]>
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" |
709e3fe8 | 35 | #include "xbzrle.h" |
7b1e1a22 | 36 | #include "ram.h" |
6666c96a | 37 | #include "migration.h" |
f2a8f0a6 | 38 | #include "migration/register.h" |
7b1e1a22 | 39 | #include "migration/misc.h" |
08a0aee1 | 40 | #include "qemu-file.h" |
be07b0ac | 41 | #include "postcopy-ram.h" |
53d37d36 | 42 | #include "page_cache.h" |
56e93d26 | 43 | #include "qemu/error-report.h" |
e688df6b | 44 | #include "qapi/error.h" |
ab7cbb0b | 45 | #include "qapi/qapi-types-migration.h" |
9af23989 | 46 | #include "qapi/qapi-events-migration.h" |
8acabf69 | 47 | #include "qapi/qmp/qerror.h" |
56e93d26 | 48 | #include "trace.h" |
56e93d26 | 49 | #include "exec/ram_addr.h" |
f9494614 | 50 | #include "exec/target_page.h" |
56e93d26 | 51 | #include "qemu/rcu_queue.h" |
a91246c9 | 52 | #include "migration/colo.h" |
53d37d36 | 53 | #include "block.h" |
b0c3cf94 | 54 | #include "sysemu/cpu-throttle.h" |
edd090c7 | 55 | #include "savevm.h" |
b9ee2f7d | 56 | #include "qemu/iov.h" |
d32ca5ad | 57 | #include "multifd.h" |
278e2f55 AG |
58 | #include "sysemu/runstate.h" |
59 | ||
e5fdf920 LS |
60 | #include "hw/boards.h" /* for machine_dump_guest_core() */ |
61 | ||
278e2f55 AG |
62 | #if defined(__linux__) |
63 | #include "qemu/userfaultfd.h" | |
64 | #endif /* defined(__linux__) */ | |
56e93d26 | 65 | |
56e93d26 JQ |
66 | /***********************************************************/ |
67 | /* ram save/restore */ | |
68 | ||
bb890ed5 JQ |
69 | /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it |
70 | * worked for pages that where filled with the same char. We switched | |
71 | * it to only search for the zero value. And to avoid confusion with | |
72 | * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it. | |
73 | */ | |
74 | ||
56e93d26 | 75 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ |
bb890ed5 | 76 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
77 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
78 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
79 | #define RAM_SAVE_FLAG_EOS 0x10 | |
80 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
81 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
82 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
83 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
84 | ||
9360447d JQ |
85 | XBZRLECacheStats xbzrle_counters; |
86 | ||
56e93d26 JQ |
87 | /* struct contains XBZRLE cache and a static page |
88 | used by the compression */ | |
89 | static struct { | |
90 | /* buffer used for XBZRLE encoding */ | |
91 | uint8_t *encoded_buf; | |
92 | /* buffer for storing page content */ | |
93 | uint8_t *current_buf; | |
94 | /* Cache for XBZRLE, Protected by lock. */ | |
95 | PageCache *cache; | |
96 | QemuMutex lock; | |
c00e0928 JQ |
97 | /* it will store a page full of zeros */ |
98 | uint8_t *zero_target_page; | |
f265e0e4 JQ |
99 | /* buffer used for XBZRLE decoding */ |
100 | uint8_t *decoded_buf; | |
56e93d26 JQ |
101 | } XBZRLE; |
102 | ||
56e93d26 JQ |
103 | static void XBZRLE_cache_lock(void) |
104 | { | |
f4c51a6b | 105 | if (migrate_use_xbzrle()) { |
56e93d26 | 106 | qemu_mutex_lock(&XBZRLE.lock); |
f4c51a6b | 107 | } |
56e93d26 JQ |
108 | } |
109 | ||
110 | static void XBZRLE_cache_unlock(void) | |
111 | { | |
f4c51a6b | 112 | if (migrate_use_xbzrle()) { |
56e93d26 | 113 | qemu_mutex_unlock(&XBZRLE.lock); |
f4c51a6b | 114 | } |
56e93d26 JQ |
115 | } |
116 | ||
3d0684b2 JQ |
117 | /** |
118 | * xbzrle_cache_resize: resize the xbzrle cache | |
119 | * | |
cbde7be9 | 120 | * This function is called from migrate_params_apply in main |
3d0684b2 JQ |
121 | * thread, possibly while a migration is in progress. A running |
122 | * migration may be using the cache and might finish during this call, | |
123 | * hence changes to the cache are protected by XBZRLE.lock(). | |
124 | * | |
c9dede2d | 125 | * Returns 0 for success or -1 for error |
3d0684b2 JQ |
126 | * |
127 | * @new_size: new cache size | |
8acabf69 | 128 | * @errp: set *errp if the check failed, with reason |
56e93d26 | 129 | */ |
8b9407a0 | 130 | int xbzrle_cache_resize(uint64_t new_size, Error **errp) |
56e93d26 JQ |
131 | { |
132 | PageCache *new_cache; | |
c9dede2d | 133 | int64_t ret = 0; |
56e93d26 | 134 | |
8acabf69 JQ |
135 | /* Check for truncation */ |
136 | if (new_size != (size_t)new_size) { | |
137 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
138 | "exceeding address space"); | |
139 | return -1; | |
140 | } | |
141 | ||
2a313e5c JQ |
142 | if (new_size == migrate_xbzrle_cache_size()) { |
143 | /* nothing to do */ | |
c9dede2d | 144 | return 0; |
2a313e5c JQ |
145 | } |
146 | ||
56e93d26 JQ |
147 | XBZRLE_cache_lock(); |
148 | ||
149 | if (XBZRLE.cache != NULL) { | |
80f8dfde | 150 | new_cache = cache_init(new_size, TARGET_PAGE_SIZE, errp); |
56e93d26 | 151 | if (!new_cache) { |
56e93d26 JQ |
152 | ret = -1; |
153 | goto out; | |
154 | } | |
155 | ||
156 | cache_fini(XBZRLE.cache); | |
157 | XBZRLE.cache = new_cache; | |
158 | } | |
56e93d26 JQ |
159 | out: |
160 | XBZRLE_cache_unlock(); | |
161 | return ret; | |
162 | } | |
163 | ||
3ded54b1 | 164 | bool ramblock_is_ignored(RAMBlock *block) |
fbd162e6 YK |
165 | { |
166 | return !qemu_ram_is_migratable(block) || | |
167 | (migrate_ignore_shared() && qemu_ram_is_shared(block)); | |
168 | } | |
169 | ||
343f632c DDAG |
170 | #undef RAMBLOCK_FOREACH |
171 | ||
fbd162e6 YK |
172 | int foreach_not_ignored_block(RAMBlockIterFunc func, void *opaque) |
173 | { | |
174 | RAMBlock *block; | |
175 | int ret = 0; | |
176 | ||
89ac5a1d DDAG |
177 | RCU_READ_LOCK_GUARD(); |
178 | ||
fbd162e6 YK |
179 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
180 | ret = func(block, opaque); | |
181 | if (ret) { | |
182 | break; | |
183 | } | |
184 | } | |
fbd162e6 YK |
185 | return ret; |
186 | } | |
187 | ||
f9494614 AP |
188 | static void ramblock_recv_map_init(void) |
189 | { | |
190 | RAMBlock *rb; | |
191 | ||
fbd162e6 | 192 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
193 | assert(!rb->receivedmap); |
194 | rb->receivedmap = bitmap_new(rb->max_length >> qemu_target_page_bits()); | |
195 | } | |
196 | } | |
197 | ||
198 | int ramblock_recv_bitmap_test(RAMBlock *rb, void *host_addr) | |
199 | { | |
200 | return test_bit(ramblock_recv_bitmap_offset(host_addr, rb), | |
201 | rb->receivedmap); | |
202 | } | |
203 | ||
1cba9f6e DDAG |
204 | bool ramblock_recv_bitmap_test_byte_offset(RAMBlock *rb, uint64_t byte_offset) |
205 | { | |
206 | return test_bit(byte_offset >> TARGET_PAGE_BITS, rb->receivedmap); | |
207 | } | |
208 | ||
f9494614 AP |
209 | void ramblock_recv_bitmap_set(RAMBlock *rb, void *host_addr) |
210 | { | |
211 | set_bit_atomic(ramblock_recv_bitmap_offset(host_addr, rb), rb->receivedmap); | |
212 | } | |
213 | ||
214 | void ramblock_recv_bitmap_set_range(RAMBlock *rb, void *host_addr, | |
215 | size_t nr) | |
216 | { | |
217 | bitmap_set_atomic(rb->receivedmap, | |
218 | ramblock_recv_bitmap_offset(host_addr, rb), | |
219 | nr); | |
220 | } | |
221 | ||
a335debb PX |
222 | #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL) |
223 | ||
224 | /* | |
225 | * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes). | |
226 | * | |
227 | * Returns >0 if success with sent bytes, or <0 if error. | |
228 | */ | |
229 | int64_t ramblock_recv_bitmap_send(QEMUFile *file, | |
230 | const char *block_name) | |
231 | { | |
232 | RAMBlock *block = qemu_ram_block_by_name(block_name); | |
233 | unsigned long *le_bitmap, nbits; | |
234 | uint64_t size; | |
235 | ||
236 | if (!block) { | |
237 | error_report("%s: invalid block name: %s", __func__, block_name); | |
238 | return -1; | |
239 | } | |
240 | ||
898ba906 | 241 | nbits = block->postcopy_length >> TARGET_PAGE_BITS; |
a335debb PX |
242 | |
243 | /* | |
244 | * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit | |
245 | * machines we may need 4 more bytes for padding (see below | |
246 | * comment). So extend it a bit before hand. | |
247 | */ | |
248 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
249 | ||
250 | /* | |
251 | * Always use little endian when sending the bitmap. This is | |
252 | * required that when source and destination VMs are not using the | |
3a4452d8 | 253 | * same endianness. (Note: big endian won't work.) |
a335debb PX |
254 | */ |
255 | bitmap_to_le(le_bitmap, block->receivedmap, nbits); | |
256 | ||
257 | /* Size of the bitmap, in bytes */ | |
a725ef9f | 258 | size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
259 | |
260 | /* | |
261 | * size is always aligned to 8 bytes for 64bit machines, but it | |
262 | * may not be true for 32bit machines. We need this padding to | |
263 | * make sure the migration can survive even between 32bit and | |
264 | * 64bit machines. | |
265 | */ | |
266 | size = ROUND_UP(size, 8); | |
267 | ||
268 | qemu_put_be64(file, size); | |
269 | qemu_put_buffer(file, (const uint8_t *)le_bitmap, size); | |
270 | /* | |
271 | * Mark as an end, in case the middle part is screwed up due to | |
3a4452d8 | 272 | * some "mysterious" reason. |
a335debb PX |
273 | */ |
274 | qemu_put_be64(file, RAMBLOCK_RECV_BITMAP_ENDING); | |
275 | qemu_fflush(file); | |
276 | ||
bf269906 | 277 | g_free(le_bitmap); |
a335debb PX |
278 | |
279 | if (qemu_file_get_error(file)) { | |
280 | return qemu_file_get_error(file); | |
281 | } | |
282 | ||
283 | return size + sizeof(size); | |
284 | } | |
285 | ||
ec481c6c JQ |
286 | /* |
287 | * An outstanding page request, on the source, having been received | |
288 | * and queued | |
289 | */ | |
290 | struct RAMSrcPageRequest { | |
291 | RAMBlock *rb; | |
292 | hwaddr offset; | |
293 | hwaddr len; | |
294 | ||
295 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
296 | }; | |
297 | ||
6f37bb8b JQ |
298 | /* State of RAM for migration */ |
299 | struct RAMState { | |
204b88b8 JQ |
300 | /* QEMUFile used for this migration */ |
301 | QEMUFile *f; | |
278e2f55 AG |
302 | /* UFFD file descriptor, used in 'write-tracking' migration */ |
303 | int uffdio_fd; | |
6f37bb8b JQ |
304 | /* Last block that we have visited searching for dirty pages */ |
305 | RAMBlock *last_seen_block; | |
306 | /* Last block from where we have sent data */ | |
307 | RAMBlock *last_sent_block; | |
269ace29 JQ |
308 | /* Last dirty target page we have sent */ |
309 | ram_addr_t last_page; | |
6f37bb8b JQ |
310 | /* last ram version we have seen */ |
311 | uint32_t last_version; | |
8d820d6f JQ |
312 | /* How many times we have dirty too many pages */ |
313 | int dirty_rate_high_cnt; | |
f664da80 JQ |
314 | /* these variables are used for bitmap sync */ |
315 | /* last time we did a full bitmap_sync */ | |
316 | int64_t time_last_bitmap_sync; | |
eac74159 | 317 | /* bytes transferred at start_time */ |
c4bdf0cf | 318 | uint64_t bytes_xfer_prev; |
a66cd90c | 319 | /* number of dirty pages since start_time */ |
68908ed6 | 320 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
321 | /* xbzrle misses since the beginning of the period */ |
322 | uint64_t xbzrle_cache_miss_prev; | |
e460a4b1 WW |
323 | /* Amount of xbzrle pages since the beginning of the period */ |
324 | uint64_t xbzrle_pages_prev; | |
325 | /* Amount of xbzrle encoded bytes since the beginning of the period */ | |
326 | uint64_t xbzrle_bytes_prev; | |
1a373522 DH |
327 | /* Start using XBZRLE (e.g., after the first round). */ |
328 | bool xbzrle_enabled; | |
05931ec5 JQ |
329 | /* Are we on the last stage of migration */ |
330 | bool last_stage; | |
76e03000 XG |
331 | /* compression statistics since the beginning of the period */ |
332 | /* amount of count that no free thread to compress data */ | |
333 | uint64_t compress_thread_busy_prev; | |
334 | /* amount bytes after compression */ | |
335 | uint64_t compressed_size_prev; | |
336 | /* amount of compressed pages */ | |
337 | uint64_t compress_pages_prev; | |
338 | ||
be8b02ed XG |
339 | /* total handled target pages at the beginning of period */ |
340 | uint64_t target_page_count_prev; | |
341 | /* total handled target pages since start */ | |
342 | uint64_t target_page_count; | |
9360447d | 343 | /* number of dirty bits in the bitmap */ |
2dfaf12e | 344 | uint64_t migration_dirty_pages; |
386a907b | 345 | /* Protects modification of the bitmap and migration dirty pages */ |
108cfae0 | 346 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
347 | /* The RAMBlock used in the last src_page_requests */ |
348 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
349 | /* Queue of outstanding page requests from the destination */ |
350 | QemuMutex src_page_req_mutex; | |
b58deb34 | 351 | QSIMPLEQ_HEAD(, RAMSrcPageRequest) src_page_requests; |
6f37bb8b JQ |
352 | }; |
353 | typedef struct RAMState RAMState; | |
354 | ||
53518d94 | 355 | static RAMState *ram_state; |
6f37bb8b | 356 | |
bd227060 WW |
357 | static NotifierWithReturnList precopy_notifier_list; |
358 | ||
a1fe28df PX |
359 | /* Whether postcopy has queued requests? */ |
360 | static bool postcopy_has_request(RAMState *rs) | |
361 | { | |
362 | return !QSIMPLEQ_EMPTY_ATOMIC(&rs->src_page_requests); | |
363 | } | |
364 | ||
bd227060 WW |
365 | void precopy_infrastructure_init(void) |
366 | { | |
367 | notifier_with_return_list_init(&precopy_notifier_list); | |
368 | } | |
369 | ||
370 | void precopy_add_notifier(NotifierWithReturn *n) | |
371 | { | |
372 | notifier_with_return_list_add(&precopy_notifier_list, n); | |
373 | } | |
374 | ||
375 | void precopy_remove_notifier(NotifierWithReturn *n) | |
376 | { | |
377 | notifier_with_return_remove(n); | |
378 | } | |
379 | ||
380 | int precopy_notify(PrecopyNotifyReason reason, Error **errp) | |
381 | { | |
382 | PrecopyNotifyData pnd; | |
383 | pnd.reason = reason; | |
384 | pnd.errp = errp; | |
385 | ||
386 | return notifier_with_return_list_notify(&precopy_notifier_list, &pnd); | |
387 | } | |
388 | ||
9edabd4d | 389 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 390 | { |
bae416e5 DDAG |
391 | return ram_state ? (ram_state->migration_dirty_pages * TARGET_PAGE_SIZE) : |
392 | 0; | |
2f4fde93 JQ |
393 | } |
394 | ||
9360447d | 395 | MigrationStats ram_counters; |
96506894 | 396 | |
4c2d0f6d DE |
397 | static void ram_transferred_add(uint64_t bytes) |
398 | { | |
ae680668 DE |
399 | if (runstate_is_running()) { |
400 | ram_counters.precopy_bytes += bytes; | |
401 | } else if (migration_in_postcopy()) { | |
402 | ram_counters.postcopy_bytes += bytes; | |
403 | } else { | |
404 | ram_counters.downtime_bytes += bytes; | |
405 | } | |
4c2d0f6d DE |
406 | ram_counters.transferred += bytes; |
407 | } | |
408 | ||
b8fb8cb7 DDAG |
409 | /* used by the search for pages to send */ |
410 | struct PageSearchStatus { | |
411 | /* Current block being searched */ | |
412 | RAMBlock *block; | |
a935e30f JQ |
413 | /* Current page to search from */ |
414 | unsigned long page; | |
b8fb8cb7 DDAG |
415 | /* Set once we wrap around */ |
416 | bool complete_round; | |
ea2faf0c PX |
417 | /* Whether current page is explicitly requested by postcopy */ |
418 | bool postcopy_requested; | |
b8fb8cb7 DDAG |
419 | }; |
420 | typedef struct PageSearchStatus PageSearchStatus; | |
421 | ||
76e03000 XG |
422 | CompressionStats compression_counters; |
423 | ||
56e93d26 | 424 | struct CompressParam { |
56e93d26 | 425 | bool done; |
90e56fb4 | 426 | bool quit; |
5e5fdcff | 427 | bool zero_page; |
56e93d26 JQ |
428 | QEMUFile *file; |
429 | QemuMutex mutex; | |
430 | QemuCond cond; | |
431 | RAMBlock *block; | |
432 | ram_addr_t offset; | |
34ab9e97 XG |
433 | |
434 | /* internally used fields */ | |
dcaf446e | 435 | z_stream stream; |
34ab9e97 | 436 | uint8_t *originbuf; |
56e93d26 JQ |
437 | }; |
438 | typedef struct CompressParam CompressParam; | |
439 | ||
440 | struct DecompressParam { | |
73a8912b | 441 | bool done; |
90e56fb4 | 442 | bool quit; |
56e93d26 JQ |
443 | QemuMutex mutex; |
444 | QemuCond cond; | |
445 | void *des; | |
d341d9f3 | 446 | uint8_t *compbuf; |
56e93d26 | 447 | int len; |
797ca154 | 448 | z_stream stream; |
56e93d26 JQ |
449 | }; |
450 | typedef struct DecompressParam DecompressParam; | |
451 | ||
452 | static CompressParam *comp_param; | |
453 | static QemuThread *compress_threads; | |
454 | /* comp_done_cond is used to wake up the migration thread when | |
455 | * one of the compression threads has finished the compression. | |
456 | * comp_done_lock is used to co-work with comp_done_cond. | |
457 | */ | |
0d9f9a5c LL |
458 | static QemuMutex comp_done_lock; |
459 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
460 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
461 | static const QEMUFileOps empty_ops = { }; | |
462 | ||
34ab9e97 | 463 | static QEMUFile *decomp_file; |
56e93d26 JQ |
464 | static DecompressParam *decomp_param; |
465 | static QemuThread *decompress_threads; | |
73a8912b LL |
466 | static QemuMutex decomp_done_lock; |
467 | static QemuCond decomp_done_cond; | |
56e93d26 | 468 | |
5e5fdcff | 469 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 470 | ram_addr_t offset, uint8_t *source_buf); |
56e93d26 JQ |
471 | |
472 | static void *do_data_compress(void *opaque) | |
473 | { | |
474 | CompressParam *param = opaque; | |
a7a9a88f LL |
475 | RAMBlock *block; |
476 | ram_addr_t offset; | |
5e5fdcff | 477 | bool zero_page; |
56e93d26 | 478 | |
a7a9a88f | 479 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 480 | while (!param->quit) { |
a7a9a88f LL |
481 | if (param->block) { |
482 | block = param->block; | |
483 | offset = param->offset; | |
484 | param->block = NULL; | |
485 | qemu_mutex_unlock(¶m->mutex); | |
486 | ||
5e5fdcff XG |
487 | zero_page = do_compress_ram_page(param->file, ¶m->stream, |
488 | block, offset, param->originbuf); | |
a7a9a88f | 489 | |
0d9f9a5c | 490 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 491 | param->done = true; |
5e5fdcff | 492 | param->zero_page = zero_page; |
0d9f9a5c LL |
493 | qemu_cond_signal(&comp_done_cond); |
494 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
495 | |
496 | qemu_mutex_lock(¶m->mutex); | |
497 | } else { | |
56e93d26 JQ |
498 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
499 | } | |
56e93d26 | 500 | } |
a7a9a88f | 501 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
502 | |
503 | return NULL; | |
504 | } | |
505 | ||
f0afa331 | 506 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
507 | { |
508 | int i, thread_count; | |
509 | ||
05306935 | 510 | if (!migrate_use_compression() || !comp_param) { |
56e93d26 JQ |
511 | return; |
512 | } | |
05306935 | 513 | |
56e93d26 JQ |
514 | thread_count = migrate_compress_threads(); |
515 | for (i = 0; i < thread_count; i++) { | |
dcaf446e XG |
516 | /* |
517 | * we use it as a indicator which shows if the thread is | |
518 | * properly init'd or not | |
519 | */ | |
520 | if (!comp_param[i].file) { | |
521 | break; | |
522 | } | |
05306935 FL |
523 | |
524 | qemu_mutex_lock(&comp_param[i].mutex); | |
525 | comp_param[i].quit = true; | |
526 | qemu_cond_signal(&comp_param[i].cond); | |
527 | qemu_mutex_unlock(&comp_param[i].mutex); | |
528 | ||
56e93d26 | 529 | qemu_thread_join(compress_threads + i); |
56e93d26 JQ |
530 | qemu_mutex_destroy(&comp_param[i].mutex); |
531 | qemu_cond_destroy(&comp_param[i].cond); | |
dcaf446e | 532 | deflateEnd(&comp_param[i].stream); |
34ab9e97 | 533 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
534 | qemu_fclose(comp_param[i].file); |
535 | comp_param[i].file = NULL; | |
56e93d26 | 536 | } |
0d9f9a5c LL |
537 | qemu_mutex_destroy(&comp_done_lock); |
538 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
539 | g_free(compress_threads); |
540 | g_free(comp_param); | |
56e93d26 JQ |
541 | compress_threads = NULL; |
542 | comp_param = NULL; | |
56e93d26 JQ |
543 | } |
544 | ||
dcaf446e | 545 | static int compress_threads_save_setup(void) |
56e93d26 JQ |
546 | { |
547 | int i, thread_count; | |
548 | ||
549 | if (!migrate_use_compression()) { | |
dcaf446e | 550 | return 0; |
56e93d26 | 551 | } |
56e93d26 JQ |
552 | thread_count = migrate_compress_threads(); |
553 | compress_threads = g_new0(QemuThread, thread_count); | |
554 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
555 | qemu_cond_init(&comp_done_cond); |
556 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 557 | for (i = 0; i < thread_count; i++) { |
34ab9e97 XG |
558 | comp_param[i].originbuf = g_try_malloc(TARGET_PAGE_SIZE); |
559 | if (!comp_param[i].originbuf) { | |
560 | goto exit; | |
561 | } | |
562 | ||
dcaf446e XG |
563 | if (deflateInit(&comp_param[i].stream, |
564 | migrate_compress_level()) != Z_OK) { | |
34ab9e97 | 565 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
566 | goto exit; |
567 | } | |
568 | ||
e110aa91 C |
569 | /* comp_param[i].file is just used as a dummy buffer to save data, |
570 | * set its ops to empty. | |
56e93d26 | 571 | */ |
c6ad5be7 | 572 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops, false); |
56e93d26 | 573 | comp_param[i].done = true; |
90e56fb4 | 574 | comp_param[i].quit = false; |
56e93d26 JQ |
575 | qemu_mutex_init(&comp_param[i].mutex); |
576 | qemu_cond_init(&comp_param[i].cond); | |
577 | qemu_thread_create(compress_threads + i, "compress", | |
578 | do_data_compress, comp_param + i, | |
579 | QEMU_THREAD_JOINABLE); | |
580 | } | |
dcaf446e XG |
581 | return 0; |
582 | ||
583 | exit: | |
584 | compress_threads_save_cleanup(); | |
585 | return -1; | |
56e93d26 JQ |
586 | } |
587 | ||
588 | /** | |
3d0684b2 | 589 | * save_page_header: write page header to wire |
56e93d26 JQ |
590 | * |
591 | * If this is the 1st block, it also writes the block identification | |
592 | * | |
3d0684b2 | 593 | * Returns the number of bytes written |
56e93d26 JQ |
594 | * |
595 | * @f: QEMUFile where to send the data | |
596 | * @block: block that contains the page we want to send | |
597 | * @offset: offset inside the block for the page | |
598 | * in the lower bits, it contains flags | |
599 | */ | |
2bf3aa85 JQ |
600 | static size_t save_page_header(RAMState *rs, QEMUFile *f, RAMBlock *block, |
601 | ram_addr_t offset) | |
56e93d26 | 602 | { |
9f5f380b | 603 | size_t size, len; |
56e93d26 | 604 | |
24795694 JQ |
605 | if (block == rs->last_sent_block) { |
606 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
607 | } | |
2bf3aa85 | 608 | qemu_put_be64(f, offset); |
56e93d26 JQ |
609 | size = 8; |
610 | ||
611 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 612 | len = strlen(block->idstr); |
2bf3aa85 JQ |
613 | qemu_put_byte(f, len); |
614 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 615 | size += 1 + len; |
24795694 | 616 | rs->last_sent_block = block; |
56e93d26 JQ |
617 | } |
618 | return size; | |
619 | } | |
620 | ||
3d0684b2 | 621 | /** |
179a8080 | 622 | * mig_throttle_guest_down: throttle down the guest |
3d0684b2 JQ |
623 | * |
624 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
625 | * writes. If guest dirty memory rate is reduced below the rate at | |
626 | * which we can transfer pages to the destination then we should be | |
627 | * able to complete migration. Some workloads dirty memory way too | |
628 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 | 629 | */ |
cbbf8182 KZ |
630 | static void mig_throttle_guest_down(uint64_t bytes_dirty_period, |
631 | uint64_t bytes_dirty_threshold) | |
070afca2 JH |
632 | { |
633 | MigrationState *s = migrate_get_current(); | |
2594f56d | 634 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
cbbf8182 KZ |
635 | uint64_t pct_increment = s->parameters.cpu_throttle_increment; |
636 | bool pct_tailslow = s->parameters.cpu_throttle_tailslow; | |
4cbc9c7f | 637 | int pct_max = s->parameters.max_cpu_throttle; |
070afca2 | 638 | |
cbbf8182 KZ |
639 | uint64_t throttle_now = cpu_throttle_get_percentage(); |
640 | uint64_t cpu_now, cpu_ideal, throttle_inc; | |
641 | ||
070afca2 JH |
642 | /* We have not started throttling yet. Let's start it. */ |
643 | if (!cpu_throttle_active()) { | |
644 | cpu_throttle_set(pct_initial); | |
645 | } else { | |
646 | /* Throttling already on, just increase the rate */ | |
cbbf8182 KZ |
647 | if (!pct_tailslow) { |
648 | throttle_inc = pct_increment; | |
649 | } else { | |
650 | /* Compute the ideal CPU percentage used by Guest, which may | |
651 | * make the dirty rate match the dirty rate threshold. */ | |
652 | cpu_now = 100 - throttle_now; | |
653 | cpu_ideal = cpu_now * (bytes_dirty_threshold * 1.0 / | |
654 | bytes_dirty_period); | |
655 | throttle_inc = MIN(cpu_now - cpu_ideal, pct_increment); | |
656 | } | |
657 | cpu_throttle_set(MIN(throttle_now + throttle_inc, pct_max)); | |
070afca2 JH |
658 | } |
659 | } | |
660 | ||
91fe9a8d RL |
661 | void mig_throttle_counter_reset(void) |
662 | { | |
663 | RAMState *rs = ram_state; | |
664 | ||
665 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
666 | rs->num_dirty_pages_period = 0; | |
667 | rs->bytes_xfer_prev = ram_counters.transferred; | |
668 | } | |
669 | ||
3d0684b2 JQ |
670 | /** |
671 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
672 | * | |
6f37bb8b | 673 | * @rs: current RAM state |
3d0684b2 JQ |
674 | * @current_addr: address for the zero page |
675 | * | |
676 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
677 | * The important thing is that a stale (not-yet-0'd) page be replaced |
678 | * by the new data. | |
679 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 680 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 681 | */ |
6f37bb8b | 682 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 683 | { |
1a373522 | 684 | if (!rs->xbzrle_enabled) { |
56e93d26 JQ |
685 | return; |
686 | } | |
687 | ||
688 | /* We don't care if this fails to allocate a new cache page | |
689 | * as long as it updated an old one */ | |
c00e0928 | 690 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 691 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
692 | } |
693 | ||
694 | #define ENCODING_FLAG_XBZRLE 0x1 | |
695 | ||
696 | /** | |
697 | * save_xbzrle_page: compress and send current page | |
698 | * | |
699 | * Returns: 1 means that we wrote the page | |
700 | * 0 means that page is identical to the one already sent | |
701 | * -1 means that xbzrle would be longer than normal | |
702 | * | |
5a987738 | 703 | * @rs: current RAM state |
3d0684b2 JQ |
704 | * @current_data: pointer to the address of the page contents |
705 | * @current_addr: addr of the page | |
56e93d26 JQ |
706 | * @block: block that contains the page we want to send |
707 | * @offset: offset inside the block for the page | |
56e93d26 | 708 | */ |
204b88b8 | 709 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 710 | ram_addr_t current_addr, RAMBlock *block, |
05931ec5 | 711 | ram_addr_t offset) |
56e93d26 JQ |
712 | { |
713 | int encoded_len = 0, bytes_xbzrle; | |
714 | uint8_t *prev_cached_page; | |
715 | ||
9360447d JQ |
716 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
717 | ram_counters.dirty_sync_count)) { | |
718 | xbzrle_counters.cache_miss++; | |
05931ec5 | 719 | if (!rs->last_stage) { |
56e93d26 | 720 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, |
9360447d | 721 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
722 | return -1; |
723 | } else { | |
724 | /* update *current_data when the page has been | |
725 | inserted into cache */ | |
726 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
727 | } | |
728 | } | |
729 | return -1; | |
730 | } | |
731 | ||
e460a4b1 WW |
732 | /* |
733 | * Reaching here means the page has hit the xbzrle cache, no matter what | |
734 | * encoding result it is (normal encoding, overflow or skipping the page), | |
3a4452d8 | 735 | * count the page as encoded. This is used to calculate the encoding rate. |
e460a4b1 WW |
736 | * |
737 | * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB, | |
738 | * 2nd page turns out to be skipped (i.e. no new bytes written to the | |
739 | * page), the overall encoding rate will be 8KB / 2KB = 4, which has the | |
740 | * skipped page included. In this way, the encoding rate can tell if the | |
741 | * guest page is good for xbzrle encoding. | |
742 | */ | |
743 | xbzrle_counters.pages++; | |
56e93d26 JQ |
744 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); |
745 | ||
746 | /* save current buffer into memory */ | |
747 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
748 | ||
749 | /* XBZRLE encoding (if there is no overflow) */ | |
750 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
751 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
752 | TARGET_PAGE_SIZE); | |
ca353803 WY |
753 | |
754 | /* | |
755 | * Update the cache contents, so that it corresponds to the data | |
756 | * sent, in all cases except where we skip the page. | |
757 | */ | |
05931ec5 | 758 | if (!rs->last_stage && encoded_len != 0) { |
ca353803 WY |
759 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); |
760 | /* | |
761 | * In the case where we couldn't compress, ensure that the caller | |
762 | * sends the data from the cache, since the guest might have | |
763 | * changed the RAM since we copied it. | |
764 | */ | |
765 | *current_data = prev_cached_page; | |
766 | } | |
767 | ||
56e93d26 | 768 | if (encoded_len == 0) { |
55c4446b | 769 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
770 | return 0; |
771 | } else if (encoded_len == -1) { | |
55c4446b | 772 | trace_save_xbzrle_page_overflow(); |
9360447d | 773 | xbzrle_counters.overflow++; |
e460a4b1 | 774 | xbzrle_counters.bytes += TARGET_PAGE_SIZE; |
56e93d26 JQ |
775 | return -1; |
776 | } | |
777 | ||
56e93d26 | 778 | /* Send XBZRLE based compressed page */ |
2bf3aa85 | 779 | bytes_xbzrle = save_page_header(rs, rs->f, block, |
204b88b8 JQ |
780 | offset | RAM_SAVE_FLAG_XBZRLE); |
781 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
782 | qemu_put_be16(rs->f, encoded_len); | |
783 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 784 | bytes_xbzrle += encoded_len + 1 + 2; |
e460a4b1 WW |
785 | /* |
786 | * Like compressed_size (please see update_compress_thread_counts), | |
787 | * the xbzrle encoded bytes don't count the 8 byte header with | |
788 | * RAM_SAVE_FLAG_CONTINUE. | |
789 | */ | |
790 | xbzrle_counters.bytes += bytes_xbzrle - 8; | |
4c2d0f6d | 791 | ram_transferred_add(bytes_xbzrle); |
56e93d26 JQ |
792 | |
793 | return 1; | |
794 | } | |
795 | ||
3d0684b2 JQ |
796 | /** |
797 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 798 | * |
a5f7b1a6 | 799 | * Returns the page offset within memory region of the start of a dirty page |
3d0684b2 | 800 | * |
6f37bb8b | 801 | * @rs: current RAM state |
3d0684b2 | 802 | * @rb: RAMBlock where to search for dirty pages |
a935e30f | 803 | * @start: page where we start the search |
f3f491fc | 804 | */ |
56e93d26 | 805 | static inline |
a935e30f | 806 | unsigned long migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
f20e2865 | 807 | unsigned long start) |
56e93d26 | 808 | { |
6b6712ef JQ |
809 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
810 | unsigned long *bitmap = rb->bmap; | |
56e93d26 | 811 | |
fbd162e6 | 812 | if (ramblock_is_ignored(rb)) { |
b895de50 CLG |
813 | return size; |
814 | } | |
815 | ||
1a373522 | 816 | return find_next_bit(bitmap, size, start); |
56e93d26 JQ |
817 | } |
818 | ||
1230a25f | 819 | static void migration_clear_memory_region_dirty_bitmap(RAMBlock *rb, |
3143577d WW |
820 | unsigned long page) |
821 | { | |
822 | uint8_t shift; | |
823 | hwaddr size, start; | |
824 | ||
825 | if (!rb->clear_bmap || !clear_bmap_test_and_clear(rb, page)) { | |
826 | return; | |
827 | } | |
828 | ||
829 | shift = rb->clear_bmap_shift; | |
830 | /* | |
831 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
832 | * can make things easier sometimes since then start address | |
833 | * of the small chunk will always be 64 pages aligned so the | |
834 | * bitmap will always be aligned to unsigned long. We should | |
835 | * even be able to remove this restriction but I'm simply | |
836 | * keeping it. | |
837 | */ | |
838 | assert(shift >= 6); | |
839 | ||
840 | size = 1ULL << (TARGET_PAGE_BITS + shift); | |
7648297d | 841 | start = QEMU_ALIGN_DOWN((ram_addr_t)page << TARGET_PAGE_BITS, size); |
3143577d WW |
842 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); |
843 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
844 | } | |
845 | ||
846 | static void | |
1230a25f | 847 | migration_clear_memory_region_dirty_bitmap_range(RAMBlock *rb, |
3143577d WW |
848 | unsigned long start, |
849 | unsigned long npages) | |
850 | { | |
851 | unsigned long i, chunk_pages = 1UL << rb->clear_bmap_shift; | |
852 | unsigned long chunk_start = QEMU_ALIGN_DOWN(start, chunk_pages); | |
853 | unsigned long chunk_end = QEMU_ALIGN_UP(start + npages, chunk_pages); | |
854 | ||
855 | /* | |
856 | * Clear pages from start to start + npages - 1, so the end boundary is | |
857 | * exclusive. | |
858 | */ | |
859 | for (i = chunk_start; i < chunk_end; i += chunk_pages) { | |
1230a25f | 860 | migration_clear_memory_region_dirty_bitmap(rb, i); |
3143577d WW |
861 | } |
862 | } | |
863 | ||
a6a83cef RL |
864 | /* |
865 | * colo_bitmap_find_diry:find contiguous dirty pages from start | |
866 | * | |
867 | * Returns the page offset within memory region of the start of the contiguout | |
868 | * dirty page | |
869 | * | |
870 | * @rs: current RAM state | |
871 | * @rb: RAMBlock where to search for dirty pages | |
872 | * @start: page where we start the search | |
873 | * @num: the number of contiguous dirty pages | |
874 | */ | |
875 | static inline | |
876 | unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, | |
877 | unsigned long start, unsigned long *num) | |
878 | { | |
879 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; | |
880 | unsigned long *bitmap = rb->bmap; | |
881 | unsigned long first, next; | |
882 | ||
883 | *num = 0; | |
884 | ||
885 | if (ramblock_is_ignored(rb)) { | |
886 | return size; | |
887 | } | |
888 | ||
889 | first = find_next_bit(bitmap, size, start); | |
890 | if (first >= size) { | |
891 | return first; | |
892 | } | |
893 | next = find_next_zero_bit(bitmap, size, first + 1); | |
894 | assert(next >= first); | |
895 | *num = next - first; | |
896 | return first; | |
897 | } | |
898 | ||
06b10688 | 899 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
900 | RAMBlock *rb, |
901 | unsigned long page) | |
a82d593b DDAG |
902 | { |
903 | bool ret; | |
a82d593b | 904 | |
002cad6b PX |
905 | /* |
906 | * Clear dirty bitmap if needed. This _must_ be called before we | |
907 | * send any of the page in the chunk because we need to make sure | |
908 | * we can capture further page content changes when we sync dirty | |
909 | * log the next time. So as long as we are going to send any of | |
910 | * the page in the chunk we clear the remote dirty bitmap for all. | |
911 | * Clearing it earlier won't be a problem, but too late will. | |
912 | */ | |
1230a25f | 913 | migration_clear_memory_region_dirty_bitmap(rb, page); |
002cad6b | 914 | |
6b6712ef | 915 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b | 916 | if (ret) { |
0d8ec885 | 917 | rs->migration_dirty_pages--; |
a82d593b | 918 | } |
386a907b | 919 | |
a82d593b DDAG |
920 | return ret; |
921 | } | |
922 | ||
be39b4cd DH |
923 | static void dirty_bitmap_clear_section(MemoryRegionSection *section, |
924 | void *opaque) | |
925 | { | |
926 | const hwaddr offset = section->offset_within_region; | |
927 | const hwaddr size = int128_get64(section->size); | |
928 | const unsigned long start = offset >> TARGET_PAGE_BITS; | |
929 | const unsigned long npages = size >> TARGET_PAGE_BITS; | |
930 | RAMBlock *rb = section->mr->ram_block; | |
931 | uint64_t *cleared_bits = opaque; | |
932 | ||
933 | /* | |
934 | * We don't grab ram_state->bitmap_mutex because we expect to run | |
935 | * only when starting migration or during postcopy recovery where | |
936 | * we don't have concurrent access. | |
937 | */ | |
938 | if (!migration_in_postcopy() && !migrate_background_snapshot()) { | |
939 | migration_clear_memory_region_dirty_bitmap_range(rb, start, npages); | |
940 | } | |
941 | *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages); | |
942 | bitmap_clear(rb->bmap, start, npages); | |
943 | } | |
944 | ||
945 | /* | |
946 | * Exclude all dirty pages from migration that fall into a discarded range as | |
947 | * managed by a RamDiscardManager responsible for the mapped memory region of | |
948 | * the RAMBlock. Clear the corresponding bits in the dirty bitmaps. | |
949 | * | |
950 | * Discarded pages ("logically unplugged") have undefined content and must | |
951 | * not get migrated, because even reading these pages for migration might | |
952 | * result in undesired behavior. | |
953 | * | |
954 | * Returns the number of cleared bits in the RAMBlock dirty bitmap. | |
955 | * | |
956 | * Note: The result is only stable while migrating (precopy/postcopy). | |
957 | */ | |
958 | static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb) | |
959 | { | |
960 | uint64_t cleared_bits = 0; | |
961 | ||
962 | if (rb->mr && rb->bmap && memory_region_has_ram_discard_manager(rb->mr)) { | |
963 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
964 | MemoryRegionSection section = { | |
965 | .mr = rb->mr, | |
966 | .offset_within_region = 0, | |
967 | .size = int128_make64(qemu_ram_get_used_length(rb)), | |
968 | }; | |
969 | ||
970 | ram_discard_manager_replay_discarded(rdm, §ion, | |
971 | dirty_bitmap_clear_section, | |
972 | &cleared_bits); | |
973 | } | |
974 | return cleared_bits; | |
975 | } | |
976 | ||
9470c5e0 DH |
977 | /* |
978 | * Check if a host-page aligned page falls into a discarded range as managed by | |
979 | * a RamDiscardManager responsible for the mapped memory region of the RAMBlock. | |
980 | * | |
981 | * Note: The result is only stable while migrating (precopy/postcopy). | |
982 | */ | |
983 | bool ramblock_page_is_discarded(RAMBlock *rb, ram_addr_t start) | |
984 | { | |
985 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
986 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
987 | MemoryRegionSection section = { | |
988 | .mr = rb->mr, | |
989 | .offset_within_region = start, | |
990 | .size = int128_make64(qemu_ram_pagesize(rb)), | |
991 | }; | |
992 | ||
993 | return !ram_discard_manager_is_populated(rdm, §ion); | |
994 | } | |
995 | return false; | |
996 | } | |
997 | ||
267691b6 | 998 | /* Called with RCU critical section */ |
7a3e9571 | 999 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 1000 | { |
fb613580 KZ |
1001 | uint64_t new_dirty_pages = |
1002 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length); | |
1003 | ||
1004 | rs->migration_dirty_pages += new_dirty_pages; | |
1005 | rs->num_dirty_pages_period += new_dirty_pages; | |
56e93d26 JQ |
1006 | } |
1007 | ||
3d0684b2 JQ |
1008 | /** |
1009 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
1010 | * | |
1011 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
1012 | * | |
1013 | * For VMs with just normal pages this is equivalent to the host page | |
1014 | * size. If it's got some huge pages then it's the OR of all the | |
1015 | * different page sizes. | |
e8ca1db2 DDAG |
1016 | */ |
1017 | uint64_t ram_pagesize_summary(void) | |
1018 | { | |
1019 | RAMBlock *block; | |
1020 | uint64_t summary = 0; | |
1021 | ||
fbd162e6 | 1022 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
1023 | summary |= block->page_size; |
1024 | } | |
1025 | ||
1026 | return summary; | |
1027 | } | |
1028 | ||
aecbfe9c XG |
1029 | uint64_t ram_get_total_transferred_pages(void) |
1030 | { | |
1031 | return ram_counters.normal + ram_counters.duplicate + | |
1032 | compression_counters.pages + xbzrle_counters.pages; | |
1033 | } | |
1034 | ||
b734035b XG |
1035 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
1036 | { | |
be8b02ed | 1037 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 1038 | double compressed_size; |
b734035b XG |
1039 | |
1040 | /* calculate period counters */ | |
1041 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 | |
1042 | / (end_time - rs->time_last_bitmap_sync); | |
1043 | ||
be8b02ed | 1044 | if (!page_count) { |
b734035b XG |
1045 | return; |
1046 | } | |
1047 | ||
1048 | if (migrate_use_xbzrle()) { | |
e460a4b1 WW |
1049 | double encoded_size, unencoded_size; |
1050 | ||
b734035b | 1051 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - |
be8b02ed | 1052 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b | 1053 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
e460a4b1 WW |
1054 | unencoded_size = (xbzrle_counters.pages - rs->xbzrle_pages_prev) * |
1055 | TARGET_PAGE_SIZE; | |
1056 | encoded_size = xbzrle_counters.bytes - rs->xbzrle_bytes_prev; | |
92271402 | 1057 | if (xbzrle_counters.pages == rs->xbzrle_pages_prev || !encoded_size) { |
e460a4b1 | 1058 | xbzrle_counters.encoding_rate = 0; |
e460a4b1 WW |
1059 | } else { |
1060 | xbzrle_counters.encoding_rate = unencoded_size / encoded_size; | |
1061 | } | |
1062 | rs->xbzrle_pages_prev = xbzrle_counters.pages; | |
1063 | rs->xbzrle_bytes_prev = xbzrle_counters.bytes; | |
b734035b | 1064 | } |
76e03000 XG |
1065 | |
1066 | if (migrate_use_compression()) { | |
1067 | compression_counters.busy_rate = (double)(compression_counters.busy - | |
1068 | rs->compress_thread_busy_prev) / page_count; | |
1069 | rs->compress_thread_busy_prev = compression_counters.busy; | |
1070 | ||
1071 | compressed_size = compression_counters.compressed_size - | |
1072 | rs->compressed_size_prev; | |
1073 | if (compressed_size) { | |
1074 | double uncompressed_size = (compression_counters.pages - | |
1075 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
1076 | ||
1077 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
1078 | compression_counters.compression_rate = | |
1079 | uncompressed_size / compressed_size; | |
1080 | ||
1081 | rs->compress_pages_prev = compression_counters.pages; | |
1082 | rs->compressed_size_prev = compression_counters.compressed_size; | |
1083 | } | |
1084 | } | |
b734035b XG |
1085 | } |
1086 | ||
dc14a470 KZ |
1087 | static void migration_trigger_throttle(RAMState *rs) |
1088 | { | |
1089 | MigrationState *s = migrate_get_current(); | |
1090 | uint64_t threshold = s->parameters.throttle_trigger_threshold; | |
1091 | ||
1092 | uint64_t bytes_xfer_period = ram_counters.transferred - rs->bytes_xfer_prev; | |
1093 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; | |
1094 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
1095 | ||
1096 | /* During block migration the auto-converge logic incorrectly detects | |
1097 | * that ram migration makes no progress. Avoid this by disabling the | |
1098 | * throttling logic during the bulk phase of block migration. */ | |
1099 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
1100 | /* The following detection logic can be refined later. For now: | |
1101 | Check to see if the ratio between dirtied bytes and the approx. | |
1102 | amount of bytes that just got transferred since the last time | |
1103 | we were in this routine reaches the threshold. If that happens | |
1104 | twice, start or increase throttling. */ | |
1105 | ||
1106 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
1107 | (++rs->dirty_rate_high_cnt >= 2)) { | |
1108 | trace_migration_throttle(); | |
1109 | rs->dirty_rate_high_cnt = 0; | |
cbbf8182 KZ |
1110 | mig_throttle_guest_down(bytes_dirty_period, |
1111 | bytes_dirty_threshold); | |
dc14a470 KZ |
1112 | } |
1113 | } | |
1114 | } | |
1115 | ||
8d820d6f | 1116 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
1117 | { |
1118 | RAMBlock *block; | |
56e93d26 | 1119 | int64_t end_time; |
56e93d26 | 1120 | |
9360447d | 1121 | ram_counters.dirty_sync_count++; |
56e93d26 | 1122 | |
f664da80 JQ |
1123 | if (!rs->time_last_bitmap_sync) { |
1124 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
1125 | } |
1126 | ||
1127 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 1128 | memory_global_dirty_log_sync(); |
56e93d26 | 1129 | |
108cfae0 | 1130 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
1131 | WITH_RCU_READ_LOCK_GUARD() { |
1132 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1133 | ramblock_sync_dirty_bitmap(rs, block); | |
1134 | } | |
1135 | ram_counters.remaining = ram_bytes_remaining(); | |
56e93d26 | 1136 | } |
108cfae0 | 1137 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 1138 | |
9458a9a1 | 1139 | memory_global_after_dirty_log_sync(); |
a66cd90c | 1140 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 1141 | |
56e93d26 JQ |
1142 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
1143 | ||
1144 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 1145 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 1146 | migration_trigger_throttle(rs); |
070afca2 | 1147 | |
b734035b XG |
1148 | migration_update_rates(rs, end_time); |
1149 | ||
be8b02ed | 1150 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
1151 | |
1152 | /* reset period counters */ | |
f664da80 | 1153 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 1154 | rs->num_dirty_pages_period = 0; |
dc14a470 | 1155 | rs->bytes_xfer_prev = ram_counters.transferred; |
56e93d26 | 1156 | } |
4addcd4f | 1157 | if (migrate_use_events()) { |
3ab72385 | 1158 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count); |
4addcd4f | 1159 | } |
56e93d26 JQ |
1160 | } |
1161 | ||
bd227060 WW |
1162 | static void migration_bitmap_sync_precopy(RAMState *rs) |
1163 | { | |
1164 | Error *local_err = NULL; | |
1165 | ||
1166 | /* | |
1167 | * The current notifier usage is just an optimization to migration, so we | |
1168 | * don't stop the normal migration process in the error case. | |
1169 | */ | |
1170 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
1171 | error_report_err(local_err); | |
b4a1733c | 1172 | local_err = NULL; |
bd227060 WW |
1173 | } |
1174 | ||
1175 | migration_bitmap_sync(rs); | |
1176 | ||
1177 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
1178 | error_report_err(local_err); | |
1179 | } | |
1180 | } | |
1181 | ||
47fe16ff JQ |
1182 | static void ram_release_page(const char *rbname, uint64_t offset) |
1183 | { | |
1184 | if (!migrate_release_ram() || !migration_in_postcopy()) { | |
1185 | return; | |
1186 | } | |
1187 | ||
1188 | ram_discard_range(rbname, offset, TARGET_PAGE_SIZE); | |
1189 | } | |
1190 | ||
6c97ec5f XG |
1191 | /** |
1192 | * save_zero_page_to_file: send the zero page to the file | |
1193 | * | |
1194 | * Returns the size of data written to the file, 0 means the page is not | |
1195 | * a zero page | |
1196 | * | |
1197 | * @rs: current RAM state | |
1198 | * @file: the file where the data is saved | |
1199 | * @block: block that contains the page we want to send | |
1200 | * @offset: offset inside the block for the page | |
1201 | */ | |
1202 | static int save_zero_page_to_file(RAMState *rs, QEMUFile *file, | |
1203 | RAMBlock *block, ram_addr_t offset) | |
1204 | { | |
1205 | uint8_t *p = block->host + offset; | |
1206 | int len = 0; | |
1207 | ||
bad452a7 | 1208 | if (buffer_is_zero(p, TARGET_PAGE_SIZE)) { |
6c97ec5f XG |
1209 | len += save_page_header(rs, file, block, offset | RAM_SAVE_FLAG_ZERO); |
1210 | qemu_put_byte(file, 0); | |
1211 | len += 1; | |
47fe16ff | 1212 | ram_release_page(block->idstr, offset); |
6c97ec5f XG |
1213 | } |
1214 | return len; | |
1215 | } | |
1216 | ||
56e93d26 | 1217 | /** |
3d0684b2 | 1218 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1219 | * |
3d0684b2 | 1220 | * Returns the number of pages written. |
56e93d26 | 1221 | * |
f7ccd61b | 1222 | * @rs: current RAM state |
56e93d26 JQ |
1223 | * @block: block that contains the page we want to send |
1224 | * @offset: offset inside the block for the page | |
56e93d26 | 1225 | */ |
7faccdc3 | 1226 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset) |
56e93d26 | 1227 | { |
6c97ec5f | 1228 | int len = save_zero_page_to_file(rs, rs->f, block, offset); |
56e93d26 | 1229 | |
6c97ec5f | 1230 | if (len) { |
9360447d | 1231 | ram_counters.duplicate++; |
4c2d0f6d | 1232 | ram_transferred_add(len); |
6c97ec5f | 1233 | return 1; |
56e93d26 | 1234 | } |
6c97ec5f | 1235 | return -1; |
56e93d26 JQ |
1236 | } |
1237 | ||
059ff0fb XG |
1238 | /* |
1239 | * @pages: the number of pages written by the control path, | |
1240 | * < 0 - error | |
1241 | * > 0 - number of pages written | |
1242 | * | |
1243 | * Return true if the pages has been saved, otherwise false is returned. | |
1244 | */ | |
1245 | static bool control_save_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, | |
1246 | int *pages) | |
1247 | { | |
1248 | uint64_t bytes_xmit = 0; | |
1249 | int ret; | |
1250 | ||
1251 | *pages = -1; | |
1252 | ret = ram_control_save_page(rs->f, block->offset, offset, TARGET_PAGE_SIZE, | |
1253 | &bytes_xmit); | |
1254 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { | |
1255 | return false; | |
1256 | } | |
1257 | ||
1258 | if (bytes_xmit) { | |
4c2d0f6d | 1259 | ram_transferred_add(bytes_xmit); |
059ff0fb XG |
1260 | *pages = 1; |
1261 | } | |
1262 | ||
1263 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1264 | return true; | |
1265 | } | |
1266 | ||
1267 | if (bytes_xmit > 0) { | |
1268 | ram_counters.normal++; | |
1269 | } else if (bytes_xmit == 0) { | |
1270 | ram_counters.duplicate++; | |
1271 | } | |
1272 | ||
1273 | return true; | |
1274 | } | |
1275 | ||
65dacaa0 XG |
1276 | /* |
1277 | * directly send the page to the stream | |
1278 | * | |
1279 | * Returns the number of pages written. | |
1280 | * | |
1281 | * @rs: current RAM state | |
1282 | * @block: block that contains the page we want to send | |
1283 | * @offset: offset inside the block for the page | |
1284 | * @buf: the page to be sent | |
1285 | * @async: send to page asyncly | |
1286 | */ | |
1287 | static int save_normal_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, | |
1288 | uint8_t *buf, bool async) | |
1289 | { | |
4c2d0f6d DE |
1290 | ram_transferred_add(save_page_header(rs, rs->f, block, |
1291 | offset | RAM_SAVE_FLAG_PAGE)); | |
65dacaa0 XG |
1292 | if (async) { |
1293 | qemu_put_buffer_async(rs->f, buf, TARGET_PAGE_SIZE, | |
f912ec5b | 1294 | migrate_release_ram() && |
65dacaa0 XG |
1295 | migration_in_postcopy()); |
1296 | } else { | |
1297 | qemu_put_buffer(rs->f, buf, TARGET_PAGE_SIZE); | |
1298 | } | |
4c2d0f6d | 1299 | ram_transferred_add(TARGET_PAGE_SIZE); |
65dacaa0 XG |
1300 | ram_counters.normal++; |
1301 | return 1; | |
1302 | } | |
1303 | ||
56e93d26 | 1304 | /** |
3d0684b2 | 1305 | * ram_save_page: send the given page to the stream |
56e93d26 | 1306 | * |
3d0684b2 | 1307 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1308 | * < 0 - error |
1309 | * >=0 - Number of pages written - this might legally be 0 | |
1310 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1311 | * |
6f37bb8b | 1312 | * @rs: current RAM state |
56e93d26 JQ |
1313 | * @block: block that contains the page we want to send |
1314 | * @offset: offset inside the block for the page | |
56e93d26 | 1315 | */ |
05931ec5 | 1316 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss) |
56e93d26 JQ |
1317 | { |
1318 | int pages = -1; | |
56e93d26 | 1319 | uint8_t *p; |
56e93d26 | 1320 | bool send_async = true; |
a08f6890 | 1321 | RAMBlock *block = pss->block; |
8bba004c | 1322 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1323 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1324 | |
2f68e399 | 1325 | p = block->host + offset; |
1db9d8e5 | 1326 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1327 | |
56e93d26 | 1328 | XBZRLE_cache_lock(); |
1a373522 | 1329 | if (rs->xbzrle_enabled && !migration_in_postcopy()) { |
059ff0fb | 1330 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
05931ec5 JQ |
1331 | offset); |
1332 | if (!rs->last_stage) { | |
059ff0fb XG |
1333 | /* Can't send this cached data async, since the cache page |
1334 | * might get updated before it gets to the wire | |
56e93d26 | 1335 | */ |
059ff0fb | 1336 | send_async = false; |
56e93d26 JQ |
1337 | } |
1338 | } | |
1339 | ||
1340 | /* XBZRLE overflow or normal page */ | |
1341 | if (pages == -1) { | |
65dacaa0 | 1342 | pages = save_normal_page(rs, block, offset, p, send_async); |
56e93d26 JQ |
1343 | } |
1344 | ||
1345 | XBZRLE_cache_unlock(); | |
1346 | ||
1347 | return pages; | |
1348 | } | |
1349 | ||
b9ee2f7d JQ |
1350 | static int ram_save_multifd_page(RAMState *rs, RAMBlock *block, |
1351 | ram_addr_t offset) | |
1352 | { | |
67a4c891 | 1353 | if (multifd_queue_page(rs->f, block, offset) < 0) { |
713f762a IR |
1354 | return -1; |
1355 | } | |
b9ee2f7d JQ |
1356 | ram_counters.normal++; |
1357 | ||
1358 | return 1; | |
1359 | } | |
1360 | ||
5e5fdcff | 1361 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 1362 | ram_addr_t offset, uint8_t *source_buf) |
56e93d26 | 1363 | { |
53518d94 | 1364 | RAMState *rs = ram_state; |
20d549cb | 1365 | uint8_t *p = block->host + offset; |
6ef3771c | 1366 | int ret; |
56e93d26 | 1367 | |
5e5fdcff | 1368 | if (save_zero_page_to_file(rs, f, block, offset)) { |
e7f2e190 | 1369 | return true; |
5e5fdcff XG |
1370 | } |
1371 | ||
6ef3771c | 1372 | save_page_header(rs, f, block, offset | RAM_SAVE_FLAG_COMPRESS_PAGE); |
34ab9e97 XG |
1373 | |
1374 | /* | |
1375 | * copy it to a internal buffer to avoid it being modified by VM | |
1376 | * so that we can catch up the error during compression and | |
1377 | * decompression | |
1378 | */ | |
1379 | memcpy(source_buf, p, TARGET_PAGE_SIZE); | |
6ef3771c XG |
1380 | ret = qemu_put_compression_data(f, stream, source_buf, TARGET_PAGE_SIZE); |
1381 | if (ret < 0) { | |
1382 | qemu_file_set_error(migrate_get_current()->to_dst_file, ret); | |
b3be2896 | 1383 | error_report("compressed data failed!"); |
b3be2896 | 1384 | } |
e7f2e190 | 1385 | return false; |
5e5fdcff XG |
1386 | } |
1387 | ||
1388 | static void | |
1389 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1390 | { | |
4c2d0f6d | 1391 | ram_transferred_add(bytes_xmit); |
76e03000 | 1392 | |
5e5fdcff XG |
1393 | if (param->zero_page) { |
1394 | ram_counters.duplicate++; | |
76e03000 | 1395 | return; |
5e5fdcff | 1396 | } |
76e03000 XG |
1397 | |
1398 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1399 | compression_counters.compressed_size += bytes_xmit - 8; | |
1400 | compression_counters.pages++; | |
56e93d26 JQ |
1401 | } |
1402 | ||
32b05495 XG |
1403 | static bool save_page_use_compression(RAMState *rs); |
1404 | ||
ce25d337 | 1405 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
1406 | { |
1407 | int idx, len, thread_count; | |
1408 | ||
32b05495 | 1409 | if (!save_page_use_compression(rs)) { |
56e93d26 JQ |
1410 | return; |
1411 | } | |
1412 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1413 | |
0d9f9a5c | 1414 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1415 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1416 | while (!comp_param[idx].done) { |
0d9f9a5c | 1417 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1418 | } |
a7a9a88f | 1419 | } |
0d9f9a5c | 1420 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1421 | |
1422 | for (idx = 0; idx < thread_count; idx++) { | |
1423 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1424 | if (!comp_param[idx].quit) { |
ce25d337 | 1425 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
5e5fdcff XG |
1426 | /* |
1427 | * it's safe to fetch zero_page without holding comp_done_lock | |
1428 | * as there is no further request submitted to the thread, | |
1429 | * i.e, the thread should be waiting for a request at this point. | |
1430 | */ | |
1431 | update_compress_thread_counts(&comp_param[idx], len); | |
56e93d26 | 1432 | } |
a7a9a88f | 1433 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1434 | } |
1435 | } | |
1436 | ||
1437 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1438 | ram_addr_t offset) | |
1439 | { | |
1440 | param->block = block; | |
1441 | param->offset = offset; | |
1442 | } | |
1443 | ||
ce25d337 JQ |
1444 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
1445 | ram_addr_t offset) | |
56e93d26 JQ |
1446 | { |
1447 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1d58872a | 1448 | bool wait = migrate_compress_wait_thread(); |
56e93d26 JQ |
1449 | |
1450 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1451 | qemu_mutex_lock(&comp_done_lock); |
1d58872a XG |
1452 | retry: |
1453 | for (idx = 0; idx < thread_count; idx++) { | |
1454 | if (comp_param[idx].done) { | |
1455 | comp_param[idx].done = false; | |
1456 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); | |
1457 | qemu_mutex_lock(&comp_param[idx].mutex); | |
1458 | set_compress_params(&comp_param[idx], block, offset); | |
1459 | qemu_cond_signal(&comp_param[idx].cond); | |
1460 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
1461 | pages = 1; | |
5e5fdcff | 1462 | update_compress_thread_counts(&comp_param[idx], bytes_xmit); |
56e93d26 | 1463 | break; |
56e93d26 JQ |
1464 | } |
1465 | } | |
1d58872a XG |
1466 | |
1467 | /* | |
1468 | * wait for the free thread if the user specifies 'compress-wait-thread', | |
1469 | * otherwise we will post the page out in the main thread as normal page. | |
1470 | */ | |
1471 | if (pages < 0 && wait) { | |
1472 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); | |
1473 | goto retry; | |
1474 | } | |
0d9f9a5c | 1475 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1476 | |
1477 | return pages; | |
1478 | } | |
1479 | ||
3d0684b2 JQ |
1480 | /** |
1481 | * find_dirty_block: find the next dirty page and update any state | |
1482 | * associated with the search process. | |
b9e60928 | 1483 | * |
a5f7b1a6 | 1484 | * Returns true if a page is found |
b9e60928 | 1485 | * |
6f37bb8b | 1486 | * @rs: current RAM state |
3d0684b2 JQ |
1487 | * @pss: data about the state of the current dirty page scan |
1488 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1489 | */ |
f20e2865 | 1490 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1491 | { |
ea2faf0c PX |
1492 | /* This is not a postcopy requested page */ |
1493 | pss->postcopy_requested = false; | |
1494 | ||
f20e2865 | 1495 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
6f37bb8b | 1496 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1497 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1498 | /* |
1499 | * We've been once around the RAM and haven't found anything. | |
1500 | * Give up. | |
1501 | */ | |
1502 | *again = false; | |
1503 | return false; | |
1504 | } | |
542147f4 DH |
1505 | if (!offset_in_ramblock(pss->block, |
1506 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)) { | |
b9e60928 | 1507 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1508 | pss->page = 0; |
b9e60928 DDAG |
1509 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1510 | if (!pss->block) { | |
48df9d80 XG |
1511 | /* |
1512 | * If memory migration starts over, we will meet a dirtied page | |
1513 | * which may still exists in compression threads's ring, so we | |
1514 | * should flush the compressed data to make sure the new page | |
1515 | * is not overwritten by the old one in the destination. | |
1516 | * | |
1517 | * Also If xbzrle is on, stop using the data compression at this | |
1518 | * point. In theory, xbzrle can do better than compression. | |
1519 | */ | |
1520 | flush_compressed_data(rs); | |
1521 | ||
b9e60928 DDAG |
1522 | /* Hit the end of the list */ |
1523 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1524 | /* Flag that we've looped */ | |
1525 | pss->complete_round = true; | |
1a373522 DH |
1526 | /* After the first round, enable XBZRLE. */ |
1527 | if (migrate_use_xbzrle()) { | |
1528 | rs->xbzrle_enabled = true; | |
1529 | } | |
b9e60928 DDAG |
1530 | } |
1531 | /* Didn't find anything this time, but try again on the new block */ | |
1532 | *again = true; | |
1533 | return false; | |
1534 | } else { | |
1535 | /* Can go around again, but... */ | |
1536 | *again = true; | |
1537 | /* We've found something so probably don't need to */ | |
1538 | return true; | |
1539 | } | |
1540 | } | |
1541 | ||
3d0684b2 JQ |
1542 | /** |
1543 | * unqueue_page: gets a page of the queue | |
1544 | * | |
a82d593b | 1545 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1546 | * |
3d0684b2 JQ |
1547 | * Returns the block of the page (or NULL if none available) |
1548 | * | |
ec481c6c | 1549 | * @rs: current RAM state |
3d0684b2 | 1550 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1551 | */ |
f20e2865 | 1552 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b | 1553 | { |
a1fe28df | 1554 | struct RAMSrcPageRequest *entry; |
a82d593b | 1555 | RAMBlock *block = NULL; |
cfd66f30 | 1556 | size_t page_size; |
a82d593b | 1557 | |
a1fe28df | 1558 | if (!postcopy_has_request(rs)) { |
ae526e32 XG |
1559 | return NULL; |
1560 | } | |
1561 | ||
6e8a355d | 1562 | QEMU_LOCK_GUARD(&rs->src_page_req_mutex); |
a1fe28df PX |
1563 | |
1564 | /* | |
1565 | * This should _never_ change even after we take the lock, because no one | |
1566 | * should be taking anything off the request list other than us. | |
1567 | */ | |
1568 | assert(postcopy_has_request(rs)); | |
1569 | ||
1570 | entry = QSIMPLEQ_FIRST(&rs->src_page_requests); | |
1571 | block = entry->rb; | |
1572 | *offset = entry->offset; | |
cfd66f30 PX |
1573 | page_size = qemu_ram_pagesize(block); |
1574 | /* Each page request should only be multiple page size of the ramblock */ | |
1575 | assert((entry->len % page_size) == 0); | |
a1fe28df | 1576 | |
cfd66f30 PX |
1577 | if (entry->len > page_size) { |
1578 | entry->len -= page_size; | |
1579 | entry->offset += page_size; | |
a1fe28df PX |
1580 | } else { |
1581 | memory_region_unref(block->mr); | |
1582 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); | |
1583 | g_free(entry); | |
1584 | migration_consume_urgent_request(); | |
a82d593b | 1585 | } |
a82d593b | 1586 | |
cfd66f30 PX |
1587 | trace_unqueue_page(block->idstr, *offset, |
1588 | test_bit((*offset >> TARGET_PAGE_BITS), block->bmap)); | |
1589 | ||
a82d593b DDAG |
1590 | return block; |
1591 | } | |
1592 | ||
278e2f55 AG |
1593 | #if defined(__linux__) |
1594 | /** | |
1595 | * poll_fault_page: try to get next UFFD write fault page and, if pending fault | |
1596 | * is found, return RAM block pointer and page offset | |
1597 | * | |
1598 | * Returns pointer to the RAMBlock containing faulting page, | |
1599 | * NULL if no write faults are pending | |
1600 | * | |
1601 | * @rs: current RAM state | |
1602 | * @offset: page offset from the beginning of the block | |
1603 | */ | |
1604 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1605 | { | |
1606 | struct uffd_msg uffd_msg; | |
1607 | void *page_address; | |
82ea3e3b | 1608 | RAMBlock *block; |
278e2f55 AG |
1609 | int res; |
1610 | ||
1611 | if (!migrate_background_snapshot()) { | |
1612 | return NULL; | |
1613 | } | |
1614 | ||
1615 | res = uffd_read_events(rs->uffdio_fd, &uffd_msg, 1); | |
1616 | if (res <= 0) { | |
1617 | return NULL; | |
1618 | } | |
1619 | ||
1620 | page_address = (void *)(uintptr_t) uffd_msg.arg.pagefault.address; | |
82ea3e3b AG |
1621 | block = qemu_ram_block_from_host(page_address, false, offset); |
1622 | assert(block && (block->flags & RAM_UF_WRITEPROTECT) != 0); | |
1623 | return block; | |
278e2f55 AG |
1624 | } |
1625 | ||
1626 | /** | |
1627 | * ram_save_release_protection: release UFFD write protection after | |
1628 | * a range of pages has been saved | |
1629 | * | |
1630 | * @rs: current RAM state | |
1631 | * @pss: page-search-status structure | |
1632 | * @start_page: index of the first page in the range relative to pss->block | |
1633 | * | |
1634 | * Returns 0 on success, negative value in case of an error | |
1635 | */ | |
1636 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1637 | unsigned long start_page) | |
1638 | { | |
1639 | int res = 0; | |
1640 | ||
1641 | /* Check if page is from UFFD-managed region. */ | |
1642 | if (pss->block->flags & RAM_UF_WRITEPROTECT) { | |
1643 | void *page_address = pss->block->host + (start_page << TARGET_PAGE_BITS); | |
258f5c98 | 1644 | uint64_t run_length = (pss->page - start_page) << TARGET_PAGE_BITS; |
278e2f55 AG |
1645 | |
1646 | /* Flush async buffers before un-protect. */ | |
1647 | qemu_fflush(rs->f); | |
1648 | /* Un-protect memory range. */ | |
1649 | res = uffd_change_protection(rs->uffdio_fd, page_address, run_length, | |
1650 | false, false); | |
1651 | } | |
1652 | ||
1653 | return res; | |
1654 | } | |
1655 | ||
1656 | /* ram_write_tracking_available: check if kernel supports required UFFD features | |
1657 | * | |
1658 | * Returns true if supports, false otherwise | |
1659 | */ | |
1660 | bool ram_write_tracking_available(void) | |
1661 | { | |
1662 | uint64_t uffd_features; | |
1663 | int res; | |
1664 | ||
1665 | res = uffd_query_features(&uffd_features); | |
1666 | return (res == 0 && | |
1667 | (uffd_features & UFFD_FEATURE_PAGEFAULT_FLAG_WP) != 0); | |
1668 | } | |
1669 | ||
1670 | /* ram_write_tracking_compatible: check if guest configuration is | |
1671 | * compatible with 'write-tracking' | |
1672 | * | |
1673 | * Returns true if compatible, false otherwise | |
1674 | */ | |
1675 | bool ram_write_tracking_compatible(void) | |
1676 | { | |
1677 | const uint64_t uffd_ioctls_mask = BIT(_UFFDIO_WRITEPROTECT); | |
1678 | int uffd_fd; | |
82ea3e3b | 1679 | RAMBlock *block; |
278e2f55 AG |
1680 | bool ret = false; |
1681 | ||
1682 | /* Open UFFD file descriptor */ | |
1683 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, false); | |
1684 | if (uffd_fd < 0) { | |
1685 | return false; | |
1686 | } | |
1687 | ||
1688 | RCU_READ_LOCK_GUARD(); | |
1689 | ||
82ea3e3b | 1690 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 AG |
1691 | uint64_t uffd_ioctls; |
1692 | ||
1693 | /* Nothing to do with read-only and MMIO-writable regions */ | |
82ea3e3b | 1694 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1695 | continue; |
1696 | } | |
1697 | /* Try to register block memory via UFFD-IO to track writes */ | |
82ea3e3b | 1698 | if (uffd_register_memory(uffd_fd, block->host, block->max_length, |
278e2f55 AG |
1699 | UFFDIO_REGISTER_MODE_WP, &uffd_ioctls)) { |
1700 | goto out; | |
1701 | } | |
1702 | if ((uffd_ioctls & uffd_ioctls_mask) != uffd_ioctls_mask) { | |
1703 | goto out; | |
1704 | } | |
1705 | } | |
1706 | ret = true; | |
1707 | ||
1708 | out: | |
1709 | uffd_close_fd(uffd_fd); | |
1710 | return ret; | |
1711 | } | |
1712 | ||
f7b9dcfb DH |
1713 | static inline void populate_read_range(RAMBlock *block, ram_addr_t offset, |
1714 | ram_addr_t size) | |
1715 | { | |
1716 | /* | |
1717 | * We read one byte of each page; this will preallocate page tables if | |
1718 | * required and populate the shared zeropage on MAP_PRIVATE anonymous memory | |
1719 | * where no page was populated yet. This might require adaption when | |
1720 | * supporting other mappings, like shmem. | |
1721 | */ | |
1722 | for (; offset < size; offset += block->page_size) { | |
1723 | char tmp = *((char *)block->host + offset); | |
1724 | ||
1725 | /* Don't optimize the read out */ | |
1726 | asm volatile("" : "+r" (tmp)); | |
1727 | } | |
1728 | } | |
1729 | ||
6fee3a1f DH |
1730 | static inline int populate_read_section(MemoryRegionSection *section, |
1731 | void *opaque) | |
1732 | { | |
1733 | const hwaddr size = int128_get64(section->size); | |
1734 | hwaddr offset = section->offset_within_region; | |
1735 | RAMBlock *block = section->mr->ram_block; | |
1736 | ||
1737 | populate_read_range(block, offset, size); | |
1738 | return 0; | |
1739 | } | |
1740 | ||
eeccb99c | 1741 | /* |
f7b9dcfb DH |
1742 | * ram_block_populate_read: preallocate page tables and populate pages in the |
1743 | * RAM block by reading a byte of each page. | |
eeccb99c AG |
1744 | * |
1745 | * Since it's solely used for userfault_fd WP feature, here we just | |
1746 | * hardcode page size to qemu_real_host_page_size. | |
1747 | * | |
82ea3e3b | 1748 | * @block: RAM block to populate |
eeccb99c | 1749 | */ |
6fee3a1f | 1750 | static void ram_block_populate_read(RAMBlock *rb) |
eeccb99c | 1751 | { |
6fee3a1f DH |
1752 | /* |
1753 | * Skip populating all pages that fall into a discarded range as managed by | |
1754 | * a RamDiscardManager responsible for the mapped memory region of the | |
1755 | * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock | |
1756 | * must not get populated automatically. We don't have to track | |
1757 | * modifications via userfaultfd WP reliably, because these pages will | |
1758 | * not be part of the migration stream either way -- see | |
1759 | * ramblock_dirty_bitmap_exclude_discarded_pages(). | |
1760 | * | |
1761 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1762 | */ | |
1763 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1764 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1765 | MemoryRegionSection section = { | |
1766 | .mr = rb->mr, | |
1767 | .offset_within_region = 0, | |
1768 | .size = rb->mr->size, | |
1769 | }; | |
1770 | ||
1771 | ram_discard_manager_replay_populated(rdm, §ion, | |
1772 | populate_read_section, NULL); | |
1773 | } else { | |
1774 | populate_read_range(rb, 0, rb->used_length); | |
1775 | } | |
eeccb99c AG |
1776 | } |
1777 | ||
1778 | /* | |
1779 | * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking | |
1780 | */ | |
1781 | void ram_write_tracking_prepare(void) | |
1782 | { | |
82ea3e3b | 1783 | RAMBlock *block; |
eeccb99c AG |
1784 | |
1785 | RCU_READ_LOCK_GUARD(); | |
1786 | ||
82ea3e3b | 1787 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
eeccb99c | 1788 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1789 | if (block->mr->readonly || block->mr->rom_device) { |
eeccb99c AG |
1790 | continue; |
1791 | } | |
1792 | ||
1793 | /* | |
1794 | * Populate pages of the RAM block before enabling userfault_fd | |
1795 | * write protection. | |
1796 | * | |
1797 | * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with | |
1798 | * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip | |
1799 | * pages with pte_none() entries in page table. | |
1800 | */ | |
f7b9dcfb | 1801 | ram_block_populate_read(block); |
eeccb99c AG |
1802 | } |
1803 | } | |
1804 | ||
278e2f55 AG |
1805 | /* |
1806 | * ram_write_tracking_start: start UFFD-WP memory tracking | |
1807 | * | |
1808 | * Returns 0 for success or negative value in case of error | |
1809 | */ | |
1810 | int ram_write_tracking_start(void) | |
1811 | { | |
1812 | int uffd_fd; | |
1813 | RAMState *rs = ram_state; | |
82ea3e3b | 1814 | RAMBlock *block; |
278e2f55 AG |
1815 | |
1816 | /* Open UFFD file descriptor */ | |
1817 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, true); | |
1818 | if (uffd_fd < 0) { | |
1819 | return uffd_fd; | |
1820 | } | |
1821 | rs->uffdio_fd = uffd_fd; | |
1822 | ||
1823 | RCU_READ_LOCK_GUARD(); | |
1824 | ||
82ea3e3b | 1825 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 | 1826 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1827 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1828 | continue; |
1829 | } | |
1830 | ||
1831 | /* Register block memory with UFFD to track writes */ | |
82ea3e3b AG |
1832 | if (uffd_register_memory(rs->uffdio_fd, block->host, |
1833 | block->max_length, UFFDIO_REGISTER_MODE_WP, NULL)) { | |
278e2f55 AG |
1834 | goto fail; |
1835 | } | |
1836 | /* Apply UFFD write protection to the block memory range */ | |
82ea3e3b AG |
1837 | if (uffd_change_protection(rs->uffdio_fd, block->host, |
1838 | block->max_length, true, false)) { | |
278e2f55 AG |
1839 | goto fail; |
1840 | } | |
82ea3e3b AG |
1841 | block->flags |= RAM_UF_WRITEPROTECT; |
1842 | memory_region_ref(block->mr); | |
278e2f55 | 1843 | |
82ea3e3b AG |
1844 | trace_ram_write_tracking_ramblock_start(block->idstr, block->page_size, |
1845 | block->host, block->max_length); | |
278e2f55 AG |
1846 | } |
1847 | ||
1848 | return 0; | |
1849 | ||
1850 | fail: | |
1851 | error_report("ram_write_tracking_start() failed: restoring initial memory state"); | |
1852 | ||
82ea3e3b AG |
1853 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1854 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1855 | continue; |
1856 | } | |
1857 | /* | |
1858 | * In case some memory block failed to be write-protected | |
1859 | * remove protection and unregister all succeeded RAM blocks | |
1860 | */ | |
82ea3e3b AG |
1861 | uffd_change_protection(rs->uffdio_fd, block->host, block->max_length, |
1862 | false, false); | |
1863 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); | |
278e2f55 | 1864 | /* Cleanup flags and remove reference */ |
82ea3e3b AG |
1865 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1866 | memory_region_unref(block->mr); | |
278e2f55 AG |
1867 | } |
1868 | ||
1869 | uffd_close_fd(uffd_fd); | |
1870 | rs->uffdio_fd = -1; | |
1871 | return -1; | |
1872 | } | |
1873 | ||
1874 | /** | |
1875 | * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection | |
1876 | */ | |
1877 | void ram_write_tracking_stop(void) | |
1878 | { | |
1879 | RAMState *rs = ram_state; | |
82ea3e3b | 1880 | RAMBlock *block; |
278e2f55 AG |
1881 | |
1882 | RCU_READ_LOCK_GUARD(); | |
1883 | ||
82ea3e3b AG |
1884 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1885 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1886 | continue; |
1887 | } | |
1888 | /* Remove protection and unregister all affected RAM blocks */ | |
82ea3e3b AG |
1889 | uffd_change_protection(rs->uffdio_fd, block->host, block->max_length, |
1890 | false, false); | |
1891 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); | |
278e2f55 | 1892 | |
82ea3e3b AG |
1893 | trace_ram_write_tracking_ramblock_stop(block->idstr, block->page_size, |
1894 | block->host, block->max_length); | |
278e2f55 AG |
1895 | |
1896 | /* Cleanup flags and remove reference */ | |
82ea3e3b AG |
1897 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1898 | memory_region_unref(block->mr); | |
278e2f55 AG |
1899 | } |
1900 | ||
1901 | /* Finally close UFFD file descriptor */ | |
1902 | uffd_close_fd(rs->uffdio_fd); | |
1903 | rs->uffdio_fd = -1; | |
1904 | } | |
1905 | ||
1906 | #else | |
1907 | /* No target OS support, stubs just fail or ignore */ | |
1908 | ||
1909 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1910 | { | |
1911 | (void) rs; | |
1912 | (void) offset; | |
1913 | ||
1914 | return NULL; | |
1915 | } | |
1916 | ||
1917 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1918 | unsigned long start_page) | |
1919 | { | |
1920 | (void) rs; | |
1921 | (void) pss; | |
1922 | (void) start_page; | |
1923 | ||
1924 | return 0; | |
1925 | } | |
1926 | ||
1927 | bool ram_write_tracking_available(void) | |
1928 | { | |
1929 | return false; | |
1930 | } | |
1931 | ||
1932 | bool ram_write_tracking_compatible(void) | |
1933 | { | |
1934 | assert(0); | |
1935 | return false; | |
1936 | } | |
1937 | ||
1938 | int ram_write_tracking_start(void) | |
1939 | { | |
1940 | assert(0); | |
1941 | return -1; | |
1942 | } | |
1943 | ||
1944 | void ram_write_tracking_stop(void) | |
1945 | { | |
1946 | assert(0); | |
1947 | } | |
1948 | #endif /* defined(__linux__) */ | |
1949 | ||
3d0684b2 | 1950 | /** |
ff1543af | 1951 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
1952 | * |
1953 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1954 | * |
a5f7b1a6 | 1955 | * Returns true if a queued page is found |
a82d593b | 1956 | * |
6f37bb8b | 1957 | * @rs: current RAM state |
3d0684b2 | 1958 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1959 | */ |
f20e2865 | 1960 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1961 | { |
1962 | RAMBlock *block; | |
1963 | ram_addr_t offset; | |
a82d593b | 1964 | |
cfd66f30 | 1965 | block = unqueue_page(rs, &offset); |
a82d593b | 1966 | |
278e2f55 AG |
1967 | if (!block) { |
1968 | /* | |
1969 | * Poll write faults too if background snapshot is enabled; that's | |
1970 | * when we have vcpus got blocked by the write protected pages. | |
1971 | */ | |
1972 | block = poll_fault_page(rs, &offset); | |
1973 | } | |
1974 | ||
a82d593b | 1975 | if (block) { |
a82d593b DDAG |
1976 | /* |
1977 | * We want the background search to continue from the queued page | |
1978 | * since the guest is likely to want other pages near to the page | |
1979 | * it just requested. | |
1980 | */ | |
1981 | pss->block = block; | |
a935e30f | 1982 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
1983 | |
1984 | /* | |
1985 | * This unqueued page would break the "one round" check, even is | |
1986 | * really rare. | |
1987 | */ | |
1988 | pss->complete_round = false; | |
ea2faf0c | 1989 | pss->postcopy_requested = true; |
a82d593b DDAG |
1990 | } |
1991 | ||
1992 | return !!block; | |
1993 | } | |
1994 | ||
6c595cde | 1995 | /** |
5e58f968 JQ |
1996 | * migration_page_queue_free: drop any remaining pages in the ram |
1997 | * request queue | |
6c595cde | 1998 | * |
3d0684b2 JQ |
1999 | * It should be empty at the end anyway, but in error cases there may |
2000 | * be some left. in case that there is any page left, we drop it. | |
2001 | * | |
6c595cde | 2002 | */ |
83c13382 | 2003 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 2004 | { |
ec481c6c | 2005 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
2006 | /* This queue generally should be empty - but in the case of a failed |
2007 | * migration might have some droppings in. | |
2008 | */ | |
89ac5a1d | 2009 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 2010 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 2011 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 2012 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
2013 | g_free(mspr); |
2014 | } | |
6c595cde DDAG |
2015 | } |
2016 | ||
2017 | /** | |
3d0684b2 JQ |
2018 | * ram_save_queue_pages: queue the page for transmission |
2019 | * | |
2020 | * A request from postcopy destination for example. | |
2021 | * | |
2022 | * Returns zero on success or negative on error | |
2023 | * | |
3d0684b2 JQ |
2024 | * @rbname: Name of the RAMBLock of the request. NULL means the |
2025 | * same that last one. | |
2026 | * @start: starting address from the start of the RAMBlock | |
2027 | * @len: length (in bytes) to send | |
6c595cde | 2028 | */ |
96506894 | 2029 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
2030 | { |
2031 | RAMBlock *ramblock; | |
53518d94 | 2032 | RAMState *rs = ram_state; |
6c595cde | 2033 | |
9360447d | 2034 | ram_counters.postcopy_requests++; |
89ac5a1d DDAG |
2035 | RCU_READ_LOCK_GUARD(); |
2036 | ||
6c595cde DDAG |
2037 | if (!rbname) { |
2038 | /* Reuse last RAMBlock */ | |
68a098f3 | 2039 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
2040 | |
2041 | if (!ramblock) { | |
2042 | /* | |
2043 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
2044 | * it's the 1st request. | |
2045 | */ | |
2046 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 2047 | return -1; |
6c595cde DDAG |
2048 | } |
2049 | } else { | |
2050 | ramblock = qemu_ram_block_by_name(rbname); | |
2051 | ||
2052 | if (!ramblock) { | |
2053 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
2054 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 2055 | return -1; |
6c595cde | 2056 | } |
68a098f3 | 2057 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
2058 | } |
2059 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
542147f4 | 2060 | if (!offset_in_ramblock(ramblock, start + len - 1)) { |
9458ad6b JQ |
2061 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
2062 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 2063 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 2064 | return -1; |
6c595cde DDAG |
2065 | } |
2066 | ||
ec481c6c | 2067 | struct RAMSrcPageRequest *new_entry = |
b21e2380 | 2068 | g_new0(struct RAMSrcPageRequest, 1); |
6c595cde DDAG |
2069 | new_entry->rb = ramblock; |
2070 | new_entry->offset = start; | |
2071 | new_entry->len = len; | |
2072 | ||
2073 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
2074 | qemu_mutex_lock(&rs->src_page_req_mutex); |
2075 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 2076 | migration_make_urgent_request(); |
ec481c6c | 2077 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
2078 | |
2079 | return 0; | |
6c595cde DDAG |
2080 | } |
2081 | ||
d7400a34 XG |
2082 | static bool save_page_use_compression(RAMState *rs) |
2083 | { | |
2084 | if (!migrate_use_compression()) { | |
2085 | return false; | |
2086 | } | |
2087 | ||
2088 | /* | |
1a373522 DH |
2089 | * If xbzrle is enabled (e.g., after first round of migration), stop |
2090 | * using the data compression. In theory, xbzrle can do better than | |
2091 | * compression. | |
d7400a34 | 2092 | */ |
1a373522 DH |
2093 | if (rs->xbzrle_enabled) { |
2094 | return false; | |
d7400a34 XG |
2095 | } |
2096 | ||
1a373522 | 2097 | return true; |
d7400a34 XG |
2098 | } |
2099 | ||
5e5fdcff XG |
2100 | /* |
2101 | * try to compress the page before posting it out, return true if the page | |
2102 | * has been properly handled by compression, otherwise needs other | |
2103 | * paths to handle it | |
2104 | */ | |
2105 | static bool save_compress_page(RAMState *rs, RAMBlock *block, ram_addr_t offset) | |
2106 | { | |
2107 | if (!save_page_use_compression(rs)) { | |
2108 | return false; | |
2109 | } | |
2110 | ||
2111 | /* | |
2112 | * When starting the process of a new block, the first page of | |
2113 | * the block should be sent out before other pages in the same | |
2114 | * block, and all the pages in last block should have been sent | |
2115 | * out, keeping this order is important, because the 'cont' flag | |
2116 | * is used to avoid resending the block name. | |
2117 | * | |
2118 | * We post the fist page as normal page as compression will take | |
2119 | * much CPU resource. | |
2120 | */ | |
2121 | if (block != rs->last_sent_block) { | |
2122 | flush_compressed_data(rs); | |
2123 | return false; | |
2124 | } | |
2125 | ||
2126 | if (compress_page_with_multi_thread(rs, block, offset) > 0) { | |
2127 | return true; | |
2128 | } | |
2129 | ||
76e03000 | 2130 | compression_counters.busy++; |
5e5fdcff XG |
2131 | return false; |
2132 | } | |
2133 | ||
a82d593b | 2134 | /** |
3d0684b2 | 2135 | * ram_save_target_page: save one target page |
a82d593b | 2136 | * |
3d0684b2 | 2137 | * Returns the number of pages written |
a82d593b | 2138 | * |
6f37bb8b | 2139 | * @rs: current RAM state |
3d0684b2 | 2140 | * @pss: data about the page we want to send |
a82d593b | 2141 | */ |
05931ec5 | 2142 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2143 | { |
a8ec91f9 | 2144 | RAMBlock *block = pss->block; |
8bba004c | 2145 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
2146 | int res; |
2147 | ||
2148 | if (control_save_page(rs, block, offset, &res)) { | |
2149 | return res; | |
2150 | } | |
2151 | ||
5e5fdcff XG |
2152 | if (save_compress_page(rs, block, offset)) { |
2153 | return 1; | |
d7400a34 XG |
2154 | } |
2155 | ||
2156 | res = save_zero_page(rs, block, offset); | |
2157 | if (res > 0) { | |
2158 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
2159 | * page would be stale | |
2160 | */ | |
2161 | if (!save_page_use_compression(rs)) { | |
2162 | XBZRLE_cache_lock(); | |
2163 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
2164 | XBZRLE_cache_unlock(); | |
2165 | } | |
d7400a34 XG |
2166 | return res; |
2167 | } | |
2168 | ||
da3f56cb | 2169 | /* |
c6b3a2e0 WY |
2170 | * Do not use multifd for: |
2171 | * 1. Compression as the first page in the new block should be posted out | |
2172 | * before sending the compressed page | |
2173 | * 2. In postcopy as one whole host page should be placed | |
da3f56cb | 2174 | */ |
c6b3a2e0 WY |
2175 | if (!save_page_use_compression(rs) && migrate_use_multifd() |
2176 | && !migration_in_postcopy()) { | |
b9ee2f7d | 2177 | return ram_save_multifd_page(rs, block, offset); |
a82d593b DDAG |
2178 | } |
2179 | ||
05931ec5 | 2180 | return ram_save_page(rs, pss); |
a82d593b DDAG |
2181 | } |
2182 | ||
2183 | /** | |
3d0684b2 | 2184 | * ram_save_host_page: save a whole host page |
a82d593b | 2185 | * |
3d0684b2 JQ |
2186 | * Starting at *offset send pages up to the end of the current host |
2187 | * page. It's valid for the initial offset to point into the middle of | |
2188 | * a host page in which case the remainder of the hostpage is sent. | |
2189 | * Only dirty target pages are sent. Note that the host page size may | |
2190 | * be a huge page for this block. | |
1eb3fc0a DDAG |
2191 | * The saving stops at the boundary of the used_length of the block |
2192 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 2193 | * |
3d0684b2 JQ |
2194 | * Returns the number of pages written or negative on error |
2195 | * | |
6f37bb8b | 2196 | * @rs: current RAM state |
3d0684b2 | 2197 | * @pss: data about the page we want to send |
a82d593b | 2198 | */ |
05931ec5 | 2199 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
2200 | { |
2201 | int tmppages, pages = 0; | |
a935e30f JQ |
2202 | size_t pagesize_bits = |
2203 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
ba1b7c81 KJ |
2204 | unsigned long hostpage_boundary = |
2205 | QEMU_ALIGN_UP(pss->page + 1, pagesize_bits); | |
278e2f55 AG |
2206 | unsigned long start_page = pss->page; |
2207 | int res; | |
4c011c37 | 2208 | |
fbd162e6 | 2209 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
2210 | error_report("block %s should not be migrated !", pss->block->idstr); |
2211 | return 0; | |
2212 | } | |
2213 | ||
a82d593b | 2214 | do { |
1faa5665 | 2215 | /* Check the pages is dirty and if it is send it */ |
ba1b7c81 | 2216 | if (migration_bitmap_clear_dirty(rs, pss->block, pss->page)) { |
05931ec5 | 2217 | tmppages = ram_save_target_page(rs, pss); |
ba1b7c81 KJ |
2218 | if (tmppages < 0) { |
2219 | return tmppages; | |
2220 | } | |
a82d593b | 2221 | |
ba1b7c81 KJ |
2222 | pages += tmppages; |
2223 | /* | |
2224 | * Allow rate limiting to happen in the middle of huge pages if | |
2225 | * something is sent in the current iteration. | |
2226 | */ | |
2227 | if (pagesize_bits > 1 && tmppages > 0) { | |
2228 | migration_rate_limit(); | |
2229 | } | |
23feba90 | 2230 | } |
ba1b7c81 KJ |
2231 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
2232 | } while ((pss->page < hostpage_boundary) && | |
8bba004c AR |
2233 | offset_in_ramblock(pss->block, |
2234 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)); | |
ba1b7c81 | 2235 | /* The offset we leave with is the min boundary of host page and block */ |
258f5c98 | 2236 | pss->page = MIN(pss->page, hostpage_boundary); |
278e2f55 AG |
2237 | |
2238 | res = ram_save_release_protection(rs, pss, start_page); | |
2239 | return (res < 0 ? res : pages); | |
a82d593b | 2240 | } |
6c595cde | 2241 | |
56e93d26 | 2242 | /** |
3d0684b2 | 2243 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
2244 | * |
2245 | * Called within an RCU critical section. | |
2246 | * | |
e8f3735f XG |
2247 | * Returns the number of pages written where zero means no dirty pages, |
2248 | * or negative on error | |
56e93d26 | 2249 | * |
6f37bb8b | 2250 | * @rs: current RAM state |
a82d593b DDAG |
2251 | * |
2252 | * On systems where host-page-size > target-page-size it will send all the | |
2253 | * pages in a host page that are dirty. | |
56e93d26 | 2254 | */ |
05931ec5 | 2255 | static int ram_find_and_save_block(RAMState *rs) |
56e93d26 | 2256 | { |
b8fb8cb7 | 2257 | PageSearchStatus pss; |
56e93d26 | 2258 | int pages = 0; |
b9e60928 | 2259 | bool again, found; |
56e93d26 | 2260 | |
0827b9e9 AA |
2261 | /* No dirty page as there is zero RAM */ |
2262 | if (!ram_bytes_total()) { | |
2263 | return pages; | |
2264 | } | |
2265 | ||
6f37bb8b | 2266 | pss.block = rs->last_seen_block; |
a935e30f | 2267 | pss.page = rs->last_page; |
b8fb8cb7 DDAG |
2268 | pss.complete_round = false; |
2269 | ||
2270 | if (!pss.block) { | |
2271 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
2272 | } | |
56e93d26 | 2273 | |
b9e60928 | 2274 | do { |
a82d593b | 2275 | again = true; |
f20e2865 | 2276 | found = get_queued_page(rs, &pss); |
b9e60928 | 2277 | |
a82d593b DDAG |
2278 | if (!found) { |
2279 | /* priority queue empty, so just search for something dirty */ | |
f20e2865 | 2280 | found = find_dirty_block(rs, &pss, &again); |
a82d593b | 2281 | } |
f3f491fc | 2282 | |
a82d593b | 2283 | if (found) { |
05931ec5 | 2284 | pages = ram_save_host_page(rs, &pss); |
56e93d26 | 2285 | } |
b9e60928 | 2286 | } while (!pages && again); |
56e93d26 | 2287 | |
6f37bb8b | 2288 | rs->last_seen_block = pss.block; |
a935e30f | 2289 | rs->last_page = pss.page; |
56e93d26 JQ |
2290 | |
2291 | return pages; | |
2292 | } | |
2293 | ||
2294 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
2295 | { | |
2296 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 2297 | |
56e93d26 | 2298 | if (zero) { |
9360447d | 2299 | ram_counters.duplicate += pages; |
56e93d26 | 2300 | } else { |
9360447d | 2301 | ram_counters.normal += pages; |
4c2d0f6d | 2302 | ram_transferred_add(size); |
56e93d26 JQ |
2303 | qemu_update_position(f, size); |
2304 | } | |
2305 | } | |
2306 | ||
fbd162e6 | 2307 | static uint64_t ram_bytes_total_common(bool count_ignored) |
56e93d26 JQ |
2308 | { |
2309 | RAMBlock *block; | |
2310 | uint64_t total = 0; | |
2311 | ||
89ac5a1d DDAG |
2312 | RCU_READ_LOCK_GUARD(); |
2313 | ||
fbd162e6 YK |
2314 | if (count_ignored) { |
2315 | RAMBLOCK_FOREACH_MIGRATABLE(block) { | |
2316 | total += block->used_length; | |
2317 | } | |
2318 | } else { | |
2319 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2320 | total += block->used_length; | |
2321 | } | |
99e15582 | 2322 | } |
56e93d26 JQ |
2323 | return total; |
2324 | } | |
2325 | ||
fbd162e6 YK |
2326 | uint64_t ram_bytes_total(void) |
2327 | { | |
2328 | return ram_bytes_total_common(false); | |
2329 | } | |
2330 | ||
f265e0e4 | 2331 | static void xbzrle_load_setup(void) |
56e93d26 | 2332 | { |
f265e0e4 | 2333 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2334 | } |
2335 | ||
f265e0e4 JQ |
2336 | static void xbzrle_load_cleanup(void) |
2337 | { | |
2338 | g_free(XBZRLE.decoded_buf); | |
2339 | XBZRLE.decoded_buf = NULL; | |
2340 | } | |
2341 | ||
7d7c96be PX |
2342 | static void ram_state_cleanup(RAMState **rsp) |
2343 | { | |
b9ccaf6d DDAG |
2344 | if (*rsp) { |
2345 | migration_page_queue_free(*rsp); | |
2346 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
2347 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
2348 | g_free(*rsp); | |
2349 | *rsp = NULL; | |
2350 | } | |
7d7c96be PX |
2351 | } |
2352 | ||
84593a08 PX |
2353 | static void xbzrle_cleanup(void) |
2354 | { | |
2355 | XBZRLE_cache_lock(); | |
2356 | if (XBZRLE.cache) { | |
2357 | cache_fini(XBZRLE.cache); | |
2358 | g_free(XBZRLE.encoded_buf); | |
2359 | g_free(XBZRLE.current_buf); | |
2360 | g_free(XBZRLE.zero_target_page); | |
2361 | XBZRLE.cache = NULL; | |
2362 | XBZRLE.encoded_buf = NULL; | |
2363 | XBZRLE.current_buf = NULL; | |
2364 | XBZRLE.zero_target_page = NULL; | |
2365 | } | |
2366 | XBZRLE_cache_unlock(); | |
2367 | } | |
2368 | ||
f265e0e4 | 2369 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 2370 | { |
53518d94 | 2371 | RAMState **rsp = opaque; |
6b6712ef | 2372 | RAMBlock *block; |
eb859c53 | 2373 | |
278e2f55 AG |
2374 | /* We don't use dirty log with background snapshots */ |
2375 | if (!migrate_background_snapshot()) { | |
2376 | /* caller have hold iothread lock or is in a bh, so there is | |
2377 | * no writing race against the migration bitmap | |
2378 | */ | |
63b41db4 HH |
2379 | if (global_dirty_tracking & GLOBAL_DIRTY_MIGRATION) { |
2380 | /* | |
2381 | * do not stop dirty log without starting it, since | |
2382 | * memory_global_dirty_log_stop will assert that | |
2383 | * memory_global_dirty_log_start/stop used in pairs | |
2384 | */ | |
2385 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); | |
2386 | } | |
278e2f55 | 2387 | } |
6b6712ef | 2388 | |
fbd162e6 | 2389 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
2390 | g_free(block->clear_bmap); |
2391 | block->clear_bmap = NULL; | |
6b6712ef JQ |
2392 | g_free(block->bmap); |
2393 | block->bmap = NULL; | |
56e93d26 JQ |
2394 | } |
2395 | ||
84593a08 | 2396 | xbzrle_cleanup(); |
f0afa331 | 2397 | compress_threads_save_cleanup(); |
7d7c96be | 2398 | ram_state_cleanup(rsp); |
56e93d26 JQ |
2399 | } |
2400 | ||
6f37bb8b | 2401 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 2402 | { |
6f37bb8b JQ |
2403 | rs->last_seen_block = NULL; |
2404 | rs->last_sent_block = NULL; | |
269ace29 | 2405 | rs->last_page = 0; |
6f37bb8b | 2406 | rs->last_version = ram_list.version; |
1a373522 | 2407 | rs->xbzrle_enabled = false; |
56e93d26 JQ |
2408 | } |
2409 | ||
2410 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
2411 | ||
e0b266f0 DDAG |
2412 | /* **** functions for postcopy ***** */ |
2413 | ||
ced1c616 PB |
2414 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
2415 | { | |
2416 | struct RAMBlock *block; | |
ced1c616 | 2417 | |
fbd162e6 | 2418 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2419 | unsigned long *bitmap = block->bmap; |
2420 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
2421 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
2422 | |
2423 | while (run_start < range) { | |
2424 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
2425 | ram_discard_range(block->idstr, |
2426 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
2427 | ((ram_addr_t)(run_end - run_start)) | |
2428 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
2429 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
2430 | } | |
2431 | } | |
2432 | } | |
2433 | ||
3d0684b2 JQ |
2434 | /** |
2435 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
2436 | * | |
e0b266f0 | 2437 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
2438 | * |
2439 | * @ms: current migration state | |
89dab31b | 2440 | * @block: RAMBlock to discard |
e0b266f0 | 2441 | */ |
9e7d1223 | 2442 | static void postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 2443 | { |
6b6712ef | 2444 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 2445 | unsigned long current; |
1e7cf8c3 | 2446 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 2447 | |
6b6712ef | 2448 | for (current = 0; current < end; ) { |
1e7cf8c3 | 2449 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 2450 | unsigned long zero, discard_length; |
e0b266f0 | 2451 | |
33a5cb62 WY |
2452 | if (one >= end) { |
2453 | break; | |
2454 | } | |
e0b266f0 | 2455 | |
1e7cf8c3 | 2456 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
2457 | |
2458 | if (zero >= end) { | |
2459 | discard_length = end - one; | |
e0b266f0 | 2460 | } else { |
33a5cb62 WY |
2461 | discard_length = zero - one; |
2462 | } | |
810cf2bb | 2463 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 2464 | current = one + discard_length; |
e0b266f0 | 2465 | } |
e0b266f0 DDAG |
2466 | } |
2467 | ||
f30c2e5b PX |
2468 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block); |
2469 | ||
3d0684b2 JQ |
2470 | /** |
2471 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
2472 | * | |
e0b266f0 DDAG |
2473 | * Utility for the outgoing postcopy code. |
2474 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
2475 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
2476 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
2477 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
2478 | * |
2479 | * @ms: current migration state | |
e0b266f0 | 2480 | */ |
739fcc1b | 2481 | static void postcopy_each_ram_send_discard(MigrationState *ms) |
e0b266f0 DDAG |
2482 | { |
2483 | struct RAMBlock *block; | |
e0b266f0 | 2484 | |
fbd162e6 | 2485 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 2486 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 | 2487 | |
f30c2e5b PX |
2488 | /* |
2489 | * Deal with TPS != HPS and huge pages. It discard any partially sent | |
2490 | * host-page size chunks, mark any partially dirty host-page size | |
2491 | * chunks as all dirty. In this case the host-page is the host-page | |
2492 | * for the particular RAMBlock, i.e. it might be a huge page. | |
2493 | */ | |
2494 | postcopy_chunk_hostpages_pass(ms, block); | |
2495 | ||
e0b266f0 DDAG |
2496 | /* |
2497 | * Postcopy sends chunks of bitmap over the wire, but it | |
2498 | * just needs indexes at this point, avoids it having | |
2499 | * target page specific code. | |
2500 | */ | |
739fcc1b | 2501 | postcopy_send_discard_bm_ram(ms, block); |
810cf2bb | 2502 | postcopy_discard_send_finish(ms); |
e0b266f0 | 2503 | } |
e0b266f0 DDAG |
2504 | } |
2505 | ||
3d0684b2 | 2506 | /** |
8324ef86 | 2507 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2508 | * |
2509 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2510 | * canonicalize the two bitmaps, that are similar, but one is | |
2511 | * inverted. | |
99e314eb | 2512 | * |
3d0684b2 JQ |
2513 | * Postcopy requires that all target pages in a hostpage are dirty or |
2514 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2515 | * |
3d0684b2 | 2516 | * @ms: current migration state |
3d0684b2 | 2517 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2518 | */ |
1e7cf8c3 | 2519 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2520 | { |
53518d94 | 2521 | RAMState *rs = ram_state; |
6b6712ef | 2522 | unsigned long *bitmap = block->bmap; |
29c59172 | 2523 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2524 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2525 | unsigned long run_start; |
2526 | ||
29c59172 DDAG |
2527 | if (block->page_size == TARGET_PAGE_SIZE) { |
2528 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2529 | return; | |
2530 | } | |
2531 | ||
1e7cf8c3 WY |
2532 | /* Find a dirty page */ |
2533 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2534 | |
6b6712ef | 2535 | while (run_start < pages) { |
99e314eb DDAG |
2536 | |
2537 | /* | |
2538 | * If the start of this run of pages is in the middle of a host | |
2539 | * page, then we need to fixup this host page. | |
2540 | */ | |
9dec3cc3 | 2541 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2542 | /* Find the end of this run */ |
1e7cf8c3 | 2543 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2544 | /* |
2545 | * If the end isn't at the start of a host page, then the | |
2546 | * run doesn't finish at the end of a host page | |
2547 | * and we need to discard. | |
2548 | */ | |
99e314eb DDAG |
2549 | } |
2550 | ||
9dec3cc3 | 2551 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2552 | unsigned long page; |
dad45ab2 WY |
2553 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2554 | host_ratio); | |
2555 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2556 | |
99e314eb DDAG |
2557 | /* Clean up the bitmap */ |
2558 | for (page = fixup_start_addr; | |
2559 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2560 | /* |
2561 | * Remark them as dirty, updating the count for any pages | |
2562 | * that weren't previously dirty. | |
2563 | */ | |
0d8ec885 | 2564 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2565 | } |
2566 | } | |
2567 | ||
1e7cf8c3 WY |
2568 | /* Find the next dirty page for the next iteration */ |
2569 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2570 | } |
2571 | } | |
2572 | ||
3d0684b2 JQ |
2573 | /** |
2574 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2575 | * | |
e0b266f0 DDAG |
2576 | * Transmit the set of pages to be discarded after precopy to the target |
2577 | * these are pages that: | |
2578 | * a) Have been previously transmitted but are now dirty again | |
2579 | * b) Pages that have never been transmitted, this ensures that | |
2580 | * any pages on the destination that have been mapped by background | |
2581 | * tasks get discarded (transparent huge pages is the specific concern) | |
2582 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2583 | * |
2584 | * @ms: current migration state | |
e0b266f0 | 2585 | */ |
739fcc1b | 2586 | void ram_postcopy_send_discard_bitmap(MigrationState *ms) |
e0b266f0 | 2587 | { |
53518d94 | 2588 | RAMState *rs = ram_state; |
e0b266f0 | 2589 | |
89ac5a1d | 2590 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2591 | |
2592 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 2593 | migration_bitmap_sync(rs); |
e0b266f0 | 2594 | |
6b6712ef JQ |
2595 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
2596 | rs->last_seen_block = NULL; | |
2597 | rs->last_sent_block = NULL; | |
2598 | rs->last_page = 0; | |
e0b266f0 | 2599 | |
739fcc1b | 2600 | postcopy_each_ram_send_discard(ms); |
e0b266f0 | 2601 | |
739fcc1b | 2602 | trace_ram_postcopy_send_discard_bitmap(); |
e0b266f0 DDAG |
2603 | } |
2604 | ||
3d0684b2 JQ |
2605 | /** |
2606 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2607 | * |
3d0684b2 | 2608 | * Returns zero on success |
e0b266f0 | 2609 | * |
36449157 JQ |
2610 | * @rbname: name of the RAMBlock of the request. NULL means the |
2611 | * same that last one. | |
3d0684b2 JQ |
2612 | * @start: RAMBlock starting page |
2613 | * @length: RAMBlock size | |
e0b266f0 | 2614 | */ |
aaa2064c | 2615 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 2616 | { |
36449157 | 2617 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2618 | |
89ac5a1d | 2619 | RCU_READ_LOCK_GUARD(); |
36449157 | 2620 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2621 | |
2622 | if (!rb) { | |
36449157 | 2623 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 2624 | return -1; |
e0b266f0 DDAG |
2625 | } |
2626 | ||
814bb08f PX |
2627 | /* |
2628 | * On source VM, we don't need to update the received bitmap since | |
2629 | * we don't even have one. | |
2630 | */ | |
2631 | if (rb->receivedmap) { | |
2632 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
2633 | length >> qemu_target_page_bits()); | |
2634 | } | |
2635 | ||
03acb4e9 | 2636 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2637 | } |
2638 | ||
84593a08 PX |
2639 | /* |
2640 | * For every allocation, we will try not to crash the VM if the | |
2641 | * allocation failed. | |
2642 | */ | |
2643 | static int xbzrle_init(void) | |
2644 | { | |
2645 | Error *local_err = NULL; | |
2646 | ||
2647 | if (!migrate_use_xbzrle()) { | |
2648 | return 0; | |
2649 | } | |
2650 | ||
2651 | XBZRLE_cache_lock(); | |
2652 | ||
2653 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
2654 | if (!XBZRLE.zero_target_page) { | |
2655 | error_report("%s: Error allocating zero page", __func__); | |
2656 | goto err_out; | |
2657 | } | |
2658 | ||
2659 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
2660 | TARGET_PAGE_SIZE, &local_err); | |
2661 | if (!XBZRLE.cache) { | |
2662 | error_report_err(local_err); | |
2663 | goto free_zero_page; | |
2664 | } | |
2665 | ||
2666 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2667 | if (!XBZRLE.encoded_buf) { | |
2668 | error_report("%s: Error allocating encoded_buf", __func__); | |
2669 | goto free_cache; | |
2670 | } | |
2671 | ||
2672 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2673 | if (!XBZRLE.current_buf) { | |
2674 | error_report("%s: Error allocating current_buf", __func__); | |
2675 | goto free_encoded_buf; | |
2676 | } | |
2677 | ||
2678 | /* We are all good */ | |
2679 | XBZRLE_cache_unlock(); | |
2680 | return 0; | |
2681 | ||
2682 | free_encoded_buf: | |
2683 | g_free(XBZRLE.encoded_buf); | |
2684 | XBZRLE.encoded_buf = NULL; | |
2685 | free_cache: | |
2686 | cache_fini(XBZRLE.cache); | |
2687 | XBZRLE.cache = NULL; | |
2688 | free_zero_page: | |
2689 | g_free(XBZRLE.zero_target_page); | |
2690 | XBZRLE.zero_target_page = NULL; | |
2691 | err_out: | |
2692 | XBZRLE_cache_unlock(); | |
2693 | return -ENOMEM; | |
2694 | } | |
2695 | ||
53518d94 | 2696 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2697 | { |
7d00ee6a PX |
2698 | *rsp = g_try_new0(RAMState, 1); |
2699 | ||
2700 | if (!*rsp) { | |
2701 | error_report("%s: Init ramstate fail", __func__); | |
2702 | return -1; | |
2703 | } | |
53518d94 JQ |
2704 | |
2705 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
2706 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
2707 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
56e93d26 | 2708 | |
7d00ee6a | 2709 | /* |
40c4d4a8 IR |
2710 | * Count the total number of pages used by ram blocks not including any |
2711 | * gaps due to alignment or unplugs. | |
03158519 | 2712 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 2713 | */ |
40c4d4a8 | 2714 | (*rsp)->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
7d00ee6a PX |
2715 | ram_state_reset(*rsp); |
2716 | ||
2717 | return 0; | |
2718 | } | |
2719 | ||
d6eff5d7 | 2720 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 2721 | { |
002cad6b | 2722 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
2723 | RAMBlock *block; |
2724 | unsigned long pages; | |
002cad6b | 2725 | uint8_t shift; |
56e93d26 | 2726 | |
0827b9e9 AA |
2727 | /* Skip setting bitmap if there is no RAM */ |
2728 | if (ram_bytes_total()) { | |
002cad6b PX |
2729 | shift = ms->clear_bitmap_shift; |
2730 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
2731 | error_report("clear_bitmap_shift (%u) too big, using " | |
2732 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
2733 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
2734 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
2735 | error_report("clear_bitmap_shift (%u) too small, using " | |
2736 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
2737 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
2738 | } | |
2739 | ||
fbd162e6 | 2740 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 2741 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
2742 | /* |
2743 | * The initial dirty bitmap for migration must be set with all | |
2744 | * ones to make sure we'll migrate every guest RAM page to | |
2745 | * destination. | |
40c4d4a8 IR |
2746 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
2747 | * new migration after a failed migration, ram_list. | |
2748 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
2749 | * guest memory. | |
03158519 | 2750 | */ |
6b6712ef | 2751 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 2752 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
2753 | block->clear_bmap_shift = shift; |
2754 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 2755 | } |
f3f491fc | 2756 | } |
d6eff5d7 PX |
2757 | } |
2758 | ||
be39b4cd DH |
2759 | static void migration_bitmap_clear_discarded_pages(RAMState *rs) |
2760 | { | |
2761 | unsigned long pages; | |
2762 | RAMBlock *rb; | |
2763 | ||
2764 | RCU_READ_LOCK_GUARD(); | |
2765 | ||
2766 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { | |
2767 | pages = ramblock_dirty_bitmap_clear_discarded_pages(rb); | |
2768 | rs->migration_dirty_pages -= pages; | |
2769 | } | |
2770 | } | |
2771 | ||
d6eff5d7 PX |
2772 | static void ram_init_bitmaps(RAMState *rs) |
2773 | { | |
2774 | /* For memory_global_dirty_log_start below. */ | |
2775 | qemu_mutex_lock_iothread(); | |
2776 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 2777 | |
89ac5a1d DDAG |
2778 | WITH_RCU_READ_LOCK_GUARD() { |
2779 | ram_list_init_bitmaps(); | |
278e2f55 AG |
2780 | /* We don't use dirty log with background snapshots */ |
2781 | if (!migrate_background_snapshot()) { | |
63b41db4 | 2782 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
278e2f55 AG |
2783 | migration_bitmap_sync_precopy(rs); |
2784 | } | |
89ac5a1d | 2785 | } |
56e93d26 | 2786 | qemu_mutex_unlock_ramlist(); |
49877834 | 2787 | qemu_mutex_unlock_iothread(); |
be39b4cd DH |
2788 | |
2789 | /* | |
2790 | * After an eventual first bitmap sync, fixup the initial bitmap | |
2791 | * containing all 1s to exclude any discarded pages from migration. | |
2792 | */ | |
2793 | migration_bitmap_clear_discarded_pages(rs); | |
d6eff5d7 PX |
2794 | } |
2795 | ||
2796 | static int ram_init_all(RAMState **rsp) | |
2797 | { | |
2798 | if (ram_state_init(rsp)) { | |
2799 | return -1; | |
2800 | } | |
2801 | ||
2802 | if (xbzrle_init()) { | |
2803 | ram_state_cleanup(rsp); | |
2804 | return -1; | |
2805 | } | |
2806 | ||
2807 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
2808 | |
2809 | return 0; | |
2810 | } | |
2811 | ||
08614f34 PX |
2812 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
2813 | { | |
2814 | RAMBlock *block; | |
2815 | uint64_t pages = 0; | |
2816 | ||
2817 | /* | |
2818 | * Postcopy is not using xbzrle/compression, so no need for that. | |
2819 | * Also, since source are already halted, we don't need to care | |
2820 | * about dirty page logging as well. | |
2821 | */ | |
2822 | ||
fbd162e6 | 2823 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
2824 | pages += bitmap_count_one(block->bmap, |
2825 | block->used_length >> TARGET_PAGE_BITS); | |
2826 | } | |
2827 | ||
2828 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
2829 | rs->migration_dirty_pages = pages; | |
2830 | ||
1a373522 | 2831 | ram_state_reset(rs); |
08614f34 PX |
2832 | |
2833 | /* Update RAMState cache of output QEMUFile */ | |
2834 | rs->f = out; | |
2835 | ||
2836 | trace_ram_state_resume_prepare(pages); | |
2837 | } | |
2838 | ||
6bcb05fc WW |
2839 | /* |
2840 | * This function clears bits of the free pages reported by the caller from the | |
2841 | * migration dirty bitmap. @addr is the host address corresponding to the | |
2842 | * start of the continuous guest free pages, and @len is the total bytes of | |
2843 | * those pages. | |
2844 | */ | |
2845 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
2846 | { | |
2847 | RAMBlock *block; | |
2848 | ram_addr_t offset; | |
2849 | size_t used_len, start, npages; | |
2850 | MigrationState *s = migrate_get_current(); | |
2851 | ||
2852 | /* This function is currently expected to be used during live migration */ | |
2853 | if (!migration_is_setup_or_active(s->state)) { | |
2854 | return; | |
2855 | } | |
2856 | ||
2857 | for (; len > 0; len -= used_len, addr += used_len) { | |
2858 | block = qemu_ram_block_from_host(addr, false, &offset); | |
2859 | if (unlikely(!block || offset >= block->used_length)) { | |
2860 | /* | |
2861 | * The implementation might not support RAMBlock resize during | |
2862 | * live migration, but it could happen in theory with future | |
2863 | * updates. So we add a check here to capture that case. | |
2864 | */ | |
2865 | error_report_once("%s unexpected error", __func__); | |
2866 | return; | |
2867 | } | |
2868 | ||
2869 | if (len <= block->used_length - offset) { | |
2870 | used_len = len; | |
2871 | } else { | |
2872 | used_len = block->used_length - offset; | |
2873 | } | |
2874 | ||
2875 | start = offset >> TARGET_PAGE_BITS; | |
2876 | npages = used_len >> TARGET_PAGE_BITS; | |
2877 | ||
2878 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3143577d WW |
2879 | /* |
2880 | * The skipped free pages are equavalent to be sent from clear_bmap's | |
2881 | * perspective, so clear the bits from the memory region bitmap which | |
2882 | * are initially set. Otherwise those skipped pages will be sent in | |
2883 | * the next round after syncing from the memory region bitmap. | |
2884 | */ | |
1230a25f | 2885 | migration_clear_memory_region_dirty_bitmap_range(block, start, npages); |
6bcb05fc WW |
2886 | ram_state->migration_dirty_pages -= |
2887 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
2888 | bitmap_clear(block->bmap, start, npages); | |
2889 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
2890 | } | |
2891 | } | |
2892 | ||
3d0684b2 JQ |
2893 | /* |
2894 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2895 | * long-running RCU critical section. When rcu-reclaims in the code |
2896 | * start to become numerous it will be necessary to reduce the | |
2897 | * granularity of these critical sections. | |
2898 | */ | |
2899 | ||
3d0684b2 JQ |
2900 | /** |
2901 | * ram_save_setup: Setup RAM for migration | |
2902 | * | |
2903 | * Returns zero to indicate success and negative for error | |
2904 | * | |
2905 | * @f: QEMUFile where to send the data | |
2906 | * @opaque: RAMState pointer | |
2907 | */ | |
a91246c9 HZ |
2908 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2909 | { | |
53518d94 | 2910 | RAMState **rsp = opaque; |
a91246c9 HZ |
2911 | RAMBlock *block; |
2912 | ||
dcaf446e XG |
2913 | if (compress_threads_save_setup()) { |
2914 | return -1; | |
2915 | } | |
2916 | ||
a91246c9 HZ |
2917 | /* migration has already setup the bitmap, reuse it. */ |
2918 | if (!migration_in_colo_state()) { | |
7d00ee6a | 2919 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 2920 | compress_threads_save_cleanup(); |
a91246c9 | 2921 | return -1; |
53518d94 | 2922 | } |
a91246c9 | 2923 | } |
53518d94 | 2924 | (*rsp)->f = f; |
a91246c9 | 2925 | |
0e6ebd48 DDAG |
2926 | WITH_RCU_READ_LOCK_GUARD() { |
2927 | qemu_put_be64(f, ram_bytes_total_common(true) | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 2928 | |
0e6ebd48 DDAG |
2929 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
2930 | qemu_put_byte(f, strlen(block->idstr)); | |
2931 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2932 | qemu_put_be64(f, block->used_length); | |
2933 | if (migrate_postcopy_ram() && block->page_size != | |
2934 | qemu_host_page_size) { | |
2935 | qemu_put_be64(f, block->page_size); | |
2936 | } | |
2937 | if (migrate_ignore_shared()) { | |
2938 | qemu_put_be64(f, block->mr->addr); | |
2939 | } | |
fbd162e6 | 2940 | } |
56e93d26 JQ |
2941 | } |
2942 | ||
56e93d26 JQ |
2943 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
2944 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2945 | ||
99f2c6fb | 2946 | multifd_send_sync_main(f); |
56e93d26 | 2947 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 2948 | qemu_fflush(f); |
56e93d26 JQ |
2949 | |
2950 | return 0; | |
2951 | } | |
2952 | ||
3d0684b2 JQ |
2953 | /** |
2954 | * ram_save_iterate: iterative stage for migration | |
2955 | * | |
2956 | * Returns zero to indicate success and negative for error | |
2957 | * | |
2958 | * @f: QEMUFile where to send the data | |
2959 | * @opaque: RAMState pointer | |
2960 | */ | |
56e93d26 JQ |
2961 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2962 | { | |
53518d94 JQ |
2963 | RAMState **temp = opaque; |
2964 | RAMState *rs = *temp; | |
3d4095b2 | 2965 | int ret = 0; |
56e93d26 JQ |
2966 | int i; |
2967 | int64_t t0; | |
5c90308f | 2968 | int done = 0; |
56e93d26 | 2969 | |
b2557345 PL |
2970 | if (blk_mig_bulk_active()) { |
2971 | /* Avoid transferring ram during bulk phase of block migration as | |
2972 | * the bulk phase will usually take a long time and transferring | |
2973 | * ram updates during that time is pointless. */ | |
2974 | goto out; | |
2975 | } | |
2976 | ||
63268c49 PX |
2977 | /* |
2978 | * We'll take this lock a little bit long, but it's okay for two reasons. | |
2979 | * Firstly, the only possible other thread to take it is who calls | |
2980 | * qemu_guest_free_page_hint(), which should be rare; secondly, see | |
2981 | * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which | |
2982 | * guarantees that we'll at least released it in a regular basis. | |
2983 | */ | |
2984 | qemu_mutex_lock(&rs->bitmap_mutex); | |
89ac5a1d DDAG |
2985 | WITH_RCU_READ_LOCK_GUARD() { |
2986 | if (ram_list.version != rs->last_version) { | |
2987 | ram_state_reset(rs); | |
2988 | } | |
56e93d26 | 2989 | |
89ac5a1d DDAG |
2990 | /* Read version before ram_list.blocks */ |
2991 | smp_rmb(); | |
56e93d26 | 2992 | |
89ac5a1d | 2993 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 2994 | |
89ac5a1d DDAG |
2995 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
2996 | i = 0; | |
2997 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
a1fe28df | 2998 | postcopy_has_request(rs)) { |
89ac5a1d | 2999 | int pages; |
e03a34f8 | 3000 | |
89ac5a1d DDAG |
3001 | if (qemu_file_get_error(f)) { |
3002 | break; | |
3003 | } | |
e8f3735f | 3004 | |
05931ec5 | 3005 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3006 | /* no more pages to sent */ |
3007 | if (pages == 0) { | |
3008 | done = 1; | |
3009 | break; | |
3010 | } | |
e8f3735f | 3011 | |
89ac5a1d DDAG |
3012 | if (pages < 0) { |
3013 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
3014 | break; |
3015 | } | |
89ac5a1d DDAG |
3016 | |
3017 | rs->target_page_count += pages; | |
3018 | ||
644acf99 WY |
3019 | /* |
3020 | * During postcopy, it is necessary to make sure one whole host | |
3021 | * page is sent in one chunk. | |
3022 | */ | |
3023 | if (migrate_postcopy_ram()) { | |
3024 | flush_compressed_data(rs); | |
3025 | } | |
3026 | ||
89ac5a1d DDAG |
3027 | /* |
3028 | * we want to check in the 1st loop, just in case it was the 1st | |
3029 | * time and we had to sync the dirty bitmap. | |
3030 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
3031 | * some iterations | |
3032 | */ | |
3033 | if ((i & 63) == 0) { | |
3034 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
3035 | 1000000; | |
3036 | if (t1 > MAX_WAIT) { | |
3037 | trace_ram_save_iterate_big_wait(t1, i); | |
3038 | break; | |
3039 | } | |
3040 | } | |
3041 | i++; | |
56e93d26 | 3042 | } |
56e93d26 | 3043 | } |
63268c49 | 3044 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 JQ |
3045 | |
3046 | /* | |
3047 | * Must occur before EOS (or any QEMUFile operation) | |
3048 | * because of RDMA protocol. | |
3049 | */ | |
3050 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
3051 | ||
b2557345 | 3052 | out: |
b69a0227 JQ |
3053 | if (ret >= 0 |
3054 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
99f2c6fb | 3055 | multifd_send_sync_main(rs->f); |
3d4095b2 JQ |
3056 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3057 | qemu_fflush(f); | |
4c2d0f6d | 3058 | ram_transferred_add(8); |
56e93d26 | 3059 | |
3d4095b2 JQ |
3060 | ret = qemu_file_get_error(f); |
3061 | } | |
56e93d26 JQ |
3062 | if (ret < 0) { |
3063 | return ret; | |
3064 | } | |
3065 | ||
5c90308f | 3066 | return done; |
56e93d26 JQ |
3067 | } |
3068 | ||
3d0684b2 JQ |
3069 | /** |
3070 | * ram_save_complete: function called to send the remaining amount of ram | |
3071 | * | |
e8f3735f | 3072 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
3073 | * |
3074 | * Called with iothread lock | |
3075 | * | |
3076 | * @f: QEMUFile where to send the data | |
3077 | * @opaque: RAMState pointer | |
3078 | */ | |
56e93d26 JQ |
3079 | static int ram_save_complete(QEMUFile *f, void *opaque) |
3080 | { | |
53518d94 JQ |
3081 | RAMState **temp = opaque; |
3082 | RAMState *rs = *temp; | |
e8f3735f | 3083 | int ret = 0; |
6f37bb8b | 3084 | |
05931ec5 JQ |
3085 | rs->last_stage = !migration_in_colo_state(); |
3086 | ||
89ac5a1d DDAG |
3087 | WITH_RCU_READ_LOCK_GUARD() { |
3088 | if (!migration_in_postcopy()) { | |
3089 | migration_bitmap_sync_precopy(rs); | |
3090 | } | |
56e93d26 | 3091 | |
89ac5a1d | 3092 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 3093 | |
89ac5a1d | 3094 | /* try transferring iterative blocks of memory */ |
56e93d26 | 3095 | |
89ac5a1d DDAG |
3096 | /* flush all remaining blocks regardless of rate limiting */ |
3097 | while (true) { | |
3098 | int pages; | |
56e93d26 | 3099 | |
05931ec5 | 3100 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3101 | /* no more blocks to sent */ |
3102 | if (pages == 0) { | |
3103 | break; | |
3104 | } | |
3105 | if (pages < 0) { | |
3106 | ret = pages; | |
3107 | break; | |
3108 | } | |
e8f3735f | 3109 | } |
56e93d26 | 3110 | |
89ac5a1d DDAG |
3111 | flush_compressed_data(rs); |
3112 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); | |
3113 | } | |
d09a6fde | 3114 | |
3d4095b2 | 3115 | if (ret >= 0) { |
99f2c6fb | 3116 | multifd_send_sync_main(rs->f); |
3d4095b2 JQ |
3117 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3118 | qemu_fflush(f); | |
3119 | } | |
56e93d26 | 3120 | |
e8f3735f | 3121 | return ret; |
56e93d26 JQ |
3122 | } |
3123 | ||
c31b098f | 3124 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
47995026 VSO |
3125 | uint64_t *res_precopy_only, |
3126 | uint64_t *res_compatible, | |
3127 | uint64_t *res_postcopy_only) | |
56e93d26 | 3128 | { |
53518d94 JQ |
3129 | RAMState **temp = opaque; |
3130 | RAMState *rs = *temp; | |
56e93d26 JQ |
3131 | uint64_t remaining_size; |
3132 | ||
9edabd4d | 3133 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3134 | |
5727309d | 3135 | if (!migration_in_postcopy() && |
663e6c1d | 3136 | remaining_size < max_size) { |
56e93d26 | 3137 | qemu_mutex_lock_iothread(); |
89ac5a1d DDAG |
3138 | WITH_RCU_READ_LOCK_GUARD() { |
3139 | migration_bitmap_sync_precopy(rs); | |
3140 | } | |
56e93d26 | 3141 | qemu_mutex_unlock_iothread(); |
9edabd4d | 3142 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3143 | } |
c31b098f | 3144 | |
86e1167e VSO |
3145 | if (migrate_postcopy_ram()) { |
3146 | /* We can do postcopy, and all the data is postcopiable */ | |
47995026 | 3147 | *res_compatible += remaining_size; |
86e1167e | 3148 | } else { |
47995026 | 3149 | *res_precopy_only += remaining_size; |
86e1167e | 3150 | } |
56e93d26 JQ |
3151 | } |
3152 | ||
3153 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
3154 | { | |
3155 | unsigned int xh_len; | |
3156 | int xh_flags; | |
063e760a | 3157 | uint8_t *loaded_data; |
56e93d26 | 3158 | |
56e93d26 JQ |
3159 | /* extract RLE header */ |
3160 | xh_flags = qemu_get_byte(f); | |
3161 | xh_len = qemu_get_be16(f); | |
3162 | ||
3163 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
3164 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
3165 | return -1; | |
3166 | } | |
3167 | ||
3168 | if (xh_len > TARGET_PAGE_SIZE) { | |
3169 | error_report("Failed to load XBZRLE page - len overflow!"); | |
3170 | return -1; | |
3171 | } | |
f265e0e4 | 3172 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 3173 | /* load data and decode */ |
f265e0e4 | 3174 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 3175 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
3176 | |
3177 | /* decode RLE */ | |
063e760a | 3178 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
3179 | TARGET_PAGE_SIZE) == -1) { |
3180 | error_report("Failed to load XBZRLE page - decode error!"); | |
3181 | return -1; | |
3182 | } | |
3183 | ||
3184 | return 0; | |
3185 | } | |
3186 | ||
3d0684b2 JQ |
3187 | /** |
3188 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
3189 | * | |
3190 | * Must be called from within a rcu critical section. | |
3191 | * | |
56e93d26 | 3192 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 3193 | * |
755e8d7c | 3194 | * @mis: the migration incoming state pointer |
3d0684b2 JQ |
3195 | * @f: QEMUFile where to read the data from |
3196 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 3197 | */ |
755e8d7c PX |
3198 | static inline RAMBlock *ram_block_from_stream(MigrationIncomingState *mis, |
3199 | QEMUFile *f, int flags) | |
56e93d26 | 3200 | { |
755e8d7c | 3201 | RAMBlock *block = mis->last_recv_block; |
56e93d26 JQ |
3202 | char id[256]; |
3203 | uint8_t len; | |
3204 | ||
3205 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 3206 | if (!block) { |
56e93d26 JQ |
3207 | error_report("Ack, bad migration stream!"); |
3208 | return NULL; | |
3209 | } | |
4c4bad48 | 3210 | return block; |
56e93d26 JQ |
3211 | } |
3212 | ||
3213 | len = qemu_get_byte(f); | |
3214 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3215 | id[len] = 0; | |
3216 | ||
e3dd7493 | 3217 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
3218 | if (!block) { |
3219 | error_report("Can't find block %s", id); | |
3220 | return NULL; | |
56e93d26 JQ |
3221 | } |
3222 | ||
fbd162e6 | 3223 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
3224 | error_report("block %s should not be migrated !", id); |
3225 | return NULL; | |
3226 | } | |
3227 | ||
755e8d7c PX |
3228 | mis->last_recv_block = block; |
3229 | ||
4c4bad48 HZ |
3230 | return block; |
3231 | } | |
3232 | ||
3233 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
3234 | ram_addr_t offset) | |
3235 | { | |
3236 | if (!offset_in_ramblock(block, offset)) { | |
3237 | return NULL; | |
3238 | } | |
3239 | ||
3240 | return block->host + offset; | |
56e93d26 JQ |
3241 | } |
3242 | ||
6a23f639 DH |
3243 | static void *host_page_from_ram_block_offset(RAMBlock *block, |
3244 | ram_addr_t offset) | |
3245 | { | |
3246 | /* Note: Explicitly no check against offset_in_ramblock(). */ | |
3247 | return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block->host + offset), | |
3248 | block->page_size); | |
3249 | } | |
3250 | ||
3251 | static ram_addr_t host_page_offset_from_ram_block_offset(RAMBlock *block, | |
3252 | ram_addr_t offset) | |
3253 | { | |
3254 | return ((uintptr_t)block->host + offset) & (block->page_size - 1); | |
3255 | } | |
3256 | ||
13af18f2 | 3257 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 3258 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
3259 | { |
3260 | if (!offset_in_ramblock(block, offset)) { | |
3261 | return NULL; | |
3262 | } | |
3263 | if (!block->colo_cache) { | |
3264 | error_report("%s: colo_cache is NULL in block :%s", | |
3265 | __func__, block->idstr); | |
3266 | return NULL; | |
3267 | } | |
7d9acafa ZC |
3268 | |
3269 | /* | |
3270 | * During colo checkpoint, we need bitmap of these migrated pages. | |
3271 | * It help us to decide which pages in ram cache should be flushed | |
3272 | * into VM's RAM later. | |
3273 | */ | |
8af66371 HZ |
3274 | if (record_bitmap && |
3275 | !test_and_set_bit(offset >> TARGET_PAGE_BITS, block->bmap)) { | |
7d9acafa ZC |
3276 | ram_state->migration_dirty_pages++; |
3277 | } | |
13af18f2 ZC |
3278 | return block->colo_cache + offset; |
3279 | } | |
3280 | ||
3d0684b2 JQ |
3281 | /** |
3282 | * ram_handle_compressed: handle the zero page case | |
3283 | * | |
56e93d26 JQ |
3284 | * If a page (or a whole RDMA chunk) has been |
3285 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
3286 | * |
3287 | * @host: host address for the zero page | |
3288 | * @ch: what the page is filled from. We only support zero | |
3289 | * @size: size of the zero page | |
56e93d26 JQ |
3290 | */ |
3291 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
3292 | { | |
bad452a7 | 3293 | if (ch != 0 || !buffer_is_zero(host, size)) { |
56e93d26 JQ |
3294 | memset(host, ch, size); |
3295 | } | |
3296 | } | |
3297 | ||
797ca154 XG |
3298 | /* return the size after decompression, or negative value on error */ |
3299 | static int | |
3300 | qemu_uncompress_data(z_stream *stream, uint8_t *dest, size_t dest_len, | |
3301 | const uint8_t *source, size_t source_len) | |
3302 | { | |
3303 | int err; | |
3304 | ||
3305 | err = inflateReset(stream); | |
3306 | if (err != Z_OK) { | |
3307 | return -1; | |
3308 | } | |
3309 | ||
3310 | stream->avail_in = source_len; | |
3311 | stream->next_in = (uint8_t *)source; | |
3312 | stream->avail_out = dest_len; | |
3313 | stream->next_out = dest; | |
3314 | ||
3315 | err = inflate(stream, Z_NO_FLUSH); | |
3316 | if (err != Z_STREAM_END) { | |
3317 | return -1; | |
3318 | } | |
3319 | ||
3320 | return stream->total_out; | |
3321 | } | |
3322 | ||
56e93d26 JQ |
3323 | static void *do_data_decompress(void *opaque) |
3324 | { | |
3325 | DecompressParam *param = opaque; | |
3326 | unsigned long pagesize; | |
33d151f4 | 3327 | uint8_t *des; |
34ab9e97 | 3328 | int len, ret; |
56e93d26 | 3329 | |
33d151f4 | 3330 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 3331 | while (!param->quit) { |
33d151f4 LL |
3332 | if (param->des) { |
3333 | des = param->des; | |
3334 | len = param->len; | |
3335 | param->des = 0; | |
3336 | qemu_mutex_unlock(¶m->mutex); | |
3337 | ||
56e93d26 | 3338 | pagesize = TARGET_PAGE_SIZE; |
34ab9e97 XG |
3339 | |
3340 | ret = qemu_uncompress_data(¶m->stream, des, pagesize, | |
3341 | param->compbuf, len); | |
f548222c | 3342 | if (ret < 0 && migrate_get_current()->decompress_error_check) { |
34ab9e97 XG |
3343 | error_report("decompress data failed"); |
3344 | qemu_file_set_error(decomp_file, ret); | |
3345 | } | |
73a8912b | 3346 | |
33d151f4 LL |
3347 | qemu_mutex_lock(&decomp_done_lock); |
3348 | param->done = true; | |
3349 | qemu_cond_signal(&decomp_done_cond); | |
3350 | qemu_mutex_unlock(&decomp_done_lock); | |
3351 | ||
3352 | qemu_mutex_lock(¶m->mutex); | |
3353 | } else { | |
3354 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
3355 | } | |
56e93d26 | 3356 | } |
33d151f4 | 3357 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
3358 | |
3359 | return NULL; | |
3360 | } | |
3361 | ||
34ab9e97 | 3362 | static int wait_for_decompress_done(void) |
5533b2e9 LL |
3363 | { |
3364 | int idx, thread_count; | |
3365 | ||
3366 | if (!migrate_use_compression()) { | |
34ab9e97 | 3367 | return 0; |
5533b2e9 LL |
3368 | } |
3369 | ||
3370 | thread_count = migrate_decompress_threads(); | |
3371 | qemu_mutex_lock(&decomp_done_lock); | |
3372 | for (idx = 0; idx < thread_count; idx++) { | |
3373 | while (!decomp_param[idx].done) { | |
3374 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
3375 | } | |
3376 | } | |
3377 | qemu_mutex_unlock(&decomp_done_lock); | |
34ab9e97 | 3378 | return qemu_file_get_error(decomp_file); |
5533b2e9 LL |
3379 | } |
3380 | ||
f0afa331 | 3381 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
3382 | { |
3383 | int i, thread_count; | |
3384 | ||
3416ab5b JQ |
3385 | if (!migrate_use_compression()) { |
3386 | return; | |
3387 | } | |
56e93d26 JQ |
3388 | thread_count = migrate_decompress_threads(); |
3389 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3390 | /* |
3391 | * we use it as a indicator which shows if the thread is | |
3392 | * properly init'd or not | |
3393 | */ | |
3394 | if (!decomp_param[i].compbuf) { | |
3395 | break; | |
3396 | } | |
3397 | ||
56e93d26 | 3398 | qemu_mutex_lock(&decomp_param[i].mutex); |
90e56fb4 | 3399 | decomp_param[i].quit = true; |
56e93d26 JQ |
3400 | qemu_cond_signal(&decomp_param[i].cond); |
3401 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
3402 | } | |
3403 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3404 | if (!decomp_param[i].compbuf) { |
3405 | break; | |
3406 | } | |
3407 | ||
56e93d26 JQ |
3408 | qemu_thread_join(decompress_threads + i); |
3409 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
3410 | qemu_cond_destroy(&decomp_param[i].cond); | |
797ca154 | 3411 | inflateEnd(&decomp_param[i].stream); |
56e93d26 | 3412 | g_free(decomp_param[i].compbuf); |
797ca154 | 3413 | decomp_param[i].compbuf = NULL; |
56e93d26 JQ |
3414 | } |
3415 | g_free(decompress_threads); | |
3416 | g_free(decomp_param); | |
56e93d26 JQ |
3417 | decompress_threads = NULL; |
3418 | decomp_param = NULL; | |
34ab9e97 | 3419 | decomp_file = NULL; |
56e93d26 JQ |
3420 | } |
3421 | ||
34ab9e97 | 3422 | static int compress_threads_load_setup(QEMUFile *f) |
797ca154 XG |
3423 | { |
3424 | int i, thread_count; | |
3425 | ||
3426 | if (!migrate_use_compression()) { | |
3427 | return 0; | |
3428 | } | |
3429 | ||
3430 | thread_count = migrate_decompress_threads(); | |
3431 | decompress_threads = g_new0(QemuThread, thread_count); | |
3432 | decomp_param = g_new0(DecompressParam, thread_count); | |
3433 | qemu_mutex_init(&decomp_done_lock); | |
3434 | qemu_cond_init(&decomp_done_cond); | |
34ab9e97 | 3435 | decomp_file = f; |
797ca154 XG |
3436 | for (i = 0; i < thread_count; i++) { |
3437 | if (inflateInit(&decomp_param[i].stream) != Z_OK) { | |
3438 | goto exit; | |
3439 | } | |
3440 | ||
3441 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
3442 | qemu_mutex_init(&decomp_param[i].mutex); | |
3443 | qemu_cond_init(&decomp_param[i].cond); | |
3444 | decomp_param[i].done = true; | |
3445 | decomp_param[i].quit = false; | |
3446 | qemu_thread_create(decompress_threads + i, "decompress", | |
3447 | do_data_decompress, decomp_param + i, | |
3448 | QEMU_THREAD_JOINABLE); | |
3449 | } | |
3450 | return 0; | |
3451 | exit: | |
3452 | compress_threads_load_cleanup(); | |
3453 | return -1; | |
3454 | } | |
3455 | ||
c1bc6626 | 3456 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
3457 | void *host, int len) |
3458 | { | |
3459 | int idx, thread_count; | |
3460 | ||
3461 | thread_count = migrate_decompress_threads(); | |
37396950 | 3462 | QEMU_LOCK_GUARD(&decomp_done_lock); |
56e93d26 JQ |
3463 | while (true) { |
3464 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 3465 | if (decomp_param[idx].done) { |
33d151f4 LL |
3466 | decomp_param[idx].done = false; |
3467 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 3468 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
3469 | decomp_param[idx].des = host; |
3470 | decomp_param[idx].len = len; | |
33d151f4 LL |
3471 | qemu_cond_signal(&decomp_param[idx].cond); |
3472 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
3473 | break; |
3474 | } | |
3475 | } | |
3476 | if (idx < thread_count) { | |
3477 | break; | |
73a8912b LL |
3478 | } else { |
3479 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
3480 | } |
3481 | } | |
3482 | } | |
3483 | ||
b70cb3b4 RL |
3484 | static void colo_init_ram_state(void) |
3485 | { | |
3486 | ram_state_init(&ram_state); | |
b70cb3b4 RL |
3487 | } |
3488 | ||
13af18f2 ZC |
3489 | /* |
3490 | * colo cache: this is for secondary VM, we cache the whole | |
3491 | * memory of the secondary VM, it is need to hold the global lock | |
3492 | * to call this helper. | |
3493 | */ | |
3494 | int colo_init_ram_cache(void) | |
3495 | { | |
3496 | RAMBlock *block; | |
3497 | ||
44901b5a PB |
3498 | WITH_RCU_READ_LOCK_GUARD() { |
3499 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3500 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
8dbe22c6 | 3501 | NULL, false, false); |
44901b5a PB |
3502 | if (!block->colo_cache) { |
3503 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
3504 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
3505 | block->used_length); | |
3506 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3507 | if (block->colo_cache) { | |
3508 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3509 | block->colo_cache = NULL; | |
3510 | } | |
89ac5a1d | 3511 | } |
44901b5a | 3512 | return -errno; |
89ac5a1d | 3513 | } |
e5fdf920 LS |
3514 | if (!machine_dump_guest_core(current_machine)) { |
3515 | qemu_madvise(block->colo_cache, block->used_length, | |
3516 | QEMU_MADV_DONTDUMP); | |
3517 | } | |
13af18f2 | 3518 | } |
13af18f2 | 3519 | } |
44901b5a | 3520 | |
7d9acafa ZC |
3521 | /* |
3522 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3523 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3524 | * we use the same name 'ram_bitmap' as for migration. | |
3525 | */ | |
3526 | if (ram_bytes_total()) { | |
3527 | RAMBlock *block; | |
3528 | ||
fbd162e6 | 3529 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3530 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3531 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3532 | } |
3533 | } | |
7d9acafa | 3534 | |
b70cb3b4 | 3535 | colo_init_ram_state(); |
13af18f2 | 3536 | return 0; |
13af18f2 ZC |
3537 | } |
3538 | ||
0393031a HZ |
3539 | /* TODO: duplicated with ram_init_bitmaps */ |
3540 | void colo_incoming_start_dirty_log(void) | |
3541 | { | |
3542 | RAMBlock *block = NULL; | |
3543 | /* For memory_global_dirty_log_start below. */ | |
3544 | qemu_mutex_lock_iothread(); | |
3545 | qemu_mutex_lock_ramlist(); | |
3546 | ||
3547 | memory_global_dirty_log_sync(); | |
3548 | WITH_RCU_READ_LOCK_GUARD() { | |
3549 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3550 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3551 | /* Discard this dirty bitmap record */ | |
3552 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3553 | } | |
63b41db4 | 3554 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
0393031a HZ |
3555 | } |
3556 | ram_state->migration_dirty_pages = 0; | |
3557 | qemu_mutex_unlock_ramlist(); | |
3558 | qemu_mutex_unlock_iothread(); | |
3559 | } | |
3560 | ||
13af18f2 ZC |
3561 | /* It is need to hold the global lock to call this helper */ |
3562 | void colo_release_ram_cache(void) | |
3563 | { | |
3564 | RAMBlock *block; | |
3565 | ||
63b41db4 | 3566 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); |
fbd162e6 | 3567 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3568 | g_free(block->bmap); |
3569 | block->bmap = NULL; | |
3570 | } | |
3571 | ||
89ac5a1d DDAG |
3572 | WITH_RCU_READ_LOCK_GUARD() { |
3573 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3574 | if (block->colo_cache) { | |
3575 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3576 | block->colo_cache = NULL; | |
3577 | } | |
13af18f2 ZC |
3578 | } |
3579 | } | |
0393031a | 3580 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
3581 | } |
3582 | ||
f265e0e4 JQ |
3583 | /** |
3584 | * ram_load_setup: Setup RAM for migration incoming side | |
3585 | * | |
3586 | * Returns zero to indicate success and negative for error | |
3587 | * | |
3588 | * @f: QEMUFile where to receive the data | |
3589 | * @opaque: RAMState pointer | |
3590 | */ | |
3591 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
3592 | { | |
34ab9e97 | 3593 | if (compress_threads_load_setup(f)) { |
797ca154 XG |
3594 | return -1; |
3595 | } | |
3596 | ||
f265e0e4 | 3597 | xbzrle_load_setup(); |
f9494614 | 3598 | ramblock_recv_map_init(); |
13af18f2 | 3599 | |
f265e0e4 JQ |
3600 | return 0; |
3601 | } | |
3602 | ||
3603 | static int ram_load_cleanup(void *opaque) | |
3604 | { | |
f9494614 | 3605 | RAMBlock *rb; |
56eb90af | 3606 | |
fbd162e6 | 3607 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 3608 | qemu_ram_block_writeback(rb); |
56eb90af JH |
3609 | } |
3610 | ||
f265e0e4 | 3611 | xbzrle_load_cleanup(); |
f0afa331 | 3612 | compress_threads_load_cleanup(); |
f9494614 | 3613 | |
fbd162e6 | 3614 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
3615 | g_free(rb->receivedmap); |
3616 | rb->receivedmap = NULL; | |
3617 | } | |
13af18f2 | 3618 | |
f265e0e4 JQ |
3619 | return 0; |
3620 | } | |
3621 | ||
3d0684b2 JQ |
3622 | /** |
3623 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
3624 | * | |
3625 | * Returns 0 for success and negative if there was one error | |
3626 | * | |
3627 | * @mis: current migration incoming state | |
3628 | * | |
3629 | * Allocate data structures etc needed by incoming migration with | |
3630 | * postcopy-ram. postcopy-ram's similarly names | |
3631 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
3632 | */ |
3633 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
3634 | { | |
c136180c | 3635 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
3636 | } |
3637 | ||
3d0684b2 JQ |
3638 | /** |
3639 | * ram_load_postcopy: load a page in postcopy case | |
3640 | * | |
3641 | * Returns 0 for success or -errno in case of error | |
3642 | * | |
a7180877 DDAG |
3643 | * Called in postcopy mode by ram_load(). |
3644 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
3645 | * |
3646 | * @f: QEMUFile where to send the data | |
a7180877 | 3647 | */ |
929068ec | 3648 | int ram_load_postcopy(QEMUFile *f) |
a7180877 DDAG |
3649 | { |
3650 | int flags = 0, ret = 0; | |
3651 | bool place_needed = false; | |
1aa83678 | 3652 | bool matches_target_page_size = false; |
a7180877 | 3653 | MigrationIncomingState *mis = migration_incoming_get_current(); |
77dadc3f PX |
3654 | /* Currently we only use channel 0. TODO: use all the channels */ |
3655 | PostcopyTmpPage *tmp_page = &mis->postcopy_tmp_pages[0]; | |
a7180877 DDAG |
3656 | |
3657 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
3658 | ram_addr_t addr; | |
a7180877 DDAG |
3659 | void *page_buffer = NULL; |
3660 | void *place_source = NULL; | |
df9ff5e1 | 3661 | RAMBlock *block = NULL; |
a7180877 | 3662 | uint8_t ch; |
644acf99 | 3663 | int len; |
a7180877 DDAG |
3664 | |
3665 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
3666 | |
3667 | /* | |
3668 | * If qemu file error, we should stop here, and then "addr" | |
3669 | * may be invalid | |
3670 | */ | |
3671 | ret = qemu_file_get_error(f); | |
3672 | if (ret) { | |
3673 | break; | |
3674 | } | |
3675 | ||
a7180877 DDAG |
3676 | flags = addr & ~TARGET_PAGE_MASK; |
3677 | addr &= TARGET_PAGE_MASK; | |
3678 | ||
3679 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
644acf99 WY |
3680 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
3681 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
755e8d7c | 3682 | block = ram_block_from_stream(mis, f, flags); |
6a23f639 DH |
3683 | if (!block) { |
3684 | ret = -EINVAL; | |
3685 | break; | |
3686 | } | |
4c4bad48 | 3687 | |
898ba906 DH |
3688 | /* |
3689 | * Relying on used_length is racy and can result in false positives. | |
3690 | * We might place pages beyond used_length in case RAM was shrunk | |
3691 | * while in postcopy, which is fine - trying to place via | |
3692 | * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault. | |
3693 | */ | |
3694 | if (!block->host || addr >= block->postcopy_length) { | |
a7180877 DDAG |
3695 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); |
3696 | ret = -EINVAL; | |
3697 | break; | |
3698 | } | |
77dadc3f | 3699 | tmp_page->target_pages++; |
1aa83678 | 3700 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 3701 | /* |
28abd200 DDAG |
3702 | * Postcopy requires that we place whole host pages atomically; |
3703 | * these may be huge pages for RAMBlocks that are backed by | |
3704 | * hugetlbfs. | |
a7180877 DDAG |
3705 | * To make it atomic, the data is read into a temporary page |
3706 | * that's moved into place later. | |
3707 | * The migration protocol uses, possibly smaller, target-pages | |
3708 | * however the source ensures it always sends all the components | |
91ba442f | 3709 | * of a host page in one chunk. |
a7180877 | 3710 | */ |
77dadc3f | 3711 | page_buffer = tmp_page->tmp_huge_page + |
6a23f639 DH |
3712 | host_page_offset_from_ram_block_offset(block, addr); |
3713 | /* If all TP are zero then we can optimise the place */ | |
77dadc3f PX |
3714 | if (tmp_page->target_pages == 1) { |
3715 | tmp_page->host_addr = | |
3716 | host_page_from_ram_block_offset(block, addr); | |
3717 | } else if (tmp_page->host_addr != | |
3718 | host_page_from_ram_block_offset(block, addr)) { | |
c53b7ddc | 3719 | /* not the 1st TP within the HP */ |
cfc7dc8a PX |
3720 | error_report("Non-same host page detected. " |
3721 | "Target host page %p, received host page %p " | |
3722 | "(rb %s offset 0x"RAM_ADDR_FMT" target_pages %d)", | |
3723 | tmp_page->host_addr, | |
3724 | host_page_from_ram_block_offset(block, addr), | |
3725 | block->idstr, addr, tmp_page->target_pages); | |
6a23f639 DH |
3726 | ret = -EINVAL; |
3727 | break; | |
a7180877 DDAG |
3728 | } |
3729 | ||
3730 | /* | |
3731 | * If it's the last part of a host page then we place the host | |
3732 | * page | |
3733 | */ | |
77dadc3f PX |
3734 | if (tmp_page->target_pages == |
3735 | (block->page_size / TARGET_PAGE_SIZE)) { | |
4cbb3c63 | 3736 | place_needed = true; |
4cbb3c63 | 3737 | } |
77dadc3f | 3738 | place_source = tmp_page->tmp_huge_page; |
a7180877 DDAG |
3739 | } |
3740 | ||
3741 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 3742 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 3743 | ch = qemu_get_byte(f); |
2e36bc1b WY |
3744 | /* |
3745 | * Can skip to set page_buffer when | |
3746 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
3747 | */ | |
3748 | if (ch || !matches_target_page_size) { | |
3749 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
3750 | } | |
a7180877 | 3751 | if (ch) { |
77dadc3f | 3752 | tmp_page->all_zero = false; |
a7180877 DDAG |
3753 | } |
3754 | break; | |
3755 | ||
3756 | case RAM_SAVE_FLAG_PAGE: | |
77dadc3f | 3757 | tmp_page->all_zero = false; |
1aa83678 PX |
3758 | if (!matches_target_page_size) { |
3759 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
3760 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
3761 | } else { | |
1aa83678 PX |
3762 | /* |
3763 | * For small pages that matches target page size, we | |
3764 | * avoid the qemu_file copy. Instead we directly use | |
3765 | * the buffer of QEMUFile to place the page. Note: we | |
3766 | * cannot do any QEMUFile operation before using that | |
3767 | * buffer to make sure the buffer is valid when | |
3768 | * placing the page. | |
a7180877 DDAG |
3769 | */ |
3770 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
3771 | TARGET_PAGE_SIZE); | |
3772 | } | |
3773 | break; | |
644acf99 | 3774 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
77dadc3f | 3775 | tmp_page->all_zero = false; |
644acf99 WY |
3776 | len = qemu_get_be32(f); |
3777 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
3778 | error_report("Invalid compressed data length: %d", len); | |
3779 | ret = -EINVAL; | |
3780 | break; | |
3781 | } | |
3782 | decompress_data_with_multi_threads(f, page_buffer, len); | |
3783 | break; | |
3784 | ||
a7180877 DDAG |
3785 | case RAM_SAVE_FLAG_EOS: |
3786 | /* normal exit */ | |
6df264ac | 3787 | multifd_recv_sync_main(); |
a7180877 DDAG |
3788 | break; |
3789 | default: | |
29fccade | 3790 | error_report("Unknown combination of migration flags: 0x%x" |
a7180877 DDAG |
3791 | " (postcopy mode)", flags); |
3792 | ret = -EINVAL; | |
7a9ddfbf PX |
3793 | break; |
3794 | } | |
3795 | ||
644acf99 WY |
3796 | /* Got the whole host page, wait for decompress before placing. */ |
3797 | if (place_needed) { | |
3798 | ret |= wait_for_decompress_done(); | |
3799 | } | |
3800 | ||
7a9ddfbf PX |
3801 | /* Detect for any possible file errors */ |
3802 | if (!ret && qemu_file_get_error(f)) { | |
3803 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
3804 | } |
3805 | ||
7a9ddfbf | 3806 | if (!ret && place_needed) { |
77dadc3f PX |
3807 | if (tmp_page->all_zero) { |
3808 | ret = postcopy_place_page_zero(mis, tmp_page->host_addr, block); | |
a7180877 | 3809 | } else { |
77dadc3f PX |
3810 | ret = postcopy_place_page(mis, tmp_page->host_addr, |
3811 | place_source, block); | |
a7180877 | 3812 | } |
ddf35bdf | 3813 | place_needed = false; |
77dadc3f | 3814 | postcopy_temp_page_reset(tmp_page); |
a7180877 | 3815 | } |
a7180877 DDAG |
3816 | } |
3817 | ||
3818 | return ret; | |
3819 | } | |
3820 | ||
acab30b8 DHB |
3821 | static bool postcopy_is_advised(void) |
3822 | { | |
3823 | PostcopyState ps = postcopy_state_get(); | |
3824 | return ps >= POSTCOPY_INCOMING_ADVISE && ps < POSTCOPY_INCOMING_END; | |
3825 | } | |
3826 | ||
3827 | static bool postcopy_is_running(void) | |
3828 | { | |
3829 | PostcopyState ps = postcopy_state_get(); | |
3830 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
3831 | } | |
3832 | ||
e6f4aa18 ZC |
3833 | /* |
3834 | * Flush content of RAM cache into SVM's memory. | |
3835 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
3836 | */ | |
24fa16f8 | 3837 | void colo_flush_ram_cache(void) |
e6f4aa18 ZC |
3838 | { |
3839 | RAMBlock *block = NULL; | |
3840 | void *dst_host; | |
3841 | void *src_host; | |
3842 | unsigned long offset = 0; | |
3843 | ||
d1955d22 | 3844 | memory_global_dirty_log_sync(); |
89ac5a1d DDAG |
3845 | WITH_RCU_READ_LOCK_GUARD() { |
3846 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3847 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3848 | } | |
d1955d22 | 3849 | } |
d1955d22 | 3850 | |
e6f4aa18 | 3851 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
3852 | WITH_RCU_READ_LOCK_GUARD() { |
3853 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 3854 | |
89ac5a1d | 3855 | while (block) { |
a6a83cef | 3856 | unsigned long num = 0; |
e6f4aa18 | 3857 | |
a6a83cef | 3858 | offset = colo_bitmap_find_dirty(ram_state, block, offset, &num); |
542147f4 DH |
3859 | if (!offset_in_ramblock(block, |
3860 | ((ram_addr_t)offset) << TARGET_PAGE_BITS)) { | |
89ac5a1d | 3861 | offset = 0; |
a6a83cef | 3862 | num = 0; |
89ac5a1d DDAG |
3863 | block = QLIST_NEXT_RCU(block, next); |
3864 | } else { | |
a6a83cef RL |
3865 | unsigned long i = 0; |
3866 | ||
3867 | for (i = 0; i < num; i++) { | |
3868 | migration_bitmap_clear_dirty(ram_state, block, offset + i); | |
3869 | } | |
8bba004c AR |
3870 | dst_host = block->host |
3871 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
3872 | src_host = block->colo_cache | |
3873 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
a6a83cef RL |
3874 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE * num); |
3875 | offset += num; | |
89ac5a1d | 3876 | } |
e6f4aa18 ZC |
3877 | } |
3878 | } | |
e6f4aa18 ZC |
3879 | trace_colo_flush_ram_cache_end(); |
3880 | } | |
3881 | ||
10da4a36 WY |
3882 | /** |
3883 | * ram_load_precopy: load pages in precopy case | |
3884 | * | |
3885 | * Returns 0 for success or -errno in case of error | |
3886 | * | |
3887 | * Called in precopy mode by ram_load(). | |
3888 | * rcu_read_lock is taken prior to this being called. | |
3889 | * | |
3890 | * @f: QEMUFile where to send the data | |
3891 | */ | |
3892 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 3893 | { |
755e8d7c | 3894 | MigrationIncomingState *mis = migration_incoming_get_current(); |
e65cec5e | 3895 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 3896 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
acab30b8 | 3897 | bool postcopy_advised = postcopy_is_advised(); |
edc60127 JQ |
3898 | if (!migrate_use_compression()) { |
3899 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
3900 | } | |
a7180877 | 3901 | |
10da4a36 | 3902 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 3903 | ram_addr_t addr, total_ram_bytes; |
0393031a | 3904 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
3905 | uint8_t ch; |
3906 | ||
e65cec5e YK |
3907 | /* |
3908 | * Yield periodically to let main loop run, but an iteration of | |
3909 | * the main loop is expensive, so do it each some iterations | |
3910 | */ | |
3911 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
3912 | aio_co_schedule(qemu_get_current_aio_context(), | |
3913 | qemu_coroutine_self()); | |
3914 | qemu_coroutine_yield(); | |
3915 | } | |
3916 | i++; | |
3917 | ||
56e93d26 JQ |
3918 | addr = qemu_get_be64(f); |
3919 | flags = addr & ~TARGET_PAGE_MASK; | |
3920 | addr &= TARGET_PAGE_MASK; | |
3921 | ||
edc60127 JQ |
3922 | if (flags & invalid_flags) { |
3923 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
3924 | error_report("Received an unexpected compressed page"); | |
3925 | } | |
3926 | ||
3927 | ret = -EINVAL; | |
3928 | break; | |
3929 | } | |
3930 | ||
bb890ed5 | 3931 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 3932 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
755e8d7c | 3933 | RAMBlock *block = ram_block_from_stream(mis, f, flags); |
4c4bad48 | 3934 | |
0393031a | 3935 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 3936 | /* |
0393031a HZ |
3937 | * After going into COLO stage, we should not load the page |
3938 | * into SVM's memory directly, we put them into colo_cache firstly. | |
3939 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
3940 | * Previously, we copied all these memory in preparing stage of COLO | |
3941 | * while we need to stop VM, which is a time-consuming process. | |
3942 | * Here we optimize it by a trick, back-up every page while in | |
3943 | * migration process while COLO is enabled, though it affects the | |
3944 | * speed of the migration, but it obviously reduce the downtime of | |
3945 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 3946 | */ |
0393031a HZ |
3947 | if (migration_incoming_colo_enabled()) { |
3948 | if (migration_incoming_in_colo_state()) { | |
3949 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 3950 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
3951 | } else { |
3952 | /* | |
3953 | * In migration stage but before COLO stage, | |
3954 | * Put all pages into both cache and SVM's memory. | |
3955 | */ | |
8af66371 | 3956 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 3957 | } |
13af18f2 | 3958 | } |
a776aa15 DDAG |
3959 | if (!host) { |
3960 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
3961 | ret = -EINVAL; | |
3962 | break; | |
3963 | } | |
13af18f2 ZC |
3964 | if (!migration_incoming_in_colo_state()) { |
3965 | ramblock_recv_bitmap_set(block, host); | |
3966 | } | |
3967 | ||
1db9d8e5 | 3968 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
3969 | } |
3970 | ||
56e93d26 JQ |
3971 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
3972 | case RAM_SAVE_FLAG_MEM_SIZE: | |
3973 | /* Synchronize RAM block list */ | |
3974 | total_ram_bytes = addr; | |
3975 | while (!ret && total_ram_bytes) { | |
3976 | RAMBlock *block; | |
56e93d26 JQ |
3977 | char id[256]; |
3978 | ram_addr_t length; | |
3979 | ||
3980 | len = qemu_get_byte(f); | |
3981 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3982 | id[len] = 0; | |
3983 | length = qemu_get_be64(f); | |
3984 | ||
e3dd7493 | 3985 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
3986 | if (block && !qemu_ram_is_migratable(block)) { |
3987 | error_report("block %s should not be migrated !", id); | |
3988 | ret = -EINVAL; | |
3989 | } else if (block) { | |
e3dd7493 DDAG |
3990 | if (length != block->used_length) { |
3991 | Error *local_err = NULL; | |
56e93d26 | 3992 | |
fa53a0e5 | 3993 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
3994 | &local_err); |
3995 | if (local_err) { | |
3996 | error_report_err(local_err); | |
56e93d26 | 3997 | } |
56e93d26 | 3998 | } |
ef08fb38 | 3999 | /* For postcopy we need to check hugepage sizes match */ |
e846b746 | 4000 | if (postcopy_advised && migrate_postcopy_ram() && |
ef08fb38 DDAG |
4001 | block->page_size != qemu_host_page_size) { |
4002 | uint64_t remote_page_size = qemu_get_be64(f); | |
4003 | if (remote_page_size != block->page_size) { | |
4004 | error_report("Mismatched RAM page size %s " | |
4005 | "(local) %zd != %" PRId64, | |
4006 | id, block->page_size, | |
4007 | remote_page_size); | |
4008 | ret = -EINVAL; | |
4009 | } | |
4010 | } | |
fbd162e6 YK |
4011 | if (migrate_ignore_shared()) { |
4012 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
4013 | if (ramblock_is_ignored(block) && |
4014 | block->mr->addr != addr) { | |
4015 | error_report("Mismatched GPAs for block %s " | |
4016 | "%" PRId64 "!= %" PRId64, | |
4017 | id, (uint64_t)addr, | |
4018 | (uint64_t)block->mr->addr); | |
4019 | ret = -EINVAL; | |
4020 | } | |
4021 | } | |
e3dd7493 DDAG |
4022 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
4023 | block->idstr); | |
4024 | } else { | |
56e93d26 JQ |
4025 | error_report("Unknown ramblock \"%s\", cannot " |
4026 | "accept migration", id); | |
4027 | ret = -EINVAL; | |
4028 | } | |
4029 | ||
4030 | total_ram_bytes -= length; | |
4031 | } | |
4032 | break; | |
a776aa15 | 4033 | |
bb890ed5 | 4034 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
4035 | ch = qemu_get_byte(f); |
4036 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
4037 | break; | |
a776aa15 | 4038 | |
56e93d26 | 4039 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
4040 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
4041 | break; | |
56e93d26 | 4042 | |
a776aa15 | 4043 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
4044 | len = qemu_get_be32(f); |
4045 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4046 | error_report("Invalid compressed data length: %d", len); | |
4047 | ret = -EINVAL; | |
4048 | break; | |
4049 | } | |
c1bc6626 | 4050 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 4051 | break; |
a776aa15 | 4052 | |
56e93d26 | 4053 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
4054 | if (load_xbzrle(f, addr, host) < 0) { |
4055 | error_report("Failed to decompress XBZRLE page at " | |
4056 | RAM_ADDR_FMT, addr); | |
4057 | ret = -EINVAL; | |
4058 | break; | |
4059 | } | |
4060 | break; | |
4061 | case RAM_SAVE_FLAG_EOS: | |
4062 | /* normal exit */ | |
6df264ac | 4063 | multifd_recv_sync_main(); |
56e93d26 JQ |
4064 | break; |
4065 | default: | |
4066 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 4067 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 | 4068 | } else { |
29fccade | 4069 | error_report("Unknown combination of migration flags: 0x%x", |
56e93d26 JQ |
4070 | flags); |
4071 | ret = -EINVAL; | |
4072 | } | |
4073 | } | |
4074 | if (!ret) { | |
4075 | ret = qemu_file_get_error(f); | |
4076 | } | |
0393031a HZ |
4077 | if (!ret && host_bak) { |
4078 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
4079 | } | |
56e93d26 JQ |
4080 | } |
4081 | ||
ca1a6b70 | 4082 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
4083 | return ret; |
4084 | } | |
4085 | ||
4086 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
4087 | { | |
4088 | int ret = 0; | |
4089 | static uint64_t seq_iter; | |
4090 | /* | |
4091 | * If system is running in postcopy mode, page inserts to host memory must | |
4092 | * be atomic | |
4093 | */ | |
4094 | bool postcopy_running = postcopy_is_running(); | |
4095 | ||
4096 | seq_iter++; | |
4097 | ||
4098 | if (version_id != 4) { | |
4099 | return -EINVAL; | |
4100 | } | |
4101 | ||
4102 | /* | |
4103 | * This RCU critical section can be very long running. | |
4104 | * When RCU reclaims in the code start to become numerous, | |
4105 | * it will be necessary to reduce the granularity of this | |
4106 | * critical section. | |
4107 | */ | |
89ac5a1d DDAG |
4108 | WITH_RCU_READ_LOCK_GUARD() { |
4109 | if (postcopy_running) { | |
4110 | ret = ram_load_postcopy(f); | |
4111 | } else { | |
4112 | ret = ram_load_precopy(f); | |
4113 | } | |
10da4a36 | 4114 | } |
55c4446b | 4115 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 | 4116 | |
56e93d26 JQ |
4117 | return ret; |
4118 | } | |
4119 | ||
c6467627 VSO |
4120 | static bool ram_has_postcopy(void *opaque) |
4121 | { | |
469dd51b | 4122 | RAMBlock *rb; |
fbd162e6 | 4123 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
4124 | if (ramblock_is_pmem(rb)) { |
4125 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
4126 | "is not supported now!", rb->idstr, rb->host); | |
4127 | return false; | |
4128 | } | |
4129 | } | |
4130 | ||
c6467627 VSO |
4131 | return migrate_postcopy_ram(); |
4132 | } | |
4133 | ||
edd090c7 PX |
4134 | /* Sync all the dirty bitmap with destination VM. */ |
4135 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
4136 | { | |
4137 | RAMBlock *block; | |
4138 | QEMUFile *file = s->to_dst_file; | |
4139 | int ramblock_count = 0; | |
4140 | ||
4141 | trace_ram_dirty_bitmap_sync_start(); | |
4142 | ||
fbd162e6 | 4143 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
4144 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
4145 | trace_ram_dirty_bitmap_request(block->idstr); | |
4146 | ramblock_count++; | |
4147 | } | |
4148 | ||
4149 | trace_ram_dirty_bitmap_sync_wait(); | |
4150 | ||
4151 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
4152 | while (ramblock_count--) { | |
4153 | qemu_sem_wait(&s->rp_state.rp_sem); | |
4154 | } | |
4155 | ||
4156 | trace_ram_dirty_bitmap_sync_complete(); | |
4157 | ||
4158 | return 0; | |
4159 | } | |
4160 | ||
4161 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
4162 | { | |
4163 | qemu_sem_post(&s->rp_state.rp_sem); | |
4164 | } | |
4165 | ||
a335debb PX |
4166 | /* |
4167 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
4168 | * This is only used when the postcopy migration is paused but wants | |
4169 | * to resume from a middle point. | |
4170 | */ | |
4171 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
4172 | { | |
4173 | int ret = -EINVAL; | |
43044ac0 | 4174 | /* from_dst_file is always valid because we're within rp_thread */ |
a335debb PX |
4175 | QEMUFile *file = s->rp_state.from_dst_file; |
4176 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 4177 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
4178 | uint64_t size, end_mark; |
4179 | ||
4180 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
4181 | ||
4182 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
4183 | error_report("%s: incorrect state %s", __func__, | |
4184 | MigrationStatus_str(s->state)); | |
4185 | return -EINVAL; | |
4186 | } | |
4187 | ||
4188 | /* | |
4189 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3a4452d8 | 4190 | * need the endianness conversion, and the paddings. |
a335debb PX |
4191 | */ |
4192 | local_size = ROUND_UP(local_size, 8); | |
4193 | ||
4194 | /* Add paddings */ | |
4195 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
4196 | ||
4197 | size = qemu_get_be64(file); | |
4198 | ||
4199 | /* The size of the bitmap should match with our ramblock */ | |
4200 | if (size != local_size) { | |
4201 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
4202 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
4203 | block->idstr, size, local_size); | |
4204 | ret = -EINVAL; | |
4205 | goto out; | |
4206 | } | |
4207 | ||
4208 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
4209 | end_mark = qemu_get_be64(file); | |
4210 | ||
4211 | ret = qemu_file_get_error(file); | |
4212 | if (ret || size != local_size) { | |
4213 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
4214 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
4215 | __func__, block->idstr, ret, local_size, size); | |
4216 | ret = -EIO; | |
4217 | goto out; | |
4218 | } | |
4219 | ||
4220 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
af3bbbe9 | 4221 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64, |
a335debb PX |
4222 | __func__, block->idstr, end_mark); |
4223 | ret = -EINVAL; | |
4224 | goto out; | |
4225 | } | |
4226 | ||
4227 | /* | |
3a4452d8 | 4228 | * Endianness conversion. We are during postcopy (though paused). |
a335debb PX |
4229 | * The dirty bitmap won't change. We can directly modify it. |
4230 | */ | |
4231 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
4232 | ||
4233 | /* | |
4234 | * What we received is "received bitmap". Revert it as the initial | |
4235 | * dirty bitmap for this ramblock. | |
4236 | */ | |
4237 | bitmap_complement(block->bmap, block->bmap, nbits); | |
4238 | ||
be39b4cd DH |
4239 | /* Clear dirty bits of discarded ranges that we don't want to migrate. */ |
4240 | ramblock_dirty_bitmap_clear_discarded_pages(block); | |
4241 | ||
4242 | /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */ | |
a335debb PX |
4243 | trace_ram_dirty_bitmap_reload_complete(block->idstr); |
4244 | ||
edd090c7 PX |
4245 | /* |
4246 | * We succeeded to sync bitmap for current ramblock. If this is | |
4247 | * the last one to sync, we need to notify the main send thread. | |
4248 | */ | |
4249 | ram_dirty_bitmap_reload_notify(s); | |
4250 | ||
a335debb PX |
4251 | ret = 0; |
4252 | out: | |
bf269906 | 4253 | g_free(le_bitmap); |
a335debb PX |
4254 | return ret; |
4255 | } | |
4256 | ||
edd090c7 PX |
4257 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
4258 | { | |
4259 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 4260 | int ret; |
edd090c7 | 4261 | |
08614f34 PX |
4262 | ret = ram_dirty_bitmap_sync_all(s, rs); |
4263 | if (ret) { | |
4264 | return ret; | |
4265 | } | |
4266 | ||
4267 | ram_state_resume_prepare(rs, s->to_dst_file); | |
4268 | ||
4269 | return 0; | |
edd090c7 PX |
4270 | } |
4271 | ||
56e93d26 | 4272 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 4273 | .save_setup = ram_save_setup, |
56e93d26 | 4274 | .save_live_iterate = ram_save_iterate, |
763c906b | 4275 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 4276 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 4277 | .has_postcopy = ram_has_postcopy, |
56e93d26 JQ |
4278 | .save_live_pending = ram_save_pending, |
4279 | .load_state = ram_load, | |
f265e0e4 JQ |
4280 | .save_cleanup = ram_save_cleanup, |
4281 | .load_setup = ram_load_setup, | |
4282 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 4283 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
4284 | }; |
4285 | ||
c7c0e724 DH |
4286 | static void ram_mig_ram_block_resized(RAMBlockNotifier *n, void *host, |
4287 | size_t old_size, size_t new_size) | |
4288 | { | |
cc61c703 | 4289 | PostcopyState ps = postcopy_state_get(); |
c7c0e724 DH |
4290 | ram_addr_t offset; |
4291 | RAMBlock *rb = qemu_ram_block_from_host(host, false, &offset); | |
4292 | Error *err = NULL; | |
4293 | ||
4294 | if (ramblock_is_ignored(rb)) { | |
4295 | return; | |
4296 | } | |
4297 | ||
4298 | if (!migration_is_idle()) { | |
4299 | /* | |
4300 | * Precopy code on the source cannot deal with the size of RAM blocks | |
4301 | * changing at random points in time - especially after sending the | |
4302 | * RAM block sizes in the migration stream, they must no longer change. | |
4303 | * Abort and indicate a proper reason. | |
4304 | */ | |
4305 | error_setg(&err, "RAM block '%s' resized during precopy.", rb->idstr); | |
458fecca | 4306 | migration_cancel(err); |
c7c0e724 | 4307 | error_free(err); |
c7c0e724 | 4308 | } |
cc61c703 DH |
4309 | |
4310 | switch (ps) { | |
4311 | case POSTCOPY_INCOMING_ADVISE: | |
4312 | /* | |
4313 | * Update what ram_postcopy_incoming_init()->init_range() does at the | |
4314 | * time postcopy was advised. Syncing RAM blocks with the source will | |
4315 | * result in RAM resizes. | |
4316 | */ | |
4317 | if (old_size < new_size) { | |
4318 | if (ram_discard_range(rb->idstr, old_size, new_size - old_size)) { | |
4319 | error_report("RAM block '%s' discard of resized RAM failed", | |
4320 | rb->idstr); | |
4321 | } | |
4322 | } | |
898ba906 | 4323 | rb->postcopy_length = new_size; |
cc61c703 DH |
4324 | break; |
4325 | case POSTCOPY_INCOMING_NONE: | |
4326 | case POSTCOPY_INCOMING_RUNNING: | |
4327 | case POSTCOPY_INCOMING_END: | |
4328 | /* | |
4329 | * Once our guest is running, postcopy does no longer care about | |
4330 | * resizes. When growing, the new memory was not available on the | |
4331 | * source, no handler needed. | |
4332 | */ | |
4333 | break; | |
4334 | default: | |
4335 | error_report("RAM block '%s' resized during postcopy state: %d", | |
4336 | rb->idstr, ps); | |
4337 | exit(-1); | |
4338 | } | |
c7c0e724 DH |
4339 | } |
4340 | ||
4341 | static RAMBlockNotifier ram_mig_ram_notifier = { | |
4342 | .ram_block_resized = ram_mig_ram_block_resized, | |
4343 | }; | |
4344 | ||
56e93d26 JQ |
4345 | void ram_mig_init(void) |
4346 | { | |
4347 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 4348 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
c7c0e724 | 4349 | ram_block_notifier_add(&ram_mig_ram_notifier); |
56e93d26 | 4350 | } |