]>
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; | |
8d80e195 JQ |
333 | /* total ram size in bytes */ |
334 | uint64_t ram_bytes_total; | |
6f37bb8b JQ |
335 | /* Last block that we have visited searching for dirty pages */ |
336 | RAMBlock *last_seen_block; | |
269ace29 JQ |
337 | /* Last dirty target page we have sent */ |
338 | ram_addr_t last_page; | |
6f37bb8b JQ |
339 | /* last ram version we have seen */ |
340 | uint32_t last_version; | |
8d820d6f JQ |
341 | /* How many times we have dirty too many pages */ |
342 | int dirty_rate_high_cnt; | |
f664da80 JQ |
343 | /* these variables are used for bitmap sync */ |
344 | /* last time we did a full bitmap_sync */ | |
345 | int64_t time_last_bitmap_sync; | |
eac74159 | 346 | /* bytes transferred at start_time */ |
c4bdf0cf | 347 | uint64_t bytes_xfer_prev; |
a66cd90c | 348 | /* number of dirty pages since start_time */ |
68908ed6 | 349 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
350 | /* xbzrle misses since the beginning of the period */ |
351 | uint64_t xbzrle_cache_miss_prev; | |
e460a4b1 WW |
352 | /* Amount of xbzrle pages since the beginning of the period */ |
353 | uint64_t xbzrle_pages_prev; | |
354 | /* Amount of xbzrle encoded bytes since the beginning of the period */ | |
355 | uint64_t xbzrle_bytes_prev; | |
1a373522 DH |
356 | /* Start using XBZRLE (e.g., after the first round). */ |
357 | bool xbzrle_enabled; | |
05931ec5 JQ |
358 | /* Are we on the last stage of migration */ |
359 | bool last_stage; | |
76e03000 XG |
360 | /* compression statistics since the beginning of the period */ |
361 | /* amount of count that no free thread to compress data */ | |
362 | uint64_t compress_thread_busy_prev; | |
363 | /* amount bytes after compression */ | |
364 | uint64_t compressed_size_prev; | |
365 | /* amount of compressed pages */ | |
366 | uint64_t compress_pages_prev; | |
367 | ||
be8b02ed XG |
368 | /* total handled target pages at the beginning of period */ |
369 | uint64_t target_page_count_prev; | |
370 | /* total handled target pages since start */ | |
371 | uint64_t target_page_count; | |
9360447d | 372 | /* number of dirty bits in the bitmap */ |
2dfaf12e | 373 | uint64_t migration_dirty_pages; |
f1668764 PX |
374 | /* |
375 | * Protects: | |
376 | * - dirty/clear bitmap | |
377 | * - migration_dirty_pages | |
378 | * - pss structures | |
379 | */ | |
108cfae0 | 380 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
381 | /* The RAMBlock used in the last src_page_requests */ |
382 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
383 | /* Queue of outstanding page requests from the destination */ |
384 | QemuMutex src_page_req_mutex; | |
b58deb34 | 385 | QSIMPLEQ_HEAD(, RAMSrcPageRequest) src_page_requests; |
6f37bb8b JQ |
386 | }; |
387 | typedef struct RAMState RAMState; | |
388 | ||
53518d94 | 389 | static RAMState *ram_state; |
6f37bb8b | 390 | |
bd227060 WW |
391 | static NotifierWithReturnList precopy_notifier_list; |
392 | ||
a1fe28df PX |
393 | /* Whether postcopy has queued requests? */ |
394 | static bool postcopy_has_request(RAMState *rs) | |
395 | { | |
396 | return !QSIMPLEQ_EMPTY_ATOMIC(&rs->src_page_requests); | |
397 | } | |
398 | ||
bd227060 WW |
399 | void precopy_infrastructure_init(void) |
400 | { | |
401 | notifier_with_return_list_init(&precopy_notifier_list); | |
402 | } | |
403 | ||
404 | void precopy_add_notifier(NotifierWithReturn *n) | |
405 | { | |
406 | notifier_with_return_list_add(&precopy_notifier_list, n); | |
407 | } | |
408 | ||
409 | void precopy_remove_notifier(NotifierWithReturn *n) | |
410 | { | |
411 | notifier_with_return_remove(n); | |
412 | } | |
413 | ||
414 | int precopy_notify(PrecopyNotifyReason reason, Error **errp) | |
415 | { | |
416 | PrecopyNotifyData pnd; | |
417 | pnd.reason = reason; | |
418 | pnd.errp = errp; | |
419 | ||
420 | return notifier_with_return_list_notify(&precopy_notifier_list, &pnd); | |
421 | } | |
422 | ||
9edabd4d | 423 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 424 | { |
bae416e5 DDAG |
425 | return ram_state ? (ram_state->migration_dirty_pages * TARGET_PAGE_SIZE) : |
426 | 0; | |
2f4fde93 JQ |
427 | } |
428 | ||
23b7576d PX |
429 | /* |
430 | * NOTE: not all stats in ram_counters are used in reality. See comments | |
431 | * for struct MigrationAtomicStats. The ultimate result of ram migration | |
432 | * counters will be a merged version with both ram_counters and the atomic | |
433 | * fields in ram_atomic_counters. | |
434 | */ | |
9360447d | 435 | MigrationStats ram_counters; |
23b7576d | 436 | MigrationAtomicStats ram_atomic_counters; |
96506894 | 437 | |
26a26069 | 438 | void ram_transferred_add(uint64_t bytes) |
4c2d0f6d | 439 | { |
ae680668 DE |
440 | if (runstate_is_running()) { |
441 | ram_counters.precopy_bytes += bytes; | |
442 | } else if (migration_in_postcopy()) { | |
23b7576d | 443 | stat64_add(&ram_atomic_counters.postcopy_bytes, bytes); |
ae680668 DE |
444 | } else { |
445 | ram_counters.downtime_bytes += bytes; | |
446 | } | |
23b7576d | 447 | stat64_add(&ram_atomic_counters.transferred, bytes); |
4c2d0f6d DE |
448 | } |
449 | ||
d59c40cc LB |
450 | void dirty_sync_missed_zero_copy(void) |
451 | { | |
452 | ram_counters.dirty_sync_missed_zero_copy++; | |
453 | } | |
454 | ||
4010ba38 JQ |
455 | struct MigrationOps { |
456 | int (*ram_save_target_page)(RAMState *rs, PageSearchStatus *pss); | |
457 | }; | |
458 | typedef struct MigrationOps MigrationOps; | |
459 | ||
460 | MigrationOps *migration_ops; | |
461 | ||
76e03000 XG |
462 | CompressionStats compression_counters; |
463 | ||
56e93d26 | 464 | struct CompressParam { |
56e93d26 | 465 | bool done; |
90e56fb4 | 466 | bool quit; |
5e5fdcff | 467 | bool zero_page; |
56e93d26 JQ |
468 | QEMUFile *file; |
469 | QemuMutex mutex; | |
470 | QemuCond cond; | |
471 | RAMBlock *block; | |
472 | ram_addr_t offset; | |
34ab9e97 XG |
473 | |
474 | /* internally used fields */ | |
dcaf446e | 475 | z_stream stream; |
34ab9e97 | 476 | uint8_t *originbuf; |
56e93d26 JQ |
477 | }; |
478 | typedef struct CompressParam CompressParam; | |
479 | ||
480 | struct DecompressParam { | |
73a8912b | 481 | bool done; |
90e56fb4 | 482 | bool quit; |
56e93d26 JQ |
483 | QemuMutex mutex; |
484 | QemuCond cond; | |
485 | void *des; | |
d341d9f3 | 486 | uint8_t *compbuf; |
56e93d26 | 487 | int len; |
797ca154 | 488 | z_stream stream; |
56e93d26 JQ |
489 | }; |
490 | typedef struct DecompressParam DecompressParam; | |
491 | ||
492 | static CompressParam *comp_param; | |
493 | static QemuThread *compress_threads; | |
494 | /* comp_done_cond is used to wake up the migration thread when | |
495 | * one of the compression threads has finished the compression. | |
496 | * comp_done_lock is used to co-work with comp_done_cond. | |
497 | */ | |
0d9f9a5c LL |
498 | static QemuMutex comp_done_lock; |
499 | static QemuCond comp_done_cond; | |
56e93d26 | 500 | |
34ab9e97 | 501 | static QEMUFile *decomp_file; |
56e93d26 JQ |
502 | static DecompressParam *decomp_param; |
503 | static QemuThread *decompress_threads; | |
73a8912b LL |
504 | static QemuMutex decomp_done_lock; |
505 | static QemuCond decomp_done_cond; | |
56e93d26 | 506 | |
93589827 PX |
507 | static int ram_save_host_page_urgent(PageSearchStatus *pss); |
508 | ||
5e5fdcff | 509 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 510 | ram_addr_t offset, uint8_t *source_buf); |
56e93d26 | 511 | |
ebd88a49 PX |
512 | /* NOTE: page is the PFN not real ram_addr_t. */ |
513 | static void pss_init(PageSearchStatus *pss, RAMBlock *rb, ram_addr_t page) | |
514 | { | |
515 | pss->block = rb; | |
516 | pss->page = page; | |
517 | pss->complete_round = false; | |
518 | } | |
519 | ||
93589827 PX |
520 | /* |
521 | * Check whether two PSSs are actively sending the same page. Return true | |
522 | * if it is, false otherwise. | |
523 | */ | |
524 | static bool pss_overlap(PageSearchStatus *pss1, PageSearchStatus *pss2) | |
525 | { | |
526 | return pss1->host_page_sending && pss2->host_page_sending && | |
527 | (pss1->host_page_start == pss2->host_page_start); | |
528 | } | |
529 | ||
56e93d26 JQ |
530 | static void *do_data_compress(void *opaque) |
531 | { | |
532 | CompressParam *param = opaque; | |
a7a9a88f LL |
533 | RAMBlock *block; |
534 | ram_addr_t offset; | |
5e5fdcff | 535 | bool zero_page; |
56e93d26 | 536 | |
a7a9a88f | 537 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 538 | while (!param->quit) { |
a7a9a88f LL |
539 | if (param->block) { |
540 | block = param->block; | |
541 | offset = param->offset; | |
542 | param->block = NULL; | |
543 | qemu_mutex_unlock(¶m->mutex); | |
544 | ||
5e5fdcff XG |
545 | zero_page = do_compress_ram_page(param->file, ¶m->stream, |
546 | block, offset, param->originbuf); | |
a7a9a88f | 547 | |
0d9f9a5c | 548 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 549 | param->done = true; |
5e5fdcff | 550 | param->zero_page = zero_page; |
0d9f9a5c LL |
551 | qemu_cond_signal(&comp_done_cond); |
552 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
553 | |
554 | qemu_mutex_lock(¶m->mutex); | |
555 | } else { | |
56e93d26 JQ |
556 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
557 | } | |
56e93d26 | 558 | } |
a7a9a88f | 559 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
560 | |
561 | return NULL; | |
562 | } | |
563 | ||
f0afa331 | 564 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
565 | { |
566 | int i, thread_count; | |
567 | ||
05306935 | 568 | if (!migrate_use_compression() || !comp_param) { |
56e93d26 JQ |
569 | return; |
570 | } | |
05306935 | 571 | |
56e93d26 JQ |
572 | thread_count = migrate_compress_threads(); |
573 | for (i = 0; i < thread_count; i++) { | |
dcaf446e XG |
574 | /* |
575 | * we use it as a indicator which shows if the thread is | |
576 | * properly init'd or not | |
577 | */ | |
578 | if (!comp_param[i].file) { | |
579 | break; | |
580 | } | |
05306935 FL |
581 | |
582 | qemu_mutex_lock(&comp_param[i].mutex); | |
583 | comp_param[i].quit = true; | |
584 | qemu_cond_signal(&comp_param[i].cond); | |
585 | qemu_mutex_unlock(&comp_param[i].mutex); | |
586 | ||
56e93d26 | 587 | qemu_thread_join(compress_threads + i); |
56e93d26 JQ |
588 | qemu_mutex_destroy(&comp_param[i].mutex); |
589 | qemu_cond_destroy(&comp_param[i].cond); | |
dcaf446e | 590 | deflateEnd(&comp_param[i].stream); |
34ab9e97 | 591 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
592 | qemu_fclose(comp_param[i].file); |
593 | comp_param[i].file = NULL; | |
56e93d26 | 594 | } |
0d9f9a5c LL |
595 | qemu_mutex_destroy(&comp_done_lock); |
596 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
597 | g_free(compress_threads); |
598 | g_free(comp_param); | |
56e93d26 JQ |
599 | compress_threads = NULL; |
600 | comp_param = NULL; | |
56e93d26 JQ |
601 | } |
602 | ||
dcaf446e | 603 | static int compress_threads_save_setup(void) |
56e93d26 JQ |
604 | { |
605 | int i, thread_count; | |
606 | ||
607 | if (!migrate_use_compression()) { | |
dcaf446e | 608 | return 0; |
56e93d26 | 609 | } |
56e93d26 JQ |
610 | thread_count = migrate_compress_threads(); |
611 | compress_threads = g_new0(QemuThread, thread_count); | |
612 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
613 | qemu_cond_init(&comp_done_cond); |
614 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 615 | for (i = 0; i < thread_count; i++) { |
34ab9e97 XG |
616 | comp_param[i].originbuf = g_try_malloc(TARGET_PAGE_SIZE); |
617 | if (!comp_param[i].originbuf) { | |
618 | goto exit; | |
619 | } | |
620 | ||
dcaf446e XG |
621 | if (deflateInit(&comp_param[i].stream, |
622 | migrate_compress_level()) != Z_OK) { | |
34ab9e97 | 623 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
624 | goto exit; |
625 | } | |
626 | ||
e110aa91 C |
627 | /* comp_param[i].file is just used as a dummy buffer to save data, |
628 | * set its ops to empty. | |
56e93d26 | 629 | */ |
77ef2dc1 | 630 | comp_param[i].file = qemu_file_new_output( |
c0e0825c | 631 | QIO_CHANNEL(qio_channel_null_new())); |
56e93d26 | 632 | comp_param[i].done = true; |
90e56fb4 | 633 | comp_param[i].quit = false; |
56e93d26 JQ |
634 | qemu_mutex_init(&comp_param[i].mutex); |
635 | qemu_cond_init(&comp_param[i].cond); | |
636 | qemu_thread_create(compress_threads + i, "compress", | |
637 | do_data_compress, comp_param + i, | |
638 | QEMU_THREAD_JOINABLE); | |
639 | } | |
dcaf446e XG |
640 | return 0; |
641 | ||
642 | exit: | |
643 | compress_threads_save_cleanup(); | |
644 | return -1; | |
56e93d26 JQ |
645 | } |
646 | ||
647 | /** | |
3d0684b2 | 648 | * save_page_header: write page header to wire |
56e93d26 JQ |
649 | * |
650 | * If this is the 1st block, it also writes the block identification | |
651 | * | |
3d0684b2 | 652 | * Returns the number of bytes written |
56e93d26 | 653 | * |
ec6f3ab9 | 654 | * @pss: current PSS channel status |
56e93d26 JQ |
655 | * @block: block that contains the page we want to send |
656 | * @offset: offset inside the block for the page | |
657 | * in the lower bits, it contains flags | |
658 | */ | |
ec6f3ab9 | 659 | static size_t save_page_header(PageSearchStatus *pss, RAMBlock *block, |
2bf3aa85 | 660 | ram_addr_t offset) |
56e93d26 | 661 | { |
9f5f380b | 662 | size_t size, len; |
ec6f3ab9 PX |
663 | bool same_block = (block == pss->last_sent_block); |
664 | QEMUFile *f = pss->pss_channel; | |
56e93d26 | 665 | |
10661f11 | 666 | if (same_block) { |
24795694 JQ |
667 | offset |= RAM_SAVE_FLAG_CONTINUE; |
668 | } | |
2bf3aa85 | 669 | qemu_put_be64(f, offset); |
56e93d26 JQ |
670 | size = 8; |
671 | ||
10661f11 | 672 | if (!same_block) { |
9f5f380b | 673 | len = strlen(block->idstr); |
2bf3aa85 JQ |
674 | qemu_put_byte(f, len); |
675 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 676 | size += 1 + len; |
ec6f3ab9 | 677 | pss->last_sent_block = block; |
56e93d26 JQ |
678 | } |
679 | return size; | |
680 | } | |
681 | ||
3d0684b2 | 682 | /** |
179a8080 | 683 | * mig_throttle_guest_down: throttle down the guest |
3d0684b2 JQ |
684 | * |
685 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
686 | * writes. If guest dirty memory rate is reduced below the rate at | |
687 | * which we can transfer pages to the destination then we should be | |
688 | * able to complete migration. Some workloads dirty memory way too | |
689 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 | 690 | */ |
cbbf8182 KZ |
691 | static void mig_throttle_guest_down(uint64_t bytes_dirty_period, |
692 | uint64_t bytes_dirty_threshold) | |
070afca2 JH |
693 | { |
694 | MigrationState *s = migrate_get_current(); | |
2594f56d | 695 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
cbbf8182 KZ |
696 | uint64_t pct_increment = s->parameters.cpu_throttle_increment; |
697 | bool pct_tailslow = s->parameters.cpu_throttle_tailslow; | |
4cbc9c7f | 698 | int pct_max = s->parameters.max_cpu_throttle; |
070afca2 | 699 | |
cbbf8182 KZ |
700 | uint64_t throttle_now = cpu_throttle_get_percentage(); |
701 | uint64_t cpu_now, cpu_ideal, throttle_inc; | |
702 | ||
070afca2 JH |
703 | /* We have not started throttling yet. Let's start it. */ |
704 | if (!cpu_throttle_active()) { | |
705 | cpu_throttle_set(pct_initial); | |
706 | } else { | |
707 | /* Throttling already on, just increase the rate */ | |
cbbf8182 KZ |
708 | if (!pct_tailslow) { |
709 | throttle_inc = pct_increment; | |
710 | } else { | |
711 | /* Compute the ideal CPU percentage used by Guest, which may | |
712 | * make the dirty rate match the dirty rate threshold. */ | |
713 | cpu_now = 100 - throttle_now; | |
714 | cpu_ideal = cpu_now * (bytes_dirty_threshold * 1.0 / | |
715 | bytes_dirty_period); | |
716 | throttle_inc = MIN(cpu_now - cpu_ideal, pct_increment); | |
717 | } | |
718 | cpu_throttle_set(MIN(throttle_now + throttle_inc, pct_max)); | |
070afca2 JH |
719 | } |
720 | } | |
721 | ||
91fe9a8d RL |
722 | void mig_throttle_counter_reset(void) |
723 | { | |
724 | RAMState *rs = ram_state; | |
725 | ||
726 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
727 | rs->num_dirty_pages_period = 0; | |
23b7576d | 728 | rs->bytes_xfer_prev = stat64_get(&ram_atomic_counters.transferred); |
91fe9a8d RL |
729 | } |
730 | ||
3d0684b2 JQ |
731 | /** |
732 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
733 | * | |
6f37bb8b | 734 | * @rs: current RAM state |
3d0684b2 JQ |
735 | * @current_addr: address for the zero page |
736 | * | |
737 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
738 | * The important thing is that a stale (not-yet-0'd) page be replaced |
739 | * by the new data. | |
740 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 741 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 742 | */ |
6f37bb8b | 743 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 744 | { |
56e93d26 JQ |
745 | /* We don't care if this fails to allocate a new cache page |
746 | * as long as it updated an old one */ | |
c00e0928 | 747 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 748 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
749 | } |
750 | ||
751 | #define ENCODING_FLAG_XBZRLE 0x1 | |
752 | ||
753 | /** | |
754 | * save_xbzrle_page: compress and send current page | |
755 | * | |
756 | * Returns: 1 means that we wrote the page | |
757 | * 0 means that page is identical to the one already sent | |
758 | * -1 means that xbzrle would be longer than normal | |
759 | * | |
5a987738 | 760 | * @rs: current RAM state |
ec6f3ab9 | 761 | * @pss: current PSS channel |
3d0684b2 JQ |
762 | * @current_data: pointer to the address of the page contents |
763 | * @current_addr: addr of the page | |
56e93d26 JQ |
764 | * @block: block that contains the page we want to send |
765 | * @offset: offset inside the block for the page | |
56e93d26 | 766 | */ |
ec6f3ab9 | 767 | static int save_xbzrle_page(RAMState *rs, PageSearchStatus *pss, |
61717ea9 PX |
768 | uint8_t **current_data, ram_addr_t current_addr, |
769 | RAMBlock *block, ram_addr_t offset) | |
56e93d26 JQ |
770 | { |
771 | int encoded_len = 0, bytes_xbzrle; | |
772 | uint8_t *prev_cached_page; | |
ec6f3ab9 | 773 | QEMUFile *file = pss->pss_channel; |
56e93d26 | 774 | |
9360447d JQ |
775 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
776 | ram_counters.dirty_sync_count)) { | |
777 | xbzrle_counters.cache_miss++; | |
05931ec5 | 778 | if (!rs->last_stage) { |
56e93d26 | 779 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, |
9360447d | 780 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
781 | return -1; |
782 | } else { | |
783 | /* update *current_data when the page has been | |
784 | inserted into cache */ | |
785 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
786 | } | |
787 | } | |
788 | return -1; | |
789 | } | |
790 | ||
e460a4b1 WW |
791 | /* |
792 | * Reaching here means the page has hit the xbzrle cache, no matter what | |
793 | * encoding result it is (normal encoding, overflow or skipping the page), | |
3a4452d8 | 794 | * count the page as encoded. This is used to calculate the encoding rate. |
e460a4b1 WW |
795 | * |
796 | * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB, | |
797 | * 2nd page turns out to be skipped (i.e. no new bytes written to the | |
798 | * page), the overall encoding rate will be 8KB / 2KB = 4, which has the | |
799 | * skipped page included. In this way, the encoding rate can tell if the | |
800 | * guest page is good for xbzrle encoding. | |
801 | */ | |
802 | xbzrle_counters.pages++; | |
56e93d26 JQ |
803 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); |
804 | ||
805 | /* save current buffer into memory */ | |
806 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
807 | ||
808 | /* XBZRLE encoding (if there is no overflow) */ | |
809 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
810 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
811 | TARGET_PAGE_SIZE); | |
ca353803 WY |
812 | |
813 | /* | |
814 | * Update the cache contents, so that it corresponds to the data | |
815 | * sent, in all cases except where we skip the page. | |
816 | */ | |
05931ec5 | 817 | if (!rs->last_stage && encoded_len != 0) { |
ca353803 WY |
818 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); |
819 | /* | |
820 | * In the case where we couldn't compress, ensure that the caller | |
821 | * sends the data from the cache, since the guest might have | |
822 | * changed the RAM since we copied it. | |
823 | */ | |
824 | *current_data = prev_cached_page; | |
825 | } | |
826 | ||
56e93d26 | 827 | if (encoded_len == 0) { |
55c4446b | 828 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
829 | return 0; |
830 | } else if (encoded_len == -1) { | |
55c4446b | 831 | trace_save_xbzrle_page_overflow(); |
9360447d | 832 | xbzrle_counters.overflow++; |
e460a4b1 | 833 | xbzrle_counters.bytes += TARGET_PAGE_SIZE; |
56e93d26 JQ |
834 | return -1; |
835 | } | |
836 | ||
56e93d26 | 837 | /* Send XBZRLE based compressed page */ |
ec6f3ab9 | 838 | bytes_xbzrle = save_page_header(pss, block, |
204b88b8 | 839 | offset | RAM_SAVE_FLAG_XBZRLE); |
61717ea9 PX |
840 | qemu_put_byte(file, ENCODING_FLAG_XBZRLE); |
841 | qemu_put_be16(file, encoded_len); | |
842 | qemu_put_buffer(file, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 843 | bytes_xbzrle += encoded_len + 1 + 2; |
e460a4b1 WW |
844 | /* |
845 | * Like compressed_size (please see update_compress_thread_counts), | |
846 | * the xbzrle encoded bytes don't count the 8 byte header with | |
847 | * RAM_SAVE_FLAG_CONTINUE. | |
848 | */ | |
849 | xbzrle_counters.bytes += bytes_xbzrle - 8; | |
4c2d0f6d | 850 | ram_transferred_add(bytes_xbzrle); |
56e93d26 JQ |
851 | |
852 | return 1; | |
853 | } | |
854 | ||
3d0684b2 | 855 | /** |
d9e474ea | 856 | * pss_find_next_dirty: find the next dirty page of current ramblock |
f3f491fc | 857 | * |
d9e474ea PX |
858 | * This function updates pss->page to point to the next dirty page index |
859 | * within the ramblock to migrate, or the end of ramblock when nothing | |
860 | * found. Note that when pss->host_page_sending==true it means we're | |
861 | * during sending a host page, so we won't look for dirty page that is | |
862 | * outside the host page boundary. | |
3d0684b2 | 863 | * |
d9e474ea | 864 | * @pss: the current page search status |
f3f491fc | 865 | */ |
d9e474ea | 866 | static void pss_find_next_dirty(PageSearchStatus *pss) |
56e93d26 | 867 | { |
d9e474ea | 868 | RAMBlock *rb = pss->block; |
6b6712ef JQ |
869 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
870 | unsigned long *bitmap = rb->bmap; | |
56e93d26 | 871 | |
fbd162e6 | 872 | if (ramblock_is_ignored(rb)) { |
d9e474ea PX |
873 | /* Points directly to the end, so we know no dirty page */ |
874 | pss->page = size; | |
875 | return; | |
876 | } | |
877 | ||
878 | /* | |
879 | * If during sending a host page, only look for dirty pages within the | |
880 | * current host page being send. | |
881 | */ | |
882 | if (pss->host_page_sending) { | |
883 | assert(pss->host_page_end); | |
884 | size = MIN(size, pss->host_page_end); | |
b895de50 CLG |
885 | } |
886 | ||
d9e474ea | 887 | pss->page = find_next_bit(bitmap, size, pss->page); |
56e93d26 JQ |
888 | } |
889 | ||
1230a25f | 890 | static void migration_clear_memory_region_dirty_bitmap(RAMBlock *rb, |
3143577d WW |
891 | unsigned long page) |
892 | { | |
893 | uint8_t shift; | |
894 | hwaddr size, start; | |
895 | ||
896 | if (!rb->clear_bmap || !clear_bmap_test_and_clear(rb, page)) { | |
897 | return; | |
898 | } | |
899 | ||
900 | shift = rb->clear_bmap_shift; | |
901 | /* | |
902 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
903 | * can make things easier sometimes since then start address | |
904 | * of the small chunk will always be 64 pages aligned so the | |
905 | * bitmap will always be aligned to unsigned long. We should | |
906 | * even be able to remove this restriction but I'm simply | |
907 | * keeping it. | |
908 | */ | |
909 | assert(shift >= 6); | |
910 | ||
911 | size = 1ULL << (TARGET_PAGE_BITS + shift); | |
7648297d | 912 | start = QEMU_ALIGN_DOWN((ram_addr_t)page << TARGET_PAGE_BITS, size); |
3143577d WW |
913 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); |
914 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
915 | } | |
916 | ||
917 | static void | |
1230a25f | 918 | migration_clear_memory_region_dirty_bitmap_range(RAMBlock *rb, |
3143577d WW |
919 | unsigned long start, |
920 | unsigned long npages) | |
921 | { | |
922 | unsigned long i, chunk_pages = 1UL << rb->clear_bmap_shift; | |
923 | unsigned long chunk_start = QEMU_ALIGN_DOWN(start, chunk_pages); | |
924 | unsigned long chunk_end = QEMU_ALIGN_UP(start + npages, chunk_pages); | |
925 | ||
926 | /* | |
927 | * Clear pages from start to start + npages - 1, so the end boundary is | |
928 | * exclusive. | |
929 | */ | |
930 | for (i = chunk_start; i < chunk_end; i += chunk_pages) { | |
1230a25f | 931 | migration_clear_memory_region_dirty_bitmap(rb, i); |
3143577d WW |
932 | } |
933 | } | |
934 | ||
a6a83cef RL |
935 | /* |
936 | * colo_bitmap_find_diry:find contiguous dirty pages from start | |
937 | * | |
938 | * Returns the page offset within memory region of the start of the contiguout | |
939 | * dirty page | |
940 | * | |
941 | * @rs: current RAM state | |
942 | * @rb: RAMBlock where to search for dirty pages | |
943 | * @start: page where we start the search | |
944 | * @num: the number of contiguous dirty pages | |
945 | */ | |
946 | static inline | |
947 | unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, | |
948 | unsigned long start, unsigned long *num) | |
949 | { | |
950 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; | |
951 | unsigned long *bitmap = rb->bmap; | |
952 | unsigned long first, next; | |
953 | ||
954 | *num = 0; | |
955 | ||
956 | if (ramblock_is_ignored(rb)) { | |
957 | return size; | |
958 | } | |
959 | ||
960 | first = find_next_bit(bitmap, size, start); | |
961 | if (first >= size) { | |
962 | return first; | |
963 | } | |
964 | next = find_next_zero_bit(bitmap, size, first + 1); | |
965 | assert(next >= first); | |
966 | *num = next - first; | |
967 | return first; | |
968 | } | |
969 | ||
06b10688 | 970 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
971 | RAMBlock *rb, |
972 | unsigned long page) | |
a82d593b DDAG |
973 | { |
974 | bool ret; | |
a82d593b | 975 | |
002cad6b PX |
976 | /* |
977 | * Clear dirty bitmap if needed. This _must_ be called before we | |
978 | * send any of the page in the chunk because we need to make sure | |
979 | * we can capture further page content changes when we sync dirty | |
980 | * log the next time. So as long as we are going to send any of | |
981 | * the page in the chunk we clear the remote dirty bitmap for all. | |
982 | * Clearing it earlier won't be a problem, but too late will. | |
983 | */ | |
1230a25f | 984 | migration_clear_memory_region_dirty_bitmap(rb, page); |
002cad6b | 985 | |
6b6712ef | 986 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b | 987 | if (ret) { |
0d8ec885 | 988 | rs->migration_dirty_pages--; |
a82d593b | 989 | } |
386a907b | 990 | |
a82d593b DDAG |
991 | return ret; |
992 | } | |
993 | ||
be39b4cd DH |
994 | static void dirty_bitmap_clear_section(MemoryRegionSection *section, |
995 | void *opaque) | |
996 | { | |
997 | const hwaddr offset = section->offset_within_region; | |
998 | const hwaddr size = int128_get64(section->size); | |
999 | const unsigned long start = offset >> TARGET_PAGE_BITS; | |
1000 | const unsigned long npages = size >> TARGET_PAGE_BITS; | |
1001 | RAMBlock *rb = section->mr->ram_block; | |
1002 | uint64_t *cleared_bits = opaque; | |
1003 | ||
1004 | /* | |
1005 | * We don't grab ram_state->bitmap_mutex because we expect to run | |
1006 | * only when starting migration or during postcopy recovery where | |
1007 | * we don't have concurrent access. | |
1008 | */ | |
1009 | if (!migration_in_postcopy() && !migrate_background_snapshot()) { | |
1010 | migration_clear_memory_region_dirty_bitmap_range(rb, start, npages); | |
1011 | } | |
1012 | *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages); | |
1013 | bitmap_clear(rb->bmap, start, npages); | |
1014 | } | |
1015 | ||
1016 | /* | |
1017 | * Exclude all dirty pages from migration that fall into a discarded range as | |
1018 | * managed by a RamDiscardManager responsible for the mapped memory region of | |
1019 | * the RAMBlock. Clear the corresponding bits in the dirty bitmaps. | |
1020 | * | |
1021 | * Discarded pages ("logically unplugged") have undefined content and must | |
1022 | * not get migrated, because even reading these pages for migration might | |
1023 | * result in undesired behavior. | |
1024 | * | |
1025 | * Returns the number of cleared bits in the RAMBlock dirty bitmap. | |
1026 | * | |
1027 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1028 | */ | |
1029 | static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb) | |
1030 | { | |
1031 | uint64_t cleared_bits = 0; | |
1032 | ||
1033 | if (rb->mr && rb->bmap && memory_region_has_ram_discard_manager(rb->mr)) { | |
1034 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1035 | MemoryRegionSection section = { | |
1036 | .mr = rb->mr, | |
1037 | .offset_within_region = 0, | |
1038 | .size = int128_make64(qemu_ram_get_used_length(rb)), | |
1039 | }; | |
1040 | ||
1041 | ram_discard_manager_replay_discarded(rdm, §ion, | |
1042 | dirty_bitmap_clear_section, | |
1043 | &cleared_bits); | |
1044 | } | |
1045 | return cleared_bits; | |
1046 | } | |
1047 | ||
9470c5e0 DH |
1048 | /* |
1049 | * Check if a host-page aligned page falls into a discarded range as managed by | |
1050 | * a RamDiscardManager responsible for the mapped memory region of the RAMBlock. | |
1051 | * | |
1052 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1053 | */ | |
1054 | bool ramblock_page_is_discarded(RAMBlock *rb, ram_addr_t start) | |
1055 | { | |
1056 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1057 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1058 | MemoryRegionSection section = { | |
1059 | .mr = rb->mr, | |
1060 | .offset_within_region = start, | |
1061 | .size = int128_make64(qemu_ram_pagesize(rb)), | |
1062 | }; | |
1063 | ||
1064 | return !ram_discard_manager_is_populated(rdm, §ion); | |
1065 | } | |
1066 | return false; | |
1067 | } | |
1068 | ||
267691b6 | 1069 | /* Called with RCU critical section */ |
7a3e9571 | 1070 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 1071 | { |
fb613580 KZ |
1072 | uint64_t new_dirty_pages = |
1073 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length); | |
1074 | ||
1075 | rs->migration_dirty_pages += new_dirty_pages; | |
1076 | rs->num_dirty_pages_period += new_dirty_pages; | |
56e93d26 JQ |
1077 | } |
1078 | ||
3d0684b2 JQ |
1079 | /** |
1080 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
1081 | * | |
1082 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
1083 | * | |
1084 | * For VMs with just normal pages this is equivalent to the host page | |
1085 | * size. If it's got some huge pages then it's the OR of all the | |
1086 | * different page sizes. | |
e8ca1db2 DDAG |
1087 | */ |
1088 | uint64_t ram_pagesize_summary(void) | |
1089 | { | |
1090 | RAMBlock *block; | |
1091 | uint64_t summary = 0; | |
1092 | ||
fbd162e6 | 1093 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
1094 | summary |= block->page_size; |
1095 | } | |
1096 | ||
1097 | return summary; | |
1098 | } | |
1099 | ||
aecbfe9c XG |
1100 | uint64_t ram_get_total_transferred_pages(void) |
1101 | { | |
23b7576d PX |
1102 | return stat64_get(&ram_atomic_counters.normal) + |
1103 | stat64_get(&ram_atomic_counters.duplicate) + | |
1104 | compression_counters.pages + xbzrle_counters.pages; | |
aecbfe9c XG |
1105 | } |
1106 | ||
b734035b XG |
1107 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
1108 | { | |
be8b02ed | 1109 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 1110 | double compressed_size; |
b734035b XG |
1111 | |
1112 | /* calculate period counters */ | |
1113 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 | |
1114 | / (end_time - rs->time_last_bitmap_sync); | |
1115 | ||
be8b02ed | 1116 | if (!page_count) { |
b734035b XG |
1117 | return; |
1118 | } | |
1119 | ||
1120 | if (migrate_use_xbzrle()) { | |
e460a4b1 WW |
1121 | double encoded_size, unencoded_size; |
1122 | ||
b734035b | 1123 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - |
be8b02ed | 1124 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b | 1125 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
e460a4b1 WW |
1126 | unencoded_size = (xbzrle_counters.pages - rs->xbzrle_pages_prev) * |
1127 | TARGET_PAGE_SIZE; | |
1128 | encoded_size = xbzrle_counters.bytes - rs->xbzrle_bytes_prev; | |
92271402 | 1129 | if (xbzrle_counters.pages == rs->xbzrle_pages_prev || !encoded_size) { |
e460a4b1 | 1130 | xbzrle_counters.encoding_rate = 0; |
e460a4b1 WW |
1131 | } else { |
1132 | xbzrle_counters.encoding_rate = unencoded_size / encoded_size; | |
1133 | } | |
1134 | rs->xbzrle_pages_prev = xbzrle_counters.pages; | |
1135 | rs->xbzrle_bytes_prev = xbzrle_counters.bytes; | |
b734035b | 1136 | } |
76e03000 XG |
1137 | |
1138 | if (migrate_use_compression()) { | |
1139 | compression_counters.busy_rate = (double)(compression_counters.busy - | |
1140 | rs->compress_thread_busy_prev) / page_count; | |
1141 | rs->compress_thread_busy_prev = compression_counters.busy; | |
1142 | ||
1143 | compressed_size = compression_counters.compressed_size - | |
1144 | rs->compressed_size_prev; | |
1145 | if (compressed_size) { | |
1146 | double uncompressed_size = (compression_counters.pages - | |
1147 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
1148 | ||
1149 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
1150 | compression_counters.compression_rate = | |
1151 | uncompressed_size / compressed_size; | |
1152 | ||
1153 | rs->compress_pages_prev = compression_counters.pages; | |
1154 | rs->compressed_size_prev = compression_counters.compressed_size; | |
1155 | } | |
1156 | } | |
b734035b XG |
1157 | } |
1158 | ||
dc14a470 KZ |
1159 | static void migration_trigger_throttle(RAMState *rs) |
1160 | { | |
1161 | MigrationState *s = migrate_get_current(); | |
1162 | uint64_t threshold = s->parameters.throttle_trigger_threshold; | |
23b7576d PX |
1163 | uint64_t bytes_xfer_period = |
1164 | stat64_get(&ram_atomic_counters.transferred) - rs->bytes_xfer_prev; | |
dc14a470 KZ |
1165 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; |
1166 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
1167 | ||
1168 | /* During block migration the auto-converge logic incorrectly detects | |
1169 | * that ram migration makes no progress. Avoid this by disabling the | |
1170 | * throttling logic during the bulk phase of block migration. */ | |
1171 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
1172 | /* The following detection logic can be refined later. For now: | |
1173 | Check to see if the ratio between dirtied bytes and the approx. | |
1174 | amount of bytes that just got transferred since the last time | |
1175 | we were in this routine reaches the threshold. If that happens | |
1176 | twice, start or increase throttling. */ | |
1177 | ||
1178 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
1179 | (++rs->dirty_rate_high_cnt >= 2)) { | |
1180 | trace_migration_throttle(); | |
1181 | rs->dirty_rate_high_cnt = 0; | |
cbbf8182 KZ |
1182 | mig_throttle_guest_down(bytes_dirty_period, |
1183 | bytes_dirty_threshold); | |
dc14a470 KZ |
1184 | } |
1185 | } | |
1186 | } | |
1187 | ||
8d820d6f | 1188 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
1189 | { |
1190 | RAMBlock *block; | |
56e93d26 | 1191 | int64_t end_time; |
56e93d26 | 1192 | |
9360447d | 1193 | ram_counters.dirty_sync_count++; |
56e93d26 | 1194 | |
f664da80 JQ |
1195 | if (!rs->time_last_bitmap_sync) { |
1196 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
1197 | } |
1198 | ||
1199 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 1200 | memory_global_dirty_log_sync(); |
56e93d26 | 1201 | |
108cfae0 | 1202 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
1203 | WITH_RCU_READ_LOCK_GUARD() { |
1204 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1205 | ramblock_sync_dirty_bitmap(rs, block); | |
1206 | } | |
1207 | ram_counters.remaining = ram_bytes_remaining(); | |
56e93d26 | 1208 | } |
108cfae0 | 1209 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 1210 | |
9458a9a1 | 1211 | memory_global_after_dirty_log_sync(); |
a66cd90c | 1212 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 1213 | |
56e93d26 JQ |
1214 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
1215 | ||
1216 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 1217 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 1218 | migration_trigger_throttle(rs); |
070afca2 | 1219 | |
b734035b XG |
1220 | migration_update_rates(rs, end_time); |
1221 | ||
be8b02ed | 1222 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
1223 | |
1224 | /* reset period counters */ | |
f664da80 | 1225 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 1226 | rs->num_dirty_pages_period = 0; |
23b7576d | 1227 | rs->bytes_xfer_prev = stat64_get(&ram_atomic_counters.transferred); |
56e93d26 | 1228 | } |
4addcd4f | 1229 | if (migrate_use_events()) { |
3ab72385 | 1230 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count); |
4addcd4f | 1231 | } |
56e93d26 JQ |
1232 | } |
1233 | ||
bd227060 WW |
1234 | static void migration_bitmap_sync_precopy(RAMState *rs) |
1235 | { | |
1236 | Error *local_err = NULL; | |
1237 | ||
1238 | /* | |
1239 | * The current notifier usage is just an optimization to migration, so we | |
1240 | * don't stop the normal migration process in the error case. | |
1241 | */ | |
1242 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
1243 | error_report_err(local_err); | |
b4a1733c | 1244 | local_err = NULL; |
bd227060 WW |
1245 | } |
1246 | ||
1247 | migration_bitmap_sync(rs); | |
1248 | ||
1249 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
1250 | error_report_err(local_err); | |
1251 | } | |
1252 | } | |
1253 | ||
a4dbaf8e | 1254 | void ram_release_page(const char *rbname, uint64_t offset) |
47fe16ff JQ |
1255 | { |
1256 | if (!migrate_release_ram() || !migration_in_postcopy()) { | |
1257 | return; | |
1258 | } | |
1259 | ||
1260 | ram_discard_range(rbname, offset, TARGET_PAGE_SIZE); | |
1261 | } | |
1262 | ||
6c97ec5f XG |
1263 | /** |
1264 | * save_zero_page_to_file: send the zero page to the file | |
1265 | * | |
1266 | * Returns the size of data written to the file, 0 means the page is not | |
1267 | * a zero page | |
1268 | * | |
ec6f3ab9 | 1269 | * @pss: current PSS channel |
6c97ec5f XG |
1270 | * @block: block that contains the page we want to send |
1271 | * @offset: offset inside the block for the page | |
1272 | */ | |
ec6f3ab9 | 1273 | static int save_zero_page_to_file(PageSearchStatus *pss, |
6c97ec5f XG |
1274 | RAMBlock *block, ram_addr_t offset) |
1275 | { | |
1276 | uint8_t *p = block->host + offset; | |
ec6f3ab9 | 1277 | QEMUFile *file = pss->pss_channel; |
6c97ec5f XG |
1278 | int len = 0; |
1279 | ||
bad452a7 | 1280 | if (buffer_is_zero(p, TARGET_PAGE_SIZE)) { |
ec6f3ab9 | 1281 | len += save_page_header(pss, block, offset | RAM_SAVE_FLAG_ZERO); |
6c97ec5f XG |
1282 | qemu_put_byte(file, 0); |
1283 | len += 1; | |
47fe16ff | 1284 | ram_release_page(block->idstr, offset); |
6c97ec5f XG |
1285 | } |
1286 | return len; | |
1287 | } | |
1288 | ||
56e93d26 | 1289 | /** |
3d0684b2 | 1290 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1291 | * |
3d0684b2 | 1292 | * Returns the number of pages written. |
56e93d26 | 1293 | * |
ec6f3ab9 | 1294 | * @pss: current PSS channel |
56e93d26 JQ |
1295 | * @block: block that contains the page we want to send |
1296 | * @offset: offset inside the block for the page | |
56e93d26 | 1297 | */ |
ec6f3ab9 | 1298 | static int save_zero_page(PageSearchStatus *pss, RAMBlock *block, |
61717ea9 | 1299 | ram_addr_t offset) |
56e93d26 | 1300 | { |
ec6f3ab9 | 1301 | int len = save_zero_page_to_file(pss, block, offset); |
56e93d26 | 1302 | |
6c97ec5f | 1303 | if (len) { |
23b7576d | 1304 | stat64_add(&ram_atomic_counters.duplicate, 1); |
4c2d0f6d | 1305 | ram_transferred_add(len); |
6c97ec5f | 1306 | return 1; |
56e93d26 | 1307 | } |
6c97ec5f | 1308 | return -1; |
56e93d26 JQ |
1309 | } |
1310 | ||
059ff0fb XG |
1311 | /* |
1312 | * @pages: the number of pages written by the control path, | |
1313 | * < 0 - error | |
1314 | * > 0 - number of pages written | |
1315 | * | |
1316 | * Return true if the pages has been saved, otherwise false is returned. | |
1317 | */ | |
61717ea9 PX |
1318 | static bool control_save_page(PageSearchStatus *pss, RAMBlock *block, |
1319 | ram_addr_t offset, int *pages) | |
059ff0fb XG |
1320 | { |
1321 | uint64_t bytes_xmit = 0; | |
1322 | int ret; | |
1323 | ||
1324 | *pages = -1; | |
61717ea9 PX |
1325 | ret = ram_control_save_page(pss->pss_channel, block->offset, offset, |
1326 | TARGET_PAGE_SIZE, &bytes_xmit); | |
059ff0fb XG |
1327 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { |
1328 | return false; | |
1329 | } | |
1330 | ||
1331 | if (bytes_xmit) { | |
4c2d0f6d | 1332 | ram_transferred_add(bytes_xmit); |
059ff0fb XG |
1333 | *pages = 1; |
1334 | } | |
1335 | ||
1336 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1337 | return true; | |
1338 | } | |
1339 | ||
1340 | if (bytes_xmit > 0) { | |
23b7576d | 1341 | stat64_add(&ram_atomic_counters.normal, 1); |
059ff0fb | 1342 | } else if (bytes_xmit == 0) { |
23b7576d | 1343 | stat64_add(&ram_atomic_counters.duplicate, 1); |
059ff0fb XG |
1344 | } |
1345 | ||
1346 | return true; | |
1347 | } | |
1348 | ||
65dacaa0 XG |
1349 | /* |
1350 | * directly send the page to the stream | |
1351 | * | |
1352 | * Returns the number of pages written. | |
1353 | * | |
ec6f3ab9 | 1354 | * @pss: current PSS channel |
65dacaa0 XG |
1355 | * @block: block that contains the page we want to send |
1356 | * @offset: offset inside the block for the page | |
1357 | * @buf: the page to be sent | |
1358 | * @async: send to page asyncly | |
1359 | */ | |
ec6f3ab9 | 1360 | static int save_normal_page(PageSearchStatus *pss, RAMBlock *block, |
61717ea9 | 1361 | ram_addr_t offset, uint8_t *buf, bool async) |
65dacaa0 | 1362 | { |
ec6f3ab9 PX |
1363 | QEMUFile *file = pss->pss_channel; |
1364 | ||
1365 | ram_transferred_add(save_page_header(pss, block, | |
4c2d0f6d | 1366 | offset | RAM_SAVE_FLAG_PAGE)); |
65dacaa0 | 1367 | if (async) { |
61717ea9 | 1368 | qemu_put_buffer_async(file, buf, TARGET_PAGE_SIZE, |
f912ec5b | 1369 | migrate_release_ram() && |
65dacaa0 XG |
1370 | migration_in_postcopy()); |
1371 | } else { | |
61717ea9 | 1372 | qemu_put_buffer(file, buf, TARGET_PAGE_SIZE); |
65dacaa0 | 1373 | } |
4c2d0f6d | 1374 | ram_transferred_add(TARGET_PAGE_SIZE); |
23b7576d | 1375 | stat64_add(&ram_atomic_counters.normal, 1); |
65dacaa0 XG |
1376 | return 1; |
1377 | } | |
1378 | ||
56e93d26 | 1379 | /** |
3d0684b2 | 1380 | * ram_save_page: send the given page to the stream |
56e93d26 | 1381 | * |
3d0684b2 | 1382 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1383 | * < 0 - error |
1384 | * >=0 - Number of pages written - this might legally be 0 | |
1385 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1386 | * |
6f37bb8b | 1387 | * @rs: current RAM state |
56e93d26 JQ |
1388 | * @block: block that contains the page we want to send |
1389 | * @offset: offset inside the block for the page | |
56e93d26 | 1390 | */ |
05931ec5 | 1391 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss) |
56e93d26 JQ |
1392 | { |
1393 | int pages = -1; | |
56e93d26 | 1394 | uint8_t *p; |
56e93d26 | 1395 | bool send_async = true; |
a08f6890 | 1396 | RAMBlock *block = pss->block; |
8bba004c | 1397 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1398 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1399 | |
2f68e399 | 1400 | p = block->host + offset; |
1db9d8e5 | 1401 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1402 | |
56e93d26 | 1403 | XBZRLE_cache_lock(); |
1a373522 | 1404 | if (rs->xbzrle_enabled && !migration_in_postcopy()) { |
ec6f3ab9 | 1405 | pages = save_xbzrle_page(rs, pss, &p, current_addr, |
61717ea9 | 1406 | block, offset); |
05931ec5 | 1407 | if (!rs->last_stage) { |
059ff0fb XG |
1408 | /* Can't send this cached data async, since the cache page |
1409 | * might get updated before it gets to the wire | |
56e93d26 | 1410 | */ |
059ff0fb | 1411 | send_async = false; |
56e93d26 JQ |
1412 | } |
1413 | } | |
1414 | ||
1415 | /* XBZRLE overflow or normal page */ | |
1416 | if (pages == -1) { | |
ec6f3ab9 | 1417 | pages = save_normal_page(pss, block, offset, p, send_async); |
56e93d26 JQ |
1418 | } |
1419 | ||
1420 | XBZRLE_cache_unlock(); | |
1421 | ||
1422 | return pages; | |
1423 | } | |
1424 | ||
61717ea9 | 1425 | static int ram_save_multifd_page(QEMUFile *file, RAMBlock *block, |
b9ee2f7d JQ |
1426 | ram_addr_t offset) |
1427 | { | |
61717ea9 | 1428 | if (multifd_queue_page(file, block, offset) < 0) { |
713f762a IR |
1429 | return -1; |
1430 | } | |
23b7576d | 1431 | stat64_add(&ram_atomic_counters.normal, 1); |
b9ee2f7d JQ |
1432 | |
1433 | return 1; | |
1434 | } | |
1435 | ||
5e5fdcff | 1436 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 1437 | ram_addr_t offset, uint8_t *source_buf) |
56e93d26 | 1438 | { |
53518d94 | 1439 | RAMState *rs = ram_state; |
ec6f3ab9 | 1440 | PageSearchStatus *pss = &rs->pss[RAM_CHANNEL_PRECOPY]; |
20d549cb | 1441 | uint8_t *p = block->host + offset; |
6ef3771c | 1442 | int ret; |
56e93d26 | 1443 | |
ec6f3ab9 | 1444 | if (save_zero_page_to_file(pss, block, offset)) { |
e7f2e190 | 1445 | return true; |
5e5fdcff XG |
1446 | } |
1447 | ||
ec6f3ab9 | 1448 | save_page_header(pss, block, offset | RAM_SAVE_FLAG_COMPRESS_PAGE); |
34ab9e97 XG |
1449 | |
1450 | /* | |
1451 | * copy it to a internal buffer to avoid it being modified by VM | |
1452 | * so that we can catch up the error during compression and | |
1453 | * decompression | |
1454 | */ | |
1455 | memcpy(source_buf, p, TARGET_PAGE_SIZE); | |
6ef3771c XG |
1456 | ret = qemu_put_compression_data(f, stream, source_buf, TARGET_PAGE_SIZE); |
1457 | if (ret < 0) { | |
1458 | qemu_file_set_error(migrate_get_current()->to_dst_file, ret); | |
b3be2896 | 1459 | error_report("compressed data failed!"); |
b3be2896 | 1460 | } |
e7f2e190 | 1461 | return false; |
5e5fdcff XG |
1462 | } |
1463 | ||
1464 | static void | |
1465 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1466 | { | |
4c2d0f6d | 1467 | ram_transferred_add(bytes_xmit); |
76e03000 | 1468 | |
5e5fdcff | 1469 | if (param->zero_page) { |
23b7576d | 1470 | stat64_add(&ram_atomic_counters.duplicate, 1); |
76e03000 | 1471 | return; |
5e5fdcff | 1472 | } |
76e03000 XG |
1473 | |
1474 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1475 | compression_counters.compressed_size += bytes_xmit - 8; | |
1476 | compression_counters.pages++; | |
56e93d26 JQ |
1477 | } |
1478 | ||
32b05495 XG |
1479 | static bool save_page_use_compression(RAMState *rs); |
1480 | ||
ce25d337 | 1481 | static void flush_compressed_data(RAMState *rs) |
56e93d26 | 1482 | { |
eaa238ab | 1483 | MigrationState *ms = migrate_get_current(); |
56e93d26 JQ |
1484 | int idx, len, thread_count; |
1485 | ||
32b05495 | 1486 | if (!save_page_use_compression(rs)) { |
56e93d26 JQ |
1487 | return; |
1488 | } | |
1489 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1490 | |
0d9f9a5c | 1491 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1492 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1493 | while (!comp_param[idx].done) { |
0d9f9a5c | 1494 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1495 | } |
a7a9a88f | 1496 | } |
0d9f9a5c | 1497 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1498 | |
1499 | for (idx = 0; idx < thread_count; idx++) { | |
1500 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1501 | if (!comp_param[idx].quit) { |
eaa238ab | 1502 | len = qemu_put_qemu_file(ms->to_dst_file, comp_param[idx].file); |
5e5fdcff XG |
1503 | /* |
1504 | * it's safe to fetch zero_page without holding comp_done_lock | |
1505 | * as there is no further request submitted to the thread, | |
1506 | * i.e, the thread should be waiting for a request at this point. | |
1507 | */ | |
1508 | update_compress_thread_counts(&comp_param[idx], len); | |
56e93d26 | 1509 | } |
a7a9a88f | 1510 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1511 | } |
1512 | } | |
1513 | ||
1514 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1515 | ram_addr_t offset) | |
1516 | { | |
1517 | param->block = block; | |
1518 | param->offset = offset; | |
1519 | } | |
1520 | ||
eaa238ab | 1521 | static int compress_page_with_multi_thread(RAMBlock *block, ram_addr_t offset) |
56e93d26 JQ |
1522 | { |
1523 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1d58872a | 1524 | bool wait = migrate_compress_wait_thread(); |
eaa238ab | 1525 | MigrationState *ms = migrate_get_current(); |
56e93d26 JQ |
1526 | |
1527 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1528 | qemu_mutex_lock(&comp_done_lock); |
1d58872a XG |
1529 | retry: |
1530 | for (idx = 0; idx < thread_count; idx++) { | |
1531 | if (comp_param[idx].done) { | |
1532 | comp_param[idx].done = false; | |
eaa238ab PX |
1533 | bytes_xmit = qemu_put_qemu_file(ms->to_dst_file, |
1534 | comp_param[idx].file); | |
1d58872a XG |
1535 | qemu_mutex_lock(&comp_param[idx].mutex); |
1536 | set_compress_params(&comp_param[idx], block, offset); | |
1537 | qemu_cond_signal(&comp_param[idx].cond); | |
1538 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
1539 | pages = 1; | |
5e5fdcff | 1540 | update_compress_thread_counts(&comp_param[idx], bytes_xmit); |
56e93d26 | 1541 | break; |
56e93d26 JQ |
1542 | } |
1543 | } | |
1d58872a XG |
1544 | |
1545 | /* | |
1546 | * wait for the free thread if the user specifies 'compress-wait-thread', | |
1547 | * otherwise we will post the page out in the main thread as normal page. | |
1548 | */ | |
1549 | if (pages < 0 && wait) { | |
1550 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); | |
1551 | goto retry; | |
1552 | } | |
0d9f9a5c | 1553 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1554 | |
1555 | return pages; | |
1556 | } | |
1557 | ||
31e2ac74 JQ |
1558 | #define PAGE_ALL_CLEAN 0 |
1559 | #define PAGE_TRY_AGAIN 1 | |
1560 | #define PAGE_DIRTY_FOUND 2 | |
3d0684b2 JQ |
1561 | /** |
1562 | * find_dirty_block: find the next dirty page and update any state | |
1563 | * associated with the search process. | |
b9e60928 | 1564 | * |
31e2ac74 JQ |
1565 | * Returns: |
1566 | * PAGE_ALL_CLEAN: no dirty page found, give up | |
1567 | * PAGE_TRY_AGAIN: no dirty page found, retry for next block | |
1568 | * PAGE_DIRTY_FOUND: dirty page found | |
b9e60928 | 1569 | * |
6f37bb8b | 1570 | * @rs: current RAM state |
3d0684b2 JQ |
1571 | * @pss: data about the state of the current dirty page scan |
1572 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1573 | */ |
31e2ac74 | 1574 | static int find_dirty_block(RAMState *rs, PageSearchStatus *pss) |
b9e60928 | 1575 | { |
d9e474ea PX |
1576 | /* Update pss->page for the next dirty bit in ramblock */ |
1577 | pss_find_next_dirty(pss); | |
1578 | ||
6f37bb8b | 1579 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1580 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1581 | /* |
1582 | * We've been once around the RAM and haven't found anything. | |
1583 | * Give up. | |
1584 | */ | |
31e2ac74 | 1585 | return PAGE_ALL_CLEAN; |
b9e60928 | 1586 | } |
542147f4 DH |
1587 | if (!offset_in_ramblock(pss->block, |
1588 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)) { | |
b9e60928 | 1589 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1590 | pss->page = 0; |
b9e60928 DDAG |
1591 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1592 | if (!pss->block) { | |
48df9d80 XG |
1593 | /* |
1594 | * If memory migration starts over, we will meet a dirtied page | |
1595 | * which may still exists in compression threads's ring, so we | |
1596 | * should flush the compressed data to make sure the new page | |
1597 | * is not overwritten by the old one in the destination. | |
1598 | * | |
1599 | * Also If xbzrle is on, stop using the data compression at this | |
1600 | * point. In theory, xbzrle can do better than compression. | |
1601 | */ | |
1602 | flush_compressed_data(rs); | |
1603 | ||
b9e60928 DDAG |
1604 | /* Hit the end of the list */ |
1605 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1606 | /* Flag that we've looped */ | |
1607 | pss->complete_round = true; | |
1a373522 DH |
1608 | /* After the first round, enable XBZRLE. */ |
1609 | if (migrate_use_xbzrle()) { | |
1610 | rs->xbzrle_enabled = true; | |
1611 | } | |
b9e60928 DDAG |
1612 | } |
1613 | /* Didn't find anything this time, but try again on the new block */ | |
31e2ac74 | 1614 | return PAGE_TRY_AGAIN; |
b9e60928 | 1615 | } else { |
31e2ac74 JQ |
1616 | /* We've found something */ |
1617 | return PAGE_DIRTY_FOUND; | |
b9e60928 DDAG |
1618 | } |
1619 | } | |
1620 | ||
3d0684b2 JQ |
1621 | /** |
1622 | * unqueue_page: gets a page of the queue | |
1623 | * | |
a82d593b | 1624 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1625 | * |
3d0684b2 JQ |
1626 | * Returns the block of the page (or NULL if none available) |
1627 | * | |
ec481c6c | 1628 | * @rs: current RAM state |
3d0684b2 | 1629 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1630 | */ |
f20e2865 | 1631 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b | 1632 | { |
a1fe28df | 1633 | struct RAMSrcPageRequest *entry; |
a82d593b DDAG |
1634 | RAMBlock *block = NULL; |
1635 | ||
a1fe28df | 1636 | if (!postcopy_has_request(rs)) { |
ae526e32 XG |
1637 | return NULL; |
1638 | } | |
1639 | ||
6e8a355d | 1640 | QEMU_LOCK_GUARD(&rs->src_page_req_mutex); |
a1fe28df PX |
1641 | |
1642 | /* | |
1643 | * This should _never_ change even after we take the lock, because no one | |
1644 | * should be taking anything off the request list other than us. | |
1645 | */ | |
1646 | assert(postcopy_has_request(rs)); | |
1647 | ||
1648 | entry = QSIMPLEQ_FIRST(&rs->src_page_requests); | |
1649 | block = entry->rb; | |
1650 | *offset = entry->offset; | |
1651 | ||
777f53c7 TH |
1652 | if (entry->len > TARGET_PAGE_SIZE) { |
1653 | entry->len -= TARGET_PAGE_SIZE; | |
1654 | entry->offset += TARGET_PAGE_SIZE; | |
a1fe28df PX |
1655 | } else { |
1656 | memory_region_unref(block->mr); | |
1657 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); | |
1658 | g_free(entry); | |
1659 | migration_consume_urgent_request(); | |
a82d593b | 1660 | } |
a82d593b DDAG |
1661 | |
1662 | return block; | |
1663 | } | |
1664 | ||
278e2f55 AG |
1665 | #if defined(__linux__) |
1666 | /** | |
1667 | * poll_fault_page: try to get next UFFD write fault page and, if pending fault | |
1668 | * is found, return RAM block pointer and page offset | |
1669 | * | |
1670 | * Returns pointer to the RAMBlock containing faulting page, | |
1671 | * NULL if no write faults are pending | |
1672 | * | |
1673 | * @rs: current RAM state | |
1674 | * @offset: page offset from the beginning of the block | |
1675 | */ | |
1676 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1677 | { | |
1678 | struct uffd_msg uffd_msg; | |
1679 | void *page_address; | |
82ea3e3b | 1680 | RAMBlock *block; |
278e2f55 AG |
1681 | int res; |
1682 | ||
1683 | if (!migrate_background_snapshot()) { | |
1684 | return NULL; | |
1685 | } | |
1686 | ||
1687 | res = uffd_read_events(rs->uffdio_fd, &uffd_msg, 1); | |
1688 | if (res <= 0) { | |
1689 | return NULL; | |
1690 | } | |
1691 | ||
1692 | page_address = (void *)(uintptr_t) uffd_msg.arg.pagefault.address; | |
82ea3e3b AG |
1693 | block = qemu_ram_block_from_host(page_address, false, offset); |
1694 | assert(block && (block->flags & RAM_UF_WRITEPROTECT) != 0); | |
1695 | return block; | |
278e2f55 AG |
1696 | } |
1697 | ||
1698 | /** | |
1699 | * ram_save_release_protection: release UFFD write protection after | |
1700 | * a range of pages has been saved | |
1701 | * | |
1702 | * @rs: current RAM state | |
1703 | * @pss: page-search-status structure | |
1704 | * @start_page: index of the first page in the range relative to pss->block | |
1705 | * | |
1706 | * Returns 0 on success, negative value in case of an error | |
1707 | */ | |
1708 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1709 | unsigned long start_page) | |
1710 | { | |
1711 | int res = 0; | |
1712 | ||
1713 | /* Check if page is from UFFD-managed region. */ | |
1714 | if (pss->block->flags & RAM_UF_WRITEPROTECT) { | |
1715 | void *page_address = pss->block->host + (start_page << TARGET_PAGE_BITS); | |
258f5c98 | 1716 | uint64_t run_length = (pss->page - start_page) << TARGET_PAGE_BITS; |
278e2f55 AG |
1717 | |
1718 | /* Flush async buffers before un-protect. */ | |
61717ea9 | 1719 | qemu_fflush(pss->pss_channel); |
278e2f55 AG |
1720 | /* Un-protect memory range. */ |
1721 | res = uffd_change_protection(rs->uffdio_fd, page_address, run_length, | |
1722 | false, false); | |
1723 | } | |
1724 | ||
1725 | return res; | |
1726 | } | |
1727 | ||
1728 | /* ram_write_tracking_available: check if kernel supports required UFFD features | |
1729 | * | |
1730 | * Returns true if supports, false otherwise | |
1731 | */ | |
1732 | bool ram_write_tracking_available(void) | |
1733 | { | |
1734 | uint64_t uffd_features; | |
1735 | int res; | |
1736 | ||
1737 | res = uffd_query_features(&uffd_features); | |
1738 | return (res == 0 && | |
1739 | (uffd_features & UFFD_FEATURE_PAGEFAULT_FLAG_WP) != 0); | |
1740 | } | |
1741 | ||
1742 | /* ram_write_tracking_compatible: check if guest configuration is | |
1743 | * compatible with 'write-tracking' | |
1744 | * | |
1745 | * Returns true if compatible, false otherwise | |
1746 | */ | |
1747 | bool ram_write_tracking_compatible(void) | |
1748 | { | |
1749 | const uint64_t uffd_ioctls_mask = BIT(_UFFDIO_WRITEPROTECT); | |
1750 | int uffd_fd; | |
82ea3e3b | 1751 | RAMBlock *block; |
278e2f55 AG |
1752 | bool ret = false; |
1753 | ||
1754 | /* Open UFFD file descriptor */ | |
1755 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, false); | |
1756 | if (uffd_fd < 0) { | |
1757 | return false; | |
1758 | } | |
1759 | ||
1760 | RCU_READ_LOCK_GUARD(); | |
1761 | ||
82ea3e3b | 1762 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 AG |
1763 | uint64_t uffd_ioctls; |
1764 | ||
1765 | /* Nothing to do with read-only and MMIO-writable regions */ | |
82ea3e3b | 1766 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1767 | continue; |
1768 | } | |
1769 | /* Try to register block memory via UFFD-IO to track writes */ | |
82ea3e3b | 1770 | if (uffd_register_memory(uffd_fd, block->host, block->max_length, |
278e2f55 AG |
1771 | UFFDIO_REGISTER_MODE_WP, &uffd_ioctls)) { |
1772 | goto out; | |
1773 | } | |
1774 | if ((uffd_ioctls & uffd_ioctls_mask) != uffd_ioctls_mask) { | |
1775 | goto out; | |
1776 | } | |
1777 | } | |
1778 | ret = true; | |
1779 | ||
1780 | out: | |
1781 | uffd_close_fd(uffd_fd); | |
1782 | return ret; | |
1783 | } | |
1784 | ||
f7b9dcfb DH |
1785 | static inline void populate_read_range(RAMBlock *block, ram_addr_t offset, |
1786 | ram_addr_t size) | |
1787 | { | |
5f19a449 DH |
1788 | const ram_addr_t end = offset + size; |
1789 | ||
f7b9dcfb DH |
1790 | /* |
1791 | * We read one byte of each page; this will preallocate page tables if | |
1792 | * required and populate the shared zeropage on MAP_PRIVATE anonymous memory | |
1793 | * where no page was populated yet. This might require adaption when | |
1794 | * supporting other mappings, like shmem. | |
1795 | */ | |
5f19a449 | 1796 | for (; offset < end; offset += block->page_size) { |
f7b9dcfb DH |
1797 | char tmp = *((char *)block->host + offset); |
1798 | ||
1799 | /* Don't optimize the read out */ | |
1800 | asm volatile("" : "+r" (tmp)); | |
1801 | } | |
1802 | } | |
1803 | ||
6fee3a1f DH |
1804 | static inline int populate_read_section(MemoryRegionSection *section, |
1805 | void *opaque) | |
1806 | { | |
1807 | const hwaddr size = int128_get64(section->size); | |
1808 | hwaddr offset = section->offset_within_region; | |
1809 | RAMBlock *block = section->mr->ram_block; | |
1810 | ||
1811 | populate_read_range(block, offset, size); | |
1812 | return 0; | |
1813 | } | |
1814 | ||
eeccb99c | 1815 | /* |
f7b9dcfb DH |
1816 | * ram_block_populate_read: preallocate page tables and populate pages in the |
1817 | * RAM block by reading a byte of each page. | |
eeccb99c AG |
1818 | * |
1819 | * Since it's solely used for userfault_fd WP feature, here we just | |
1820 | * hardcode page size to qemu_real_host_page_size. | |
1821 | * | |
82ea3e3b | 1822 | * @block: RAM block to populate |
eeccb99c | 1823 | */ |
6fee3a1f | 1824 | static void ram_block_populate_read(RAMBlock *rb) |
eeccb99c | 1825 | { |
6fee3a1f DH |
1826 | /* |
1827 | * Skip populating all pages that fall into a discarded range as managed by | |
1828 | * a RamDiscardManager responsible for the mapped memory region of the | |
1829 | * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock | |
1830 | * must not get populated automatically. We don't have to track | |
1831 | * modifications via userfaultfd WP reliably, because these pages will | |
1832 | * not be part of the migration stream either way -- see | |
1833 | * ramblock_dirty_bitmap_exclude_discarded_pages(). | |
1834 | * | |
1835 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1836 | */ | |
1837 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1838 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1839 | MemoryRegionSection section = { | |
1840 | .mr = rb->mr, | |
1841 | .offset_within_region = 0, | |
1842 | .size = rb->mr->size, | |
1843 | }; | |
1844 | ||
1845 | ram_discard_manager_replay_populated(rdm, §ion, | |
1846 | populate_read_section, NULL); | |
1847 | } else { | |
1848 | populate_read_range(rb, 0, rb->used_length); | |
1849 | } | |
eeccb99c AG |
1850 | } |
1851 | ||
1852 | /* | |
1853 | * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking | |
1854 | */ | |
1855 | void ram_write_tracking_prepare(void) | |
1856 | { | |
82ea3e3b | 1857 | RAMBlock *block; |
eeccb99c AG |
1858 | |
1859 | RCU_READ_LOCK_GUARD(); | |
1860 | ||
82ea3e3b | 1861 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
eeccb99c | 1862 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1863 | if (block->mr->readonly || block->mr->rom_device) { |
eeccb99c AG |
1864 | continue; |
1865 | } | |
1866 | ||
1867 | /* | |
1868 | * Populate pages of the RAM block before enabling userfault_fd | |
1869 | * write protection. | |
1870 | * | |
1871 | * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with | |
1872 | * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip | |
1873 | * pages with pte_none() entries in page table. | |
1874 | */ | |
f7b9dcfb | 1875 | ram_block_populate_read(block); |
eeccb99c AG |
1876 | } |
1877 | } | |
1878 | ||
e41c5770 DH |
1879 | static inline int uffd_protect_section(MemoryRegionSection *section, |
1880 | void *opaque) | |
1881 | { | |
1882 | const hwaddr size = int128_get64(section->size); | |
1883 | const hwaddr offset = section->offset_within_region; | |
1884 | RAMBlock *rb = section->mr->ram_block; | |
1885 | int uffd_fd = (uintptr_t)opaque; | |
1886 | ||
1887 | return uffd_change_protection(uffd_fd, rb->host + offset, size, true, | |
1888 | false); | |
1889 | } | |
1890 | ||
1891 | static int ram_block_uffd_protect(RAMBlock *rb, int uffd_fd) | |
1892 | { | |
1893 | assert(rb->flags & RAM_UF_WRITEPROTECT); | |
1894 | ||
1895 | /* See ram_block_populate_read() */ | |
1896 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1897 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1898 | MemoryRegionSection section = { | |
1899 | .mr = rb->mr, | |
1900 | .offset_within_region = 0, | |
1901 | .size = rb->mr->size, | |
1902 | }; | |
1903 | ||
1904 | return ram_discard_manager_replay_populated(rdm, §ion, | |
1905 | uffd_protect_section, | |
1906 | (void *)(uintptr_t)uffd_fd); | |
1907 | } | |
1908 | return uffd_change_protection(uffd_fd, rb->host, | |
1909 | rb->used_length, true, false); | |
1910 | } | |
1911 | ||
278e2f55 AG |
1912 | /* |
1913 | * ram_write_tracking_start: start UFFD-WP memory tracking | |
1914 | * | |
1915 | * Returns 0 for success or negative value in case of error | |
1916 | */ | |
1917 | int ram_write_tracking_start(void) | |
1918 | { | |
1919 | int uffd_fd; | |
1920 | RAMState *rs = ram_state; | |
82ea3e3b | 1921 | RAMBlock *block; |
278e2f55 AG |
1922 | |
1923 | /* Open UFFD file descriptor */ | |
1924 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, true); | |
1925 | if (uffd_fd < 0) { | |
1926 | return uffd_fd; | |
1927 | } | |
1928 | rs->uffdio_fd = uffd_fd; | |
1929 | ||
1930 | RCU_READ_LOCK_GUARD(); | |
1931 | ||
82ea3e3b | 1932 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 | 1933 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1934 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1935 | continue; |
1936 | } | |
1937 | ||
1938 | /* Register block memory with UFFD to track writes */ | |
82ea3e3b AG |
1939 | if (uffd_register_memory(rs->uffdio_fd, block->host, |
1940 | block->max_length, UFFDIO_REGISTER_MODE_WP, NULL)) { | |
278e2f55 AG |
1941 | goto fail; |
1942 | } | |
72ef3a37 DH |
1943 | block->flags |= RAM_UF_WRITEPROTECT; |
1944 | memory_region_ref(block->mr); | |
1945 | ||
278e2f55 | 1946 | /* Apply UFFD write protection to the block memory range */ |
e41c5770 | 1947 | if (ram_block_uffd_protect(block, uffd_fd)) { |
278e2f55 AG |
1948 | goto fail; |
1949 | } | |
278e2f55 | 1950 | |
82ea3e3b AG |
1951 | trace_ram_write_tracking_ramblock_start(block->idstr, block->page_size, |
1952 | block->host, block->max_length); | |
278e2f55 AG |
1953 | } |
1954 | ||
1955 | return 0; | |
1956 | ||
1957 | fail: | |
1958 | error_report("ram_write_tracking_start() failed: restoring initial memory state"); | |
1959 | ||
82ea3e3b AG |
1960 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1961 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1962 | continue; |
1963 | } | |
82ea3e3b | 1964 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 1965 | /* Cleanup flags and remove reference */ |
82ea3e3b AG |
1966 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1967 | memory_region_unref(block->mr); | |
278e2f55 AG |
1968 | } |
1969 | ||
1970 | uffd_close_fd(uffd_fd); | |
1971 | rs->uffdio_fd = -1; | |
1972 | return -1; | |
1973 | } | |
1974 | ||
1975 | /** | |
1976 | * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection | |
1977 | */ | |
1978 | void ram_write_tracking_stop(void) | |
1979 | { | |
1980 | RAMState *rs = ram_state; | |
82ea3e3b | 1981 | RAMBlock *block; |
278e2f55 AG |
1982 | |
1983 | RCU_READ_LOCK_GUARD(); | |
1984 | ||
82ea3e3b AG |
1985 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1986 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1987 | continue; |
1988 | } | |
82ea3e3b | 1989 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 1990 | |
82ea3e3b AG |
1991 | trace_ram_write_tracking_ramblock_stop(block->idstr, block->page_size, |
1992 | block->host, block->max_length); | |
278e2f55 AG |
1993 | |
1994 | /* Cleanup flags and remove reference */ | |
82ea3e3b AG |
1995 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1996 | memory_region_unref(block->mr); | |
278e2f55 AG |
1997 | } |
1998 | ||
1999 | /* Finally close UFFD file descriptor */ | |
2000 | uffd_close_fd(rs->uffdio_fd); | |
2001 | rs->uffdio_fd = -1; | |
2002 | } | |
2003 | ||
2004 | #else | |
2005 | /* No target OS support, stubs just fail or ignore */ | |
2006 | ||
2007 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
2008 | { | |
2009 | (void) rs; | |
2010 | (void) offset; | |
2011 | ||
2012 | return NULL; | |
2013 | } | |
2014 | ||
2015 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
2016 | unsigned long start_page) | |
2017 | { | |
2018 | (void) rs; | |
2019 | (void) pss; | |
2020 | (void) start_page; | |
2021 | ||
2022 | return 0; | |
2023 | } | |
2024 | ||
2025 | bool ram_write_tracking_available(void) | |
2026 | { | |
2027 | return false; | |
2028 | } | |
2029 | ||
2030 | bool ram_write_tracking_compatible(void) | |
2031 | { | |
2032 | assert(0); | |
2033 | return false; | |
2034 | } | |
2035 | ||
2036 | int ram_write_tracking_start(void) | |
2037 | { | |
2038 | assert(0); | |
2039 | return -1; | |
2040 | } | |
2041 | ||
2042 | void ram_write_tracking_stop(void) | |
2043 | { | |
2044 | assert(0); | |
2045 | } | |
2046 | #endif /* defined(__linux__) */ | |
2047 | ||
3d0684b2 | 2048 | /** |
ff1543af | 2049 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
2050 | * |
2051 | * Skips pages that are already sent (!dirty) | |
a82d593b | 2052 | * |
a5f7b1a6 | 2053 | * Returns true if a queued page is found |
a82d593b | 2054 | * |
6f37bb8b | 2055 | * @rs: current RAM state |
3d0684b2 | 2056 | * @pss: data about the state of the current dirty page scan |
a82d593b | 2057 | */ |
f20e2865 | 2058 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
2059 | { |
2060 | RAMBlock *block; | |
2061 | ram_addr_t offset; | |
777f53c7 TH |
2062 | bool dirty; |
2063 | ||
2064 | do { | |
2065 | block = unqueue_page(rs, &offset); | |
2066 | /* | |
2067 | * We're sending this page, and since it's postcopy nothing else | |
2068 | * will dirty it, and we must make sure it doesn't get sent again | |
2069 | * even if this queue request was received after the background | |
2070 | * search already sent it. | |
2071 | */ | |
2072 | if (block) { | |
2073 | unsigned long page; | |
2074 | ||
2075 | page = offset >> TARGET_PAGE_BITS; | |
2076 | dirty = test_bit(page, block->bmap); | |
2077 | if (!dirty) { | |
2078 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, | |
2079 | page); | |
2080 | } else { | |
2081 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); | |
2082 | } | |
2083 | } | |
a82d593b | 2084 | |
777f53c7 | 2085 | } while (block && !dirty); |
a82d593b | 2086 | |
b062106d | 2087 | if (!block) { |
278e2f55 AG |
2088 | /* |
2089 | * Poll write faults too if background snapshot is enabled; that's | |
2090 | * when we have vcpus got blocked by the write protected pages. | |
2091 | */ | |
2092 | block = poll_fault_page(rs, &offset); | |
2093 | } | |
2094 | ||
a82d593b | 2095 | if (block) { |
a82d593b DDAG |
2096 | /* |
2097 | * We want the background search to continue from the queued page | |
2098 | * since the guest is likely to want other pages near to the page | |
2099 | * it just requested. | |
2100 | */ | |
2101 | pss->block = block; | |
a935e30f | 2102 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
2103 | |
2104 | /* | |
2105 | * This unqueued page would break the "one round" check, even is | |
2106 | * really rare. | |
2107 | */ | |
2108 | pss->complete_round = false; | |
a82d593b DDAG |
2109 | } |
2110 | ||
2111 | return !!block; | |
2112 | } | |
2113 | ||
6c595cde | 2114 | /** |
5e58f968 JQ |
2115 | * migration_page_queue_free: drop any remaining pages in the ram |
2116 | * request queue | |
6c595cde | 2117 | * |
3d0684b2 JQ |
2118 | * It should be empty at the end anyway, but in error cases there may |
2119 | * be some left. in case that there is any page left, we drop it. | |
2120 | * | |
6c595cde | 2121 | */ |
83c13382 | 2122 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 2123 | { |
ec481c6c | 2124 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
2125 | /* This queue generally should be empty - but in the case of a failed |
2126 | * migration might have some droppings in. | |
2127 | */ | |
89ac5a1d | 2128 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 2129 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 2130 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 2131 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
2132 | g_free(mspr); |
2133 | } | |
6c595cde DDAG |
2134 | } |
2135 | ||
2136 | /** | |
3d0684b2 JQ |
2137 | * ram_save_queue_pages: queue the page for transmission |
2138 | * | |
2139 | * A request from postcopy destination for example. | |
2140 | * | |
2141 | * Returns zero on success or negative on error | |
2142 | * | |
3d0684b2 JQ |
2143 | * @rbname: Name of the RAMBLock of the request. NULL means the |
2144 | * same that last one. | |
2145 | * @start: starting address from the start of the RAMBlock | |
2146 | * @len: length (in bytes) to send | |
6c595cde | 2147 | */ |
96506894 | 2148 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
2149 | { |
2150 | RAMBlock *ramblock; | |
53518d94 | 2151 | RAMState *rs = ram_state; |
6c595cde | 2152 | |
9360447d | 2153 | ram_counters.postcopy_requests++; |
89ac5a1d DDAG |
2154 | RCU_READ_LOCK_GUARD(); |
2155 | ||
6c595cde DDAG |
2156 | if (!rbname) { |
2157 | /* Reuse last RAMBlock */ | |
68a098f3 | 2158 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
2159 | |
2160 | if (!ramblock) { | |
2161 | /* | |
2162 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
2163 | * it's the 1st request. | |
2164 | */ | |
2165 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 2166 | return -1; |
6c595cde DDAG |
2167 | } |
2168 | } else { | |
2169 | ramblock = qemu_ram_block_by_name(rbname); | |
2170 | ||
2171 | if (!ramblock) { | |
2172 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
2173 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 2174 | return -1; |
6c595cde | 2175 | } |
68a098f3 | 2176 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
2177 | } |
2178 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
542147f4 | 2179 | if (!offset_in_ramblock(ramblock, start + len - 1)) { |
9458ad6b JQ |
2180 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
2181 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 2182 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 2183 | return -1; |
6c595cde DDAG |
2184 | } |
2185 | ||
93589827 PX |
2186 | /* |
2187 | * When with postcopy preempt, we send back the page directly in the | |
2188 | * rp-return thread. | |
2189 | */ | |
2190 | if (postcopy_preempt_active()) { | |
2191 | ram_addr_t page_start = start >> TARGET_PAGE_BITS; | |
2192 | size_t page_size = qemu_ram_pagesize(ramblock); | |
2193 | PageSearchStatus *pss = &ram_state->pss[RAM_CHANNEL_POSTCOPY]; | |
2194 | int ret = 0; | |
2195 | ||
2196 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2197 | ||
2198 | pss_init(pss, ramblock, page_start); | |
2199 | /* | |
2200 | * Always use the preempt channel, and make sure it's there. It's | |
2201 | * safe to access without lock, because when rp-thread is running | |
2202 | * we should be the only one who operates on the qemufile | |
2203 | */ | |
2204 | pss->pss_channel = migrate_get_current()->postcopy_qemufile_src; | |
93589827 PX |
2205 | assert(pss->pss_channel); |
2206 | ||
2207 | /* | |
2208 | * It must be either one or multiple of host page size. Just | |
2209 | * assert; if something wrong we're mostly split brain anyway. | |
2210 | */ | |
2211 | assert(len % page_size == 0); | |
2212 | while (len) { | |
2213 | if (ram_save_host_page_urgent(pss)) { | |
2214 | error_report("%s: ram_save_host_page_urgent() failed: " | |
2215 | "ramblock=%s, start_addr=0x"RAM_ADDR_FMT, | |
2216 | __func__, ramblock->idstr, start); | |
2217 | ret = -1; | |
2218 | break; | |
2219 | } | |
2220 | /* | |
2221 | * NOTE: after ram_save_host_page_urgent() succeeded, pss->page | |
2222 | * will automatically be moved and point to the next host page | |
2223 | * we're going to send, so no need to update here. | |
2224 | * | |
2225 | * Normally QEMU never sends >1 host page in requests, so | |
2226 | * logically we don't even need that as the loop should only | |
2227 | * run once, but just to be consistent. | |
2228 | */ | |
2229 | len -= page_size; | |
2230 | }; | |
2231 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
2232 | ||
2233 | return ret; | |
2234 | } | |
2235 | ||
ec481c6c | 2236 | struct RAMSrcPageRequest *new_entry = |
b21e2380 | 2237 | g_new0(struct RAMSrcPageRequest, 1); |
6c595cde DDAG |
2238 | new_entry->rb = ramblock; |
2239 | new_entry->offset = start; | |
2240 | new_entry->len = len; | |
2241 | ||
2242 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
2243 | qemu_mutex_lock(&rs->src_page_req_mutex); |
2244 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 2245 | migration_make_urgent_request(); |
ec481c6c | 2246 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
2247 | |
2248 | return 0; | |
6c595cde DDAG |
2249 | } |
2250 | ||
d7400a34 XG |
2251 | static bool save_page_use_compression(RAMState *rs) |
2252 | { | |
2253 | if (!migrate_use_compression()) { | |
2254 | return false; | |
2255 | } | |
2256 | ||
2257 | /* | |
1a373522 DH |
2258 | * If xbzrle is enabled (e.g., after first round of migration), stop |
2259 | * using the data compression. In theory, xbzrle can do better than | |
2260 | * compression. | |
d7400a34 | 2261 | */ |
1a373522 DH |
2262 | if (rs->xbzrle_enabled) { |
2263 | return false; | |
d7400a34 XG |
2264 | } |
2265 | ||
1a373522 | 2266 | return true; |
d7400a34 XG |
2267 | } |
2268 | ||
5e5fdcff XG |
2269 | /* |
2270 | * try to compress the page before posting it out, return true if the page | |
2271 | * has been properly handled by compression, otherwise needs other | |
2272 | * paths to handle it | |
2273 | */ | |
ec6f3ab9 PX |
2274 | static bool save_compress_page(RAMState *rs, PageSearchStatus *pss, |
2275 | RAMBlock *block, ram_addr_t offset) | |
5e5fdcff XG |
2276 | { |
2277 | if (!save_page_use_compression(rs)) { | |
2278 | return false; | |
2279 | } | |
2280 | ||
2281 | /* | |
2282 | * When starting the process of a new block, the first page of | |
2283 | * the block should be sent out before other pages in the same | |
2284 | * block, and all the pages in last block should have been sent | |
2285 | * out, keeping this order is important, because the 'cont' flag | |
2286 | * is used to avoid resending the block name. | |
2287 | * | |
2288 | * We post the fist page as normal page as compression will take | |
2289 | * much CPU resource. | |
2290 | */ | |
ec6f3ab9 | 2291 | if (block != pss->last_sent_block) { |
5e5fdcff XG |
2292 | flush_compressed_data(rs); |
2293 | return false; | |
2294 | } | |
2295 | ||
eaa238ab | 2296 | if (compress_page_with_multi_thread(block, offset) > 0) { |
5e5fdcff XG |
2297 | return true; |
2298 | } | |
2299 | ||
76e03000 | 2300 | compression_counters.busy++; |
5e5fdcff XG |
2301 | return false; |
2302 | } | |
2303 | ||
a82d593b | 2304 | /** |
4010ba38 | 2305 | * ram_save_target_page_legacy: save one target page |
a82d593b | 2306 | * |
3d0684b2 | 2307 | * Returns the number of pages written |
a82d593b | 2308 | * |
6f37bb8b | 2309 | * @rs: current RAM state |
3d0684b2 | 2310 | * @pss: data about the page we want to send |
a82d593b | 2311 | */ |
4010ba38 | 2312 | static int ram_save_target_page_legacy(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2313 | { |
a8ec91f9 | 2314 | RAMBlock *block = pss->block; |
8bba004c | 2315 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
2316 | int res; |
2317 | ||
61717ea9 | 2318 | if (control_save_page(pss, block, offset, &res)) { |
a8ec91f9 XG |
2319 | return res; |
2320 | } | |
2321 | ||
ec6f3ab9 | 2322 | if (save_compress_page(rs, pss, block, offset)) { |
5e5fdcff | 2323 | return 1; |
d7400a34 XG |
2324 | } |
2325 | ||
ec6f3ab9 | 2326 | res = save_zero_page(pss, block, offset); |
d7400a34 XG |
2327 | if (res > 0) { |
2328 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
2329 | * page would be stale | |
2330 | */ | |
ef5c3d13 | 2331 | if (rs->xbzrle_enabled) { |
d7400a34 XG |
2332 | XBZRLE_cache_lock(); |
2333 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
2334 | XBZRLE_cache_unlock(); | |
2335 | } | |
d7400a34 XG |
2336 | return res; |
2337 | } | |
2338 | ||
da3f56cb | 2339 | /* |
6f39c90b PX |
2340 | * Do not use multifd in postcopy as one whole host page should be |
2341 | * placed. Meanwhile postcopy requires atomic update of pages, so even | |
2342 | * if host page size == guest page size the dest guest during run may | |
2343 | * still see partially copied pages which is data corruption. | |
da3f56cb | 2344 | */ |
6f39c90b | 2345 | if (migrate_use_multifd() && !migration_in_postcopy()) { |
61717ea9 | 2346 | return ram_save_multifd_page(pss->pss_channel, block, offset); |
a82d593b DDAG |
2347 | } |
2348 | ||
05931ec5 | 2349 | return ram_save_page(rs, pss); |
a82d593b DDAG |
2350 | } |
2351 | ||
d9e474ea PX |
2352 | /* Should be called before sending a host page */ |
2353 | static void pss_host_page_prepare(PageSearchStatus *pss) | |
2354 | { | |
2355 | /* How many guest pages are there in one host page? */ | |
2356 | size_t guest_pfns = qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
2357 | ||
2358 | pss->host_page_sending = true; | |
301d7ffe PX |
2359 | if (guest_pfns <= 1) { |
2360 | /* | |
2361 | * This covers both when guest psize == host psize, or when guest | |
2362 | * has larger psize than the host (guest_pfns==0). | |
2363 | * | |
2364 | * For the latter, we always send one whole guest page per | |
2365 | * iteration of the host page (example: an Alpha VM on x86 host | |
2366 | * will have guest psize 8K while host psize 4K). | |
2367 | */ | |
2368 | pss->host_page_start = pss->page; | |
2369 | pss->host_page_end = pss->page + 1; | |
2370 | } else { | |
2371 | /* | |
2372 | * The host page spans over multiple guest pages, we send them | |
2373 | * within the same host page iteration. | |
2374 | */ | |
2375 | pss->host_page_start = ROUND_DOWN(pss->page, guest_pfns); | |
2376 | pss->host_page_end = ROUND_UP(pss->page + 1, guest_pfns); | |
2377 | } | |
d9e474ea PX |
2378 | } |
2379 | ||
2380 | /* | |
2381 | * Whether the page pointed by PSS is within the host page being sent. | |
2382 | * Must be called after a previous pss_host_page_prepare(). | |
2383 | */ | |
2384 | static bool pss_within_range(PageSearchStatus *pss) | |
2385 | { | |
2386 | ram_addr_t ram_addr; | |
2387 | ||
2388 | assert(pss->host_page_sending); | |
2389 | ||
2390 | /* Over host-page boundary? */ | |
2391 | if (pss->page >= pss->host_page_end) { | |
2392 | return false; | |
2393 | } | |
2394 | ||
2395 | ram_addr = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; | |
2396 | ||
2397 | return offset_in_ramblock(pss->block, ram_addr); | |
2398 | } | |
2399 | ||
2400 | static void pss_host_page_finish(PageSearchStatus *pss) | |
2401 | { | |
2402 | pss->host_page_sending = false; | |
2403 | /* This is not needed, but just to reset it */ | |
2404 | pss->host_page_start = pss->host_page_end = 0; | |
2405 | } | |
2406 | ||
93589827 PX |
2407 | /* |
2408 | * Send an urgent host page specified by `pss'. Need to be called with | |
2409 | * bitmap_mutex held. | |
2410 | * | |
2411 | * Returns 0 if save host page succeeded, false otherwise. | |
2412 | */ | |
2413 | static int ram_save_host_page_urgent(PageSearchStatus *pss) | |
2414 | { | |
2415 | bool page_dirty, sent = false; | |
2416 | RAMState *rs = ram_state; | |
2417 | int ret = 0; | |
2418 | ||
2419 | trace_postcopy_preempt_send_host_page(pss->block->idstr, pss->page); | |
2420 | pss_host_page_prepare(pss); | |
2421 | ||
2422 | /* | |
2423 | * If precopy is sending the same page, let it be done in precopy, or | |
2424 | * we could send the same page in two channels and none of them will | |
2425 | * receive the whole page. | |
2426 | */ | |
2427 | if (pss_overlap(pss, &ram_state->pss[RAM_CHANNEL_PRECOPY])) { | |
2428 | trace_postcopy_preempt_hit(pss->block->idstr, | |
2429 | pss->page << TARGET_PAGE_BITS); | |
2430 | return 0; | |
2431 | } | |
2432 | ||
2433 | do { | |
2434 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); | |
2435 | ||
2436 | if (page_dirty) { | |
2437 | /* Be strict to return code; it must be 1, or what else? */ | |
4010ba38 | 2438 | if (migration_ops->ram_save_target_page(rs, pss) != 1) { |
93589827 PX |
2439 | error_report_once("%s: ram_save_target_page failed", __func__); |
2440 | ret = -1; | |
2441 | goto out; | |
2442 | } | |
2443 | sent = true; | |
2444 | } | |
2445 | pss_find_next_dirty(pss); | |
2446 | } while (pss_within_range(pss)); | |
2447 | out: | |
2448 | pss_host_page_finish(pss); | |
2449 | /* For urgent requests, flush immediately if sent */ | |
2450 | if (sent) { | |
2451 | qemu_fflush(pss->pss_channel); | |
2452 | } | |
2453 | return ret; | |
2454 | } | |
2455 | ||
a82d593b | 2456 | /** |
3d0684b2 | 2457 | * ram_save_host_page: save a whole host page |
a82d593b | 2458 | * |
3d0684b2 JQ |
2459 | * Starting at *offset send pages up to the end of the current host |
2460 | * page. It's valid for the initial offset to point into the middle of | |
2461 | * a host page in which case the remainder of the hostpage is sent. | |
2462 | * Only dirty target pages are sent. Note that the host page size may | |
2463 | * be a huge page for this block. | |
f3321554 | 2464 | * |
1eb3fc0a DDAG |
2465 | * The saving stops at the boundary of the used_length of the block |
2466 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 2467 | * |
f3321554 PX |
2468 | * The caller must be with ram_state.bitmap_mutex held to call this |
2469 | * function. Note that this function can temporarily release the lock, but | |
2470 | * when the function is returned it'll make sure the lock is still held. | |
2471 | * | |
3d0684b2 JQ |
2472 | * Returns the number of pages written or negative on error |
2473 | * | |
6f37bb8b | 2474 | * @rs: current RAM state |
3d0684b2 | 2475 | * @pss: data about the page we want to send |
a82d593b | 2476 | */ |
05931ec5 | 2477 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2478 | { |
f3321554 | 2479 | bool page_dirty, preempt_active = postcopy_preempt_active(); |
a82d593b | 2480 | int tmppages, pages = 0; |
a935e30f JQ |
2481 | size_t pagesize_bits = |
2482 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
278e2f55 AG |
2483 | unsigned long start_page = pss->page; |
2484 | int res; | |
4c011c37 | 2485 | |
fbd162e6 | 2486 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
2487 | error_report("block %s should not be migrated !", pss->block->idstr); |
2488 | return 0; | |
2489 | } | |
2490 | ||
d9e474ea PX |
2491 | /* Update host page boundary information */ |
2492 | pss_host_page_prepare(pss); | |
2493 | ||
a82d593b | 2494 | do { |
f3321554 | 2495 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); |
a82d593b | 2496 | |
f3321554 PX |
2497 | /* Check the pages is dirty and if it is send it */ |
2498 | if (page_dirty) { | |
ba1b7c81 | 2499 | /* |
f3321554 PX |
2500 | * Properly yield the lock only in postcopy preempt mode |
2501 | * because both migration thread and rp-return thread can | |
2502 | * operate on the bitmaps. | |
ba1b7c81 | 2503 | */ |
f3321554 PX |
2504 | if (preempt_active) { |
2505 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
ba1b7c81 | 2506 | } |
4010ba38 | 2507 | tmppages = migration_ops->ram_save_target_page(rs, pss); |
f3321554 PX |
2508 | if (tmppages >= 0) { |
2509 | pages += tmppages; | |
2510 | /* | |
2511 | * Allow rate limiting to happen in the middle of huge pages if | |
2512 | * something is sent in the current iteration. | |
2513 | */ | |
2514 | if (pagesize_bits > 1 && tmppages > 0) { | |
2515 | migration_rate_limit(); | |
2516 | } | |
2517 | } | |
2518 | if (preempt_active) { | |
2519 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2520 | } | |
2521 | } else { | |
2522 | tmppages = 0; | |
23feba90 | 2523 | } |
f3321554 PX |
2524 | |
2525 | if (tmppages < 0) { | |
d9e474ea | 2526 | pss_host_page_finish(pss); |
f3321554 PX |
2527 | return tmppages; |
2528 | } | |
2529 | ||
d9e474ea PX |
2530 | pss_find_next_dirty(pss); |
2531 | } while (pss_within_range(pss)); | |
2532 | ||
2533 | pss_host_page_finish(pss); | |
278e2f55 AG |
2534 | |
2535 | res = ram_save_release_protection(rs, pss, start_page); | |
2536 | return (res < 0 ? res : pages); | |
a82d593b | 2537 | } |
6c595cde | 2538 | |
56e93d26 | 2539 | /** |
3d0684b2 | 2540 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
2541 | * |
2542 | * Called within an RCU critical section. | |
2543 | * | |
e8f3735f XG |
2544 | * Returns the number of pages written where zero means no dirty pages, |
2545 | * or negative on error | |
56e93d26 | 2546 | * |
6f37bb8b | 2547 | * @rs: current RAM state |
a82d593b DDAG |
2548 | * |
2549 | * On systems where host-page-size > target-page-size it will send all the | |
2550 | * pages in a host page that are dirty. | |
56e93d26 | 2551 | */ |
05931ec5 | 2552 | static int ram_find_and_save_block(RAMState *rs) |
56e93d26 | 2553 | { |
f1668764 | 2554 | PageSearchStatus *pss = &rs->pss[RAM_CHANNEL_PRECOPY]; |
56e93d26 | 2555 | int pages = 0; |
56e93d26 | 2556 | |
0827b9e9 | 2557 | /* No dirty page as there is zero RAM */ |
8d80e195 | 2558 | if (!rs->ram_bytes_total) { |
0827b9e9 AA |
2559 | return pages; |
2560 | } | |
2561 | ||
4934a5dd PX |
2562 | /* |
2563 | * Always keep last_seen_block/last_page valid during this procedure, | |
2564 | * because find_dirty_block() relies on these values (e.g., we compare | |
2565 | * last_seen_block with pss.block to see whether we searched all the | |
2566 | * ramblocks) to detect the completion of migration. Having NULL value | |
2567 | * of last_seen_block can conditionally cause below loop to run forever. | |
2568 | */ | |
2569 | if (!rs->last_seen_block) { | |
2570 | rs->last_seen_block = QLIST_FIRST_RCU(&ram_list.blocks); | |
2571 | rs->last_page = 0; | |
2572 | } | |
2573 | ||
f1668764 | 2574 | pss_init(pss, rs->last_seen_block, rs->last_page); |
b8fb8cb7 | 2575 | |
31e2ac74 | 2576 | while (true){ |
51efd36f | 2577 | if (!get_queued_page(rs, pss)) { |
b062106d | 2578 | /* priority queue empty, so just search for something dirty */ |
31e2ac74 JQ |
2579 | int res = find_dirty_block(rs, pss); |
2580 | if (res != PAGE_DIRTY_FOUND) { | |
2581 | if (res == PAGE_ALL_CLEAN) { | |
51efd36f | 2582 | break; |
31e2ac74 JQ |
2583 | } else if (res == PAGE_TRY_AGAIN) { |
2584 | continue; | |
51efd36f JQ |
2585 | } |
2586 | } | |
56e93d26 | 2587 | } |
51efd36f | 2588 | pages = ram_save_host_page(rs, pss); |
31e2ac74 JQ |
2589 | if (pages) { |
2590 | break; | |
2591 | } | |
2592 | } | |
56e93d26 | 2593 | |
f1668764 PX |
2594 | rs->last_seen_block = pss->block; |
2595 | rs->last_page = pss->page; | |
56e93d26 JQ |
2596 | |
2597 | return pages; | |
2598 | } | |
2599 | ||
2600 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
2601 | { | |
2602 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 2603 | |
56e93d26 | 2604 | if (zero) { |
23b7576d | 2605 | stat64_add(&ram_atomic_counters.duplicate, pages); |
56e93d26 | 2606 | } else { |
23b7576d | 2607 | stat64_add(&ram_atomic_counters.normal, pages); |
4c2d0f6d | 2608 | ram_transferred_add(size); |
1a93bd2f | 2609 | qemu_file_credit_transfer(f, size); |
56e93d26 JQ |
2610 | } |
2611 | } | |
2612 | ||
8008a272 | 2613 | static uint64_t ram_bytes_total_with_ignored(void) |
56e93d26 JQ |
2614 | { |
2615 | RAMBlock *block; | |
2616 | uint64_t total = 0; | |
2617 | ||
89ac5a1d DDAG |
2618 | RCU_READ_LOCK_GUARD(); |
2619 | ||
8008a272 JQ |
2620 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
2621 | total += block->used_length; | |
99e15582 | 2622 | } |
56e93d26 JQ |
2623 | return total; |
2624 | } | |
2625 | ||
fbd162e6 YK |
2626 | uint64_t ram_bytes_total(void) |
2627 | { | |
8008a272 JQ |
2628 | RAMBlock *block; |
2629 | uint64_t total = 0; | |
2630 | ||
2631 | RCU_READ_LOCK_GUARD(); | |
2632 | ||
2633 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2634 | total += block->used_length; | |
2635 | } | |
2636 | return total; | |
fbd162e6 YK |
2637 | } |
2638 | ||
f265e0e4 | 2639 | static void xbzrle_load_setup(void) |
56e93d26 | 2640 | { |
f265e0e4 | 2641 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2642 | } |
2643 | ||
f265e0e4 JQ |
2644 | static void xbzrle_load_cleanup(void) |
2645 | { | |
2646 | g_free(XBZRLE.decoded_buf); | |
2647 | XBZRLE.decoded_buf = NULL; | |
2648 | } | |
2649 | ||
7d7c96be PX |
2650 | static void ram_state_cleanup(RAMState **rsp) |
2651 | { | |
b9ccaf6d DDAG |
2652 | if (*rsp) { |
2653 | migration_page_queue_free(*rsp); | |
2654 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
2655 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
2656 | g_free(*rsp); | |
2657 | *rsp = NULL; | |
2658 | } | |
7d7c96be PX |
2659 | } |
2660 | ||
84593a08 PX |
2661 | static void xbzrle_cleanup(void) |
2662 | { | |
2663 | XBZRLE_cache_lock(); | |
2664 | if (XBZRLE.cache) { | |
2665 | cache_fini(XBZRLE.cache); | |
2666 | g_free(XBZRLE.encoded_buf); | |
2667 | g_free(XBZRLE.current_buf); | |
2668 | g_free(XBZRLE.zero_target_page); | |
2669 | XBZRLE.cache = NULL; | |
2670 | XBZRLE.encoded_buf = NULL; | |
2671 | XBZRLE.current_buf = NULL; | |
2672 | XBZRLE.zero_target_page = NULL; | |
2673 | } | |
2674 | XBZRLE_cache_unlock(); | |
2675 | } | |
2676 | ||
f265e0e4 | 2677 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 2678 | { |
53518d94 | 2679 | RAMState **rsp = opaque; |
6b6712ef | 2680 | RAMBlock *block; |
eb859c53 | 2681 | |
278e2f55 AG |
2682 | /* We don't use dirty log with background snapshots */ |
2683 | if (!migrate_background_snapshot()) { | |
2684 | /* caller have hold iothread lock or is in a bh, so there is | |
2685 | * no writing race against the migration bitmap | |
2686 | */ | |
63b41db4 HH |
2687 | if (global_dirty_tracking & GLOBAL_DIRTY_MIGRATION) { |
2688 | /* | |
2689 | * do not stop dirty log without starting it, since | |
2690 | * memory_global_dirty_log_stop will assert that | |
2691 | * memory_global_dirty_log_start/stop used in pairs | |
2692 | */ | |
2693 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); | |
2694 | } | |
278e2f55 | 2695 | } |
6b6712ef | 2696 | |
fbd162e6 | 2697 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
2698 | g_free(block->clear_bmap); |
2699 | block->clear_bmap = NULL; | |
6b6712ef JQ |
2700 | g_free(block->bmap); |
2701 | block->bmap = NULL; | |
56e93d26 JQ |
2702 | } |
2703 | ||
84593a08 | 2704 | xbzrle_cleanup(); |
f0afa331 | 2705 | compress_threads_save_cleanup(); |
7d7c96be | 2706 | ram_state_cleanup(rsp); |
4010ba38 JQ |
2707 | g_free(migration_ops); |
2708 | migration_ops = NULL; | |
56e93d26 JQ |
2709 | } |
2710 | ||
6f37bb8b | 2711 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 2712 | { |
ec6f3ab9 PX |
2713 | int i; |
2714 | ||
2715 | for (i = 0; i < RAM_CHANNEL_MAX; i++) { | |
2716 | rs->pss[i].last_sent_block = NULL; | |
2717 | } | |
2718 | ||
6f37bb8b | 2719 | rs->last_seen_block = NULL; |
269ace29 | 2720 | rs->last_page = 0; |
6f37bb8b | 2721 | rs->last_version = ram_list.version; |
1a373522 | 2722 | rs->xbzrle_enabled = false; |
56e93d26 JQ |
2723 | } |
2724 | ||
2725 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
2726 | ||
e0b266f0 DDAG |
2727 | /* **** functions for postcopy ***** */ |
2728 | ||
ced1c616 PB |
2729 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
2730 | { | |
2731 | struct RAMBlock *block; | |
ced1c616 | 2732 | |
fbd162e6 | 2733 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2734 | unsigned long *bitmap = block->bmap; |
2735 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
2736 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
2737 | |
2738 | while (run_start < range) { | |
2739 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
2740 | ram_discard_range(block->idstr, |
2741 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
2742 | ((ram_addr_t)(run_end - run_start)) | |
2743 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
2744 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
2745 | } | |
2746 | } | |
2747 | } | |
2748 | ||
3d0684b2 JQ |
2749 | /** |
2750 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
2751 | * | |
e0b266f0 | 2752 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
2753 | * |
2754 | * @ms: current migration state | |
89dab31b | 2755 | * @block: RAMBlock to discard |
e0b266f0 | 2756 | */ |
9e7d1223 | 2757 | static void postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 2758 | { |
6b6712ef | 2759 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 2760 | unsigned long current; |
1e7cf8c3 | 2761 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 2762 | |
6b6712ef | 2763 | for (current = 0; current < end; ) { |
1e7cf8c3 | 2764 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 2765 | unsigned long zero, discard_length; |
e0b266f0 | 2766 | |
33a5cb62 WY |
2767 | if (one >= end) { |
2768 | break; | |
2769 | } | |
e0b266f0 | 2770 | |
1e7cf8c3 | 2771 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
2772 | |
2773 | if (zero >= end) { | |
2774 | discard_length = end - one; | |
e0b266f0 | 2775 | } else { |
33a5cb62 WY |
2776 | discard_length = zero - one; |
2777 | } | |
810cf2bb | 2778 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 2779 | current = one + discard_length; |
e0b266f0 | 2780 | } |
e0b266f0 DDAG |
2781 | } |
2782 | ||
f30c2e5b PX |
2783 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block); |
2784 | ||
3d0684b2 JQ |
2785 | /** |
2786 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
2787 | * | |
e0b266f0 DDAG |
2788 | * Utility for the outgoing postcopy code. |
2789 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
2790 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
2791 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
2792 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
2793 | * |
2794 | * @ms: current migration state | |
e0b266f0 | 2795 | */ |
739fcc1b | 2796 | static void postcopy_each_ram_send_discard(MigrationState *ms) |
e0b266f0 DDAG |
2797 | { |
2798 | struct RAMBlock *block; | |
e0b266f0 | 2799 | |
fbd162e6 | 2800 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 2801 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 | 2802 | |
f30c2e5b PX |
2803 | /* |
2804 | * Deal with TPS != HPS and huge pages. It discard any partially sent | |
2805 | * host-page size chunks, mark any partially dirty host-page size | |
2806 | * chunks as all dirty. In this case the host-page is the host-page | |
2807 | * for the particular RAMBlock, i.e. it might be a huge page. | |
2808 | */ | |
2809 | postcopy_chunk_hostpages_pass(ms, block); | |
2810 | ||
e0b266f0 DDAG |
2811 | /* |
2812 | * Postcopy sends chunks of bitmap over the wire, but it | |
2813 | * just needs indexes at this point, avoids it having | |
2814 | * target page specific code. | |
2815 | */ | |
739fcc1b | 2816 | postcopy_send_discard_bm_ram(ms, block); |
810cf2bb | 2817 | postcopy_discard_send_finish(ms); |
e0b266f0 | 2818 | } |
e0b266f0 DDAG |
2819 | } |
2820 | ||
3d0684b2 | 2821 | /** |
8324ef86 | 2822 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2823 | * |
2824 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2825 | * canonicalize the two bitmaps, that are similar, but one is | |
2826 | * inverted. | |
99e314eb | 2827 | * |
3d0684b2 JQ |
2828 | * Postcopy requires that all target pages in a hostpage are dirty or |
2829 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2830 | * |
3d0684b2 | 2831 | * @ms: current migration state |
3d0684b2 | 2832 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2833 | */ |
1e7cf8c3 | 2834 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2835 | { |
53518d94 | 2836 | RAMState *rs = ram_state; |
6b6712ef | 2837 | unsigned long *bitmap = block->bmap; |
29c59172 | 2838 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2839 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2840 | unsigned long run_start; |
2841 | ||
29c59172 DDAG |
2842 | if (block->page_size == TARGET_PAGE_SIZE) { |
2843 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2844 | return; | |
2845 | } | |
2846 | ||
1e7cf8c3 WY |
2847 | /* Find a dirty page */ |
2848 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2849 | |
6b6712ef | 2850 | while (run_start < pages) { |
99e314eb DDAG |
2851 | |
2852 | /* | |
2853 | * If the start of this run of pages is in the middle of a host | |
2854 | * page, then we need to fixup this host page. | |
2855 | */ | |
9dec3cc3 | 2856 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2857 | /* Find the end of this run */ |
1e7cf8c3 | 2858 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2859 | /* |
2860 | * If the end isn't at the start of a host page, then the | |
2861 | * run doesn't finish at the end of a host page | |
2862 | * and we need to discard. | |
2863 | */ | |
99e314eb DDAG |
2864 | } |
2865 | ||
9dec3cc3 | 2866 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2867 | unsigned long page; |
dad45ab2 WY |
2868 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2869 | host_ratio); | |
2870 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2871 | |
99e314eb DDAG |
2872 | /* Clean up the bitmap */ |
2873 | for (page = fixup_start_addr; | |
2874 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2875 | /* |
2876 | * Remark them as dirty, updating the count for any pages | |
2877 | * that weren't previously dirty. | |
2878 | */ | |
0d8ec885 | 2879 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2880 | } |
2881 | } | |
2882 | ||
1e7cf8c3 WY |
2883 | /* Find the next dirty page for the next iteration */ |
2884 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2885 | } |
2886 | } | |
2887 | ||
3d0684b2 JQ |
2888 | /** |
2889 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2890 | * | |
e0b266f0 DDAG |
2891 | * Transmit the set of pages to be discarded after precopy to the target |
2892 | * these are pages that: | |
2893 | * a) Have been previously transmitted but are now dirty again | |
2894 | * b) Pages that have never been transmitted, this ensures that | |
2895 | * any pages on the destination that have been mapped by background | |
2896 | * tasks get discarded (transparent huge pages is the specific concern) | |
2897 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2898 | * |
2899 | * @ms: current migration state | |
e0b266f0 | 2900 | */ |
739fcc1b | 2901 | void ram_postcopy_send_discard_bitmap(MigrationState *ms) |
e0b266f0 | 2902 | { |
53518d94 | 2903 | RAMState *rs = ram_state; |
e0b266f0 | 2904 | |
89ac5a1d | 2905 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2906 | |
2907 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 2908 | migration_bitmap_sync(rs); |
e0b266f0 | 2909 | |
6b6712ef | 2910 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
ec6f3ab9 | 2911 | rs->pss[RAM_CHANNEL_PRECOPY].last_sent_block = NULL; |
6b6712ef | 2912 | rs->last_seen_block = NULL; |
6b6712ef | 2913 | rs->last_page = 0; |
e0b266f0 | 2914 | |
739fcc1b | 2915 | postcopy_each_ram_send_discard(ms); |
e0b266f0 | 2916 | |
739fcc1b | 2917 | trace_ram_postcopy_send_discard_bitmap(); |
e0b266f0 DDAG |
2918 | } |
2919 | ||
3d0684b2 JQ |
2920 | /** |
2921 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2922 | * |
3d0684b2 | 2923 | * Returns zero on success |
e0b266f0 | 2924 | * |
36449157 JQ |
2925 | * @rbname: name of the RAMBlock of the request. NULL means the |
2926 | * same that last one. | |
3d0684b2 JQ |
2927 | * @start: RAMBlock starting page |
2928 | * @length: RAMBlock size | |
e0b266f0 | 2929 | */ |
aaa2064c | 2930 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 2931 | { |
36449157 | 2932 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2933 | |
89ac5a1d | 2934 | RCU_READ_LOCK_GUARD(); |
36449157 | 2935 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2936 | |
2937 | if (!rb) { | |
36449157 | 2938 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 2939 | return -1; |
e0b266f0 DDAG |
2940 | } |
2941 | ||
814bb08f PX |
2942 | /* |
2943 | * On source VM, we don't need to update the received bitmap since | |
2944 | * we don't even have one. | |
2945 | */ | |
2946 | if (rb->receivedmap) { | |
2947 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
2948 | length >> qemu_target_page_bits()); | |
2949 | } | |
2950 | ||
03acb4e9 | 2951 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2952 | } |
2953 | ||
84593a08 PX |
2954 | /* |
2955 | * For every allocation, we will try not to crash the VM if the | |
2956 | * allocation failed. | |
2957 | */ | |
2958 | static int xbzrle_init(void) | |
2959 | { | |
2960 | Error *local_err = NULL; | |
2961 | ||
2962 | if (!migrate_use_xbzrle()) { | |
2963 | return 0; | |
2964 | } | |
2965 | ||
2966 | XBZRLE_cache_lock(); | |
2967 | ||
2968 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
2969 | if (!XBZRLE.zero_target_page) { | |
2970 | error_report("%s: Error allocating zero page", __func__); | |
2971 | goto err_out; | |
2972 | } | |
2973 | ||
2974 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
2975 | TARGET_PAGE_SIZE, &local_err); | |
2976 | if (!XBZRLE.cache) { | |
2977 | error_report_err(local_err); | |
2978 | goto free_zero_page; | |
2979 | } | |
2980 | ||
2981 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2982 | if (!XBZRLE.encoded_buf) { | |
2983 | error_report("%s: Error allocating encoded_buf", __func__); | |
2984 | goto free_cache; | |
2985 | } | |
2986 | ||
2987 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2988 | if (!XBZRLE.current_buf) { | |
2989 | error_report("%s: Error allocating current_buf", __func__); | |
2990 | goto free_encoded_buf; | |
2991 | } | |
2992 | ||
2993 | /* We are all good */ | |
2994 | XBZRLE_cache_unlock(); | |
2995 | return 0; | |
2996 | ||
2997 | free_encoded_buf: | |
2998 | g_free(XBZRLE.encoded_buf); | |
2999 | XBZRLE.encoded_buf = NULL; | |
3000 | free_cache: | |
3001 | cache_fini(XBZRLE.cache); | |
3002 | XBZRLE.cache = NULL; | |
3003 | free_zero_page: | |
3004 | g_free(XBZRLE.zero_target_page); | |
3005 | XBZRLE.zero_target_page = NULL; | |
3006 | err_out: | |
3007 | XBZRLE_cache_unlock(); | |
3008 | return -ENOMEM; | |
3009 | } | |
3010 | ||
53518d94 | 3011 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 3012 | { |
7d00ee6a PX |
3013 | *rsp = g_try_new0(RAMState, 1); |
3014 | ||
3015 | if (!*rsp) { | |
3016 | error_report("%s: Init ramstate fail", __func__); | |
3017 | return -1; | |
3018 | } | |
53518d94 JQ |
3019 | |
3020 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
3021 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
3022 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
8d80e195 | 3023 | (*rsp)->ram_bytes_total = ram_bytes_total(); |
56e93d26 | 3024 | |
7d00ee6a | 3025 | /* |
40c4d4a8 IR |
3026 | * Count the total number of pages used by ram blocks not including any |
3027 | * gaps due to alignment or unplugs. | |
03158519 | 3028 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 3029 | */ |
8d80e195 | 3030 | (*rsp)->migration_dirty_pages = (*rsp)->ram_bytes_total >> TARGET_PAGE_BITS; |
7d00ee6a PX |
3031 | ram_state_reset(*rsp); |
3032 | ||
3033 | return 0; | |
3034 | } | |
3035 | ||
d6eff5d7 | 3036 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 3037 | { |
002cad6b | 3038 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
3039 | RAMBlock *block; |
3040 | unsigned long pages; | |
002cad6b | 3041 | uint8_t shift; |
56e93d26 | 3042 | |
0827b9e9 AA |
3043 | /* Skip setting bitmap if there is no RAM */ |
3044 | if (ram_bytes_total()) { | |
002cad6b PX |
3045 | shift = ms->clear_bitmap_shift; |
3046 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
3047 | error_report("clear_bitmap_shift (%u) too big, using " | |
3048 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
3049 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
3050 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
3051 | error_report("clear_bitmap_shift (%u) too small, using " | |
3052 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
3053 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
3054 | } | |
3055 | ||
fbd162e6 | 3056 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 3057 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
3058 | /* |
3059 | * The initial dirty bitmap for migration must be set with all | |
3060 | * ones to make sure we'll migrate every guest RAM page to | |
3061 | * destination. | |
40c4d4a8 IR |
3062 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
3063 | * new migration after a failed migration, ram_list. | |
3064 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
3065 | * guest memory. | |
03158519 | 3066 | */ |
6b6712ef | 3067 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 3068 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
3069 | block->clear_bmap_shift = shift; |
3070 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 3071 | } |
f3f491fc | 3072 | } |
d6eff5d7 PX |
3073 | } |
3074 | ||
be39b4cd DH |
3075 | static void migration_bitmap_clear_discarded_pages(RAMState *rs) |
3076 | { | |
3077 | unsigned long pages; | |
3078 | RAMBlock *rb; | |
3079 | ||
3080 | RCU_READ_LOCK_GUARD(); | |
3081 | ||
3082 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { | |
3083 | pages = ramblock_dirty_bitmap_clear_discarded_pages(rb); | |
3084 | rs->migration_dirty_pages -= pages; | |
3085 | } | |
3086 | } | |
3087 | ||
d6eff5d7 PX |
3088 | static void ram_init_bitmaps(RAMState *rs) |
3089 | { | |
3090 | /* For memory_global_dirty_log_start below. */ | |
3091 | qemu_mutex_lock_iothread(); | |
3092 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 3093 | |
89ac5a1d DDAG |
3094 | WITH_RCU_READ_LOCK_GUARD() { |
3095 | ram_list_init_bitmaps(); | |
278e2f55 AG |
3096 | /* We don't use dirty log with background snapshots */ |
3097 | if (!migrate_background_snapshot()) { | |
63b41db4 | 3098 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
278e2f55 AG |
3099 | migration_bitmap_sync_precopy(rs); |
3100 | } | |
89ac5a1d | 3101 | } |
56e93d26 | 3102 | qemu_mutex_unlock_ramlist(); |
49877834 | 3103 | qemu_mutex_unlock_iothread(); |
be39b4cd DH |
3104 | |
3105 | /* | |
3106 | * After an eventual first bitmap sync, fixup the initial bitmap | |
3107 | * containing all 1s to exclude any discarded pages from migration. | |
3108 | */ | |
3109 | migration_bitmap_clear_discarded_pages(rs); | |
d6eff5d7 PX |
3110 | } |
3111 | ||
3112 | static int ram_init_all(RAMState **rsp) | |
3113 | { | |
3114 | if (ram_state_init(rsp)) { | |
3115 | return -1; | |
3116 | } | |
3117 | ||
3118 | if (xbzrle_init()) { | |
3119 | ram_state_cleanup(rsp); | |
3120 | return -1; | |
3121 | } | |
3122 | ||
3123 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
3124 | |
3125 | return 0; | |
3126 | } | |
3127 | ||
08614f34 PX |
3128 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
3129 | { | |
3130 | RAMBlock *block; | |
3131 | uint64_t pages = 0; | |
3132 | ||
3133 | /* | |
3134 | * Postcopy is not using xbzrle/compression, so no need for that. | |
3135 | * Also, since source are already halted, we don't need to care | |
3136 | * about dirty page logging as well. | |
3137 | */ | |
3138 | ||
fbd162e6 | 3139 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
3140 | pages += bitmap_count_one(block->bmap, |
3141 | block->used_length >> TARGET_PAGE_BITS); | |
3142 | } | |
3143 | ||
3144 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
3145 | rs->migration_dirty_pages = pages; | |
3146 | ||
1a373522 | 3147 | ram_state_reset(rs); |
08614f34 PX |
3148 | |
3149 | /* Update RAMState cache of output QEMUFile */ | |
7f401b80 | 3150 | rs->pss[RAM_CHANNEL_PRECOPY].pss_channel = out; |
08614f34 PX |
3151 | |
3152 | trace_ram_state_resume_prepare(pages); | |
3153 | } | |
3154 | ||
6bcb05fc WW |
3155 | /* |
3156 | * This function clears bits of the free pages reported by the caller from the | |
3157 | * migration dirty bitmap. @addr is the host address corresponding to the | |
3158 | * start of the continuous guest free pages, and @len is the total bytes of | |
3159 | * those pages. | |
3160 | */ | |
3161 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
3162 | { | |
3163 | RAMBlock *block; | |
3164 | ram_addr_t offset; | |
3165 | size_t used_len, start, npages; | |
3166 | MigrationState *s = migrate_get_current(); | |
3167 | ||
3168 | /* This function is currently expected to be used during live migration */ | |
3169 | if (!migration_is_setup_or_active(s->state)) { | |
3170 | return; | |
3171 | } | |
3172 | ||
3173 | for (; len > 0; len -= used_len, addr += used_len) { | |
3174 | block = qemu_ram_block_from_host(addr, false, &offset); | |
3175 | if (unlikely(!block || offset >= block->used_length)) { | |
3176 | /* | |
3177 | * The implementation might not support RAMBlock resize during | |
3178 | * live migration, but it could happen in theory with future | |
3179 | * updates. So we add a check here to capture that case. | |
3180 | */ | |
3181 | error_report_once("%s unexpected error", __func__); | |
3182 | return; | |
3183 | } | |
3184 | ||
3185 | if (len <= block->used_length - offset) { | |
3186 | used_len = len; | |
3187 | } else { | |
3188 | used_len = block->used_length - offset; | |
3189 | } | |
3190 | ||
3191 | start = offset >> TARGET_PAGE_BITS; | |
3192 | npages = used_len >> TARGET_PAGE_BITS; | |
3193 | ||
3194 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3143577d WW |
3195 | /* |
3196 | * The skipped free pages are equavalent to be sent from clear_bmap's | |
3197 | * perspective, so clear the bits from the memory region bitmap which | |
3198 | * are initially set. Otherwise those skipped pages will be sent in | |
3199 | * the next round after syncing from the memory region bitmap. | |
3200 | */ | |
1230a25f | 3201 | migration_clear_memory_region_dirty_bitmap_range(block, start, npages); |
6bcb05fc WW |
3202 | ram_state->migration_dirty_pages -= |
3203 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
3204 | bitmap_clear(block->bmap, start, npages); | |
3205 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
3206 | } | |
3207 | } | |
3208 | ||
3d0684b2 JQ |
3209 | /* |
3210 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
3211 | * long-running RCU critical section. When rcu-reclaims in the code |
3212 | * start to become numerous it will be necessary to reduce the | |
3213 | * granularity of these critical sections. | |
3214 | */ | |
3215 | ||
3d0684b2 JQ |
3216 | /** |
3217 | * ram_save_setup: Setup RAM for migration | |
3218 | * | |
3219 | * Returns zero to indicate success and negative for error | |
3220 | * | |
3221 | * @f: QEMUFile where to send the data | |
3222 | * @opaque: RAMState pointer | |
3223 | */ | |
a91246c9 HZ |
3224 | static int ram_save_setup(QEMUFile *f, void *opaque) |
3225 | { | |
53518d94 | 3226 | RAMState **rsp = opaque; |
a91246c9 | 3227 | RAMBlock *block; |
33d70973 | 3228 | int ret; |
a91246c9 | 3229 | |
dcaf446e XG |
3230 | if (compress_threads_save_setup()) { |
3231 | return -1; | |
3232 | } | |
3233 | ||
a91246c9 HZ |
3234 | /* migration has already setup the bitmap, reuse it. */ |
3235 | if (!migration_in_colo_state()) { | |
7d00ee6a | 3236 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 3237 | compress_threads_save_cleanup(); |
a91246c9 | 3238 | return -1; |
53518d94 | 3239 | } |
a91246c9 | 3240 | } |
7f401b80 | 3241 | (*rsp)->pss[RAM_CHANNEL_PRECOPY].pss_channel = f; |
a91246c9 | 3242 | |
0e6ebd48 | 3243 | WITH_RCU_READ_LOCK_GUARD() { |
8008a272 JQ |
3244 | qemu_put_be64(f, ram_bytes_total_with_ignored() |
3245 | | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 3246 | |
0e6ebd48 DDAG |
3247 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
3248 | qemu_put_byte(f, strlen(block->idstr)); | |
3249 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
3250 | qemu_put_be64(f, block->used_length); | |
3251 | if (migrate_postcopy_ram() && block->page_size != | |
3252 | qemu_host_page_size) { | |
3253 | qemu_put_be64(f, block->page_size); | |
3254 | } | |
3255 | if (migrate_ignore_shared()) { | |
3256 | qemu_put_be64(f, block->mr->addr); | |
3257 | } | |
fbd162e6 | 3258 | } |
56e93d26 JQ |
3259 | } |
3260 | ||
56e93d26 JQ |
3261 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
3262 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
3263 | ||
4010ba38 JQ |
3264 | migration_ops = g_malloc0(sizeof(MigrationOps)); |
3265 | migration_ops->ram_save_target_page = ram_save_target_page_legacy; | |
33d70973 LB |
3266 | ret = multifd_send_sync_main(f); |
3267 | if (ret < 0) { | |
3268 | return ret; | |
3269 | } | |
3270 | ||
56e93d26 | 3271 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 3272 | qemu_fflush(f); |
56e93d26 JQ |
3273 | |
3274 | return 0; | |
3275 | } | |
3276 | ||
3d0684b2 JQ |
3277 | /** |
3278 | * ram_save_iterate: iterative stage for migration | |
3279 | * | |
3280 | * Returns zero to indicate success and negative for error | |
3281 | * | |
3282 | * @f: QEMUFile where to send the data | |
3283 | * @opaque: RAMState pointer | |
3284 | */ | |
56e93d26 JQ |
3285 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
3286 | { | |
53518d94 JQ |
3287 | RAMState **temp = opaque; |
3288 | RAMState *rs = *temp; | |
3d4095b2 | 3289 | int ret = 0; |
56e93d26 JQ |
3290 | int i; |
3291 | int64_t t0; | |
5c90308f | 3292 | int done = 0; |
56e93d26 | 3293 | |
b2557345 PL |
3294 | if (blk_mig_bulk_active()) { |
3295 | /* Avoid transferring ram during bulk phase of block migration as | |
3296 | * the bulk phase will usually take a long time and transferring | |
3297 | * ram updates during that time is pointless. */ | |
3298 | goto out; | |
3299 | } | |
3300 | ||
63268c49 PX |
3301 | /* |
3302 | * We'll take this lock a little bit long, but it's okay for two reasons. | |
3303 | * Firstly, the only possible other thread to take it is who calls | |
3304 | * qemu_guest_free_page_hint(), which should be rare; secondly, see | |
3305 | * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which | |
3306 | * guarantees that we'll at least released it in a regular basis. | |
3307 | */ | |
3308 | qemu_mutex_lock(&rs->bitmap_mutex); | |
89ac5a1d DDAG |
3309 | WITH_RCU_READ_LOCK_GUARD() { |
3310 | if (ram_list.version != rs->last_version) { | |
3311 | ram_state_reset(rs); | |
3312 | } | |
56e93d26 | 3313 | |
89ac5a1d DDAG |
3314 | /* Read version before ram_list.blocks */ |
3315 | smp_rmb(); | |
56e93d26 | 3316 | |
89ac5a1d | 3317 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 3318 | |
89ac5a1d DDAG |
3319 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
3320 | i = 0; | |
3321 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
a1fe28df | 3322 | postcopy_has_request(rs)) { |
89ac5a1d | 3323 | int pages; |
e03a34f8 | 3324 | |
89ac5a1d DDAG |
3325 | if (qemu_file_get_error(f)) { |
3326 | break; | |
3327 | } | |
e8f3735f | 3328 | |
05931ec5 | 3329 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3330 | /* no more pages to sent */ |
3331 | if (pages == 0) { | |
3332 | done = 1; | |
3333 | break; | |
3334 | } | |
e8f3735f | 3335 | |
89ac5a1d DDAG |
3336 | if (pages < 0) { |
3337 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
3338 | break; |
3339 | } | |
89ac5a1d DDAG |
3340 | |
3341 | rs->target_page_count += pages; | |
3342 | ||
644acf99 WY |
3343 | /* |
3344 | * During postcopy, it is necessary to make sure one whole host | |
3345 | * page is sent in one chunk. | |
3346 | */ | |
3347 | if (migrate_postcopy_ram()) { | |
3348 | flush_compressed_data(rs); | |
3349 | } | |
3350 | ||
89ac5a1d DDAG |
3351 | /* |
3352 | * we want to check in the 1st loop, just in case it was the 1st | |
3353 | * time and we had to sync the dirty bitmap. | |
3354 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
3355 | * some iterations | |
3356 | */ | |
3357 | if ((i & 63) == 0) { | |
3358 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
3359 | 1000000; | |
3360 | if (t1 > MAX_WAIT) { | |
3361 | trace_ram_save_iterate_big_wait(t1, i); | |
3362 | break; | |
3363 | } | |
3364 | } | |
3365 | i++; | |
56e93d26 | 3366 | } |
56e93d26 | 3367 | } |
63268c49 | 3368 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 JQ |
3369 | |
3370 | /* | |
3371 | * Must occur before EOS (or any QEMUFile operation) | |
3372 | * because of RDMA protocol. | |
3373 | */ | |
3374 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
3375 | ||
b2557345 | 3376 | out: |
b69a0227 JQ |
3377 | if (ret >= 0 |
3378 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
7f401b80 | 3379 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); |
33d70973 LB |
3380 | if (ret < 0) { |
3381 | return ret; | |
3382 | } | |
3383 | ||
3d4095b2 JQ |
3384 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3385 | qemu_fflush(f); | |
4c2d0f6d | 3386 | ram_transferred_add(8); |
56e93d26 | 3387 | |
3d4095b2 JQ |
3388 | ret = qemu_file_get_error(f); |
3389 | } | |
56e93d26 JQ |
3390 | if (ret < 0) { |
3391 | return ret; | |
3392 | } | |
3393 | ||
5c90308f | 3394 | return done; |
56e93d26 JQ |
3395 | } |
3396 | ||
3d0684b2 JQ |
3397 | /** |
3398 | * ram_save_complete: function called to send the remaining amount of ram | |
3399 | * | |
e8f3735f | 3400 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
3401 | * |
3402 | * Called with iothread lock | |
3403 | * | |
3404 | * @f: QEMUFile where to send the data | |
3405 | * @opaque: RAMState pointer | |
3406 | */ | |
56e93d26 JQ |
3407 | static int ram_save_complete(QEMUFile *f, void *opaque) |
3408 | { | |
53518d94 JQ |
3409 | RAMState **temp = opaque; |
3410 | RAMState *rs = *temp; | |
e8f3735f | 3411 | int ret = 0; |
6f37bb8b | 3412 | |
05931ec5 JQ |
3413 | rs->last_stage = !migration_in_colo_state(); |
3414 | ||
89ac5a1d DDAG |
3415 | WITH_RCU_READ_LOCK_GUARD() { |
3416 | if (!migration_in_postcopy()) { | |
3417 | migration_bitmap_sync_precopy(rs); | |
3418 | } | |
56e93d26 | 3419 | |
89ac5a1d | 3420 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 3421 | |
89ac5a1d | 3422 | /* try transferring iterative blocks of memory */ |
56e93d26 | 3423 | |
89ac5a1d | 3424 | /* flush all remaining blocks regardless of rate limiting */ |
c13221b5 | 3425 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
3426 | while (true) { |
3427 | int pages; | |
56e93d26 | 3428 | |
05931ec5 | 3429 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3430 | /* no more blocks to sent */ |
3431 | if (pages == 0) { | |
3432 | break; | |
3433 | } | |
3434 | if (pages < 0) { | |
3435 | ret = pages; | |
3436 | break; | |
3437 | } | |
e8f3735f | 3438 | } |
c13221b5 | 3439 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 3440 | |
89ac5a1d DDAG |
3441 | flush_compressed_data(rs); |
3442 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); | |
3443 | } | |
d09a6fde | 3444 | |
33d70973 LB |
3445 | if (ret < 0) { |
3446 | return ret; | |
3d4095b2 | 3447 | } |
56e93d26 | 3448 | |
7f401b80 | 3449 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); |
33d70973 LB |
3450 | if (ret < 0) { |
3451 | return ret; | |
3452 | } | |
3453 | ||
3454 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); | |
3455 | qemu_fflush(f); | |
3456 | ||
3457 | return 0; | |
56e93d26 JQ |
3458 | } |
3459 | ||
fd70385d | 3460 | static void ram_state_pending_estimate(void *opaque, |
c8df4a7a JQ |
3461 | uint64_t *res_precopy_only, |
3462 | uint64_t *res_compatible, | |
3463 | uint64_t *res_postcopy_only) | |
56e93d26 | 3464 | { |
53518d94 JQ |
3465 | RAMState **temp = opaque; |
3466 | RAMState *rs = *temp; | |
56e93d26 | 3467 | |
c8df4a7a | 3468 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3469 | |
c8df4a7a JQ |
3470 | if (migrate_postcopy_ram()) { |
3471 | /* We can do postcopy, and all the data is postcopiable */ | |
3472 | *res_postcopy_only += remaining_size; | |
3473 | } else { | |
3474 | *res_precopy_only += remaining_size; | |
3475 | } | |
3476 | } | |
3477 | ||
fd70385d | 3478 | static void ram_state_pending_exact(void *opaque, |
c8df4a7a JQ |
3479 | uint64_t *res_precopy_only, |
3480 | uint64_t *res_compatible, | |
3481 | uint64_t *res_postcopy_only) | |
3482 | { | |
3483 | RAMState **temp = opaque; | |
3484 | RAMState *rs = *temp; | |
3485 | ||
3486 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; | |
3487 | ||
3488 | if (!migration_in_postcopy()) { | |
56e93d26 | 3489 | qemu_mutex_lock_iothread(); |
89ac5a1d DDAG |
3490 | WITH_RCU_READ_LOCK_GUARD() { |
3491 | migration_bitmap_sync_precopy(rs); | |
3492 | } | |
56e93d26 | 3493 | qemu_mutex_unlock_iothread(); |
9edabd4d | 3494 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3495 | } |
c31b098f | 3496 | |
86e1167e VSO |
3497 | if (migrate_postcopy_ram()) { |
3498 | /* We can do postcopy, and all the data is postcopiable */ | |
47995026 | 3499 | *res_compatible += remaining_size; |
86e1167e | 3500 | } else { |
47995026 | 3501 | *res_precopy_only += remaining_size; |
86e1167e | 3502 | } |
56e93d26 JQ |
3503 | } |
3504 | ||
3505 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
3506 | { | |
3507 | unsigned int xh_len; | |
3508 | int xh_flags; | |
063e760a | 3509 | uint8_t *loaded_data; |
56e93d26 | 3510 | |
56e93d26 JQ |
3511 | /* extract RLE header */ |
3512 | xh_flags = qemu_get_byte(f); | |
3513 | xh_len = qemu_get_be16(f); | |
3514 | ||
3515 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
3516 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
3517 | return -1; | |
3518 | } | |
3519 | ||
3520 | if (xh_len > TARGET_PAGE_SIZE) { | |
3521 | error_report("Failed to load XBZRLE page - len overflow!"); | |
3522 | return -1; | |
3523 | } | |
f265e0e4 | 3524 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 3525 | /* load data and decode */ |
f265e0e4 | 3526 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 3527 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
3528 | |
3529 | /* decode RLE */ | |
063e760a | 3530 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
3531 | TARGET_PAGE_SIZE) == -1) { |
3532 | error_report("Failed to load XBZRLE page - decode error!"); | |
3533 | return -1; | |
3534 | } | |
3535 | ||
3536 | return 0; | |
3537 | } | |
3538 | ||
3d0684b2 JQ |
3539 | /** |
3540 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
3541 | * | |
3542 | * Must be called from within a rcu critical section. | |
3543 | * | |
56e93d26 | 3544 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 3545 | * |
755e8d7c | 3546 | * @mis: the migration incoming state pointer |
3d0684b2 JQ |
3547 | * @f: QEMUFile where to read the data from |
3548 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
c01b16ed | 3549 | * @channel: the channel we're using |
a7180877 | 3550 | */ |
755e8d7c | 3551 | static inline RAMBlock *ram_block_from_stream(MigrationIncomingState *mis, |
c01b16ed PX |
3552 | QEMUFile *f, int flags, |
3553 | int channel) | |
56e93d26 | 3554 | { |
c01b16ed | 3555 | RAMBlock *block = mis->last_recv_block[channel]; |
56e93d26 JQ |
3556 | char id[256]; |
3557 | uint8_t len; | |
3558 | ||
3559 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 3560 | if (!block) { |
56e93d26 JQ |
3561 | error_report("Ack, bad migration stream!"); |
3562 | return NULL; | |
3563 | } | |
4c4bad48 | 3564 | return block; |
56e93d26 JQ |
3565 | } |
3566 | ||
3567 | len = qemu_get_byte(f); | |
3568 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3569 | id[len] = 0; | |
3570 | ||
e3dd7493 | 3571 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
3572 | if (!block) { |
3573 | error_report("Can't find block %s", id); | |
3574 | return NULL; | |
56e93d26 JQ |
3575 | } |
3576 | ||
fbd162e6 | 3577 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
3578 | error_report("block %s should not be migrated !", id); |
3579 | return NULL; | |
3580 | } | |
3581 | ||
c01b16ed | 3582 | mis->last_recv_block[channel] = block; |
755e8d7c | 3583 | |
4c4bad48 HZ |
3584 | return block; |
3585 | } | |
3586 | ||
3587 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
3588 | ram_addr_t offset) | |
3589 | { | |
3590 | if (!offset_in_ramblock(block, offset)) { | |
3591 | return NULL; | |
3592 | } | |
3593 | ||
3594 | return block->host + offset; | |
56e93d26 JQ |
3595 | } |
3596 | ||
6a23f639 DH |
3597 | static void *host_page_from_ram_block_offset(RAMBlock *block, |
3598 | ram_addr_t offset) | |
3599 | { | |
3600 | /* Note: Explicitly no check against offset_in_ramblock(). */ | |
3601 | return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block->host + offset), | |
3602 | block->page_size); | |
3603 | } | |
3604 | ||
3605 | static ram_addr_t host_page_offset_from_ram_block_offset(RAMBlock *block, | |
3606 | ram_addr_t offset) | |
3607 | { | |
3608 | return ((uintptr_t)block->host + offset) & (block->page_size - 1); | |
3609 | } | |
3610 | ||
13af18f2 | 3611 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 3612 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
3613 | { |
3614 | if (!offset_in_ramblock(block, offset)) { | |
3615 | return NULL; | |
3616 | } | |
3617 | if (!block->colo_cache) { | |
3618 | error_report("%s: colo_cache is NULL in block :%s", | |
3619 | __func__, block->idstr); | |
3620 | return NULL; | |
3621 | } | |
7d9acafa ZC |
3622 | |
3623 | /* | |
3624 | * During colo checkpoint, we need bitmap of these migrated pages. | |
3625 | * It help us to decide which pages in ram cache should be flushed | |
3626 | * into VM's RAM later. | |
3627 | */ | |
8af66371 HZ |
3628 | if (record_bitmap && |
3629 | !test_and_set_bit(offset >> TARGET_PAGE_BITS, block->bmap)) { | |
7d9acafa ZC |
3630 | ram_state->migration_dirty_pages++; |
3631 | } | |
13af18f2 ZC |
3632 | return block->colo_cache + offset; |
3633 | } | |
3634 | ||
3d0684b2 JQ |
3635 | /** |
3636 | * ram_handle_compressed: handle the zero page case | |
3637 | * | |
56e93d26 JQ |
3638 | * If a page (or a whole RDMA chunk) has been |
3639 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
3640 | * |
3641 | * @host: host address for the zero page | |
3642 | * @ch: what the page is filled from. We only support zero | |
3643 | * @size: size of the zero page | |
56e93d26 JQ |
3644 | */ |
3645 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
3646 | { | |
bad452a7 | 3647 | if (ch != 0 || !buffer_is_zero(host, size)) { |
56e93d26 JQ |
3648 | memset(host, ch, size); |
3649 | } | |
3650 | } | |
3651 | ||
797ca154 XG |
3652 | /* return the size after decompression, or negative value on error */ |
3653 | static int | |
3654 | qemu_uncompress_data(z_stream *stream, uint8_t *dest, size_t dest_len, | |
3655 | const uint8_t *source, size_t source_len) | |
3656 | { | |
3657 | int err; | |
3658 | ||
3659 | err = inflateReset(stream); | |
3660 | if (err != Z_OK) { | |
3661 | return -1; | |
3662 | } | |
3663 | ||
3664 | stream->avail_in = source_len; | |
3665 | stream->next_in = (uint8_t *)source; | |
3666 | stream->avail_out = dest_len; | |
3667 | stream->next_out = dest; | |
3668 | ||
3669 | err = inflate(stream, Z_NO_FLUSH); | |
3670 | if (err != Z_STREAM_END) { | |
3671 | return -1; | |
3672 | } | |
3673 | ||
3674 | return stream->total_out; | |
3675 | } | |
3676 | ||
56e93d26 JQ |
3677 | static void *do_data_decompress(void *opaque) |
3678 | { | |
3679 | DecompressParam *param = opaque; | |
3680 | unsigned long pagesize; | |
33d151f4 | 3681 | uint8_t *des; |
34ab9e97 | 3682 | int len, ret; |
56e93d26 | 3683 | |
33d151f4 | 3684 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 3685 | while (!param->quit) { |
33d151f4 LL |
3686 | if (param->des) { |
3687 | des = param->des; | |
3688 | len = param->len; | |
3689 | param->des = 0; | |
3690 | qemu_mutex_unlock(¶m->mutex); | |
3691 | ||
56e93d26 | 3692 | pagesize = TARGET_PAGE_SIZE; |
34ab9e97 XG |
3693 | |
3694 | ret = qemu_uncompress_data(¶m->stream, des, pagesize, | |
3695 | param->compbuf, len); | |
f548222c | 3696 | if (ret < 0 && migrate_get_current()->decompress_error_check) { |
34ab9e97 XG |
3697 | error_report("decompress data failed"); |
3698 | qemu_file_set_error(decomp_file, ret); | |
3699 | } | |
73a8912b | 3700 | |
33d151f4 LL |
3701 | qemu_mutex_lock(&decomp_done_lock); |
3702 | param->done = true; | |
3703 | qemu_cond_signal(&decomp_done_cond); | |
3704 | qemu_mutex_unlock(&decomp_done_lock); | |
3705 | ||
3706 | qemu_mutex_lock(¶m->mutex); | |
3707 | } else { | |
3708 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
3709 | } | |
56e93d26 | 3710 | } |
33d151f4 | 3711 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
3712 | |
3713 | return NULL; | |
3714 | } | |
3715 | ||
34ab9e97 | 3716 | static int wait_for_decompress_done(void) |
5533b2e9 LL |
3717 | { |
3718 | int idx, thread_count; | |
3719 | ||
3720 | if (!migrate_use_compression()) { | |
34ab9e97 | 3721 | return 0; |
5533b2e9 LL |
3722 | } |
3723 | ||
3724 | thread_count = migrate_decompress_threads(); | |
3725 | qemu_mutex_lock(&decomp_done_lock); | |
3726 | for (idx = 0; idx < thread_count; idx++) { | |
3727 | while (!decomp_param[idx].done) { | |
3728 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
3729 | } | |
3730 | } | |
3731 | qemu_mutex_unlock(&decomp_done_lock); | |
34ab9e97 | 3732 | return qemu_file_get_error(decomp_file); |
5533b2e9 LL |
3733 | } |
3734 | ||
f0afa331 | 3735 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
3736 | { |
3737 | int i, thread_count; | |
3738 | ||
3416ab5b JQ |
3739 | if (!migrate_use_compression()) { |
3740 | return; | |
3741 | } | |
56e93d26 JQ |
3742 | thread_count = migrate_decompress_threads(); |
3743 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3744 | /* |
3745 | * we use it as a indicator which shows if the thread is | |
3746 | * properly init'd or not | |
3747 | */ | |
3748 | if (!decomp_param[i].compbuf) { | |
3749 | break; | |
3750 | } | |
3751 | ||
56e93d26 | 3752 | qemu_mutex_lock(&decomp_param[i].mutex); |
90e56fb4 | 3753 | decomp_param[i].quit = true; |
56e93d26 JQ |
3754 | qemu_cond_signal(&decomp_param[i].cond); |
3755 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
3756 | } | |
3757 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3758 | if (!decomp_param[i].compbuf) { |
3759 | break; | |
3760 | } | |
3761 | ||
56e93d26 JQ |
3762 | qemu_thread_join(decompress_threads + i); |
3763 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
3764 | qemu_cond_destroy(&decomp_param[i].cond); | |
797ca154 | 3765 | inflateEnd(&decomp_param[i].stream); |
56e93d26 | 3766 | g_free(decomp_param[i].compbuf); |
797ca154 | 3767 | decomp_param[i].compbuf = NULL; |
56e93d26 JQ |
3768 | } |
3769 | g_free(decompress_threads); | |
3770 | g_free(decomp_param); | |
56e93d26 JQ |
3771 | decompress_threads = NULL; |
3772 | decomp_param = NULL; | |
34ab9e97 | 3773 | decomp_file = NULL; |
56e93d26 JQ |
3774 | } |
3775 | ||
34ab9e97 | 3776 | static int compress_threads_load_setup(QEMUFile *f) |
797ca154 XG |
3777 | { |
3778 | int i, thread_count; | |
3779 | ||
3780 | if (!migrate_use_compression()) { | |
3781 | return 0; | |
3782 | } | |
3783 | ||
3784 | thread_count = migrate_decompress_threads(); | |
3785 | decompress_threads = g_new0(QemuThread, thread_count); | |
3786 | decomp_param = g_new0(DecompressParam, thread_count); | |
3787 | qemu_mutex_init(&decomp_done_lock); | |
3788 | qemu_cond_init(&decomp_done_cond); | |
34ab9e97 | 3789 | decomp_file = f; |
797ca154 XG |
3790 | for (i = 0; i < thread_count; i++) { |
3791 | if (inflateInit(&decomp_param[i].stream) != Z_OK) { | |
3792 | goto exit; | |
3793 | } | |
3794 | ||
3795 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
3796 | qemu_mutex_init(&decomp_param[i].mutex); | |
3797 | qemu_cond_init(&decomp_param[i].cond); | |
3798 | decomp_param[i].done = true; | |
3799 | decomp_param[i].quit = false; | |
3800 | qemu_thread_create(decompress_threads + i, "decompress", | |
3801 | do_data_decompress, decomp_param + i, | |
3802 | QEMU_THREAD_JOINABLE); | |
3803 | } | |
3804 | return 0; | |
3805 | exit: | |
3806 | compress_threads_load_cleanup(); | |
3807 | return -1; | |
3808 | } | |
3809 | ||
c1bc6626 | 3810 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
3811 | void *host, int len) |
3812 | { | |
3813 | int idx, thread_count; | |
3814 | ||
3815 | thread_count = migrate_decompress_threads(); | |
37396950 | 3816 | QEMU_LOCK_GUARD(&decomp_done_lock); |
56e93d26 JQ |
3817 | while (true) { |
3818 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 3819 | if (decomp_param[idx].done) { |
33d151f4 LL |
3820 | decomp_param[idx].done = false; |
3821 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 3822 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
3823 | decomp_param[idx].des = host; |
3824 | decomp_param[idx].len = len; | |
33d151f4 LL |
3825 | qemu_cond_signal(&decomp_param[idx].cond); |
3826 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
3827 | break; |
3828 | } | |
3829 | } | |
3830 | if (idx < thread_count) { | |
3831 | break; | |
73a8912b LL |
3832 | } else { |
3833 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
3834 | } |
3835 | } | |
3836 | } | |
3837 | ||
b70cb3b4 RL |
3838 | static void colo_init_ram_state(void) |
3839 | { | |
3840 | ram_state_init(&ram_state); | |
b70cb3b4 RL |
3841 | } |
3842 | ||
13af18f2 ZC |
3843 | /* |
3844 | * colo cache: this is for secondary VM, we cache the whole | |
3845 | * memory of the secondary VM, it is need to hold the global lock | |
3846 | * to call this helper. | |
3847 | */ | |
3848 | int colo_init_ram_cache(void) | |
3849 | { | |
3850 | RAMBlock *block; | |
3851 | ||
44901b5a PB |
3852 | WITH_RCU_READ_LOCK_GUARD() { |
3853 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3854 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
8dbe22c6 | 3855 | NULL, false, false); |
44901b5a PB |
3856 | if (!block->colo_cache) { |
3857 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
3858 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
3859 | block->used_length); | |
3860 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3861 | if (block->colo_cache) { | |
3862 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3863 | block->colo_cache = NULL; | |
3864 | } | |
89ac5a1d | 3865 | } |
44901b5a | 3866 | return -errno; |
89ac5a1d | 3867 | } |
e5fdf920 LS |
3868 | if (!machine_dump_guest_core(current_machine)) { |
3869 | qemu_madvise(block->colo_cache, block->used_length, | |
3870 | QEMU_MADV_DONTDUMP); | |
3871 | } | |
13af18f2 | 3872 | } |
13af18f2 | 3873 | } |
44901b5a | 3874 | |
7d9acafa ZC |
3875 | /* |
3876 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3877 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3878 | * we use the same name 'ram_bitmap' as for migration. | |
3879 | */ | |
3880 | if (ram_bytes_total()) { | |
3881 | RAMBlock *block; | |
3882 | ||
fbd162e6 | 3883 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3884 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3885 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3886 | } |
3887 | } | |
7d9acafa | 3888 | |
b70cb3b4 | 3889 | colo_init_ram_state(); |
13af18f2 | 3890 | return 0; |
13af18f2 ZC |
3891 | } |
3892 | ||
0393031a HZ |
3893 | /* TODO: duplicated with ram_init_bitmaps */ |
3894 | void colo_incoming_start_dirty_log(void) | |
3895 | { | |
3896 | RAMBlock *block = NULL; | |
3897 | /* For memory_global_dirty_log_start below. */ | |
3898 | qemu_mutex_lock_iothread(); | |
3899 | qemu_mutex_lock_ramlist(); | |
3900 | ||
3901 | memory_global_dirty_log_sync(); | |
3902 | WITH_RCU_READ_LOCK_GUARD() { | |
3903 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3904 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3905 | /* Discard this dirty bitmap record */ | |
3906 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3907 | } | |
63b41db4 | 3908 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
0393031a HZ |
3909 | } |
3910 | ram_state->migration_dirty_pages = 0; | |
3911 | qemu_mutex_unlock_ramlist(); | |
3912 | qemu_mutex_unlock_iothread(); | |
3913 | } | |
3914 | ||
13af18f2 ZC |
3915 | /* It is need to hold the global lock to call this helper */ |
3916 | void colo_release_ram_cache(void) | |
3917 | { | |
3918 | RAMBlock *block; | |
3919 | ||
63b41db4 | 3920 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); |
fbd162e6 | 3921 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3922 | g_free(block->bmap); |
3923 | block->bmap = NULL; | |
3924 | } | |
3925 | ||
89ac5a1d DDAG |
3926 | WITH_RCU_READ_LOCK_GUARD() { |
3927 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3928 | if (block->colo_cache) { | |
3929 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3930 | block->colo_cache = NULL; | |
3931 | } | |
13af18f2 ZC |
3932 | } |
3933 | } | |
0393031a | 3934 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
3935 | } |
3936 | ||
f265e0e4 JQ |
3937 | /** |
3938 | * ram_load_setup: Setup RAM for migration incoming side | |
3939 | * | |
3940 | * Returns zero to indicate success and negative for error | |
3941 | * | |
3942 | * @f: QEMUFile where to receive the data | |
3943 | * @opaque: RAMState pointer | |
3944 | */ | |
3945 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
3946 | { | |
34ab9e97 | 3947 | if (compress_threads_load_setup(f)) { |
797ca154 XG |
3948 | return -1; |
3949 | } | |
3950 | ||
f265e0e4 | 3951 | xbzrle_load_setup(); |
f9494614 | 3952 | ramblock_recv_map_init(); |
13af18f2 | 3953 | |
f265e0e4 JQ |
3954 | return 0; |
3955 | } | |
3956 | ||
3957 | static int ram_load_cleanup(void *opaque) | |
3958 | { | |
f9494614 | 3959 | RAMBlock *rb; |
56eb90af | 3960 | |
fbd162e6 | 3961 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 3962 | qemu_ram_block_writeback(rb); |
56eb90af JH |
3963 | } |
3964 | ||
f265e0e4 | 3965 | xbzrle_load_cleanup(); |
f0afa331 | 3966 | compress_threads_load_cleanup(); |
f9494614 | 3967 | |
fbd162e6 | 3968 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
3969 | g_free(rb->receivedmap); |
3970 | rb->receivedmap = NULL; | |
3971 | } | |
13af18f2 | 3972 | |
f265e0e4 JQ |
3973 | return 0; |
3974 | } | |
3975 | ||
3d0684b2 JQ |
3976 | /** |
3977 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
3978 | * | |
3979 | * Returns 0 for success and negative if there was one error | |
3980 | * | |
3981 | * @mis: current migration incoming state | |
3982 | * | |
3983 | * Allocate data structures etc needed by incoming migration with | |
3984 | * postcopy-ram. postcopy-ram's similarly names | |
3985 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
3986 | */ |
3987 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
3988 | { | |
c136180c | 3989 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
3990 | } |
3991 | ||
3d0684b2 JQ |
3992 | /** |
3993 | * ram_load_postcopy: load a page in postcopy case | |
3994 | * | |
3995 | * Returns 0 for success or -errno in case of error | |
3996 | * | |
a7180877 DDAG |
3997 | * Called in postcopy mode by ram_load(). |
3998 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
3999 | * |
4000 | * @f: QEMUFile where to send the data | |
36f62f11 | 4001 | * @channel: the channel to use for loading |
a7180877 | 4002 | */ |
36f62f11 | 4003 | int ram_load_postcopy(QEMUFile *f, int channel) |
a7180877 DDAG |
4004 | { |
4005 | int flags = 0, ret = 0; | |
4006 | bool place_needed = false; | |
1aa83678 | 4007 | bool matches_target_page_size = false; |
a7180877 | 4008 | MigrationIncomingState *mis = migration_incoming_get_current(); |
36f62f11 | 4009 | PostcopyTmpPage *tmp_page = &mis->postcopy_tmp_pages[channel]; |
a7180877 DDAG |
4010 | |
4011 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
4012 | ram_addr_t addr; | |
a7180877 DDAG |
4013 | void *page_buffer = NULL; |
4014 | void *place_source = NULL; | |
df9ff5e1 | 4015 | RAMBlock *block = NULL; |
a7180877 | 4016 | uint8_t ch; |
644acf99 | 4017 | int len; |
a7180877 DDAG |
4018 | |
4019 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
4020 | |
4021 | /* | |
4022 | * If qemu file error, we should stop here, and then "addr" | |
4023 | * may be invalid | |
4024 | */ | |
4025 | ret = qemu_file_get_error(f); | |
4026 | if (ret) { | |
4027 | break; | |
4028 | } | |
4029 | ||
a7180877 DDAG |
4030 | flags = addr & ~TARGET_PAGE_MASK; |
4031 | addr &= TARGET_PAGE_MASK; | |
4032 | ||
36f62f11 | 4033 | trace_ram_load_postcopy_loop(channel, (uint64_t)addr, flags); |
644acf99 WY |
4034 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
4035 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
c01b16ed | 4036 | block = ram_block_from_stream(mis, f, flags, channel); |
6a23f639 DH |
4037 | if (!block) { |
4038 | ret = -EINVAL; | |
4039 | break; | |
4040 | } | |
4c4bad48 | 4041 | |
898ba906 DH |
4042 | /* |
4043 | * Relying on used_length is racy and can result in false positives. | |
4044 | * We might place pages beyond used_length in case RAM was shrunk | |
4045 | * while in postcopy, which is fine - trying to place via | |
4046 | * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault. | |
4047 | */ | |
4048 | if (!block->host || addr >= block->postcopy_length) { | |
a7180877 DDAG |
4049 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); |
4050 | ret = -EINVAL; | |
4051 | break; | |
4052 | } | |
77dadc3f | 4053 | tmp_page->target_pages++; |
1aa83678 | 4054 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 4055 | /* |
28abd200 DDAG |
4056 | * Postcopy requires that we place whole host pages atomically; |
4057 | * these may be huge pages for RAMBlocks that are backed by | |
4058 | * hugetlbfs. | |
a7180877 DDAG |
4059 | * To make it atomic, the data is read into a temporary page |
4060 | * that's moved into place later. | |
4061 | * The migration protocol uses, possibly smaller, target-pages | |
4062 | * however the source ensures it always sends all the components | |
91ba442f | 4063 | * of a host page in one chunk. |
a7180877 | 4064 | */ |
77dadc3f | 4065 | page_buffer = tmp_page->tmp_huge_page + |
6a23f639 DH |
4066 | host_page_offset_from_ram_block_offset(block, addr); |
4067 | /* If all TP are zero then we can optimise the place */ | |
77dadc3f PX |
4068 | if (tmp_page->target_pages == 1) { |
4069 | tmp_page->host_addr = | |
4070 | host_page_from_ram_block_offset(block, addr); | |
4071 | } else if (tmp_page->host_addr != | |
4072 | host_page_from_ram_block_offset(block, addr)) { | |
c53b7ddc | 4073 | /* not the 1st TP within the HP */ |
36f62f11 | 4074 | error_report("Non-same host page detected on channel %d: " |
cfc7dc8a PX |
4075 | "Target host page %p, received host page %p " |
4076 | "(rb %s offset 0x"RAM_ADDR_FMT" target_pages %d)", | |
36f62f11 | 4077 | channel, tmp_page->host_addr, |
cfc7dc8a PX |
4078 | host_page_from_ram_block_offset(block, addr), |
4079 | block->idstr, addr, tmp_page->target_pages); | |
6a23f639 DH |
4080 | ret = -EINVAL; |
4081 | break; | |
a7180877 DDAG |
4082 | } |
4083 | ||
4084 | /* | |
4085 | * If it's the last part of a host page then we place the host | |
4086 | * page | |
4087 | */ | |
77dadc3f PX |
4088 | if (tmp_page->target_pages == |
4089 | (block->page_size / TARGET_PAGE_SIZE)) { | |
4cbb3c63 | 4090 | place_needed = true; |
4cbb3c63 | 4091 | } |
77dadc3f | 4092 | place_source = tmp_page->tmp_huge_page; |
a7180877 DDAG |
4093 | } |
4094 | ||
4095 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 4096 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 4097 | ch = qemu_get_byte(f); |
2e36bc1b WY |
4098 | /* |
4099 | * Can skip to set page_buffer when | |
4100 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
4101 | */ | |
4102 | if (ch || !matches_target_page_size) { | |
4103 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
4104 | } | |
a7180877 | 4105 | if (ch) { |
77dadc3f | 4106 | tmp_page->all_zero = false; |
a7180877 DDAG |
4107 | } |
4108 | break; | |
4109 | ||
4110 | case RAM_SAVE_FLAG_PAGE: | |
77dadc3f | 4111 | tmp_page->all_zero = false; |
1aa83678 PX |
4112 | if (!matches_target_page_size) { |
4113 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
4114 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
4115 | } else { | |
1aa83678 PX |
4116 | /* |
4117 | * For small pages that matches target page size, we | |
4118 | * avoid the qemu_file copy. Instead we directly use | |
4119 | * the buffer of QEMUFile to place the page. Note: we | |
4120 | * cannot do any QEMUFile operation before using that | |
4121 | * buffer to make sure the buffer is valid when | |
4122 | * placing the page. | |
a7180877 DDAG |
4123 | */ |
4124 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
4125 | TARGET_PAGE_SIZE); | |
4126 | } | |
4127 | break; | |
644acf99 | 4128 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
77dadc3f | 4129 | tmp_page->all_zero = false; |
644acf99 WY |
4130 | len = qemu_get_be32(f); |
4131 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4132 | error_report("Invalid compressed data length: %d", len); | |
4133 | ret = -EINVAL; | |
4134 | break; | |
4135 | } | |
4136 | decompress_data_with_multi_threads(f, page_buffer, len); | |
4137 | break; | |
4138 | ||
a7180877 DDAG |
4139 | case RAM_SAVE_FLAG_EOS: |
4140 | /* normal exit */ | |
6df264ac | 4141 | multifd_recv_sync_main(); |
a7180877 DDAG |
4142 | break; |
4143 | default: | |
29fccade | 4144 | error_report("Unknown combination of migration flags: 0x%x" |
a7180877 DDAG |
4145 | " (postcopy mode)", flags); |
4146 | ret = -EINVAL; | |
7a9ddfbf PX |
4147 | break; |
4148 | } | |
4149 | ||
644acf99 WY |
4150 | /* Got the whole host page, wait for decompress before placing. */ |
4151 | if (place_needed) { | |
4152 | ret |= wait_for_decompress_done(); | |
4153 | } | |
4154 | ||
7a9ddfbf PX |
4155 | /* Detect for any possible file errors */ |
4156 | if (!ret && qemu_file_get_error(f)) { | |
4157 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
4158 | } |
4159 | ||
7a9ddfbf | 4160 | if (!ret && place_needed) { |
77dadc3f PX |
4161 | if (tmp_page->all_zero) { |
4162 | ret = postcopy_place_page_zero(mis, tmp_page->host_addr, block); | |
a7180877 | 4163 | } else { |
77dadc3f PX |
4164 | ret = postcopy_place_page(mis, tmp_page->host_addr, |
4165 | place_source, block); | |
a7180877 | 4166 | } |
ddf35bdf | 4167 | place_needed = false; |
77dadc3f | 4168 | postcopy_temp_page_reset(tmp_page); |
a7180877 | 4169 | } |
a7180877 DDAG |
4170 | } |
4171 | ||
4172 | return ret; | |
4173 | } | |
4174 | ||
acab30b8 DHB |
4175 | static bool postcopy_is_running(void) |
4176 | { | |
4177 | PostcopyState ps = postcopy_state_get(); | |
4178 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
4179 | } | |
4180 | ||
e6f4aa18 ZC |
4181 | /* |
4182 | * Flush content of RAM cache into SVM's memory. | |
4183 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
4184 | */ | |
24fa16f8 | 4185 | void colo_flush_ram_cache(void) |
e6f4aa18 ZC |
4186 | { |
4187 | RAMBlock *block = NULL; | |
4188 | void *dst_host; | |
4189 | void *src_host; | |
4190 | unsigned long offset = 0; | |
4191 | ||
d1955d22 | 4192 | memory_global_dirty_log_sync(); |
89ac5a1d DDAG |
4193 | WITH_RCU_READ_LOCK_GUARD() { |
4194 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
4195 | ramblock_sync_dirty_bitmap(ram_state, block); | |
4196 | } | |
d1955d22 | 4197 | } |
d1955d22 | 4198 | |
e6f4aa18 | 4199 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
4200 | WITH_RCU_READ_LOCK_GUARD() { |
4201 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 4202 | |
89ac5a1d | 4203 | while (block) { |
a6a83cef | 4204 | unsigned long num = 0; |
e6f4aa18 | 4205 | |
a6a83cef | 4206 | offset = colo_bitmap_find_dirty(ram_state, block, offset, &num); |
542147f4 DH |
4207 | if (!offset_in_ramblock(block, |
4208 | ((ram_addr_t)offset) << TARGET_PAGE_BITS)) { | |
89ac5a1d | 4209 | offset = 0; |
a6a83cef | 4210 | num = 0; |
89ac5a1d DDAG |
4211 | block = QLIST_NEXT_RCU(block, next); |
4212 | } else { | |
a6a83cef RL |
4213 | unsigned long i = 0; |
4214 | ||
4215 | for (i = 0; i < num; i++) { | |
4216 | migration_bitmap_clear_dirty(ram_state, block, offset + i); | |
4217 | } | |
8bba004c AR |
4218 | dst_host = block->host |
4219 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
4220 | src_host = block->colo_cache | |
4221 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
a6a83cef RL |
4222 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE * num); |
4223 | offset += num; | |
89ac5a1d | 4224 | } |
e6f4aa18 ZC |
4225 | } |
4226 | } | |
e6f4aa18 ZC |
4227 | trace_colo_flush_ram_cache_end(); |
4228 | } | |
4229 | ||
10da4a36 WY |
4230 | /** |
4231 | * ram_load_precopy: load pages in precopy case | |
4232 | * | |
4233 | * Returns 0 for success or -errno in case of error | |
4234 | * | |
4235 | * Called in precopy mode by ram_load(). | |
4236 | * rcu_read_lock is taken prior to this being called. | |
4237 | * | |
4238 | * @f: QEMUFile where to send the data | |
4239 | */ | |
4240 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 4241 | { |
755e8d7c | 4242 | MigrationIncomingState *mis = migration_incoming_get_current(); |
e65cec5e | 4243 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 4244 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
80fe315c | 4245 | bool postcopy_advised = migration_incoming_postcopy_advised(); |
edc60127 JQ |
4246 | if (!migrate_use_compression()) { |
4247 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
4248 | } | |
a7180877 | 4249 | |
10da4a36 | 4250 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 4251 | ram_addr_t addr, total_ram_bytes; |
0393031a | 4252 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
4253 | uint8_t ch; |
4254 | ||
e65cec5e YK |
4255 | /* |
4256 | * Yield periodically to let main loop run, but an iteration of | |
4257 | * the main loop is expensive, so do it each some iterations | |
4258 | */ | |
4259 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
4260 | aio_co_schedule(qemu_get_current_aio_context(), | |
4261 | qemu_coroutine_self()); | |
4262 | qemu_coroutine_yield(); | |
4263 | } | |
4264 | i++; | |
4265 | ||
56e93d26 JQ |
4266 | addr = qemu_get_be64(f); |
4267 | flags = addr & ~TARGET_PAGE_MASK; | |
4268 | addr &= TARGET_PAGE_MASK; | |
4269 | ||
edc60127 JQ |
4270 | if (flags & invalid_flags) { |
4271 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
4272 | error_report("Received an unexpected compressed page"); | |
4273 | } | |
4274 | ||
4275 | ret = -EINVAL; | |
4276 | break; | |
4277 | } | |
4278 | ||
bb890ed5 | 4279 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 4280 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
c01b16ed PX |
4281 | RAMBlock *block = ram_block_from_stream(mis, f, flags, |
4282 | RAM_CHANNEL_PRECOPY); | |
4c4bad48 | 4283 | |
0393031a | 4284 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 4285 | /* |
0393031a HZ |
4286 | * After going into COLO stage, we should not load the page |
4287 | * into SVM's memory directly, we put them into colo_cache firstly. | |
4288 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
4289 | * Previously, we copied all these memory in preparing stage of COLO | |
4290 | * while we need to stop VM, which is a time-consuming process. | |
4291 | * Here we optimize it by a trick, back-up every page while in | |
4292 | * migration process while COLO is enabled, though it affects the | |
4293 | * speed of the migration, but it obviously reduce the downtime of | |
4294 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 4295 | */ |
0393031a HZ |
4296 | if (migration_incoming_colo_enabled()) { |
4297 | if (migration_incoming_in_colo_state()) { | |
4298 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 4299 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
4300 | } else { |
4301 | /* | |
4302 | * In migration stage but before COLO stage, | |
4303 | * Put all pages into both cache and SVM's memory. | |
4304 | */ | |
8af66371 | 4305 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 4306 | } |
13af18f2 | 4307 | } |
a776aa15 DDAG |
4308 | if (!host) { |
4309 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
4310 | ret = -EINVAL; | |
4311 | break; | |
4312 | } | |
13af18f2 ZC |
4313 | if (!migration_incoming_in_colo_state()) { |
4314 | ramblock_recv_bitmap_set(block, host); | |
4315 | } | |
4316 | ||
1db9d8e5 | 4317 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
4318 | } |
4319 | ||
56e93d26 JQ |
4320 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
4321 | case RAM_SAVE_FLAG_MEM_SIZE: | |
4322 | /* Synchronize RAM block list */ | |
4323 | total_ram_bytes = addr; | |
4324 | while (!ret && total_ram_bytes) { | |
4325 | RAMBlock *block; | |
56e93d26 JQ |
4326 | char id[256]; |
4327 | ram_addr_t length; | |
4328 | ||
4329 | len = qemu_get_byte(f); | |
4330 | qemu_get_buffer(f, (uint8_t *)id, len); | |
4331 | id[len] = 0; | |
4332 | length = qemu_get_be64(f); | |
4333 | ||
e3dd7493 | 4334 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
4335 | if (block && !qemu_ram_is_migratable(block)) { |
4336 | error_report("block %s should not be migrated !", id); | |
4337 | ret = -EINVAL; | |
4338 | } else if (block) { | |
e3dd7493 DDAG |
4339 | if (length != block->used_length) { |
4340 | Error *local_err = NULL; | |
56e93d26 | 4341 | |
fa53a0e5 | 4342 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
4343 | &local_err); |
4344 | if (local_err) { | |
4345 | error_report_err(local_err); | |
56e93d26 | 4346 | } |
56e93d26 | 4347 | } |
ef08fb38 | 4348 | /* For postcopy we need to check hugepage sizes match */ |
e846b746 | 4349 | if (postcopy_advised && migrate_postcopy_ram() && |
ef08fb38 DDAG |
4350 | block->page_size != qemu_host_page_size) { |
4351 | uint64_t remote_page_size = qemu_get_be64(f); | |
4352 | if (remote_page_size != block->page_size) { | |
4353 | error_report("Mismatched RAM page size %s " | |
4354 | "(local) %zd != %" PRId64, | |
4355 | id, block->page_size, | |
4356 | remote_page_size); | |
4357 | ret = -EINVAL; | |
4358 | } | |
4359 | } | |
fbd162e6 YK |
4360 | if (migrate_ignore_shared()) { |
4361 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
4362 | if (ramblock_is_ignored(block) && |
4363 | block->mr->addr != addr) { | |
4364 | error_report("Mismatched GPAs for block %s " | |
4365 | "%" PRId64 "!= %" PRId64, | |
4366 | id, (uint64_t)addr, | |
4367 | (uint64_t)block->mr->addr); | |
4368 | ret = -EINVAL; | |
4369 | } | |
4370 | } | |
e3dd7493 DDAG |
4371 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
4372 | block->idstr); | |
4373 | } else { | |
56e93d26 JQ |
4374 | error_report("Unknown ramblock \"%s\", cannot " |
4375 | "accept migration", id); | |
4376 | ret = -EINVAL; | |
4377 | } | |
4378 | ||
4379 | total_ram_bytes -= length; | |
4380 | } | |
4381 | break; | |
a776aa15 | 4382 | |
bb890ed5 | 4383 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
4384 | ch = qemu_get_byte(f); |
4385 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
4386 | break; | |
a776aa15 | 4387 | |
56e93d26 | 4388 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
4389 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
4390 | break; | |
56e93d26 | 4391 | |
a776aa15 | 4392 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
4393 | len = qemu_get_be32(f); |
4394 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4395 | error_report("Invalid compressed data length: %d", len); | |
4396 | ret = -EINVAL; | |
4397 | break; | |
4398 | } | |
c1bc6626 | 4399 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 4400 | break; |
a776aa15 | 4401 | |
56e93d26 | 4402 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
4403 | if (load_xbzrle(f, addr, host) < 0) { |
4404 | error_report("Failed to decompress XBZRLE page at " | |
4405 | RAM_ADDR_FMT, addr); | |
4406 | ret = -EINVAL; | |
4407 | break; | |
4408 | } | |
4409 | break; | |
4410 | case RAM_SAVE_FLAG_EOS: | |
4411 | /* normal exit */ | |
6df264ac | 4412 | multifd_recv_sync_main(); |
56e93d26 JQ |
4413 | break; |
4414 | default: | |
4415 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 4416 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 | 4417 | } else { |
29fccade | 4418 | error_report("Unknown combination of migration flags: 0x%x", |
56e93d26 JQ |
4419 | flags); |
4420 | ret = -EINVAL; | |
4421 | } | |
4422 | } | |
4423 | if (!ret) { | |
4424 | ret = qemu_file_get_error(f); | |
4425 | } | |
0393031a HZ |
4426 | if (!ret && host_bak) { |
4427 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
4428 | } | |
56e93d26 JQ |
4429 | } |
4430 | ||
ca1a6b70 | 4431 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
4432 | return ret; |
4433 | } | |
4434 | ||
4435 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
4436 | { | |
4437 | int ret = 0; | |
4438 | static uint64_t seq_iter; | |
4439 | /* | |
4440 | * If system is running in postcopy mode, page inserts to host memory must | |
4441 | * be atomic | |
4442 | */ | |
4443 | bool postcopy_running = postcopy_is_running(); | |
4444 | ||
4445 | seq_iter++; | |
4446 | ||
4447 | if (version_id != 4) { | |
4448 | return -EINVAL; | |
4449 | } | |
4450 | ||
4451 | /* | |
4452 | * This RCU critical section can be very long running. | |
4453 | * When RCU reclaims in the code start to become numerous, | |
4454 | * it will be necessary to reduce the granularity of this | |
4455 | * critical section. | |
4456 | */ | |
89ac5a1d DDAG |
4457 | WITH_RCU_READ_LOCK_GUARD() { |
4458 | if (postcopy_running) { | |
36f62f11 PX |
4459 | /* |
4460 | * Note! Here RAM_CHANNEL_PRECOPY is the precopy channel of | |
4461 | * postcopy migration, we have another RAM_CHANNEL_POSTCOPY to | |
4462 | * service fast page faults. | |
4463 | */ | |
4464 | ret = ram_load_postcopy(f, RAM_CHANNEL_PRECOPY); | |
89ac5a1d DDAG |
4465 | } else { |
4466 | ret = ram_load_precopy(f); | |
4467 | } | |
10da4a36 | 4468 | } |
55c4446b | 4469 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 | 4470 | |
56e93d26 JQ |
4471 | return ret; |
4472 | } | |
4473 | ||
c6467627 VSO |
4474 | static bool ram_has_postcopy(void *opaque) |
4475 | { | |
469dd51b | 4476 | RAMBlock *rb; |
fbd162e6 | 4477 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
4478 | if (ramblock_is_pmem(rb)) { |
4479 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
4480 | "is not supported now!", rb->idstr, rb->host); | |
4481 | return false; | |
4482 | } | |
4483 | } | |
4484 | ||
c6467627 VSO |
4485 | return migrate_postcopy_ram(); |
4486 | } | |
4487 | ||
edd090c7 PX |
4488 | /* Sync all the dirty bitmap with destination VM. */ |
4489 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
4490 | { | |
4491 | RAMBlock *block; | |
4492 | QEMUFile *file = s->to_dst_file; | |
4493 | int ramblock_count = 0; | |
4494 | ||
4495 | trace_ram_dirty_bitmap_sync_start(); | |
4496 | ||
fbd162e6 | 4497 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
4498 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
4499 | trace_ram_dirty_bitmap_request(block->idstr); | |
4500 | ramblock_count++; | |
4501 | } | |
4502 | ||
4503 | trace_ram_dirty_bitmap_sync_wait(); | |
4504 | ||
4505 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
4506 | while (ramblock_count--) { | |
4507 | qemu_sem_wait(&s->rp_state.rp_sem); | |
4508 | } | |
4509 | ||
4510 | trace_ram_dirty_bitmap_sync_complete(); | |
4511 | ||
4512 | return 0; | |
4513 | } | |
4514 | ||
4515 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
4516 | { | |
4517 | qemu_sem_post(&s->rp_state.rp_sem); | |
4518 | } | |
4519 | ||
a335debb PX |
4520 | /* |
4521 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
4522 | * This is only used when the postcopy migration is paused but wants | |
4523 | * to resume from a middle point. | |
4524 | */ | |
4525 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
4526 | { | |
4527 | int ret = -EINVAL; | |
43044ac0 | 4528 | /* from_dst_file is always valid because we're within rp_thread */ |
a335debb PX |
4529 | QEMUFile *file = s->rp_state.from_dst_file; |
4530 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 4531 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
4532 | uint64_t size, end_mark; |
4533 | ||
4534 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
4535 | ||
4536 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
4537 | error_report("%s: incorrect state %s", __func__, | |
4538 | MigrationStatus_str(s->state)); | |
4539 | return -EINVAL; | |
4540 | } | |
4541 | ||
4542 | /* | |
4543 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3a4452d8 | 4544 | * need the endianness conversion, and the paddings. |
a335debb PX |
4545 | */ |
4546 | local_size = ROUND_UP(local_size, 8); | |
4547 | ||
4548 | /* Add paddings */ | |
4549 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
4550 | ||
4551 | size = qemu_get_be64(file); | |
4552 | ||
4553 | /* The size of the bitmap should match with our ramblock */ | |
4554 | if (size != local_size) { | |
4555 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
4556 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
4557 | block->idstr, size, local_size); | |
4558 | ret = -EINVAL; | |
4559 | goto out; | |
4560 | } | |
4561 | ||
4562 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
4563 | end_mark = qemu_get_be64(file); | |
4564 | ||
4565 | ret = qemu_file_get_error(file); | |
4566 | if (ret || size != local_size) { | |
4567 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
4568 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
4569 | __func__, block->idstr, ret, local_size, size); | |
4570 | ret = -EIO; | |
4571 | goto out; | |
4572 | } | |
4573 | ||
4574 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
af3bbbe9 | 4575 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64, |
a335debb PX |
4576 | __func__, block->idstr, end_mark); |
4577 | ret = -EINVAL; | |
4578 | goto out; | |
4579 | } | |
4580 | ||
4581 | /* | |
3a4452d8 | 4582 | * Endianness conversion. We are during postcopy (though paused). |
a335debb PX |
4583 | * The dirty bitmap won't change. We can directly modify it. |
4584 | */ | |
4585 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
4586 | ||
4587 | /* | |
4588 | * What we received is "received bitmap". Revert it as the initial | |
4589 | * dirty bitmap for this ramblock. | |
4590 | */ | |
4591 | bitmap_complement(block->bmap, block->bmap, nbits); | |
4592 | ||
be39b4cd DH |
4593 | /* Clear dirty bits of discarded ranges that we don't want to migrate. */ |
4594 | ramblock_dirty_bitmap_clear_discarded_pages(block); | |
4595 | ||
4596 | /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */ | |
a335debb PX |
4597 | trace_ram_dirty_bitmap_reload_complete(block->idstr); |
4598 | ||
edd090c7 PX |
4599 | /* |
4600 | * We succeeded to sync bitmap for current ramblock. If this is | |
4601 | * the last one to sync, we need to notify the main send thread. | |
4602 | */ | |
4603 | ram_dirty_bitmap_reload_notify(s); | |
4604 | ||
a335debb PX |
4605 | ret = 0; |
4606 | out: | |
bf269906 | 4607 | g_free(le_bitmap); |
a335debb PX |
4608 | return ret; |
4609 | } | |
4610 | ||
edd090c7 PX |
4611 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
4612 | { | |
4613 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 4614 | int ret; |
edd090c7 | 4615 | |
08614f34 PX |
4616 | ret = ram_dirty_bitmap_sync_all(s, rs); |
4617 | if (ret) { | |
4618 | return ret; | |
4619 | } | |
4620 | ||
4621 | ram_state_resume_prepare(rs, s->to_dst_file); | |
4622 | ||
4623 | return 0; | |
edd090c7 PX |
4624 | } |
4625 | ||
36f62f11 PX |
4626 | void postcopy_preempt_shutdown_file(MigrationState *s) |
4627 | { | |
4628 | qemu_put_be64(s->postcopy_qemufile_src, RAM_SAVE_FLAG_EOS); | |
4629 | qemu_fflush(s->postcopy_qemufile_src); | |
4630 | } | |
4631 | ||
56e93d26 | 4632 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 4633 | .save_setup = ram_save_setup, |
56e93d26 | 4634 | .save_live_iterate = ram_save_iterate, |
763c906b | 4635 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 4636 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 4637 | .has_postcopy = ram_has_postcopy, |
c8df4a7a JQ |
4638 | .state_pending_exact = ram_state_pending_exact, |
4639 | .state_pending_estimate = ram_state_pending_estimate, | |
56e93d26 | 4640 | .load_state = ram_load, |
f265e0e4 JQ |
4641 | .save_cleanup = ram_save_cleanup, |
4642 | .load_setup = ram_load_setup, | |
4643 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 4644 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
4645 | }; |
4646 | ||
c7c0e724 DH |
4647 | static void ram_mig_ram_block_resized(RAMBlockNotifier *n, void *host, |
4648 | size_t old_size, size_t new_size) | |
4649 | { | |
cc61c703 | 4650 | PostcopyState ps = postcopy_state_get(); |
c7c0e724 DH |
4651 | ram_addr_t offset; |
4652 | RAMBlock *rb = qemu_ram_block_from_host(host, false, &offset); | |
4653 | Error *err = NULL; | |
4654 | ||
4655 | if (ramblock_is_ignored(rb)) { | |
4656 | return; | |
4657 | } | |
4658 | ||
4659 | if (!migration_is_idle()) { | |
4660 | /* | |
4661 | * Precopy code on the source cannot deal with the size of RAM blocks | |
4662 | * changing at random points in time - especially after sending the | |
4663 | * RAM block sizes in the migration stream, they must no longer change. | |
4664 | * Abort and indicate a proper reason. | |
4665 | */ | |
4666 | error_setg(&err, "RAM block '%s' resized during precopy.", rb->idstr); | |
458fecca | 4667 | migration_cancel(err); |
c7c0e724 | 4668 | error_free(err); |
c7c0e724 | 4669 | } |
cc61c703 DH |
4670 | |
4671 | switch (ps) { | |
4672 | case POSTCOPY_INCOMING_ADVISE: | |
4673 | /* | |
4674 | * Update what ram_postcopy_incoming_init()->init_range() does at the | |
4675 | * time postcopy was advised. Syncing RAM blocks with the source will | |
4676 | * result in RAM resizes. | |
4677 | */ | |
4678 | if (old_size < new_size) { | |
4679 | if (ram_discard_range(rb->idstr, old_size, new_size - old_size)) { | |
4680 | error_report("RAM block '%s' discard of resized RAM failed", | |
4681 | rb->idstr); | |
4682 | } | |
4683 | } | |
898ba906 | 4684 | rb->postcopy_length = new_size; |
cc61c703 DH |
4685 | break; |
4686 | case POSTCOPY_INCOMING_NONE: | |
4687 | case POSTCOPY_INCOMING_RUNNING: | |
4688 | case POSTCOPY_INCOMING_END: | |
4689 | /* | |
4690 | * Once our guest is running, postcopy does no longer care about | |
4691 | * resizes. When growing, the new memory was not available on the | |
4692 | * source, no handler needed. | |
4693 | */ | |
4694 | break; | |
4695 | default: | |
4696 | error_report("RAM block '%s' resized during postcopy state: %d", | |
4697 | rb->idstr, ps); | |
4698 | exit(-1); | |
4699 | } | |
c7c0e724 DH |
4700 | } |
4701 | ||
4702 | static RAMBlockNotifier ram_mig_ram_notifier = { | |
4703 | .ram_block_resized = ram_mig_ram_block_resized, | |
4704 | }; | |
4705 | ||
56e93d26 JQ |
4706 | void ram_mig_init(void) |
4707 | { | |
4708 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 4709 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
c7c0e724 | 4710 | ram_block_notifier_add(&ram_mig_ram_notifier); |
56e93d26 | 4711 | } |