]>
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 JQ |
138 | XBZRLE_cache_lock(); |
139 | ||
140 | if (XBZRLE.cache != NULL) { | |
141 | if (pow2floor(new_size) == migrate_xbzrle_cache_size()) { | |
142 | goto out_new_size; | |
143 | } | |
80f8dfde | 144 | new_cache = cache_init(new_size, TARGET_PAGE_SIZE, errp); |
56e93d26 | 145 | if (!new_cache) { |
56e93d26 JQ |
146 | ret = -1; |
147 | goto out; | |
148 | } | |
149 | ||
150 | cache_fini(XBZRLE.cache); | |
151 | XBZRLE.cache = new_cache; | |
152 | } | |
153 | ||
154 | out_new_size: | |
155 | ret = pow2floor(new_size); | |
156 | out: | |
157 | XBZRLE_cache_unlock(); | |
158 | return ret; | |
159 | } | |
160 | ||
ec481c6c JQ |
161 | /* |
162 | * An outstanding page request, on the source, having been received | |
163 | * and queued | |
164 | */ | |
165 | struct RAMSrcPageRequest { | |
166 | RAMBlock *rb; | |
167 | hwaddr offset; | |
168 | hwaddr len; | |
169 | ||
170 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
171 | }; | |
172 | ||
6f37bb8b JQ |
173 | /* State of RAM for migration */ |
174 | struct RAMState { | |
204b88b8 JQ |
175 | /* QEMUFile used for this migration */ |
176 | QEMUFile *f; | |
6f37bb8b JQ |
177 | /* Last block that we have visited searching for dirty pages */ |
178 | RAMBlock *last_seen_block; | |
179 | /* Last block from where we have sent data */ | |
180 | RAMBlock *last_sent_block; | |
269ace29 JQ |
181 | /* Last dirty target page we have sent */ |
182 | ram_addr_t last_page; | |
6f37bb8b JQ |
183 | /* last ram version we have seen */ |
184 | uint32_t last_version; | |
185 | /* We are in the first round */ | |
186 | bool ram_bulk_stage; | |
8d820d6f JQ |
187 | /* How many times we have dirty too many pages */ |
188 | int dirty_rate_high_cnt; | |
f664da80 JQ |
189 | /* these variables are used for bitmap sync */ |
190 | /* last time we did a full bitmap_sync */ | |
191 | int64_t time_last_bitmap_sync; | |
eac74159 | 192 | /* bytes transferred at start_time */ |
c4bdf0cf | 193 | uint64_t bytes_xfer_prev; |
a66cd90c | 194 | /* number of dirty pages since start_time */ |
68908ed6 | 195 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
196 | /* xbzrle misses since the beginning of the period */ |
197 | uint64_t xbzrle_cache_miss_prev; | |
36040d9c JQ |
198 | /* number of iterations at the beginning of period */ |
199 | uint64_t iterations_prev; | |
23b28c3c JQ |
200 | /* Iterations since start */ |
201 | uint64_t iterations; | |
9360447d | 202 | /* number of dirty bits in the bitmap */ |
2dfaf12e PX |
203 | uint64_t migration_dirty_pages; |
204 | /* protects modification of the bitmap */ | |
108cfae0 | 205 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
206 | /* The RAMBlock used in the last src_page_requests */ |
207 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
208 | /* Queue of outstanding page requests from the destination */ |
209 | QemuMutex src_page_req_mutex; | |
210 | QSIMPLEQ_HEAD(src_page_requests, RAMSrcPageRequest) src_page_requests; | |
6f37bb8b JQ |
211 | }; |
212 | typedef struct RAMState RAMState; | |
213 | ||
53518d94 | 214 | static RAMState *ram_state; |
6f37bb8b | 215 | |
9edabd4d | 216 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 217 | { |
53518d94 | 218 | return ram_state->migration_dirty_pages * TARGET_PAGE_SIZE; |
2f4fde93 JQ |
219 | } |
220 | ||
9360447d | 221 | MigrationStats ram_counters; |
96506894 | 222 | |
b8fb8cb7 DDAG |
223 | /* used by the search for pages to send */ |
224 | struct PageSearchStatus { | |
225 | /* Current block being searched */ | |
226 | RAMBlock *block; | |
a935e30f JQ |
227 | /* Current page to search from */ |
228 | unsigned long page; | |
b8fb8cb7 DDAG |
229 | /* Set once we wrap around */ |
230 | bool complete_round; | |
231 | }; | |
232 | typedef struct PageSearchStatus PageSearchStatus; | |
233 | ||
56e93d26 | 234 | struct CompressParam { |
56e93d26 | 235 | bool done; |
90e56fb4 | 236 | bool quit; |
56e93d26 JQ |
237 | QEMUFile *file; |
238 | QemuMutex mutex; | |
239 | QemuCond cond; | |
240 | RAMBlock *block; | |
241 | ram_addr_t offset; | |
242 | }; | |
243 | typedef struct CompressParam CompressParam; | |
244 | ||
245 | struct DecompressParam { | |
73a8912b | 246 | bool done; |
90e56fb4 | 247 | bool quit; |
56e93d26 JQ |
248 | QemuMutex mutex; |
249 | QemuCond cond; | |
250 | void *des; | |
d341d9f3 | 251 | uint8_t *compbuf; |
56e93d26 JQ |
252 | int len; |
253 | }; | |
254 | typedef struct DecompressParam DecompressParam; | |
255 | ||
256 | static CompressParam *comp_param; | |
257 | static QemuThread *compress_threads; | |
258 | /* comp_done_cond is used to wake up the migration thread when | |
259 | * one of the compression threads has finished the compression. | |
260 | * comp_done_lock is used to co-work with comp_done_cond. | |
261 | */ | |
0d9f9a5c LL |
262 | static QemuMutex comp_done_lock; |
263 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
264 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
265 | static const QEMUFileOps empty_ops = { }; | |
266 | ||
56e93d26 JQ |
267 | static DecompressParam *decomp_param; |
268 | static QemuThread *decompress_threads; | |
73a8912b LL |
269 | static QemuMutex decomp_done_lock; |
270 | static QemuCond decomp_done_cond; | |
56e93d26 | 271 | |
a7a9a88f LL |
272 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
273 | ram_addr_t offset); | |
56e93d26 JQ |
274 | |
275 | static void *do_data_compress(void *opaque) | |
276 | { | |
277 | CompressParam *param = opaque; | |
a7a9a88f LL |
278 | RAMBlock *block; |
279 | ram_addr_t offset; | |
56e93d26 | 280 | |
a7a9a88f | 281 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 282 | while (!param->quit) { |
a7a9a88f LL |
283 | if (param->block) { |
284 | block = param->block; | |
285 | offset = param->offset; | |
286 | param->block = NULL; | |
287 | qemu_mutex_unlock(¶m->mutex); | |
288 | ||
289 | do_compress_ram_page(param->file, block, offset); | |
290 | ||
0d9f9a5c | 291 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 292 | param->done = true; |
0d9f9a5c LL |
293 | qemu_cond_signal(&comp_done_cond); |
294 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
295 | |
296 | qemu_mutex_lock(¶m->mutex); | |
297 | } else { | |
56e93d26 JQ |
298 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
299 | } | |
56e93d26 | 300 | } |
a7a9a88f | 301 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
302 | |
303 | return NULL; | |
304 | } | |
305 | ||
306 | static inline void terminate_compression_threads(void) | |
307 | { | |
308 | int idx, thread_count; | |
309 | ||
310 | thread_count = migrate_compress_threads(); | |
3d0684b2 | 311 | |
56e93d26 JQ |
312 | for (idx = 0; idx < thread_count; idx++) { |
313 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 314 | comp_param[idx].quit = true; |
56e93d26 JQ |
315 | qemu_cond_signal(&comp_param[idx].cond); |
316 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
317 | } | |
318 | } | |
319 | ||
f0afa331 | 320 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
321 | { |
322 | int i, thread_count; | |
323 | ||
324 | if (!migrate_use_compression()) { | |
325 | return; | |
326 | } | |
327 | terminate_compression_threads(); | |
328 | thread_count = migrate_compress_threads(); | |
329 | for (i = 0; i < thread_count; i++) { | |
330 | qemu_thread_join(compress_threads + i); | |
331 | qemu_fclose(comp_param[i].file); | |
332 | qemu_mutex_destroy(&comp_param[i].mutex); | |
333 | qemu_cond_destroy(&comp_param[i].cond); | |
334 | } | |
0d9f9a5c LL |
335 | qemu_mutex_destroy(&comp_done_lock); |
336 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
337 | g_free(compress_threads); |
338 | g_free(comp_param); | |
56e93d26 JQ |
339 | compress_threads = NULL; |
340 | comp_param = NULL; | |
56e93d26 JQ |
341 | } |
342 | ||
f0afa331 | 343 | static void compress_threads_save_setup(void) |
56e93d26 JQ |
344 | { |
345 | int i, thread_count; | |
346 | ||
347 | if (!migrate_use_compression()) { | |
348 | return; | |
349 | } | |
56e93d26 JQ |
350 | thread_count = migrate_compress_threads(); |
351 | compress_threads = g_new0(QemuThread, thread_count); | |
352 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
353 | qemu_cond_init(&comp_done_cond); |
354 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 355 | for (i = 0; i < thread_count; i++) { |
e110aa91 C |
356 | /* comp_param[i].file is just used as a dummy buffer to save data, |
357 | * set its ops to empty. | |
56e93d26 JQ |
358 | */ |
359 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops); | |
360 | comp_param[i].done = true; | |
90e56fb4 | 361 | comp_param[i].quit = false; |
56e93d26 JQ |
362 | qemu_mutex_init(&comp_param[i].mutex); |
363 | qemu_cond_init(&comp_param[i].cond); | |
364 | qemu_thread_create(compress_threads + i, "compress", | |
365 | do_data_compress, comp_param + i, | |
366 | QEMU_THREAD_JOINABLE); | |
367 | } | |
368 | } | |
369 | ||
f986c3d2 JQ |
370 | /* Multiple fd's */ |
371 | ||
372 | struct MultiFDSendParams { | |
373 | uint8_t id; | |
374 | char *name; | |
375 | QemuThread thread; | |
376 | QemuSemaphore sem; | |
377 | QemuMutex mutex; | |
378 | bool quit; | |
379 | }; | |
380 | typedef struct MultiFDSendParams MultiFDSendParams; | |
381 | ||
382 | struct { | |
383 | MultiFDSendParams *params; | |
384 | /* number of created threads */ | |
385 | int count; | |
386 | } *multifd_send_state; | |
387 | ||
388 | static void terminate_multifd_send_threads(Error *errp) | |
389 | { | |
390 | int i; | |
391 | ||
392 | for (i = 0; i < multifd_send_state->count; i++) { | |
393 | MultiFDSendParams *p = &multifd_send_state->params[i]; | |
394 | ||
395 | qemu_mutex_lock(&p->mutex); | |
396 | p->quit = true; | |
397 | qemu_sem_post(&p->sem); | |
398 | qemu_mutex_unlock(&p->mutex); | |
399 | } | |
400 | } | |
401 | ||
402 | int multifd_save_cleanup(Error **errp) | |
403 | { | |
404 | int i; | |
405 | int ret = 0; | |
406 | ||
407 | if (!migrate_use_multifd()) { | |
408 | return 0; | |
409 | } | |
410 | terminate_multifd_send_threads(NULL); | |
411 | for (i = 0; i < multifd_send_state->count; i++) { | |
412 | MultiFDSendParams *p = &multifd_send_state->params[i]; | |
413 | ||
414 | qemu_thread_join(&p->thread); | |
415 | qemu_mutex_destroy(&p->mutex); | |
416 | qemu_sem_destroy(&p->sem); | |
417 | g_free(p->name); | |
418 | p->name = NULL; | |
419 | } | |
420 | g_free(multifd_send_state->params); | |
421 | multifd_send_state->params = NULL; | |
422 | g_free(multifd_send_state); | |
423 | multifd_send_state = NULL; | |
424 | return ret; | |
425 | } | |
426 | ||
427 | static void *multifd_send_thread(void *opaque) | |
428 | { | |
429 | MultiFDSendParams *p = opaque; | |
430 | ||
431 | while (true) { | |
432 | qemu_mutex_lock(&p->mutex); | |
433 | if (p->quit) { | |
434 | qemu_mutex_unlock(&p->mutex); | |
435 | break; | |
436 | } | |
437 | qemu_mutex_unlock(&p->mutex); | |
438 | qemu_sem_wait(&p->sem); | |
439 | } | |
440 | ||
441 | return NULL; | |
442 | } | |
443 | ||
444 | int multifd_save_setup(void) | |
445 | { | |
446 | int thread_count; | |
447 | uint8_t i; | |
448 | ||
449 | if (!migrate_use_multifd()) { | |
450 | return 0; | |
451 | } | |
452 | thread_count = migrate_multifd_channels(); | |
453 | multifd_send_state = g_malloc0(sizeof(*multifd_send_state)); | |
454 | multifd_send_state->params = g_new0(MultiFDSendParams, thread_count); | |
455 | multifd_send_state->count = 0; | |
456 | for (i = 0; i < thread_count; i++) { | |
457 | MultiFDSendParams *p = &multifd_send_state->params[i]; | |
458 | ||
459 | qemu_mutex_init(&p->mutex); | |
460 | qemu_sem_init(&p->sem, 0); | |
461 | p->quit = false; | |
462 | p->id = i; | |
463 | p->name = g_strdup_printf("multifdsend_%d", i); | |
464 | qemu_thread_create(&p->thread, p->name, multifd_send_thread, p, | |
465 | QEMU_THREAD_JOINABLE); | |
466 | ||
467 | multifd_send_state->count++; | |
468 | } | |
469 | return 0; | |
470 | } | |
471 | ||
472 | struct MultiFDRecvParams { | |
473 | uint8_t id; | |
474 | char *name; | |
475 | QemuThread thread; | |
476 | QemuSemaphore sem; | |
477 | QemuMutex mutex; | |
478 | bool quit; | |
479 | }; | |
480 | typedef struct MultiFDRecvParams MultiFDRecvParams; | |
481 | ||
482 | struct { | |
483 | MultiFDRecvParams *params; | |
484 | /* number of created threads */ | |
485 | int count; | |
486 | } *multifd_recv_state; | |
487 | ||
488 | static void terminate_multifd_recv_threads(Error *errp) | |
489 | { | |
490 | int i; | |
491 | ||
492 | for (i = 0; i < multifd_recv_state->count; i++) { | |
493 | MultiFDRecvParams *p = &multifd_recv_state->params[i]; | |
494 | ||
495 | qemu_mutex_lock(&p->mutex); | |
496 | p->quit = true; | |
497 | qemu_sem_post(&p->sem); | |
498 | qemu_mutex_unlock(&p->mutex); | |
499 | } | |
500 | } | |
501 | ||
502 | int multifd_load_cleanup(Error **errp) | |
503 | { | |
504 | int i; | |
505 | int ret = 0; | |
506 | ||
507 | if (!migrate_use_multifd()) { | |
508 | return 0; | |
509 | } | |
510 | terminate_multifd_recv_threads(NULL); | |
511 | for (i = 0; i < multifd_recv_state->count; i++) { | |
512 | MultiFDRecvParams *p = &multifd_recv_state->params[i]; | |
513 | ||
514 | qemu_thread_join(&p->thread); | |
515 | qemu_mutex_destroy(&p->mutex); | |
516 | qemu_sem_destroy(&p->sem); | |
517 | g_free(p->name); | |
518 | p->name = NULL; | |
519 | } | |
520 | g_free(multifd_recv_state->params); | |
521 | multifd_recv_state->params = NULL; | |
522 | g_free(multifd_recv_state); | |
523 | multifd_recv_state = NULL; | |
524 | ||
525 | return ret; | |
526 | } | |
527 | ||
528 | static void *multifd_recv_thread(void *opaque) | |
529 | { | |
530 | MultiFDRecvParams *p = opaque; | |
531 | ||
532 | while (true) { | |
533 | qemu_mutex_lock(&p->mutex); | |
534 | if (p->quit) { | |
535 | qemu_mutex_unlock(&p->mutex); | |
536 | break; | |
537 | } | |
538 | qemu_mutex_unlock(&p->mutex); | |
539 | qemu_sem_wait(&p->sem); | |
540 | } | |
541 | ||
542 | return NULL; | |
543 | } | |
544 | ||
545 | int multifd_load_setup(void) | |
546 | { | |
547 | int thread_count; | |
548 | uint8_t i; | |
549 | ||
550 | if (!migrate_use_multifd()) { | |
551 | return 0; | |
552 | } | |
553 | thread_count = migrate_multifd_channels(); | |
554 | multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state)); | |
555 | multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count); | |
556 | multifd_recv_state->count = 0; | |
557 | for (i = 0; i < thread_count; i++) { | |
558 | MultiFDRecvParams *p = &multifd_recv_state->params[i]; | |
559 | ||
560 | qemu_mutex_init(&p->mutex); | |
561 | qemu_sem_init(&p->sem, 0); | |
562 | p->quit = false; | |
563 | p->id = i; | |
564 | p->name = g_strdup_printf("multifdrecv_%d", i); | |
565 | qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p, | |
566 | QEMU_THREAD_JOINABLE); | |
567 | multifd_recv_state->count++; | |
568 | } | |
569 | return 0; | |
570 | } | |
571 | ||
56e93d26 | 572 | /** |
3d0684b2 | 573 | * save_page_header: write page header to wire |
56e93d26 JQ |
574 | * |
575 | * If this is the 1st block, it also writes the block identification | |
576 | * | |
3d0684b2 | 577 | * Returns the number of bytes written |
56e93d26 JQ |
578 | * |
579 | * @f: QEMUFile where to send the data | |
580 | * @block: block that contains the page we want to send | |
581 | * @offset: offset inside the block for the page | |
582 | * in the lower bits, it contains flags | |
583 | */ | |
2bf3aa85 JQ |
584 | static size_t save_page_header(RAMState *rs, QEMUFile *f, RAMBlock *block, |
585 | ram_addr_t offset) | |
56e93d26 | 586 | { |
9f5f380b | 587 | size_t size, len; |
56e93d26 | 588 | |
24795694 JQ |
589 | if (block == rs->last_sent_block) { |
590 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
591 | } | |
2bf3aa85 | 592 | qemu_put_be64(f, offset); |
56e93d26 JQ |
593 | size = 8; |
594 | ||
595 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 596 | len = strlen(block->idstr); |
2bf3aa85 JQ |
597 | qemu_put_byte(f, len); |
598 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 599 | size += 1 + len; |
24795694 | 600 | rs->last_sent_block = block; |
56e93d26 JQ |
601 | } |
602 | return size; | |
603 | } | |
604 | ||
3d0684b2 JQ |
605 | /** |
606 | * mig_throttle_guest_down: throotle down the guest | |
607 | * | |
608 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
609 | * writes. If guest dirty memory rate is reduced below the rate at | |
610 | * which we can transfer pages to the destination then we should be | |
611 | * able to complete migration. Some workloads dirty memory way too | |
612 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 JH |
613 | */ |
614 | static void mig_throttle_guest_down(void) | |
615 | { | |
616 | MigrationState *s = migrate_get_current(); | |
2594f56d DB |
617 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
618 | uint64_t pct_icrement = s->parameters.cpu_throttle_increment; | |
070afca2 JH |
619 | |
620 | /* We have not started throttling yet. Let's start it. */ | |
621 | if (!cpu_throttle_active()) { | |
622 | cpu_throttle_set(pct_initial); | |
623 | } else { | |
624 | /* Throttling already on, just increase the rate */ | |
625 | cpu_throttle_set(cpu_throttle_get_percentage() + pct_icrement); | |
626 | } | |
627 | } | |
628 | ||
3d0684b2 JQ |
629 | /** |
630 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
631 | * | |
6f37bb8b | 632 | * @rs: current RAM state |
3d0684b2 JQ |
633 | * @current_addr: address for the zero page |
634 | * | |
635 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
636 | * The important thing is that a stale (not-yet-0'd) page be replaced |
637 | * by the new data. | |
638 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 639 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 640 | */ |
6f37bb8b | 641 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 642 | { |
6f37bb8b | 643 | if (rs->ram_bulk_stage || !migrate_use_xbzrle()) { |
56e93d26 JQ |
644 | return; |
645 | } | |
646 | ||
647 | /* We don't care if this fails to allocate a new cache page | |
648 | * as long as it updated an old one */ | |
c00e0928 | 649 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 650 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
651 | } |
652 | ||
653 | #define ENCODING_FLAG_XBZRLE 0x1 | |
654 | ||
655 | /** | |
656 | * save_xbzrle_page: compress and send current page | |
657 | * | |
658 | * Returns: 1 means that we wrote the page | |
659 | * 0 means that page is identical to the one already sent | |
660 | * -1 means that xbzrle would be longer than normal | |
661 | * | |
5a987738 | 662 | * @rs: current RAM state |
3d0684b2 JQ |
663 | * @current_data: pointer to the address of the page contents |
664 | * @current_addr: addr of the page | |
56e93d26 JQ |
665 | * @block: block that contains the page we want to send |
666 | * @offset: offset inside the block for the page | |
667 | * @last_stage: if we are at the completion stage | |
56e93d26 | 668 | */ |
204b88b8 | 669 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 670 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 671 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
672 | { |
673 | int encoded_len = 0, bytes_xbzrle; | |
674 | uint8_t *prev_cached_page; | |
675 | ||
9360447d JQ |
676 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
677 | ram_counters.dirty_sync_count)) { | |
678 | xbzrle_counters.cache_miss++; | |
56e93d26 JQ |
679 | if (!last_stage) { |
680 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
9360447d | 681 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
682 | return -1; |
683 | } else { | |
684 | /* update *current_data when the page has been | |
685 | inserted into cache */ | |
686 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
687 | } | |
688 | } | |
689 | return -1; | |
690 | } | |
691 | ||
692 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); | |
693 | ||
694 | /* save current buffer into memory */ | |
695 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
696 | ||
697 | /* XBZRLE encoding (if there is no overflow) */ | |
698 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
699 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
700 | TARGET_PAGE_SIZE); | |
701 | if (encoded_len == 0) { | |
55c4446b | 702 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
703 | return 0; |
704 | } else if (encoded_len == -1) { | |
55c4446b | 705 | trace_save_xbzrle_page_overflow(); |
9360447d | 706 | xbzrle_counters.overflow++; |
56e93d26 JQ |
707 | /* update data in the cache */ |
708 | if (!last_stage) { | |
709 | memcpy(prev_cached_page, *current_data, TARGET_PAGE_SIZE); | |
710 | *current_data = prev_cached_page; | |
711 | } | |
712 | return -1; | |
713 | } | |
714 | ||
715 | /* we need to update the data in the cache, in order to get the same data */ | |
716 | if (!last_stage) { | |
717 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
718 | } | |
719 | ||
720 | /* Send XBZRLE based compressed page */ | |
2bf3aa85 | 721 | bytes_xbzrle = save_page_header(rs, rs->f, block, |
204b88b8 JQ |
722 | offset | RAM_SAVE_FLAG_XBZRLE); |
723 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
724 | qemu_put_be16(rs->f, encoded_len); | |
725 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 726 | bytes_xbzrle += encoded_len + 1 + 2; |
9360447d JQ |
727 | xbzrle_counters.pages++; |
728 | xbzrle_counters.bytes += bytes_xbzrle; | |
729 | ram_counters.transferred += bytes_xbzrle; | |
56e93d26 JQ |
730 | |
731 | return 1; | |
732 | } | |
733 | ||
3d0684b2 JQ |
734 | /** |
735 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 736 | * |
3d0684b2 JQ |
737 | * Called with rcu_read_lock() to protect migration_bitmap |
738 | * | |
739 | * Returns the byte offset within memory region of the start of a dirty page | |
740 | * | |
6f37bb8b | 741 | * @rs: current RAM state |
3d0684b2 | 742 | * @rb: RAMBlock where to search for dirty pages |
a935e30f | 743 | * @start: page where we start the search |
f3f491fc | 744 | */ |
56e93d26 | 745 | static inline |
a935e30f | 746 | unsigned long migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
f20e2865 | 747 | unsigned long start) |
56e93d26 | 748 | { |
6b6712ef JQ |
749 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
750 | unsigned long *bitmap = rb->bmap; | |
56e93d26 JQ |
751 | unsigned long next; |
752 | ||
6b6712ef JQ |
753 | if (rs->ram_bulk_stage && start > 0) { |
754 | next = start + 1; | |
56e93d26 | 755 | } else { |
6b6712ef | 756 | next = find_next_bit(bitmap, size, start); |
56e93d26 JQ |
757 | } |
758 | ||
6b6712ef | 759 | return next; |
56e93d26 JQ |
760 | } |
761 | ||
06b10688 | 762 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
763 | RAMBlock *rb, |
764 | unsigned long page) | |
a82d593b DDAG |
765 | { |
766 | bool ret; | |
a82d593b | 767 | |
6b6712ef | 768 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b DDAG |
769 | |
770 | if (ret) { | |
0d8ec885 | 771 | rs->migration_dirty_pages--; |
a82d593b DDAG |
772 | } |
773 | return ret; | |
774 | } | |
775 | ||
15440dd5 JQ |
776 | static void migration_bitmap_sync_range(RAMState *rs, RAMBlock *rb, |
777 | ram_addr_t start, ram_addr_t length) | |
56e93d26 | 778 | { |
0d8ec885 | 779 | rs->migration_dirty_pages += |
6b6712ef | 780 | cpu_physical_memory_sync_dirty_bitmap(rb, start, length, |
0d8ec885 | 781 | &rs->num_dirty_pages_period); |
56e93d26 JQ |
782 | } |
783 | ||
3d0684b2 JQ |
784 | /** |
785 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
786 | * | |
787 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
788 | * | |
789 | * For VMs with just normal pages this is equivalent to the host page | |
790 | * size. If it's got some huge pages then it's the OR of all the | |
791 | * different page sizes. | |
e8ca1db2 DDAG |
792 | */ |
793 | uint64_t ram_pagesize_summary(void) | |
794 | { | |
795 | RAMBlock *block; | |
796 | uint64_t summary = 0; | |
797 | ||
99e15582 | 798 | RAMBLOCK_FOREACH(block) { |
e8ca1db2 DDAG |
799 | summary |= block->page_size; |
800 | } | |
801 | ||
802 | return summary; | |
803 | } | |
804 | ||
8d820d6f | 805 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
806 | { |
807 | RAMBlock *block; | |
56e93d26 | 808 | int64_t end_time; |
c4bdf0cf | 809 | uint64_t bytes_xfer_now; |
56e93d26 | 810 | |
9360447d | 811 | ram_counters.dirty_sync_count++; |
56e93d26 | 812 | |
f664da80 JQ |
813 | if (!rs->time_last_bitmap_sync) { |
814 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
815 | } |
816 | ||
817 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 818 | memory_global_dirty_log_sync(); |
56e93d26 | 819 | |
108cfae0 | 820 | qemu_mutex_lock(&rs->bitmap_mutex); |
56e93d26 | 821 | rcu_read_lock(); |
99e15582 | 822 | RAMBLOCK_FOREACH(block) { |
15440dd5 | 823 | migration_bitmap_sync_range(rs, block, 0, block->used_length); |
56e93d26 JQ |
824 | } |
825 | rcu_read_unlock(); | |
108cfae0 | 826 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 827 | |
a66cd90c | 828 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 829 | |
56e93d26 JQ |
830 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
831 | ||
832 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 833 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
d693c6f1 | 834 | /* calculate period counters */ |
9360447d | 835 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 |
d693c6f1 | 836 | / (end_time - rs->time_last_bitmap_sync); |
9360447d | 837 | bytes_xfer_now = ram_counters.transferred; |
d693c6f1 | 838 | |
9ac78b61 PL |
839 | /* During block migration the auto-converge logic incorrectly detects |
840 | * that ram migration makes no progress. Avoid this by disabling the | |
841 | * throttling logic during the bulk phase of block migration. */ | |
842 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
56e93d26 JQ |
843 | /* The following detection logic can be refined later. For now: |
844 | Check to see if the dirtied bytes is 50% more than the approx. | |
845 | amount of bytes that just got transferred since the last time we | |
070afca2 JH |
846 | were in this routine. If that happens twice, start or increase |
847 | throttling */ | |
070afca2 | 848 | |
d693c6f1 | 849 | if ((rs->num_dirty_pages_period * TARGET_PAGE_SIZE > |
eac74159 | 850 | (bytes_xfer_now - rs->bytes_xfer_prev) / 2) && |
b4a3c64b | 851 | (++rs->dirty_rate_high_cnt >= 2)) { |
56e93d26 | 852 | trace_migration_throttle(); |
8d820d6f | 853 | rs->dirty_rate_high_cnt = 0; |
070afca2 | 854 | mig_throttle_guest_down(); |
d693c6f1 | 855 | } |
56e93d26 | 856 | } |
070afca2 | 857 | |
56e93d26 | 858 | if (migrate_use_xbzrle()) { |
23b28c3c | 859 | if (rs->iterations_prev != rs->iterations) { |
9360447d JQ |
860 | xbzrle_counters.cache_miss_rate = |
861 | (double)(xbzrle_counters.cache_miss - | |
b5833fde | 862 | rs->xbzrle_cache_miss_prev) / |
23b28c3c | 863 | (rs->iterations - rs->iterations_prev); |
56e93d26 | 864 | } |
23b28c3c | 865 | rs->iterations_prev = rs->iterations; |
9360447d | 866 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
56e93d26 | 867 | } |
d693c6f1 FF |
868 | |
869 | /* reset period counters */ | |
f664da80 | 870 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 871 | rs->num_dirty_pages_period = 0; |
d2a4d85a | 872 | rs->bytes_xfer_prev = bytes_xfer_now; |
56e93d26 | 873 | } |
4addcd4f | 874 | if (migrate_use_events()) { |
9360447d | 875 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count, NULL); |
4addcd4f | 876 | } |
56e93d26 JQ |
877 | } |
878 | ||
879 | /** | |
3d0684b2 | 880 | * save_zero_page: send the zero page to the stream |
56e93d26 | 881 | * |
3d0684b2 | 882 | * Returns the number of pages written. |
56e93d26 | 883 | * |
f7ccd61b | 884 | * @rs: current RAM state |
56e93d26 JQ |
885 | * @block: block that contains the page we want to send |
886 | * @offset: offset inside the block for the page | |
887 | * @p: pointer to the page | |
56e93d26 | 888 | */ |
ce25d337 JQ |
889 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, |
890 | uint8_t *p) | |
56e93d26 JQ |
891 | { |
892 | int pages = -1; | |
893 | ||
894 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
9360447d JQ |
895 | ram_counters.duplicate++; |
896 | ram_counters.transferred += | |
bb890ed5 | 897 | save_page_header(rs, rs->f, block, offset | RAM_SAVE_FLAG_ZERO); |
ce25d337 | 898 | qemu_put_byte(rs->f, 0); |
9360447d | 899 | ram_counters.transferred += 1; |
56e93d26 JQ |
900 | pages = 1; |
901 | } | |
902 | ||
903 | return pages; | |
904 | } | |
905 | ||
5727309d | 906 | static void ram_release_pages(const char *rbname, uint64_t offset, int pages) |
53f09a10 | 907 | { |
5727309d | 908 | if (!migrate_release_ram() || !migration_in_postcopy()) { |
53f09a10 PB |
909 | return; |
910 | } | |
911 | ||
aaa2064c | 912 | ram_discard_range(rbname, offset, pages << TARGET_PAGE_BITS); |
53f09a10 PB |
913 | } |
914 | ||
56e93d26 | 915 | /** |
3d0684b2 | 916 | * ram_save_page: send the given page to the stream |
56e93d26 | 917 | * |
3d0684b2 | 918 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
919 | * < 0 - error |
920 | * >=0 - Number of pages written - this might legally be 0 | |
921 | * if xbzrle noticed the page was the same. | |
56e93d26 | 922 | * |
6f37bb8b | 923 | * @rs: current RAM state |
56e93d26 JQ |
924 | * @block: block that contains the page we want to send |
925 | * @offset: offset inside the block for the page | |
926 | * @last_stage: if we are at the completion stage | |
56e93d26 | 927 | */ |
a0a8aa14 | 928 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
929 | { |
930 | int pages = -1; | |
931 | uint64_t bytes_xmit; | |
932 | ram_addr_t current_addr; | |
56e93d26 JQ |
933 | uint8_t *p; |
934 | int ret; | |
935 | bool send_async = true; | |
a08f6890 | 936 | RAMBlock *block = pss->block; |
a935e30f | 937 | ram_addr_t offset = pss->page << TARGET_PAGE_BITS; |
56e93d26 | 938 | |
2f68e399 | 939 | p = block->host + offset; |
1db9d8e5 | 940 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 JQ |
941 | |
942 | /* In doubt sent page as normal */ | |
943 | bytes_xmit = 0; | |
ce25d337 | 944 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
945 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
946 | if (bytes_xmit) { | |
9360447d | 947 | ram_counters.transferred += bytes_xmit; |
56e93d26 JQ |
948 | pages = 1; |
949 | } | |
950 | ||
951 | XBZRLE_cache_lock(); | |
952 | ||
953 | current_addr = block->offset + offset; | |
954 | ||
56e93d26 JQ |
955 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
956 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
957 | if (bytes_xmit > 0) { | |
9360447d | 958 | ram_counters.normal++; |
56e93d26 | 959 | } else if (bytes_xmit == 0) { |
9360447d | 960 | ram_counters.duplicate++; |
56e93d26 JQ |
961 | } |
962 | } | |
963 | } else { | |
ce25d337 | 964 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 JQ |
965 | if (pages > 0) { |
966 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
967 | * page would be stale | |
968 | */ | |
6f37bb8b | 969 | xbzrle_cache_zero_page(rs, current_addr); |
a935e30f | 970 | ram_release_pages(block->idstr, offset, pages); |
6f37bb8b | 971 | } else if (!rs->ram_bulk_stage && |
5727309d | 972 | !migration_in_postcopy() && migrate_use_xbzrle()) { |
204b88b8 | 973 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
072c2511 | 974 | offset, last_stage); |
56e93d26 JQ |
975 | if (!last_stage) { |
976 | /* Can't send this cached data async, since the cache page | |
977 | * might get updated before it gets to the wire | |
978 | */ | |
979 | send_async = false; | |
980 | } | |
981 | } | |
982 | } | |
983 | ||
984 | /* XBZRLE overflow or normal page */ | |
985 | if (pages == -1) { | |
9360447d JQ |
986 | ram_counters.transferred += |
987 | save_page_header(rs, rs->f, block, offset | RAM_SAVE_FLAG_PAGE); | |
56e93d26 | 988 | if (send_async) { |
ce25d337 | 989 | qemu_put_buffer_async(rs->f, p, TARGET_PAGE_SIZE, |
53f09a10 | 990 | migrate_release_ram() & |
5727309d | 991 | migration_in_postcopy()); |
56e93d26 | 992 | } else { |
ce25d337 | 993 | qemu_put_buffer(rs->f, p, TARGET_PAGE_SIZE); |
56e93d26 | 994 | } |
9360447d | 995 | ram_counters.transferred += TARGET_PAGE_SIZE; |
56e93d26 | 996 | pages = 1; |
9360447d | 997 | ram_counters.normal++; |
56e93d26 JQ |
998 | } |
999 | ||
1000 | XBZRLE_cache_unlock(); | |
1001 | ||
1002 | return pages; | |
1003 | } | |
1004 | ||
a7a9a88f LL |
1005 | static int do_compress_ram_page(QEMUFile *f, RAMBlock *block, |
1006 | ram_addr_t offset) | |
56e93d26 | 1007 | { |
53518d94 | 1008 | RAMState *rs = ram_state; |
56e93d26 | 1009 | int bytes_sent, blen; |
a7a9a88f | 1010 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
56e93d26 | 1011 | |
2bf3aa85 | 1012 | bytes_sent = save_page_header(rs, f, block, offset | |
56e93d26 | 1013 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
a7a9a88f | 1014 | blen = qemu_put_compression_data(f, p, TARGET_PAGE_SIZE, |
56e93d26 | 1015 | migrate_compress_level()); |
b3be2896 LL |
1016 | if (blen < 0) { |
1017 | bytes_sent = 0; | |
1018 | qemu_file_set_error(migrate_get_current()->to_dst_file, blen); | |
1019 | error_report("compressed data failed!"); | |
1020 | } else { | |
1021 | bytes_sent += blen; | |
5727309d | 1022 | ram_release_pages(block->idstr, offset & TARGET_PAGE_MASK, 1); |
b3be2896 | 1023 | } |
56e93d26 JQ |
1024 | |
1025 | return bytes_sent; | |
1026 | } | |
1027 | ||
ce25d337 | 1028 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
1029 | { |
1030 | int idx, len, thread_count; | |
1031 | ||
1032 | if (!migrate_use_compression()) { | |
1033 | return; | |
1034 | } | |
1035 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1036 | |
0d9f9a5c | 1037 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1038 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1039 | while (!comp_param[idx].done) { |
0d9f9a5c | 1040 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1041 | } |
a7a9a88f | 1042 | } |
0d9f9a5c | 1043 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1044 | |
1045 | for (idx = 0; idx < thread_count; idx++) { | |
1046 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1047 | if (!comp_param[idx].quit) { |
ce25d337 | 1048 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
9360447d | 1049 | ram_counters.transferred += len; |
56e93d26 | 1050 | } |
a7a9a88f | 1051 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1052 | } |
1053 | } | |
1054 | ||
1055 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1056 | ram_addr_t offset) | |
1057 | { | |
1058 | param->block = block; | |
1059 | param->offset = offset; | |
1060 | } | |
1061 | ||
ce25d337 JQ |
1062 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
1063 | ram_addr_t offset) | |
56e93d26 JQ |
1064 | { |
1065 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1066 | ||
1067 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1068 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 JQ |
1069 | while (true) { |
1070 | for (idx = 0; idx < thread_count; idx++) { | |
1071 | if (comp_param[idx].done) { | |
a7a9a88f | 1072 | comp_param[idx].done = false; |
ce25d337 | 1073 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
a7a9a88f | 1074 | qemu_mutex_lock(&comp_param[idx].mutex); |
56e93d26 | 1075 | set_compress_params(&comp_param[idx], block, offset); |
a7a9a88f LL |
1076 | qemu_cond_signal(&comp_param[idx].cond); |
1077 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
56e93d26 | 1078 | pages = 1; |
9360447d JQ |
1079 | ram_counters.normal++; |
1080 | ram_counters.transferred += bytes_xmit; | |
56e93d26 JQ |
1081 | break; |
1082 | } | |
1083 | } | |
1084 | if (pages > 0) { | |
1085 | break; | |
1086 | } else { | |
0d9f9a5c | 1087 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 JQ |
1088 | } |
1089 | } | |
0d9f9a5c | 1090 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1091 | |
1092 | return pages; | |
1093 | } | |
1094 | ||
1095 | /** | |
1096 | * ram_save_compressed_page: compress the given page and send it to the stream | |
1097 | * | |
3d0684b2 | 1098 | * Returns the number of pages written. |
56e93d26 | 1099 | * |
6f37bb8b | 1100 | * @rs: current RAM state |
56e93d26 JQ |
1101 | * @block: block that contains the page we want to send |
1102 | * @offset: offset inside the block for the page | |
1103 | * @last_stage: if we are at the completion stage | |
56e93d26 | 1104 | */ |
a0a8aa14 JQ |
1105 | static int ram_save_compressed_page(RAMState *rs, PageSearchStatus *pss, |
1106 | bool last_stage) | |
56e93d26 JQ |
1107 | { |
1108 | int pages = -1; | |
fc50438e | 1109 | uint64_t bytes_xmit = 0; |
56e93d26 | 1110 | uint8_t *p; |
fc50438e | 1111 | int ret, blen; |
a08f6890 | 1112 | RAMBlock *block = pss->block; |
a935e30f | 1113 | ram_addr_t offset = pss->page << TARGET_PAGE_BITS; |
56e93d26 | 1114 | |
2f68e399 | 1115 | p = block->host + offset; |
56e93d26 | 1116 | |
ce25d337 | 1117 | ret = ram_control_save_page(rs->f, block->offset, |
56e93d26 JQ |
1118 | offset, TARGET_PAGE_SIZE, &bytes_xmit); |
1119 | if (bytes_xmit) { | |
9360447d | 1120 | ram_counters.transferred += bytes_xmit; |
56e93d26 JQ |
1121 | pages = 1; |
1122 | } | |
56e93d26 JQ |
1123 | if (ret != RAM_SAVE_CONTROL_NOT_SUPP) { |
1124 | if (ret != RAM_SAVE_CONTROL_DELAYED) { | |
1125 | if (bytes_xmit > 0) { | |
9360447d | 1126 | ram_counters.normal++; |
56e93d26 | 1127 | } else if (bytes_xmit == 0) { |
9360447d | 1128 | ram_counters.duplicate++; |
56e93d26 JQ |
1129 | } |
1130 | } | |
1131 | } else { | |
1132 | /* When starting the process of a new block, the first page of | |
1133 | * the block should be sent out before other pages in the same | |
1134 | * block, and all the pages in last block should have been sent | |
1135 | * out, keeping this order is important, because the 'cont' flag | |
1136 | * is used to avoid resending the block name. | |
1137 | */ | |
6f37bb8b | 1138 | if (block != rs->last_sent_block) { |
ce25d337 JQ |
1139 | flush_compressed_data(rs); |
1140 | pages = save_zero_page(rs, block, offset, p); | |
56e93d26 | 1141 | if (pages == -1) { |
fc50438e | 1142 | /* Make sure the first page is sent out before other pages */ |
2bf3aa85 | 1143 | bytes_xmit = save_page_header(rs, rs->f, block, offset | |
fc50438e | 1144 | RAM_SAVE_FLAG_COMPRESS_PAGE); |
ce25d337 | 1145 | blen = qemu_put_compression_data(rs->f, p, TARGET_PAGE_SIZE, |
fc50438e LL |
1146 | migrate_compress_level()); |
1147 | if (blen > 0) { | |
9360447d JQ |
1148 | ram_counters.transferred += bytes_xmit + blen; |
1149 | ram_counters.normal++; | |
b3be2896 | 1150 | pages = 1; |
fc50438e | 1151 | } else { |
ce25d337 | 1152 | qemu_file_set_error(rs->f, blen); |
fc50438e | 1153 | error_report("compressed data failed!"); |
b3be2896 | 1154 | } |
56e93d26 | 1155 | } |
53f09a10 | 1156 | if (pages > 0) { |
a935e30f | 1157 | ram_release_pages(block->idstr, offset, pages); |
53f09a10 | 1158 | } |
56e93d26 | 1159 | } else { |
ce25d337 | 1160 | pages = save_zero_page(rs, block, offset, p); |
56e93d26 | 1161 | if (pages == -1) { |
ce25d337 | 1162 | pages = compress_page_with_multi_thread(rs, block, offset); |
53f09a10 | 1163 | } else { |
a935e30f | 1164 | ram_release_pages(block->idstr, offset, pages); |
56e93d26 JQ |
1165 | } |
1166 | } | |
1167 | } | |
1168 | ||
1169 | return pages; | |
1170 | } | |
1171 | ||
3d0684b2 JQ |
1172 | /** |
1173 | * find_dirty_block: find the next dirty page and update any state | |
1174 | * associated with the search process. | |
b9e60928 | 1175 | * |
3d0684b2 | 1176 | * Returns if a page is found |
b9e60928 | 1177 | * |
6f37bb8b | 1178 | * @rs: current RAM state |
3d0684b2 JQ |
1179 | * @pss: data about the state of the current dirty page scan |
1180 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1181 | */ |
f20e2865 | 1182 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1183 | { |
f20e2865 | 1184 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
6f37bb8b | 1185 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1186 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1187 | /* |
1188 | * We've been once around the RAM and haven't found anything. | |
1189 | * Give up. | |
1190 | */ | |
1191 | *again = false; | |
1192 | return false; | |
1193 | } | |
a935e30f | 1194 | if ((pss->page << TARGET_PAGE_BITS) >= pss->block->used_length) { |
b9e60928 | 1195 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1196 | pss->page = 0; |
b9e60928 DDAG |
1197 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1198 | if (!pss->block) { | |
1199 | /* Hit the end of the list */ | |
1200 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1201 | /* Flag that we've looped */ | |
1202 | pss->complete_round = true; | |
6f37bb8b | 1203 | rs->ram_bulk_stage = false; |
b9e60928 DDAG |
1204 | if (migrate_use_xbzrle()) { |
1205 | /* If xbzrle is on, stop using the data compression at this | |
1206 | * point. In theory, xbzrle can do better than compression. | |
1207 | */ | |
ce25d337 | 1208 | flush_compressed_data(rs); |
b9e60928 DDAG |
1209 | } |
1210 | } | |
1211 | /* Didn't find anything this time, but try again on the new block */ | |
1212 | *again = true; | |
1213 | return false; | |
1214 | } else { | |
1215 | /* Can go around again, but... */ | |
1216 | *again = true; | |
1217 | /* We've found something so probably don't need to */ | |
1218 | return true; | |
1219 | } | |
1220 | } | |
1221 | ||
3d0684b2 JQ |
1222 | /** |
1223 | * unqueue_page: gets a page of the queue | |
1224 | * | |
a82d593b | 1225 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1226 | * |
3d0684b2 JQ |
1227 | * Returns the block of the page (or NULL if none available) |
1228 | * | |
ec481c6c | 1229 | * @rs: current RAM state |
3d0684b2 | 1230 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1231 | */ |
f20e2865 | 1232 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b DDAG |
1233 | { |
1234 | RAMBlock *block = NULL; | |
1235 | ||
ec481c6c JQ |
1236 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1237 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
1238 | struct RAMSrcPageRequest *entry = | |
1239 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1240 | block = entry->rb; |
1241 | *offset = entry->offset; | |
a82d593b DDAG |
1242 | |
1243 | if (entry->len > TARGET_PAGE_SIZE) { | |
1244 | entry->len -= TARGET_PAGE_SIZE; | |
1245 | entry->offset += TARGET_PAGE_SIZE; | |
1246 | } else { | |
1247 | memory_region_unref(block->mr); | |
ec481c6c | 1248 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b DDAG |
1249 | g_free(entry); |
1250 | } | |
1251 | } | |
ec481c6c | 1252 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
a82d593b DDAG |
1253 | |
1254 | return block; | |
1255 | } | |
1256 | ||
3d0684b2 JQ |
1257 | /** |
1258 | * get_queued_page: unqueue a page from the postocpy requests | |
1259 | * | |
1260 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1261 | * |
3d0684b2 | 1262 | * Returns if a queued page is found |
a82d593b | 1263 | * |
6f37bb8b | 1264 | * @rs: current RAM state |
3d0684b2 | 1265 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1266 | */ |
f20e2865 | 1267 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1268 | { |
1269 | RAMBlock *block; | |
1270 | ram_addr_t offset; | |
1271 | bool dirty; | |
1272 | ||
1273 | do { | |
f20e2865 | 1274 | block = unqueue_page(rs, &offset); |
a82d593b DDAG |
1275 | /* |
1276 | * We're sending this page, and since it's postcopy nothing else | |
1277 | * will dirty it, and we must make sure it doesn't get sent again | |
1278 | * even if this queue request was received after the background | |
1279 | * search already sent it. | |
1280 | */ | |
1281 | if (block) { | |
f20e2865 JQ |
1282 | unsigned long page; |
1283 | ||
6b6712ef JQ |
1284 | page = offset >> TARGET_PAGE_BITS; |
1285 | dirty = test_bit(page, block->bmap); | |
a82d593b | 1286 | if (!dirty) { |
06b10688 | 1287 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, |
6b6712ef | 1288 | page, test_bit(page, block->unsentmap)); |
a82d593b | 1289 | } else { |
f20e2865 | 1290 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); |
a82d593b DDAG |
1291 | } |
1292 | } | |
1293 | ||
1294 | } while (block && !dirty); | |
1295 | ||
1296 | if (block) { | |
1297 | /* | |
1298 | * As soon as we start servicing pages out of order, then we have | |
1299 | * to kill the bulk stage, since the bulk stage assumes | |
1300 | * in (migration_bitmap_find_and_reset_dirty) that every page is | |
1301 | * dirty, that's no longer true. | |
1302 | */ | |
6f37bb8b | 1303 | rs->ram_bulk_stage = false; |
a82d593b DDAG |
1304 | |
1305 | /* | |
1306 | * We want the background search to continue from the queued page | |
1307 | * since the guest is likely to want other pages near to the page | |
1308 | * it just requested. | |
1309 | */ | |
1310 | pss->block = block; | |
a935e30f | 1311 | pss->page = offset >> TARGET_PAGE_BITS; |
a82d593b DDAG |
1312 | } |
1313 | ||
1314 | return !!block; | |
1315 | } | |
1316 | ||
6c595cde | 1317 | /** |
5e58f968 JQ |
1318 | * migration_page_queue_free: drop any remaining pages in the ram |
1319 | * request queue | |
6c595cde | 1320 | * |
3d0684b2 JQ |
1321 | * It should be empty at the end anyway, but in error cases there may |
1322 | * be some left. in case that there is any page left, we drop it. | |
1323 | * | |
6c595cde | 1324 | */ |
83c13382 | 1325 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 1326 | { |
ec481c6c | 1327 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
1328 | /* This queue generally should be empty - but in the case of a failed |
1329 | * migration might have some droppings in. | |
1330 | */ | |
1331 | rcu_read_lock(); | |
ec481c6c | 1332 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1333 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1334 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1335 | g_free(mspr); |
1336 | } | |
1337 | rcu_read_unlock(); | |
1338 | } | |
1339 | ||
1340 | /** | |
3d0684b2 JQ |
1341 | * ram_save_queue_pages: queue the page for transmission |
1342 | * | |
1343 | * A request from postcopy destination for example. | |
1344 | * | |
1345 | * Returns zero on success or negative on error | |
1346 | * | |
3d0684b2 JQ |
1347 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1348 | * same that last one. | |
1349 | * @start: starting address from the start of the RAMBlock | |
1350 | * @len: length (in bytes) to send | |
6c595cde | 1351 | */ |
96506894 | 1352 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1353 | { |
1354 | RAMBlock *ramblock; | |
53518d94 | 1355 | RAMState *rs = ram_state; |
6c595cde | 1356 | |
9360447d | 1357 | ram_counters.postcopy_requests++; |
6c595cde DDAG |
1358 | rcu_read_lock(); |
1359 | if (!rbname) { | |
1360 | /* Reuse last RAMBlock */ | |
68a098f3 | 1361 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1362 | |
1363 | if (!ramblock) { | |
1364 | /* | |
1365 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1366 | * it's the 1st request. | |
1367 | */ | |
1368 | error_report("ram_save_queue_pages no previous block"); | |
1369 | goto err; | |
1370 | } | |
1371 | } else { | |
1372 | ramblock = qemu_ram_block_by_name(rbname); | |
1373 | ||
1374 | if (!ramblock) { | |
1375 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1376 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
1377 | goto err; | |
1378 | } | |
68a098f3 | 1379 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1380 | } |
1381 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
1382 | if (start+len > ramblock->used_length) { | |
9458ad6b JQ |
1383 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1384 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde DDAG |
1385 | __func__, start, len, ramblock->used_length); |
1386 | goto err; | |
1387 | } | |
1388 | ||
ec481c6c JQ |
1389 | struct RAMSrcPageRequest *new_entry = |
1390 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
1391 | new_entry->rb = ramblock; |
1392 | new_entry->offset = start; | |
1393 | new_entry->len = len; | |
1394 | ||
1395 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
1396 | qemu_mutex_lock(&rs->src_page_req_mutex); |
1397 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
1398 | qemu_mutex_unlock(&rs->src_page_req_mutex); | |
6c595cde DDAG |
1399 | rcu_read_unlock(); |
1400 | ||
1401 | return 0; | |
1402 | ||
1403 | err: | |
1404 | rcu_read_unlock(); | |
1405 | return -1; | |
1406 | } | |
1407 | ||
a82d593b | 1408 | /** |
3d0684b2 | 1409 | * ram_save_target_page: save one target page |
a82d593b | 1410 | * |
3d0684b2 | 1411 | * Returns the number of pages written |
a82d593b | 1412 | * |
6f37bb8b | 1413 | * @rs: current RAM state |
3d0684b2 | 1414 | * @ms: current migration state |
3d0684b2 | 1415 | * @pss: data about the page we want to send |
a82d593b | 1416 | * @last_stage: if we are at the completion stage |
a82d593b | 1417 | */ |
a0a8aa14 | 1418 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1419 | bool last_stage) |
a82d593b DDAG |
1420 | { |
1421 | int res = 0; | |
1422 | ||
1423 | /* Check the pages is dirty and if it is send it */ | |
f20e2865 | 1424 | if (migration_bitmap_clear_dirty(rs, pss->block, pss->page)) { |
6d358d94 JQ |
1425 | /* |
1426 | * If xbzrle is on, stop using the data compression after first | |
1427 | * round of migration even if compression is enabled. In theory, | |
1428 | * xbzrle can do better than compression. | |
1429 | */ | |
6b6712ef JQ |
1430 | if (migrate_use_compression() && |
1431 | (rs->ram_bulk_stage || !migrate_use_xbzrle())) { | |
a0a8aa14 | 1432 | res = ram_save_compressed_page(rs, pss, last_stage); |
a82d593b | 1433 | } else { |
a0a8aa14 | 1434 | res = ram_save_page(rs, pss, last_stage); |
a82d593b DDAG |
1435 | } |
1436 | ||
1437 | if (res < 0) { | |
1438 | return res; | |
1439 | } | |
6b6712ef JQ |
1440 | if (pss->block->unsentmap) { |
1441 | clear_bit(pss->page, pss->block->unsentmap); | |
a82d593b DDAG |
1442 | } |
1443 | } | |
1444 | ||
1445 | return res; | |
1446 | } | |
1447 | ||
1448 | /** | |
3d0684b2 | 1449 | * ram_save_host_page: save a whole host page |
a82d593b | 1450 | * |
3d0684b2 JQ |
1451 | * Starting at *offset send pages up to the end of the current host |
1452 | * page. It's valid for the initial offset to point into the middle of | |
1453 | * a host page in which case the remainder of the hostpage is sent. | |
1454 | * Only dirty target pages are sent. Note that the host page size may | |
1455 | * be a huge page for this block. | |
1eb3fc0a DDAG |
1456 | * The saving stops at the boundary of the used_length of the block |
1457 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 1458 | * |
3d0684b2 JQ |
1459 | * Returns the number of pages written or negative on error |
1460 | * | |
6f37bb8b | 1461 | * @rs: current RAM state |
3d0684b2 | 1462 | * @ms: current migration state |
3d0684b2 | 1463 | * @pss: data about the page we want to send |
a82d593b | 1464 | * @last_stage: if we are at the completion stage |
a82d593b | 1465 | */ |
a0a8aa14 | 1466 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 1467 | bool last_stage) |
a82d593b DDAG |
1468 | { |
1469 | int tmppages, pages = 0; | |
a935e30f JQ |
1470 | size_t pagesize_bits = |
1471 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
4c011c37 | 1472 | |
a82d593b | 1473 | do { |
f20e2865 | 1474 | tmppages = ram_save_target_page(rs, pss, last_stage); |
a82d593b DDAG |
1475 | if (tmppages < 0) { |
1476 | return tmppages; | |
1477 | } | |
1478 | ||
1479 | pages += tmppages; | |
a935e30f | 1480 | pss->page++; |
1eb3fc0a DDAG |
1481 | } while ((pss->page & (pagesize_bits - 1)) && |
1482 | offset_in_ramblock(pss->block, pss->page << TARGET_PAGE_BITS)); | |
a82d593b DDAG |
1483 | |
1484 | /* The offset we leave with is the last one we looked at */ | |
a935e30f | 1485 | pss->page--; |
a82d593b DDAG |
1486 | return pages; |
1487 | } | |
6c595cde | 1488 | |
56e93d26 | 1489 | /** |
3d0684b2 | 1490 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
1491 | * |
1492 | * Called within an RCU critical section. | |
1493 | * | |
3d0684b2 | 1494 | * Returns the number of pages written where zero means no dirty pages |
56e93d26 | 1495 | * |
6f37bb8b | 1496 | * @rs: current RAM state |
56e93d26 | 1497 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
1498 | * |
1499 | * On systems where host-page-size > target-page-size it will send all the | |
1500 | * pages in a host page that are dirty. | |
56e93d26 JQ |
1501 | */ |
1502 | ||
ce25d337 | 1503 | static int ram_find_and_save_block(RAMState *rs, bool last_stage) |
56e93d26 | 1504 | { |
b8fb8cb7 | 1505 | PageSearchStatus pss; |
56e93d26 | 1506 | int pages = 0; |
b9e60928 | 1507 | bool again, found; |
56e93d26 | 1508 | |
0827b9e9 AA |
1509 | /* No dirty page as there is zero RAM */ |
1510 | if (!ram_bytes_total()) { | |
1511 | return pages; | |
1512 | } | |
1513 | ||
6f37bb8b | 1514 | pss.block = rs->last_seen_block; |
a935e30f | 1515 | pss.page = rs->last_page; |
b8fb8cb7 DDAG |
1516 | pss.complete_round = false; |
1517 | ||
1518 | if (!pss.block) { | |
1519 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1520 | } | |
56e93d26 | 1521 | |
b9e60928 | 1522 | do { |
a82d593b | 1523 | again = true; |
f20e2865 | 1524 | found = get_queued_page(rs, &pss); |
b9e60928 | 1525 | |
a82d593b DDAG |
1526 | if (!found) { |
1527 | /* priority queue empty, so just search for something dirty */ | |
f20e2865 | 1528 | found = find_dirty_block(rs, &pss, &again); |
a82d593b | 1529 | } |
f3f491fc | 1530 | |
a82d593b | 1531 | if (found) { |
f20e2865 | 1532 | pages = ram_save_host_page(rs, &pss, last_stage); |
56e93d26 | 1533 | } |
b9e60928 | 1534 | } while (!pages && again); |
56e93d26 | 1535 | |
6f37bb8b | 1536 | rs->last_seen_block = pss.block; |
a935e30f | 1537 | rs->last_page = pss.page; |
56e93d26 JQ |
1538 | |
1539 | return pages; | |
1540 | } | |
1541 | ||
1542 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
1543 | { | |
1544 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 1545 | |
56e93d26 | 1546 | if (zero) { |
9360447d | 1547 | ram_counters.duplicate += pages; |
56e93d26 | 1548 | } else { |
9360447d JQ |
1549 | ram_counters.normal += pages; |
1550 | ram_counters.transferred += size; | |
56e93d26 JQ |
1551 | qemu_update_position(f, size); |
1552 | } | |
1553 | } | |
1554 | ||
56e93d26 JQ |
1555 | uint64_t ram_bytes_total(void) |
1556 | { | |
1557 | RAMBlock *block; | |
1558 | uint64_t total = 0; | |
1559 | ||
1560 | rcu_read_lock(); | |
99e15582 | 1561 | RAMBLOCK_FOREACH(block) { |
56e93d26 | 1562 | total += block->used_length; |
99e15582 | 1563 | } |
56e93d26 JQ |
1564 | rcu_read_unlock(); |
1565 | return total; | |
1566 | } | |
1567 | ||
f265e0e4 | 1568 | static void xbzrle_load_setup(void) |
56e93d26 | 1569 | { |
f265e0e4 | 1570 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
1571 | } |
1572 | ||
f265e0e4 JQ |
1573 | static void xbzrle_load_cleanup(void) |
1574 | { | |
1575 | g_free(XBZRLE.decoded_buf); | |
1576 | XBZRLE.decoded_buf = NULL; | |
1577 | } | |
1578 | ||
7d7c96be PX |
1579 | static void ram_state_cleanup(RAMState **rsp) |
1580 | { | |
1581 | migration_page_queue_free(*rsp); | |
1582 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
1583 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
1584 | g_free(*rsp); | |
1585 | *rsp = NULL; | |
1586 | } | |
1587 | ||
f265e0e4 | 1588 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 1589 | { |
53518d94 | 1590 | RAMState **rsp = opaque; |
6b6712ef | 1591 | RAMBlock *block; |
eb859c53 | 1592 | |
2ff64038 LZ |
1593 | /* caller have hold iothread lock or is in a bh, so there is |
1594 | * no writing race against this migration_bitmap | |
1595 | */ | |
6b6712ef JQ |
1596 | memory_global_dirty_log_stop(); |
1597 | ||
1598 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
1599 | g_free(block->bmap); | |
1600 | block->bmap = NULL; | |
1601 | g_free(block->unsentmap); | |
1602 | block->unsentmap = NULL; | |
56e93d26 JQ |
1603 | } |
1604 | ||
1605 | XBZRLE_cache_lock(); | |
1606 | if (XBZRLE.cache) { | |
1607 | cache_fini(XBZRLE.cache); | |
1608 | g_free(XBZRLE.encoded_buf); | |
1609 | g_free(XBZRLE.current_buf); | |
c00e0928 | 1610 | g_free(XBZRLE.zero_target_page); |
56e93d26 JQ |
1611 | XBZRLE.cache = NULL; |
1612 | XBZRLE.encoded_buf = NULL; | |
1613 | XBZRLE.current_buf = NULL; | |
c00e0928 | 1614 | XBZRLE.zero_target_page = NULL; |
56e93d26 JQ |
1615 | } |
1616 | XBZRLE_cache_unlock(); | |
f0afa331 | 1617 | compress_threads_save_cleanup(); |
7d7c96be | 1618 | ram_state_cleanup(rsp); |
56e93d26 JQ |
1619 | } |
1620 | ||
6f37bb8b | 1621 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 1622 | { |
6f37bb8b JQ |
1623 | rs->last_seen_block = NULL; |
1624 | rs->last_sent_block = NULL; | |
269ace29 | 1625 | rs->last_page = 0; |
6f37bb8b JQ |
1626 | rs->last_version = ram_list.version; |
1627 | rs->ram_bulk_stage = true; | |
56e93d26 JQ |
1628 | } |
1629 | ||
1630 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
1631 | ||
4f2e4252 DDAG |
1632 | /* |
1633 | * 'expected' is the value you expect the bitmap mostly to be full | |
1634 | * of; it won't bother printing lines that are all this value. | |
1635 | * If 'todump' is null the migration bitmap is dumped. | |
1636 | */ | |
6b6712ef JQ |
1637 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected, |
1638 | unsigned long pages) | |
4f2e4252 | 1639 | { |
4f2e4252 DDAG |
1640 | int64_t cur; |
1641 | int64_t linelen = 128; | |
1642 | char linebuf[129]; | |
1643 | ||
6b6712ef | 1644 | for (cur = 0; cur < pages; cur += linelen) { |
4f2e4252 DDAG |
1645 | int64_t curb; |
1646 | bool found = false; | |
1647 | /* | |
1648 | * Last line; catch the case where the line length | |
1649 | * is longer than remaining ram | |
1650 | */ | |
6b6712ef JQ |
1651 | if (cur + linelen > pages) { |
1652 | linelen = pages - cur; | |
4f2e4252 DDAG |
1653 | } |
1654 | for (curb = 0; curb < linelen; curb++) { | |
1655 | bool thisbit = test_bit(cur + curb, todump); | |
1656 | linebuf[curb] = thisbit ? '1' : '.'; | |
1657 | found = found || (thisbit != expected); | |
1658 | } | |
1659 | if (found) { | |
1660 | linebuf[curb] = '\0'; | |
1661 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
1662 | } | |
1663 | } | |
1664 | } | |
1665 | ||
e0b266f0 DDAG |
1666 | /* **** functions for postcopy ***** */ |
1667 | ||
ced1c616 PB |
1668 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
1669 | { | |
1670 | struct RAMBlock *block; | |
ced1c616 | 1671 | |
99e15582 | 1672 | RAMBLOCK_FOREACH(block) { |
6b6712ef JQ |
1673 | unsigned long *bitmap = block->bmap; |
1674 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
1675 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
1676 | |
1677 | while (run_start < range) { | |
1678 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
aaa2064c | 1679 | ram_discard_range(block->idstr, run_start << TARGET_PAGE_BITS, |
ced1c616 PB |
1680 | (run_end - run_start) << TARGET_PAGE_BITS); |
1681 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); | |
1682 | } | |
1683 | } | |
1684 | } | |
1685 | ||
3d0684b2 JQ |
1686 | /** |
1687 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
1688 | * | |
1689 | * Returns zero on success | |
1690 | * | |
e0b266f0 DDAG |
1691 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
1692 | * Note: At this point the 'unsentmap' is the processed bitmap combined | |
1693 | * with the dirtymap; so a '1' means it's either dirty or unsent. | |
3d0684b2 JQ |
1694 | * |
1695 | * @ms: current migration state | |
1696 | * @pds: state for postcopy | |
1697 | * @start: RAMBlock starting page | |
1698 | * @length: RAMBlock size | |
e0b266f0 DDAG |
1699 | */ |
1700 | static int postcopy_send_discard_bm_ram(MigrationState *ms, | |
1701 | PostcopyDiscardState *pds, | |
6b6712ef | 1702 | RAMBlock *block) |
e0b266f0 | 1703 | { |
6b6712ef | 1704 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 1705 | unsigned long current; |
6b6712ef | 1706 | unsigned long *unsentmap = block->unsentmap; |
e0b266f0 | 1707 | |
6b6712ef | 1708 | for (current = 0; current < end; ) { |
e0b266f0 DDAG |
1709 | unsigned long one = find_next_bit(unsentmap, end, current); |
1710 | ||
1711 | if (one <= end) { | |
1712 | unsigned long zero = find_next_zero_bit(unsentmap, end, one + 1); | |
1713 | unsigned long discard_length; | |
1714 | ||
1715 | if (zero >= end) { | |
1716 | discard_length = end - one; | |
1717 | } else { | |
1718 | discard_length = zero - one; | |
1719 | } | |
d688c62d DDAG |
1720 | if (discard_length) { |
1721 | postcopy_discard_send_range(ms, pds, one, discard_length); | |
1722 | } | |
e0b266f0 DDAG |
1723 | current = one + discard_length; |
1724 | } else { | |
1725 | current = one; | |
1726 | } | |
1727 | } | |
1728 | ||
1729 | return 0; | |
1730 | } | |
1731 | ||
3d0684b2 JQ |
1732 | /** |
1733 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
1734 | * | |
1735 | * Returns 0 for success or negative for error | |
1736 | * | |
e0b266f0 DDAG |
1737 | * Utility for the outgoing postcopy code. |
1738 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
1739 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
1740 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
1741 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
1742 | * |
1743 | * @ms: current migration state | |
e0b266f0 DDAG |
1744 | */ |
1745 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
1746 | { | |
1747 | struct RAMBlock *block; | |
1748 | int ret; | |
1749 | ||
99e15582 | 1750 | RAMBLOCK_FOREACH(block) { |
6b6712ef JQ |
1751 | PostcopyDiscardState *pds = |
1752 | postcopy_discard_send_init(ms, block->idstr); | |
e0b266f0 DDAG |
1753 | |
1754 | /* | |
1755 | * Postcopy sends chunks of bitmap over the wire, but it | |
1756 | * just needs indexes at this point, avoids it having | |
1757 | * target page specific code. | |
1758 | */ | |
6b6712ef | 1759 | ret = postcopy_send_discard_bm_ram(ms, pds, block); |
e0b266f0 DDAG |
1760 | postcopy_discard_send_finish(ms, pds); |
1761 | if (ret) { | |
1762 | return ret; | |
1763 | } | |
1764 | } | |
1765 | ||
1766 | return 0; | |
1767 | } | |
1768 | ||
3d0684b2 JQ |
1769 | /** |
1770 | * postcopy_chunk_hostpages_pass: canocalize bitmap in hostpages | |
1771 | * | |
1772 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
1773 | * canonicalize the two bitmaps, that are similar, but one is | |
1774 | * inverted. | |
99e314eb | 1775 | * |
3d0684b2 JQ |
1776 | * Postcopy requires that all target pages in a hostpage are dirty or |
1777 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 1778 | * |
3d0684b2 JQ |
1779 | * @ms: current migration state |
1780 | * @unsent_pass: if true we need to canonicalize partially unsent host pages | |
1781 | * otherwise we need to canonicalize partially dirty host pages | |
1782 | * @block: block that contains the page we want to canonicalize | |
1783 | * @pds: state for postcopy | |
99e314eb DDAG |
1784 | */ |
1785 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, bool unsent_pass, | |
1786 | RAMBlock *block, | |
1787 | PostcopyDiscardState *pds) | |
1788 | { | |
53518d94 | 1789 | RAMState *rs = ram_state; |
6b6712ef JQ |
1790 | unsigned long *bitmap = block->bmap; |
1791 | unsigned long *unsentmap = block->unsentmap; | |
29c59172 | 1792 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 1793 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
1794 | unsigned long run_start; |
1795 | ||
29c59172 DDAG |
1796 | if (block->page_size == TARGET_PAGE_SIZE) { |
1797 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
1798 | return; | |
1799 | } | |
1800 | ||
99e314eb DDAG |
1801 | if (unsent_pass) { |
1802 | /* Find a sent page */ | |
6b6712ef | 1803 | run_start = find_next_zero_bit(unsentmap, pages, 0); |
99e314eb DDAG |
1804 | } else { |
1805 | /* Find a dirty page */ | |
6b6712ef | 1806 | run_start = find_next_bit(bitmap, pages, 0); |
99e314eb DDAG |
1807 | } |
1808 | ||
6b6712ef | 1809 | while (run_start < pages) { |
99e314eb DDAG |
1810 | bool do_fixup = false; |
1811 | unsigned long fixup_start_addr; | |
1812 | unsigned long host_offset; | |
1813 | ||
1814 | /* | |
1815 | * If the start of this run of pages is in the middle of a host | |
1816 | * page, then we need to fixup this host page. | |
1817 | */ | |
1818 | host_offset = run_start % host_ratio; | |
1819 | if (host_offset) { | |
1820 | do_fixup = true; | |
1821 | run_start -= host_offset; | |
1822 | fixup_start_addr = run_start; | |
1823 | /* For the next pass */ | |
1824 | run_start = run_start + host_ratio; | |
1825 | } else { | |
1826 | /* Find the end of this run */ | |
1827 | unsigned long run_end; | |
1828 | if (unsent_pass) { | |
6b6712ef | 1829 | run_end = find_next_bit(unsentmap, pages, run_start + 1); |
99e314eb | 1830 | } else { |
6b6712ef | 1831 | run_end = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
1832 | } |
1833 | /* | |
1834 | * If the end isn't at the start of a host page, then the | |
1835 | * run doesn't finish at the end of a host page | |
1836 | * and we need to discard. | |
1837 | */ | |
1838 | host_offset = run_end % host_ratio; | |
1839 | if (host_offset) { | |
1840 | do_fixup = true; | |
1841 | fixup_start_addr = run_end - host_offset; | |
1842 | /* | |
1843 | * This host page has gone, the next loop iteration starts | |
1844 | * from after the fixup | |
1845 | */ | |
1846 | run_start = fixup_start_addr + host_ratio; | |
1847 | } else { | |
1848 | /* | |
1849 | * No discards on this iteration, next loop starts from | |
1850 | * next sent/dirty page | |
1851 | */ | |
1852 | run_start = run_end + 1; | |
1853 | } | |
1854 | } | |
1855 | ||
1856 | if (do_fixup) { | |
1857 | unsigned long page; | |
1858 | ||
1859 | /* Tell the destination to discard this page */ | |
1860 | if (unsent_pass || !test_bit(fixup_start_addr, unsentmap)) { | |
1861 | /* For the unsent_pass we: | |
1862 | * discard partially sent pages | |
1863 | * For the !unsent_pass (dirty) we: | |
1864 | * discard partially dirty pages that were sent | |
1865 | * (any partially sent pages were already discarded | |
1866 | * by the previous unsent_pass) | |
1867 | */ | |
1868 | postcopy_discard_send_range(ms, pds, fixup_start_addr, | |
1869 | host_ratio); | |
1870 | } | |
1871 | ||
1872 | /* Clean up the bitmap */ | |
1873 | for (page = fixup_start_addr; | |
1874 | page < fixup_start_addr + host_ratio; page++) { | |
1875 | /* All pages in this host page are now not sent */ | |
1876 | set_bit(page, unsentmap); | |
1877 | ||
1878 | /* | |
1879 | * Remark them as dirty, updating the count for any pages | |
1880 | * that weren't previously dirty. | |
1881 | */ | |
0d8ec885 | 1882 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
1883 | } |
1884 | } | |
1885 | ||
1886 | if (unsent_pass) { | |
1887 | /* Find the next sent page for the next iteration */ | |
6b6712ef | 1888 | run_start = find_next_zero_bit(unsentmap, pages, run_start); |
99e314eb DDAG |
1889 | } else { |
1890 | /* Find the next dirty page for the next iteration */ | |
6b6712ef | 1891 | run_start = find_next_bit(bitmap, pages, run_start); |
99e314eb DDAG |
1892 | } |
1893 | } | |
1894 | } | |
1895 | ||
3d0684b2 JQ |
1896 | /** |
1897 | * postcopy_chuck_hostpages: discrad any partially sent host page | |
1898 | * | |
99e314eb DDAG |
1899 | * Utility for the outgoing postcopy code. |
1900 | * | |
1901 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
1902 | * dirty host-page size chunks as all dirty. In this case the host-page |
1903 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 1904 | * |
3d0684b2 JQ |
1905 | * Returns zero on success |
1906 | * | |
1907 | * @ms: current migration state | |
6b6712ef | 1908 | * @block: block we want to work with |
99e314eb | 1909 | */ |
6b6712ef | 1910 | static int postcopy_chunk_hostpages(MigrationState *ms, RAMBlock *block) |
99e314eb | 1911 | { |
6b6712ef JQ |
1912 | PostcopyDiscardState *pds = |
1913 | postcopy_discard_send_init(ms, block->idstr); | |
99e314eb | 1914 | |
6b6712ef JQ |
1915 | /* First pass: Discard all partially sent host pages */ |
1916 | postcopy_chunk_hostpages_pass(ms, true, block, pds); | |
1917 | /* | |
1918 | * Second pass: Ensure that all partially dirty host pages are made | |
1919 | * fully dirty. | |
1920 | */ | |
1921 | postcopy_chunk_hostpages_pass(ms, false, block, pds); | |
99e314eb | 1922 | |
6b6712ef | 1923 | postcopy_discard_send_finish(ms, pds); |
99e314eb DDAG |
1924 | return 0; |
1925 | } | |
1926 | ||
3d0684b2 JQ |
1927 | /** |
1928 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
1929 | * | |
1930 | * Returns zero on success | |
1931 | * | |
e0b266f0 DDAG |
1932 | * Transmit the set of pages to be discarded after precopy to the target |
1933 | * these are pages that: | |
1934 | * a) Have been previously transmitted but are now dirty again | |
1935 | * b) Pages that have never been transmitted, this ensures that | |
1936 | * any pages on the destination that have been mapped by background | |
1937 | * tasks get discarded (transparent huge pages is the specific concern) | |
1938 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
1939 | * |
1940 | * @ms: current migration state | |
e0b266f0 DDAG |
1941 | */ |
1942 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
1943 | { | |
53518d94 | 1944 | RAMState *rs = ram_state; |
6b6712ef | 1945 | RAMBlock *block; |
e0b266f0 | 1946 | int ret; |
e0b266f0 DDAG |
1947 | |
1948 | rcu_read_lock(); | |
1949 | ||
1950 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 1951 | migration_bitmap_sync(rs); |
e0b266f0 | 1952 | |
6b6712ef JQ |
1953 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
1954 | rs->last_seen_block = NULL; | |
1955 | rs->last_sent_block = NULL; | |
1956 | rs->last_page = 0; | |
e0b266f0 | 1957 | |
6b6712ef JQ |
1958 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { |
1959 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; | |
1960 | unsigned long *bitmap = block->bmap; | |
1961 | unsigned long *unsentmap = block->unsentmap; | |
1962 | ||
1963 | if (!unsentmap) { | |
1964 | /* We don't have a safe way to resize the sentmap, so | |
1965 | * if the bitmap was resized it will be NULL at this | |
1966 | * point. | |
1967 | */ | |
1968 | error_report("migration ram resized during precopy phase"); | |
1969 | rcu_read_unlock(); | |
1970 | return -EINVAL; | |
1971 | } | |
1972 | /* Deal with TPS != HPS and huge pages */ | |
1973 | ret = postcopy_chunk_hostpages(ms, block); | |
1974 | if (ret) { | |
1975 | rcu_read_unlock(); | |
1976 | return ret; | |
1977 | } | |
e0b266f0 | 1978 | |
6b6712ef JQ |
1979 | /* |
1980 | * Update the unsentmap to be unsentmap = unsentmap | dirty | |
1981 | */ | |
1982 | bitmap_or(unsentmap, unsentmap, bitmap, pages); | |
e0b266f0 | 1983 | #ifdef DEBUG_POSTCOPY |
6b6712ef | 1984 | ram_debug_dump_bitmap(unsentmap, true, pages); |
e0b266f0 | 1985 | #endif |
6b6712ef JQ |
1986 | } |
1987 | trace_ram_postcopy_send_discard_bitmap(); | |
e0b266f0 DDAG |
1988 | |
1989 | ret = postcopy_each_ram_send_discard(ms); | |
1990 | rcu_read_unlock(); | |
1991 | ||
1992 | return ret; | |
1993 | } | |
1994 | ||
3d0684b2 JQ |
1995 | /** |
1996 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 1997 | * |
3d0684b2 | 1998 | * Returns zero on success |
e0b266f0 | 1999 | * |
36449157 JQ |
2000 | * @rbname: name of the RAMBlock of the request. NULL means the |
2001 | * same that last one. | |
3d0684b2 JQ |
2002 | * @start: RAMBlock starting page |
2003 | * @length: RAMBlock size | |
e0b266f0 | 2004 | */ |
aaa2064c | 2005 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 DDAG |
2006 | { |
2007 | int ret = -1; | |
2008 | ||
36449157 | 2009 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2010 | |
e0b266f0 | 2011 | rcu_read_lock(); |
36449157 | 2012 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2013 | |
2014 | if (!rb) { | |
36449157 | 2015 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
e0b266f0 DDAG |
2016 | goto err; |
2017 | } | |
2018 | ||
d3a5038c | 2019 | ret = ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2020 | |
2021 | err: | |
2022 | rcu_read_unlock(); | |
2023 | ||
2024 | return ret; | |
2025 | } | |
2026 | ||
53518d94 | 2027 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2028 | { |
7d00ee6a PX |
2029 | *rsp = g_try_new0(RAMState, 1); |
2030 | ||
2031 | if (!*rsp) { | |
2032 | error_report("%s: Init ramstate fail", __func__); | |
2033 | return -1; | |
2034 | } | |
53518d94 JQ |
2035 | |
2036 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
2037 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
2038 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
56e93d26 | 2039 | |
7d00ee6a PX |
2040 | /* |
2041 | * Count the total number of pages used by ram blocks not including any | |
2042 | * gaps due to alignment or unplugs. | |
2043 | */ | |
2044 | (*rsp)->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; | |
2045 | ||
2046 | ram_state_reset(*rsp); | |
2047 | ||
2048 | return 0; | |
2049 | } | |
2050 | ||
2051 | static int ram_init_all(RAMState **rsp) | |
2052 | { | |
2053 | Error *local_err = NULL; | |
2054 | ||
2055 | if (ram_state_init(rsp)) { | |
2056 | return -1; | |
2057 | } | |
2058 | ||
56e93d26 JQ |
2059 | if (migrate_use_xbzrle()) { |
2060 | XBZRLE_cache_lock(); | |
c00e0928 | 2061 | XBZRLE.zero_target_page = g_malloc0(TARGET_PAGE_SIZE); |
80f8dfde JQ |
2062 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), |
2063 | TARGET_PAGE_SIZE, &local_err); | |
56e93d26 JQ |
2064 | if (!XBZRLE.cache) { |
2065 | XBZRLE_cache_unlock(); | |
80f8dfde | 2066 | error_report_err(local_err); |
53518d94 JQ |
2067 | g_free(*rsp); |
2068 | *rsp = NULL; | |
56e93d26 JQ |
2069 | return -1; |
2070 | } | |
2071 | XBZRLE_cache_unlock(); | |
2072 | ||
2073 | /* We prefer not to abort if there is no memory */ | |
2074 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2075 | if (!XBZRLE.encoded_buf) { | |
2076 | error_report("Error allocating encoded_buf"); | |
53518d94 JQ |
2077 | g_free(*rsp); |
2078 | *rsp = NULL; | |
56e93d26 JQ |
2079 | return -1; |
2080 | } | |
2081 | ||
2082 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2083 | if (!XBZRLE.current_buf) { | |
2084 | error_report("Error allocating current_buf"); | |
2085 | g_free(XBZRLE.encoded_buf); | |
2086 | XBZRLE.encoded_buf = NULL; | |
53518d94 JQ |
2087 | g_free(*rsp); |
2088 | *rsp = NULL; | |
56e93d26 JQ |
2089 | return -1; |
2090 | } | |
56e93d26 JQ |
2091 | } |
2092 | ||
49877834 PB |
2093 | /* For memory_global_dirty_log_start below. */ |
2094 | qemu_mutex_lock_iothread(); | |
2095 | ||
56e93d26 JQ |
2096 | qemu_mutex_lock_ramlist(); |
2097 | rcu_read_lock(); | |
56e93d26 | 2098 | |
0827b9e9 AA |
2099 | /* Skip setting bitmap if there is no RAM */ |
2100 | if (ram_bytes_total()) { | |
6b6712ef JQ |
2101 | RAMBlock *block; |
2102 | ||
2103 | QLIST_FOREACH_RCU(block, &ram_list.blocks, next) { | |
2104 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; | |
0827b9e9 | 2105 | |
6b6712ef JQ |
2106 | block->bmap = bitmap_new(pages); |
2107 | bitmap_set(block->bmap, 0, pages); | |
2108 | if (migrate_postcopy_ram()) { | |
2109 | block->unsentmap = bitmap_new(pages); | |
2110 | bitmap_set(block->unsentmap, 0, pages); | |
2111 | } | |
0827b9e9 | 2112 | } |
f3f491fc DDAG |
2113 | } |
2114 | ||
56e93d26 | 2115 | memory_global_dirty_log_start(); |
53518d94 | 2116 | migration_bitmap_sync(*rsp); |
56e93d26 | 2117 | qemu_mutex_unlock_ramlist(); |
49877834 | 2118 | qemu_mutex_unlock_iothread(); |
a91246c9 HZ |
2119 | rcu_read_unlock(); |
2120 | ||
2121 | return 0; | |
2122 | } | |
2123 | ||
3d0684b2 JQ |
2124 | /* |
2125 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2126 | * long-running RCU critical section. When rcu-reclaims in the code |
2127 | * start to become numerous it will be necessary to reduce the | |
2128 | * granularity of these critical sections. | |
2129 | */ | |
2130 | ||
3d0684b2 JQ |
2131 | /** |
2132 | * ram_save_setup: Setup RAM for migration | |
2133 | * | |
2134 | * Returns zero to indicate success and negative for error | |
2135 | * | |
2136 | * @f: QEMUFile where to send the data | |
2137 | * @opaque: RAMState pointer | |
2138 | */ | |
a91246c9 HZ |
2139 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2140 | { | |
53518d94 | 2141 | RAMState **rsp = opaque; |
a91246c9 HZ |
2142 | RAMBlock *block; |
2143 | ||
2144 | /* migration has already setup the bitmap, reuse it. */ | |
2145 | if (!migration_in_colo_state()) { | |
7d00ee6a | 2146 | if (ram_init_all(rsp) != 0) { |
a91246c9 | 2147 | return -1; |
53518d94 | 2148 | } |
a91246c9 | 2149 | } |
53518d94 | 2150 | (*rsp)->f = f; |
a91246c9 HZ |
2151 | |
2152 | rcu_read_lock(); | |
56e93d26 JQ |
2153 | |
2154 | qemu_put_be64(f, ram_bytes_total() | RAM_SAVE_FLAG_MEM_SIZE); | |
2155 | ||
99e15582 | 2156 | RAMBLOCK_FOREACH(block) { |
56e93d26 JQ |
2157 | qemu_put_byte(f, strlen(block->idstr)); |
2158 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2159 | qemu_put_be64(f, block->used_length); | |
ef08fb38 DDAG |
2160 | if (migrate_postcopy_ram() && block->page_size != qemu_host_page_size) { |
2161 | qemu_put_be64(f, block->page_size); | |
2162 | } | |
56e93d26 JQ |
2163 | } |
2164 | ||
2165 | rcu_read_unlock(); | |
f0afa331 | 2166 | compress_threads_save_setup(); |
56e93d26 JQ |
2167 | |
2168 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); | |
2169 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2170 | ||
2171 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
2172 | ||
2173 | return 0; | |
2174 | } | |
2175 | ||
3d0684b2 JQ |
2176 | /** |
2177 | * ram_save_iterate: iterative stage for migration | |
2178 | * | |
2179 | * Returns zero to indicate success and negative for error | |
2180 | * | |
2181 | * @f: QEMUFile where to send the data | |
2182 | * @opaque: RAMState pointer | |
2183 | */ | |
56e93d26 JQ |
2184 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2185 | { | |
53518d94 JQ |
2186 | RAMState **temp = opaque; |
2187 | RAMState *rs = *temp; | |
56e93d26 JQ |
2188 | int ret; |
2189 | int i; | |
2190 | int64_t t0; | |
5c90308f | 2191 | int done = 0; |
56e93d26 JQ |
2192 | |
2193 | rcu_read_lock(); | |
6f37bb8b JQ |
2194 | if (ram_list.version != rs->last_version) { |
2195 | ram_state_reset(rs); | |
56e93d26 JQ |
2196 | } |
2197 | ||
2198 | /* Read version before ram_list.blocks */ | |
2199 | smp_rmb(); | |
2200 | ||
2201 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); | |
2202 | ||
2203 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | |
2204 | i = 0; | |
2205 | while ((ret = qemu_file_rate_limit(f)) == 0) { | |
2206 | int pages; | |
2207 | ||
ce25d337 | 2208 | pages = ram_find_and_save_block(rs, false); |
56e93d26 JQ |
2209 | /* no more pages to sent */ |
2210 | if (pages == 0) { | |
5c90308f | 2211 | done = 1; |
56e93d26 JQ |
2212 | break; |
2213 | } | |
23b28c3c | 2214 | rs->iterations++; |
070afca2 | 2215 | |
56e93d26 JQ |
2216 | /* we want to check in the 1st loop, just in case it was the 1st time |
2217 | and we had to sync the dirty bitmap. | |
2218 | qemu_get_clock_ns() is a bit expensive, so we only check each some | |
2219 | iterations | |
2220 | */ | |
2221 | if ((i & 63) == 0) { | |
2222 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / 1000000; | |
2223 | if (t1 > MAX_WAIT) { | |
55c4446b | 2224 | trace_ram_save_iterate_big_wait(t1, i); |
56e93d26 JQ |
2225 | break; |
2226 | } | |
2227 | } | |
2228 | i++; | |
2229 | } | |
ce25d337 | 2230 | flush_compressed_data(rs); |
56e93d26 JQ |
2231 | rcu_read_unlock(); |
2232 | ||
2233 | /* | |
2234 | * Must occur before EOS (or any QEMUFile operation) | |
2235 | * because of RDMA protocol. | |
2236 | */ | |
2237 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
2238 | ||
2239 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
9360447d | 2240 | ram_counters.transferred += 8; |
56e93d26 JQ |
2241 | |
2242 | ret = qemu_file_get_error(f); | |
2243 | if (ret < 0) { | |
2244 | return ret; | |
2245 | } | |
2246 | ||
5c90308f | 2247 | return done; |
56e93d26 JQ |
2248 | } |
2249 | ||
3d0684b2 JQ |
2250 | /** |
2251 | * ram_save_complete: function called to send the remaining amount of ram | |
2252 | * | |
2253 | * Returns zero to indicate success | |
2254 | * | |
2255 | * Called with iothread lock | |
2256 | * | |
2257 | * @f: QEMUFile where to send the data | |
2258 | * @opaque: RAMState pointer | |
2259 | */ | |
56e93d26 JQ |
2260 | static int ram_save_complete(QEMUFile *f, void *opaque) |
2261 | { | |
53518d94 JQ |
2262 | RAMState **temp = opaque; |
2263 | RAMState *rs = *temp; | |
6f37bb8b | 2264 | |
56e93d26 JQ |
2265 | rcu_read_lock(); |
2266 | ||
5727309d | 2267 | if (!migration_in_postcopy()) { |
8d820d6f | 2268 | migration_bitmap_sync(rs); |
663e6c1d | 2269 | } |
56e93d26 JQ |
2270 | |
2271 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); | |
2272 | ||
2273 | /* try transferring iterative blocks of memory */ | |
2274 | ||
2275 | /* flush all remaining blocks regardless of rate limiting */ | |
2276 | while (true) { | |
2277 | int pages; | |
2278 | ||
ce25d337 | 2279 | pages = ram_find_and_save_block(rs, !migration_in_colo_state()); |
56e93d26 JQ |
2280 | /* no more blocks to sent */ |
2281 | if (pages == 0) { | |
2282 | break; | |
2283 | } | |
2284 | } | |
2285 | ||
ce25d337 | 2286 | flush_compressed_data(rs); |
56e93d26 | 2287 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 JQ |
2288 | |
2289 | rcu_read_unlock(); | |
d09a6fde | 2290 | |
56e93d26 JQ |
2291 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
2292 | ||
2293 | return 0; | |
2294 | } | |
2295 | ||
c31b098f DDAG |
2296 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
2297 | uint64_t *non_postcopiable_pending, | |
2298 | uint64_t *postcopiable_pending) | |
56e93d26 | 2299 | { |
53518d94 JQ |
2300 | RAMState **temp = opaque; |
2301 | RAMState *rs = *temp; | |
56e93d26 JQ |
2302 | uint64_t remaining_size; |
2303 | ||
9edabd4d | 2304 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2305 | |
5727309d | 2306 | if (!migration_in_postcopy() && |
663e6c1d | 2307 | remaining_size < max_size) { |
56e93d26 JQ |
2308 | qemu_mutex_lock_iothread(); |
2309 | rcu_read_lock(); | |
8d820d6f | 2310 | migration_bitmap_sync(rs); |
56e93d26 JQ |
2311 | rcu_read_unlock(); |
2312 | qemu_mutex_unlock_iothread(); | |
9edabd4d | 2313 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 2314 | } |
c31b098f | 2315 | |
86e1167e VSO |
2316 | if (migrate_postcopy_ram()) { |
2317 | /* We can do postcopy, and all the data is postcopiable */ | |
2318 | *postcopiable_pending += remaining_size; | |
2319 | } else { | |
2320 | *non_postcopiable_pending += remaining_size; | |
2321 | } | |
56e93d26 JQ |
2322 | } |
2323 | ||
2324 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
2325 | { | |
2326 | unsigned int xh_len; | |
2327 | int xh_flags; | |
063e760a | 2328 | uint8_t *loaded_data; |
56e93d26 | 2329 | |
56e93d26 JQ |
2330 | /* extract RLE header */ |
2331 | xh_flags = qemu_get_byte(f); | |
2332 | xh_len = qemu_get_be16(f); | |
2333 | ||
2334 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
2335 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
2336 | return -1; | |
2337 | } | |
2338 | ||
2339 | if (xh_len > TARGET_PAGE_SIZE) { | |
2340 | error_report("Failed to load XBZRLE page - len overflow!"); | |
2341 | return -1; | |
2342 | } | |
f265e0e4 | 2343 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 2344 | /* load data and decode */ |
f265e0e4 | 2345 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 2346 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
2347 | |
2348 | /* decode RLE */ | |
063e760a | 2349 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
2350 | TARGET_PAGE_SIZE) == -1) { |
2351 | error_report("Failed to load XBZRLE page - decode error!"); | |
2352 | return -1; | |
2353 | } | |
2354 | ||
2355 | return 0; | |
2356 | } | |
2357 | ||
3d0684b2 JQ |
2358 | /** |
2359 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
2360 | * | |
2361 | * Must be called from within a rcu critical section. | |
2362 | * | |
56e93d26 | 2363 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 2364 | * |
3d0684b2 JQ |
2365 | * @f: QEMUFile where to read the data from |
2366 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 2367 | */ |
3d0684b2 | 2368 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 JQ |
2369 | { |
2370 | static RAMBlock *block = NULL; | |
2371 | char id[256]; | |
2372 | uint8_t len; | |
2373 | ||
2374 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 2375 | if (!block) { |
56e93d26 JQ |
2376 | error_report("Ack, bad migration stream!"); |
2377 | return NULL; | |
2378 | } | |
4c4bad48 | 2379 | return block; |
56e93d26 JQ |
2380 | } |
2381 | ||
2382 | len = qemu_get_byte(f); | |
2383 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2384 | id[len] = 0; | |
2385 | ||
e3dd7493 | 2386 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
2387 | if (!block) { |
2388 | error_report("Can't find block %s", id); | |
2389 | return NULL; | |
56e93d26 JQ |
2390 | } |
2391 | ||
4c4bad48 HZ |
2392 | return block; |
2393 | } | |
2394 | ||
2395 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
2396 | ram_addr_t offset) | |
2397 | { | |
2398 | if (!offset_in_ramblock(block, offset)) { | |
2399 | return NULL; | |
2400 | } | |
2401 | ||
2402 | return block->host + offset; | |
56e93d26 JQ |
2403 | } |
2404 | ||
3d0684b2 JQ |
2405 | /** |
2406 | * ram_handle_compressed: handle the zero page case | |
2407 | * | |
56e93d26 JQ |
2408 | * If a page (or a whole RDMA chunk) has been |
2409 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
2410 | * |
2411 | * @host: host address for the zero page | |
2412 | * @ch: what the page is filled from. We only support zero | |
2413 | * @size: size of the zero page | |
56e93d26 JQ |
2414 | */ |
2415 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
2416 | { | |
2417 | if (ch != 0 || !is_zero_range(host, size)) { | |
2418 | memset(host, ch, size); | |
2419 | } | |
2420 | } | |
2421 | ||
2422 | static void *do_data_decompress(void *opaque) | |
2423 | { | |
2424 | DecompressParam *param = opaque; | |
2425 | unsigned long pagesize; | |
33d151f4 LL |
2426 | uint8_t *des; |
2427 | int len; | |
56e93d26 | 2428 | |
33d151f4 | 2429 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 2430 | while (!param->quit) { |
33d151f4 LL |
2431 | if (param->des) { |
2432 | des = param->des; | |
2433 | len = param->len; | |
2434 | param->des = 0; | |
2435 | qemu_mutex_unlock(¶m->mutex); | |
2436 | ||
56e93d26 | 2437 | pagesize = TARGET_PAGE_SIZE; |
73a8912b LL |
2438 | /* uncompress() will return failed in some case, especially |
2439 | * when the page is dirted when doing the compression, it's | |
2440 | * not a problem because the dirty page will be retransferred | |
2441 | * and uncompress() won't break the data in other pages. | |
2442 | */ | |
33d151f4 LL |
2443 | uncompress((Bytef *)des, &pagesize, |
2444 | (const Bytef *)param->compbuf, len); | |
73a8912b | 2445 | |
33d151f4 LL |
2446 | qemu_mutex_lock(&decomp_done_lock); |
2447 | param->done = true; | |
2448 | qemu_cond_signal(&decomp_done_cond); | |
2449 | qemu_mutex_unlock(&decomp_done_lock); | |
2450 | ||
2451 | qemu_mutex_lock(¶m->mutex); | |
2452 | } else { | |
2453 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
2454 | } | |
56e93d26 | 2455 | } |
33d151f4 | 2456 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
2457 | |
2458 | return NULL; | |
2459 | } | |
2460 | ||
5533b2e9 LL |
2461 | static void wait_for_decompress_done(void) |
2462 | { | |
2463 | int idx, thread_count; | |
2464 | ||
2465 | if (!migrate_use_compression()) { | |
2466 | return; | |
2467 | } | |
2468 | ||
2469 | thread_count = migrate_decompress_threads(); | |
2470 | qemu_mutex_lock(&decomp_done_lock); | |
2471 | for (idx = 0; idx < thread_count; idx++) { | |
2472 | while (!decomp_param[idx].done) { | |
2473 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
2474 | } | |
2475 | } | |
2476 | qemu_mutex_unlock(&decomp_done_lock); | |
2477 | } | |
2478 | ||
f0afa331 | 2479 | static void compress_threads_load_setup(void) |
56e93d26 JQ |
2480 | { |
2481 | int i, thread_count; | |
2482 | ||
3416ab5b JQ |
2483 | if (!migrate_use_compression()) { |
2484 | return; | |
2485 | } | |
56e93d26 JQ |
2486 | thread_count = migrate_decompress_threads(); |
2487 | decompress_threads = g_new0(QemuThread, thread_count); | |
2488 | decomp_param = g_new0(DecompressParam, thread_count); | |
73a8912b LL |
2489 | qemu_mutex_init(&decomp_done_lock); |
2490 | qemu_cond_init(&decomp_done_cond); | |
56e93d26 JQ |
2491 | for (i = 0; i < thread_count; i++) { |
2492 | qemu_mutex_init(&decomp_param[i].mutex); | |
2493 | qemu_cond_init(&decomp_param[i].cond); | |
2494 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
73a8912b | 2495 | decomp_param[i].done = true; |
90e56fb4 | 2496 | decomp_param[i].quit = false; |
56e93d26 JQ |
2497 | qemu_thread_create(decompress_threads + i, "decompress", |
2498 | do_data_decompress, decomp_param + i, | |
2499 | QEMU_THREAD_JOINABLE); | |
2500 | } | |
2501 | } | |
2502 | ||
f0afa331 | 2503 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
2504 | { |
2505 | int i, thread_count; | |
2506 | ||
3416ab5b JQ |
2507 | if (!migrate_use_compression()) { |
2508 | return; | |
2509 | } | |
56e93d26 JQ |
2510 | thread_count = migrate_decompress_threads(); |
2511 | for (i = 0; i < thread_count; i++) { | |
2512 | qemu_mutex_lock(&decomp_param[i].mutex); | |
90e56fb4 | 2513 | decomp_param[i].quit = true; |
56e93d26 JQ |
2514 | qemu_cond_signal(&decomp_param[i].cond); |
2515 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
2516 | } | |
2517 | for (i = 0; i < thread_count; i++) { | |
2518 | qemu_thread_join(decompress_threads + i); | |
2519 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
2520 | qemu_cond_destroy(&decomp_param[i].cond); | |
2521 | g_free(decomp_param[i].compbuf); | |
2522 | } | |
2523 | g_free(decompress_threads); | |
2524 | g_free(decomp_param); | |
56e93d26 JQ |
2525 | decompress_threads = NULL; |
2526 | decomp_param = NULL; | |
56e93d26 JQ |
2527 | } |
2528 | ||
c1bc6626 | 2529 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
2530 | void *host, int len) |
2531 | { | |
2532 | int idx, thread_count; | |
2533 | ||
2534 | thread_count = migrate_decompress_threads(); | |
73a8912b | 2535 | qemu_mutex_lock(&decomp_done_lock); |
56e93d26 JQ |
2536 | while (true) { |
2537 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 2538 | if (decomp_param[idx].done) { |
33d151f4 LL |
2539 | decomp_param[idx].done = false; |
2540 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 2541 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
2542 | decomp_param[idx].des = host; |
2543 | decomp_param[idx].len = len; | |
33d151f4 LL |
2544 | qemu_cond_signal(&decomp_param[idx].cond); |
2545 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
2546 | break; |
2547 | } | |
2548 | } | |
2549 | if (idx < thread_count) { | |
2550 | break; | |
73a8912b LL |
2551 | } else { |
2552 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
2553 | } |
2554 | } | |
73a8912b | 2555 | qemu_mutex_unlock(&decomp_done_lock); |
56e93d26 JQ |
2556 | } |
2557 | ||
f265e0e4 JQ |
2558 | /** |
2559 | * ram_load_setup: Setup RAM for migration incoming side | |
2560 | * | |
2561 | * Returns zero to indicate success and negative for error | |
2562 | * | |
2563 | * @f: QEMUFile where to receive the data | |
2564 | * @opaque: RAMState pointer | |
2565 | */ | |
2566 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
2567 | { | |
2568 | xbzrle_load_setup(); | |
f0afa331 | 2569 | compress_threads_load_setup(); |
f265e0e4 JQ |
2570 | return 0; |
2571 | } | |
2572 | ||
2573 | static int ram_load_cleanup(void *opaque) | |
2574 | { | |
2575 | xbzrle_load_cleanup(); | |
f0afa331 | 2576 | compress_threads_load_cleanup(); |
f265e0e4 JQ |
2577 | return 0; |
2578 | } | |
2579 | ||
3d0684b2 JQ |
2580 | /** |
2581 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
2582 | * | |
2583 | * Returns 0 for success and negative if there was one error | |
2584 | * | |
2585 | * @mis: current migration incoming state | |
2586 | * | |
2587 | * Allocate data structures etc needed by incoming migration with | |
2588 | * postcopy-ram. postcopy-ram's similarly names | |
2589 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
2590 | */ |
2591 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
2592 | { | |
b8c48993 | 2593 | unsigned long ram_pages = last_ram_page(); |
1caddf8a DDAG |
2594 | |
2595 | return postcopy_ram_incoming_init(mis, ram_pages); | |
2596 | } | |
2597 | ||
3d0684b2 JQ |
2598 | /** |
2599 | * ram_load_postcopy: load a page in postcopy case | |
2600 | * | |
2601 | * Returns 0 for success or -errno in case of error | |
2602 | * | |
a7180877 DDAG |
2603 | * Called in postcopy mode by ram_load(). |
2604 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
2605 | * |
2606 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
2607 | */ |
2608 | static int ram_load_postcopy(QEMUFile *f) | |
2609 | { | |
2610 | int flags = 0, ret = 0; | |
2611 | bool place_needed = false; | |
28abd200 | 2612 | bool matching_page_sizes = false; |
a7180877 DDAG |
2613 | MigrationIncomingState *mis = migration_incoming_get_current(); |
2614 | /* Temporary page that is later 'placed' */ | |
2615 | void *postcopy_host_page = postcopy_get_tmp_page(mis); | |
c53b7ddc | 2616 | void *last_host = NULL; |
a3b6ff6d | 2617 | bool all_zero = false; |
a7180877 DDAG |
2618 | |
2619 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
2620 | ram_addr_t addr; | |
2621 | void *host = NULL; | |
2622 | void *page_buffer = NULL; | |
2623 | void *place_source = NULL; | |
df9ff5e1 | 2624 | RAMBlock *block = NULL; |
a7180877 | 2625 | uint8_t ch; |
a7180877 DDAG |
2626 | |
2627 | addr = qemu_get_be64(f); | |
2628 | flags = addr & ~TARGET_PAGE_MASK; | |
2629 | addr &= TARGET_PAGE_MASK; | |
2630 | ||
2631 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
2632 | place_needed = false; | |
bb890ed5 | 2633 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE)) { |
df9ff5e1 | 2634 | block = ram_block_from_stream(f, flags); |
4c4bad48 HZ |
2635 | |
2636 | host = host_from_ram_block_offset(block, addr); | |
a7180877 DDAG |
2637 | if (!host) { |
2638 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2639 | ret = -EINVAL; | |
2640 | break; | |
2641 | } | |
28abd200 | 2642 | matching_page_sizes = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 2643 | /* |
28abd200 DDAG |
2644 | * Postcopy requires that we place whole host pages atomically; |
2645 | * these may be huge pages for RAMBlocks that are backed by | |
2646 | * hugetlbfs. | |
a7180877 DDAG |
2647 | * To make it atomic, the data is read into a temporary page |
2648 | * that's moved into place later. | |
2649 | * The migration protocol uses, possibly smaller, target-pages | |
2650 | * however the source ensures it always sends all the components | |
2651 | * of a host page in order. | |
2652 | */ | |
2653 | page_buffer = postcopy_host_page + | |
28abd200 | 2654 | ((uintptr_t)host & (block->page_size - 1)); |
a7180877 | 2655 | /* If all TP are zero then we can optimise the place */ |
28abd200 | 2656 | if (!((uintptr_t)host & (block->page_size - 1))) { |
a7180877 | 2657 | all_zero = true; |
c53b7ddc DDAG |
2658 | } else { |
2659 | /* not the 1st TP within the HP */ | |
2660 | if (host != (last_host + TARGET_PAGE_SIZE)) { | |
9af9e0fe | 2661 | error_report("Non-sequential target page %p/%p", |
c53b7ddc DDAG |
2662 | host, last_host); |
2663 | ret = -EINVAL; | |
2664 | break; | |
2665 | } | |
a7180877 DDAG |
2666 | } |
2667 | ||
c53b7ddc | 2668 | |
a7180877 DDAG |
2669 | /* |
2670 | * If it's the last part of a host page then we place the host | |
2671 | * page | |
2672 | */ | |
2673 | place_needed = (((uintptr_t)host + TARGET_PAGE_SIZE) & | |
28abd200 | 2674 | (block->page_size - 1)) == 0; |
a7180877 DDAG |
2675 | place_source = postcopy_host_page; |
2676 | } | |
c53b7ddc | 2677 | last_host = host; |
a7180877 DDAG |
2678 | |
2679 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 2680 | case RAM_SAVE_FLAG_ZERO: |
a7180877 DDAG |
2681 | ch = qemu_get_byte(f); |
2682 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
2683 | if (ch) { | |
2684 | all_zero = false; | |
2685 | } | |
2686 | break; | |
2687 | ||
2688 | case RAM_SAVE_FLAG_PAGE: | |
2689 | all_zero = false; | |
2690 | if (!place_needed || !matching_page_sizes) { | |
2691 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); | |
2692 | } else { | |
2693 | /* Avoids the qemu_file copy during postcopy, which is | |
2694 | * going to do a copy later; can only do it when we | |
2695 | * do this read in one go (matching page sizes) | |
2696 | */ | |
2697 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
2698 | TARGET_PAGE_SIZE); | |
2699 | } | |
2700 | break; | |
2701 | case RAM_SAVE_FLAG_EOS: | |
2702 | /* normal exit */ | |
2703 | break; | |
2704 | default: | |
2705 | error_report("Unknown combination of migration flags: %#x" | |
2706 | " (postcopy mode)", flags); | |
2707 | ret = -EINVAL; | |
2708 | } | |
2709 | ||
2710 | if (place_needed) { | |
2711 | /* This gets called at the last target page in the host page */ | |
df9ff5e1 DDAG |
2712 | void *place_dest = host + TARGET_PAGE_SIZE - block->page_size; |
2713 | ||
a7180877 | 2714 | if (all_zero) { |
df9ff5e1 DDAG |
2715 | ret = postcopy_place_page_zero(mis, place_dest, |
2716 | block->page_size); | |
a7180877 | 2717 | } else { |
df9ff5e1 DDAG |
2718 | ret = postcopy_place_page(mis, place_dest, |
2719 | place_source, block->page_size); | |
a7180877 DDAG |
2720 | } |
2721 | } | |
2722 | if (!ret) { | |
2723 | ret = qemu_file_get_error(f); | |
2724 | } | |
2725 | } | |
2726 | ||
2727 | return ret; | |
2728 | } | |
2729 | ||
56e93d26 JQ |
2730 | static int ram_load(QEMUFile *f, void *opaque, int version_id) |
2731 | { | |
edc60127 | 2732 | int flags = 0, ret = 0, invalid_flags = 0; |
56e93d26 JQ |
2733 | static uint64_t seq_iter; |
2734 | int len = 0; | |
a7180877 DDAG |
2735 | /* |
2736 | * If system is running in postcopy mode, page inserts to host memory must | |
2737 | * be atomic | |
2738 | */ | |
2739 | bool postcopy_running = postcopy_state_get() >= POSTCOPY_INCOMING_LISTENING; | |
ef08fb38 DDAG |
2740 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
2741 | bool postcopy_advised = postcopy_state_get() >= POSTCOPY_INCOMING_ADVISE; | |
56e93d26 JQ |
2742 | |
2743 | seq_iter++; | |
2744 | ||
2745 | if (version_id != 4) { | |
2746 | ret = -EINVAL; | |
2747 | } | |
2748 | ||
edc60127 JQ |
2749 | if (!migrate_use_compression()) { |
2750 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
2751 | } | |
56e93d26 JQ |
2752 | /* This RCU critical section can be very long running. |
2753 | * When RCU reclaims in the code start to become numerous, | |
2754 | * it will be necessary to reduce the granularity of this | |
2755 | * critical section. | |
2756 | */ | |
2757 | rcu_read_lock(); | |
a7180877 DDAG |
2758 | |
2759 | if (postcopy_running) { | |
2760 | ret = ram_load_postcopy(f); | |
2761 | } | |
2762 | ||
2763 | while (!postcopy_running && !ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
56e93d26 | 2764 | ram_addr_t addr, total_ram_bytes; |
a776aa15 | 2765 | void *host = NULL; |
56e93d26 JQ |
2766 | uint8_t ch; |
2767 | ||
2768 | addr = qemu_get_be64(f); | |
2769 | flags = addr & ~TARGET_PAGE_MASK; | |
2770 | addr &= TARGET_PAGE_MASK; | |
2771 | ||
edc60127 JQ |
2772 | if (flags & invalid_flags) { |
2773 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
2774 | error_report("Received an unexpected compressed page"); | |
2775 | } | |
2776 | ||
2777 | ret = -EINVAL; | |
2778 | break; | |
2779 | } | |
2780 | ||
bb890ed5 | 2781 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 2782 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
4c4bad48 HZ |
2783 | RAMBlock *block = ram_block_from_stream(f, flags); |
2784 | ||
2785 | host = host_from_ram_block_offset(block, addr); | |
a776aa15 DDAG |
2786 | if (!host) { |
2787 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
2788 | ret = -EINVAL; | |
2789 | break; | |
2790 | } | |
1db9d8e5 | 2791 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
2792 | } |
2793 | ||
56e93d26 JQ |
2794 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
2795 | case RAM_SAVE_FLAG_MEM_SIZE: | |
2796 | /* Synchronize RAM block list */ | |
2797 | total_ram_bytes = addr; | |
2798 | while (!ret && total_ram_bytes) { | |
2799 | RAMBlock *block; | |
56e93d26 JQ |
2800 | char id[256]; |
2801 | ram_addr_t length; | |
2802 | ||
2803 | len = qemu_get_byte(f); | |
2804 | qemu_get_buffer(f, (uint8_t *)id, len); | |
2805 | id[len] = 0; | |
2806 | length = qemu_get_be64(f); | |
2807 | ||
e3dd7493 DDAG |
2808 | block = qemu_ram_block_by_name(id); |
2809 | if (block) { | |
2810 | if (length != block->used_length) { | |
2811 | Error *local_err = NULL; | |
56e93d26 | 2812 | |
fa53a0e5 | 2813 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
2814 | &local_err); |
2815 | if (local_err) { | |
2816 | error_report_err(local_err); | |
56e93d26 | 2817 | } |
56e93d26 | 2818 | } |
ef08fb38 DDAG |
2819 | /* For postcopy we need to check hugepage sizes match */ |
2820 | if (postcopy_advised && | |
2821 | block->page_size != qemu_host_page_size) { | |
2822 | uint64_t remote_page_size = qemu_get_be64(f); | |
2823 | if (remote_page_size != block->page_size) { | |
2824 | error_report("Mismatched RAM page size %s " | |
2825 | "(local) %zd != %" PRId64, | |
2826 | id, block->page_size, | |
2827 | remote_page_size); | |
2828 | ret = -EINVAL; | |
2829 | } | |
2830 | } | |
e3dd7493 DDAG |
2831 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
2832 | block->idstr); | |
2833 | } else { | |
56e93d26 JQ |
2834 | error_report("Unknown ramblock \"%s\", cannot " |
2835 | "accept migration", id); | |
2836 | ret = -EINVAL; | |
2837 | } | |
2838 | ||
2839 | total_ram_bytes -= length; | |
2840 | } | |
2841 | break; | |
a776aa15 | 2842 | |
bb890ed5 | 2843 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
2844 | ch = qemu_get_byte(f); |
2845 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
2846 | break; | |
a776aa15 | 2847 | |
56e93d26 | 2848 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
2849 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
2850 | break; | |
56e93d26 | 2851 | |
a776aa15 | 2852 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
2853 | len = qemu_get_be32(f); |
2854 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
2855 | error_report("Invalid compressed data length: %d", len); | |
2856 | ret = -EINVAL; | |
2857 | break; | |
2858 | } | |
c1bc6626 | 2859 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 2860 | break; |
a776aa15 | 2861 | |
56e93d26 | 2862 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
2863 | if (load_xbzrle(f, addr, host) < 0) { |
2864 | error_report("Failed to decompress XBZRLE page at " | |
2865 | RAM_ADDR_FMT, addr); | |
2866 | ret = -EINVAL; | |
2867 | break; | |
2868 | } | |
2869 | break; | |
2870 | case RAM_SAVE_FLAG_EOS: | |
2871 | /* normal exit */ | |
2872 | break; | |
2873 | default: | |
2874 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 2875 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 JQ |
2876 | } else { |
2877 | error_report("Unknown combination of migration flags: %#x", | |
2878 | flags); | |
2879 | ret = -EINVAL; | |
2880 | } | |
2881 | } | |
2882 | if (!ret) { | |
2883 | ret = qemu_file_get_error(f); | |
2884 | } | |
2885 | } | |
2886 | ||
5533b2e9 | 2887 | wait_for_decompress_done(); |
56e93d26 | 2888 | rcu_read_unlock(); |
55c4446b | 2889 | trace_ram_load_complete(ret, seq_iter); |
56e93d26 JQ |
2890 | return ret; |
2891 | } | |
2892 | ||
c6467627 VSO |
2893 | static bool ram_has_postcopy(void *opaque) |
2894 | { | |
2895 | return migrate_postcopy_ram(); | |
2896 | } | |
2897 | ||
56e93d26 | 2898 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 2899 | .save_setup = ram_save_setup, |
56e93d26 | 2900 | .save_live_iterate = ram_save_iterate, |
763c906b | 2901 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 2902 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 2903 | .has_postcopy = ram_has_postcopy, |
56e93d26 JQ |
2904 | .save_live_pending = ram_save_pending, |
2905 | .load_state = ram_load, | |
f265e0e4 JQ |
2906 | .save_cleanup = ram_save_cleanup, |
2907 | .load_setup = ram_load_setup, | |
2908 | .load_cleanup = ram_load_cleanup, | |
56e93d26 JQ |
2909 | }; |
2910 | ||
2911 | void ram_mig_init(void) | |
2912 | { | |
2913 | qemu_mutex_init(&XBZRLE.lock); | |
6f37bb8b | 2914 | register_savevm_live(NULL, "ram", 0, 4, &savevm_ram_handlers, &ram_state); |
56e93d26 | 2915 | } |