]>
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 | 29 | #include "cpu.h" |
56e93d26 | 30 | #include <zlib.h> |
4addcd4f | 31 | #include "qapi-event.h" |
f348b6d1 | 32 | #include "qemu/cutils.h" |
56e93d26 JQ |
33 | #include "qemu/bitops.h" |
34 | #include "qemu/bitmap.h" | |
7205c9ec | 35 | #include "qemu/main-loop.h" |
709e3fe8 | 36 | #include "xbzrle.h" |
7b1e1a22 | 37 | #include "ram.h" |
6666c96a | 38 | #include "migration.h" |
f2a8f0a6 | 39 | #include "migration/register.h" |
7b1e1a22 | 40 | #include "migration/misc.h" |
08a0aee1 | 41 | #include "qemu-file.h" |
be07b0ac | 42 | #include "postcopy-ram.h" |
56e93d26 | 43 | #include "migration/page_cache.h" |
56e93d26 | 44 | #include "qemu/error-report.h" |
8acabf69 | 45 | #include "qapi/qmp/qerror.h" |
56e93d26 | 46 | #include "trace.h" |
56e93d26 | 47 | #include "exec/ram_addr.h" |
56e93d26 | 48 | #include "qemu/rcu_queue.h" |
a91246c9 | 49 | #include "migration/colo.h" |
9ac78b61 | 50 | #include "migration/block.h" |
56e93d26 | 51 | |
56e93d26 JQ |
52 | /***********************************************************/ |
53 | /* ram save/restore */ | |
54 | ||
bb890ed5 JQ |
55 | /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it |
56 | * worked for pages that where filled with the same char. We switched | |
57 | * it to only search for the zero value. And to avoid confusion with | |
58 | * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it. | |
59 | */ | |
60 | ||
56e93d26 | 61 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ |
bb890ed5 | 62 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
63 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
64 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
65 | #define RAM_SAVE_FLAG_EOS 0x10 | |
66 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
67 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
68 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
69 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
70 | ||
56e93d26 JQ |
71 | static inline bool is_zero_range(uint8_t *p, uint64_t size) |
72 | { | |
a1febc49 | 73 | return buffer_is_zero(p, size); |
56e93d26 JQ |
74 | } |
75 | ||
9360447d JQ |
76 | XBZRLECacheStats xbzrle_counters; |
77 | ||
56e93d26 JQ |
78 | /* struct contains XBZRLE cache and a static page |
79 | used by the compression */ | |
80 | static struct { | |
81 | /* buffer used for XBZRLE encoding */ | |
82 | uint8_t *encoded_buf; | |
83 | /* buffer for storing page content */ | |
84 | uint8_t *current_buf; | |
85 | /* Cache for XBZRLE, Protected by lock. */ | |
86 | PageCache *cache; | |
87 | QemuMutex lock; | |
c00e0928 JQ |
88 | /* it will store a page full of zeros */ |
89 | uint8_t *zero_target_page; | |
f265e0e4 JQ |
90 | /* buffer used for XBZRLE decoding */ |
91 | uint8_t *decoded_buf; | |
56e93d26 JQ |
92 | } XBZRLE; |
93 | ||
56e93d26 JQ |
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 | |
8acabf69 | 117 | * @errp: set *errp if the check failed, with reason |
56e93d26 | 118 | */ |
8acabf69 | 119 | int64_t xbzrle_cache_resize(int64_t new_size, Error **errp) |
56e93d26 JQ |
120 | { |
121 | PageCache *new_cache; | |
122 | int64_t ret; | |
123 | ||
8acabf69 JQ |
124 | /* Check for truncation */ |
125 | if (new_size != (size_t)new_size) { | |
126 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
127 | "exceeding address space"); | |
128 | return -1; | |
129 | } | |
130 | ||
131 | /* Cache should not be larger than guest ram size */ | |
132 | if (new_size > ram_bytes_total()) { | |
133 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
134 | "exceeds guest ram size"); | |
135 | return -1; | |
136 | } | |
137 | ||
56e93d26 | 138 | if (new_size < TARGET_PAGE_SIZE) { |
8acabf69 JQ |
139 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", |
140 | "is smaller than one target page size"); | |
56e93d26 JQ |
141 | return -1; |
142 | } | |
143 | ||
144 | XBZRLE_cache_lock(); | |
145 | ||
146 | if (XBZRLE.cache != NULL) { | |
147 | if (pow2floor(new_size) == migrate_xbzrle_cache_size()) { | |
148 | goto out_new_size; | |
149 | } | |
150 | new_cache = cache_init(new_size / TARGET_PAGE_SIZE, | |
151 | TARGET_PAGE_SIZE); | |
152 | if (!new_cache) { | |
8acabf69 | 153 | error_setg(errp, "Error creating cache"); |
56e93d26 JQ |
154 | ret = -1; |
155 | goto out; | |
156 | } | |
157 | ||
158 | cache_fini(XBZRLE.cache); | |
159 | XBZRLE.cache = new_cache; | |
160 | } | |
161 | ||
162 | out_new_size: | |
163 | ret = pow2floor(new_size); | |
164 | out: | |
165 | XBZRLE_cache_unlock(); | |
166 | return ret; | |
167 | } | |
168 | ||
ec481c6c JQ |
169 | /* |
170 | * An outstanding page request, on the source, having been received | |
171 | * and queued | |
172 | */ | |
173 | struct RAMSrcPageRequest { | |
174 | RAMBlock *rb; | |
175 | hwaddr offset; | |
176 | hwaddr len; | |
177 | ||
178 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
179 | }; | |
180 | ||
6f37bb8b JQ |
181 | /* State of RAM for migration */ |
182 | struct RAMState { | |
204b88b8 JQ |
183 | /* QEMUFile used for this migration */ |
184 | QEMUFile *f; | |
6f37bb8b JQ |
185 | /* Last block that we have visited searching for dirty pages */ |
186 | RAMBlock *last_seen_block; | |
187 | /* Last block from where we have sent data */ | |
188 | RAMBlock *last_sent_block; | |
269ace29 JQ |
189 | /* Last dirty target page we have sent */ |
190 | ram_addr_t last_page; | |
6f37bb8b JQ |
191 | /* last ram version we have seen */ |
192 | uint32_t last_version; | |
193 | /* We are in the first round */ | |
194 | bool ram_bulk_stage; | |
8d820d6f JQ |
195 | /* How many times we have dirty too many pages */ |
196 | int dirty_rate_high_cnt; | |
f664da80 JQ |
197 | /* these variables are used for bitmap sync */ |
198 | /* last time we did a full bitmap_sync */ | |
199 | int64_t time_last_bitmap_sync; | |
eac74159 | 200 | /* bytes transferred at start_time */ |
c4bdf0cf | 201 | uint64_t bytes_xfer_prev; |
a66cd90c | 202 | /* number of dirty pages since start_time */ |
68908ed6 | 203 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
204 | /* xbzrle misses since the beginning of the period */ |
205 | uint64_t xbzrle_cache_miss_prev; | |
36040d9c JQ |
206 | /* number of iterations at the beginning of period */ |
207 | uint64_t iterations_prev; | |
23b28c3c JQ |
208 | /* Iterations since start */ |
209 | uint64_t iterations; | |
9360447d | 210 | /* number of dirty bits in the bitmap */ |
2dfaf12e PX |
211 | uint64_t migration_dirty_pages; |
212 | /* protects modification of the bitmap */ | |
108cfae0 | 213 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
214 | /* The RAMBlock used in the last src_page_requests */ |
215 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
216 | /* Queue of outstanding page requests from the destination */ |
217 | QemuMutex src_page_req_mutex; | |
218 | QSIMPLEQ_HEAD(src_page_requests, RAMSrcPageRequest) src_page_requests; | |
6f37bb8b JQ |
219 | }; |
220 | typedef struct RAMState RAMState; | |
221 | ||
53518d94 | 222 | static RAMState *ram_state; |
6f37bb8b | 223 | |
9edabd4d | 224 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 225 | { |
53518d94 | 226 | return ram_state->migration_dirty_pages * TARGET_PAGE_SIZE; |
2f4fde93 JQ |
227 | } |
228 | ||
9360447d | 229 | MigrationStats ram_counters; |
96506894 | 230 | |
b8fb8cb7 DDAG |
231 | /* used by the search for pages to send */ |
232 | struct PageSearchStatus { | |
233 | /* Current block being searched */ | |
234 | RAMBlock *block; | |
a935e30f JQ |
235 | /* Current page to search from */ |
236 | unsigned long page; | |
b8fb8cb7 DDAG |
237 | /* Set once we wrap around */ |
238 | bool complete_round; | |
239 | }; | |
240 | typedef struct PageSearchStatus PageSearchStatus; | |
241 | ||
56e93d26 | 242 | struct CompressParam { |
56e93d26 | 243 | bool done; |
90e56fb4 | 244 | bool quit; |
56e93d26 JQ |
245 | QEMUFile *file; |
246 | QemuMutex mutex; | |
247 | QemuCond cond; | |
248 | RAMBlock *block; | |
249 | ram_addr_t offset; | |
250 | }; | |
251 | typedef struct CompressParam CompressParam; | |
252 | ||
253 | struct DecompressParam { | |
73a8912b | 254 | bool done; |
90e56fb4 | 255 | bool quit; |
56e93d26 JQ |
256 | QemuMutex mutex; |
257 | QemuCond cond; | |
258 | void *des; | |
d341d9f3 | 259 | uint8_t *compbuf; |
56e93d26 JQ |
260 | int len; |
261 | }; | |
262 | typedef struct DecompressParam DecompressParam; | |
263 | ||
264 | static CompressParam *comp_param; | |
265 | static QemuThread *compress_threads; | |
266 | /* comp_done_cond is used to wake up the migration thread when | |
267 | * one of the compression threads has finished the compression. | |
268 | * comp_done_lock is used to co-work with comp_done_cond. | |
269 | */ | |
0d9f9a5c LL |
270 | static QemuMutex comp_done_lock; |
271 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
272 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
273 | static const QEMUFileOps empty_ops = { }; | |
274 | ||
56e93d26 JQ |
275 | static DecompressParam *decomp_param; |
276 | static QemuThread *decompress_threads; | |
73a8912b LL |
277 | static QemuMutex decomp_done_lock; |
278 | static QemuCond decomp_done_cond; | |
56e93d26 | 279 | |
a7a9a88f LL |
280 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
281 | ram_addr_t offset); | |
56e93d26 JQ |
282 | |
283 | static void *do_data_compress(void *opaque) | |
284 | { | |
285 | CompressParam *param = opaque; | |
a7a9a88f LL |
286 | RAMBlock *block; |
287 | ram_addr_t offset; | |
56e93d26 | 288 | |
a7a9a88f | 289 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 290 | while (!param->quit) { |
a7a9a88f LL |
291 | if (param->block) { |
292 | block = param->block; | |
293 | offset = param->offset; | |
294 | param->block = NULL; | |
295 | qemu_mutex_unlock(¶m->mutex); | |
296 | ||
297 | do_compress_ram_page(param->file, block, offset); | |
298 | ||
0d9f9a5c | 299 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 300 | param->done = true; |
0d9f9a5c LL |
301 | qemu_cond_signal(&comp_done_cond); |
302 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
303 | |
304 | qemu_mutex_lock(¶m->mutex); | |
305 | } else { | |
56e93d26 JQ |
306 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
307 | } | |
56e93d26 | 308 | } |
a7a9a88f | 309 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
310 | |
311 | return NULL; | |
312 | } | |
313 | ||
314 | static inline void terminate_compression_threads(void) | |
315 | { | |
316 | int idx, thread_count; | |
317 | ||
318 | thread_count = migrate_compress_threads(); | |
3d0684b2 | 319 | |
56e93d26 JQ |
320 | for (idx = 0; idx < thread_count; idx++) { |
321 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 322 | comp_param[idx].quit = true; |
56e93d26 JQ |
323 | qemu_cond_signal(&comp_param[idx].cond); |
324 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
325 | } | |
326 | } | |
327 | ||
f0afa331 | 328 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
329 | { |
330 | int i, thread_count; | |
331 | ||
332 | if (!migrate_use_compression()) { | |
333 | return; | |
334 | } | |
335 | terminate_compression_threads(); | |
336 | thread_count = migrate_compress_threads(); | |
337 | for (i = 0; i < thread_count; i++) { | |
338 | qemu_thread_join(compress_threads + i); | |
339 | qemu_fclose(comp_param[i].file); | |
340 | qemu_mutex_destroy(&comp_param[i].mutex); | |
341 | qemu_cond_destroy(&comp_param[i].cond); | |
342 | } | |
0d9f9a5c LL |
343 | qemu_mutex_destroy(&comp_done_lock); |
344 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
345 | g_free(compress_threads); |
346 | g_free(comp_param); | |
56e93d26 JQ |
347 | compress_threads = NULL; |
348 | comp_param = NULL; | |
56e93d26 JQ |
349 | } |
350 | ||
f0afa331 | 351 | static void compress_threads_save_setup(void) |
56e93d26 JQ |
352 | { |
353 | int i, thread_count; | |
354 | ||
355 | if (!migrate_use_compression()) { | |
356 | return; | |
357 | } | |
56e93d26 JQ |
358 | thread_count = migrate_compress_threads(); |
359 | compress_threads = g_new0(QemuThread, thread_count); | |
360 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
361 | qemu_cond_init(&comp_done_cond); |
362 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 363 | for (i = 0; i < thread_count; i++) { |
e110aa91 C |
364 | /* comp_param[i].file is just used as a dummy buffer to save data, |
365 | * set its ops to empty. | |
56e93d26 JQ |
366 | */ |
367 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
368 | comp_param[i].done = true; | |
90e56fb4 | 369 | comp_param[i].quit = false; |
56e93d26 JQ |
370 | qemu_mutex_init(&comp_param[i].mutex); |
371 | qemu_cond_init(&comp_param[i].cond); | |
372 | qemu_thread_create(compress_threads + i, "compress", | |
373 | do_data_compress, comp_param + i, | |
374 | QEMU_THREAD_JOINABLE); | |
375 | } | |
376 | } | |
377 | ||
f986c3d2 JQ |
378 | /* Multiple fd's */ |
379 | ||
380 | struct MultiFDSendParams { | |
381 | uint8_t id; | |
382 | char *name; | |
383 | QemuThread thread; | |
384 | QemuSemaphore sem; | |
385 | QemuMutex mutex; | |
386 | bool quit; | |
387 | }; | |
388 | typedef struct MultiFDSendParams MultiFDSendParams; | |
389 | ||
390 | struct { | |
391 | MultiFDSendParams *params; | |
392 | /* number of created threads */ | |
393 | int count; | |
394 | } *multifd_send_state; | |
395 | ||
396 | static void terminate_multifd_send_threads(Error *errp) | |
397 | { | |
398 | int i; | |
399 | ||
400 | for (i = 0; i < multifd_send_state->count; i++) { | |
401 | MultiFDSendParams *p = &multifd_send_state->params[i]; | |
402 | ||
403 | qemu_mutex_lock(&p->mutex); | |
404 | p->quit = true; | |
405 | qemu_sem_post(&p->sem); | |
406 | qemu_mutex_unlock(&p->mutex); | |
407 | } | |
408 | } | |
409 | ||
410 | int multifd_save_cleanup(Error **errp) | |
411 | { | |
412 | int i; | |
413 | int ret = 0; | |
414 | ||
415 | if (!migrate_use_multifd()) { | |
416 | return 0; | |
417 | } | |
418 | terminate_multifd_send_threads(NULL); | |
419 | for (i = 0; i < multifd_send_state->count; i++) { | |
420 | MultiFDSendParams *p = &multifd_send_state->params[i]; | |
421 | ||
422 | qemu_thread_join(&p->thread); | |
423 | qemu_mutex_destroy(&p->mutex); | |
424 | qemu_sem_destroy(&p->sem); | |
425 | g_free(p->name); | |
426 | p->name = NULL; | |
427 | } | |
428 | g_free(multifd_send_state->params); | |
429 | multifd_send_state->params = NULL; | |
430 | g_free(multifd_send_state); | |
431 | multifd_send_state = NULL; | |
432 | return ret; | |
433 | } | |
434 | ||
435 | static void *multifd_send_thread(void *opaque) | |
436 | { | |
437 | MultiFDSendParams *p = opaque; | |
438 | ||
439 | while (true) { | |
440 | qemu_mutex_lock(&p->mutex); | |
441 | if (p->quit) { | |
442 | qemu_mutex_unlock(&p->mutex); | |
443 | break; | |
444 | } | |
445 | qemu_mutex_unlock(&p->mutex); | |
446 | qemu_sem_wait(&p->sem); | |
447 | } | |
448 | ||
449 | return NULL; | |
450 | } | |
451 | ||
452 | int multifd_save_setup(void) | |
453 | { | |
454 | int thread_count; | |
455 | uint8_t i; | |
456 | ||
457 | if (!migrate_use_multifd()) { | |
458 | return 0; | |
459 | } | |
460 | thread_count = migrate_multifd_channels(); | |
461 | multifd_send_state = g_malloc0(sizeof(*multifd_send_state)); | |
462 | multifd_send_state->params = g_new0(MultiFDSendParams, thread_count); | |
463 | multifd_send_state->count = 0; | |
464 | for (i = 0; i < thread_count; i++) { | |
465 | MultiFDSendParams *p = &multifd_send_state->params[i]; | |
466 | ||
467 | qemu_mutex_init(&p->mutex); | |
468 | qemu_sem_init(&p->sem, 0); | |
469 | p->quit = false; | |
470 | p->id = i; | |
471 | p->name = g_strdup_printf("multifdsend_%d", i); | |
472 | qemu_thread_create(&p->thread, p->name, multifd_send_thread, p, | |
473 | QEMU_THREAD_JOINABLE); | |
474 | ||
475 | multifd_send_state->count++; | |
476 | } | |
477 | return 0; | |
478 | } | |
479 | ||
480 | struct MultiFDRecvParams { | |
481 | uint8_t id; | |
482 | char *name; | |
483 | QemuThread thread; | |
484 | QemuSemaphore sem; | |
485 | QemuMutex mutex; | |
486 | bool quit; | |
487 | }; | |
488 | typedef struct MultiFDRecvParams MultiFDRecvParams; | |
489 | ||
490 | struct { | |
491 | MultiFDRecvParams *params; | |
492 | /* number of created threads */ | |
493 | int count; | |
494 | } *multifd_recv_state; | |
495 | ||
496 | static void terminate_multifd_recv_threads(Error *errp) | |
497 | { | |
498 | int i; | |
499 | ||
500 | for (i = 0; i < multifd_recv_state->count; i++) { | |
501 | MultiFDRecvParams *p = &multifd_recv_state->params[i]; | |
502 | ||
503 | qemu_mutex_lock(&p->mutex); | |
504 | p->quit = true; | |
505 | qemu_sem_post(&p->sem); | |
506 | qemu_mutex_unlock(&p->mutex); | |
507 | } | |
508 | } | |
509 | ||
510 | int multifd_load_cleanup(Error **errp) | |
511 | { | |
512 | int i; | |
513 | int ret = 0; | |
514 | ||
515 | if (!migrate_use_multifd()) { | |
516 | return 0; | |
517 | } | |
518 | terminate_multifd_recv_threads(NULL); | |
519 | for (i = 0; i < multifd_recv_state->count; i++) { | |
520 | MultiFDRecvParams *p = &multifd_recv_state->params[i]; | |
521 | ||
522 | qemu_thread_join(&p->thread); | |
523 | qemu_mutex_destroy(&p->mutex); | |
524 | qemu_sem_destroy(&p->sem); | |
525 | g_free(p->name); | |
526 | p->name = NULL; | |
527 | } | |
528 | g_free(multifd_recv_state->params); | |
529 | multifd_recv_state->params = NULL; | |
530 | g_free(multifd_recv_state); | |
531 | multifd_recv_state = NULL; | |
532 | ||
533 | return ret; | |
534 | } | |
535 | ||
536 | static void *multifd_recv_thread(void *opaque) | |
537 | { | |
538 | MultiFDRecvParams *p = opaque; | |
539 | ||
540 | while (true) { | |
541 | qemu_mutex_lock(&p->mutex); | |
542 | if (p->quit) { | |
543 | qemu_mutex_unlock(&p->mutex); | |
544 | break; | |
545 | } | |
546 | qemu_mutex_unlock(&p->mutex); | |
547 | qemu_sem_wait(&p->sem); | |
548 | } | |
549 | ||
550 | return NULL; | |
551 | } | |
552 | ||
553 | int multifd_load_setup(void) | |
554 | { | |
555 | int thread_count; | |
556 | uint8_t i; | |
557 | ||
558 | if (!migrate_use_multifd()) { | |
559 | return 0; | |
560 | } | |
561 | thread_count = migrate_multifd_channels(); | |
562 | multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state)); | |
563 | multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count); | |
564 | multifd_recv_state->count = 0; | |
565 | for (i = 0; i < thread_count; i++) { | |
566 | MultiFDRecvParams *p = &multifd_recv_state->params[i]; | |
567 | ||
568 | qemu_mutex_init(&p->mutex); | |
569 | qemu_sem_init(&p->sem, 0); | |
570 | p->quit = false; | |
571 | p->id = i; | |
572 | p->name = g_strdup_printf("multifdrecv_%d", i); | |
573 | qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p, | |
574 | QEMU_THREAD_JOINABLE); | |
575 | multifd_recv_state->count++; | |
576 | } | |
577 | return 0; | |
578 | } | |
579 | ||
56e93d26 | 580 | /** |
3d0684b2 | 581 | * save_page_header: write page header to wire |
56e93d26 JQ |
582 | * |
583 | * If this is the 1st block, it also writes the block identification | |
584 | * | |
3d0684b2 | 585 | * Returns the number of bytes written |
56e93d26 JQ |
586 | * |
587 | * @f: QEMUFile where to send the data | |
588 | * @block: block that contains the page we want to send | |
589 | * @offset: offset inside the block for the page | |
590 | * in the lower bits, it contains flags | |
591 | */ | |
2bf3aa85 JQ |
592 | static size_t save_page_header(RAMState *rs, QEMUFile *f, RAMBlock *block, |
593 | ram_addr_t offset) | |
56e93d26 | 594 | { |
9f5f380b | 595 | size_t size, len; |
56e93d26 | 596 | |
24795694 JQ |
597 | if (block == rs->last_sent_block) { |
598 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
599 | } | |
2bf3aa85 | 600 | qemu_put_be64(f, offset); |
56e93d26 JQ |
601 | size = 8; |
602 | ||
603 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 604 | len = strlen(block->idstr); |
2bf3aa85 JQ |
605 | qemu_put_byte(f, len); |
606 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 607 | size += 1 + len; |
24795694 | 608 | rs->last_sent_block = block; |
56e93d26 JQ |
609 | } |
610 | return size; | |
611 | } | |
612 | ||
3d0684b2 JQ |
613 | /** |
614 | * mig_throttle_guest_down: throotle down the guest | |
615 | * | |
616 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
617 | * writes. If guest dirty memory rate is reduced below the rate at | |
618 | * which we can transfer pages to the destination then we should be | |
619 | * able to complete migration. Some workloads dirty memory way too | |
620 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
621 | */ |
622 | static void mig_throttle_guest_down(void) | |
623 | { | |
624 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
625 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
626 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
627 | |
628 | /* We have not started throttling yet. Let's start it. */ | |
629 | if (!cpu_throttle_active()) { | |
630 | cpu_throttle_set(pct_initial); | |
631 | } else { | |
632 | /* Throttling already on, just increase the rate */ | |
633 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
634 | } | |
635 | } | |
636 | ||
3d0684b2 JQ |
637 | /** |
638 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
639 | * | |
6f37bb8b | 640 | * @rs: current RAM state |
3d0684b2 JQ |
641 | * @current_addr: address for the zero page |
642 | * | |
643 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
644 | * The important thing is that a stale (not-yet-0'd) page be replaced |
645 | * by the new data. | |
646 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 647 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 648 | */ |
6f37bb8b | 649 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 650 | { |
6f37bb8b | 651 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
652 | return; |
653 | } | |
654 | ||
655 | /* We don't care if this fails to allocate a new cache page | |
656 | * as long as it updated an old one */ | |
c00e0928 | 657 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 658 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
659 | } |
660 | ||
661 | #define ENCODING_FLAG_XBZRLE 0x1 | |
662 | ||
663 | /** | |
664 | * save_xbzrle_page: compress and send current page | |
665 | * | |
666 | * Returns: 1 means that we wrote the page | |
667 | * 0 means that page is identical to the one already sent | |
668 | * -1 means that xbzrle would be longer than normal | |
669 | * | |
5a987738 | 670 | * @rs: current RAM state |
3d0684b2 JQ |
671 | * @current_data: pointer to the address of the page contents |
672 | * @current_addr: addr of the page | |
56e93d26 JQ |
673 | * @block: block that contains the page we want to send |
674 | * @offset: offset inside the block for the page | |
675 | * @last_stage: if we are at the completion stage | |
56e93d26 | 676 | */ |
204b88b8 | 677 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 678 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 679 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
680 | { |
681 | int encoded_len = 0, bytes_xbzrle; | |
682 | uint8_t *prev_cached_page; | |
683 | ||
9360447d JQ |
684 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
685 | ram_counters.dirty_sync_count)) { | |
686 | xbzrle_counters.cache_miss++; | |
56e93d26 JQ |
687 | if (!last_stage) { |
688 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
9360447d | 689 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
690 | return -1; |
691 | } else { | |
692 | /* update *current_data when the page has been | |
693 | inserted into cache */ | |
694 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
695 | } | |
696 | } | |
697 | return -1; | |
698 | } | |
699 | ||
700 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
701 | ||
702 | /* save current buffer into memory */ | |
703 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
704 | ||
705 | /* XBZRLE encoding (if there is no overflow) */ | |
706 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
707 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
708 | TARGET_PAGE_SIZE); | |
709 | if (encoded_len == 0) { | |
55c4446b | 710 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
711 | return 0; |
712 | } else if (encoded_len == -1) { | |
55c4446b | 713 | trace_save_xbzrle_page_overflow(); |
9360447d | 714 | xbzrle_counters.overflow++; |
56e93d26 JQ |
715 | /* update data in the cache */ |
716 | if (!last_stage) { | |
717 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
718 | *current_data = prev_cached_page; | |
719 | } | |
720 | return -1; | |
721 | } | |
722 | ||
723 | /* we need to update the data in the cache, in order to get the same data */ | |
724 | if (!last_stage) { | |
725 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
726 | } | |
727 | ||
728 | /* Send XBZRLE based compressed page */ | |
2bf3aa85 | 729 | bytes_xbzrle = save_page_header(rs, rs->f, block, |
204b88b8 JQ |
730 | offset | RAM_SAVE_FLAG_XBZRLE); |
731 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
732 | qemu_put_be16(rs->f, encoded_len); | |
733 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 734 | bytes_xbzrle += encoded_len + 1 + 2; |
9360447d JQ |
735 | xbzrle_counters.pages++; |
736 | xbzrle_counters.bytes += bytes_xbzrle; | |
737 | ram_counters.transferred += bytes_xbzrle; | |
56e93d26 JQ |
738 | |
739 | return 1; | |
740 | } | |
741 | ||
3d0684b2 JQ |
742 | /** |
743 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 744 | * |
3d0684b2 JQ |
745 | * Called with rcu_read_lock() to protect migration_bitmap |
746 | * | |
747 | * Returns the byte offset within memory region of the start of a dirty page | |
748 | * | |
6f37bb8b | 749 | * @rs: current RAM state |
3d0684b2 | 750 | * @rb: RAMBlock where to search for dirty pages |
a935e30f | 751 | * @start: page where we start the search |
f3f491fc | 752 | */ |
56e93d26 | 753 | static inline |
a935e30f | 754 | unsigned long migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
f20e2865 | 755 | unsigned long start) |
56e93d26 | 756 | { |
6b6712ef JQ |
757 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
758 | unsigned long *bitmap = rb->bmap; | |
56e93d26 JQ |
759 | unsigned long next; |
760 | ||
6b6712ef JQ |
761 | if (rs->ram_bulk_stage && start > 0) { |
762 | next = start + 1; | |
56e93d26 | 763 | } else { |
6b6712ef | 764 | next = find_next_bit(bitmap, size, start); |
56e93d26 JQ |
765 | } |
766 | ||
6b6712ef | 767 | return next; |
56e93d26 JQ |
768 | } |
769 | ||
06b10688 | 770 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
771 | RAMBlock *rb, |
772 | unsigned long page) | |
a82d593b DDAG |
773 | { |
774 | bool ret; | |
a82d593b | 775 | |
6b6712ef | 776 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b DDAG |
777 | |
778 | if (ret) { | |
0d8ec885 | 779 | rs->migration_dirty_pages--; |
a82d593b DDAG |
780 | } |
781 | return ret; | |
782 | } | |
783 | ||
15440dd5 JQ |
784 | static void migration_bitmap_sync_range(RAMState *rs, RAMBlock *rb, |
785 | ram_addr_t start, ram_addr_t length) | |
56e93d26 | 786 | { |
0d8ec885 | 787 | rs->migration_dirty_pages += |
6b6712ef | 788 | cpu_physical_memory_sync_dirty_bitmap(rb, start, length, |
0d8ec885 | 789 | &rs->num_dirty_pages_period); |
56e93d26 JQ |
790 | } |
791 | ||
3d0684b2 JQ |
792 | /** |
793 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
794 | * | |
795 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
796 | * | |
797 | * For VMs with just normal pages this is equivalent to the host page | |
798 | * size. If it's got some huge pages then it's the OR of all the | |
799 | * different page sizes. | |
e8ca1db2 DDAG |
800 | */ |
801 | uint64_t ram_pagesize_summary(void) | |
802 | { | |
803 | RAMBlock *block; | |
804 | uint64_t summary = 0; | |
805 | ||
99e15582 | 806 | RAMBLOCK_FOREACH(block) { |
e8ca1db2 DDAG |
807 | summary |= block->page_size; |
808 | } | |
809 | ||
810 | return summary; | |
811 | } | |
812 | ||
8d820d6f | 813 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
814 | { |
815 | RAMBlock *block; | |
56e93d26 | 816 | int64_t end_time; |
c4bdf0cf | 817 | uint64_t bytes_xfer_now; |
56e93d26 | 818 | |
9360447d | 819 | ram_counters.dirty_sync_count++; |
56e93d26 | 820 | |
f664da80 JQ |
821 | if (!rs->time_last_bitmap_sync) { |
822 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
823 | } |
824 | ||
825 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 826 | memory_global_dirty_log_sync(); |
56e93d26 | 827 | |
108cfae0 | 828 | qemu_mutex_lock(&rs->bitmap_mutex); |
56e93d26 | 829 | rcu_read_lock(); |
99e15582 | 830 | RAMBLOCK_FOREACH(block) { |
15440dd5 | 831 | migration_bitmap_sync_range(rs, block, 0, block->used_length); |
56e93d26 JQ |
832 | } |
833 | rcu_read_unlock(); | |
108cfae0 | 834 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 835 | |
a66cd90c | 836 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 837 | |
56e93d26 JQ |
838 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
839 | ||
840 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 841 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
d693c6f1 | 842 | /* calculate period counters */ |
9360447d | 843 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 |
d693c6f1 | 844 | / (end_time - rs->time_last_bitmap_sync); |
9360447d | 845 | bytes_xfer_now = ram_counters.transferred; |
d693c6f1 | 846 | |
9ac78b61 PL |
847 | /* During block migration the auto-converge logic incorrectly detects |
848 | * that ram migration makes no progress. Avoid this by disabling the | |
849 | * throttling logic during the bulk phase of block migration. */ | |
850 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
56e93d26 JQ |
851 | /* The following detection logic can be refined later. For now: |
852 | Check to see if the dirtied bytes is 50% more than the approx. | |
853 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
854 | were in this routine. If that happens twice, start or increase |
855 | throttling */ | |
070afca2 | 856 | |
d693c6f1 | 857 | if ((rs->num_dirty_pages_period * TARGET_PAGE_SIZE > |
eac74159 | 858 | (bytes_xfer_now - rs->bytes_xfer_prev) / 2) && |
b4a3c64b | 859 | (++rs->dirty_rate_high_cnt >= 2)) { |
56e93d26 | 860 | trace_migration_throttle(); |
8d820d6f | 861 | rs->dirty_rate_high_cnt = 0; |
070afca2 | 862 | mig_throttle_guest_down(); |
d693c6f1 | 863 | } |
56e93d26 | 864 | } |
070afca2 | 865 | |
56e93d26 | 866 | if (migrate_use_xbzrle()) { |
23b28c3c | 867 | if (rs->iterations_prev != rs->iterations) { |
9360447d JQ |
868 | xbzrle_counters.cache_miss_rate = |
869 | (double)(xbzrle_counters.cache_miss - | |
b5833fde | 870 | rs->xbzrle_cache_miss_prev) / |
23b28c3c | 871 | (rs->iterations - rs->iterations_prev); |
56e93d26 | 872 | } |
23b28c3c | 873 | rs->iterations_prev = rs->iterations; |
9360447d | 874 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
56e93d26 | 875 | } |
d693c6f1 FF |
876 | |
877 | /* reset period counters */ | |
f664da80 | 878 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 879 | rs->num_dirty_pages_period = 0; |
d2a4d85a | 880 | rs->bytes_xfer_prev = bytes_xfer_now; |
56e93d26 | 881 | } |
4addcd4f | 882 | if (migrate_use_events()) { |
9360447d | 883 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count, NULL); |
4addcd4f | 884 | } |
56e93d26 JQ |
885 | } |
886 | ||
887 | /** | |
3d0684b2 | 888 | * save_zero_page: send the zero page to the stream |
56e93d26 | 889 | * |
3d0684b2 | 890 | * Returns the number of pages written. |
56e93d26 | 891 | * |
f7ccd61b | 892 | * @rs: current RAM state |
56e93d26 JQ |
893 | * @block: block that contains the page we want to send |
894 | * @offset: offset inside the block for the page | |
895 | * @p: pointer to the page | |
56e93d26 | 896 | */ |
ce25d337 JQ |
897 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, |
898 | uint8_t *p) | |
56e93d26 JQ |
899 | { |
900 | int pages = -1; | |
901 | ||
902 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
9360447d JQ |
903 | ram_counters.duplicate++; |
904 | ram_counters.transferred += | |
bb890ed5 | 905 | save_page_header(rs, rs->f, block, offset | RAM_SAVE_FLAG_ZERO); |
ce25d337 | 906 | qemu_put_byte(rs->f, 0); |
9360447d | 907 | ram_counters.transferred += 1; |
56e93d26 JQ |
908 | pages = 1; |
909 | } | |
910 | ||
911 | return pages; | |
912 | } | |
913 | ||
5727309d | 914 | static void ram_release_pages(const char *rbname, uint64_t offset, int pages) |
53f09a10 | 915 | { |
5727309d | 916 | if (!migrate_release_ram() || !migration_in_postcopy()) { |
53f09a10 PB |
917 | return; |
918 | } | |
919 | ||
aaa2064c | 920 | ram_discard_range(rbname, offset, pages << TARGET_PAGE_BITS); |
53f09a10 PB |
921 | } |
922 | ||
56e93d26 | 923 | /** |
3d0684b2 | 924 | * ram_save_page: send the given page to the stream |
56e93d26 | 925 | * |
3d0684b2 | 926 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
927 | * < 0 - error |
928 | * >=0 - Number of pages written - this might legally be 0 | |
929 | * if xbzrle noticed the page was the same. | |
56e93d26 | 930 | * |
6f37bb8b | 931 | * @rs: current RAM state |
56e93d26 JQ |
932 | * @block: block that contains the page we want to send |
933 | * @offset: offset inside the block for the page | |
934 | * @last_stage: if we are at the completion stage | |
56e93d26 | 935 | */ |
a0a8aa14 | 936 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
937 | { |
938 | int pages = -1; | |
939 | uint64_t bytes_xmit; | |
940 | ram_addr_t current_addr; | |
56e93d26 JQ |
941 | uint8_t *p; |
942 | int ret; | |
943 | bool send_async = true; | |
a08f6890 | 944 | RAMBlock *block = pss->block; |
a935e30f | 945 | ram_addr_t offset = pss->page << TARGET_PAGE_BITS; |
56e93d26 | 946 | |
2f68e399 | 947 | p = block->host + offset; |
1db9d8e5 | 948 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 JQ |
949 | |
950 | /* In doubt sent page as normal */ | |
951 | bytes_xmit = 0; | |
ce25d337 | 952 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
953 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
954 | if (bytes_xmit) { | |
9360447d | 955 | ram_counters.transferred += bytes_xmit; |
56e93d26 JQ |
956 | pages = 1; |
957 | } | |
958 | ||
959 | XBZRLE_cache_lock(); | |
960 | ||
961 | current_addr = block->offset + offset; | |
962 | ||
56e93d26 JQ |
963 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
964 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
965 | if (bytes_xmit > 0) { | |
9360447d | 966 | ram_counters.normal++; |
56e93d26 | 967 | } else if (bytes_xmit == 0) { |
9360447d | 968 | ram_counters.duplicate++; |
56e93d26 JQ |
969 | } |
970 | } | |
971 | } else { | |
ce25d337 | 972 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 JQ |
973 | if (pages > 0) { |
974 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
975 | * page would be stale | |
976 | */ | |
6f37bb8b | 977 | xbzrle_cache_zero_page(rs, current_addr); |
a935e30f | 978 | ram_release_pages(block->idstr, offset, pages); |
6f37bb8b | 979 | } else if (!rs->ram_bulk_stage && |
5727309d | 980 | !migration_in_postcopy() && migrate_use_xbzrle()) { |
204b88b8 | 981 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
072c2511 | 982 | offset, last_stage); |
56e93d26 JQ |
983 | if (!last_stage) { |
984 | /* Can't send this cached data async, since the cache page | |
985 | * might get updated before it gets to the wire | |
986 | */ | |
987 | send_async = false; | |
988 | } | |
989 | } | |
990 | } | |
991 | ||
992 | /* XBZRLE overflow or normal page */ | |
993 | if (pages == -1) { | |
9360447d JQ |
994 | ram_counters.transferred += |
995 | save_page_header(rs, rs->f, block, offset | RAM_SAVE_FLAG_PAGE); | |
56e93d26 | 996 | if (send_async) { |
ce25d337 | 997 | qemu_put_buffer_async(rs->f, p, TARGET_PAGE_SIZE, |
53f09a10 | 998 | migrate_release_ram() & |
5727309d | 999 | migration_in_postcopy()); |
56e93d26 | 1000 | } else { |
ce25d337 | 1001 | qemu_put_buffer(rs->f, p, TARGET_PAGE_SIZE); |
56e93d26 | 1002 | } |
9360447d | 1003 | ram_counters.transferred += TARGET_PAGE_SIZE; |
56e93d26 | 1004 | pages = 1; |
9360447d | 1005 | ram_counters.normal++; |
56e93d26 JQ |
1006 | } |
1007 | ||
1008 | XBZRLE_cache_unlock(); | |
1009 | ||
1010 | return pages; | |
1011 | } | |
1012 | ||
a7a9a88f LL |
1013 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
1014 | ram_addr_t offset) | |
56e93d26 | 1015 | { |
53518d94 | 1016 | RAMState *rs = ram_state; |
56e93d26 | 1017 | int bytes_sent, blen; |
a7a9a88f | 1018 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 | 1019 | |
2bf3aa85 | 1020 | bytes_sent = save_page_header(rs, f, block, offset | |
56e93d26 | 1021 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
a7a9a88f | 1022 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, |
56e93d26 | 1023 | migrate_compress_level()); |
b3be2896 LL |
1024 | if (blen < 0) { |
1025 | bytes_sent = 0; | |
1026 | qemu_file_set_error(migrate_get_current()->to_dst_file, blen); | |
1027 | error_report("compressed data failed!"); | |
1028 | } else { | |
1029 | bytes_sent += blen; | |
5727309d | 1030 | ram_release_pages(block->idstr, offset & TARGET_PAGE_MASK, 1); |
b3be2896 | 1031 | } |
56e93d26 JQ |
1032 | |
1033 | return bytes_sent; | |
1034 | } | |
1035 | ||
ce25d337 | 1036 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
1037 | { |
1038 | int idx, len, thread_count; | |
1039 | ||
1040 | if (!migrate_use_compression()) { | |
1041 | return; | |
1042 | } | |
1043 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1044 | |
0d9f9a5c | 1045 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1046 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1047 | while (!comp_param[idx].done) { |
0d9f9a5c | 1048 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1049 | } |
a7a9a88f | 1050 | } |
0d9f9a5c | 1051 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1052 | |
1053 | for (idx = 0; idx < thread_count; idx++) { | |
1054 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1055 | if (!comp_param[idx].quit) { |
ce25d337 | 1056 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
9360447d | 1057 | ram_counters.transferred += len; |
56e93d26 | 1058 | } |
a7a9a88f | 1059 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1060 | } |
1061 | } | |
1062 | ||
1063 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1064 | ram_addr_t offset) | |
1065 | { | |
1066 | param->block = block; | |
1067 | param->offset = offset; | |
1068 | } | |
1069 | ||
ce25d337 JQ |
1070 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
1071 | ram_addr_t offset) | |
56e93d26 JQ |
1072 | { |
1073 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1074 | ||
1075 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1076 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 JQ |
1077 | while (true) { |
1078 | for (idx = 0; idx < thread_count; idx++) { | |
1079 | if (comp_param[idx].done) { | |
a7a9a88f | 1080 | comp_param[idx].done = false; |
ce25d337 | 1081 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
a7a9a88f | 1082 | qemu_mutex_lock(&comp_param[idx].mutex); |
56e93d26 | 1083 | set_compress_params(&comp_param[idx], block, offset); |
a7a9a88f LL |
1084 | qemu_cond_signal(&comp_param[idx].cond); |
1085 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
56e93d26 | 1086 | pages = 1; |
9360447d JQ |
1087 | ram_counters.normal++; |
1088 | ram_counters.transferred += bytes_xmit; | |
56e93d26 JQ |
1089 | break; |
1090 | } | |
1091 | } | |
1092 | if (pages > 0) { | |
1093 | break; | |
1094 | } else { | |
0d9f9a5c | 1095 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 JQ |
1096 | } |
1097 | } | |
0d9f9a5c | 1098 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1099 | |
1100 | return pages; | |
1101 | } | |
1102 | ||
1103 | /** | |
1104 | * ram_save_compressed_page: compress the given page and send it to the stream | |
1105 | * | |
3d0684b2 | 1106 | * Returns the number of pages written. |
56e93d26 | 1107 | * |
6f37bb8b | 1108 | * @rs: current RAM state |
56e93d26 JQ |
1109 | * @block: block that contains the page we want to send |
1110 | * @offset: offset inside the block for the page | |
1111 | * @last_stage: if we are at the completion stage | |
56e93d26 | 1112 | */ |
a0a8aa14 JQ |
1113 | static int ram_save_compressed_page(RAMState *rs, PageSearchStatus *pss, |
1114 | bool last_stage) | |
56e93d26 JQ |
1115 | { |
1116 | int pages = -1; | |
fc50438e | 1117 | uint64_t bytes_xmit = 0; |
56e93d26 | 1118 | uint8_t *p; |
fc50438e | 1119 | int ret, blen; |
a08f6890 | 1120 | RAMBlock *block = pss->block; |
a935e30f | 1121 | ram_addr_t offset = pss->page << TARGET_PAGE_BITS; |
56e93d26 | 1122 | |
2f68e399 | 1123 | p = block->host + offset; |
56e93d26 | 1124 | |
ce25d337 | 1125 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
1126 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
1127 | if (bytes_xmit) { | |
9360447d | 1128 | ram_counters.transferred += bytes_xmit; |
56e93d26 JQ |
1129 | pages = 1; |
1130 | } | |
56e93d26 JQ |
1131 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
1132 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
1133 | if (bytes_xmit > 0) { | |
9360447d | 1134 | ram_counters.normal++; |
56e93d26 | 1135 | } else if (bytes_xmit == 0) { |
9360447d | 1136 | ram_counters.duplicate++; |
56e93d26 JQ |
1137 | } |
1138 | } | |
1139 | } else { | |
1140 | /* When starting the process of a new block, the first page of | |
1141 | * the block should be sent out before other pages in the same | |
1142 | * block, and all the pages in last block should have been sent | |
1143 | * out, keeping this order is important, because the 'cont' flag | |
1144 | * is used to avoid resending the block name. | |
1145 | */ | |
6f37bb8b | 1146 | if (block != rs->last_sent_block) { |
ce25d337 JQ |
1147 | flush_compressed_data(rs); |
1148 | pages = save_zero_page(rs, block, offset, p); | |
56e93d26 | 1149 | if (pages == -1) { |
fc50438e | 1150 | /* Make sure the first page is sent out before other pages */ |
2bf3aa85 | 1151 | bytes_xmit = save_page_header(rs, rs->f, block, offset | |
fc50438e | 1152 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
ce25d337 | 1153 | blen = qemu_put_compression_data(rs->f, p, TARGET_PAGE_SIZE, |
fc50438e LL |
1154 | migrate_compress_level()); |
1155 | if (blen > 0) { | |
9360447d JQ |
1156 | ram_counters.transferred += bytes_xmit + blen; |
1157 | ram_counters.normal++; | |
b3be2896 | 1158 | pages = 1; |
fc50438e | 1159 | } else { |
ce25d337 | 1160 | qemu_file_set_error(rs->f, blen); |
fc50438e | 1161 | error_report("compressed data failed!"); |
b3be2896 | 1162 | } |
56e93d26 | 1163 | } |
53f09a10 | 1164 | if (pages > 0) { |
a935e30f | 1165 | ram_release_pages(block->idstr, offset, pages); |
53f09a10 | 1166 | } |
56e93d26 | 1167 | } else { |
ce25d337 | 1168 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 | 1169 | if (pages == -1) { |
ce25d337 | 1170 | pages = compress_page_with_multi_thread(rs, block, offset); |
53f09a10 | 1171 | } else { |
a935e30f | 1172 | ram_release_pages(block->idstr, offset, pages); |
56e93d26 JQ |
1173 | } |
1174 | } | |
1175 | } | |
1176 | ||
1177 | return pages; | |
1178 | } | |
1179 | ||
3d0684b2 JQ |
1180 | /** |
1181 | * find_dirty_block: find the next dirty page and update any state | |
1182 | * associated with the search process. | |
b9e60928 | 1183 | * |
3d0684b2 | 1184 | * Returns if a page is found |
b9e60928 | 1185 | * |
6f37bb8b | 1186 | * @rs: current RAM state |
3d0684b2 JQ |
1187 | * @pss: data about the state of the current dirty page scan |
1188 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1189 | */ |
f20e2865 | 1190 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1191 | { |
f20e2865 | 1192 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
6f37bb8b | 1193 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1194 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1195 | /* |
1196 | * We've been once around the RAM and haven't found anything. | |
1197 | * Give up. | |
1198 | */ | |
1199 | *again = false; | |
1200 | return false; | |
1201 | } | |
a935e30f | 1202 | if ((pss->page << TARGET_PAGE_BITS) >= pss->block->used_length) { |
b9e60928 | 1203 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1204 | pss->page = 0; |
b9e60928 DDAG |
1205 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1206 | if (!pss->block) { | |
1207 | /* Hit the end of the list */ | |
1208 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1209 | /* Flag that we've looped */ | |
1210 | pss->complete_round = true; | |
6f37bb8b | 1211 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1212 | if (migrate_use_xbzrle()) { |
1213 | /* If xbzrle is on, stop using the data compression at this | |
1214 | * point. In theory, xbzrle can do better than compression. | |
1215 | */ | |
ce25d337 | 1216 | flush_compressed_data(rs); |
b9e60928 DDAG |
1217 | } |
1218 | } | |
1219 | /* Didn't find anything this time, but try again on the new block */ | |
1220 | *again = true; | |
1221 | return false; | |
1222 | } else { | |
1223 | /* Can go around again, but... */ | |
1224 | *again = true; | |
1225 | /* We've found something so probably don't need to */ | |
1226 | return true; | |
1227 | } | |
1228 | } | |
1229 | ||
3d0684b2 JQ |
1230 | /** |
1231 | * unqueue_page: gets a page of the queue | |
1232 | * | |
a82d593b | 1233 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1234 | * |
3d0684b2 JQ |
1235 | * Returns the block of the page (or NULL if none available) |
1236 | * | |
ec481c6c | 1237 | * @rs: current RAM state |
3d0684b2 | 1238 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1239 | */ |
f20e2865 | 1240 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b DDAG |
1241 | { |
1242 | RAMBlock *block = NULL; | |
1243 | ||
ec481c6c JQ |
1244 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1245 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
1246 | struct RAMSrcPageRequest *entry = | |
1247 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1248 | block = entry->rb; |
1249 | *offset = entry->offset; | |
a82d593b DDAG |
1250 | |
1251 | if (entry->len > TARGET_PAGE_SIZE) { | |
1252 | entry->len -= TARGET_PAGE_SIZE; | |
1253 | entry->offset += TARGET_PAGE_SIZE; | |
1254 | } else { | |
1255 | memory_region_unref(block->mr); | |
ec481c6c | 1256 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b DDAG |
1257 | g_free(entry); |
1258 | } | |
1259 | } | |
ec481c6c | 1260 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
a82d593b DDAG |
1261 | |
1262 | return block; | |
1263 | } | |
1264 | ||
3d0684b2 JQ |
1265 | /** |
1266 | * get_queued_page: unqueue a page from the postocpy requests | |
1267 | * | |
1268 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1269 | * |
3d0684b2 | 1270 | * Returns if a queued page is found |
a82d593b | 1271 | * |
6f37bb8b | 1272 | * @rs: current RAM state |
3d0684b2 | 1273 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1274 | */ |
f20e2865 | 1275 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1276 | { |
1277 | RAMBlock *block; | |
1278 | ram_addr_t offset; | |
1279 | bool dirty; | |
1280 | ||
1281 | do { | |
f20e2865 | 1282 | block = unqueue_page(rs, &offset); |
a82d593b DDAG |
1283 | /* |
1284 | * We're sending this page, and since it's postcopy nothing else | |
1285 | * will dirty it, and we must make sure it doesn't get sent again | |
1286 | * even if this queue request was received after the background | |
1287 | * search already sent it. | |
1288 | */ | |
1289 | if (block) { | |
f20e2865 JQ |
1290 | unsigned long page; |
1291 | ||
6b6712ef JQ |
1292 | page = offset >> TARGET_PAGE_BITS; |
1293 | dirty = test_bit(page, block->bmap); | |
a82d593b | 1294 | if (!dirty) { |
06b10688 | 1295 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, |
6b6712ef | 1296 | page, test_bit(page, block->unsentmap)); |
a82d593b | 1297 | } else { |
f20e2865 | 1298 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); |
a82d593b DDAG |
1299 | } |
1300 | } | |
1301 | ||
1302 | } while (block && !dirty); | |
1303 | ||
1304 | if (block) { | |
1305 | /* | |
1306 | * As soon as we start servicing pages out of order, then we have | |
1307 | * to kill the bulk stage, since the bulk stage assumes | |
1308 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1309 | * dirty, that's no longer true. | |
1310 | */ | |
6f37bb8b | 1311 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1312 | |
1313 | /* | |
1314 | * We want the background search to continue from the queued page | |
1315 | * since the guest is likely to want other pages near to the page | |
1316 | * it just requested. | |
1317 | */ | |
1318 | pss->block = block; | |
a935e30f | 1319 | pss->page = offset >> TARGET_PAGE_BITS; |
a82d593b DDAG |
1320 | } |
1321 | ||
1322 | return !!block; | |
1323 | } | |
1324 | ||
6c595cde | 1325 | /** |
5e58f968 JQ |
1326 | * migration_page_queue_free: drop any remaining pages in the ram |
1327 | * request queue | |
6c595cde | 1328 | * |
3d0684b2 JQ |
1329 | * It should be empty at the end anyway, but in error cases there may |
1330 | * be some left. in case that there is any page left, we drop it. | |
1331 | * | |
6c595cde | 1332 | */ |
83c13382 | 1333 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 1334 | { |
ec481c6c | 1335 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
1336 | /* This queue generally should be empty - but in the case of a failed |
1337 | * migration might have some droppings in. | |
1338 | */ | |
1339 | rcu_read_lock(); | |
ec481c6c | 1340 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1341 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1342 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1343 | g_free(mspr); |
1344 | } | |
1345 | rcu_read_unlock(); | |
1346 | } | |
1347 | ||
1348 | /** | |
3d0684b2 JQ |
1349 | * ram_save_queue_pages: queue the page for transmission |
1350 | * | |
1351 | * A request from postcopy destination for example. | |
1352 | * | |
1353 | * Returns zero on success or negative on error | |
1354 | * | |
3d0684b2 JQ |
1355 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1356 | * same that last one. | |
1357 | * @start: starting address from the start of the RAMBlock | |
1358 | * @len: length (in bytes) to send | |
6c595cde | 1359 | */ |
96506894 | 1360 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1361 | { |
1362 | RAMBlock *ramblock; | |
53518d94 | 1363 | RAMState *rs = ram_state; |
6c595cde | 1364 | |
9360447d | 1365 | ram_counters.postcopy_requests++; |
6c595cde DDAG |
1366 | rcu_read_lock(); |
1367 | if (!rbname) { | |
1368 | /* Reuse last RAMBlock */ | |
68a098f3 | 1369 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1370 | |
1371 | if (!ramblock) { | |
1372 | /* | |
1373 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1374 | * it's the 1st request. | |
1375 | */ | |
1376 | error_report("ram_save_queue_pages no previous block"); | |
1377 | goto err; | |
1378 | } | |
1379 | } else { | |
1380 | ramblock = qemu_ram_block_by_name(rbname); | |
1381 | ||
1382 | if (!ramblock) { | |
1383 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1384 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1385 | goto err; | |
1386 | } | |
68a098f3 | 1387 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1388 | } |
1389 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1390 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1391 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1392 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1393 | __func__, start, len, ramblock->used_length); |
1394 | goto err; | |
1395 | } | |
1396 | ||
ec481c6c JQ |
1397 | struct RAMSrcPageRequest *new_entry = |
1398 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
1399 | new_entry->rb = ramblock; |
1400 | new_entry->offset = start; | |
1401 | new_entry->len = len; | |
1402 | ||
1403 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
1404 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1405 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
1406 | qemu_mutex_unlock(&rs->src_page_req_mutex); | |
6c595cde DDAG |
1407 | rcu_read_unlock(); |
1408 | ||
1409 | return 0; | |
1410 | ||
1411 | err: | |
1412 | rcu_read_unlock(); | |
1413 | return -1; | |
1414 | } | |
1415 | ||
a82d593b | 1416 | /** |
3d0684b2 | 1417 | * ram_save_target_page: save one target page |
a82d593b | 1418 | * |
3d0684b2 | 1419 | * Returns the number of pages written |
a82d593b | 1420 | * |
6f37bb8b | 1421 | * @rs: current RAM state |
3d0684b2 | 1422 | * @ms: current migration state |
3d0684b2 | 1423 | * @pss: data about the page we want to send |
a82d593b | 1424 | * @last_stage: if we are at the completion stage |
a82d593b | 1425 | */ |
a0a8aa14 | 1426 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1427 | bool last_stage) |
a82d593b DDAG |
1428 | { |
1429 | int res = 0; | |
1430 | ||
1431 | /* Check the pages is dirty and if it is send it */ | |
f20e2865 | 1432 | if (migration_bitmap_clear_dirty(rs, pss->block, pss->page)) { |
6d358d94 JQ |
1433 | /* |
1434 | * If xbzrle is on, stop using the data compression after first | |
1435 | * round of migration even if compression is enabled. In theory, | |
1436 | * xbzrle can do better than compression. | |
1437 | */ | |
6b6712ef JQ |
1438 | if (migrate_use_compression() && |
1439 | (rs->ram_bulk_stage || !migrate_use_xbzrle())) { | |
a0a8aa14 | 1440 | res = ram_save_compressed_page(rs, pss, last_stage); |
a82d593b | 1441 | } else { |
a0a8aa14 | 1442 | res = ram_save_page(rs, pss, last_stage); |
a82d593b DDAG |
1443 | } |
1444 | ||
1445 | if (res < 0) { | |
1446 | return res; | |
1447 | } | |
6b6712ef JQ |
1448 | if (pss->block->unsentmap) { |
1449 | clear_bit(pss->page, pss->block->unsentmap); | |
a82d593b DDAG |
1450 | } |
1451 | } | |
1452 | ||
1453 | return res; | |
1454 | } | |
1455 | ||
1456 | /** | |
3d0684b2 | 1457 | * ram_save_host_page: save a whole host page |
a82d593b | 1458 | * |
3d0684b2 JQ |
1459 | * Starting at *offset send pages up to the end of the current host |
1460 | * page. It's valid for the initial offset to point into the middle of | |
1461 | * a host page in which case the remainder of the hostpage is sent. | |
1462 | * Only dirty target pages are sent. Note that the host page size may | |
1463 | * be a huge page for this block. | |
1eb3fc0a DDAG |
1464 | * The saving stops at the boundary of the used_length of the block |
1465 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 1466 | * |
3d0684b2 JQ |
1467 | * Returns the number of pages written or negative on error |
1468 | * | |
6f37bb8b | 1469 | * @rs: current RAM state |
3d0684b2 | 1470 | * @ms: current migration state |
3d0684b2 | 1471 | * @pss: data about the page we want to send |
a82d593b | 1472 | * @last_stage: if we are at the completion stage |
a82d593b | 1473 | */ |
a0a8aa14 | 1474 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1475 | bool last_stage) |
a82d593b DDAG |
1476 | { |
1477 | int tmppages, pages = 0; | |
a935e30f JQ |
1478 | size_t pagesize_bits = |
1479 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
4c011c37 | 1480 | |
a82d593b | 1481 | do { |
f20e2865 | 1482 | tmppages = ram_save_target_page(rs, pss, last_stage); |
a82d593b DDAG |
1483 | if (tmppages < 0) { |
1484 | return tmppages; | |
1485 | } | |
1486 | ||
1487 | pages += tmppages; | |
a935e30f | 1488 | pss->page++; |
1eb3fc0a DDAG |
1489 | } while ((pss->page & (pagesize_bits - 1)) && |
1490 | offset_in_ramblock(pss->block, pss->page << TARGET_PAGE_BITS)); | |
a82d593b DDAG |
1491 | |
1492 | /* The offset we leave with is the last one we looked at */ | |
a935e30f | 1493 | pss->page--; |
a82d593b DDAG |
1494 | return pages; |
1495 | } | |
6c595cde | 1496 | |
56e93d26 | 1497 | /** |
3d0684b2 | 1498 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1499 | * |
1500 | * Called within an RCU critical section. | |
1501 | * | |
3d0684b2 | 1502 | * Returns the number of pages written where zero means no dirty pages |
56e93d26 | 1503 | * |
6f37bb8b | 1504 | * @rs: current RAM state |
56e93d26 | 1505 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1506 | * |
1507 | * On systems where host-page-size > target-page-size it will send all the | |
1508 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1509 | */ |
1510 | ||
ce25d337 | 1511 | static int ram_find_and_save_block(RAMState *rs, bool last_stage) |
56e93d26 | 1512 | { |
b8fb8cb7 | 1513 | PageSearchStatus pss; |
56e93d26 | 1514 | int pages = 0; |
b9e60928 | 1515 | bool again, found; |
56e93d26 | 1516 | |
0827b9e9 AA |
1517 | /* No dirty page as there is zero RAM */ |
1518 | if (!ram_bytes_total()) { | |
1519 | return pages; | |
1520 | } | |
1521 | ||
6f37bb8b | 1522 | pss.block = rs->last_seen_block; |
a935e30f | 1523 | pss.page = rs->last_page; |
b8fb8cb7 DDAG |
1524 | pss.complete_round = false; |
1525 | ||
1526 | if (!pss.block) { | |
1527 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1528 | } | |
56e93d26 | 1529 | |
b9e60928 | 1530 | do { |
a82d593b | 1531 | again = true; |
f20e2865 | 1532 | found = get_queued_page(rs, &pss); |
b9e60928 | 1533 | |
a82d593b DDAG |
1534 | if (!found) { |
1535 | /* priority queue empty, so just search for something dirty */ | |
f20e2865 | 1536 | found = find_dirty_block(rs, &pss, &again); |
a82d593b | 1537 | } |
f3f491fc | 1538 | |
a82d593b | 1539 | if (found) { |
f20e2865 | 1540 | pages = ram_save_host_page(rs, &pss, last_stage); |
56e93d26 | 1541 | } |
b9e60928 | 1542 | } while (!pages && again); |
56e93d26 | 1543 | |
6f37bb8b | 1544 | rs->last_seen_block = pss.block; |
a935e30f | 1545 | rs->last_page = pss.page; |
56e93d26 JQ |
1546 | |
1547 | return pages; | |
1548 | } | |
1549 | ||
1550 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1551 | { | |
1552 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 1553 | |
56e93d26 | 1554 | if (zero) { |
9360447d | 1555 | ram_counters.duplicate += pages; |
56e93d26 | 1556 | } else { |
9360447d JQ |
1557 | ram_counters.normal += pages; |
1558 | ram_counters.transferred += size; | |
56e93d26 JQ |
1559 | qemu_update_position(f, size); |
1560 | } | |
1561 | } | |
1562 | ||
56e93d26 JQ |
1563 | uint64_t ram_bytes_total(void) |
1564 | { | |
1565 | RAMBlock *block; | |
1566 | uint64_t total = 0; | |
1567 | ||
1568 | rcu_read_lock(); | |
99e15582 | 1569 | RAMBLOCK_FOREACH(block) { |
56e93d26 | 1570 | total += block->used_length; |
99e15582 | 1571 | } |
56e93d26 JQ |
1572 | rcu_read_unlock(); |
1573 | return total; | |
1574 | } | |
1575 | ||
f265e0e4 | 1576 | static void xbzrle_load_setup(void) |
56e93d26 | 1577 | { |
f265e0e4 | 1578 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
1579 | } |
1580 | ||
f265e0e4 JQ |
1581 | static void xbzrle_load_cleanup(void) |
1582 | { | |
1583 | g_free(XBZRLE.decoded_buf); | |
1584 | XBZRLE.decoded_buf = NULL; | |
1585 | } | |
1586 | ||
1587 | static void ram_save_cleanup(void *opaque) | |
56e93d26 | 1588 | { |
53518d94 | 1589 | RAMState **rsp = opaque; |
6b6712ef | 1590 | RAMBlock *block; |
eb859c53 | 1591 | |
2ff64038 LZ |
1592 | /* caller have hold iothread lock or is in a bh, so there is |
1593 | * no writing race against this migration_bitmap | |
1594 | */ | |
6b6712ef JQ |
1595 | memory_global_dirty_log_stop(); |
1596 | ||
1597 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1598 | g_free(block->bmap); | |
1599 | block->bmap = NULL; | |
1600 | g_free(block->unsentmap); | |
1601 | block->unsentmap = NULL; | |
56e93d26 JQ |
1602 | } |
1603 | ||
1604 | XBZRLE_cache_lock(); | |
1605 | if (XBZRLE.cache) { | |
1606 | cache_fini(XBZRLE.cache); | |
1607 | g_free(XBZRLE.encoded_buf); | |
1608 | g_free(XBZRLE.current_buf); | |
c00e0928 | 1609 | g_free(XBZRLE.zero_target_page); |
56e93d26 JQ |
1610 | XBZRLE.cache = NULL; |
1611 | XBZRLE.encoded_buf = NULL; | |
1612 | XBZRLE.current_buf = NULL; | |
c00e0928 | 1613 | XBZRLE.zero_target_page = NULL; |
56e93d26 JQ |
1614 | } |
1615 | XBZRLE_cache_unlock(); | |
53518d94 | 1616 | migration_page_queue_free(*rsp); |
f0afa331 | 1617 | compress_threads_save_cleanup(); |
53518d94 JQ |
1618 | g_free(*rsp); |
1619 | *rsp = NULL; | |
56e93d26 JQ |
1620 | } |
1621 | ||
6f37bb8b | 1622 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1623 | { |
6f37bb8b JQ |
1624 | rs->last_seen_block = NULL; |
1625 | rs->last_sent_block = NULL; | |
269ace29 | 1626 | rs->last_page = 0; |
6f37bb8b JQ |
1627 | rs->last_version = ram_list.version; |
1628 | rs->ram_bulk_stage = true; | |
56e93d26 JQ |
1629 | } |
1630 | ||
1631 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1632 | ||
4f2e4252 DDAG |
1633 | /* |
1634 | * 'expected' is the value you expect the bitmap mostly to be full | |
1635 | * of; it won't bother printing lines that are all this value. | |
1636 | * If 'todump' is null the migration bitmap is dumped. | |
1637 | */ | |
6b6712ef JQ |
1638 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected, |
1639 | unsigned long pages) | |
4f2e4252 | 1640 | { |
4f2e4252 DDAG |
1641 | int64_t cur; |
1642 | int64_t linelen = 128; | |
1643 | char linebuf[129]; | |
1644 | ||
6b6712ef | 1645 | for (cur = 0; cur < pages; cur += linelen) { |
4f2e4252 DDAG |
1646 | int64_t curb; |
1647 | bool found = false; | |
1648 | /* | |
1649 | * Last line; catch the case where the line length | |
1650 | * is longer than remaining ram | |
1651 | */ | |
6b6712ef JQ |
1652 | if (cur + linelen > pages) { |
1653 | linelen = pages - cur; | |
4f2e4252 DDAG |
1654 | } |
1655 | for (curb = 0; curb < linelen; curb++) { | |
1656 | bool thisbit = test_bit(cur + curb, todump); | |
1657 | linebuf[curb] = thisbit ? '1' : '.'; | |
1658 | found = found || (thisbit != expected); | |
1659 | } | |
1660 | if (found) { | |
1661 | linebuf[curb] = '\0'; | |
1662 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1663 | } | |
1664 | } | |
1665 | } | |
1666 | ||
e0b266f0 DDAG |
1667 | /* **** functions for postcopy ***** */ |
1668 | ||
ced1c616 PB |
1669 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1670 | { | |
1671 | struct RAMBlock *block; | |
ced1c616 | 1672 | |
99e15582 | 1673 | RAMBLOCK_FOREACH(block) { |
6b6712ef JQ |
1674 | unsigned long *bitmap = block->bmap; |
1675 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
1676 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
1677 | |
1678 | while (run_start < range) { | |
1679 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
aaa2064c | 1680 | ram_discard_range(block->idstr, run_start << TARGET_PAGE_BITS, |
ced1c616 PB |
1681 | (run_end - run_start) << TARGET_PAGE_BITS); |
1682 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); | |
1683 | } | |
1684 | } | |
1685 | } | |
1686 | ||
3d0684b2 JQ |
1687 | /** |
1688 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1689 | * | |
1690 | * Returns zero on success | |
1691 | * | |
e0b266f0 DDAG |
1692 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
1693 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1694 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
3d0684b2 JQ |
1695 | * |
1696 | * @ms: current migration state | |
1697 | * @pds: state for postcopy | |
1698 | * @start: RAMBlock starting page | |
1699 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1700 | */ |
1701 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1702 | PostcopyDiscardState *pds, | |
6b6712ef | 1703 | RAMBlock *block) |
e0b266f0 | 1704 | { |
6b6712ef | 1705 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 1706 | unsigned long current; |
6b6712ef | 1707 | unsigned long *unsentmap = block->unsentmap; |
e0b266f0 | 1708 | |
6b6712ef | 1709 | for (current = 0; current < end; ) { |
e0b266f0 DDAG |
1710 | unsigned long one = find_next_bit(unsentmap, end, current); |
1711 | ||
1712 | if (one <= end) { | |
1713 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1714 | unsigned long discard_length; | |
1715 | ||
1716 | if (zero >= end) { | |
1717 | discard_length = end - one; | |
1718 | } else { | |
1719 | discard_length = zero - one; | |
1720 | } | |
d688c62d DDAG |
1721 | if (discard_length) { |
1722 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1723 | } | |
e0b266f0 DDAG |
1724 | current = one + discard_length; |
1725 | } else { | |
1726 | current = one; | |
1727 | } | |
1728 | } | |
1729 | ||
1730 | return 0; | |
1731 | } | |
1732 | ||
3d0684b2 JQ |
1733 | /** |
1734 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1735 | * | |
1736 | * Returns 0 for success or negative for error | |
1737 | * | |
e0b266f0 DDAG |
1738 | * Utility for the outgoing postcopy code. |
1739 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1740 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1741 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1742 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1743 | * |
1744 | * @ms: current migration state | |
e0b266f0 DDAG |
1745 | */ |
1746 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1747 | { | |
1748 | struct RAMBlock *block; | |
1749 | int ret; | |
1750 | ||
99e15582 | 1751 | RAMBLOCK_FOREACH(block) { |
6b6712ef JQ |
1752 | PostcopyDiscardState *pds = |
1753 | postcopy_discard_send_init(ms, block->idstr); | |
e0b266f0 DDAG |
1754 | |
1755 | /* | |
1756 | * Postcopy sends chunks of bitmap over the wire, but it | |
1757 | * just needs indexes at this point, avoids it having | |
1758 | * target page specific code. | |
1759 | */ | |
6b6712ef | 1760 | ret = postcopy_send_discard_bm_ram(ms, pds, block); |
e0b266f0 DDAG |
1761 | postcopy_discard_send_finish(ms, pds); |
1762 | if (ret) { | |
1763 | return ret; | |
1764 | } | |
1765 | } | |
1766 | ||
1767 | return 0; | |
1768 | } | |
1769 | ||
3d0684b2 JQ |
1770 | /** |
1771 | * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages | |
1772 | * | |
1773 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
1774 | * canonicalize the two bitmaps, that are similar, but one is | |
1775 | * inverted. | |
99e314eb | 1776 | * |
3d0684b2 JQ |
1777 | * Postcopy requires that all target pages in a hostpage are dirty or |
1778 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 1779 | * |
3d0684b2 JQ |
1780 | * @ms: current migration state |
1781 | * @unsent_pass: if true we need to canonicalize partially unsent host pages | |
1782 | * otherwise we need to canonicalize partially dirty host pages | |
1783 | * @block: block that contains the page we want to canonicalize | |
1784 | * @pds: state for postcopy | |
99e314eb DDAG |
1785 | */ |
1786 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1787 | RAMBlock *block, | |
1788 | PostcopyDiscardState *pds) | |
1789 | { | |
53518d94 | 1790 | RAMState *rs = ram_state; |
6b6712ef JQ |
1791 | unsigned long *bitmap = block->bmap; |
1792 | unsigned long *unsentmap = block->unsentmap; | |
29c59172 | 1793 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 1794 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
1795 | unsigned long run_start; |
1796 | ||
29c59172 DDAG |
1797 | if (block->page_size == TARGET_PAGE_SIZE) { |
1798 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
1799 | return; | |
1800 | } | |
1801 | ||
99e314eb DDAG |
1802 | if (unsent_pass) { |
1803 | /* Find a sent page */ | |
6b6712ef | 1804 | run_start = find_next_zero_bit(unsentmap, pages, 0); |
99e314eb DDAG |
1805 | } else { |
1806 | /* Find a dirty page */ | |
6b6712ef | 1807 | run_start = find_next_bit(bitmap, pages, 0); |
99e314eb DDAG |
1808 | } |
1809 | ||
6b6712ef | 1810 | while (run_start < pages) { |
99e314eb DDAG |
1811 | bool do_fixup = false; |
1812 | unsigned long fixup_start_addr; | |
1813 | unsigned long host_offset; | |
1814 | ||
1815 | /* | |
1816 | * If the start of this run of pages is in the middle of a host | |
1817 | * page, then we need to fixup this host page. | |
1818 | */ | |
1819 | host_offset = run_start % host_ratio; | |
1820 | if (host_offset) { | |
1821 | do_fixup = true; | |
1822 | run_start -= host_offset; | |
1823 | fixup_start_addr = run_start; | |
1824 | /* For the next pass */ | |
1825 | run_start = run_start + host_ratio; | |
1826 | } else { | |
1827 | /* Find the end of this run */ | |
1828 | unsigned long run_end; | |
1829 | if (unsent_pass) { | |
6b6712ef | 1830 | run_end = find_next_bit(unsentmap, pages, run_start + 1); |
99e314eb | 1831 | } else { |
6b6712ef | 1832 | run_end = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
1833 | } |
1834 | /* | |
1835 | * If the end isn't at the start of a host page, then the | |
1836 | * run doesn't finish at the end of a host page | |
1837 | * and we need to discard. | |
1838 | */ | |
1839 | host_offset = run_end % host_ratio; | |
1840 | if (host_offset) { | |
1841 | do_fixup = true; | |
1842 | fixup_start_addr = run_end - host_offset; | |
1843 | /* | |
1844 | * This host page has gone, the next loop iteration starts | |
1845 | * from after the fixup | |
1846 | */ | |
1847 | run_start = fixup_start_addr + host_ratio; | |
1848 | } else { | |
1849 | /* | |
1850 | * No discards on this iteration, next loop starts from | |
1851 | * next sent/dirty page | |
1852 | */ | |
1853 | run_start = run_end + 1; | |
1854 | } | |
1855 | } | |
1856 | ||
1857 | if (do_fixup) { | |
1858 | unsigned long page; | |
1859 | ||
1860 | /* Tell the destination to discard this page */ | |
1861 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1862 | /* For the unsent_pass we: | |
1863 | * discard partially sent pages | |
1864 | * For the !unsent_pass (dirty) we: | |
1865 | * discard partially dirty pages that were sent | |
1866 | * (any partially sent pages were already discarded | |
1867 | * by the previous unsent_pass) | |
1868 | */ | |
1869 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1870 | host_ratio); | |
1871 | } | |
1872 | ||
1873 | /* Clean up the bitmap */ | |
1874 | for (page = fixup_start_addr; | |
1875 | page < fixup_start_addr + host_ratio; page++) { | |
1876 | /* All pages in this host page are now not sent */ | |
1877 | set_bit(page, unsentmap); | |
1878 | ||
1879 | /* | |
1880 | * Remark them as dirty, updating the count for any pages | |
1881 | * that weren't previously dirty. | |
1882 | */ | |
0d8ec885 | 1883 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
1884 | } |
1885 | } | |
1886 | ||
1887 | if (unsent_pass) { | |
1888 | /* Find the next sent page for the next iteration */ | |
6b6712ef | 1889 | run_start = find_next_zero_bit(unsentmap, pages, run_start); |
99e314eb DDAG |
1890 | } else { |
1891 | /* Find the next dirty page for the next iteration */ | |
6b6712ef | 1892 | run_start = find_next_bit(bitmap, pages, run_start); |
99e314eb DDAG |
1893 | } |
1894 | } | |
1895 | } | |
1896 | ||
3d0684b2 JQ |
1897 | /** |
1898 | * postcopy_chuck_hostpages: discrad any partially sent host page | |
1899 | * | |
99e314eb DDAG |
1900 | * Utility for the outgoing postcopy code. |
1901 | * | |
1902 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
1903 | * dirty host-page size chunks as all dirty. In this case the host-page |
1904 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 1905 | * |
3d0684b2 JQ |
1906 | * Returns zero on success |
1907 | * | |
1908 | * @ms: current migration state | |
6b6712ef | 1909 | * @block: block we want to work with |
99e314eb | 1910 | */ |
6b6712ef | 1911 | static int postcopy_chunk_hostpages(MigrationState *ms, RAMBlock *block) |
99e314eb | 1912 | { |
6b6712ef JQ |
1913 | PostcopyDiscardState *pds = |
1914 | postcopy_discard_send_init(ms, block->idstr); | |
99e314eb | 1915 | |
6b6712ef JQ |
1916 | /* First pass: Discard all partially sent host pages */ |
1917 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1918 | /* | |
1919 | * Second pass: Ensure that all partially dirty host pages are made | |
1920 | * fully dirty. | |
1921 | */ | |
1922 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
99e314eb | 1923 | |
6b6712ef | 1924 | postcopy_discard_send_finish(ms, pds); |
99e314eb DDAG |
1925 | return 0; |
1926 | } | |
1927 | ||
3d0684b2 JQ |
1928 | /** |
1929 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
1930 | * | |
1931 | * Returns zero on success | |
1932 | * | |
e0b266f0 DDAG |
1933 | * Transmit the set of pages to be discarded after precopy to the target |
1934 | * these are pages that: | |
1935 | * a) Have been previously transmitted but are now dirty again | |
1936 | * b) Pages that have never been transmitted, this ensures that | |
1937 | * any pages on the destination that have been mapped by background | |
1938 | * tasks get discarded (transparent huge pages is the specific concern) | |
1939 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
1940 | * |
1941 | * @ms: current migration state | |
e0b266f0 DDAG |
1942 | */ |
1943 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1944 | { | |
53518d94 | 1945 | RAMState *rs = ram_state; |
6b6712ef | 1946 | RAMBlock *block; |
e0b266f0 | 1947 | int ret; |
e0b266f0 DDAG |
1948 | |
1949 | rcu_read_lock(); | |
1950 | ||
1951 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 1952 | migration_bitmap_sync(rs); |
e0b266f0 | 1953 | |
6b6712ef JQ |
1954 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
1955 | rs->last_seen_block = NULL; | |
1956 | rs->last_sent_block = NULL; | |
1957 | rs->last_page = 0; | |
e0b266f0 | 1958 | |
6b6712ef JQ |
1959 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { |
1960 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; | |
1961 | unsigned long *bitmap = block->bmap; | |
1962 | unsigned long *unsentmap = block->unsentmap; | |
1963 | ||
1964 | if (!unsentmap) { | |
1965 | /* We don't have a safe way to resize the sentmap, so | |
1966 | * if the bitmap was resized it will be NULL at this | |
1967 | * point. | |
1968 | */ | |
1969 | error_report("migration ram resized during precopy phase"); | |
1970 | rcu_read_unlock(); | |
1971 | return -EINVAL; | |
1972 | } | |
1973 | /* Deal with TPS != HPS and huge pages */ | |
1974 | ret = postcopy_chunk_hostpages(ms, block); | |
1975 | if (ret) { | |
1976 | rcu_read_unlock(); | |
1977 | return ret; | |
1978 | } | |
e0b266f0 | 1979 | |
6b6712ef JQ |
1980 | /* |
1981 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1982 | */ | |
1983 | bitmap_or(unsentmap, unsentmap, bitmap, pages); | |
e0b266f0 | 1984 | #ifdef DEBUG_POSTCOPY |
6b6712ef | 1985 | ram_debug_dump_bitmap(unsentmap, true, pages); |
e0b266f0 | 1986 | #endif |
6b6712ef JQ |
1987 | } |
1988 | trace_ram_postcopy_send_discard_bitmap(); | |
e0b266f0 DDAG |
1989 | |
1990 | ret = postcopy_each_ram_send_discard(ms); | |
1991 | rcu_read_unlock(); | |
1992 | ||
1993 | return ret; | |
1994 | } | |
1995 | ||
3d0684b2 JQ |
1996 | /** |
1997 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 1998 | * |
3d0684b2 | 1999 | * Returns zero on success |
e0b266f0 | 2000 | * |
36449157 JQ |
2001 | * @rbname: name of the RAMBlock of the request. NULL means the |
2002 | * same that last one. | |
3d0684b2 JQ |
2003 | * @start: RAMBlock starting page |
2004 | * @length: RAMBlock size | |
e0b266f0 | 2005 | */ |
aaa2064c | 2006 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 DDAG |
2007 | { |
2008 | int ret = -1; | |
2009 | ||
36449157 | 2010 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2011 | |
e0b266f0 | 2012 | rcu_read_lock(); |
36449157 | 2013 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2014 | |
2015 | if (!rb) { | |
36449157 | 2016 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
e0b266f0 DDAG |
2017 | goto err; |
2018 | } | |
2019 | ||
d3a5038c | 2020 | ret = ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2021 | |
2022 | err: | |
2023 | rcu_read_unlock(); | |
2024 | ||
2025 | return ret; | |
2026 | } | |
2027 | ||
53518d94 | 2028 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2029 | { |
53518d94 JQ |
2030 | *rsp = g_new0(RAMState, 1); |
2031 | ||
2032 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
2033 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
2034 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
56e93d26 JQ |
2035 | |
2036 | if (migrate_use_xbzrle()) { | |
2037 | XBZRLE_cache_lock(); | |
c00e0928 | 2038 | XBZRLE.zero_target_page = g_malloc0(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2039 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size() / |
2040 | TARGET_PAGE_SIZE, | |
2041 | TARGET_PAGE_SIZE); | |
2042 | if (!XBZRLE.cache) { | |
2043 | XBZRLE_cache_unlock(); | |
2044 | error_report("Error creating cache"); | |
53518d94 JQ |
2045 | g_free(*rsp); |
2046 | *rsp = NULL; | |
56e93d26 JQ |
2047 | return -1; |
2048 | } | |
2049 | XBZRLE_cache_unlock(); | |
2050 | ||
2051 | /* We prefer not to abort if there is no memory */ | |
2052 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2053 | if (!XBZRLE.encoded_buf) { | |
2054 | error_report("Error allocating encoded_buf"); | |
53518d94 JQ |
2055 | g_free(*rsp); |
2056 | *rsp = NULL; | |
56e93d26 JQ |
2057 | return -1; |
2058 | } | |
2059 | ||
2060 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2061 | if (!XBZRLE.current_buf) { | |
2062 | error_report("Error allocating current_buf"); | |
2063 | g_free(XBZRLE.encoded_buf); | |
2064 | XBZRLE.encoded_buf = NULL; | |
53518d94 JQ |
2065 | g_free(*rsp); |
2066 | *rsp = NULL; | |
56e93d26 JQ |
2067 | return -1; |
2068 | } | |
56e93d26 JQ |
2069 | } |
2070 | ||
49877834 PB |
2071 | /* For memory_global_dirty_log_start below. */ |
2072 | qemu_mutex_lock_iothread(); | |
2073 | ||
56e93d26 JQ |
2074 | qemu_mutex_lock_ramlist(); |
2075 | rcu_read_lock(); | |
53518d94 | 2076 | ram_state_reset(*rsp); |
56e93d26 | 2077 | |
0827b9e9 AA |
2078 | /* Skip setting bitmap if there is no RAM */ |
2079 | if (ram_bytes_total()) { | |
6b6712ef JQ |
2080 | RAMBlock *block; |
2081 | ||
2082 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
2083 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; | |
0827b9e9 | 2084 | |
6b6712ef JQ |
2085 | block->bmap = bitmap_new(pages); |
2086 | bitmap_set(block->bmap, 0, pages); | |
2087 | if (migrate_postcopy_ram()) { | |
2088 | block->unsentmap = bitmap_new(pages); | |
2089 | bitmap_set(block->unsentmap, 0, pages); | |
2090 | } | |
0827b9e9 | 2091 | } |
f3f491fc DDAG |
2092 | } |
2093 | ||
56e93d26 JQ |
2094 | /* |
2095 | * Count the total number of pages used by ram blocks not including any | |
2096 | * gaps due to alignment or unplugs. | |
2097 | */ | |
53518d94 | 2098 | (*rsp)->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
56e93d26 JQ |
2099 | |
2100 | memory_global_dirty_log_start(); | |
53518d94 | 2101 | migration_bitmap_sync(*rsp); |
56e93d26 | 2102 | qemu_mutex_unlock_ramlist(); |
49877834 | 2103 | qemu_mutex_unlock_iothread(); |
a91246c9 HZ |
2104 | rcu_read_unlock(); |
2105 | ||
2106 | return 0; | |
2107 | } | |
2108 | ||
3d0684b2 JQ |
2109 | /* |
2110 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2111 | * long-running RCU critical section. When rcu-reclaims in the code |
2112 | * start to become numerous it will be necessary to reduce the | |
2113 | * granularity of these critical sections. | |
2114 | */ | |
2115 | ||
3d0684b2 JQ |
2116 | /** |
2117 | * ram_save_setup: Setup RAM for migration | |
2118 | * | |
2119 | * Returns zero to indicate success and negative for error | |
2120 | * | |
2121 | * @f: QEMUFile where to send the data | |
2122 | * @opaque: RAMState pointer | |
2123 | */ | |
a91246c9 HZ |
2124 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2125 | { | |
53518d94 | 2126 | RAMState **rsp = opaque; |
a91246c9 HZ |
2127 | RAMBlock *block; |
2128 | ||
2129 | /* migration has already setup the bitmap, reuse it. */ | |
2130 | if (!migration_in_colo_state()) { | |
53518d94 | 2131 | if (ram_state_init(rsp) != 0) { |
a91246c9 | 2132 | return -1; |
53518d94 | 2133 | } |
a91246c9 | 2134 | } |
53518d94 | 2135 | (*rsp)->f = f; |
a91246c9 HZ |
2136 | |
2137 | rcu_read_lock(); | |
56e93d26 JQ |
2138 | |
2139 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
2140 | ||
99e15582 | 2141 | RAMBLOCK_FOREACH(block) { |
56e93d26 JQ |
2142 | qemu_put_byte(f, strlen(block->idstr)); |
2143 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2144 | qemu_put_be64(f, block->used_length); | |
ef08fb38 DDAG |
2145 | if (migrate_postcopy_ram() && block->page_size != qemu_host_page_size) { |
2146 | qemu_put_be64(f, block->page_size); | |
2147 | } | |
56e93d26 JQ |
2148 | } |
2149 | ||
2150 | rcu_read_unlock(); | |
f0afa331 | 2151 | compress_threads_save_setup(); |
56e93d26 JQ |
2152 | |
2153 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
2154 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2155 | ||
2156 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2157 | ||
2158 | return 0; | |
2159 | } | |
2160 | ||
3d0684b2 JQ |
2161 | /** |
2162 | * ram_save_iterate: iterative stage for migration | |
2163 | * | |
2164 | * Returns zero to indicate success and negative for error | |
2165 | * | |
2166 | * @f: QEMUFile where to send the data | |
2167 | * @opaque: RAMState pointer | |
2168 | */ | |
56e93d26 JQ |
2169 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2170 | { | |
53518d94 JQ |
2171 | RAMState **temp = opaque; |
2172 | RAMState *rs = *temp; | |
56e93d26 JQ |
2173 | int ret; |
2174 | int i; | |
2175 | int64_t t0; | |
5c90308f | 2176 | int done = 0; |
56e93d26 JQ |
2177 | |
2178 | rcu_read_lock(); | |
6f37bb8b JQ |
2179 | if (ram_list.version != rs->last_version) { |
2180 | ram_state_reset(rs); | |
56e93d26 JQ |
2181 | } |
2182 | ||
2183 | /* Read version before ram_list.blocks */ | |
2184 | smp_rmb(); | |
2185 | ||
2186 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
2187 | ||
2188 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
2189 | i = 0; | |
2190 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
2191 | int pages; | |
2192 | ||
ce25d337 | 2193 | pages = ram_find_and_save_block(rs, false); |
56e93d26 JQ |
2194 | /* no more pages to sent */ |
2195 | if (pages == 0) { | |
5c90308f | 2196 | done = 1; |
56e93d26 JQ |
2197 | break; |
2198 | } | |
23b28c3c | 2199 | rs->iterations++; |
070afca2 | 2200 | |
56e93d26 JQ |
2201 | /* we want to check in the 1st loop, just in case it was the 1st time |
2202 | and we had to sync the dirty bitmap. | |
2203 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2204 | iterations | |
2205 | */ | |
2206 | if ((i & 63) == 0) { | |
2207 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2208 | if (t1 > MAX_WAIT) { | |
55c4446b | 2209 | trace_ram_save_iterate_big_wait(t1, i); |
56e93d26 JQ |
2210 | break; |
2211 | } | |
2212 | } | |
2213 | i++; | |
2214 | } | |
ce25d337 | 2215 | flush_compressed_data(rs); |
56e93d26 JQ |
2216 | rcu_read_unlock(); |
2217 | ||
2218 | /* | |
2219 | * Must occur before EOS (or any QEMUFile operation) | |
2220 | * because of RDMA protocol. | |
2221 | */ | |
2222 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2223 | ||
2224 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
9360447d | 2225 | ram_counters.transferred += 8; |
56e93d26 JQ |
2226 | |
2227 | ret = qemu_file_get_error(f); | |
2228 | if (ret < 0) { | |
2229 | return ret; | |
2230 | } | |
2231 | ||
5c90308f | 2232 | return done; |
56e93d26 JQ |
2233 | } |
2234 | ||
3d0684b2 JQ |
2235 | /** |
2236 | * ram_save_complete: function called to send the remaining amount of ram | |
2237 | * | |
2238 | * Returns zero to indicate success | |
2239 | * | |
2240 | * Called with iothread lock | |
2241 | * | |
2242 | * @f: QEMUFile where to send the data | |
2243 | * @opaque: RAMState pointer | |
2244 | */ | |
56e93d26 JQ |
2245 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2246 | { | |
53518d94 JQ |
2247 | RAMState **temp = opaque; |
2248 | RAMState *rs = *temp; | |
6f37bb8b | 2249 | |
56e93d26 JQ |
2250 | rcu_read_lock(); |
2251 | ||
5727309d | 2252 | if (!migration_in_postcopy()) { |
8d820d6f | 2253 | migration_bitmap_sync(rs); |
663e6c1d | 2254 | } |
56e93d26 JQ |
2255 | |
2256 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2257 | ||
2258 | /* try transferring iterative blocks of memory */ | |
2259 | ||
2260 | /* flush all remaining blocks regardless of rate limiting */ | |
2261 | while (true) { | |
2262 | int pages; | |
2263 | ||
ce25d337 | 2264 | pages = ram_find_and_save_block(rs, !migration_in_colo_state()); |
56e93d26 JQ |
2265 | /* no more blocks to sent */ |
2266 | if (pages == 0) { | |
2267 | break; | |
2268 | } | |
2269 | } | |
2270 | ||
ce25d337 | 2271 | flush_compressed_data(rs); |
56e93d26 | 2272 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 JQ |
2273 | |
2274 | rcu_read_unlock(); | |
d09a6fde | 2275 | |
56e93d26 JQ |
2276 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2277 | ||
2278 | return 0; | |
2279 | } | |
2280 | ||
c31b098f DDAG |
2281 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2282 | uint64_t *non_postcopiable_pending, | |
2283 | uint64_t *postcopiable_pending) | |
56e93d26 | 2284 | { |
53518d94 JQ |
2285 | RAMState **temp = opaque; |
2286 | RAMState *rs = *temp; | |
56e93d26 JQ |
2287 | uint64_t remaining_size; |
2288 | ||
9edabd4d | 2289 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2290 | |
5727309d | 2291 | if (!migration_in_postcopy() && |
663e6c1d | 2292 | remaining_size < max_size) { |
56e93d26 JQ |
2293 | qemu_mutex_lock_iothread(); |
2294 | rcu_read_lock(); | |
8d820d6f | 2295 | migration_bitmap_sync(rs); |
56e93d26 JQ |
2296 | rcu_read_unlock(); |
2297 | qemu_mutex_unlock_iothread(); | |
9edabd4d | 2298 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2299 | } |
c31b098f | 2300 | |
86e1167e VSO |
2301 | if (migrate_postcopy_ram()) { |
2302 | /* We can do postcopy, and all the data is postcopiable */ | |
2303 | *postcopiable_pending += remaining_size; | |
2304 | } else { | |
2305 | *non_postcopiable_pending += remaining_size; | |
2306 | } | |
56e93d26 JQ |
2307 | } |
2308 | ||
2309 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2310 | { | |
2311 | unsigned int xh_len; | |
2312 | int xh_flags; | |
063e760a | 2313 | uint8_t *loaded_data; |
56e93d26 | 2314 | |
56e93d26 JQ |
2315 | /* extract RLE header */ |
2316 | xh_flags = qemu_get_byte(f); | |
2317 | xh_len = qemu_get_be16(f); | |
2318 | ||
2319 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2320 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2321 | return -1; | |
2322 | } | |
2323 | ||
2324 | if (xh_len > TARGET_PAGE_SIZE) { | |
2325 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2326 | return -1; | |
2327 | } | |
f265e0e4 | 2328 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 2329 | /* load data and decode */ |
f265e0e4 | 2330 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 2331 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2332 | |
2333 | /* decode RLE */ | |
063e760a | 2334 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2335 | TARGET_PAGE_SIZE) == -1) { |
2336 | error_report("Failed to load XBZRLE page - decode error!"); | |
2337 | return -1; | |
2338 | } | |
2339 | ||
2340 | return 0; | |
2341 | } | |
2342 | ||
3d0684b2 JQ |
2343 | /** |
2344 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2345 | * | |
2346 | * Must be called from within a rcu critical section. | |
2347 | * | |
56e93d26 | 2348 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2349 | * |
3d0684b2 JQ |
2350 | * @f: QEMUFile where to read the data from |
2351 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2352 | */ |
3d0684b2 | 2353 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2354 | { |
2355 | static RAMBlock *block = NULL; | |
2356 | char id[256]; | |
2357 | uint8_t len; | |
2358 | ||
2359 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2360 | if (!block) { |
56e93d26 JQ |
2361 | error_report("Ack, bad migration stream!"); |
2362 | return NULL; | |
2363 | } | |
4c4bad48 | 2364 | return block; |
56e93d26 JQ |
2365 | } |
2366 | ||
2367 | len = qemu_get_byte(f); | |
2368 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2369 | id[len] = 0; | |
2370 | ||
e3dd7493 | 2371 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2372 | if (!block) { |
2373 | error_report("Can't find block %s", id); | |
2374 | return NULL; | |
56e93d26 JQ |
2375 | } |
2376 | ||
4c4bad48 HZ |
2377 | return block; |
2378 | } | |
2379 | ||
2380 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2381 | ram_addr_t offset) | |
2382 | { | |
2383 | if (!offset_in_ramblock(block, offset)) { | |
2384 | return NULL; | |
2385 | } | |
2386 | ||
2387 | return block->host + offset; | |
56e93d26 JQ |
2388 | } |
2389 | ||
3d0684b2 JQ |
2390 | /** |
2391 | * ram_handle_compressed: handle the zero page case | |
2392 | * | |
56e93d26 JQ |
2393 | * If a page (or a whole RDMA chunk) has been |
2394 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2395 | * |
2396 | * @host: host address for the zero page | |
2397 | * @ch: what the page is filled from. We only support zero | |
2398 | * @size: size of the zero page | |
56e93d26 JQ |
2399 | */ |
2400 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2401 | { | |
2402 | if (ch != 0 || !is_zero_range(host, size)) { | |
2403 | memset(host, ch, size); | |
2404 | } | |
2405 | } | |
2406 | ||
2407 | static void *do_data_decompress(void *opaque) | |
2408 | { | |
2409 | DecompressParam *param = opaque; | |
2410 | unsigned long pagesize; | |
33d151f4 LL |
2411 | uint8_t *des; |
2412 | int len; | |
56e93d26 | 2413 | |
33d151f4 | 2414 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2415 | while (!param->quit) { |
33d151f4 LL |
2416 | if (param->des) { |
2417 | des = param->des; | |
2418 | len = param->len; | |
2419 | param->des = 0; | |
2420 | qemu_mutex_unlock(¶m->mutex); | |
2421 | ||
56e93d26 | 2422 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2423 | /* uncompress() will return failed in some case, especially |
2424 | * when the page is dirted when doing the compression, it's | |
2425 | * not a problem because the dirty page will be retransferred | |
2426 | * and uncompress() won't break the data in other pages. | |
2427 | */ | |
33d151f4 LL |
2428 | uncompress((Bytef *)des, &pagesize, |
2429 | (const Bytef *)param->compbuf, len); | |
73a8912b | 2430 | |
33d151f4 LL |
2431 | qemu_mutex_lock(&decomp_done_lock); |
2432 | param->done = true; | |
2433 | qemu_cond_signal(&decomp_done_cond); | |
2434 | qemu_mutex_unlock(&decomp_done_lock); | |
2435 | ||
2436 | qemu_mutex_lock(¶m->mutex); | |
2437 | } else { | |
2438 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2439 | } | |
56e93d26 | 2440 | } |
33d151f4 | 2441 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2442 | |
2443 | return NULL; | |
2444 | } | |
2445 | ||
5533b2e9 LL |
2446 | static void wait_for_decompress_done(void) |
2447 | { | |
2448 | int idx, thread_count; | |
2449 | ||
2450 | if (!migrate_use_compression()) { | |
2451 | return; | |
2452 | } | |
2453 | ||
2454 | thread_count = migrate_decompress_threads(); | |
2455 | qemu_mutex_lock(&decomp_done_lock); | |
2456 | for (idx = 0; idx < thread_count; idx++) { | |
2457 | while (!decomp_param[idx].done) { | |
2458 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2459 | } | |
2460 | } | |
2461 | qemu_mutex_unlock(&decomp_done_lock); | |
2462 | } | |
2463 | ||
f0afa331 | 2464 | static void compress_threads_load_setup(void) |
56e93d26 JQ |
2465 | { |
2466 | int i, thread_count; | |
2467 | ||
3416ab5b JQ |
2468 | if (!migrate_use_compression()) { |
2469 | return; | |
2470 | } | |
56e93d26 JQ |
2471 | thread_count = migrate_decompress_threads(); |
2472 | decompress_threads = g_new0(QemuThread, thread_count); | |
2473 | decomp_param = g_new0(DecompressParam, thread_count); | |
73a8912b LL |
2474 | qemu_mutex_init(&decomp_done_lock); |
2475 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2476 | for (i = 0; i < thread_count; i++) { |
2477 | qemu_mutex_init(&decomp_param[i].mutex); | |
2478 | qemu_cond_init(&decomp_param[i].cond); | |
2479 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2480 | decomp_param[i].done = true; |
90e56fb4 | 2481 | decomp_param[i].quit = false; |
56e93d26 JQ |
2482 | qemu_thread_create(decompress_threads + i, "decompress", |
2483 | do_data_decompress, decomp_param + i, | |
2484 | QEMU_THREAD_JOINABLE); | |
2485 | } | |
2486 | } | |
2487 | ||
f0afa331 | 2488 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
2489 | { |
2490 | int i, thread_count; | |
2491 | ||
3416ab5b JQ |
2492 | if (!migrate_use_compression()) { |
2493 | return; | |
2494 | } | |
56e93d26 JQ |
2495 | thread_count = migrate_decompress_threads(); |
2496 | for (i = 0; i < thread_count; i++) { | |
2497 | qemu_mutex_lock(&decomp_param[i].mutex); | |
90e56fb4 | 2498 | decomp_param[i].quit = true; |
56e93d26 JQ |
2499 | qemu_cond_signal(&decomp_param[i].cond); |
2500 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2501 | } | |
2502 | for (i = 0; i < thread_count; i++) { | |
2503 | qemu_thread_join(decompress_threads + i); | |
2504 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2505 | qemu_cond_destroy(&decomp_param[i].cond); | |
2506 | g_free(decomp_param[i].compbuf); | |
2507 | } | |
2508 | g_free(decompress_threads); | |
2509 | g_free(decomp_param); | |
56e93d26 JQ |
2510 | decompress_threads = NULL; |
2511 | decomp_param = NULL; | |
56e93d26 JQ |
2512 | } |
2513 | ||
c1bc6626 | 2514 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2515 | void *host, int len) |
2516 | { | |
2517 | int idx, thread_count; | |
2518 | ||
2519 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2520 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2521 | while (true) { |
2522 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2523 | if (decomp_param[idx].done) { |
33d151f4 LL |
2524 | decomp_param[idx].done = false; |
2525 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2526 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2527 | decomp_param[idx].des = host; |
2528 | decomp_param[idx].len = len; | |
33d151f4 LL |
2529 | qemu_cond_signal(&decomp_param[idx].cond); |
2530 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2531 | break; |
2532 | } | |
2533 | } | |
2534 | if (idx < thread_count) { | |
2535 | break; | |
73a8912b LL |
2536 | } else { |
2537 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2538 | } |
2539 | } | |
73a8912b | 2540 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2541 | } |
2542 | ||
f265e0e4 JQ |
2543 | /** |
2544 | * ram_load_setup: Setup RAM for migration incoming side | |
2545 | * | |
2546 | * Returns zero to indicate success and negative for error | |
2547 | * | |
2548 | * @f: QEMUFile where to receive the data | |
2549 | * @opaque: RAMState pointer | |
2550 | */ | |
2551 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
2552 | { | |
2553 | xbzrle_load_setup(); | |
f0afa331 | 2554 | compress_threads_load_setup(); |
f265e0e4 JQ |
2555 | return 0; |
2556 | } | |
2557 | ||
2558 | static int ram_load_cleanup(void *opaque) | |
2559 | { | |
2560 | xbzrle_load_cleanup(); | |
f0afa331 | 2561 | compress_threads_load_cleanup(); |
f265e0e4 JQ |
2562 | return 0; |
2563 | } | |
2564 | ||
3d0684b2 JQ |
2565 | /** |
2566 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
2567 | * | |
2568 | * Returns 0 for success and negative if there was one error | |
2569 | * | |
2570 | * @mis: current migration incoming state | |
2571 | * | |
2572 | * Allocate data structures etc needed by incoming migration with | |
2573 | * postcopy-ram. postcopy-ram's similarly names | |
2574 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
2575 | */ |
2576 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2577 | { | |
b8c48993 | 2578 | unsigned long ram_pages = last_ram_page(); |
1caddf8a DDAG |
2579 | |
2580 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2581 | } | |
2582 | ||
3d0684b2 JQ |
2583 | /** |
2584 | * ram_load_postcopy: load a page in postcopy case | |
2585 | * | |
2586 | * Returns 0 for success or -errno in case of error | |
2587 | * | |
a7180877 DDAG |
2588 | * Called in postcopy mode by ram_load(). |
2589 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
2590 | * |
2591 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
2592 | */ |
2593 | static int ram_load_postcopy(QEMUFile *f) | |
2594 | { | |
2595 | int flags = 0, ret = 0; | |
2596 | bool place_needed = false; | |
28abd200 | 2597 | bool matching_page_sizes = false; |
a7180877 DDAG |
2598 | MigrationIncomingState *mis = migration_incoming_get_current(); |
2599 | /* Temporary page that is later 'placed' */ | |
2600 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2601 | void *last_host = NULL; |
a3b6ff6d | 2602 | bool all_zero = false; |
a7180877 DDAG |
2603 | |
2604 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2605 | ram_addr_t addr; | |
2606 | void *host = NULL; | |
2607 | void *page_buffer = NULL; | |
2608 | void *place_source = NULL; | |
df9ff5e1 | 2609 | RAMBlock *block = NULL; |
a7180877 | 2610 | uint8_t ch; |
a7180877 DDAG |
2611 | |
2612 | addr = qemu_get_be64(f); | |
2613 | flags = addr & ~TARGET_PAGE_MASK; | |
2614 | addr &= TARGET_PAGE_MASK; | |
2615 | ||
2616 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2617 | place_needed = false; | |
bb890ed5 | 2618 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE)) { |
df9ff5e1 | 2619 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
2620 | |
2621 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2622 | if (!host) { |
2623 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2624 | ret = -EINVAL; | |
2625 | break; | |
2626 | } | |
28abd200 | 2627 | matching_page_sizes = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 2628 | /* |
28abd200 DDAG |
2629 | * Postcopy requires that we place whole host pages atomically; |
2630 | * these may be huge pages for RAMBlocks that are backed by | |
2631 | * hugetlbfs. | |
a7180877 DDAG |
2632 | * To make it atomic, the data is read into a temporary page |
2633 | * that's moved into place later. | |
2634 | * The migration protocol uses, possibly smaller, target-pages | |
2635 | * however the source ensures it always sends all the components | |
2636 | * of a host page in order. | |
2637 | */ | |
2638 | page_buffer = postcopy_host_page + | |
28abd200 | 2639 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 2640 | /* If all TP are zero then we can optimise the place */ |
28abd200 | 2641 | if (!((uintptr_t)host & (block->page_size - 1))) { |
a7180877 | 2642 | all_zero = true; |
c53b7ddc DDAG |
2643 | } else { |
2644 | /* not the 1st TP within the HP */ | |
2645 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2646 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2647 | host, last_host); |
2648 | ret = -EINVAL; | |
2649 | break; | |
2650 | } | |
a7180877 DDAG |
2651 | } |
2652 | ||
c53b7ddc | 2653 | |
a7180877 DDAG |
2654 | /* |
2655 | * If it's the last part of a host page then we place the host | |
2656 | * page | |
2657 | */ | |
2658 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
28abd200 | 2659 | (block->page_size - 1)) == 0; |
a7180877 DDAG |
2660 | place_source = postcopy_host_page; |
2661 | } | |
c53b7ddc | 2662 | last_host = host; |
a7180877 DDAG |
2663 | |
2664 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 2665 | case RAM_SAVE_FLAG_ZERO: |
a7180877 DDAG |
2666 | ch = qemu_get_byte(f); |
2667 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2668 | if (ch) { | |
2669 | all_zero = false; | |
2670 | } | |
2671 | break; | |
2672 | ||
2673 | case RAM_SAVE_FLAG_PAGE: | |
2674 | all_zero = false; | |
2675 | if (!place_needed || !matching_page_sizes) { | |
2676 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2677 | } else { | |
2678 | /* Avoids the qemu_file copy during postcopy, which is | |
2679 | * going to do a copy later; can only do it when we | |
2680 | * do this read in one go (matching page sizes) | |
2681 | */ | |
2682 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2683 | TARGET_PAGE_SIZE); | |
2684 | } | |
2685 | break; | |
2686 | case RAM_SAVE_FLAG_EOS: | |
2687 | /* normal exit */ | |
2688 | break; | |
2689 | default: | |
2690 | error_report("Unknown combination of migration flags: %#x" | |
2691 | " (postcopy mode)", flags); | |
2692 | ret = -EINVAL; | |
2693 | } | |
2694 | ||
2695 | if (place_needed) { | |
2696 | /* This gets called at the last target page in the host page */ | |
df9ff5e1 DDAG |
2697 | void *place_dest = host + TARGET_PAGE_SIZE - block->page_size; |
2698 | ||
a7180877 | 2699 | if (all_zero) { |
df9ff5e1 DDAG |
2700 | ret = postcopy_place_page_zero(mis, place_dest, |
2701 | block->page_size); | |
a7180877 | 2702 | } else { |
df9ff5e1 DDAG |
2703 | ret = postcopy_place_page(mis, place_dest, |
2704 | place_source, block->page_size); | |
a7180877 DDAG |
2705 | } |
2706 | } | |
2707 | if (!ret) { | |
2708 | ret = qemu_file_get_error(f); | |
2709 | } | |
2710 | } | |
2711 | ||
2712 | return ret; | |
2713 | } | |
2714 | ||
56e93d26 JQ |
2715 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2716 | { | |
edc60127 | 2717 | int flags = 0, ret = 0, invalid_flags = 0; |
56e93d26 JQ |
2718 | static uint64_t seq_iter; |
2719 | int len = 0; | |
a7180877 DDAG |
2720 | /* |
2721 | * If system is running in postcopy mode, page inserts to host memory must | |
2722 | * be atomic | |
2723 | */ | |
2724 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
ef08fb38 DDAG |
2725 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
2726 | bool postcopy_advised = postcopy_state_get() >= POSTCOPY_INCOMING_ADVISE; | |
56e93d26 JQ |
2727 | |
2728 | seq_iter++; | |
2729 | ||
2730 | if (version_id != 4) { | |
2731 | ret = -EINVAL; | |
2732 | } | |
2733 | ||
edc60127 JQ |
2734 | if (!migrate_use_compression()) { |
2735 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
2736 | } | |
56e93d26 JQ |
2737 | /* This RCU critical section can be very long running. |
2738 | * When RCU reclaims in the code start to become numerous, | |
2739 | * it will be necessary to reduce the granularity of this | |
2740 | * critical section. | |
2741 | */ | |
2742 | rcu_read_lock(); | |
a7180877 DDAG |
2743 | |
2744 | if (postcopy_running) { | |
2745 | ret = ram_load_postcopy(f); | |
2746 | } | |
2747 | ||
2748 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2749 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2750 | void *host = NULL; |
56e93d26 JQ |
2751 | uint8_t ch; |
2752 | ||
2753 | addr = qemu_get_be64(f); | |
2754 | flags = addr & ~TARGET_PAGE_MASK; | |
2755 | addr &= TARGET_PAGE_MASK; | |
2756 | ||
edc60127 JQ |
2757 | if (flags & invalid_flags) { |
2758 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
2759 | error_report("Received an unexpected compressed page"); | |
2760 | } | |
2761 | ||
2762 | ret = -EINVAL; | |
2763 | break; | |
2764 | } | |
2765 | ||
bb890ed5 | 2766 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 2767 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
4c4bad48 HZ |
2768 | RAMBlock *block = ram_block_from_stream(f, flags); |
2769 | ||
2770 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2771 | if (!host) { |
2772 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2773 | ret = -EINVAL; | |
2774 | break; | |
2775 | } | |
1db9d8e5 | 2776 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
2777 | } |
2778 | ||
56e93d26 JQ |
2779 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2780 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2781 | /* Synchronize RAM block list */ | |
2782 | total_ram_bytes = addr; | |
2783 | while (!ret && total_ram_bytes) { | |
2784 | RAMBlock *block; | |
56e93d26 JQ |
2785 | char id[256]; |
2786 | ram_addr_t length; | |
2787 | ||
2788 | len = qemu_get_byte(f); | |
2789 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2790 | id[len] = 0; | |
2791 | length = qemu_get_be64(f); | |
2792 | ||
e3dd7493 DDAG |
2793 | block = qemu_ram_block_by_name(id); |
2794 | if (block) { | |
2795 | if (length != block->used_length) { | |
2796 | Error *local_err = NULL; | |
56e93d26 | 2797 | |
fa53a0e5 | 2798 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2799 | &local_err); |
2800 | if (local_err) { | |
2801 | error_report_err(local_err); | |
56e93d26 | 2802 | } |
56e93d26 | 2803 | } |
ef08fb38 DDAG |
2804 | /* For postcopy we need to check hugepage sizes match */ |
2805 | if (postcopy_advised && | |
2806 | block->page_size != qemu_host_page_size) { | |
2807 | uint64_t remote_page_size = qemu_get_be64(f); | |
2808 | if (remote_page_size != block->page_size) { | |
2809 | error_report("Mismatched RAM page size %s " | |
2810 | "(local) %zd != %" PRId64, | |
2811 | id, block->page_size, | |
2812 | remote_page_size); | |
2813 | ret = -EINVAL; | |
2814 | } | |
2815 | } | |
e3dd7493 DDAG |
2816 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2817 | block->idstr); | |
2818 | } else { | |
56e93d26 JQ |
2819 | error_report("Unknown ramblock \"%s\", cannot " |
2820 | "accept migration", id); | |
2821 | ret = -EINVAL; | |
2822 | } | |
2823 | ||
2824 | total_ram_bytes -= length; | |
2825 | } | |
2826 | break; | |
a776aa15 | 2827 | |
bb890ed5 | 2828 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
2829 | ch = qemu_get_byte(f); |
2830 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2831 | break; | |
a776aa15 | 2832 | |
56e93d26 | 2833 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2834 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2835 | break; | |
56e93d26 | 2836 | |
a776aa15 | 2837 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2838 | len = qemu_get_be32(f); |
2839 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2840 | error_report("Invalid compressed data length: %d", len); | |
2841 | ret = -EINVAL; | |
2842 | break; | |
2843 | } | |
c1bc6626 | 2844 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2845 | break; |
a776aa15 | 2846 | |
56e93d26 | 2847 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2848 | if (load_xbzrle(f, addr, host) < 0) { |
2849 | error_report("Failed to decompress XBZRLE page at " | |
2850 | RAM_ADDR_FMT, addr); | |
2851 | ret = -EINVAL; | |
2852 | break; | |
2853 | } | |
2854 | break; | |
2855 | case RAM_SAVE_FLAG_EOS: | |
2856 | /* normal exit */ | |
2857 | break; | |
2858 | default: | |
2859 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2860 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2861 | } else { |
2862 | error_report("Unknown combination of migration flags: %#x", | |
2863 | flags); | |
2864 | ret = -EINVAL; | |
2865 | } | |
2866 | } | |
2867 | if (!ret) { | |
2868 | ret = qemu_file_get_error(f); | |
2869 | } | |
2870 | } | |
2871 | ||
5533b2e9 | 2872 | wait_for_decompress_done(); |
56e93d26 | 2873 | rcu_read_unlock(); |
55c4446b | 2874 | trace_ram_load_complete(ret, seq_iter); |
56e93d26 JQ |
2875 | return ret; |
2876 | } | |
2877 | ||
c6467627 VSO |
2878 | static bool ram_has_postcopy(void *opaque) |
2879 | { | |
2880 | return migrate_postcopy_ram(); | |
2881 | } | |
2882 | ||
56e93d26 | 2883 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 2884 | .save_setup = ram_save_setup, |
56e93d26 | 2885 | .save_live_iterate = ram_save_iterate, |
763c906b | 2886 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2887 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 2888 | .has_postcopy = ram_has_postcopy, |
56e93d26 JQ |
2889 | .save_live_pending = ram_save_pending, |
2890 | .load_state = ram_load, | |
f265e0e4 JQ |
2891 | .save_cleanup = ram_save_cleanup, |
2892 | .load_setup = ram_load_setup, | |
2893 | .load_cleanup = ram_load_cleanup, | |
56e93d26 JQ |
2894 | }; |
2895 | ||
2896 | void ram_mig_init(void) | |
2897 | { | |
2898 | qemu_mutex_init(&XBZRLE.lock); | |
6f37bb8b | 2899 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 2900 | } |