]>
Commit | Line | Data |
---|---|---|
56e93d26 JQ |
1 | /* |
2 | * QEMU System Emulator | |
3 | * | |
4 | * Copyright (c) 2003-2008 Fabrice Bellard | |
76cc7b58 JQ |
5 | * Copyright (c) 2011-2015 Red Hat Inc |
6 | * | |
7 | * Authors: | |
8 | * Juan Quintela <quintela@redhat.com> | |
56e93d26 JQ |
9 | * |
10 | * Permission is hereby granted, free of charge, to any person obtaining a copy | |
11 | * of this software and associated documentation files (the "Software"), to deal | |
12 | * in the Software without restriction, including without limitation the rights | |
13 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell | |
14 | * copies of the Software, and to permit persons to whom the Software is | |
15 | * furnished to do so, subject to the following conditions: | |
16 | * | |
17 | * The above copyright notice and this permission notice shall be included in | |
18 | * all copies or substantial portions of the Software. | |
19 | * | |
20 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
21 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
22 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
23 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | |
24 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | |
25 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | |
26 | * THE SOFTWARE. | |
27 | */ | |
e688df6b | 28 | |
1393a485 | 29 | #include "qemu/osdep.h" |
f348b6d1 | 30 | #include "qemu/cutils.h" |
56e93d26 JQ |
31 | #include "qemu/bitops.h" |
32 | #include "qemu/bitmap.h" | |
b85ea5fa | 33 | #include "qemu/madvise.h" |
7205c9ec | 34 | #include "qemu/main-loop.h" |
709e3fe8 | 35 | #include "xbzrle.h" |
b5ca3368 | 36 | #include "ram-compress.h" |
7b1e1a22 | 37 | #include "ram.h" |
6666c96a | 38 | #include "migration.h" |
947701cc | 39 | #include "migration-stats.h" |
f2a8f0a6 | 40 | #include "migration/register.h" |
7b1e1a22 | 41 | #include "migration/misc.h" |
08a0aee1 | 42 | #include "qemu-file.h" |
be07b0ac | 43 | #include "postcopy-ram.h" |
53d37d36 | 44 | #include "page_cache.h" |
56e93d26 | 45 | #include "qemu/error-report.h" |
e688df6b | 46 | #include "qapi/error.h" |
ab7cbb0b | 47 | #include "qapi/qapi-types-migration.h" |
9af23989 | 48 | #include "qapi/qapi-events-migration.h" |
8acabf69 | 49 | #include "qapi/qmp/qerror.h" |
56e93d26 | 50 | #include "trace.h" |
56e93d26 | 51 | #include "exec/ram_addr.h" |
f9494614 | 52 | #include "exec/target_page.h" |
56e93d26 | 53 | #include "qemu/rcu_queue.h" |
a91246c9 | 54 | #include "migration/colo.h" |
53d37d36 | 55 | #include "block.h" |
b0c3cf94 | 56 | #include "sysemu/cpu-throttle.h" |
edd090c7 | 57 | #include "savevm.h" |
b9ee2f7d | 58 | #include "qemu/iov.h" |
d32ca5ad | 59 | #include "multifd.h" |
278e2f55 | 60 | #include "sysemu/runstate.h" |
1f0776f1 | 61 | #include "options.h" |
278e2f55 | 62 | |
e5fdf920 LS |
63 | #include "hw/boards.h" /* for machine_dump_guest_core() */ |
64 | ||
278e2f55 AG |
65 | #if defined(__linux__) |
66 | #include "qemu/userfaultfd.h" | |
67 | #endif /* defined(__linux__) */ | |
56e93d26 | 68 | |
56e93d26 JQ |
69 | /***********************************************************/ |
70 | /* ram save/restore */ | |
71 | ||
7b548761 JQ |
72 | /* |
73 | * RAM_SAVE_FLAG_ZERO used to be named RAM_SAVE_FLAG_COMPRESS, it | |
74 | * worked for pages that were filled with the same char. We switched | |
bb890ed5 | 75 | * it to only search for the zero value. And to avoid confusion with |
7b548761 | 76 | * RAM_SAVE_FLAG_COMPRESS_PAGE just rename it. |
bb890ed5 | 77 | */ |
7b548761 JQ |
78 | /* |
79 | * RAM_SAVE_FLAG_FULL was obsoleted in 2009, it can be reused now | |
80 | */ | |
81 | #define RAM_SAVE_FLAG_FULL 0x01 | |
bb890ed5 | 82 | #define RAM_SAVE_FLAG_ZERO 0x02 |
56e93d26 JQ |
83 | #define RAM_SAVE_FLAG_MEM_SIZE 0x04 |
84 | #define RAM_SAVE_FLAG_PAGE 0x08 | |
85 | #define RAM_SAVE_FLAG_EOS 0x10 | |
86 | #define RAM_SAVE_FLAG_CONTINUE 0x20 | |
87 | #define RAM_SAVE_FLAG_XBZRLE 0x40 | |
7b548761 | 88 | /* 0x80 is reserved in qemu-file.h for RAM_SAVE_FLAG_HOOK */ |
56e93d26 | 89 | #define RAM_SAVE_FLAG_COMPRESS_PAGE 0x100 |
294e5a40 | 90 | #define RAM_SAVE_FLAG_MULTIFD_FLUSH 0x200 |
7b548761 | 91 | /* We can't use any flag that is bigger than 0x200 */ |
56e93d26 | 92 | |
04ffce13 | 93 | int (*xbzrle_encode_buffer_func)(uint8_t *, uint8_t *, int, |
94 | uint8_t *, int) = xbzrle_encode_buffer; | |
95 | #if defined(CONFIG_AVX512BW_OPT) | |
96 | #include "qemu/cpuid.h" | |
97 | static void __attribute__((constructor)) init_cpu_flag(void) | |
98 | { | |
99 | unsigned max = __get_cpuid_max(0, NULL); | |
100 | int a, b, c, d; | |
101 | if (max >= 1) { | |
102 | __cpuid(1, a, b, c, d); | |
103 | /* We must check that AVX is not just available, but usable. */ | |
104 | if ((c & bit_OSXSAVE) && (c & bit_AVX) && max >= 7) { | |
105 | int bv; | |
106 | __asm("xgetbv" : "=a"(bv), "=d"(d) : "c"(0)); | |
107 | __cpuid_count(7, 0, a, b, c, d); | |
108 | /* 0xe6: | |
109 | * XCR0[7:5] = 111b (OPMASK state, upper 256-bit of ZMM0-ZMM15 | |
110 | * and ZMM16-ZMM31 state are enabled by OS) | |
111 | * XCR0[2:1] = 11b (XMM state and YMM state are enabled by OS) | |
112 | */ | |
113 | if ((bv & 0xe6) == 0xe6 && (b & bit_AVX512BW)) { | |
114 | xbzrle_encode_buffer_func = xbzrle_encode_buffer_avx512; | |
115 | } | |
116 | } | |
117 | } | |
118 | } | |
119 | #endif | |
120 | ||
9360447d JQ |
121 | XBZRLECacheStats xbzrle_counters; |
122 | ||
f1668764 PX |
123 | /* used by the search for pages to send */ |
124 | struct PageSearchStatus { | |
125 | /* The migration channel used for a specific host page */ | |
126 | QEMUFile *pss_channel; | |
ec6f3ab9 PX |
127 | /* Last block from where we have sent data */ |
128 | RAMBlock *last_sent_block; | |
f1668764 PX |
129 | /* Current block being searched */ |
130 | RAMBlock *block; | |
131 | /* Current page to search from */ | |
132 | unsigned long page; | |
133 | /* Set once we wrap around */ | |
134 | bool complete_round; | |
f1668764 PX |
135 | /* Whether we're sending a host page */ |
136 | bool host_page_sending; | |
137 | /* The start/end of current host page. Invalid if host_page_sending==false */ | |
138 | unsigned long host_page_start; | |
139 | unsigned long host_page_end; | |
140 | }; | |
141 | typedef struct PageSearchStatus PageSearchStatus; | |
142 | ||
56e93d26 JQ |
143 | /* struct contains XBZRLE cache and a static page |
144 | used by the compression */ | |
145 | static struct { | |
146 | /* buffer used for XBZRLE encoding */ | |
147 | uint8_t *encoded_buf; | |
148 | /* buffer for storing page content */ | |
149 | uint8_t *current_buf; | |
150 | /* Cache for XBZRLE, Protected by lock. */ | |
151 | PageCache *cache; | |
152 | QemuMutex lock; | |
c00e0928 JQ |
153 | /* it will store a page full of zeros */ |
154 | uint8_t *zero_target_page; | |
f265e0e4 JQ |
155 | /* buffer used for XBZRLE decoding */ |
156 | uint8_t *decoded_buf; | |
56e93d26 JQ |
157 | } XBZRLE; |
158 | ||
56e93d26 JQ |
159 | static void XBZRLE_cache_lock(void) |
160 | { | |
87dca0c9 | 161 | if (migrate_xbzrle()) { |
56e93d26 | 162 | qemu_mutex_lock(&XBZRLE.lock); |
f4c51a6b | 163 | } |
56e93d26 JQ |
164 | } |
165 | ||
166 | static void XBZRLE_cache_unlock(void) | |
167 | { | |
87dca0c9 | 168 | if (migrate_xbzrle()) { |
56e93d26 | 169 | qemu_mutex_unlock(&XBZRLE.lock); |
f4c51a6b | 170 | } |
56e93d26 JQ |
171 | } |
172 | ||
3d0684b2 JQ |
173 | /** |
174 | * xbzrle_cache_resize: resize the xbzrle cache | |
175 | * | |
cbde7be9 | 176 | * This function is called from migrate_params_apply in main |
3d0684b2 JQ |
177 | * thread, possibly while a migration is in progress. A running |
178 | * migration may be using the cache and might finish during this call, | |
179 | * hence changes to the cache are protected by XBZRLE.lock(). | |
180 | * | |
c9dede2d | 181 | * Returns 0 for success or -1 for error |
3d0684b2 JQ |
182 | * |
183 | * @new_size: new cache size | |
8acabf69 | 184 | * @errp: set *errp if the check failed, with reason |
56e93d26 | 185 | */ |
8b9407a0 | 186 | int xbzrle_cache_resize(uint64_t new_size, Error **errp) |
56e93d26 JQ |
187 | { |
188 | PageCache *new_cache; | |
c9dede2d | 189 | int64_t ret = 0; |
56e93d26 | 190 | |
8acabf69 JQ |
191 | /* Check for truncation */ |
192 | if (new_size != (size_t)new_size) { | |
193 | error_setg(errp, QERR_INVALID_PARAMETER_VALUE, "cache size", | |
194 | "exceeding address space"); | |
195 | return -1; | |
196 | } | |
197 | ||
2a313e5c JQ |
198 | if (new_size == migrate_xbzrle_cache_size()) { |
199 | /* nothing to do */ | |
c9dede2d | 200 | return 0; |
2a313e5c JQ |
201 | } |
202 | ||
56e93d26 JQ |
203 | XBZRLE_cache_lock(); |
204 | ||
205 | if (XBZRLE.cache != NULL) { | |
80f8dfde | 206 | new_cache = cache_init(new_size, TARGET_PAGE_SIZE, errp); |
56e93d26 | 207 | if (!new_cache) { |
56e93d26 JQ |
208 | ret = -1; |
209 | goto out; | |
210 | } | |
211 | ||
212 | cache_fini(XBZRLE.cache); | |
213 | XBZRLE.cache = new_cache; | |
214 | } | |
56e93d26 JQ |
215 | out: |
216 | XBZRLE_cache_unlock(); | |
217 | return ret; | |
218 | } | |
219 | ||
20123ee1 PX |
220 | static bool postcopy_preempt_active(void) |
221 | { | |
222 | return migrate_postcopy_preempt() && migration_in_postcopy(); | |
223 | } | |
224 | ||
3ded54b1 | 225 | bool ramblock_is_ignored(RAMBlock *block) |
fbd162e6 YK |
226 | { |
227 | return !qemu_ram_is_migratable(block) || | |
228 | (migrate_ignore_shared() && qemu_ram_is_shared(block)); | |
229 | } | |
230 | ||
343f632c DDAG |
231 | #undef RAMBLOCK_FOREACH |
232 | ||
fbd162e6 YK |
233 | int foreach_not_ignored_block(RAMBlockIterFunc func, void *opaque) |
234 | { | |
235 | RAMBlock *block; | |
236 | int ret = 0; | |
237 | ||
89ac5a1d DDAG |
238 | RCU_READ_LOCK_GUARD(); |
239 | ||
fbd162e6 YK |
240 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
241 | ret = func(block, opaque); | |
242 | if (ret) { | |
243 | break; | |
244 | } | |
245 | } | |
fbd162e6 YK |
246 | return ret; |
247 | } | |
248 | ||
f9494614 AP |
249 | static void ramblock_recv_map_init(void) |
250 | { | |
251 | RAMBlock *rb; | |
252 | ||
fbd162e6 | 253 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
254 | assert(!rb->receivedmap); |
255 | rb->receivedmap = bitmap_new(rb->max_length >> qemu_target_page_bits()); | |
256 | } | |
257 | } | |
258 | ||
259 | int ramblock_recv_bitmap_test(RAMBlock *rb, void *host_addr) | |
260 | { | |
261 | return test_bit(ramblock_recv_bitmap_offset(host_addr, rb), | |
262 | rb->receivedmap); | |
263 | } | |
264 | ||
1cba9f6e DDAG |
265 | bool ramblock_recv_bitmap_test_byte_offset(RAMBlock *rb, uint64_t byte_offset) |
266 | { | |
267 | return test_bit(byte_offset >> TARGET_PAGE_BITS, rb->receivedmap); | |
268 | } | |
269 | ||
f9494614 AP |
270 | void ramblock_recv_bitmap_set(RAMBlock *rb, void *host_addr) |
271 | { | |
272 | set_bit_atomic(ramblock_recv_bitmap_offset(host_addr, rb), rb->receivedmap); | |
273 | } | |
274 | ||
275 | void ramblock_recv_bitmap_set_range(RAMBlock *rb, void *host_addr, | |
276 | size_t nr) | |
277 | { | |
278 | bitmap_set_atomic(rb->receivedmap, | |
279 | ramblock_recv_bitmap_offset(host_addr, rb), | |
280 | nr); | |
281 | } | |
282 | ||
a335debb PX |
283 | #define RAMBLOCK_RECV_BITMAP_ENDING (0x0123456789abcdefULL) |
284 | ||
285 | /* | |
286 | * Format: bitmap_size (8 bytes) + whole_bitmap (N bytes). | |
287 | * | |
288 | * Returns >0 if success with sent bytes, or <0 if error. | |
289 | */ | |
290 | int64_t ramblock_recv_bitmap_send(QEMUFile *file, | |
291 | const char *block_name) | |
292 | { | |
293 | RAMBlock *block = qemu_ram_block_by_name(block_name); | |
294 | unsigned long *le_bitmap, nbits; | |
295 | uint64_t size; | |
296 | ||
297 | if (!block) { | |
298 | error_report("%s: invalid block name: %s", __func__, block_name); | |
299 | return -1; | |
300 | } | |
301 | ||
898ba906 | 302 | nbits = block->postcopy_length >> TARGET_PAGE_BITS; |
a335debb PX |
303 | |
304 | /* | |
305 | * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit | |
306 | * machines we may need 4 more bytes for padding (see below | |
307 | * comment). So extend it a bit before hand. | |
308 | */ | |
309 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
310 | ||
311 | /* | |
312 | * Always use little endian when sending the bitmap. This is | |
313 | * required that when source and destination VMs are not using the | |
3a4452d8 | 314 | * same endianness. (Note: big endian won't work.) |
a335debb PX |
315 | */ |
316 | bitmap_to_le(le_bitmap, block->receivedmap, nbits); | |
317 | ||
318 | /* Size of the bitmap, in bytes */ | |
a725ef9f | 319 | size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
320 | |
321 | /* | |
322 | * size is always aligned to 8 bytes for 64bit machines, but it | |
323 | * may not be true for 32bit machines. We need this padding to | |
324 | * make sure the migration can survive even between 32bit and | |
325 | * 64bit machines. | |
326 | */ | |
327 | size = ROUND_UP(size, 8); | |
328 | ||
329 | qemu_put_be64(file, size); | |
330 | qemu_put_buffer(file, (const uint8_t *)le_bitmap, size); | |
331 | /* | |
332 | * Mark as an end, in case the middle part is screwed up due to | |
3a4452d8 | 333 | * some "mysterious" reason. |
a335debb PX |
334 | */ |
335 | qemu_put_be64(file, RAMBLOCK_RECV_BITMAP_ENDING); | |
336 | qemu_fflush(file); | |
337 | ||
bf269906 | 338 | g_free(le_bitmap); |
a335debb PX |
339 | |
340 | if (qemu_file_get_error(file)) { | |
341 | return qemu_file_get_error(file); | |
342 | } | |
343 | ||
344 | return size + sizeof(size); | |
345 | } | |
346 | ||
ec481c6c JQ |
347 | /* |
348 | * An outstanding page request, on the source, having been received | |
349 | * and queued | |
350 | */ | |
351 | struct RAMSrcPageRequest { | |
352 | RAMBlock *rb; | |
353 | hwaddr offset; | |
354 | hwaddr len; | |
355 | ||
356 | QSIMPLEQ_ENTRY(RAMSrcPageRequest) next_req; | |
357 | }; | |
358 | ||
6f37bb8b JQ |
359 | /* State of RAM for migration */ |
360 | struct RAMState { | |
f1668764 PX |
361 | /* |
362 | * PageSearchStatus structures for the channels when send pages. | |
363 | * Protected by the bitmap_mutex. | |
364 | */ | |
365 | PageSearchStatus pss[RAM_CHANNEL_MAX]; | |
278e2f55 AG |
366 | /* UFFD file descriptor, used in 'write-tracking' migration */ |
367 | int uffdio_fd; | |
8d80e195 JQ |
368 | /* total ram size in bytes */ |
369 | uint64_t ram_bytes_total; | |
6f37bb8b JQ |
370 | /* Last block that we have visited searching for dirty pages */ |
371 | RAMBlock *last_seen_block; | |
269ace29 JQ |
372 | /* Last dirty target page we have sent */ |
373 | ram_addr_t last_page; | |
6f37bb8b JQ |
374 | /* last ram version we have seen */ |
375 | uint32_t last_version; | |
8d820d6f JQ |
376 | /* How many times we have dirty too many pages */ |
377 | int dirty_rate_high_cnt; | |
f664da80 JQ |
378 | /* these variables are used for bitmap sync */ |
379 | /* last time we did a full bitmap_sync */ | |
380 | int64_t time_last_bitmap_sync; | |
eac74159 | 381 | /* bytes transferred at start_time */ |
c4bdf0cf | 382 | uint64_t bytes_xfer_prev; |
a66cd90c | 383 | /* number of dirty pages since start_time */ |
68908ed6 | 384 | uint64_t num_dirty_pages_period; |
b5833fde JQ |
385 | /* xbzrle misses since the beginning of the period */ |
386 | uint64_t xbzrle_cache_miss_prev; | |
e460a4b1 WW |
387 | /* Amount of xbzrle pages since the beginning of the period */ |
388 | uint64_t xbzrle_pages_prev; | |
389 | /* Amount of xbzrle encoded bytes since the beginning of the period */ | |
390 | uint64_t xbzrle_bytes_prev; | |
f3095cc8 JQ |
391 | /* Are we really using XBZRLE (e.g., after the first round). */ |
392 | bool xbzrle_started; | |
05931ec5 JQ |
393 | /* Are we on the last stage of migration */ |
394 | bool last_stage; | |
76e03000 XG |
395 | /* compression statistics since the beginning of the period */ |
396 | /* amount of count that no free thread to compress data */ | |
397 | uint64_t compress_thread_busy_prev; | |
398 | /* amount bytes after compression */ | |
399 | uint64_t compressed_size_prev; | |
400 | /* amount of compressed pages */ | |
401 | uint64_t compress_pages_prev; | |
402 | ||
be8b02ed XG |
403 | /* total handled target pages at the beginning of period */ |
404 | uint64_t target_page_count_prev; | |
405 | /* total handled target pages since start */ | |
406 | uint64_t target_page_count; | |
9360447d | 407 | /* number of dirty bits in the bitmap */ |
2dfaf12e | 408 | uint64_t migration_dirty_pages; |
f1668764 PX |
409 | /* |
410 | * Protects: | |
411 | * - dirty/clear bitmap | |
412 | * - migration_dirty_pages | |
413 | * - pss structures | |
414 | */ | |
108cfae0 | 415 | QemuMutex bitmap_mutex; |
68a098f3 JQ |
416 | /* The RAMBlock used in the last src_page_requests */ |
417 | RAMBlock *last_req_rb; | |
ec481c6c JQ |
418 | /* Queue of outstanding page requests from the destination */ |
419 | QemuMutex src_page_req_mutex; | |
b58deb34 | 420 | QSIMPLEQ_HEAD(, RAMSrcPageRequest) src_page_requests; |
6f37bb8b JQ |
421 | }; |
422 | typedef struct RAMState RAMState; | |
423 | ||
53518d94 | 424 | static RAMState *ram_state; |
6f37bb8b | 425 | |
bd227060 WW |
426 | static NotifierWithReturnList precopy_notifier_list; |
427 | ||
a1fe28df PX |
428 | /* Whether postcopy has queued requests? */ |
429 | static bool postcopy_has_request(RAMState *rs) | |
430 | { | |
431 | return !QSIMPLEQ_EMPTY_ATOMIC(&rs->src_page_requests); | |
432 | } | |
433 | ||
bd227060 WW |
434 | void precopy_infrastructure_init(void) |
435 | { | |
436 | notifier_with_return_list_init(&precopy_notifier_list); | |
437 | } | |
438 | ||
439 | void precopy_add_notifier(NotifierWithReturn *n) | |
440 | { | |
441 | notifier_with_return_list_add(&precopy_notifier_list, n); | |
442 | } | |
443 | ||
444 | void precopy_remove_notifier(NotifierWithReturn *n) | |
445 | { | |
446 | notifier_with_return_remove(n); | |
447 | } | |
448 | ||
449 | int precopy_notify(PrecopyNotifyReason reason, Error **errp) | |
450 | { | |
451 | PrecopyNotifyData pnd; | |
452 | pnd.reason = reason; | |
453 | pnd.errp = errp; | |
454 | ||
455 | return notifier_with_return_list_notify(&precopy_notifier_list, &pnd); | |
456 | } | |
457 | ||
9edabd4d | 458 | uint64_t ram_bytes_remaining(void) |
2f4fde93 | 459 | { |
bae416e5 DDAG |
460 | return ram_state ? (ram_state->migration_dirty_pages * TARGET_PAGE_SIZE) : |
461 | 0; | |
2f4fde93 JQ |
462 | } |
463 | ||
26a26069 | 464 | void ram_transferred_add(uint64_t bytes) |
4c2d0f6d | 465 | { |
ae680668 | 466 | if (runstate_is_running()) { |
aff3f660 | 467 | stat64_add(&mig_stats.precopy_bytes, bytes); |
ae680668 | 468 | } else if (migration_in_postcopy()) { |
aff3f660 | 469 | stat64_add(&mig_stats.postcopy_bytes, bytes); |
ae680668 | 470 | } else { |
aff3f660 | 471 | stat64_add(&mig_stats.downtime_bytes, bytes); |
ae680668 | 472 | } |
aff3f660 | 473 | stat64_add(&mig_stats.transferred, bytes); |
4c2d0f6d DE |
474 | } |
475 | ||
4010ba38 JQ |
476 | struct MigrationOps { |
477 | int (*ram_save_target_page)(RAMState *rs, PageSearchStatus *pss); | |
478 | }; | |
479 | typedef struct MigrationOps MigrationOps; | |
480 | ||
481 | MigrationOps *migration_ops; | |
482 | ||
93589827 PX |
483 | static int ram_save_host_page_urgent(PageSearchStatus *pss); |
484 | ||
ebd88a49 PX |
485 | /* NOTE: page is the PFN not real ram_addr_t. */ |
486 | static void pss_init(PageSearchStatus *pss, RAMBlock *rb, ram_addr_t page) | |
487 | { | |
488 | pss->block = rb; | |
489 | pss->page = page; | |
490 | pss->complete_round = false; | |
491 | } | |
492 | ||
93589827 PX |
493 | /* |
494 | * Check whether two PSSs are actively sending the same page. Return true | |
495 | * if it is, false otherwise. | |
496 | */ | |
497 | static bool pss_overlap(PageSearchStatus *pss1, PageSearchStatus *pss2) | |
498 | { | |
499 | return pss1->host_page_sending && pss2->host_page_sending && | |
500 | (pss1->host_page_start == pss2->host_page_start); | |
501 | } | |
502 | ||
56e93d26 | 503 | /** |
3d0684b2 | 504 | * save_page_header: write page header to wire |
56e93d26 JQ |
505 | * |
506 | * If this is the 1st block, it also writes the block identification | |
507 | * | |
3d0684b2 | 508 | * Returns the number of bytes written |
56e93d26 | 509 | * |
ec6f3ab9 | 510 | * @pss: current PSS channel status |
56e93d26 JQ |
511 | * @block: block that contains the page we want to send |
512 | * @offset: offset inside the block for the page | |
513 | * in the lower bits, it contains flags | |
514 | */ | |
37502df3 LS |
515 | static size_t save_page_header(PageSearchStatus *pss, QEMUFile *f, |
516 | RAMBlock *block, ram_addr_t offset) | |
56e93d26 | 517 | { |
9f5f380b | 518 | size_t size, len; |
ec6f3ab9 | 519 | bool same_block = (block == pss->last_sent_block); |
56e93d26 | 520 | |
10661f11 | 521 | if (same_block) { |
24795694 JQ |
522 | offset |= RAM_SAVE_FLAG_CONTINUE; |
523 | } | |
2bf3aa85 | 524 | qemu_put_be64(f, offset); |
56e93d26 JQ |
525 | size = 8; |
526 | ||
10661f11 | 527 | if (!same_block) { |
9f5f380b | 528 | len = strlen(block->idstr); |
2bf3aa85 JQ |
529 | qemu_put_byte(f, len); |
530 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 531 | size += 1 + len; |
ec6f3ab9 | 532 | pss->last_sent_block = block; |
56e93d26 JQ |
533 | } |
534 | return size; | |
535 | } | |
536 | ||
3d0684b2 | 537 | /** |
179a8080 | 538 | * mig_throttle_guest_down: throttle down the guest |
3d0684b2 JQ |
539 | * |
540 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
541 | * writes. If guest dirty memory rate is reduced below the rate at | |
542 | * which we can transfer pages to the destination then we should be | |
543 | * able to complete migration. Some workloads dirty memory way too | |
544 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 | 545 | */ |
cbbf8182 KZ |
546 | static void mig_throttle_guest_down(uint64_t bytes_dirty_period, |
547 | uint64_t bytes_dirty_threshold) | |
070afca2 | 548 | { |
2a8ec380 | 549 | uint64_t pct_initial = migrate_cpu_throttle_initial(); |
9605c2ac | 550 | uint64_t pct_increment = migrate_cpu_throttle_increment(); |
873f674c | 551 | bool pct_tailslow = migrate_cpu_throttle_tailslow(); |
24155bd0 | 552 | int pct_max = migrate_max_cpu_throttle(); |
070afca2 | 553 | |
cbbf8182 KZ |
554 | uint64_t throttle_now = cpu_throttle_get_percentage(); |
555 | uint64_t cpu_now, cpu_ideal, throttle_inc; | |
556 | ||
070afca2 JH |
557 | /* We have not started throttling yet. Let's start it. */ |
558 | if (!cpu_throttle_active()) { | |
559 | cpu_throttle_set(pct_initial); | |
560 | } else { | |
561 | /* Throttling already on, just increase the rate */ | |
cbbf8182 KZ |
562 | if (!pct_tailslow) { |
563 | throttle_inc = pct_increment; | |
564 | } else { | |
565 | /* Compute the ideal CPU percentage used by Guest, which may | |
566 | * make the dirty rate match the dirty rate threshold. */ | |
567 | cpu_now = 100 - throttle_now; | |
568 | cpu_ideal = cpu_now * (bytes_dirty_threshold * 1.0 / | |
569 | bytes_dirty_period); | |
570 | throttle_inc = MIN(cpu_now - cpu_ideal, pct_increment); | |
571 | } | |
572 | cpu_throttle_set(MIN(throttle_now + throttle_inc, pct_max)); | |
070afca2 JH |
573 | } |
574 | } | |
575 | ||
91fe9a8d RL |
576 | void mig_throttle_counter_reset(void) |
577 | { | |
578 | RAMState *rs = ram_state; | |
579 | ||
580 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
581 | rs->num_dirty_pages_period = 0; | |
aff3f660 | 582 | rs->bytes_xfer_prev = stat64_get(&mig_stats.transferred); |
91fe9a8d RL |
583 | } |
584 | ||
3d0684b2 JQ |
585 | /** |
586 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
587 | * | |
6f37bb8b | 588 | * @rs: current RAM state |
3d0684b2 JQ |
589 | * @current_addr: address for the zero page |
590 | * | |
591 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
592 | * The important thing is that a stale (not-yet-0'd) page be replaced |
593 | * by the new data. | |
594 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 595 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 596 | */ |
6f37bb8b | 597 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 598 | { |
56e93d26 JQ |
599 | /* We don't care if this fails to allocate a new cache page |
600 | * as long as it updated an old one */ | |
c00e0928 | 601 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
aff3f660 | 602 | stat64_get(&mig_stats.dirty_sync_count)); |
56e93d26 JQ |
603 | } |
604 | ||
605 | #define ENCODING_FLAG_XBZRLE 0x1 | |
606 | ||
607 | /** | |
608 | * save_xbzrle_page: compress and send current page | |
609 | * | |
610 | * Returns: 1 means that we wrote the page | |
611 | * 0 means that page is identical to the one already sent | |
612 | * -1 means that xbzrle would be longer than normal | |
613 | * | |
5a987738 | 614 | * @rs: current RAM state |
ec6f3ab9 | 615 | * @pss: current PSS channel |
3d0684b2 JQ |
616 | * @current_data: pointer to the address of the page contents |
617 | * @current_addr: addr of the page | |
56e93d26 JQ |
618 | * @block: block that contains the page we want to send |
619 | * @offset: offset inside the block for the page | |
56e93d26 | 620 | */ |
ec6f3ab9 | 621 | static int save_xbzrle_page(RAMState *rs, PageSearchStatus *pss, |
61717ea9 PX |
622 | uint8_t **current_data, ram_addr_t current_addr, |
623 | RAMBlock *block, ram_addr_t offset) | |
56e93d26 JQ |
624 | { |
625 | int encoded_len = 0, bytes_xbzrle; | |
626 | uint8_t *prev_cached_page; | |
ec6f3ab9 | 627 | QEMUFile *file = pss->pss_channel; |
aff3f660 | 628 | uint64_t generation = stat64_get(&mig_stats.dirty_sync_count); |
56e93d26 | 629 | |
536b5a4e | 630 | if (!cache_is_cached(XBZRLE.cache, current_addr, generation)) { |
9360447d | 631 | xbzrle_counters.cache_miss++; |
05931ec5 | 632 | if (!rs->last_stage) { |
56e93d26 | 633 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, |
536b5a4e | 634 | generation) == -1) { |
56e93d26 JQ |
635 | return -1; |
636 | } else { | |
637 | /* update *current_data when the page has been | |
638 | inserted into cache */ | |
639 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
640 | } | |
641 | } | |
642 | return -1; | |
643 | } | |
644 | ||
e460a4b1 WW |
645 | /* |
646 | * Reaching here means the page has hit the xbzrle cache, no matter what | |
647 | * encoding result it is (normal encoding, overflow or skipping the page), | |
3a4452d8 | 648 | * count the page as encoded. This is used to calculate the encoding rate. |
e460a4b1 WW |
649 | * |
650 | * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB, | |
651 | * 2nd page turns out to be skipped (i.e. no new bytes written to the | |
652 | * page), the overall encoding rate will be 8KB / 2KB = 4, which has the | |
653 | * skipped page included. In this way, the encoding rate can tell if the | |
654 | * guest page is good for xbzrle encoding. | |
655 | */ | |
656 | xbzrle_counters.pages++; | |
56e93d26 JQ |
657 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); |
658 | ||
659 | /* save current buffer into memory */ | |
660 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
661 | ||
662 | /* XBZRLE encoding (if there is no overflow) */ | |
04ffce13 | 663 | encoded_len = xbzrle_encode_buffer_func(prev_cached_page, XBZRLE.current_buf, |
664 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
665 | TARGET_PAGE_SIZE); | |
ca353803 WY |
666 | |
667 | /* | |
668 | * Update the cache contents, so that it corresponds to the data | |
669 | * sent, in all cases except where we skip the page. | |
670 | */ | |
05931ec5 | 671 | if (!rs->last_stage && encoded_len != 0) { |
ca353803 WY |
672 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); |
673 | /* | |
674 | * In the case where we couldn't compress, ensure that the caller | |
675 | * sends the data from the cache, since the guest might have | |
676 | * changed the RAM since we copied it. | |
677 | */ | |
678 | *current_data = prev_cached_page; | |
679 | } | |
680 | ||
56e93d26 | 681 | if (encoded_len == 0) { |
55c4446b | 682 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
683 | return 0; |
684 | } else if (encoded_len == -1) { | |
55c4446b | 685 | trace_save_xbzrle_page_overflow(); |
9360447d | 686 | xbzrle_counters.overflow++; |
e460a4b1 | 687 | xbzrle_counters.bytes += TARGET_PAGE_SIZE; |
56e93d26 JQ |
688 | return -1; |
689 | } | |
690 | ||
56e93d26 | 691 | /* Send XBZRLE based compressed page */ |
37502df3 | 692 | bytes_xbzrle = save_page_header(pss, pss->pss_channel, block, |
204b88b8 | 693 | offset | RAM_SAVE_FLAG_XBZRLE); |
61717ea9 PX |
694 | qemu_put_byte(file, ENCODING_FLAG_XBZRLE); |
695 | qemu_put_be16(file, encoded_len); | |
696 | qemu_put_buffer(file, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 697 | bytes_xbzrle += encoded_len + 1 + 2; |
e460a4b1 WW |
698 | /* |
699 | * Like compressed_size (please see update_compress_thread_counts), | |
700 | * the xbzrle encoded bytes don't count the 8 byte header with | |
701 | * RAM_SAVE_FLAG_CONTINUE. | |
702 | */ | |
703 | xbzrle_counters.bytes += bytes_xbzrle - 8; | |
4c2d0f6d | 704 | ram_transferred_add(bytes_xbzrle); |
56e93d26 JQ |
705 | |
706 | return 1; | |
707 | } | |
708 | ||
3d0684b2 | 709 | /** |
d9e474ea | 710 | * pss_find_next_dirty: find the next dirty page of current ramblock |
f3f491fc | 711 | * |
d9e474ea PX |
712 | * This function updates pss->page to point to the next dirty page index |
713 | * within the ramblock to migrate, or the end of ramblock when nothing | |
714 | * found. Note that when pss->host_page_sending==true it means we're | |
715 | * during sending a host page, so we won't look for dirty page that is | |
716 | * outside the host page boundary. | |
3d0684b2 | 717 | * |
d9e474ea | 718 | * @pss: the current page search status |
f3f491fc | 719 | */ |
d9e474ea | 720 | static void pss_find_next_dirty(PageSearchStatus *pss) |
56e93d26 | 721 | { |
d9e474ea | 722 | RAMBlock *rb = pss->block; |
6b6712ef JQ |
723 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
724 | unsigned long *bitmap = rb->bmap; | |
56e93d26 | 725 | |
fbd162e6 | 726 | if (ramblock_is_ignored(rb)) { |
d9e474ea PX |
727 | /* Points directly to the end, so we know no dirty page */ |
728 | pss->page = size; | |
729 | return; | |
730 | } | |
731 | ||
732 | /* | |
733 | * If during sending a host page, only look for dirty pages within the | |
734 | * current host page being send. | |
735 | */ | |
736 | if (pss->host_page_sending) { | |
737 | assert(pss->host_page_end); | |
738 | size = MIN(size, pss->host_page_end); | |
b895de50 CLG |
739 | } |
740 | ||
d9e474ea | 741 | pss->page = find_next_bit(bitmap, size, pss->page); |
56e93d26 JQ |
742 | } |
743 | ||
1230a25f | 744 | static void migration_clear_memory_region_dirty_bitmap(RAMBlock *rb, |
3143577d WW |
745 | unsigned long page) |
746 | { | |
747 | uint8_t shift; | |
748 | hwaddr size, start; | |
749 | ||
750 | if (!rb->clear_bmap || !clear_bmap_test_and_clear(rb, page)) { | |
751 | return; | |
752 | } | |
753 | ||
754 | shift = rb->clear_bmap_shift; | |
755 | /* | |
756 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
757 | * can make things easier sometimes since then start address | |
758 | * of the small chunk will always be 64 pages aligned so the | |
759 | * bitmap will always be aligned to unsigned long. We should | |
760 | * even be able to remove this restriction but I'm simply | |
761 | * keeping it. | |
762 | */ | |
763 | assert(shift >= 6); | |
764 | ||
765 | size = 1ULL << (TARGET_PAGE_BITS + shift); | |
7648297d | 766 | start = QEMU_ALIGN_DOWN((ram_addr_t)page << TARGET_PAGE_BITS, size); |
3143577d WW |
767 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); |
768 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
769 | } | |
770 | ||
771 | static void | |
1230a25f | 772 | migration_clear_memory_region_dirty_bitmap_range(RAMBlock *rb, |
3143577d WW |
773 | unsigned long start, |
774 | unsigned long npages) | |
775 | { | |
776 | unsigned long i, chunk_pages = 1UL << rb->clear_bmap_shift; | |
777 | unsigned long chunk_start = QEMU_ALIGN_DOWN(start, chunk_pages); | |
778 | unsigned long chunk_end = QEMU_ALIGN_UP(start + npages, chunk_pages); | |
779 | ||
780 | /* | |
781 | * Clear pages from start to start + npages - 1, so the end boundary is | |
782 | * exclusive. | |
783 | */ | |
784 | for (i = chunk_start; i < chunk_end; i += chunk_pages) { | |
1230a25f | 785 | migration_clear_memory_region_dirty_bitmap(rb, i); |
3143577d WW |
786 | } |
787 | } | |
788 | ||
a6a83cef RL |
789 | /* |
790 | * colo_bitmap_find_diry:find contiguous dirty pages from start | |
791 | * | |
792 | * Returns the page offset within memory region of the start of the contiguout | |
793 | * dirty page | |
794 | * | |
795 | * @rs: current RAM state | |
796 | * @rb: RAMBlock where to search for dirty pages | |
797 | * @start: page where we start the search | |
798 | * @num: the number of contiguous dirty pages | |
799 | */ | |
800 | static inline | |
801 | unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, | |
802 | unsigned long start, unsigned long *num) | |
803 | { | |
804 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; | |
805 | unsigned long *bitmap = rb->bmap; | |
806 | unsigned long first, next; | |
807 | ||
808 | *num = 0; | |
809 | ||
810 | if (ramblock_is_ignored(rb)) { | |
811 | return size; | |
812 | } | |
813 | ||
814 | first = find_next_bit(bitmap, size, start); | |
815 | if (first >= size) { | |
816 | return first; | |
817 | } | |
818 | next = find_next_zero_bit(bitmap, size, first + 1); | |
819 | assert(next >= first); | |
820 | *num = next - first; | |
821 | return first; | |
822 | } | |
823 | ||
06b10688 | 824 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
825 | RAMBlock *rb, |
826 | unsigned long page) | |
a82d593b DDAG |
827 | { |
828 | bool ret; | |
a82d593b | 829 | |
002cad6b PX |
830 | /* |
831 | * Clear dirty bitmap if needed. This _must_ be called before we | |
832 | * send any of the page in the chunk because we need to make sure | |
833 | * we can capture further page content changes when we sync dirty | |
834 | * log the next time. So as long as we are going to send any of | |
835 | * the page in the chunk we clear the remote dirty bitmap for all. | |
836 | * Clearing it earlier won't be a problem, but too late will. | |
837 | */ | |
1230a25f | 838 | migration_clear_memory_region_dirty_bitmap(rb, page); |
002cad6b | 839 | |
6b6712ef | 840 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b | 841 | if (ret) { |
0d8ec885 | 842 | rs->migration_dirty_pages--; |
a82d593b | 843 | } |
386a907b | 844 | |
a82d593b DDAG |
845 | return ret; |
846 | } | |
847 | ||
be39b4cd DH |
848 | static void dirty_bitmap_clear_section(MemoryRegionSection *section, |
849 | void *opaque) | |
850 | { | |
851 | const hwaddr offset = section->offset_within_region; | |
852 | const hwaddr size = int128_get64(section->size); | |
853 | const unsigned long start = offset >> TARGET_PAGE_BITS; | |
854 | const unsigned long npages = size >> TARGET_PAGE_BITS; | |
855 | RAMBlock *rb = section->mr->ram_block; | |
856 | uint64_t *cleared_bits = opaque; | |
857 | ||
858 | /* | |
859 | * We don't grab ram_state->bitmap_mutex because we expect to run | |
860 | * only when starting migration or during postcopy recovery where | |
861 | * we don't have concurrent access. | |
862 | */ | |
863 | if (!migration_in_postcopy() && !migrate_background_snapshot()) { | |
864 | migration_clear_memory_region_dirty_bitmap_range(rb, start, npages); | |
865 | } | |
866 | *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages); | |
867 | bitmap_clear(rb->bmap, start, npages); | |
868 | } | |
869 | ||
870 | /* | |
871 | * Exclude all dirty pages from migration that fall into a discarded range as | |
872 | * managed by a RamDiscardManager responsible for the mapped memory region of | |
873 | * the RAMBlock. Clear the corresponding bits in the dirty bitmaps. | |
874 | * | |
875 | * Discarded pages ("logically unplugged") have undefined content and must | |
876 | * not get migrated, because even reading these pages for migration might | |
877 | * result in undesired behavior. | |
878 | * | |
879 | * Returns the number of cleared bits in the RAMBlock dirty bitmap. | |
880 | * | |
881 | * Note: The result is only stable while migrating (precopy/postcopy). | |
882 | */ | |
883 | static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb) | |
884 | { | |
885 | uint64_t cleared_bits = 0; | |
886 | ||
887 | if (rb->mr && rb->bmap && memory_region_has_ram_discard_manager(rb->mr)) { | |
888 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
889 | MemoryRegionSection section = { | |
890 | .mr = rb->mr, | |
891 | .offset_within_region = 0, | |
892 | .size = int128_make64(qemu_ram_get_used_length(rb)), | |
893 | }; | |
894 | ||
895 | ram_discard_manager_replay_discarded(rdm, §ion, | |
896 | dirty_bitmap_clear_section, | |
897 | &cleared_bits); | |
898 | } | |
899 | return cleared_bits; | |
900 | } | |
901 | ||
9470c5e0 DH |
902 | /* |
903 | * Check if a host-page aligned page falls into a discarded range as managed by | |
904 | * a RamDiscardManager responsible for the mapped memory region of the RAMBlock. | |
905 | * | |
906 | * Note: The result is only stable while migrating (precopy/postcopy). | |
907 | */ | |
908 | bool ramblock_page_is_discarded(RAMBlock *rb, ram_addr_t start) | |
909 | { | |
910 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
911 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
912 | MemoryRegionSection section = { | |
913 | .mr = rb->mr, | |
914 | .offset_within_region = start, | |
915 | .size = int128_make64(qemu_ram_pagesize(rb)), | |
916 | }; | |
917 | ||
918 | return !ram_discard_manager_is_populated(rdm, §ion); | |
919 | } | |
920 | return false; | |
921 | } | |
922 | ||
267691b6 | 923 | /* Called with RCU critical section */ |
7a3e9571 | 924 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 925 | { |
fb613580 KZ |
926 | uint64_t new_dirty_pages = |
927 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length); | |
928 | ||
929 | rs->migration_dirty_pages += new_dirty_pages; | |
930 | rs->num_dirty_pages_period += new_dirty_pages; | |
56e93d26 JQ |
931 | } |
932 | ||
3d0684b2 JQ |
933 | /** |
934 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
935 | * | |
936 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
937 | * | |
938 | * For VMs with just normal pages this is equivalent to the host page | |
939 | * size. If it's got some huge pages then it's the OR of all the | |
940 | * different page sizes. | |
e8ca1db2 DDAG |
941 | */ |
942 | uint64_t ram_pagesize_summary(void) | |
943 | { | |
944 | RAMBlock *block; | |
945 | uint64_t summary = 0; | |
946 | ||
fbd162e6 | 947 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
948 | summary |= block->page_size; |
949 | } | |
950 | ||
951 | return summary; | |
952 | } | |
953 | ||
aecbfe9c XG |
954 | uint64_t ram_get_total_transferred_pages(void) |
955 | { | |
aff3f660 JQ |
956 | return stat64_get(&mig_stats.normal_pages) + |
957 | stat64_get(&mig_stats.zero_pages) + | |
23b7576d | 958 | compression_counters.pages + xbzrle_counters.pages; |
aecbfe9c XG |
959 | } |
960 | ||
b734035b XG |
961 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
962 | { | |
be8b02ed | 963 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 964 | double compressed_size; |
b734035b XG |
965 | |
966 | /* calculate period counters */ | |
aff3f660 | 967 | stat64_set(&mig_stats.dirty_pages_rate, |
72f8e587 JQ |
968 | rs->num_dirty_pages_period * 1000 / |
969 | (end_time - rs->time_last_bitmap_sync)); | |
b734035b | 970 | |
be8b02ed | 971 | if (!page_count) { |
b734035b XG |
972 | return; |
973 | } | |
974 | ||
87dca0c9 | 975 | if (migrate_xbzrle()) { |
e460a4b1 WW |
976 | double encoded_size, unencoded_size; |
977 | ||
b734035b | 978 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - |
be8b02ed | 979 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b | 980 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
e460a4b1 WW |
981 | unencoded_size = (xbzrle_counters.pages - rs->xbzrle_pages_prev) * |
982 | TARGET_PAGE_SIZE; | |
983 | encoded_size = xbzrle_counters.bytes - rs->xbzrle_bytes_prev; | |
92271402 | 984 | if (xbzrle_counters.pages == rs->xbzrle_pages_prev || !encoded_size) { |
e460a4b1 | 985 | xbzrle_counters.encoding_rate = 0; |
e460a4b1 WW |
986 | } else { |
987 | xbzrle_counters.encoding_rate = unencoded_size / encoded_size; | |
988 | } | |
989 | rs->xbzrle_pages_prev = xbzrle_counters.pages; | |
990 | rs->xbzrle_bytes_prev = xbzrle_counters.bytes; | |
b734035b | 991 | } |
76e03000 | 992 | |
a7a94d14 | 993 | if (migrate_compress()) { |
76e03000 XG |
994 | compression_counters.busy_rate = (double)(compression_counters.busy - |
995 | rs->compress_thread_busy_prev) / page_count; | |
996 | rs->compress_thread_busy_prev = compression_counters.busy; | |
997 | ||
998 | compressed_size = compression_counters.compressed_size - | |
999 | rs->compressed_size_prev; | |
1000 | if (compressed_size) { | |
1001 | double uncompressed_size = (compression_counters.pages - | |
1002 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
1003 | ||
1004 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
1005 | compression_counters.compression_rate = | |
1006 | uncompressed_size / compressed_size; | |
1007 | ||
1008 | rs->compress_pages_prev = compression_counters.pages; | |
1009 | rs->compressed_size_prev = compression_counters.compressed_size; | |
1010 | } | |
1011 | } | |
b734035b XG |
1012 | } |
1013 | ||
dc14a470 KZ |
1014 | static void migration_trigger_throttle(RAMState *rs) |
1015 | { | |
6499efdb | 1016 | uint64_t threshold = migrate_throttle_trigger_threshold(); |
23b7576d | 1017 | uint64_t bytes_xfer_period = |
aff3f660 | 1018 | stat64_get(&mig_stats.transferred) - rs->bytes_xfer_prev; |
dc14a470 KZ |
1019 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; |
1020 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
1021 | ||
1022 | /* During block migration the auto-converge logic incorrectly detects | |
1023 | * that ram migration makes no progress. Avoid this by disabling the | |
1024 | * throttling logic during the bulk phase of block migration. */ | |
1025 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
1026 | /* The following detection logic can be refined later. For now: | |
1027 | Check to see if the ratio between dirtied bytes and the approx. | |
1028 | amount of bytes that just got transferred since the last time | |
1029 | we were in this routine reaches the threshold. If that happens | |
1030 | twice, start or increase throttling. */ | |
1031 | ||
1032 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
1033 | (++rs->dirty_rate_high_cnt >= 2)) { | |
1034 | trace_migration_throttle(); | |
1035 | rs->dirty_rate_high_cnt = 0; | |
cbbf8182 KZ |
1036 | mig_throttle_guest_down(bytes_dirty_period, |
1037 | bytes_dirty_threshold); | |
dc14a470 KZ |
1038 | } |
1039 | } | |
1040 | } | |
1041 | ||
1e493be5 | 1042 | static void migration_bitmap_sync(RAMState *rs, bool last_stage) |
56e93d26 JQ |
1043 | { |
1044 | RAMBlock *block; | |
56e93d26 | 1045 | int64_t end_time; |
56e93d26 | 1046 | |
aff3f660 | 1047 | stat64_add(&mig_stats.dirty_sync_count, 1); |
56e93d26 | 1048 | |
f664da80 JQ |
1049 | if (!rs->time_last_bitmap_sync) { |
1050 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
1051 | } |
1052 | ||
1053 | trace_migration_bitmap_sync_start(); | |
1e493be5 | 1054 | memory_global_dirty_log_sync(last_stage); |
56e93d26 | 1055 | |
108cfae0 | 1056 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
1057 | WITH_RCU_READ_LOCK_GUARD() { |
1058 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1059 | ramblock_sync_dirty_bitmap(rs, block); | |
1060 | } | |
aff3f660 | 1061 | stat64_set(&mig_stats.dirty_bytes_last_sync, ram_bytes_remaining()); |
56e93d26 | 1062 | } |
108cfae0 | 1063 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 1064 | |
9458a9a1 | 1065 | memory_global_after_dirty_log_sync(); |
a66cd90c | 1066 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 1067 | |
56e93d26 JQ |
1068 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
1069 | ||
1070 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 1071 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 1072 | migration_trigger_throttle(rs); |
070afca2 | 1073 | |
b734035b XG |
1074 | migration_update_rates(rs, end_time); |
1075 | ||
be8b02ed | 1076 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
1077 | |
1078 | /* reset period counters */ | |
f664da80 | 1079 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 1080 | rs->num_dirty_pages_period = 0; |
aff3f660 | 1081 | rs->bytes_xfer_prev = stat64_get(&mig_stats.transferred); |
56e93d26 | 1082 | } |
b890902c | 1083 | if (migrate_events()) { |
aff3f660 | 1084 | uint64_t generation = stat64_get(&mig_stats.dirty_sync_count); |
536b5a4e | 1085 | qapi_event_send_migration_pass(generation); |
4addcd4f | 1086 | } |
56e93d26 JQ |
1087 | } |
1088 | ||
1e493be5 | 1089 | static void migration_bitmap_sync_precopy(RAMState *rs, bool last_stage) |
bd227060 WW |
1090 | { |
1091 | Error *local_err = NULL; | |
1092 | ||
1093 | /* | |
1094 | * The current notifier usage is just an optimization to migration, so we | |
1095 | * don't stop the normal migration process in the error case. | |
1096 | */ | |
1097 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
1098 | error_report_err(local_err); | |
b4a1733c | 1099 | local_err = NULL; |
bd227060 WW |
1100 | } |
1101 | ||
1e493be5 | 1102 | migration_bitmap_sync(rs, last_stage); |
bd227060 WW |
1103 | |
1104 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
1105 | error_report_err(local_err); | |
1106 | } | |
1107 | } | |
1108 | ||
a4dbaf8e | 1109 | void ram_release_page(const char *rbname, uint64_t offset) |
47fe16ff JQ |
1110 | { |
1111 | if (!migrate_release_ram() || !migration_in_postcopy()) { | |
1112 | return; | |
1113 | } | |
1114 | ||
1115 | ram_discard_range(rbname, offset, TARGET_PAGE_SIZE); | |
1116 | } | |
1117 | ||
6c97ec5f XG |
1118 | /** |
1119 | * save_zero_page_to_file: send the zero page to the file | |
1120 | * | |
1121 | * Returns the size of data written to the file, 0 means the page is not | |
1122 | * a zero page | |
1123 | * | |
ec6f3ab9 | 1124 | * @pss: current PSS channel |
6c97ec5f XG |
1125 | * @block: block that contains the page we want to send |
1126 | * @offset: offset inside the block for the page | |
1127 | */ | |
37502df3 | 1128 | static int save_zero_page_to_file(PageSearchStatus *pss, QEMUFile *file, |
6c97ec5f XG |
1129 | RAMBlock *block, ram_addr_t offset) |
1130 | { | |
1131 | uint8_t *p = block->host + offset; | |
1132 | int len = 0; | |
1133 | ||
bad452a7 | 1134 | if (buffer_is_zero(p, TARGET_PAGE_SIZE)) { |
37502df3 | 1135 | len += save_page_header(pss, file, block, offset | RAM_SAVE_FLAG_ZERO); |
6c97ec5f XG |
1136 | qemu_put_byte(file, 0); |
1137 | len += 1; | |
47fe16ff | 1138 | ram_release_page(block->idstr, offset); |
6c97ec5f XG |
1139 | } |
1140 | return len; | |
1141 | } | |
1142 | ||
56e93d26 | 1143 | /** |
3d0684b2 | 1144 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1145 | * |
3d0684b2 | 1146 | * Returns the number of pages written. |
56e93d26 | 1147 | * |
ec6f3ab9 | 1148 | * @pss: current PSS channel |
56e93d26 JQ |
1149 | * @block: block that contains the page we want to send |
1150 | * @offset: offset inside the block for the page | |
56e93d26 | 1151 | */ |
37502df3 | 1152 | static int save_zero_page(PageSearchStatus *pss, QEMUFile *f, RAMBlock *block, |
61717ea9 | 1153 | ram_addr_t offset) |
56e93d26 | 1154 | { |
37502df3 | 1155 | int len = save_zero_page_to_file(pss, f, block, offset); |
56e93d26 | 1156 | |
6c97ec5f | 1157 | if (len) { |
aff3f660 | 1158 | stat64_add(&mig_stats.zero_pages, 1); |
4c2d0f6d | 1159 | ram_transferred_add(len); |
6c97ec5f | 1160 | return 1; |
56e93d26 | 1161 | } |
6c97ec5f | 1162 | return -1; |
56e93d26 JQ |
1163 | } |
1164 | ||
059ff0fb XG |
1165 | /* |
1166 | * @pages: the number of pages written by the control path, | |
1167 | * < 0 - error | |
1168 | * > 0 - number of pages written | |
1169 | * | |
1170 | * Return true if the pages has been saved, otherwise false is returned. | |
1171 | */ | |
61717ea9 PX |
1172 | static bool control_save_page(PageSearchStatus *pss, RAMBlock *block, |
1173 | ram_addr_t offset, int *pages) | |
059ff0fb XG |
1174 | { |
1175 | uint64_t bytes_xmit = 0; | |
1176 | int ret; | |
1177 | ||
1178 | *pages = -1; | |
61717ea9 PX |
1179 | ret = ram_control_save_page(pss->pss_channel, block->offset, offset, |
1180 | TARGET_PAGE_SIZE, &bytes_xmit); | |
059ff0fb XG |
1181 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { |
1182 | return false; | |
1183 | } | |
1184 | ||
1185 | if (bytes_xmit) { | |
4c2d0f6d | 1186 | ram_transferred_add(bytes_xmit); |
059ff0fb XG |
1187 | *pages = 1; |
1188 | } | |
1189 | ||
1190 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1191 | return true; | |
1192 | } | |
1193 | ||
1194 | if (bytes_xmit > 0) { | |
aff3f660 | 1195 | stat64_add(&mig_stats.normal_pages, 1); |
059ff0fb | 1196 | } else if (bytes_xmit == 0) { |
aff3f660 | 1197 | stat64_add(&mig_stats.zero_pages, 1); |
059ff0fb XG |
1198 | } |
1199 | ||
1200 | return true; | |
1201 | } | |
1202 | ||
65dacaa0 XG |
1203 | /* |
1204 | * directly send the page to the stream | |
1205 | * | |
1206 | * Returns the number of pages written. | |
1207 | * | |
ec6f3ab9 | 1208 | * @pss: current PSS channel |
65dacaa0 XG |
1209 | * @block: block that contains the page we want to send |
1210 | * @offset: offset inside the block for the page | |
1211 | * @buf: the page to be sent | |
1212 | * @async: send to page asyncly | |
1213 | */ | |
ec6f3ab9 | 1214 | static int save_normal_page(PageSearchStatus *pss, RAMBlock *block, |
61717ea9 | 1215 | ram_addr_t offset, uint8_t *buf, bool async) |
65dacaa0 | 1216 | { |
ec6f3ab9 PX |
1217 | QEMUFile *file = pss->pss_channel; |
1218 | ||
37502df3 | 1219 | ram_transferred_add(save_page_header(pss, pss->pss_channel, block, |
4c2d0f6d | 1220 | offset | RAM_SAVE_FLAG_PAGE)); |
65dacaa0 | 1221 | if (async) { |
61717ea9 | 1222 | qemu_put_buffer_async(file, buf, TARGET_PAGE_SIZE, |
f912ec5b | 1223 | migrate_release_ram() && |
65dacaa0 XG |
1224 | migration_in_postcopy()); |
1225 | } else { | |
61717ea9 | 1226 | qemu_put_buffer(file, buf, TARGET_PAGE_SIZE); |
65dacaa0 | 1227 | } |
4c2d0f6d | 1228 | ram_transferred_add(TARGET_PAGE_SIZE); |
aff3f660 | 1229 | stat64_add(&mig_stats.normal_pages, 1); |
65dacaa0 XG |
1230 | return 1; |
1231 | } | |
1232 | ||
56e93d26 | 1233 | /** |
3d0684b2 | 1234 | * ram_save_page: send the given page to the stream |
56e93d26 | 1235 | * |
3d0684b2 | 1236 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1237 | * < 0 - error |
1238 | * >=0 - Number of pages written - this might legally be 0 | |
1239 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1240 | * |
6f37bb8b | 1241 | * @rs: current RAM state |
56e93d26 JQ |
1242 | * @block: block that contains the page we want to send |
1243 | * @offset: offset inside the block for the page | |
56e93d26 | 1244 | */ |
05931ec5 | 1245 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss) |
56e93d26 JQ |
1246 | { |
1247 | int pages = -1; | |
56e93d26 | 1248 | uint8_t *p; |
56e93d26 | 1249 | bool send_async = true; |
a08f6890 | 1250 | RAMBlock *block = pss->block; |
8bba004c | 1251 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1252 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1253 | |
2f68e399 | 1254 | p = block->host + offset; |
1db9d8e5 | 1255 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1256 | |
56e93d26 | 1257 | XBZRLE_cache_lock(); |
f3095cc8 | 1258 | if (rs->xbzrle_started && !migration_in_postcopy()) { |
ec6f3ab9 | 1259 | pages = save_xbzrle_page(rs, pss, &p, current_addr, |
61717ea9 | 1260 | block, offset); |
05931ec5 | 1261 | if (!rs->last_stage) { |
059ff0fb XG |
1262 | /* Can't send this cached data async, since the cache page |
1263 | * might get updated before it gets to the wire | |
56e93d26 | 1264 | */ |
059ff0fb | 1265 | send_async = false; |
56e93d26 JQ |
1266 | } |
1267 | } | |
1268 | ||
1269 | /* XBZRLE overflow or normal page */ | |
1270 | if (pages == -1) { | |
ec6f3ab9 | 1271 | pages = save_normal_page(pss, block, offset, p, send_async); |
56e93d26 JQ |
1272 | } |
1273 | ||
1274 | XBZRLE_cache_unlock(); | |
1275 | ||
1276 | return pages; | |
1277 | } | |
1278 | ||
61717ea9 | 1279 | static int ram_save_multifd_page(QEMUFile *file, RAMBlock *block, |
b9ee2f7d JQ |
1280 | ram_addr_t offset) |
1281 | { | |
61717ea9 | 1282 | if (multifd_queue_page(file, block, offset) < 0) { |
713f762a IR |
1283 | return -1; |
1284 | } | |
aff3f660 | 1285 | stat64_add(&mig_stats.normal_pages, 1); |
b9ee2f7d JQ |
1286 | |
1287 | return 1; | |
1288 | } | |
1289 | ||
5e5fdcff XG |
1290 | static void |
1291 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1292 | { | |
4c2d0f6d | 1293 | ram_transferred_add(bytes_xmit); |
76e03000 | 1294 | |
97274a87 | 1295 | if (param->result == RES_ZEROPAGE) { |
aff3f660 | 1296 | stat64_add(&mig_stats.zero_pages, 1); |
76e03000 | 1297 | return; |
5e5fdcff | 1298 | } |
76e03000 XG |
1299 | |
1300 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1301 | compression_counters.compressed_size += bytes_xmit - 8; | |
1302 | compression_counters.pages++; | |
56e93d26 JQ |
1303 | } |
1304 | ||
32b05495 XG |
1305 | static bool save_page_use_compression(RAMState *rs); |
1306 | ||
3e81763e | 1307 | static int send_queued_data(CompressParam *param) |
56e93d26 | 1308 | { |
3e81763e | 1309 | PageSearchStatus *pss = &ram_state->pss[RAM_CHANNEL_PRECOPY]; |
eaa238ab | 1310 | MigrationState *ms = migrate_get_current(); |
3e81763e LS |
1311 | QEMUFile *file = ms->to_dst_file; |
1312 | int len = 0; | |
1313 | ||
1314 | RAMBlock *block = param->block; | |
1315 | ram_addr_t offset = param->offset; | |
1316 | ||
1317 | if (param->result == RES_NONE) { | |
1318 | return 0; | |
1319 | } | |
1320 | ||
1321 | assert(block == pss->last_sent_block); | |
1322 | ||
1323 | if (param->result == RES_ZEROPAGE) { | |
4024cc85 | 1324 | assert(qemu_file_buffer_empty(param->file)); |
3e81763e LS |
1325 | len += save_page_header(pss, file, block, offset | RAM_SAVE_FLAG_ZERO); |
1326 | qemu_put_byte(file, 0); | |
1327 | len += 1; | |
1328 | ram_release_page(block->idstr, offset); | |
1329 | } else if (param->result == RES_COMPRESS) { | |
4024cc85 | 1330 | assert(!qemu_file_buffer_empty(param->file)); |
3e81763e LS |
1331 | len += save_page_header(pss, file, block, |
1332 | offset | RAM_SAVE_FLAG_COMPRESS_PAGE); | |
1333 | len += qemu_put_qemu_file(file, param->file); | |
1334 | } else { | |
1335 | abort(); | |
1336 | } | |
1337 | ||
680628d2 LS |
1338 | update_compress_thread_counts(param, len); |
1339 | ||
3e81763e LS |
1340 | return len; |
1341 | } | |
1342 | ||
ef4f5f5d LS |
1343 | static void ram_flush_compressed_data(RAMState *rs) |
1344 | { | |
1345 | if (!save_page_use_compression(rs)) { | |
1346 | return; | |
1347 | } | |
1348 | ||
1349 | flush_compressed_data(send_queued_data); | |
1350 | } | |
1351 | ||
31e2ac74 JQ |
1352 | #define PAGE_ALL_CLEAN 0 |
1353 | #define PAGE_TRY_AGAIN 1 | |
1354 | #define PAGE_DIRTY_FOUND 2 | |
3d0684b2 JQ |
1355 | /** |
1356 | * find_dirty_block: find the next dirty page and update any state | |
1357 | * associated with the search process. | |
b9e60928 | 1358 | * |
31e2ac74 | 1359 | * Returns: |
294e5a40 | 1360 | * <0: An error happened |
31e2ac74 JQ |
1361 | * PAGE_ALL_CLEAN: no dirty page found, give up |
1362 | * PAGE_TRY_AGAIN: no dirty page found, retry for next block | |
1363 | * PAGE_DIRTY_FOUND: dirty page found | |
b9e60928 | 1364 | * |
6f37bb8b | 1365 | * @rs: current RAM state |
3d0684b2 JQ |
1366 | * @pss: data about the state of the current dirty page scan |
1367 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1368 | */ |
31e2ac74 | 1369 | static int find_dirty_block(RAMState *rs, PageSearchStatus *pss) |
b9e60928 | 1370 | { |
d9e474ea PX |
1371 | /* Update pss->page for the next dirty bit in ramblock */ |
1372 | pss_find_next_dirty(pss); | |
1373 | ||
6f37bb8b | 1374 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1375 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1376 | /* |
1377 | * We've been once around the RAM and haven't found anything. | |
1378 | * Give up. | |
1379 | */ | |
31e2ac74 | 1380 | return PAGE_ALL_CLEAN; |
b9e60928 | 1381 | } |
542147f4 DH |
1382 | if (!offset_in_ramblock(pss->block, |
1383 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)) { | |
b9e60928 | 1384 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1385 | pss->page = 0; |
b9e60928 DDAG |
1386 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1387 | if (!pss->block) { | |
294e5a40 JQ |
1388 | if (!migrate_multifd_flush_after_each_section()) { |
1389 | QEMUFile *f = rs->pss[RAM_CHANNEL_PRECOPY].pss_channel; | |
1390 | int ret = multifd_send_sync_main(f); | |
1391 | if (ret < 0) { | |
1392 | return ret; | |
1393 | } | |
1394 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); | |
1395 | qemu_fflush(f); | |
1396 | } | |
48df9d80 XG |
1397 | /* |
1398 | * If memory migration starts over, we will meet a dirtied page | |
1399 | * which may still exists in compression threads's ring, so we | |
1400 | * should flush the compressed data to make sure the new page | |
1401 | * is not overwritten by the old one in the destination. | |
1402 | * | |
1403 | * Also If xbzrle is on, stop using the data compression at this | |
1404 | * point. In theory, xbzrle can do better than compression. | |
1405 | */ | |
ef4f5f5d | 1406 | ram_flush_compressed_data(rs); |
48df9d80 | 1407 | |
b9e60928 DDAG |
1408 | /* Hit the end of the list */ |
1409 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1410 | /* Flag that we've looped */ | |
1411 | pss->complete_round = true; | |
1a373522 | 1412 | /* After the first round, enable XBZRLE. */ |
87dca0c9 | 1413 | if (migrate_xbzrle()) { |
f3095cc8 | 1414 | rs->xbzrle_started = true; |
1a373522 | 1415 | } |
b9e60928 DDAG |
1416 | } |
1417 | /* Didn't find anything this time, but try again on the new block */ | |
31e2ac74 | 1418 | return PAGE_TRY_AGAIN; |
b9e60928 | 1419 | } else { |
31e2ac74 JQ |
1420 | /* We've found something */ |
1421 | return PAGE_DIRTY_FOUND; | |
b9e60928 DDAG |
1422 | } |
1423 | } | |
1424 | ||
3d0684b2 JQ |
1425 | /** |
1426 | * unqueue_page: gets a page of the queue | |
1427 | * | |
a82d593b | 1428 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1429 | * |
3d0684b2 JQ |
1430 | * Returns the block of the page (or NULL if none available) |
1431 | * | |
ec481c6c | 1432 | * @rs: current RAM state |
3d0684b2 | 1433 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1434 | */ |
f20e2865 | 1435 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b | 1436 | { |
a1fe28df | 1437 | struct RAMSrcPageRequest *entry; |
a82d593b DDAG |
1438 | RAMBlock *block = NULL; |
1439 | ||
a1fe28df | 1440 | if (!postcopy_has_request(rs)) { |
ae526e32 XG |
1441 | return NULL; |
1442 | } | |
1443 | ||
6e8a355d | 1444 | QEMU_LOCK_GUARD(&rs->src_page_req_mutex); |
a1fe28df PX |
1445 | |
1446 | /* | |
1447 | * This should _never_ change even after we take the lock, because no one | |
1448 | * should be taking anything off the request list other than us. | |
1449 | */ | |
1450 | assert(postcopy_has_request(rs)); | |
1451 | ||
1452 | entry = QSIMPLEQ_FIRST(&rs->src_page_requests); | |
1453 | block = entry->rb; | |
1454 | *offset = entry->offset; | |
1455 | ||
777f53c7 TH |
1456 | if (entry->len > TARGET_PAGE_SIZE) { |
1457 | entry->len -= TARGET_PAGE_SIZE; | |
1458 | entry->offset += TARGET_PAGE_SIZE; | |
a1fe28df PX |
1459 | } else { |
1460 | memory_region_unref(block->mr); | |
1461 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); | |
1462 | g_free(entry); | |
1463 | migration_consume_urgent_request(); | |
a82d593b | 1464 | } |
a82d593b DDAG |
1465 | |
1466 | return block; | |
1467 | } | |
1468 | ||
278e2f55 AG |
1469 | #if defined(__linux__) |
1470 | /** | |
1471 | * poll_fault_page: try to get next UFFD write fault page and, if pending fault | |
1472 | * is found, return RAM block pointer and page offset | |
1473 | * | |
1474 | * Returns pointer to the RAMBlock containing faulting page, | |
1475 | * NULL if no write faults are pending | |
1476 | * | |
1477 | * @rs: current RAM state | |
1478 | * @offset: page offset from the beginning of the block | |
1479 | */ | |
1480 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1481 | { | |
1482 | struct uffd_msg uffd_msg; | |
1483 | void *page_address; | |
82ea3e3b | 1484 | RAMBlock *block; |
278e2f55 AG |
1485 | int res; |
1486 | ||
1487 | if (!migrate_background_snapshot()) { | |
1488 | return NULL; | |
1489 | } | |
1490 | ||
1491 | res = uffd_read_events(rs->uffdio_fd, &uffd_msg, 1); | |
1492 | if (res <= 0) { | |
1493 | return NULL; | |
1494 | } | |
1495 | ||
1496 | page_address = (void *)(uintptr_t) uffd_msg.arg.pagefault.address; | |
82ea3e3b AG |
1497 | block = qemu_ram_block_from_host(page_address, false, offset); |
1498 | assert(block && (block->flags & RAM_UF_WRITEPROTECT) != 0); | |
1499 | return block; | |
278e2f55 AG |
1500 | } |
1501 | ||
1502 | /** | |
1503 | * ram_save_release_protection: release UFFD write protection after | |
1504 | * a range of pages has been saved | |
1505 | * | |
1506 | * @rs: current RAM state | |
1507 | * @pss: page-search-status structure | |
1508 | * @start_page: index of the first page in the range relative to pss->block | |
1509 | * | |
1510 | * Returns 0 on success, negative value in case of an error | |
1511 | */ | |
1512 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1513 | unsigned long start_page) | |
1514 | { | |
1515 | int res = 0; | |
1516 | ||
1517 | /* Check if page is from UFFD-managed region. */ | |
1518 | if (pss->block->flags & RAM_UF_WRITEPROTECT) { | |
1519 | void *page_address = pss->block->host + (start_page << TARGET_PAGE_BITS); | |
258f5c98 | 1520 | uint64_t run_length = (pss->page - start_page) << TARGET_PAGE_BITS; |
278e2f55 AG |
1521 | |
1522 | /* Flush async buffers before un-protect. */ | |
61717ea9 | 1523 | qemu_fflush(pss->pss_channel); |
278e2f55 AG |
1524 | /* Un-protect memory range. */ |
1525 | res = uffd_change_protection(rs->uffdio_fd, page_address, run_length, | |
1526 | false, false); | |
1527 | } | |
1528 | ||
1529 | return res; | |
1530 | } | |
1531 | ||
1532 | /* ram_write_tracking_available: check if kernel supports required UFFD features | |
1533 | * | |
1534 | * Returns true if supports, false otherwise | |
1535 | */ | |
1536 | bool ram_write_tracking_available(void) | |
1537 | { | |
1538 | uint64_t uffd_features; | |
1539 | int res; | |
1540 | ||
1541 | res = uffd_query_features(&uffd_features); | |
1542 | return (res == 0 && | |
1543 | (uffd_features & UFFD_FEATURE_PAGEFAULT_FLAG_WP) != 0); | |
1544 | } | |
1545 | ||
1546 | /* ram_write_tracking_compatible: check if guest configuration is | |
1547 | * compatible with 'write-tracking' | |
1548 | * | |
1549 | * Returns true if compatible, false otherwise | |
1550 | */ | |
1551 | bool ram_write_tracking_compatible(void) | |
1552 | { | |
1553 | const uint64_t uffd_ioctls_mask = BIT(_UFFDIO_WRITEPROTECT); | |
1554 | int uffd_fd; | |
82ea3e3b | 1555 | RAMBlock *block; |
278e2f55 AG |
1556 | bool ret = false; |
1557 | ||
1558 | /* Open UFFD file descriptor */ | |
1559 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, false); | |
1560 | if (uffd_fd < 0) { | |
1561 | return false; | |
1562 | } | |
1563 | ||
1564 | RCU_READ_LOCK_GUARD(); | |
1565 | ||
82ea3e3b | 1566 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 AG |
1567 | uint64_t uffd_ioctls; |
1568 | ||
1569 | /* Nothing to do with read-only and MMIO-writable regions */ | |
82ea3e3b | 1570 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1571 | continue; |
1572 | } | |
1573 | /* Try to register block memory via UFFD-IO to track writes */ | |
82ea3e3b | 1574 | if (uffd_register_memory(uffd_fd, block->host, block->max_length, |
278e2f55 AG |
1575 | UFFDIO_REGISTER_MODE_WP, &uffd_ioctls)) { |
1576 | goto out; | |
1577 | } | |
1578 | if ((uffd_ioctls & uffd_ioctls_mask) != uffd_ioctls_mask) { | |
1579 | goto out; | |
1580 | } | |
1581 | } | |
1582 | ret = true; | |
1583 | ||
1584 | out: | |
1585 | uffd_close_fd(uffd_fd); | |
1586 | return ret; | |
1587 | } | |
1588 | ||
f7b9dcfb DH |
1589 | static inline void populate_read_range(RAMBlock *block, ram_addr_t offset, |
1590 | ram_addr_t size) | |
1591 | { | |
5f19a449 DH |
1592 | const ram_addr_t end = offset + size; |
1593 | ||
f7b9dcfb DH |
1594 | /* |
1595 | * We read one byte of each page; this will preallocate page tables if | |
1596 | * required and populate the shared zeropage on MAP_PRIVATE anonymous memory | |
1597 | * where no page was populated yet. This might require adaption when | |
1598 | * supporting other mappings, like shmem. | |
1599 | */ | |
5f19a449 | 1600 | for (; offset < end; offset += block->page_size) { |
f7b9dcfb DH |
1601 | char tmp = *((char *)block->host + offset); |
1602 | ||
1603 | /* Don't optimize the read out */ | |
1604 | asm volatile("" : "+r" (tmp)); | |
1605 | } | |
1606 | } | |
1607 | ||
6fee3a1f DH |
1608 | static inline int populate_read_section(MemoryRegionSection *section, |
1609 | void *opaque) | |
1610 | { | |
1611 | const hwaddr size = int128_get64(section->size); | |
1612 | hwaddr offset = section->offset_within_region; | |
1613 | RAMBlock *block = section->mr->ram_block; | |
1614 | ||
1615 | populate_read_range(block, offset, size); | |
1616 | return 0; | |
1617 | } | |
1618 | ||
eeccb99c | 1619 | /* |
f7b9dcfb DH |
1620 | * ram_block_populate_read: preallocate page tables and populate pages in the |
1621 | * RAM block by reading a byte of each page. | |
eeccb99c AG |
1622 | * |
1623 | * Since it's solely used for userfault_fd WP feature, here we just | |
1624 | * hardcode page size to qemu_real_host_page_size. | |
1625 | * | |
82ea3e3b | 1626 | * @block: RAM block to populate |
eeccb99c | 1627 | */ |
6fee3a1f | 1628 | static void ram_block_populate_read(RAMBlock *rb) |
eeccb99c | 1629 | { |
6fee3a1f DH |
1630 | /* |
1631 | * Skip populating all pages that fall into a discarded range as managed by | |
1632 | * a RamDiscardManager responsible for the mapped memory region of the | |
1633 | * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock | |
1634 | * must not get populated automatically. We don't have to track | |
1635 | * modifications via userfaultfd WP reliably, because these pages will | |
1636 | * not be part of the migration stream either way -- see | |
1637 | * ramblock_dirty_bitmap_exclude_discarded_pages(). | |
1638 | * | |
1639 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1640 | */ | |
1641 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1642 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1643 | MemoryRegionSection section = { | |
1644 | .mr = rb->mr, | |
1645 | .offset_within_region = 0, | |
1646 | .size = rb->mr->size, | |
1647 | }; | |
1648 | ||
1649 | ram_discard_manager_replay_populated(rdm, §ion, | |
1650 | populate_read_section, NULL); | |
1651 | } else { | |
1652 | populate_read_range(rb, 0, rb->used_length); | |
1653 | } | |
eeccb99c AG |
1654 | } |
1655 | ||
1656 | /* | |
1657 | * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking | |
1658 | */ | |
1659 | void ram_write_tracking_prepare(void) | |
1660 | { | |
82ea3e3b | 1661 | RAMBlock *block; |
eeccb99c AG |
1662 | |
1663 | RCU_READ_LOCK_GUARD(); | |
1664 | ||
82ea3e3b | 1665 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
eeccb99c | 1666 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1667 | if (block->mr->readonly || block->mr->rom_device) { |
eeccb99c AG |
1668 | continue; |
1669 | } | |
1670 | ||
1671 | /* | |
1672 | * Populate pages of the RAM block before enabling userfault_fd | |
1673 | * write protection. | |
1674 | * | |
1675 | * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with | |
1676 | * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip | |
1677 | * pages with pte_none() entries in page table. | |
1678 | */ | |
f7b9dcfb | 1679 | ram_block_populate_read(block); |
eeccb99c AG |
1680 | } |
1681 | } | |
1682 | ||
e41c5770 DH |
1683 | static inline int uffd_protect_section(MemoryRegionSection *section, |
1684 | void *opaque) | |
1685 | { | |
1686 | const hwaddr size = int128_get64(section->size); | |
1687 | const hwaddr offset = section->offset_within_region; | |
1688 | RAMBlock *rb = section->mr->ram_block; | |
1689 | int uffd_fd = (uintptr_t)opaque; | |
1690 | ||
1691 | return uffd_change_protection(uffd_fd, rb->host + offset, size, true, | |
1692 | false); | |
1693 | } | |
1694 | ||
1695 | static int ram_block_uffd_protect(RAMBlock *rb, int uffd_fd) | |
1696 | { | |
1697 | assert(rb->flags & RAM_UF_WRITEPROTECT); | |
1698 | ||
1699 | /* See ram_block_populate_read() */ | |
1700 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1701 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1702 | MemoryRegionSection section = { | |
1703 | .mr = rb->mr, | |
1704 | .offset_within_region = 0, | |
1705 | .size = rb->mr->size, | |
1706 | }; | |
1707 | ||
1708 | return ram_discard_manager_replay_populated(rdm, §ion, | |
1709 | uffd_protect_section, | |
1710 | (void *)(uintptr_t)uffd_fd); | |
1711 | } | |
1712 | return uffd_change_protection(uffd_fd, rb->host, | |
1713 | rb->used_length, true, false); | |
1714 | } | |
1715 | ||
278e2f55 AG |
1716 | /* |
1717 | * ram_write_tracking_start: start UFFD-WP memory tracking | |
1718 | * | |
1719 | * Returns 0 for success or negative value in case of error | |
1720 | */ | |
1721 | int ram_write_tracking_start(void) | |
1722 | { | |
1723 | int uffd_fd; | |
1724 | RAMState *rs = ram_state; | |
82ea3e3b | 1725 | RAMBlock *block; |
278e2f55 AG |
1726 | |
1727 | /* Open UFFD file descriptor */ | |
1728 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, true); | |
1729 | if (uffd_fd < 0) { | |
1730 | return uffd_fd; | |
1731 | } | |
1732 | rs->uffdio_fd = uffd_fd; | |
1733 | ||
1734 | RCU_READ_LOCK_GUARD(); | |
1735 | ||
82ea3e3b | 1736 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 | 1737 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1738 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1739 | continue; |
1740 | } | |
1741 | ||
1742 | /* Register block memory with UFFD to track writes */ | |
82ea3e3b AG |
1743 | if (uffd_register_memory(rs->uffdio_fd, block->host, |
1744 | block->max_length, UFFDIO_REGISTER_MODE_WP, NULL)) { | |
278e2f55 AG |
1745 | goto fail; |
1746 | } | |
72ef3a37 DH |
1747 | block->flags |= RAM_UF_WRITEPROTECT; |
1748 | memory_region_ref(block->mr); | |
1749 | ||
278e2f55 | 1750 | /* Apply UFFD write protection to the block memory range */ |
e41c5770 | 1751 | if (ram_block_uffd_protect(block, uffd_fd)) { |
278e2f55 AG |
1752 | goto fail; |
1753 | } | |
278e2f55 | 1754 | |
82ea3e3b AG |
1755 | trace_ram_write_tracking_ramblock_start(block->idstr, block->page_size, |
1756 | block->host, block->max_length); | |
278e2f55 AG |
1757 | } |
1758 | ||
1759 | return 0; | |
1760 | ||
1761 | fail: | |
1762 | error_report("ram_write_tracking_start() failed: restoring initial memory state"); | |
1763 | ||
82ea3e3b AG |
1764 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1765 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1766 | continue; |
1767 | } | |
82ea3e3b | 1768 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 1769 | /* Cleanup flags and remove reference */ |
82ea3e3b AG |
1770 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1771 | memory_region_unref(block->mr); | |
278e2f55 AG |
1772 | } |
1773 | ||
1774 | uffd_close_fd(uffd_fd); | |
1775 | rs->uffdio_fd = -1; | |
1776 | return -1; | |
1777 | } | |
1778 | ||
1779 | /** | |
1780 | * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection | |
1781 | */ | |
1782 | void ram_write_tracking_stop(void) | |
1783 | { | |
1784 | RAMState *rs = ram_state; | |
82ea3e3b | 1785 | RAMBlock *block; |
278e2f55 AG |
1786 | |
1787 | RCU_READ_LOCK_GUARD(); | |
1788 | ||
82ea3e3b AG |
1789 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
1790 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
1791 | continue; |
1792 | } | |
82ea3e3b | 1793 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 1794 | |
82ea3e3b AG |
1795 | trace_ram_write_tracking_ramblock_stop(block->idstr, block->page_size, |
1796 | block->host, block->max_length); | |
278e2f55 AG |
1797 | |
1798 | /* Cleanup flags and remove reference */ | |
82ea3e3b AG |
1799 | block->flags &= ~RAM_UF_WRITEPROTECT; |
1800 | memory_region_unref(block->mr); | |
278e2f55 AG |
1801 | } |
1802 | ||
1803 | /* Finally close UFFD file descriptor */ | |
1804 | uffd_close_fd(rs->uffdio_fd); | |
1805 | rs->uffdio_fd = -1; | |
1806 | } | |
1807 | ||
1808 | #else | |
1809 | /* No target OS support, stubs just fail or ignore */ | |
1810 | ||
1811 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1812 | { | |
1813 | (void) rs; | |
1814 | (void) offset; | |
1815 | ||
1816 | return NULL; | |
1817 | } | |
1818 | ||
1819 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1820 | unsigned long start_page) | |
1821 | { | |
1822 | (void) rs; | |
1823 | (void) pss; | |
1824 | (void) start_page; | |
1825 | ||
1826 | return 0; | |
1827 | } | |
1828 | ||
1829 | bool ram_write_tracking_available(void) | |
1830 | { | |
1831 | return false; | |
1832 | } | |
1833 | ||
1834 | bool ram_write_tracking_compatible(void) | |
1835 | { | |
1836 | assert(0); | |
1837 | return false; | |
1838 | } | |
1839 | ||
1840 | int ram_write_tracking_start(void) | |
1841 | { | |
1842 | assert(0); | |
1843 | return -1; | |
1844 | } | |
1845 | ||
1846 | void ram_write_tracking_stop(void) | |
1847 | { | |
1848 | assert(0); | |
1849 | } | |
1850 | #endif /* defined(__linux__) */ | |
1851 | ||
3d0684b2 | 1852 | /** |
ff1543af | 1853 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
1854 | * |
1855 | * Skips pages that are already sent (!dirty) | |
a82d593b | 1856 | * |
a5f7b1a6 | 1857 | * Returns true if a queued page is found |
a82d593b | 1858 | * |
6f37bb8b | 1859 | * @rs: current RAM state |
3d0684b2 | 1860 | * @pss: data about the state of the current dirty page scan |
a82d593b | 1861 | */ |
f20e2865 | 1862 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
1863 | { |
1864 | RAMBlock *block; | |
1865 | ram_addr_t offset; | |
777f53c7 TH |
1866 | bool dirty; |
1867 | ||
1868 | do { | |
1869 | block = unqueue_page(rs, &offset); | |
1870 | /* | |
1871 | * We're sending this page, and since it's postcopy nothing else | |
1872 | * will dirty it, and we must make sure it doesn't get sent again | |
1873 | * even if this queue request was received after the background | |
1874 | * search already sent it. | |
1875 | */ | |
1876 | if (block) { | |
1877 | unsigned long page; | |
1878 | ||
1879 | page = offset >> TARGET_PAGE_BITS; | |
1880 | dirty = test_bit(page, block->bmap); | |
1881 | if (!dirty) { | |
1882 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, | |
1883 | page); | |
1884 | } else { | |
1885 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); | |
1886 | } | |
1887 | } | |
a82d593b | 1888 | |
777f53c7 | 1889 | } while (block && !dirty); |
a82d593b | 1890 | |
b062106d | 1891 | if (!block) { |
278e2f55 AG |
1892 | /* |
1893 | * Poll write faults too if background snapshot is enabled; that's | |
1894 | * when we have vcpus got blocked by the write protected pages. | |
1895 | */ | |
1896 | block = poll_fault_page(rs, &offset); | |
1897 | } | |
1898 | ||
a82d593b | 1899 | if (block) { |
a82d593b DDAG |
1900 | /* |
1901 | * We want the background search to continue from the queued page | |
1902 | * since the guest is likely to want other pages near to the page | |
1903 | * it just requested. | |
1904 | */ | |
1905 | pss->block = block; | |
a935e30f | 1906 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
1907 | |
1908 | /* | |
1909 | * This unqueued page would break the "one round" check, even is | |
1910 | * really rare. | |
1911 | */ | |
1912 | pss->complete_round = false; | |
a82d593b DDAG |
1913 | } |
1914 | ||
1915 | return !!block; | |
1916 | } | |
1917 | ||
6c595cde | 1918 | /** |
5e58f968 JQ |
1919 | * migration_page_queue_free: drop any remaining pages in the ram |
1920 | * request queue | |
6c595cde | 1921 | * |
3d0684b2 JQ |
1922 | * It should be empty at the end anyway, but in error cases there may |
1923 | * be some left. in case that there is any page left, we drop it. | |
1924 | * | |
6c595cde | 1925 | */ |
83c13382 | 1926 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 1927 | { |
ec481c6c | 1928 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
1929 | /* This queue generally should be empty - but in the case of a failed |
1930 | * migration might have some droppings in. | |
1931 | */ | |
89ac5a1d | 1932 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 1933 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 1934 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 1935 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
1936 | g_free(mspr); |
1937 | } | |
6c595cde DDAG |
1938 | } |
1939 | ||
1940 | /** | |
3d0684b2 JQ |
1941 | * ram_save_queue_pages: queue the page for transmission |
1942 | * | |
1943 | * A request from postcopy destination for example. | |
1944 | * | |
1945 | * Returns zero on success or negative on error | |
1946 | * | |
3d0684b2 JQ |
1947 | * @rbname: Name of the RAMBLock of the request. NULL means the |
1948 | * same that last one. | |
1949 | * @start: starting address from the start of the RAMBlock | |
1950 | * @len: length (in bytes) to send | |
6c595cde | 1951 | */ |
96506894 | 1952 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
1953 | { |
1954 | RAMBlock *ramblock; | |
53518d94 | 1955 | RAMState *rs = ram_state; |
6c595cde | 1956 | |
aff3f660 | 1957 | stat64_add(&mig_stats.postcopy_requests, 1); |
89ac5a1d DDAG |
1958 | RCU_READ_LOCK_GUARD(); |
1959 | ||
6c595cde DDAG |
1960 | if (!rbname) { |
1961 | /* Reuse last RAMBlock */ | |
68a098f3 | 1962 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
1963 | |
1964 | if (!ramblock) { | |
1965 | /* | |
1966 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
1967 | * it's the 1st request. | |
1968 | */ | |
1969 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 1970 | return -1; |
6c595cde DDAG |
1971 | } |
1972 | } else { | |
1973 | ramblock = qemu_ram_block_by_name(rbname); | |
1974 | ||
1975 | if (!ramblock) { | |
1976 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
1977 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 1978 | return -1; |
6c595cde | 1979 | } |
68a098f3 | 1980 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
1981 | } |
1982 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
542147f4 | 1983 | if (!offset_in_ramblock(ramblock, start + len - 1)) { |
9458ad6b JQ |
1984 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
1985 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 1986 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 1987 | return -1; |
6c595cde DDAG |
1988 | } |
1989 | ||
93589827 PX |
1990 | /* |
1991 | * When with postcopy preempt, we send back the page directly in the | |
1992 | * rp-return thread. | |
1993 | */ | |
1994 | if (postcopy_preempt_active()) { | |
1995 | ram_addr_t page_start = start >> TARGET_PAGE_BITS; | |
1996 | size_t page_size = qemu_ram_pagesize(ramblock); | |
1997 | PageSearchStatus *pss = &ram_state->pss[RAM_CHANNEL_POSTCOPY]; | |
1998 | int ret = 0; | |
1999 | ||
2000 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2001 | ||
2002 | pss_init(pss, ramblock, page_start); | |
2003 | /* | |
2004 | * Always use the preempt channel, and make sure it's there. It's | |
2005 | * safe to access without lock, because when rp-thread is running | |
2006 | * we should be the only one who operates on the qemufile | |
2007 | */ | |
2008 | pss->pss_channel = migrate_get_current()->postcopy_qemufile_src; | |
93589827 PX |
2009 | assert(pss->pss_channel); |
2010 | ||
2011 | /* | |
2012 | * It must be either one or multiple of host page size. Just | |
2013 | * assert; if something wrong we're mostly split brain anyway. | |
2014 | */ | |
2015 | assert(len % page_size == 0); | |
2016 | while (len) { | |
2017 | if (ram_save_host_page_urgent(pss)) { | |
2018 | error_report("%s: ram_save_host_page_urgent() failed: " | |
2019 | "ramblock=%s, start_addr=0x"RAM_ADDR_FMT, | |
2020 | __func__, ramblock->idstr, start); | |
2021 | ret = -1; | |
2022 | break; | |
2023 | } | |
2024 | /* | |
2025 | * NOTE: after ram_save_host_page_urgent() succeeded, pss->page | |
2026 | * will automatically be moved and point to the next host page | |
2027 | * we're going to send, so no need to update here. | |
2028 | * | |
2029 | * Normally QEMU never sends >1 host page in requests, so | |
2030 | * logically we don't even need that as the loop should only | |
2031 | * run once, but just to be consistent. | |
2032 | */ | |
2033 | len -= page_size; | |
2034 | }; | |
2035 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
2036 | ||
2037 | return ret; | |
2038 | } | |
2039 | ||
ec481c6c | 2040 | struct RAMSrcPageRequest *new_entry = |
b21e2380 | 2041 | g_new0(struct RAMSrcPageRequest, 1); |
6c595cde DDAG |
2042 | new_entry->rb = ramblock; |
2043 | new_entry->offset = start; | |
2044 | new_entry->len = len; | |
2045 | ||
2046 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
2047 | qemu_mutex_lock(&rs->src_page_req_mutex); |
2048 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 2049 | migration_make_urgent_request(); |
ec481c6c | 2050 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
2051 | |
2052 | return 0; | |
6c595cde DDAG |
2053 | } |
2054 | ||
d7400a34 XG |
2055 | static bool save_page_use_compression(RAMState *rs) |
2056 | { | |
a7a94d14 | 2057 | if (!migrate_compress()) { |
d7400a34 XG |
2058 | return false; |
2059 | } | |
2060 | ||
2061 | /* | |
1a373522 DH |
2062 | * If xbzrle is enabled (e.g., after first round of migration), stop |
2063 | * using the data compression. In theory, xbzrle can do better than | |
2064 | * compression. | |
d7400a34 | 2065 | */ |
f3095cc8 | 2066 | if (rs->xbzrle_started) { |
1a373522 | 2067 | return false; |
d7400a34 XG |
2068 | } |
2069 | ||
1a373522 | 2070 | return true; |
d7400a34 XG |
2071 | } |
2072 | ||
5e5fdcff XG |
2073 | /* |
2074 | * try to compress the page before posting it out, return true if the page | |
2075 | * has been properly handled by compression, otherwise needs other | |
2076 | * paths to handle it | |
2077 | */ | |
ec6f3ab9 PX |
2078 | static bool save_compress_page(RAMState *rs, PageSearchStatus *pss, |
2079 | RAMBlock *block, ram_addr_t offset) | |
5e5fdcff XG |
2080 | { |
2081 | if (!save_page_use_compression(rs)) { | |
2082 | return false; | |
2083 | } | |
2084 | ||
2085 | /* | |
2086 | * When starting the process of a new block, the first page of | |
2087 | * the block should be sent out before other pages in the same | |
2088 | * block, and all the pages in last block should have been sent | |
2089 | * out, keeping this order is important, because the 'cont' flag | |
2090 | * is used to avoid resending the block name. | |
2091 | * | |
2092 | * We post the fist page as normal page as compression will take | |
2093 | * much CPU resource. | |
2094 | */ | |
ec6f3ab9 | 2095 | if (block != pss->last_sent_block) { |
ef4f5f5d | 2096 | ram_flush_compressed_data(rs); |
5e5fdcff XG |
2097 | return false; |
2098 | } | |
2099 | ||
ef4f5f5d | 2100 | if (compress_page_with_multi_thread(block, offset, send_queued_data) > 0) { |
5e5fdcff XG |
2101 | return true; |
2102 | } | |
2103 | ||
76e03000 | 2104 | compression_counters.busy++; |
5e5fdcff XG |
2105 | return false; |
2106 | } | |
2107 | ||
a82d593b | 2108 | /** |
4010ba38 | 2109 | * ram_save_target_page_legacy: save one target page |
a82d593b | 2110 | * |
3d0684b2 | 2111 | * Returns the number of pages written |
a82d593b | 2112 | * |
6f37bb8b | 2113 | * @rs: current RAM state |
3d0684b2 | 2114 | * @pss: data about the page we want to send |
a82d593b | 2115 | */ |
4010ba38 | 2116 | static int ram_save_target_page_legacy(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2117 | { |
a8ec91f9 | 2118 | RAMBlock *block = pss->block; |
8bba004c | 2119 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
2120 | int res; |
2121 | ||
61717ea9 | 2122 | if (control_save_page(pss, block, offset, &res)) { |
a8ec91f9 XG |
2123 | return res; |
2124 | } | |
2125 | ||
ec6f3ab9 | 2126 | if (save_compress_page(rs, pss, block, offset)) { |
5e5fdcff | 2127 | return 1; |
d7400a34 XG |
2128 | } |
2129 | ||
37502df3 | 2130 | res = save_zero_page(pss, pss->pss_channel, block, offset); |
d7400a34 XG |
2131 | if (res > 0) { |
2132 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
2133 | * page would be stale | |
2134 | */ | |
f3095cc8 | 2135 | if (rs->xbzrle_started) { |
d7400a34 XG |
2136 | XBZRLE_cache_lock(); |
2137 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
2138 | XBZRLE_cache_unlock(); | |
2139 | } | |
d7400a34 XG |
2140 | return res; |
2141 | } | |
2142 | ||
da3f56cb | 2143 | /* |
6f39c90b PX |
2144 | * Do not use multifd in postcopy as one whole host page should be |
2145 | * placed. Meanwhile postcopy requires atomic update of pages, so even | |
2146 | * if host page size == guest page size the dest guest during run may | |
2147 | * still see partially copied pages which is data corruption. | |
da3f56cb | 2148 | */ |
51b07548 | 2149 | if (migrate_multifd() && !migration_in_postcopy()) { |
61717ea9 | 2150 | return ram_save_multifd_page(pss->pss_channel, block, offset); |
a82d593b DDAG |
2151 | } |
2152 | ||
05931ec5 | 2153 | return ram_save_page(rs, pss); |
a82d593b DDAG |
2154 | } |
2155 | ||
d9e474ea PX |
2156 | /* Should be called before sending a host page */ |
2157 | static void pss_host_page_prepare(PageSearchStatus *pss) | |
2158 | { | |
2159 | /* How many guest pages are there in one host page? */ | |
2160 | size_t guest_pfns = qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
2161 | ||
2162 | pss->host_page_sending = true; | |
301d7ffe PX |
2163 | if (guest_pfns <= 1) { |
2164 | /* | |
2165 | * This covers both when guest psize == host psize, or when guest | |
2166 | * has larger psize than the host (guest_pfns==0). | |
2167 | * | |
2168 | * For the latter, we always send one whole guest page per | |
2169 | * iteration of the host page (example: an Alpha VM on x86 host | |
2170 | * will have guest psize 8K while host psize 4K). | |
2171 | */ | |
2172 | pss->host_page_start = pss->page; | |
2173 | pss->host_page_end = pss->page + 1; | |
2174 | } else { | |
2175 | /* | |
2176 | * The host page spans over multiple guest pages, we send them | |
2177 | * within the same host page iteration. | |
2178 | */ | |
2179 | pss->host_page_start = ROUND_DOWN(pss->page, guest_pfns); | |
2180 | pss->host_page_end = ROUND_UP(pss->page + 1, guest_pfns); | |
2181 | } | |
d9e474ea PX |
2182 | } |
2183 | ||
2184 | /* | |
2185 | * Whether the page pointed by PSS is within the host page being sent. | |
2186 | * Must be called after a previous pss_host_page_prepare(). | |
2187 | */ | |
2188 | static bool pss_within_range(PageSearchStatus *pss) | |
2189 | { | |
2190 | ram_addr_t ram_addr; | |
2191 | ||
2192 | assert(pss->host_page_sending); | |
2193 | ||
2194 | /* Over host-page boundary? */ | |
2195 | if (pss->page >= pss->host_page_end) { | |
2196 | return false; | |
2197 | } | |
2198 | ||
2199 | ram_addr = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; | |
2200 | ||
2201 | return offset_in_ramblock(pss->block, ram_addr); | |
2202 | } | |
2203 | ||
2204 | static void pss_host_page_finish(PageSearchStatus *pss) | |
2205 | { | |
2206 | pss->host_page_sending = false; | |
2207 | /* This is not needed, but just to reset it */ | |
2208 | pss->host_page_start = pss->host_page_end = 0; | |
2209 | } | |
2210 | ||
93589827 PX |
2211 | /* |
2212 | * Send an urgent host page specified by `pss'. Need to be called with | |
2213 | * bitmap_mutex held. | |
2214 | * | |
2215 | * Returns 0 if save host page succeeded, false otherwise. | |
2216 | */ | |
2217 | static int ram_save_host_page_urgent(PageSearchStatus *pss) | |
2218 | { | |
2219 | bool page_dirty, sent = false; | |
2220 | RAMState *rs = ram_state; | |
2221 | int ret = 0; | |
2222 | ||
2223 | trace_postcopy_preempt_send_host_page(pss->block->idstr, pss->page); | |
2224 | pss_host_page_prepare(pss); | |
2225 | ||
2226 | /* | |
2227 | * If precopy is sending the same page, let it be done in precopy, or | |
2228 | * we could send the same page in two channels and none of them will | |
2229 | * receive the whole page. | |
2230 | */ | |
2231 | if (pss_overlap(pss, &ram_state->pss[RAM_CHANNEL_PRECOPY])) { | |
2232 | trace_postcopy_preempt_hit(pss->block->idstr, | |
2233 | pss->page << TARGET_PAGE_BITS); | |
2234 | return 0; | |
2235 | } | |
2236 | ||
2237 | do { | |
2238 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); | |
2239 | ||
2240 | if (page_dirty) { | |
2241 | /* Be strict to return code; it must be 1, or what else? */ | |
4010ba38 | 2242 | if (migration_ops->ram_save_target_page(rs, pss) != 1) { |
93589827 PX |
2243 | error_report_once("%s: ram_save_target_page failed", __func__); |
2244 | ret = -1; | |
2245 | goto out; | |
2246 | } | |
2247 | sent = true; | |
2248 | } | |
2249 | pss_find_next_dirty(pss); | |
2250 | } while (pss_within_range(pss)); | |
2251 | out: | |
2252 | pss_host_page_finish(pss); | |
2253 | /* For urgent requests, flush immediately if sent */ | |
2254 | if (sent) { | |
2255 | qemu_fflush(pss->pss_channel); | |
2256 | } | |
2257 | return ret; | |
2258 | } | |
2259 | ||
a82d593b | 2260 | /** |
3d0684b2 | 2261 | * ram_save_host_page: save a whole host page |
a82d593b | 2262 | * |
3d0684b2 JQ |
2263 | * Starting at *offset send pages up to the end of the current host |
2264 | * page. It's valid for the initial offset to point into the middle of | |
2265 | * a host page in which case the remainder of the hostpage is sent. | |
2266 | * Only dirty target pages are sent. Note that the host page size may | |
2267 | * be a huge page for this block. | |
f3321554 | 2268 | * |
1eb3fc0a DDAG |
2269 | * The saving stops at the boundary of the used_length of the block |
2270 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 2271 | * |
f3321554 PX |
2272 | * The caller must be with ram_state.bitmap_mutex held to call this |
2273 | * function. Note that this function can temporarily release the lock, but | |
2274 | * when the function is returned it'll make sure the lock is still held. | |
2275 | * | |
3d0684b2 JQ |
2276 | * Returns the number of pages written or negative on error |
2277 | * | |
6f37bb8b | 2278 | * @rs: current RAM state |
3d0684b2 | 2279 | * @pss: data about the page we want to send |
a82d593b | 2280 | */ |
05931ec5 | 2281 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2282 | { |
f3321554 | 2283 | bool page_dirty, preempt_active = postcopy_preempt_active(); |
a82d593b | 2284 | int tmppages, pages = 0; |
a935e30f JQ |
2285 | size_t pagesize_bits = |
2286 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
278e2f55 AG |
2287 | unsigned long start_page = pss->page; |
2288 | int res; | |
4c011c37 | 2289 | |
fbd162e6 | 2290 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
2291 | error_report("block %s should not be migrated !", pss->block->idstr); |
2292 | return 0; | |
2293 | } | |
2294 | ||
d9e474ea PX |
2295 | /* Update host page boundary information */ |
2296 | pss_host_page_prepare(pss); | |
2297 | ||
a82d593b | 2298 | do { |
f3321554 | 2299 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); |
a82d593b | 2300 | |
f3321554 PX |
2301 | /* Check the pages is dirty and if it is send it */ |
2302 | if (page_dirty) { | |
ba1b7c81 | 2303 | /* |
f3321554 PX |
2304 | * Properly yield the lock only in postcopy preempt mode |
2305 | * because both migration thread and rp-return thread can | |
2306 | * operate on the bitmaps. | |
ba1b7c81 | 2307 | */ |
f3321554 PX |
2308 | if (preempt_active) { |
2309 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
ba1b7c81 | 2310 | } |
4010ba38 | 2311 | tmppages = migration_ops->ram_save_target_page(rs, pss); |
f3321554 PX |
2312 | if (tmppages >= 0) { |
2313 | pages += tmppages; | |
2314 | /* | |
2315 | * Allow rate limiting to happen in the middle of huge pages if | |
2316 | * something is sent in the current iteration. | |
2317 | */ | |
2318 | if (pagesize_bits > 1 && tmppages > 0) { | |
2319 | migration_rate_limit(); | |
2320 | } | |
2321 | } | |
2322 | if (preempt_active) { | |
2323 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2324 | } | |
2325 | } else { | |
2326 | tmppages = 0; | |
23feba90 | 2327 | } |
f3321554 PX |
2328 | |
2329 | if (tmppages < 0) { | |
d9e474ea | 2330 | pss_host_page_finish(pss); |
f3321554 PX |
2331 | return tmppages; |
2332 | } | |
2333 | ||
d9e474ea PX |
2334 | pss_find_next_dirty(pss); |
2335 | } while (pss_within_range(pss)); | |
2336 | ||
2337 | pss_host_page_finish(pss); | |
278e2f55 AG |
2338 | |
2339 | res = ram_save_release_protection(rs, pss, start_page); | |
2340 | return (res < 0 ? res : pages); | |
a82d593b | 2341 | } |
6c595cde | 2342 | |
56e93d26 | 2343 | /** |
3d0684b2 | 2344 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
2345 | * |
2346 | * Called within an RCU critical section. | |
2347 | * | |
e8f3735f XG |
2348 | * Returns the number of pages written where zero means no dirty pages, |
2349 | * or negative on error | |
56e93d26 | 2350 | * |
6f37bb8b | 2351 | * @rs: current RAM state |
a82d593b DDAG |
2352 | * |
2353 | * On systems where host-page-size > target-page-size it will send all the | |
2354 | * pages in a host page that are dirty. | |
56e93d26 | 2355 | */ |
05931ec5 | 2356 | static int ram_find_and_save_block(RAMState *rs) |
56e93d26 | 2357 | { |
f1668764 | 2358 | PageSearchStatus *pss = &rs->pss[RAM_CHANNEL_PRECOPY]; |
56e93d26 | 2359 | int pages = 0; |
56e93d26 | 2360 | |
0827b9e9 | 2361 | /* No dirty page as there is zero RAM */ |
8d80e195 | 2362 | if (!rs->ram_bytes_total) { |
0827b9e9 AA |
2363 | return pages; |
2364 | } | |
2365 | ||
4934a5dd PX |
2366 | /* |
2367 | * Always keep last_seen_block/last_page valid during this procedure, | |
2368 | * because find_dirty_block() relies on these values (e.g., we compare | |
2369 | * last_seen_block with pss.block to see whether we searched all the | |
2370 | * ramblocks) to detect the completion of migration. Having NULL value | |
2371 | * of last_seen_block can conditionally cause below loop to run forever. | |
2372 | */ | |
2373 | if (!rs->last_seen_block) { | |
2374 | rs->last_seen_block = QLIST_FIRST_RCU(&ram_list.blocks); | |
2375 | rs->last_page = 0; | |
2376 | } | |
2377 | ||
f1668764 | 2378 | pss_init(pss, rs->last_seen_block, rs->last_page); |
b8fb8cb7 | 2379 | |
31e2ac74 | 2380 | while (true){ |
51efd36f | 2381 | if (!get_queued_page(rs, pss)) { |
b062106d | 2382 | /* priority queue empty, so just search for something dirty */ |
31e2ac74 JQ |
2383 | int res = find_dirty_block(rs, pss); |
2384 | if (res != PAGE_DIRTY_FOUND) { | |
2385 | if (res == PAGE_ALL_CLEAN) { | |
51efd36f | 2386 | break; |
31e2ac74 JQ |
2387 | } else if (res == PAGE_TRY_AGAIN) { |
2388 | continue; | |
294e5a40 JQ |
2389 | } else if (res < 0) { |
2390 | pages = res; | |
2391 | break; | |
51efd36f JQ |
2392 | } |
2393 | } | |
56e93d26 | 2394 | } |
51efd36f | 2395 | pages = ram_save_host_page(rs, pss); |
31e2ac74 JQ |
2396 | if (pages) { |
2397 | break; | |
2398 | } | |
2399 | } | |
56e93d26 | 2400 | |
f1668764 PX |
2401 | rs->last_seen_block = pss->block; |
2402 | rs->last_page = pss->page; | |
56e93d26 JQ |
2403 | |
2404 | return pages; | |
2405 | } | |
2406 | ||
8008a272 | 2407 | static uint64_t ram_bytes_total_with_ignored(void) |
56e93d26 JQ |
2408 | { |
2409 | RAMBlock *block; | |
2410 | uint64_t total = 0; | |
2411 | ||
89ac5a1d DDAG |
2412 | RCU_READ_LOCK_GUARD(); |
2413 | ||
8008a272 JQ |
2414 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
2415 | total += block->used_length; | |
99e15582 | 2416 | } |
56e93d26 JQ |
2417 | return total; |
2418 | } | |
2419 | ||
fbd162e6 YK |
2420 | uint64_t ram_bytes_total(void) |
2421 | { | |
8008a272 JQ |
2422 | RAMBlock *block; |
2423 | uint64_t total = 0; | |
2424 | ||
2425 | RCU_READ_LOCK_GUARD(); | |
2426 | ||
2427 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2428 | total += block->used_length; | |
2429 | } | |
2430 | return total; | |
fbd162e6 YK |
2431 | } |
2432 | ||
f265e0e4 | 2433 | static void xbzrle_load_setup(void) |
56e93d26 | 2434 | { |
f265e0e4 | 2435 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2436 | } |
2437 | ||
f265e0e4 JQ |
2438 | static void xbzrle_load_cleanup(void) |
2439 | { | |
2440 | g_free(XBZRLE.decoded_buf); | |
2441 | XBZRLE.decoded_buf = NULL; | |
2442 | } | |
2443 | ||
7d7c96be PX |
2444 | static void ram_state_cleanup(RAMState **rsp) |
2445 | { | |
b9ccaf6d DDAG |
2446 | if (*rsp) { |
2447 | migration_page_queue_free(*rsp); | |
2448 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
2449 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
2450 | g_free(*rsp); | |
2451 | *rsp = NULL; | |
2452 | } | |
7d7c96be PX |
2453 | } |
2454 | ||
84593a08 PX |
2455 | static void xbzrle_cleanup(void) |
2456 | { | |
2457 | XBZRLE_cache_lock(); | |
2458 | if (XBZRLE.cache) { | |
2459 | cache_fini(XBZRLE.cache); | |
2460 | g_free(XBZRLE.encoded_buf); | |
2461 | g_free(XBZRLE.current_buf); | |
2462 | g_free(XBZRLE.zero_target_page); | |
2463 | XBZRLE.cache = NULL; | |
2464 | XBZRLE.encoded_buf = NULL; | |
2465 | XBZRLE.current_buf = NULL; | |
2466 | XBZRLE.zero_target_page = NULL; | |
2467 | } | |
2468 | XBZRLE_cache_unlock(); | |
2469 | } | |
2470 | ||
f265e0e4 | 2471 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 2472 | { |
53518d94 | 2473 | RAMState **rsp = opaque; |
6b6712ef | 2474 | RAMBlock *block; |
eb859c53 | 2475 | |
278e2f55 AG |
2476 | /* We don't use dirty log with background snapshots */ |
2477 | if (!migrate_background_snapshot()) { | |
2478 | /* caller have hold iothread lock or is in a bh, so there is | |
2479 | * no writing race against the migration bitmap | |
2480 | */ | |
63b41db4 HH |
2481 | if (global_dirty_tracking & GLOBAL_DIRTY_MIGRATION) { |
2482 | /* | |
2483 | * do not stop dirty log without starting it, since | |
2484 | * memory_global_dirty_log_stop will assert that | |
2485 | * memory_global_dirty_log_start/stop used in pairs | |
2486 | */ | |
2487 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); | |
2488 | } | |
278e2f55 | 2489 | } |
6b6712ef | 2490 | |
fbd162e6 | 2491 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
2492 | g_free(block->clear_bmap); |
2493 | block->clear_bmap = NULL; | |
6b6712ef JQ |
2494 | g_free(block->bmap); |
2495 | block->bmap = NULL; | |
56e93d26 JQ |
2496 | } |
2497 | ||
84593a08 | 2498 | xbzrle_cleanup(); |
f0afa331 | 2499 | compress_threads_save_cleanup(); |
7d7c96be | 2500 | ram_state_cleanup(rsp); |
4010ba38 JQ |
2501 | g_free(migration_ops); |
2502 | migration_ops = NULL; | |
56e93d26 JQ |
2503 | } |
2504 | ||
6f37bb8b | 2505 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 2506 | { |
ec6f3ab9 PX |
2507 | int i; |
2508 | ||
2509 | for (i = 0; i < RAM_CHANNEL_MAX; i++) { | |
2510 | rs->pss[i].last_sent_block = NULL; | |
2511 | } | |
2512 | ||
6f37bb8b | 2513 | rs->last_seen_block = NULL; |
269ace29 | 2514 | rs->last_page = 0; |
6f37bb8b | 2515 | rs->last_version = ram_list.version; |
f3095cc8 | 2516 | rs->xbzrle_started = false; |
56e93d26 JQ |
2517 | } |
2518 | ||
2519 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
2520 | ||
e0b266f0 DDAG |
2521 | /* **** functions for postcopy ***** */ |
2522 | ||
ced1c616 PB |
2523 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
2524 | { | |
2525 | struct RAMBlock *block; | |
ced1c616 | 2526 | |
fbd162e6 | 2527 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2528 | unsigned long *bitmap = block->bmap; |
2529 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
2530 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
2531 | |
2532 | while (run_start < range) { | |
2533 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
2534 | ram_discard_range(block->idstr, |
2535 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
2536 | ((ram_addr_t)(run_end - run_start)) | |
2537 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
2538 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
2539 | } | |
2540 | } | |
2541 | } | |
2542 | ||
3d0684b2 JQ |
2543 | /** |
2544 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
2545 | * | |
e0b266f0 | 2546 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
2547 | * |
2548 | * @ms: current migration state | |
89dab31b | 2549 | * @block: RAMBlock to discard |
e0b266f0 | 2550 | */ |
9e7d1223 | 2551 | static void postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 2552 | { |
6b6712ef | 2553 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 2554 | unsigned long current; |
1e7cf8c3 | 2555 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 2556 | |
6b6712ef | 2557 | for (current = 0; current < end; ) { |
1e7cf8c3 | 2558 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 2559 | unsigned long zero, discard_length; |
e0b266f0 | 2560 | |
33a5cb62 WY |
2561 | if (one >= end) { |
2562 | break; | |
2563 | } | |
e0b266f0 | 2564 | |
1e7cf8c3 | 2565 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
2566 | |
2567 | if (zero >= end) { | |
2568 | discard_length = end - one; | |
e0b266f0 | 2569 | } else { |
33a5cb62 WY |
2570 | discard_length = zero - one; |
2571 | } | |
810cf2bb | 2572 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 2573 | current = one + discard_length; |
e0b266f0 | 2574 | } |
e0b266f0 DDAG |
2575 | } |
2576 | ||
f30c2e5b PX |
2577 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block); |
2578 | ||
3d0684b2 JQ |
2579 | /** |
2580 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
2581 | * | |
e0b266f0 DDAG |
2582 | * Utility for the outgoing postcopy code. |
2583 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
2584 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
2585 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
2586 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
2587 | * |
2588 | * @ms: current migration state | |
e0b266f0 | 2589 | */ |
739fcc1b | 2590 | static void postcopy_each_ram_send_discard(MigrationState *ms) |
e0b266f0 DDAG |
2591 | { |
2592 | struct RAMBlock *block; | |
e0b266f0 | 2593 | |
fbd162e6 | 2594 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 2595 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 | 2596 | |
f30c2e5b PX |
2597 | /* |
2598 | * Deal with TPS != HPS and huge pages. It discard any partially sent | |
2599 | * host-page size chunks, mark any partially dirty host-page size | |
2600 | * chunks as all dirty. In this case the host-page is the host-page | |
2601 | * for the particular RAMBlock, i.e. it might be a huge page. | |
2602 | */ | |
2603 | postcopy_chunk_hostpages_pass(ms, block); | |
2604 | ||
e0b266f0 DDAG |
2605 | /* |
2606 | * Postcopy sends chunks of bitmap over the wire, but it | |
2607 | * just needs indexes at this point, avoids it having | |
2608 | * target page specific code. | |
2609 | */ | |
739fcc1b | 2610 | postcopy_send_discard_bm_ram(ms, block); |
810cf2bb | 2611 | postcopy_discard_send_finish(ms); |
e0b266f0 | 2612 | } |
e0b266f0 DDAG |
2613 | } |
2614 | ||
3d0684b2 | 2615 | /** |
8324ef86 | 2616 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2617 | * |
2618 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2619 | * canonicalize the two bitmaps, that are similar, but one is | |
2620 | * inverted. | |
99e314eb | 2621 | * |
3d0684b2 JQ |
2622 | * Postcopy requires that all target pages in a hostpage are dirty or |
2623 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2624 | * |
3d0684b2 | 2625 | * @ms: current migration state |
3d0684b2 | 2626 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2627 | */ |
1e7cf8c3 | 2628 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2629 | { |
53518d94 | 2630 | RAMState *rs = ram_state; |
6b6712ef | 2631 | unsigned long *bitmap = block->bmap; |
29c59172 | 2632 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2633 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2634 | unsigned long run_start; |
2635 | ||
29c59172 DDAG |
2636 | if (block->page_size == TARGET_PAGE_SIZE) { |
2637 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2638 | return; | |
2639 | } | |
2640 | ||
1e7cf8c3 WY |
2641 | /* Find a dirty page */ |
2642 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2643 | |
6b6712ef | 2644 | while (run_start < pages) { |
99e314eb DDAG |
2645 | |
2646 | /* | |
2647 | * If the start of this run of pages is in the middle of a host | |
2648 | * page, then we need to fixup this host page. | |
2649 | */ | |
9dec3cc3 | 2650 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2651 | /* Find the end of this run */ |
1e7cf8c3 | 2652 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2653 | /* |
2654 | * If the end isn't at the start of a host page, then the | |
2655 | * run doesn't finish at the end of a host page | |
2656 | * and we need to discard. | |
2657 | */ | |
99e314eb DDAG |
2658 | } |
2659 | ||
9dec3cc3 | 2660 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2661 | unsigned long page; |
dad45ab2 WY |
2662 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2663 | host_ratio); | |
2664 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2665 | |
99e314eb DDAG |
2666 | /* Clean up the bitmap */ |
2667 | for (page = fixup_start_addr; | |
2668 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2669 | /* |
2670 | * Remark them as dirty, updating the count for any pages | |
2671 | * that weren't previously dirty. | |
2672 | */ | |
0d8ec885 | 2673 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2674 | } |
2675 | } | |
2676 | ||
1e7cf8c3 WY |
2677 | /* Find the next dirty page for the next iteration */ |
2678 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2679 | } |
2680 | } | |
2681 | ||
3d0684b2 JQ |
2682 | /** |
2683 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2684 | * | |
e0b266f0 DDAG |
2685 | * Transmit the set of pages to be discarded after precopy to the target |
2686 | * these are pages that: | |
2687 | * a) Have been previously transmitted but are now dirty again | |
2688 | * b) Pages that have never been transmitted, this ensures that | |
2689 | * any pages on the destination that have been mapped by background | |
2690 | * tasks get discarded (transparent huge pages is the specific concern) | |
2691 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2692 | * |
2693 | * @ms: current migration state | |
e0b266f0 | 2694 | */ |
739fcc1b | 2695 | void ram_postcopy_send_discard_bitmap(MigrationState *ms) |
e0b266f0 | 2696 | { |
53518d94 | 2697 | RAMState *rs = ram_state; |
e0b266f0 | 2698 | |
89ac5a1d | 2699 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2700 | |
2701 | /* This should be our last sync, the src is now paused */ | |
1e493be5 | 2702 | migration_bitmap_sync(rs, false); |
e0b266f0 | 2703 | |
6b6712ef | 2704 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
ec6f3ab9 | 2705 | rs->pss[RAM_CHANNEL_PRECOPY].last_sent_block = NULL; |
6b6712ef | 2706 | rs->last_seen_block = NULL; |
6b6712ef | 2707 | rs->last_page = 0; |
e0b266f0 | 2708 | |
739fcc1b | 2709 | postcopy_each_ram_send_discard(ms); |
e0b266f0 | 2710 | |
739fcc1b | 2711 | trace_ram_postcopy_send_discard_bitmap(); |
e0b266f0 DDAG |
2712 | } |
2713 | ||
3d0684b2 JQ |
2714 | /** |
2715 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2716 | * |
3d0684b2 | 2717 | * Returns zero on success |
e0b266f0 | 2718 | * |
36449157 JQ |
2719 | * @rbname: name of the RAMBlock of the request. NULL means the |
2720 | * same that last one. | |
3d0684b2 JQ |
2721 | * @start: RAMBlock starting page |
2722 | * @length: RAMBlock size | |
e0b266f0 | 2723 | */ |
aaa2064c | 2724 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 2725 | { |
36449157 | 2726 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 2727 | |
89ac5a1d | 2728 | RCU_READ_LOCK_GUARD(); |
36449157 | 2729 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
2730 | |
2731 | if (!rb) { | |
36449157 | 2732 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 2733 | return -1; |
e0b266f0 DDAG |
2734 | } |
2735 | ||
814bb08f PX |
2736 | /* |
2737 | * On source VM, we don't need to update the received bitmap since | |
2738 | * we don't even have one. | |
2739 | */ | |
2740 | if (rb->receivedmap) { | |
2741 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
2742 | length >> qemu_target_page_bits()); | |
2743 | } | |
2744 | ||
03acb4e9 | 2745 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
2746 | } |
2747 | ||
84593a08 PX |
2748 | /* |
2749 | * For every allocation, we will try not to crash the VM if the | |
2750 | * allocation failed. | |
2751 | */ | |
2752 | static int xbzrle_init(void) | |
2753 | { | |
2754 | Error *local_err = NULL; | |
2755 | ||
87dca0c9 | 2756 | if (!migrate_xbzrle()) { |
84593a08 PX |
2757 | return 0; |
2758 | } | |
2759 | ||
2760 | XBZRLE_cache_lock(); | |
2761 | ||
2762 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
2763 | if (!XBZRLE.zero_target_page) { | |
2764 | error_report("%s: Error allocating zero page", __func__); | |
2765 | goto err_out; | |
2766 | } | |
2767 | ||
2768 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
2769 | TARGET_PAGE_SIZE, &local_err); | |
2770 | if (!XBZRLE.cache) { | |
2771 | error_report_err(local_err); | |
2772 | goto free_zero_page; | |
2773 | } | |
2774 | ||
2775 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
2776 | if (!XBZRLE.encoded_buf) { | |
2777 | error_report("%s: Error allocating encoded_buf", __func__); | |
2778 | goto free_cache; | |
2779 | } | |
2780 | ||
2781 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
2782 | if (!XBZRLE.current_buf) { | |
2783 | error_report("%s: Error allocating current_buf", __func__); | |
2784 | goto free_encoded_buf; | |
2785 | } | |
2786 | ||
2787 | /* We are all good */ | |
2788 | XBZRLE_cache_unlock(); | |
2789 | return 0; | |
2790 | ||
2791 | free_encoded_buf: | |
2792 | g_free(XBZRLE.encoded_buf); | |
2793 | XBZRLE.encoded_buf = NULL; | |
2794 | free_cache: | |
2795 | cache_fini(XBZRLE.cache); | |
2796 | XBZRLE.cache = NULL; | |
2797 | free_zero_page: | |
2798 | g_free(XBZRLE.zero_target_page); | |
2799 | XBZRLE.zero_target_page = NULL; | |
2800 | err_out: | |
2801 | XBZRLE_cache_unlock(); | |
2802 | return -ENOMEM; | |
2803 | } | |
2804 | ||
53518d94 | 2805 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 2806 | { |
7d00ee6a PX |
2807 | *rsp = g_try_new0(RAMState, 1); |
2808 | ||
2809 | if (!*rsp) { | |
2810 | error_report("%s: Init ramstate fail", __func__); | |
2811 | return -1; | |
2812 | } | |
53518d94 JQ |
2813 | |
2814 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
2815 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
2816 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
8d80e195 | 2817 | (*rsp)->ram_bytes_total = ram_bytes_total(); |
56e93d26 | 2818 | |
7d00ee6a | 2819 | /* |
40c4d4a8 IR |
2820 | * Count the total number of pages used by ram blocks not including any |
2821 | * gaps due to alignment or unplugs. | |
03158519 | 2822 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 2823 | */ |
8d80e195 | 2824 | (*rsp)->migration_dirty_pages = (*rsp)->ram_bytes_total >> TARGET_PAGE_BITS; |
7d00ee6a PX |
2825 | ram_state_reset(*rsp); |
2826 | ||
2827 | return 0; | |
2828 | } | |
2829 | ||
d6eff5d7 | 2830 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 2831 | { |
002cad6b | 2832 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
2833 | RAMBlock *block; |
2834 | unsigned long pages; | |
002cad6b | 2835 | uint8_t shift; |
56e93d26 | 2836 | |
0827b9e9 AA |
2837 | /* Skip setting bitmap if there is no RAM */ |
2838 | if (ram_bytes_total()) { | |
002cad6b PX |
2839 | shift = ms->clear_bitmap_shift; |
2840 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
2841 | error_report("clear_bitmap_shift (%u) too big, using " | |
2842 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
2843 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
2844 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
2845 | error_report("clear_bitmap_shift (%u) too small, using " | |
2846 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
2847 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
2848 | } | |
2849 | ||
fbd162e6 | 2850 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 2851 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
2852 | /* |
2853 | * The initial dirty bitmap for migration must be set with all | |
2854 | * ones to make sure we'll migrate every guest RAM page to | |
2855 | * destination. | |
40c4d4a8 IR |
2856 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
2857 | * new migration after a failed migration, ram_list. | |
2858 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
2859 | * guest memory. | |
03158519 | 2860 | */ |
6b6712ef | 2861 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 2862 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
2863 | block->clear_bmap_shift = shift; |
2864 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 2865 | } |
f3f491fc | 2866 | } |
d6eff5d7 PX |
2867 | } |
2868 | ||
be39b4cd DH |
2869 | static void migration_bitmap_clear_discarded_pages(RAMState *rs) |
2870 | { | |
2871 | unsigned long pages; | |
2872 | RAMBlock *rb; | |
2873 | ||
2874 | RCU_READ_LOCK_GUARD(); | |
2875 | ||
2876 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { | |
2877 | pages = ramblock_dirty_bitmap_clear_discarded_pages(rb); | |
2878 | rs->migration_dirty_pages -= pages; | |
2879 | } | |
2880 | } | |
2881 | ||
d6eff5d7 PX |
2882 | static void ram_init_bitmaps(RAMState *rs) |
2883 | { | |
2884 | /* For memory_global_dirty_log_start below. */ | |
2885 | qemu_mutex_lock_iothread(); | |
2886 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 2887 | |
89ac5a1d DDAG |
2888 | WITH_RCU_READ_LOCK_GUARD() { |
2889 | ram_list_init_bitmaps(); | |
278e2f55 AG |
2890 | /* We don't use dirty log with background snapshots */ |
2891 | if (!migrate_background_snapshot()) { | |
63b41db4 | 2892 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
1e493be5 | 2893 | migration_bitmap_sync_precopy(rs, false); |
278e2f55 | 2894 | } |
89ac5a1d | 2895 | } |
56e93d26 | 2896 | qemu_mutex_unlock_ramlist(); |
49877834 | 2897 | qemu_mutex_unlock_iothread(); |
be39b4cd DH |
2898 | |
2899 | /* | |
2900 | * After an eventual first bitmap sync, fixup the initial bitmap | |
2901 | * containing all 1s to exclude any discarded pages from migration. | |
2902 | */ | |
2903 | migration_bitmap_clear_discarded_pages(rs); | |
d6eff5d7 PX |
2904 | } |
2905 | ||
2906 | static int ram_init_all(RAMState **rsp) | |
2907 | { | |
2908 | if (ram_state_init(rsp)) { | |
2909 | return -1; | |
2910 | } | |
2911 | ||
2912 | if (xbzrle_init()) { | |
2913 | ram_state_cleanup(rsp); | |
2914 | return -1; | |
2915 | } | |
2916 | ||
2917 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
2918 | |
2919 | return 0; | |
2920 | } | |
2921 | ||
08614f34 PX |
2922 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
2923 | { | |
2924 | RAMBlock *block; | |
2925 | uint64_t pages = 0; | |
2926 | ||
2927 | /* | |
2928 | * Postcopy is not using xbzrle/compression, so no need for that. | |
2929 | * Also, since source are already halted, we don't need to care | |
2930 | * about dirty page logging as well. | |
2931 | */ | |
2932 | ||
fbd162e6 | 2933 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
2934 | pages += bitmap_count_one(block->bmap, |
2935 | block->used_length >> TARGET_PAGE_BITS); | |
2936 | } | |
2937 | ||
2938 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
2939 | rs->migration_dirty_pages = pages; | |
2940 | ||
1a373522 | 2941 | ram_state_reset(rs); |
08614f34 PX |
2942 | |
2943 | /* Update RAMState cache of output QEMUFile */ | |
7f401b80 | 2944 | rs->pss[RAM_CHANNEL_PRECOPY].pss_channel = out; |
08614f34 PX |
2945 | |
2946 | trace_ram_state_resume_prepare(pages); | |
2947 | } | |
2948 | ||
6bcb05fc WW |
2949 | /* |
2950 | * This function clears bits of the free pages reported by the caller from the | |
2951 | * migration dirty bitmap. @addr is the host address corresponding to the | |
2952 | * start of the continuous guest free pages, and @len is the total bytes of | |
2953 | * those pages. | |
2954 | */ | |
2955 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
2956 | { | |
2957 | RAMBlock *block; | |
2958 | ram_addr_t offset; | |
2959 | size_t used_len, start, npages; | |
2960 | MigrationState *s = migrate_get_current(); | |
2961 | ||
2962 | /* This function is currently expected to be used during live migration */ | |
2963 | if (!migration_is_setup_or_active(s->state)) { | |
2964 | return; | |
2965 | } | |
2966 | ||
2967 | for (; len > 0; len -= used_len, addr += used_len) { | |
2968 | block = qemu_ram_block_from_host(addr, false, &offset); | |
2969 | if (unlikely(!block || offset >= block->used_length)) { | |
2970 | /* | |
2971 | * The implementation might not support RAMBlock resize during | |
2972 | * live migration, but it could happen in theory with future | |
2973 | * updates. So we add a check here to capture that case. | |
2974 | */ | |
2975 | error_report_once("%s unexpected error", __func__); | |
2976 | return; | |
2977 | } | |
2978 | ||
2979 | if (len <= block->used_length - offset) { | |
2980 | used_len = len; | |
2981 | } else { | |
2982 | used_len = block->used_length - offset; | |
2983 | } | |
2984 | ||
2985 | start = offset >> TARGET_PAGE_BITS; | |
2986 | npages = used_len >> TARGET_PAGE_BITS; | |
2987 | ||
2988 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3143577d WW |
2989 | /* |
2990 | * The skipped free pages are equavalent to be sent from clear_bmap's | |
2991 | * perspective, so clear the bits from the memory region bitmap which | |
2992 | * are initially set. Otherwise those skipped pages will be sent in | |
2993 | * the next round after syncing from the memory region bitmap. | |
2994 | */ | |
1230a25f | 2995 | migration_clear_memory_region_dirty_bitmap_range(block, start, npages); |
6bcb05fc WW |
2996 | ram_state->migration_dirty_pages -= |
2997 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
2998 | bitmap_clear(block->bmap, start, npages); | |
2999 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
3000 | } | |
3001 | } | |
3002 | ||
3d0684b2 JQ |
3003 | /* |
3004 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
3005 | * long-running RCU critical section. When rcu-reclaims in the code |
3006 | * start to become numerous it will be necessary to reduce the | |
3007 | * granularity of these critical sections. | |
3008 | */ | |
3009 | ||
3d0684b2 JQ |
3010 | /** |
3011 | * ram_save_setup: Setup RAM for migration | |
3012 | * | |
3013 | * Returns zero to indicate success and negative for error | |
3014 | * | |
3015 | * @f: QEMUFile where to send the data | |
3016 | * @opaque: RAMState pointer | |
3017 | */ | |
a91246c9 HZ |
3018 | static int ram_save_setup(QEMUFile *f, void *opaque) |
3019 | { | |
53518d94 | 3020 | RAMState **rsp = opaque; |
a91246c9 | 3021 | RAMBlock *block; |
33d70973 | 3022 | int ret; |
a91246c9 | 3023 | |
dcaf446e XG |
3024 | if (compress_threads_save_setup()) { |
3025 | return -1; | |
3026 | } | |
3027 | ||
a91246c9 HZ |
3028 | /* migration has already setup the bitmap, reuse it. */ |
3029 | if (!migration_in_colo_state()) { | |
7d00ee6a | 3030 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 3031 | compress_threads_save_cleanup(); |
a91246c9 | 3032 | return -1; |
53518d94 | 3033 | } |
a91246c9 | 3034 | } |
7f401b80 | 3035 | (*rsp)->pss[RAM_CHANNEL_PRECOPY].pss_channel = f; |
a91246c9 | 3036 | |
0e6ebd48 | 3037 | WITH_RCU_READ_LOCK_GUARD() { |
8008a272 JQ |
3038 | qemu_put_be64(f, ram_bytes_total_with_ignored() |
3039 | | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 3040 | |
0e6ebd48 DDAG |
3041 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
3042 | qemu_put_byte(f, strlen(block->idstr)); | |
3043 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
3044 | qemu_put_be64(f, block->used_length); | |
3045 | if (migrate_postcopy_ram() && block->page_size != | |
3046 | qemu_host_page_size) { | |
3047 | qemu_put_be64(f, block->page_size); | |
3048 | } | |
3049 | if (migrate_ignore_shared()) { | |
3050 | qemu_put_be64(f, block->mr->addr); | |
3051 | } | |
fbd162e6 | 3052 | } |
56e93d26 JQ |
3053 | } |
3054 | ||
56e93d26 JQ |
3055 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
3056 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
3057 | ||
4010ba38 JQ |
3058 | migration_ops = g_malloc0(sizeof(MigrationOps)); |
3059 | migration_ops->ram_save_target_page = ram_save_target_page_legacy; | |
8ebb6ecc | 3060 | ret = multifd_send_sync_main(f); |
33d70973 LB |
3061 | if (ret < 0) { |
3062 | return ret; | |
3063 | } | |
3064 | ||
294e5a40 JQ |
3065 | if (!migrate_multifd_flush_after_each_section()) { |
3066 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); | |
3067 | } | |
3068 | ||
56e93d26 | 3069 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 3070 | qemu_fflush(f); |
56e93d26 JQ |
3071 | |
3072 | return 0; | |
3073 | } | |
3074 | ||
3d0684b2 JQ |
3075 | /** |
3076 | * ram_save_iterate: iterative stage for migration | |
3077 | * | |
3078 | * Returns zero to indicate success and negative for error | |
3079 | * | |
3080 | * @f: QEMUFile where to send the data | |
3081 | * @opaque: RAMState pointer | |
3082 | */ | |
56e93d26 JQ |
3083 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
3084 | { | |
53518d94 JQ |
3085 | RAMState **temp = opaque; |
3086 | RAMState *rs = *temp; | |
3d4095b2 | 3087 | int ret = 0; |
56e93d26 JQ |
3088 | int i; |
3089 | int64_t t0; | |
5c90308f | 3090 | int done = 0; |
56e93d26 | 3091 | |
b2557345 PL |
3092 | if (blk_mig_bulk_active()) { |
3093 | /* Avoid transferring ram during bulk phase of block migration as | |
3094 | * the bulk phase will usually take a long time and transferring | |
3095 | * ram updates during that time is pointless. */ | |
3096 | goto out; | |
3097 | } | |
3098 | ||
63268c49 PX |
3099 | /* |
3100 | * We'll take this lock a little bit long, but it's okay for two reasons. | |
3101 | * Firstly, the only possible other thread to take it is who calls | |
3102 | * qemu_guest_free_page_hint(), which should be rare; secondly, see | |
3103 | * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which | |
3104 | * guarantees that we'll at least released it in a regular basis. | |
3105 | */ | |
3106 | qemu_mutex_lock(&rs->bitmap_mutex); | |
89ac5a1d DDAG |
3107 | WITH_RCU_READ_LOCK_GUARD() { |
3108 | if (ram_list.version != rs->last_version) { | |
3109 | ram_state_reset(rs); | |
3110 | } | |
56e93d26 | 3111 | |
89ac5a1d DDAG |
3112 | /* Read version before ram_list.blocks */ |
3113 | smp_rmb(); | |
56e93d26 | 3114 | |
89ac5a1d | 3115 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 3116 | |
89ac5a1d DDAG |
3117 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
3118 | i = 0; | |
3119 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
a1fe28df | 3120 | postcopy_has_request(rs)) { |
89ac5a1d | 3121 | int pages; |
e03a34f8 | 3122 | |
89ac5a1d DDAG |
3123 | if (qemu_file_get_error(f)) { |
3124 | break; | |
3125 | } | |
e8f3735f | 3126 | |
05931ec5 | 3127 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3128 | /* no more pages to sent */ |
3129 | if (pages == 0) { | |
3130 | done = 1; | |
3131 | break; | |
3132 | } | |
e8f3735f | 3133 | |
89ac5a1d DDAG |
3134 | if (pages < 0) { |
3135 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
3136 | break; |
3137 | } | |
89ac5a1d DDAG |
3138 | |
3139 | rs->target_page_count += pages; | |
3140 | ||
644acf99 WY |
3141 | /* |
3142 | * During postcopy, it is necessary to make sure one whole host | |
3143 | * page is sent in one chunk. | |
3144 | */ | |
3145 | if (migrate_postcopy_ram()) { | |
ef4f5f5d | 3146 | ram_flush_compressed_data(rs); |
644acf99 WY |
3147 | } |
3148 | ||
89ac5a1d DDAG |
3149 | /* |
3150 | * we want to check in the 1st loop, just in case it was the 1st | |
3151 | * time and we had to sync the dirty bitmap. | |
3152 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
3153 | * some iterations | |
3154 | */ | |
3155 | if ((i & 63) == 0) { | |
3156 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
3157 | 1000000; | |
3158 | if (t1 > MAX_WAIT) { | |
3159 | trace_ram_save_iterate_big_wait(t1, i); | |
3160 | break; | |
3161 | } | |
3162 | } | |
3163 | i++; | |
56e93d26 | 3164 | } |
56e93d26 | 3165 | } |
63268c49 | 3166 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 JQ |
3167 | |
3168 | /* | |
3169 | * Must occur before EOS (or any QEMUFile operation) | |
3170 | * because of RDMA protocol. | |
3171 | */ | |
3172 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
3173 | ||
b2557345 | 3174 | out: |
b69a0227 JQ |
3175 | if (ret >= 0 |
3176 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
b05292c2 JQ |
3177 | if (migrate_multifd_flush_after_each_section()) { |
3178 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); | |
3179 | if (ret < 0) { | |
3180 | return ret; | |
3181 | } | |
33d70973 LB |
3182 | } |
3183 | ||
3d4095b2 JQ |
3184 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3185 | qemu_fflush(f); | |
4c2d0f6d | 3186 | ram_transferred_add(8); |
56e93d26 | 3187 | |
3d4095b2 JQ |
3188 | ret = qemu_file_get_error(f); |
3189 | } | |
56e93d26 JQ |
3190 | if (ret < 0) { |
3191 | return ret; | |
3192 | } | |
3193 | ||
5c90308f | 3194 | return done; |
56e93d26 JQ |
3195 | } |
3196 | ||
3d0684b2 JQ |
3197 | /** |
3198 | * ram_save_complete: function called to send the remaining amount of ram | |
3199 | * | |
e8f3735f | 3200 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
3201 | * |
3202 | * Called with iothread lock | |
3203 | * | |
3204 | * @f: QEMUFile where to send the data | |
3205 | * @opaque: RAMState pointer | |
3206 | */ | |
56e93d26 JQ |
3207 | static int ram_save_complete(QEMUFile *f, void *opaque) |
3208 | { | |
53518d94 JQ |
3209 | RAMState **temp = opaque; |
3210 | RAMState *rs = *temp; | |
e8f3735f | 3211 | int ret = 0; |
6f37bb8b | 3212 | |
05931ec5 JQ |
3213 | rs->last_stage = !migration_in_colo_state(); |
3214 | ||
89ac5a1d DDAG |
3215 | WITH_RCU_READ_LOCK_GUARD() { |
3216 | if (!migration_in_postcopy()) { | |
1e493be5 | 3217 | migration_bitmap_sync_precopy(rs, true); |
89ac5a1d | 3218 | } |
56e93d26 | 3219 | |
89ac5a1d | 3220 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 3221 | |
89ac5a1d | 3222 | /* try transferring iterative blocks of memory */ |
56e93d26 | 3223 | |
89ac5a1d | 3224 | /* flush all remaining blocks regardless of rate limiting */ |
c13221b5 | 3225 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
3226 | while (true) { |
3227 | int pages; | |
56e93d26 | 3228 | |
05931ec5 | 3229 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3230 | /* no more blocks to sent */ |
3231 | if (pages == 0) { | |
3232 | break; | |
3233 | } | |
3234 | if (pages < 0) { | |
3235 | ret = pages; | |
3236 | break; | |
3237 | } | |
e8f3735f | 3238 | } |
c13221b5 | 3239 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 3240 | |
ef4f5f5d | 3241 | ram_flush_compressed_data(rs); |
89ac5a1d DDAG |
3242 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
3243 | } | |
d09a6fde | 3244 | |
33d70973 LB |
3245 | if (ret < 0) { |
3246 | return ret; | |
3d4095b2 | 3247 | } |
56e93d26 | 3248 | |
7f401b80 | 3249 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); |
33d70973 LB |
3250 | if (ret < 0) { |
3251 | return ret; | |
3252 | } | |
3253 | ||
294e5a40 JQ |
3254 | if (!migrate_multifd_flush_after_each_section()) { |
3255 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); | |
3256 | } | |
33d70973 LB |
3257 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3258 | qemu_fflush(f); | |
3259 | ||
3260 | return 0; | |
56e93d26 JQ |
3261 | } |
3262 | ||
24beea4e JQ |
3263 | static void ram_state_pending_estimate(void *opaque, uint64_t *must_precopy, |
3264 | uint64_t *can_postcopy) | |
56e93d26 | 3265 | { |
53518d94 JQ |
3266 | RAMState **temp = opaque; |
3267 | RAMState *rs = *temp; | |
56e93d26 | 3268 | |
c8df4a7a | 3269 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3270 | |
c8df4a7a JQ |
3271 | if (migrate_postcopy_ram()) { |
3272 | /* We can do postcopy, and all the data is postcopiable */ | |
24beea4e | 3273 | *can_postcopy += remaining_size; |
c8df4a7a | 3274 | } else { |
24beea4e | 3275 | *must_precopy += remaining_size; |
c8df4a7a JQ |
3276 | } |
3277 | } | |
3278 | ||
24beea4e JQ |
3279 | static void ram_state_pending_exact(void *opaque, uint64_t *must_precopy, |
3280 | uint64_t *can_postcopy) | |
c8df4a7a | 3281 | { |
28ef5339 | 3282 | MigrationState *s = migrate_get_current(); |
c8df4a7a JQ |
3283 | RAMState **temp = opaque; |
3284 | RAMState *rs = *temp; | |
3285 | ||
3286 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; | |
3287 | ||
28ef5339 | 3288 | if (!migration_in_postcopy() && remaining_size < s->threshold_size) { |
56e93d26 | 3289 | qemu_mutex_lock_iothread(); |
89ac5a1d | 3290 | WITH_RCU_READ_LOCK_GUARD() { |
1e493be5 | 3291 | migration_bitmap_sync_precopy(rs, false); |
89ac5a1d | 3292 | } |
56e93d26 | 3293 | qemu_mutex_unlock_iothread(); |
9edabd4d | 3294 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3295 | } |
c31b098f | 3296 | |
86e1167e VSO |
3297 | if (migrate_postcopy_ram()) { |
3298 | /* We can do postcopy, and all the data is postcopiable */ | |
24beea4e | 3299 | *can_postcopy += remaining_size; |
86e1167e | 3300 | } else { |
24beea4e | 3301 | *must_precopy += remaining_size; |
86e1167e | 3302 | } |
56e93d26 JQ |
3303 | } |
3304 | ||
3305 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
3306 | { | |
3307 | unsigned int xh_len; | |
3308 | int xh_flags; | |
063e760a | 3309 | uint8_t *loaded_data; |
56e93d26 | 3310 | |
56e93d26 JQ |
3311 | /* extract RLE header */ |
3312 | xh_flags = qemu_get_byte(f); | |
3313 | xh_len = qemu_get_be16(f); | |
3314 | ||
3315 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
3316 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
3317 | return -1; | |
3318 | } | |
3319 | ||
3320 | if (xh_len > TARGET_PAGE_SIZE) { | |
3321 | error_report("Failed to load XBZRLE page - len overflow!"); | |
3322 | return -1; | |
3323 | } | |
f265e0e4 | 3324 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 3325 | /* load data and decode */ |
f265e0e4 | 3326 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 3327 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
3328 | |
3329 | /* decode RLE */ | |
063e760a | 3330 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
3331 | TARGET_PAGE_SIZE) == -1) { |
3332 | error_report("Failed to load XBZRLE page - decode error!"); | |
3333 | return -1; | |
3334 | } | |
3335 | ||
3336 | return 0; | |
3337 | } | |
3338 | ||
3d0684b2 JQ |
3339 | /** |
3340 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
3341 | * | |
3342 | * Must be called from within a rcu critical section. | |
3343 | * | |
56e93d26 | 3344 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 3345 | * |
755e8d7c | 3346 | * @mis: the migration incoming state pointer |
3d0684b2 JQ |
3347 | * @f: QEMUFile where to read the data from |
3348 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
c01b16ed | 3349 | * @channel: the channel we're using |
a7180877 | 3350 | */ |
755e8d7c | 3351 | static inline RAMBlock *ram_block_from_stream(MigrationIncomingState *mis, |
c01b16ed PX |
3352 | QEMUFile *f, int flags, |
3353 | int channel) | |
56e93d26 | 3354 | { |
c01b16ed | 3355 | RAMBlock *block = mis->last_recv_block[channel]; |
56e93d26 JQ |
3356 | char id[256]; |
3357 | uint8_t len; | |
3358 | ||
3359 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 3360 | if (!block) { |
56e93d26 JQ |
3361 | error_report("Ack, bad migration stream!"); |
3362 | return NULL; | |
3363 | } | |
4c4bad48 | 3364 | return block; |
56e93d26 JQ |
3365 | } |
3366 | ||
3367 | len = qemu_get_byte(f); | |
3368 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3369 | id[len] = 0; | |
3370 | ||
e3dd7493 | 3371 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
3372 | if (!block) { |
3373 | error_report("Can't find block %s", id); | |
3374 | return NULL; | |
56e93d26 JQ |
3375 | } |
3376 | ||
fbd162e6 | 3377 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
3378 | error_report("block %s should not be migrated !", id); |
3379 | return NULL; | |
3380 | } | |
3381 | ||
c01b16ed | 3382 | mis->last_recv_block[channel] = block; |
755e8d7c | 3383 | |
4c4bad48 HZ |
3384 | return block; |
3385 | } | |
3386 | ||
3387 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
3388 | ram_addr_t offset) | |
3389 | { | |
3390 | if (!offset_in_ramblock(block, offset)) { | |
3391 | return NULL; | |
3392 | } | |
3393 | ||
3394 | return block->host + offset; | |
56e93d26 JQ |
3395 | } |
3396 | ||
6a23f639 DH |
3397 | static void *host_page_from_ram_block_offset(RAMBlock *block, |
3398 | ram_addr_t offset) | |
3399 | { | |
3400 | /* Note: Explicitly no check against offset_in_ramblock(). */ | |
3401 | return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block->host + offset), | |
3402 | block->page_size); | |
3403 | } | |
3404 | ||
3405 | static ram_addr_t host_page_offset_from_ram_block_offset(RAMBlock *block, | |
3406 | ram_addr_t offset) | |
3407 | { | |
3408 | return ((uintptr_t)block->host + offset) & (block->page_size - 1); | |
3409 | } | |
3410 | ||
871cfc54 LS |
3411 | void colo_record_bitmap(RAMBlock *block, ram_addr_t *normal, uint32_t pages) |
3412 | { | |
3413 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3414 | for (int i = 0; i < pages; i++) { | |
3415 | ram_addr_t offset = normal[i]; | |
3416 | ram_state->migration_dirty_pages += !test_and_set_bit( | |
3417 | offset >> TARGET_PAGE_BITS, | |
3418 | block->bmap); | |
3419 | } | |
3420 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
3421 | } | |
3422 | ||
13af18f2 | 3423 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 3424 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
3425 | { |
3426 | if (!offset_in_ramblock(block, offset)) { | |
3427 | return NULL; | |
3428 | } | |
3429 | if (!block->colo_cache) { | |
3430 | error_report("%s: colo_cache is NULL in block :%s", | |
3431 | __func__, block->idstr); | |
3432 | return NULL; | |
3433 | } | |
7d9acafa ZC |
3434 | |
3435 | /* | |
3436 | * During colo checkpoint, we need bitmap of these migrated pages. | |
3437 | * It help us to decide which pages in ram cache should be flushed | |
3438 | * into VM's RAM later. | |
3439 | */ | |
871cfc54 LS |
3440 | if (record_bitmap) { |
3441 | colo_record_bitmap(block, &offset, 1); | |
7d9acafa | 3442 | } |
13af18f2 ZC |
3443 | return block->colo_cache + offset; |
3444 | } | |
3445 | ||
3d0684b2 JQ |
3446 | /** |
3447 | * ram_handle_compressed: handle the zero page case | |
3448 | * | |
56e93d26 JQ |
3449 | * If a page (or a whole RDMA chunk) has been |
3450 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
3451 | * |
3452 | * @host: host address for the zero page | |
3453 | * @ch: what the page is filled from. We only support zero | |
3454 | * @size: size of the zero page | |
56e93d26 JQ |
3455 | */ |
3456 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
3457 | { | |
bad452a7 | 3458 | if (ch != 0 || !buffer_is_zero(host, size)) { |
56e93d26 JQ |
3459 | memset(host, ch, size); |
3460 | } | |
3461 | } | |
3462 | ||
b70cb3b4 RL |
3463 | static void colo_init_ram_state(void) |
3464 | { | |
3465 | ram_state_init(&ram_state); | |
b70cb3b4 RL |
3466 | } |
3467 | ||
13af18f2 ZC |
3468 | /* |
3469 | * colo cache: this is for secondary VM, we cache the whole | |
3470 | * memory of the secondary VM, it is need to hold the global lock | |
3471 | * to call this helper. | |
3472 | */ | |
3473 | int colo_init_ram_cache(void) | |
3474 | { | |
3475 | RAMBlock *block; | |
3476 | ||
44901b5a PB |
3477 | WITH_RCU_READ_LOCK_GUARD() { |
3478 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3479 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
8dbe22c6 | 3480 | NULL, false, false); |
44901b5a PB |
3481 | if (!block->colo_cache) { |
3482 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
3483 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
3484 | block->used_length); | |
3485 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3486 | if (block->colo_cache) { | |
3487 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3488 | block->colo_cache = NULL; | |
3489 | } | |
89ac5a1d | 3490 | } |
44901b5a | 3491 | return -errno; |
89ac5a1d | 3492 | } |
e5fdf920 LS |
3493 | if (!machine_dump_guest_core(current_machine)) { |
3494 | qemu_madvise(block->colo_cache, block->used_length, | |
3495 | QEMU_MADV_DONTDUMP); | |
3496 | } | |
13af18f2 | 3497 | } |
13af18f2 | 3498 | } |
44901b5a | 3499 | |
7d9acafa ZC |
3500 | /* |
3501 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3502 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3503 | * we use the same name 'ram_bitmap' as for migration. | |
3504 | */ | |
3505 | if (ram_bytes_total()) { | |
3506 | RAMBlock *block; | |
3507 | ||
fbd162e6 | 3508 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3509 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3510 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3511 | } |
3512 | } | |
7d9acafa | 3513 | |
b70cb3b4 | 3514 | colo_init_ram_state(); |
13af18f2 | 3515 | return 0; |
13af18f2 ZC |
3516 | } |
3517 | ||
0393031a HZ |
3518 | /* TODO: duplicated with ram_init_bitmaps */ |
3519 | void colo_incoming_start_dirty_log(void) | |
3520 | { | |
3521 | RAMBlock *block = NULL; | |
3522 | /* For memory_global_dirty_log_start below. */ | |
3523 | qemu_mutex_lock_iothread(); | |
3524 | qemu_mutex_lock_ramlist(); | |
3525 | ||
1e493be5 | 3526 | memory_global_dirty_log_sync(false); |
0393031a HZ |
3527 | WITH_RCU_READ_LOCK_GUARD() { |
3528 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3529 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3530 | /* Discard this dirty bitmap record */ | |
3531 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3532 | } | |
63b41db4 | 3533 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
0393031a HZ |
3534 | } |
3535 | ram_state->migration_dirty_pages = 0; | |
3536 | qemu_mutex_unlock_ramlist(); | |
3537 | qemu_mutex_unlock_iothread(); | |
3538 | } | |
3539 | ||
13af18f2 ZC |
3540 | /* It is need to hold the global lock to call this helper */ |
3541 | void colo_release_ram_cache(void) | |
3542 | { | |
3543 | RAMBlock *block; | |
3544 | ||
63b41db4 | 3545 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); |
fbd162e6 | 3546 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3547 | g_free(block->bmap); |
3548 | block->bmap = NULL; | |
3549 | } | |
3550 | ||
89ac5a1d DDAG |
3551 | WITH_RCU_READ_LOCK_GUARD() { |
3552 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3553 | if (block->colo_cache) { | |
3554 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3555 | block->colo_cache = NULL; | |
3556 | } | |
13af18f2 ZC |
3557 | } |
3558 | } | |
0393031a | 3559 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
3560 | } |
3561 | ||
f265e0e4 JQ |
3562 | /** |
3563 | * ram_load_setup: Setup RAM for migration incoming side | |
3564 | * | |
3565 | * Returns zero to indicate success and negative for error | |
3566 | * | |
3567 | * @f: QEMUFile where to receive the data | |
3568 | * @opaque: RAMState pointer | |
3569 | */ | |
3570 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
3571 | { | |
3572 | xbzrle_load_setup(); | |
f9494614 | 3573 | ramblock_recv_map_init(); |
13af18f2 | 3574 | |
f265e0e4 JQ |
3575 | return 0; |
3576 | } | |
3577 | ||
3578 | static int ram_load_cleanup(void *opaque) | |
3579 | { | |
f9494614 | 3580 | RAMBlock *rb; |
56eb90af | 3581 | |
fbd162e6 | 3582 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 3583 | qemu_ram_block_writeback(rb); |
56eb90af JH |
3584 | } |
3585 | ||
f265e0e4 | 3586 | xbzrle_load_cleanup(); |
f9494614 | 3587 | |
fbd162e6 | 3588 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
3589 | g_free(rb->receivedmap); |
3590 | rb->receivedmap = NULL; | |
3591 | } | |
13af18f2 | 3592 | |
f265e0e4 JQ |
3593 | return 0; |
3594 | } | |
3595 | ||
3d0684b2 JQ |
3596 | /** |
3597 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
3598 | * | |
3599 | * Returns 0 for success and negative if there was one error | |
3600 | * | |
3601 | * @mis: current migration incoming state | |
3602 | * | |
3603 | * Allocate data structures etc needed by incoming migration with | |
3604 | * postcopy-ram. postcopy-ram's similarly names | |
3605 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
3606 | */ |
3607 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
3608 | { | |
c136180c | 3609 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
3610 | } |
3611 | ||
3d0684b2 JQ |
3612 | /** |
3613 | * ram_load_postcopy: load a page in postcopy case | |
3614 | * | |
3615 | * Returns 0 for success or -errno in case of error | |
3616 | * | |
a7180877 DDAG |
3617 | * Called in postcopy mode by ram_load(). |
3618 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
3619 | * |
3620 | * @f: QEMUFile where to send the data | |
36f62f11 | 3621 | * @channel: the channel to use for loading |
a7180877 | 3622 | */ |
36f62f11 | 3623 | int ram_load_postcopy(QEMUFile *f, int channel) |
a7180877 DDAG |
3624 | { |
3625 | int flags = 0, ret = 0; | |
3626 | bool place_needed = false; | |
1aa83678 | 3627 | bool matches_target_page_size = false; |
a7180877 | 3628 | MigrationIncomingState *mis = migration_incoming_get_current(); |
36f62f11 | 3629 | PostcopyTmpPage *tmp_page = &mis->postcopy_tmp_pages[channel]; |
a7180877 DDAG |
3630 | |
3631 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
3632 | ram_addr_t addr; | |
a7180877 DDAG |
3633 | void *page_buffer = NULL; |
3634 | void *place_source = NULL; | |
df9ff5e1 | 3635 | RAMBlock *block = NULL; |
a7180877 | 3636 | uint8_t ch; |
644acf99 | 3637 | int len; |
a7180877 DDAG |
3638 | |
3639 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
3640 | |
3641 | /* | |
3642 | * If qemu file error, we should stop here, and then "addr" | |
3643 | * may be invalid | |
3644 | */ | |
3645 | ret = qemu_file_get_error(f); | |
3646 | if (ret) { | |
3647 | break; | |
3648 | } | |
3649 | ||
a7180877 DDAG |
3650 | flags = addr & ~TARGET_PAGE_MASK; |
3651 | addr &= TARGET_PAGE_MASK; | |
3652 | ||
36f62f11 | 3653 | trace_ram_load_postcopy_loop(channel, (uint64_t)addr, flags); |
644acf99 WY |
3654 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
3655 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
c01b16ed | 3656 | block = ram_block_from_stream(mis, f, flags, channel); |
6a23f639 DH |
3657 | if (!block) { |
3658 | ret = -EINVAL; | |
3659 | break; | |
3660 | } | |
4c4bad48 | 3661 | |
898ba906 DH |
3662 | /* |
3663 | * Relying on used_length is racy and can result in false positives. | |
3664 | * We might place pages beyond used_length in case RAM was shrunk | |
3665 | * while in postcopy, which is fine - trying to place via | |
3666 | * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault. | |
3667 | */ | |
3668 | if (!block->host || addr >= block->postcopy_length) { | |
a7180877 DDAG |
3669 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); |
3670 | ret = -EINVAL; | |
3671 | break; | |
3672 | } | |
77dadc3f | 3673 | tmp_page->target_pages++; |
1aa83678 | 3674 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 3675 | /* |
28abd200 DDAG |
3676 | * Postcopy requires that we place whole host pages atomically; |
3677 | * these may be huge pages for RAMBlocks that are backed by | |
3678 | * hugetlbfs. | |
a7180877 DDAG |
3679 | * To make it atomic, the data is read into a temporary page |
3680 | * that's moved into place later. | |
3681 | * The migration protocol uses, possibly smaller, target-pages | |
3682 | * however the source ensures it always sends all the components | |
91ba442f | 3683 | * of a host page in one chunk. |
a7180877 | 3684 | */ |
77dadc3f | 3685 | page_buffer = tmp_page->tmp_huge_page + |
6a23f639 DH |
3686 | host_page_offset_from_ram_block_offset(block, addr); |
3687 | /* If all TP are zero then we can optimise the place */ | |
77dadc3f PX |
3688 | if (tmp_page->target_pages == 1) { |
3689 | tmp_page->host_addr = | |
3690 | host_page_from_ram_block_offset(block, addr); | |
3691 | } else if (tmp_page->host_addr != | |
3692 | host_page_from_ram_block_offset(block, addr)) { | |
c53b7ddc | 3693 | /* not the 1st TP within the HP */ |
36f62f11 | 3694 | error_report("Non-same host page detected on channel %d: " |
cfc7dc8a PX |
3695 | "Target host page %p, received host page %p " |
3696 | "(rb %s offset 0x"RAM_ADDR_FMT" target_pages %d)", | |
36f62f11 | 3697 | channel, tmp_page->host_addr, |
cfc7dc8a PX |
3698 | host_page_from_ram_block_offset(block, addr), |
3699 | block->idstr, addr, tmp_page->target_pages); | |
6a23f639 DH |
3700 | ret = -EINVAL; |
3701 | break; | |
a7180877 DDAG |
3702 | } |
3703 | ||
3704 | /* | |
3705 | * If it's the last part of a host page then we place the host | |
3706 | * page | |
3707 | */ | |
77dadc3f PX |
3708 | if (tmp_page->target_pages == |
3709 | (block->page_size / TARGET_PAGE_SIZE)) { | |
4cbb3c63 | 3710 | place_needed = true; |
4cbb3c63 | 3711 | } |
77dadc3f | 3712 | place_source = tmp_page->tmp_huge_page; |
a7180877 DDAG |
3713 | } |
3714 | ||
3715 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 3716 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 3717 | ch = qemu_get_byte(f); |
2e36bc1b WY |
3718 | /* |
3719 | * Can skip to set page_buffer when | |
3720 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
3721 | */ | |
3722 | if (ch || !matches_target_page_size) { | |
3723 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
3724 | } | |
a7180877 | 3725 | if (ch) { |
77dadc3f | 3726 | tmp_page->all_zero = false; |
a7180877 DDAG |
3727 | } |
3728 | break; | |
3729 | ||
3730 | case RAM_SAVE_FLAG_PAGE: | |
77dadc3f | 3731 | tmp_page->all_zero = false; |
1aa83678 PX |
3732 | if (!matches_target_page_size) { |
3733 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
3734 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
3735 | } else { | |
1aa83678 PX |
3736 | /* |
3737 | * For small pages that matches target page size, we | |
3738 | * avoid the qemu_file copy. Instead we directly use | |
3739 | * the buffer of QEMUFile to place the page. Note: we | |
3740 | * cannot do any QEMUFile operation before using that | |
3741 | * buffer to make sure the buffer is valid when | |
3742 | * placing the page. | |
a7180877 DDAG |
3743 | */ |
3744 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
3745 | TARGET_PAGE_SIZE); | |
3746 | } | |
3747 | break; | |
644acf99 | 3748 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
77dadc3f | 3749 | tmp_page->all_zero = false; |
644acf99 WY |
3750 | len = qemu_get_be32(f); |
3751 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
3752 | error_report("Invalid compressed data length: %d", len); | |
3753 | ret = -EINVAL; | |
3754 | break; | |
3755 | } | |
3756 | decompress_data_with_multi_threads(f, page_buffer, len); | |
3757 | break; | |
294e5a40 JQ |
3758 | case RAM_SAVE_FLAG_MULTIFD_FLUSH: |
3759 | multifd_recv_sync_main(); | |
3760 | break; | |
a7180877 DDAG |
3761 | case RAM_SAVE_FLAG_EOS: |
3762 | /* normal exit */ | |
b05292c2 JQ |
3763 | if (migrate_multifd_flush_after_each_section()) { |
3764 | multifd_recv_sync_main(); | |
3765 | } | |
a7180877 DDAG |
3766 | break; |
3767 | default: | |
29fccade | 3768 | error_report("Unknown combination of migration flags: 0x%x" |
a7180877 DDAG |
3769 | " (postcopy mode)", flags); |
3770 | ret = -EINVAL; | |
7a9ddfbf PX |
3771 | break; |
3772 | } | |
3773 | ||
644acf99 WY |
3774 | /* Got the whole host page, wait for decompress before placing. */ |
3775 | if (place_needed) { | |
3776 | ret |= wait_for_decompress_done(); | |
3777 | } | |
3778 | ||
7a9ddfbf PX |
3779 | /* Detect for any possible file errors */ |
3780 | if (!ret && qemu_file_get_error(f)) { | |
3781 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
3782 | } |
3783 | ||
7a9ddfbf | 3784 | if (!ret && place_needed) { |
77dadc3f PX |
3785 | if (tmp_page->all_zero) { |
3786 | ret = postcopy_place_page_zero(mis, tmp_page->host_addr, block); | |
a7180877 | 3787 | } else { |
77dadc3f PX |
3788 | ret = postcopy_place_page(mis, tmp_page->host_addr, |
3789 | place_source, block); | |
a7180877 | 3790 | } |
ddf35bdf | 3791 | place_needed = false; |
77dadc3f | 3792 | postcopy_temp_page_reset(tmp_page); |
a7180877 | 3793 | } |
a7180877 DDAG |
3794 | } |
3795 | ||
3796 | return ret; | |
3797 | } | |
3798 | ||
acab30b8 DHB |
3799 | static bool postcopy_is_running(void) |
3800 | { | |
3801 | PostcopyState ps = postcopy_state_get(); | |
3802 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
3803 | } | |
3804 | ||
e6f4aa18 ZC |
3805 | /* |
3806 | * Flush content of RAM cache into SVM's memory. | |
3807 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
3808 | */ | |
24fa16f8 | 3809 | void colo_flush_ram_cache(void) |
e6f4aa18 ZC |
3810 | { |
3811 | RAMBlock *block = NULL; | |
3812 | void *dst_host; | |
3813 | void *src_host; | |
3814 | unsigned long offset = 0; | |
3815 | ||
1e493be5 | 3816 | memory_global_dirty_log_sync(false); |
9d638407 | 3817 | qemu_mutex_lock(&ram_state->bitmap_mutex); |
89ac5a1d DDAG |
3818 | WITH_RCU_READ_LOCK_GUARD() { |
3819 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3820 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3821 | } | |
d1955d22 | 3822 | } |
d1955d22 | 3823 | |
e6f4aa18 | 3824 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
3825 | WITH_RCU_READ_LOCK_GUARD() { |
3826 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 3827 | |
89ac5a1d | 3828 | while (block) { |
a6a83cef | 3829 | unsigned long num = 0; |
e6f4aa18 | 3830 | |
a6a83cef | 3831 | offset = colo_bitmap_find_dirty(ram_state, block, offset, &num); |
542147f4 DH |
3832 | if (!offset_in_ramblock(block, |
3833 | ((ram_addr_t)offset) << TARGET_PAGE_BITS)) { | |
89ac5a1d | 3834 | offset = 0; |
a6a83cef | 3835 | num = 0; |
89ac5a1d DDAG |
3836 | block = QLIST_NEXT_RCU(block, next); |
3837 | } else { | |
a6a83cef RL |
3838 | unsigned long i = 0; |
3839 | ||
3840 | for (i = 0; i < num; i++) { | |
3841 | migration_bitmap_clear_dirty(ram_state, block, offset + i); | |
3842 | } | |
8bba004c AR |
3843 | dst_host = block->host |
3844 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
3845 | src_host = block->colo_cache | |
3846 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
a6a83cef RL |
3847 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE * num); |
3848 | offset += num; | |
89ac5a1d | 3849 | } |
e6f4aa18 ZC |
3850 | } |
3851 | } | |
9d638407 | 3852 | qemu_mutex_unlock(&ram_state->bitmap_mutex); |
e6f4aa18 ZC |
3853 | trace_colo_flush_ram_cache_end(); |
3854 | } | |
3855 | ||
10da4a36 WY |
3856 | /** |
3857 | * ram_load_precopy: load pages in precopy case | |
3858 | * | |
3859 | * Returns 0 for success or -errno in case of error | |
3860 | * | |
3861 | * Called in precopy mode by ram_load(). | |
3862 | * rcu_read_lock is taken prior to this being called. | |
3863 | * | |
3864 | * @f: QEMUFile where to send the data | |
3865 | */ | |
3866 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 3867 | { |
755e8d7c | 3868 | MigrationIncomingState *mis = migration_incoming_get_current(); |
e65cec5e | 3869 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 3870 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
80fe315c | 3871 | bool postcopy_advised = migration_incoming_postcopy_advised(); |
a7a94d14 | 3872 | if (!migrate_compress()) { |
edc60127 JQ |
3873 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; |
3874 | } | |
a7180877 | 3875 | |
10da4a36 | 3876 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 3877 | ram_addr_t addr, total_ram_bytes; |
0393031a | 3878 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
3879 | uint8_t ch; |
3880 | ||
e65cec5e YK |
3881 | /* |
3882 | * Yield periodically to let main loop run, but an iteration of | |
3883 | * the main loop is expensive, so do it each some iterations | |
3884 | */ | |
3885 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
3886 | aio_co_schedule(qemu_get_current_aio_context(), | |
3887 | qemu_coroutine_self()); | |
3888 | qemu_coroutine_yield(); | |
3889 | } | |
3890 | i++; | |
3891 | ||
56e93d26 JQ |
3892 | addr = qemu_get_be64(f); |
3893 | flags = addr & ~TARGET_PAGE_MASK; | |
3894 | addr &= TARGET_PAGE_MASK; | |
3895 | ||
edc60127 JQ |
3896 | if (flags & invalid_flags) { |
3897 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
3898 | error_report("Received an unexpected compressed page"); | |
3899 | } | |
3900 | ||
3901 | ret = -EINVAL; | |
3902 | break; | |
3903 | } | |
3904 | ||
bb890ed5 | 3905 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 3906 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
c01b16ed PX |
3907 | RAMBlock *block = ram_block_from_stream(mis, f, flags, |
3908 | RAM_CHANNEL_PRECOPY); | |
4c4bad48 | 3909 | |
0393031a | 3910 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 3911 | /* |
0393031a HZ |
3912 | * After going into COLO stage, we should not load the page |
3913 | * into SVM's memory directly, we put them into colo_cache firstly. | |
3914 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
3915 | * Previously, we copied all these memory in preparing stage of COLO | |
3916 | * while we need to stop VM, which is a time-consuming process. | |
3917 | * Here we optimize it by a trick, back-up every page while in | |
3918 | * migration process while COLO is enabled, though it affects the | |
3919 | * speed of the migration, but it obviously reduce the downtime of | |
3920 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 3921 | */ |
0393031a HZ |
3922 | if (migration_incoming_colo_enabled()) { |
3923 | if (migration_incoming_in_colo_state()) { | |
3924 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 3925 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
3926 | } else { |
3927 | /* | |
3928 | * In migration stage but before COLO stage, | |
3929 | * Put all pages into both cache and SVM's memory. | |
3930 | */ | |
8af66371 | 3931 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 3932 | } |
13af18f2 | 3933 | } |
a776aa15 DDAG |
3934 | if (!host) { |
3935 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
3936 | ret = -EINVAL; | |
3937 | break; | |
3938 | } | |
13af18f2 ZC |
3939 | if (!migration_incoming_in_colo_state()) { |
3940 | ramblock_recv_bitmap_set(block, host); | |
3941 | } | |
3942 | ||
1db9d8e5 | 3943 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
3944 | } |
3945 | ||
56e93d26 JQ |
3946 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
3947 | case RAM_SAVE_FLAG_MEM_SIZE: | |
3948 | /* Synchronize RAM block list */ | |
3949 | total_ram_bytes = addr; | |
3950 | while (!ret && total_ram_bytes) { | |
3951 | RAMBlock *block; | |
56e93d26 JQ |
3952 | char id[256]; |
3953 | ram_addr_t length; | |
3954 | ||
3955 | len = qemu_get_byte(f); | |
3956 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3957 | id[len] = 0; | |
3958 | length = qemu_get_be64(f); | |
3959 | ||
e3dd7493 | 3960 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
3961 | if (block && !qemu_ram_is_migratable(block)) { |
3962 | error_report("block %s should not be migrated !", id); | |
3963 | ret = -EINVAL; | |
3964 | } else if (block) { | |
e3dd7493 DDAG |
3965 | if (length != block->used_length) { |
3966 | Error *local_err = NULL; | |
56e93d26 | 3967 | |
fa53a0e5 | 3968 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
3969 | &local_err); |
3970 | if (local_err) { | |
3971 | error_report_err(local_err); | |
56e93d26 | 3972 | } |
56e93d26 | 3973 | } |
ef08fb38 | 3974 | /* For postcopy we need to check hugepage sizes match */ |
e846b746 | 3975 | if (postcopy_advised && migrate_postcopy_ram() && |
ef08fb38 DDAG |
3976 | block->page_size != qemu_host_page_size) { |
3977 | uint64_t remote_page_size = qemu_get_be64(f); | |
3978 | if (remote_page_size != block->page_size) { | |
3979 | error_report("Mismatched RAM page size %s " | |
3980 | "(local) %zd != %" PRId64, | |
3981 | id, block->page_size, | |
3982 | remote_page_size); | |
3983 | ret = -EINVAL; | |
3984 | } | |
3985 | } | |
fbd162e6 YK |
3986 | if (migrate_ignore_shared()) { |
3987 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
3988 | if (ramblock_is_ignored(block) && |
3989 | block->mr->addr != addr) { | |
3990 | error_report("Mismatched GPAs for block %s " | |
3991 | "%" PRId64 "!= %" PRId64, | |
3992 | id, (uint64_t)addr, | |
3993 | (uint64_t)block->mr->addr); | |
3994 | ret = -EINVAL; | |
3995 | } | |
3996 | } | |
e3dd7493 DDAG |
3997 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
3998 | block->idstr); | |
3999 | } else { | |
56e93d26 JQ |
4000 | error_report("Unknown ramblock \"%s\", cannot " |
4001 | "accept migration", id); | |
4002 | ret = -EINVAL; | |
4003 | } | |
4004 | ||
4005 | total_ram_bytes -= length; | |
4006 | } | |
4007 | break; | |
a776aa15 | 4008 | |
bb890ed5 | 4009 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
4010 | ch = qemu_get_byte(f); |
4011 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
4012 | break; | |
a776aa15 | 4013 | |
56e93d26 | 4014 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
4015 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
4016 | break; | |
56e93d26 | 4017 | |
a776aa15 | 4018 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
4019 | len = qemu_get_be32(f); |
4020 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4021 | error_report("Invalid compressed data length: %d", len); | |
4022 | ret = -EINVAL; | |
4023 | break; | |
4024 | } | |
c1bc6626 | 4025 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 4026 | break; |
a776aa15 | 4027 | |
56e93d26 | 4028 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
4029 | if (load_xbzrle(f, addr, host) < 0) { |
4030 | error_report("Failed to decompress XBZRLE page at " | |
4031 | RAM_ADDR_FMT, addr); | |
4032 | ret = -EINVAL; | |
4033 | break; | |
4034 | } | |
4035 | break; | |
294e5a40 JQ |
4036 | case RAM_SAVE_FLAG_MULTIFD_FLUSH: |
4037 | multifd_recv_sync_main(); | |
4038 | break; | |
56e93d26 JQ |
4039 | case RAM_SAVE_FLAG_EOS: |
4040 | /* normal exit */ | |
b05292c2 JQ |
4041 | if (migrate_multifd_flush_after_each_section()) { |
4042 | multifd_recv_sync_main(); | |
4043 | } | |
56e93d26 | 4044 | break; |
5f1e7540 JQ |
4045 | case RAM_SAVE_FLAG_HOOK: |
4046 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); | |
4047 | break; | |
56e93d26 | 4048 | default: |
5f1e7540 JQ |
4049 | error_report("Unknown combination of migration flags: 0x%x", flags); |
4050 | ret = -EINVAL; | |
56e93d26 JQ |
4051 | } |
4052 | if (!ret) { | |
4053 | ret = qemu_file_get_error(f); | |
4054 | } | |
0393031a HZ |
4055 | if (!ret && host_bak) { |
4056 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
4057 | } | |
56e93d26 JQ |
4058 | } |
4059 | ||
ca1a6b70 | 4060 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
4061 | return ret; |
4062 | } | |
4063 | ||
4064 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
4065 | { | |
4066 | int ret = 0; | |
4067 | static uint64_t seq_iter; | |
4068 | /* | |
4069 | * If system is running in postcopy mode, page inserts to host memory must | |
4070 | * be atomic | |
4071 | */ | |
4072 | bool postcopy_running = postcopy_is_running(); | |
4073 | ||
4074 | seq_iter++; | |
4075 | ||
4076 | if (version_id != 4) { | |
4077 | return -EINVAL; | |
4078 | } | |
4079 | ||
4080 | /* | |
4081 | * This RCU critical section can be very long running. | |
4082 | * When RCU reclaims in the code start to become numerous, | |
4083 | * it will be necessary to reduce the granularity of this | |
4084 | * critical section. | |
4085 | */ | |
89ac5a1d DDAG |
4086 | WITH_RCU_READ_LOCK_GUARD() { |
4087 | if (postcopy_running) { | |
36f62f11 PX |
4088 | /* |
4089 | * Note! Here RAM_CHANNEL_PRECOPY is the precopy channel of | |
4090 | * postcopy migration, we have another RAM_CHANNEL_POSTCOPY to | |
4091 | * service fast page faults. | |
4092 | */ | |
4093 | ret = ram_load_postcopy(f, RAM_CHANNEL_PRECOPY); | |
89ac5a1d DDAG |
4094 | } else { |
4095 | ret = ram_load_precopy(f); | |
4096 | } | |
10da4a36 | 4097 | } |
55c4446b | 4098 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 | 4099 | |
56e93d26 JQ |
4100 | return ret; |
4101 | } | |
4102 | ||
c6467627 VSO |
4103 | static bool ram_has_postcopy(void *opaque) |
4104 | { | |
469dd51b | 4105 | RAMBlock *rb; |
fbd162e6 | 4106 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
4107 | if (ramblock_is_pmem(rb)) { |
4108 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
4109 | "is not supported now!", rb->idstr, rb->host); | |
4110 | return false; | |
4111 | } | |
4112 | } | |
4113 | ||
c6467627 VSO |
4114 | return migrate_postcopy_ram(); |
4115 | } | |
4116 | ||
edd090c7 PX |
4117 | /* Sync all the dirty bitmap with destination VM. */ |
4118 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
4119 | { | |
4120 | RAMBlock *block; | |
4121 | QEMUFile *file = s->to_dst_file; | |
4122 | int ramblock_count = 0; | |
4123 | ||
4124 | trace_ram_dirty_bitmap_sync_start(); | |
4125 | ||
fbd162e6 | 4126 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
4127 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
4128 | trace_ram_dirty_bitmap_request(block->idstr); | |
4129 | ramblock_count++; | |
4130 | } | |
4131 | ||
4132 | trace_ram_dirty_bitmap_sync_wait(); | |
4133 | ||
4134 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
4135 | while (ramblock_count--) { | |
4136 | qemu_sem_wait(&s->rp_state.rp_sem); | |
4137 | } | |
4138 | ||
4139 | trace_ram_dirty_bitmap_sync_complete(); | |
4140 | ||
4141 | return 0; | |
4142 | } | |
4143 | ||
4144 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
4145 | { | |
4146 | qemu_sem_post(&s->rp_state.rp_sem); | |
4147 | } | |
4148 | ||
a335debb PX |
4149 | /* |
4150 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
4151 | * This is only used when the postcopy migration is paused but wants | |
4152 | * to resume from a middle point. | |
4153 | */ | |
4154 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
4155 | { | |
4156 | int ret = -EINVAL; | |
43044ac0 | 4157 | /* from_dst_file is always valid because we're within rp_thread */ |
a335debb PX |
4158 | QEMUFile *file = s->rp_state.from_dst_file; |
4159 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 4160 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
4161 | uint64_t size, end_mark; |
4162 | ||
4163 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
4164 | ||
4165 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
4166 | error_report("%s: incorrect state %s", __func__, | |
4167 | MigrationStatus_str(s->state)); | |
4168 | return -EINVAL; | |
4169 | } | |
4170 | ||
4171 | /* | |
4172 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3a4452d8 | 4173 | * need the endianness conversion, and the paddings. |
a335debb PX |
4174 | */ |
4175 | local_size = ROUND_UP(local_size, 8); | |
4176 | ||
4177 | /* Add paddings */ | |
4178 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
4179 | ||
4180 | size = qemu_get_be64(file); | |
4181 | ||
4182 | /* The size of the bitmap should match with our ramblock */ | |
4183 | if (size != local_size) { | |
4184 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
4185 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
4186 | block->idstr, size, local_size); | |
4187 | ret = -EINVAL; | |
4188 | goto out; | |
4189 | } | |
4190 | ||
4191 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
4192 | end_mark = qemu_get_be64(file); | |
4193 | ||
4194 | ret = qemu_file_get_error(file); | |
4195 | if (ret || size != local_size) { | |
4196 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
4197 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
4198 | __func__, block->idstr, ret, local_size, size); | |
4199 | ret = -EIO; | |
4200 | goto out; | |
4201 | } | |
4202 | ||
4203 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
af3bbbe9 | 4204 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64, |
a335debb PX |
4205 | __func__, block->idstr, end_mark); |
4206 | ret = -EINVAL; | |
4207 | goto out; | |
4208 | } | |
4209 | ||
4210 | /* | |
3a4452d8 | 4211 | * Endianness conversion. We are during postcopy (though paused). |
a335debb PX |
4212 | * The dirty bitmap won't change. We can directly modify it. |
4213 | */ | |
4214 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
4215 | ||
4216 | /* | |
4217 | * What we received is "received bitmap". Revert it as the initial | |
4218 | * dirty bitmap for this ramblock. | |
4219 | */ | |
4220 | bitmap_complement(block->bmap, block->bmap, nbits); | |
4221 | ||
be39b4cd DH |
4222 | /* Clear dirty bits of discarded ranges that we don't want to migrate. */ |
4223 | ramblock_dirty_bitmap_clear_discarded_pages(block); | |
4224 | ||
4225 | /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */ | |
a335debb PX |
4226 | trace_ram_dirty_bitmap_reload_complete(block->idstr); |
4227 | ||
edd090c7 PX |
4228 | /* |
4229 | * We succeeded to sync bitmap for current ramblock. If this is | |
4230 | * the last one to sync, we need to notify the main send thread. | |
4231 | */ | |
4232 | ram_dirty_bitmap_reload_notify(s); | |
4233 | ||
a335debb PX |
4234 | ret = 0; |
4235 | out: | |
bf269906 | 4236 | g_free(le_bitmap); |
a335debb PX |
4237 | return ret; |
4238 | } | |
4239 | ||
edd090c7 PX |
4240 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
4241 | { | |
4242 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 4243 | int ret; |
edd090c7 | 4244 | |
08614f34 PX |
4245 | ret = ram_dirty_bitmap_sync_all(s, rs); |
4246 | if (ret) { | |
4247 | return ret; | |
4248 | } | |
4249 | ||
4250 | ram_state_resume_prepare(rs, s->to_dst_file); | |
4251 | ||
4252 | return 0; | |
edd090c7 PX |
4253 | } |
4254 | ||
36f62f11 PX |
4255 | void postcopy_preempt_shutdown_file(MigrationState *s) |
4256 | { | |
4257 | qemu_put_be64(s->postcopy_qemufile_src, RAM_SAVE_FLAG_EOS); | |
4258 | qemu_fflush(s->postcopy_qemufile_src); | |
4259 | } | |
4260 | ||
56e93d26 | 4261 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 4262 | .save_setup = ram_save_setup, |
56e93d26 | 4263 | .save_live_iterate = ram_save_iterate, |
763c906b | 4264 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 4265 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 4266 | .has_postcopy = ram_has_postcopy, |
c8df4a7a JQ |
4267 | .state_pending_exact = ram_state_pending_exact, |
4268 | .state_pending_estimate = ram_state_pending_estimate, | |
56e93d26 | 4269 | .load_state = ram_load, |
f265e0e4 JQ |
4270 | .save_cleanup = ram_save_cleanup, |
4271 | .load_setup = ram_load_setup, | |
4272 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 4273 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
4274 | }; |
4275 | ||
c7c0e724 DH |
4276 | static void ram_mig_ram_block_resized(RAMBlockNotifier *n, void *host, |
4277 | size_t old_size, size_t new_size) | |
4278 | { | |
cc61c703 | 4279 | PostcopyState ps = postcopy_state_get(); |
c7c0e724 DH |
4280 | ram_addr_t offset; |
4281 | RAMBlock *rb = qemu_ram_block_from_host(host, false, &offset); | |
4282 | Error *err = NULL; | |
4283 | ||
4284 | if (ramblock_is_ignored(rb)) { | |
4285 | return; | |
4286 | } | |
4287 | ||
4288 | if (!migration_is_idle()) { | |
4289 | /* | |
4290 | * Precopy code on the source cannot deal with the size of RAM blocks | |
4291 | * changing at random points in time - especially after sending the | |
4292 | * RAM block sizes in the migration stream, they must no longer change. | |
4293 | * Abort and indicate a proper reason. | |
4294 | */ | |
4295 | error_setg(&err, "RAM block '%s' resized during precopy.", rb->idstr); | |
458fecca | 4296 | migration_cancel(err); |
c7c0e724 | 4297 | error_free(err); |
c7c0e724 | 4298 | } |
cc61c703 DH |
4299 | |
4300 | switch (ps) { | |
4301 | case POSTCOPY_INCOMING_ADVISE: | |
4302 | /* | |
4303 | * Update what ram_postcopy_incoming_init()->init_range() does at the | |
4304 | * time postcopy was advised. Syncing RAM blocks with the source will | |
4305 | * result in RAM resizes. | |
4306 | */ | |
4307 | if (old_size < new_size) { | |
4308 | if (ram_discard_range(rb->idstr, old_size, new_size - old_size)) { | |
4309 | error_report("RAM block '%s' discard of resized RAM failed", | |
4310 | rb->idstr); | |
4311 | } | |
4312 | } | |
898ba906 | 4313 | rb->postcopy_length = new_size; |
cc61c703 DH |
4314 | break; |
4315 | case POSTCOPY_INCOMING_NONE: | |
4316 | case POSTCOPY_INCOMING_RUNNING: | |
4317 | case POSTCOPY_INCOMING_END: | |
4318 | /* | |
4319 | * Once our guest is running, postcopy does no longer care about | |
4320 | * resizes. When growing, the new memory was not available on the | |
4321 | * source, no handler needed. | |
4322 | */ | |
4323 | break; | |
4324 | default: | |
4325 | error_report("RAM block '%s' resized during postcopy state: %d", | |
4326 | rb->idstr, ps); | |
4327 | exit(-1); | |
4328 | } | |
c7c0e724 DH |
4329 | } |
4330 | ||
4331 | static RAMBlockNotifier ram_mig_ram_notifier = { | |
4332 | .ram_block_resized = ram_mig_ram_block_resized, | |
4333 | }; | |
4334 | ||
56e93d26 JQ |
4335 | void ram_mig_init(void) |
4336 | { | |
4337 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 4338 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
c7c0e724 | 4339 | ram_block_notifier_add(&ram_mig_ram_notifier); |
56e93d26 | 4340 | } |