]>
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" | |
7205c9ec | 33 | #include "qemu/main-loop.h" |
709e3fe8 | 34 | #include "xbzrle.h" |
7b1e1a22 | 35 | #include "ram.h" |
6666c96a | 36 | #include "migration.h" |
f2a8f0a6 | 37 | #include "migration/register.h" |
7b1e1a22 | 38 | #include "migration/misc.h" |
08a0aee1 | 39 | #include "qemu-file.h" |
be07b0ac | 40 | #include "postcopy-ram.h" |
53d37d36 | 41 | #include "page_cache.h" |
56e93d26 | 42 | #include "qemu/error-report.h" |
e688df6b | 43 | #include "qapi/error.h" |
ab7cbb0b | 44 | #include "qapi/qapi-types-migration.h" |
9af23989 | 45 | #include "qapi/qapi-events-migration.h" |
8acabf69 | 46 | #include "qapi/qmp/qerror.h" |
56e93d26 | 47 | #include "trace.h" |
56e93d26 | 48 | #include "exec/ram_addr.h" |
f9494614 | 49 | #include "exec/target_page.h" |
56e93d26 | 50 | #include "qemu/rcu_queue.h" |
a91246c9 | 51 | #include "migration/colo.h" |
53d37d36 | 52 | #include "block.h" |
b0c3cf94 | 53 | #include "sysemu/cpu-throttle.h" |
edd090c7 | 54 | #include "savevm.h" |
b9ee2f7d | 55 | #include "qemu/iov.h" |
d32ca5ad | 56 | #include "multifd.h" |
278e2f55 AG |
57 | #include "sysemu/runstate.h" |
58 | ||
59 | #if defined(__linux__) | |
60 | #include "qemu/userfaultfd.h" | |
61 | #endif /* defined(__linux__) */ | |
56e93d26 | 62 | |
56e93d26 JQ |
63 | /***********************************************************/ |
64 | /* ram save/restore */ | |
65 | ||
bb890ed5 JQ |
66 | /* RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it |
67 | * worked for pages that where filled with the same char. We switched | |
68 | * it to only search for the zero value. And to avoid confusion with | |
69 | * RAM_SSAVE_FLAG_COMPRESS_PAGE just rename it. | |
70 | */ | |
71 | ||
56e93d26 | 72 | #define RAM_SAVE_FLAG_FULL 0x01 /* Obsolete, not used anymore */ |
bb890ed5 | 73 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
74 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
75 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
76 | #define RAM_SAVE_FLAG_EOS 0x10 | |
77 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
78 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
79 | /* 0x80 is reserved in migration.h start with 0x100 next */ | |
80 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 | |
81 | ||
56e93d26 JQ |
82 | static inline bool is_zero_range(uint8_t *p, uint64_t size) |
83 | { | |
a1febc49 | 84 | return buffer_is_zero(p, size); |
56e93d26 JQ |
85 | } |
86 | ||
9360447d JQ |
87 | XBZRLECacheStats xbzrle_counters; |
88 | ||
56e93d26 JQ |
89 | /* struct contains XBZRLE cache and a static page |
90 | used by the compression */ | |
91 | static struct { | |
92 | /* buffer used for XBZRLE encoding */ | |
93 | uint8_t *encoded_buf; | |
94 | /* buffer for storing page content */ | |
95 | uint8_t *current_buf; | |
96 | /* Cache for XBZRLE, Protected by lock. */ | |
97 | PageCache *cache; | |
98 | QemuMutex lock; | |
c00e0928 JQ |
99 | /* it will store a page full of zeros */ |
100 | uint8_t *zero_target_page; | |
f265e0e4 JQ |
101 | /* buffer used for XBZRLE decoding */ |
102 | uint8_t *decoded_buf; | |
56e93d26 JQ |
103 | } XBZRLE; |
104 | ||
56e93d26 JQ |
105 | static void XBZRLE_cache_lock(void) |
106 | { | |
f4c51a6b | 107 | if (migrate_use_xbzrle()) { |
56e93d26 | 108 | qemu_mutex_lock(&XBZRLE.lock); |
f4c51a6b | 109 | } |
56e93d26 JQ |
110 | } |
111 | ||
112 | static void XBZRLE_cache_unlock(void) | |
113 | { | |
f4c51a6b | 114 | if (migrate_use_xbzrle()) { |
56e93d26 | 115 | qemu_mutex_unlock(&XBZRLE.lock); |
f4c51a6b | 116 | } |
56e93d26 JQ |
117 | } |
118 | ||
3d0684b2 JQ |
119 | /** |
120 | * xbzrle_cache_resize: resize the xbzrle cache | |
121 | * | |
cbde7be9 | 122 | * This function is called from migrate_params_apply in main |
3d0684b2 JQ |
123 | * thread, possibly while a migration is in progress. A running |
124 | * migration may be using the cache and might finish during this call, | |
125 | * hence changes to the cache are protected by XBZRLE.lock(). | |
126 | * | |
c9dede2d | 127 | * Returns 0 for success or -1 for error |
3d0684b2 JQ |
128 | * |
129 | * @new_size: new cache size | |
8acabf69 | 130 | * @errp: set *errp if the check failed, with reason |
56e93d26 | 131 | */ |
8b9407a0 | 132 | int xbzrle_cache_resize(uint64_t new_size, Error **errp) |
56e93d26 JQ |
133 | { |
134 | PageCache *new_cache; | |
c9dede2d | 135 | int64_t ret = 0; |
56e93d26 | 136 | |
8acabf69 JQ |
137 | /* Check for truncation */ |
138 | if (new_size != (size_t)new_size) { | |
139 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
140 | "exceeding address space"); | |
141 | return -1; | |
142 | } | |
143 | ||
2a313e5c JQ |
144 | if (new_size == migrate_xbzrle_cache_size()) { |
145 | /* nothing to do */ | |
c9dede2d | 146 | return 0; |
2a313e5c JQ |
147 | } |
148 | ||
56e93d26 JQ |
149 | XBZRLE_cache_lock(); |
150 | ||
151 | if (XBZRLE.cache != NULL) { | |
80f8dfde | 152 | new_cache = cache_init(new_size, TARGET_PAGE_SIZE, errp); |
56e93d26 | 153 | if (!new_cache) { |
56e93d26 JQ |
154 | ret = -1; |
155 | goto out; | |
156 | } | |
157 | ||
158 | cache_fini(XBZRLE.cache); | |
159 | XBZRLE.cache = new_cache; | |
160 | } | |
56e93d26 JQ |
161 | out: |
162 | XBZRLE_cache_unlock(); | |
163 | return ret; | |
164 | } | |
165 | ||
3ded54b1 | 166 | bool ramblock_is_ignored(RAMBlock *block) |
fbd162e6 YK |
167 | { |
168 | return !qemu_ram_is_migratable(block) || | |
169 | (migrate_ignore_shared() && qemu_ram_is_shared(block)); | |
170 | } | |
171 | ||
343f632c DDAG |
172 | #undef RAMBLOCK_FOREACH |
173 | ||
fbd162e6 YK |
174 | int foreach_not_ignored_block(RAMBlockIterFunc func, void *opaque) |
175 | { | |
176 | RAMBlock *block; | |
177 | int ret = 0; | |
178 | ||
89ac5a1d DDAG |
179 | RCU_READ_LOCK_GUARD(); |
180 | ||
fbd162e6 YK |
181 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
182 | ret = func(block, opaque); | |
183 | if (ret) { | |
184 | break; | |
185 | } | |
186 | } | |
fbd162e6 YK |
187 | return ret; |
188 | } | |
189 | ||
f9494614 AP |
190 | static void ramblock_recv_map_init(void) |
191 | { | |
192 | RAMBlock *rb; | |
193 | ||
fbd162e6 | 194 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
195 | assert(!rb->receivedmap); |
196 | rb->receivedmap = bitmap_new(rb->max_length >> qemu_target_page_bits()); | |
197 | } | |
198 | } | |
199 | ||
200 | int ramblock_recv_bitmap_test(RAMBlock *rb, void *host_addr) | |
201 | { | |
202 | return test_bit(ramblock_recv_bitmap_offset(host_addr, rb), | |
203 | rb->receivedmap); | |
204 | } | |
205 | ||
1cba9f6e DDAG |
206 | bool ramblock_recv_bitmap_test_byte_offset(RAMBlock *rb, uint64_t byte_offset) |
207 | { | |
208 | return test_bit(byte_offset >> TARGET_PAGE_BITS, rb->receivedmap); | |
209 | } | |
210 | ||
f9494614 AP |
211 | void ramblock_recv_bitmap_set(RAMBlock *rb, void *host_addr) |
212 | { | |
213 | set_bit_atomic(ramblock_recv_bitmap_offset(host_addr, rb), rb->receivedmap); | |
214 | } | |
215 | ||
216 | void ramblock_recv_bitmap_set_range(RAMBlock *rb, void *host_addr, | |
217 | size_t nr) | |
218 | { | |
219 | bitmap_set_atomic(rb->receivedmap, | |
220 | ramblock_recv_bitmap_offset(host_addr, rb), | |
221 | nr); | |
222 | } | |
223 | ||
a335debb PX |
224 | #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL) |
225 | ||
226 | /* | |
227 | * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes). | |
228 | * | |
229 | * Returns >0 if success with sent bytes, or <0 if error. | |
230 | */ | |
231 | int64_t ramblock_recv_bitmap_send(QEMUFile *file, | |
232 | const char *block_name) | |
233 | { | |
234 | RAMBlock *block = qemu_ram_block_by_name(block_name); | |
235 | unsigned long *le_bitmap, nbits; | |
236 | uint64_t size; | |
237 | ||
238 | if (!block) { | |
239 | error_report("%s: invalid block name: %s", __func__, block_name); | |
240 | return -1; | |
241 | } | |
242 | ||
898ba906 | 243 | nbits = block->postcopy_length >> TARGET_PAGE_BITS; |
a335debb PX |
244 | |
245 | /* | |
246 | * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit | |
247 | * machines we may need 4 more bytes for padding (see below | |
248 | * comment). So extend it a bit before hand. | |
249 | */ | |
250 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
251 | ||
252 | /* | |
253 | * Always use little endian when sending the bitmap. This is | |
254 | * required that when source and destination VMs are not using the | |
3a4452d8 | 255 | * same endianness. (Note: big endian won't work.) |
a335debb PX |
256 | */ |
257 | bitmap_to_le(le_bitmap, block->receivedmap, nbits); | |
258 | ||
259 | /* Size of the bitmap, in bytes */ | |
a725ef9f | 260 | size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
261 | |
262 | /* | |
263 | * size is always aligned to 8 bytes for 64bit machines, but it | |
264 | * may not be true for 32bit machines. We need this padding to | |
265 | * make sure the migration can survive even between 32bit and | |
266 | * 64bit machines. | |
267 | */ | |
268 | size = ROUND_UP(size, 8); | |
269 | ||
270 | qemu_put_be64(file, size); | |
271 | qemu_put_buffer(file, (const uint8_t *)le_bitmap, size); | |
272 | /* | |
273 | * Mark as an end, in case the middle part is screwed up due to | |
3a4452d8 | 274 | * some "mysterious" reason. |
a335debb PX |
275 | */ |
276 | qemu_put_be64(file, RAMBLOCK_RECV_BITMAP_ENDING); | |
277 | qemu_fflush(file); | |
278 | ||
bf269906 | 279 | g_free(le_bitmap); |
a335debb PX |
280 | |
281 | if (qemu_file_get_error(file)) { | |
282 | return qemu_file_get_error(file); | |
283 | } | |
284 | ||
285 | return size + sizeof(size); | |
286 | } | |
287 | ||
ec481c6c JQ |
288 | /* |
289 | * An outstanding page request, on the source, having been received | |
290 | * and queued | |
291 | */ | |
292 | struct RAMSrcPageRequest { | |
293 | RAMBlock *rb; | |
294 | hwaddr offset; | |
295 | hwaddr len; | |
296 | ||
297 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
298 | }; | |
299 | ||
6f37bb8b JQ |
300 | /* State of RAM for migration */ |
301 | struct RAMState { | |
204b88b8 JQ |
302 | /* QEMUFile used for this migration */ |
303 | QEMUFile *f; | |
278e2f55 AG |
304 | /* UFFD file descriptor, used in 'write-tracking' migration */ |
305 | int uffdio_fd; | |
6f37bb8b JQ |
306 | /* Last block that we have visited searching for dirty pages */ |
307 | RAMBlock *last_seen_block; | |
308 | /* Last block from where we have sent data */ | |
309 | RAMBlock *last_sent_block; | |
269ace29 JQ |
310 | /* Last dirty target page we have sent */ |
311 | ram_addr_t last_page; | |
6f37bb8b JQ |
312 | /* last ram version we have seen */ |
313 | uint32_t last_version; | |
8d820d6f JQ |
314 | /* How many times we have dirty too many pages */ |
315 | int dirty_rate_high_cnt; | |
f664da80 JQ |
316 | /* these variables are used for bitmap sync */ |
317 | /* last time we did a full bitmap_sync */ | |
318 | int64_t time_last_bitmap_sync; | |
eac74159 | 319 | /* bytes transferred at start_time */ |
c4bdf0cf | 320 | uint64_t bytes_xfer_prev; |
a66cd90c | 321 | /* number of dirty pages since start_time */ |
68908ed6 | 322 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
323 | /* xbzrle misses since the beginning of the period */ |
324 | uint64_t xbzrle_cache_miss_prev; | |
e460a4b1 WW |
325 | /* Amount of xbzrle pages since the beginning of the period */ |
326 | uint64_t xbzrle_pages_prev; | |
327 | /* Amount of xbzrle encoded bytes since the beginning of the period */ | |
328 | uint64_t xbzrle_bytes_prev; | |
1a373522 DH |
329 | /* Start using XBZRLE (e.g., after the first round). */ |
330 | bool xbzrle_enabled; | |
76e03000 XG |
331 | |
332 | /* compression statistics since the beginning of the period */ | |
333 | /* amount of count that no free thread to compress data */ | |
334 | uint64_t compress_thread_busy_prev; | |
335 | /* amount bytes after compression */ | |
336 | uint64_t compressed_size_prev; | |
337 | /* amount of compressed pages */ | |
338 | uint64_t compress_pages_prev; | |
339 | ||
be8b02ed XG |
340 | /* total handled target pages at the beginning of period */ |
341 | uint64_t target_page_count_prev; | |
342 | /* total handled target pages since start */ | |
343 | uint64_t target_page_count; | |
9360447d | 344 | /* number of dirty bits in the bitmap */ |
2dfaf12e | 345 | uint64_t migration_dirty_pages; |
386a907b | 346 | /* Protects modification of the bitmap and migration dirty pages */ |
108cfae0 | 347 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
348 | /* The RAMBlock used in the last src_page_requests */ |
349 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
350 | /* Queue of outstanding page requests from the destination */ |
351 | QemuMutex src_page_req_mutex; | |
b58deb34 | 352 | QSIMPLEQ_HEAD(, RAMSrcPageRequest) src_page_requests; |
6f37bb8b JQ |
353 | }; |
354 | typedef struct RAMState RAMState; | |
355 | ||
53518d94 | 356 | static RAMState *ram_state; |
6f37bb8b | 357 | |
bd227060 WW |
358 | static NotifierWithReturnList precopy_notifier_list; |
359 | ||
360 | void precopy_infrastructure_init(void) | |
361 | { | |
362 | notifier_with_return_list_init(&precopy_notifier_list); | |
363 | } | |
364 | ||
365 | void precopy_add_notifier(NotifierWithReturn *n) | |
366 | { | |
367 | notifier_with_return_list_add(&precopy_notifier_list, n); | |
368 | } | |
369 | ||
370 | void precopy_remove_notifier(NotifierWithReturn *n) | |
371 | { | |
372 | notifier_with_return_remove(n); | |
373 | } | |
374 | ||
375 | int precopy_notify(PrecopyNotifyReason reason, Error **errp) | |
376 | { | |
377 | PrecopyNotifyData pnd; | |
378 | pnd.reason = reason; | |
379 | pnd.errp = errp; | |
380 | ||
381 | return notifier_with_return_list_notify(&precopy_notifier_list, &pnd); | |
382 | } | |
383 | ||
9edabd4d | 384 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 385 | { |
bae416e5 DDAG |
386 | return ram_state ? (ram_state->migration_dirty_pages * TARGET_PAGE_SIZE) : |
387 | 0; | |
2f4fde93 JQ |
388 | } |
389 | ||
9360447d | 390 | MigrationStats ram_counters; |
96506894 | 391 | |
b8fb8cb7 DDAG |
392 | /* used by the search for pages to send */ |
393 | struct PageSearchStatus { | |
394 | /* Current block being searched */ | |
395 | RAMBlock *block; | |
a935e30f JQ |
396 | /* Current page to search from */ |
397 | unsigned long page; | |
b8fb8cb7 DDAG |
398 | /* Set once we wrap around */ |
399 | bool complete_round; | |
400 | }; | |
401 | typedef struct PageSearchStatus PageSearchStatus; | |
402 | ||
76e03000 XG |
403 | CompressionStats compression_counters; |
404 | ||
56e93d26 | 405 | struct CompressParam { |
56e93d26 | 406 | bool done; |
90e56fb4 | 407 | bool quit; |
5e5fdcff | 408 | bool zero_page; |
56e93d26 JQ |
409 | QEMUFile *file; |
410 | QemuMutex mutex; | |
411 | QemuCond cond; | |
412 | RAMBlock *block; | |
413 | ram_addr_t offset; | |
34ab9e97 XG |
414 | |
415 | /* internally used fields */ | |
dcaf446e | 416 | z_stream stream; |
34ab9e97 | 417 | uint8_t *originbuf; |
56e93d26 JQ |
418 | }; |
419 | typedef struct CompressParam CompressParam; | |
420 | ||
421 | struct DecompressParam { | |
73a8912b | 422 | bool done; |
90e56fb4 | 423 | bool quit; |
56e93d26 JQ |
424 | QemuMutex mutex; |
425 | QemuCond cond; | |
426 | void *des; | |
d341d9f3 | 427 | uint8_t *compbuf; |
56e93d26 | 428 | int len; |
797ca154 | 429 | z_stream stream; |
56e93d26 JQ |
430 | }; |
431 | typedef struct DecompressParam DecompressParam; | |
432 | ||
433 | static CompressParam *comp_param; | |
434 | static QemuThread *compress_threads; | |
435 | /* comp_done_cond is used to wake up the migration thread when | |
436 | * one of the compression threads has finished the compression. | |
437 | * comp_done_lock is used to co-work with comp_done_cond. | |
438 | */ | |
0d9f9a5c LL |
439 | static QemuMutex comp_done_lock; |
440 | static QemuCond comp_done_cond; | |
56e93d26 JQ |
441 | /* The empty QEMUFileOps will be used by file in CompressParam */ |
442 | static const QEMUFileOps empty_ops = { }; | |
443 | ||
34ab9e97 | 444 | static QEMUFile *decomp_file; |
56e93d26 JQ |
445 | static DecompressParam *decomp_param; |
446 | static QemuThread *decompress_threads; | |
73a8912b LL |
447 | static QemuMutex decomp_done_lock; |
448 | static QemuCond decomp_done_cond; | |
56e93d26 | 449 | |
5e5fdcff | 450 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 451 | ram_addr_t offset, uint8_t *source_buf); |
56e93d26 JQ |
452 | |
453 | static void *do_data_compress(void *opaque) | |
454 | { | |
455 | CompressParam *param = opaque; | |
a7a9a88f LL |
456 | RAMBlock *block; |
457 | ram_addr_t offset; | |
5e5fdcff | 458 | bool zero_page; |
56e93d26 | 459 | |
a7a9a88f | 460 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 461 | while (!param->quit) { |
a7a9a88f LL |
462 | if (param->block) { |
463 | block = param->block; | |
464 | offset = param->offset; | |
465 | param->block = NULL; | |
466 | qemu_mutex_unlock(¶m->mutex); | |
467 | ||
5e5fdcff XG |
468 | zero_page = do_compress_ram_page(param->file, ¶m->stream, |
469 | block, offset, param->originbuf); | |
a7a9a88f | 470 | |
0d9f9a5c | 471 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 472 | param->done = true; |
5e5fdcff | 473 | param->zero_page = zero_page; |
0d9f9a5c LL |
474 | qemu_cond_signal(&comp_done_cond); |
475 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
476 | |
477 | qemu_mutex_lock(¶m->mutex); | |
478 | } else { | |
56e93d26 JQ |
479 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
480 | } | |
56e93d26 | 481 | } |
a7a9a88f | 482 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
483 | |
484 | return NULL; | |
485 | } | |
486 | ||
f0afa331 | 487 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
488 | { |
489 | int i, thread_count; | |
490 | ||
05306935 | 491 | if (!migrate_use_compression() || !comp_param) { |
56e93d26 JQ |
492 | return; |
493 | } | |
05306935 | 494 | |
56e93d26 JQ |
495 | thread_count = migrate_compress_threads(); |
496 | for (i = 0; i < thread_count; i++) { | |
dcaf446e XG |
497 | /* |
498 | * we use it as a indicator which shows if the thread is | |
499 | * properly init'd or not | |
500 | */ | |
501 | if (!comp_param[i].file) { | |
502 | break; | |
503 | } | |
05306935 FL |
504 | |
505 | qemu_mutex_lock(&comp_param[i].mutex); | |
506 | comp_param[i].quit = true; | |
507 | qemu_cond_signal(&comp_param[i].cond); | |
508 | qemu_mutex_unlock(&comp_param[i].mutex); | |
509 | ||
56e93d26 | 510 | qemu_thread_join(compress_threads + i); |
56e93d26 JQ |
511 | qemu_mutex_destroy(&comp_param[i].mutex); |
512 | qemu_cond_destroy(&comp_param[i].cond); | |
dcaf446e | 513 | deflateEnd(&comp_param[i].stream); |
34ab9e97 | 514 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
515 | qemu_fclose(comp_param[i].file); |
516 | comp_param[i].file = NULL; | |
56e93d26 | 517 | } |
0d9f9a5c LL |
518 | qemu_mutex_destroy(&comp_done_lock); |
519 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
520 | g_free(compress_threads); |
521 | g_free(comp_param); | |
56e93d26 JQ |
522 | compress_threads = NULL; |
523 | comp_param = NULL; | |
56e93d26 JQ |
524 | } |
525 | ||
dcaf446e | 526 | static int compress_threads_save_setup(void) |
56e93d26 JQ |
527 | { |
528 | int i, thread_count; | |
529 | ||
530 | if (!migrate_use_compression()) { | |
dcaf446e | 531 | return 0; |
56e93d26 | 532 | } |
56e93d26 JQ |
533 | thread_count = migrate_compress_threads(); |
534 | compress_threads = g_new0(QemuThread, thread_count); | |
535 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
536 | qemu_cond_init(&comp_done_cond); |
537 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 538 | for (i = 0; i < thread_count; i++) { |
34ab9e97 XG |
539 | comp_param[i].originbuf = g_try_malloc(TARGET_PAGE_SIZE); |
540 | if (!comp_param[i].originbuf) { | |
541 | goto exit; | |
542 | } | |
543 | ||
dcaf446e XG |
544 | if (deflateInit(&comp_param[i].stream, |
545 | migrate_compress_level()) != Z_OK) { | |
34ab9e97 | 546 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
547 | goto exit; |
548 | } | |
549 | ||
e110aa91 C |
550 | /* comp_param[i].file is just used as a dummy buffer to save data, |
551 | * set its ops to empty. | |
56e93d26 | 552 | */ |
c6ad5be7 | 553 | comp_param[i].file = qemu_fopen_ops(NULL, &empty_ops, false); |
56e93d26 | 554 | comp_param[i].done = true; |
90e56fb4 | 555 | comp_param[i].quit = false; |
56e93d26 JQ |
556 | qemu_mutex_init(&comp_param[i].mutex); |
557 | qemu_cond_init(&comp_param[i].cond); | |
558 | qemu_thread_create(compress_threads + i, "compress", | |
559 | do_data_compress, comp_param + i, | |
560 | QEMU_THREAD_JOINABLE); | |
561 | } | |
dcaf446e XG |
562 | return 0; |
563 | ||
564 | exit: | |
565 | compress_threads_save_cleanup(); | |
566 | return -1; | |
56e93d26 JQ |
567 | } |
568 | ||
569 | /** | |
3d0684b2 | 570 | * save_page_header: write page header to wire |
56e93d26 JQ |
571 | * |
572 | * If this is the 1st block, it also writes the block identification | |
573 | * | |
3d0684b2 | 574 | * Returns the number of bytes written |
56e93d26 JQ |
575 | * |
576 | * @f: QEMUFile where to send the data | |
577 | * @block: block that contains the page we want to send | |
578 | * @offset: offset inside the block for the page | |
579 | * in the lower bits, it contains flags | |
580 | */ | |
2bf3aa85 JQ |
581 | static size_t save_page_header(RAMState *rs, QEMUFile *f, RAMBlock *block, |
582 | ram_addr_t offset) | |
56e93d26 | 583 | { |
9f5f380b | 584 | size_t size, len; |
56e93d26 | 585 | |
24795694 JQ |
586 | if (block == rs->last_sent_block) { |
587 | offset |= RAM_SAVE_FLAG_CONTINUE; | |
588 | } | |
2bf3aa85 | 589 | qemu_put_be64(f, offset); |
56e93d26 JQ |
590 | size = 8; |
591 | ||
592 | if (!(offset & RAM_SAVE_FLAG_CONTINUE)) { | |
9f5f380b | 593 | len = strlen(block->idstr); |
2bf3aa85 JQ |
594 | qemu_put_byte(f, len); |
595 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 596 | size += 1 + len; |
24795694 | 597 | rs->last_sent_block = block; |
56e93d26 JQ |
598 | } |
599 | return size; | |
600 | } | |
601 | ||
3d0684b2 | 602 | /** |
179a8080 | 603 | * mig_throttle_guest_down: throttle down the guest |
3d0684b2 JQ |
604 | * |
605 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
606 | * writes. If guest dirty memory rate is reduced below the rate at | |
607 | * which we can transfer pages to the destination then we should be | |
608 | * able to complete migration. Some workloads dirty memory way too | |
609 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 | 610 | */ |
cbbf8182 KZ |
611 | static void mig_throttle_guest_down(uint64_t bytes_dirty_period, |
612 | uint64_t bytes_dirty_threshold) | |
070afca2 JH |
613 | { |
614 | MigrationState *s = migrate_get_current(); | |
2594f56d | 615 | uint64_t pct_initial = s->parameters.cpu_throttle_initial; |
cbbf8182 KZ |
616 | uint64_t pct_increment = s->parameters.cpu_throttle_increment; |
617 | bool pct_tailslow = s->parameters.cpu_throttle_tailslow; | |
4cbc9c7f | 618 | int pct_max = s->parameters.max_cpu_throttle; |
070afca2 | 619 | |
cbbf8182 KZ |
620 | uint64_t throttle_now = cpu_throttle_get_percentage(); |
621 | uint64_t cpu_now, cpu_ideal, throttle_inc; | |
622 | ||
070afca2 JH |
623 | /* We have not started throttling yet. Let's start it. */ |
624 | if (!cpu_throttle_active()) { | |
625 | cpu_throttle_set(pct_initial); | |
626 | } else { | |
627 | /* Throttling already on, just increase the rate */ | |
cbbf8182 KZ |
628 | if (!pct_tailslow) { |
629 | throttle_inc = pct_increment; | |
630 | } else { | |
631 | /* Compute the ideal CPU percentage used by Guest, which may | |
632 | * make the dirty rate match the dirty rate threshold. */ | |
633 | cpu_now = 100 - throttle_now; | |
634 | cpu_ideal = cpu_now * (bytes_dirty_threshold * 1.0 / | |
635 | bytes_dirty_period); | |
636 | throttle_inc = MIN(cpu_now - cpu_ideal, pct_increment); | |
637 | } | |
638 | cpu_throttle_set(MIN(throttle_now + throttle_inc, pct_max)); | |
070afca2 JH |
639 | } |
640 | } | |
641 | ||
3d0684b2 JQ |
642 | /** |
643 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
644 | * | |
6f37bb8b | 645 | * @rs: current RAM state |
3d0684b2 JQ |
646 | * @current_addr: address for the zero page |
647 | * | |
648 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
649 | * The important thing is that a stale (not-yet-0'd) page be replaced |
650 | * by the new data. | |
651 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 652 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 653 | */ |
6f37bb8b | 654 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 655 | { |
1a373522 | 656 | if (!rs->xbzrle_enabled) { |
56e93d26 JQ |
657 | return; |
658 | } | |
659 | ||
660 | /* We don't care if this fails to allocate a new cache page | |
661 | * as long as it updated an old one */ | |
c00e0928 | 662 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
9360447d | 663 | ram_counters.dirty_sync_count); |
56e93d26 JQ |
664 | } |
665 | ||
666 | #define ENCODING_FLAG_XBZRLE 0x1 | |
667 | ||
668 | /** | |
669 | * save_xbzrle_page: compress and send current page | |
670 | * | |
671 | * Returns: 1 means that we wrote the page | |
672 | * 0 means that page is identical to the one already sent | |
673 | * -1 means that xbzrle would be longer than normal | |
674 | * | |
5a987738 | 675 | * @rs: current RAM state |
3d0684b2 JQ |
676 | * @current_data: pointer to the address of the page contents |
677 | * @current_addr: addr of the page | |
56e93d26 JQ |
678 | * @block: block that contains the page we want to send |
679 | * @offset: offset inside the block for the page | |
680 | * @last_stage: if we are at the completion stage | |
56e93d26 | 681 | */ |
204b88b8 | 682 | static int save_xbzrle_page(RAMState *rs, uint8_t **current_data, |
56e93d26 | 683 | ram_addr_t current_addr, RAMBlock *block, |
072c2511 | 684 | ram_addr_t offset, bool last_stage) |
56e93d26 JQ |
685 | { |
686 | int encoded_len = 0, bytes_xbzrle; | |
687 | uint8_t *prev_cached_page; | |
688 | ||
9360447d JQ |
689 | if (!cache_is_cached(XBZRLE.cache, current_addr, |
690 | ram_counters.dirty_sync_count)) { | |
691 | xbzrle_counters.cache_miss++; | |
56e93d26 JQ |
692 | if (!last_stage) { |
693 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, | |
9360447d | 694 | ram_counters.dirty_sync_count) == -1) { |
56e93d26 JQ |
695 | return -1; |
696 | } else { | |
697 | /* update *current_data when the page has been | |
698 | inserted into cache */ | |
699 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
700 | } | |
701 | } | |
702 | return -1; | |
703 | } | |
704 | ||
e460a4b1 WW |
705 | /* |
706 | * Reaching here means the page has hit the xbzrle cache, no matter what | |
707 | * encoding result it is (normal encoding, overflow or skipping the page), | |
3a4452d8 | 708 | * count the page as encoded. This is used to calculate the encoding rate. |
e460a4b1 WW |
709 | * |
710 | * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB, | |
711 | * 2nd page turns out to be skipped (i.e. no new bytes written to the | |
712 | * page), the overall encoding rate will be 8KB / 2KB = 4, which has the | |
713 | * skipped page included. In this way, the encoding rate can tell if the | |
714 | * guest page is good for xbzrle encoding. | |
715 | */ | |
716 | xbzrle_counters.pages++; | |
56e93d26 JQ |
717 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); |
718 | ||
719 | /* save current buffer into memory */ | |
720 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
721 | ||
722 | /* XBZRLE encoding (if there is no overflow) */ | |
723 | encoded_len = xbzrle_encode_buffer(prev_cached_page, XBZRLE.current_buf, | |
724 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
725 | TARGET_PAGE_SIZE); | |
ca353803 WY |
726 | |
727 | /* | |
728 | * Update the cache contents, so that it corresponds to the data | |
729 | * sent, in all cases except where we skip the page. | |
730 | */ | |
731 | if (!last_stage && encoded_len != 0) { | |
732 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); | |
733 | /* | |
734 | * In the case where we couldn't compress, ensure that the caller | |
735 | * sends the data from the cache, since the guest might have | |
736 | * changed the RAM since we copied it. | |
737 | */ | |
738 | *current_data = prev_cached_page; | |
739 | } | |
740 | ||
56e93d26 | 741 | if (encoded_len == 0) { |
55c4446b | 742 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
743 | return 0; |
744 | } else if (encoded_len == -1) { | |
55c4446b | 745 | trace_save_xbzrle_page_overflow(); |
9360447d | 746 | xbzrle_counters.overflow++; |
e460a4b1 | 747 | xbzrle_counters.bytes += TARGET_PAGE_SIZE; |
56e93d26 JQ |
748 | return -1; |
749 | } | |
750 | ||
56e93d26 | 751 | /* Send XBZRLE based compressed page */ |
2bf3aa85 | 752 | bytes_xbzrle = save_page_header(rs, rs->f, block, |
204b88b8 JQ |
753 | offset | RAM_SAVE_FLAG_XBZRLE); |
754 | qemu_put_byte(rs->f, ENCODING_FLAG_XBZRLE); | |
755 | qemu_put_be16(rs->f, encoded_len); | |
756 | qemu_put_buffer(rs->f, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 757 | bytes_xbzrle += encoded_len + 1 + 2; |
e460a4b1 WW |
758 | /* |
759 | * Like compressed_size (please see update_compress_thread_counts), | |
760 | * the xbzrle encoded bytes don't count the 8 byte header with | |
761 | * RAM_SAVE_FLAG_CONTINUE. | |
762 | */ | |
763 | xbzrle_counters.bytes += bytes_xbzrle - 8; | |
9360447d | 764 | ram_counters.transferred += bytes_xbzrle; |
56e93d26 JQ |
765 | |
766 | return 1; | |
767 | } | |
768 | ||
3d0684b2 JQ |
769 | /** |
770 | * migration_bitmap_find_dirty: find the next dirty page from start | |
f3f491fc | 771 | * |
a5f7b1a6 | 772 | * Returns the page offset within memory region of the start of a dirty page |
3d0684b2 | 773 | * |
6f37bb8b | 774 | * @rs: current RAM state |
3d0684b2 | 775 | * @rb: RAMBlock where to search for dirty pages |
a935e30f | 776 | * @start: page where we start the search |
f3f491fc | 777 | */ |
56e93d26 | 778 | static inline |
a935e30f | 779 | unsigned long migration_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, |
f20e2865 | 780 | unsigned long start) |
56e93d26 | 781 | { |
6b6712ef JQ |
782 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
783 | unsigned long *bitmap = rb->bmap; | |
56e93d26 | 784 | |
fbd162e6 | 785 | if (ramblock_is_ignored(rb)) { |
b895de50 CLG |
786 | return size; |
787 | } | |
788 | ||
1a373522 | 789 | return find_next_bit(bitmap, size, start); |
56e93d26 JQ |
790 | } |
791 | ||
1230a25f | 792 | static void migration_clear_memory_region_dirty_bitmap(RAMBlock *rb, |
3143577d WW |
793 | unsigned long page) |
794 | { | |
795 | uint8_t shift; | |
796 | hwaddr size, start; | |
797 | ||
798 | if (!rb->clear_bmap || !clear_bmap_test_and_clear(rb, page)) { | |
799 | return; | |
800 | } | |
801 | ||
802 | shift = rb->clear_bmap_shift; | |
803 | /* | |
804 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
805 | * can make things easier sometimes since then start address | |
806 | * of the small chunk will always be 64 pages aligned so the | |
807 | * bitmap will always be aligned to unsigned long. We should | |
808 | * even be able to remove this restriction but I'm simply | |
809 | * keeping it. | |
810 | */ | |
811 | assert(shift >= 6); | |
812 | ||
813 | size = 1ULL << (TARGET_PAGE_BITS + shift); | |
7648297d | 814 | start = QEMU_ALIGN_DOWN((ram_addr_t)page << TARGET_PAGE_BITS, size); |
3143577d WW |
815 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); |
816 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
817 | } | |
818 | ||
819 | static void | |
1230a25f | 820 | migration_clear_memory_region_dirty_bitmap_range(RAMBlock *rb, |
3143577d WW |
821 | unsigned long start, |
822 | unsigned long npages) | |
823 | { | |
824 | unsigned long i, chunk_pages = 1UL << rb->clear_bmap_shift; | |
825 | unsigned long chunk_start = QEMU_ALIGN_DOWN(start, chunk_pages); | |
826 | unsigned long chunk_end = QEMU_ALIGN_UP(start + npages, chunk_pages); | |
827 | ||
828 | /* | |
829 | * Clear pages from start to start + npages - 1, so the end boundary is | |
830 | * exclusive. | |
831 | */ | |
832 | for (i = chunk_start; i < chunk_end; i += chunk_pages) { | |
1230a25f | 833 | migration_clear_memory_region_dirty_bitmap(rb, i); |
3143577d WW |
834 | } |
835 | } | |
836 | ||
06b10688 | 837 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
838 | RAMBlock *rb, |
839 | unsigned long page) | |
a82d593b DDAG |
840 | { |
841 | bool ret; | |
a82d593b | 842 | |
002cad6b PX |
843 | /* |
844 | * Clear dirty bitmap if needed. This _must_ be called before we | |
845 | * send any of the page in the chunk because we need to make sure | |
846 | * we can capture further page content changes when we sync dirty | |
847 | * log the next time. So as long as we are going to send any of | |
848 | * the page in the chunk we clear the remote dirty bitmap for all. | |
849 | * Clearing it earlier won't be a problem, but too late will. | |
850 | */ | |
1230a25f | 851 | migration_clear_memory_region_dirty_bitmap(rb, page); |
002cad6b | 852 | |
6b6712ef | 853 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b | 854 | if (ret) { |
0d8ec885 | 855 | rs->migration_dirty_pages--; |
a82d593b | 856 | } |
386a907b | 857 | |
a82d593b DDAG |
858 | return ret; |
859 | } | |
860 | ||
be39b4cd DH |
861 | static void dirty_bitmap_clear_section(MemoryRegionSection *section, |
862 | void *opaque) | |
863 | { | |
864 | const hwaddr offset = section->offset_within_region; | |
865 | const hwaddr size = int128_get64(section->size); | |
866 | const unsigned long start = offset >> TARGET_PAGE_BITS; | |
867 | const unsigned long npages = size >> TARGET_PAGE_BITS; | |
868 | RAMBlock *rb = section->mr->ram_block; | |
869 | uint64_t *cleared_bits = opaque; | |
870 | ||
871 | /* | |
872 | * We don't grab ram_state->bitmap_mutex because we expect to run | |
873 | * only when starting migration or during postcopy recovery where | |
874 | * we don't have concurrent access. | |
875 | */ | |
876 | if (!migration_in_postcopy() && !migrate_background_snapshot()) { | |
877 | migration_clear_memory_region_dirty_bitmap_range(rb, start, npages); | |
878 | } | |
879 | *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages); | |
880 | bitmap_clear(rb->bmap, start, npages); | |
881 | } | |
882 | ||
883 | /* | |
884 | * Exclude all dirty pages from migration that fall into a discarded range as | |
885 | * managed by a RamDiscardManager responsible for the mapped memory region of | |
886 | * the RAMBlock. Clear the corresponding bits in the dirty bitmaps. | |
887 | * | |
888 | * Discarded pages ("logically unplugged") have undefined content and must | |
889 | * not get migrated, because even reading these pages for migration might | |
890 | * result in undesired behavior. | |
891 | * | |
892 | * Returns the number of cleared bits in the RAMBlock dirty bitmap. | |
893 | * | |
894 | * Note: The result is only stable while migrating (precopy/postcopy). | |
895 | */ | |
896 | static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb) | |
897 | { | |
898 | uint64_t cleared_bits = 0; | |
899 | ||
900 | if (rb->mr && rb->bmap && memory_region_has_ram_discard_manager(rb->mr)) { | |
901 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
902 | MemoryRegionSection section = { | |
903 | .mr = rb->mr, | |
904 | .offset_within_region = 0, | |
905 | .size = int128_make64(qemu_ram_get_used_length(rb)), | |
906 | }; | |
907 | ||
908 | ram_discard_manager_replay_discarded(rdm, §ion, | |
909 | dirty_bitmap_clear_section, | |
910 | &cleared_bits); | |
911 | } | |
912 | return cleared_bits; | |
913 | } | |
914 | ||
9470c5e0 DH |
915 | /* |
916 | * Check if a host-page aligned page falls into a discarded range as managed by | |
917 | * a RamDiscardManager responsible for the mapped memory region of the RAMBlock. | |
918 | * | |
919 | * Note: The result is only stable while migrating (precopy/postcopy). | |
920 | */ | |
921 | bool ramblock_page_is_discarded(RAMBlock *rb, ram_addr_t start) | |
922 | { | |
923 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
924 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
925 | MemoryRegionSection section = { | |
926 | .mr = rb->mr, | |
927 | .offset_within_region = start, | |
928 | .size = int128_make64(qemu_ram_pagesize(rb)), | |
929 | }; | |
930 | ||
931 | return !ram_discard_manager_is_populated(rdm, §ion); | |
932 | } | |
933 | return false; | |
934 | } | |
935 | ||
267691b6 | 936 | /* Called with RCU critical section */ |
7a3e9571 | 937 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 938 | { |
fb613580 KZ |
939 | uint64_t new_dirty_pages = |
940 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length); | |
941 | ||
942 | rs->migration_dirty_pages += new_dirty_pages; | |
943 | rs->num_dirty_pages_period += new_dirty_pages; | |
56e93d26 JQ |
944 | } |
945 | ||
3d0684b2 JQ |
946 | /** |
947 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
948 | * | |
949 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
950 | * | |
951 | * For VMs with just normal pages this is equivalent to the host page | |
952 | * size. If it's got some huge pages then it's the OR of all the | |
953 | * different page sizes. | |
e8ca1db2 DDAG |
954 | */ |
955 | uint64_t ram_pagesize_summary(void) | |
956 | { | |
957 | RAMBlock *block; | |
958 | uint64_t summary = 0; | |
959 | ||
fbd162e6 | 960 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
961 | summary |= block->page_size; |
962 | } | |
963 | ||
964 | return summary; | |
965 | } | |
966 | ||
aecbfe9c XG |
967 | uint64_t ram_get_total_transferred_pages(void) |
968 | { | |
969 | return ram_counters.normal + ram_counters.duplicate + | |
970 | compression_counters.pages + xbzrle_counters.pages; | |
971 | } | |
972 | ||
b734035b XG |
973 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
974 | { | |
be8b02ed | 975 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 976 | double compressed_size; |
b734035b XG |
977 | |
978 | /* calculate period counters */ | |
979 | ram_counters.dirty_pages_rate = rs->num_dirty_pages_period * 1000 | |
980 | / (end_time - rs->time_last_bitmap_sync); | |
981 | ||
be8b02ed | 982 | if (!page_count) { |
b734035b XG |
983 | return; |
984 | } | |
985 | ||
986 | if (migrate_use_xbzrle()) { | |
e460a4b1 WW |
987 | double encoded_size, unencoded_size; |
988 | ||
b734035b | 989 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - |
be8b02ed | 990 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b | 991 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
e460a4b1 WW |
992 | unencoded_size = (xbzrle_counters.pages - rs->xbzrle_pages_prev) * |
993 | TARGET_PAGE_SIZE; | |
994 | encoded_size = xbzrle_counters.bytes - rs->xbzrle_bytes_prev; | |
92271402 | 995 | if (xbzrle_counters.pages == rs->xbzrle_pages_prev || !encoded_size) { |
e460a4b1 | 996 | xbzrle_counters.encoding_rate = 0; |
e460a4b1 WW |
997 | } else { |
998 | xbzrle_counters.encoding_rate = unencoded_size / encoded_size; | |
999 | } | |
1000 | rs->xbzrle_pages_prev = xbzrle_counters.pages; | |
1001 | rs->xbzrle_bytes_prev = xbzrle_counters.bytes; | |
b734035b | 1002 | } |
76e03000 XG |
1003 | |
1004 | if (migrate_use_compression()) { | |
1005 | compression_counters.busy_rate = (double)(compression_counters.busy - | |
1006 | rs->compress_thread_busy_prev) / page_count; | |
1007 | rs->compress_thread_busy_prev = compression_counters.busy; | |
1008 | ||
1009 | compressed_size = compression_counters.compressed_size - | |
1010 | rs->compressed_size_prev; | |
1011 | if (compressed_size) { | |
1012 | double uncompressed_size = (compression_counters.pages - | |
1013 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
1014 | ||
1015 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
1016 | compression_counters.compression_rate = | |
1017 | uncompressed_size / compressed_size; | |
1018 | ||
1019 | rs->compress_pages_prev = compression_counters.pages; | |
1020 | rs->compressed_size_prev = compression_counters.compressed_size; | |
1021 | } | |
1022 | } | |
b734035b XG |
1023 | } |
1024 | ||
dc14a470 KZ |
1025 | static void migration_trigger_throttle(RAMState *rs) |
1026 | { | |
1027 | MigrationState *s = migrate_get_current(); | |
1028 | uint64_t threshold = s->parameters.throttle_trigger_threshold; | |
1029 | ||
1030 | uint64_t bytes_xfer_period = ram_counters.transferred - rs->bytes_xfer_prev; | |
1031 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; | |
1032 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
1033 | ||
1034 | /* During block migration the auto-converge logic incorrectly detects | |
1035 | * that ram migration makes no progress. Avoid this by disabling the | |
1036 | * throttling logic during the bulk phase of block migration. */ | |
1037 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
1038 | /* The following detection logic can be refined later. For now: | |
1039 | Check to see if the ratio between dirtied bytes and the approx. | |
1040 | amount of bytes that just got transferred since the last time | |
1041 | we were in this routine reaches the threshold. If that happens | |
1042 | twice, start or increase throttling. */ | |
1043 | ||
1044 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
1045 | (++rs->dirty_rate_high_cnt >= 2)) { | |
1046 | trace_migration_throttle(); | |
1047 | rs->dirty_rate_high_cnt = 0; | |
cbbf8182 KZ |
1048 | mig_throttle_guest_down(bytes_dirty_period, |
1049 | bytes_dirty_threshold); | |
dc14a470 KZ |
1050 | } |
1051 | } | |
1052 | } | |
1053 | ||
8d820d6f | 1054 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
1055 | { |
1056 | RAMBlock *block; | |
56e93d26 | 1057 | int64_t end_time; |
56e93d26 | 1058 | |
9360447d | 1059 | ram_counters.dirty_sync_count++; |
56e93d26 | 1060 | |
f664da80 JQ |
1061 | if (!rs->time_last_bitmap_sync) { |
1062 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
1063 | } |
1064 | ||
1065 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 1066 | memory_global_dirty_log_sync(); |
56e93d26 | 1067 | |
108cfae0 | 1068 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
1069 | WITH_RCU_READ_LOCK_GUARD() { |
1070 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1071 | ramblock_sync_dirty_bitmap(rs, block); | |
1072 | } | |
1073 | ram_counters.remaining = ram_bytes_remaining(); | |
56e93d26 | 1074 | } |
108cfae0 | 1075 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 1076 | |
9458a9a1 | 1077 | memory_global_after_dirty_log_sync(); |
a66cd90c | 1078 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 1079 | |
56e93d26 JQ |
1080 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
1081 | ||
1082 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 1083 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 1084 | migration_trigger_throttle(rs); |
070afca2 | 1085 | |
b734035b XG |
1086 | migration_update_rates(rs, end_time); |
1087 | ||
be8b02ed | 1088 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
1089 | |
1090 | /* reset period counters */ | |
f664da80 | 1091 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 1092 | rs->num_dirty_pages_period = 0; |
dc14a470 | 1093 | rs->bytes_xfer_prev = ram_counters.transferred; |
56e93d26 | 1094 | } |
4addcd4f | 1095 | if (migrate_use_events()) { |
3ab72385 | 1096 | qapi_event_send_migration_pass(ram_counters.dirty_sync_count); |
4addcd4f | 1097 | } |
56e93d26 JQ |
1098 | } |
1099 | ||
bd227060 WW |
1100 | static void migration_bitmap_sync_precopy(RAMState *rs) |
1101 | { | |
1102 | Error *local_err = NULL; | |
1103 | ||
1104 | /* | |
1105 | * The current notifier usage is just an optimization to migration, so we | |
1106 | * don't stop the normal migration process in the error case. | |
1107 | */ | |
1108 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
1109 | error_report_err(local_err); | |
b4a1733c | 1110 | local_err = NULL; |
bd227060 WW |
1111 | } |
1112 | ||
1113 | migration_bitmap_sync(rs); | |
1114 | ||
1115 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
1116 | error_report_err(local_err); | |
1117 | } | |
1118 | } | |
1119 | ||
6c97ec5f XG |
1120 | /** |
1121 | * save_zero_page_to_file: send the zero page to the file | |
1122 | * | |
1123 | * Returns the size of data written to the file, 0 means the page is not | |
1124 | * a zero page | |
1125 | * | |
1126 | * @rs: current RAM state | |
1127 | * @file: the file where the data is saved | |
1128 | * @block: block that contains the page we want to send | |
1129 | * @offset: offset inside the block for the page | |
1130 | */ | |
1131 | static int save_zero_page_to_file(RAMState *rs, QEMUFile *file, | |
1132 | RAMBlock *block, ram_addr_t offset) | |
1133 | { | |
1134 | uint8_t *p = block->host + offset; | |
1135 | int len = 0; | |
1136 | ||
1137 | if (is_zero_range(p, TARGET_PAGE_SIZE)) { | |
1138 | len += save_page_header(rs, file, block, offset | RAM_SAVE_FLAG_ZERO); | |
1139 | qemu_put_byte(file, 0); | |
1140 | len += 1; | |
1141 | } | |
1142 | return len; | |
1143 | } | |
1144 | ||
56e93d26 | 1145 | /** |
3d0684b2 | 1146 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1147 | * |
3d0684b2 | 1148 | * Returns the number of pages written. |
56e93d26 | 1149 | * |
f7ccd61b | 1150 | * @rs: current RAM state |
56e93d26 JQ |
1151 | * @block: block that contains the page we want to send |
1152 | * @offset: offset inside the block for the page | |
56e93d26 | 1153 | */ |
7faccdc3 | 1154 | static int save_zero_page(RAMState *rs, RAMBlock *block, ram_addr_t offset) |
56e93d26 | 1155 | { |
6c97ec5f | 1156 | int len = save_zero_page_to_file(rs, rs->f, block, offset); |
56e93d26 | 1157 | |
6c97ec5f | 1158 | if (len) { |
9360447d | 1159 | ram_counters.duplicate++; |
6c97ec5f XG |
1160 | ram_counters.transferred += len; |
1161 | return 1; | |
56e93d26 | 1162 | } |
6c97ec5f | 1163 | return -1; |
56e93d26 JQ |
1164 | } |
1165 | ||
5727309d | 1166 | static void ram_release_pages(const char *rbname, uint64_t offset, int pages) |
53f09a10 | 1167 | { |
5727309d | 1168 | if (!migrate_release_ram() || !migration_in_postcopy()) { |
53f09a10 PB |
1169 | return; |
1170 | } | |
1171 | ||
8bba004c | 1172 | ram_discard_range(rbname, offset, ((ram_addr_t)pages) << TARGET_PAGE_BITS); |
53f09a10 PB |
1173 | } |
1174 | ||
059ff0fb XG |
1175 | /* |
1176 | * @pages: the number of pages written by the control path, | |
1177 | * < 0 - error | |
1178 | * > 0 - number of pages written | |
1179 | * | |
1180 | * Return true if the pages has been saved, otherwise false is returned. | |
1181 | */ | |
1182 | static bool control_save_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, | |
1183 | int *pages) | |
1184 | { | |
1185 | uint64_t bytes_xmit = 0; | |
1186 | int ret; | |
1187 | ||
1188 | *pages = -1; | |
1189 | ret = ram_control_save_page(rs->f, block->offset, offset, TARGET_PAGE_SIZE, | |
1190 | &bytes_xmit); | |
1191 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { | |
1192 | return false; | |
1193 | } | |
1194 | ||
1195 | if (bytes_xmit) { | |
1196 | ram_counters.transferred += bytes_xmit; | |
1197 | *pages = 1; | |
1198 | } | |
1199 | ||
1200 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1201 | return true; | |
1202 | } | |
1203 | ||
1204 | if (bytes_xmit > 0) { | |
1205 | ram_counters.normal++; | |
1206 | } else if (bytes_xmit == 0) { | |
1207 | ram_counters.duplicate++; | |
1208 | } | |
1209 | ||
1210 | return true; | |
1211 | } | |
1212 | ||
65dacaa0 XG |
1213 | /* |
1214 | * directly send the page to the stream | |
1215 | * | |
1216 | * Returns the number of pages written. | |
1217 | * | |
1218 | * @rs: current RAM state | |
1219 | * @block: block that contains the page we want to send | |
1220 | * @offset: offset inside the block for the page | |
1221 | * @buf: the page to be sent | |
1222 | * @async: send to page asyncly | |
1223 | */ | |
1224 | static int save_normal_page(RAMState *rs, RAMBlock *block, ram_addr_t offset, | |
1225 | uint8_t *buf, bool async) | |
1226 | { | |
1227 | ram_counters.transferred += save_page_header(rs, rs->f, block, | |
1228 | offset | RAM_SAVE_FLAG_PAGE); | |
1229 | if (async) { | |
1230 | qemu_put_buffer_async(rs->f, buf, TARGET_PAGE_SIZE, | |
1231 | migrate_release_ram() & | |
1232 | migration_in_postcopy()); | |
1233 | } else { | |
1234 | qemu_put_buffer(rs->f, buf, TARGET_PAGE_SIZE); | |
1235 | } | |
1236 | ram_counters.transferred += TARGET_PAGE_SIZE; | |
1237 | ram_counters.normal++; | |
1238 | return 1; | |
1239 | } | |
1240 | ||
56e93d26 | 1241 | /** |
3d0684b2 | 1242 | * ram_save_page: send the given page to the stream |
56e93d26 | 1243 | * |
3d0684b2 | 1244 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1245 | * < 0 - error |
1246 | * >=0 - Number of pages written - this might legally be 0 | |
1247 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1248 | * |
6f37bb8b | 1249 | * @rs: current RAM state |
56e93d26 JQ |
1250 | * @block: block that contains the page we want to send |
1251 | * @offset: offset inside the block for the page | |
1252 | * @last_stage: if we are at the completion stage | |
56e93d26 | 1253 | */ |
a0a8aa14 | 1254 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss, bool last_stage) |
56e93d26 JQ |
1255 | { |
1256 | int pages = -1; | |
56e93d26 | 1257 | uint8_t *p; |
56e93d26 | 1258 | bool send_async = true; |
a08f6890 | 1259 | RAMBlock *block = pss->block; |
8bba004c | 1260 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1261 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1262 | |
2f68e399 | 1263 | p = block->host + offset; |
1db9d8e5 | 1264 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1265 | |
56e93d26 | 1266 | XBZRLE_cache_lock(); |
1a373522 | 1267 | if (rs->xbzrle_enabled && !migration_in_postcopy()) { |
059ff0fb XG |
1268 | pages = save_xbzrle_page(rs, &p, current_addr, block, |
1269 | offset, last_stage); | |
1270 | if (!last_stage) { | |
1271 | /* Can't send this cached data async, since the cache page | |
1272 | * might get updated before it gets to the wire | |
56e93d26 | 1273 | */ |
059ff0fb | 1274 | send_async = false; |
56e93d26 JQ |
1275 | } |
1276 | } | |
1277 | ||
1278 | /* XBZRLE overflow or normal page */ | |
1279 | if (pages == -1) { | |
65dacaa0 | 1280 | pages = save_normal_page(rs, block, offset, p, send_async); |
56e93d26 JQ |
1281 | } |
1282 | ||
1283 | XBZRLE_cache_unlock(); | |
1284 | ||
1285 | return pages; | |
1286 | } | |
1287 | ||
b9ee2f7d JQ |
1288 | static int ram_save_multifd_page(RAMState *rs, RAMBlock *block, |
1289 | ram_addr_t offset) | |
1290 | { | |
67a4c891 | 1291 | if (multifd_queue_page(rs->f, block, offset) < 0) { |
713f762a IR |
1292 | return -1; |
1293 | } | |
b9ee2f7d JQ |
1294 | ram_counters.normal++; |
1295 | ||
1296 | return 1; | |
1297 | } | |
1298 | ||
5e5fdcff | 1299 | static bool do_compress_ram_page(QEMUFile *f, z_stream *stream, RAMBlock *block, |
6ef3771c | 1300 | ram_addr_t offset, uint8_t *source_buf) |
56e93d26 | 1301 | { |
53518d94 | 1302 | RAMState *rs = ram_state; |
a7a9a88f | 1303 | uint8_t *p = block->host + (offset & TARGET_PAGE_MASK); |
5e5fdcff | 1304 | bool zero_page = false; |
6ef3771c | 1305 | int ret; |
56e93d26 | 1306 | |
5e5fdcff XG |
1307 | if (save_zero_page_to_file(rs, f, block, offset)) { |
1308 | zero_page = true; | |
1309 | goto exit; | |
1310 | } | |
1311 | ||
6ef3771c | 1312 | save_page_header(rs, f, block, offset | RAM_SAVE_FLAG_COMPRESS_PAGE); |
34ab9e97 XG |
1313 | |
1314 | /* | |
1315 | * copy it to a internal buffer to avoid it being modified by VM | |
1316 | * so that we can catch up the error during compression and | |
1317 | * decompression | |
1318 | */ | |
1319 | memcpy(source_buf, p, TARGET_PAGE_SIZE); | |
6ef3771c XG |
1320 | ret = qemu_put_compression_data(f, stream, source_buf, TARGET_PAGE_SIZE); |
1321 | if (ret < 0) { | |
1322 | qemu_file_set_error(migrate_get_current()->to_dst_file, ret); | |
b3be2896 | 1323 | error_report("compressed data failed!"); |
5e5fdcff | 1324 | return false; |
b3be2896 | 1325 | } |
56e93d26 | 1326 | |
5e5fdcff | 1327 | exit: |
6ef3771c | 1328 | ram_release_pages(block->idstr, offset & TARGET_PAGE_MASK, 1); |
5e5fdcff XG |
1329 | return zero_page; |
1330 | } | |
1331 | ||
1332 | static void | |
1333 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1334 | { | |
76e03000 XG |
1335 | ram_counters.transferred += bytes_xmit; |
1336 | ||
5e5fdcff XG |
1337 | if (param->zero_page) { |
1338 | ram_counters.duplicate++; | |
76e03000 | 1339 | return; |
5e5fdcff | 1340 | } |
76e03000 XG |
1341 | |
1342 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1343 | compression_counters.compressed_size += bytes_xmit - 8; | |
1344 | compression_counters.pages++; | |
56e93d26 JQ |
1345 | } |
1346 | ||
32b05495 XG |
1347 | static bool save_page_use_compression(RAMState *rs); |
1348 | ||
ce25d337 | 1349 | static void flush_compressed_data(RAMState *rs) |
56e93d26 JQ |
1350 | { |
1351 | int idx, len, thread_count; | |
1352 | ||
32b05495 | 1353 | if (!save_page_use_compression(rs)) { |
56e93d26 JQ |
1354 | return; |
1355 | } | |
1356 | thread_count = migrate_compress_threads(); | |
a7a9a88f | 1357 | |
0d9f9a5c | 1358 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1359 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1360 | while (!comp_param[idx].done) { |
0d9f9a5c | 1361 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1362 | } |
a7a9a88f | 1363 | } |
0d9f9a5c | 1364 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1365 | |
1366 | for (idx = 0; idx < thread_count; idx++) { | |
1367 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1368 | if (!comp_param[idx].quit) { |
ce25d337 | 1369 | len = qemu_put_qemu_file(rs->f, comp_param[idx].file); |
5e5fdcff XG |
1370 | /* |
1371 | * it's safe to fetch zero_page without holding comp_done_lock | |
1372 | * as there is no further request submitted to the thread, | |
1373 | * i.e, the thread should be waiting for a request at this point. | |
1374 | */ | |
1375 | update_compress_thread_counts(&comp_param[idx], len); | |
56e93d26 | 1376 | } |
a7a9a88f | 1377 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1378 | } |
1379 | } | |
1380 | ||
1381 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, | |
1382 | ram_addr_t offset) | |
1383 | { | |
1384 | param->block = block; | |
1385 | param->offset = offset; | |
1386 | } | |
1387 | ||
ce25d337 JQ |
1388 | static int compress_page_with_multi_thread(RAMState *rs, RAMBlock *block, |
1389 | ram_addr_t offset) | |
56e93d26 JQ |
1390 | { |
1391 | int idx, thread_count, bytes_xmit = -1, pages = -1; | |
1d58872a | 1392 | bool wait = migrate_compress_wait_thread(); |
56e93d26 JQ |
1393 | |
1394 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1395 | qemu_mutex_lock(&comp_done_lock); |
1d58872a XG |
1396 | retry: |
1397 | for (idx = 0; idx < thread_count; idx++) { | |
1398 | if (comp_param[idx].done) { | |
1399 | comp_param[idx].done = false; | |
1400 | bytes_xmit = qemu_put_qemu_file(rs->f, comp_param[idx].file); | |
1401 | qemu_mutex_lock(&comp_param[idx].mutex); | |
1402 | set_compress_params(&comp_param[idx], block, offset); | |
1403 | qemu_cond_signal(&comp_param[idx].cond); | |
1404 | qemu_mutex_unlock(&comp_param[idx].mutex); | |
1405 | pages = 1; | |
5e5fdcff | 1406 | update_compress_thread_counts(&comp_param[idx], bytes_xmit); |
56e93d26 | 1407 | break; |
56e93d26 JQ |
1408 | } |
1409 | } | |
1d58872a XG |
1410 | |
1411 | /* | |
1412 | * wait for the free thread if the user specifies 'compress-wait-thread', | |
1413 | * otherwise we will post the page out in the main thread as normal page. | |
1414 | */ | |
1415 | if (pages < 0 && wait) { | |
1416 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); | |
1417 | goto retry; | |
1418 | } | |
0d9f9a5c | 1419 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1420 | |
1421 | return pages; | |
1422 | } | |
1423 | ||
3d0684b2 JQ |
1424 | /** |
1425 | * find_dirty_block: find the next dirty page and update any state | |
1426 | * associated with the search process. | |
b9e60928 | 1427 | * |
a5f7b1a6 | 1428 | * Returns true if a page is found |
b9e60928 | 1429 | * |
6f37bb8b | 1430 | * @rs: current RAM state |
3d0684b2 JQ |
1431 | * @pss: data about the state of the current dirty page scan |
1432 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1433 | */ |
f20e2865 | 1434 | static bool find_dirty_block(RAMState *rs, PageSearchStatus *pss, bool *again) |
b9e60928 | 1435 | { |
f20e2865 | 1436 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
6f37bb8b | 1437 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1438 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1439 | /* |
1440 | * We've been once around the RAM and haven't found anything. | |
1441 | * Give up. | |
1442 | */ | |
1443 | *again = false; | |
1444 | return false; | |
1445 | } | |
542147f4 DH |
1446 | if (!offset_in_ramblock(pss->block, |
1447 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)) { | |
b9e60928 | 1448 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1449 | pss->page = 0; |
b9e60928 DDAG |
1450 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1451 | if (!pss->block) { | |
48df9d80 XG |
1452 | /* |
1453 | * If memory migration starts over, we will meet a dirtied page | |
1454 | * which may still exists in compression threads's ring, so we | |
1455 | * should flush the compressed data to make sure the new page | |
1456 | * is not overwritten by the old one in the destination. | |
1457 | * | |
1458 | * Also If xbzrle is on, stop using the data compression at this | |
1459 | * point. In theory, xbzrle can do better than compression. | |
1460 | */ | |
1461 | flush_compressed_data(rs); | |
1462 | ||
b9e60928 DDAG |
1463 | /* Hit the end of the list */ |
1464 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1465 | /* Flag that we've looped */ | |
1466 | pss->complete_round = true; | |
1a373522 DH |
1467 | /* After the first round, enable XBZRLE. */ |
1468 | if (migrate_use_xbzrle()) { | |
1469 | rs->xbzrle_enabled = true; | |
1470 | } | |
b9e60928 DDAG |
1471 | } |
1472 | /* Didn't find anything this time, but try again on the new block */ | |
1473 | *again = true; | |
1474 | return false; | |
1475 | } else { | |
1476 | /* Can go around again, but... */ | |
1477 | *again = true; | |
1478 | /* We've found something so probably don't need to */ | |
1479 | return true; | |
1480 | } | |
1481 | } | |
1482 | ||
3d0684b2 JQ |
1483 | /** |
1484 | * unqueue_page: gets a page of the queue | |
1485 | * | |
a82d593b | 1486 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1487 | * |
3d0684b2 JQ |
1488 | * Returns the block of the page (or NULL if none available) |
1489 | * | |
ec481c6c | 1490 | * @rs: current RAM state |
3d0684b2 | 1491 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1492 | */ |
f20e2865 | 1493 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b DDAG |
1494 | { |
1495 | RAMBlock *block = NULL; | |
1496 | ||
ae526e32 XG |
1497 | if (QSIMPLEQ_EMPTY_ATOMIC(&rs->src_page_requests)) { |
1498 | return NULL; | |
1499 | } | |
1500 | ||
6e8a355d | 1501 | QEMU_LOCK_GUARD(&rs->src_page_req_mutex); |
ec481c6c JQ |
1502 | if (!QSIMPLEQ_EMPTY(&rs->src_page_requests)) { |
1503 | struct RAMSrcPageRequest *entry = | |
1504 | QSIMPLEQ_FIRST(&rs->src_page_requests); | |
a82d593b DDAG |
1505 | block = entry->rb; |
1506 | *offset = entry->offset; | |
a82d593b DDAG |
1507 | |
1508 | if (entry->len > TARGET_PAGE_SIZE) { | |
1509 | entry->len -= TARGET_PAGE_SIZE; | |
1510 | entry->offset += TARGET_PAGE_SIZE; | |
1511 | } else { | |
1512 | memory_region_unref(block->mr); | |
ec481c6c | 1513 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
a82d593b | 1514 | g_free(entry); |
e03a34f8 | 1515 | migration_consume_urgent_request(); |
a82d593b DDAG |
1516 | } |
1517 | } | |
a82d593b DDAG |
1518 | |
1519 | return block; | |
1520 | } | |
1521 | ||
278e2f55 AG |
1522 | #if defined(__linux__) |
1523 | /** | |
1524 | * poll_fault_page: try to get next UFFD write fault page and, if pending fault | |
1525 | * is found, return RAM block pointer and page offset | |
1526 | * | |
1527 | * Returns pointer to the RAMBlock containing faulting page, | |
1528 | * NULL if no write faults are pending | |
1529 | * | |
1530 | * @rs: current RAM state | |
1531 | * @offset: page offset from the beginning of the block | |
1532 | */ | |
1533 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1534 | { | |
1535 | struct uffd_msg uffd_msg; | |
1536 | void *page_address; | |
82ea3e3b | 1537 | RAMBlock *block; |
278e2f55 AG |
1538 | int res; |
1539 | ||
1540 | if (!migrate_background_snapshot()) { | |
1541 | return NULL; | |
1542 | } | |
1543 | ||
1544 | res = uffd_read_events(rs->uffdio_fd, &uffd_msg, 1); | |
1545 | if (res <= 0) { | |
1546 | return NULL; | |
1547 | } | |
1548 | ||
1549 | page_address = (void *)(uintptr_t) uffd_msg.arg.pagefault.address; | |
82ea3e3b AG |
1550 | block = qemu_ram_block_from_host(page_address, false, offset); |
1551 | assert(block && (block->flags & RAM_UF_WRITEPROTECT) != 0); | |
1552 | return block; | |
278e2f55 AG |
1553 | } |
1554 | ||
1555 | /** | |
1556 | * ram_save_release_protection: release UFFD write protection after | |
1557 | * a range of pages has been saved | |
1558 | * | |
1559 | * @rs: current RAM state | |
1560 | * @pss: page-search-status structure | |
1561 | * @start_page: index of the first page in the range relative to pss->block | |
1562 | * | |
1563 | * Returns 0 on success, negative value in case of an error | |
1564 | */ | |
1565 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1566 | unsigned long start_page) | |
1567 | { | |
1568 | int res = 0; | |
1569 | ||
1570 | /* Check if page is from UFFD-managed region. */ | |
1571 | if (pss->block->flags & RAM_UF_WRITEPROTECT) { | |
1572 | void *page_address = pss->block->host + (start_page << TARGET_PAGE_BITS); | |
1573 | uint64_t run_length = (pss->page - start_page + 1) << TARGET_PAGE_BITS; | |
1574 | ||
1575 | /* Flush async buffers before un-protect. */ | |
1576 | qemu_fflush(rs->f); | |
1577 | /* Un-protect memory range. */ | |
1578 | res = uffd_change_protection(rs->uffdio_fd, page_address, run_length, | |
1579 | false, false); | |
1580 | } | |
1581 | ||
1582 | return res; | |
1583 | } | |
1584 | ||
1585 | /* ram_write_tracking_available: check if kernel supports required UFFD features | |
1586 | * | |
1587 | * Returns true if supports, false otherwise | |
1588 | */ | |
1589 | bool ram_write_tracking_available(void) | |
1590 | { | |
1591 | uint64_t uffd_features; | |
1592 | int res; | |
1593 | ||
1594 | res = uffd_query_features(&uffd_features); | |
1595 | return (res == 0 && | |
1596 | (uffd_features & UFFD_FEATURE_PAGEFAULT_FLAG_WP) != 0); | |
1597 | } | |
1598 | ||
1599 | /* ram_write_tracking_compatible: check if guest configuration is | |
1600 | * compatible with 'write-tracking' | |
1601 | * | |
1602 | * Returns true if compatible, false otherwise | |
1603 | */ | |
1604 | bool ram_write_tracking_compatible(void) | |
1605 | { | |
1606 | const uint64_t uffd_ioctls_mask = BIT(_UFFDIO_WRITEPROTECT); | |
1607 | int uffd_fd; | |
82ea3e3b | 1608 | RAMBlock *block; |
278e2f55 AG |
1609 | bool ret = false; |
1610 | ||
1611 | /* Open UFFD file descriptor */ | |
1612 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, false); | |
1613 | if (uffd_fd < 0) { | |
1614 | return false; | |
1615 | } | |
1616 | ||
1617 | RCU_READ_LOCK_GUARD(); | |
1618 | ||
82ea3e3b | 1619 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 AG |
1620 | uint64_t uffd_ioctls; |
1621 | ||
1622 | /* Nothing to do with read-only and MMIO-writable regions */ | |
82ea3e3b | 1623 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1624 | continue; |
1625 | } | |
1626 | /* Try to register block memory via UFFD-IO to track writes */ | |
82ea3e3b | 1627 | if (uffd_register_memory(uffd_fd, block->host, block->max_length, |
278e2f55 AG |
1628 | UFFDIO_REGISTER_MODE_WP, &uffd_ioctls)) { |
1629 | goto out; | |
1630 | } | |
1631 | if ((uffd_ioctls & uffd_ioctls_mask) != uffd_ioctls_mask) { | |
1632 | goto out; | |
1633 | } | |
1634 | } | |
1635 | ret = true; | |
1636 | ||
1637 | out: | |
1638 | uffd_close_fd(uffd_fd); | |
1639 | return ret; | |
1640 | } | |
1641 | ||
f7b9dcfb DH |
1642 | static inline void populate_read_range(RAMBlock *block, ram_addr_t offset, |
1643 | ram_addr_t size) | |
1644 | { | |
1645 | /* | |
1646 | * We read one byte of each page; this will preallocate page tables if | |
1647 | * required and populate the shared zeropage on MAP_PRIVATE anonymous memory | |
1648 | * where no page was populated yet. This might require adaption when | |
1649 | * supporting other mappings, like shmem. | |
1650 | */ | |
1651 | for (; offset < size; offset += block->page_size) { | |
1652 | char tmp = *((char *)block->host + offset); | |
1653 | ||
1654 | /* Don't optimize the read out */ | |
1655 | asm volatile("" : "+r" (tmp)); | |
1656 | } | |
1657 | } | |
1658 | ||
6fee3a1f DH |
1659 | static inline int populate_read_section(MemoryRegionSection *section, |
1660 | void *opaque) | |
1661 | { | |
1662 | const hwaddr size = int128_get64(section->size); | |
1663 | hwaddr offset = section->offset_within_region; | |
1664 | RAMBlock *block = section->mr->ram_block; | |
1665 | ||
1666 | populate_read_range(block, offset, size); | |
1667 | return 0; | |
1668 | } | |
1669 | ||
eeccb99c | 1670 | /* |
f7b9dcfb DH |
1671 | * ram_block_populate_read: preallocate page tables and populate pages in the |
1672 | * RAM block by reading a byte of each page. | |
eeccb99c AG |
1673 | * |
1674 | * Since it's solely used for userfault_fd WP feature, here we just | |
1675 | * hardcode page size to qemu_real_host_page_size. | |
1676 | * | |
82ea3e3b | 1677 | * @block: RAM block to populate |
eeccb99c | 1678 | */ |
6fee3a1f | 1679 | static void ram_block_populate_read(RAMBlock *rb) |
eeccb99c | 1680 | { |
6fee3a1f DH |
1681 | /* |
1682 | * Skip populating all pages that fall into a discarded range as managed by | |
1683 | * a RamDiscardManager responsible for the mapped memory region of the | |
1684 | * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock | |
1685 | * must not get populated automatically. We don't have to track | |
1686 | * modifications via userfaultfd WP reliably, because these pages will | |
1687 | * not be part of the migration stream either way -- see | |
1688 | * ramblock_dirty_bitmap_exclude_discarded_pages(). | |
1689 | * | |
1690 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1691 | */ | |
1692 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1693 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1694 | MemoryRegionSection section = { | |
1695 | .mr = rb->mr, | |
1696 | .offset_within_region = 0, | |
1697 | .size = rb->mr->size, | |
1698 | }; | |
1699 | ||
1700 | ram_discard_manager_replay_populated(rdm, §ion, | |
1701 | populate_read_section, NULL); | |
1702 | } else { | |
1703 | populate_read_range(rb, 0, rb->used_length); | |
1704 | } | |
eeccb99c AG |
1705 | } |
1706 | ||
1707 | /* | |
1708 | * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking | |
1709 | */ | |
1710 | void ram_write_tracking_prepare(void) | |
1711 | { | |
82ea3e3b | 1712 | RAMBlock *block; |
eeccb99c AG |
1713 | |
1714 | RCU_READ_LOCK_GUARD(); | |
1715 | ||
82ea3e3b | 1716 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
eeccb99c | 1717 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1718 | if (block->mr->readonly || block->mr->rom_device) { |
eeccb99c AG |
1719 | continue; |
1720 | } | |
1721 | ||
1722 | /* | |
1723 | * Populate pages of the RAM block before enabling userfault_fd | |
1724 | * write protection. | |
1725 | * | |
1726 | * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with | |
1727 | * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip | |
1728 | * pages with pte_none() entries in page table. | |
1729 | */ | |
f7b9dcfb | 1730 | ram_block_populate_read(block); |
eeccb99c AG |
1731 | } |
1732 | } | |
1733 | ||
278e2f55 AG |
1734 | /* |
1735 | * ram_write_tracking_start: start UFFD-WP memory tracking | |
1736 | * | |
1737 | * Returns 0 for success or negative value in case of error | |
1738 | */ | |
1739 | int ram_write_tracking_start(void) | |
1740 | { | |
1741 | int uffd_fd; | |
1742 | RAMState *rs = ram_state; | |
82ea3e3b | 1743 | RAMBlock *block; |
278e2f55 AG |
1744 | |
1745 | /* Open UFFD file descriptor */ | |
1746 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, true); | |
1747 | if (uffd_fd < 0) { | |
1748 | return uffd_fd; | |
1749 | } | |
1750 | rs->uffdio_fd = uffd_fd; | |
1751 | ||
1752 | RCU_READ_LOCK_GUARD(); | |
1753 | ||
82ea3e3b | 1754 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 | 1755 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1756 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1757 | continue; |
1758 | } | |
1759 | ||
1760 | /* Register block memory with UFFD to track writes */ | |
82ea3e3b AG |
1761 | if (uffd_register_memory(rs->uffdio_fd, block->host, |
1762 | block->max_length, UFFDIO_REGISTER_MODE_WP, NULL)) { | |
278e2f55 AG |
1763 | goto fail; |
1764 | } | |
1765 | /* Apply UFFD write protection to the block memory range */ | |
82ea3e3b AG |
1766 | if (uffd_change_protection(rs->uffdio_fd, block->host, |
1767 | block->max_length, true, false)) { | |
278e2f55 AG |
1768 | goto fail; |
1769 | } | |
82ea3e3b AG |
1770 | block->flags |= RAM_UF_WRITEPROTECT; |
1771 | memory_region_ref(block->mr); | |
278e2f55 | 1772 | |
82ea3e3b AG |
1773 | trace_ram_write_tracking_ramblock_start(block->idstr, block->page_size, |
1774 | block->host, block->max_length); | |
278e2f55 AG |
1775 | } |
1776 | ||
1777 | return 0; | |
1778 | ||
1779 | fail: | |
1780 | error_report("ram_write_tracking_start() failed: restoring initial memory state"); | |
1781 | ||
82ea3e3b AG |
1782 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1783 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1784 | continue; |
1785 | } | |
1786 | /* | |
1787 | * In case some memory block failed to be write-protected | |
1788 | * remove protection and unregister all succeeded RAM blocks | |
1789 | */ | |
82ea3e3b AG |
1790 | uffd_change_protection(rs->uffdio_fd, block->host, block->max_length, |
1791 | false, false); | |
1792 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); | |
278e2f55 | 1793 | /* Cleanup flags and remove reference */ |
82ea3e3b AG |
1794 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1795 | memory_region_unref(block->mr); | |
278e2f55 AG |
1796 | } |
1797 | ||
1798 | uffd_close_fd(uffd_fd); | |
1799 | rs->uffdio_fd = -1; | |
1800 | return -1; | |
1801 | } | |
1802 | ||
1803 | /** | |
1804 | * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection | |
1805 | */ | |
1806 | void ram_write_tracking_stop(void) | |
1807 | { | |
1808 | RAMState *rs = ram_state; | |
82ea3e3b | 1809 | RAMBlock *block; |
278e2f55 AG |
1810 | |
1811 | RCU_READ_LOCK_GUARD(); | |
1812 | ||
82ea3e3b AG |
1813 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1814 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1815 | continue; |
1816 | } | |
1817 | /* Remove protection and unregister all affected RAM blocks */ | |
82ea3e3b AG |
1818 | uffd_change_protection(rs->uffdio_fd, block->host, block->max_length, |
1819 | false, false); | |
1820 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); | |
278e2f55 | 1821 | |
82ea3e3b AG |
1822 | trace_ram_write_tracking_ramblock_stop(block->idstr, block->page_size, |
1823 | block->host, block->max_length); | |
278e2f55 AG |
1824 | |
1825 | /* Cleanup flags and remove reference */ | |
82ea3e3b AG |
1826 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1827 | memory_region_unref(block->mr); | |
278e2f55 AG |
1828 | } |
1829 | ||
1830 | /* Finally close UFFD file descriptor */ | |
1831 | uffd_close_fd(rs->uffdio_fd); | |
1832 | rs->uffdio_fd = -1; | |
1833 | } | |
1834 | ||
1835 | #else | |
1836 | /* No target OS support, stubs just fail or ignore */ | |
1837 | ||
1838 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1839 | { | |
1840 | (void) rs; | |
1841 | (void) offset; | |
1842 | ||
1843 | return NULL; | |
1844 | } | |
1845 | ||
1846 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1847 | unsigned long start_page) | |
1848 | { | |
1849 | (void) rs; | |
1850 | (void) pss; | |
1851 | (void) start_page; | |
1852 | ||
1853 | return 0; | |
1854 | } | |
1855 | ||
1856 | bool ram_write_tracking_available(void) | |
1857 | { | |
1858 | return false; | |
1859 | } | |
1860 | ||
1861 | bool ram_write_tracking_compatible(void) | |
1862 | { | |
1863 | assert(0); | |
1864 | return false; | |
1865 | } | |
1866 | ||
1867 | int ram_write_tracking_start(void) | |
1868 | { | |
1869 | assert(0); | |
1870 | return -1; | |
1871 | } | |
1872 | ||
1873 | void ram_write_tracking_stop(void) | |
1874 | { | |
1875 | assert(0); | |
1876 | } | |
1877 | #endif /* defined(__linux__) */ | |
1878 | ||
3d0684b2 | 1879 | /** |
ff1543af | 1880 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
1881 | * |
1882 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1883 | * |
a5f7b1a6 | 1884 | * Returns true if a queued page is found |
a82d593b | 1885 | * |
6f37bb8b | 1886 | * @rs: current RAM state |
3d0684b2 | 1887 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1888 | */ |
f20e2865 | 1889 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1890 | { |
1891 | RAMBlock *block; | |
1892 | ram_addr_t offset; | |
1893 | bool dirty; | |
1894 | ||
1895 | do { | |
f20e2865 | 1896 | block = unqueue_page(rs, &offset); |
a82d593b DDAG |
1897 | /* |
1898 | * We're sending this page, and since it's postcopy nothing else | |
1899 | * will dirty it, and we must make sure it doesn't get sent again | |
1900 | * even if this queue request was received after the background | |
1901 | * search already sent it. | |
1902 | */ | |
1903 | if (block) { | |
f20e2865 JQ |
1904 | unsigned long page; |
1905 | ||
6b6712ef JQ |
1906 | page = offset >> TARGET_PAGE_BITS; |
1907 | dirty = test_bit(page, block->bmap); | |
a82d593b | 1908 | if (!dirty) { |
06b10688 | 1909 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, |
64737606 | 1910 | page); |
a82d593b | 1911 | } else { |
f20e2865 | 1912 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); |
a82d593b DDAG |
1913 | } |
1914 | } | |
1915 | ||
1916 | } while (block && !dirty); | |
1917 | ||
278e2f55 AG |
1918 | if (!block) { |
1919 | /* | |
1920 | * Poll write faults too if background snapshot is enabled; that's | |
1921 | * when we have vcpus got blocked by the write protected pages. | |
1922 | */ | |
1923 | block = poll_fault_page(rs, &offset); | |
1924 | } | |
1925 | ||
a82d593b | 1926 | if (block) { |
a82d593b DDAG |
1927 | /* |
1928 | * We want the background search to continue from the queued page | |
1929 | * since the guest is likely to want other pages near to the page | |
1930 | * it just requested. | |
1931 | */ | |
1932 | pss->block = block; | |
a935e30f | 1933 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
1934 | |
1935 | /* | |
1936 | * This unqueued page would break the "one round" check, even is | |
1937 | * really rare. | |
1938 | */ | |
1939 | pss->complete_round = false; | |
a82d593b DDAG |
1940 | } |
1941 | ||
1942 | return !!block; | |
1943 | } | |
1944 | ||
6c595cde | 1945 | /** |
5e58f968 JQ |
1946 | * migration_page_queue_free: drop any remaining pages in the ram |
1947 | * request queue | |
6c595cde | 1948 | * |
3d0684b2 JQ |
1949 | * It should be empty at the end anyway, but in error cases there may |
1950 | * be some left. in case that there is any page left, we drop it. | |
1951 | * | |
6c595cde | 1952 | */ |
83c13382 | 1953 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 1954 | { |
ec481c6c | 1955 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
1956 | /* This queue generally should be empty - but in the case of a failed |
1957 | * migration might have some droppings in. | |
1958 | */ | |
89ac5a1d | 1959 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 1960 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1961 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1962 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1963 | g_free(mspr); |
1964 | } | |
6c595cde DDAG |
1965 | } |
1966 | ||
1967 | /** | |
3d0684b2 JQ |
1968 | * ram_save_queue_pages: queue the page for transmission |
1969 | * | |
1970 | * A request from postcopy destination for example. | |
1971 | * | |
1972 | * Returns zero on success or negative on error | |
1973 | * | |
3d0684b2 JQ |
1974 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1975 | * same that last one. | |
1976 | * @start: starting address from the start of the RAMBlock | |
1977 | * @len: length (in bytes) to send | |
6c595cde | 1978 | */ |
96506894 | 1979 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1980 | { |
1981 | RAMBlock *ramblock; | |
53518d94 | 1982 | RAMState *rs = ram_state; |
6c595cde | 1983 | |
9360447d | 1984 | ram_counters.postcopy_requests++; |
89ac5a1d DDAG |
1985 | RCU_READ_LOCK_GUARD(); |
1986 | ||
6c595cde DDAG |
1987 | if (!rbname) { |
1988 | /* Reuse last RAMBlock */ | |
68a098f3 | 1989 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1990 | |
1991 | if (!ramblock) { | |
1992 | /* | |
1993 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1994 | * it's the 1st request. | |
1995 | */ | |
1996 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 1997 | return -1; |
6c595cde DDAG |
1998 | } |
1999 | } else { | |
2000 | ramblock = qemu_ram_block_by_name(rbname); | |
2001 | ||
2002 | if (!ramblock) { | |
2003 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
2004 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 2005 | return -1; |
6c595cde | 2006 | } |
68a098f3 | 2007 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
2008 | } |
2009 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
542147f4 | 2010 | if (!offset_in_ramblock(ramblock, start + len - 1)) { |
9458ad6b JQ |
2011 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
2012 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 2013 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 2014 | return -1; |
6c595cde DDAG |
2015 | } |
2016 | ||
ec481c6c JQ |
2017 | struct RAMSrcPageRequest *new_entry = |
2018 | g_malloc0(sizeof(struct RAMSrcPageRequest)); | |
6c595cde DDAG |
2019 | new_entry->rb = ramblock; |
2020 | new_entry->offset = start; | |
2021 | new_entry->len = len; | |
2022 | ||
2023 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
2024 | qemu_mutex_lock(&rs->src_page_req_mutex); |
2025 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 2026 | migration_make_urgent_request(); |
ec481c6c | 2027 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
2028 | |
2029 | return 0; | |
6c595cde DDAG |
2030 | } |
2031 | ||
d7400a34 XG |
2032 | static bool save_page_use_compression(RAMState *rs) |
2033 | { | |
2034 | if (!migrate_use_compression()) { | |
2035 | return false; | |
2036 | } | |
2037 | ||
2038 | /* | |
1a373522 DH |
2039 | * If xbzrle is enabled (e.g., after first round of migration), stop |
2040 | * using the data compression. In theory, xbzrle can do better than | |
2041 | * compression. | |
d7400a34 | 2042 | */ |
1a373522 DH |
2043 | if (rs->xbzrle_enabled) { |
2044 | return false; | |
d7400a34 XG |
2045 | } |
2046 | ||
1a373522 | 2047 | return true; |
d7400a34 XG |
2048 | } |
2049 | ||
5e5fdcff XG |
2050 | /* |
2051 | * try to compress the page before posting it out, return true if the page | |
2052 | * has been properly handled by compression, otherwise needs other | |
2053 | * paths to handle it | |
2054 | */ | |
2055 | static bool save_compress_page(RAMState *rs, RAMBlock *block, ram_addr_t offset) | |
2056 | { | |
2057 | if (!save_page_use_compression(rs)) { | |
2058 | return false; | |
2059 | } | |
2060 | ||
2061 | /* | |
2062 | * When starting the process of a new block, the first page of | |
2063 | * the block should be sent out before other pages in the same | |
2064 | * block, and all the pages in last block should have been sent | |
2065 | * out, keeping this order is important, because the 'cont' flag | |
2066 | * is used to avoid resending the block name. | |
2067 | * | |
2068 | * We post the fist page as normal page as compression will take | |
2069 | * much CPU resource. | |
2070 | */ | |
2071 | if (block != rs->last_sent_block) { | |
2072 | flush_compressed_data(rs); | |
2073 | return false; | |
2074 | } | |
2075 | ||
2076 | if (compress_page_with_multi_thread(rs, block, offset) > 0) { | |
2077 | return true; | |
2078 | } | |
2079 | ||
76e03000 | 2080 | compression_counters.busy++; |
5e5fdcff XG |
2081 | return false; |
2082 | } | |
2083 | ||
a82d593b | 2084 | /** |
3d0684b2 | 2085 | * ram_save_target_page: save one target page |
a82d593b | 2086 | * |
3d0684b2 | 2087 | * Returns the number of pages written |
a82d593b | 2088 | * |
6f37bb8b | 2089 | * @rs: current RAM state |
3d0684b2 | 2090 | * @pss: data about the page we want to send |
a82d593b | 2091 | * @last_stage: if we are at the completion stage |
a82d593b | 2092 | */ |
a0a8aa14 | 2093 | static int ram_save_target_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 2094 | bool last_stage) |
a82d593b | 2095 | { |
a8ec91f9 | 2096 | RAMBlock *block = pss->block; |
8bba004c | 2097 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
2098 | int res; |
2099 | ||
2100 | if (control_save_page(rs, block, offset, &res)) { | |
2101 | return res; | |
2102 | } | |
2103 | ||
5e5fdcff XG |
2104 | if (save_compress_page(rs, block, offset)) { |
2105 | return 1; | |
d7400a34 XG |
2106 | } |
2107 | ||
2108 | res = save_zero_page(rs, block, offset); | |
2109 | if (res > 0) { | |
2110 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
2111 | * page would be stale | |
2112 | */ | |
2113 | if (!save_page_use_compression(rs)) { | |
2114 | XBZRLE_cache_lock(); | |
2115 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
2116 | XBZRLE_cache_unlock(); | |
2117 | } | |
2118 | ram_release_pages(block->idstr, offset, res); | |
2119 | return res; | |
2120 | } | |
2121 | ||
da3f56cb | 2122 | /* |
c6b3a2e0 WY |
2123 | * Do not use multifd for: |
2124 | * 1. Compression as the first page in the new block should be posted out | |
2125 | * before sending the compressed page | |
2126 | * 2. In postcopy as one whole host page should be placed | |
da3f56cb | 2127 | */ |
c6b3a2e0 WY |
2128 | if (!save_page_use_compression(rs) && migrate_use_multifd() |
2129 | && !migration_in_postcopy()) { | |
b9ee2f7d | 2130 | return ram_save_multifd_page(rs, block, offset); |
a82d593b DDAG |
2131 | } |
2132 | ||
1faa5665 | 2133 | return ram_save_page(rs, pss, last_stage); |
a82d593b DDAG |
2134 | } |
2135 | ||
2136 | /** | |
3d0684b2 | 2137 | * ram_save_host_page: save a whole host page |
a82d593b | 2138 | * |
3d0684b2 JQ |
2139 | * Starting at *offset send pages up to the end of the current host |
2140 | * page. It's valid for the initial offset to point into the middle of | |
2141 | * a host page in which case the remainder of the hostpage is sent. | |
2142 | * Only dirty target pages are sent. Note that the host page size may | |
2143 | * be a huge page for this block. | |
1eb3fc0a DDAG |
2144 | * The saving stops at the boundary of the used_length of the block |
2145 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 2146 | * |
3d0684b2 JQ |
2147 | * Returns the number of pages written or negative on error |
2148 | * | |
6f37bb8b | 2149 | * @rs: current RAM state |
3d0684b2 | 2150 | * @ms: current migration state |
3d0684b2 | 2151 | * @pss: data about the page we want to send |
a82d593b | 2152 | * @last_stage: if we are at the completion stage |
a82d593b | 2153 | */ |
a0a8aa14 | 2154 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss, |
f20e2865 | 2155 | bool last_stage) |
a82d593b DDAG |
2156 | { |
2157 | int tmppages, pages = 0; | |
a935e30f JQ |
2158 | size_t pagesize_bits = |
2159 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
ba1b7c81 KJ |
2160 | unsigned long hostpage_boundary = |
2161 | QEMU_ALIGN_UP(pss->page + 1, pagesize_bits); | |
278e2f55 AG |
2162 | unsigned long start_page = pss->page; |
2163 | int res; | |
4c011c37 | 2164 | |
fbd162e6 | 2165 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
2166 | error_report("block %s should not be migrated !", pss->block->idstr); |
2167 | return 0; | |
2168 | } | |
2169 | ||
a82d593b | 2170 | do { |
1faa5665 | 2171 | /* Check the pages is dirty and if it is send it */ |
ba1b7c81 KJ |
2172 | if (migration_bitmap_clear_dirty(rs, pss->block, pss->page)) { |
2173 | tmppages = ram_save_target_page(rs, pss, last_stage); | |
2174 | if (tmppages < 0) { | |
2175 | return tmppages; | |
2176 | } | |
a82d593b | 2177 | |
ba1b7c81 KJ |
2178 | pages += tmppages; |
2179 | /* | |
2180 | * Allow rate limiting to happen in the middle of huge pages if | |
2181 | * something is sent in the current iteration. | |
2182 | */ | |
2183 | if (pagesize_bits > 1 && tmppages > 0) { | |
2184 | migration_rate_limit(); | |
2185 | } | |
23feba90 | 2186 | } |
ba1b7c81 KJ |
2187 | pss->page = migration_bitmap_find_dirty(rs, pss->block, pss->page); |
2188 | } while ((pss->page < hostpage_boundary) && | |
8bba004c AR |
2189 | offset_in_ramblock(pss->block, |
2190 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)); | |
ba1b7c81 KJ |
2191 | /* The offset we leave with is the min boundary of host page and block */ |
2192 | pss->page = MIN(pss->page, hostpage_boundary) - 1; | |
278e2f55 AG |
2193 | |
2194 | res = ram_save_release_protection(rs, pss, start_page); | |
2195 | return (res < 0 ? res : pages); | |
a82d593b | 2196 | } |
6c595cde | 2197 | |
56e93d26 | 2198 | /** |
3d0684b2 | 2199 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
2200 | * |
2201 | * Called within an RCU critical section. | |
2202 | * | |
e8f3735f XG |
2203 | * Returns the number of pages written where zero means no dirty pages, |
2204 | * or negative on error | |
56e93d26 | 2205 | * |
6f37bb8b | 2206 | * @rs: current RAM state |
56e93d26 | 2207 | * @last_stage: if we are at the completion stage |
a82d593b DDAG |
2208 | * |
2209 | * On systems where host-page-size > target-page-size it will send all the | |
2210 | * pages in a host page that are dirty. | |
56e93d26 JQ |
2211 | */ |
2212 | ||
ce25d337 | 2213 | static int ram_find_and_save_block(RAMState *rs, bool last_stage) |
56e93d26 | 2214 | { |
b8fb8cb7 | 2215 | PageSearchStatus pss; |
56e93d26 | 2216 | int pages = 0; |
b9e60928 | 2217 | bool again, found; |
56e93d26 | 2218 | |
0827b9e9 AA |
2219 | /* No dirty page as there is zero RAM */ |
2220 | if (!ram_bytes_total()) { | |
2221 | return pages; | |
2222 | } | |
2223 | ||
6f37bb8b | 2224 | pss.block = rs->last_seen_block; |
a935e30f | 2225 | pss.page = rs->last_page; |
b8fb8cb7 DDAG |
2226 | pss.complete_round = false; |
2227 | ||
2228 | if (!pss.block) { | |
2229 | pss.block = QLIST_FIRST_RCU(&ram_list.blocks); | |
2230 | } | |
56e93d26 | 2231 | |
b9e60928 | 2232 | do { |
a82d593b | 2233 | again = true; |
f20e2865 | 2234 | found = get_queued_page(rs, &pss); |
b9e60928 | 2235 | |
a82d593b DDAG |
2236 | if (!found) { |
2237 | /* priority queue empty, so just search for something dirty */ | |
f20e2865 | 2238 | found = find_dirty_block(rs, &pss, &again); |
a82d593b | 2239 | } |
f3f491fc | 2240 | |
a82d593b | 2241 | if (found) { |
f20e2865 | 2242 | pages = ram_save_host_page(rs, &pss, last_stage); |
56e93d26 | 2243 | } |
b9e60928 | 2244 | } while (!pages && again); |
56e93d26 | 2245 | |
6f37bb8b | 2246 | rs->last_seen_block = pss.block; |
a935e30f | 2247 | rs->last_page = pss.page; |
56e93d26 JQ |
2248 | |
2249 | return pages; | |
2250 | } | |
2251 | ||
2252 | void acct_update_position(QEMUFile *f, size_t size, bool zero) | |
2253 | { | |
2254 | uint64_t pages = size / TARGET_PAGE_SIZE; | |
f7ccd61b | 2255 | |
56e93d26 | 2256 | if (zero) { |
9360447d | 2257 | ram_counters.duplicate += pages; |
56e93d26 | 2258 | } else { |
9360447d JQ |
2259 | ram_counters.normal += pages; |
2260 | ram_counters.transferred += size; | |
56e93d26 JQ |
2261 | qemu_update_position(f, size); |
2262 | } | |
2263 | } | |
2264 | ||
fbd162e6 | 2265 | static uint64_t ram_bytes_total_common(bool count_ignored) |
56e93d26 JQ |
2266 | { |
2267 | RAMBlock *block; | |
2268 | uint64_t total = 0; | |
2269 | ||
89ac5a1d DDAG |
2270 | RCU_READ_LOCK_GUARD(); |
2271 | ||
fbd162e6 YK |
2272 | if (count_ignored) { |
2273 | RAMBLOCK_FOREACH_MIGRATABLE(block) { | |
2274 | total += block->used_length; | |
2275 | } | |
2276 | } else { | |
2277 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2278 | total += block->used_length; | |
2279 | } | |
99e15582 | 2280 | } |
56e93d26 JQ |
2281 | return total; |
2282 | } | |
2283 | ||
fbd162e6 YK |
2284 | uint64_t ram_bytes_total(void) |
2285 | { | |
2286 | return ram_bytes_total_common(false); | |
2287 | } | |
2288 | ||
f265e0e4 | 2289 | static void xbzrle_load_setup(void) |
56e93d26 | 2290 | { |
f265e0e4 | 2291 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2292 | } |
2293 | ||
f265e0e4 JQ |
2294 | static void xbzrle_load_cleanup(void) |
2295 | { | |
2296 | g_free(XBZRLE.decoded_buf); | |
2297 | XBZRLE.decoded_buf = NULL; | |
2298 | } | |
2299 | ||
7d7c96be PX |
2300 | static void ram_state_cleanup(RAMState **rsp) |
2301 | { | |
b9ccaf6d DDAG |
2302 | if (*rsp) { |
2303 | migration_page_queue_free(*rsp); | |
2304 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
2305 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
2306 | g_free(*rsp); | |
2307 | *rsp = NULL; | |
2308 | } | |
7d7c96be PX |
2309 | } |
2310 | ||
84593a08 PX |
2311 | static void xbzrle_cleanup(void) |
2312 | { | |
2313 | XBZRLE_cache_lock(); | |
2314 | if (XBZRLE.cache) { | |
2315 | cache_fini(XBZRLE.cache); | |
2316 | g_free(XBZRLE.encoded_buf); | |
2317 | g_free(XBZRLE.current_buf); | |
2318 | g_free(XBZRLE.zero_target_page); | |
2319 | XBZRLE.cache = NULL; | |
2320 | XBZRLE.encoded_buf = NULL; | |
2321 | XBZRLE.current_buf = NULL; | |
2322 | XBZRLE.zero_target_page = NULL; | |
2323 | } | |
2324 | XBZRLE_cache_unlock(); | |
2325 | } | |
2326 | ||
f265e0e4 | 2327 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 2328 | { |
53518d94 | 2329 | RAMState **rsp = opaque; |
6b6712ef | 2330 | RAMBlock *block; |
eb859c53 | 2331 | |
278e2f55 AG |
2332 | /* We don't use dirty log with background snapshots */ |
2333 | if (!migrate_background_snapshot()) { | |
2334 | /* caller have hold iothread lock or is in a bh, so there is | |
2335 | * no writing race against the migration bitmap | |
2336 | */ | |
63b41db4 HH |
2337 | if (global_dirty_tracking & GLOBAL_DIRTY_MIGRATION) { |
2338 | /* | |
2339 | * do not stop dirty log without starting it, since | |
2340 | * memory_global_dirty_log_stop will assert that | |
2341 | * memory_global_dirty_log_start/stop used in pairs | |
2342 | */ | |
2343 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); | |
2344 | } | |
278e2f55 | 2345 | } |
6b6712ef | 2346 | |
fbd162e6 | 2347 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
2348 | g_free(block->clear_bmap); |
2349 | block->clear_bmap = NULL; | |
6b6712ef JQ |
2350 | g_free(block->bmap); |
2351 | block->bmap = NULL; | |
56e93d26 JQ |
2352 | } |
2353 | ||
84593a08 | 2354 | xbzrle_cleanup(); |
f0afa331 | 2355 | compress_threads_save_cleanup(); |
7d7c96be | 2356 | ram_state_cleanup(rsp); |
56e93d26 JQ |
2357 | } |
2358 | ||
6f37bb8b | 2359 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 2360 | { |
6f37bb8b JQ |
2361 | rs->last_seen_block = NULL; |
2362 | rs->last_sent_block = NULL; | |
269ace29 | 2363 | rs->last_page = 0; |
6f37bb8b | 2364 | rs->last_version = ram_list.version; |
1a373522 | 2365 | rs->xbzrle_enabled = false; |
56e93d26 JQ |
2366 | } |
2367 | ||
2368 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
2369 | ||
4f2e4252 DDAG |
2370 | /* |
2371 | * 'expected' is the value you expect the bitmap mostly to be full | |
2372 | * of; it won't bother printing lines that are all this value. | |
2373 | * If 'todump' is null the migration bitmap is dumped. | |
2374 | */ | |
6b6712ef JQ |
2375 | void ram_debug_dump_bitmap(unsigned long *todump, bool expected, |
2376 | unsigned long pages) | |
4f2e4252 | 2377 | { |
4f2e4252 DDAG |
2378 | int64_t cur; |
2379 | int64_t linelen = 128; | |
2380 | char linebuf[129]; | |
2381 | ||
6b6712ef | 2382 | for (cur = 0; cur < pages; cur += linelen) { |
4f2e4252 DDAG |
2383 | int64_t curb; |
2384 | bool found = false; | |
2385 | /* | |
2386 | * Last line; catch the case where the line length | |
2387 | * is longer than remaining ram | |
2388 | */ | |
6b6712ef JQ |
2389 | if (cur + linelen > pages) { |
2390 | linelen = pages - cur; | |
4f2e4252 DDAG |
2391 | } |
2392 | for (curb = 0; curb < linelen; curb++) { | |
2393 | bool thisbit = test_bit(cur + curb, todump); | |
2394 | linebuf[curb] = thisbit ? '1' : '.'; | |
2395 | found = found || (thisbit != expected); | |
2396 | } | |
2397 | if (found) { | |
2398 | linebuf[curb] = '\0'; | |
2399 | fprintf(stderr, "0x%08" PRIx64 " : %s\n", cur, linebuf); | |
2400 | } | |
2401 | } | |
2402 | } | |
2403 | ||
e0b266f0 DDAG |
2404 | /* **** functions for postcopy ***** */ |
2405 | ||
ced1c616 PB |
2406 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
2407 | { | |
2408 | struct RAMBlock *block; | |
ced1c616 | 2409 | |
fbd162e6 | 2410 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2411 | unsigned long *bitmap = block->bmap; |
2412 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
2413 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
2414 | |
2415 | while (run_start < range) { | |
2416 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
2417 | ram_discard_range(block->idstr, |
2418 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
2419 | ((ram_addr_t)(run_end - run_start)) | |
2420 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
2421 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
2422 | } | |
2423 | } | |
2424 | } | |
2425 | ||
3d0684b2 JQ |
2426 | /** |
2427 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
2428 | * | |
2429 | * Returns zero on success | |
2430 | * | |
e0b266f0 | 2431 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
2432 | * |
2433 | * @ms: current migration state | |
89dab31b | 2434 | * @block: RAMBlock to discard |
e0b266f0 | 2435 | */ |
810cf2bb | 2436 | static int postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 2437 | { |
6b6712ef | 2438 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 2439 | unsigned long current; |
1e7cf8c3 | 2440 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 2441 | |
6b6712ef | 2442 | for (current = 0; current < end; ) { |
1e7cf8c3 | 2443 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 2444 | unsigned long zero, discard_length; |
e0b266f0 | 2445 | |
33a5cb62 WY |
2446 | if (one >= end) { |
2447 | break; | |
2448 | } | |
e0b266f0 | 2449 | |
1e7cf8c3 | 2450 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
2451 | |
2452 | if (zero >= end) { | |
2453 | discard_length = end - one; | |
e0b266f0 | 2454 | } else { |
33a5cb62 WY |
2455 | discard_length = zero - one; |
2456 | } | |
810cf2bb | 2457 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 2458 | current = one + discard_length; |
e0b266f0 DDAG |
2459 | } |
2460 | ||
2461 | return 0; | |
2462 | } | |
2463 | ||
3d0684b2 JQ |
2464 | /** |
2465 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
2466 | * | |
2467 | * Returns 0 for success or negative for error | |
2468 | * | |
e0b266f0 DDAG |
2469 | * Utility for the outgoing postcopy code. |
2470 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
2471 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
2472 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
2473 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
2474 | * |
2475 | * @ms: current migration state | |
e0b266f0 DDAG |
2476 | */ |
2477 | static int postcopy_each_ram_send_discard(MigrationState *ms) | |
2478 | { | |
2479 | struct RAMBlock *block; | |
2480 | int ret; | |
2481 | ||
fbd162e6 | 2482 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 2483 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 DDAG |
2484 | |
2485 | /* | |
2486 | * Postcopy sends chunks of bitmap over the wire, but it | |
2487 | * just needs indexes at this point, avoids it having | |
2488 | * target page specific code. | |
2489 | */ | |
810cf2bb WY |
2490 | ret = postcopy_send_discard_bm_ram(ms, block); |
2491 | postcopy_discard_send_finish(ms); | |
e0b266f0 DDAG |
2492 | if (ret) { |
2493 | return ret; | |
2494 | } | |
2495 | } | |
2496 | ||
2497 | return 0; | |
2498 | } | |
2499 | ||
3d0684b2 | 2500 | /** |
8324ef86 | 2501 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2502 | * |
2503 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2504 | * canonicalize the two bitmaps, that are similar, but one is | |
2505 | * inverted. | |
99e314eb | 2506 | * |
3d0684b2 JQ |
2507 | * Postcopy requires that all target pages in a hostpage are dirty or |
2508 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2509 | * |
3d0684b2 | 2510 | * @ms: current migration state |
3d0684b2 | 2511 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2512 | */ |
1e7cf8c3 | 2513 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2514 | { |
53518d94 | 2515 | RAMState *rs = ram_state; |
6b6712ef | 2516 | unsigned long *bitmap = block->bmap; |
29c59172 | 2517 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2518 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2519 | unsigned long run_start; |
2520 | ||
29c59172 DDAG |
2521 | if (block->page_size == TARGET_PAGE_SIZE) { |
2522 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2523 | return; | |
2524 | } | |
2525 | ||
1e7cf8c3 WY |
2526 | /* Find a dirty page */ |
2527 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2528 | |
6b6712ef | 2529 | while (run_start < pages) { |
99e314eb DDAG |
2530 | |
2531 | /* | |
2532 | * If the start of this run of pages is in the middle of a host | |
2533 | * page, then we need to fixup this host page. | |
2534 | */ | |
9dec3cc3 | 2535 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2536 | /* Find the end of this run */ |
1e7cf8c3 | 2537 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2538 | /* |
2539 | * If the end isn't at the start of a host page, then the | |
2540 | * run doesn't finish at the end of a host page | |
2541 | * and we need to discard. | |
2542 | */ | |
99e314eb DDAG |
2543 | } |
2544 | ||
9dec3cc3 | 2545 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2546 | unsigned long page; |
dad45ab2 WY |
2547 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2548 | host_ratio); | |
2549 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2550 | |
99e314eb DDAG |
2551 | /* Clean up the bitmap */ |
2552 | for (page = fixup_start_addr; | |
2553 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2554 | /* |
2555 | * Remark them as dirty, updating the count for any pages | |
2556 | * that weren't previously dirty. | |
2557 | */ | |
0d8ec885 | 2558 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2559 | } |
2560 | } | |
2561 | ||
1e7cf8c3 WY |
2562 | /* Find the next dirty page for the next iteration */ |
2563 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2564 | } |
2565 | } | |
2566 | ||
3d0684b2 | 2567 | /** |
89dab31b | 2568 | * postcopy_chunk_hostpages: discard any partially sent host page |
3d0684b2 | 2569 | * |
99e314eb DDAG |
2570 | * Utility for the outgoing postcopy code. |
2571 | * | |
2572 | * Discard any partially sent host-page size chunks, mark any partially | |
29c59172 DDAG |
2573 | * dirty host-page size chunks as all dirty. In this case the host-page |
2574 | * is the host-page for the particular RAMBlock, i.e. it might be a huge page | |
99e314eb | 2575 | * |
3d0684b2 JQ |
2576 | * Returns zero on success |
2577 | * | |
2578 | * @ms: current migration state | |
6b6712ef | 2579 | * @block: block we want to work with |
99e314eb | 2580 | */ |
6b6712ef | 2581 | static int postcopy_chunk_hostpages(MigrationState *ms, RAMBlock *block) |
99e314eb | 2582 | { |
810cf2bb | 2583 | postcopy_discard_send_init(ms, block->idstr); |
99e314eb | 2584 | |
6b6712ef | 2585 | /* |
1e7cf8c3 | 2586 | * Ensure that all partially dirty host pages are made fully dirty. |
6b6712ef | 2587 | */ |
1e7cf8c3 | 2588 | postcopy_chunk_hostpages_pass(ms, block); |
99e314eb | 2589 | |
810cf2bb | 2590 | postcopy_discard_send_finish(ms); |
99e314eb DDAG |
2591 | return 0; |
2592 | } | |
2593 | ||
3d0684b2 JQ |
2594 | /** |
2595 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2596 | * | |
2597 | * Returns zero on success | |
2598 | * | |
e0b266f0 DDAG |
2599 | * Transmit the set of pages to be discarded after precopy to the target |
2600 | * these are pages that: | |
2601 | * a) Have been previously transmitted but are now dirty again | |
2602 | * b) Pages that have never been transmitted, this ensures that | |
2603 | * any pages on the destination that have been mapped by background | |
2604 | * tasks get discarded (transparent huge pages is the specific concern) | |
2605 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2606 | * |
2607 | * @ms: current migration state | |
e0b266f0 DDAG |
2608 | */ |
2609 | int ram_postcopy_send_discard_bitmap(MigrationState *ms) | |
2610 | { | |
53518d94 | 2611 | RAMState *rs = ram_state; |
6b6712ef | 2612 | RAMBlock *block; |
e0b266f0 | 2613 | int ret; |
e0b266f0 | 2614 | |
89ac5a1d | 2615 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2616 | |
2617 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 2618 | migration_bitmap_sync(rs); |
e0b266f0 | 2619 | |
6b6712ef JQ |
2620 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
2621 | rs->last_seen_block = NULL; | |
2622 | rs->last_sent_block = NULL; | |
2623 | rs->last_page = 0; | |
e0b266f0 | 2624 | |
fbd162e6 | 2625 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2626 | /* Deal with TPS != HPS and huge pages */ |
2627 | ret = postcopy_chunk_hostpages(ms, block); | |
2628 | if (ret) { | |
6b6712ef JQ |
2629 | return ret; |
2630 | } | |
e0b266f0 | 2631 | |
e0b266f0 | 2632 | #ifdef DEBUG_POSTCOPY |
1e7cf8c3 WY |
2633 | ram_debug_dump_bitmap(block->bmap, true, |
2634 | block->used_length >> TARGET_PAGE_BITS); | |
e0b266f0 | 2635 | #endif |
6b6712ef JQ |
2636 | } |
2637 | trace_ram_postcopy_send_discard_bitmap(); | |
e0b266f0 | 2638 | |
b3ac2b94 | 2639 | return postcopy_each_ram_send_discard(ms); |
e0b266f0 DDAG |
2640 | } |
2641 | ||
3d0684b2 JQ |
2642 | /** |
2643 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2644 | * |
3d0684b2 | 2645 | * Returns zero on success |
e0b266f0 | 2646 | * |
36449157 JQ |
2647 | * @rbname: name of the RAMBlock of the request. NULL means the |
2648 | * same that last one. | |
3d0684b2 JQ |
2649 | * @start: RAMBlock starting page |
2650 | * @length: RAMBlock size | |
e0b266f0 | 2651 | */ |
aaa2064c | 2652 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 2653 | { |
36449157 | 2654 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2655 | |
89ac5a1d | 2656 | RCU_READ_LOCK_GUARD(); |
36449157 | 2657 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2658 | |
2659 | if (!rb) { | |
36449157 | 2660 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 2661 | return -1; |
e0b266f0 DDAG |
2662 | } |
2663 | ||
814bb08f PX |
2664 | /* |
2665 | * On source VM, we don't need to update the received bitmap since | |
2666 | * we don't even have one. | |
2667 | */ | |
2668 | if (rb->receivedmap) { | |
2669 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
2670 | length >> qemu_target_page_bits()); | |
2671 | } | |
2672 | ||
03acb4e9 | 2673 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2674 | } |
2675 | ||
84593a08 PX |
2676 | /* |
2677 | * For every allocation, we will try not to crash the VM if the | |
2678 | * allocation failed. | |
2679 | */ | |
2680 | static int xbzrle_init(void) | |
2681 | { | |
2682 | Error *local_err = NULL; | |
2683 | ||
2684 | if (!migrate_use_xbzrle()) { | |
2685 | return 0; | |
2686 | } | |
2687 | ||
2688 | XBZRLE_cache_lock(); | |
2689 | ||
2690 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
2691 | if (!XBZRLE.zero_target_page) { | |
2692 | error_report("%s: Error allocating zero page", __func__); | |
2693 | goto err_out; | |
2694 | } | |
2695 | ||
2696 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
2697 | TARGET_PAGE_SIZE, &local_err); | |
2698 | if (!XBZRLE.cache) { | |
2699 | error_report_err(local_err); | |
2700 | goto free_zero_page; | |
2701 | } | |
2702 | ||
2703 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2704 | if (!XBZRLE.encoded_buf) { | |
2705 | error_report("%s: Error allocating encoded_buf", __func__); | |
2706 | goto free_cache; | |
2707 | } | |
2708 | ||
2709 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2710 | if (!XBZRLE.current_buf) { | |
2711 | error_report("%s: Error allocating current_buf", __func__); | |
2712 | goto free_encoded_buf; | |
2713 | } | |
2714 | ||
2715 | /* We are all good */ | |
2716 | XBZRLE_cache_unlock(); | |
2717 | return 0; | |
2718 | ||
2719 | free_encoded_buf: | |
2720 | g_free(XBZRLE.encoded_buf); | |
2721 | XBZRLE.encoded_buf = NULL; | |
2722 | free_cache: | |
2723 | cache_fini(XBZRLE.cache); | |
2724 | XBZRLE.cache = NULL; | |
2725 | free_zero_page: | |
2726 | g_free(XBZRLE.zero_target_page); | |
2727 | XBZRLE.zero_target_page = NULL; | |
2728 | err_out: | |
2729 | XBZRLE_cache_unlock(); | |
2730 | return -ENOMEM; | |
2731 | } | |
2732 | ||
53518d94 | 2733 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2734 | { |
7d00ee6a PX |
2735 | *rsp = g_try_new0(RAMState, 1); |
2736 | ||
2737 | if (!*rsp) { | |
2738 | error_report("%s: Init ramstate fail", __func__); | |
2739 | return -1; | |
2740 | } | |
53518d94 JQ |
2741 | |
2742 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
2743 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
2744 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
56e93d26 | 2745 | |
7d00ee6a | 2746 | /* |
40c4d4a8 IR |
2747 | * Count the total number of pages used by ram blocks not including any |
2748 | * gaps due to alignment or unplugs. | |
03158519 | 2749 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 2750 | */ |
40c4d4a8 | 2751 | (*rsp)->migration_dirty_pages = ram_bytes_total() >> TARGET_PAGE_BITS; |
7d00ee6a PX |
2752 | ram_state_reset(*rsp); |
2753 | ||
2754 | return 0; | |
2755 | } | |
2756 | ||
d6eff5d7 | 2757 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 2758 | { |
002cad6b | 2759 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
2760 | RAMBlock *block; |
2761 | unsigned long pages; | |
002cad6b | 2762 | uint8_t shift; |
56e93d26 | 2763 | |
0827b9e9 AA |
2764 | /* Skip setting bitmap if there is no RAM */ |
2765 | if (ram_bytes_total()) { | |
002cad6b PX |
2766 | shift = ms->clear_bitmap_shift; |
2767 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
2768 | error_report("clear_bitmap_shift (%u) too big, using " | |
2769 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
2770 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
2771 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
2772 | error_report("clear_bitmap_shift (%u) too small, using " | |
2773 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
2774 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
2775 | } | |
2776 | ||
fbd162e6 | 2777 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 2778 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
2779 | /* |
2780 | * The initial dirty bitmap for migration must be set with all | |
2781 | * ones to make sure we'll migrate every guest RAM page to | |
2782 | * destination. | |
40c4d4a8 IR |
2783 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
2784 | * new migration after a failed migration, ram_list. | |
2785 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
2786 | * guest memory. | |
03158519 | 2787 | */ |
6b6712ef | 2788 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 2789 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
2790 | block->clear_bmap_shift = shift; |
2791 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 2792 | } |
f3f491fc | 2793 | } |
d6eff5d7 PX |
2794 | } |
2795 | ||
be39b4cd DH |
2796 | static void migration_bitmap_clear_discarded_pages(RAMState *rs) |
2797 | { | |
2798 | unsigned long pages; | |
2799 | RAMBlock *rb; | |
2800 | ||
2801 | RCU_READ_LOCK_GUARD(); | |
2802 | ||
2803 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { | |
2804 | pages = ramblock_dirty_bitmap_clear_discarded_pages(rb); | |
2805 | rs->migration_dirty_pages -= pages; | |
2806 | } | |
2807 | } | |
2808 | ||
d6eff5d7 PX |
2809 | static void ram_init_bitmaps(RAMState *rs) |
2810 | { | |
2811 | /* For memory_global_dirty_log_start below. */ | |
2812 | qemu_mutex_lock_iothread(); | |
2813 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 2814 | |
89ac5a1d DDAG |
2815 | WITH_RCU_READ_LOCK_GUARD() { |
2816 | ram_list_init_bitmaps(); | |
278e2f55 AG |
2817 | /* We don't use dirty log with background snapshots */ |
2818 | if (!migrate_background_snapshot()) { | |
63b41db4 | 2819 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
278e2f55 AG |
2820 | migration_bitmap_sync_precopy(rs); |
2821 | } | |
89ac5a1d | 2822 | } |
56e93d26 | 2823 | qemu_mutex_unlock_ramlist(); |
49877834 | 2824 | qemu_mutex_unlock_iothread(); |
be39b4cd DH |
2825 | |
2826 | /* | |
2827 | * After an eventual first bitmap sync, fixup the initial bitmap | |
2828 | * containing all 1s to exclude any discarded pages from migration. | |
2829 | */ | |
2830 | migration_bitmap_clear_discarded_pages(rs); | |
d6eff5d7 PX |
2831 | } |
2832 | ||
2833 | static int ram_init_all(RAMState **rsp) | |
2834 | { | |
2835 | if (ram_state_init(rsp)) { | |
2836 | return -1; | |
2837 | } | |
2838 | ||
2839 | if (xbzrle_init()) { | |
2840 | ram_state_cleanup(rsp); | |
2841 | return -1; | |
2842 | } | |
2843 | ||
2844 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
2845 | |
2846 | return 0; | |
2847 | } | |
2848 | ||
08614f34 PX |
2849 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
2850 | { | |
2851 | RAMBlock *block; | |
2852 | uint64_t pages = 0; | |
2853 | ||
2854 | /* | |
2855 | * Postcopy is not using xbzrle/compression, so no need for that. | |
2856 | * Also, since source are already halted, we don't need to care | |
2857 | * about dirty page logging as well. | |
2858 | */ | |
2859 | ||
fbd162e6 | 2860 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
2861 | pages += bitmap_count_one(block->bmap, |
2862 | block->used_length >> TARGET_PAGE_BITS); | |
2863 | } | |
2864 | ||
2865 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
2866 | rs->migration_dirty_pages = pages; | |
2867 | ||
1a373522 | 2868 | ram_state_reset(rs); |
08614f34 PX |
2869 | |
2870 | /* Update RAMState cache of output QEMUFile */ | |
2871 | rs->f = out; | |
2872 | ||
2873 | trace_ram_state_resume_prepare(pages); | |
2874 | } | |
2875 | ||
6bcb05fc WW |
2876 | /* |
2877 | * This function clears bits of the free pages reported by the caller from the | |
2878 | * migration dirty bitmap. @addr is the host address corresponding to the | |
2879 | * start of the continuous guest free pages, and @len is the total bytes of | |
2880 | * those pages. | |
2881 | */ | |
2882 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
2883 | { | |
2884 | RAMBlock *block; | |
2885 | ram_addr_t offset; | |
2886 | size_t used_len, start, npages; | |
2887 | MigrationState *s = migrate_get_current(); | |
2888 | ||
2889 | /* This function is currently expected to be used during live migration */ | |
2890 | if (!migration_is_setup_or_active(s->state)) { | |
2891 | return; | |
2892 | } | |
2893 | ||
2894 | for (; len > 0; len -= used_len, addr += used_len) { | |
2895 | block = qemu_ram_block_from_host(addr, false, &offset); | |
2896 | if (unlikely(!block || offset >= block->used_length)) { | |
2897 | /* | |
2898 | * The implementation might not support RAMBlock resize during | |
2899 | * live migration, but it could happen in theory with future | |
2900 | * updates. So we add a check here to capture that case. | |
2901 | */ | |
2902 | error_report_once("%s unexpected error", __func__); | |
2903 | return; | |
2904 | } | |
2905 | ||
2906 | if (len <= block->used_length - offset) { | |
2907 | used_len = len; | |
2908 | } else { | |
2909 | used_len = block->used_length - offset; | |
2910 | } | |
2911 | ||
2912 | start = offset >> TARGET_PAGE_BITS; | |
2913 | npages = used_len >> TARGET_PAGE_BITS; | |
2914 | ||
2915 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3143577d WW |
2916 | /* |
2917 | * The skipped free pages are equavalent to be sent from clear_bmap's | |
2918 | * perspective, so clear the bits from the memory region bitmap which | |
2919 | * are initially set. Otherwise those skipped pages will be sent in | |
2920 | * the next round after syncing from the memory region bitmap. | |
2921 | */ | |
1230a25f | 2922 | migration_clear_memory_region_dirty_bitmap_range(block, start, npages); |
6bcb05fc WW |
2923 | ram_state->migration_dirty_pages -= |
2924 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
2925 | bitmap_clear(block->bmap, start, npages); | |
2926 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
2927 | } | |
2928 | } | |
2929 | ||
3d0684b2 JQ |
2930 | /* |
2931 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
2932 | * long-running RCU critical section. When rcu-reclaims in the code |
2933 | * start to become numerous it will be necessary to reduce the | |
2934 | * granularity of these critical sections. | |
2935 | */ | |
2936 | ||
3d0684b2 JQ |
2937 | /** |
2938 | * ram_save_setup: Setup RAM for migration | |
2939 | * | |
2940 | * Returns zero to indicate success and negative for error | |
2941 | * | |
2942 | * @f: QEMUFile where to send the data | |
2943 | * @opaque: RAMState pointer | |
2944 | */ | |
a91246c9 HZ |
2945 | static int ram_save_setup(QEMUFile *f, void *opaque) |
2946 | { | |
53518d94 | 2947 | RAMState **rsp = opaque; |
a91246c9 HZ |
2948 | RAMBlock *block; |
2949 | ||
dcaf446e XG |
2950 | if (compress_threads_save_setup()) { |
2951 | return -1; | |
2952 | } | |
2953 | ||
a91246c9 HZ |
2954 | /* migration has already setup the bitmap, reuse it. */ |
2955 | if (!migration_in_colo_state()) { | |
7d00ee6a | 2956 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 2957 | compress_threads_save_cleanup(); |
a91246c9 | 2958 | return -1; |
53518d94 | 2959 | } |
a91246c9 | 2960 | } |
53518d94 | 2961 | (*rsp)->f = f; |
a91246c9 | 2962 | |
0e6ebd48 DDAG |
2963 | WITH_RCU_READ_LOCK_GUARD() { |
2964 | qemu_put_be64(f, ram_bytes_total_common(true) | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 2965 | |
0e6ebd48 DDAG |
2966 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
2967 | qemu_put_byte(f, strlen(block->idstr)); | |
2968 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
2969 | qemu_put_be64(f, block->used_length); | |
2970 | if (migrate_postcopy_ram() && block->page_size != | |
2971 | qemu_host_page_size) { | |
2972 | qemu_put_be64(f, block->page_size); | |
2973 | } | |
2974 | if (migrate_ignore_shared()) { | |
2975 | qemu_put_be64(f, block->mr->addr); | |
2976 | } | |
fbd162e6 | 2977 | } |
56e93d26 JQ |
2978 | } |
2979 | ||
56e93d26 JQ |
2980 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
2981 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
2982 | ||
99f2c6fb | 2983 | multifd_send_sync_main(f); |
56e93d26 | 2984 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 2985 | qemu_fflush(f); |
56e93d26 JQ |
2986 | |
2987 | return 0; | |
2988 | } | |
2989 | ||
3d0684b2 JQ |
2990 | /** |
2991 | * ram_save_iterate: iterative stage for migration | |
2992 | * | |
2993 | * Returns zero to indicate success and negative for error | |
2994 | * | |
2995 | * @f: QEMUFile where to send the data | |
2996 | * @opaque: RAMState pointer | |
2997 | */ | |
56e93d26 JQ |
2998 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
2999 | { | |
53518d94 JQ |
3000 | RAMState **temp = opaque; |
3001 | RAMState *rs = *temp; | |
3d4095b2 | 3002 | int ret = 0; |
56e93d26 JQ |
3003 | int i; |
3004 | int64_t t0; | |
5c90308f | 3005 | int done = 0; |
56e93d26 | 3006 | |
b2557345 PL |
3007 | if (blk_mig_bulk_active()) { |
3008 | /* Avoid transferring ram during bulk phase of block migration as | |
3009 | * the bulk phase will usually take a long time and transferring | |
3010 | * ram updates during that time is pointless. */ | |
3011 | goto out; | |
3012 | } | |
3013 | ||
63268c49 PX |
3014 | /* |
3015 | * We'll take this lock a little bit long, but it's okay for two reasons. | |
3016 | * Firstly, the only possible other thread to take it is who calls | |
3017 | * qemu_guest_free_page_hint(), which should be rare; secondly, see | |
3018 | * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which | |
3019 | * guarantees that we'll at least released it in a regular basis. | |
3020 | */ | |
3021 | qemu_mutex_lock(&rs->bitmap_mutex); | |
89ac5a1d DDAG |
3022 | WITH_RCU_READ_LOCK_GUARD() { |
3023 | if (ram_list.version != rs->last_version) { | |
3024 | ram_state_reset(rs); | |
3025 | } | |
56e93d26 | 3026 | |
89ac5a1d DDAG |
3027 | /* Read version before ram_list.blocks */ |
3028 | smp_rmb(); | |
56e93d26 | 3029 | |
89ac5a1d | 3030 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 3031 | |
89ac5a1d DDAG |
3032 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
3033 | i = 0; | |
3034 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
3035 | !QSIMPLEQ_EMPTY(&rs->src_page_requests)) { | |
3036 | int pages; | |
e03a34f8 | 3037 | |
89ac5a1d DDAG |
3038 | if (qemu_file_get_error(f)) { |
3039 | break; | |
3040 | } | |
e8f3735f | 3041 | |
89ac5a1d DDAG |
3042 | pages = ram_find_and_save_block(rs, false); |
3043 | /* no more pages to sent */ | |
3044 | if (pages == 0) { | |
3045 | done = 1; | |
3046 | break; | |
3047 | } | |
e8f3735f | 3048 | |
89ac5a1d DDAG |
3049 | if (pages < 0) { |
3050 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
3051 | break; |
3052 | } | |
89ac5a1d DDAG |
3053 | |
3054 | rs->target_page_count += pages; | |
3055 | ||
644acf99 WY |
3056 | /* |
3057 | * During postcopy, it is necessary to make sure one whole host | |
3058 | * page is sent in one chunk. | |
3059 | */ | |
3060 | if (migrate_postcopy_ram()) { | |
3061 | flush_compressed_data(rs); | |
3062 | } | |
3063 | ||
89ac5a1d DDAG |
3064 | /* |
3065 | * we want to check in the 1st loop, just in case it was the 1st | |
3066 | * time and we had to sync the dirty bitmap. | |
3067 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
3068 | * some iterations | |
3069 | */ | |
3070 | if ((i & 63) == 0) { | |
3071 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
3072 | 1000000; | |
3073 | if (t1 > MAX_WAIT) { | |
3074 | trace_ram_save_iterate_big_wait(t1, i); | |
3075 | break; | |
3076 | } | |
3077 | } | |
3078 | i++; | |
56e93d26 | 3079 | } |
56e93d26 | 3080 | } |
63268c49 | 3081 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 JQ |
3082 | |
3083 | /* | |
3084 | * Must occur before EOS (or any QEMUFile operation) | |
3085 | * because of RDMA protocol. | |
3086 | */ | |
3087 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
3088 | ||
b2557345 | 3089 | out: |
b69a0227 JQ |
3090 | if (ret >= 0 |
3091 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
99f2c6fb | 3092 | multifd_send_sync_main(rs->f); |
3d4095b2 JQ |
3093 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3094 | qemu_fflush(f); | |
3095 | ram_counters.transferred += 8; | |
56e93d26 | 3096 | |
3d4095b2 JQ |
3097 | ret = qemu_file_get_error(f); |
3098 | } | |
56e93d26 JQ |
3099 | if (ret < 0) { |
3100 | return ret; | |
3101 | } | |
3102 | ||
5c90308f | 3103 | return done; |
56e93d26 JQ |
3104 | } |
3105 | ||
3d0684b2 JQ |
3106 | /** |
3107 | * ram_save_complete: function called to send the remaining amount of ram | |
3108 | * | |
e8f3735f | 3109 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
3110 | * |
3111 | * Called with iothread lock | |
3112 | * | |
3113 | * @f: QEMUFile where to send the data | |
3114 | * @opaque: RAMState pointer | |
3115 | */ | |
56e93d26 JQ |
3116 | static int ram_save_complete(QEMUFile *f, void *opaque) |
3117 | { | |
53518d94 JQ |
3118 | RAMState **temp = opaque; |
3119 | RAMState *rs = *temp; | |
e8f3735f | 3120 | int ret = 0; |
6f37bb8b | 3121 | |
89ac5a1d DDAG |
3122 | WITH_RCU_READ_LOCK_GUARD() { |
3123 | if (!migration_in_postcopy()) { | |
3124 | migration_bitmap_sync_precopy(rs); | |
3125 | } | |
56e93d26 | 3126 | |
89ac5a1d | 3127 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 3128 | |
89ac5a1d | 3129 | /* try transferring iterative blocks of memory */ |
56e93d26 | 3130 | |
89ac5a1d DDAG |
3131 | /* flush all remaining blocks regardless of rate limiting */ |
3132 | while (true) { | |
3133 | int pages; | |
56e93d26 | 3134 | |
89ac5a1d DDAG |
3135 | pages = ram_find_and_save_block(rs, !migration_in_colo_state()); |
3136 | /* no more blocks to sent */ | |
3137 | if (pages == 0) { | |
3138 | break; | |
3139 | } | |
3140 | if (pages < 0) { | |
3141 | ret = pages; | |
3142 | break; | |
3143 | } | |
e8f3735f | 3144 | } |
56e93d26 | 3145 | |
89ac5a1d DDAG |
3146 | flush_compressed_data(rs); |
3147 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); | |
3148 | } | |
d09a6fde | 3149 | |
3d4095b2 | 3150 | if (ret >= 0) { |
99f2c6fb | 3151 | multifd_send_sync_main(rs->f); |
3d4095b2 JQ |
3152 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3153 | qemu_fflush(f); | |
3154 | } | |
56e93d26 | 3155 | |
e8f3735f | 3156 | return ret; |
56e93d26 JQ |
3157 | } |
3158 | ||
c31b098f | 3159 | static void ram_save_pending(QEMUFile *f, void *opaque, uint64_t max_size, |
47995026 VSO |
3160 | uint64_t *res_precopy_only, |
3161 | uint64_t *res_compatible, | |
3162 | uint64_t *res_postcopy_only) | |
56e93d26 | 3163 | { |
53518d94 JQ |
3164 | RAMState **temp = opaque; |
3165 | RAMState *rs = *temp; | |
56e93d26 JQ |
3166 | uint64_t remaining_size; |
3167 | ||
9edabd4d | 3168 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3169 | |
5727309d | 3170 | if (!migration_in_postcopy() && |
663e6c1d | 3171 | remaining_size < max_size) { |
56e93d26 | 3172 | qemu_mutex_lock_iothread(); |
89ac5a1d DDAG |
3173 | WITH_RCU_READ_LOCK_GUARD() { |
3174 | migration_bitmap_sync_precopy(rs); | |
3175 | } | |
56e93d26 | 3176 | qemu_mutex_unlock_iothread(); |
9edabd4d | 3177 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3178 | } |
c31b098f | 3179 | |
86e1167e VSO |
3180 | if (migrate_postcopy_ram()) { |
3181 | /* We can do postcopy, and all the data is postcopiable */ | |
47995026 | 3182 | *res_compatible += remaining_size; |
86e1167e | 3183 | } else { |
47995026 | 3184 | *res_precopy_only += remaining_size; |
86e1167e | 3185 | } |
56e93d26 JQ |
3186 | } |
3187 | ||
3188 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
3189 | { | |
3190 | unsigned int xh_len; | |
3191 | int xh_flags; | |
063e760a | 3192 | uint8_t *loaded_data; |
56e93d26 | 3193 | |
56e93d26 JQ |
3194 | /* extract RLE header */ |
3195 | xh_flags = qemu_get_byte(f); | |
3196 | xh_len = qemu_get_be16(f); | |
3197 | ||
3198 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
3199 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
3200 | return -1; | |
3201 | } | |
3202 | ||
3203 | if (xh_len > TARGET_PAGE_SIZE) { | |
3204 | error_report("Failed to load XBZRLE page - len overflow!"); | |
3205 | return -1; | |
3206 | } | |
f265e0e4 | 3207 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 3208 | /* load data and decode */ |
f265e0e4 | 3209 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 3210 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
3211 | |
3212 | /* decode RLE */ | |
063e760a | 3213 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
3214 | TARGET_PAGE_SIZE) == -1) { |
3215 | error_report("Failed to load XBZRLE page - decode error!"); | |
3216 | return -1; | |
3217 | } | |
3218 | ||
3219 | return 0; | |
3220 | } | |
3221 | ||
3d0684b2 JQ |
3222 | /** |
3223 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
3224 | * | |
3225 | * Must be called from within a rcu critical section. | |
3226 | * | |
56e93d26 | 3227 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 3228 | * |
3d0684b2 JQ |
3229 | * @f: QEMUFile where to read the data from |
3230 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
a7180877 | 3231 | */ |
3d0684b2 | 3232 | static inline RAMBlock *ram_block_from_stream(QEMUFile *f, int flags) |
56e93d26 | 3233 | { |
49324e93 | 3234 | static RAMBlock *block; |
56e93d26 JQ |
3235 | char id[256]; |
3236 | uint8_t len; | |
3237 | ||
3238 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 3239 | if (!block) { |
56e93d26 JQ |
3240 | error_report("Ack, bad migration stream!"); |
3241 | return NULL; | |
3242 | } | |
4c4bad48 | 3243 | return block; |
56e93d26 JQ |
3244 | } |
3245 | ||
3246 | len = qemu_get_byte(f); | |
3247 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3248 | id[len] = 0; | |
3249 | ||
e3dd7493 | 3250 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
3251 | if (!block) { |
3252 | error_report("Can't find block %s", id); | |
3253 | return NULL; | |
56e93d26 JQ |
3254 | } |
3255 | ||
fbd162e6 | 3256 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
3257 | error_report("block %s should not be migrated !", id); |
3258 | return NULL; | |
3259 | } | |
3260 | ||
4c4bad48 HZ |
3261 | return block; |
3262 | } | |
3263 | ||
3264 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
3265 | ram_addr_t offset) | |
3266 | { | |
3267 | if (!offset_in_ramblock(block, offset)) { | |
3268 | return NULL; | |
3269 | } | |
3270 | ||
3271 | return block->host + offset; | |
56e93d26 JQ |
3272 | } |
3273 | ||
6a23f639 DH |
3274 | static void *host_page_from_ram_block_offset(RAMBlock *block, |
3275 | ram_addr_t offset) | |
3276 | { | |
3277 | /* Note: Explicitly no check against offset_in_ramblock(). */ | |
3278 | return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block->host + offset), | |
3279 | block->page_size); | |
3280 | } | |
3281 | ||
3282 | static ram_addr_t host_page_offset_from_ram_block_offset(RAMBlock *block, | |
3283 | ram_addr_t offset) | |
3284 | { | |
3285 | return ((uintptr_t)block->host + offset) & (block->page_size - 1); | |
3286 | } | |
3287 | ||
13af18f2 | 3288 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 3289 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
3290 | { |
3291 | if (!offset_in_ramblock(block, offset)) { | |
3292 | return NULL; | |
3293 | } | |
3294 | if (!block->colo_cache) { | |
3295 | error_report("%s: colo_cache is NULL in block :%s", | |
3296 | __func__, block->idstr); | |
3297 | return NULL; | |
3298 | } | |
7d9acafa ZC |
3299 | |
3300 | /* | |
3301 | * During colo checkpoint, we need bitmap of these migrated pages. | |
3302 | * It help us to decide which pages in ram cache should be flushed | |
3303 | * into VM's RAM later. | |
3304 | */ | |
8af66371 HZ |
3305 | if (record_bitmap && |
3306 | !test_and_set_bit(offset >> TARGET_PAGE_BITS, block->bmap)) { | |
7d9acafa ZC |
3307 | ram_state->migration_dirty_pages++; |
3308 | } | |
13af18f2 ZC |
3309 | return block->colo_cache + offset; |
3310 | } | |
3311 | ||
3d0684b2 JQ |
3312 | /** |
3313 | * ram_handle_compressed: handle the zero page case | |
3314 | * | |
56e93d26 JQ |
3315 | * If a page (or a whole RDMA chunk) has been |
3316 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
3317 | * |
3318 | * @host: host address for the zero page | |
3319 | * @ch: what the page is filled from. We only support zero | |
3320 | * @size: size of the zero page | |
56e93d26 JQ |
3321 | */ |
3322 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
3323 | { | |
3324 | if (ch != 0 || !is_zero_range(host, size)) { | |
3325 | memset(host, ch, size); | |
3326 | } | |
3327 | } | |
3328 | ||
797ca154 XG |
3329 | /* return the size after decompression, or negative value on error */ |
3330 | static int | |
3331 | qemu_uncompress_data(z_stream *stream, uint8_t *dest, size_t dest_len, | |
3332 | const uint8_t *source, size_t source_len) | |
3333 | { | |
3334 | int err; | |
3335 | ||
3336 | err = inflateReset(stream); | |
3337 | if (err != Z_OK) { | |
3338 | return -1; | |
3339 | } | |
3340 | ||
3341 | stream->avail_in = source_len; | |
3342 | stream->next_in = (uint8_t *)source; | |
3343 | stream->avail_out = dest_len; | |
3344 | stream->next_out = dest; | |
3345 | ||
3346 | err = inflate(stream, Z_NO_FLUSH); | |
3347 | if (err != Z_STREAM_END) { | |
3348 | return -1; | |
3349 | } | |
3350 | ||
3351 | return stream->total_out; | |
3352 | } | |
3353 | ||
56e93d26 JQ |
3354 | static void *do_data_decompress(void *opaque) |
3355 | { | |
3356 | DecompressParam *param = opaque; | |
3357 | unsigned long pagesize; | |
33d151f4 | 3358 | uint8_t *des; |
34ab9e97 | 3359 | int len, ret; |
56e93d26 | 3360 | |
33d151f4 | 3361 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 3362 | while (!param->quit) { |
33d151f4 LL |
3363 | if (param->des) { |
3364 | des = param->des; | |
3365 | len = param->len; | |
3366 | param->des = 0; | |
3367 | qemu_mutex_unlock(¶m->mutex); | |
3368 | ||
56e93d26 | 3369 | pagesize = TARGET_PAGE_SIZE; |
34ab9e97 XG |
3370 | |
3371 | ret = qemu_uncompress_data(¶m->stream, des, pagesize, | |
3372 | param->compbuf, len); | |
f548222c | 3373 | if (ret < 0 && migrate_get_current()->decompress_error_check) { |
34ab9e97 XG |
3374 | error_report("decompress data failed"); |
3375 | qemu_file_set_error(decomp_file, ret); | |
3376 | } | |
73a8912b | 3377 | |
33d151f4 LL |
3378 | qemu_mutex_lock(&decomp_done_lock); |
3379 | param->done = true; | |
3380 | qemu_cond_signal(&decomp_done_cond); | |
3381 | qemu_mutex_unlock(&decomp_done_lock); | |
3382 | ||
3383 | qemu_mutex_lock(¶m->mutex); | |
3384 | } else { | |
3385 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
3386 | } | |
56e93d26 | 3387 | } |
33d151f4 | 3388 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
3389 | |
3390 | return NULL; | |
3391 | } | |
3392 | ||
34ab9e97 | 3393 | static int wait_for_decompress_done(void) |
5533b2e9 LL |
3394 | { |
3395 | int idx, thread_count; | |
3396 | ||
3397 | if (!migrate_use_compression()) { | |
34ab9e97 | 3398 | return 0; |
5533b2e9 LL |
3399 | } |
3400 | ||
3401 | thread_count = migrate_decompress_threads(); | |
3402 | qemu_mutex_lock(&decomp_done_lock); | |
3403 | for (idx = 0; idx < thread_count; idx++) { | |
3404 | while (!decomp_param[idx].done) { | |
3405 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
3406 | } | |
3407 | } | |
3408 | qemu_mutex_unlock(&decomp_done_lock); | |
34ab9e97 | 3409 | return qemu_file_get_error(decomp_file); |
5533b2e9 LL |
3410 | } |
3411 | ||
f0afa331 | 3412 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
3413 | { |
3414 | int i, thread_count; | |
3415 | ||
3416ab5b JQ |
3416 | if (!migrate_use_compression()) { |
3417 | return; | |
3418 | } | |
56e93d26 JQ |
3419 | thread_count = migrate_decompress_threads(); |
3420 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3421 | /* |
3422 | * we use it as a indicator which shows if the thread is | |
3423 | * properly init'd or not | |
3424 | */ | |
3425 | if (!decomp_param[i].compbuf) { | |
3426 | break; | |
3427 | } | |
3428 | ||
56e93d26 | 3429 | qemu_mutex_lock(&decomp_param[i].mutex); |
90e56fb4 | 3430 | decomp_param[i].quit = true; |
56e93d26 JQ |
3431 | qemu_cond_signal(&decomp_param[i].cond); |
3432 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
3433 | } | |
3434 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3435 | if (!decomp_param[i].compbuf) { |
3436 | break; | |
3437 | } | |
3438 | ||
56e93d26 JQ |
3439 | qemu_thread_join(decompress_threads + i); |
3440 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
3441 | qemu_cond_destroy(&decomp_param[i].cond); | |
797ca154 | 3442 | inflateEnd(&decomp_param[i].stream); |
56e93d26 | 3443 | g_free(decomp_param[i].compbuf); |
797ca154 | 3444 | decomp_param[i].compbuf = NULL; |
56e93d26 JQ |
3445 | } |
3446 | g_free(decompress_threads); | |
3447 | g_free(decomp_param); | |
56e93d26 JQ |
3448 | decompress_threads = NULL; |
3449 | decomp_param = NULL; | |
34ab9e97 | 3450 | decomp_file = NULL; |
56e93d26 JQ |
3451 | } |
3452 | ||
34ab9e97 | 3453 | static int compress_threads_load_setup(QEMUFile *f) |
797ca154 XG |
3454 | { |
3455 | int i, thread_count; | |
3456 | ||
3457 | if (!migrate_use_compression()) { | |
3458 | return 0; | |
3459 | } | |
3460 | ||
3461 | thread_count = migrate_decompress_threads(); | |
3462 | decompress_threads = g_new0(QemuThread, thread_count); | |
3463 | decomp_param = g_new0(DecompressParam, thread_count); | |
3464 | qemu_mutex_init(&decomp_done_lock); | |
3465 | qemu_cond_init(&decomp_done_cond); | |
34ab9e97 | 3466 | decomp_file = f; |
797ca154 XG |
3467 | for (i = 0; i < thread_count; i++) { |
3468 | if (inflateInit(&decomp_param[i].stream) != Z_OK) { | |
3469 | goto exit; | |
3470 | } | |
3471 | ||
3472 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
3473 | qemu_mutex_init(&decomp_param[i].mutex); | |
3474 | qemu_cond_init(&decomp_param[i].cond); | |
3475 | decomp_param[i].done = true; | |
3476 | decomp_param[i].quit = false; | |
3477 | qemu_thread_create(decompress_threads + i, "decompress", | |
3478 | do_data_decompress, decomp_param + i, | |
3479 | QEMU_THREAD_JOINABLE); | |
3480 | } | |
3481 | return 0; | |
3482 | exit: | |
3483 | compress_threads_load_cleanup(); | |
3484 | return -1; | |
3485 | } | |
3486 | ||
c1bc6626 | 3487 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
3488 | void *host, int len) |
3489 | { | |
3490 | int idx, thread_count; | |
3491 | ||
3492 | thread_count = migrate_decompress_threads(); | |
37396950 | 3493 | QEMU_LOCK_GUARD(&decomp_done_lock); |
56e93d26 JQ |
3494 | while (true) { |
3495 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 3496 | if (decomp_param[idx].done) { |
33d151f4 LL |
3497 | decomp_param[idx].done = false; |
3498 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 3499 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
3500 | decomp_param[idx].des = host; |
3501 | decomp_param[idx].len = len; | |
33d151f4 LL |
3502 | qemu_cond_signal(&decomp_param[idx].cond); |
3503 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
3504 | break; |
3505 | } | |
3506 | } | |
3507 | if (idx < thread_count) { | |
3508 | break; | |
73a8912b LL |
3509 | } else { |
3510 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
3511 | } |
3512 | } | |
3513 | } | |
3514 | ||
b70cb3b4 RL |
3515 | static void colo_init_ram_state(void) |
3516 | { | |
3517 | ram_state_init(&ram_state); | |
b70cb3b4 RL |
3518 | } |
3519 | ||
13af18f2 ZC |
3520 | /* |
3521 | * colo cache: this is for secondary VM, we cache the whole | |
3522 | * memory of the secondary VM, it is need to hold the global lock | |
3523 | * to call this helper. | |
3524 | */ | |
3525 | int colo_init_ram_cache(void) | |
3526 | { | |
3527 | RAMBlock *block; | |
3528 | ||
44901b5a PB |
3529 | WITH_RCU_READ_LOCK_GUARD() { |
3530 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3531 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
8dbe22c6 | 3532 | NULL, false, false); |
44901b5a PB |
3533 | if (!block->colo_cache) { |
3534 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
3535 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
3536 | block->used_length); | |
3537 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3538 | if (block->colo_cache) { | |
3539 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3540 | block->colo_cache = NULL; | |
3541 | } | |
89ac5a1d | 3542 | } |
44901b5a | 3543 | return -errno; |
89ac5a1d | 3544 | } |
13af18f2 | 3545 | } |
13af18f2 | 3546 | } |
44901b5a | 3547 | |
7d9acafa ZC |
3548 | /* |
3549 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3550 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3551 | * we use the same name 'ram_bitmap' as for migration. | |
3552 | */ | |
3553 | if (ram_bytes_total()) { | |
3554 | RAMBlock *block; | |
3555 | ||
fbd162e6 | 3556 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3557 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3558 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3559 | } |
3560 | } | |
7d9acafa | 3561 | |
b70cb3b4 | 3562 | colo_init_ram_state(); |
13af18f2 | 3563 | return 0; |
13af18f2 ZC |
3564 | } |
3565 | ||
0393031a HZ |
3566 | /* TODO: duplicated with ram_init_bitmaps */ |
3567 | void colo_incoming_start_dirty_log(void) | |
3568 | { | |
3569 | RAMBlock *block = NULL; | |
3570 | /* For memory_global_dirty_log_start below. */ | |
3571 | qemu_mutex_lock_iothread(); | |
3572 | qemu_mutex_lock_ramlist(); | |
3573 | ||
3574 | memory_global_dirty_log_sync(); | |
3575 | WITH_RCU_READ_LOCK_GUARD() { | |
3576 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3577 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3578 | /* Discard this dirty bitmap record */ | |
3579 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3580 | } | |
63b41db4 | 3581 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
0393031a HZ |
3582 | } |
3583 | ram_state->migration_dirty_pages = 0; | |
3584 | qemu_mutex_unlock_ramlist(); | |
3585 | qemu_mutex_unlock_iothread(); | |
3586 | } | |
3587 | ||
13af18f2 ZC |
3588 | /* It is need to hold the global lock to call this helper */ |
3589 | void colo_release_ram_cache(void) | |
3590 | { | |
3591 | RAMBlock *block; | |
3592 | ||
63b41db4 | 3593 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); |
fbd162e6 | 3594 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3595 | g_free(block->bmap); |
3596 | block->bmap = NULL; | |
3597 | } | |
3598 | ||
89ac5a1d DDAG |
3599 | WITH_RCU_READ_LOCK_GUARD() { |
3600 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3601 | if (block->colo_cache) { | |
3602 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3603 | block->colo_cache = NULL; | |
3604 | } | |
13af18f2 ZC |
3605 | } |
3606 | } | |
0393031a | 3607 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
3608 | } |
3609 | ||
f265e0e4 JQ |
3610 | /** |
3611 | * ram_load_setup: Setup RAM for migration incoming side | |
3612 | * | |
3613 | * Returns zero to indicate success and negative for error | |
3614 | * | |
3615 | * @f: QEMUFile where to receive the data | |
3616 | * @opaque: RAMState pointer | |
3617 | */ | |
3618 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
3619 | { | |
34ab9e97 | 3620 | if (compress_threads_load_setup(f)) { |
797ca154 XG |
3621 | return -1; |
3622 | } | |
3623 | ||
f265e0e4 | 3624 | xbzrle_load_setup(); |
f9494614 | 3625 | ramblock_recv_map_init(); |
13af18f2 | 3626 | |
f265e0e4 JQ |
3627 | return 0; |
3628 | } | |
3629 | ||
3630 | static int ram_load_cleanup(void *opaque) | |
3631 | { | |
f9494614 | 3632 | RAMBlock *rb; |
56eb90af | 3633 | |
fbd162e6 | 3634 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 3635 | qemu_ram_block_writeback(rb); |
56eb90af JH |
3636 | } |
3637 | ||
f265e0e4 | 3638 | xbzrle_load_cleanup(); |
f0afa331 | 3639 | compress_threads_load_cleanup(); |
f9494614 | 3640 | |
fbd162e6 | 3641 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
3642 | g_free(rb->receivedmap); |
3643 | rb->receivedmap = NULL; | |
3644 | } | |
13af18f2 | 3645 | |
f265e0e4 JQ |
3646 | return 0; |
3647 | } | |
3648 | ||
3d0684b2 JQ |
3649 | /** |
3650 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
3651 | * | |
3652 | * Returns 0 for success and negative if there was one error | |
3653 | * | |
3654 | * @mis: current migration incoming state | |
3655 | * | |
3656 | * Allocate data structures etc needed by incoming migration with | |
3657 | * postcopy-ram. postcopy-ram's similarly names | |
3658 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
3659 | */ |
3660 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
3661 | { | |
c136180c | 3662 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
3663 | } |
3664 | ||
3d0684b2 JQ |
3665 | /** |
3666 | * ram_load_postcopy: load a page in postcopy case | |
3667 | * | |
3668 | * Returns 0 for success or -errno in case of error | |
3669 | * | |
a7180877 DDAG |
3670 | * Called in postcopy mode by ram_load(). |
3671 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
3672 | * |
3673 | * @f: QEMUFile where to send the data | |
a7180877 DDAG |
3674 | */ |
3675 | static int ram_load_postcopy(QEMUFile *f) | |
3676 | { | |
3677 | int flags = 0, ret = 0; | |
3678 | bool place_needed = false; | |
1aa83678 | 3679 | bool matches_target_page_size = false; |
a7180877 DDAG |
3680 | MigrationIncomingState *mis = migration_incoming_get_current(); |
3681 | /* Temporary page that is later 'placed' */ | |
3414322a | 3682 | void *postcopy_host_page = mis->postcopy_tmp_page; |
6a23f639 | 3683 | void *host_page = NULL; |
ddf35bdf | 3684 | bool all_zero = true; |
4cbb3c63 | 3685 | int target_pages = 0; |
a7180877 DDAG |
3686 | |
3687 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
3688 | ram_addr_t addr; | |
a7180877 DDAG |
3689 | void *page_buffer = NULL; |
3690 | void *place_source = NULL; | |
df9ff5e1 | 3691 | RAMBlock *block = NULL; |
a7180877 | 3692 | uint8_t ch; |
644acf99 | 3693 | int len; |
a7180877 DDAG |
3694 | |
3695 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
3696 | |
3697 | /* | |
3698 | * If qemu file error, we should stop here, and then "addr" | |
3699 | * may be invalid | |
3700 | */ | |
3701 | ret = qemu_file_get_error(f); | |
3702 | if (ret) { | |
3703 | break; | |
3704 | } | |
3705 | ||
a7180877 DDAG |
3706 | flags = addr & ~TARGET_PAGE_MASK; |
3707 | addr &= TARGET_PAGE_MASK; | |
3708 | ||
3709 | trace_ram_load_postcopy_loop((uint64_t)addr, flags); | |
644acf99 WY |
3710 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
3711 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
df9ff5e1 | 3712 | block = ram_block_from_stream(f, flags); |
6a23f639 DH |
3713 | if (!block) { |
3714 | ret = -EINVAL; | |
3715 | break; | |
3716 | } | |
4c4bad48 | 3717 | |
898ba906 DH |
3718 | /* |
3719 | * Relying on used_length is racy and can result in false positives. | |
3720 | * We might place pages beyond used_length in case RAM was shrunk | |
3721 | * while in postcopy, which is fine - trying to place via | |
3722 | * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault. | |
3723 | */ | |
3724 | if (!block->host || addr >= block->postcopy_length) { | |
a7180877 DDAG |
3725 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); |
3726 | ret = -EINVAL; | |
3727 | break; | |
3728 | } | |
4cbb3c63 | 3729 | target_pages++; |
1aa83678 | 3730 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 3731 | /* |
28abd200 DDAG |
3732 | * Postcopy requires that we place whole host pages atomically; |
3733 | * these may be huge pages for RAMBlocks that are backed by | |
3734 | * hugetlbfs. | |
a7180877 DDAG |
3735 | * To make it atomic, the data is read into a temporary page |
3736 | * that's moved into place later. | |
3737 | * The migration protocol uses, possibly smaller, target-pages | |
3738 | * however the source ensures it always sends all the components | |
91ba442f | 3739 | * of a host page in one chunk. |
a7180877 DDAG |
3740 | */ |
3741 | page_buffer = postcopy_host_page + | |
6a23f639 DH |
3742 | host_page_offset_from_ram_block_offset(block, addr); |
3743 | /* If all TP are zero then we can optimise the place */ | |
e5e73b0f | 3744 | if (target_pages == 1) { |
6a23f639 DH |
3745 | host_page = host_page_from_ram_block_offset(block, addr); |
3746 | } else if (host_page != host_page_from_ram_block_offset(block, | |
3747 | addr)) { | |
c53b7ddc | 3748 | /* not the 1st TP within the HP */ |
6a23f639 DH |
3749 | error_report("Non-same host page %p/%p", host_page, |
3750 | host_page_from_ram_block_offset(block, addr)); | |
3751 | ret = -EINVAL; | |
3752 | break; | |
a7180877 DDAG |
3753 | } |
3754 | ||
3755 | /* | |
3756 | * If it's the last part of a host page then we place the host | |
3757 | * page | |
3758 | */ | |
4cbb3c63 WY |
3759 | if (target_pages == (block->page_size / TARGET_PAGE_SIZE)) { |
3760 | place_needed = true; | |
4cbb3c63 | 3761 | } |
a7180877 DDAG |
3762 | place_source = postcopy_host_page; |
3763 | } | |
3764 | ||
3765 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 3766 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 3767 | ch = qemu_get_byte(f); |
2e36bc1b WY |
3768 | /* |
3769 | * Can skip to set page_buffer when | |
3770 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
3771 | */ | |
3772 | if (ch || !matches_target_page_size) { | |
3773 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
3774 | } | |
a7180877 DDAG |
3775 | if (ch) { |
3776 | all_zero = false; | |
3777 | } | |
3778 | break; | |
3779 | ||
3780 | case RAM_SAVE_FLAG_PAGE: | |
3781 | all_zero = false; | |
1aa83678 PX |
3782 | if (!matches_target_page_size) { |
3783 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
3784 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
3785 | } else { | |
1aa83678 PX |
3786 | /* |
3787 | * For small pages that matches target page size, we | |
3788 | * avoid the qemu_file copy. Instead we directly use | |
3789 | * the buffer of QEMUFile to place the page. Note: we | |
3790 | * cannot do any QEMUFile operation before using that | |
3791 | * buffer to make sure the buffer is valid when | |
3792 | * placing the page. | |
a7180877 DDAG |
3793 | */ |
3794 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
3795 | TARGET_PAGE_SIZE); | |
3796 | } | |
3797 | break; | |
644acf99 WY |
3798 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
3799 | all_zero = false; | |
3800 | len = qemu_get_be32(f); | |
3801 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
3802 | error_report("Invalid compressed data length: %d", len); | |
3803 | ret = -EINVAL; | |
3804 | break; | |
3805 | } | |
3806 | decompress_data_with_multi_threads(f, page_buffer, len); | |
3807 | break; | |
3808 | ||
a7180877 DDAG |
3809 | case RAM_SAVE_FLAG_EOS: |
3810 | /* normal exit */ | |
6df264ac | 3811 | multifd_recv_sync_main(); |
a7180877 DDAG |
3812 | break; |
3813 | default: | |
29fccade | 3814 | error_report("Unknown combination of migration flags: 0x%x" |
a7180877 DDAG |
3815 | " (postcopy mode)", flags); |
3816 | ret = -EINVAL; | |
7a9ddfbf PX |
3817 | break; |
3818 | } | |
3819 | ||
644acf99 WY |
3820 | /* Got the whole host page, wait for decompress before placing. */ |
3821 | if (place_needed) { | |
3822 | ret |= wait_for_decompress_done(); | |
3823 | } | |
3824 | ||
7a9ddfbf PX |
3825 | /* Detect for any possible file errors */ |
3826 | if (!ret && qemu_file_get_error(f)) { | |
3827 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
3828 | } |
3829 | ||
7a9ddfbf | 3830 | if (!ret && place_needed) { |
a7180877 | 3831 | if (all_zero) { |
6a23f639 | 3832 | ret = postcopy_place_page_zero(mis, host_page, block); |
a7180877 | 3833 | } else { |
6a23f639 DH |
3834 | ret = postcopy_place_page(mis, host_page, place_source, |
3835 | block); | |
a7180877 | 3836 | } |
ddf35bdf DH |
3837 | place_needed = false; |
3838 | target_pages = 0; | |
3839 | /* Assume we have a zero page until we detect something different */ | |
3840 | all_zero = true; | |
a7180877 | 3841 | } |
a7180877 DDAG |
3842 | } |
3843 | ||
3844 | return ret; | |
3845 | } | |
3846 | ||
acab30b8 DHB |
3847 | static bool postcopy_is_advised(void) |
3848 | { | |
3849 | PostcopyState ps = postcopy_state_get(); | |
3850 | return ps >= POSTCOPY_INCOMING_ADVISE && ps < POSTCOPY_INCOMING_END; | |
3851 | } | |
3852 | ||
3853 | static bool postcopy_is_running(void) | |
3854 | { | |
3855 | PostcopyState ps = postcopy_state_get(); | |
3856 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
3857 | } | |
3858 | ||
e6f4aa18 ZC |
3859 | /* |
3860 | * Flush content of RAM cache into SVM's memory. | |
3861 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
3862 | */ | |
24fa16f8 | 3863 | void colo_flush_ram_cache(void) |
e6f4aa18 ZC |
3864 | { |
3865 | RAMBlock *block = NULL; | |
3866 | void *dst_host; | |
3867 | void *src_host; | |
3868 | unsigned long offset = 0; | |
3869 | ||
d1955d22 | 3870 | memory_global_dirty_log_sync(); |
63268c49 | 3871 | qemu_mutex_lock(&ram_state->bitmap_mutex); |
89ac5a1d DDAG |
3872 | WITH_RCU_READ_LOCK_GUARD() { |
3873 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3874 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3875 | } | |
d1955d22 | 3876 | } |
d1955d22 | 3877 | |
e6f4aa18 | 3878 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
3879 | WITH_RCU_READ_LOCK_GUARD() { |
3880 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 3881 | |
89ac5a1d DDAG |
3882 | while (block) { |
3883 | offset = migration_bitmap_find_dirty(ram_state, block, offset); | |
e6f4aa18 | 3884 | |
542147f4 DH |
3885 | if (!offset_in_ramblock(block, |
3886 | ((ram_addr_t)offset) << TARGET_PAGE_BITS)) { | |
89ac5a1d DDAG |
3887 | offset = 0; |
3888 | block = QLIST_NEXT_RCU(block, next); | |
3889 | } else { | |
3890 | migration_bitmap_clear_dirty(ram_state, block, offset); | |
8bba004c AR |
3891 | dst_host = block->host |
3892 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
3893 | src_host = block->colo_cache | |
3894 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
89ac5a1d DDAG |
3895 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE); |
3896 | } | |
e6f4aa18 ZC |
3897 | } |
3898 | } | |
e6f4aa18 | 3899 | trace_colo_flush_ram_cache_end(); |
63268c49 | 3900 | qemu_mutex_unlock(&ram_state->bitmap_mutex); |
e6f4aa18 ZC |
3901 | } |
3902 | ||
10da4a36 WY |
3903 | /** |
3904 | * ram_load_precopy: load pages in precopy case | |
3905 | * | |
3906 | * Returns 0 for success or -errno in case of error | |
3907 | * | |
3908 | * Called in precopy mode by ram_load(). | |
3909 | * rcu_read_lock is taken prior to this being called. | |
3910 | * | |
3911 | * @f: QEMUFile where to send the data | |
3912 | */ | |
3913 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 3914 | { |
e65cec5e | 3915 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 3916 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
acab30b8 | 3917 | bool postcopy_advised = postcopy_is_advised(); |
edc60127 JQ |
3918 | if (!migrate_use_compression()) { |
3919 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; | |
3920 | } | |
a7180877 | 3921 | |
10da4a36 | 3922 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 3923 | ram_addr_t addr, total_ram_bytes; |
0393031a | 3924 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
3925 | uint8_t ch; |
3926 | ||
e65cec5e YK |
3927 | /* |
3928 | * Yield periodically to let main loop run, but an iteration of | |
3929 | * the main loop is expensive, so do it each some iterations | |
3930 | */ | |
3931 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
3932 | aio_co_schedule(qemu_get_current_aio_context(), | |
3933 | qemu_coroutine_self()); | |
3934 | qemu_coroutine_yield(); | |
3935 | } | |
3936 | i++; | |
3937 | ||
56e93d26 JQ |
3938 | addr = qemu_get_be64(f); |
3939 | flags = addr & ~TARGET_PAGE_MASK; | |
3940 | addr &= TARGET_PAGE_MASK; | |
3941 | ||
edc60127 JQ |
3942 | if (flags & invalid_flags) { |
3943 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
3944 | error_report("Received an unexpected compressed page"); | |
3945 | } | |
3946 | ||
3947 | ret = -EINVAL; | |
3948 | break; | |
3949 | } | |
3950 | ||
bb890ed5 | 3951 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 3952 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
4c4bad48 HZ |
3953 | RAMBlock *block = ram_block_from_stream(f, flags); |
3954 | ||
0393031a | 3955 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 3956 | /* |
0393031a HZ |
3957 | * After going into COLO stage, we should not load the page |
3958 | * into SVM's memory directly, we put them into colo_cache firstly. | |
3959 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
3960 | * Previously, we copied all these memory in preparing stage of COLO | |
3961 | * while we need to stop VM, which is a time-consuming process. | |
3962 | * Here we optimize it by a trick, back-up every page while in | |
3963 | * migration process while COLO is enabled, though it affects the | |
3964 | * speed of the migration, but it obviously reduce the downtime of | |
3965 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 3966 | */ |
0393031a HZ |
3967 | if (migration_incoming_colo_enabled()) { |
3968 | if (migration_incoming_in_colo_state()) { | |
3969 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 3970 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
3971 | } else { |
3972 | /* | |
3973 | * In migration stage but before COLO stage, | |
3974 | * Put all pages into both cache and SVM's memory. | |
3975 | */ | |
8af66371 | 3976 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 3977 | } |
13af18f2 | 3978 | } |
a776aa15 DDAG |
3979 | if (!host) { |
3980 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
3981 | ret = -EINVAL; | |
3982 | break; | |
3983 | } | |
13af18f2 ZC |
3984 | if (!migration_incoming_in_colo_state()) { |
3985 | ramblock_recv_bitmap_set(block, host); | |
3986 | } | |
3987 | ||
1db9d8e5 | 3988 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
3989 | } |
3990 | ||
56e93d26 JQ |
3991 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
3992 | case RAM_SAVE_FLAG_MEM_SIZE: | |
3993 | /* Synchronize RAM block list */ | |
3994 | total_ram_bytes = addr; | |
3995 | while (!ret && total_ram_bytes) { | |
3996 | RAMBlock *block; | |
56e93d26 JQ |
3997 | char id[256]; |
3998 | ram_addr_t length; | |
3999 | ||
4000 | len = qemu_get_byte(f); | |
4001 | qemu_get_buffer(f, (uint8_t *)id, len); | |
4002 | id[len] = 0; | |
4003 | length = qemu_get_be64(f); | |
4004 | ||
e3dd7493 | 4005 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
4006 | if (block && !qemu_ram_is_migratable(block)) { |
4007 | error_report("block %s should not be migrated !", id); | |
4008 | ret = -EINVAL; | |
4009 | } else if (block) { | |
e3dd7493 DDAG |
4010 | if (length != block->used_length) { |
4011 | Error *local_err = NULL; | |
56e93d26 | 4012 | |
fa53a0e5 | 4013 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
4014 | &local_err); |
4015 | if (local_err) { | |
4016 | error_report_err(local_err); | |
56e93d26 | 4017 | } |
56e93d26 | 4018 | } |
ef08fb38 | 4019 | /* For postcopy we need to check hugepage sizes match */ |
e846b746 | 4020 | if (postcopy_advised && migrate_postcopy_ram() && |
ef08fb38 DDAG |
4021 | block->page_size != qemu_host_page_size) { |
4022 | uint64_t remote_page_size = qemu_get_be64(f); | |
4023 | if (remote_page_size != block->page_size) { | |
4024 | error_report("Mismatched RAM page size %s " | |
4025 | "(local) %zd != %" PRId64, | |
4026 | id, block->page_size, | |
4027 | remote_page_size); | |
4028 | ret = -EINVAL; | |
4029 | } | |
4030 | } | |
fbd162e6 YK |
4031 | if (migrate_ignore_shared()) { |
4032 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
4033 | if (ramblock_is_ignored(block) && |
4034 | block->mr->addr != addr) { | |
4035 | error_report("Mismatched GPAs for block %s " | |
4036 | "%" PRId64 "!= %" PRId64, | |
4037 | id, (uint64_t)addr, | |
4038 | (uint64_t)block->mr->addr); | |
4039 | ret = -EINVAL; | |
4040 | } | |
4041 | } | |
e3dd7493 DDAG |
4042 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
4043 | block->idstr); | |
4044 | } else { | |
56e93d26 JQ |
4045 | error_report("Unknown ramblock \"%s\", cannot " |
4046 | "accept migration", id); | |
4047 | ret = -EINVAL; | |
4048 | } | |
4049 | ||
4050 | total_ram_bytes -= length; | |
4051 | } | |
4052 | break; | |
a776aa15 | 4053 | |
bb890ed5 | 4054 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
4055 | ch = qemu_get_byte(f); |
4056 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
4057 | break; | |
a776aa15 | 4058 | |
56e93d26 | 4059 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
4060 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
4061 | break; | |
56e93d26 | 4062 | |
a776aa15 | 4063 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
4064 | len = qemu_get_be32(f); |
4065 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4066 | error_report("Invalid compressed data length: %d", len); | |
4067 | ret = -EINVAL; | |
4068 | break; | |
4069 | } | |
c1bc6626 | 4070 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 4071 | break; |
a776aa15 | 4072 | |
56e93d26 | 4073 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
4074 | if (load_xbzrle(f, addr, host) < 0) { |
4075 | error_report("Failed to decompress XBZRLE page at " | |
4076 | RAM_ADDR_FMT, addr); | |
4077 | ret = -EINVAL; | |
4078 | break; | |
4079 | } | |
4080 | break; | |
4081 | case RAM_SAVE_FLAG_EOS: | |
4082 | /* normal exit */ | |
6df264ac | 4083 | multifd_recv_sync_main(); |
56e93d26 JQ |
4084 | break; |
4085 | default: | |
4086 | if (flags & RAM_SAVE_FLAG_HOOK) { | |
632e3a5c | 4087 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); |
56e93d26 | 4088 | } else { |
29fccade | 4089 | error_report("Unknown combination of migration flags: 0x%x", |
56e93d26 JQ |
4090 | flags); |
4091 | ret = -EINVAL; | |
4092 | } | |
4093 | } | |
4094 | if (!ret) { | |
4095 | ret = qemu_file_get_error(f); | |
4096 | } | |
0393031a HZ |
4097 | if (!ret && host_bak) { |
4098 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
4099 | } | |
56e93d26 JQ |
4100 | } |
4101 | ||
ca1a6b70 | 4102 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
4103 | return ret; |
4104 | } | |
4105 | ||
4106 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
4107 | { | |
4108 | int ret = 0; | |
4109 | static uint64_t seq_iter; | |
4110 | /* | |
4111 | * If system is running in postcopy mode, page inserts to host memory must | |
4112 | * be atomic | |
4113 | */ | |
4114 | bool postcopy_running = postcopy_is_running(); | |
4115 | ||
4116 | seq_iter++; | |
4117 | ||
4118 | if (version_id != 4) { | |
4119 | return -EINVAL; | |
4120 | } | |
4121 | ||
4122 | /* | |
4123 | * This RCU critical section can be very long running. | |
4124 | * When RCU reclaims in the code start to become numerous, | |
4125 | * it will be necessary to reduce the granularity of this | |
4126 | * critical section. | |
4127 | */ | |
89ac5a1d DDAG |
4128 | WITH_RCU_READ_LOCK_GUARD() { |
4129 | if (postcopy_running) { | |
4130 | ret = ram_load_postcopy(f); | |
4131 | } else { | |
4132 | ret = ram_load_precopy(f); | |
4133 | } | |
10da4a36 | 4134 | } |
55c4446b | 4135 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 | 4136 | |
56e93d26 JQ |
4137 | return ret; |
4138 | } | |
4139 | ||
c6467627 VSO |
4140 | static bool ram_has_postcopy(void *opaque) |
4141 | { | |
469dd51b | 4142 | RAMBlock *rb; |
fbd162e6 | 4143 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
4144 | if (ramblock_is_pmem(rb)) { |
4145 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
4146 | "is not supported now!", rb->idstr, rb->host); | |
4147 | return false; | |
4148 | } | |
4149 | } | |
4150 | ||
c6467627 VSO |
4151 | return migrate_postcopy_ram(); |
4152 | } | |
4153 | ||
edd090c7 PX |
4154 | /* Sync all the dirty bitmap with destination VM. */ |
4155 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
4156 | { | |
4157 | RAMBlock *block; | |
4158 | QEMUFile *file = s->to_dst_file; | |
4159 | int ramblock_count = 0; | |
4160 | ||
4161 | trace_ram_dirty_bitmap_sync_start(); | |
4162 | ||
fbd162e6 | 4163 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
4164 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
4165 | trace_ram_dirty_bitmap_request(block->idstr); | |
4166 | ramblock_count++; | |
4167 | } | |
4168 | ||
4169 | trace_ram_dirty_bitmap_sync_wait(); | |
4170 | ||
4171 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
4172 | while (ramblock_count--) { | |
4173 | qemu_sem_wait(&s->rp_state.rp_sem); | |
4174 | } | |
4175 | ||
4176 | trace_ram_dirty_bitmap_sync_complete(); | |
4177 | ||
4178 | return 0; | |
4179 | } | |
4180 | ||
4181 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
4182 | { | |
4183 | qemu_sem_post(&s->rp_state.rp_sem); | |
4184 | } | |
4185 | ||
a335debb PX |
4186 | /* |
4187 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
4188 | * This is only used when the postcopy migration is paused but wants | |
4189 | * to resume from a middle point. | |
4190 | */ | |
4191 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
4192 | { | |
4193 | int ret = -EINVAL; | |
43044ac0 | 4194 | /* from_dst_file is always valid because we're within rp_thread */ |
a335debb PX |
4195 | QEMUFile *file = s->rp_state.from_dst_file; |
4196 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 4197 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
4198 | uint64_t size, end_mark; |
4199 | ||
4200 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
4201 | ||
4202 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
4203 | error_report("%s: incorrect state %s", __func__, | |
4204 | MigrationStatus_str(s->state)); | |
4205 | return -EINVAL; | |
4206 | } | |
4207 | ||
4208 | /* | |
4209 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3a4452d8 | 4210 | * need the endianness conversion, and the paddings. |
a335debb PX |
4211 | */ |
4212 | local_size = ROUND_UP(local_size, 8); | |
4213 | ||
4214 | /* Add paddings */ | |
4215 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
4216 | ||
4217 | size = qemu_get_be64(file); | |
4218 | ||
4219 | /* The size of the bitmap should match with our ramblock */ | |
4220 | if (size != local_size) { | |
4221 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
4222 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
4223 | block->idstr, size, local_size); | |
4224 | ret = -EINVAL; | |
4225 | goto out; | |
4226 | } | |
4227 | ||
4228 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
4229 | end_mark = qemu_get_be64(file); | |
4230 | ||
4231 | ret = qemu_file_get_error(file); | |
4232 | if (ret || size != local_size) { | |
4233 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
4234 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
4235 | __func__, block->idstr, ret, local_size, size); | |
4236 | ret = -EIO; | |
4237 | goto out; | |
4238 | } | |
4239 | ||
4240 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
af3bbbe9 | 4241 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64, |
a335debb PX |
4242 | __func__, block->idstr, end_mark); |
4243 | ret = -EINVAL; | |
4244 | goto out; | |
4245 | } | |
4246 | ||
4247 | /* | |
3a4452d8 | 4248 | * Endianness conversion. We are during postcopy (though paused). |
a335debb PX |
4249 | * The dirty bitmap won't change. We can directly modify it. |
4250 | */ | |
4251 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
4252 | ||
4253 | /* | |
4254 | * What we received is "received bitmap". Revert it as the initial | |
4255 | * dirty bitmap for this ramblock. | |
4256 | */ | |
4257 | bitmap_complement(block->bmap, block->bmap, nbits); | |
4258 | ||
be39b4cd DH |
4259 | /* Clear dirty bits of discarded ranges that we don't want to migrate. */ |
4260 | ramblock_dirty_bitmap_clear_discarded_pages(block); | |
4261 | ||
4262 | /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */ | |
a335debb PX |
4263 | trace_ram_dirty_bitmap_reload_complete(block->idstr); |
4264 | ||
edd090c7 PX |
4265 | /* |
4266 | * We succeeded to sync bitmap for current ramblock. If this is | |
4267 | * the last one to sync, we need to notify the main send thread. | |
4268 | */ | |
4269 | ram_dirty_bitmap_reload_notify(s); | |
4270 | ||
a335debb PX |
4271 | ret = 0; |
4272 | out: | |
bf269906 | 4273 | g_free(le_bitmap); |
a335debb PX |
4274 | return ret; |
4275 | } | |
4276 | ||
edd090c7 PX |
4277 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
4278 | { | |
4279 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 4280 | int ret; |
edd090c7 | 4281 | |
08614f34 PX |
4282 | ret = ram_dirty_bitmap_sync_all(s, rs); |
4283 | if (ret) { | |
4284 | return ret; | |
4285 | } | |
4286 | ||
4287 | ram_state_resume_prepare(rs, s->to_dst_file); | |
4288 | ||
4289 | return 0; | |
edd090c7 PX |
4290 | } |
4291 | ||
56e93d26 | 4292 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 4293 | .save_setup = ram_save_setup, |
56e93d26 | 4294 | .save_live_iterate = ram_save_iterate, |
763c906b | 4295 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 4296 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 4297 | .has_postcopy = ram_has_postcopy, |
56e93d26 JQ |
4298 | .save_live_pending = ram_save_pending, |
4299 | .load_state = ram_load, | |
f265e0e4 JQ |
4300 | .save_cleanup = ram_save_cleanup, |
4301 | .load_setup = ram_load_setup, | |
4302 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 4303 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
4304 | }; |
4305 | ||
c7c0e724 DH |
4306 | static void ram_mig_ram_block_resized(RAMBlockNotifier *n, void *host, |
4307 | size_t old_size, size_t new_size) | |
4308 | { | |
cc61c703 | 4309 | PostcopyState ps = postcopy_state_get(); |
c7c0e724 DH |
4310 | ram_addr_t offset; |
4311 | RAMBlock *rb = qemu_ram_block_from_host(host, false, &offset); | |
4312 | Error *err = NULL; | |
4313 | ||
4314 | if (ramblock_is_ignored(rb)) { | |
4315 | return; | |
4316 | } | |
4317 | ||
4318 | if (!migration_is_idle()) { | |
4319 | /* | |
4320 | * Precopy code on the source cannot deal with the size of RAM blocks | |
4321 | * changing at random points in time - especially after sending the | |
4322 | * RAM block sizes in the migration stream, they must no longer change. | |
4323 | * Abort and indicate a proper reason. | |
4324 | */ | |
4325 | error_setg(&err, "RAM block '%s' resized during precopy.", rb->idstr); | |
458fecca | 4326 | migration_cancel(err); |
c7c0e724 | 4327 | error_free(err); |
c7c0e724 | 4328 | } |
cc61c703 DH |
4329 | |
4330 | switch (ps) { | |
4331 | case POSTCOPY_INCOMING_ADVISE: | |
4332 | /* | |
4333 | * Update what ram_postcopy_incoming_init()->init_range() does at the | |
4334 | * time postcopy was advised. Syncing RAM blocks with the source will | |
4335 | * result in RAM resizes. | |
4336 | */ | |
4337 | if (old_size < new_size) { | |
4338 | if (ram_discard_range(rb->idstr, old_size, new_size - old_size)) { | |
4339 | error_report("RAM block '%s' discard of resized RAM failed", | |
4340 | rb->idstr); | |
4341 | } | |
4342 | } | |
898ba906 | 4343 | rb->postcopy_length = new_size; |
cc61c703 DH |
4344 | break; |
4345 | case POSTCOPY_INCOMING_NONE: | |
4346 | case POSTCOPY_INCOMING_RUNNING: | |
4347 | case POSTCOPY_INCOMING_END: | |
4348 | /* | |
4349 | * Once our guest is running, postcopy does no longer care about | |
4350 | * resizes. When growing, the new memory was not available on the | |
4351 | * source, no handler needed. | |
4352 | */ | |
4353 | break; | |
4354 | default: | |
4355 | error_report("RAM block '%s' resized during postcopy state: %d", | |
4356 | rb->idstr, ps); | |
4357 | exit(-1); | |
4358 | } | |
c7c0e724 DH |
4359 | } |
4360 | ||
4361 | static RAMBlockNotifier ram_mig_ram_notifier = { | |
4362 | .ram_block_resized = ram_mig_ram_block_resized, | |
4363 | }; | |
4364 | ||
56e93d26 JQ |
4365 | void ram_mig_init(void) |
4366 | { | |
4367 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 4368 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
c7c0e724 | 4369 | ram_block_notifier_add(&ram_mig_ram_notifier); |
56e93d26 | 4370 | } |