_MASK; | |
2542 | ||
2543 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2544 | place_needed = false; | |
2545 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE)) { | |
df9ff5e1 | 2546 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
2547 | |
2548 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2549 | if (!host) { |
2550 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2551 | ret = -EINVAL; | |
2552 | break; | |
2553 | } | |
28abd200 | 2554 | matching_page_sizes = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 2555 | /* |
28abd200 DDAG |
2556 | * Postcopy requires that we place whole host pages atomically; |
2557 | * these may be huge pages for RAMBlocks that are backed by | |
2558 | * hugetlbfs. | |
a7180877 DDAG |
2559 | * To make it atomic, the data is read into a temporary page |
2560 | * that's moved into place later. | |
2561 | * The migration protocol uses, possibly smaller, target-pages | |
2562 | * however the source ensures it always sends all the components | |
2563 | * of a host page in order. | |
2564 | */ | |
2565 | page_buffer = postcopy_host_page + | |
28abd200 | 2566 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 2567 | /* If all TP are zero then we can optimise the place */ |
28abd200 | 2568 | if (!((uintptr_t)host & (block->page_size - 1))) { |
a7180877 | 2569 | all_zero = true; |
c53b7ddc DDAG |
2570 | } else { |
2571 | /* not the 1st TP within the HP */ | |
2572 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2573 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2574 | host, last_host); |
2575 | ret = -EINVAL; | |
2576 | break; | |
2577 | } | |
a7180877 DDAG |
2578 | } |
2579 | ||
c53b7ddc | 2580 | |
a7180877 DDAG |
2581 | /* |
2582 | * If it's the last part of a host page then we place the host | |
2583 | * page | |
2584 | */ | |
2585 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
28abd200 | 2586 | (block->page_size - 1)) == 0; |
a7180877 DDAG |
2587 | place_source = postcopy_host_page; |
2588 | } | |
c53b7ddc | 2589 | last_host = host; |
a7180877 DDAG |
2590 | |
2591 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
2592 | case RAM_SAVE_FLAG_COMPRESS: | |
2593 | ch = qemu_get_byte(f); | |
2594 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2595 | if (ch) { | |
2596 | all_zero = false; | |
2597 | } | |
2598 | break; | |
2599 | ||
2600 | case RAM_SAVE_FLAG_PAGE: | |
2601 | all_zero = false; | |
2602 | if (!place_needed || !matching_page_sizes) { | |
2603 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2604 | } else { | |
2605 | /* Avoids the qemu_file copy during postcopy, which is | |
2606 | * going to do a copy later; can only do it when we | |
2607 | * do this read in one go (matching page sizes) | |
2608 | */ | |
2609 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2610 | TARGET_PAGE_SIZE); | |
2611 | } | |
2612 | break; | |
2613 | case RAM_SAVE_FLAG_EOS: | |
2614 | /* normal exit */ | |
2615 | break; | |
2616 | default: | |
2617 | error_report("Unknown combination of migration flags: %#x" | |
2618 | " (postcopy mode)", flags); | |
2619 | ret = -EINVAL; | |
2620 | } | |
2621 | ||
2622 | if (place_needed) { | |
2623 | /* This gets called at the last target page in the host page */ | |
df9ff5e1 DDAG |
2624 | void *place_dest = host + TARGET_PAGE_SIZE - block->page_size; |
2625 | ||
a7180877 | 2626 | if (all_zero) { |
df9ff5e1 DDAG |
2627 | ret = postcopy_place_page_zero(mis, place_dest, |
2628 | block->page_size); | |
a7180877 | 2629 | } else { |
df9ff5e1 DDAG |
2630 | ret = postcopy_place_page(mis, place_dest, |
2631 | place_source, block->page_size); | |
a7180877 DDAG |
2632 | } |
2633 | } | |
2634 | if (!ret) { | |
2635 | ret = qemu_file_get_error(f); | |
2636 | } | |
2637 | } | |
2638 | ||
2639 | return ret; | |
2640 | } | |
2641 | ||
56e93d26 JQ |
2642 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2643 | { | |
2644 | int flags = 0, ret = 0; | |
2645 | static uint64_t seq_iter; | |
2646 | int len = 0; | |
a7180877 DDAG |
2647 | /* |
2648 | * If system is running in postcopy mode, page inserts to host memory must | |
2649 | * be atomic | |
2650 | */ | |
2651 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
ef08fb38 DDAG |
2652 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
2653 | bool postcopy_advised = postcopy_state_get() >= POSTCOPY_INCOMING_ADVISE; | |
56e93d26 JQ |
2654 | |
2655 | seq_iter++; | |
2656 | ||
2657 | if (version_id != 4) { | |
2658 | ret = -EINVAL; | |
2659 | } | |
2660 | ||
2661 | /* This RCU critical section can be very long running. | |
2662 | * When RCU reclaims in the code start to become numerous, | |
2663 | * it will be necessary to reduce the granularity of this | |
2664 | * critical section. | |
2665 | */ | |
2666 | rcu_read_lock(); | |
a7180877 DDAG |
2667 | |
2668 | if (postcopy_running) { | |
2669 | ret = ram_load_postcopy(f); | |
2670 | } | |
2671 | ||
2672 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2673 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2674 | void *host = NULL; |
56e93d26 JQ |
2675 | uint8_t ch; |
2676 | ||
2677 | addr = qemu_get_be64(f); | |
2678 | flags = addr & ~TARGET_PAGE_MASK; | |
2679 | addr &= TARGET_PAGE_MASK; | |
2680 | ||
a776aa15 DDAG |
2681 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE | |
2682 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { | |
4c4bad48 HZ |
2683 | RAMBlock *block = ram_block_from_stream(f, flags); |
2684 | ||
2685 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2686 | if (!host) { |
2687 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2688 | ret = -EINVAL; | |
2689 | break; | |
2690 | } | |
2691 | } | |
2692 | ||
56e93d26 JQ |
2693 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2694 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2695 | /* Synchronize RAM block list */ | |
2696 | total_ram_bytes = addr; | |
2697 | while (!ret && total_ram_bytes) { | |
2698 | RAMBlock *block; | |
56e93d26 JQ |
2699 | char id[256]; |
2700 | ram_addr_t length; | |
2701 | ||
2702 | len = qemu_get_byte(f); | |
2703 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2704 | id[len] = 0; | |
2705 | length = qemu_get_be64(f); | |
2706 | ||
e3dd7493 DDAG |
2707 | block = qemu_ram_block_by_name(id); |
2708 | if (block) { | |
2709 | if (length != block->used_length) { | |
2710 | Error *local_err = NULL; | |
56e93d26 | 2711 | |
fa53a0e5 | 2712 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2713 | &local_err); |
2714 | if (local_err) { | |
2715 | error_report_err(local_err); | |
56e93d26 | 2716 | } |
56e93d26 | 2717 | } |
ef08fb38 DDAG |
2718 | /* For postcopy we need to check hugepage sizes match */ |
2719 | if (postcopy_advised && | |
2720 | block->page_size != qemu_host_page_size) { | |
2721 | uint64_t remote_page_size = qemu_get_be64(f); | |
2722 | if (remote_page_size != block->page_size) { | |
2723 | error_report("Mismatched RAM page size %s " | |
2724 | "(local) %zd != %" PRId64, | |
2725 | id, block->page_size, | |
2726 | remote_page_size); | |
2727 | ret = -EINVAL; | |
2728 | } | |
2729 | } | |
e3dd7493 DDAG |
2730 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2731 | block->idstr); | |
2732 | } else { | |
56e93d26 JQ |
2733 | error_report("Unknown ramblock \"%s\", cannot " |
2734 | "accept migration", id); | |
2735 | ret = -EINVAL; | |
2736 | } | |
2737 | ||
2738 | total_ram_bytes -= length; | |
2739 | } | |
2740 | break; | |
a776aa15 | 2741 | |
56e93d26 | 2742 | case RAM_SAVE_FLAG_COMPRESS: |
56e93d26 JQ |
2743 | ch = qemu_get_byte(f); |
2744 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2745 | break; | |
a776aa15 | 2746 | |
56e93d26 | 2747 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2748 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2749 | break; | |
56e93d26 | 2750 | |
a776aa15 | 2751 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2752 | len = qemu_get_be32(f); |
2753 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2754 | error_report("Invalid compressed data length: %d", len); | |
2755 | ret = -EINVAL; | |
2756 | break; | |
2757 | } | |
c1bc6626 | 2758 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2759 | break; |
a776aa15 | 2760 | |
56e93d26 | 2761 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2762 | if (load_xbzrle(f, addr, host) < 0) { |
2763 | error_report("Failed to decompress XBZRLE page at " | |
2764 | RAM_ADDR_FMT, addr); | |
2765 | ret = -EINVAL; | |
2766 | break; | |
2767 | } | |
2768 | break; | |
2769 | case RAM_SAVE_FLAG_EOS: | |
2770 | /* normal exit */ | |
2771 | break; | |
2772 | default: | |
2773 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2774 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2775 | } else { |
2776 | error_report("Unknown combination of migration flags: %#x", | |
2777 | flags); | |
2778 | ret = -EINVAL; | |
2779 | } | |
2780 | } | |
2781 | if (!ret) { | |
2782 | ret = qemu_file_get_error(f); | |
2783 | } | |
2784 | } | |
2785 | ||
5533b2e9 | 2786 | wait_for_decompress_done(); |
56e93d26 | 2787 | rcu_read_unlock(); |
55c4446b | 2788 | trace_ram_load_complete(ret, seq_iter); |
56e93d26 JQ |
2789 | return ret; |
2790 | } | |
2791 | ||
2792 | static SaveVMHandlers savevm_ram_handlers = { | |
2793 | .save_live_setup = ram_save_setup, | |
2794 | .save_live_iterate = ram_save_iterate, | |
763c906b | 2795 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2796 | .save_live_complete_precopy = ram_save_complete, |
56e93d26 JQ |
2797 | .save_live_pending = ram_save_pending, |
2798 | .load_state = ram_load, | |
6ad2a215 | 2799 | .cleanup = ram_migration_cleanup, |
56e93d26 JQ |
2800 | }; |
2801 | ||
2802 | void ram_mig_init(void) | |
2803 | { | |
2804 | qemu_mutex_init(&XBZRLE.lock); | |
6f37bb8b | 2805 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 2806 | } |