]>
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" | |
709e3fe8 | 38 | #include "xbzrle.h" |
7b1e1a22 | 39 | #include "ram.h" |
56e93d26 | 40 | #include "migration/migration.h" |
7b1e1a22 | 41 | #include "migration/misc.h" |
08a0aee1 | 42 | #include "qemu-file.h" |
987772d9 | 43 | #include "migration/vmstate.h" |
be07b0ac | 44 | #include "postcopy-ram.h" |
56e93d26 JQ |
45 | #include "exec/address-spaces.h" |
46 | #include "migration/page_cache.h" | |
56e93d26 | 47 | #include "qemu/error-report.h" |
56e93d26 | 48 | #include "trace.h" |
56e93d26 | 49 | #include "exec/ram_addr.h" |
56e93d26 | 50 | #include "qemu/rcu_queue.h" |
a91246c9 | 51 | #include "migration/colo.h" |
56e93d26 | 52 | |
56e93d26 JQ |
53 | /***********************************************************/ |
54 | /* ram save/restore */ | |
55 | ||
bb890ed5 JQ |
56 | /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it |
57 | * worked for pages that where filled with the same char. We switched | |
58 | * it to only search for the zero value. And to avoid confusion with | |
59 | * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it. | |
60 | */ | |
61 | ||
56e93d26 | 62 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ |
bb890ed5 | 63 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
64 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
65 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
66 | #define RAM_SAVE_FLAG_EOS 0x10 | |
67 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
68 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
69 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
70 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
71 | ||
56e93d26 JQ |
72 | static inline bool is_zero_range(uint8_t *p, uint64_t size) |
73 | { | |
a1febc49 | 74 | return buffer_is_zero(p, size); |
56e93d26 JQ |
75 | } |
76 | ||
77 | /* struct contains XBZRLE cache and a static page | |
78 | used by the compression */ | |
79 | static struct { | |
80 | /* buffer used for XBZRLE encoding */ | |
81 | uint8_t *encoded_buf; | |
82 | /* buffer for storing page content */ | |
83 | uint8_t *current_buf; | |
84 | /* Cache for XBZRLE, Protected by lock. */ | |
85 | PageCache *cache; | |
86 | QemuMutex lock; | |
c00e0928 JQ |
87 | /* it will store a page full of zeros */ |
88 | uint8_t *zero_target_page; | |
56e93d26 JQ |
89 | } XBZRLE; |
90 | ||
91 | /* buffer used for XBZRLE decoding */ | |
92 | static uint8_t *xbzrle_decoded_buf; | |
93 | ||
94 | static void XBZRLE_cache_lock(void) | |
95 | { | |
96 | if (migrate_use_xbzrle()) | |
97 | qemu_mutex_lock(&XBZRLE.lock); | |
98 | } | |
99 | ||
100 | static void XBZRLE_cache_unlock(void) | |
101 | { | |
102 | if (migrate_use_xbzrle()) | |
103 | qemu_mutex_unlock(&XBZRLE.lock); | |
104 | } | |
105 | ||
3d0684b2 JQ |
106 | /** |
107 | * xbzrle_cache_resize: resize the xbzrle cache | |
108 | * | |
109 | * This function is called from qmp_migrate_set_cache_size in main | |
110 | * thread, possibly while a migration is in progress. A running | |
111 | * migration may be using the cache and might finish during this call, | |
112 | * hence changes to the cache are protected by XBZRLE.lock(). | |
113 | * | |
114 | * Returns the new_size or negative in case of error. | |
115 | * | |
116 | * @new_size: new cache size | |
56e93d26 JQ |
117 | */ |
118 | int64_t xbzrle_cache_resize(int64_t new_size) | |
119 | { | |
120 | PageCache *new_cache; | |
121 | int64_t ret; | |
122 | ||
123 | if (new_size < TARGET_PAGE_SIZE) { | |
124 | return -1; | |
125 | } | |
126 | ||
127 | XBZRLE_cache_lock(); | |
128 | ||
129 | if (XBZRLE.cache != NULL) { | |
130 | if (pow2floor(new_size) == migrate_xbzrle_cache_size()) { | |
131 | goto out_new_size; | |
132 | } | |
133 | new_cache = cache_init(new_size / TARGET_PAGE_SIZE, | |
134 | TARGET_PAGE_SIZE); | |
135 | if (!new_cache) { | |
136 | error_report("Error creating cache"); | |
137 | ret = -1; | |
138 | goto out; | |
139 | } | |
140 | ||
141 | cache_fini(XBZRLE.cache); | |
142 | XBZRLE.cache = new_cache; | |
143 | } | |
144 | ||
145 | out_new_size: | |
146 | ret = pow2floor(new_size); | |
147 | out: | |
148 | XBZRLE_cache_unlock(); | |
149 | return ret; | |
150 | } | |
151 | ||
ec481c6c JQ |
152 | /* |
153 | * An outstanding page request, on the source, having been received | |
154 | * and queued | |
155 | */ | |
156 | struct RAMSrcPageRequest { | |
157 | RAMBlock *rb; | |
158 | hwaddr offset; | |
159 | hwaddr len; | |
160 | ||
161 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
162 | }; | |
163 | ||
6f37bb8b JQ |
164 | /* State of RAM for migration */ |
165 | struct RAMState { | |
204b88b8 JQ |
166 | /* QEMUFile used for this migration */ |
167 | QEMUFile *f; | |
6f37bb8b JQ |
168 | /* Last block that we have visited searching for dirty pages */ |
169 | RAMBlock *last_seen_block; | |
170 | /* Last block from where we have sent data */ | |
171 | RAMBlock *last_sent_block; | |
269ace29 JQ |
172 | /* Last dirty target page we have sent */ |
173 | ram_addr_t last_page; | |
6f37bb8b JQ |
174 | /* last ram version we have seen */ |
175 | uint32_t last_version; | |
176 | /* We are in the first round */ | |
177 | bool ram_bulk_stage; | |
8d820d6f JQ |
178 | /* How many times we have dirty too many pages */ |
179 | int dirty_rate_high_cnt; | |
5a987738 JQ |
180 | /* How many times we have synchronized the bitmap */ |
181 | uint64_t bitmap_sync_count; | |
f664da80 JQ |
182 | /* these variables are used for bitmap sync */ |
183 | /* last time we did a full bitmap_sync */ | |
184 | int64_t time_last_bitmap_sync; | |
eac74159 | 185 | /* bytes transferred at start_time */ |
c4bdf0cf | 186 | uint64_t bytes_xfer_prev; |
a66cd90c | 187 | /* number of dirty pages since start_time */ |
68908ed6 | 188 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
189 | /* xbzrle misses since the beginning of the period */ |
190 | uint64_t xbzrle_cache_miss_prev; | |
36040d9c JQ |
191 | /* number of iterations at the beginning of period */ |
192 | uint64_t iterations_prev; | |
f7ccd61b JQ |
193 | /* Accounting fields */ |
194 | /* number of zero pages. It used to be pages filled by the same char. */ | |
195 | uint64_t zero_pages; | |
b4d1c6e7 JQ |
196 | /* number of normal transferred pages */ |
197 | uint64_t norm_pages; | |
23b28c3c JQ |
198 | /* Iterations since start */ |
199 | uint64_t iterations; | |
f36ada95 JQ |
200 | /* xbzrle transmitted bytes. Notice that this is with |
201 | * compression, they can't be calculated from the pages */ | |
07ed50a2 | 202 | uint64_t xbzrle_bytes; |
f36ada95 JQ |
203 | /* xbzrle transmmited pages */ |
204 | uint64_t xbzrle_pages; | |
544c36f1 JQ |
205 | /* xbzrle number of cache miss */ |
206 | uint64_t xbzrle_cache_miss; | |
b07016b6 JQ |
207 | /* xbzrle miss rate */ |
208 | double xbzrle_cache_miss_rate; | |
180f61f7 JQ |
209 | /* xbzrle number of overflows */ |
210 | uint64_t xbzrle_overflows; | |
0d8ec885 JQ |
211 | /* number of dirty bits in the bitmap */ |
212 | uint64_t migration_dirty_pages; | |
2f4fde93 JQ |
213 | /* total number of bytes transferred */ |
214 | uint64_t bytes_transferred; | |
47ad8619 JQ |
215 | /* number of dirtied pages in the last second */ |
216 | uint64_t dirty_pages_rate; | |
96506894 JQ |
217 | /* Count of requests incoming from destination */ |
218 | uint64_t postcopy_requests; | |
108cfae0 JQ |
219 | /* protects modification of the bitmap */ |
220 | QemuMutex bitmap_mutex; | |
68a098f3 JQ |
221 | /* The RAMBlock used in the last src_page_requests */ |
222 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
223 | /* Queue of outstanding page requests from the destination */ |
224 | QemuMutex src_page_req_mutex; | |
225 | QSIMPLEQ_HEAD(src_page_requests, RAMSrcPageRequest) src_page_requests; | |
6f37bb8b JQ |
226 | }; |
227 | typedef struct RAMState RAMState; | |
228 | ||
229 | static RAMState ram_state; | |
230 | ||
56e93d26 JQ |
231 | uint64_t dup_mig_pages_transferred(void) |
232 | { | |
f7ccd61b | 233 | return ram_state.zero_pages; |
56e93d26 JQ |
234 | } |
235 | ||
56e93d26 JQ |
236 | uint64_t norm_mig_pages_transferred(void) |
237 | { | |
b4d1c6e7 | 238 | return ram_state.norm_pages; |
56e93d26 JQ |
239 | } |
240 | ||
241 | uint64_t xbzrle_mig_bytes_transferred(void) | |
242 | { | |
07ed50a2 | 243 | return ram_state.xbzrle_bytes; |
56e93d26 JQ |
244 | } |
245 | ||
246 | uint64_t xbzrle_mig_pages_transferred(void) | |
247 | { | |
f36ada95 | 248 | return ram_state.xbzrle_pages; |
56e93d26 JQ |
249 | } |
250 | ||
251 | uint64_t xbzrle_mig_pages_cache_miss(void) | |
252 | { | |
544c36f1 | 253 | return ram_state.xbzrle_cache_miss; |
56e93d26 JQ |
254 | } |
255 | ||
256 | double xbzrle_mig_cache_miss_rate(void) | |
257 | { | |
b07016b6 | 258 | return ram_state.xbzrle_cache_miss_rate; |
56e93d26 JQ |
259 | } |
260 | ||
261 | uint64_t xbzrle_mig_pages_overflow(void) | |
262 | { | |
180f61f7 | 263 | return ram_state.xbzrle_overflows; |
56e93d26 JQ |
264 | } |
265 | ||
9edabd4d | 266 | uint64_t ram_bytes_transferred(void) |
0d8ec885 | 267 | { |
9edabd4d | 268 | return ram_state.bytes_transferred; |
0d8ec885 JQ |
269 | } |
270 | ||
9edabd4d | 271 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 272 | { |
9edabd4d | 273 | return ram_state.migration_dirty_pages * TARGET_PAGE_SIZE; |
2f4fde93 JQ |
274 | } |
275 | ||
42d219d3 JQ |
276 | uint64_t ram_dirty_sync_count(void) |
277 | { | |
278 | return ram_state.bitmap_sync_count; | |
279 | } | |
280 | ||
47ad8619 JQ |
281 | uint64_t ram_dirty_pages_rate(void) |
282 | { | |
283 | return ram_state.dirty_pages_rate; | |
284 | } | |
285 | ||
96506894 JQ |
286 | uint64_t ram_postcopy_requests(void) |
287 | { | |
288 | return ram_state.postcopy_requests; | |
289 | } | |
290 | ||
b8fb8cb7 DDAG |
291 | /* used by the search for pages to send */ |
292 | struct PageSearchStatus { | |
293 | /* Current block being searched */ | |
294 | RAMBlock *block; | |
a935e30f JQ |
295 | /* Current page to search from */ |
296 | unsigned long page; | |
b8fb8cb7 DDAG |
297 | /* Set once we wrap around */ |
298 | bool complete_round; | |
299 | }; | |
300 | typedef struct PageSearchStatus PageSearchStatus; | |
301 | ||
56e93d26 | 302 | struct CompressParam { |
56e93d26 | 303 | bool done; |
90e56fb4 | 304 | bool quit; |
56e93d26 JQ |
305 | QEMUFile *file; |
306 | QemuMutex mutex; | |
307 | QemuCond cond; | |
308 | RAMBlock *block; | |
309 | ram_addr_t offset; | |
310 | }; | |
311 | typedef struct CompressParam CompressParam; | |
312 | ||
313 | struct DecompressParam { | |
73a8912b | 314 | bool done; |
90e56fb4 | 315 | bool quit; |
56e93d26 JQ |
316 | QemuMutex mutex; |
317 | QemuCond cond; | |
318 | void *des; | |
d341d9f3 | 319 | uint8_t *compbuf; |
56e93d26 JQ |
320 | int len; |
321 | }; | |
322 | typedef struct DecompressParam DecompressParam; | |
323 | ||
324 | static CompressParam *comp_param; | |
325 | static QemuThread *compress_threads; | |
326 | /* comp_done_cond is used to wake up the migration thread when | |
327 | * one of the compression threads has finished the compression. | |
328 | * comp_done_lock is used to co-work with comp_done_cond. | |
329 | */ | |
0d9f9a5c LL |
330 | static QemuMutex comp_done_lock; |
331 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
332 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
333 | static const QEMUFileOps empty_ops = { }; | |
334 | ||
56e93d26 JQ |
335 | static DecompressParam *decomp_param; |
336 | static QemuThread *decompress_threads; | |
73a8912b LL |
337 | static QemuMutex decomp_done_lock; |
338 | static QemuCond decomp_done_cond; | |
56e93d26 | 339 | |
a7a9a88f LL |
340 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
341 | ram_addr_t offset); | |
56e93d26 JQ |
342 | |
343 | static void *do_data_compress(void *opaque) | |
344 | { | |
345 | CompressParam *param = opaque; | |
a7a9a88f LL |
346 | RAMBlock *block; |
347 | ram_addr_t offset; | |
56e93d26 | 348 | |
a7a9a88f | 349 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 350 | while (!param->quit) { |
a7a9a88f LL |
351 | if (param->block) { |
352 | block = param->block; | |
353 | offset = param->offset; | |
354 | param->block = NULL; | |
355 | qemu_mutex_unlock(¶m->mutex); | |
356 | ||
357 | do_compress_ram_page(param->file, block, offset); | |
358 | ||
0d9f9a5c | 359 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 360 | param->done = true; |
0d9f9a5c LL |
361 | qemu_cond_signal(&comp_done_cond); |
362 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
363 | |
364 | qemu_mutex_lock(¶m->mutex); | |
365 | } else { | |
56e93d26 JQ |
366 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
367 | } | |
56e93d26 | 368 | } |
a7a9a88f | 369 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
370 | |
371 | return NULL; | |
372 | } | |
373 | ||
374 | static inline void terminate_compression_threads(void) | |
375 | { | |
376 | int idx, thread_count; | |
377 | ||
378 | thread_count = migrate_compress_threads(); | |
3d0684b2 | 379 | |
56e93d26 JQ |
380 | for (idx = 0; idx < thread_count; idx++) { |
381 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 382 | comp_param[idx].quit = true; |
56e93d26 JQ |
383 | qemu_cond_signal(&comp_param[idx].cond); |
384 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
385 | } | |
386 | } | |
387 | ||
388 | void migrate_compress_threads_join(void) | |
389 | { | |
390 | int i, thread_count; | |
391 | ||
392 | if (!migrate_use_compression()) { | |
393 | return; | |
394 | } | |
395 | terminate_compression_threads(); | |
396 | thread_count = migrate_compress_threads(); | |
397 | for (i = 0; i < thread_count; i++) { | |
398 | qemu_thread_join(compress_threads + i); | |
399 | qemu_fclose(comp_param[i].file); | |
400 | qemu_mutex_destroy(&comp_param[i].mutex); | |
401 | qemu_cond_destroy(&comp_param[i].cond); | |
402 | } | |
0d9f9a5c LL |
403 | qemu_mutex_destroy(&comp_done_lock); |
404 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
405 | g_free(compress_threads); |
406 | g_free(comp_param); | |
56e93d26 JQ |
407 | compress_threads = NULL; |
408 | comp_param = NULL; | |
56e93d26 JQ |
409 | } |
410 | ||
411 | void migrate_compress_threads_create(void) | |
412 | { | |
413 | int i, thread_count; | |
414 | ||
415 | if (!migrate_use_compression()) { | |
416 | return; | |
417 | } | |
56e93d26 JQ |
418 | thread_count = migrate_compress_threads(); |
419 | compress_threads = g_new0(QemuThread, thread_count); | |
420 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
421 | qemu_cond_init(&comp_done_cond); |
422 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 423 | for (i = 0; i < thread_count; i++) { |
e110aa91 C |
424 | /* comp_param[i].file is just used as a dummy buffer to save data, |
425 | * set its ops to empty. | |
56e93d26 JQ |
426 | */ |
427 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
428 | comp_param[i].done = true; | |
90e56fb4 | 429 | comp_param[i].quit = false; |
56e93d26 JQ |
430 | qemu_mutex_init(&comp_param[i].mutex); |
431 | qemu_cond_init(&comp_param[i].cond); | |
432 | qemu_thread_create(compress_threads + i, "compress", | |
433 | do_data_compress, comp_param + i, | |
434 | QEMU_THREAD_JOINABLE); | |
435 | } | |
436 | } | |
437 | ||
438 | /** | |
3d0684b2 | 439 | * save_page_header: write page header to wire |
56e93d26 JQ |
440 | * |
441 | * If this is the 1st block, it also writes the block identification | |
442 | * | |
3d0684b2 | 443 | * Returns the number of bytes written |
56e93d26 JQ |
444 | * |
445 | * @f: QEMUFile where to send the data | |
446 | * @block: block that contains the page we want to send | |
447 | * @offset: offset inside the block for the page | |
448 | * in the lower bits, it contains flags | |
449 | */ | |
2bf3aa85 JQ |
450 | static size_t save_page_header(RAMState *rs, QEMUFile *f, RAMBlock *block, |
451 | ram_addr_t offset) | |
56e93d26 | 452 | { |
9f5f380b | 453 | size_t size, len; |
56e93d26 | 454 | |
24795694 JQ |
455 | if (block == rs->last_sent_block) { |
456 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
457 | } | |
2bf3aa85 | 458 | qemu_put_be64(f, offset); |
56e93d26 JQ |
459 | size = 8; |
460 | ||
461 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 462 | len = strlen(block->idstr); |
2bf3aa85 JQ |
463 | qemu_put_byte(f, len); |
464 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 465 | size += 1 + len; |
24795694 | 466 | rs->last_sent_block = block; |
56e93d26 JQ |
467 | } |
468 | return size; | |
469 | } | |
470 | ||
3d0684b2 JQ |
471 | /** |
472 | * mig_throttle_guest_down: throotle down the guest | |
473 | * | |
474 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
475 | * writes. If guest dirty memory rate is reduced below the rate at | |
476 | * which we can transfer pages to the destination then we should be | |
477 | * able to complete migration. Some workloads dirty memory way too | |
478 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
479 | */ |
480 | static void mig_throttle_guest_down(void) | |
481 | { | |
482 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
483 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
484 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
485 | |
486 | /* We have not started throttling yet. Let's start it. */ | |
487 | if (!cpu_throttle_active()) { | |
488 | cpu_throttle_set(pct_initial); | |
489 | } else { | |
490 | /* Throttling already on, just increase the rate */ | |
491 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
492 | } | |
493 | } | |
494 | ||
3d0684b2 JQ |
495 | /** |
496 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
497 | * | |
6f37bb8b | 498 | * @rs: current RAM state |
3d0684b2 JQ |
499 | * @current_addr: address for the zero page |
500 | * | |
501 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
502 | * The important thing is that a stale (not-yet-0'd) page be replaced |
503 | * by the new data. | |
504 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 505 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 506 | */ |
6f37bb8b | 507 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 508 | { |
6f37bb8b | 509 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
510 | return; |
511 | } | |
512 | ||
513 | /* We don't care if this fails to allocate a new cache page | |
514 | * as long as it updated an old one */ | |
c00e0928 | 515 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
5a987738 | 516 | rs->bitmap_sync_count); |
56e93d26 JQ |
517 | } |
518 | ||
519 | #define ENCODING_FLAG_XBZRLE 0x1 | |
520 | ||
521 | /** | |
522 | * save_xbzrle_page: compress and send current page | |
523 | * | |
524 | * Returns: 1 means that we wrote the page | |
525 | * 0 means that page is identical to the one already sent | |
526 | * -1 means that xbzrle would be longer than normal | |
527 | * | |
5a987738 | 528 | * @rs: current RAM state |
3d0684b2 JQ |
529 | * @current_data: pointer to the address of the page contents |
530 | * @current_addr: addr of the page | |
56e93d26 JQ |
531 | * @block: block that contains the page we want to send |
532 | * @offset: offset inside the block for the page | |
533 | * @last_stage: if we are at the completion stage | |
56e93d26 | 534 | */ |
204b88b8 | 535 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 536 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 537 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
538 | { |
539 | int encoded_len = 0, bytes_xbzrle; | |
540 | uint8_t *prev_cached_page; | |
541 | ||
5a987738 | 542 | if (!cache_is_cached(XBZRLE.cache, current_addr, rs->bitmap_sync_count)) { |
544c36f1 | 543 | rs->xbzrle_cache_miss++; |
56e93d26 JQ |
544 | if (!last_stage) { |
545 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
5a987738 | 546 | rs->bitmap_sync_count) == -1) { |
56e93d26 JQ |
547 | return -1; |
548 | } else { | |
549 | /* update *current_data when the page has been | |
550 | inserted into cache */ | |
551 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
552 | } | |
553 | } | |
554 | return -1; | |
555 | } | |
556 | ||
557 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
558 | ||
559 | /* save current buffer into memory */ | |
560 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
561 | ||
562 | /* XBZRLE encoding (if there is no overflow) */ | |
563 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
564 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
565 | TARGET_PAGE_SIZE); | |
566 | if (encoded_len == 0) { | |
55c4446b | 567 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
568 | return 0; |
569 | } else if (encoded_len == -1) { | |
55c4446b | 570 | trace_save_xbzrle_page_overflow(); |
180f61f7 | 571 | rs->xbzrle_overflows++; |
56e93d26 JQ |
572 | /* update data in the cache */ |
573 | if (!last_stage) { | |
574 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
575 | *current_data = prev_cached_page; | |
576 | } | |
577 | return -1; | |
578 | } | |
579 | ||
580 | /* we need to update the data in the cache, in order to get the same data */ | |
581 | if (!last_stage) { | |
582 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
583 | } | |
584 | ||
585 | /* Send XBZRLE based compressed page */ | |
2bf3aa85 | 586 | bytes_xbzrle = save_page_header(rs, rs->f, block, |
204b88b8 JQ |
587 | offset | RAM_SAVE_FLAG_XBZRLE); |
588 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
589 | qemu_put_be16(rs->f, encoded_len); | |
590 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 591 | bytes_xbzrle += encoded_len + 1 + 2; |
f36ada95 | 592 | rs->xbzrle_pages++; |
07ed50a2 | 593 | rs->xbzrle_bytes += bytes_xbzrle; |
072c2511 | 594 | rs->bytes_transferred += bytes_xbzrle; |
56e93d26 JQ |
595 | |
596 | return 1; | |
597 | } | |
598 | ||
3d0684b2 JQ |
599 | /** |
600 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 601 | * |
3d0684b2 JQ |
602 | * Called with rcu_read_lock() to protect migration_bitmap |
603 | * | |
604 | * Returns the byte offset within memory region of the start of a dirty page | |
605 | * | |
6f37bb8b | 606 | * @rs: current RAM state |
3d0684b2 | 607 | * @rb: RAMBlock where to search for dirty pages |
a935e30f | 608 | * @start: page where we start the search |
f3f491fc | 609 | */ |
56e93d26 | 610 | static inline |
a935e30f | 611 | unsigned long migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
f20e2865 | 612 | unsigned long start) |
56e93d26 | 613 | { |
6b6712ef JQ |
614 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
615 | unsigned long *bitmap = rb->bmap; | |
56e93d26 JQ |
616 | unsigned long next; |
617 | ||
6b6712ef JQ |
618 | if (rs->ram_bulk_stage && start > 0) { |
619 | next = start + 1; | |
56e93d26 | 620 | } else { |
6b6712ef | 621 | next = find_next_bit(bitmap, size, start); |
56e93d26 JQ |
622 | } |
623 | ||
6b6712ef | 624 | return next; |
56e93d26 JQ |
625 | } |
626 | ||
06b10688 | 627 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
628 | RAMBlock *rb, |
629 | unsigned long page) | |
a82d593b DDAG |
630 | { |
631 | bool ret; | |
a82d593b | 632 | |
6b6712ef | 633 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b DDAG |
634 | |
635 | if (ret) { | |
0d8ec885 | 636 | rs->migration_dirty_pages--; |
a82d593b DDAG |
637 | } |
638 | return ret; | |
639 | } | |
640 | ||
15440dd5 JQ |
641 | static void migration_bitmap_sync_range(RAMState *rs, RAMBlock *rb, |
642 | ram_addr_t start, ram_addr_t length) | |
56e93d26 | 643 | { |
0d8ec885 | 644 | rs->migration_dirty_pages += |
6b6712ef | 645 | cpu_physical_memory_sync_dirty_bitmap(rb, start, length, |
0d8ec885 | 646 | &rs->num_dirty_pages_period); |
56e93d26 JQ |
647 | } |
648 | ||
3d0684b2 JQ |
649 | /** |
650 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
651 | * | |
652 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
653 | * | |
654 | * For VMs with just normal pages this is equivalent to the host page | |
655 | * size. If it's got some huge pages then it's the OR of all the | |
656 | * different page sizes. | |
e8ca1db2 DDAG |
657 | */ |
658 | uint64_t ram_pagesize_summary(void) | |
659 | { | |
660 | RAMBlock *block; | |
661 | uint64_t summary = 0; | |
662 | ||
99e15582 | 663 | RAMBLOCK_FOREACH(block) { |
e8ca1db2 DDAG |
664 | summary |= block->page_size; |
665 | } | |
666 | ||
667 | return summary; | |
668 | } | |
669 | ||
8d820d6f | 670 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
671 | { |
672 | RAMBlock *block; | |
56e93d26 | 673 | int64_t end_time; |
c4bdf0cf | 674 | uint64_t bytes_xfer_now; |
56e93d26 | 675 | |
5a987738 | 676 | rs->bitmap_sync_count++; |
56e93d26 | 677 | |
f664da80 JQ |
678 | if (!rs->time_last_bitmap_sync) { |
679 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
680 | } |
681 | ||
682 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 683 | memory_global_dirty_log_sync(); |
56e93d26 | 684 | |
108cfae0 | 685 | qemu_mutex_lock(&rs->bitmap_mutex); |
56e93d26 | 686 | rcu_read_lock(); |
99e15582 | 687 | RAMBLOCK_FOREACH(block) { |
15440dd5 | 688 | migration_bitmap_sync_range(rs, block, 0, block->used_length); |
56e93d26 JQ |
689 | } |
690 | rcu_read_unlock(); | |
108cfae0 | 691 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 692 | |
a66cd90c | 693 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 694 | |
56e93d26 JQ |
695 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
696 | ||
697 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 698 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
d693c6f1 FF |
699 | /* calculate period counters */ |
700 | rs->dirty_pages_rate = rs->num_dirty_pages_period * 1000 | |
701 | / (end_time - rs->time_last_bitmap_sync); | |
d2a4d85a | 702 | bytes_xfer_now = ram_bytes_transferred(); |
d693c6f1 | 703 | |
56e93d26 JQ |
704 | if (migrate_auto_converge()) { |
705 | /* The following detection logic can be refined later. For now: | |
706 | Check to see if the dirtied bytes is 50% more than the approx. | |
707 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
708 | were in this routine. If that happens twice, start or increase |
709 | throttling */ | |
070afca2 | 710 | |
d693c6f1 | 711 | if ((rs->num_dirty_pages_period * TARGET_PAGE_SIZE > |
eac74159 | 712 | (bytes_xfer_now - rs->bytes_xfer_prev) / 2) && |
b4a3c64b | 713 | (++rs->dirty_rate_high_cnt >= 2)) { |
56e93d26 | 714 | trace_migration_throttle(); |
8d820d6f | 715 | rs->dirty_rate_high_cnt = 0; |
070afca2 | 716 | mig_throttle_guest_down(); |
d693c6f1 | 717 | } |
56e93d26 | 718 | } |
070afca2 | 719 | |
56e93d26 | 720 | if (migrate_use_xbzrle()) { |
23b28c3c | 721 | if (rs->iterations_prev != rs->iterations) { |
b07016b6 | 722 | rs->xbzrle_cache_miss_rate = |
544c36f1 | 723 | (double)(rs->xbzrle_cache_miss - |
b5833fde | 724 | rs->xbzrle_cache_miss_prev) / |
23b28c3c | 725 | (rs->iterations - rs->iterations_prev); |
56e93d26 | 726 | } |
23b28c3c | 727 | rs->iterations_prev = rs->iterations; |
544c36f1 | 728 | rs->xbzrle_cache_miss_prev = rs->xbzrle_cache_miss; |
56e93d26 | 729 | } |
d693c6f1 FF |
730 | |
731 | /* reset period counters */ | |
f664da80 | 732 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 733 | rs->num_dirty_pages_period = 0; |
d2a4d85a | 734 | rs->bytes_xfer_prev = bytes_xfer_now; |
56e93d26 | 735 | } |
4addcd4f | 736 | if (migrate_use_events()) { |
5a987738 | 737 | qapi_event_send_migration_pass(rs->bitmap_sync_count, NULL); |
4addcd4f | 738 | } |
56e93d26 JQ |
739 | } |
740 | ||
741 | /** | |
3d0684b2 | 742 | * save_zero_page: send the zero page to the stream |
56e93d26 | 743 | * |
3d0684b2 | 744 | * Returns the number of pages written. |
56e93d26 | 745 | * |
f7ccd61b | 746 | * @rs: current RAM state |
56e93d26 JQ |
747 | * @block: block that contains the page we want to send |
748 | * @offset: offset inside the block for the page | |
749 | * @p: pointer to the page | |
56e93d26 | 750 | */ |
ce25d337 JQ |
751 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, |
752 | uint8_t *p) | |
56e93d26 JQ |
753 | { |
754 | int pages = -1; | |
755 | ||
756 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
f7ccd61b | 757 | rs->zero_pages++; |
072c2511 | 758 | rs->bytes_transferred += |
bb890ed5 | 759 | save_page_header(rs, rs->f, block, offset | RAM_SAVE_FLAG_ZERO); |
ce25d337 | 760 | qemu_put_byte(rs->f, 0); |
072c2511 | 761 | rs->bytes_transferred += 1; |
56e93d26 JQ |
762 | pages = 1; |
763 | } | |
764 | ||
765 | return pages; | |
766 | } | |
767 | ||
5727309d | 768 | static void ram_release_pages(const char *rbname, uint64_t offset, int pages) |
53f09a10 | 769 | { |
5727309d | 770 | if (!migrate_release_ram() || !migration_in_postcopy()) { |
53f09a10 PB |
771 | return; |
772 | } | |
773 | ||
aaa2064c | 774 | ram_discard_range(rbname, offset, pages << TARGET_PAGE_BITS); |
53f09a10 PB |
775 | } |
776 | ||
56e93d26 | 777 | /** |
3d0684b2 | 778 | * ram_save_page: send the given page to the stream |
56e93d26 | 779 | * |
3d0684b2 | 780 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
781 | * < 0 - error |
782 | * >=0 - Number of pages written - this might legally be 0 | |
783 | * if xbzrle noticed the page was the same. | |
56e93d26 | 784 | * |
6f37bb8b | 785 | * @rs: current RAM state |
56e93d26 JQ |
786 | * @block: block that contains the page we want to send |
787 | * @offset: offset inside the block for the page | |
788 | * @last_stage: if we are at the completion stage | |
56e93d26 | 789 | */ |
a0a8aa14 | 790 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
791 | { |
792 | int pages = -1; | |
793 | uint64_t bytes_xmit; | |
794 | ram_addr_t current_addr; | |
56e93d26 JQ |
795 | uint8_t *p; |
796 | int ret; | |
797 | bool send_async = true; | |
a08f6890 | 798 | RAMBlock *block = pss->block; |
a935e30f | 799 | ram_addr_t offset = pss->page << TARGET_PAGE_BITS; |
56e93d26 | 800 | |
2f68e399 | 801 | p = block->host + offset; |
1db9d8e5 | 802 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 JQ |
803 | |
804 | /* In doubt sent page as normal */ | |
805 | bytes_xmit = 0; | |
ce25d337 | 806 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
807 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
808 | if (bytes_xmit) { | |
072c2511 | 809 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
810 | pages = 1; |
811 | } | |
812 | ||
813 | XBZRLE_cache_lock(); | |
814 | ||
815 | current_addr = block->offset + offset; | |
816 | ||
56e93d26 JQ |
817 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
818 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
819 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 820 | rs->norm_pages++; |
56e93d26 | 821 | } else if (bytes_xmit == 0) { |
f7ccd61b | 822 | rs->zero_pages++; |
56e93d26 JQ |
823 | } |
824 | } | |
825 | } else { | |
ce25d337 | 826 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 JQ |
827 | if (pages > 0) { |
828 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
829 | * page would be stale | |
830 | */ | |
6f37bb8b | 831 | xbzrle_cache_zero_page(rs, current_addr); |
a935e30f | 832 | ram_release_pages(block->idstr, offset, pages); |
6f37bb8b | 833 | } else if (!rs->ram_bulk_stage && |
5727309d | 834 | !migration_in_postcopy() && migrate_use_xbzrle()) { |
204b88b8 | 835 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
072c2511 | 836 | offset, last_stage); |
56e93d26 JQ |
837 | if (!last_stage) { |
838 | /* Can't send this cached data async, since the cache page | |
839 | * might get updated before it gets to the wire | |
840 | */ | |
841 | send_async = false; | |
842 | } | |
843 | } | |
844 | } | |
845 | ||
846 | /* XBZRLE overflow or normal page */ | |
847 | if (pages == -1) { | |
2bf3aa85 | 848 | rs->bytes_transferred += save_page_header(rs, rs->f, block, |
24795694 | 849 | offset | RAM_SAVE_FLAG_PAGE); |
56e93d26 | 850 | if (send_async) { |
ce25d337 | 851 | qemu_put_buffer_async(rs->f, p, TARGET_PAGE_SIZE, |
53f09a10 | 852 | migrate_release_ram() & |
5727309d | 853 | migration_in_postcopy()); |
56e93d26 | 854 | } else { |
ce25d337 | 855 | qemu_put_buffer(rs->f, p, TARGET_PAGE_SIZE); |
56e93d26 | 856 | } |
072c2511 | 857 | rs->bytes_transferred += TARGET_PAGE_SIZE; |
56e93d26 | 858 | pages = 1; |
b4d1c6e7 | 859 | rs->norm_pages++; |
56e93d26 JQ |
860 | } |
861 | ||
862 | XBZRLE_cache_unlock(); | |
863 | ||
864 | return pages; | |
865 | } | |
866 | ||
a7a9a88f LL |
867 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
868 | ram_addr_t offset) | |
56e93d26 | 869 | { |
24795694 | 870 | RAMState *rs = &ram_state; |
56e93d26 | 871 | int bytes_sent, blen; |
a7a9a88f | 872 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 | 873 | |
2bf3aa85 | 874 | bytes_sent = save_page_header(rs, f, block, offset | |
56e93d26 | 875 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
a7a9a88f | 876 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, |
56e93d26 | 877 | migrate_compress_level()); |
b3be2896 LL |
878 | if (blen < 0) { |
879 | bytes_sent = 0; | |
880 | qemu_file_set_error(migrate_get_current()->to_dst_file, blen); | |
881 | error_report("compressed data failed!"); | |
882 | } else { | |
883 | bytes_sent += blen; | |
5727309d | 884 | ram_release_pages(block->idstr, offset & TARGET_PAGE_MASK, 1); |
b3be2896 | 885 | } |
56e93d26 JQ |
886 | |
887 | return bytes_sent; | |
888 | } | |
889 | ||
ce25d337 | 890 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
891 | { |
892 | int idx, len, thread_count; | |
893 | ||
894 | if (!migrate_use_compression()) { | |
895 | return; | |
896 | } | |
897 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 898 | |
0d9f9a5c | 899 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 900 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 901 | while (!comp_param[idx].done) { |
0d9f9a5c | 902 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 903 | } |
a7a9a88f | 904 | } |
0d9f9a5c | 905 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
906 | |
907 | for (idx = 0; idx < thread_count; idx++) { | |
908 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 909 | if (!comp_param[idx].quit) { |
ce25d337 | 910 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
2f4fde93 | 911 | rs->bytes_transferred += len; |
56e93d26 | 912 | } |
a7a9a88f | 913 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
914 | } |
915 | } | |
916 | ||
917 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
918 | ram_addr_t offset) | |
919 | { | |
920 | param->block = block; | |
921 | param->offset = offset; | |
922 | } | |
923 | ||
ce25d337 JQ |
924 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
925 | ram_addr_t offset) | |
56e93d26 JQ |
926 | { |
927 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
928 | ||
929 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 930 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 JQ |
931 | while (true) { |
932 | for (idx = 0; idx < thread_count; idx++) { | |
933 | if (comp_param[idx].done) { | |
a7a9a88f | 934 | comp_param[idx].done = false; |
ce25d337 | 935 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
a7a9a88f | 936 | qemu_mutex_lock(&comp_param[idx].mutex); |
56e93d26 | 937 | set_compress_params(&comp_param[idx], block, offset); |
a7a9a88f LL |
938 | qemu_cond_signal(&comp_param[idx].cond); |
939 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
56e93d26 | 940 | pages = 1; |
b4d1c6e7 | 941 | rs->norm_pages++; |
072c2511 | 942 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
943 | break; |
944 | } | |
945 | } | |
946 | if (pages > 0) { | |
947 | break; | |
948 | } else { | |
0d9f9a5c | 949 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 JQ |
950 | } |
951 | } | |
0d9f9a5c | 952 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
953 | |
954 | return pages; | |
955 | } | |
956 | ||
957 | /** | |
958 | * ram_save_compressed_page: compress the given page and send it to the stream | |
959 | * | |
3d0684b2 | 960 | * Returns the number of pages written. |
56e93d26 | 961 | * |
6f37bb8b | 962 | * @rs: current RAM state |
56e93d26 JQ |
963 | * @block: block that contains the page we want to send |
964 | * @offset: offset inside the block for the page | |
965 | * @last_stage: if we are at the completion stage | |
56e93d26 | 966 | */ |
a0a8aa14 JQ |
967 | static int ram_save_compressed_page(RAMState *rs, PageSearchStatus *pss, |
968 | bool last_stage) | |
56e93d26 JQ |
969 | { |
970 | int pages = -1; | |
fc50438e | 971 | uint64_t bytes_xmit = 0; |
56e93d26 | 972 | uint8_t *p; |
fc50438e | 973 | int ret, blen; |
a08f6890 | 974 | RAMBlock *block = pss->block; |
a935e30f | 975 | ram_addr_t offset = pss->page << TARGET_PAGE_BITS; |
56e93d26 | 976 | |
2f68e399 | 977 | p = block->host + offset; |
56e93d26 | 978 | |
ce25d337 | 979 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
980 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
981 | if (bytes_xmit) { | |
072c2511 | 982 | rs->bytes_transferred += bytes_xmit; |
56e93d26 JQ |
983 | pages = 1; |
984 | } | |
56e93d26 JQ |
985 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
986 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
987 | if (bytes_xmit > 0) { | |
b4d1c6e7 | 988 | rs->norm_pages++; |
56e93d26 | 989 | } else if (bytes_xmit == 0) { |
f7ccd61b | 990 | rs->zero_pages++; |
56e93d26 JQ |
991 | } |
992 | } | |
993 | } else { | |
994 | /* When starting the process of a new block, the first page of | |
995 | * the block should be sent out before other pages in the same | |
996 | * block, and all the pages in last block should have been sent | |
997 | * out, keeping this order is important, because the 'cont' flag | |
998 | * is used to avoid resending the block name. | |
999 | */ | |
6f37bb8b | 1000 | if (block != rs->last_sent_block) { |
ce25d337 JQ |
1001 | flush_compressed_data(rs); |
1002 | pages = save_zero_page(rs, block, offset, p); | |
56e93d26 | 1003 | if (pages == -1) { |
fc50438e | 1004 | /* Make sure the first page is sent out before other pages */ |
2bf3aa85 | 1005 | bytes_xmit = save_page_header(rs, rs->f, block, offset | |
fc50438e | 1006 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
ce25d337 | 1007 | blen = qemu_put_compression_data(rs->f, p, TARGET_PAGE_SIZE, |
fc50438e LL |
1008 | migrate_compress_level()); |
1009 | if (blen > 0) { | |
072c2511 | 1010 | rs->bytes_transferred += bytes_xmit + blen; |
b4d1c6e7 | 1011 | rs->norm_pages++; |
b3be2896 | 1012 | pages = 1; |
fc50438e | 1013 | } else { |
ce25d337 | 1014 | qemu_file_set_error(rs->f, blen); |
fc50438e | 1015 | error_report("compressed data failed!"); |
b3be2896 | 1016 | } |
56e93d26 | 1017 | } |
53f09a10 | 1018 | if (pages > 0) { |
a935e30f | 1019 | ram_release_pages(block->idstr, offset, pages); |
53f09a10 | 1020 | } |
56e93d26 | 1021 | } else { |
ce25d337 | 1022 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 | 1023 | if (pages == -1) { |
ce25d337 | 1024 | pages = compress_page_with_multi_thread(rs, block, offset); |
53f09a10 | 1025 | } else { |
a935e30f | 1026 | ram_release_pages(block->idstr, offset, pages); |
56e93d26 JQ |
1027 | } |
1028 | } | |
1029 | } | |
1030 | ||
1031 | return pages; | |
1032 | } | |
1033 | ||
3d0684b2 JQ |
1034 | /** |
1035 | * find_dirty_block: find the next dirty page and update any state | |
1036 | * associated with the search process. | |
b9e60928 | 1037 | * |
3d0684b2 | 1038 | * Returns if a page is found |
b9e60928 | 1039 | * |
6f37bb8b | 1040 | * @rs: current RAM state |
3d0684b2 JQ |
1041 | * @pss: data about the state of the current dirty page scan |
1042 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1043 | */ |
f20e2865 | 1044 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1045 | { |
f20e2865 | 1046 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
6f37bb8b | 1047 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1048 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1049 | /* |
1050 | * We've been once around the RAM and haven't found anything. | |
1051 | * Give up. | |
1052 | */ | |
1053 | *again = false; | |
1054 | return false; | |
1055 | } | |
a935e30f | 1056 | if ((pss->page << TARGET_PAGE_BITS) >= pss->block->used_length) { |
b9e60928 | 1057 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1058 | pss->page = 0; |
b9e60928 DDAG |
1059 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1060 | if (!pss->block) { | |
1061 | /* Hit the end of the list */ | |
1062 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1063 | /* Flag that we've looped */ | |
1064 | pss->complete_round = true; | |
6f37bb8b | 1065 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1066 | if (migrate_use_xbzrle()) { |
1067 | /* If xbzrle is on, stop using the data compression at this | |
1068 | * point. In theory, xbzrle can do better than compression. | |
1069 | */ | |
ce25d337 | 1070 | flush_compressed_data(rs); |
b9e60928 DDAG |
1071 | } |
1072 | } | |
1073 | /* Didn't find anything this time, but try again on the new block */ | |
1074 | *again = true; | |
1075 | return false; | |
1076 | } else { | |
1077 | /* Can go around again, but... */ | |
1078 | *again = true; | |
1079 | /* We've found something so probably don't need to */ | |
1080 | return true; | |
1081 | } | |
1082 | } | |
1083 | ||
3d0684b2 JQ |
1084 | /** |
1085 | * unqueue_page: gets a page of the queue | |
1086 | * | |
a82d593b | 1087 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1088 | * |
3d0684b2 JQ |
1089 | * Returns the block of the page (or NULL if none available) |
1090 | * | |
ec481c6c | 1091 | * @rs: current RAM state |
3d0684b2 | 1092 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1093 | */ |
f20e2865 | 1094 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b DDAG |
1095 | { |
1096 | RAMBlock *block = NULL; | |
1097 | ||
ec481c6c JQ |
1098 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1099 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
1100 | struct RAMSrcPageRequest *entry = | |
1101 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1102 | block = entry->rb; |
1103 | *offset = entry->offset; | |
a82d593b DDAG |
1104 | |
1105 | if (entry->len > TARGET_PAGE_SIZE) { | |
1106 | entry->len -= TARGET_PAGE_SIZE; | |
1107 | entry->offset += TARGET_PAGE_SIZE; | |
1108 | } else { | |
1109 | memory_region_unref(block->mr); | |
ec481c6c | 1110 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b DDAG |
1111 | g_free(entry); |
1112 | } | |
1113 | } | |
ec481c6c | 1114 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
a82d593b DDAG |
1115 | |
1116 | return block; | |
1117 | } | |
1118 | ||
3d0684b2 JQ |
1119 | /** |
1120 | * get_queued_page: unqueue a page from the postocpy requests | |
1121 | * | |
1122 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1123 | * |
3d0684b2 | 1124 | * Returns if a queued page is found |
a82d593b | 1125 | * |
6f37bb8b | 1126 | * @rs: current RAM state |
3d0684b2 | 1127 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1128 | */ |
f20e2865 | 1129 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1130 | { |
1131 | RAMBlock *block; | |
1132 | ram_addr_t offset; | |
1133 | bool dirty; | |
1134 | ||
1135 | do { | |
f20e2865 | 1136 | block = unqueue_page(rs, &offset); |
a82d593b DDAG |
1137 | /* |
1138 | * We're sending this page, and since it's postcopy nothing else | |
1139 | * will dirty it, and we must make sure it doesn't get sent again | |
1140 | * even if this queue request was received after the background | |
1141 | * search already sent it. | |
1142 | */ | |
1143 | if (block) { | |
f20e2865 JQ |
1144 | unsigned long page; |
1145 | ||
6b6712ef JQ |
1146 | page = offset >> TARGET_PAGE_BITS; |
1147 | dirty = test_bit(page, block->bmap); | |
a82d593b | 1148 | if (!dirty) { |
06b10688 | 1149 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, |
6b6712ef | 1150 | page, test_bit(page, block->unsentmap)); |
a82d593b | 1151 | } else { |
f20e2865 | 1152 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); |
a82d593b DDAG |
1153 | } |
1154 | } | |
1155 | ||
1156 | } while (block && !dirty); | |
1157 | ||
1158 | if (block) { | |
1159 | /* | |
1160 | * As soon as we start servicing pages out of order, then we have | |
1161 | * to kill the bulk stage, since the bulk stage assumes | |
1162 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1163 | * dirty, that's no longer true. | |
1164 | */ | |
6f37bb8b | 1165 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1166 | |
1167 | /* | |
1168 | * We want the background search to continue from the queued page | |
1169 | * since the guest is likely to want other pages near to the page | |
1170 | * it just requested. | |
1171 | */ | |
1172 | pss->block = block; | |
a935e30f | 1173 | pss->page = offset >> TARGET_PAGE_BITS; |
a82d593b DDAG |
1174 | } |
1175 | ||
1176 | return !!block; | |
1177 | } | |
1178 | ||
6c595cde | 1179 | /** |
5e58f968 JQ |
1180 | * migration_page_queue_free: drop any remaining pages in the ram |
1181 | * request queue | |
6c595cde | 1182 | * |
3d0684b2 JQ |
1183 | * It should be empty at the end anyway, but in error cases there may |
1184 | * be some left. in case that there is any page left, we drop it. | |
1185 | * | |
6c595cde | 1186 | */ |
83c13382 | 1187 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 1188 | { |
ec481c6c | 1189 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
1190 | /* This queue generally should be empty - but in the case of a failed |
1191 | * migration might have some droppings in. | |
1192 | */ | |
1193 | rcu_read_lock(); | |
ec481c6c | 1194 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1195 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1196 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1197 | g_free(mspr); |
1198 | } | |
1199 | rcu_read_unlock(); | |
1200 | } | |
1201 | ||
1202 | /** | |
3d0684b2 JQ |
1203 | * ram_save_queue_pages: queue the page for transmission |
1204 | * | |
1205 | * A request from postcopy destination for example. | |
1206 | * | |
1207 | * Returns zero on success or negative on error | |
1208 | * | |
3d0684b2 JQ |
1209 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1210 | * same that last one. | |
1211 | * @start: starting address from the start of the RAMBlock | |
1212 | * @len: length (in bytes) to send | |
6c595cde | 1213 | */ |
96506894 | 1214 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1215 | { |
1216 | RAMBlock *ramblock; | |
68a098f3 | 1217 | RAMState *rs = &ram_state; |
6c595cde | 1218 | |
96506894 | 1219 | rs->postcopy_requests++; |
6c595cde DDAG |
1220 | rcu_read_lock(); |
1221 | if (!rbname) { | |
1222 | /* Reuse last RAMBlock */ | |
68a098f3 | 1223 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1224 | |
1225 | if (!ramblock) { | |
1226 | /* | |
1227 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1228 | * it's the 1st request. | |
1229 | */ | |
1230 | error_report("ram_save_queue_pages no previous block"); | |
1231 | goto err; | |
1232 | } | |
1233 | } else { | |
1234 | ramblock = qemu_ram_block_by_name(rbname); | |
1235 | ||
1236 | if (!ramblock) { | |
1237 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1238 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1239 | goto err; | |
1240 | } | |
68a098f3 | 1241 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1242 | } |
1243 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1244 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1245 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1246 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1247 | __func__, start, len, ramblock->used_length); |
1248 | goto err; | |
1249 | } | |
1250 | ||
ec481c6c JQ |
1251 | struct RAMSrcPageRequest *new_entry = |
1252 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
1253 | new_entry->rb = ramblock; |
1254 | new_entry->offset = start; | |
1255 | new_entry->len = len; | |
1256 | ||
1257 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
1258 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1259 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
1260 | qemu_mutex_unlock(&rs->src_page_req_mutex); | |
6c595cde DDAG |
1261 | rcu_read_unlock(); |
1262 | ||
1263 | return 0; | |
1264 | ||
1265 | err: | |
1266 | rcu_read_unlock(); | |
1267 | return -1; | |
1268 | } | |
1269 | ||
a82d593b | 1270 | /** |
3d0684b2 | 1271 | * ram_save_target_page: save one target page |
a82d593b | 1272 | * |
3d0684b2 | 1273 | * Returns the number of pages written |
a82d593b | 1274 | * |
6f37bb8b | 1275 | * @rs: current RAM state |
3d0684b2 | 1276 | * @ms: current migration state |
3d0684b2 | 1277 | * @pss: data about the page we want to send |
a82d593b | 1278 | * @last_stage: if we are at the completion stage |
a82d593b | 1279 | */ |
a0a8aa14 | 1280 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1281 | bool last_stage) |
a82d593b DDAG |
1282 | { |
1283 | int res = 0; | |
1284 | ||
1285 | /* Check the pages is dirty and if it is send it */ | |
f20e2865 | 1286 | if (migration_bitmap_clear_dirty(rs, pss->block, pss->page)) { |
6d358d94 JQ |
1287 | /* |
1288 | * If xbzrle is on, stop using the data compression after first | |
1289 | * round of migration even if compression is enabled. In theory, | |
1290 | * xbzrle can do better than compression. | |
1291 | */ | |
6b6712ef JQ |
1292 | if (migrate_use_compression() && |
1293 | (rs->ram_bulk_stage || !migrate_use_xbzrle())) { | |
a0a8aa14 | 1294 | res = ram_save_compressed_page(rs, pss, last_stage); |
a82d593b | 1295 | } else { |
a0a8aa14 | 1296 | res = ram_save_page(rs, pss, last_stage); |
a82d593b DDAG |
1297 | } |
1298 | ||
1299 | if (res < 0) { | |
1300 | return res; | |
1301 | } | |
6b6712ef JQ |
1302 | if (pss->block->unsentmap) { |
1303 | clear_bit(pss->page, pss->block->unsentmap); | |
a82d593b DDAG |
1304 | } |
1305 | } | |
1306 | ||
1307 | return res; | |
1308 | } | |
1309 | ||
1310 | /** | |
3d0684b2 | 1311 | * ram_save_host_page: save a whole host page |
a82d593b | 1312 | * |
3d0684b2 JQ |
1313 | * Starting at *offset send pages up to the end of the current host |
1314 | * page. It's valid for the initial offset to point into the middle of | |
1315 | * a host page in which case the remainder of the hostpage is sent. | |
1316 | * Only dirty target pages are sent. Note that the host page size may | |
1317 | * be a huge page for this block. | |
1eb3fc0a DDAG |
1318 | * The saving stops at the boundary of the used_length of the block |
1319 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 1320 | * |
3d0684b2 JQ |
1321 | * Returns the number of pages written or negative on error |
1322 | * | |
6f37bb8b | 1323 | * @rs: current RAM state |
3d0684b2 | 1324 | * @ms: current migration state |
3d0684b2 | 1325 | * @pss: data about the page we want to send |
a82d593b | 1326 | * @last_stage: if we are at the completion stage |
a82d593b | 1327 | */ |
a0a8aa14 | 1328 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1329 | bool last_stage) |
a82d593b DDAG |
1330 | { |
1331 | int tmppages, pages = 0; | |
a935e30f JQ |
1332 | size_t pagesize_bits = |
1333 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
4c011c37 | 1334 | |
a82d593b | 1335 | do { |
f20e2865 | 1336 | tmppages = ram_save_target_page(rs, pss, last_stage); |
a82d593b DDAG |
1337 | if (tmppages < 0) { |
1338 | return tmppages; | |
1339 | } | |
1340 | ||
1341 | pages += tmppages; | |
a935e30f | 1342 | pss->page++; |
1eb3fc0a DDAG |
1343 | } while ((pss->page & (pagesize_bits - 1)) && |
1344 | offset_in_ramblock(pss->block, pss->page << TARGET_PAGE_BITS)); | |
a82d593b DDAG |
1345 | |
1346 | /* The offset we leave with is the last one we looked at */ | |
a935e30f | 1347 | pss->page--; |
a82d593b DDAG |
1348 | return pages; |
1349 | } | |
6c595cde | 1350 | |
56e93d26 | 1351 | /** |
3d0684b2 | 1352 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1353 | * |
1354 | * Called within an RCU critical section. | |
1355 | * | |
3d0684b2 | 1356 | * Returns the number of pages written where zero means no dirty pages |
56e93d26 | 1357 | * |
6f37bb8b | 1358 | * @rs: current RAM state |
56e93d26 | 1359 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1360 | * |
1361 | * On systems where host-page-size > target-page-size it will send all the | |
1362 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1363 | */ |
1364 | ||
ce25d337 | 1365 | static int ram_find_and_save_block(RAMState *rs, bool last_stage) |
56e93d26 | 1366 | { |
b8fb8cb7 | 1367 | PageSearchStatus pss; |
56e93d26 | 1368 | int pages = 0; |
b9e60928 | 1369 | bool again, found; |
56e93d26 | 1370 | |
0827b9e9 AA |
1371 | /* No dirty page as there is zero RAM */ |
1372 | if (!ram_bytes_total()) { | |
1373 | return pages; | |
1374 | } | |
1375 | ||
6f37bb8b | 1376 | pss.block = rs->last_seen_block; |
a935e30f | 1377 | pss.page = rs->last_page; |
b8fb8cb7 DDAG |
1378 | pss.complete_round = false; |
1379 | ||
1380 | if (!pss.block) { | |
1381 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1382 | } | |
56e93d26 | 1383 | |
b9e60928 | 1384 | do { |
a82d593b | 1385 | again = true; |
f20e2865 | 1386 | found = get_queued_page(rs, &pss); |
b9e60928 | 1387 | |
a82d593b DDAG |
1388 | if (!found) { |
1389 | /* priority queue empty, so just search for something dirty */ | |
f20e2865 | 1390 | found = find_dirty_block(rs, &pss, &again); |
a82d593b | 1391 | } |
f3f491fc | 1392 | |
a82d593b | 1393 | if (found) { |
f20e2865 | 1394 | pages = ram_save_host_page(rs, &pss, last_stage); |
56e93d26 | 1395 | } |
b9e60928 | 1396 | } while (!pages && again); |
56e93d26 | 1397 | |
6f37bb8b | 1398 | rs->last_seen_block = pss.block; |
a935e30f | 1399 | rs->last_page = pss.page; |
56e93d26 JQ |
1400 | |
1401 | return pages; | |
1402 | } | |
1403 | ||
1404 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1405 | { | |
1406 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b JQ |
1407 | RAMState *rs = &ram_state; |
1408 | ||
56e93d26 | 1409 | if (zero) { |
f7ccd61b | 1410 | rs->zero_pages += pages; |
56e93d26 | 1411 | } else { |
b4d1c6e7 | 1412 | rs->norm_pages += pages; |
2f4fde93 | 1413 | rs->bytes_transferred += size; |
56e93d26 JQ |
1414 | qemu_update_position(f, size); |
1415 | } | |
1416 | } | |
1417 | ||
56e93d26 JQ |
1418 | uint64_t ram_bytes_total(void) |
1419 | { | |
1420 | RAMBlock *block; | |
1421 | uint64_t total = 0; | |
1422 | ||
1423 | rcu_read_lock(); | |
99e15582 | 1424 | RAMBLOCK_FOREACH(block) { |
56e93d26 | 1425 | total += block->used_length; |
99e15582 | 1426 | } |
56e93d26 JQ |
1427 | rcu_read_unlock(); |
1428 | return total; | |
1429 | } | |
1430 | ||
1431 | void free_xbzrle_decoded_buf(void) | |
1432 | { | |
1433 | g_free(xbzrle_decoded_buf); | |
1434 | xbzrle_decoded_buf = NULL; | |
1435 | } | |
1436 | ||
6ad2a215 | 1437 | static void ram_migration_cleanup(void *opaque) |
56e93d26 | 1438 | { |
83c13382 | 1439 | RAMState *rs = opaque; |
6b6712ef | 1440 | RAMBlock *block; |
eb859c53 | 1441 | |
2ff64038 LZ |
1442 | /* caller have hold iothread lock or is in a bh, so there is |
1443 | * no writing race against this migration_bitmap | |
1444 | */ | |
6b6712ef JQ |
1445 | memory_global_dirty_log_stop(); |
1446 | ||
1447 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1448 | g_free(block->bmap); | |
1449 | block->bmap = NULL; | |
1450 | g_free(block->unsentmap); | |
1451 | block->unsentmap = NULL; | |
56e93d26 JQ |
1452 | } |
1453 | ||
1454 | XBZRLE_cache_lock(); | |
1455 | if (XBZRLE.cache) { | |
1456 | cache_fini(XBZRLE.cache); | |
1457 | g_free(XBZRLE.encoded_buf); | |
1458 | g_free(XBZRLE.current_buf); | |
c00e0928 | 1459 | g_free(XBZRLE.zero_target_page); |
56e93d26 JQ |
1460 | XBZRLE.cache = NULL; |
1461 | XBZRLE.encoded_buf = NULL; | |
1462 | XBZRLE.current_buf = NULL; | |
c00e0928 | 1463 | XBZRLE.zero_target_page = NULL; |
56e93d26 JQ |
1464 | } |
1465 | XBZRLE_cache_unlock(); | |
83c13382 | 1466 | migration_page_queue_free(rs); |
56e93d26 JQ |
1467 | } |
1468 | ||
6f37bb8b | 1469 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1470 | { |
6f37bb8b JQ |
1471 | rs->last_seen_block = NULL; |
1472 | rs->last_sent_block = NULL; | |
269ace29 | 1473 | rs->last_page = 0; |
6f37bb8b JQ |
1474 | rs->last_version = ram_list.version; |
1475 | rs->ram_bulk_stage = true; | |
56e93d26 JQ |
1476 | } |
1477 | ||
1478 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1479 | ||
4f2e4252 DDAG |
1480 | /* |
1481 | * 'expected' is the value you expect the bitmap mostly to be full | |
1482 | * of; it won't bother printing lines that are all this value. | |
1483 | * If 'todump' is null the migration bitmap is dumped. | |
1484 | */ | |
6b6712ef JQ |
1485 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected, |
1486 | unsigned long pages) | |
4f2e4252 | 1487 | { |
4f2e4252 DDAG |
1488 | int64_t cur; |
1489 | int64_t linelen = 128; | |
1490 | char linebuf[129]; | |
1491 | ||
6b6712ef | 1492 | for (cur = 0; cur < pages; cur += linelen) { |
4f2e4252 DDAG |
1493 | int64_t curb; |
1494 | bool found = false; | |
1495 | /* | |
1496 | * Last line; catch the case where the line length | |
1497 | * is longer than remaining ram | |
1498 | */ | |
6b6712ef JQ |
1499 | if (cur + linelen > pages) { |
1500 | linelen = pages - cur; | |
4f2e4252 DDAG |
1501 | } |
1502 | for (curb = 0; curb < linelen; curb++) { | |
1503 | bool thisbit = test_bit(cur + curb, todump); | |
1504 | linebuf[curb] = thisbit ? '1' : '.'; | |
1505 | found = found || (thisbit != expected); | |
1506 | } | |
1507 | if (found) { | |
1508 | linebuf[curb] = '\0'; | |
1509 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1510 | } | |
1511 | } | |
1512 | } | |
1513 | ||
e0b266f0 DDAG |
1514 | /* **** functions for postcopy ***** */ |
1515 | ||
ced1c616 PB |
1516 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1517 | { | |
1518 | struct RAMBlock *block; | |
ced1c616 | 1519 | |
99e15582 | 1520 | RAMBLOCK_FOREACH(block) { |
6b6712ef JQ |
1521 | unsigned long *bitmap = block->bmap; |
1522 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
1523 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
1524 | |
1525 | while (run_start < range) { | |
1526 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
aaa2064c | 1527 | ram_discard_range(block->idstr, run_start << TARGET_PAGE_BITS, |
ced1c616 PB |
1528 | (run_end - run_start) << TARGET_PAGE_BITS); |
1529 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); | |
1530 | } | |
1531 | } | |
1532 | } | |
1533 | ||
3d0684b2 JQ |
1534 | /** |
1535 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1536 | * | |
1537 | * Returns zero on success | |
1538 | * | |
e0b266f0 DDAG |
1539 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
1540 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1541 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
3d0684b2 JQ |
1542 | * |
1543 | * @ms: current migration state | |
1544 | * @pds: state for postcopy | |
1545 | * @start: RAMBlock starting page | |
1546 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1547 | */ |
1548 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1549 | PostcopyDiscardState *pds, | |
6b6712ef | 1550 | RAMBlock *block) |
e0b266f0 | 1551 | { |
6b6712ef | 1552 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 1553 | unsigned long current; |
6b6712ef | 1554 | unsigned long *unsentmap = block->unsentmap; |
e0b266f0 | 1555 | |
6b6712ef | 1556 | for (current = 0; current < end; ) { |
e0b266f0 DDAG |
1557 | unsigned long one = find_next_bit(unsentmap, end, current); |
1558 | ||
1559 | if (one <= end) { | |
1560 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1561 | unsigned long discard_length; | |
1562 | ||
1563 | if (zero >= end) { | |
1564 | discard_length = end - one; | |
1565 | } else { | |
1566 | discard_length = zero - one; | |
1567 | } | |
d688c62d DDAG |
1568 | if (discard_length) { |
1569 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1570 | } | |
e0b266f0 DDAG |
1571 | current = one + discard_length; |
1572 | } else { | |
1573 | current = one; | |
1574 | } | |
1575 | } | |
1576 | ||
1577 | return 0; | |
1578 | } | |
1579 | ||
3d0684b2 JQ |
1580 | /** |
1581 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1582 | * | |
1583 | * Returns 0 for success or negative for error | |
1584 | * | |
e0b266f0 DDAG |
1585 | * Utility for the outgoing postcopy code. |
1586 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1587 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1588 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1589 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1590 | * |
1591 | * @ms: current migration state | |
e0b266f0 DDAG |
1592 | */ |
1593 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1594 | { | |
1595 | struct RAMBlock *block; | |
1596 | int ret; | |
1597 | ||
99e15582 | 1598 | RAMBLOCK_FOREACH(block) { |
6b6712ef JQ |
1599 | PostcopyDiscardState *pds = |
1600 | postcopy_discard_send_init(ms, block->idstr); | |
e0b266f0 DDAG |
1601 | |
1602 | /* | |
1603 | * Postcopy sends chunks of bitmap over the wire, but it | |
1604 | * just needs indexes at this point, avoids it having | |
1605 | * target page specific code. | |
1606 | */ | |
6b6712ef | 1607 | ret = postcopy_send_discard_bm_ram(ms, pds, block); |
e0b266f0 DDAG |
1608 | postcopy_discard_send_finish(ms, pds); |
1609 | if (ret) { | |
1610 | return ret; | |
1611 | } | |
1612 | } | |
1613 | ||
1614 | return 0; | |
1615 | } | |
1616 | ||
3d0684b2 JQ |
1617 | /** |
1618 | * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages | |
1619 | * | |
1620 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
1621 | * canonicalize the two bitmaps, that are similar, but one is | |
1622 | * inverted. | |
99e314eb | 1623 | * |
3d0684b2 JQ |
1624 | * Postcopy requires that all target pages in a hostpage are dirty or |
1625 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 1626 | * |
3d0684b2 JQ |
1627 | * @ms: current migration state |
1628 | * @unsent_pass: if true we need to canonicalize partially unsent host pages | |
1629 | * otherwise we need to canonicalize partially dirty host pages | |
1630 | * @block: block that contains the page we want to canonicalize | |
1631 | * @pds: state for postcopy | |
99e314eb DDAG |
1632 | */ |
1633 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1634 | RAMBlock *block, | |
1635 | PostcopyDiscardState *pds) | |
1636 | { | |
0d8ec885 | 1637 | RAMState *rs = &ram_state; |
6b6712ef JQ |
1638 | unsigned long *bitmap = block->bmap; |
1639 | unsigned long *unsentmap = block->unsentmap; | |
29c59172 | 1640 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 1641 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
1642 | unsigned long run_start; |
1643 | ||
29c59172 DDAG |
1644 | if (block->page_size == TARGET_PAGE_SIZE) { |
1645 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
1646 | return; | |
1647 | } | |
1648 | ||
99e314eb DDAG |
1649 | if (unsent_pass) { |
1650 | /* Find a sent page */ | |
6b6712ef | 1651 | run_start = find_next_zero_bit(unsentmap, pages, 0); |
99e314eb DDAG |
1652 | } else { |
1653 | /* Find a dirty page */ | |
6b6712ef | 1654 | run_start = find_next_bit(bitmap, pages, 0); |
99e314eb DDAG |
1655 | } |
1656 | ||
6b6712ef | 1657 | while (run_start < pages) { |
99e314eb DDAG |
1658 | bool do_fixup = false; |
1659 | unsigned long fixup_start_addr; | |
1660 | unsigned long host_offset; | |
1661 | ||
1662 | /* | |
1663 | * If the start of this run of pages is in the middle of a host | |
1664 | * page, then we need to fixup this host page. | |
1665 | */ | |
1666 | host_offset = run_start % host_ratio; | |
1667 | if (host_offset) { | |
1668 | do_fixup = true; | |
1669 | run_start -= host_offset; | |
1670 | fixup_start_addr = run_start; | |
1671 | /* For the next pass */ | |
1672 | run_start = run_start + host_ratio; | |
1673 | } else { | |
1674 | /* Find the end of this run */ | |
1675 | unsigned long run_end; | |
1676 | if (unsent_pass) { | |
6b6712ef | 1677 | run_end = find_next_bit(unsentmap, pages, run_start + 1); |
99e314eb | 1678 | } else { |
6b6712ef | 1679 | run_end = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
1680 | } |
1681 | /* | |
1682 | * If the end isn't at the start of a host page, then the | |
1683 | * run doesn't finish at the end of a host page | |
1684 | * and we need to discard. | |
1685 | */ | |
1686 | host_offset = run_end % host_ratio; | |
1687 | if (host_offset) { | |
1688 | do_fixup = true; | |
1689 | fixup_start_addr = run_end - host_offset; | |
1690 | /* | |
1691 | * This host page has gone, the next loop iteration starts | |
1692 | * from after the fixup | |
1693 | */ | |
1694 | run_start = fixup_start_addr + host_ratio; | |
1695 | } else { | |
1696 | /* | |
1697 | * No discards on this iteration, next loop starts from | |
1698 | * next sent/dirty page | |
1699 | */ | |
1700 | run_start = run_end + 1; | |
1701 | } | |
1702 | } | |
1703 | ||
1704 | if (do_fixup) { | |
1705 | unsigned long page; | |
1706 | ||
1707 | /* Tell the destination to discard this page */ | |
1708 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1709 | /* For the unsent_pass we: | |
1710 | * discard partially sent pages | |
1711 | * For the !unsent_pass (dirty) we: | |
1712 | * discard partially dirty pages that were sent | |
1713 | * (any partially sent pages were already discarded | |
1714 | * by the previous unsent_pass) | |
1715 | */ | |
1716 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1717 | host_ratio); | |
1718 | } | |
1719 | ||
1720 | /* Clean up the bitmap */ | |
1721 | for (page = fixup_start_addr; | |
1722 | page < fixup_start_addr + host_ratio; page++) { | |
1723 | /* All pages in this host page are now not sent */ | |
1724 | set_bit(page, unsentmap); | |
1725 | ||
1726 | /* | |
1727 | * Remark them as dirty, updating the count for any pages | |
1728 | * that weren't previously dirty. | |
1729 | */ | |
0d8ec885 | 1730 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
1731 | } |
1732 | } | |
1733 | ||
1734 | if (unsent_pass) { | |
1735 | /* Find the next sent page for the next iteration */ | |
6b6712ef | 1736 | run_start = find_next_zero_bit(unsentmap, pages, run_start); |
99e314eb DDAG |
1737 | } else { |
1738 | /* Find the next dirty page for the next iteration */ | |
6b6712ef | 1739 | run_start = find_next_bit(bitmap, pages, run_start); |
99e314eb DDAG |
1740 | } |
1741 | } | |
1742 | } | |
1743 | ||
3d0684b2 JQ |
1744 | /** |
1745 | * postcopy_chuck_hostpages: discrad any partially sent host page | |
1746 | * | |
99e314eb DDAG |
1747 | * Utility for the outgoing postcopy code. |
1748 | * | |
1749 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
1750 | * dirty host-page size chunks as all dirty. In this case the host-page |
1751 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 1752 | * |
3d0684b2 JQ |
1753 | * Returns zero on success |
1754 | * | |
1755 | * @ms: current migration state | |
6b6712ef | 1756 | * @block: block we want to work with |
99e314eb | 1757 | */ |
6b6712ef | 1758 | static int postcopy_chunk_hostpages(MigrationState *ms, RAMBlock *block) |
99e314eb | 1759 | { |
6b6712ef JQ |
1760 | PostcopyDiscardState *pds = |
1761 | postcopy_discard_send_init(ms, block->idstr); | |
99e314eb | 1762 | |
6b6712ef JQ |
1763 | /* First pass: Discard all partially sent host pages */ |
1764 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1765 | /* | |
1766 | * Second pass: Ensure that all partially dirty host pages are made | |
1767 | * fully dirty. | |
1768 | */ | |
1769 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
99e314eb | 1770 | |
6b6712ef | 1771 | postcopy_discard_send_finish(ms, pds); |
99e314eb DDAG |
1772 | return 0; |
1773 | } | |
1774 | ||
3d0684b2 JQ |
1775 | /** |
1776 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
1777 | * | |
1778 | * Returns zero on success | |
1779 | * | |
e0b266f0 DDAG |
1780 | * Transmit the set of pages to be discarded after precopy to the target |
1781 | * these are pages that: | |
1782 | * a) Have been previously transmitted but are now dirty again | |
1783 | * b) Pages that have never been transmitted, this ensures that | |
1784 | * any pages on the destination that have been mapped by background | |
1785 | * tasks get discarded (transparent huge pages is the specific concern) | |
1786 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
1787 | * |
1788 | * @ms: current migration state | |
e0b266f0 DDAG |
1789 | */ |
1790 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1791 | { | |
eb859c53 | 1792 | RAMState *rs = &ram_state; |
6b6712ef | 1793 | RAMBlock *block; |
e0b266f0 | 1794 | int ret; |
e0b266f0 DDAG |
1795 | |
1796 | rcu_read_lock(); | |
1797 | ||
1798 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 1799 | migration_bitmap_sync(rs); |
e0b266f0 | 1800 | |
6b6712ef JQ |
1801 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
1802 | rs->last_seen_block = NULL; | |
1803 | rs->last_sent_block = NULL; | |
1804 | rs->last_page = 0; | |
e0b266f0 | 1805 | |
6b6712ef JQ |
1806 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { |
1807 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; | |
1808 | unsigned long *bitmap = block->bmap; | |
1809 | unsigned long *unsentmap = block->unsentmap; | |
1810 | ||
1811 | if (!unsentmap) { | |
1812 | /* We don't have a safe way to resize the sentmap, so | |
1813 | * if the bitmap was resized it will be NULL at this | |
1814 | * point. | |
1815 | */ | |
1816 | error_report("migration ram resized during precopy phase"); | |
1817 | rcu_read_unlock(); | |
1818 | return -EINVAL; | |
1819 | } | |
1820 | /* Deal with TPS != HPS and huge pages */ | |
1821 | ret = postcopy_chunk_hostpages(ms, block); | |
1822 | if (ret) { | |
1823 | rcu_read_unlock(); | |
1824 | return ret; | |
1825 | } | |
e0b266f0 | 1826 | |
6b6712ef JQ |
1827 | /* |
1828 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1829 | */ | |
1830 | bitmap_or(unsentmap, unsentmap, bitmap, pages); | |
e0b266f0 | 1831 | #ifdef DEBUG_POSTCOPY |
6b6712ef | 1832 | ram_debug_dump_bitmap(unsentmap, true, pages); |
e0b266f0 | 1833 | #endif |
6b6712ef JQ |
1834 | } |
1835 | trace_ram_postcopy_send_discard_bitmap(); | |
e0b266f0 DDAG |
1836 | |
1837 | ret = postcopy_each_ram_send_discard(ms); | |
1838 | rcu_read_unlock(); | |
1839 | ||
1840 | return ret; | |
1841 | } | |
1842 | ||
3d0684b2 JQ |
1843 | /** |
1844 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 1845 | * |
3d0684b2 | 1846 | * Returns zero on success |
e0b266f0 | 1847 | * |
36449157 JQ |
1848 | * @rbname: name of the RAMBlock of the request. NULL means the |
1849 | * same that last one. | |
3d0684b2 JQ |
1850 | * @start: RAMBlock starting page |
1851 | * @length: RAMBlock size | |
e0b266f0 | 1852 | */ |
aaa2064c | 1853 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 DDAG |
1854 | { |
1855 | int ret = -1; | |
1856 | ||
36449157 | 1857 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 1858 | |
e0b266f0 | 1859 | rcu_read_lock(); |
36449157 | 1860 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
1861 | |
1862 | if (!rb) { | |
36449157 | 1863 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
e0b266f0 DDAG |
1864 | goto err; |
1865 | } | |
1866 | ||
d3a5038c | 1867 | ret = ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
1868 | |
1869 | err: | |
1870 | rcu_read_unlock(); | |
1871 | ||
1872 | return ret; | |
1873 | } | |
1874 | ||
ceb4d168 | 1875 | static int ram_state_init(RAMState *rs) |
56e93d26 | 1876 | { |
ceb4d168 | 1877 | memset(rs, 0, sizeof(*rs)); |
108cfae0 | 1878 | qemu_mutex_init(&rs->bitmap_mutex); |
ec481c6c JQ |
1879 | qemu_mutex_init(&rs->src_page_req_mutex); |
1880 | QSIMPLEQ_INIT(&rs->src_page_requests); | |
56e93d26 JQ |
1881 | |
1882 | if (migrate_use_xbzrle()) { | |
1883 | XBZRLE_cache_lock(); | |
c00e0928 | 1884 | XBZRLE.zero_target_page = g_malloc0(TARGET_PAGE_SIZE); |
56e93d26 JQ |
1885 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() / |
1886 | TARGET_PAGE_SIZE, | |
1887 | TARGET_PAGE_SIZE); | |
1888 | if (!XBZRLE.cache) { | |
1889 | XBZRLE_cache_unlock(); | |
1890 | error_report("Error creating cache"); | |
1891 | return -1; | |
1892 | } | |
1893 | XBZRLE_cache_unlock(); | |
1894 | ||
1895 | /* We prefer not to abort if there is no memory */ | |
1896 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
1897 | if (!XBZRLE.encoded_buf) { | |
1898 | error_report("Error allocating encoded_buf"); | |
1899 | return -1; | |
1900 | } | |
1901 | ||
1902 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
1903 | if (!XBZRLE.current_buf) { | |
1904 | error_report("Error allocating current_buf"); | |
1905 | g_free(XBZRLE.encoded_buf); | |
1906 | XBZRLE.encoded_buf = NULL; | |
1907 | return -1; | |
1908 | } | |
56e93d26 JQ |
1909 | } |
1910 | ||
49877834 PB |
1911 | /* For memory_global_dirty_log_start below. */ |
1912 | qemu_mutex_lock_iothread(); | |
1913 | ||
56e93d26 JQ |
1914 | qemu_mutex_lock_ramlist(); |
1915 | rcu_read_lock(); | |
6f37bb8b | 1916 | ram_state_reset(rs); |
56e93d26 | 1917 | |
0827b9e9 AA |
1918 | /* Skip setting bitmap if there is no RAM */ |
1919 | if (ram_bytes_total()) { | |
6b6712ef JQ |
1920 | RAMBlock *block; |
1921 | ||
1922 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1923 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; | |
0827b9e9 | 1924 | |
6b6712ef JQ |
1925 | block->bmap = bitmap_new(pages); |
1926 | bitmap_set(block->bmap, 0, pages); | |
1927 | if (migrate_postcopy_ram()) { | |
1928 | block->unsentmap = bitmap_new(pages); | |
1929 | bitmap_set(block->unsentmap, 0, pages); | |
1930 | } | |
0827b9e9 | 1931 | } |
f3f491fc DDAG |
1932 | } |
1933 | ||
56e93d26 JQ |
1934 | /* |
1935 | * Count the total number of pages used by ram blocks not including any | |
1936 | * gaps due to alignment or unplugs. | |
1937 | */ | |
0d8ec885 | 1938 | rs->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
56e93d26 JQ |
1939 | |
1940 | memory_global_dirty_log_start(); | |
8d820d6f | 1941 | migration_bitmap_sync(rs); |
56e93d26 | 1942 | qemu_mutex_unlock_ramlist(); |
49877834 | 1943 | qemu_mutex_unlock_iothread(); |
a91246c9 HZ |
1944 | rcu_read_unlock(); |
1945 | ||
1946 | return 0; | |
1947 | } | |
1948 | ||
3d0684b2 JQ |
1949 | /* |
1950 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
1951 | * long-running RCU critical section. When rcu-reclaims in the code |
1952 | * start to become numerous it will be necessary to reduce the | |
1953 | * granularity of these critical sections. | |
1954 | */ | |
1955 | ||
3d0684b2 JQ |
1956 | /** |
1957 | * ram_save_setup: Setup RAM for migration | |
1958 | * | |
1959 | * Returns zero to indicate success and negative for error | |
1960 | * | |
1961 | * @f: QEMUFile where to send the data | |
1962 | * @opaque: RAMState pointer | |
1963 | */ | |
a91246c9 HZ |
1964 | static int ram_save_setup(QEMUFile *f, void *opaque) |
1965 | { | |
6f37bb8b | 1966 | RAMState *rs = opaque; |
a91246c9 HZ |
1967 | RAMBlock *block; |
1968 | ||
1969 | /* migration has already setup the bitmap, reuse it. */ | |
1970 | if (!migration_in_colo_state()) { | |
ceb4d168 | 1971 | if (ram_state_init(rs) < 0) { |
a91246c9 HZ |
1972 | return -1; |
1973 | } | |
1974 | } | |
204b88b8 | 1975 | rs->f = f; |
a91246c9 HZ |
1976 | |
1977 | rcu_read_lock(); | |
56e93d26 JQ |
1978 | |
1979 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
1980 | ||
99e15582 | 1981 | RAMBLOCK_FOREACH(block) { |
56e93d26 JQ |
1982 | qemu_put_byte(f, strlen(block->idstr)); |
1983 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
1984 | qemu_put_be64(f, block->used_length); | |
ef08fb38 DDAG |
1985 | if (migrate_postcopy_ram() && block->page_size != qemu_host_page_size) { |
1986 | qemu_put_be64(f, block->page_size); | |
1987 | } | |
56e93d26 JQ |
1988 | } |
1989 | ||
1990 | rcu_read_unlock(); | |
1991 | ||
1992 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
1993 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
1994 | ||
1995 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
1996 | ||
1997 | return 0; | |
1998 | } | |
1999 | ||
3d0684b2 JQ |
2000 | /** |
2001 | * ram_save_iterate: iterative stage for migration | |
2002 | * | |
2003 | * Returns zero to indicate success and negative for error | |
2004 | * | |
2005 | * @f: QEMUFile where to send the data | |
2006 | * @opaque: RAMState pointer | |
2007 | */ | |
56e93d26 JQ |
2008 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2009 | { | |
6f37bb8b | 2010 | RAMState *rs = opaque; |
56e93d26 JQ |
2011 | int ret; |
2012 | int i; | |
2013 | int64_t t0; | |
5c90308f | 2014 | int done = 0; |
56e93d26 JQ |
2015 | |
2016 | rcu_read_lock(); | |
6f37bb8b JQ |
2017 | if (ram_list.version != rs->last_version) { |
2018 | ram_state_reset(rs); | |
56e93d26 JQ |
2019 | } |
2020 | ||
2021 | /* Read version before ram_list.blocks */ | |
2022 | smp_rmb(); | |
2023 | ||
2024 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
2025 | ||
2026 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
2027 | i = 0; | |
2028 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
2029 | int pages; | |
2030 | ||
ce25d337 | 2031 | pages = ram_find_and_save_block(rs, false); |
56e93d26 JQ |
2032 | /* no more pages to sent */ |
2033 | if (pages == 0) { | |
5c90308f | 2034 | done = 1; |
56e93d26 JQ |
2035 | break; |
2036 | } | |
23b28c3c | 2037 | rs->iterations++; |
070afca2 | 2038 | |
56e93d26 JQ |
2039 | /* we want to check in the 1st loop, just in case it was the 1st time |
2040 | and we had to sync the dirty bitmap. | |
2041 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2042 | iterations | |
2043 | */ | |
2044 | if ((i & 63) == 0) { | |
2045 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2046 | if (t1 > MAX_WAIT) { | |
55c4446b | 2047 | trace_ram_save_iterate_big_wait(t1, i); |
56e93d26 JQ |
2048 | break; |
2049 | } | |
2050 | } | |
2051 | i++; | |
2052 | } | |
ce25d337 | 2053 | flush_compressed_data(rs); |
56e93d26 JQ |
2054 | rcu_read_unlock(); |
2055 | ||
2056 | /* | |
2057 | * Must occur before EOS (or any QEMUFile operation) | |
2058 | * because of RDMA protocol. | |
2059 | */ | |
2060 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2061 | ||
2062 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2f4fde93 | 2063 | rs->bytes_transferred += 8; |
56e93d26 JQ |
2064 | |
2065 | ret = qemu_file_get_error(f); | |
2066 | if (ret < 0) { | |
2067 | return ret; | |
2068 | } | |
2069 | ||
5c90308f | 2070 | return done; |
56e93d26 JQ |
2071 | } |
2072 | ||
3d0684b2 JQ |
2073 | /** |
2074 | * ram_save_complete: function called to send the remaining amount of ram | |
2075 | * | |
2076 | * Returns zero to indicate success | |
2077 | * | |
2078 | * Called with iothread lock | |
2079 | * | |
2080 | * @f: QEMUFile where to send the data | |
2081 | * @opaque: RAMState pointer | |
2082 | */ | |
56e93d26 JQ |
2083 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2084 | { | |
6f37bb8b JQ |
2085 | RAMState *rs = opaque; |
2086 | ||
56e93d26 JQ |
2087 | rcu_read_lock(); |
2088 | ||
5727309d | 2089 | if (!migration_in_postcopy()) { |
8d820d6f | 2090 | migration_bitmap_sync(rs); |
663e6c1d | 2091 | } |
56e93d26 JQ |
2092 | |
2093 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2094 | ||
2095 | /* try transferring iterative blocks of memory */ | |
2096 | ||
2097 | /* flush all remaining blocks regardless of rate limiting */ | |
2098 | while (true) { | |
2099 | int pages; | |
2100 | ||
ce25d337 | 2101 | pages = ram_find_and_save_block(rs, !migration_in_colo_state()); |
56e93d26 JQ |
2102 | /* no more blocks to sent */ |
2103 | if (pages == 0) { | |
2104 | break; | |
2105 | } | |
2106 | } | |
2107 | ||
ce25d337 | 2108 | flush_compressed_data(rs); |
56e93d26 | 2109 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 JQ |
2110 | |
2111 | rcu_read_unlock(); | |
d09a6fde | 2112 | |
56e93d26 JQ |
2113 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2114 | ||
2115 | return 0; | |
2116 | } | |
2117 | ||
c31b098f DDAG |
2118 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2119 | uint64_t *non_postcopiable_pending, | |
2120 | uint64_t *postcopiable_pending) | |
56e93d26 | 2121 | { |
8d820d6f | 2122 | RAMState *rs = opaque; |
56e93d26 JQ |
2123 | uint64_t remaining_size; |
2124 | ||
9edabd4d | 2125 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2126 | |
5727309d | 2127 | if (!migration_in_postcopy() && |
663e6c1d | 2128 | remaining_size < max_size) { |
56e93d26 JQ |
2129 | qemu_mutex_lock_iothread(); |
2130 | rcu_read_lock(); | |
8d820d6f | 2131 | migration_bitmap_sync(rs); |
56e93d26 JQ |
2132 | rcu_read_unlock(); |
2133 | qemu_mutex_unlock_iothread(); | |
9edabd4d | 2134 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2135 | } |
c31b098f DDAG |
2136 | |
2137 | /* We can do postcopy, and all the data is postcopiable */ | |
2138 | *postcopiable_pending += remaining_size; | |
56e93d26 JQ |
2139 | } |
2140 | ||
2141 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2142 | { | |
2143 | unsigned int xh_len; | |
2144 | int xh_flags; | |
063e760a | 2145 | uint8_t *loaded_data; |
56e93d26 JQ |
2146 | |
2147 | if (!xbzrle_decoded_buf) { | |
2148 | xbzrle_decoded_buf = g_malloc(TARGET_PAGE_SIZE); | |
2149 | } | |
063e760a | 2150 | loaded_data = xbzrle_decoded_buf; |
56e93d26 JQ |
2151 | |
2152 | /* extract RLE header */ | |
2153 | xh_flags = qemu_get_byte(f); | |
2154 | xh_len = qemu_get_be16(f); | |
2155 | ||
2156 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2157 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2158 | return -1; | |
2159 | } | |
2160 | ||
2161 | if (xh_len > TARGET_PAGE_SIZE) { | |
2162 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2163 | return -1; | |
2164 | } | |
2165 | /* load data and decode */ | |
063e760a | 2166 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2167 | |
2168 | /* decode RLE */ | |
063e760a | 2169 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2170 | TARGET_PAGE_SIZE) == -1) { |
2171 | error_report("Failed to load XBZRLE page - decode error!"); | |
2172 | return -1; | |
2173 | } | |
2174 | ||
2175 | return 0; | |
2176 | } | |
2177 | ||
3d0684b2 JQ |
2178 | /** |
2179 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2180 | * | |
2181 | * Must be called from within a rcu critical section. | |
2182 | * | |
56e93d26 | 2183 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2184 | * |
3d0684b2 JQ |
2185 | * @f: QEMUFile where to read the data from |
2186 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2187 | */ |
3d0684b2 | 2188 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2189 | { |
2190 | static RAMBlock *block = NULL; | |
2191 | char id[256]; | |
2192 | uint8_t len; | |
2193 | ||
2194 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2195 | if (!block) { |
56e93d26 JQ |
2196 | error_report("Ack, bad migration stream!"); |
2197 | return NULL; | |
2198 | } | |
4c4bad48 | 2199 | return block; |
56e93d26 JQ |
2200 | } |
2201 | ||
2202 | len = qemu_get_byte(f); | |
2203 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2204 | id[len] = 0; | |
2205 | ||
e3dd7493 | 2206 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2207 | if (!block) { |
2208 | error_report("Can't find block %s", id); | |
2209 | return NULL; | |
56e93d26 JQ |
2210 | } |
2211 | ||
4c4bad48 HZ |
2212 | return block; |
2213 | } | |
2214 | ||
2215 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2216 | ram_addr_t offset) | |
2217 | { | |
2218 | if (!offset_in_ramblock(block, offset)) { | |
2219 | return NULL; | |
2220 | } | |
2221 | ||
2222 | return block->host + offset; | |
56e93d26 JQ |
2223 | } |
2224 | ||
3d0684b2 JQ |
2225 | /** |
2226 | * ram_handle_compressed: handle the zero page case | |
2227 | * | |
56e93d26 JQ |
2228 | * If a page (or a whole RDMA chunk) has been |
2229 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2230 | * |
2231 | * @host: host address for the zero page | |
2232 | * @ch: what the page is filled from. We only support zero | |
2233 | * @size: size of the zero page | |
56e93d26 JQ |
2234 | */ |
2235 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2236 | { | |
2237 | if (ch != 0 || !is_zero_range(host, size)) { | |
2238 | memset(host, ch, size); | |
2239 | } | |
2240 | } | |
2241 | ||
2242 | static void *do_data_decompress(void *opaque) | |
2243 | { | |
2244 | DecompressParam *param = opaque; | |
2245 | unsigned long pagesize; | |
33d151f4 LL |
2246 | uint8_t *des; |
2247 | int len; | |
56e93d26 | 2248 | |
33d151f4 | 2249 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2250 | while (!param->quit) { |
33d151f4 LL |
2251 | if (param->des) { |
2252 | des = param->des; | |
2253 | len = param->len; | |
2254 | param->des = 0; | |
2255 | qemu_mutex_unlock(¶m->mutex); | |
2256 | ||
56e93d26 | 2257 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2258 | /* uncompress() will return failed in some case, especially |
2259 | * when the page is dirted when doing the compression, it's | |
2260 | * not a problem because the dirty page will be retransferred | |
2261 | * and uncompress() won't break the data in other pages. | |
2262 | */ | |
33d151f4 LL |
2263 | uncompress((Bytef *)des, &pagesize, |
2264 | (const Bytef *)param->compbuf, len); | |
73a8912b | 2265 | |
33d151f4 LL |
2266 | qemu_mutex_lock(&decomp_done_lock); |
2267 | param->done = true; | |
2268 | qemu_cond_signal(&decomp_done_cond); | |
2269 | qemu_mutex_unlock(&decomp_done_lock); | |
2270 | ||
2271 | qemu_mutex_lock(¶m->mutex); | |
2272 | } else { | |
2273 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2274 | } | |
56e93d26 | 2275 | } |
33d151f4 | 2276 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2277 | |
2278 | return NULL; | |
2279 | } | |
2280 | ||
5533b2e9 LL |
2281 | static void wait_for_decompress_done(void) |
2282 | { | |
2283 | int idx, thread_count; | |
2284 | ||
2285 | if (!migrate_use_compression()) { | |
2286 | return; | |
2287 | } | |
2288 | ||
2289 | thread_count = migrate_decompress_threads(); | |
2290 | qemu_mutex_lock(&decomp_done_lock); | |
2291 | for (idx = 0; idx < thread_count; idx++) { | |
2292 | while (!decomp_param[idx].done) { | |
2293 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2294 | } | |
2295 | } | |
2296 | qemu_mutex_unlock(&decomp_done_lock); | |
2297 | } | |
2298 | ||
56e93d26 JQ |
2299 | void migrate_decompress_threads_create(void) |
2300 | { | |
2301 | int i, thread_count; | |
2302 | ||
2303 | thread_count = migrate_decompress_threads(); | |
2304 | decompress_threads = g_new0(QemuThread, thread_count); | |
2305 | decomp_param = g_new0(DecompressParam, thread_count); | |
73a8912b LL |
2306 | qemu_mutex_init(&decomp_done_lock); |
2307 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2308 | for (i = 0; i < thread_count; i++) { |
2309 | qemu_mutex_init(&decomp_param[i].mutex); | |
2310 | qemu_cond_init(&decomp_param[i].cond); | |
2311 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2312 | decomp_param[i].done = true; |
90e56fb4 | 2313 | decomp_param[i].quit = false; |
56e93d26 JQ |
2314 | qemu_thread_create(decompress_threads + i, "decompress", |
2315 | do_data_decompress, decomp_param + i, | |
2316 | QEMU_THREAD_JOINABLE); | |
2317 | } | |
2318 | } | |
2319 | ||
2320 | void migrate_decompress_threads_join(void) | |
2321 | { | |
2322 | int i, thread_count; | |
2323 | ||
56e93d26 JQ |
2324 | thread_count = migrate_decompress_threads(); |
2325 | for (i = 0; i < thread_count; i++) { | |
2326 | qemu_mutex_lock(&decomp_param[i].mutex); | |
90e56fb4 | 2327 | decomp_param[i].quit = true; |
56e93d26 JQ |
2328 | qemu_cond_signal(&decomp_param[i].cond); |
2329 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2330 | } | |
2331 | for (i = 0; i < thread_count; i++) { | |
2332 | qemu_thread_join(decompress_threads + i); | |
2333 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2334 | qemu_cond_destroy(&decomp_param[i].cond); | |
2335 | g_free(decomp_param[i].compbuf); | |
2336 | } | |
2337 | g_free(decompress_threads); | |
2338 | g_free(decomp_param); | |
56e93d26 JQ |
2339 | decompress_threads = NULL; |
2340 | decomp_param = NULL; | |
56e93d26 JQ |
2341 | } |
2342 | ||
c1bc6626 | 2343 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2344 | void *host, int len) |
2345 | { | |
2346 | int idx, thread_count; | |
2347 | ||
2348 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2349 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2350 | while (true) { |
2351 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2352 | if (decomp_param[idx].done) { |
33d151f4 LL |
2353 | decomp_param[idx].done = false; |
2354 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2355 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2356 | decomp_param[idx].des = host; |
2357 | decomp_param[idx].len = len; | |
33d151f4 LL |
2358 | qemu_cond_signal(&decomp_param[idx].cond); |
2359 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2360 | break; |
2361 | } | |
2362 | } | |
2363 | if (idx < thread_count) { | |
2364 | break; | |
73a8912b LL |
2365 | } else { |
2366 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2367 | } |
2368 | } | |
73a8912b | 2369 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2370 | } |
2371 | ||
3d0684b2 JQ |
2372 | /** |
2373 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
2374 | * | |
2375 | * Returns 0 for success and negative if there was one error | |
2376 | * | |
2377 | * @mis: current migration incoming state | |
2378 | * | |
2379 | * Allocate data structures etc needed by incoming migration with | |
2380 | * postcopy-ram. postcopy-ram's similarly names | |
2381 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
2382 | */ |
2383 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2384 | { | |
b8c48993 | 2385 | unsigned long ram_pages = last_ram_page(); |
1caddf8a DDAG |
2386 | |
2387 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2388 | } | |
2389 | ||
3d0684b2 JQ |
2390 | /** |
2391 | * ram_load_postcopy: load a page in postcopy case | |
2392 | * | |
2393 | * Returns 0 for success or -errno in case of error | |
2394 | * | |
a7180877 DDAG |
2395 | * Called in postcopy mode by ram_load(). |
2396 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
2397 | * |
2398 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
2399 | */ |
2400 | static int ram_load_postcopy(QEMUFile *f) | |
2401 | { | |
2402 | int flags = 0, ret = 0; | |
2403 | bool place_needed = false; | |
28abd200 | 2404 | bool matching_page_sizes = false; |
a7180877 DDAG |
2405 | MigrationIncomingState *mis = migration_incoming_get_current(); |
2406 | /* Temporary page that is later 'placed' */ | |
2407 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2408 | void *last_host = NULL; |
a3b6ff6d | 2409 | bool all_zero = false; |
a7180877 DDAG |
2410 | |
2411 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2412 | ram_addr_t addr; | |
2413 | void *host = NULL; | |
2414 | void *page_buffer = NULL; | |
2415 | void *place_source = NULL; | |
df9ff5e1 | 2416 | RAMBlock *block = NULL; |
a7180877 | 2417 | uint8_t ch; |
a7180877 DDAG |
2418 | |
2419 | addr = qemu_get_be64(f); | |
2420 | flags = addr & ~TARGET_PAGE_MASK; | |
2421 | addr &= TARGET_PAGE_MASK; | |
2422 | ||
2423 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2424 | place_needed = false; | |
bb890ed5 | 2425 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE)) { |
df9ff5e1 | 2426 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
2427 | |
2428 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2429 | if (!host) { |
2430 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2431 | ret = -EINVAL; | |
2432 | break; | |
2433 | } | |
28abd200 | 2434 | matching_page_sizes = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 2435 | /* |
28abd200 DDAG |
2436 | * Postcopy requires that we place whole host pages atomically; |
2437 | * these may be huge pages for RAMBlocks that are backed by | |
2438 | * hugetlbfs. | |
a7180877 DDAG |
2439 | * To make it atomic, the data is read into a temporary page |
2440 | * that's moved into place later. | |
2441 | * The migration protocol uses, possibly smaller, target-pages | |
2442 | * however the source ensures it always sends all the components | |
2443 | * of a host page in order. | |
2444 | */ | |
2445 | page_buffer = postcopy_host_page + | |
28abd200 | 2446 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 2447 | /* If all TP are zero then we can optimise the place */ |
28abd200 | 2448 | if (!((uintptr_t)host & (block->page_size - 1))) { |
a7180877 | 2449 | all_zero = true; |
c53b7ddc DDAG |
2450 | } else { |
2451 | /* not the 1st TP within the HP */ | |
2452 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2453 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2454 | host, last_host); |
2455 | ret = -EINVAL; | |
2456 | break; | |
2457 | } | |
a7180877 DDAG |
2458 | } |
2459 | ||
c53b7ddc | 2460 | |
a7180877 DDAG |
2461 | /* |
2462 | * If it's the last part of a host page then we place the host | |
2463 | * page | |
2464 | */ | |
2465 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
28abd200 | 2466 | (block->page_size - 1)) == 0; |
a7180877 DDAG |
2467 | place_source = postcopy_host_page; |
2468 | } | |
c53b7ddc | 2469 | last_host = host; |
a7180877 DDAG |
2470 | |
2471 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 2472 | case RAM_SAVE_FLAG_ZERO: |
a7180877 DDAG |
2473 | ch = qemu_get_byte(f); |
2474 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2475 | if (ch) { | |
2476 | all_zero = false; | |
2477 | } | |
2478 | break; | |
2479 | ||
2480 | case RAM_SAVE_FLAG_PAGE: | |
2481 | all_zero = false; | |
2482 | if (!place_needed || !matching_page_sizes) { | |
2483 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2484 | } else { | |
2485 | /* Avoids the qemu_file copy during postcopy, which is | |
2486 | * going to do a copy later; can only do it when we | |
2487 | * do this read in one go (matching page sizes) | |
2488 | */ | |
2489 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2490 | TARGET_PAGE_SIZE); | |
2491 | } | |
2492 | break; | |
2493 | case RAM_SAVE_FLAG_EOS: | |
2494 | /* normal exit */ | |
2495 | break; | |
2496 | default: | |
2497 | error_report("Unknown combination of migration flags: %#x" | |
2498 | " (postcopy mode)", flags); | |
2499 | ret = -EINVAL; | |
2500 | } | |
2501 | ||
2502 | if (place_needed) { | |
2503 | /* This gets called at the last target page in the host page */ | |
df9ff5e1 DDAG |
2504 | void *place_dest = host + TARGET_PAGE_SIZE - block->page_size; |
2505 | ||
a7180877 | 2506 | if (all_zero) { |
df9ff5e1 DDAG |
2507 | ret = postcopy_place_page_zero(mis, place_dest, |
2508 | block->page_size); | |
a7180877 | 2509 | } else { |
df9ff5e1 DDAG |
2510 | ret = postcopy_place_page(mis, place_dest, |
2511 | place_source, block->page_size); | |
a7180877 DDAG |
2512 | } |
2513 | } | |
2514 | if (!ret) { | |
2515 | ret = qemu_file_get_error(f); | |
2516 | } | |
2517 | } | |
2518 | ||
2519 | return ret; | |
2520 | } | |
2521 | ||
56e93d26 JQ |
2522 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2523 | { | |
2524 | int flags = 0, ret = 0; | |
2525 | static uint64_t seq_iter; | |
2526 | int len = 0; | |
a7180877 DDAG |
2527 | /* |
2528 | * If system is running in postcopy mode, page inserts to host memory must | |
2529 | * be atomic | |
2530 | */ | |
2531 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
ef08fb38 DDAG |
2532 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
2533 | bool postcopy_advised = postcopy_state_get() >= POSTCOPY_INCOMING_ADVISE; | |
56e93d26 JQ |
2534 | |
2535 | seq_iter++; | |
2536 | ||
2537 | if (version_id != 4) { | |
2538 | ret = -EINVAL; | |
2539 | } | |
2540 | ||
2541 | /* This RCU critical section can be very long running. | |
2542 | * When RCU reclaims in the code start to become numerous, | |
2543 | * it will be necessary to reduce the granularity of this | |
2544 | * critical section. | |
2545 | */ | |
2546 | rcu_read_lock(); | |
a7180877 DDAG |
2547 | |
2548 | if (postcopy_running) { | |
2549 | ret = ram_load_postcopy(f); | |
2550 | } | |
2551 | ||
2552 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2553 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2554 | void *host = NULL; |
56e93d26 JQ |
2555 | uint8_t ch; |
2556 | ||
2557 | addr = qemu_get_be64(f); | |
2558 | flags = addr & ~TARGET_PAGE_MASK; | |
2559 | addr &= TARGET_PAGE_MASK; | |
2560 | ||
bb890ed5 | 2561 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 2562 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
4c4bad48 HZ |
2563 | RAMBlock *block = ram_block_from_stream(f, flags); |
2564 | ||
2565 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2566 | if (!host) { |
2567 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2568 | ret = -EINVAL; | |
2569 | break; | |
2570 | } | |
1db9d8e5 | 2571 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
2572 | } |
2573 | ||
56e93d26 JQ |
2574 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2575 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2576 | /* Synchronize RAM block list */ | |
2577 | total_ram_bytes = addr; | |
2578 | while (!ret && total_ram_bytes) { | |
2579 | RAMBlock *block; | |
56e93d26 JQ |
2580 | char id[256]; |
2581 | ram_addr_t length; | |
2582 | ||
2583 | len = qemu_get_byte(f); | |
2584 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2585 | id[len] = 0; | |
2586 | length = qemu_get_be64(f); | |
2587 | ||
e3dd7493 DDAG |
2588 | block = qemu_ram_block_by_name(id); |
2589 | if (block) { | |
2590 | if (length != block->used_length) { | |
2591 | Error *local_err = NULL; | |
56e93d26 | 2592 | |
fa53a0e5 | 2593 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2594 | &local_err); |
2595 | if (local_err) { | |
2596 | error_report_err(local_err); | |
56e93d26 | 2597 | } |
56e93d26 | 2598 | } |
ef08fb38 DDAG |
2599 | /* For postcopy we need to check hugepage sizes match */ |
2600 | if (postcopy_advised && | |
2601 | block->page_size != qemu_host_page_size) { | |
2602 | uint64_t remote_page_size = qemu_get_be64(f); | |
2603 | if (remote_page_size != block->page_size) { | |
2604 | error_report("Mismatched RAM page size %s " | |
2605 | "(local) %zd != %" PRId64, | |
2606 | id, block->page_size, | |
2607 | remote_page_size); | |
2608 | ret = -EINVAL; | |
2609 | } | |
2610 | } | |
e3dd7493 DDAG |
2611 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2612 | block->idstr); | |
2613 | } else { | |
56e93d26 JQ |
2614 | error_report("Unknown ramblock \"%s\", cannot " |
2615 | "accept migration", id); | |
2616 | ret = -EINVAL; | |
2617 | } | |
2618 | ||
2619 | total_ram_bytes -= length; | |
2620 | } | |
2621 | break; | |
a776aa15 | 2622 | |
bb890ed5 | 2623 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
2624 | ch = qemu_get_byte(f); |
2625 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2626 | break; | |
a776aa15 | 2627 | |
56e93d26 | 2628 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2629 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2630 | break; | |
56e93d26 | 2631 | |
a776aa15 | 2632 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2633 | len = qemu_get_be32(f); |
2634 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2635 | error_report("Invalid compressed data length: %d", len); | |
2636 | ret = -EINVAL; | |
2637 | break; | |
2638 | } | |
c1bc6626 | 2639 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2640 | break; |
a776aa15 | 2641 | |
56e93d26 | 2642 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2643 | if (load_xbzrle(f, addr, host) < 0) { |
2644 | error_report("Failed to decompress XBZRLE page at " | |
2645 | RAM_ADDR_FMT, addr); | |
2646 | ret = -EINVAL; | |
2647 | break; | |
2648 | } | |
2649 | break; | |
2650 | case RAM_SAVE_FLAG_EOS: | |
2651 | /* normal exit */ | |
2652 | break; | |
2653 | default: | |
2654 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2655 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2656 | } else { |
2657 | error_report("Unknown combination of migration flags: %#x", | |
2658 | flags); | |
2659 | ret = -EINVAL; | |
2660 | } | |
2661 | } | |
2662 | if (!ret) { | |
2663 | ret = qemu_file_get_error(f); | |
2664 | } | |
2665 | } | |
2666 | ||
5533b2e9 | 2667 | wait_for_decompress_done(); |
56e93d26 | 2668 | rcu_read_unlock(); |
55c4446b | 2669 | trace_ram_load_complete(ret, seq_iter); |
56e93d26 JQ |
2670 | return ret; |
2671 | } | |
2672 | ||
2673 | static SaveVMHandlers savevm_ram_handlers = { | |
2674 | .save_live_setup = ram_save_setup, | |
2675 | .save_live_iterate = ram_save_iterate, | |
763c906b | 2676 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2677 | .save_live_complete_precopy = ram_save_complete, |
56e93d26 JQ |
2678 | .save_live_pending = ram_save_pending, |
2679 | .load_state = ram_load, | |
6ad2a215 | 2680 | .cleanup = ram_migration_cleanup, |
56e93d26 JQ |
2681 | }; |
2682 | ||
2683 | void ram_mig_init(void) | |
2684 | { | |
2685 | qemu_mutex_init(&XBZRLE.lock); | |
6f37bb8b | 2686 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 2687 | } |