]>
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; | |
269ace29 JQ |
174 | /* Last dirty target page we have sent */ |
175 | ram_addr_t last_page; | |
6f37bb8b JQ |
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 | ||
56e93d26 JQ |
339 | static DecompressParam *decomp_param; |
340 | static QemuThread *decompress_threads; | |
73a8912b LL |
341 | static QemuMutex decomp_done_lock; |
342 | static QemuCond decomp_done_cond; | |
56e93d26 | 343 | |
a7a9a88f LL |
344 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
345 | ram_addr_t offset); | |
56e93d26 JQ |
346 | |
347 | static void *do_data_compress(void *opaque) | |
348 | { | |
349 | CompressParam *param = opaque; | |
a7a9a88f LL |
350 | RAMBlock *block; |
351 | ram_addr_t offset; | |
56e93d26 | 352 | |
a7a9a88f | 353 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 354 | while (!param->quit) { |
a7a9a88f LL |
355 | if (param->block) { |
356 | block = param->block; | |
357 | offset = param->offset; | |
358 | param->block = NULL; | |
359 | qemu_mutex_unlock(¶m->mutex); | |
360 | ||
361 | do_compress_ram_page(param->file, block, offset); | |
362 | ||
0d9f9a5c | 363 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 364 | param->done = true; |
0d9f9a5c LL |
365 | qemu_cond_signal(&comp_done_cond); |
366 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
367 | |
368 | qemu_mutex_lock(¶m->mutex); | |
369 | } else { | |
56e93d26 JQ |
370 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
371 | } | |
56e93d26 | 372 | } |
a7a9a88f | 373 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
374 | |
375 | return NULL; | |
376 | } | |
377 | ||
378 | static inline void terminate_compression_threads(void) | |
379 | { | |
380 | int idx, thread_count; | |
381 | ||
382 | thread_count = migrate_compress_threads(); | |
3d0684b2 | 383 | |
56e93d26 JQ |
384 | for (idx = 0; idx < thread_count; idx++) { |
385 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 386 | comp_param[idx].quit = true; |
56e93d26 JQ |
387 | qemu_cond_signal(&comp_param[idx].cond); |
388 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
389 | } | |
390 | } | |
391 | ||
392 | void migrate_compress_threads_join(void) | |
393 | { | |
394 | int i, thread_count; | |
395 | ||
396 | if (!migrate_use_compression()) { | |
397 | return; | |
398 | } | |
399 | terminate_compression_threads(); | |
400 | thread_count = migrate_compress_threads(); | |
401 | for (i = 0; i < thread_count; i++) { | |
402 | qemu_thread_join(compress_threads + i); | |
403 | qemu_fclose(comp_param[i].file); | |
404 | qemu_mutex_destroy(&comp_param[i].mutex); | |
405 | qemu_cond_destroy(&comp_param[i].cond); | |
406 | } | |
0d9f9a5c LL |
407 | qemu_mutex_destroy(&comp_done_lock); |
408 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
409 | g_free(compress_threads); |
410 | g_free(comp_param); | |
56e93d26 JQ |
411 | compress_threads = NULL; |
412 | comp_param = NULL; | |
56e93d26 JQ |
413 | } |
414 | ||
415 | void migrate_compress_threads_create(void) | |
416 | { | |
417 | int i, thread_count; | |
418 | ||
419 | if (!migrate_use_compression()) { | |
420 | return; | |
421 | } | |
56e93d26 JQ |
422 | thread_count = migrate_compress_threads(); |
423 | compress_threads = g_new0(QemuThread, thread_count); | |
424 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
425 | qemu_cond_init(&comp_done_cond); |
426 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 427 | for (i = 0; i < thread_count; i++) { |
e110aa91 C |
428 | /* comp_param[i].file is just used as a dummy buffer to save data, |
429 | * set its ops to empty. | |
56e93d26 JQ |
430 | */ |
431 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
432 | comp_param[i].done = true; | |
90e56fb4 | 433 | comp_param[i].quit = false; |
56e93d26 JQ |
434 | qemu_mutex_init(&comp_param[i].mutex); |
435 | qemu_cond_init(&comp_param[i].cond); | |
436 | qemu_thread_create(compress_threads + i, "compress", | |
437 | do_data_compress, comp_param + i, | |
438 | QEMU_THREAD_JOINABLE); | |
439 | } | |
440 | } | |
441 | ||
442 | /** | |
3d0684b2 | 443 | * save_page_header: write page header to wire |
56e93d26 JQ |
444 | * |
445 | * If this is the 1st block, it also writes the block identification | |
446 | * | |
3d0684b2 | 447 | * Returns the number of bytes written |
56e93d26 JQ |
448 | * |
449 | * @f: QEMUFile where to send the data | |
450 | * @block: block that contains the page we want to send | |
451 | * @offset: offset inside the block for the page | |
452 | * in the lower bits, it contains flags | |
453 | */ | |
24795694 | 454 | static size_t save_page_header(RAMState *rs, RAMBlock *block, ram_addr_t offset) |
56e93d26 | 455 | { |
9f5f380b | 456 | size_t size, len; |
56e93d26 | 457 | |
24795694 JQ |
458 | if (block == rs->last_sent_block) { |
459 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
460 | } | |
461 | qemu_put_be64(rs->f, offset); | |
56e93d26 JQ |
462 | size = 8; |
463 | ||
464 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 465 | len = strlen(block->idstr); |
24795694 JQ |
466 | qemu_put_byte(rs->f, len); |
467 | qemu_put_buffer(rs->f, (uint8_t *)block->idstr, len); | |
9f5f380b | 468 | size += 1 + len; |
24795694 | 469 | rs->last_sent_block = block; |
56e93d26 JQ |
470 | } |
471 | return size; | |
472 | } | |
473 | ||
3d0684b2 JQ |
474 | /** |
475 | * mig_throttle_guest_down: throotle down the guest | |
476 | * | |
477 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
478 | * writes. If guest dirty memory rate is reduced below the rate at | |
479 | * which we can transfer pages to the destination then we should be | |
480 | * able to complete migration. Some workloads dirty memory way too | |
481 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
482 | */ |
483 | static void mig_throttle_guest_down(void) | |
484 | { | |
485 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
486 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
487 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
488 | |
489 | /* We have not started throttling yet. Let's start it. */ | |
490 | if (!cpu_throttle_active()) { | |
491 | cpu_throttle_set(pct_initial); | |
492 | } else { | |
493 | /* Throttling already on, just increase the rate */ | |
494 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
495 | } | |
496 | } | |
497 | ||
3d0684b2 JQ |
498 | /** |
499 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
500 | * | |
6f37bb8b | 501 | * @rs: current RAM state |
3d0684b2 JQ |
502 | * @current_addr: address for the zero page |
503 | * | |
504 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
505 | * The important thing is that a stale (not-yet-0'd) page be replaced |
506 | * by the new data. | |
507 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 508 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 509 | */ |
6f37bb8b | 510 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 511 | { |
6f37bb8b | 512 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
513 | return; |
514 | } | |
515 | ||
516 | /* We don't care if this fails to allocate a new cache page | |
517 | * as long as it updated an old one */ | |
518 | cache_insert(XBZRLE.cache, current_addr, ZERO_TARGET_PAGE, | |
5a987738 | 519 | rs->bitmap_sync_count); |
56e93d26 JQ |
520 | } |
521 | ||
522 | #define ENCODING_FLAG_XBZRLE 0x1 | |
523 | ||
524 | /** | |
525 | * save_xbzrle_page: compress and send current page | |
526 | * | |
527 | * Returns: 1 means that we wrote the page | |
528 | * 0 means that page is identical to the one already sent | |
529 | * -1 means that xbzrle would be longer than normal | |
530 | * | |
5a987738 | 531 | * @rs: current RAM state |
3d0684b2 JQ |
532 | * @current_data: pointer to the address of the page contents |
533 | * @current_addr: addr of the page | |
56e93d26 JQ |
534 | * @block: block that contains the page we want to send |
535 | * @offset: offset inside the block for the page | |
536 | * @last_stage: if we are at the completion stage | |
56e93d26 | 537 | */ |
204b88b8 | 538 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 539 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 540 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
541 | { |
542 | int encoded_len = 0, bytes_xbzrle; | |
543 | uint8_t *prev_cached_page; | |
544 | ||
5a987738 | 545 | if (!cache_is_cached(XBZRLE.cache, current_addr, rs->bitmap_sync_count)) { |
544c36f1 | 546 | rs->xbzrle_cache_miss++; |
56e93d26 JQ |
547 | if (!last_stage) { |
548 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
5a987738 | 549 | rs->bitmap_sync_count) == -1) { |
56e93d26 JQ |
550 | return -1; |
551 | } else { | |
552 | /* update *current_data when the page has been | |
553 | inserted into cache */ | |
554 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
555 | } | |
556 | } | |
557 | return -1; | |
558 | } | |
559 | ||
560 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
561 | ||
562 | /* save current buffer into memory */ | |
563 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
564 | ||
565 | /* XBZRLE encoding (if there is no overflow) */ | |
566 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
567 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
568 | TARGET_PAGE_SIZE); | |
569 | if (encoded_len == 0) { | |
55c4446b | 570 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
571 | return 0; |
572 | } else if (encoded_len == -1) { | |
55c4446b | 573 | trace_save_xbzrle_page_overflow(); |
180f61f7 | 574 | rs->xbzrle_overflows++; |
56e93d26 JQ |
575 | /* update data in the cache */ |
576 | if (!last_stage) { | |
577 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
578 | *current_data = prev_cached_page; | |
579 | } | |
580 | return -1; | |
581 | } | |
582 | ||
583 | /* we need to update the data in the cache, in order to get the same data */ | |
584 | if (!last_stage) { | |
585 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
586 | } | |
587 | ||
588 | /* Send XBZRLE based compressed page */ | |
24795694 | 589 | bytes_xbzrle = save_page_header(rs, block, |
204b88b8 JQ |
590 | offset | RAM_SAVE_FLAG_XBZRLE); |
591 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
592 | qemu_put_be16(rs->f, encoded_len); | |
593 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 594 | bytes_xbzrle += encoded_len + 1 + 2; |
f36ada95 | 595 | rs->xbzrle_pages++; |
07ed50a2 | 596 | rs->xbzrle_bytes += bytes_xbzrle; |
072c2511 | 597 | rs->bytes_transferred += bytes_xbzrle; |
56e93d26 JQ |
598 | |
599 | return 1; | |
600 | } | |
601 | ||
3d0684b2 JQ |
602 | /** |
603 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 604 | * |
3d0684b2 JQ |
605 | * Called with rcu_read_lock() to protect migration_bitmap |
606 | * | |
607 | * Returns the byte offset within memory region of the start of a dirty page | |
608 | * | |
6f37bb8b | 609 | * @rs: current RAM state |
3d0684b2 JQ |
610 | * @rb: RAMBlock where to search for dirty pages |
611 | * @start: starting address (typically so we can continue from previous page) | |
06b10688 | 612 | * @page_abs: pointer into where to store the dirty page |
f3f491fc | 613 | */ |
56e93d26 | 614 | static inline |
6f37bb8b | 615 | ram_addr_t migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
a82d593b | 616 | ram_addr_t start, |
06b10688 | 617 | unsigned long *page_abs) |
56e93d26 | 618 | { |
2f68e399 | 619 | unsigned long base = rb->offset >> TARGET_PAGE_BITS; |
56e93d26 | 620 | unsigned long nr = base + (start >> TARGET_PAGE_BITS); |
2f68e399 DDAG |
621 | uint64_t rb_size = rb->used_length; |
622 | unsigned long size = base + (rb_size >> TARGET_PAGE_BITS); | |
2ff64038 | 623 | unsigned long *bitmap; |
56e93d26 JQ |
624 | |
625 | unsigned long next; | |
626 | ||
eb859c53 | 627 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
6f37bb8b | 628 | if (rs->ram_bulk_stage && nr > base) { |
56e93d26 JQ |
629 | next = nr + 1; |
630 | } else { | |
2ff64038 | 631 | next = find_next_bit(bitmap, size, nr); |
56e93d26 JQ |
632 | } |
633 | ||
06b10688 | 634 | *page_abs = next; |
56e93d26 JQ |
635 | return (next - base) << TARGET_PAGE_BITS; |
636 | } | |
637 | ||
06b10688 JQ |
638 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
639 | unsigned long page_abs) | |
a82d593b DDAG |
640 | { |
641 | bool ret; | |
eb859c53 | 642 | unsigned long *bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
a82d593b | 643 | |
06b10688 | 644 | ret = test_and_clear_bit(page_abs, bitmap); |
a82d593b DDAG |
645 | |
646 | if (ret) { | |
0d8ec885 | 647 | rs->migration_dirty_pages--; |
a82d593b DDAG |
648 | } |
649 | return ret; | |
650 | } | |
651 | ||
15440dd5 JQ |
652 | static void migration_bitmap_sync_range(RAMState *rs, RAMBlock *rb, |
653 | ram_addr_t start, ram_addr_t length) | |
56e93d26 | 654 | { |
2ff64038 | 655 | unsigned long *bitmap; |
eb859c53 | 656 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
0d8ec885 | 657 | rs->migration_dirty_pages += |
15440dd5 | 658 | cpu_physical_memory_sync_dirty_bitmap(bitmap, rb, start, length, |
0d8ec885 | 659 | &rs->num_dirty_pages_period); |
56e93d26 JQ |
660 | } |
661 | ||
3d0684b2 JQ |
662 | /** |
663 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
664 | * | |
665 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
666 | * | |
667 | * For VMs with just normal pages this is equivalent to the host page | |
668 | * size. If it's got some huge pages then it's the OR of all the | |
669 | * different page sizes. | |
e8ca1db2 DDAG |
670 | */ |
671 | uint64_t ram_pagesize_summary(void) | |
672 | { | |
673 | RAMBlock *block; | |
674 | uint64_t summary = 0; | |
675 | ||
676 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
677 | summary |= block->page_size; | |
678 | } | |
679 | ||
680 | return summary; | |
681 | } | |
682 | ||
8d820d6f | 683 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
684 | { |
685 | RAMBlock *block; | |
56e93d26 | 686 | int64_t end_time; |
c4bdf0cf | 687 | uint64_t bytes_xfer_now; |
56e93d26 | 688 | |
5a987738 | 689 | rs->bitmap_sync_count++; |
56e93d26 | 690 | |
eac74159 JQ |
691 | if (!rs->bytes_xfer_prev) { |
692 | rs->bytes_xfer_prev = ram_bytes_transferred(); | |
56e93d26 JQ |
693 | } |
694 | ||
f664da80 JQ |
695 | if (!rs->time_last_bitmap_sync) { |
696 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
697 | } |
698 | ||
699 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 700 | memory_global_dirty_log_sync(); |
56e93d26 | 701 | |
108cfae0 | 702 | qemu_mutex_lock(&rs->bitmap_mutex); |
56e93d26 JQ |
703 | rcu_read_lock(); |
704 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
15440dd5 | 705 | migration_bitmap_sync_range(rs, block, 0, block->used_length); |
56e93d26 JQ |
706 | } |
707 | rcu_read_unlock(); | |
108cfae0 | 708 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 709 | |
a66cd90c | 710 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 711 | |
56e93d26 JQ |
712 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
713 | ||
714 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 715 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
56e93d26 JQ |
716 | if (migrate_auto_converge()) { |
717 | /* The following detection logic can be refined later. For now: | |
718 | Check to see if the dirtied bytes is 50% more than the approx. | |
719 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
720 | were in this routine. If that happens twice, start or increase |
721 | throttling */ | |
56e93d26 | 722 | bytes_xfer_now = ram_bytes_transferred(); |
070afca2 | 723 | |
47ad8619 | 724 | if (rs->dirty_pages_rate && |
a66cd90c | 725 | (rs->num_dirty_pages_period * TARGET_PAGE_SIZE > |
eac74159 | 726 | (bytes_xfer_now - rs->bytes_xfer_prev) / 2) && |
8d820d6f | 727 | (rs->dirty_rate_high_cnt++ >= 2)) { |
56e93d26 | 728 | trace_migration_throttle(); |
8d820d6f | 729 | rs->dirty_rate_high_cnt = 0; |
070afca2 | 730 | mig_throttle_guest_down(); |
56e93d26 | 731 | } |
eac74159 | 732 | rs->bytes_xfer_prev = bytes_xfer_now; |
56e93d26 | 733 | } |
070afca2 | 734 | |
56e93d26 | 735 | if (migrate_use_xbzrle()) { |
23b28c3c | 736 | if (rs->iterations_prev != rs->iterations) { |
b07016b6 | 737 | rs->xbzrle_cache_miss_rate = |
544c36f1 | 738 | (double)(rs->xbzrle_cache_miss - |
b5833fde | 739 | rs->xbzrle_cache_miss_prev) / |
23b28c3c | 740 | (rs->iterations - rs->iterations_prev); |
56e93d26 | 741 | } |
23b28c3c | 742 | rs->iterations_prev = rs->iterations; |
544c36f1 | 743 | rs->xbzrle_cache_miss_prev = rs->xbzrle_cache_miss; |
56e93d26 | 744 | } |
47ad8619 | 745 | rs->dirty_pages_rate = rs->num_dirty_pages_period * 1000 |
f664da80 | 746 | / (end_time - rs->time_last_bitmap_sync); |
f664da80 | 747 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 748 | rs->num_dirty_pages_period = 0; |
56e93d26 | 749 | } |
4addcd4f | 750 | if (migrate_use_events()) { |
5a987738 | 751 | qapi_event_send_migration_pass(rs->bitmap_sync_count, NULL); |
4addcd4f | 752 | } |
56e93d26 JQ |
753 | } |
754 | ||
755 | /** | |
3d0684b2 | 756 | * save_zero_page: send the zero page to the stream |
56e93d26 | 757 | * |
3d0684b2 | 758 | * Returns the number of pages written. |
56e93d26 | 759 | * |
f7ccd61b | 760 | * @rs: current RAM state |
56e93d26 JQ |
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 | */ |
ce25d337 JQ |
765 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, |
766 | 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 | 772 | rs->bytes_transferred += |
24795694 | 773 | save_page_header(rs, block, offset | RAM_SAVE_FLAG_COMPRESS); |
ce25d337 | 774 | qemu_put_byte(rs->f, 0); |
072c2511 | 775 | rs->bytes_transferred += 1; |
56e93d26 JQ |
776 | pages = 1; |
777 | } | |
778 | ||
779 | return pages; | |
780 | } | |
781 | ||
5727309d | 782 | static void ram_release_pages(const char *rbname, uint64_t offset, int pages) |
53f09a10 | 783 | { |
5727309d | 784 | if (!migrate_release_ram() || !migration_in_postcopy()) { |
53f09a10 PB |
785 | return; |
786 | } | |
787 | ||
aaa2064c | 788 | ram_discard_range(rbname, offset, pages << TARGET_PAGE_BITS); |
53f09a10 PB |
789 | } |
790 | ||
56e93d26 | 791 | /** |
3d0684b2 | 792 | * ram_save_page: send the given page to the stream |
56e93d26 | 793 | * |
3d0684b2 | 794 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
795 | * < 0 - error |
796 | * >=0 - Number of pages written - this might legally be 0 | |
797 | * if xbzrle noticed the page was the same. | |
56e93d26 | 798 | * |
6f37bb8b | 799 | * @rs: current RAM state |
56e93d26 JQ |
800 | * @block: block that contains the page we want to send |
801 | * @offset: offset inside the block for the page | |
802 | * @last_stage: if we are at the completion stage | |
56e93d26 | 803 | */ |
a0a8aa14 | 804 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
805 | { |
806 | int pages = -1; | |
807 | uint64_t bytes_xmit; | |
808 | ram_addr_t current_addr; | |
56e93d26 JQ |
809 | uint8_t *p; |
810 | int ret; | |
811 | bool send_async = true; | |
a08f6890 HZ |
812 | RAMBlock *block = pss->block; |
813 | ram_addr_t offset = pss->offset; | |
56e93d26 | 814 | |
2f68e399 | 815 | p = block->host + offset; |
56e93d26 JQ |
816 | |
817 | /* In doubt sent page as normal */ | |
818 | bytes_xmit = 0; | |
ce25d337 | 819 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
820 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
821 | if (bytes_xmit) { | |
072c2511 | 822 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
823 | pages = 1; |
824 | } | |
825 | ||
826 | XBZRLE_cache_lock(); | |
827 | ||
828 | current_addr = block->offset + offset; | |
829 | ||
56e93d26 JQ |
830 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
831 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
832 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 833 | rs->norm_pages++; |
56e93d26 | 834 | } else if (bytes_xmit == 0) { |
f7ccd61b | 835 | rs->zero_pages++; |
56e93d26 JQ |
836 | } |
837 | } | |
838 | } else { | |
ce25d337 | 839 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 JQ |
840 | if (pages > 0) { |
841 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
842 | * page would be stale | |
843 | */ | |
6f37bb8b | 844 | xbzrle_cache_zero_page(rs, current_addr); |
5727309d | 845 | ram_release_pages(block->idstr, pss->offset, pages); |
6f37bb8b | 846 | } else if (!rs->ram_bulk_stage && |
5727309d | 847 | !migration_in_postcopy() && migrate_use_xbzrle()) { |
204b88b8 | 848 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
072c2511 | 849 | offset, last_stage); |
56e93d26 JQ |
850 | if (!last_stage) { |
851 | /* Can't send this cached data async, since the cache page | |
852 | * might get updated before it gets to the wire | |
853 | */ | |
854 | send_async = false; | |
855 | } | |
856 | } | |
857 | } | |
858 | ||
859 | /* XBZRLE overflow or normal page */ | |
860 | if (pages == -1) { | |
24795694 JQ |
861 | rs->bytes_transferred += save_page_header(rs, block, |
862 | offset | RAM_SAVE_FLAG_PAGE); | |
56e93d26 | 863 | if (send_async) { |
ce25d337 | 864 | qemu_put_buffer_async(rs->f, p, TARGET_PAGE_SIZE, |
53f09a10 | 865 | migrate_release_ram() & |
5727309d | 866 | migration_in_postcopy()); |
56e93d26 | 867 | } else { |
ce25d337 | 868 | qemu_put_buffer(rs->f, p, TARGET_PAGE_SIZE); |
56e93d26 | 869 | } |
072c2511 | 870 | rs->bytes_transferred += TARGET_PAGE_SIZE; |
56e93d26 | 871 | pages = 1; |
b4d1c6e7 | 872 | rs->norm_pages++; |
56e93d26 JQ |
873 | } |
874 | ||
875 | XBZRLE_cache_unlock(); | |
876 | ||
877 | return pages; | |
878 | } | |
879 | ||
a7a9a88f LL |
880 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
881 | ram_addr_t offset) | |
56e93d26 | 882 | { |
24795694 | 883 | RAMState *rs = &ram_state; |
56e93d26 | 884 | int bytes_sent, blen; |
a7a9a88f | 885 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 | 886 | |
24795694 | 887 | bytes_sent = save_page_header(rs, block, offset | |
56e93d26 | 888 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
a7a9a88f | 889 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, |
56e93d26 | 890 | migrate_compress_level()); |
b3be2896 LL |
891 | if (blen < 0) { |
892 | bytes_sent = 0; | |
893 | qemu_file_set_error(migrate_get_current()->to_dst_file, blen); | |
894 | error_report("compressed data failed!"); | |
895 | } else { | |
896 | bytes_sent += blen; | |
5727309d | 897 | ram_release_pages(block->idstr, offset & TARGET_PAGE_MASK, 1); |
b3be2896 | 898 | } |
56e93d26 JQ |
899 | |
900 | return bytes_sent; | |
901 | } | |
902 | ||
ce25d337 | 903 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
904 | { |
905 | int idx, len, thread_count; | |
906 | ||
907 | if (!migrate_use_compression()) { | |
908 | return; | |
909 | } | |
910 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 911 | |
0d9f9a5c | 912 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 913 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 914 | while (!comp_param[idx].done) { |
0d9f9a5c | 915 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 916 | } |
a7a9a88f | 917 | } |
0d9f9a5c | 918 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
919 | |
920 | for (idx = 0; idx < thread_count; idx++) { | |
921 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 922 | if (!comp_param[idx].quit) { |
ce25d337 | 923 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
2f4fde93 | 924 | rs->bytes_transferred += len; |
56e93d26 | 925 | } |
a7a9a88f | 926 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
927 | } |
928 | } | |
929 | ||
930 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
931 | ram_addr_t offset) | |
932 | { | |
933 | param->block = block; | |
934 | param->offset = offset; | |
935 | } | |
936 | ||
ce25d337 JQ |
937 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
938 | ram_addr_t offset) | |
56e93d26 JQ |
939 | { |
940 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
941 | ||
942 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 943 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 JQ |
944 | while (true) { |
945 | for (idx = 0; idx < thread_count; idx++) { | |
946 | if (comp_param[idx].done) { | |
a7a9a88f | 947 | comp_param[idx].done = false; |
ce25d337 | 948 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
a7a9a88f | 949 | qemu_mutex_lock(&comp_param[idx].mutex); |
56e93d26 | 950 | set_compress_params(&comp_param[idx], block, offset); |
a7a9a88f LL |
951 | qemu_cond_signal(&comp_param[idx].cond); |
952 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
56e93d26 | 953 | pages = 1; |
b4d1c6e7 | 954 | rs->norm_pages++; |
072c2511 | 955 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
956 | break; |
957 | } | |
958 | } | |
959 | if (pages > 0) { | |
960 | break; | |
961 | } else { | |
0d9f9a5c | 962 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 JQ |
963 | } |
964 | } | |
0d9f9a5c | 965 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
966 | |
967 | return pages; | |
968 | } | |
969 | ||
970 | /** | |
971 | * ram_save_compressed_page: compress the given page and send it to the stream | |
972 | * | |
3d0684b2 | 973 | * Returns the number of pages written. |
56e93d26 | 974 | * |
6f37bb8b | 975 | * @rs: current RAM state |
56e93d26 JQ |
976 | * @block: block that contains the page we want to send |
977 | * @offset: offset inside the block for the page | |
978 | * @last_stage: if we are at the completion stage | |
56e93d26 | 979 | */ |
a0a8aa14 JQ |
980 | static int ram_save_compressed_page(RAMState *rs, PageSearchStatus *pss, |
981 | bool last_stage) | |
56e93d26 JQ |
982 | { |
983 | int pages = -1; | |
fc50438e | 984 | uint64_t bytes_xmit = 0; |
56e93d26 | 985 | uint8_t *p; |
fc50438e | 986 | int ret, blen; |
a08f6890 HZ |
987 | RAMBlock *block = pss->block; |
988 | ram_addr_t offset = pss->offset; | |
56e93d26 | 989 | |
2f68e399 | 990 | p = block->host + offset; |
56e93d26 | 991 | |
ce25d337 | 992 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
993 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
994 | if (bytes_xmit) { | |
072c2511 | 995 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
996 | pages = 1; |
997 | } | |
56e93d26 JQ |
998 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
999 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
1000 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 1001 | rs->norm_pages++; |
56e93d26 | 1002 | } else if (bytes_xmit == 0) { |
f7ccd61b | 1003 | rs->zero_pages++; |
56e93d26 JQ |
1004 | } |
1005 | } | |
1006 | } else { | |
1007 | /* When starting the process of a new block, the first page of | |
1008 | * the block should be sent out before other pages in the same | |
1009 | * block, and all the pages in last block should have been sent | |
1010 | * out, keeping this order is important, because the 'cont' flag | |
1011 | * is used to avoid resending the block name. | |
1012 | */ | |
6f37bb8b | 1013 | if (block != rs->last_sent_block) { |
ce25d337 JQ |
1014 | flush_compressed_data(rs); |
1015 | pages = save_zero_page(rs, block, offset, p); | |
56e93d26 | 1016 | if (pages == -1) { |
fc50438e | 1017 | /* Make sure the first page is sent out before other pages */ |
24795694 | 1018 | bytes_xmit = save_page_header(rs, block, offset | |
fc50438e | 1019 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
ce25d337 | 1020 | blen = qemu_put_compression_data(rs->f, p, TARGET_PAGE_SIZE, |
fc50438e LL |
1021 | migrate_compress_level()); |
1022 | if (blen > 0) { | |
072c2511 | 1023 | rs->bytes_transferred += bytes_xmit + blen; |
b4d1c6e7 | 1024 | rs->norm_pages++; |
b3be2896 | 1025 | pages = 1; |
fc50438e | 1026 | } else { |
ce25d337 | 1027 | qemu_file_set_error(rs->f, blen); |
fc50438e | 1028 | error_report("compressed data failed!"); |
b3be2896 | 1029 | } |
56e93d26 | 1030 | } |
53f09a10 | 1031 | if (pages > 0) { |
5727309d | 1032 | ram_release_pages(block->idstr, pss->offset, pages); |
53f09a10 | 1033 | } |
56e93d26 | 1034 | } else { |
ce25d337 | 1035 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 | 1036 | if (pages == -1) { |
ce25d337 | 1037 | pages = compress_page_with_multi_thread(rs, block, offset); |
53f09a10 | 1038 | } else { |
5727309d | 1039 | ram_release_pages(block->idstr, pss->offset, pages); |
56e93d26 JQ |
1040 | } |
1041 | } | |
1042 | } | |
1043 | ||
1044 | return pages; | |
1045 | } | |
1046 | ||
3d0684b2 JQ |
1047 | /** |
1048 | * find_dirty_block: find the next dirty page and update any state | |
1049 | * associated with the search process. | |
b9e60928 | 1050 | * |
3d0684b2 | 1051 | * Returns if a page is found |
b9e60928 | 1052 | * |
6f37bb8b | 1053 | * @rs: current RAM state |
3d0684b2 JQ |
1054 | * @pss: data about the state of the current dirty page scan |
1055 | * @again: set to false if the search has scanned the whole of RAM | |
06b10688 | 1056 | * @page_abs: pointer into where to store the dirty page |
b9e60928 | 1057 | */ |
ce25d337 | 1058 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, |
06b10688 | 1059 | bool *again, unsigned long *page_abs) |
b9e60928 | 1060 | { |
6f37bb8b | 1061 | pss->offset = migration_bitmap_find_dirty(rs, pss->block, pss->offset, |
06b10688 | 1062 | page_abs); |
6f37bb8b | 1063 | if (pss->complete_round && pss->block == rs->last_seen_block && |
269ace29 | 1064 | (pss->offset >> TARGET_PAGE_BITS) >= rs->last_page) { |
b9e60928 DDAG |
1065 | /* |
1066 | * We've been once around the RAM and haven't found anything. | |
1067 | * Give up. | |
1068 | */ | |
1069 | *again = false; | |
1070 | return false; | |
1071 | } | |
1072 | if (pss->offset >= pss->block->used_length) { | |
1073 | /* Didn't find anything in this RAM Block */ | |
1074 | pss->offset = 0; | |
1075 | pss->block = QLIST_NEXT_RCU(pss->block, next); | |
1076 | if (!pss->block) { | |
1077 | /* Hit the end of the list */ | |
1078 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1079 | /* Flag that we've looped */ | |
1080 | pss->complete_round = true; | |
6f37bb8b | 1081 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1082 | if (migrate_use_xbzrle()) { |
1083 | /* If xbzrle is on, stop using the data compression at this | |
1084 | * point. In theory, xbzrle can do better than compression. | |
1085 | */ | |
ce25d337 | 1086 | flush_compressed_data(rs); |
b9e60928 DDAG |
1087 | } |
1088 | } | |
1089 | /* Didn't find anything this time, but try again on the new block */ | |
1090 | *again = true; | |
1091 | return false; | |
1092 | } else { | |
1093 | /* Can go around again, but... */ | |
1094 | *again = true; | |
1095 | /* We've found something so probably don't need to */ | |
1096 | return true; | |
1097 | } | |
1098 | } | |
1099 | ||
3d0684b2 JQ |
1100 | /** |
1101 | * unqueue_page: gets a page of the queue | |
1102 | * | |
a82d593b | 1103 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1104 | * |
3d0684b2 JQ |
1105 | * Returns the block of the page (or NULL if none available) |
1106 | * | |
ec481c6c | 1107 | * @rs: current RAM state |
3d0684b2 | 1108 | * @offset: used to return the offset within the RAMBlock |
06b10688 | 1109 | * @page_abs: pointer into where to store the dirty page |
a82d593b | 1110 | */ |
ec481c6c | 1111 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset, |
06b10688 | 1112 | unsigned long *page_abs) |
a82d593b DDAG |
1113 | { |
1114 | RAMBlock *block = NULL; | |
1115 | ||
ec481c6c JQ |
1116 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1117 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
1118 | struct RAMSrcPageRequest *entry = | |
1119 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1120 | block = entry->rb; |
1121 | *offset = entry->offset; | |
06b10688 | 1122 | *page_abs = (entry->offset + entry->rb->offset) >> TARGET_PAGE_BITS; |
a82d593b DDAG |
1123 | |
1124 | if (entry->len > TARGET_PAGE_SIZE) { | |
1125 | entry->len -= TARGET_PAGE_SIZE; | |
1126 | entry->offset += TARGET_PAGE_SIZE; | |
1127 | } else { | |
1128 | memory_region_unref(block->mr); | |
ec481c6c | 1129 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b DDAG |
1130 | g_free(entry); |
1131 | } | |
1132 | } | |
ec481c6c | 1133 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
a82d593b DDAG |
1134 | |
1135 | return block; | |
1136 | } | |
1137 | ||
3d0684b2 JQ |
1138 | /** |
1139 | * get_queued_page: unqueue a page from the postocpy requests | |
1140 | * | |
1141 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1142 | * |
3d0684b2 | 1143 | * Returns if a queued page is found |
a82d593b | 1144 | * |
6f37bb8b | 1145 | * @rs: current RAM state |
3d0684b2 | 1146 | * @pss: data about the state of the current dirty page scan |
06b10688 | 1147 | * @page_abs: pointer into where to store the dirty page |
a82d593b | 1148 | */ |
ec481c6c | 1149 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss, |
06b10688 | 1150 | unsigned long *page_abs) |
a82d593b DDAG |
1151 | { |
1152 | RAMBlock *block; | |
1153 | ram_addr_t offset; | |
1154 | bool dirty; | |
1155 | ||
1156 | do { | |
06b10688 | 1157 | block = unqueue_page(rs, &offset, page_abs); |
a82d593b DDAG |
1158 | /* |
1159 | * We're sending this page, and since it's postcopy nothing else | |
1160 | * will dirty it, and we must make sure it doesn't get sent again | |
1161 | * even if this queue request was received after the background | |
1162 | * search already sent it. | |
1163 | */ | |
1164 | if (block) { | |
1165 | unsigned long *bitmap; | |
eb859c53 | 1166 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
06b10688 | 1167 | dirty = test_bit(*page_abs, bitmap); |
a82d593b | 1168 | if (!dirty) { |
06b10688 JQ |
1169 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, |
1170 | *page_abs, | |
1171 | test_bit(*page_abs, | |
1172 | atomic_rcu_read(&rs->ram_bitmap)->unsentmap)); | |
a82d593b | 1173 | } else { |
06b10688 JQ |
1174 | trace_get_queued_page(block->idstr, (uint64_t)offset, |
1175 | *page_abs); | |
a82d593b DDAG |
1176 | } |
1177 | } | |
1178 | ||
1179 | } while (block && !dirty); | |
1180 | ||
1181 | if (block) { | |
1182 | /* | |
1183 | * As soon as we start servicing pages out of order, then we have | |
1184 | * to kill the bulk stage, since the bulk stage assumes | |
1185 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1186 | * dirty, that's no longer true. | |
1187 | */ | |
6f37bb8b | 1188 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1189 | |
1190 | /* | |
1191 | * We want the background search to continue from the queued page | |
1192 | * since the guest is likely to want other pages near to the page | |
1193 | * it just requested. | |
1194 | */ | |
1195 | pss->block = block; | |
1196 | pss->offset = offset; | |
1197 | } | |
1198 | ||
1199 | return !!block; | |
1200 | } | |
1201 | ||
6c595cde | 1202 | /** |
5e58f968 JQ |
1203 | * migration_page_queue_free: drop any remaining pages in the ram |
1204 | * request queue | |
6c595cde | 1205 | * |
3d0684b2 JQ |
1206 | * It should be empty at the end anyway, but in error cases there may |
1207 | * be some left. in case that there is any page left, we drop it. | |
1208 | * | |
6c595cde | 1209 | */ |
ec481c6c | 1210 | void migration_page_queue_free(void) |
6c595cde | 1211 | { |
ec481c6c JQ |
1212 | struct RAMSrcPageRequest *mspr, *next_mspr; |
1213 | RAMState *rs = &ram_state; | |
6c595cde DDAG |
1214 | /* This queue generally should be empty - but in the case of a failed |
1215 | * migration might have some droppings in. | |
1216 | */ | |
1217 | rcu_read_lock(); | |
ec481c6c | 1218 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1219 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1220 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1221 | g_free(mspr); |
1222 | } | |
1223 | rcu_read_unlock(); | |
1224 | } | |
1225 | ||
1226 | /** | |
3d0684b2 JQ |
1227 | * ram_save_queue_pages: queue the page for transmission |
1228 | * | |
1229 | * A request from postcopy destination for example. | |
1230 | * | |
1231 | * Returns zero on success or negative on error | |
1232 | * | |
3d0684b2 JQ |
1233 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1234 | * same that last one. | |
1235 | * @start: starting address from the start of the RAMBlock | |
1236 | * @len: length (in bytes) to send | |
6c595cde | 1237 | */ |
96506894 | 1238 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1239 | { |
1240 | RAMBlock *ramblock; | |
68a098f3 | 1241 | RAMState *rs = &ram_state; |
6c595cde | 1242 | |
96506894 | 1243 | rs->postcopy_requests++; |
6c595cde DDAG |
1244 | rcu_read_lock(); |
1245 | if (!rbname) { | |
1246 | /* Reuse last RAMBlock */ | |
68a098f3 | 1247 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1248 | |
1249 | if (!ramblock) { | |
1250 | /* | |
1251 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1252 | * it's the 1st request. | |
1253 | */ | |
1254 | error_report("ram_save_queue_pages no previous block"); | |
1255 | goto err; | |
1256 | } | |
1257 | } else { | |
1258 | ramblock = qemu_ram_block_by_name(rbname); | |
1259 | ||
1260 | if (!ramblock) { | |
1261 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1262 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1263 | goto err; | |
1264 | } | |
68a098f3 | 1265 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1266 | } |
1267 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1268 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1269 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1270 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1271 | __func__, start, len, ramblock->used_length); |
1272 | goto err; | |
1273 | } | |
1274 | ||
ec481c6c JQ |
1275 | struct RAMSrcPageRequest *new_entry = |
1276 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
1277 | new_entry->rb = ramblock; |
1278 | new_entry->offset = start; | |
1279 | new_entry->len = len; | |
1280 | ||
1281 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
1282 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1283 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
1284 | qemu_mutex_unlock(&rs->src_page_req_mutex); | |
6c595cde DDAG |
1285 | rcu_read_unlock(); |
1286 | ||
1287 | return 0; | |
1288 | ||
1289 | err: | |
1290 | rcu_read_unlock(); | |
1291 | return -1; | |
1292 | } | |
1293 | ||
a82d593b | 1294 | /** |
3d0684b2 | 1295 | * ram_save_target_page: save one target page |
a82d593b | 1296 | * |
3d0684b2 | 1297 | * Returns the number of pages written |
a82d593b | 1298 | * |
6f37bb8b | 1299 | * @rs: current RAM state |
3d0684b2 | 1300 | * @ms: current migration state |
3d0684b2 | 1301 | * @pss: data about the page we want to send |
a82d593b | 1302 | * @last_stage: if we are at the completion stage |
06b10688 | 1303 | * @page_abs: page number of the dirty page |
a82d593b | 1304 | */ |
a0a8aa14 | 1305 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss, |
06b10688 | 1306 | bool last_stage, unsigned long page_abs) |
a82d593b DDAG |
1307 | { |
1308 | int res = 0; | |
1309 | ||
1310 | /* Check the pages is dirty and if it is send it */ | |
06b10688 | 1311 | if (migration_bitmap_clear_dirty(rs, page_abs)) { |
a82d593b | 1312 | unsigned long *unsentmap; |
6d358d94 JQ |
1313 | /* |
1314 | * If xbzrle is on, stop using the data compression after first | |
1315 | * round of migration even if compression is enabled. In theory, | |
1316 | * xbzrle can do better than compression. | |
1317 | */ | |
1318 | ||
1319 | if (migrate_use_compression() | |
1320 | && (rs->ram_bulk_stage || !migrate_use_xbzrle())) { | |
a0a8aa14 | 1321 | res = ram_save_compressed_page(rs, pss, last_stage); |
a82d593b | 1322 | } else { |
a0a8aa14 | 1323 | res = ram_save_page(rs, pss, last_stage); |
a82d593b DDAG |
1324 | } |
1325 | ||
1326 | if (res < 0) { | |
1327 | return res; | |
1328 | } | |
eb859c53 | 1329 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
a82d593b | 1330 | if (unsentmap) { |
06b10688 | 1331 | clear_bit(page_abs, unsentmap); |
a82d593b DDAG |
1332 | } |
1333 | } | |
1334 | ||
1335 | return res; | |
1336 | } | |
1337 | ||
1338 | /** | |
3d0684b2 | 1339 | * ram_save_host_page: save a whole host page |
a82d593b | 1340 | * |
3d0684b2 JQ |
1341 | * Starting at *offset send pages up to the end of the current host |
1342 | * page. It's valid for the initial offset to point into the middle of | |
1343 | * a host page in which case the remainder of the hostpage is sent. | |
1344 | * Only dirty target pages are sent. Note that the host page size may | |
1345 | * be a huge page for this block. | |
a82d593b | 1346 | * |
3d0684b2 JQ |
1347 | * Returns the number of pages written or negative on error |
1348 | * | |
6f37bb8b | 1349 | * @rs: current RAM state |
3d0684b2 | 1350 | * @ms: current migration state |
3d0684b2 | 1351 | * @pss: data about the page we want to send |
a82d593b | 1352 | * @last_stage: if we are at the completion stage |
06b10688 | 1353 | * @page_abs: Page number of the dirty page |
a82d593b | 1354 | */ |
a0a8aa14 | 1355 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss, |
a08f6890 | 1356 | bool last_stage, |
06b10688 | 1357 | unsigned long page_abs) |
a82d593b DDAG |
1358 | { |
1359 | int tmppages, pages = 0; | |
4c011c37 DDAG |
1360 | size_t pagesize = qemu_ram_pagesize(pss->block); |
1361 | ||
a82d593b | 1362 | do { |
06b10688 | 1363 | tmppages = ram_save_target_page(rs, pss, last_stage, page_abs); |
a82d593b DDAG |
1364 | if (tmppages < 0) { |
1365 | return tmppages; | |
1366 | } | |
1367 | ||
1368 | pages += tmppages; | |
a08f6890 | 1369 | pss->offset += TARGET_PAGE_SIZE; |
06b10688 | 1370 | page_abs++; |
4c011c37 | 1371 | } while (pss->offset & (pagesize - 1)); |
a82d593b DDAG |
1372 | |
1373 | /* The offset we leave with is the last one we looked at */ | |
a08f6890 | 1374 | pss->offset -= TARGET_PAGE_SIZE; |
a82d593b DDAG |
1375 | return pages; |
1376 | } | |
6c595cde | 1377 | |
56e93d26 | 1378 | /** |
3d0684b2 | 1379 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1380 | * |
1381 | * Called within an RCU critical section. | |
1382 | * | |
3d0684b2 | 1383 | * Returns the number of pages written where zero means no dirty pages |
56e93d26 | 1384 | * |
6f37bb8b | 1385 | * @rs: current RAM state |
56e93d26 | 1386 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1387 | * |
1388 | * On systems where host-page-size > target-page-size it will send all the | |
1389 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1390 | */ |
1391 | ||
ce25d337 | 1392 | static int ram_find_and_save_block(RAMState *rs, bool last_stage) |
56e93d26 | 1393 | { |
b8fb8cb7 | 1394 | PageSearchStatus pss; |
56e93d26 | 1395 | int pages = 0; |
b9e60928 | 1396 | bool again, found; |
06b10688 | 1397 | unsigned long page_abs; /* Page number of the dirty page */ |
56e93d26 | 1398 | |
0827b9e9 AA |
1399 | /* No dirty page as there is zero RAM */ |
1400 | if (!ram_bytes_total()) { | |
1401 | return pages; | |
1402 | } | |
1403 | ||
6f37bb8b | 1404 | pss.block = rs->last_seen_block; |
269ace29 | 1405 | pss.offset = rs->last_page << TARGET_PAGE_BITS; |
b8fb8cb7 DDAG |
1406 | pss.complete_round = false; |
1407 | ||
1408 | if (!pss.block) { | |
1409 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1410 | } | |
56e93d26 | 1411 | |
b9e60928 | 1412 | do { |
a82d593b | 1413 | again = true; |
06b10688 | 1414 | found = get_queued_page(rs, &pss, &page_abs); |
b9e60928 | 1415 | |
a82d593b DDAG |
1416 | if (!found) { |
1417 | /* priority queue empty, so just search for something dirty */ | |
06b10688 | 1418 | found = find_dirty_block(rs, &pss, &again, &page_abs); |
a82d593b | 1419 | } |
f3f491fc | 1420 | |
a82d593b | 1421 | if (found) { |
06b10688 | 1422 | pages = ram_save_host_page(rs, &pss, last_stage, page_abs); |
56e93d26 | 1423 | } |
b9e60928 | 1424 | } while (!pages && again); |
56e93d26 | 1425 | |
6f37bb8b | 1426 | rs->last_seen_block = pss.block; |
269ace29 | 1427 | rs->last_page = pss.offset >> TARGET_PAGE_BITS; |
56e93d26 JQ |
1428 | |
1429 | return pages; | |
1430 | } | |
1431 | ||
1432 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1433 | { | |
1434 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b JQ |
1435 | RAMState *rs = &ram_state; |
1436 | ||
56e93d26 | 1437 | if (zero) { |
f7ccd61b | 1438 | rs->zero_pages += pages; |
56e93d26 | 1439 | } else { |
b4d1c6e7 | 1440 | rs->norm_pages += pages; |
2f4fde93 | 1441 | rs->bytes_transferred += size; |
56e93d26 JQ |
1442 | qemu_update_position(f, size); |
1443 | } | |
1444 | } | |
1445 | ||
56e93d26 JQ |
1446 | uint64_t ram_bytes_total(void) |
1447 | { | |
1448 | RAMBlock *block; | |
1449 | uint64_t total = 0; | |
1450 | ||
1451 | rcu_read_lock(); | |
1452 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) | |
1453 | total += block->used_length; | |
1454 | rcu_read_unlock(); | |
1455 | return total; | |
1456 | } | |
1457 | ||
1458 | void free_xbzrle_decoded_buf(void) | |
1459 | { | |
1460 | g_free(xbzrle_decoded_buf); | |
1461 | xbzrle_decoded_buf = NULL; | |
1462 | } | |
1463 | ||
eb859c53 | 1464 | static void migration_bitmap_free(struct RAMBitmap *bmap) |
60be6340 DL |
1465 | { |
1466 | g_free(bmap->bmap); | |
f3f491fc | 1467 | g_free(bmap->unsentmap); |
60be6340 DL |
1468 | g_free(bmap); |
1469 | } | |
1470 | ||
6ad2a215 | 1471 | static void ram_migration_cleanup(void *opaque) |
56e93d26 | 1472 | { |
eb859c53 JQ |
1473 | RAMState *rs = opaque; |
1474 | ||
2ff64038 LZ |
1475 | /* caller have hold iothread lock or is in a bh, so there is |
1476 | * no writing race against this migration_bitmap | |
1477 | */ | |
eb859c53 JQ |
1478 | struct RAMBitmap *bitmap = rs->ram_bitmap; |
1479 | atomic_rcu_set(&rs->ram_bitmap, NULL); | |
2ff64038 | 1480 | if (bitmap) { |
56e93d26 | 1481 | memory_global_dirty_log_stop(); |
60be6340 | 1482 | call_rcu(bitmap, migration_bitmap_free, rcu); |
56e93d26 JQ |
1483 | } |
1484 | ||
1485 | XBZRLE_cache_lock(); | |
1486 | if (XBZRLE.cache) { | |
1487 | cache_fini(XBZRLE.cache); | |
1488 | g_free(XBZRLE.encoded_buf); | |
1489 | g_free(XBZRLE.current_buf); | |
adb65dec | 1490 | g_free(ZERO_TARGET_PAGE); |
56e93d26 JQ |
1491 | XBZRLE.cache = NULL; |
1492 | XBZRLE.encoded_buf = NULL; | |
1493 | XBZRLE.current_buf = NULL; | |
1494 | } | |
1495 | XBZRLE_cache_unlock(); | |
1496 | } | |
1497 | ||
6f37bb8b | 1498 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1499 | { |
6f37bb8b JQ |
1500 | rs->last_seen_block = NULL; |
1501 | rs->last_sent_block = NULL; | |
269ace29 | 1502 | rs->last_page = 0; |
6f37bb8b JQ |
1503 | rs->last_version = ram_list.version; |
1504 | rs->ram_bulk_stage = true; | |
56e93d26 JQ |
1505 | } |
1506 | ||
1507 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1508 | ||
dd631697 LZ |
1509 | void migration_bitmap_extend(ram_addr_t old, ram_addr_t new) |
1510 | { | |
0d8ec885 | 1511 | RAMState *rs = &ram_state; |
108cfae0 | 1512 | |
dd631697 LZ |
1513 | /* called in qemu main thread, so there is |
1514 | * no writing race against this migration_bitmap | |
1515 | */ | |
eb859c53 JQ |
1516 | if (rs->ram_bitmap) { |
1517 | struct RAMBitmap *old_bitmap = rs->ram_bitmap, *bitmap; | |
1518 | bitmap = g_new(struct RAMBitmap, 1); | |
60be6340 | 1519 | bitmap->bmap = bitmap_new(new); |
dd631697 LZ |
1520 | |
1521 | /* prevent migration_bitmap content from being set bit | |
1522 | * by migration_bitmap_sync_range() at the same time. | |
1523 | * it is safe to migration if migration_bitmap is cleared bit | |
1524 | * at the same time. | |
1525 | */ | |
108cfae0 | 1526 | qemu_mutex_lock(&rs->bitmap_mutex); |
60be6340 DL |
1527 | bitmap_copy(bitmap->bmap, old_bitmap->bmap, old); |
1528 | bitmap_set(bitmap->bmap, old, new - old); | |
f3f491fc DDAG |
1529 | |
1530 | /* We don't have a way to safely extend the sentmap | |
1531 | * with RCU; so mark it as missing, entry to postcopy | |
1532 | * will fail. | |
1533 | */ | |
1534 | bitmap->unsentmap = NULL; | |
1535 | ||
eb859c53 | 1536 | atomic_rcu_set(&rs->ram_bitmap, bitmap); |
108cfae0 | 1537 | qemu_mutex_unlock(&rs->bitmap_mutex); |
0d8ec885 | 1538 | rs->migration_dirty_pages += new - old; |
60be6340 | 1539 | call_rcu(old_bitmap, migration_bitmap_free, rcu); |
dd631697 LZ |
1540 | } |
1541 | } | |
56e93d26 | 1542 | |
4f2e4252 DDAG |
1543 | /* |
1544 | * 'expected' is the value you expect the bitmap mostly to be full | |
1545 | * of; it won't bother printing lines that are all this value. | |
1546 | * If 'todump' is null the migration bitmap is dumped. | |
1547 | */ | |
1548 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected) | |
1549 | { | |
1550 | int64_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
eb859c53 | 1551 | RAMState *rs = &ram_state; |
4f2e4252 DDAG |
1552 | int64_t cur; |
1553 | int64_t linelen = 128; | |
1554 | char linebuf[129]; | |
1555 | ||
1556 | if (!todump) { | |
eb859c53 | 1557 | todump = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
4f2e4252 DDAG |
1558 | } |
1559 | ||
1560 | for (cur = 0; cur < ram_pages; cur += linelen) { | |
1561 | int64_t curb; | |
1562 | bool found = false; | |
1563 | /* | |
1564 | * Last line; catch the case where the line length | |
1565 | * is longer than remaining ram | |
1566 | */ | |
1567 | if (cur + linelen > ram_pages) { | |
1568 | linelen = ram_pages - cur; | |
1569 | } | |
1570 | for (curb = 0; curb < linelen; curb++) { | |
1571 | bool thisbit = test_bit(cur + curb, todump); | |
1572 | linebuf[curb] = thisbit ? '1' : '.'; | |
1573 | found = found || (thisbit != expected); | |
1574 | } | |
1575 | if (found) { | |
1576 | linebuf[curb] = '\0'; | |
1577 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1578 | } | |
1579 | } | |
1580 | } | |
1581 | ||
e0b266f0 DDAG |
1582 | /* **** functions for postcopy ***** */ |
1583 | ||
ced1c616 PB |
1584 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1585 | { | |
eb859c53 | 1586 | RAMState *rs = &ram_state; |
ced1c616 | 1587 | struct RAMBlock *block; |
eb859c53 | 1588 | unsigned long *bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
ced1c616 PB |
1589 | |
1590 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1591 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1592 | unsigned long range = first + (block->used_length >> TARGET_PAGE_BITS); | |
1593 | unsigned long run_start = find_next_zero_bit(bitmap, range, first); | |
1594 | ||
1595 | while (run_start < range) { | |
1596 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
aaa2064c | 1597 | ram_discard_range(block->idstr, run_start << TARGET_PAGE_BITS, |
ced1c616 PB |
1598 | (run_end - run_start) << TARGET_PAGE_BITS); |
1599 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); | |
1600 | } | |
1601 | } | |
1602 | } | |
1603 | ||
3d0684b2 JQ |
1604 | /** |
1605 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1606 | * | |
1607 | * Returns zero on success | |
1608 | * | |
e0b266f0 DDAG |
1609 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
1610 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1611 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
3d0684b2 JQ |
1612 | * |
1613 | * @ms: current migration state | |
1614 | * @pds: state for postcopy | |
1615 | * @start: RAMBlock starting page | |
1616 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1617 | */ |
1618 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1619 | PostcopyDiscardState *pds, | |
1620 | unsigned long start, | |
1621 | unsigned long length) | |
1622 | { | |
eb859c53 | 1623 | RAMState *rs = &ram_state; |
e0b266f0 DDAG |
1624 | unsigned long end = start + length; /* one after the end */ |
1625 | unsigned long current; | |
1626 | unsigned long *unsentmap; | |
1627 | ||
eb859c53 | 1628 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
e0b266f0 DDAG |
1629 | for (current = start; current < end; ) { |
1630 | unsigned long one = find_next_bit(unsentmap, end, current); | |
1631 | ||
1632 | if (one <= end) { | |
1633 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1634 | unsigned long discard_length; | |
1635 | ||
1636 | if (zero >= end) { | |
1637 | discard_length = end - one; | |
1638 | } else { | |
1639 | discard_length = zero - one; | |
1640 | } | |
d688c62d DDAG |
1641 | if (discard_length) { |
1642 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1643 | } | |
e0b266f0 DDAG |
1644 | current = one + discard_length; |
1645 | } else { | |
1646 | current = one; | |
1647 | } | |
1648 | } | |
1649 | ||
1650 | return 0; | |
1651 | } | |
1652 | ||
3d0684b2 JQ |
1653 | /** |
1654 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1655 | * | |
1656 | * Returns 0 for success or negative for error | |
1657 | * | |
e0b266f0 DDAG |
1658 | * Utility for the outgoing postcopy code. |
1659 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1660 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1661 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1662 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1663 | * |
1664 | * @ms: current migration state | |
e0b266f0 DDAG |
1665 | */ |
1666 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1667 | { | |
1668 | struct RAMBlock *block; | |
1669 | int ret; | |
1670 | ||
1671 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1672 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1673 | PostcopyDiscardState *pds = postcopy_discard_send_init(ms, | |
1674 | first, | |
1675 | block->idstr); | |
1676 | ||
1677 | /* | |
1678 | * Postcopy sends chunks of bitmap over the wire, but it | |
1679 | * just needs indexes at this point, avoids it having | |
1680 | * target page specific code. | |
1681 | */ | |
1682 | ret = postcopy_send_discard_bm_ram(ms, pds, first, | |
1683 | block->used_length >> TARGET_PAGE_BITS); | |
1684 | postcopy_discard_send_finish(ms, pds); | |
1685 | if (ret) { | |
1686 | return ret; | |
1687 | } | |
1688 | } | |
1689 | ||
1690 | return 0; | |
1691 | } | |
1692 | ||
3d0684b2 JQ |
1693 | /** |
1694 | * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages | |
1695 | * | |
1696 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
1697 | * canonicalize the two bitmaps, that are similar, but one is | |
1698 | * inverted. | |
99e314eb | 1699 | * |
3d0684b2 JQ |
1700 | * Postcopy requires that all target pages in a hostpage are dirty or |
1701 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 1702 | * |
3d0684b2 JQ |
1703 | * @ms: current migration state |
1704 | * @unsent_pass: if true we need to canonicalize partially unsent host pages | |
1705 | * otherwise we need to canonicalize partially dirty host pages | |
1706 | * @block: block that contains the page we want to canonicalize | |
1707 | * @pds: state for postcopy | |
99e314eb DDAG |
1708 | */ |
1709 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1710 | RAMBlock *block, | |
1711 | PostcopyDiscardState *pds) | |
1712 | { | |
0d8ec885 | 1713 | RAMState *rs = &ram_state; |
99e314eb DDAG |
1714 | unsigned long *bitmap; |
1715 | unsigned long *unsentmap; | |
29c59172 | 1716 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
99e314eb DDAG |
1717 | unsigned long first = block->offset >> TARGET_PAGE_BITS; |
1718 | unsigned long len = block->used_length >> TARGET_PAGE_BITS; | |
1719 | unsigned long last = first + (len - 1); | |
1720 | unsigned long run_start; | |
1721 | ||
29c59172 DDAG |
1722 | if (block->page_size == TARGET_PAGE_SIZE) { |
1723 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
1724 | return; | |
1725 | } | |
1726 | ||
eb859c53 JQ |
1727 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
1728 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; | |
99e314eb DDAG |
1729 | |
1730 | if (unsent_pass) { | |
1731 | /* Find a sent page */ | |
1732 | run_start = find_next_zero_bit(unsentmap, last + 1, first); | |
1733 | } else { | |
1734 | /* Find a dirty page */ | |
1735 | run_start = find_next_bit(bitmap, last + 1, first); | |
1736 | } | |
1737 | ||
1738 | while (run_start <= last) { | |
1739 | bool do_fixup = false; | |
1740 | unsigned long fixup_start_addr; | |
1741 | unsigned long host_offset; | |
1742 | ||
1743 | /* | |
1744 | * If the start of this run of pages is in the middle of a host | |
1745 | * page, then we need to fixup this host page. | |
1746 | */ | |
1747 | host_offset = run_start % host_ratio; | |
1748 | if (host_offset) { | |
1749 | do_fixup = true; | |
1750 | run_start -= host_offset; | |
1751 | fixup_start_addr = run_start; | |
1752 | /* For the next pass */ | |
1753 | run_start = run_start + host_ratio; | |
1754 | } else { | |
1755 | /* Find the end of this run */ | |
1756 | unsigned long run_end; | |
1757 | if (unsent_pass) { | |
1758 | run_end = find_next_bit(unsentmap, last + 1, run_start + 1); | |
1759 | } else { | |
1760 | run_end = find_next_zero_bit(bitmap, last + 1, run_start + 1); | |
1761 | } | |
1762 | /* | |
1763 | * If the end isn't at the start of a host page, then the | |
1764 | * run doesn't finish at the end of a host page | |
1765 | * and we need to discard. | |
1766 | */ | |
1767 | host_offset = run_end % host_ratio; | |
1768 | if (host_offset) { | |
1769 | do_fixup = true; | |
1770 | fixup_start_addr = run_end - host_offset; | |
1771 | /* | |
1772 | * This host page has gone, the next loop iteration starts | |
1773 | * from after the fixup | |
1774 | */ | |
1775 | run_start = fixup_start_addr + host_ratio; | |
1776 | } else { | |
1777 | /* | |
1778 | * No discards on this iteration, next loop starts from | |
1779 | * next sent/dirty page | |
1780 | */ | |
1781 | run_start = run_end + 1; | |
1782 | } | |
1783 | } | |
1784 | ||
1785 | if (do_fixup) { | |
1786 | unsigned long page; | |
1787 | ||
1788 | /* Tell the destination to discard this page */ | |
1789 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1790 | /* For the unsent_pass we: | |
1791 | * discard partially sent pages | |
1792 | * For the !unsent_pass (dirty) we: | |
1793 | * discard partially dirty pages that were sent | |
1794 | * (any partially sent pages were already discarded | |
1795 | * by the previous unsent_pass) | |
1796 | */ | |
1797 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1798 | host_ratio); | |
1799 | } | |
1800 | ||
1801 | /* Clean up the bitmap */ | |
1802 | for (page = fixup_start_addr; | |
1803 | page < fixup_start_addr + host_ratio; page++) { | |
1804 | /* All pages in this host page are now not sent */ | |
1805 | set_bit(page, unsentmap); | |
1806 | ||
1807 | /* | |
1808 | * Remark them as dirty, updating the count for any pages | |
1809 | * that weren't previously dirty. | |
1810 | */ | |
0d8ec885 | 1811 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
1812 | } |
1813 | } | |
1814 | ||
1815 | if (unsent_pass) { | |
1816 | /* Find the next sent page for the next iteration */ | |
1817 | run_start = find_next_zero_bit(unsentmap, last + 1, | |
1818 | run_start); | |
1819 | } else { | |
1820 | /* Find the next dirty page for the next iteration */ | |
1821 | run_start = find_next_bit(bitmap, last + 1, run_start); | |
1822 | } | |
1823 | } | |
1824 | } | |
1825 | ||
3d0684b2 JQ |
1826 | /** |
1827 | * postcopy_chuck_hostpages: discrad any partially sent host page | |
1828 | * | |
99e314eb DDAG |
1829 | * Utility for the outgoing postcopy code. |
1830 | * | |
1831 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
1832 | * dirty host-page size chunks as all dirty. In this case the host-page |
1833 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 1834 | * |
3d0684b2 JQ |
1835 | * Returns zero on success |
1836 | * | |
1837 | * @ms: current migration state | |
99e314eb DDAG |
1838 | */ |
1839 | static int postcopy_chunk_hostpages(MigrationState *ms) | |
1840 | { | |
6f37bb8b | 1841 | RAMState *rs = &ram_state; |
99e314eb DDAG |
1842 | struct RAMBlock *block; |
1843 | ||
99e314eb | 1844 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
6f37bb8b JQ |
1845 | rs->last_seen_block = NULL; |
1846 | rs->last_sent_block = NULL; | |
269ace29 | 1847 | rs->last_page = 0; |
99e314eb DDAG |
1848 | |
1849 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1850 | unsigned long first = block->offset >> TARGET_PAGE_BITS; | |
1851 | ||
1852 | PostcopyDiscardState *pds = | |
1853 | postcopy_discard_send_init(ms, first, block->idstr); | |
1854 | ||
1855 | /* First pass: Discard all partially sent host pages */ | |
1856 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1857 | /* | |
1858 | * Second pass: Ensure that all partially dirty host pages are made | |
1859 | * fully dirty. | |
1860 | */ | |
1861 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
1862 | ||
1863 | postcopy_discard_send_finish(ms, pds); | |
1864 | } /* ram_list loop */ | |
1865 | ||
1866 | return 0; | |
1867 | } | |
1868 | ||
3d0684b2 JQ |
1869 | /** |
1870 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
1871 | * | |
1872 | * Returns zero on success | |
1873 | * | |
e0b266f0 DDAG |
1874 | * Transmit the set of pages to be discarded after precopy to the target |
1875 | * these are pages that: | |
1876 | * a) Have been previously transmitted but are now dirty again | |
1877 | * b) Pages that have never been transmitted, this ensures that | |
1878 | * any pages on the destination that have been mapped by background | |
1879 | * tasks get discarded (transparent huge pages is the specific concern) | |
1880 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
1881 | * |
1882 | * @ms: current migration state | |
e0b266f0 DDAG |
1883 | */ |
1884 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1885 | { | |
eb859c53 | 1886 | RAMState *rs = &ram_state; |
e0b266f0 DDAG |
1887 | int ret; |
1888 | unsigned long *bitmap, *unsentmap; | |
1889 | ||
1890 | rcu_read_lock(); | |
1891 | ||
1892 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 1893 | migration_bitmap_sync(rs); |
e0b266f0 | 1894 | |
eb859c53 | 1895 | unsentmap = atomic_rcu_read(&rs->ram_bitmap)->unsentmap; |
e0b266f0 DDAG |
1896 | if (!unsentmap) { |
1897 | /* We don't have a safe way to resize the sentmap, so | |
1898 | * if the bitmap was resized it will be NULL at this | |
1899 | * point. | |
1900 | */ | |
1901 | error_report("migration ram resized during precopy phase"); | |
1902 | rcu_read_unlock(); | |
1903 | return -EINVAL; | |
1904 | } | |
1905 | ||
29c59172 | 1906 | /* Deal with TPS != HPS and huge pages */ |
99e314eb DDAG |
1907 | ret = postcopy_chunk_hostpages(ms); |
1908 | if (ret) { | |
1909 | rcu_read_unlock(); | |
1910 | return ret; | |
1911 | } | |
1912 | ||
e0b266f0 DDAG |
1913 | /* |
1914 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1915 | */ | |
eb859c53 | 1916 | bitmap = atomic_rcu_read(&rs->ram_bitmap)->bmap; |
e0b266f0 DDAG |
1917 | bitmap_or(unsentmap, unsentmap, bitmap, |
1918 | last_ram_offset() >> TARGET_PAGE_BITS); | |
1919 | ||
1920 | ||
1921 | trace_ram_postcopy_send_discard_bitmap(); | |
1922 | #ifdef DEBUG_POSTCOPY | |
1923 | ram_debug_dump_bitmap(unsentmap, true); | |
1924 | #endif | |
1925 | ||
1926 | ret = postcopy_each_ram_send_discard(ms); | |
1927 | rcu_read_unlock(); | |
1928 | ||
1929 | return ret; | |
1930 | } | |
1931 | ||
3d0684b2 JQ |
1932 | /** |
1933 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 1934 | * |
3d0684b2 | 1935 | * Returns zero on success |
e0b266f0 | 1936 | * |
36449157 JQ |
1937 | * @rbname: name of the RAMBlock of the request. NULL means the |
1938 | * same that last one. | |
3d0684b2 JQ |
1939 | * @start: RAMBlock starting page |
1940 | * @length: RAMBlock size | |
e0b266f0 | 1941 | */ |
aaa2064c | 1942 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 DDAG |
1943 | { |
1944 | int ret = -1; | |
1945 | ||
36449157 | 1946 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 1947 | |
e0b266f0 | 1948 | rcu_read_lock(); |
36449157 | 1949 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
1950 | |
1951 | if (!rb) { | |
36449157 | 1952 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
e0b266f0 DDAG |
1953 | goto err; |
1954 | } | |
1955 | ||
d3a5038c | 1956 | ret = ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
1957 | |
1958 | err: | |
1959 | rcu_read_unlock(); | |
1960 | ||
1961 | return ret; | |
1962 | } | |
1963 | ||
ceb4d168 | 1964 | static int ram_state_init(RAMState *rs) |
56e93d26 | 1965 | { |
56e93d26 JQ |
1966 | int64_t ram_bitmap_pages; /* Size of bitmap in pages, including gaps */ |
1967 | ||
ceb4d168 | 1968 | memset(rs, 0, sizeof(*rs)); |
108cfae0 | 1969 | qemu_mutex_init(&rs->bitmap_mutex); |
ec481c6c JQ |
1970 | qemu_mutex_init(&rs->src_page_req_mutex); |
1971 | QSIMPLEQ_INIT(&rs->src_page_requests); | |
56e93d26 JQ |
1972 | |
1973 | if (migrate_use_xbzrle()) { | |
1974 | XBZRLE_cache_lock(); | |
adb65dec | 1975 | ZERO_TARGET_PAGE = g_malloc0(TARGET_PAGE_SIZE); |
56e93d26 JQ |
1976 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() / |
1977 | TARGET_PAGE_SIZE, | |
1978 | TARGET_PAGE_SIZE); | |
1979 | if (!XBZRLE.cache) { | |
1980 | XBZRLE_cache_unlock(); | |
1981 | error_report("Error creating cache"); | |
1982 | return -1; | |
1983 | } | |
1984 | XBZRLE_cache_unlock(); | |
1985 | ||
1986 | /* We prefer not to abort if there is no memory */ | |
1987 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
1988 | if (!XBZRLE.encoded_buf) { | |
1989 | error_report("Error allocating encoded_buf"); | |
1990 | return -1; | |
1991 | } | |
1992 | ||
1993 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
1994 | if (!XBZRLE.current_buf) { | |
1995 | error_report("Error allocating current_buf"); | |
1996 | g_free(XBZRLE.encoded_buf); | |
1997 | XBZRLE.encoded_buf = NULL; | |
1998 | return -1; | |
1999 | } | |
56e93d26 JQ |
2000 | } |
2001 | ||
49877834 PB |
2002 | /* For memory_global_dirty_log_start below. */ |
2003 | qemu_mutex_lock_iothread(); | |
2004 | ||
56e93d26 JQ |
2005 | qemu_mutex_lock_ramlist(); |
2006 | rcu_read_lock(); | |
6f37bb8b | 2007 | ram_state_reset(rs); |
56e93d26 | 2008 | |
eb859c53 | 2009 | rs->ram_bitmap = g_new0(struct RAMBitmap, 1); |
0827b9e9 AA |
2010 | /* Skip setting bitmap if there is no RAM */ |
2011 | if (ram_bytes_total()) { | |
2012 | ram_bitmap_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
eb859c53 JQ |
2013 | rs->ram_bitmap->bmap = bitmap_new(ram_bitmap_pages); |
2014 | bitmap_set(rs->ram_bitmap->bmap, 0, ram_bitmap_pages); | |
0827b9e9 AA |
2015 | |
2016 | if (migrate_postcopy_ram()) { | |
eb859c53 JQ |
2017 | rs->ram_bitmap->unsentmap = bitmap_new(ram_bitmap_pages); |
2018 | bitmap_set(rs->ram_bitmap->unsentmap, 0, ram_bitmap_pages); | |
0827b9e9 | 2019 | } |
f3f491fc DDAG |
2020 | } |
2021 | ||
56e93d26 JQ |
2022 | /* |
2023 | * Count the total number of pages used by ram blocks not including any | |
2024 | * gaps due to alignment or unplugs. | |
2025 | */ | |
0d8ec885 | 2026 | rs->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
56e93d26 JQ |
2027 | |
2028 | memory_global_dirty_log_start(); | |
8d820d6f | 2029 | migration_bitmap_sync(rs); |
56e93d26 | 2030 | qemu_mutex_unlock_ramlist(); |
49877834 | 2031 | qemu_mutex_unlock_iothread(); |
a91246c9 HZ |
2032 | rcu_read_unlock(); |
2033 | ||
2034 | return 0; | |
2035 | } | |
2036 | ||
3d0684b2 JQ |
2037 | /* |
2038 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2039 | * long-running RCU critical section. When rcu-reclaims in the code |
2040 | * start to become numerous it will be necessary to reduce the | |
2041 | * granularity of these critical sections. | |
2042 | */ | |
2043 | ||
3d0684b2 JQ |
2044 | /** |
2045 | * ram_save_setup: Setup RAM for migration | |
2046 | * | |
2047 | * Returns zero to indicate success and negative for error | |
2048 | * | |
2049 | * @f: QEMUFile where to send the data | |
2050 | * @opaque: RAMState pointer | |
2051 | */ | |
a91246c9 HZ |
2052 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2053 | { | |
6f37bb8b | 2054 | RAMState *rs = opaque; |
a91246c9 HZ |
2055 | RAMBlock *block; |
2056 | ||
2057 | /* migration has already setup the bitmap, reuse it. */ | |
2058 | if (!migration_in_colo_state()) { | |
ceb4d168 | 2059 | if (ram_state_init(rs) < 0) { |
a91246c9 HZ |
2060 | return -1; |
2061 | } | |
2062 | } | |
204b88b8 | 2063 | rs->f = f; |
a91246c9 HZ |
2064 | |
2065 | rcu_read_lock(); | |
56e93d26 JQ |
2066 | |
2067 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
2068 | ||
2069 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
2070 | qemu_put_byte(f, strlen(block->idstr)); | |
2071 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2072 | qemu_put_be64(f, block->used_length); | |
ef08fb38 DDAG |
2073 | if (migrate_postcopy_ram() && block->page_size != qemu_host_page_size) { |
2074 | qemu_put_be64(f, block->page_size); | |
2075 | } | |
56e93d26 JQ |
2076 | } |
2077 | ||
2078 | rcu_read_unlock(); | |
2079 | ||
2080 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
2081 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2082 | ||
2083 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2084 | ||
2085 | return 0; | |
2086 | } | |
2087 | ||
3d0684b2 JQ |
2088 | /** |
2089 | * ram_save_iterate: iterative stage for migration | |
2090 | * | |
2091 | * Returns zero to indicate success and negative for error | |
2092 | * | |
2093 | * @f: QEMUFile where to send the data | |
2094 | * @opaque: RAMState pointer | |
2095 | */ | |
56e93d26 JQ |
2096 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2097 | { | |
6f37bb8b | 2098 | RAMState *rs = opaque; |
56e93d26 JQ |
2099 | int ret; |
2100 | int i; | |
2101 | int64_t t0; | |
5c90308f | 2102 | int done = 0; |
56e93d26 JQ |
2103 | |
2104 | rcu_read_lock(); | |
6f37bb8b JQ |
2105 | if (ram_list.version != rs->last_version) { |
2106 | ram_state_reset(rs); | |
56e93d26 JQ |
2107 | } |
2108 | ||
2109 | /* Read version before ram_list.blocks */ | |
2110 | smp_rmb(); | |
2111 | ||
2112 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
2113 | ||
2114 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
2115 | i = 0; | |
2116 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
2117 | int pages; | |
2118 | ||
ce25d337 | 2119 | pages = ram_find_and_save_block(rs, false); |
56e93d26 JQ |
2120 | /* no more pages to sent */ |
2121 | if (pages == 0) { | |
5c90308f | 2122 | done = 1; |
56e93d26 JQ |
2123 | break; |
2124 | } | |
23b28c3c | 2125 | rs->iterations++; |
070afca2 | 2126 | |
56e93d26 JQ |
2127 | /* we want to check in the 1st loop, just in case it was the 1st time |
2128 | and we had to sync the dirty bitmap. | |
2129 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2130 | iterations | |
2131 | */ | |
2132 | if ((i & 63) == 0) { | |
2133 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2134 | if (t1 > MAX_WAIT) { | |
55c4446b | 2135 | trace_ram_save_iterate_big_wait(t1, i); |
56e93d26 JQ |
2136 | break; |
2137 | } | |
2138 | } | |
2139 | i++; | |
2140 | } | |
ce25d337 | 2141 | flush_compressed_data(rs); |
56e93d26 JQ |
2142 | rcu_read_unlock(); |
2143 | ||
2144 | /* | |
2145 | * Must occur before EOS (or any QEMUFile operation) | |
2146 | * because of RDMA protocol. | |
2147 | */ | |
2148 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2149 | ||
2150 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2f4fde93 | 2151 | rs->bytes_transferred += 8; |
56e93d26 JQ |
2152 | |
2153 | ret = qemu_file_get_error(f); | |
2154 | if (ret < 0) { | |
2155 | return ret; | |
2156 | } | |
2157 | ||
5c90308f | 2158 | return done; |
56e93d26 JQ |
2159 | } |
2160 | ||
3d0684b2 JQ |
2161 | /** |
2162 | * ram_save_complete: function called to send the remaining amount of ram | |
2163 | * | |
2164 | * Returns zero to indicate success | |
2165 | * | |
2166 | * Called with iothread lock | |
2167 | * | |
2168 | * @f: QEMUFile where to send the data | |
2169 | * @opaque: RAMState pointer | |
2170 | */ | |
56e93d26 JQ |
2171 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2172 | { | |
6f37bb8b JQ |
2173 | RAMState *rs = opaque; |
2174 | ||
56e93d26 JQ |
2175 | rcu_read_lock(); |
2176 | ||
5727309d | 2177 | if (!migration_in_postcopy()) { |
8d820d6f | 2178 | migration_bitmap_sync(rs); |
663e6c1d | 2179 | } |
56e93d26 JQ |
2180 | |
2181 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2182 | ||
2183 | /* try transferring iterative blocks of memory */ | |
2184 | ||
2185 | /* flush all remaining blocks regardless of rate limiting */ | |
2186 | while (true) { | |
2187 | int pages; | |
2188 | ||
ce25d337 | 2189 | pages = ram_find_and_save_block(rs, !migration_in_colo_state()); |
56e93d26 JQ |
2190 | /* no more blocks to sent */ |
2191 | if (pages == 0) { | |
2192 | break; | |
2193 | } | |
2194 | } | |
2195 | ||
ce25d337 | 2196 | flush_compressed_data(rs); |
56e93d26 | 2197 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 JQ |
2198 | |
2199 | rcu_read_unlock(); | |
d09a6fde | 2200 | |
56e93d26 JQ |
2201 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2202 | ||
2203 | return 0; | |
2204 | } | |
2205 | ||
c31b098f DDAG |
2206 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2207 | uint64_t *non_postcopiable_pending, | |
2208 | uint64_t *postcopiable_pending) | |
56e93d26 | 2209 | { |
8d820d6f | 2210 | RAMState *rs = opaque; |
56e93d26 JQ |
2211 | uint64_t remaining_size; |
2212 | ||
9edabd4d | 2213 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2214 | |
5727309d | 2215 | if (!migration_in_postcopy() && |
663e6c1d | 2216 | remaining_size < max_size) { |
56e93d26 JQ |
2217 | qemu_mutex_lock_iothread(); |
2218 | rcu_read_lock(); | |
8d820d6f | 2219 | migration_bitmap_sync(rs); |
56e93d26 JQ |
2220 | rcu_read_unlock(); |
2221 | qemu_mutex_unlock_iothread(); | |
9edabd4d | 2222 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2223 | } |
c31b098f DDAG |
2224 | |
2225 | /* We can do postcopy, and all the data is postcopiable */ | |
2226 | *postcopiable_pending += remaining_size; | |
56e93d26 JQ |
2227 | } |
2228 | ||
2229 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2230 | { | |
2231 | unsigned int xh_len; | |
2232 | int xh_flags; | |
063e760a | 2233 | uint8_t *loaded_data; |
56e93d26 JQ |
2234 | |
2235 | if (!xbzrle_decoded_buf) { | |
2236 | xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE); | |
2237 | } | |
063e760a | 2238 | loaded_data = xbzrle_decoded_buf; |
56e93d26 JQ |
2239 | |
2240 | /* extract RLE header */ | |
2241 | xh_flags = qemu_get_byte(f); | |
2242 | xh_len = qemu_get_be16(f); | |
2243 | ||
2244 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2245 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2246 | return -1; | |
2247 | } | |
2248 | ||
2249 | if (xh_len > TARGET_PAGE_SIZE) { | |
2250 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2251 | return -1; | |
2252 | } | |
2253 | /* load data and decode */ | |
063e760a | 2254 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2255 | |
2256 | /* decode RLE */ | |
063e760a | 2257 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2258 | TARGET_PAGE_SIZE) == -1) { |
2259 | error_report("Failed to load XBZRLE page - decode error!"); | |
2260 | return -1; | |
2261 | } | |
2262 | ||
2263 | return 0; | |
2264 | } | |
2265 | ||
3d0684b2 JQ |
2266 | /** |
2267 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2268 | * | |
2269 | * Must be called from within a rcu critical section. | |
2270 | * | |
56e93d26 | 2271 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2272 | * |
3d0684b2 JQ |
2273 | * @f: QEMUFile where to read the data from |
2274 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2275 | */ |
3d0684b2 | 2276 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2277 | { |
2278 | static RAMBlock *block = NULL; | |
2279 | char id[256]; | |
2280 | uint8_t len; | |
2281 | ||
2282 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2283 | if (!block) { |
56e93d26 JQ |
2284 | error_report("Ack, bad migration stream!"); |
2285 | return NULL; | |
2286 | } | |
4c4bad48 | 2287 | return block; |
56e93d26 JQ |
2288 | } |
2289 | ||
2290 | len = qemu_get_byte(f); | |
2291 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2292 | id[len] = 0; | |
2293 | ||
e3dd7493 | 2294 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2295 | if (!block) { |
2296 | error_report("Can't find block %s", id); | |
2297 | return NULL; | |
56e93d26 JQ |
2298 | } |
2299 | ||
4c4bad48 HZ |
2300 | return block; |
2301 | } | |
2302 | ||
2303 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2304 | ram_addr_t offset) | |
2305 | { | |
2306 | if (!offset_in_ramblock(block, offset)) { | |
2307 | return NULL; | |
2308 | } | |
2309 | ||
2310 | return block->host + offset; | |
56e93d26 JQ |
2311 | } |
2312 | ||
3d0684b2 JQ |
2313 | /** |
2314 | * ram_handle_compressed: handle the zero page case | |
2315 | * | |
56e93d26 JQ |
2316 | * If a page (or a whole RDMA chunk) has been |
2317 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2318 | * |
2319 | * @host: host address for the zero page | |
2320 | * @ch: what the page is filled from. We only support zero | |
2321 | * @size: size of the zero page | |
56e93d26 JQ |
2322 | */ |
2323 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2324 | { | |
2325 | if (ch != 0 || !is_zero_range(host, size)) { | |
2326 | memset(host, ch, size); | |
2327 | } | |
2328 | } | |
2329 | ||
2330 | static void *do_data_decompress(void *opaque) | |
2331 | { | |
2332 | DecompressParam *param = opaque; | |
2333 | unsigned long pagesize; | |
33d151f4 LL |
2334 | uint8_t *des; |
2335 | int len; | |
56e93d26 | 2336 | |
33d151f4 | 2337 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2338 | while (!param->quit) { |
33d151f4 LL |
2339 | if (param->des) { |
2340 | des = param->des; | |
2341 | len = param->len; | |
2342 | param->des = 0; | |
2343 | qemu_mutex_unlock(¶m->mutex); | |
2344 | ||
56e93d26 | 2345 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2346 | /* uncompress() will return failed in some case, especially |
2347 | * when the page is dirted when doing the compression, it's | |
2348 | * not a problem because the dirty page will be retransferred | |
2349 | * and uncompress() won't break the data in other pages. | |
2350 | */ | |
33d151f4 LL |
2351 | uncompress((Bytef *)des, &pagesize, |
2352 | (const Bytef *)param->compbuf, len); | |
73a8912b | 2353 | |
33d151f4 LL |
2354 | qemu_mutex_lock(&decomp_done_lock); |
2355 | param->done = true; | |
2356 | qemu_cond_signal(&decomp_done_cond); | |
2357 | qemu_mutex_unlock(&decomp_done_lock); | |
2358 | ||
2359 | qemu_mutex_lock(¶m->mutex); | |
2360 | } else { | |
2361 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2362 | } | |
56e93d26 | 2363 | } |
33d151f4 | 2364 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2365 | |
2366 | return NULL; | |
2367 | } | |
2368 | ||
5533b2e9 LL |
2369 | static void wait_for_decompress_done(void) |
2370 | { | |
2371 | int idx, thread_count; | |
2372 | ||
2373 | if (!migrate_use_compression()) { | |
2374 | return; | |
2375 | } | |
2376 | ||
2377 | thread_count = migrate_decompress_threads(); | |
2378 | qemu_mutex_lock(&decomp_done_lock); | |
2379 | for (idx = 0; idx < thread_count; idx++) { | |
2380 | while (!decomp_param[idx].done) { | |
2381 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2382 | } | |
2383 | } | |
2384 | qemu_mutex_unlock(&decomp_done_lock); | |
2385 | } | |
2386 | ||
56e93d26 JQ |
2387 | void migrate_decompress_threads_create(void) |
2388 | { | |
2389 | int i, thread_count; | |
2390 | ||
2391 | thread_count = migrate_decompress_threads(); | |
2392 | decompress_threads = g_new0(QemuThread, thread_count); | |
2393 | decomp_param = g_new0(DecompressParam, thread_count); | |
73a8912b LL |
2394 | qemu_mutex_init(&decomp_done_lock); |
2395 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2396 | for (i = 0; i < thread_count; i++) { |
2397 | qemu_mutex_init(&decomp_param[i].mutex); | |
2398 | qemu_cond_init(&decomp_param[i].cond); | |
2399 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2400 | decomp_param[i].done = true; |
90e56fb4 | 2401 | decomp_param[i].quit = false; |
56e93d26 JQ |
2402 | qemu_thread_create(decompress_threads + i, "decompress", |
2403 | do_data_decompress, decomp_param + i, | |
2404 | QEMU_THREAD_JOINABLE); | |
2405 | } | |
2406 | } | |
2407 | ||
2408 | void migrate_decompress_threads_join(void) | |
2409 | { | |
2410 | int i, thread_count; | |
2411 | ||
56e93d26 JQ |
2412 | thread_count = migrate_decompress_threads(); |
2413 | for (i = 0; i < thread_count; i++) { | |
2414 | qemu_mutex_lock(&decomp_param[i].mutex); | |
90e56fb4 | 2415 | decomp_param[i].quit = true; |
56e93d26 JQ |
2416 | qemu_cond_signal(&decomp_param[i].cond); |
2417 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2418 | } | |
2419 | for (i = 0; i < thread_count; i++) { | |
2420 | qemu_thread_join(decompress_threads + i); | |
2421 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2422 | qemu_cond_destroy(&decomp_param[i].cond); | |
2423 | g_free(decomp_param[i].compbuf); | |
2424 | } | |
2425 | g_free(decompress_threads); | |
2426 | g_free(decomp_param); | |
56e93d26 JQ |
2427 | decompress_threads = NULL; |
2428 | decomp_param = NULL; | |
56e93d26 JQ |
2429 | } |
2430 | ||
c1bc6626 | 2431 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2432 | void *host, int len) |
2433 | { | |
2434 | int idx, thread_count; | |
2435 | ||
2436 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2437 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2438 | while (true) { |
2439 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2440 | if (decomp_param[idx].done) { |
33d151f4 LL |
2441 | decomp_param[idx].done = false; |
2442 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2443 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2444 | decomp_param[idx].des = host; |
2445 | decomp_param[idx].len = len; | |
33d151f4 LL |
2446 | qemu_cond_signal(&decomp_param[idx].cond); |
2447 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2448 | break; |
2449 | } | |
2450 | } | |
2451 | if (idx < thread_count) { | |
2452 | break; | |
73a8912b LL |
2453 | } else { |
2454 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2455 | } |
2456 | } | |
73a8912b | 2457 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2458 | } |
2459 | ||
3d0684b2 JQ |
2460 | /** |
2461 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
2462 | * | |
2463 | * Returns 0 for success and negative if there was one error | |
2464 | * | |
2465 | * @mis: current migration incoming state | |
2466 | * | |
2467 | * Allocate data structures etc needed by incoming migration with | |
2468 | * postcopy-ram. postcopy-ram's similarly names | |
2469 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
2470 | */ |
2471 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2472 | { | |
2473 | size_t ram_pages = last_ram_offset() >> TARGET_PAGE_BITS; | |
2474 | ||
2475 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2476 | } | |
2477 | ||
3d0684b2 JQ |
2478 | /** |
2479 | * ram_load_postcopy: load a page in postcopy case | |
2480 | * | |
2481 | * Returns 0 for success or -errno in case of error | |
2482 | * | |
a7180877 DDAG |
2483 | * Called in postcopy mode by ram_load(). |
2484 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
2485 | * |
2486 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
2487 | */ |
2488 | static int ram_load_postcopy(QEMUFile *f) | |
2489 | { | |
2490 | int flags = 0, ret = 0; | |
2491 | bool place_needed = false; | |
28abd200 | 2492 | bool matching_page_sizes = false; |
a7180877 DDAG |
2493 | MigrationIncomingState *mis = migration_incoming_get_current(); |
2494 | /* Temporary page that is later 'placed' */ | |
2495 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2496 | void *last_host = NULL; |
a3b6ff6d | 2497 | bool all_zero = false; |
a7180877 DDAG |
2498 | |
2499 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2500 | ram_addr_t addr; | |
2501 | void *host = NULL; | |
2502 | void *page_buffer = NULL; | |
2503 | void *place_source = NULL; | |
df9ff5e1 | 2504 | RAMBlock *block = NULL; |
a7180877 | 2505 | uint8_t ch; |
a7180877 DDAG |
2506 | |
2507 | addr = qemu_get_be64(f); | |
2508 | flags = addr & ~TARGET_PAGE_MASK; | |
2509 | addr &= TARGET_PAGE_MASK; | |
2510 | ||
2511 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2512 | place_needed = false; | |
2513 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE)) { | |
df9ff5e1 | 2514 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
2515 | |
2516 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2517 | if (!host) { |
2518 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2519 | ret = -EINVAL; | |
2520 | break; | |
2521 | } | |
28abd200 | 2522 | matching_page_sizes = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 2523 | /* |
28abd200 DDAG |
2524 | * Postcopy requires that we place whole host pages atomically; |
2525 | * these may be huge pages for RAMBlocks that are backed by | |
2526 | * hugetlbfs. | |
a7180877 DDAG |
2527 | * To make it atomic, the data is read into a temporary page |
2528 | * that's moved into place later. | |
2529 | * The migration protocol uses, possibly smaller, target-pages | |
2530 | * however the source ensures it always sends all the components | |
2531 | * of a host page in order. | |
2532 | */ | |
2533 | page_buffer = postcopy_host_page + | |
28abd200 | 2534 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 2535 | /* If all TP are zero then we can optimise the place */ |
28abd200 | 2536 | if (!((uintptr_t)host & (block->page_size - 1))) { |
a7180877 | 2537 | all_zero = true; |
c53b7ddc DDAG |
2538 | } else { |
2539 | /* not the 1st TP within the HP */ | |
2540 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2541 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2542 | host, last_host); |
2543 | ret = -EINVAL; | |
2544 | break; | |
2545 | } | |
a7180877 DDAG |
2546 | } |
2547 | ||
c53b7ddc | 2548 | |
a7180877 DDAG |
2549 | /* |
2550 | * If it's the last part of a host page then we place the host | |
2551 | * page | |
2552 | */ | |
2553 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
28abd200 | 2554 | (block->page_size - 1)) == 0; |
a7180877 DDAG |
2555 | place_source = postcopy_host_page; |
2556 | } | |
c53b7ddc | 2557 | last_host = host; |
a7180877 DDAG |
2558 | |
2559 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
2560 | case RAM_SAVE_FLAG_COMPRESS: | |
2561 | ch = qemu_get_byte(f); | |
2562 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2563 | if (ch) { | |
2564 | all_zero = false; | |
2565 | } | |
2566 | break; | |
2567 | ||
2568 | case RAM_SAVE_FLAG_PAGE: | |
2569 | all_zero = false; | |
2570 | if (!place_needed || !matching_page_sizes) { | |
2571 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2572 | } else { | |
2573 | /* Avoids the qemu_file copy during postcopy, which is | |
2574 | * going to do a copy later; can only do it when we | |
2575 | * do this read in one go (matching page sizes) | |
2576 | */ | |
2577 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2578 | TARGET_PAGE_SIZE); | |
2579 | } | |
2580 | break; | |
2581 | case RAM_SAVE_FLAG_EOS: | |
2582 | /* normal exit */ | |
2583 | break; | |
2584 | default: | |
2585 | error_report("Unknown combination of migration flags: %#x" | |
2586 | " (postcopy mode)", flags); | |
2587 | ret = -EINVAL; | |
2588 | } | |
2589 | ||
2590 | if (place_needed) { | |
2591 | /* This gets called at the last target page in the host page */ | |
df9ff5e1 DDAG |
2592 | void *place_dest = host + TARGET_PAGE_SIZE - block->page_size; |
2593 | ||
a7180877 | 2594 | if (all_zero) { |
df9ff5e1 DDAG |
2595 | ret = postcopy_place_page_zero(mis, place_dest, |
2596 | block->page_size); | |
a7180877 | 2597 | } else { |
df9ff5e1 DDAG |
2598 | ret = postcopy_place_page(mis, place_dest, |
2599 | place_source, block->page_size); | |
a7180877 DDAG |
2600 | } |
2601 | } | |
2602 | if (!ret) { | |
2603 | ret = qemu_file_get_error(f); | |
2604 | } | |
2605 | } | |
2606 | ||
2607 | return ret; | |
2608 | } | |
2609 | ||
56e93d26 JQ |
2610 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2611 | { | |
2612 | int flags = 0, ret = 0; | |
2613 | static uint64_t seq_iter; | |
2614 | int len = 0; | |
a7180877 DDAG |
2615 | /* |
2616 | * If system is running in postcopy mode, page inserts to host memory must | |
2617 | * be atomic | |
2618 | */ | |
2619 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
ef08fb38 DDAG |
2620 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
2621 | bool postcopy_advised = postcopy_state_get() >= POSTCOPY_INCOMING_ADVISE; | |
56e93d26 JQ |
2622 | |
2623 | seq_iter++; | |
2624 | ||
2625 | if (version_id != 4) { | |
2626 | ret = -EINVAL; | |
2627 | } | |
2628 | ||
2629 | /* This RCU critical section can be very long running. | |
2630 | * When RCU reclaims in the code start to become numerous, | |
2631 | * it will be necessary to reduce the granularity of this | |
2632 | * critical section. | |
2633 | */ | |
2634 | rcu_read_lock(); | |
a7180877 DDAG |
2635 | |
2636 | if (postcopy_running) { | |
2637 | ret = ram_load_postcopy(f); | |
2638 | } | |
2639 | ||
2640 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2641 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2642 | void *host = NULL; |
56e93d26 JQ |
2643 | uint8_t ch; |
2644 | ||
2645 | addr = qemu_get_be64(f); | |
2646 | flags = addr & ~TARGET_PAGE_MASK; | |
2647 | addr &= TARGET_PAGE_MASK; | |
2648 | ||
a776aa15 DDAG |
2649 | if (flags & (RAM_SAVE_FLAG_COMPRESS | RAM_SAVE_FLAG_PAGE | |
2650 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { | |
4c4bad48 HZ |
2651 | RAMBlock *block = ram_block_from_stream(f, flags); |
2652 | ||
2653 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2654 | if (!host) { |
2655 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2656 | ret = -EINVAL; | |
2657 | break; | |
2658 | } | |
2659 | } | |
2660 | ||
56e93d26 JQ |
2661 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2662 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2663 | /* Synchronize RAM block list */ | |
2664 | total_ram_bytes = addr; | |
2665 | while (!ret && total_ram_bytes) { | |
2666 | RAMBlock *block; | |
56e93d26 JQ |
2667 | char id[256]; |
2668 | ram_addr_t length; | |
2669 | ||
2670 | len = qemu_get_byte(f); | |
2671 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2672 | id[len] = 0; | |
2673 | length = qemu_get_be64(f); | |
2674 | ||
e3dd7493 DDAG |
2675 | block = qemu_ram_block_by_name(id); |
2676 | if (block) { | |
2677 | if (length != block->used_length) { | |
2678 | Error *local_err = NULL; | |
56e93d26 | 2679 | |
fa53a0e5 | 2680 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2681 | &local_err); |
2682 | if (local_err) { | |
2683 | error_report_err(local_err); | |
56e93d26 | 2684 | } |
56e93d26 | 2685 | } |
ef08fb38 DDAG |
2686 | /* For postcopy we need to check hugepage sizes match */ |
2687 | if (postcopy_advised && | |
2688 | block->page_size != qemu_host_page_size) { | |
2689 | uint64_t remote_page_size = qemu_get_be64(f); | |
2690 | if (remote_page_size != block->page_size) { | |
2691 | error_report("Mismatched RAM page size %s " | |
2692 | "(local) %zd != %" PRId64, | |
2693 | id, block->page_size, | |
2694 | remote_page_size); | |
2695 | ret = -EINVAL; | |
2696 | } | |
2697 | } | |
e3dd7493 DDAG |
2698 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2699 | block->idstr); | |
2700 | } else { | |
56e93d26 JQ |
2701 | error_report("Unknown ramblock \"%s\", cannot " |
2702 | "accept migration", id); | |
2703 | ret = -EINVAL; | |
2704 | } | |
2705 | ||
2706 | total_ram_bytes -= length; | |
2707 | } | |
2708 | break; | |
a776aa15 | 2709 | |
56e93d26 | 2710 | case RAM_SAVE_FLAG_COMPRESS: |
56e93d26 JQ |
2711 | ch = qemu_get_byte(f); |
2712 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2713 | break; | |
a776aa15 | 2714 | |
56e93d26 | 2715 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2716 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2717 | break; | |
56e93d26 | 2718 | |
a776aa15 | 2719 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2720 | len = qemu_get_be32(f); |
2721 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2722 | error_report("Invalid compressed data length: %d", len); | |
2723 | ret = -EINVAL; | |
2724 | break; | |
2725 | } | |
c1bc6626 | 2726 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2727 | break; |
a776aa15 | 2728 | |
56e93d26 | 2729 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2730 | if (load_xbzrle(f, addr, host) < 0) { |
2731 | error_report("Failed to decompress XBZRLE page at " | |
2732 | RAM_ADDR_FMT, addr); | |
2733 | ret = -EINVAL; | |
2734 | break; | |
2735 | } | |
2736 | break; | |
2737 | case RAM_SAVE_FLAG_EOS: | |
2738 | /* normal exit */ | |
2739 | break; | |
2740 | default: | |
2741 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2742 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2743 | } else { |
2744 | error_report("Unknown combination of migration flags: %#x", | |
2745 | flags); | |
2746 | ret = -EINVAL; | |
2747 | } | |
2748 | } | |
2749 | if (!ret) { | |
2750 | ret = qemu_file_get_error(f); | |
2751 | } | |
2752 | } | |
2753 | ||
5533b2e9 | 2754 | wait_for_decompress_done(); |
56e93d26 | 2755 | rcu_read_unlock(); |
55c4446b | 2756 | trace_ram_load_complete(ret, seq_iter); |
56e93d26 JQ |
2757 | return ret; |
2758 | } | |
2759 | ||
2760 | static SaveVMHandlers savevm_ram_handlers = { | |
2761 | .save_live_setup = ram_save_setup, | |
2762 | .save_live_iterate = ram_save_iterate, | |
763c906b | 2763 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2764 | .save_live_complete_precopy = ram_save_complete, |
56e93d26 JQ |
2765 | .save_live_pending = ram_save_pending, |
2766 | .load_state = ram_load, | |
6ad2a215 | 2767 | .cleanup = ram_migration_cleanup, |
56e93d26 JQ |
2768 | }; |
2769 | ||
2770 | void ram_mig_init(void) | |
2771 | { | |
2772 | qemu_mutex_init(&XBZRLE.lock); | |
6f37bb8b | 2773 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 2774 | } |