]>
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 | */ | |
e688df6b | 28 | |
1393a485 | 29 | #include "qemu/osdep.h" |
f348b6d1 | 30 | #include "qemu/cutils.h" |
56e93d26 JQ |
31 | #include "qemu/bitops.h" |
32 | #include "qemu/bitmap.h" | |
b85ea5fa | 33 | #include "qemu/madvise.h" |
7205c9ec | 34 | #include "qemu/main-loop.h" |
c0e0825c | 35 | #include "io/channel-null.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" |
53d37d36 | 43 | #include "page_cache.h" |
56e93d26 | 44 | #include "qemu/error-report.h" |
e688df6b | 45 | #include "qapi/error.h" |
ab7cbb0b | 46 | #include "qapi/qapi-types-migration.h" |
9af23989 | 47 | #include "qapi/qapi-events-migration.h" |
8acabf69 | 48 | #include "qapi/qmp/qerror.h" |
56e93d26 | 49 | #include "trace.h" |
56e93d26 | 50 | #include "exec/ram_addr.h" |
f9494614 | 51 | #include "exec/target_page.h" |
56e93d26 | 52 | #include "qemu/rcu_queue.h" |
a91246c9 | 53 | #include "migration/colo.h" |
53d37d36 | 54 | #include "block.h" |
b0c3cf94 | 55 | #include "sysemu/cpu-throttle.h" |
edd090c7 | 56 | #include "savevm.h" |
b9ee2f7d | 57 | #include "qemu/iov.h" |
d32ca5ad | 58 | #include "multifd.h" |
278e2f55 AG |
59 | #include "sysemu/runstate.h" |
60 | ||
e5fdf920 LS |
61 | #include "hw/boards.h" /* for machine_dump_guest_core() */ |
62 | ||
278e2f55 AG |
63 | #if defined(__linux__) |
64 | #include "qemu/userfaultfd.h" | |
65 | #endif /* defined(__linux__) */ | |
56e93d26 | 66 | |
56e93d26 JQ |
67 | /***********************************************************/ |
68 | /* ram save/restore */ | |
69 | ||
bb890ed5 JQ |
70 | /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it |
71 | * worked for pages that where filled with the same char. We switched | |
72 | * it to only search for the zero value. And to avoid confusion with | |
73 | * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it. | |
74 | */ | |
75 | ||
56e93d26 | 76 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ |
bb890ed5 | 77 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
78 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
79 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
80 | #define RAM_SAVE_FLAG_EOS 0x10 | |
81 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
82 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
83 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
84 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
85 | ||
9360447d JQ |
86 | XBZRLECacheStats xbzrle_counters; |
87 | ||
f1668764 PX |
88 | /* used by the search for pages to send */ |
89 | struct PageSearchStatus { | |
90 | /* The migration channel used for a specific host page */ | |
91 | QEMUFile *pss_channel; | |
ec6f3ab9 PX |
92 | /* Last block from where we have sent data */ |
93 | RAMBlock *last_sent_block; | |
f1668764 PX |
94 | /* Current block being searched */ |
95 | RAMBlock *block; | |
96 | /* Current page to search from */ | |
97 | unsigned long page; | |
98 | /* Set once we wrap around */ | |
99 | bool complete_round; | |
f1668764 PX |
100 | /* Whether we're sending a host page */ |
101 | bool host_page_sending; | |
102 | /* The start/end of current host page. Invalid if host_page_sending==false */ | |
103 | unsigned long host_page_start; | |
104 | unsigned long host_page_end; | |
105 | }; | |
106 | typedef struct PageSearchStatus PageSearchStatus; | |
107 | ||
56e93d26 JQ |
108 | /* struct contains XBZRLE cache and a static page |
109 | used by the compression */ | |
110 | static struct { | |
111 | /* buffer used for XBZRLE encoding */ | |
112 | uint8_t *encoded_buf; | |
113 | /* buffer for storing page content */ | |
114 | uint8_t *current_buf; | |
115 | /* Cache for XBZRLE, Protected by lock. */ | |
116 | PageCache *cache; | |
117 | QemuMutex lock; | |
c00e0928 JQ |
118 | /* it will store a page full of zeros */ |
119 | uint8_t *zero_target_page; | |
f265e0e4 JQ |
120 | /* buffer used for XBZRLE decoding */ |
121 | uint8_t *decoded_buf; | |
56e93d26 JQ |
122 | } XBZRLE; |
123 | ||
56e93d26 JQ |
124 | static void XBZRLE_cache_lock(void) |
125 | { | |
f4c51a6b | 126 | if (migrate_use_xbzrle()) { |
56e93d26 | 127 | qemu_mutex_lock(&XBZRLE.lock); |
f4c51a6b | 128 | } |
56e93d26 JQ |
129 | } |
130 | ||
131 | static void XBZRLE_cache_unlock(void) | |
132 | { | |
f4c51a6b | 133 | if (migrate_use_xbzrle()) { |
56e93d26 | 134 | qemu_mutex_unlock(&XBZRLE.lock); |
f4c51a6b | 135 | } |
56e93d26 JQ |
136 | } |
137 | ||
3d0684b2 JQ |
138 | /** |
139 | * xbzrle_cache_resize: resize the xbzrle cache | |
140 | * | |
cbde7be9 | 141 | * This function is called from migrate_params_apply in main |
3d0684b2 JQ |
142 | * thread, possibly while a migration is in progress. A running |
143 | * migration may be using the cache and might finish during this call, | |
144 | * hence changes to the cache are protected by XBZRLE.lock(). | |
145 | * | |
c9dede2d | 146 | * Returns 0 for success or -1 for error |
3d0684b2 JQ |
147 | * |
148 | * @new_size: new cache size | |
8acabf69 | 149 | * @errp: set *errp if the check failed, with reason |
56e93d26 | 150 | */ |
8b9407a0 | 151 | int xbzrle_cache_resize(uint64_t new_size, Error **errp) |
56e93d26 JQ |
152 | { |
153 | PageCache *new_cache; | |
c9dede2d | 154 | int64_t ret = 0; |
56e93d26 | 155 | |
8acabf69 JQ |
156 | /* Check for truncation */ |
157 | if (new_size != (size_t)new_size) { | |
158 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
159 | "exceeding address space"); | |
160 | return -1; | |
161 | } | |
162 | ||
2a313e5c JQ |
163 | if (new_size == migrate_xbzrle_cache_size()) { |
164 | /* nothing to do */ | |
c9dede2d | 165 | return 0; |
2a313e5c JQ |
166 | } |
167 | ||
56e93d26 JQ |
168 | XBZRLE_cache_lock(); |
169 | ||
170 | if (XBZRLE.cache != NULL) { | |
80f8dfde | 171 | new_cache = cache_init(new_size, TARGET_PAGE_SIZE, errp); |
56e93d26 | 172 | if (!new_cache) { |
56e93d26 JQ |
173 | ret = -1; |
174 | goto out; | |
175 | } | |
176 | ||
177 | cache_fini(XBZRLE.cache); | |
178 | XBZRLE.cache = new_cache; | |
179 | } | |
56e93d26 JQ |
180 | out: |
181 | XBZRLE_cache_unlock(); | |
182 | return ret; | |
183 | } | |
184 | ||
20123ee1 PX |
185 | static bool postcopy_preempt_active(void) |
186 | { | |
187 | return migrate_postcopy_preempt() && migration_in_postcopy(); | |
188 | } | |
189 | ||
3ded54b1 | 190 | bool ramblock_is_ignored(RAMBlock *block) |
fbd162e6 YK |
191 | { |
192 | return !qemu_ram_is_migratable(block) || | |
193 | (migrate_ignore_shared() && qemu_ram_is_shared(block)); | |
194 | } | |
195 | ||
343f632c DDAG |
196 | #undef RAMBLOCK_FOREACH |
197 | ||
fbd162e6 YK |
198 | int foreach_not_ignored_block(RAMBlockIterFunc func, void *opaque) |
199 | { | |
200 | RAMBlock *block; | |
201 | int ret = 0; | |
202 | ||
89ac5a1d DDAG |
203 | RCU_READ_LOCK_GUARD(); |
204 | ||
fbd162e6 YK |
205 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
206 | ret = func(block, opaque); | |
207 | if (ret) { | |
208 | break; | |
209 | } | |
210 | } | |
fbd162e6 YK |
211 | return ret; |
212 | } | |
213 | ||
f9494614 AP |
214 | static void ramblock_recv_map_init(void) |
215 | { | |
216 | RAMBlock *rb; | |
217 | ||
fbd162e6 | 218 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
219 | assert(!rb->receivedmap); |
220 | rb->receivedmap = bitmap_new(rb->max_length >> qemu_target_page_bits()); | |
221 | } | |
222 | } | |
223 | ||
224 | int ramblock_recv_bitmap_test(RAMBlock *rb, void *host_addr) | |
225 | { | |
226 | return test_bit(ramblock_recv_bitmap_offset(host_addr, rb), | |
227 | rb->receivedmap); | |
228 | } | |
229 | ||
1cba9f6e DDAG |
230 | bool ramblock_recv_bitmap_test_byte_offset(RAMBlock *rb, uint64_t byte_offset) |
231 | { | |
232 | return test_bit(byte_offset >> TARGET_PAGE_BITS, rb->receivedmap); | |
233 | } | |
234 | ||
f9494614 AP |
235 | void ramblock_recv_bitmap_set(RAMBlock *rb, void *host_addr) |
236 | { | |
237 | set_bit_atomic(ramblock_recv_bitmap_offset(host_addr, rb), rb->receivedmap); | |
238 | } | |
239 | ||
240 | void ramblock_recv_bitmap_set_range(RAMBlock *rb, void *host_addr, | |
241 | size_t nr) | |
242 | { | |
243 | bitmap_set_atomic(rb->receivedmap, | |
244 | ramblock_recv_bitmap_offset(host_addr, rb), | |
245 | nr); | |
246 | } | |
247 | ||
a335debb PX |
248 | #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL) |
249 | ||
250 | /* | |
251 | * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes). | |
252 | * | |
253 | * Returns >0 if success with sent bytes, or <0 if error. | |
254 | */ | |
255 | int64_t ramblock_recv_bitmap_send(QEMUFile *file, | |
256 | const char *block_name) | |
257 | { | |
258 | RAMBlock *block = qemu_ram_block_by_name(block_name); | |
259 | unsigned long *le_bitmap, nbits; | |
260 | uint64_t size; | |
261 | ||
262 | if (!block) { | |
263 | error_report("%s: invalid block name: %s", __func__, block_name); | |
264 | return -1; | |
265 | } | |
266 | ||
898ba906 | 267 | nbits = block->postcopy_length >> TARGET_PAGE_BITS; |
a335debb PX |
268 | |
269 | /* | |
270 | * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit | |
271 | * machines we may need 4 more bytes for padding (see below | |
272 | * comment). So extend it a bit before hand. | |
273 | */ | |
274 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
275 | ||
276 | /* | |
277 | * Always use little endian when sending the bitmap. This is | |
278 | * required that when source and destination VMs are not using the | |
3a4452d8 | 279 | * same endianness. (Note: big endian won't work.) |
a335debb PX |
280 | */ |
281 | bitmap_to_le(le_bitmap, block->receivedmap, nbits); | |
282 | ||
283 | /* Size of the bitmap, in bytes */ | |
a725ef9f | 284 | size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
285 | |
286 | /* | |
287 | * size is always aligned to 8 bytes for 64bit machines, but it | |
288 | * may not be true for 32bit machines. We need this padding to | |
289 | * make sure the migration can survive even between 32bit and | |
290 | * 64bit machines. | |
291 | */ | |
292 | size = ROUND_UP(size, 8); | |
293 | ||
294 | qemu_put_be64(file, size); | |
295 | qemu_put_buffer(file, (const uint8_t *)le_bitmap, size); | |
296 | /* | |
297 | * Mark as an end, in case the middle part is screwed up due to | |
3a4452d8 | 298 | * some "mysterious" reason. |
a335debb PX |
299 | */ |
300 | qemu_put_be64(file, RAMBLOCK_RECV_BITMAP_ENDING); | |
301 | qemu_fflush(file); | |
302 | ||
bf269906 | 303 | g_free(le_bitmap); |
a335debb PX |
304 | |
305 | if (qemu_file_get_error(file)) { | |
306 | return qemu_file_get_error(file); | |
307 | } | |
308 | ||
309 | return size + sizeof(size); | |
310 | } | |
311 | ||
ec481c6c JQ |
312 | /* |
313 | * An outstanding page request, on the source, having been received | |
314 | * and queued | |
315 | */ | |
316 | struct RAMSrcPageRequest { | |
317 | RAMBlock *rb; | |
318 | hwaddr offset; | |
319 | hwaddr len; | |
320 | ||
321 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
322 | }; | |
323 | ||
6f37bb8b JQ |
324 | /* State of RAM for migration */ |
325 | struct RAMState { | |
f1668764 PX |
326 | /* |
327 | * PageSearchStatus structures for the channels when send pages. | |
328 | * Protected by the bitmap_mutex. | |
329 | */ | |
330 | PageSearchStatus pss[RAM_CHANNEL_MAX]; | |
278e2f55 AG |
331 | /* UFFD file descriptor, used in 'write-tracking' migration */ |
332 | int uffdio_fd; | |
6f37bb8b JQ |
333 | /* Last block that we have visited searching for dirty pages */ |
334 | RAMBlock *last_seen_block; | |
269ace29 JQ |
335 | /* Last dirty target page we have sent */ |
336 | ram_addr_t last_page; | |
6f37bb8b JQ |
337 | /* last ram version we have seen */ |
338 | uint32_t last_version; | |
8d820d6f JQ |
339 | /* How many times we have dirty too many pages */ |
340 | int dirty_rate_high_cnt; | |
f664da80 JQ |
341 | /* these variables are used for bitmap sync */ |
342 | /* last time we did a full bitmap_sync */ | |
343 | int64_t time_last_bitmap_sync; | |
eac74159 | 344 | /* bytes transferred at start_time */ |
c4bdf0cf | 345 | uint64_t bytes_xfer_prev; |
a66cd90c | 346 | /* number of dirty pages since start_time */ |
68908ed6 | 347 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
348 | /* xbzrle misses since the beginning of the period */ |
349 | uint64_t xbzrle_cache_miss_prev; | |
e460a4b1 WW |
350 | /* Amount of xbzrle pages since the beginning of the period */ |
351 | uint64_t xbzrle_pages_prev; | |
352 | /* Amount of xbzrle encoded bytes since the beginning of the period */ | |
353 | uint64_t xbzrle_bytes_prev; | |
1a373522 DH |
354 | /* Start using XBZRLE (e.g., after the first round). */ |
355 | bool xbzrle_enabled; | |
05931ec5 JQ |
356 | /* Are we on the last stage of migration */ |
357 | bool last_stage; | |
76e03000 XG |
358 | /* compression statistics since the beginning of the period */ |
359 | /* amount of count that no free thread to compress data */ | |
360 | uint64_t compress_thread_busy_prev; | |
361 | /* amount bytes after compression */ | |
362 | uint64_t compressed_size_prev; | |
363 | /* amount of compressed pages */ | |
364 | uint64_t compress_pages_prev; | |
365 | ||
be8b02ed XG |
366 | /* total handled target pages at the beginning of period */ |
367 | uint64_t target_page_count_prev; | |
368 | /* total handled target pages since start */ | |
369 | uint64_t target_page_count; | |
9360447d | 370 | /* number of dirty bits in the bitmap */ |
2dfaf12e | 371 | uint64_t migration_dirty_pages; |
f1668764 PX |
372 | /* |
373 | * Protects: | |
374 | * - dirty/clear bitmap | |
375 | * - migration_dirty_pages | |
376 | * - pss structures | |
377 | */ | |
108cfae0 | 378 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
379 | /* The RAMBlock used in the last src_page_requests */ |
380 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
381 | /* Queue of outstanding page requests from the destination */ |
382 | QemuMutex src_page_req_mutex; | |
b58deb34 | 383 | QSIMPLEQ_HEAD(, RAMSrcPageRequest) src_page_requests; |
6f37bb8b JQ |
384 | }; |
385 | typedef struct RAMState RAMState; | |
386 | ||
53518d94 | 387 | static RAMState *ram_state; |
6f37bb8b | 388 | |
bd227060 WW |
389 | static NotifierWithReturnList precopy_notifier_list; |
390 | ||
a1fe28df PX |
391 | /* Whether postcopy has queued requests? */ |
392 | static bool postcopy_has_request(RAMState *rs) | |
393 | { | |
394 | return !QSIMPLEQ_EMPTY_ATOMIC(&rs->src_page_requests); | |
395 | } | |
396 | ||
bd227060 WW |
397 | void precopy_infrastructure_init(void) |
398 | { | |
399 | notifier_with_return_list_init(&precopy_notifier_list); | |
400 | } | |
401 | ||
402 | void precopy_add_notifier(NotifierWithReturn *n) | |
403 | { | |
404 | notifier_with_return_list_add(&precopy_notifier_list, n); | |
405 | } | |
406 | ||
407 | void precopy_remove_notifier(NotifierWithReturn *n) | |
408 | { | |
409 | notifier_with_return_remove(n); | |
410 | } | |
411 | ||
412 | int precopy_notify(PrecopyNotifyReason reason, Error **errp) | |
413 | { | |
414 | PrecopyNotifyData pnd; | |
415 | pnd.reason = reason; | |
416 | pnd.errp = errp; | |
417 | ||
418 | return notifier_with_return_list_notify(&precopy_notifier_list, &pnd); | |
419 | } | |
420 | ||
9edabd4d | 421 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 422 | { |
bae416e5 DDAG |
423 | return ram_state ? (ram_state->migration_dirty_pages * TARGET_PAGE_SIZE) : |
424 | 0; | |
2f4fde93 JQ |
425 | } |
426 | ||
23b7576d PX |
427 | /* |
428 | * NOTE: not all stats in ram_counters are used in reality. See comments | |
429 | * for struct MigrationAtomicStats. The ultimate result of ram migration | |
430 | * counters will be a merged version with both ram_counters and the atomic | |
431 | * fields in ram_atomic_counters. | |
432 | */ | |
9360447d | 433 | MigrationStats ram_counters; |
23b7576d | 434 | MigrationAtomicStats ram_atomic_counters; |
96506894 | 435 | |
26a26069 | 436 | void ram_transferred_add(uint64_t bytes) |
4c2d0f6d | 437 | { |
ae680668 DE |
438 | if (runstate_is_running()) { |
439 | ram_counters.precopy_bytes += bytes; | |
440 | } else if (migration_in_postcopy()) { | |
23b7576d | 441 | stat64_add(&ram_atomic_counters.postcopy_bytes, bytes); |
ae680668 DE |
442 | } else { |
443 | ram_counters.downtime_bytes += bytes; | |
444 | } | |
23b7576d | 445 | stat64_add(&ram_atomic_counters.transferred, bytes); |
4c2d0f6d DE |
446 | } |
447 | ||
d59c40cc LB |
448 | void dirty_sync_missed_zero_copy(void) |
449 | { | |
450 | ram_counters.dirty_sync_missed_zero_copy++; | |
451 | } | |
452 | ||
76e03000 XG |
453 | CompressionStats compression_counters; |
454 | ||
56e93d26 | 455 | struct CompressParam { |
56e93d26 | 456 | bool done; |
90e56fb4 | 457 | bool quit; |
5e5fdcff | 458 | bool zero_page; |
56e93d26 JQ |
459 | QEMUFile *file; |
460 | QemuMutex mutex; | |
461 | QemuCond cond; | |
462 | RAMBlock *block; | |
463 | ram_addr_t offset; | |
34ab9e97 XG |
464 | |
465 | /* internally used fields */ | |
dcaf446e | 466 | z_stream stream; |
34ab9e97 | 467 | uint8_t *originbuf; |
56e93d26 JQ |
468 | }; |
469 | typedef struct CompressParam CompressParam; | |
470 | ||
471 | struct DecompressParam { | |
73a8912b | 472 | bool done; |
90e56fb4 | 473 | bool quit; |
56e93d26 JQ |
474 | QemuMutex mutex; |
475 | QemuCond cond; | |
476 | void *des; | |
d341d9f3 | 477 | uint8_t *compbuf; |
56e93d26 | 478 | int len; |
797ca154 | 479 | z_stream stream; |
56e93d26 JQ |
480 | }; |
481 | typedef struct DecompressParam DecompressParam; | |
482 | ||
483 | static CompressParam *comp_param; | |
484 | static QemuThread *compress_threads; | |
485 | /* comp_done_cond is used to wake up the migration thread when | |
486 | * one of the compression threads has finished the compression. | |
487 | * comp_done_lock is used to co-work with comp_done_cond. | |
488 | */ | |
0d9f9a5c LL |
489 | static QemuMutex comp_done_lock; |
490 | static QemuCond comp_done_cond; | |
56e93d26 | 491 | |
34ab9e97 | 492 | static QEMUFile *decomp_file; |
56e93d26 JQ |
493 | static DecompressParam *decomp_param; |
494 | static QemuThread *decompress_threads; | |
73a8912b LL |
495 | static QemuMutex decomp_done_lock; |
496 | static QemuCond decomp_done_cond; | |
56e93d26 | 497 | |
93589827 PX |
498 | static int ram_save_host_page_urgent(PageSearchStatus *pss); |
499 | ||
5e5fdcff | 500 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 501 | ram_addr_t offset, uint8_t *source_buf); |
56e93d26 | 502 | |
ebd88a49 PX |
503 | /* NOTE: page is the PFN not real ram_addr_t. */ |
504 | static void pss_init(PageSearchStatus *pss, RAMBlock *rb, ram_addr_t page) | |
505 | { | |
506 | pss->block = rb; | |
507 | pss->page = page; | |
508 | pss->complete_round = false; | |
509 | } | |
510 | ||
93589827 PX |
511 | /* |
512 | * Check whether two PSSs are actively sending the same page. Return true | |
513 | * if it is, false otherwise. | |
514 | */ | |
515 | static bool pss_overlap(PageSearchStatus *pss1, PageSearchStatus *pss2) | |
516 | { | |
517 | return pss1->host_page_sending && pss2->host_page_sending && | |
518 | (pss1->host_page_start == pss2->host_page_start); | |
519 | } | |
520 | ||
56e93d26 JQ |
521 | static void *do_data_compress(void *opaque) |
522 | { | |
523 | CompressParam *param = opaque; | |
a7a9a88f LL |
524 | RAMBlock *block; |
525 | ram_addr_t offset; | |
5e5fdcff | 526 | bool zero_page; |
56e93d26 | 527 | |
a7a9a88f | 528 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 529 | while (!param->quit) { |
a7a9a88f LL |
530 | if (param->block) { |
531 | block = param->block; | |
532 | offset = param->offset; | |
533 | param->block = NULL; | |
534 | qemu_mutex_unlock(¶m->mutex); | |
535 | ||
5e5fdcff XG |
536 | zero_page = do_compress_ram_page(param->file, ¶m->stream, |
537 | block, offset, param->originbuf); | |
a7a9a88f | 538 | |
0d9f9a5c | 539 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 540 | param->done = true; |
5e5fdcff | 541 | param->zero_page = zero_page; |
0d9f9a5c LL |
542 | qemu_cond_signal(&comp_done_cond); |
543 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
544 | |
545 | qemu_mutex_lock(¶m->mutex); | |
546 | } else { | |
56e93d26 JQ |
547 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
548 | } | |
56e93d26 | 549 | } |
a7a9a88f | 550 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
551 | |
552 | return NULL; | |
553 | } | |
554 | ||
f0afa331 | 555 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
556 | { |
557 | int i, thread_count; | |
558 | ||
05306935 | 559 | if (!migrate_use_compression() || !comp_param) { |
56e93d26 JQ |
560 | return; |
561 | } | |
05306935 | 562 | |
56e93d26 JQ |
563 | thread_count = migrate_compress_threads(); |
564 | for (i = 0; i < thread_count; i++) { | |
dcaf446e XG |
565 | /* |
566 | * we use it as a indicator which shows if the thread is | |
567 | * properly init'd or not | |
568 | */ | |
569 | if (!comp_param[i].file) { | |
570 | break; | |
571 | } | |
05306935 FL |
572 | |
573 | qemu_mutex_lock(&comp_param[i].mutex); | |
574 | comp_param[i].quit = true; | |
575 | qemu_cond_signal(&comp_param[i].cond); | |
576 | qemu_mutex_unlock(&comp_param[i].mutex); | |
577 | ||
56e93d26 | 578 | qemu_thread_join(compress_threads + i); |
56e93d26 JQ |
579 | qemu_mutex_destroy(&comp_param[i].mutex); |
580 | qemu_cond_destroy(&comp_param[i].cond); | |
dcaf446e | 581 | deflateEnd(&comp_param[i].stream); |
34ab9e97 | 582 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
583 | qemu_fclose(comp_param[i].file); |
584 | comp_param[i].file = NULL; | |
56e93d26 | 585 | } |
0d9f9a5c LL |
586 | qemu_mutex_destroy(&comp_done_lock); |
587 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
588 | g_free(compress_threads); |
589 | g_free(comp_param); | |
56e93d26 JQ |
590 | compress_threads = NULL; |
591 | comp_param = NULL; | |
56e93d26 JQ |
592 | } |
593 | ||
dcaf446e | 594 | static int compress_threads_save_setup(void) |
56e93d26 JQ |
595 | { |
596 | int i, thread_count; | |
597 | ||
598 | if (!migrate_use_compression()) { | |
dcaf446e | 599 | return 0; |
56e93d26 | 600 | } |
56e93d26 JQ |
601 | thread_count = migrate_compress_threads(); |
602 | compress_threads = g_new0(QemuThread, thread_count); | |
603 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
604 | qemu_cond_init(&comp_done_cond); |
605 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 606 | for (i = 0; i < thread_count; i++) { |
34ab9e97 XG |
607 | comp_param[i].originbuf = g_try_malloc(TARGET_PAGE_SIZE); |
608 | if (!comp_param[i].originbuf) { | |
609 | goto exit; | |
610 | } | |
611 | ||
dcaf446e XG |
612 | if (deflateInit(&comp_param[i].stream, |
613 | migrate_compress_level()) != Z_OK) { | |
34ab9e97 | 614 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
615 | goto exit; |
616 | } | |
617 | ||
e110aa91 C |
618 | /* comp_param[i].file is just used as a dummy buffer to save data, |
619 | * set its ops to empty. | |
56e93d26 | 620 | */ |
77ef2dc1 | 621 | comp_param[i].file = qemu_file_new_output( |
c0e0825c | 622 | QIO_CHANNEL(qio_channel_null_new())); |
56e93d26 | 623 | comp_param[i].done = true; |
90e56fb4 | 624 | comp_param[i].quit = false; |
56e93d26 JQ |
625 | qemu_mutex_init(&comp_param[i].mutex); |
626 | qemu_cond_init(&comp_param[i].cond); | |
627 | qemu_thread_create(compress_threads + i, "compress", | |
628 | do_data_compress, comp_param + i, | |
629 | QEMU_THREAD_JOINABLE); | |
630 | } | |
dcaf446e XG |
631 | return 0; |
632 | ||
633 | exit: | |
634 | compress_threads_save_cleanup(); | |
635 | return -1; | |
56e93d26 JQ |
636 | } |
637 | ||
638 | /** | |
3d0684b2 | 639 | * save_page_header: write page header to wire |
56e93d26 JQ |
640 | * |
641 | * If this is the 1st block, it also writes the block identification | |
642 | * | |
3d0684b2 | 643 | * Returns the number of bytes written |
56e93d26 | 644 | * |
ec6f3ab9 | 645 | * @pss: current PSS channel status |
56e93d26 JQ |
646 | * @block: block that contains the page we want to send |
647 | * @offset: offset inside the block for the page | |
648 | * in the lower bits, it contains flags | |
649 | */ | |
ec6f3ab9 | 650 | static size_t save_page_header(PageSearchStatus *pss, RAMBlock *block, |
2bf3aa85 | 651 | ram_addr_t offset) |
56e93d26 | 652 | { |
9f5f380b | 653 | size_t size, len; |
ec6f3ab9 PX |
654 | bool same_block = (block == pss->last_sent_block); |
655 | QEMUFile *f = pss->pss_channel; | |
56e93d26 | 656 | |
10661f11 | 657 | if (same_block) { |
24795694 JQ |
658 | offset |= RAM_SAVE_FLAG_CONTINUE; |
659 | } | |
2bf3aa85 | 660 | qemu_put_be64(f, offset); |
56e93d26 JQ |
661 | size = 8; |
662 | ||
10661f11 | 663 | if (!same_block) { |
9f5f380b | 664 | len = strlen(block->idstr); |
2bf3aa85 JQ |
665 | qemu_put_byte(f, len); |
666 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 667 | size += 1 + len; |
ec6f3ab9 | 668 | pss->last_sent_block = block; |
56e93d26 JQ |
669 | } |
670 | return size; | |
671 | } | |
672 | ||
3d0684b2 | 673 | /** |
179a8080 | 674 | * mig_throttle_guest_down: throttle down the guest |
3d0684b2 JQ |
675 | * |
676 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
677 | * writes. If guest dirty memory rate is reduced below the rate at | |
678 | * which we can transfer pages to the destination then we should be | |
679 | * able to complete migration. Some workloads dirty memory way too | |
680 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 | 681 | */ |
cbbf8182 KZ |
682 | static void mig_throttle_guest_down(uint64_t bytes_dirty_period, |
683 | uint64_t bytes_dirty_threshold) | |
070afca2 JH |
684 | { |
685 | MigrationState *s = migrate_get_current(); | |
2594f56d | 686 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
cbbf8182 KZ |
687 | uint64_t pct_increment = s->parameters.cpu_throttle_increment; |
688 | bool pct_tailslow = s->parameters.cpu_throttle_tailslow; | |
4cbc9c7f | 689 | int pct_max = s->parameters.max_cpu_throttle; |
070afca2 | 690 | |
cbbf8182 KZ |
691 | uint64_t throttle_now = cpu_throttle_get_percentage(); |
692 | uint64_t cpu_now, cpu_ideal, throttle_inc; | |
693 | ||
070afca2 JH |
694 | /* We have not started throttling yet. Let's start it. */ |
695 | if (!cpu_throttle_active()) { | |
696 | cpu_throttle_set(pct_initial); | |
697 | } else { | |
698 | /* Throttling already on, just increase the rate */ | |
cbbf8182 KZ |
699 | if (!pct_tailslow) { |
700 | throttle_inc = pct_increment; | |
701 | } else { | |
702 | /* Compute the ideal CPU percentage used by Guest, which may | |
703 | * make the dirty rate match the dirty rate threshold. */ | |
704 | cpu_now = 100 - throttle_now; | |
705 | cpu_ideal = cpu_now * (bytes_dirty_threshold * 1.0 / | |
706 | bytes_dirty_period); | |
707 | throttle_inc = MIN(cpu_now - cpu_ideal, pct_increment); | |
708 | } | |
709 | cpu_throttle_set(MIN(throttle_now + throttle_inc, pct_max)); | |
070afca2 JH |
710 | } |
711 | } | |
712 | ||
91fe9a8d RL |
713 | void mig_throttle_counter_reset(void) |
714 | { | |
715 | RAMState *rs = ram_state; | |
716 | ||
717 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
718 | rs->num_dirty_pages_period = 0; | |
23b7576d | 719 | rs->bytes_xfer_prev = stat64_get(&ram_atomic_counters.transferred); |
91fe9a8d RL |
720 | } |
721 | ||
3d0684b2 JQ |
722 | /** |
723 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
724 | * | |
6f37bb8b | 725 | * @rs: current RAM state |
3d0684b2 JQ |
726 | * @current_addr: address for the zero page |
727 | * | |
728 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
729 | * The important thing is that a stale (not-yet-0'd) page be replaced |
730 | * by the new data. | |
731 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 732 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 733 | */ |
6f37bb8b | 734 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 735 | { |
56e93d26 JQ |
736 | /* We don't care if this fails to allocate a new cache page |
737 | * as long as it updated an old one */ | |
c00e0928 | 738 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 739 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
740 | } |
741 | ||
742 | #define ENCODING_FLAG_XBZRLE 0x1 | |
743 | ||
744 | /** | |
745 | * save_xbzrle_page: compress and send current page | |
746 | * | |
747 | * Returns: 1 means that we wrote the page | |
748 | * 0 means that page is identical to the one already sent | |
749 | * -1 means that xbzrle would be longer than normal | |
750 | * | |
5a987738 | 751 | * @rs: current RAM state |
ec6f3ab9 | 752 | * @pss: current PSS channel |
3d0684b2 JQ |
753 | * @current_data: pointer to the address of the page contents |
754 | * @current_addr: addr of the page | |
56e93d26 JQ |
755 | * @block: block that contains the page we want to send |
756 | * @offset: offset inside the block for the page | |
56e93d26 | 757 | */ |
ec6f3ab9 | 758 | static int save_xbzrle_page(RAMState *rs, PageSearchStatus *pss, |
61717ea9 PX |
759 | uint8_t **current_data, ram_addr_t current_addr, |
760 | RAMBlock *block, ram_addr_t offset) | |
56e93d26 JQ |
761 | { |
762 | int encoded_len = 0, bytes_xbzrle; | |
763 | uint8_t *prev_cached_page; | |
ec6f3ab9 | 764 | QEMUFile *file = pss->pss_channel; |
56e93d26 | 765 | |
9360447d JQ |
766 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
767 | ram_counters.dirty_sync_count)) { | |
768 | xbzrle_counters.cache_miss++; | |
05931ec5 | 769 | if (!rs->last_stage) { |
56e93d26 | 770 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, |
9360447d | 771 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
772 | return -1; |
773 | } else { | |
774 | /* update *current_data when the page has been | |
775 | inserted into cache */ | |
776 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
777 | } | |
778 | } | |
779 | return -1; | |
780 | } | |
781 | ||
e460a4b1 WW |
782 | /* |
783 | * Reaching here means the page has hit the xbzrle cache, no matter what | |
784 | * encoding result it is (normal encoding, overflow or skipping the page), | |
3a4452d8 | 785 | * count the page as encoded. This is used to calculate the encoding rate. |
e460a4b1 WW |
786 | * |
787 | * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB, | |
788 | * 2nd page turns out to be skipped (i.e. no new bytes written to the | |
789 | * page), the overall encoding rate will be 8KB / 2KB = 4, which has the | |
790 | * skipped page included. In this way, the encoding rate can tell if the | |
791 | * guest page is good for xbzrle encoding. | |
792 | */ | |
793 | xbzrle_counters.pages++; | |
56e93d26 JQ |
794 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); |
795 | ||
796 | /* save current buffer into memory */ | |
797 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
798 | ||
799 | /* XBZRLE encoding (if there is no overflow) */ | |
800 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
801 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
802 | TARGET_PAGE_SIZE); | |
ca353803 WY |
803 | |
804 | /* | |
805 | * Update the cache contents, so that it corresponds to the data | |
806 | * sent, in all cases except where we skip the page. | |
807 | */ | |
05931ec5 | 808 | if (!rs->last_stage && encoded_len != 0) { |
ca353803 WY |
809 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); |
810 | /* | |
811 | * In the case where we couldn't compress, ensure that the caller | |
812 | * sends the data from the cache, since the guest might have | |
813 | * changed the RAM since we copied it. | |
814 | */ | |
815 | *current_data = prev_cached_page; | |
816 | } | |
817 | ||
56e93d26 | 818 | if (encoded_len == 0) { |
55c4446b | 819 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
820 | return 0; |
821 | } else if (encoded_len == -1) { | |
55c4446b | 822 | trace_save_xbzrle_page_overflow(); |
9360447d | 823 | xbzrle_counters.overflow++; |
e460a4b1 | 824 | xbzrle_counters.bytes += TARGET_PAGE_SIZE; |
56e93d26 JQ |
825 | return -1; |
826 | } | |
827 | ||
56e93d26 | 828 | /* Send XBZRLE based compressed page */ |
ec6f3ab9 | 829 | bytes_xbzrle = save_page_header(pss, block, |
204b88b8 | 830 | offset | RAM_SAVE_FLAG_XBZRLE); |
61717ea9 PX |
831 | qemu_put_byte(file, ENCODING_FLAG_XBZRLE); |
832 | qemu_put_be16(file, encoded_len); | |
833 | qemu_put_buffer(file, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 834 | bytes_xbzrle += encoded_len + 1 + 2; |
e460a4b1 WW |
835 | /* |
836 | * Like compressed_size (please see update_compress_thread_counts), | |
837 | * the xbzrle encoded bytes don't count the 8 byte header with | |
838 | * RAM_SAVE_FLAG_CONTINUE. | |
839 | */ | |
840 | xbzrle_counters.bytes += bytes_xbzrle - 8; | |
4c2d0f6d | 841 | ram_transferred_add(bytes_xbzrle); |
56e93d26 JQ |
842 | |
843 | return 1; | |
844 | } | |
845 | ||
3d0684b2 | 846 | /** |
d9e474ea | 847 | * pss_find_next_dirty: find the next dirty page of current ramblock |
f3f491fc | 848 | * |
d9e474ea PX |
849 | * This function updates pss->page to point to the next dirty page index |
850 | * within the ramblock to migrate, or the end of ramblock when nothing | |
851 | * found. Note that when pss->host_page_sending==true it means we're | |
852 | * during sending a host page, so we won't look for dirty page that is | |
853 | * outside the host page boundary. | |
3d0684b2 | 854 | * |
d9e474ea | 855 | * @pss: the current page search status |
f3f491fc | 856 | */ |
d9e474ea | 857 | static void pss_find_next_dirty(PageSearchStatus *pss) |
56e93d26 | 858 | { |
d9e474ea | 859 | RAMBlock *rb = pss->block; |
6b6712ef JQ |
860 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
861 | unsigned long *bitmap = rb->bmap; | |
56e93d26 | 862 | |
fbd162e6 | 863 | if (ramblock_is_ignored(rb)) { |
d9e474ea PX |
864 | /* Points directly to the end, so we know no dirty page */ |
865 | pss->page = size; | |
866 | return; | |
867 | } | |
868 | ||
869 | /* | |
870 | * If during sending a host page, only look for dirty pages within the | |
871 | * current host page being send. | |
872 | */ | |
873 | if (pss->host_page_sending) { | |
874 | assert(pss->host_page_end); | |
875 | size = MIN(size, pss->host_page_end); | |
b895de50 CLG |
876 | } |
877 | ||
d9e474ea | 878 | pss->page = find_next_bit(bitmap, size, pss->page); |
56e93d26 JQ |
879 | } |
880 | ||
1230a25f | 881 | static void migration_clear_memory_region_dirty_bitmap(RAMBlock *rb, |
3143577d WW |
882 | unsigned long page) |
883 | { | |
884 | uint8_t shift; | |
885 | hwaddr size, start; | |
886 | ||
887 | if (!rb->clear_bmap || !clear_bmap_test_and_clear(rb, page)) { | |
888 | return; | |
889 | } | |
890 | ||
891 | shift = rb->clear_bmap_shift; | |
892 | /* | |
893 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
894 | * can make things easier sometimes since then start address | |
895 | * of the small chunk will always be 64 pages aligned so the | |
896 | * bitmap will always be aligned to unsigned long. We should | |
897 | * even be able to remove this restriction but I'm simply | |
898 | * keeping it. | |
899 | */ | |
900 | assert(shift >= 6); | |
901 | ||
902 | size = 1ULL << (TARGET_PAGE_BITS + shift); | |
7648297d | 903 | start = QEMU_ALIGN_DOWN((ram_addr_t)page << TARGET_PAGE_BITS, size); |
3143577d WW |
904 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); |
905 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
906 | } | |
907 | ||
908 | static void | |
1230a25f | 909 | migration_clear_memory_region_dirty_bitmap_range(RAMBlock *rb, |
3143577d WW |
910 | unsigned long start, |
911 | unsigned long npages) | |
912 | { | |
913 | unsigned long i, chunk_pages = 1UL << rb->clear_bmap_shift; | |
914 | unsigned long chunk_start = QEMU_ALIGN_DOWN(start, chunk_pages); | |
915 | unsigned long chunk_end = QEMU_ALIGN_UP(start + npages, chunk_pages); | |
916 | ||
917 | /* | |
918 | * Clear pages from start to start + npages - 1, so the end boundary is | |
919 | * exclusive. | |
920 | */ | |
921 | for (i = chunk_start; i < chunk_end; i += chunk_pages) { | |
1230a25f | 922 | migration_clear_memory_region_dirty_bitmap(rb, i); |
3143577d WW |
923 | } |
924 | } | |
925 | ||
a6a83cef RL |
926 | /* |
927 | * colo_bitmap_find_diry:find contiguous dirty pages from start | |
928 | * | |
929 | * Returns the page offset within memory region of the start of the contiguout | |
930 | * dirty page | |
931 | * | |
932 | * @rs: current RAM state | |
933 | * @rb: RAMBlock where to search for dirty pages | |
934 | * @start: page where we start the search | |
935 | * @num: the number of contiguous dirty pages | |
936 | */ | |
937 | static inline | |
938 | unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, | |
939 | unsigned long start, unsigned long *num) | |
940 | { | |
941 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; | |
942 | unsigned long *bitmap = rb->bmap; | |
943 | unsigned long first, next; | |
944 | ||
945 | *num = 0; | |
946 | ||
947 | if (ramblock_is_ignored(rb)) { | |
948 | return size; | |
949 | } | |
950 | ||
951 | first = find_next_bit(bitmap, size, start); | |
952 | if (first >= size) { | |
953 | return first; | |
954 | } | |
955 | next = find_next_zero_bit(bitmap, size, first + 1); | |
956 | assert(next >= first); | |
957 | *num = next - first; | |
958 | return first; | |
959 | } | |
960 | ||
06b10688 | 961 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
962 | RAMBlock *rb, |
963 | unsigned long page) | |
a82d593b DDAG |
964 | { |
965 | bool ret; | |
a82d593b | 966 | |
002cad6b PX |
967 | /* |
968 | * Clear dirty bitmap if needed. This _must_ be called before we | |
969 | * send any of the page in the chunk because we need to make sure | |
970 | * we can capture further page content changes when we sync dirty | |
971 | * log the next time. So as long as we are going to send any of | |
972 | * the page in the chunk we clear the remote dirty bitmap for all. | |
973 | * Clearing it earlier won't be a problem, but too late will. | |
974 | */ | |
1230a25f | 975 | migration_clear_memory_region_dirty_bitmap(rb, page); |
002cad6b | 976 | |
6b6712ef | 977 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b | 978 | if (ret) { |
0d8ec885 | 979 | rs->migration_dirty_pages--; |
a82d593b | 980 | } |
386a907b | 981 | |
a82d593b DDAG |
982 | return ret; |
983 | } | |
984 | ||
be39b4cd DH |
985 | static void dirty_bitmap_clear_section(MemoryRegionSection *section, |
986 | void *opaque) | |
987 | { | |
988 | const hwaddr offset = section->offset_within_region; | |
989 | const hwaddr size = int128_get64(section->size); | |
990 | const unsigned long start = offset >> TARGET_PAGE_BITS; | |
991 | const unsigned long npages = size >> TARGET_PAGE_BITS; | |
992 | RAMBlock *rb = section->mr->ram_block; | |
993 | uint64_t *cleared_bits = opaque; | |
994 | ||
995 | /* | |
996 | * We don't grab ram_state->bitmap_mutex because we expect to run | |
997 | * only when starting migration or during postcopy recovery where | |
998 | * we don't have concurrent access. | |
999 | */ | |
1000 | if (!migration_in_postcopy() && !migrate_background_snapshot()) { | |
1001 | migration_clear_memory_region_dirty_bitmap_range(rb, start, npages); | |
1002 | } | |
1003 | *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages); | |
1004 | bitmap_clear(rb->bmap, start, npages); | |
1005 | } | |
1006 | ||
1007 | /* | |
1008 | * Exclude all dirty pages from migration that fall into a discarded range as | |
1009 | * managed by a RamDiscardManager responsible for the mapped memory region of | |
1010 | * the RAMBlock. Clear the corresponding bits in the dirty bitmaps. | |
1011 | * | |
1012 | * Discarded pages ("logically unplugged") have undefined content and must | |
1013 | * not get migrated, because even reading these pages for migration might | |
1014 | * result in undesired behavior. | |
1015 | * | |
1016 | * Returns the number of cleared bits in the RAMBlock dirty bitmap. | |
1017 | * | |
1018 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1019 | */ | |
1020 | static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb) | |
1021 | { | |
1022 | uint64_t cleared_bits = 0; | |
1023 | ||
1024 | if (rb->mr && rb->bmap && memory_region_has_ram_discard_manager(rb->mr)) { | |
1025 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1026 | MemoryRegionSection section = { | |
1027 | .mr = rb->mr, | |
1028 | .offset_within_region = 0, | |
1029 | .size = int128_make64(qemu_ram_get_used_length(rb)), | |
1030 | }; | |
1031 | ||
1032 | ram_discard_manager_replay_discarded(rdm, §ion, | |
1033 | dirty_bitmap_clear_section, | |
1034 | &cleared_bits); | |
1035 | } | |
1036 | return cleared_bits; | |
1037 | } | |
1038 | ||
9470c5e0 DH |
1039 | /* |
1040 | * Check if a host-page aligned page falls into a discarded range as managed by | |
1041 | * a RamDiscardManager responsible for the mapped memory region of the RAMBlock. | |
1042 | * | |
1043 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1044 | */ | |
1045 | bool ramblock_page_is_discarded(RAMBlock *rb, ram_addr_t start) | |
1046 | { | |
1047 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1048 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1049 | MemoryRegionSection section = { | |
1050 | .mr = rb->mr, | |
1051 | .offset_within_region = start, | |
1052 | .size = int128_make64(qemu_ram_pagesize(rb)), | |
1053 | }; | |
1054 | ||
1055 | return !ram_discard_manager_is_populated(rdm, §ion); | |
1056 | } | |
1057 | return false; | |
1058 | } | |
1059 | ||
267691b6 | 1060 | /* Called with RCU critical section */ |
7a3e9571 | 1061 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 1062 | { |
fb613580 KZ |
1063 | uint64_t new_dirty_pages = |
1064 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length); | |
1065 | ||
1066 | rs->migration_dirty_pages += new_dirty_pages; | |
1067 | rs->num_dirty_pages_period += new_dirty_pages; | |
56e93d26 JQ |
1068 | } |
1069 | ||
3d0684b2 JQ |
1070 | /** |
1071 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
1072 | * | |
1073 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
1074 | * | |
1075 | * For VMs with just normal pages this is equivalent to the host page | |
1076 | * size. If it's got some huge pages then it's the OR of all the | |
1077 | * different page sizes. | |
e8ca1db2 DDAG |
1078 | */ |
1079 | uint64_t ram_pagesize_summary(void) | |
1080 | { | |
1081 | RAMBlock *block; | |
1082 | uint64_t summary = 0; | |
1083 | ||
fbd162e6 | 1084 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
1085 | summary |= block->page_size; |
1086 | } | |
1087 | ||
1088 | return summary; | |
1089 | } | |
1090 | ||
aecbfe9c XG |
1091 | uint64_t ram_get_total_transferred_pages(void) |
1092 | { | |
23b7576d PX |
1093 | return stat64_get(&ram_atomic_counters.normal) + |
1094 | stat64_get(&ram_atomic_counters.duplicate) + | |
1095 | compression_counters.pages + xbzrle_counters.pages; | |
aecbfe9c XG |
1096 | } |
1097 | ||
b734035b XG |
1098 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
1099 | { | |
be8b02ed | 1100 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 1101 | double compressed_size; |
b734035b XG |
1102 | |
1103 | /* calculate period counters */ | |
1104 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 | |
1105 | / (end_time - rs->time_last_bitmap_sync); | |
1106 | ||
be8b02ed | 1107 | if (!page_count) { |
b734035b XG |
1108 | return; |
1109 | } | |
1110 | ||
1111 | if (migrate_use_xbzrle()) { | |
e460a4b1 WW |
1112 | double encoded_size, unencoded_size; |
1113 | ||
b734035b | 1114 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - |
be8b02ed | 1115 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b | 1116 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
e460a4b1 WW |
1117 | unencoded_size = (xbzrle_counters.pages - rs->xbzrle_pages_prev) * |
1118 | TARGET_PAGE_SIZE; | |
1119 | encoded_size = xbzrle_counters.bytes - rs->xbzrle_bytes_prev; | |
92271402 | 1120 | if (xbzrle_counters.pages == rs->xbzrle_pages_prev || !encoded_size) { |
e460a4b1 | 1121 | xbzrle_counters.encoding_rate = 0; |
e460a4b1 WW |
1122 | } else { |
1123 | xbzrle_counters.encoding_rate = unencoded_size / encoded_size; | |
1124 | } | |
1125 | rs->xbzrle_pages_prev = xbzrle_counters.pages; | |
1126 | rs->xbzrle_bytes_prev = xbzrle_counters.bytes; | |
b734035b | 1127 | } |
76e03000 XG |
1128 | |
1129 | if (migrate_use_compression()) { | |
1130 | compression_counters.busy_rate = (double)(compression_counters.busy - | |
1131 | rs->compress_thread_busy_prev) / page_count; | |
1132 | rs->compress_thread_busy_prev = compression_counters.busy; | |
1133 | ||
1134 | compressed_size = compression_counters.compressed_size - | |
1135 | rs->compressed_size_prev; | |
1136 | if (compressed_size) { | |
1137 | double uncompressed_size = (compression_counters.pages - | |
1138 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
1139 | ||
1140 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
1141 | compression_counters.compression_rate = | |
1142 | uncompressed_size / compressed_size; | |
1143 | ||
1144 | rs->compress_pages_prev = compression_counters.pages; | |
1145 | rs->compressed_size_prev = compression_counters.compressed_size; | |
1146 | } | |
1147 | } | |
b734035b XG |
1148 | } |
1149 | ||
dc14a470 KZ |
1150 | static void migration_trigger_throttle(RAMState *rs) |
1151 | { | |
1152 | MigrationState *s = migrate_get_current(); | |
1153 | uint64_t threshold = s->parameters.throttle_trigger_threshold; | |
23b7576d PX |
1154 | uint64_t bytes_xfer_period = |
1155 | stat64_get(&ram_atomic_counters.transferred) - rs->bytes_xfer_prev; | |
dc14a470 KZ |
1156 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; |
1157 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
1158 | ||
1159 | /* During block migration the auto-converge logic incorrectly detects | |
1160 | * that ram migration makes no progress. Avoid this by disabling the | |
1161 | * throttling logic during the bulk phase of block migration. */ | |
1162 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
1163 | /* The following detection logic can be refined later. For now: | |
1164 | Check to see if the ratio between dirtied bytes and the approx. | |
1165 | amount of bytes that just got transferred since the last time | |
1166 | we were in this routine reaches the threshold. If that happens | |
1167 | twice, start or increase throttling. */ | |
1168 | ||
1169 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
1170 | (++rs->dirty_rate_high_cnt >= 2)) { | |
1171 | trace_migration_throttle(); | |
1172 | rs->dirty_rate_high_cnt = 0; | |
cbbf8182 KZ |
1173 | mig_throttle_guest_down(bytes_dirty_period, |
1174 | bytes_dirty_threshold); | |
dc14a470 KZ |
1175 | } |
1176 | } | |
1177 | } | |
1178 | ||
8d820d6f | 1179 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
1180 | { |
1181 | RAMBlock *block; | |
56e93d26 | 1182 | int64_t end_time; |
56e93d26 | 1183 | |
9360447d | 1184 | ram_counters.dirty_sync_count++; |
56e93d26 | 1185 | |
f664da80 JQ |
1186 | if (!rs->time_last_bitmap_sync) { |
1187 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
1188 | } |
1189 | ||
1190 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 1191 | memory_global_dirty_log_sync(); |
56e93d26 | 1192 | |
108cfae0 | 1193 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
1194 | WITH_RCU_READ_LOCK_GUARD() { |
1195 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1196 | ramblock_sync_dirty_bitmap(rs, block); | |
1197 | } | |
1198 | ram_counters.remaining = ram_bytes_remaining(); | |
56e93d26 | 1199 | } |
108cfae0 | 1200 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 1201 | |
9458a9a1 | 1202 | memory_global_after_dirty_log_sync(); |
a66cd90c | 1203 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 1204 | |
56e93d26 JQ |
1205 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
1206 | ||
1207 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 1208 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 1209 | migration_trigger_throttle(rs); |
070afca2 | 1210 | |
b734035b XG |
1211 | migration_update_rates(rs, end_time); |
1212 | ||
be8b02ed | 1213 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
1214 | |
1215 | /* reset period counters */ | |
f664da80 | 1216 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 1217 | rs->num_dirty_pages_period = 0; |
23b7576d | 1218 | rs->bytes_xfer_prev = stat64_get(&ram_atomic_counters.transferred); |
56e93d26 | 1219 | } |
4addcd4f | 1220 | if (migrate_use_events()) { |
3ab72385 | 1221 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count); |
4addcd4f | 1222 | } |
56e93d26 JQ |
1223 | } |
1224 | ||
bd227060 WW |
1225 | static void migration_bitmap_sync_precopy(RAMState *rs) |
1226 | { | |
1227 | Error *local_err = NULL; | |
1228 | ||
1229 | /* | |
1230 | * The current notifier usage is just an optimization to migration, so we | |
1231 | * don't stop the normal migration process in the error case. | |
1232 | */ | |
1233 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
1234 | error_report_err(local_err); | |
b4a1733c | 1235 | local_err = NULL; |
bd227060 WW |
1236 | } |
1237 | ||
1238 | migration_bitmap_sync(rs); | |
1239 | ||
1240 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
1241 | error_report_err(local_err); | |
1242 | } | |
1243 | } | |
1244 | ||
a4dbaf8e | 1245 | void ram_release_page(const char *rbname, uint64_t offset) |
47fe16ff JQ |
1246 | { |
1247 | if (!migrate_release_ram() || !migration_in_postcopy()) { | |
1248 | return; | |
1249 | } | |
1250 | ||
1251 | ram_discard_range(rbname, offset, TARGET_PAGE_SIZE); | |
1252 | } | |
1253 | ||
6c97ec5f XG |
1254 | /** |
1255 | * save_zero_page_to_file: send the zero page to the file | |
1256 | * | |
1257 | * Returns the size of data written to the file, 0 means the page is not | |
1258 | * a zero page | |
1259 | * | |
ec6f3ab9 | 1260 | * @pss: current PSS channel |
6c97ec5f XG |
1261 | * @block: block that contains the page we want to send |
1262 | * @offset: offset inside the block for the page | |
1263 | */ | |
ec6f3ab9 | 1264 | static int save_zero_page_to_file(PageSearchStatus *pss, |
6c97ec5f XG |
1265 | RAMBlock *block, ram_addr_t offset) |
1266 | { | |
1267 | uint8_t *p = block->host + offset; | |
ec6f3ab9 | 1268 | QEMUFile *file = pss->pss_channel; |
6c97ec5f XG |
1269 | int len = 0; |
1270 | ||
bad452a7 | 1271 | if (buffer_is_zero(p, TARGET_PAGE_SIZE)) { |
ec6f3ab9 | 1272 | len += save_page_header(pss, block, offset | RAM_SAVE_FLAG_ZERO); |
6c97ec5f XG |
1273 | qemu_put_byte(file, 0); |
1274 | len += 1; | |
47fe16ff | 1275 | ram_release_page(block->idstr, offset); |
6c97ec5f XG |
1276 | } |
1277 | return len; | |
1278 | } | |
1279 | ||
56e93d26 | 1280 | /** |
3d0684b2 | 1281 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1282 | * |
3d0684b2 | 1283 | * Returns the number of pages written. |
56e93d26 | 1284 | * |
ec6f3ab9 | 1285 | * @pss: current PSS channel |
56e93d26 JQ |
1286 | * @block: block that contains the page we want to send |
1287 | * @offset: offset inside the block for the page | |
56e93d26 | 1288 | */ |
ec6f3ab9 | 1289 | static int save_zero_page(PageSearchStatus *pss, RAMBlock *block, |
61717ea9 | 1290 | ram_addr_t offset) |
56e93d26 | 1291 | { |
ec6f3ab9 | 1292 | int len = save_zero_page_to_file(pss, block, offset); |
56e93d26 | 1293 | |
6c97ec5f | 1294 | if (len) { |
23b7576d | 1295 | stat64_add(&ram_atomic_counters.duplicate, 1); |
4c2d0f6d | 1296 | ram_transferred_add(len); |
6c97ec5f | 1297 | return 1; |
56e93d26 | 1298 | } |
6c97ec5f | 1299 | return -1; |
56e93d26 JQ |
1300 | } |
1301 | ||
059ff0fb XG |
1302 | /* |
1303 | * @pages: the number of pages written by the control path, | |
1304 | * < 0 - error | |
1305 | * > 0 - number of pages written | |
1306 | * | |
1307 | * Return true if the pages has been saved, otherwise false is returned. | |
1308 | */ | |
61717ea9 PX |
1309 | static bool control_save_page(PageSearchStatus *pss, RAMBlock *block, |
1310 | ram_addr_t offset, int *pages) | |
059ff0fb XG |
1311 | { |
1312 | uint64_t bytes_xmit = 0; | |
1313 | int ret; | |
1314 | ||
1315 | *pages = -1; | |
61717ea9 PX |
1316 | ret = ram_control_save_page(pss->pss_channel, block->offset, offset, |
1317 | TARGET_PAGE_SIZE, &bytes_xmit); | |
059ff0fb XG |
1318 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { |
1319 | return false; | |
1320 | } | |
1321 | ||
1322 | if (bytes_xmit) { | |
4c2d0f6d | 1323 | ram_transferred_add(bytes_xmit); |
059ff0fb XG |
1324 | *pages = 1; |
1325 | } | |
1326 | ||
1327 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1328 | return true; | |
1329 | } | |
1330 | ||
1331 | if (bytes_xmit > 0) { | |
23b7576d | 1332 | stat64_add(&ram_atomic_counters.normal, 1); |
059ff0fb | 1333 | } else if (bytes_xmit == 0) { |
23b7576d | 1334 | stat64_add(&ram_atomic_counters.duplicate, 1); |
059ff0fb XG |
1335 | } |
1336 | ||
1337 | return true; | |
1338 | } | |
1339 | ||
65dacaa0 XG |
1340 | /* |
1341 | * directly send the page to the stream | |
1342 | * | |
1343 | * Returns the number of pages written. | |
1344 | * | |
ec6f3ab9 | 1345 | * @pss: current PSS channel |
65dacaa0 XG |
1346 | * @block: block that contains the page we want to send |
1347 | * @offset: offset inside the block for the page | |
1348 | * @buf: the page to be sent | |
1349 | * @async: send to page asyncly | |
1350 | */ | |
ec6f3ab9 | 1351 | static int save_normal_page(PageSearchStatus *pss, RAMBlock *block, |
61717ea9 | 1352 | ram_addr_t offset, uint8_t *buf, bool async) |
65dacaa0 | 1353 | { |
ec6f3ab9 PX |
1354 | QEMUFile *file = pss->pss_channel; |
1355 | ||
1356 | ram_transferred_add(save_page_header(pss, block, | |
4c2d0f6d | 1357 | offset | RAM_SAVE_FLAG_PAGE)); |
65dacaa0 | 1358 | if (async) { |
61717ea9 | 1359 | qemu_put_buffer_async(file, buf, TARGET_PAGE_SIZE, |
f912ec5b | 1360 | migrate_release_ram() && |
65dacaa0 XG |
1361 | migration_in_postcopy()); |
1362 | } else { | |
61717ea9 | 1363 | qemu_put_buffer(file, buf, TARGET_PAGE_SIZE); |
65dacaa0 | 1364 | } |
4c2d0f6d | 1365 | ram_transferred_add(TARGET_PAGE_SIZE); |
23b7576d | 1366 | stat64_add(&ram_atomic_counters.normal, 1); |
65dacaa0 XG |
1367 | return 1; |
1368 | } | |
1369 | ||
56e93d26 | 1370 | /** |
3d0684b2 | 1371 | * ram_save_page: send the given page to the stream |
56e93d26 | 1372 | * |
3d0684b2 | 1373 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1374 | * < 0 - error |
1375 | * >=0 - Number of pages written - this might legally be 0 | |
1376 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1377 | * |
6f37bb8b | 1378 | * @rs: current RAM state |
56e93d26 JQ |
1379 | * @block: block that contains the page we want to send |
1380 | * @offset: offset inside the block for the page | |
56e93d26 | 1381 | */ |
05931ec5 | 1382 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss) |
56e93d26 JQ |
1383 | { |
1384 | int pages = -1; | |
56e93d26 | 1385 | uint8_t *p; |
56e93d26 | 1386 | bool send_async = true; |
a08f6890 | 1387 | RAMBlock *block = pss->block; |
8bba004c | 1388 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1389 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1390 | |
2f68e399 | 1391 | p = block->host + offset; |
1db9d8e5 | 1392 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1393 | |
56e93d26 | 1394 | XBZRLE_cache_lock(); |
1a373522 | 1395 | if (rs->xbzrle_enabled && !migration_in_postcopy()) { |
ec6f3ab9 | 1396 | pages = save_xbzrle_page(rs, pss, &p, current_addr, |
61717ea9 | 1397 | block, offset); |
05931ec5 | 1398 | if (!rs->last_stage) { |
059ff0fb XG |
1399 | /* Can't send this cached data async, since the cache page |
1400 | * might get updated before it gets to the wire | |
56e93d26 | 1401 | */ |
059ff0fb | 1402 | send_async = false; |
56e93d26 JQ |
1403 | } |
1404 | } | |
1405 | ||
1406 | /* XBZRLE overflow or normal page */ | |
1407 | if (pages == -1) { | |
ec6f3ab9 | 1408 | pages = save_normal_page(pss, block, offset, p, send_async); |
56e93d26 JQ |
1409 | } |
1410 | ||
1411 | XBZRLE_cache_unlock(); | |
1412 | ||
1413 | return pages; | |
1414 | } | |
1415 | ||
61717ea9 | 1416 | static int ram_save_multifd_page(QEMUFile *file, RAMBlock *block, |
b9ee2f7d JQ |
1417 | ram_addr_t offset) |
1418 | { | |
61717ea9 | 1419 | if (multifd_queue_page(file, block, offset) < 0) { |
713f762a IR |
1420 | return -1; |
1421 | } | |
23b7576d | 1422 | stat64_add(&ram_atomic_counters.normal, 1); |
b9ee2f7d JQ |
1423 | |
1424 | return 1; | |
1425 | } | |
1426 | ||
5e5fdcff | 1427 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 1428 | ram_addr_t offset, uint8_t *source_buf) |
56e93d26 | 1429 | { |
53518d94 | 1430 | RAMState *rs = ram_state; |
ec6f3ab9 | 1431 | PageSearchStatus *pss = &rs->pss[RAM_CHANNEL_PRECOPY]; |
20d549cb | 1432 | uint8_t *p = block->host + offset; |
6ef3771c | 1433 | int ret; |
56e93d26 | 1434 | |
ec6f3ab9 | 1435 | if (save_zero_page_to_file(pss, block, offset)) { |
e7f2e190 | 1436 | return true; |
5e5fdcff XG |
1437 | } |
1438 | ||
ec6f3ab9 | 1439 | save_page_header(pss, block, offset | RAM_SAVE_FLAG_COMPRESS_PAGE); |
34ab9e97 XG |
1440 | |
1441 | /* | |
1442 | * copy it to a internal buffer to avoid it being modified by VM | |
1443 | * so that we can catch up the error during compression and | |
1444 | * decompression | |
1445 | */ | |
1446 | memcpy(source_buf, p, TARGET_PAGE_SIZE); | |
6ef3771c XG |
1447 | ret = qemu_put_compression_data(f, stream, source_buf, TARGET_PAGE_SIZE); |
1448 | if (ret < 0) { | |
1449 | qemu_file_set_error(migrate_get_current()->to_dst_file, ret); | |
b3be2896 | 1450 | error_report("compressed data failed!"); |
b3be2896 | 1451 | } |
e7f2e190 | 1452 | return false; |
5e5fdcff XG |
1453 | } |
1454 | ||
1455 | static void | |
1456 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1457 | { | |
4c2d0f6d | 1458 | ram_transferred_add(bytes_xmit); |
76e03000 | 1459 | |
5e5fdcff | 1460 | if (param->zero_page) { |
23b7576d | 1461 | stat64_add(&ram_atomic_counters.duplicate, 1); |
76e03000 | 1462 | return; |
5e5fdcff | 1463 | } |
76e03000 XG |
1464 | |
1465 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1466 | compression_counters.compressed_size += bytes_xmit - 8; | |
1467 | compression_counters.pages++; | |
56e93d26 JQ |
1468 | } |
1469 | ||
32b05495 XG |
1470 | static bool save_page_use_compression(RAMState *rs); |
1471 | ||
ce25d337 | 1472 | static void flush_compressed_data(RAMState *rs) |
56e93d26 | 1473 | { |
eaa238ab | 1474 | MigrationState *ms = migrate_get_current(); |
56e93d26 JQ |
1475 | int idx, len, thread_count; |
1476 | ||
32b05495 | 1477 | if (!save_page_use_compression(rs)) { |
56e93d26 JQ |
1478 | return; |
1479 | } | |
1480 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1481 | |
0d9f9a5c | 1482 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1483 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1484 | while (!comp_param[idx].done) { |
0d9f9a5c | 1485 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1486 | } |
a7a9a88f | 1487 | } |
0d9f9a5c | 1488 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1489 | |
1490 | for (idx = 0; idx < thread_count; idx++) { | |
1491 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1492 | if (!comp_param[idx].quit) { |
eaa238ab | 1493 | len = qemu_put_qemu_file(ms->to_dst_file, comp_param[idx].file); |
5e5fdcff XG |
1494 | /* |
1495 | * it's safe to fetch zero_page without holding comp_done_lock | |
1496 | * as there is no further request submitted to the thread, | |
1497 | * i.e, the thread should be waiting for a request at this point. | |
1498 | */ | |
1499 | update_compress_thread_counts(&comp_param[idx], len); | |
56e93d26 | 1500 | } |
a7a9a88f | 1501 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1502 | } |
1503 | } | |
1504 | ||
1505 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1506 | ram_addr_t offset) | |
1507 | { | |
1508 | param->block = block; | |
1509 | param->offset = offset; | |
1510 | } | |
1511 | ||
eaa238ab | 1512 | static int compress_page_with_multi_thread(RAMBlock *block, ram_addr_t offset) |
56e93d26 JQ |
1513 | { |
1514 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1d58872a | 1515 | bool wait = migrate_compress_wait_thread(); |
eaa238ab | 1516 | MigrationState *ms = migrate_get_current(); |
56e93d26 JQ |
1517 | |
1518 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1519 | qemu_mutex_lock(&comp_done_lock); |
1d58872a XG |
1520 | retry: |
1521 | for (idx = 0; idx < thread_count; idx++) { | |
1522 | if (comp_param[idx].done) { | |
1523 | comp_param[idx].done = false; | |
eaa238ab PX |
1524 | bytes_xmit = qemu_put_qemu_file(ms->to_dst_file, |
1525 | comp_param[idx].file); | |
1d58872a XG |
1526 | qemu_mutex_lock(&comp_param[idx].mutex); |
1527 | set_compress_params(&comp_param[idx], block, offset); | |
1528 | qemu_cond_signal(&comp_param[idx].cond); | |
1529 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
1530 | pages = 1; | |
5e5fdcff | 1531 | update_compress_thread_counts(&comp_param[idx], bytes_xmit); |
56e93d26 | 1532 | break; |
56e93d26 JQ |
1533 | } |
1534 | } | |
1d58872a XG |
1535 | |
1536 | /* | |
1537 | * wait for the free thread if the user specifies 'compress-wait-thread', | |
1538 | * otherwise we will post the page out in the main thread as normal page. | |
1539 | */ | |
1540 | if (pages < 0 && wait) { | |
1541 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); | |
1542 | goto retry; | |
1543 | } | |
0d9f9a5c | 1544 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1545 | |
1546 | return pages; | |
1547 | } | |
1548 | ||
3d0684b2 JQ |
1549 | /** |
1550 | * find_dirty_block: find the next dirty page and update any state | |
1551 | * associated with the search process. | |
b9e60928 | 1552 | * |
a5f7b1a6 | 1553 | * Returns true if a page is found |
b9e60928 | 1554 | * |
6f37bb8b | 1555 | * @rs: current RAM state |
3d0684b2 JQ |
1556 | * @pss: data about the state of the current dirty page scan |
1557 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1558 | */ |
f20e2865 | 1559 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1560 | { |
d9e474ea PX |
1561 | /* Update pss->page for the next dirty bit in ramblock */ |
1562 | pss_find_next_dirty(pss); | |
1563 | ||
6f37bb8b | 1564 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1565 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1566 | /* |
1567 | * We've been once around the RAM and haven't found anything. | |
1568 | * Give up. | |
1569 | */ | |
1570 | *again = false; | |
1571 | return false; | |
1572 | } | |
542147f4 DH |
1573 | if (!offset_in_ramblock(pss->block, |
1574 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)) { | |
b9e60928 | 1575 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1576 | pss->page = 0; |
b9e60928 DDAG |
1577 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1578 | if (!pss->block) { | |
48df9d80 XG |
1579 | /* |
1580 | * If memory migration starts over, we will meet a dirtied page | |
1581 | * which may still exists in compression threads's ring, so we | |
1582 | * should flush the compressed data to make sure the new page | |
1583 | * is not overwritten by the old one in the destination. | |
1584 | * | |
1585 | * Also If xbzrle is on, stop using the data compression at this | |
1586 | * point. In theory, xbzrle can do better than compression. | |
1587 | */ | |
1588 | flush_compressed_data(rs); | |
1589 | ||
b9e60928 DDAG |
1590 | /* Hit the end of the list */ |
1591 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1592 | /* Flag that we've looped */ | |
1593 | pss->complete_round = true; | |
1a373522 DH |
1594 | /* After the first round, enable XBZRLE. */ |
1595 | if (migrate_use_xbzrle()) { | |
1596 | rs->xbzrle_enabled = true; | |
1597 | } | |
b9e60928 DDAG |
1598 | } |
1599 | /* Didn't find anything this time, but try again on the new block */ | |
1600 | *again = true; | |
1601 | return false; | |
1602 | } else { | |
1603 | /* Can go around again, but... */ | |
1604 | *again = true; | |
1605 | /* We've found something so probably don't need to */ | |
1606 | return true; | |
1607 | } | |
1608 | } | |
1609 | ||
3d0684b2 JQ |
1610 | /** |
1611 | * unqueue_page: gets a page of the queue | |
1612 | * | |
a82d593b | 1613 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1614 | * |
3d0684b2 JQ |
1615 | * Returns the block of the page (or NULL if none available) |
1616 | * | |
ec481c6c | 1617 | * @rs: current RAM state |
3d0684b2 | 1618 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1619 | */ |
f20e2865 | 1620 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b | 1621 | { |
a1fe28df | 1622 | struct RAMSrcPageRequest *entry; |
a82d593b DDAG |
1623 | RAMBlock *block = NULL; |
1624 | ||
a1fe28df | 1625 | if (!postcopy_has_request(rs)) { |
ae526e32 XG |
1626 | return NULL; |
1627 | } | |
1628 | ||
6e8a355d | 1629 | QEMU_LOCK_GUARD(&rs->src_page_req_mutex); |
a1fe28df PX |
1630 | |
1631 | /* | |
1632 | * This should _never_ change even after we take the lock, because no one | |
1633 | * should be taking anything off the request list other than us. | |
1634 | */ | |
1635 | assert(postcopy_has_request(rs)); | |
1636 | ||
1637 | entry = QSIMPLEQ_FIRST(&rs->src_page_requests); | |
1638 | block = entry->rb; | |
1639 | *offset = entry->offset; | |
1640 | ||
777f53c7 TH |
1641 | if (entry->len > TARGET_PAGE_SIZE) { |
1642 | entry->len -= TARGET_PAGE_SIZE; | |
1643 | entry->offset += TARGET_PAGE_SIZE; | |
a1fe28df PX |
1644 | } else { |
1645 | memory_region_unref(block->mr); | |
1646 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); | |
1647 | g_free(entry); | |
1648 | migration_consume_urgent_request(); | |
a82d593b | 1649 | } |
a82d593b DDAG |
1650 | |
1651 | return block; | |
1652 | } | |
1653 | ||
278e2f55 AG |
1654 | #if defined(__linux__) |
1655 | /** | |
1656 | * poll_fault_page: try to get next UFFD write fault page and, if pending fault | |
1657 | * is found, return RAM block pointer and page offset | |
1658 | * | |
1659 | * Returns pointer to the RAMBlock containing faulting page, | |
1660 | * NULL if no write faults are pending | |
1661 | * | |
1662 | * @rs: current RAM state | |
1663 | * @offset: page offset from the beginning of the block | |
1664 | */ | |
1665 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1666 | { | |
1667 | struct uffd_msg uffd_msg; | |
1668 | void *page_address; | |
82ea3e3b | 1669 | RAMBlock *block; |
278e2f55 AG |
1670 | int res; |
1671 | ||
1672 | if (!migrate_background_snapshot()) { | |
1673 | return NULL; | |
1674 | } | |
1675 | ||
1676 | res = uffd_read_events(rs->uffdio_fd, &uffd_msg, 1); | |
1677 | if (res <= 0) { | |
1678 | return NULL; | |
1679 | } | |
1680 | ||
1681 | page_address = (void *)(uintptr_t) uffd_msg.arg.pagefault.address; | |
82ea3e3b AG |
1682 | block = qemu_ram_block_from_host(page_address, false, offset); |
1683 | assert(block && (block->flags & RAM_UF_WRITEPROTECT) != 0); | |
1684 | return block; | |
278e2f55 AG |
1685 | } |
1686 | ||
1687 | /** | |
1688 | * ram_save_release_protection: release UFFD write protection after | |
1689 | * a range of pages has been saved | |
1690 | * | |
1691 | * @rs: current RAM state | |
1692 | * @pss: page-search-status structure | |
1693 | * @start_page: index of the first page in the range relative to pss->block | |
1694 | * | |
1695 | * Returns 0 on success, negative value in case of an error | |
1696 | */ | |
1697 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1698 | unsigned long start_page) | |
1699 | { | |
1700 | int res = 0; | |
1701 | ||
1702 | /* Check if page is from UFFD-managed region. */ | |
1703 | if (pss->block->flags & RAM_UF_WRITEPROTECT) { | |
1704 | void *page_address = pss->block->host + (start_page << TARGET_PAGE_BITS); | |
258f5c98 | 1705 | uint64_t run_length = (pss->page - start_page) << TARGET_PAGE_BITS; |
278e2f55 AG |
1706 | |
1707 | /* Flush async buffers before un-protect. */ | |
61717ea9 | 1708 | qemu_fflush(pss->pss_channel); |
278e2f55 AG |
1709 | /* Un-protect memory range. */ |
1710 | res = uffd_change_protection(rs->uffdio_fd, page_address, run_length, | |
1711 | false, false); | |
1712 | } | |
1713 | ||
1714 | return res; | |
1715 | } | |
1716 | ||
1717 | /* ram_write_tracking_available: check if kernel supports required UFFD features | |
1718 | * | |
1719 | * Returns true if supports, false otherwise | |
1720 | */ | |
1721 | bool ram_write_tracking_available(void) | |
1722 | { | |
1723 | uint64_t uffd_features; | |
1724 | int res; | |
1725 | ||
1726 | res = uffd_query_features(&uffd_features); | |
1727 | return (res == 0 && | |
1728 | (uffd_features & UFFD_FEATURE_PAGEFAULT_FLAG_WP) != 0); | |
1729 | } | |
1730 | ||
1731 | /* ram_write_tracking_compatible: check if guest configuration is | |
1732 | * compatible with 'write-tracking' | |
1733 | * | |
1734 | * Returns true if compatible, false otherwise | |
1735 | */ | |
1736 | bool ram_write_tracking_compatible(void) | |
1737 | { | |
1738 | const uint64_t uffd_ioctls_mask = BIT(_UFFDIO_WRITEPROTECT); | |
1739 | int uffd_fd; | |
82ea3e3b | 1740 | RAMBlock *block; |
278e2f55 AG |
1741 | bool ret = false; |
1742 | ||
1743 | /* Open UFFD file descriptor */ | |
1744 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, false); | |
1745 | if (uffd_fd < 0) { | |
1746 | return false; | |
1747 | } | |
1748 | ||
1749 | RCU_READ_LOCK_GUARD(); | |
1750 | ||
82ea3e3b | 1751 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 AG |
1752 | uint64_t uffd_ioctls; |
1753 | ||
1754 | /* Nothing to do with read-only and MMIO-writable regions */ | |
82ea3e3b | 1755 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1756 | continue; |
1757 | } | |
1758 | /* Try to register block memory via UFFD-IO to track writes */ | |
82ea3e3b | 1759 | if (uffd_register_memory(uffd_fd, block->host, block->max_length, |
278e2f55 AG |
1760 | UFFDIO_REGISTER_MODE_WP, &uffd_ioctls)) { |
1761 | goto out; | |
1762 | } | |
1763 | if ((uffd_ioctls & uffd_ioctls_mask) != uffd_ioctls_mask) { | |
1764 | goto out; | |
1765 | } | |
1766 | } | |
1767 | ret = true; | |
1768 | ||
1769 | out: | |
1770 | uffd_close_fd(uffd_fd); | |
1771 | return ret; | |
1772 | } | |
1773 | ||
f7b9dcfb DH |
1774 | static inline void populate_read_range(RAMBlock *block, ram_addr_t offset, |
1775 | ram_addr_t size) | |
1776 | { | |
5f19a449 DH |
1777 | const ram_addr_t end = offset + size; |
1778 | ||
f7b9dcfb DH |
1779 | /* |
1780 | * We read one byte of each page; this will preallocate page tables if | |
1781 | * required and populate the shared zeropage on MAP_PRIVATE anonymous memory | |
1782 | * where no page was populated yet. This might require adaption when | |
1783 | * supporting other mappings, like shmem. | |
1784 | */ | |
5f19a449 | 1785 | for (; offset < end; offset += block->page_size) { |
f7b9dcfb DH |
1786 | char tmp = *((char *)block->host + offset); |
1787 | ||
1788 | /* Don't optimize the read out */ | |
1789 | asm volatile("" : "+r" (tmp)); | |
1790 | } | |
1791 | } | |
1792 | ||
6fee3a1f DH |
1793 | static inline int populate_read_section(MemoryRegionSection *section, |
1794 | void *opaque) | |
1795 | { | |
1796 | const hwaddr size = int128_get64(section->size); | |
1797 | hwaddr offset = section->offset_within_region; | |
1798 | RAMBlock *block = section->mr->ram_block; | |
1799 | ||
1800 | populate_read_range(block, offset, size); | |
1801 | return 0; | |
1802 | } | |
1803 | ||
eeccb99c | 1804 | /* |
f7b9dcfb DH |
1805 | * ram_block_populate_read: preallocate page tables and populate pages in the |
1806 | * RAM block by reading a byte of each page. | |
eeccb99c AG |
1807 | * |
1808 | * Since it's solely used for userfault_fd WP feature, here we just | |
1809 | * hardcode page size to qemu_real_host_page_size. | |
1810 | * | |
82ea3e3b | 1811 | * @block: RAM block to populate |
eeccb99c | 1812 | */ |
6fee3a1f | 1813 | static void ram_block_populate_read(RAMBlock *rb) |
eeccb99c | 1814 | { |
6fee3a1f DH |
1815 | /* |
1816 | * Skip populating all pages that fall into a discarded range as managed by | |
1817 | * a RamDiscardManager responsible for the mapped memory region of the | |
1818 | * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock | |
1819 | * must not get populated automatically. We don't have to track | |
1820 | * modifications via userfaultfd WP reliably, because these pages will | |
1821 | * not be part of the migration stream either way -- see | |
1822 | * ramblock_dirty_bitmap_exclude_discarded_pages(). | |
1823 | * | |
1824 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1825 | */ | |
1826 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1827 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1828 | MemoryRegionSection section = { | |
1829 | .mr = rb->mr, | |
1830 | .offset_within_region = 0, | |
1831 | .size = rb->mr->size, | |
1832 | }; | |
1833 | ||
1834 | ram_discard_manager_replay_populated(rdm, §ion, | |
1835 | populate_read_section, NULL); | |
1836 | } else { | |
1837 | populate_read_range(rb, 0, rb->used_length); | |
1838 | } | |
eeccb99c AG |
1839 | } |
1840 | ||
1841 | /* | |
1842 | * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking | |
1843 | */ | |
1844 | void ram_write_tracking_prepare(void) | |
1845 | { | |
82ea3e3b | 1846 | RAMBlock *block; |
eeccb99c AG |
1847 | |
1848 | RCU_READ_LOCK_GUARD(); | |
1849 | ||
82ea3e3b | 1850 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
eeccb99c | 1851 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1852 | if (block->mr->readonly || block->mr->rom_device) { |
eeccb99c AG |
1853 | continue; |
1854 | } | |
1855 | ||
1856 | /* | |
1857 | * Populate pages of the RAM block before enabling userfault_fd | |
1858 | * write protection. | |
1859 | * | |
1860 | * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with | |
1861 | * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip | |
1862 | * pages with pte_none() entries in page table. | |
1863 | */ | |
f7b9dcfb | 1864 | ram_block_populate_read(block); |
eeccb99c AG |
1865 | } |
1866 | } | |
1867 | ||
e41c5770 DH |
1868 | static inline int uffd_protect_section(MemoryRegionSection *section, |
1869 | void *opaque) | |
1870 | { | |
1871 | const hwaddr size = int128_get64(section->size); | |
1872 | const hwaddr offset = section->offset_within_region; | |
1873 | RAMBlock *rb = section->mr->ram_block; | |
1874 | int uffd_fd = (uintptr_t)opaque; | |
1875 | ||
1876 | return uffd_change_protection(uffd_fd, rb->host + offset, size, true, | |
1877 | false); | |
1878 | } | |
1879 | ||
1880 | static int ram_block_uffd_protect(RAMBlock *rb, int uffd_fd) | |
1881 | { | |
1882 | assert(rb->flags & RAM_UF_WRITEPROTECT); | |
1883 | ||
1884 | /* See ram_block_populate_read() */ | |
1885 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1886 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1887 | MemoryRegionSection section = { | |
1888 | .mr = rb->mr, | |
1889 | .offset_within_region = 0, | |
1890 | .size = rb->mr->size, | |
1891 | }; | |
1892 | ||
1893 | return ram_discard_manager_replay_populated(rdm, §ion, | |
1894 | uffd_protect_section, | |
1895 | (void *)(uintptr_t)uffd_fd); | |
1896 | } | |
1897 | return uffd_change_protection(uffd_fd, rb->host, | |
1898 | rb->used_length, true, false); | |
1899 | } | |
1900 | ||
278e2f55 AG |
1901 | /* |
1902 | * ram_write_tracking_start: start UFFD-WP memory tracking | |
1903 | * | |
1904 | * Returns 0 for success or negative value in case of error | |
1905 | */ | |
1906 | int ram_write_tracking_start(void) | |
1907 | { | |
1908 | int uffd_fd; | |
1909 | RAMState *rs = ram_state; | |
82ea3e3b | 1910 | RAMBlock *block; |
278e2f55 AG |
1911 | |
1912 | /* Open UFFD file descriptor */ | |
1913 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, true); | |
1914 | if (uffd_fd < 0) { | |
1915 | return uffd_fd; | |
1916 | } | |
1917 | rs->uffdio_fd = uffd_fd; | |
1918 | ||
1919 | RCU_READ_LOCK_GUARD(); | |
1920 | ||
82ea3e3b | 1921 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 | 1922 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1923 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1924 | continue; |
1925 | } | |
1926 | ||
1927 | /* Register block memory with UFFD to track writes */ | |
82ea3e3b AG |
1928 | if (uffd_register_memory(rs->uffdio_fd, block->host, |
1929 | block->max_length, UFFDIO_REGISTER_MODE_WP, NULL)) { | |
278e2f55 AG |
1930 | goto fail; |
1931 | } | |
72ef3a37 DH |
1932 | block->flags |= RAM_UF_WRITEPROTECT; |
1933 | memory_region_ref(block->mr); | |
1934 | ||
278e2f55 | 1935 | /* Apply UFFD write protection to the block memory range */ |
e41c5770 | 1936 | if (ram_block_uffd_protect(block, uffd_fd)) { |
278e2f55 AG |
1937 | goto fail; |
1938 | } | |
278e2f55 | 1939 | |
82ea3e3b AG |
1940 | trace_ram_write_tracking_ramblock_start(block->idstr, block->page_size, |
1941 | block->host, block->max_length); | |
278e2f55 AG |
1942 | } |
1943 | ||
1944 | return 0; | |
1945 | ||
1946 | fail: | |
1947 | error_report("ram_write_tracking_start() failed: restoring initial memory state"); | |
1948 | ||
82ea3e3b AG |
1949 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1950 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1951 | continue; |
1952 | } | |
82ea3e3b | 1953 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 1954 | /* Cleanup flags and remove reference */ |
82ea3e3b AG |
1955 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1956 | memory_region_unref(block->mr); | |
278e2f55 AG |
1957 | } |
1958 | ||
1959 | uffd_close_fd(uffd_fd); | |
1960 | rs->uffdio_fd = -1; | |
1961 | return -1; | |
1962 | } | |
1963 | ||
1964 | /** | |
1965 | * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection | |
1966 | */ | |
1967 | void ram_write_tracking_stop(void) | |
1968 | { | |
1969 | RAMState *rs = ram_state; | |
82ea3e3b | 1970 | RAMBlock *block; |
278e2f55 AG |
1971 | |
1972 | RCU_READ_LOCK_GUARD(); | |
1973 | ||
82ea3e3b AG |
1974 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1975 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1976 | continue; |
1977 | } | |
82ea3e3b | 1978 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 1979 | |
82ea3e3b AG |
1980 | trace_ram_write_tracking_ramblock_stop(block->idstr, block->page_size, |
1981 | block->host, block->max_length); | |
278e2f55 AG |
1982 | |
1983 | /* Cleanup flags and remove reference */ | |
82ea3e3b AG |
1984 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1985 | memory_region_unref(block->mr); | |
278e2f55 AG |
1986 | } |
1987 | ||
1988 | /* Finally close UFFD file descriptor */ | |
1989 | uffd_close_fd(rs->uffdio_fd); | |
1990 | rs->uffdio_fd = -1; | |
1991 | } | |
1992 | ||
1993 | #else | |
1994 | /* No target OS support, stubs just fail or ignore */ | |
1995 | ||
1996 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1997 | { | |
1998 | (void) rs; | |
1999 | (void) offset; | |
2000 | ||
2001 | return NULL; | |
2002 | } | |
2003 | ||
2004 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
2005 | unsigned long start_page) | |
2006 | { | |
2007 | (void) rs; | |
2008 | (void) pss; | |
2009 | (void) start_page; | |
2010 | ||
2011 | return 0; | |
2012 | } | |
2013 | ||
2014 | bool ram_write_tracking_available(void) | |
2015 | { | |
2016 | return false; | |
2017 | } | |
2018 | ||
2019 | bool ram_write_tracking_compatible(void) | |
2020 | { | |
2021 | assert(0); | |
2022 | return false; | |
2023 | } | |
2024 | ||
2025 | int ram_write_tracking_start(void) | |
2026 | { | |
2027 | assert(0); | |
2028 | return -1; | |
2029 | } | |
2030 | ||
2031 | void ram_write_tracking_stop(void) | |
2032 | { | |
2033 | assert(0); | |
2034 | } | |
2035 | #endif /* defined(__linux__) */ | |
2036 | ||
3d0684b2 | 2037 | /** |
ff1543af | 2038 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
2039 | * |
2040 | * Skips pages that are already sent (!dirty) | |
a82d593b | 2041 | * |
a5f7b1a6 | 2042 | * Returns true if a queued page is found |
a82d593b | 2043 | * |
6f37bb8b | 2044 | * @rs: current RAM state |
3d0684b2 | 2045 | * @pss: data about the state of the current dirty page scan |
a82d593b | 2046 | */ |
f20e2865 | 2047 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
2048 | { |
2049 | RAMBlock *block; | |
2050 | ram_addr_t offset; | |
777f53c7 TH |
2051 | bool dirty; |
2052 | ||
2053 | do { | |
2054 | block = unqueue_page(rs, &offset); | |
2055 | /* | |
2056 | * We're sending this page, and since it's postcopy nothing else | |
2057 | * will dirty it, and we must make sure it doesn't get sent again | |
2058 | * even if this queue request was received after the background | |
2059 | * search already sent it. | |
2060 | */ | |
2061 | if (block) { | |
2062 | unsigned long page; | |
2063 | ||
2064 | page = offset >> TARGET_PAGE_BITS; | |
2065 | dirty = test_bit(page, block->bmap); | |
2066 | if (!dirty) { | |
2067 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, | |
2068 | page); | |
2069 | } else { | |
2070 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); | |
2071 | } | |
2072 | } | |
a82d593b | 2073 | |
777f53c7 | 2074 | } while (block && !dirty); |
a82d593b | 2075 | |
b062106d | 2076 | if (!block) { |
278e2f55 AG |
2077 | /* |
2078 | * Poll write faults too if background snapshot is enabled; that's | |
2079 | * when we have vcpus got blocked by the write protected pages. | |
2080 | */ | |
2081 | block = poll_fault_page(rs, &offset); | |
2082 | } | |
2083 | ||
a82d593b | 2084 | if (block) { |
a82d593b DDAG |
2085 | /* |
2086 | * We want the background search to continue from the queued page | |
2087 | * since the guest is likely to want other pages near to the page | |
2088 | * it just requested. | |
2089 | */ | |
2090 | pss->block = block; | |
a935e30f | 2091 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
2092 | |
2093 | /* | |
2094 | * This unqueued page would break the "one round" check, even is | |
2095 | * really rare. | |
2096 | */ | |
2097 | pss->complete_round = false; | |
a82d593b DDAG |
2098 | } |
2099 | ||
2100 | return !!block; | |
2101 | } | |
2102 | ||
6c595cde | 2103 | /** |
5e58f968 JQ |
2104 | * migration_page_queue_free: drop any remaining pages in the ram |
2105 | * request queue | |
6c595cde | 2106 | * |
3d0684b2 JQ |
2107 | * It should be empty at the end anyway, but in error cases there may |
2108 | * be some left. in case that there is any page left, we drop it. | |
2109 | * | |
6c595cde | 2110 | */ |
83c13382 | 2111 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 2112 | { |
ec481c6c | 2113 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
2114 | /* This queue generally should be empty - but in the case of a failed |
2115 | * migration might have some droppings in. | |
2116 | */ | |
89ac5a1d | 2117 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 2118 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 2119 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 2120 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
2121 | g_free(mspr); |
2122 | } | |
6c595cde DDAG |
2123 | } |
2124 | ||
2125 | /** | |
3d0684b2 JQ |
2126 | * ram_save_queue_pages: queue the page for transmission |
2127 | * | |
2128 | * A request from postcopy destination for example. | |
2129 | * | |
2130 | * Returns zero on success or negative on error | |
2131 | * | |
3d0684b2 JQ |
2132 | * @rbname: Name of the RAMBLock of the request. NULL means the |
2133 | * same that last one. | |
2134 | * @start: starting address from the start of the RAMBlock | |
2135 | * @len: length (in bytes) to send | |
6c595cde | 2136 | */ |
96506894 | 2137 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
2138 | { |
2139 | RAMBlock *ramblock; | |
53518d94 | 2140 | RAMState *rs = ram_state; |
6c595cde | 2141 | |
9360447d | 2142 | ram_counters.postcopy_requests++; |
89ac5a1d DDAG |
2143 | RCU_READ_LOCK_GUARD(); |
2144 | ||
6c595cde DDAG |
2145 | if (!rbname) { |
2146 | /* Reuse last RAMBlock */ | |
68a098f3 | 2147 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
2148 | |
2149 | if (!ramblock) { | |
2150 | /* | |
2151 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
2152 | * it's the 1st request. | |
2153 | */ | |
2154 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 2155 | return -1; |
6c595cde DDAG |
2156 | } |
2157 | } else { | |
2158 | ramblock = qemu_ram_block_by_name(rbname); | |
2159 | ||
2160 | if (!ramblock) { | |
2161 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
2162 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 2163 | return -1; |
6c595cde | 2164 | } |
68a098f3 | 2165 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
2166 | } |
2167 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
542147f4 | 2168 | if (!offset_in_ramblock(ramblock, start + len - 1)) { |
9458ad6b JQ |
2169 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
2170 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 2171 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 2172 | return -1; |
6c595cde DDAG |
2173 | } |
2174 | ||
93589827 PX |
2175 | /* |
2176 | * When with postcopy preempt, we send back the page directly in the | |
2177 | * rp-return thread. | |
2178 | */ | |
2179 | if (postcopy_preempt_active()) { | |
2180 | ram_addr_t page_start = start >> TARGET_PAGE_BITS; | |
2181 | size_t page_size = qemu_ram_pagesize(ramblock); | |
2182 | PageSearchStatus *pss = &ram_state->pss[RAM_CHANNEL_POSTCOPY]; | |
2183 | int ret = 0; | |
2184 | ||
2185 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2186 | ||
2187 | pss_init(pss, ramblock, page_start); | |
2188 | /* | |
2189 | * Always use the preempt channel, and make sure it's there. It's | |
2190 | * safe to access without lock, because when rp-thread is running | |
2191 | * we should be the only one who operates on the qemufile | |
2192 | */ | |
2193 | pss->pss_channel = migrate_get_current()->postcopy_qemufile_src; | |
93589827 PX |
2194 | assert(pss->pss_channel); |
2195 | ||
2196 | /* | |
2197 | * It must be either one or multiple of host page size. Just | |
2198 | * assert; if something wrong we're mostly split brain anyway. | |
2199 | */ | |
2200 | assert(len % page_size == 0); | |
2201 | while (len) { | |
2202 | if (ram_save_host_page_urgent(pss)) { | |
2203 | error_report("%s: ram_save_host_page_urgent() failed: " | |
2204 | "ramblock=%s, start_addr=0x"RAM_ADDR_FMT, | |
2205 | __func__, ramblock->idstr, start); | |
2206 | ret = -1; | |
2207 | break; | |
2208 | } | |
2209 | /* | |
2210 | * NOTE: after ram_save_host_page_urgent() succeeded, pss->page | |
2211 | * will automatically be moved and point to the next host page | |
2212 | * we're going to send, so no need to update here. | |
2213 | * | |
2214 | * Normally QEMU never sends >1 host page in requests, so | |
2215 | * logically we don't even need that as the loop should only | |
2216 | * run once, but just to be consistent. | |
2217 | */ | |
2218 | len -= page_size; | |
2219 | }; | |
2220 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
2221 | ||
2222 | return ret; | |
2223 | } | |
2224 | ||
ec481c6c | 2225 | struct RAMSrcPageRequest *new_entry = |
b21e2380 | 2226 | g_new0(struct RAMSrcPageRequest, 1); |
6c595cde DDAG |
2227 | new_entry->rb = ramblock; |
2228 | new_entry->offset = start; | |
2229 | new_entry->len = len; | |
2230 | ||
2231 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
2232 | qemu_mutex_lock(&rs->src_page_req_mutex); |
2233 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 2234 | migration_make_urgent_request(); |
ec481c6c | 2235 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
2236 | |
2237 | return 0; | |
6c595cde DDAG |
2238 | } |
2239 | ||
d7400a34 XG |
2240 | static bool save_page_use_compression(RAMState *rs) |
2241 | { | |
2242 | if (!migrate_use_compression()) { | |
2243 | return false; | |
2244 | } | |
2245 | ||
2246 | /* | |
1a373522 DH |
2247 | * If xbzrle is enabled (e.g., after first round of migration), stop |
2248 | * using the data compression. In theory, xbzrle can do better than | |
2249 | * compression. | |
d7400a34 | 2250 | */ |
1a373522 DH |
2251 | if (rs->xbzrle_enabled) { |
2252 | return false; | |
d7400a34 XG |
2253 | } |
2254 | ||
1a373522 | 2255 | return true; |
d7400a34 XG |
2256 | } |
2257 | ||
5e5fdcff XG |
2258 | /* |
2259 | * try to compress the page before posting it out, return true if the page | |
2260 | * has been properly handled by compression, otherwise needs other | |
2261 | * paths to handle it | |
2262 | */ | |
ec6f3ab9 PX |
2263 | static bool save_compress_page(RAMState *rs, PageSearchStatus *pss, |
2264 | RAMBlock *block, ram_addr_t offset) | |
5e5fdcff XG |
2265 | { |
2266 | if (!save_page_use_compression(rs)) { | |
2267 | return false; | |
2268 | } | |
2269 | ||
2270 | /* | |
2271 | * When starting the process of a new block, the first page of | |
2272 | * the block should be sent out before other pages in the same | |
2273 | * block, and all the pages in last block should have been sent | |
2274 | * out, keeping this order is important, because the 'cont' flag | |
2275 | * is used to avoid resending the block name. | |
2276 | * | |
2277 | * We post the fist page as normal page as compression will take | |
2278 | * much CPU resource. | |
2279 | */ | |
ec6f3ab9 | 2280 | if (block != pss->last_sent_block) { |
5e5fdcff XG |
2281 | flush_compressed_data(rs); |
2282 | return false; | |
2283 | } | |
2284 | ||
eaa238ab | 2285 | if (compress_page_with_multi_thread(block, offset) > 0) { |
5e5fdcff XG |
2286 | return true; |
2287 | } | |
2288 | ||
76e03000 | 2289 | compression_counters.busy++; |
5e5fdcff XG |
2290 | return false; |
2291 | } | |
2292 | ||
a82d593b | 2293 | /** |
3d0684b2 | 2294 | * ram_save_target_page: save one target page |
a82d593b | 2295 | * |
3d0684b2 | 2296 | * Returns the number of pages written |
a82d593b | 2297 | * |
6f37bb8b | 2298 | * @rs: current RAM state |
3d0684b2 | 2299 | * @pss: data about the page we want to send |
a82d593b | 2300 | */ |
05931ec5 | 2301 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2302 | { |
a8ec91f9 | 2303 | RAMBlock *block = pss->block; |
8bba004c | 2304 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
2305 | int res; |
2306 | ||
61717ea9 | 2307 | if (control_save_page(pss, block, offset, &res)) { |
a8ec91f9 XG |
2308 | return res; |
2309 | } | |
2310 | ||
ec6f3ab9 | 2311 | if (save_compress_page(rs, pss, block, offset)) { |
5e5fdcff | 2312 | return 1; |
d7400a34 XG |
2313 | } |
2314 | ||
ec6f3ab9 | 2315 | res = save_zero_page(pss, block, offset); |
d7400a34 XG |
2316 | if (res > 0) { |
2317 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
2318 | * page would be stale | |
2319 | */ | |
ef5c3d13 | 2320 | if (rs->xbzrle_enabled) { |
d7400a34 XG |
2321 | XBZRLE_cache_lock(); |
2322 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
2323 | XBZRLE_cache_unlock(); | |
2324 | } | |
d7400a34 XG |
2325 | return res; |
2326 | } | |
2327 | ||
da3f56cb | 2328 | /* |
6f39c90b PX |
2329 | * Do not use multifd in postcopy as one whole host page should be |
2330 | * placed. Meanwhile postcopy requires atomic update of pages, so even | |
2331 | * if host page size == guest page size the dest guest during run may | |
2332 | * still see partially copied pages which is data corruption. | |
da3f56cb | 2333 | */ |
6f39c90b | 2334 | if (migrate_use_multifd() && !migration_in_postcopy()) { |
61717ea9 | 2335 | return ram_save_multifd_page(pss->pss_channel, block, offset); |
a82d593b DDAG |
2336 | } |
2337 | ||
05931ec5 | 2338 | return ram_save_page(rs, pss); |
a82d593b DDAG |
2339 | } |
2340 | ||
d9e474ea PX |
2341 | /* Should be called before sending a host page */ |
2342 | static void pss_host_page_prepare(PageSearchStatus *pss) | |
2343 | { | |
2344 | /* How many guest pages are there in one host page? */ | |
2345 | size_t guest_pfns = qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
2346 | ||
2347 | pss->host_page_sending = true; | |
301d7ffe PX |
2348 | if (guest_pfns <= 1) { |
2349 | /* | |
2350 | * This covers both when guest psize == host psize, or when guest | |
2351 | * has larger psize than the host (guest_pfns==0). | |
2352 | * | |
2353 | * For the latter, we always send one whole guest page per | |
2354 | * iteration of the host page (example: an Alpha VM on x86 host | |
2355 | * will have guest psize 8K while host psize 4K). | |
2356 | */ | |
2357 | pss->host_page_start = pss->page; | |
2358 | pss->host_page_end = pss->page + 1; | |
2359 | } else { | |
2360 | /* | |
2361 | * The host page spans over multiple guest pages, we send them | |
2362 | * within the same host page iteration. | |
2363 | */ | |
2364 | pss->host_page_start = ROUND_DOWN(pss->page, guest_pfns); | |
2365 | pss->host_page_end = ROUND_UP(pss->page + 1, guest_pfns); | |
2366 | } | |
d9e474ea PX |
2367 | } |
2368 | ||
2369 | /* | |
2370 | * Whether the page pointed by PSS is within the host page being sent. | |
2371 | * Must be called after a previous pss_host_page_prepare(). | |
2372 | */ | |
2373 | static bool pss_within_range(PageSearchStatus *pss) | |
2374 | { | |
2375 | ram_addr_t ram_addr; | |
2376 | ||
2377 | assert(pss->host_page_sending); | |
2378 | ||
2379 | /* Over host-page boundary? */ | |
2380 | if (pss->page >= pss->host_page_end) { | |
2381 | return false; | |
2382 | } | |
2383 | ||
2384 | ram_addr = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; | |
2385 | ||
2386 | return offset_in_ramblock(pss->block, ram_addr); | |
2387 | } | |
2388 | ||
2389 | static void pss_host_page_finish(PageSearchStatus *pss) | |
2390 | { | |
2391 | pss->host_page_sending = false; | |
2392 | /* This is not needed, but just to reset it */ | |
2393 | pss->host_page_start = pss->host_page_end = 0; | |
2394 | } | |
2395 | ||
93589827 PX |
2396 | /* |
2397 | * Send an urgent host page specified by `pss'. Need to be called with | |
2398 | * bitmap_mutex held. | |
2399 | * | |
2400 | * Returns 0 if save host page succeeded, false otherwise. | |
2401 | */ | |
2402 | static int ram_save_host_page_urgent(PageSearchStatus *pss) | |
2403 | { | |
2404 | bool page_dirty, sent = false; | |
2405 | RAMState *rs = ram_state; | |
2406 | int ret = 0; | |
2407 | ||
2408 | trace_postcopy_preempt_send_host_page(pss->block->idstr, pss->page); | |
2409 | pss_host_page_prepare(pss); | |
2410 | ||
2411 | /* | |
2412 | * If precopy is sending the same page, let it be done in precopy, or | |
2413 | * we could send the same page in two channels and none of them will | |
2414 | * receive the whole page. | |
2415 | */ | |
2416 | if (pss_overlap(pss, &ram_state->pss[RAM_CHANNEL_PRECOPY])) { | |
2417 | trace_postcopy_preempt_hit(pss->block->idstr, | |
2418 | pss->page << TARGET_PAGE_BITS); | |
2419 | return 0; | |
2420 | } | |
2421 | ||
2422 | do { | |
2423 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); | |
2424 | ||
2425 | if (page_dirty) { | |
2426 | /* Be strict to return code; it must be 1, or what else? */ | |
2427 | if (ram_save_target_page(rs, pss) != 1) { | |
2428 | error_report_once("%s: ram_save_target_page failed", __func__); | |
2429 | ret = -1; | |
2430 | goto out; | |
2431 | } | |
2432 | sent = true; | |
2433 | } | |
2434 | pss_find_next_dirty(pss); | |
2435 | } while (pss_within_range(pss)); | |
2436 | out: | |
2437 | pss_host_page_finish(pss); | |
2438 | /* For urgent requests, flush immediately if sent */ | |
2439 | if (sent) { | |
2440 | qemu_fflush(pss->pss_channel); | |
2441 | } | |
2442 | return ret; | |
2443 | } | |
2444 | ||
a82d593b | 2445 | /** |
3d0684b2 | 2446 | * ram_save_host_page: save a whole host page |
a82d593b | 2447 | * |
3d0684b2 JQ |
2448 | * Starting at *offset send pages up to the end of the current host |
2449 | * page. It's valid for the initial offset to point into the middle of | |
2450 | * a host page in which case the remainder of the hostpage is sent. | |
2451 | * Only dirty target pages are sent. Note that the host page size may | |
2452 | * be a huge page for this block. | |
f3321554 | 2453 | * |
1eb3fc0a DDAG |
2454 | * The saving stops at the boundary of the used_length of the block |
2455 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 2456 | * |
f3321554 PX |
2457 | * The caller must be with ram_state.bitmap_mutex held to call this |
2458 | * function. Note that this function can temporarily release the lock, but | |
2459 | * when the function is returned it'll make sure the lock is still held. | |
2460 | * | |
3d0684b2 JQ |
2461 | * Returns the number of pages written or negative on error |
2462 | * | |
6f37bb8b | 2463 | * @rs: current RAM state |
3d0684b2 | 2464 | * @pss: data about the page we want to send |
a82d593b | 2465 | */ |
05931ec5 | 2466 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2467 | { |
f3321554 | 2468 | bool page_dirty, preempt_active = postcopy_preempt_active(); |
a82d593b | 2469 | int tmppages, pages = 0; |
a935e30f JQ |
2470 | size_t pagesize_bits = |
2471 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
278e2f55 AG |
2472 | unsigned long start_page = pss->page; |
2473 | int res; | |
4c011c37 | 2474 | |
fbd162e6 | 2475 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
2476 | error_report("block %s should not be migrated !", pss->block->idstr); |
2477 | return 0; | |
2478 | } | |
2479 | ||
d9e474ea PX |
2480 | /* Update host page boundary information */ |
2481 | pss_host_page_prepare(pss); | |
2482 | ||
a82d593b | 2483 | do { |
f3321554 | 2484 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); |
a82d593b | 2485 | |
f3321554 PX |
2486 | /* Check the pages is dirty and if it is send it */ |
2487 | if (page_dirty) { | |
ba1b7c81 | 2488 | /* |
f3321554 PX |
2489 | * Properly yield the lock only in postcopy preempt mode |
2490 | * because both migration thread and rp-return thread can | |
2491 | * operate on the bitmaps. | |
ba1b7c81 | 2492 | */ |
f3321554 PX |
2493 | if (preempt_active) { |
2494 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
ba1b7c81 | 2495 | } |
f3321554 PX |
2496 | tmppages = ram_save_target_page(rs, pss); |
2497 | if (tmppages >= 0) { | |
2498 | pages += tmppages; | |
2499 | /* | |
2500 | * Allow rate limiting to happen in the middle of huge pages if | |
2501 | * something is sent in the current iteration. | |
2502 | */ | |
2503 | if (pagesize_bits > 1 && tmppages > 0) { | |
2504 | migration_rate_limit(); | |
2505 | } | |
2506 | } | |
2507 | if (preempt_active) { | |
2508 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2509 | } | |
2510 | } else { | |
2511 | tmppages = 0; | |
23feba90 | 2512 | } |
f3321554 PX |
2513 | |
2514 | if (tmppages < 0) { | |
d9e474ea | 2515 | pss_host_page_finish(pss); |
f3321554 PX |
2516 | return tmppages; |
2517 | } | |
2518 | ||
d9e474ea PX |
2519 | pss_find_next_dirty(pss); |
2520 | } while (pss_within_range(pss)); | |
2521 | ||
2522 | pss_host_page_finish(pss); | |
278e2f55 AG |
2523 | |
2524 | res = ram_save_release_protection(rs, pss, start_page); | |
2525 | return (res < 0 ? res : pages); | |
a82d593b | 2526 | } |
6c595cde | 2527 | |
56e93d26 | 2528 | /** |
3d0684b2 | 2529 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
2530 | * |
2531 | * Called within an RCU critical section. | |
2532 | * | |
e8f3735f XG |
2533 | * Returns the number of pages written where zero means no dirty pages, |
2534 | * or negative on error | |
56e93d26 | 2535 | * |
6f37bb8b | 2536 | * @rs: current RAM state |
a82d593b DDAG |
2537 | * |
2538 | * On systems where host-page-size > target-page-size it will send all the | |
2539 | * pages in a host page that are dirty. | |
56e93d26 | 2540 | */ |
05931ec5 | 2541 | static int ram_find_and_save_block(RAMState *rs) |
56e93d26 | 2542 | { |
f1668764 | 2543 | PageSearchStatus *pss = &rs->pss[RAM_CHANNEL_PRECOPY]; |
56e93d26 | 2544 | int pages = 0; |
56e93d26 | 2545 | |
0827b9e9 AA |
2546 | /* No dirty page as there is zero RAM */ |
2547 | if (!ram_bytes_total()) { | |
2548 | return pages; | |
2549 | } | |
2550 | ||
4934a5dd PX |
2551 | /* |
2552 | * Always keep last_seen_block/last_page valid during this procedure, | |
2553 | * because find_dirty_block() relies on these values (e.g., we compare | |
2554 | * last_seen_block with pss.block to see whether we searched all the | |
2555 | * ramblocks) to detect the completion of migration. Having NULL value | |
2556 | * of last_seen_block can conditionally cause below loop to run forever. | |
2557 | */ | |
2558 | if (!rs->last_seen_block) { | |
2559 | rs->last_seen_block = QLIST_FIRST_RCU(&ram_list.blocks); | |
2560 | rs->last_page = 0; | |
2561 | } | |
2562 | ||
f1668764 | 2563 | pss_init(pss, rs->last_seen_block, rs->last_page); |
b8fb8cb7 | 2564 | |
b9e60928 | 2565 | do { |
51efd36f | 2566 | if (!get_queued_page(rs, pss)) { |
b062106d | 2567 | /* priority queue empty, so just search for something dirty */ |
51efd36f JQ |
2568 | bool again = true; |
2569 | if (!find_dirty_block(rs, pss, &again)) { | |
2570 | if (!again) { | |
2571 | break; | |
2572 | } | |
2573 | } | |
56e93d26 | 2574 | } |
51efd36f JQ |
2575 | pages = ram_save_host_page(rs, pss); |
2576 | } while (!pages); | |
56e93d26 | 2577 | |
f1668764 PX |
2578 | rs->last_seen_block = pss->block; |
2579 | rs->last_page = pss->page; | |
56e93d26 JQ |
2580 | |
2581 | return pages; | |
2582 | } | |
2583 | ||
2584 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
2585 | { | |
2586 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 2587 | |
56e93d26 | 2588 | if (zero) { |
23b7576d | 2589 | stat64_add(&ram_atomic_counters.duplicate, pages); |
56e93d26 | 2590 | } else { |
23b7576d | 2591 | stat64_add(&ram_atomic_counters.normal, pages); |
4c2d0f6d | 2592 | ram_transferred_add(size); |
1a93bd2f | 2593 | qemu_file_credit_transfer(f, size); |
56e93d26 JQ |
2594 | } |
2595 | } | |
2596 | ||
fbd162e6 | 2597 | static uint64_t ram_bytes_total_common(bool count_ignored) |
56e93d26 JQ |
2598 | { |
2599 | RAMBlock *block; | |
2600 | uint64_t total = 0; | |
2601 | ||
89ac5a1d DDAG |
2602 | RCU_READ_LOCK_GUARD(); |
2603 | ||
fbd162e6 YK |
2604 | if (count_ignored) { |
2605 | RAMBLOCK_FOREACH_MIGRATABLE(block) { | |
2606 | total += block->used_length; | |
2607 | } | |
2608 | } else { | |
2609 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2610 | total += block->used_length; | |
2611 | } | |
99e15582 | 2612 | } |
56e93d26 JQ |
2613 | return total; |
2614 | } | |
2615 | ||
fbd162e6 YK |
2616 | uint64_t ram_bytes_total(void) |
2617 | { | |
2618 | return ram_bytes_total_common(false); | |
2619 | } | |
2620 | ||
f265e0e4 | 2621 | static void xbzrle_load_setup(void) |
56e93d26 | 2622 | { |
f265e0e4 | 2623 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2624 | } |
2625 | ||
f265e0e4 JQ |
2626 | static void xbzrle_load_cleanup(void) |
2627 | { | |
2628 | g_free(XBZRLE.decoded_buf); | |
2629 | XBZRLE.decoded_buf = NULL; | |
2630 | } | |
2631 | ||
7d7c96be PX |
2632 | static void ram_state_cleanup(RAMState **rsp) |
2633 | { | |
b9ccaf6d DDAG |
2634 | if (*rsp) { |
2635 | migration_page_queue_free(*rsp); | |
2636 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
2637 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
2638 | g_free(*rsp); | |
2639 | *rsp = NULL; | |
2640 | } | |
7d7c96be PX |
2641 | } |
2642 | ||
84593a08 PX |
2643 | static void xbzrle_cleanup(void) |
2644 | { | |
2645 | XBZRLE_cache_lock(); | |
2646 | if (XBZRLE.cache) { | |
2647 | cache_fini(XBZRLE.cache); | |
2648 | g_free(XBZRLE.encoded_buf); | |
2649 | g_free(XBZRLE.current_buf); | |
2650 | g_free(XBZRLE.zero_target_page); | |
2651 | XBZRLE.cache = NULL; | |
2652 | XBZRLE.encoded_buf = NULL; | |
2653 | XBZRLE.current_buf = NULL; | |
2654 | XBZRLE.zero_target_page = NULL; | |
2655 | } | |
2656 | XBZRLE_cache_unlock(); | |
2657 | } | |
2658 | ||
f265e0e4 | 2659 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 2660 | { |
53518d94 | 2661 | RAMState **rsp = opaque; |
6b6712ef | 2662 | RAMBlock *block; |
eb859c53 | 2663 | |
278e2f55 AG |
2664 | /* We don't use dirty log with background snapshots */ |
2665 | if (!migrate_background_snapshot()) { | |
2666 | /* caller have hold iothread lock or is in a bh, so there is | |
2667 | * no writing race against the migration bitmap | |
2668 | */ | |
63b41db4 HH |
2669 | if (global_dirty_tracking & GLOBAL_DIRTY_MIGRATION) { |
2670 | /* | |
2671 | * do not stop dirty log without starting it, since | |
2672 | * memory_global_dirty_log_stop will assert that | |
2673 | * memory_global_dirty_log_start/stop used in pairs | |
2674 | */ | |
2675 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); | |
2676 | } | |
278e2f55 | 2677 | } |
6b6712ef | 2678 | |
fbd162e6 | 2679 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
2680 | g_free(block->clear_bmap); |
2681 | block->clear_bmap = NULL; | |
6b6712ef JQ |
2682 | g_free(block->bmap); |
2683 | block->bmap = NULL; | |
56e93d26 JQ |
2684 | } |
2685 | ||
84593a08 | 2686 | xbzrle_cleanup(); |
f0afa331 | 2687 | compress_threads_save_cleanup(); |
7d7c96be | 2688 | ram_state_cleanup(rsp); |
56e93d26 JQ |
2689 | } |
2690 | ||
6f37bb8b | 2691 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 2692 | { |
ec6f3ab9 PX |
2693 | int i; |
2694 | ||
2695 | for (i = 0; i < RAM_CHANNEL_MAX; i++) { | |
2696 | rs->pss[i].last_sent_block = NULL; | |
2697 | } | |
2698 | ||
6f37bb8b | 2699 | rs->last_seen_block = NULL; |
269ace29 | 2700 | rs->last_page = 0; |
6f37bb8b | 2701 | rs->last_version = ram_list.version; |
1a373522 | 2702 | rs->xbzrle_enabled = false; |
56e93d26 JQ |
2703 | } |
2704 | ||
2705 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
2706 | ||
e0b266f0 DDAG |
2707 | /* **** functions for postcopy ***** */ |
2708 | ||
ced1c616 PB |
2709 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
2710 | { | |
2711 | struct RAMBlock *block; | |
ced1c616 | 2712 | |
fbd162e6 | 2713 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2714 | unsigned long *bitmap = block->bmap; |
2715 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
2716 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
2717 | |
2718 | while (run_start < range) { | |
2719 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
2720 | ram_discard_range(block->idstr, |
2721 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
2722 | ((ram_addr_t)(run_end - run_start)) | |
2723 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
2724 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
2725 | } | |
2726 | } | |
2727 | } | |
2728 | ||
3d0684b2 JQ |
2729 | /** |
2730 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
2731 | * | |
e0b266f0 | 2732 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
2733 | * |
2734 | * @ms: current migration state | |
89dab31b | 2735 | * @block: RAMBlock to discard |
e0b266f0 | 2736 | */ |
9e7d1223 | 2737 | static void postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 2738 | { |
6b6712ef | 2739 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 2740 | unsigned long current; |
1e7cf8c3 | 2741 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 2742 | |
6b6712ef | 2743 | for (current = 0; current < end; ) { |
1e7cf8c3 | 2744 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 2745 | unsigned long zero, discard_length; |
e0b266f0 | 2746 | |
33a5cb62 WY |
2747 | if (one >= end) { |
2748 | break; | |
2749 | } | |
e0b266f0 | 2750 | |
1e7cf8c3 | 2751 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
2752 | |
2753 | if (zero >= end) { | |
2754 | discard_length = end - one; | |
e0b266f0 | 2755 | } else { |
33a5cb62 WY |
2756 | discard_length = zero - one; |
2757 | } | |
810cf2bb | 2758 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 2759 | current = one + discard_length; |
e0b266f0 | 2760 | } |
e0b266f0 DDAG |
2761 | } |
2762 | ||
f30c2e5b PX |
2763 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block); |
2764 | ||
3d0684b2 JQ |
2765 | /** |
2766 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
2767 | * | |
e0b266f0 DDAG |
2768 | * Utility for the outgoing postcopy code. |
2769 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
2770 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
2771 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
2772 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
2773 | * |
2774 | * @ms: current migration state | |
e0b266f0 | 2775 | */ |
739fcc1b | 2776 | static void postcopy_each_ram_send_discard(MigrationState *ms) |
e0b266f0 DDAG |
2777 | { |
2778 | struct RAMBlock *block; | |
e0b266f0 | 2779 | |
fbd162e6 | 2780 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 2781 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 | 2782 | |
f30c2e5b PX |
2783 | /* |
2784 | * Deal with TPS != HPS and huge pages. It discard any partially sent | |
2785 | * host-page size chunks, mark any partially dirty host-page size | |
2786 | * chunks as all dirty. In this case the host-page is the host-page | |
2787 | * for the particular RAMBlock, i.e. it might be a huge page. | |
2788 | */ | |
2789 | postcopy_chunk_hostpages_pass(ms, block); | |
2790 | ||
e0b266f0 DDAG |
2791 | /* |
2792 | * Postcopy sends chunks of bitmap over the wire, but it | |
2793 | * just needs indexes at this point, avoids it having | |
2794 | * target page specific code. | |
2795 | */ | |
739fcc1b | 2796 | postcopy_send_discard_bm_ram(ms, block); |
810cf2bb | 2797 | postcopy_discard_send_finish(ms); |
e0b266f0 | 2798 | } |
e0b266f0 DDAG |
2799 | } |
2800 | ||
3d0684b2 | 2801 | /** |
8324ef86 | 2802 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2803 | * |
2804 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2805 | * canonicalize the two bitmaps, that are similar, but one is | |
2806 | * inverted. | |
99e314eb | 2807 | * |
3d0684b2 JQ |
2808 | * Postcopy requires that all target pages in a hostpage are dirty or |
2809 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2810 | * |
3d0684b2 | 2811 | * @ms: current migration state |
3d0684b2 | 2812 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2813 | */ |
1e7cf8c3 | 2814 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2815 | { |
53518d94 | 2816 | RAMState *rs = ram_state; |
6b6712ef | 2817 | unsigned long *bitmap = block->bmap; |
29c59172 | 2818 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2819 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2820 | unsigned long run_start; |
2821 | ||
29c59172 DDAG |
2822 | if (block->page_size == TARGET_PAGE_SIZE) { |
2823 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2824 | return; | |
2825 | } | |
2826 | ||
1e7cf8c3 WY |
2827 | /* Find a dirty page */ |
2828 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2829 | |
6b6712ef | 2830 | while (run_start < pages) { |
99e314eb DDAG |
2831 | |
2832 | /* | |
2833 | * If the start of this run of pages is in the middle of a host | |
2834 | * page, then we need to fixup this host page. | |
2835 | */ | |
9dec3cc3 | 2836 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2837 | /* Find the end of this run */ |
1e7cf8c3 | 2838 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2839 | /* |
2840 | * If the end isn't at the start of a host page, then the | |
2841 | * run doesn't finish at the end of a host page | |
2842 | * and we need to discard. | |
2843 | */ | |
99e314eb DDAG |
2844 | } |
2845 | ||
9dec3cc3 | 2846 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2847 | unsigned long page; |
dad45ab2 WY |
2848 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2849 | host_ratio); | |
2850 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2851 | |
99e314eb DDAG |
2852 | /* Clean up the bitmap */ |
2853 | for (page = fixup_start_addr; | |
2854 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2855 | /* |
2856 | * Remark them as dirty, updating the count for any pages | |
2857 | * that weren't previously dirty. | |
2858 | */ | |
0d8ec885 | 2859 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2860 | } |
2861 | } | |
2862 | ||
1e7cf8c3 WY |
2863 | /* Find the next dirty page for the next iteration */ |
2864 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2865 | } |
2866 | } | |
2867 | ||
3d0684b2 JQ |
2868 | /** |
2869 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2870 | * | |
e0b266f0 DDAG |
2871 | * Transmit the set of pages to be discarded after precopy to the target |
2872 | * these are pages that: | |
2873 | * a) Have been previously transmitted but are now dirty again | |
2874 | * b) Pages that have never been transmitted, this ensures that | |
2875 | * any pages on the destination that have been mapped by background | |
2876 | * tasks get discarded (transparent huge pages is the specific concern) | |
2877 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2878 | * |
2879 | * @ms: current migration state | |
e0b266f0 | 2880 | */ |
739fcc1b | 2881 | void ram_postcopy_send_discard_bitmap(MigrationState *ms) |
e0b266f0 | 2882 | { |
53518d94 | 2883 | RAMState *rs = ram_state; |
e0b266f0 | 2884 | |
89ac5a1d | 2885 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2886 | |
2887 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 2888 | migration_bitmap_sync(rs); |
e0b266f0 | 2889 | |
6b6712ef | 2890 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
ec6f3ab9 | 2891 | rs->pss[RAM_CHANNEL_PRECOPY].last_sent_block = NULL; |
6b6712ef | 2892 | rs->last_seen_block = NULL; |
6b6712ef | 2893 | rs->last_page = 0; |
e0b266f0 | 2894 | |
739fcc1b | 2895 | postcopy_each_ram_send_discard(ms); |
e0b266f0 | 2896 | |
739fcc1b | 2897 | trace_ram_postcopy_send_discard_bitmap(); |
e0b266f0 DDAG |
2898 | } |
2899 | ||
3d0684b2 JQ |
2900 | /** |
2901 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2902 | * |
3d0684b2 | 2903 | * Returns zero on success |
e0b266f0 | 2904 | * |
36449157 JQ |
2905 | * @rbname: name of the RAMBlock of the request. NULL means the |
2906 | * same that last one. | |
3d0684b2 JQ |
2907 | * @start: RAMBlock starting page |
2908 | * @length: RAMBlock size | |
e0b266f0 | 2909 | */ |
aaa2064c | 2910 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 2911 | { |
36449157 | 2912 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2913 | |
89ac5a1d | 2914 | RCU_READ_LOCK_GUARD(); |
36449157 | 2915 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2916 | |
2917 | if (!rb) { | |
36449157 | 2918 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 2919 | return -1; |
e0b266f0 DDAG |
2920 | } |
2921 | ||
814bb08f PX |
2922 | /* |
2923 | * On source VM, we don't need to update the received bitmap since | |
2924 | * we don't even have one. | |
2925 | */ | |
2926 | if (rb->receivedmap) { | |
2927 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
2928 | length >> qemu_target_page_bits()); | |
2929 | } | |
2930 | ||
03acb4e9 | 2931 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2932 | } |
2933 | ||
84593a08 PX |
2934 | /* |
2935 | * For every allocation, we will try not to crash the VM if the | |
2936 | * allocation failed. | |
2937 | */ | |
2938 | static int xbzrle_init(void) | |
2939 | { | |
2940 | Error *local_err = NULL; | |
2941 | ||
2942 | if (!migrate_use_xbzrle()) { | |
2943 | return 0; | |
2944 | } | |
2945 | ||
2946 | XBZRLE_cache_lock(); | |
2947 | ||
2948 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
2949 | if (!XBZRLE.zero_target_page) { | |
2950 | error_report("%s: Error allocating zero page", __func__); | |
2951 | goto err_out; | |
2952 | } | |
2953 | ||
2954 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
2955 | TARGET_PAGE_SIZE, &local_err); | |
2956 | if (!XBZRLE.cache) { | |
2957 | error_report_err(local_err); | |
2958 | goto free_zero_page; | |
2959 | } | |
2960 | ||
2961 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2962 | if (!XBZRLE.encoded_buf) { | |
2963 | error_report("%s: Error allocating encoded_buf", __func__); | |
2964 | goto free_cache; | |
2965 | } | |
2966 | ||
2967 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2968 | if (!XBZRLE.current_buf) { | |
2969 | error_report("%s: Error allocating current_buf", __func__); | |
2970 | goto free_encoded_buf; | |
2971 | } | |
2972 | ||
2973 | /* We are all good */ | |
2974 | XBZRLE_cache_unlock(); | |
2975 | return 0; | |
2976 | ||
2977 | free_encoded_buf: | |
2978 | g_free(XBZRLE.encoded_buf); | |
2979 | XBZRLE.encoded_buf = NULL; | |
2980 | free_cache: | |
2981 | cache_fini(XBZRLE.cache); | |
2982 | XBZRLE.cache = NULL; | |
2983 | free_zero_page: | |
2984 | g_free(XBZRLE.zero_target_page); | |
2985 | XBZRLE.zero_target_page = NULL; | |
2986 | err_out: | |
2987 | XBZRLE_cache_unlock(); | |
2988 | return -ENOMEM; | |
2989 | } | |
2990 | ||
53518d94 | 2991 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2992 | { |
7d00ee6a PX |
2993 | *rsp = g_try_new0(RAMState, 1); |
2994 | ||
2995 | if (!*rsp) { | |
2996 | error_report("%s: Init ramstate fail", __func__); | |
2997 | return -1; | |
2998 | } | |
53518d94 JQ |
2999 | |
3000 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
3001 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
3002 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
56e93d26 | 3003 | |
7d00ee6a | 3004 | /* |
40c4d4a8 IR |
3005 | * Count the total number of pages used by ram blocks not including any |
3006 | * gaps due to alignment or unplugs. | |
03158519 | 3007 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 3008 | */ |
40c4d4a8 | 3009 | (*rsp)->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
7d00ee6a PX |
3010 | ram_state_reset(*rsp); |
3011 | ||
3012 | return 0; | |
3013 | } | |
3014 | ||
d6eff5d7 | 3015 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 3016 | { |
002cad6b | 3017 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
3018 | RAMBlock *block; |
3019 | unsigned long pages; | |
002cad6b | 3020 | uint8_t shift; |
56e93d26 | 3021 | |
0827b9e9 AA |
3022 | /* Skip setting bitmap if there is no RAM */ |
3023 | if (ram_bytes_total()) { | |
002cad6b PX |
3024 | shift = ms->clear_bitmap_shift; |
3025 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
3026 | error_report("clear_bitmap_shift (%u) too big, using " | |
3027 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
3028 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
3029 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
3030 | error_report("clear_bitmap_shift (%u) too small, using " | |
3031 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
3032 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
3033 | } | |
3034 | ||
fbd162e6 | 3035 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 3036 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
3037 | /* |
3038 | * The initial dirty bitmap for migration must be set with all | |
3039 | * ones to make sure we'll migrate every guest RAM page to | |
3040 | * destination. | |
40c4d4a8 IR |
3041 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
3042 | * new migration after a failed migration, ram_list. | |
3043 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
3044 | * guest memory. | |
03158519 | 3045 | */ |
6b6712ef | 3046 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 3047 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
3048 | block->clear_bmap_shift = shift; |
3049 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 3050 | } |
f3f491fc | 3051 | } |
d6eff5d7 PX |
3052 | } |
3053 | ||
be39b4cd DH |
3054 | static void migration_bitmap_clear_discarded_pages(RAMState *rs) |
3055 | { | |
3056 | unsigned long pages; | |
3057 | RAMBlock *rb; | |
3058 | ||
3059 | RCU_READ_LOCK_GUARD(); | |
3060 | ||
3061 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { | |
3062 | pages = ramblock_dirty_bitmap_clear_discarded_pages(rb); | |
3063 | rs->migration_dirty_pages -= pages; | |
3064 | } | |
3065 | } | |
3066 | ||
d6eff5d7 PX |
3067 | static void ram_init_bitmaps(RAMState *rs) |
3068 | { | |
3069 | /* For memory_global_dirty_log_start below. */ | |
3070 | qemu_mutex_lock_iothread(); | |
3071 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 3072 | |
89ac5a1d DDAG |
3073 | WITH_RCU_READ_LOCK_GUARD() { |
3074 | ram_list_init_bitmaps(); | |
278e2f55 AG |
3075 | /* We don't use dirty log with background snapshots */ |
3076 | if (!migrate_background_snapshot()) { | |
63b41db4 | 3077 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
278e2f55 AG |
3078 | migration_bitmap_sync_precopy(rs); |
3079 | } | |
89ac5a1d | 3080 | } |
56e93d26 | 3081 | qemu_mutex_unlock_ramlist(); |
49877834 | 3082 | qemu_mutex_unlock_iothread(); |
be39b4cd DH |
3083 | |
3084 | /* | |
3085 | * After an eventual first bitmap sync, fixup the initial bitmap | |
3086 | * containing all 1s to exclude any discarded pages from migration. | |
3087 | */ | |
3088 | migration_bitmap_clear_discarded_pages(rs); | |
d6eff5d7 PX |
3089 | } |
3090 | ||
3091 | static int ram_init_all(RAMState **rsp) | |
3092 | { | |
3093 | if (ram_state_init(rsp)) { | |
3094 | return -1; | |
3095 | } | |
3096 | ||
3097 | if (xbzrle_init()) { | |
3098 | ram_state_cleanup(rsp); | |
3099 | return -1; | |
3100 | } | |
3101 | ||
3102 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
3103 | |
3104 | return 0; | |
3105 | } | |
3106 | ||
08614f34 PX |
3107 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
3108 | { | |
3109 | RAMBlock *block; | |
3110 | uint64_t pages = 0; | |
3111 | ||
3112 | /* | |
3113 | * Postcopy is not using xbzrle/compression, so no need for that. | |
3114 | * Also, since source are already halted, we don't need to care | |
3115 | * about dirty page logging as well. | |
3116 | */ | |
3117 | ||
fbd162e6 | 3118 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
3119 | pages += bitmap_count_one(block->bmap, |
3120 | block->used_length >> TARGET_PAGE_BITS); | |
3121 | } | |
3122 | ||
3123 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
3124 | rs->migration_dirty_pages = pages; | |
3125 | ||
1a373522 | 3126 | ram_state_reset(rs); |
08614f34 PX |
3127 | |
3128 | /* Update RAMState cache of output QEMUFile */ | |
7f401b80 | 3129 | rs->pss[RAM_CHANNEL_PRECOPY].pss_channel = out; |
08614f34 PX |
3130 | |
3131 | trace_ram_state_resume_prepare(pages); | |
3132 | } | |
3133 | ||
6bcb05fc WW |
3134 | /* |
3135 | * This function clears bits of the free pages reported by the caller from the | |
3136 | * migration dirty bitmap. @addr is the host address corresponding to the | |
3137 | * start of the continuous guest free pages, and @len is the total bytes of | |
3138 | * those pages. | |
3139 | */ | |
3140 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
3141 | { | |
3142 | RAMBlock *block; | |
3143 | ram_addr_t offset; | |
3144 | size_t used_len, start, npages; | |
3145 | MigrationState *s = migrate_get_current(); | |
3146 | ||
3147 | /* This function is currently expected to be used during live migration */ | |
3148 | if (!migration_is_setup_or_active(s->state)) { | |
3149 | return; | |
3150 | } | |
3151 | ||
3152 | for (; len > 0; len -= used_len, addr += used_len) { | |
3153 | block = qemu_ram_block_from_host(addr, false, &offset); | |
3154 | if (unlikely(!block || offset >= block->used_length)) { | |
3155 | /* | |
3156 | * The implementation might not support RAMBlock resize during | |
3157 | * live migration, but it could happen in theory with future | |
3158 | * updates. So we add a check here to capture that case. | |
3159 | */ | |
3160 | error_report_once("%s unexpected error", __func__); | |
3161 | return; | |
3162 | } | |
3163 | ||
3164 | if (len <= block->used_length - offset) { | |
3165 | used_len = len; | |
3166 | } else { | |
3167 | used_len = block->used_length - offset; | |
3168 | } | |
3169 | ||
3170 | start = offset >> TARGET_PAGE_BITS; | |
3171 | npages = used_len >> TARGET_PAGE_BITS; | |
3172 | ||
3173 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3143577d WW |
3174 | /* |
3175 | * The skipped free pages are equavalent to be sent from clear_bmap's | |
3176 | * perspective, so clear the bits from the memory region bitmap which | |
3177 | * are initially set. Otherwise those skipped pages will be sent in | |
3178 | * the next round after syncing from the memory region bitmap. | |
3179 | */ | |
1230a25f | 3180 | migration_clear_memory_region_dirty_bitmap_range(block, start, npages); |
6bcb05fc WW |
3181 | ram_state->migration_dirty_pages -= |
3182 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
3183 | bitmap_clear(block->bmap, start, npages); | |
3184 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
3185 | } | |
3186 | } | |
3187 | ||
3d0684b2 JQ |
3188 | /* |
3189 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
3190 | * long-running RCU critical section. When rcu-reclaims in the code |
3191 | * start to become numerous it will be necessary to reduce the | |
3192 | * granularity of these critical sections. | |
3193 | */ | |
3194 | ||
3d0684b2 JQ |
3195 | /** |
3196 | * ram_save_setup: Setup RAM for migration | |
3197 | * | |
3198 | * Returns zero to indicate success and negative for error | |
3199 | * | |
3200 | * @f: QEMUFile where to send the data | |
3201 | * @opaque: RAMState pointer | |
3202 | */ | |
a91246c9 HZ |
3203 | static int ram_save_setup(QEMUFile *f, void *opaque) |
3204 | { | |
53518d94 | 3205 | RAMState **rsp = opaque; |
a91246c9 | 3206 | RAMBlock *block; |
33d70973 | 3207 | int ret; |
a91246c9 | 3208 | |
dcaf446e XG |
3209 | if (compress_threads_save_setup()) { |
3210 | return -1; | |
3211 | } | |
3212 | ||
a91246c9 HZ |
3213 | /* migration has already setup the bitmap, reuse it. */ |
3214 | if (!migration_in_colo_state()) { | |
7d00ee6a | 3215 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 3216 | compress_threads_save_cleanup(); |
a91246c9 | 3217 | return -1; |
53518d94 | 3218 | } |
a91246c9 | 3219 | } |
7f401b80 | 3220 | (*rsp)->pss[RAM_CHANNEL_PRECOPY].pss_channel = f; |
a91246c9 | 3221 | |
0e6ebd48 DDAG |
3222 | WITH_RCU_READ_LOCK_GUARD() { |
3223 | qemu_put_be64(f, ram_bytes_total_common(true) | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 3224 | |
0e6ebd48 DDAG |
3225 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
3226 | qemu_put_byte(f, strlen(block->idstr)); | |
3227 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
3228 | qemu_put_be64(f, block->used_length); | |
3229 | if (migrate_postcopy_ram() && block->page_size != | |
3230 | qemu_host_page_size) { | |
3231 | qemu_put_be64(f, block->page_size); | |
3232 | } | |
3233 | if (migrate_ignore_shared()) { | |
3234 | qemu_put_be64(f, block->mr->addr); | |
3235 | } | |
fbd162e6 | 3236 | } |
56e93d26 JQ |
3237 | } |
3238 | ||
56e93d26 JQ |
3239 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
3240 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
3241 | ||
33d70973 LB |
3242 | ret = multifd_send_sync_main(f); |
3243 | if (ret < 0) { | |
3244 | return ret; | |
3245 | } | |
3246 | ||
56e93d26 | 3247 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 3248 | qemu_fflush(f); |
56e93d26 JQ |
3249 | |
3250 | return 0; | |
3251 | } | |
3252 | ||
3d0684b2 JQ |
3253 | /** |
3254 | * ram_save_iterate: iterative stage for migration | |
3255 | * | |
3256 | * Returns zero to indicate success and negative for error | |
3257 | * | |
3258 | * @f: QEMUFile where to send the data | |
3259 | * @opaque: RAMState pointer | |
3260 | */ | |
56e93d26 JQ |
3261 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
3262 | { | |
53518d94 JQ |
3263 | RAMState **temp = opaque; |
3264 | RAMState *rs = *temp; | |
3d4095b2 | 3265 | int ret = 0; |
56e93d26 JQ |
3266 | int i; |
3267 | int64_t t0; | |
5c90308f | 3268 | int done = 0; |
56e93d26 | 3269 | |
b2557345 PL |
3270 | if (blk_mig_bulk_active()) { |
3271 | /* Avoid transferring ram during bulk phase of block migration as | |
3272 | * the bulk phase will usually take a long time and transferring | |
3273 | * ram updates during that time is pointless. */ | |
3274 | goto out; | |
3275 | } | |
3276 | ||
63268c49 PX |
3277 | /* |
3278 | * We'll take this lock a little bit long, but it's okay for two reasons. | |
3279 | * Firstly, the only possible other thread to take it is who calls | |
3280 | * qemu_guest_free_page_hint(), which should be rare; secondly, see | |
3281 | * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which | |
3282 | * guarantees that we'll at least released it in a regular basis. | |
3283 | */ | |
3284 | qemu_mutex_lock(&rs->bitmap_mutex); | |
89ac5a1d DDAG |
3285 | WITH_RCU_READ_LOCK_GUARD() { |
3286 | if (ram_list.version != rs->last_version) { | |
3287 | ram_state_reset(rs); | |
3288 | } | |
56e93d26 | 3289 | |
89ac5a1d DDAG |
3290 | /* Read version before ram_list.blocks */ |
3291 | smp_rmb(); | |
56e93d26 | 3292 | |
89ac5a1d | 3293 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 3294 | |
89ac5a1d DDAG |
3295 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
3296 | i = 0; | |
3297 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
a1fe28df | 3298 | postcopy_has_request(rs)) { |
89ac5a1d | 3299 | int pages; |
e03a34f8 | 3300 | |
89ac5a1d DDAG |
3301 | if (qemu_file_get_error(f)) { |
3302 | break; | |
3303 | } | |
e8f3735f | 3304 | |
05931ec5 | 3305 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3306 | /* no more pages to sent */ |
3307 | if (pages == 0) { | |
3308 | done = 1; | |
3309 | break; | |
3310 | } | |
e8f3735f | 3311 | |
89ac5a1d DDAG |
3312 | if (pages < 0) { |
3313 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
3314 | break; |
3315 | } | |
89ac5a1d DDAG |
3316 | |
3317 | rs->target_page_count += pages; | |
3318 | ||
644acf99 WY |
3319 | /* |
3320 | * During postcopy, it is necessary to make sure one whole host | |
3321 | * page is sent in one chunk. | |
3322 | */ | |
3323 | if (migrate_postcopy_ram()) { | |
3324 | flush_compressed_data(rs); | |
3325 | } | |
3326 | ||
89ac5a1d DDAG |
3327 | /* |
3328 | * we want to check in the 1st loop, just in case it was the 1st | |
3329 | * time and we had to sync the dirty bitmap. | |
3330 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
3331 | * some iterations | |
3332 | */ | |
3333 | if ((i & 63) == 0) { | |
3334 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
3335 | 1000000; | |
3336 | if (t1 > MAX_WAIT) { | |
3337 | trace_ram_save_iterate_big_wait(t1, i); | |
3338 | break; | |
3339 | } | |
3340 | } | |
3341 | i++; | |
56e93d26 | 3342 | } |
56e93d26 | 3343 | } |
63268c49 | 3344 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 JQ |
3345 | |
3346 | /* | |
3347 | * Must occur before EOS (or any QEMUFile operation) | |
3348 | * because of RDMA protocol. | |
3349 | */ | |
3350 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
3351 | ||
b2557345 | 3352 | out: |
b69a0227 JQ |
3353 | if (ret >= 0 |
3354 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
7f401b80 | 3355 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); |
33d70973 LB |
3356 | if (ret < 0) { |
3357 | return ret; | |
3358 | } | |
3359 | ||
3d4095b2 JQ |
3360 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3361 | qemu_fflush(f); | |
4c2d0f6d | 3362 | ram_transferred_add(8); |
56e93d26 | 3363 | |
3d4095b2 JQ |
3364 | ret = qemu_file_get_error(f); |
3365 | } | |
56e93d26 JQ |
3366 | if (ret < 0) { |
3367 | return ret; | |
3368 | } | |
3369 | ||
5c90308f | 3370 | return done; |
56e93d26 JQ |
3371 | } |
3372 | ||
3d0684b2 JQ |
3373 | /** |
3374 | * ram_save_complete: function called to send the remaining amount of ram | |
3375 | * | |
e8f3735f | 3376 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
3377 | * |
3378 | * Called with iothread lock | |
3379 | * | |
3380 | * @f: QEMUFile where to send the data | |
3381 | * @opaque: RAMState pointer | |
3382 | */ | |
56e93d26 JQ |
3383 | static int ram_save_complete(QEMUFile *f, void *opaque) |
3384 | { | |
53518d94 JQ |
3385 | RAMState **temp = opaque; |
3386 | RAMState *rs = *temp; | |
e8f3735f | 3387 | int ret = 0; |
6f37bb8b | 3388 | |
05931ec5 JQ |
3389 | rs->last_stage = !migration_in_colo_state(); |
3390 | ||
89ac5a1d DDAG |
3391 | WITH_RCU_READ_LOCK_GUARD() { |
3392 | if (!migration_in_postcopy()) { | |
3393 | migration_bitmap_sync_precopy(rs); | |
3394 | } | |
56e93d26 | 3395 | |
89ac5a1d | 3396 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 3397 | |
89ac5a1d | 3398 | /* try transferring iterative blocks of memory */ |
56e93d26 | 3399 | |
89ac5a1d | 3400 | /* flush all remaining blocks regardless of rate limiting */ |
c13221b5 | 3401 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
3402 | while (true) { |
3403 | int pages; | |
56e93d26 | 3404 | |
05931ec5 | 3405 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3406 | /* no more blocks to sent */ |
3407 | if (pages == 0) { | |
3408 | break; | |
3409 | } | |
3410 | if (pages < 0) { | |
3411 | ret = pages; | |
3412 | break; | |
3413 | } | |
e8f3735f | 3414 | } |
c13221b5 | 3415 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 3416 | |
89ac5a1d DDAG |
3417 | flush_compressed_data(rs); |
3418 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); | |
3419 | } | |
d09a6fde | 3420 | |
33d70973 LB |
3421 | if (ret < 0) { |
3422 | return ret; | |
3d4095b2 | 3423 | } |
56e93d26 | 3424 | |
7f401b80 | 3425 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); |
33d70973 LB |
3426 | if (ret < 0) { |
3427 | return ret; | |
3428 | } | |
3429 | ||
3430 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
3431 | qemu_fflush(f); | |
3432 | ||
3433 | return 0; | |
56e93d26 JQ |
3434 | } |
3435 | ||
fd70385d | 3436 | static void ram_state_pending_estimate(void *opaque, |
c8df4a7a JQ |
3437 | uint64_t *res_precopy_only, |
3438 | uint64_t *res_compatible, | |
3439 | uint64_t *res_postcopy_only) | |
56e93d26 | 3440 | { |
53518d94 JQ |
3441 | RAMState **temp = opaque; |
3442 | RAMState *rs = *temp; | |
56e93d26 | 3443 | |
c8df4a7a | 3444 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3445 | |
c8df4a7a JQ |
3446 | if (migrate_postcopy_ram()) { |
3447 | /* We can do postcopy, and all the data is postcopiable */ | |
3448 | *res_postcopy_only += remaining_size; | |
3449 | } else { | |
3450 | *res_precopy_only += remaining_size; | |
3451 | } | |
3452 | } | |
3453 | ||
fd70385d | 3454 | static void ram_state_pending_exact(void *opaque, |
c8df4a7a JQ |
3455 | uint64_t *res_precopy_only, |
3456 | uint64_t *res_compatible, | |
3457 | uint64_t *res_postcopy_only) | |
3458 | { | |
3459 | RAMState **temp = opaque; | |
3460 | RAMState *rs = *temp; | |
3461 | ||
3462 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; | |
3463 | ||
3464 | if (!migration_in_postcopy()) { | |
56e93d26 | 3465 | qemu_mutex_lock_iothread(); |
89ac5a1d DDAG |
3466 | WITH_RCU_READ_LOCK_GUARD() { |
3467 | migration_bitmap_sync_precopy(rs); | |
3468 | } | |
56e93d26 | 3469 | qemu_mutex_unlock_iothread(); |
9edabd4d | 3470 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3471 | } |
c31b098f | 3472 | |
86e1167e VSO |
3473 | if (migrate_postcopy_ram()) { |
3474 | /* We can do postcopy, and all the data is postcopiable */ | |
47995026 | 3475 | *res_compatible += remaining_size; |
86e1167e | 3476 | } else { |
47995026 | 3477 | *res_precopy_only += remaining_size; |
86e1167e | 3478 | } |
56e93d26 JQ |
3479 | } |
3480 | ||
3481 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
3482 | { | |
3483 | unsigned int xh_len; | |
3484 | int xh_flags; | |
063e760a | 3485 | uint8_t *loaded_data; |
56e93d26 | 3486 | |
56e93d26 JQ |
3487 | /* extract RLE header */ |
3488 | xh_flags = qemu_get_byte(f); | |
3489 | xh_len = qemu_get_be16(f); | |
3490 | ||
3491 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
3492 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
3493 | return -1; | |
3494 | } | |
3495 | ||
3496 | if (xh_len > TARGET_PAGE_SIZE) { | |
3497 | error_report("Failed to load XBZRLE page - len overflow!"); | |
3498 | return -1; | |
3499 | } | |
f265e0e4 | 3500 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 3501 | /* load data and decode */ |
f265e0e4 | 3502 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 3503 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
3504 | |
3505 | /* decode RLE */ | |
063e760a | 3506 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
3507 | TARGET_PAGE_SIZE) == -1) { |
3508 | error_report("Failed to load XBZRLE page - decode error!"); | |
3509 | return -1; | |
3510 | } | |
3511 | ||
3512 | return 0; | |
3513 | } | |
3514 | ||
3d0684b2 JQ |
3515 | /** |
3516 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
3517 | * | |
3518 | * Must be called from within a rcu critical section. | |
3519 | * | |
56e93d26 | 3520 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 3521 | * |
755e8d7c | 3522 | * @mis: the migration incoming state pointer |
3d0684b2 JQ |
3523 | * @f: QEMUFile where to read the data from |
3524 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
c01b16ed | 3525 | * @channel: the channel we're using |
a7180877 | 3526 | */ |
755e8d7c | 3527 | static inline RAMBlock *ram_block_from_stream(MigrationIncomingState *mis, |
c01b16ed PX |
3528 | QEMUFile *f, int flags, |
3529 | int channel) | |
56e93d26 | 3530 | { |
c01b16ed | 3531 | RAMBlock *block = mis->last_recv_block[channel]; |
56e93d26 JQ |
3532 | char id[256]; |
3533 | uint8_t len; | |
3534 | ||
3535 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 3536 | if (!block) { |
56e93d26 JQ |
3537 | error_report("Ack, bad migration stream!"); |
3538 | return NULL; | |
3539 | } | |
4c4bad48 | 3540 | return block; |
56e93d26 JQ |
3541 | } |
3542 | ||
3543 | len = qemu_get_byte(f); | |
3544 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3545 | id[len] = 0; | |
3546 | ||
e3dd7493 | 3547 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
3548 | if (!block) { |
3549 | error_report("Can't find block %s", id); | |
3550 | return NULL; | |
56e93d26 JQ |
3551 | } |
3552 | ||
fbd162e6 | 3553 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
3554 | error_report("block %s should not be migrated !", id); |
3555 | return NULL; | |
3556 | } | |
3557 | ||
c01b16ed | 3558 | mis->last_recv_block[channel] = block; |
755e8d7c | 3559 | |
4c4bad48 HZ |
3560 | return block; |
3561 | } | |
3562 | ||
3563 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
3564 | ram_addr_t offset) | |
3565 | { | |
3566 | if (!offset_in_ramblock(block, offset)) { | |
3567 | return NULL; | |
3568 | } | |
3569 | ||
3570 | return block->host + offset; | |
56e93d26 JQ |
3571 | } |
3572 | ||
6a23f639 DH |
3573 | static void *host_page_from_ram_block_offset(RAMBlock *block, |
3574 | ram_addr_t offset) | |
3575 | { | |
3576 | /* Note: Explicitly no check against offset_in_ramblock(). */ | |
3577 | return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block->host + offset), | |
3578 | block->page_size); | |
3579 | } | |
3580 | ||
3581 | static ram_addr_t host_page_offset_from_ram_block_offset(RAMBlock *block, | |
3582 | ram_addr_t offset) | |
3583 | { | |
3584 | return ((uintptr_t)block->host + offset) & (block->page_size - 1); | |
3585 | } | |
3586 | ||
13af18f2 | 3587 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 3588 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
3589 | { |
3590 | if (!offset_in_ramblock(block, offset)) { | |
3591 | return NULL; | |
3592 | } | |
3593 | if (!block->colo_cache) { | |
3594 | error_report("%s: colo_cache is NULL in block :%s", | |
3595 | __func__, block->idstr); | |
3596 | return NULL; | |
3597 | } | |
7d9acafa ZC |
3598 | |
3599 | /* | |
3600 | * During colo checkpoint, we need bitmap of these migrated pages. | |
3601 | * It help us to decide which pages in ram cache should be flushed | |
3602 | * into VM's RAM later. | |
3603 | */ | |
8af66371 HZ |
3604 | if (record_bitmap && |
3605 | !test_and_set_bit(offset >> TARGET_PAGE_BITS, block->bmap)) { | |
7d9acafa ZC |
3606 | ram_state->migration_dirty_pages++; |
3607 | } | |
13af18f2 ZC |
3608 | return block->colo_cache + offset; |
3609 | } | |
3610 | ||
3d0684b2 JQ |
3611 | /** |
3612 | * ram_handle_compressed: handle the zero page case | |
3613 | * | |
56e93d26 JQ |
3614 | * If a page (or a whole RDMA chunk) has been |
3615 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
3616 | * |
3617 | * @host: host address for the zero page | |
3618 | * @ch: what the page is filled from. We only support zero | |
3619 | * @size: size of the zero page | |
56e93d26 JQ |
3620 | */ |
3621 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
3622 | { | |
bad452a7 | 3623 | if (ch != 0 || !buffer_is_zero(host, size)) { |
56e93d26 JQ |
3624 | memset(host, ch, size); |
3625 | } | |
3626 | } | |
3627 | ||
797ca154 XG |
3628 | /* return the size after decompression, or negative value on error */ |
3629 | static int | |
3630 | qemu_uncompress_data(z_stream *stream, uint8_t *dest, size_t dest_len, | |
3631 | const uint8_t *source, size_t source_len) | |
3632 | { | |
3633 | int err; | |
3634 | ||
3635 | err = inflateReset(stream); | |
3636 | if (err != Z_OK) { | |
3637 | return -1; | |
3638 | } | |
3639 | ||
3640 | stream->avail_in = source_len; | |
3641 | stream->next_in = (uint8_t *)source; | |
3642 | stream->avail_out = dest_len; | |
3643 | stream->next_out = dest; | |
3644 | ||
3645 | err = inflate(stream, Z_NO_FLUSH); | |
3646 | if (err != Z_STREAM_END) { | |
3647 | return -1; | |
3648 | } | |
3649 | ||
3650 | return stream->total_out; | |
3651 | } | |
3652 | ||
56e93d26 JQ |
3653 | static void *do_data_decompress(void *opaque) |
3654 | { | |
3655 | DecompressParam *param = opaque; | |
3656 | unsigned long pagesize; | |
33d151f4 | 3657 | uint8_t *des; |
34ab9e97 | 3658 | int len, ret; |
56e93d26 | 3659 | |
33d151f4 | 3660 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 3661 | while (!param->quit) { |
33d151f4 LL |
3662 | if (param->des) { |
3663 | des = param->des; | |
3664 | len = param->len; | |
3665 | param->des = 0; | |
3666 | qemu_mutex_unlock(¶m->mutex); | |
3667 | ||
56e93d26 | 3668 | pagesize = TARGET_PAGE_SIZE; |
34ab9e97 XG |
3669 | |
3670 | ret = qemu_uncompress_data(¶m->stream, des, pagesize, | |
3671 | param->compbuf, len); | |
f548222c | 3672 | if (ret < 0 && migrate_get_current()->decompress_error_check) { |
34ab9e97 XG |
3673 | error_report("decompress data failed"); |
3674 | qemu_file_set_error(decomp_file, ret); | |
3675 | } | |
73a8912b | 3676 | |
33d151f4 LL |
3677 | qemu_mutex_lock(&decomp_done_lock); |
3678 | param->done = true; | |
3679 | qemu_cond_signal(&decomp_done_cond); | |
3680 | qemu_mutex_unlock(&decomp_done_lock); | |
3681 | ||
3682 | qemu_mutex_lock(¶m->mutex); | |
3683 | } else { | |
3684 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
3685 | } | |
56e93d26 | 3686 | } |
33d151f4 | 3687 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
3688 | |
3689 | return NULL; | |
3690 | } | |
3691 | ||
34ab9e97 | 3692 | static int wait_for_decompress_done(void) |
5533b2e9 LL |
3693 | { |
3694 | int idx, thread_count; | |
3695 | ||
3696 | if (!migrate_use_compression()) { | |
34ab9e97 | 3697 | return 0; |
5533b2e9 LL |
3698 | } |
3699 | ||
3700 | thread_count = migrate_decompress_threads(); | |
3701 | qemu_mutex_lock(&decomp_done_lock); | |
3702 | for (idx = 0; idx < thread_count; idx++) { | |
3703 | while (!decomp_param[idx].done) { | |
3704 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
3705 | } | |
3706 | } | |
3707 | qemu_mutex_unlock(&decomp_done_lock); | |
34ab9e97 | 3708 | return qemu_file_get_error(decomp_file); |
5533b2e9 LL |
3709 | } |
3710 | ||
f0afa331 | 3711 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
3712 | { |
3713 | int i, thread_count; | |
3714 | ||
3416ab5b JQ |
3715 | if (!migrate_use_compression()) { |
3716 | return; | |
3717 | } | |
56e93d26 JQ |
3718 | thread_count = migrate_decompress_threads(); |
3719 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3720 | /* |
3721 | * we use it as a indicator which shows if the thread is | |
3722 | * properly init'd or not | |
3723 | */ | |
3724 | if (!decomp_param[i].compbuf) { | |
3725 | break; | |
3726 | } | |
3727 | ||
56e93d26 | 3728 | qemu_mutex_lock(&decomp_param[i].mutex); |
90e56fb4 | 3729 | decomp_param[i].quit = true; |
56e93d26 JQ |
3730 | qemu_cond_signal(&decomp_param[i].cond); |
3731 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
3732 | } | |
3733 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3734 | if (!decomp_param[i].compbuf) { |
3735 | break; | |
3736 | } | |
3737 | ||
56e93d26 JQ |
3738 | qemu_thread_join(decompress_threads + i); |
3739 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
3740 | qemu_cond_destroy(&decomp_param[i].cond); | |
797ca154 | 3741 | inflateEnd(&decomp_param[i].stream); |
56e93d26 | 3742 | g_free(decomp_param[i].compbuf); |
797ca154 | 3743 | decomp_param[i].compbuf = NULL; |
56e93d26 JQ |
3744 | } |
3745 | g_free(decompress_threads); | |
3746 | g_free(decomp_param); | |
56e93d26 JQ |
3747 | decompress_threads = NULL; |
3748 | decomp_param = NULL; | |
34ab9e97 | 3749 | decomp_file = NULL; |
56e93d26 JQ |
3750 | } |
3751 | ||
34ab9e97 | 3752 | static int compress_threads_load_setup(QEMUFile *f) |
797ca154 XG |
3753 | { |
3754 | int i, thread_count; | |
3755 | ||
3756 | if (!migrate_use_compression()) { | |
3757 | return 0; | |
3758 | } | |
3759 | ||
3760 | thread_count = migrate_decompress_threads(); | |
3761 | decompress_threads = g_new0(QemuThread, thread_count); | |
3762 | decomp_param = g_new0(DecompressParam, thread_count); | |
3763 | qemu_mutex_init(&decomp_done_lock); | |
3764 | qemu_cond_init(&decomp_done_cond); | |
34ab9e97 | 3765 | decomp_file = f; |
797ca154 XG |
3766 | for (i = 0; i < thread_count; i++) { |
3767 | if (inflateInit(&decomp_param[i].stream) != Z_OK) { | |
3768 | goto exit; | |
3769 | } | |
3770 | ||
3771 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
3772 | qemu_mutex_init(&decomp_param[i].mutex); | |
3773 | qemu_cond_init(&decomp_param[i].cond); | |
3774 | decomp_param[i].done = true; | |
3775 | decomp_param[i].quit = false; | |
3776 | qemu_thread_create(decompress_threads + i, "decompress", | |
3777 | do_data_decompress, decomp_param + i, | |
3778 | QEMU_THREAD_JOINABLE); | |
3779 | } | |
3780 | return 0; | |
3781 | exit: | |
3782 | compress_threads_load_cleanup(); | |
3783 | return -1; | |
3784 | } | |
3785 | ||
c1bc6626 | 3786 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
3787 | void *host, int len) |
3788 | { | |
3789 | int idx, thread_count; | |
3790 | ||
3791 | thread_count = migrate_decompress_threads(); | |
37396950 | 3792 | QEMU_LOCK_GUARD(&decomp_done_lock); |
56e93d26 JQ |
3793 | while (true) { |
3794 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 3795 | if (decomp_param[idx].done) { |
33d151f4 LL |
3796 | decomp_param[idx].done = false; |
3797 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 3798 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
3799 | decomp_param[idx].des = host; |
3800 | decomp_param[idx].len = len; | |
33d151f4 LL |
3801 | qemu_cond_signal(&decomp_param[idx].cond); |
3802 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
3803 | break; |
3804 | } | |
3805 | } | |
3806 | if (idx < thread_count) { | |
3807 | break; | |
73a8912b LL |
3808 | } else { |
3809 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
3810 | } |
3811 | } | |
3812 | } | |
3813 | ||
b70cb3b4 RL |
3814 | static void colo_init_ram_state(void) |
3815 | { | |
3816 | ram_state_init(&ram_state); | |
b70cb3b4 RL |
3817 | } |
3818 | ||
13af18f2 ZC |
3819 | /* |
3820 | * colo cache: this is for secondary VM, we cache the whole | |
3821 | * memory of the secondary VM, it is need to hold the global lock | |
3822 | * to call this helper. | |
3823 | */ | |
3824 | int colo_init_ram_cache(void) | |
3825 | { | |
3826 | RAMBlock *block; | |
3827 | ||
44901b5a PB |
3828 | WITH_RCU_READ_LOCK_GUARD() { |
3829 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3830 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
8dbe22c6 | 3831 | NULL, false, false); |
44901b5a PB |
3832 | if (!block->colo_cache) { |
3833 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
3834 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
3835 | block->used_length); | |
3836 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3837 | if (block->colo_cache) { | |
3838 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3839 | block->colo_cache = NULL; | |
3840 | } | |
89ac5a1d | 3841 | } |
44901b5a | 3842 | return -errno; |
89ac5a1d | 3843 | } |
e5fdf920 LS |
3844 | if (!machine_dump_guest_core(current_machine)) { |
3845 | qemu_madvise(block->colo_cache, block->used_length, | |
3846 | QEMU_MADV_DONTDUMP); | |
3847 | } | |
13af18f2 | 3848 | } |
13af18f2 | 3849 | } |
44901b5a | 3850 | |
7d9acafa ZC |
3851 | /* |
3852 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3853 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3854 | * we use the same name 'ram_bitmap' as for migration. | |
3855 | */ | |
3856 | if (ram_bytes_total()) { | |
3857 | RAMBlock *block; | |
3858 | ||
fbd162e6 | 3859 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3860 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3861 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3862 | } |
3863 | } | |
7d9acafa | 3864 | |
b70cb3b4 | 3865 | colo_init_ram_state(); |
13af18f2 | 3866 | return 0; |
13af18f2 ZC |
3867 | } |
3868 | ||
0393031a HZ |
3869 | /* TODO: duplicated with ram_init_bitmaps */ |
3870 | void colo_incoming_start_dirty_log(void) | |
3871 | { | |
3872 | RAMBlock *block = NULL; | |
3873 | /* For memory_global_dirty_log_start below. */ | |
3874 | qemu_mutex_lock_iothread(); | |
3875 | qemu_mutex_lock_ramlist(); | |
3876 | ||
3877 | memory_global_dirty_log_sync(); | |
3878 | WITH_RCU_READ_LOCK_GUARD() { | |
3879 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3880 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3881 | /* Discard this dirty bitmap record */ | |
3882 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3883 | } | |
63b41db4 | 3884 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
0393031a HZ |
3885 | } |
3886 | ram_state->migration_dirty_pages = 0; | |
3887 | qemu_mutex_unlock_ramlist(); | |
3888 | qemu_mutex_unlock_iothread(); | |
3889 | } | |
3890 | ||
13af18f2 ZC |
3891 | /* It is need to hold the global lock to call this helper */ |
3892 | void colo_release_ram_cache(void) | |
3893 | { | |
3894 | RAMBlock *block; | |
3895 | ||
63b41db4 | 3896 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); |
fbd162e6 | 3897 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3898 | g_free(block->bmap); |
3899 | block->bmap = NULL; | |
3900 | } | |
3901 | ||
89ac5a1d DDAG |
3902 | WITH_RCU_READ_LOCK_GUARD() { |
3903 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3904 | if (block->colo_cache) { | |
3905 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3906 | block->colo_cache = NULL; | |
3907 | } | |
13af18f2 ZC |
3908 | } |
3909 | } | |
0393031a | 3910 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
3911 | } |
3912 | ||
f265e0e4 JQ |
3913 | /** |
3914 | * ram_load_setup: Setup RAM for migration incoming side | |
3915 | * | |
3916 | * Returns zero to indicate success and negative for error | |
3917 | * | |
3918 | * @f: QEMUFile where to receive the data | |
3919 | * @opaque: RAMState pointer | |
3920 | */ | |
3921 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
3922 | { | |
34ab9e97 | 3923 | if (compress_threads_load_setup(f)) { |
797ca154 XG |
3924 | return -1; |
3925 | } | |
3926 | ||
f265e0e4 | 3927 | xbzrle_load_setup(); |
f9494614 | 3928 | ramblock_recv_map_init(); |
13af18f2 | 3929 | |
f265e0e4 JQ |
3930 | return 0; |
3931 | } | |
3932 | ||
3933 | static int ram_load_cleanup(void *opaque) | |
3934 | { | |
f9494614 | 3935 | RAMBlock *rb; |
56eb90af | 3936 | |
fbd162e6 | 3937 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 3938 | qemu_ram_block_writeback(rb); |
56eb90af JH |
3939 | } |
3940 | ||
f265e0e4 | 3941 | xbzrle_load_cleanup(); |
f0afa331 | 3942 | compress_threads_load_cleanup(); |
f9494614 | 3943 | |
fbd162e6 | 3944 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
3945 | g_free(rb->receivedmap); |
3946 | rb->receivedmap = NULL; | |
3947 | } | |
13af18f2 | 3948 | |
f265e0e4 JQ |
3949 | return 0; |
3950 | } | |
3951 | ||
3d0684b2 JQ |
3952 | /** |
3953 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
3954 | * | |
3955 | * Returns 0 for success and negative if there was one error | |
3956 | * | |
3957 | * @mis: current migration incoming state | |
3958 | * | |
3959 | * Allocate data structures etc needed by incoming migration with | |
3960 | * postcopy-ram. postcopy-ram's similarly names | |
3961 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
3962 | */ |
3963 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
3964 | { | |
c136180c | 3965 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
3966 | } |
3967 | ||
3d0684b2 JQ |
3968 | /** |
3969 | * ram_load_postcopy: load a page in postcopy case | |
3970 | * | |
3971 | * Returns 0 for success or -errno in case of error | |
3972 | * | |
a7180877 DDAG |
3973 | * Called in postcopy mode by ram_load(). |
3974 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
3975 | * |
3976 | * @f: QEMUFile where to send the data | |
36f62f11 | 3977 | * @channel: the channel to use for loading |
a7180877 | 3978 | */ |
36f62f11 | 3979 | int ram_load_postcopy(QEMUFile *f, int channel) |
a7180877 DDAG |
3980 | { |
3981 | int flags = 0, ret = 0; | |
3982 | bool place_needed = false; | |
1aa83678 | 3983 | bool matches_target_page_size = false; |
a7180877 | 3984 | MigrationIncomingState *mis = migration_incoming_get_current(); |
36f62f11 | 3985 | PostcopyTmpPage *tmp_page = &mis->postcopy_tmp_pages[channel]; |
a7180877 DDAG |
3986 | |
3987 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
3988 | ram_addr_t addr; | |
a7180877 DDAG |
3989 | void *page_buffer = NULL; |
3990 | void *place_source = NULL; | |
df9ff5e1 | 3991 | RAMBlock *block = NULL; |
a7180877 | 3992 | uint8_t ch; |
644acf99 | 3993 | int len; |
a7180877 DDAG |
3994 | |
3995 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
3996 | |
3997 | /* | |
3998 | * If qemu file error, we should stop here, and then "addr" | |
3999 | * may be invalid | |
4000 | */ | |
4001 | ret = qemu_file_get_error(f); | |
4002 | if (ret) { | |
4003 | break; | |
4004 | } | |
4005 | ||
a7180877 DDAG |
4006 | flags = addr & ~TARGET_PAGE_MASK; |
4007 | addr &= TARGET_PAGE_MASK; | |
4008 | ||
36f62f11 | 4009 | trace_ram_load_postcopy_loop(channel, (uint64_t)addr, flags); |
644acf99 WY |
4010 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
4011 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
c01b16ed | 4012 | block = ram_block_from_stream(mis, f, flags, channel); |
6a23f639 DH |
4013 | if (!block) { |
4014 | ret = -EINVAL; | |
4015 | break; | |
4016 | } | |
4c4bad48 | 4017 | |
898ba906 DH |
4018 | /* |
4019 | * Relying on used_length is racy and can result in false positives. | |
4020 | * We might place pages beyond used_length in case RAM was shrunk | |
4021 | * while in postcopy, which is fine - trying to place via | |
4022 | * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault. | |
4023 | */ | |
4024 | if (!block->host || addr >= block->postcopy_length) { | |
a7180877 DDAG |
4025 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); |
4026 | ret = -EINVAL; | |
4027 | break; | |
4028 | } | |
77dadc3f | 4029 | tmp_page->target_pages++; |
1aa83678 | 4030 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 4031 | /* |
28abd200 DDAG |
4032 | * Postcopy requires that we place whole host pages atomically; |
4033 | * these may be huge pages for RAMBlocks that are backed by | |
4034 | * hugetlbfs. | |
a7180877 DDAG |
4035 | * To make it atomic, the data is read into a temporary page |
4036 | * that's moved into place later. | |
4037 | * The migration protocol uses, possibly smaller, target-pages | |
4038 | * however the source ensures it always sends all the components | |
91ba442f | 4039 | * of a host page in one chunk. |
a7180877 | 4040 | */ |
77dadc3f | 4041 | page_buffer = tmp_page->tmp_huge_page + |
6a23f639 DH |
4042 | host_page_offset_from_ram_block_offset(block, addr); |
4043 | /* If all TP are zero then we can optimise the place */ | |
77dadc3f PX |
4044 | if (tmp_page->target_pages == 1) { |
4045 | tmp_page->host_addr = | |
4046 | host_page_from_ram_block_offset(block, addr); | |
4047 | } else if (tmp_page->host_addr != | |
4048 | host_page_from_ram_block_offset(block, addr)) { | |
c53b7ddc | 4049 | /* not the 1st TP within the HP */ |
36f62f11 | 4050 | error_report("Non-same host page detected on channel %d: " |
cfc7dc8a PX |
4051 | "Target host page %p, received host page %p " |
4052 | "(rb %s offset 0x"RAM_ADDR_FMT" target_pages %d)", | |
36f62f11 | 4053 | channel, tmp_page->host_addr, |
cfc7dc8a PX |
4054 | host_page_from_ram_block_offset(block, addr), |
4055 | block->idstr, addr, tmp_page->target_pages); | |
6a23f639 DH |
4056 | ret = -EINVAL; |
4057 | break; | |
a7180877 DDAG |
4058 | } |
4059 | ||
4060 | /* | |
4061 | * If it's the last part of a host page then we place the host | |
4062 | * page | |
4063 | */ | |
77dadc3f PX |
4064 | if (tmp_page->target_pages == |
4065 | (block->page_size / TARGET_PAGE_SIZE)) { | |
4cbb3c63 | 4066 | place_needed = true; |
4cbb3c63 | 4067 | } |
77dadc3f | 4068 | place_source = tmp_page->tmp_huge_page; |
a7180877 DDAG |
4069 | } |
4070 | ||
4071 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 4072 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 4073 | ch = qemu_get_byte(f); |
2e36bc1b WY |
4074 | /* |
4075 | * Can skip to set page_buffer when | |
4076 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
4077 | */ | |
4078 | if (ch || !matches_target_page_size) { | |
4079 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
4080 | } | |
a7180877 | 4081 | if (ch) { |
77dadc3f | 4082 | tmp_page->all_zero = false; |
a7180877 DDAG |
4083 | } |
4084 | break; | |
4085 | ||
4086 | case RAM_SAVE_FLAG_PAGE: | |
77dadc3f | 4087 | tmp_page->all_zero = false; |
1aa83678 PX |
4088 | if (!matches_target_page_size) { |
4089 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
4090 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
4091 | } else { | |
1aa83678 PX |
4092 | /* |
4093 | * For small pages that matches target page size, we | |
4094 | * avoid the qemu_file copy. Instead we directly use | |
4095 | * the buffer of QEMUFile to place the page. Note: we | |
4096 | * cannot do any QEMUFile operation before using that | |
4097 | * buffer to make sure the buffer is valid when | |
4098 | * placing the page. | |
a7180877 DDAG |
4099 | */ |
4100 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
4101 | TARGET_PAGE_SIZE); | |
4102 | } | |
4103 | break; | |
644acf99 | 4104 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
77dadc3f | 4105 | tmp_page->all_zero = false; |
644acf99 WY |
4106 | len = qemu_get_be32(f); |
4107 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4108 | error_report("Invalid compressed data length: %d", len); | |
4109 | ret = -EINVAL; | |
4110 | break; | |
4111 | } | |
4112 | decompress_data_with_multi_threads(f, page_buffer, len); | |
4113 | break; | |
4114 | ||
a7180877 DDAG |
4115 | case RAM_SAVE_FLAG_EOS: |
4116 | /* normal exit */ | |
6df264ac | 4117 | multifd_recv_sync_main(); |
a7180877 DDAG |
4118 | break; |
4119 | default: | |
29fccade | 4120 | error_report("Unknown combination of migration flags: 0x%x" |
a7180877 DDAG |
4121 | " (postcopy mode)", flags); |
4122 | ret = -EINVAL; | |
7a9ddfbf PX |
4123 | break; |
4124 | } | |
4125 | ||
644acf99 WY |
4126 | /* Got the whole host page, wait for decompress before placing. */ |
4127 | if (place_needed) { | |
4128 | ret |= wait_for_decompress_done(); | |
4129 | } | |
4130 | ||
7a9ddfbf PX |
4131 | /* Detect for any possible file errors */ |
4132 | if (!ret && qemu_file_get_error(f)) { | |
4133 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
4134 | } |
4135 | ||
7a9ddfbf | 4136 | if (!ret && place_needed) { |
77dadc3f PX |
4137 | if (tmp_page->all_zero) { |
4138 | ret = postcopy_place_page_zero(mis, tmp_page->host_addr, block); | |
a7180877 | 4139 | } else { |
77dadc3f PX |
4140 | ret = postcopy_place_page(mis, tmp_page->host_addr, |
4141 | place_source, block); | |
a7180877 | 4142 | } |
ddf35bdf | 4143 | place_needed = false; |
77dadc3f | 4144 | postcopy_temp_page_reset(tmp_page); |
a7180877 | 4145 | } |
a7180877 DDAG |
4146 | } |
4147 | ||
4148 | return ret; | |
4149 | } | |
4150 | ||
acab30b8 DHB |
4151 | static bool postcopy_is_running(void) |
4152 | { | |
4153 | PostcopyState ps = postcopy_state_get(); | |
4154 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
4155 | } | |
4156 | ||
e6f4aa18 ZC |
4157 | /* |
4158 | * Flush content of RAM cache into SVM's memory. | |
4159 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
4160 | */ | |
24fa16f8 | 4161 | void colo_flush_ram_cache(void) |
e6f4aa18 ZC |
4162 | { |
4163 | RAMBlock *block = NULL; | |
4164 | void *dst_host; | |
4165 | void *src_host; | |
4166 | unsigned long offset = 0; | |
4167 | ||
d1955d22 | 4168 | memory_global_dirty_log_sync(); |
89ac5a1d DDAG |
4169 | WITH_RCU_READ_LOCK_GUARD() { |
4170 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
4171 | ramblock_sync_dirty_bitmap(ram_state, block); | |
4172 | } | |
d1955d22 | 4173 | } |
d1955d22 | 4174 | |
e6f4aa18 | 4175 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
4176 | WITH_RCU_READ_LOCK_GUARD() { |
4177 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 4178 | |
89ac5a1d | 4179 | while (block) { |
a6a83cef | 4180 | unsigned long num = 0; |
e6f4aa18 | 4181 | |
a6a83cef | 4182 | offset = colo_bitmap_find_dirty(ram_state, block, offset, &num); |
542147f4 DH |
4183 | if (!offset_in_ramblock(block, |
4184 | ((ram_addr_t)offset) << TARGET_PAGE_BITS)) { | |
89ac5a1d | 4185 | offset = 0; |
a6a83cef | 4186 | num = 0; |
89ac5a1d DDAG |
4187 | block = QLIST_NEXT_RCU(block, next); |
4188 | } else { | |
a6a83cef RL |
4189 | unsigned long i = 0; |
4190 | ||
4191 | for (i = 0; i < num; i++) { | |
4192 | migration_bitmap_clear_dirty(ram_state, block, offset + i); | |
4193 | } | |
8bba004c AR |
4194 | dst_host = block->host |
4195 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
4196 | src_host = block->colo_cache | |
4197 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
a6a83cef RL |
4198 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE * num); |
4199 | offset += num; | |
89ac5a1d | 4200 | } |
e6f4aa18 ZC |
4201 | } |
4202 | } | |
e6f4aa18 ZC |
4203 | trace_colo_flush_ram_cache_end(); |
4204 | } | |
4205 | ||
10da4a36 WY |
4206 | /** |
4207 | * ram_load_precopy: load pages in precopy case | |
4208 | * | |
4209 | * Returns 0 for success or -errno in case of error | |
4210 | * | |
4211 | * Called in precopy mode by ram_load(). | |
4212 | * rcu_read_lock is taken prior to this being called. | |
4213 | * | |
4214 | * @f: QEMUFile where to send the data | |
4215 | */ | |
4216 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 4217 | { |
755e8d7c | 4218 | MigrationIncomingState *mis = migration_incoming_get_current(); |
e65cec5e | 4219 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 4220 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
80fe315c | 4221 | bool postcopy_advised = migration_incoming_postcopy_advised(); |
edc60127 JQ |
4222 | if (!migrate_use_compression()) { |
4223 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
4224 | } | |
a7180877 | 4225 | |
10da4a36 | 4226 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 4227 | ram_addr_t addr, total_ram_bytes; |
0393031a | 4228 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
4229 | uint8_t ch; |
4230 | ||
e65cec5e YK |
4231 | /* |
4232 | * Yield periodically to let main loop run, but an iteration of | |
4233 | * the main loop is expensive, so do it each some iterations | |
4234 | */ | |
4235 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
4236 | aio_co_schedule(qemu_get_current_aio_context(), | |
4237 | qemu_coroutine_self()); | |
4238 | qemu_coroutine_yield(); | |
4239 | } | |
4240 | i++; | |
4241 | ||
56e93d26 JQ |
4242 | addr = qemu_get_be64(f); |
4243 | flags = addr & ~TARGET_PAGE_MASK; | |
4244 | addr &= TARGET_PAGE_MASK; | |
4245 | ||
edc60127 JQ |
4246 | if (flags & invalid_flags) { |
4247 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
4248 | error_report("Received an unexpected compressed page"); | |
4249 | } | |
4250 | ||
4251 | ret = -EINVAL; | |
4252 | break; | |
4253 | } | |
4254 | ||
bb890ed5 | 4255 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 4256 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
c01b16ed PX |
4257 | RAMBlock *block = ram_block_from_stream(mis, f, flags, |
4258 | RAM_CHANNEL_PRECOPY); | |
4c4bad48 | 4259 | |
0393031a | 4260 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 4261 | /* |
0393031a HZ |
4262 | * After going into COLO stage, we should not load the page |
4263 | * into SVM's memory directly, we put them into colo_cache firstly. | |
4264 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
4265 | * Previously, we copied all these memory in preparing stage of COLO | |
4266 | * while we need to stop VM, which is a time-consuming process. | |
4267 | * Here we optimize it by a trick, back-up every page while in | |
4268 | * migration process while COLO is enabled, though it affects the | |
4269 | * speed of the migration, but it obviously reduce the downtime of | |
4270 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 4271 | */ |
0393031a HZ |
4272 | if (migration_incoming_colo_enabled()) { |
4273 | if (migration_incoming_in_colo_state()) { | |
4274 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 4275 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
4276 | } else { |
4277 | /* | |
4278 | * In migration stage but before COLO stage, | |
4279 | * Put all pages into both cache and SVM's memory. | |
4280 | */ | |
8af66371 | 4281 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 4282 | } |
13af18f2 | 4283 | } |
a776aa15 DDAG |
4284 | if (!host) { |
4285 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
4286 | ret = -EINVAL; | |
4287 | break; | |
4288 | } | |
13af18f2 ZC |
4289 | if (!migration_incoming_in_colo_state()) { |
4290 | ramblock_recv_bitmap_set(block, host); | |
4291 | } | |
4292 | ||
1db9d8e5 | 4293 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
4294 | } |
4295 | ||
56e93d26 JQ |
4296 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
4297 | case RAM_SAVE_FLAG_MEM_SIZE: | |
4298 | /* Synchronize RAM block list */ | |
4299 | total_ram_bytes = addr; | |
4300 | while (!ret && total_ram_bytes) { | |
4301 | RAMBlock *block; | |
56e93d26 JQ |
4302 | char id[256]; |
4303 | ram_addr_t length; | |
4304 | ||
4305 | len = qemu_get_byte(f); | |
4306 | qemu_get_buffer(f, (uint8_t *)id, len); | |
4307 | id[len] = 0; | |
4308 | length = qemu_get_be64(f); | |
4309 | ||
e3dd7493 | 4310 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
4311 | if (block && !qemu_ram_is_migratable(block)) { |
4312 | error_report("block %s should not be migrated !", id); | |
4313 | ret = -EINVAL; | |
4314 | } else if (block) { | |
e3dd7493 DDAG |
4315 | if (length != block->used_length) { |
4316 | Error *local_err = NULL; | |
56e93d26 | 4317 | |
fa53a0e5 | 4318 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
4319 | &local_err); |
4320 | if (local_err) { | |
4321 | error_report_err(local_err); | |
56e93d26 | 4322 | } |
56e93d26 | 4323 | } |
ef08fb38 | 4324 | /* For postcopy we need to check hugepage sizes match */ |
e846b746 | 4325 | if (postcopy_advised && migrate_postcopy_ram() && |
ef08fb38 DDAG |
4326 | block->page_size != qemu_host_page_size) { |
4327 | uint64_t remote_page_size = qemu_get_be64(f); | |
4328 | if (remote_page_size != block->page_size) { | |
4329 | error_report("Mismatched RAM page size %s " | |
4330 | "(local) %zd != %" PRId64, | |
4331 | id, block->page_size, | |
4332 | remote_page_size); | |
4333 | ret = -EINVAL; | |
4334 | } | |
4335 | } | |
fbd162e6 YK |
4336 | if (migrate_ignore_shared()) { |
4337 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
4338 | if (ramblock_is_ignored(block) && |
4339 | block->mr->addr != addr) { | |
4340 | error_report("Mismatched GPAs for block %s " | |
4341 | "%" PRId64 "!= %" PRId64, | |
4342 | id, (uint64_t)addr, | |
4343 | (uint64_t)block->mr->addr); | |
4344 | ret = -EINVAL; | |
4345 | } | |
4346 | } | |
e3dd7493 DDAG |
4347 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
4348 | block->idstr); | |
4349 | } else { | |
56e93d26 JQ |
4350 | error_report("Unknown ramblock \"%s\", cannot " |
4351 | "accept migration", id); | |
4352 | ret = -EINVAL; | |
4353 | } | |
4354 | ||
4355 | total_ram_bytes -= length; | |
4356 | } | |
4357 | break; | |
a776aa15 | 4358 | |
bb890ed5 | 4359 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
4360 | ch = qemu_get_byte(f); |
4361 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
4362 | break; | |
a776aa15 | 4363 | |
56e93d26 | 4364 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
4365 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
4366 | break; | |
56e93d26 | 4367 | |
a776aa15 | 4368 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
4369 | len = qemu_get_be32(f); |
4370 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4371 | error_report("Invalid compressed data length: %d", len); | |
4372 | ret = -EINVAL; | |
4373 | break; | |
4374 | } | |
c1bc6626 | 4375 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 4376 | break; |
a776aa15 | 4377 | |
56e93d26 | 4378 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
4379 | if (load_xbzrle(f, addr, host) < 0) { |
4380 | error_report("Failed to decompress XBZRLE page at " | |
4381 | RAM_ADDR_FMT, addr); | |
4382 | ret = -EINVAL; | |
4383 | break; | |
4384 | } | |
4385 | break; | |
4386 | case RAM_SAVE_FLAG_EOS: | |
4387 | /* normal exit */ | |
6df264ac | 4388 | multifd_recv_sync_main(); |
56e93d26 JQ |
4389 | break; |
4390 | default: | |
4391 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 4392 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 | 4393 | } else { |
29fccade | 4394 | error_report("Unknown combination of migration flags: 0x%x", |
56e93d26 JQ |
4395 | flags); |
4396 | ret = -EINVAL; | |
4397 | } | |
4398 | } | |
4399 | if (!ret) { | |
4400 | ret = qemu_file_get_error(f); | |
4401 | } | |
0393031a HZ |
4402 | if (!ret && host_bak) { |
4403 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
4404 | } | |
56e93d26 JQ |
4405 | } |
4406 | ||
ca1a6b70 | 4407 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
4408 | return ret; |
4409 | } | |
4410 | ||
4411 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
4412 | { | |
4413 | int ret = 0; | |
4414 | static uint64_t seq_iter; | |
4415 | /* | |
4416 | * If system is running in postcopy mode, page inserts to host memory must | |
4417 | * be atomic | |
4418 | */ | |
4419 | bool postcopy_running = postcopy_is_running(); | |
4420 | ||
4421 | seq_iter++; | |
4422 | ||
4423 | if (version_id != 4) { | |
4424 | return -EINVAL; | |
4425 | } | |
4426 | ||
4427 | /* | |
4428 | * This RCU critical section can be very long running. | |
4429 | * When RCU reclaims in the code start to become numerous, | |
4430 | * it will be necessary to reduce the granularity of this | |
4431 | * critical section. | |
4432 | */ | |
89ac5a1d DDAG |
4433 | WITH_RCU_READ_LOCK_GUARD() { |
4434 | if (postcopy_running) { | |
36f62f11 PX |
4435 | /* |
4436 | * Note! Here RAM_CHANNEL_PRECOPY is the precopy channel of | |
4437 | * postcopy migration, we have another RAM_CHANNEL_POSTCOPY to | |
4438 | * service fast page faults. | |
4439 | */ | |
4440 | ret = ram_load_postcopy(f, RAM_CHANNEL_PRECOPY); | |
89ac5a1d DDAG |
4441 | } else { |
4442 | ret = ram_load_precopy(f); | |
4443 | } | |
10da4a36 | 4444 | } |
55c4446b | 4445 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 | 4446 | |
56e93d26 JQ |
4447 | return ret; |
4448 | } | |
4449 | ||
c6467627 VSO |
4450 | static bool ram_has_postcopy(void *opaque) |
4451 | { | |
469dd51b | 4452 | RAMBlock *rb; |
fbd162e6 | 4453 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
4454 | if (ramblock_is_pmem(rb)) { |
4455 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
4456 | "is not supported now!", rb->idstr, rb->host); | |
4457 | return false; | |
4458 | } | |
4459 | } | |
4460 | ||
c6467627 VSO |
4461 | return migrate_postcopy_ram(); |
4462 | } | |
4463 | ||
edd090c7 PX |
4464 | /* Sync all the dirty bitmap with destination VM. */ |
4465 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
4466 | { | |
4467 | RAMBlock *block; | |
4468 | QEMUFile *file = s->to_dst_file; | |
4469 | int ramblock_count = 0; | |
4470 | ||
4471 | trace_ram_dirty_bitmap_sync_start(); | |
4472 | ||
fbd162e6 | 4473 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
4474 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
4475 | trace_ram_dirty_bitmap_request(block->idstr); | |
4476 | ramblock_count++; | |
4477 | } | |
4478 | ||
4479 | trace_ram_dirty_bitmap_sync_wait(); | |
4480 | ||
4481 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
4482 | while (ramblock_count--) { | |
4483 | qemu_sem_wait(&s->rp_state.rp_sem); | |
4484 | } | |
4485 | ||
4486 | trace_ram_dirty_bitmap_sync_complete(); | |
4487 | ||
4488 | return 0; | |
4489 | } | |
4490 | ||
4491 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
4492 | { | |
4493 | qemu_sem_post(&s->rp_state.rp_sem); | |
4494 | } | |
4495 | ||
a335debb PX |
4496 | /* |
4497 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
4498 | * This is only used when the postcopy migration is paused but wants | |
4499 | * to resume from a middle point. | |
4500 | */ | |
4501 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
4502 | { | |
4503 | int ret = -EINVAL; | |
43044ac0 | 4504 | /* from_dst_file is always valid because we're within rp_thread */ |
a335debb PX |
4505 | QEMUFile *file = s->rp_state.from_dst_file; |
4506 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 4507 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
4508 | uint64_t size, end_mark; |
4509 | ||
4510 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
4511 | ||
4512 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
4513 | error_report("%s: incorrect state %s", __func__, | |
4514 | MigrationStatus_str(s->state)); | |
4515 | return -EINVAL; | |
4516 | } | |
4517 | ||
4518 | /* | |
4519 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3a4452d8 | 4520 | * need the endianness conversion, and the paddings. |
a335debb PX |
4521 | */ |
4522 | local_size = ROUND_UP(local_size, 8); | |
4523 | ||
4524 | /* Add paddings */ | |
4525 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
4526 | ||
4527 | size = qemu_get_be64(file); | |
4528 | ||
4529 | /* The size of the bitmap should match with our ramblock */ | |
4530 | if (size != local_size) { | |
4531 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
4532 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
4533 | block->idstr, size, local_size); | |
4534 | ret = -EINVAL; | |
4535 | goto out; | |
4536 | } | |
4537 | ||
4538 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
4539 | end_mark = qemu_get_be64(file); | |
4540 | ||
4541 | ret = qemu_file_get_error(file); | |
4542 | if (ret || size != local_size) { | |
4543 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
4544 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
4545 | __func__, block->idstr, ret, local_size, size); | |
4546 | ret = -EIO; | |
4547 | goto out; | |
4548 | } | |
4549 | ||
4550 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
af3bbbe9 | 4551 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64, |
a335debb PX |
4552 | __func__, block->idstr, end_mark); |
4553 | ret = -EINVAL; | |
4554 | goto out; | |
4555 | } | |
4556 | ||
4557 | /* | |
3a4452d8 | 4558 | * Endianness conversion. We are during postcopy (though paused). |
a335debb PX |
4559 | * The dirty bitmap won't change. We can directly modify it. |
4560 | */ | |
4561 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
4562 | ||
4563 | /* | |
4564 | * What we received is "received bitmap". Revert it as the initial | |
4565 | * dirty bitmap for this ramblock. | |
4566 | */ | |
4567 | bitmap_complement(block->bmap, block->bmap, nbits); | |
4568 | ||
be39b4cd DH |
4569 | /* Clear dirty bits of discarded ranges that we don't want to migrate. */ |
4570 | ramblock_dirty_bitmap_clear_discarded_pages(block); | |
4571 | ||
4572 | /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */ | |
a335debb PX |
4573 | trace_ram_dirty_bitmap_reload_complete(block->idstr); |
4574 | ||
edd090c7 PX |
4575 | /* |
4576 | * We succeeded to sync bitmap for current ramblock. If this is | |
4577 | * the last one to sync, we need to notify the main send thread. | |
4578 | */ | |
4579 | ram_dirty_bitmap_reload_notify(s); | |
4580 | ||
a335debb PX |
4581 | ret = 0; |
4582 | out: | |
bf269906 | 4583 | g_free(le_bitmap); |
a335debb PX |
4584 | return ret; |
4585 | } | |
4586 | ||
edd090c7 PX |
4587 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
4588 | { | |
4589 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 4590 | int ret; |
edd090c7 | 4591 | |
08614f34 PX |
4592 | ret = ram_dirty_bitmap_sync_all(s, rs); |
4593 | if (ret) { | |
4594 | return ret; | |
4595 | } | |
4596 | ||
4597 | ram_state_resume_prepare(rs, s->to_dst_file); | |
4598 | ||
4599 | return 0; | |
edd090c7 PX |
4600 | } |
4601 | ||
36f62f11 PX |
4602 | void postcopy_preempt_shutdown_file(MigrationState *s) |
4603 | { | |
4604 | qemu_put_be64(s->postcopy_qemufile_src, RAM_SAVE_FLAG_EOS); | |
4605 | qemu_fflush(s->postcopy_qemufile_src); | |
4606 | } | |
4607 | ||
56e93d26 | 4608 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 4609 | .save_setup = ram_save_setup, |
56e93d26 | 4610 | .save_live_iterate = ram_save_iterate, |
763c906b | 4611 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 4612 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 4613 | .has_postcopy = ram_has_postcopy, |
c8df4a7a JQ |
4614 | .state_pending_exact = ram_state_pending_exact, |
4615 | .state_pending_estimate = ram_state_pending_estimate, | |
56e93d26 | 4616 | .load_state = ram_load, |
f265e0e4 JQ |
4617 | .save_cleanup = ram_save_cleanup, |
4618 | .load_setup = ram_load_setup, | |
4619 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 4620 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
4621 | }; |
4622 | ||
c7c0e724 DH |
4623 | static void ram_mig_ram_block_resized(RAMBlockNotifier *n, void *host, |
4624 | size_t old_size, size_t new_size) | |
4625 | { | |
cc61c703 | 4626 | PostcopyState ps = postcopy_state_get(); |
c7c0e724 DH |
4627 | ram_addr_t offset; |
4628 | RAMBlock *rb = qemu_ram_block_from_host(host, false, &offset); | |
4629 | Error *err = NULL; | |
4630 | ||
4631 | if (ramblock_is_ignored(rb)) { | |
4632 | return; | |
4633 | } | |
4634 | ||
4635 | if (!migration_is_idle()) { | |
4636 | /* | |
4637 | * Precopy code on the source cannot deal with the size of RAM blocks | |
4638 | * changing at random points in time - especially after sending the | |
4639 | * RAM block sizes in the migration stream, they must no longer change. | |
4640 | * Abort and indicate a proper reason. | |
4641 | */ | |
4642 | error_setg(&err, "RAM block '%s' resized during precopy.", rb->idstr); | |
458fecca | 4643 | migration_cancel(err); |
c7c0e724 | 4644 | error_free(err); |
c7c0e724 | 4645 | } |
cc61c703 DH |
4646 | |
4647 | switch (ps) { | |
4648 | case POSTCOPY_INCOMING_ADVISE: | |
4649 | /* | |
4650 | * Update what ram_postcopy_incoming_init()->init_range() does at the | |
4651 | * time postcopy was advised. Syncing RAM blocks with the source will | |
4652 | * result in RAM resizes. | |
4653 | */ | |
4654 | if (old_size < new_size) { | |
4655 | if (ram_discard_range(rb->idstr, old_size, new_size - old_size)) { | |
4656 | error_report("RAM block '%s' discard of resized RAM failed", | |
4657 | rb->idstr); | |
4658 | } | |
4659 | } | |
898ba906 | 4660 | rb->postcopy_length = new_size; |
cc61c703 DH |
4661 | break; |
4662 | case POSTCOPY_INCOMING_NONE: | |
4663 | case POSTCOPY_INCOMING_RUNNING: | |
4664 | case POSTCOPY_INCOMING_END: | |
4665 | /* | |
4666 | * Once our guest is running, postcopy does no longer care about | |
4667 | * resizes. When growing, the new memory was not available on the | |
4668 | * source, no handler needed. | |
4669 | */ | |
4670 | break; | |
4671 | default: | |
4672 | error_report("RAM block '%s' resized during postcopy state: %d", | |
4673 | rb->idstr, ps); | |
4674 | exit(-1); | |
4675 | } | |
c7c0e724 DH |
4676 | } |
4677 | ||
4678 | static RAMBlockNotifier ram_mig_ram_notifier = { | |
4679 | .ram_block_resized = ram_mig_ram_block_resized, | |
4680 | }; | |
4681 | ||
56e93d26 JQ |
4682 | void ram_mig_init(void) |
4683 | { | |
4684 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 4685 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
c7c0e724 | 4686 | ram_block_notifier_add(&ram_mig_ram_notifier); |
56e93d26 | 4687 | } |