]>
Commit | Line | Data |
---|---|---|
56e93d26 JQ |
1 | /* |
2 | * QEMU System Emulator | |
3 | * | |
4 | * Copyright (c) 2003-2008 Fabrice Bellard | |
76cc7b58 JQ |
5 | * Copyright (c) 2011-2015 Red Hat Inc |
6 | * | |
7 | * Authors: | |
8 | * Juan Quintela <quintela@redhat.com> | |
56e93d26 JQ |
9 | * |
10 | * Permission is hereby granted, free of charge, to any person obtaining a copy | |
11 | * of this software and associated documentation files (the "Software"), to deal | |
12 | * in the Software without restriction, including without limitation the rights | |
13 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell | |
14 | * copies of the Software, and to permit persons to whom the Software is | |
15 | * furnished to do so, subject to the following conditions: | |
16 | * | |
17 | * The above copyright notice and this permission notice shall be included in | |
18 | * all copies or substantial portions of the Software. | |
19 | * | |
20 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
21 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
22 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
23 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | |
24 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | |
25 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | |
26 | * THE SOFTWARE. | |
27 | */ | |
e688df6b | 28 | |
1393a485 | 29 | #include "qemu/osdep.h" |
f348b6d1 | 30 | #include "qemu/cutils.h" |
56e93d26 JQ |
31 | #include "qemu/bitops.h" |
32 | #include "qemu/bitmap.h" | |
b85ea5fa | 33 | #include "qemu/madvise.h" |
7205c9ec | 34 | #include "qemu/main-loop.h" |
c0e0825c | 35 | #include "io/channel-null.h" |
709e3fe8 | 36 | #include "xbzrle.h" |
7b1e1a22 | 37 | #include "ram.h" |
6666c96a | 38 | #include "migration.h" |
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 | ||
76e03000 XG |
483 | CompressionStats compression_counters; |
484 | ||
97274a87 LS |
485 | enum CompressResult { |
486 | RES_NONE = 0, | |
487 | RES_ZEROPAGE = 1, | |
488 | RES_COMPRESS = 2 | |
489 | }; | |
490 | typedef enum CompressResult CompressResult; | |
491 | ||
56e93d26 | 492 | struct CompressParam { |
56e93d26 | 493 | bool done; |
90e56fb4 | 494 | bool quit; |
10c2f7b7 | 495 | bool trigger; |
97274a87 | 496 | CompressResult result; |
56e93d26 JQ |
497 | QEMUFile *file; |
498 | QemuMutex mutex; | |
499 | QemuCond cond; | |
500 | RAMBlock *block; | |
501 | ram_addr_t offset; | |
34ab9e97 XG |
502 | |
503 | /* internally used fields */ | |
dcaf446e | 504 | z_stream stream; |
34ab9e97 | 505 | uint8_t *originbuf; |
56e93d26 JQ |
506 | }; |
507 | typedef struct CompressParam CompressParam; | |
508 | ||
509 | struct DecompressParam { | |
73a8912b | 510 | bool done; |
90e56fb4 | 511 | bool quit; |
56e93d26 JQ |
512 | QemuMutex mutex; |
513 | QemuCond cond; | |
514 | void *des; | |
d341d9f3 | 515 | uint8_t *compbuf; |
56e93d26 | 516 | int len; |
797ca154 | 517 | z_stream stream; |
56e93d26 JQ |
518 | }; |
519 | typedef struct DecompressParam DecompressParam; | |
520 | ||
521 | static CompressParam *comp_param; | |
522 | static QemuThread *compress_threads; | |
523 | /* comp_done_cond is used to wake up the migration thread when | |
524 | * one of the compression threads has finished the compression. | |
525 | * comp_done_lock is used to co-work with comp_done_cond. | |
526 | */ | |
0d9f9a5c LL |
527 | static QemuMutex comp_done_lock; |
528 | static QemuCond comp_done_cond; | |
56e93d26 | 529 | |
34ab9e97 | 530 | static QEMUFile *decomp_file; |
56e93d26 JQ |
531 | static DecompressParam *decomp_param; |
532 | static QemuThread *decompress_threads; | |
73a8912b LL |
533 | static QemuMutex decomp_done_lock; |
534 | static QemuCond decomp_done_cond; | |
56e93d26 | 535 | |
93589827 PX |
536 | static int ram_save_host_page_urgent(PageSearchStatus *pss); |
537 | ||
97274a87 LS |
538 | static CompressResult do_compress_ram_page(QEMUFile *f, z_stream *stream, |
539 | RAMBlock *block, ram_addr_t offset, | |
540 | uint8_t *source_buf); | |
56e93d26 | 541 | |
ebd88a49 PX |
542 | /* NOTE: page is the PFN not real ram_addr_t. */ |
543 | static void pss_init(PageSearchStatus *pss, RAMBlock *rb, ram_addr_t page) | |
544 | { | |
545 | pss->block = rb; | |
546 | pss->page = page; | |
547 | pss->complete_round = false; | |
548 | } | |
549 | ||
93589827 PX |
550 | /* |
551 | * Check whether two PSSs are actively sending the same page. Return true | |
552 | * if it is, false otherwise. | |
553 | */ | |
554 | static bool pss_overlap(PageSearchStatus *pss1, PageSearchStatus *pss2) | |
555 | { | |
556 | return pss1->host_page_sending && pss2->host_page_sending && | |
557 | (pss1->host_page_start == pss2->host_page_start); | |
558 | } | |
559 | ||
56e93d26 JQ |
560 | static void *do_data_compress(void *opaque) |
561 | { | |
562 | CompressParam *param = opaque; | |
a7a9a88f LL |
563 | RAMBlock *block; |
564 | ram_addr_t offset; | |
97274a87 | 565 | CompressResult result; |
56e93d26 | 566 | |
a7a9a88f | 567 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 568 | while (!param->quit) { |
10c2f7b7 | 569 | if (param->trigger) { |
a7a9a88f LL |
570 | block = param->block; |
571 | offset = param->offset; | |
10c2f7b7 | 572 | param->trigger = false; |
a7a9a88f LL |
573 | qemu_mutex_unlock(¶m->mutex); |
574 | ||
97274a87 LS |
575 | result = do_compress_ram_page(param->file, ¶m->stream, |
576 | block, offset, param->originbuf); | |
a7a9a88f | 577 | |
0d9f9a5c | 578 | qemu_mutex_lock(&comp_done_lock); |
a7a9a88f | 579 | param->done = true; |
97274a87 | 580 | param->result = result; |
0d9f9a5c LL |
581 | qemu_cond_signal(&comp_done_cond); |
582 | qemu_mutex_unlock(&comp_done_lock); | |
a7a9a88f LL |
583 | |
584 | qemu_mutex_lock(¶m->mutex); | |
585 | } else { | |
56e93d26 JQ |
586 | qemu_cond_wait(¶m->cond, ¶m->mutex); |
587 | } | |
56e93d26 | 588 | } |
a7a9a88f | 589 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
590 | |
591 | return NULL; | |
592 | } | |
593 | ||
f0afa331 | 594 | static void compress_threads_save_cleanup(void) |
56e93d26 JQ |
595 | { |
596 | int i, thread_count; | |
597 | ||
a7a94d14 | 598 | if (!migrate_compress() || !comp_param) { |
56e93d26 JQ |
599 | return; |
600 | } | |
05306935 | 601 | |
56e93d26 JQ |
602 | thread_count = migrate_compress_threads(); |
603 | for (i = 0; i < thread_count; i++) { | |
dcaf446e XG |
604 | /* |
605 | * we use it as a indicator which shows if the thread is | |
606 | * properly init'd or not | |
607 | */ | |
608 | if (!comp_param[i].file) { | |
609 | break; | |
610 | } | |
05306935 FL |
611 | |
612 | qemu_mutex_lock(&comp_param[i].mutex); | |
613 | comp_param[i].quit = true; | |
614 | qemu_cond_signal(&comp_param[i].cond); | |
615 | qemu_mutex_unlock(&comp_param[i].mutex); | |
616 | ||
56e93d26 | 617 | qemu_thread_join(compress_threads + i); |
56e93d26 JQ |
618 | qemu_mutex_destroy(&comp_param[i].mutex); |
619 | qemu_cond_destroy(&comp_param[i].cond); | |
dcaf446e | 620 | deflateEnd(&comp_param[i].stream); |
34ab9e97 | 621 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
622 | qemu_fclose(comp_param[i].file); |
623 | comp_param[i].file = NULL; | |
56e93d26 | 624 | } |
0d9f9a5c LL |
625 | qemu_mutex_destroy(&comp_done_lock); |
626 | qemu_cond_destroy(&comp_done_cond); | |
56e93d26 JQ |
627 | g_free(compress_threads); |
628 | g_free(comp_param); | |
56e93d26 JQ |
629 | compress_threads = NULL; |
630 | comp_param = NULL; | |
56e93d26 JQ |
631 | } |
632 | ||
dcaf446e | 633 | static int compress_threads_save_setup(void) |
56e93d26 JQ |
634 | { |
635 | int i, thread_count; | |
636 | ||
a7a94d14 | 637 | if (!migrate_compress()) { |
dcaf446e | 638 | return 0; |
56e93d26 | 639 | } |
56e93d26 JQ |
640 | thread_count = migrate_compress_threads(); |
641 | compress_threads = g_new0(QemuThread, thread_count); | |
642 | comp_param = g_new0(CompressParam, thread_count); | |
0d9f9a5c LL |
643 | qemu_cond_init(&comp_done_cond); |
644 | qemu_mutex_init(&comp_done_lock); | |
56e93d26 | 645 | for (i = 0; i < thread_count; i++) { |
34ab9e97 XG |
646 | comp_param[i].originbuf = g_try_malloc(TARGET_PAGE_SIZE); |
647 | if (!comp_param[i].originbuf) { | |
648 | goto exit; | |
649 | } | |
650 | ||
dcaf446e XG |
651 | if (deflateInit(&comp_param[i].stream, |
652 | migrate_compress_level()) != Z_OK) { | |
34ab9e97 | 653 | g_free(comp_param[i].originbuf); |
dcaf446e XG |
654 | goto exit; |
655 | } | |
656 | ||
e110aa91 C |
657 | /* comp_param[i].file is just used as a dummy buffer to save data, |
658 | * set its ops to empty. | |
56e93d26 | 659 | */ |
77ef2dc1 | 660 | comp_param[i].file = qemu_file_new_output( |
c0e0825c | 661 | QIO_CHANNEL(qio_channel_null_new())); |
56e93d26 | 662 | comp_param[i].done = true; |
90e56fb4 | 663 | comp_param[i].quit = false; |
56e93d26 JQ |
664 | qemu_mutex_init(&comp_param[i].mutex); |
665 | qemu_cond_init(&comp_param[i].cond); | |
666 | qemu_thread_create(compress_threads + i, "compress", | |
667 | do_data_compress, comp_param + i, | |
668 | QEMU_THREAD_JOINABLE); | |
669 | } | |
dcaf446e XG |
670 | return 0; |
671 | ||
672 | exit: | |
673 | compress_threads_save_cleanup(); | |
674 | return -1; | |
56e93d26 JQ |
675 | } |
676 | ||
677 | /** | |
3d0684b2 | 678 | * save_page_header: write page header to wire |
56e93d26 JQ |
679 | * |
680 | * If this is the 1st block, it also writes the block identification | |
681 | * | |
3d0684b2 | 682 | * Returns the number of bytes written |
56e93d26 | 683 | * |
ec6f3ab9 | 684 | * @pss: current PSS channel status |
56e93d26 JQ |
685 | * @block: block that contains the page we want to send |
686 | * @offset: offset inside the block for the page | |
687 | * in the lower bits, it contains flags | |
688 | */ | |
37502df3 LS |
689 | static size_t save_page_header(PageSearchStatus *pss, QEMUFile *f, |
690 | RAMBlock *block, ram_addr_t offset) | |
56e93d26 | 691 | { |
9f5f380b | 692 | size_t size, len; |
ec6f3ab9 | 693 | bool same_block = (block == pss->last_sent_block); |
56e93d26 | 694 | |
10661f11 | 695 | if (same_block) { |
24795694 JQ |
696 | offset |= RAM_SAVE_FLAG_CONTINUE; |
697 | } | |
2bf3aa85 | 698 | qemu_put_be64(f, offset); |
56e93d26 JQ |
699 | size = 8; |
700 | ||
10661f11 | 701 | if (!same_block) { |
9f5f380b | 702 | len = strlen(block->idstr); |
2bf3aa85 JQ |
703 | qemu_put_byte(f, len); |
704 | qemu_put_buffer(f, (uint8_t *)block->idstr, len); | |
9f5f380b | 705 | size += 1 + len; |
ec6f3ab9 | 706 | pss->last_sent_block = block; |
56e93d26 JQ |
707 | } |
708 | return size; | |
709 | } | |
710 | ||
3d0684b2 | 711 | /** |
179a8080 | 712 | * mig_throttle_guest_down: throttle down the guest |
3d0684b2 JQ |
713 | * |
714 | * Reduce amount of guest cpu execution to hopefully slow down memory | |
715 | * writes. If guest dirty memory rate is reduced below the rate at | |
716 | * which we can transfer pages to the destination then we should be | |
717 | * able to complete migration. Some workloads dirty memory way too | |
718 | * fast and will not effectively converge, even with auto-converge. | |
070afca2 | 719 | */ |
cbbf8182 KZ |
720 | static void mig_throttle_guest_down(uint64_t bytes_dirty_period, |
721 | uint64_t bytes_dirty_threshold) | |
070afca2 | 722 | { |
2a8ec380 | 723 | uint64_t pct_initial = migrate_cpu_throttle_initial(); |
9605c2ac | 724 | uint64_t pct_increment = migrate_cpu_throttle_increment(); |
873f674c | 725 | bool pct_tailslow = migrate_cpu_throttle_tailslow(); |
24155bd0 | 726 | int pct_max = migrate_max_cpu_throttle(); |
070afca2 | 727 | |
cbbf8182 KZ |
728 | uint64_t throttle_now = cpu_throttle_get_percentage(); |
729 | uint64_t cpu_now, cpu_ideal, throttle_inc; | |
730 | ||
070afca2 JH |
731 | /* We have not started throttling yet. Let's start it. */ |
732 | if (!cpu_throttle_active()) { | |
733 | cpu_throttle_set(pct_initial); | |
734 | } else { | |
735 | /* Throttling already on, just increase the rate */ | |
cbbf8182 KZ |
736 | if (!pct_tailslow) { |
737 | throttle_inc = pct_increment; | |
738 | } else { | |
739 | /* Compute the ideal CPU percentage used by Guest, which may | |
740 | * make the dirty rate match the dirty rate threshold. */ | |
741 | cpu_now = 100 - throttle_now; | |
742 | cpu_ideal = cpu_now * (bytes_dirty_threshold * 1.0 / | |
743 | bytes_dirty_period); | |
744 | throttle_inc = MIN(cpu_now - cpu_ideal, pct_increment); | |
745 | } | |
746 | cpu_throttle_set(MIN(throttle_now + throttle_inc, pct_max)); | |
070afca2 JH |
747 | } |
748 | } | |
749 | ||
91fe9a8d RL |
750 | void mig_throttle_counter_reset(void) |
751 | { | |
752 | RAMState *rs = ram_state; | |
753 | ||
754 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
755 | rs->num_dirty_pages_period = 0; | |
aff3f660 | 756 | rs->bytes_xfer_prev = stat64_get(&mig_stats.transferred); |
91fe9a8d RL |
757 | } |
758 | ||
3d0684b2 JQ |
759 | /** |
760 | * xbzrle_cache_zero_page: insert a zero page in the XBZRLE cache | |
761 | * | |
6f37bb8b | 762 | * @rs: current RAM state |
3d0684b2 JQ |
763 | * @current_addr: address for the zero page |
764 | * | |
765 | * Update the xbzrle cache to reflect a page that's been sent as all 0. | |
56e93d26 JQ |
766 | * The important thing is that a stale (not-yet-0'd) page be replaced |
767 | * by the new data. | |
768 | * As a bonus, if the page wasn't in the cache it gets added so that | |
3d0684b2 | 769 | * when a small write is made into the 0'd page it gets XBZRLE sent. |
56e93d26 | 770 | */ |
6f37bb8b | 771 | static void xbzrle_cache_zero_page(RAMState *rs, ram_addr_t current_addr) |
56e93d26 | 772 | { |
56e93d26 JQ |
773 | /* We don't care if this fails to allocate a new cache page |
774 | * as long as it updated an old one */ | |
c00e0928 | 775 | cache_insert(XBZRLE.cache, current_addr, XBZRLE.zero_target_page, |
aff3f660 | 776 | stat64_get(&mig_stats.dirty_sync_count)); |
56e93d26 JQ |
777 | } |
778 | ||
779 | #define ENCODING_FLAG_XBZRLE 0x1 | |
780 | ||
781 | /** | |
782 | * save_xbzrle_page: compress and send current page | |
783 | * | |
784 | * Returns: 1 means that we wrote the page | |
785 | * 0 means that page is identical to the one already sent | |
786 | * -1 means that xbzrle would be longer than normal | |
787 | * | |
5a987738 | 788 | * @rs: current RAM state |
ec6f3ab9 | 789 | * @pss: current PSS channel |
3d0684b2 JQ |
790 | * @current_data: pointer to the address of the page contents |
791 | * @current_addr: addr of the page | |
56e93d26 JQ |
792 | * @block: block that contains the page we want to send |
793 | * @offset: offset inside the block for the page | |
56e93d26 | 794 | */ |
ec6f3ab9 | 795 | static int save_xbzrle_page(RAMState *rs, PageSearchStatus *pss, |
61717ea9 PX |
796 | uint8_t **current_data, ram_addr_t current_addr, |
797 | RAMBlock *block, ram_addr_t offset) | |
56e93d26 JQ |
798 | { |
799 | int encoded_len = 0, bytes_xbzrle; | |
800 | uint8_t *prev_cached_page; | |
ec6f3ab9 | 801 | QEMUFile *file = pss->pss_channel; |
aff3f660 | 802 | uint64_t generation = stat64_get(&mig_stats.dirty_sync_count); |
56e93d26 | 803 | |
536b5a4e | 804 | if (!cache_is_cached(XBZRLE.cache, current_addr, generation)) { |
9360447d | 805 | xbzrle_counters.cache_miss++; |
05931ec5 | 806 | if (!rs->last_stage) { |
56e93d26 | 807 | if (cache_insert(XBZRLE.cache, current_addr, *current_data, |
536b5a4e | 808 | generation) == -1) { |
56e93d26 JQ |
809 | return -1; |
810 | } else { | |
811 | /* update *current_data when the page has been | |
812 | inserted into cache */ | |
813 | *current_data = get_cached_data(XBZRLE.cache, current_addr); | |
814 | } | |
815 | } | |
816 | return -1; | |
817 | } | |
818 | ||
e460a4b1 WW |
819 | /* |
820 | * Reaching here means the page has hit the xbzrle cache, no matter what | |
821 | * encoding result it is (normal encoding, overflow or skipping the page), | |
3a4452d8 | 822 | * count the page as encoded. This is used to calculate the encoding rate. |
e460a4b1 WW |
823 | * |
824 | * Example: 2 pages (8KB) being encoded, first page encoding generates 2KB, | |
825 | * 2nd page turns out to be skipped (i.e. no new bytes written to the | |
826 | * page), the overall encoding rate will be 8KB / 2KB = 4, which has the | |
827 | * skipped page included. In this way, the encoding rate can tell if the | |
828 | * guest page is good for xbzrle encoding. | |
829 | */ | |
830 | xbzrle_counters.pages++; | |
56e93d26 JQ |
831 | prev_cached_page = get_cached_data(XBZRLE.cache, current_addr); |
832 | ||
833 | /* save current buffer into memory */ | |
834 | memcpy(XBZRLE.current_buf, *current_data, TARGET_PAGE_SIZE); | |
835 | ||
836 | /* XBZRLE encoding (if there is no overflow) */ | |
04ffce13 | 837 | encoded_len = xbzrle_encode_buffer_func(prev_cached_page, XBZRLE.current_buf, |
838 | TARGET_PAGE_SIZE, XBZRLE.encoded_buf, | |
839 | TARGET_PAGE_SIZE); | |
ca353803 WY |
840 | |
841 | /* | |
842 | * Update the cache contents, so that it corresponds to the data | |
843 | * sent, in all cases except where we skip the page. | |
844 | */ | |
05931ec5 | 845 | if (!rs->last_stage && encoded_len != 0) { |
ca353803 WY |
846 | memcpy(prev_cached_page, XBZRLE.current_buf, TARGET_PAGE_SIZE); |
847 | /* | |
848 | * In the case where we couldn't compress, ensure that the caller | |
849 | * sends the data from the cache, since the guest might have | |
850 | * changed the RAM since we copied it. | |
851 | */ | |
852 | *current_data = prev_cached_page; | |
853 | } | |
854 | ||
56e93d26 | 855 | if (encoded_len == 0) { |
55c4446b | 856 | trace_save_xbzrle_page_skipping(); |
56e93d26 JQ |
857 | return 0; |
858 | } else if (encoded_len == -1) { | |
55c4446b | 859 | trace_save_xbzrle_page_overflow(); |
9360447d | 860 | xbzrle_counters.overflow++; |
e460a4b1 | 861 | xbzrle_counters.bytes += TARGET_PAGE_SIZE; |
56e93d26 JQ |
862 | return -1; |
863 | } | |
864 | ||
56e93d26 | 865 | /* Send XBZRLE based compressed page */ |
37502df3 | 866 | bytes_xbzrle = save_page_header(pss, pss->pss_channel, block, |
204b88b8 | 867 | offset | RAM_SAVE_FLAG_XBZRLE); |
61717ea9 PX |
868 | qemu_put_byte(file, ENCODING_FLAG_XBZRLE); |
869 | qemu_put_be16(file, encoded_len); | |
870 | qemu_put_buffer(file, XBZRLE.encoded_buf, encoded_len); | |
56e93d26 | 871 | bytes_xbzrle += encoded_len + 1 + 2; |
e460a4b1 WW |
872 | /* |
873 | * Like compressed_size (please see update_compress_thread_counts), | |
874 | * the xbzrle encoded bytes don't count the 8 byte header with | |
875 | * RAM_SAVE_FLAG_CONTINUE. | |
876 | */ | |
877 | xbzrle_counters.bytes += bytes_xbzrle - 8; | |
4c2d0f6d | 878 | ram_transferred_add(bytes_xbzrle); |
56e93d26 JQ |
879 | |
880 | return 1; | |
881 | } | |
882 | ||
3d0684b2 | 883 | /** |
d9e474ea | 884 | * pss_find_next_dirty: find the next dirty page of current ramblock |
f3f491fc | 885 | * |
d9e474ea PX |
886 | * This function updates pss->page to point to the next dirty page index |
887 | * within the ramblock to migrate, or the end of ramblock when nothing | |
888 | * found. Note that when pss->host_page_sending==true it means we're | |
889 | * during sending a host page, so we won't look for dirty page that is | |
890 | * outside the host page boundary. | |
3d0684b2 | 891 | * |
d9e474ea | 892 | * @pss: the current page search status |
f3f491fc | 893 | */ |
d9e474ea | 894 | static void pss_find_next_dirty(PageSearchStatus *pss) |
56e93d26 | 895 | { |
d9e474ea | 896 | RAMBlock *rb = pss->block; |
6b6712ef JQ |
897 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; |
898 | unsigned long *bitmap = rb->bmap; | |
56e93d26 | 899 | |
fbd162e6 | 900 | if (ramblock_is_ignored(rb)) { |
d9e474ea PX |
901 | /* Points directly to the end, so we know no dirty page */ |
902 | pss->page = size; | |
903 | return; | |
904 | } | |
905 | ||
906 | /* | |
907 | * If during sending a host page, only look for dirty pages within the | |
908 | * current host page being send. | |
909 | */ | |
910 | if (pss->host_page_sending) { | |
911 | assert(pss->host_page_end); | |
912 | size = MIN(size, pss->host_page_end); | |
b895de50 CLG |
913 | } |
914 | ||
d9e474ea | 915 | pss->page = find_next_bit(bitmap, size, pss->page); |
56e93d26 JQ |
916 | } |
917 | ||
1230a25f | 918 | static void migration_clear_memory_region_dirty_bitmap(RAMBlock *rb, |
3143577d WW |
919 | unsigned long page) |
920 | { | |
921 | uint8_t shift; | |
922 | hwaddr size, start; | |
923 | ||
924 | if (!rb->clear_bmap || !clear_bmap_test_and_clear(rb, page)) { | |
925 | return; | |
926 | } | |
927 | ||
928 | shift = rb->clear_bmap_shift; | |
929 | /* | |
930 | * CLEAR_BITMAP_SHIFT_MIN should always guarantee this... this | |
931 | * can make things easier sometimes since then start address | |
932 | * of the small chunk will always be 64 pages aligned so the | |
933 | * bitmap will always be aligned to unsigned long. We should | |
934 | * even be able to remove this restriction but I'm simply | |
935 | * keeping it. | |
936 | */ | |
937 | assert(shift >= 6); | |
938 | ||
939 | size = 1ULL << (TARGET_PAGE_BITS + shift); | |
7648297d | 940 | start = QEMU_ALIGN_DOWN((ram_addr_t)page << TARGET_PAGE_BITS, size); |
3143577d WW |
941 | trace_migration_bitmap_clear_dirty(rb->idstr, start, size, page); |
942 | memory_region_clear_dirty_bitmap(rb->mr, start, size); | |
943 | } | |
944 | ||
945 | static void | |
1230a25f | 946 | migration_clear_memory_region_dirty_bitmap_range(RAMBlock *rb, |
3143577d WW |
947 | unsigned long start, |
948 | unsigned long npages) | |
949 | { | |
950 | unsigned long i, chunk_pages = 1UL << rb->clear_bmap_shift; | |
951 | unsigned long chunk_start = QEMU_ALIGN_DOWN(start, chunk_pages); | |
952 | unsigned long chunk_end = QEMU_ALIGN_UP(start + npages, chunk_pages); | |
953 | ||
954 | /* | |
955 | * Clear pages from start to start + npages - 1, so the end boundary is | |
956 | * exclusive. | |
957 | */ | |
958 | for (i = chunk_start; i < chunk_end; i += chunk_pages) { | |
1230a25f | 959 | migration_clear_memory_region_dirty_bitmap(rb, i); |
3143577d WW |
960 | } |
961 | } | |
962 | ||
a6a83cef RL |
963 | /* |
964 | * colo_bitmap_find_diry:find contiguous dirty pages from start | |
965 | * | |
966 | * Returns the page offset within memory region of the start of the contiguout | |
967 | * dirty page | |
968 | * | |
969 | * @rs: current RAM state | |
970 | * @rb: RAMBlock where to search for dirty pages | |
971 | * @start: page where we start the search | |
972 | * @num: the number of contiguous dirty pages | |
973 | */ | |
974 | static inline | |
975 | unsigned long colo_bitmap_find_dirty(RAMState *rs, RAMBlock *rb, | |
976 | unsigned long start, unsigned long *num) | |
977 | { | |
978 | unsigned long size = rb->used_length >> TARGET_PAGE_BITS; | |
979 | unsigned long *bitmap = rb->bmap; | |
980 | unsigned long first, next; | |
981 | ||
982 | *num = 0; | |
983 | ||
984 | if (ramblock_is_ignored(rb)) { | |
985 | return size; | |
986 | } | |
987 | ||
988 | first = find_next_bit(bitmap, size, start); | |
989 | if (first >= size) { | |
990 | return first; | |
991 | } | |
992 | next = find_next_zero_bit(bitmap, size, first + 1); | |
993 | assert(next >= first); | |
994 | *num = next - first; | |
995 | return first; | |
996 | } | |
997 | ||
06b10688 | 998 | static inline bool migration_bitmap_clear_dirty(RAMState *rs, |
f20e2865 JQ |
999 | RAMBlock *rb, |
1000 | unsigned long page) | |
a82d593b DDAG |
1001 | { |
1002 | bool ret; | |
a82d593b | 1003 | |
002cad6b PX |
1004 | /* |
1005 | * Clear dirty bitmap if needed. This _must_ be called before we | |
1006 | * send any of the page in the chunk because we need to make sure | |
1007 | * we can capture further page content changes when we sync dirty | |
1008 | * log the next time. So as long as we are going to send any of | |
1009 | * the page in the chunk we clear the remote dirty bitmap for all. | |
1010 | * Clearing it earlier won't be a problem, but too late will. | |
1011 | */ | |
1230a25f | 1012 | migration_clear_memory_region_dirty_bitmap(rb, page); |
002cad6b | 1013 | |
6b6712ef | 1014 | ret = test_and_clear_bit(page, rb->bmap); |
a82d593b | 1015 | if (ret) { |
0d8ec885 | 1016 | rs->migration_dirty_pages--; |
a82d593b | 1017 | } |
386a907b | 1018 | |
a82d593b DDAG |
1019 | return ret; |
1020 | } | |
1021 | ||
be39b4cd DH |
1022 | static void dirty_bitmap_clear_section(MemoryRegionSection *section, |
1023 | void *opaque) | |
1024 | { | |
1025 | const hwaddr offset = section->offset_within_region; | |
1026 | const hwaddr size = int128_get64(section->size); | |
1027 | const unsigned long start = offset >> TARGET_PAGE_BITS; | |
1028 | const unsigned long npages = size >> TARGET_PAGE_BITS; | |
1029 | RAMBlock *rb = section->mr->ram_block; | |
1030 | uint64_t *cleared_bits = opaque; | |
1031 | ||
1032 | /* | |
1033 | * We don't grab ram_state->bitmap_mutex because we expect to run | |
1034 | * only when starting migration or during postcopy recovery where | |
1035 | * we don't have concurrent access. | |
1036 | */ | |
1037 | if (!migration_in_postcopy() && !migrate_background_snapshot()) { | |
1038 | migration_clear_memory_region_dirty_bitmap_range(rb, start, npages); | |
1039 | } | |
1040 | *cleared_bits += bitmap_count_one_with_offset(rb->bmap, start, npages); | |
1041 | bitmap_clear(rb->bmap, start, npages); | |
1042 | } | |
1043 | ||
1044 | /* | |
1045 | * Exclude all dirty pages from migration that fall into a discarded range as | |
1046 | * managed by a RamDiscardManager responsible for the mapped memory region of | |
1047 | * the RAMBlock. Clear the corresponding bits in the dirty bitmaps. | |
1048 | * | |
1049 | * Discarded pages ("logically unplugged") have undefined content and must | |
1050 | * not get migrated, because even reading these pages for migration might | |
1051 | * result in undesired behavior. | |
1052 | * | |
1053 | * Returns the number of cleared bits in the RAMBlock dirty bitmap. | |
1054 | * | |
1055 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1056 | */ | |
1057 | static uint64_t ramblock_dirty_bitmap_clear_discarded_pages(RAMBlock *rb) | |
1058 | { | |
1059 | uint64_t cleared_bits = 0; | |
1060 | ||
1061 | if (rb->mr && rb->bmap && memory_region_has_ram_discard_manager(rb->mr)) { | |
1062 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1063 | MemoryRegionSection section = { | |
1064 | .mr = rb->mr, | |
1065 | .offset_within_region = 0, | |
1066 | .size = int128_make64(qemu_ram_get_used_length(rb)), | |
1067 | }; | |
1068 | ||
1069 | ram_discard_manager_replay_discarded(rdm, §ion, | |
1070 | dirty_bitmap_clear_section, | |
1071 | &cleared_bits); | |
1072 | } | |
1073 | return cleared_bits; | |
1074 | } | |
1075 | ||
9470c5e0 DH |
1076 | /* |
1077 | * Check if a host-page aligned page falls into a discarded range as managed by | |
1078 | * a RamDiscardManager responsible for the mapped memory region of the RAMBlock. | |
1079 | * | |
1080 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1081 | */ | |
1082 | bool ramblock_page_is_discarded(RAMBlock *rb, ram_addr_t start) | |
1083 | { | |
1084 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1085 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1086 | MemoryRegionSection section = { | |
1087 | .mr = rb->mr, | |
1088 | .offset_within_region = start, | |
1089 | .size = int128_make64(qemu_ram_pagesize(rb)), | |
1090 | }; | |
1091 | ||
1092 | return !ram_discard_manager_is_populated(rdm, §ion); | |
1093 | } | |
1094 | return false; | |
1095 | } | |
1096 | ||
267691b6 | 1097 | /* Called with RCU critical section */ |
7a3e9571 | 1098 | static void ramblock_sync_dirty_bitmap(RAMState *rs, RAMBlock *rb) |
56e93d26 | 1099 | { |
fb613580 KZ |
1100 | uint64_t new_dirty_pages = |
1101 | cpu_physical_memory_sync_dirty_bitmap(rb, 0, rb->used_length); | |
1102 | ||
1103 | rs->migration_dirty_pages += new_dirty_pages; | |
1104 | rs->num_dirty_pages_period += new_dirty_pages; | |
56e93d26 JQ |
1105 | } |
1106 | ||
3d0684b2 JQ |
1107 | /** |
1108 | * ram_pagesize_summary: calculate all the pagesizes of a VM | |
1109 | * | |
1110 | * Returns a summary bitmap of the page sizes of all RAMBlocks | |
1111 | * | |
1112 | * For VMs with just normal pages this is equivalent to the host page | |
1113 | * size. If it's got some huge pages then it's the OR of all the | |
1114 | * different page sizes. | |
e8ca1db2 DDAG |
1115 | */ |
1116 | uint64_t ram_pagesize_summary(void) | |
1117 | { | |
1118 | RAMBlock *block; | |
1119 | uint64_t summary = 0; | |
1120 | ||
fbd162e6 | 1121 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
e8ca1db2 DDAG |
1122 | summary |= block->page_size; |
1123 | } | |
1124 | ||
1125 | return summary; | |
1126 | } | |
1127 | ||
aecbfe9c XG |
1128 | uint64_t ram_get_total_transferred_pages(void) |
1129 | { | |
aff3f660 JQ |
1130 | return stat64_get(&mig_stats.normal_pages) + |
1131 | stat64_get(&mig_stats.zero_pages) + | |
23b7576d | 1132 | compression_counters.pages + xbzrle_counters.pages; |
aecbfe9c XG |
1133 | } |
1134 | ||
b734035b XG |
1135 | static void migration_update_rates(RAMState *rs, int64_t end_time) |
1136 | { | |
be8b02ed | 1137 | uint64_t page_count = rs->target_page_count - rs->target_page_count_prev; |
76e03000 | 1138 | double compressed_size; |
b734035b XG |
1139 | |
1140 | /* calculate period counters */ | |
aff3f660 | 1141 | stat64_set(&mig_stats.dirty_pages_rate, |
72f8e587 JQ |
1142 | rs->num_dirty_pages_period * 1000 / |
1143 | (end_time - rs->time_last_bitmap_sync)); | |
b734035b | 1144 | |
be8b02ed | 1145 | if (!page_count) { |
b734035b XG |
1146 | return; |
1147 | } | |
1148 | ||
87dca0c9 | 1149 | if (migrate_xbzrle()) { |
e460a4b1 WW |
1150 | double encoded_size, unencoded_size; |
1151 | ||
b734035b | 1152 | xbzrle_counters.cache_miss_rate = (double)(xbzrle_counters.cache_miss - |
be8b02ed | 1153 | rs->xbzrle_cache_miss_prev) / page_count; |
b734035b | 1154 | rs->xbzrle_cache_miss_prev = xbzrle_counters.cache_miss; |
e460a4b1 WW |
1155 | unencoded_size = (xbzrle_counters.pages - rs->xbzrle_pages_prev) * |
1156 | TARGET_PAGE_SIZE; | |
1157 | encoded_size = xbzrle_counters.bytes - rs->xbzrle_bytes_prev; | |
92271402 | 1158 | if (xbzrle_counters.pages == rs->xbzrle_pages_prev || !encoded_size) { |
e460a4b1 | 1159 | xbzrle_counters.encoding_rate = 0; |
e460a4b1 WW |
1160 | } else { |
1161 | xbzrle_counters.encoding_rate = unencoded_size / encoded_size; | |
1162 | } | |
1163 | rs->xbzrle_pages_prev = xbzrle_counters.pages; | |
1164 | rs->xbzrle_bytes_prev = xbzrle_counters.bytes; | |
b734035b | 1165 | } |
76e03000 | 1166 | |
a7a94d14 | 1167 | if (migrate_compress()) { |
76e03000 XG |
1168 | compression_counters.busy_rate = (double)(compression_counters.busy - |
1169 | rs->compress_thread_busy_prev) / page_count; | |
1170 | rs->compress_thread_busy_prev = compression_counters.busy; | |
1171 | ||
1172 | compressed_size = compression_counters.compressed_size - | |
1173 | rs->compressed_size_prev; | |
1174 | if (compressed_size) { | |
1175 | double uncompressed_size = (compression_counters.pages - | |
1176 | rs->compress_pages_prev) * TARGET_PAGE_SIZE; | |
1177 | ||
1178 | /* Compression-Ratio = Uncompressed-size / Compressed-size */ | |
1179 | compression_counters.compression_rate = | |
1180 | uncompressed_size / compressed_size; | |
1181 | ||
1182 | rs->compress_pages_prev = compression_counters.pages; | |
1183 | rs->compressed_size_prev = compression_counters.compressed_size; | |
1184 | } | |
1185 | } | |
b734035b XG |
1186 | } |
1187 | ||
dc14a470 KZ |
1188 | static void migration_trigger_throttle(RAMState *rs) |
1189 | { | |
6499efdb | 1190 | uint64_t threshold = migrate_throttle_trigger_threshold(); |
23b7576d | 1191 | uint64_t bytes_xfer_period = |
aff3f660 | 1192 | stat64_get(&mig_stats.transferred) - rs->bytes_xfer_prev; |
dc14a470 KZ |
1193 | uint64_t bytes_dirty_period = rs->num_dirty_pages_period * TARGET_PAGE_SIZE; |
1194 | uint64_t bytes_dirty_threshold = bytes_xfer_period * threshold / 100; | |
1195 | ||
1196 | /* During block migration the auto-converge logic incorrectly detects | |
1197 | * that ram migration makes no progress. Avoid this by disabling the | |
1198 | * throttling logic during the bulk phase of block migration. */ | |
1199 | if (migrate_auto_converge() && !blk_mig_bulk_active()) { | |
1200 | /* The following detection logic can be refined later. For now: | |
1201 | Check to see if the ratio between dirtied bytes and the approx. | |
1202 | amount of bytes that just got transferred since the last time | |
1203 | we were in this routine reaches the threshold. If that happens | |
1204 | twice, start or increase throttling. */ | |
1205 | ||
1206 | if ((bytes_dirty_period > bytes_dirty_threshold) && | |
1207 | (++rs->dirty_rate_high_cnt >= 2)) { | |
1208 | trace_migration_throttle(); | |
1209 | rs->dirty_rate_high_cnt = 0; | |
cbbf8182 KZ |
1210 | mig_throttle_guest_down(bytes_dirty_period, |
1211 | bytes_dirty_threshold); | |
dc14a470 KZ |
1212 | } |
1213 | } | |
1214 | } | |
1215 | ||
8d820d6f | 1216 | static void migration_bitmap_sync(RAMState *rs) |
56e93d26 JQ |
1217 | { |
1218 | RAMBlock *block; | |
56e93d26 | 1219 | int64_t end_time; |
56e93d26 | 1220 | |
aff3f660 | 1221 | stat64_add(&mig_stats.dirty_sync_count, 1); |
56e93d26 | 1222 | |
f664da80 JQ |
1223 | if (!rs->time_last_bitmap_sync) { |
1224 | rs->time_last_bitmap_sync = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
56e93d26 JQ |
1225 | } |
1226 | ||
1227 | trace_migration_bitmap_sync_start(); | |
9c1f8f44 | 1228 | memory_global_dirty_log_sync(); |
56e93d26 | 1229 | |
108cfae0 | 1230 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
1231 | WITH_RCU_READ_LOCK_GUARD() { |
1232 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
1233 | ramblock_sync_dirty_bitmap(rs, block); | |
1234 | } | |
aff3f660 | 1235 | stat64_set(&mig_stats.dirty_bytes_last_sync, ram_bytes_remaining()); |
56e93d26 | 1236 | } |
108cfae0 | 1237 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 1238 | |
9458a9a1 | 1239 | memory_global_after_dirty_log_sync(); |
a66cd90c | 1240 | trace_migration_bitmap_sync_end(rs->num_dirty_pages_period); |
1ffb5dfd | 1241 | |
56e93d26 JQ |
1242 | end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); |
1243 | ||
1244 | /* more than 1 second = 1000 millisecons */ | |
f664da80 | 1245 | if (end_time > rs->time_last_bitmap_sync + 1000) { |
dc14a470 | 1246 | migration_trigger_throttle(rs); |
070afca2 | 1247 | |
b734035b XG |
1248 | migration_update_rates(rs, end_time); |
1249 | ||
be8b02ed | 1250 | rs->target_page_count_prev = rs->target_page_count; |
d693c6f1 FF |
1251 | |
1252 | /* reset period counters */ | |
f664da80 | 1253 | rs->time_last_bitmap_sync = end_time; |
a66cd90c | 1254 | rs->num_dirty_pages_period = 0; |
aff3f660 | 1255 | rs->bytes_xfer_prev = stat64_get(&mig_stats.transferred); |
56e93d26 | 1256 | } |
b890902c | 1257 | if (migrate_events()) { |
aff3f660 | 1258 | uint64_t generation = stat64_get(&mig_stats.dirty_sync_count); |
536b5a4e | 1259 | qapi_event_send_migration_pass(generation); |
4addcd4f | 1260 | } |
56e93d26 JQ |
1261 | } |
1262 | ||
bd227060 WW |
1263 | static void migration_bitmap_sync_precopy(RAMState *rs) |
1264 | { | |
1265 | Error *local_err = NULL; | |
1266 | ||
1267 | /* | |
1268 | * The current notifier usage is just an optimization to migration, so we | |
1269 | * don't stop the normal migration process in the error case. | |
1270 | */ | |
1271 | if (precopy_notify(PRECOPY_NOTIFY_BEFORE_BITMAP_SYNC, &local_err)) { | |
1272 | error_report_err(local_err); | |
b4a1733c | 1273 | local_err = NULL; |
bd227060 WW |
1274 | } |
1275 | ||
1276 | migration_bitmap_sync(rs); | |
1277 | ||
1278 | if (precopy_notify(PRECOPY_NOTIFY_AFTER_BITMAP_SYNC, &local_err)) { | |
1279 | error_report_err(local_err); | |
1280 | } | |
1281 | } | |
1282 | ||
a4dbaf8e | 1283 | void ram_release_page(const char *rbname, uint64_t offset) |
47fe16ff JQ |
1284 | { |
1285 | if (!migrate_release_ram() || !migration_in_postcopy()) { | |
1286 | return; | |
1287 | } | |
1288 | ||
1289 | ram_discard_range(rbname, offset, TARGET_PAGE_SIZE); | |
1290 | } | |
1291 | ||
6c97ec5f XG |
1292 | /** |
1293 | * save_zero_page_to_file: send the zero page to the file | |
1294 | * | |
1295 | * Returns the size of data written to the file, 0 means the page is not | |
1296 | * a zero page | |
1297 | * | |
ec6f3ab9 | 1298 | * @pss: current PSS channel |
6c97ec5f XG |
1299 | * @block: block that contains the page we want to send |
1300 | * @offset: offset inside the block for the page | |
1301 | */ | |
37502df3 | 1302 | static int save_zero_page_to_file(PageSearchStatus *pss, QEMUFile *file, |
6c97ec5f XG |
1303 | RAMBlock *block, ram_addr_t offset) |
1304 | { | |
1305 | uint8_t *p = block->host + offset; | |
1306 | int len = 0; | |
1307 | ||
bad452a7 | 1308 | if (buffer_is_zero(p, TARGET_PAGE_SIZE)) { |
37502df3 | 1309 | len += save_page_header(pss, file, block, offset | RAM_SAVE_FLAG_ZERO); |
6c97ec5f XG |
1310 | qemu_put_byte(file, 0); |
1311 | len += 1; | |
47fe16ff | 1312 | ram_release_page(block->idstr, offset); |
6c97ec5f XG |
1313 | } |
1314 | return len; | |
1315 | } | |
1316 | ||
56e93d26 | 1317 | /** |
3d0684b2 | 1318 | * save_zero_page: send the zero page to the stream |
56e93d26 | 1319 | * |
3d0684b2 | 1320 | * Returns the number of pages written. |
56e93d26 | 1321 | * |
ec6f3ab9 | 1322 | * @pss: current PSS channel |
56e93d26 JQ |
1323 | * @block: block that contains the page we want to send |
1324 | * @offset: offset inside the block for the page | |
56e93d26 | 1325 | */ |
37502df3 | 1326 | static int save_zero_page(PageSearchStatus *pss, QEMUFile *f, RAMBlock *block, |
61717ea9 | 1327 | ram_addr_t offset) |
56e93d26 | 1328 | { |
37502df3 | 1329 | int len = save_zero_page_to_file(pss, f, block, offset); |
56e93d26 | 1330 | |
6c97ec5f | 1331 | if (len) { |
aff3f660 | 1332 | stat64_add(&mig_stats.zero_pages, 1); |
4c2d0f6d | 1333 | ram_transferred_add(len); |
6c97ec5f | 1334 | return 1; |
56e93d26 | 1335 | } |
6c97ec5f | 1336 | return -1; |
56e93d26 JQ |
1337 | } |
1338 | ||
059ff0fb XG |
1339 | /* |
1340 | * @pages: the number of pages written by the control path, | |
1341 | * < 0 - error | |
1342 | * > 0 - number of pages written | |
1343 | * | |
1344 | * Return true if the pages has been saved, otherwise false is returned. | |
1345 | */ | |
61717ea9 PX |
1346 | static bool control_save_page(PageSearchStatus *pss, RAMBlock *block, |
1347 | ram_addr_t offset, int *pages) | |
059ff0fb XG |
1348 | { |
1349 | uint64_t bytes_xmit = 0; | |
1350 | int ret; | |
1351 | ||
1352 | *pages = -1; | |
61717ea9 PX |
1353 | ret = ram_control_save_page(pss->pss_channel, block->offset, offset, |
1354 | TARGET_PAGE_SIZE, &bytes_xmit); | |
059ff0fb XG |
1355 | if (ret == RAM_SAVE_CONTROL_NOT_SUPP) { |
1356 | return false; | |
1357 | } | |
1358 | ||
1359 | if (bytes_xmit) { | |
4c2d0f6d | 1360 | ram_transferred_add(bytes_xmit); |
059ff0fb XG |
1361 | *pages = 1; |
1362 | } | |
1363 | ||
1364 | if (ret == RAM_SAVE_CONTROL_DELAYED) { | |
1365 | return true; | |
1366 | } | |
1367 | ||
1368 | if (bytes_xmit > 0) { | |
aff3f660 | 1369 | stat64_add(&mig_stats.normal_pages, 1); |
059ff0fb | 1370 | } else if (bytes_xmit == 0) { |
aff3f660 | 1371 | stat64_add(&mig_stats.zero_pages, 1); |
059ff0fb XG |
1372 | } |
1373 | ||
1374 | return true; | |
1375 | } | |
1376 | ||
65dacaa0 XG |
1377 | /* |
1378 | * directly send the page to the stream | |
1379 | * | |
1380 | * Returns the number of pages written. | |
1381 | * | |
ec6f3ab9 | 1382 | * @pss: current PSS channel |
65dacaa0 XG |
1383 | * @block: block that contains the page we want to send |
1384 | * @offset: offset inside the block for the page | |
1385 | * @buf: the page to be sent | |
1386 | * @async: send to page asyncly | |
1387 | */ | |
ec6f3ab9 | 1388 | static int save_normal_page(PageSearchStatus *pss, RAMBlock *block, |
61717ea9 | 1389 | ram_addr_t offset, uint8_t *buf, bool async) |
65dacaa0 | 1390 | { |
ec6f3ab9 PX |
1391 | QEMUFile *file = pss->pss_channel; |
1392 | ||
37502df3 | 1393 | ram_transferred_add(save_page_header(pss, pss->pss_channel, block, |
4c2d0f6d | 1394 | offset | RAM_SAVE_FLAG_PAGE)); |
65dacaa0 | 1395 | if (async) { |
61717ea9 | 1396 | qemu_put_buffer_async(file, buf, TARGET_PAGE_SIZE, |
f912ec5b | 1397 | migrate_release_ram() && |
65dacaa0 XG |
1398 | migration_in_postcopy()); |
1399 | } else { | |
61717ea9 | 1400 | qemu_put_buffer(file, buf, TARGET_PAGE_SIZE); |
65dacaa0 | 1401 | } |
4c2d0f6d | 1402 | ram_transferred_add(TARGET_PAGE_SIZE); |
aff3f660 | 1403 | stat64_add(&mig_stats.normal_pages, 1); |
65dacaa0 XG |
1404 | return 1; |
1405 | } | |
1406 | ||
56e93d26 | 1407 | /** |
3d0684b2 | 1408 | * ram_save_page: send the given page to the stream |
56e93d26 | 1409 | * |
3d0684b2 | 1410 | * Returns the number of pages written. |
3fd3c4b3 DDAG |
1411 | * < 0 - error |
1412 | * >=0 - Number of pages written - this might legally be 0 | |
1413 | * if xbzrle noticed the page was the same. | |
56e93d26 | 1414 | * |
6f37bb8b | 1415 | * @rs: current RAM state |
56e93d26 JQ |
1416 | * @block: block that contains the page we want to send |
1417 | * @offset: offset inside the block for the page | |
56e93d26 | 1418 | */ |
05931ec5 | 1419 | static int ram_save_page(RAMState *rs, PageSearchStatus *pss) |
56e93d26 JQ |
1420 | { |
1421 | int pages = -1; | |
56e93d26 | 1422 | uint8_t *p; |
56e93d26 | 1423 | bool send_async = true; |
a08f6890 | 1424 | RAMBlock *block = pss->block; |
8bba004c | 1425 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
059ff0fb | 1426 | ram_addr_t current_addr = block->offset + offset; |
56e93d26 | 1427 | |
2f68e399 | 1428 | p = block->host + offset; |
1db9d8e5 | 1429 | trace_ram_save_page(block->idstr, (uint64_t)offset, p); |
56e93d26 | 1430 | |
56e93d26 | 1431 | XBZRLE_cache_lock(); |
f3095cc8 | 1432 | if (rs->xbzrle_started && !migration_in_postcopy()) { |
ec6f3ab9 | 1433 | pages = save_xbzrle_page(rs, pss, &p, current_addr, |
61717ea9 | 1434 | block, offset); |
05931ec5 | 1435 | if (!rs->last_stage) { |
059ff0fb XG |
1436 | /* Can't send this cached data async, since the cache page |
1437 | * might get updated before it gets to the wire | |
56e93d26 | 1438 | */ |
059ff0fb | 1439 | send_async = false; |
56e93d26 JQ |
1440 | } |
1441 | } | |
1442 | ||
1443 | /* XBZRLE overflow or normal page */ | |
1444 | if (pages == -1) { | |
ec6f3ab9 | 1445 | pages = save_normal_page(pss, block, offset, p, send_async); |
56e93d26 JQ |
1446 | } |
1447 | ||
1448 | XBZRLE_cache_unlock(); | |
1449 | ||
1450 | return pages; | |
1451 | } | |
1452 | ||
61717ea9 | 1453 | static int ram_save_multifd_page(QEMUFile *file, RAMBlock *block, |
b9ee2f7d JQ |
1454 | ram_addr_t offset) |
1455 | { | |
61717ea9 | 1456 | if (multifd_queue_page(file, block, offset) < 0) { |
713f762a IR |
1457 | return -1; |
1458 | } | |
aff3f660 | 1459 | stat64_add(&mig_stats.normal_pages, 1); |
b9ee2f7d JQ |
1460 | |
1461 | return 1; | |
1462 | } | |
1463 | ||
97274a87 LS |
1464 | static CompressResult do_compress_ram_page(QEMUFile *f, z_stream *stream, |
1465 | RAMBlock *block, ram_addr_t offset, | |
1466 | uint8_t *source_buf) | |
56e93d26 | 1467 | { |
20d549cb | 1468 | uint8_t *p = block->host + offset; |
6ef3771c | 1469 | int ret; |
56e93d26 | 1470 | |
3e81763e | 1471 | if (buffer_is_zero(p, TARGET_PAGE_SIZE)) { |
97274a87 | 1472 | return RES_ZEROPAGE; |
5e5fdcff XG |
1473 | } |
1474 | ||
34ab9e97 XG |
1475 | /* |
1476 | * copy it to a internal buffer to avoid it being modified by VM | |
1477 | * so that we can catch up the error during compression and | |
1478 | * decompression | |
1479 | */ | |
1480 | memcpy(source_buf, p, TARGET_PAGE_SIZE); | |
6ef3771c XG |
1481 | ret = qemu_put_compression_data(f, stream, source_buf, TARGET_PAGE_SIZE); |
1482 | if (ret < 0) { | |
1483 | qemu_file_set_error(migrate_get_current()->to_dst_file, ret); | |
b3be2896 | 1484 | error_report("compressed data failed!"); |
97274a87 | 1485 | return RES_NONE; |
b3be2896 | 1486 | } |
97274a87 | 1487 | return RES_COMPRESS; |
5e5fdcff XG |
1488 | } |
1489 | ||
1490 | static void | |
1491 | update_compress_thread_counts(const CompressParam *param, int bytes_xmit) | |
1492 | { | |
4c2d0f6d | 1493 | ram_transferred_add(bytes_xmit); |
76e03000 | 1494 | |
97274a87 | 1495 | if (param->result == RES_ZEROPAGE) { |
aff3f660 | 1496 | stat64_add(&mig_stats.zero_pages, 1); |
76e03000 | 1497 | return; |
5e5fdcff | 1498 | } |
76e03000 XG |
1499 | |
1500 | /* 8 means a header with RAM_SAVE_FLAG_CONTINUE. */ | |
1501 | compression_counters.compressed_size += bytes_xmit - 8; | |
1502 | compression_counters.pages++; | |
56e93d26 JQ |
1503 | } |
1504 | ||
32b05495 XG |
1505 | static bool save_page_use_compression(RAMState *rs); |
1506 | ||
b5cf1cd3 LS |
1507 | static inline void compress_reset_result(CompressParam *param) |
1508 | { | |
1509 | param->result = RES_NONE; | |
1510 | param->block = NULL; | |
1511 | param->offset = 0; | |
1512 | } | |
1513 | ||
3e81763e | 1514 | static int send_queued_data(CompressParam *param) |
56e93d26 | 1515 | { |
3e81763e | 1516 | PageSearchStatus *pss = &ram_state->pss[RAM_CHANNEL_PRECOPY]; |
eaa238ab | 1517 | MigrationState *ms = migrate_get_current(); |
3e81763e LS |
1518 | QEMUFile *file = ms->to_dst_file; |
1519 | int len = 0; | |
1520 | ||
1521 | RAMBlock *block = param->block; | |
1522 | ram_addr_t offset = param->offset; | |
1523 | ||
1524 | if (param->result == RES_NONE) { | |
1525 | return 0; | |
1526 | } | |
1527 | ||
1528 | assert(block == pss->last_sent_block); | |
1529 | ||
1530 | if (param->result == RES_ZEROPAGE) { | |
1531 | len += save_page_header(pss, file, block, offset | RAM_SAVE_FLAG_ZERO); | |
1532 | qemu_put_byte(file, 0); | |
1533 | len += 1; | |
1534 | ram_release_page(block->idstr, offset); | |
1535 | } else if (param->result == RES_COMPRESS) { | |
1536 | len += save_page_header(pss, file, block, | |
1537 | offset | RAM_SAVE_FLAG_COMPRESS_PAGE); | |
1538 | len += qemu_put_qemu_file(file, param->file); | |
1539 | } else { | |
1540 | abort(); | |
1541 | } | |
1542 | ||
680628d2 LS |
1543 | update_compress_thread_counts(param, len); |
1544 | ||
3e81763e LS |
1545 | return len; |
1546 | } | |
1547 | ||
ef4f5f5d | 1548 | static void flush_compressed_data(int (send_queued_data(CompressParam *))) |
3e81763e | 1549 | { |
680628d2 | 1550 | int idx, thread_count; |
56e93d26 | 1551 | |
56e93d26 | 1552 | thread_count = migrate_compress_threads(); |
a7a9a88f | 1553 | |
0d9f9a5c | 1554 | qemu_mutex_lock(&comp_done_lock); |
56e93d26 | 1555 | for (idx = 0; idx < thread_count; idx++) { |
a7a9a88f | 1556 | while (!comp_param[idx].done) { |
0d9f9a5c | 1557 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); |
56e93d26 | 1558 | } |
a7a9a88f | 1559 | } |
0d9f9a5c | 1560 | qemu_mutex_unlock(&comp_done_lock); |
a7a9a88f LL |
1561 | |
1562 | for (idx = 0; idx < thread_count; idx++) { | |
1563 | qemu_mutex_lock(&comp_param[idx].mutex); | |
90e56fb4 | 1564 | if (!comp_param[idx].quit) { |
b5cf1cd3 | 1565 | CompressParam *param = &comp_param[idx]; |
680628d2 | 1566 | send_queued_data(param); |
b5cf1cd3 | 1567 | compress_reset_result(param); |
56e93d26 | 1568 | } |
a7a9a88f | 1569 | qemu_mutex_unlock(&comp_param[idx].mutex); |
56e93d26 JQ |
1570 | } |
1571 | } | |
1572 | ||
ef4f5f5d LS |
1573 | static void ram_flush_compressed_data(RAMState *rs) |
1574 | { | |
1575 | if (!save_page_use_compression(rs)) { | |
1576 | return; | |
1577 | } | |
1578 | ||
1579 | flush_compressed_data(send_queued_data); | |
1580 | } | |
1581 | ||
56e93d26 JQ |
1582 | static inline void set_compress_params(CompressParam *param, RAMBlock *block, |
1583 | ram_addr_t offset) | |
1584 | { | |
1585 | param->block = block; | |
1586 | param->offset = offset; | |
10c2f7b7 | 1587 | param->trigger = true; |
56e93d26 JQ |
1588 | } |
1589 | ||
ef4f5f5d LS |
1590 | static int compress_page_with_multi_thread(RAMBlock *block, ram_addr_t offset, |
1591 | int (send_queued_data(CompressParam *))) | |
56e93d26 | 1592 | { |
680628d2 | 1593 | int idx, thread_count, pages = -1; |
1d58872a | 1594 | bool wait = migrate_compress_wait_thread(); |
56e93d26 JQ |
1595 | |
1596 | thread_count = migrate_compress_threads(); | |
0d9f9a5c | 1597 | qemu_mutex_lock(&comp_done_lock); |
1d58872a XG |
1598 | retry: |
1599 | for (idx = 0; idx < thread_count; idx++) { | |
1600 | if (comp_param[idx].done) { | |
b5cf1cd3 LS |
1601 | CompressParam *param = &comp_param[idx]; |
1602 | qemu_mutex_lock(¶m->mutex); | |
1603 | param->done = false; | |
680628d2 | 1604 | send_queued_data(param); |
b5cf1cd3 LS |
1605 | compress_reset_result(param); |
1606 | set_compress_params(param, block, offset); | |
1607 | ||
b5cf1cd3 LS |
1608 | qemu_cond_signal(¶m->cond); |
1609 | qemu_mutex_unlock(¶m->mutex); | |
1d58872a | 1610 | pages = 1; |
56e93d26 | 1611 | break; |
56e93d26 JQ |
1612 | } |
1613 | } | |
1d58872a XG |
1614 | |
1615 | /* | |
1616 | * wait for the free thread if the user specifies 'compress-wait-thread', | |
1617 | * otherwise we will post the page out in the main thread as normal page. | |
1618 | */ | |
1619 | if (pages < 0 && wait) { | |
1620 | qemu_cond_wait(&comp_done_cond, &comp_done_lock); | |
1621 | goto retry; | |
1622 | } | |
0d9f9a5c | 1623 | qemu_mutex_unlock(&comp_done_lock); |
56e93d26 JQ |
1624 | |
1625 | return pages; | |
1626 | } | |
1627 | ||
31e2ac74 JQ |
1628 | #define PAGE_ALL_CLEAN 0 |
1629 | #define PAGE_TRY_AGAIN 1 | |
1630 | #define PAGE_DIRTY_FOUND 2 | |
3d0684b2 JQ |
1631 | /** |
1632 | * find_dirty_block: find the next dirty page and update any state | |
1633 | * associated with the search process. | |
b9e60928 | 1634 | * |
31e2ac74 | 1635 | * Returns: |
294e5a40 | 1636 | * <0: An error happened |
31e2ac74 JQ |
1637 | * PAGE_ALL_CLEAN: no dirty page found, give up |
1638 | * PAGE_TRY_AGAIN: no dirty page found, retry for next block | |
1639 | * PAGE_DIRTY_FOUND: dirty page found | |
b9e60928 | 1640 | * |
6f37bb8b | 1641 | * @rs: current RAM state |
3d0684b2 JQ |
1642 | * @pss: data about the state of the current dirty page scan |
1643 | * @again: set to false if the search has scanned the whole of RAM | |
b9e60928 | 1644 | */ |
31e2ac74 | 1645 | static int find_dirty_block(RAMState *rs, PageSearchStatus *pss) |
b9e60928 | 1646 | { |
d9e474ea PX |
1647 | /* Update pss->page for the next dirty bit in ramblock */ |
1648 | pss_find_next_dirty(pss); | |
1649 | ||
6f37bb8b | 1650 | if (pss->complete_round && pss->block == rs->last_seen_block && |
a935e30f | 1651 | pss->page >= rs->last_page) { |
b9e60928 DDAG |
1652 | /* |
1653 | * We've been once around the RAM and haven't found anything. | |
1654 | * Give up. | |
1655 | */ | |
31e2ac74 | 1656 | return PAGE_ALL_CLEAN; |
b9e60928 | 1657 | } |
542147f4 DH |
1658 | if (!offset_in_ramblock(pss->block, |
1659 | ((ram_addr_t)pss->page) << TARGET_PAGE_BITS)) { | |
b9e60928 | 1660 | /* Didn't find anything in this RAM Block */ |
a935e30f | 1661 | pss->page = 0; |
b9e60928 DDAG |
1662 | pss->block = QLIST_NEXT_RCU(pss->block, next); |
1663 | if (!pss->block) { | |
294e5a40 JQ |
1664 | if (!migrate_multifd_flush_after_each_section()) { |
1665 | QEMUFile *f = rs->pss[RAM_CHANNEL_PRECOPY].pss_channel; | |
1666 | int ret = multifd_send_sync_main(f); | |
1667 | if (ret < 0) { | |
1668 | return ret; | |
1669 | } | |
1670 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); | |
1671 | qemu_fflush(f); | |
1672 | } | |
48df9d80 XG |
1673 | /* |
1674 | * If memory migration starts over, we will meet a dirtied page | |
1675 | * which may still exists in compression threads's ring, so we | |
1676 | * should flush the compressed data to make sure the new page | |
1677 | * is not overwritten by the old one in the destination. | |
1678 | * | |
1679 | * Also If xbzrle is on, stop using the data compression at this | |
1680 | * point. In theory, xbzrle can do better than compression. | |
1681 | */ | |
ef4f5f5d | 1682 | ram_flush_compressed_data(rs); |
48df9d80 | 1683 | |
b9e60928 DDAG |
1684 | /* Hit the end of the list */ |
1685 | pss->block = QLIST_FIRST_RCU(&ram_list.blocks); | |
1686 | /* Flag that we've looped */ | |
1687 | pss->complete_round = true; | |
1a373522 | 1688 | /* After the first round, enable XBZRLE. */ |
87dca0c9 | 1689 | if (migrate_xbzrle()) { |
f3095cc8 | 1690 | rs->xbzrle_started = true; |
1a373522 | 1691 | } |
b9e60928 DDAG |
1692 | } |
1693 | /* Didn't find anything this time, but try again on the new block */ | |
31e2ac74 | 1694 | return PAGE_TRY_AGAIN; |
b9e60928 | 1695 | } else { |
31e2ac74 JQ |
1696 | /* We've found something */ |
1697 | return PAGE_DIRTY_FOUND; | |
b9e60928 DDAG |
1698 | } |
1699 | } | |
1700 | ||
3d0684b2 JQ |
1701 | /** |
1702 | * unqueue_page: gets a page of the queue | |
1703 | * | |
a82d593b | 1704 | * Helper for 'get_queued_page' - gets a page off the queue |
a82d593b | 1705 | * |
3d0684b2 JQ |
1706 | * Returns the block of the page (or NULL if none available) |
1707 | * | |
ec481c6c | 1708 | * @rs: current RAM state |
3d0684b2 | 1709 | * @offset: used to return the offset within the RAMBlock |
a82d593b | 1710 | */ |
f20e2865 | 1711 | static RAMBlock *unqueue_page(RAMState *rs, ram_addr_t *offset) |
a82d593b | 1712 | { |
a1fe28df | 1713 | struct RAMSrcPageRequest *entry; |
a82d593b DDAG |
1714 | RAMBlock *block = NULL; |
1715 | ||
a1fe28df | 1716 | if (!postcopy_has_request(rs)) { |
ae526e32 XG |
1717 | return NULL; |
1718 | } | |
1719 | ||
6e8a355d | 1720 | QEMU_LOCK_GUARD(&rs->src_page_req_mutex); |
a1fe28df PX |
1721 | |
1722 | /* | |
1723 | * This should _never_ change even after we take the lock, because no one | |
1724 | * should be taking anything off the request list other than us. | |
1725 | */ | |
1726 | assert(postcopy_has_request(rs)); | |
1727 | ||
1728 | entry = QSIMPLEQ_FIRST(&rs->src_page_requests); | |
1729 | block = entry->rb; | |
1730 | *offset = entry->offset; | |
1731 | ||
777f53c7 TH |
1732 | if (entry->len > TARGET_PAGE_SIZE) { |
1733 | entry->len -= TARGET_PAGE_SIZE; | |
1734 | entry->offset += TARGET_PAGE_SIZE; | |
a1fe28df PX |
1735 | } else { |
1736 | memory_region_unref(block->mr); | |
1737 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); | |
1738 | g_free(entry); | |
1739 | migration_consume_urgent_request(); | |
a82d593b | 1740 | } |
a82d593b DDAG |
1741 | |
1742 | return block; | |
1743 | } | |
1744 | ||
278e2f55 AG |
1745 | #if defined(__linux__) |
1746 | /** | |
1747 | * poll_fault_page: try to get next UFFD write fault page and, if pending fault | |
1748 | * is found, return RAM block pointer and page offset | |
1749 | * | |
1750 | * Returns pointer to the RAMBlock containing faulting page, | |
1751 | * NULL if no write faults are pending | |
1752 | * | |
1753 | * @rs: current RAM state | |
1754 | * @offset: page offset from the beginning of the block | |
1755 | */ | |
1756 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
1757 | { | |
1758 | struct uffd_msg uffd_msg; | |
1759 | void *page_address; | |
82ea3e3b | 1760 | RAMBlock *block; |
278e2f55 AG |
1761 | int res; |
1762 | ||
1763 | if (!migrate_background_snapshot()) { | |
1764 | return NULL; | |
1765 | } | |
1766 | ||
1767 | res = uffd_read_events(rs->uffdio_fd, &uffd_msg, 1); | |
1768 | if (res <= 0) { | |
1769 | return NULL; | |
1770 | } | |
1771 | ||
1772 | page_address = (void *)(uintptr_t) uffd_msg.arg.pagefault.address; | |
82ea3e3b AG |
1773 | block = qemu_ram_block_from_host(page_address, false, offset); |
1774 | assert(block && (block->flags & RAM_UF_WRITEPROTECT) != 0); | |
1775 | return block; | |
278e2f55 AG |
1776 | } |
1777 | ||
1778 | /** | |
1779 | * ram_save_release_protection: release UFFD write protection after | |
1780 | * a range of pages has been saved | |
1781 | * | |
1782 | * @rs: current RAM state | |
1783 | * @pss: page-search-status structure | |
1784 | * @start_page: index of the first page in the range relative to pss->block | |
1785 | * | |
1786 | * Returns 0 on success, negative value in case of an error | |
1787 | */ | |
1788 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
1789 | unsigned long start_page) | |
1790 | { | |
1791 | int res = 0; | |
1792 | ||
1793 | /* Check if page is from UFFD-managed region. */ | |
1794 | if (pss->block->flags & RAM_UF_WRITEPROTECT) { | |
1795 | void *page_address = pss->block->host + (start_page << TARGET_PAGE_BITS); | |
258f5c98 | 1796 | uint64_t run_length = (pss->page - start_page) << TARGET_PAGE_BITS; |
278e2f55 AG |
1797 | |
1798 | /* Flush async buffers before un-protect. */ | |
61717ea9 | 1799 | qemu_fflush(pss->pss_channel); |
278e2f55 AG |
1800 | /* Un-protect memory range. */ |
1801 | res = uffd_change_protection(rs->uffdio_fd, page_address, run_length, | |
1802 | false, false); | |
1803 | } | |
1804 | ||
1805 | return res; | |
1806 | } | |
1807 | ||
1808 | /* ram_write_tracking_available: check if kernel supports required UFFD features | |
1809 | * | |
1810 | * Returns true if supports, false otherwise | |
1811 | */ | |
1812 | bool ram_write_tracking_available(void) | |
1813 | { | |
1814 | uint64_t uffd_features; | |
1815 | int res; | |
1816 | ||
1817 | res = uffd_query_features(&uffd_features); | |
1818 | return (res == 0 && | |
1819 | (uffd_features & UFFD_FEATURE_PAGEFAULT_FLAG_WP) != 0); | |
1820 | } | |
1821 | ||
1822 | /* ram_write_tracking_compatible: check if guest configuration is | |
1823 | * compatible with 'write-tracking' | |
1824 | * | |
1825 | * Returns true if compatible, false otherwise | |
1826 | */ | |
1827 | bool ram_write_tracking_compatible(void) | |
1828 | { | |
1829 | const uint64_t uffd_ioctls_mask = BIT(_UFFDIO_WRITEPROTECT); | |
1830 | int uffd_fd; | |
82ea3e3b | 1831 | RAMBlock *block; |
278e2f55 AG |
1832 | bool ret = false; |
1833 | ||
1834 | /* Open UFFD file descriptor */ | |
1835 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, false); | |
1836 | if (uffd_fd < 0) { | |
1837 | return false; | |
1838 | } | |
1839 | ||
1840 | RCU_READ_LOCK_GUARD(); | |
1841 | ||
82ea3e3b | 1842 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 AG |
1843 | uint64_t uffd_ioctls; |
1844 | ||
1845 | /* Nothing to do with read-only and MMIO-writable regions */ | |
82ea3e3b | 1846 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
1847 | continue; |
1848 | } | |
1849 | /* Try to register block memory via UFFD-IO to track writes */ | |
82ea3e3b | 1850 | if (uffd_register_memory(uffd_fd, block->host, block->max_length, |
278e2f55 AG |
1851 | UFFDIO_REGISTER_MODE_WP, &uffd_ioctls)) { |
1852 | goto out; | |
1853 | } | |
1854 | if ((uffd_ioctls & uffd_ioctls_mask) != uffd_ioctls_mask) { | |
1855 | goto out; | |
1856 | } | |
1857 | } | |
1858 | ret = true; | |
1859 | ||
1860 | out: | |
1861 | uffd_close_fd(uffd_fd); | |
1862 | return ret; | |
1863 | } | |
1864 | ||
f7b9dcfb DH |
1865 | static inline void populate_read_range(RAMBlock *block, ram_addr_t offset, |
1866 | ram_addr_t size) | |
1867 | { | |
5f19a449 DH |
1868 | const ram_addr_t end = offset + size; |
1869 | ||
f7b9dcfb DH |
1870 | /* |
1871 | * We read one byte of each page; this will preallocate page tables if | |
1872 | * required and populate the shared zeropage on MAP_PRIVATE anonymous memory | |
1873 | * where no page was populated yet. This might require adaption when | |
1874 | * supporting other mappings, like shmem. | |
1875 | */ | |
5f19a449 | 1876 | for (; offset < end; offset += block->page_size) { |
f7b9dcfb DH |
1877 | char tmp = *((char *)block->host + offset); |
1878 | ||
1879 | /* Don't optimize the read out */ | |
1880 | asm volatile("" : "+r" (tmp)); | |
1881 | } | |
1882 | } | |
1883 | ||
6fee3a1f DH |
1884 | static inline int populate_read_section(MemoryRegionSection *section, |
1885 | void *opaque) | |
1886 | { | |
1887 | const hwaddr size = int128_get64(section->size); | |
1888 | hwaddr offset = section->offset_within_region; | |
1889 | RAMBlock *block = section->mr->ram_block; | |
1890 | ||
1891 | populate_read_range(block, offset, size); | |
1892 | return 0; | |
1893 | } | |
1894 | ||
eeccb99c | 1895 | /* |
f7b9dcfb DH |
1896 | * ram_block_populate_read: preallocate page tables and populate pages in the |
1897 | * RAM block by reading a byte of each page. | |
eeccb99c AG |
1898 | * |
1899 | * Since it's solely used for userfault_fd WP feature, here we just | |
1900 | * hardcode page size to qemu_real_host_page_size. | |
1901 | * | |
82ea3e3b | 1902 | * @block: RAM block to populate |
eeccb99c | 1903 | */ |
6fee3a1f | 1904 | static void ram_block_populate_read(RAMBlock *rb) |
eeccb99c | 1905 | { |
6fee3a1f DH |
1906 | /* |
1907 | * Skip populating all pages that fall into a discarded range as managed by | |
1908 | * a RamDiscardManager responsible for the mapped memory region of the | |
1909 | * RAMBlock. Such discarded ("logically unplugged") parts of a RAMBlock | |
1910 | * must not get populated automatically. We don't have to track | |
1911 | * modifications via userfaultfd WP reliably, because these pages will | |
1912 | * not be part of the migration stream either way -- see | |
1913 | * ramblock_dirty_bitmap_exclude_discarded_pages(). | |
1914 | * | |
1915 | * Note: The result is only stable while migrating (precopy/postcopy). | |
1916 | */ | |
1917 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1918 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1919 | MemoryRegionSection section = { | |
1920 | .mr = rb->mr, | |
1921 | .offset_within_region = 0, | |
1922 | .size = rb->mr->size, | |
1923 | }; | |
1924 | ||
1925 | ram_discard_manager_replay_populated(rdm, §ion, | |
1926 | populate_read_section, NULL); | |
1927 | } else { | |
1928 | populate_read_range(rb, 0, rb->used_length); | |
1929 | } | |
eeccb99c AG |
1930 | } |
1931 | ||
1932 | /* | |
1933 | * ram_write_tracking_prepare: prepare for UFFD-WP memory tracking | |
1934 | */ | |
1935 | void ram_write_tracking_prepare(void) | |
1936 | { | |
82ea3e3b | 1937 | RAMBlock *block; |
eeccb99c AG |
1938 | |
1939 | RCU_READ_LOCK_GUARD(); | |
1940 | ||
82ea3e3b | 1941 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
eeccb99c | 1942 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 1943 | if (block->mr->readonly || block->mr->rom_device) { |
eeccb99c AG |
1944 | continue; |
1945 | } | |
1946 | ||
1947 | /* | |
1948 | * Populate pages of the RAM block before enabling userfault_fd | |
1949 | * write protection. | |
1950 | * | |
1951 | * This stage is required since ioctl(UFFDIO_WRITEPROTECT) with | |
1952 | * UFFDIO_WRITEPROTECT_MODE_WP mode setting would silently skip | |
1953 | * pages with pte_none() entries in page table. | |
1954 | */ | |
f7b9dcfb | 1955 | ram_block_populate_read(block); |
eeccb99c AG |
1956 | } |
1957 | } | |
1958 | ||
e41c5770 DH |
1959 | static inline int uffd_protect_section(MemoryRegionSection *section, |
1960 | void *opaque) | |
1961 | { | |
1962 | const hwaddr size = int128_get64(section->size); | |
1963 | const hwaddr offset = section->offset_within_region; | |
1964 | RAMBlock *rb = section->mr->ram_block; | |
1965 | int uffd_fd = (uintptr_t)opaque; | |
1966 | ||
1967 | return uffd_change_protection(uffd_fd, rb->host + offset, size, true, | |
1968 | false); | |
1969 | } | |
1970 | ||
1971 | static int ram_block_uffd_protect(RAMBlock *rb, int uffd_fd) | |
1972 | { | |
1973 | assert(rb->flags & RAM_UF_WRITEPROTECT); | |
1974 | ||
1975 | /* See ram_block_populate_read() */ | |
1976 | if (rb->mr && memory_region_has_ram_discard_manager(rb->mr)) { | |
1977 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(rb->mr); | |
1978 | MemoryRegionSection section = { | |
1979 | .mr = rb->mr, | |
1980 | .offset_within_region = 0, | |
1981 | .size = rb->mr->size, | |
1982 | }; | |
1983 | ||
1984 | return ram_discard_manager_replay_populated(rdm, §ion, | |
1985 | uffd_protect_section, | |
1986 | (void *)(uintptr_t)uffd_fd); | |
1987 | } | |
1988 | return uffd_change_protection(uffd_fd, rb->host, | |
1989 | rb->used_length, true, false); | |
1990 | } | |
1991 | ||
278e2f55 AG |
1992 | /* |
1993 | * ram_write_tracking_start: start UFFD-WP memory tracking | |
1994 | * | |
1995 | * Returns 0 for success or negative value in case of error | |
1996 | */ | |
1997 | int ram_write_tracking_start(void) | |
1998 | { | |
1999 | int uffd_fd; | |
2000 | RAMState *rs = ram_state; | |
82ea3e3b | 2001 | RAMBlock *block; |
278e2f55 AG |
2002 | |
2003 | /* Open UFFD file descriptor */ | |
2004 | uffd_fd = uffd_create_fd(UFFD_FEATURE_PAGEFAULT_FLAG_WP, true); | |
2005 | if (uffd_fd < 0) { | |
2006 | return uffd_fd; | |
2007 | } | |
2008 | rs->uffdio_fd = uffd_fd; | |
2009 | ||
2010 | RCU_READ_LOCK_GUARD(); | |
2011 | ||
82ea3e3b | 2012 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
278e2f55 | 2013 | /* Nothing to do with read-only and MMIO-writable regions */ |
82ea3e3b | 2014 | if (block->mr->readonly || block->mr->rom_device) { |
278e2f55 AG |
2015 | continue; |
2016 | } | |
2017 | ||
2018 | /* Register block memory with UFFD to track writes */ | |
82ea3e3b AG |
2019 | if (uffd_register_memory(rs->uffdio_fd, block->host, |
2020 | block->max_length, UFFDIO_REGISTER_MODE_WP, NULL)) { | |
278e2f55 AG |
2021 | goto fail; |
2022 | } | |
72ef3a37 DH |
2023 | block->flags |= RAM_UF_WRITEPROTECT; |
2024 | memory_region_ref(block->mr); | |
2025 | ||
278e2f55 | 2026 | /* Apply UFFD write protection to the block memory range */ |
e41c5770 | 2027 | if (ram_block_uffd_protect(block, uffd_fd)) { |
278e2f55 AG |
2028 | goto fail; |
2029 | } | |
278e2f55 | 2030 | |
82ea3e3b AG |
2031 | trace_ram_write_tracking_ramblock_start(block->idstr, block->page_size, |
2032 | block->host, block->max_length); | |
278e2f55 AG |
2033 | } |
2034 | ||
2035 | return 0; | |
2036 | ||
2037 | fail: | |
2038 | error_report("ram_write_tracking_start() failed: restoring initial memory state"); | |
2039 | ||
82ea3e3b AG |
2040 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
2041 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
2042 | continue; |
2043 | } | |
82ea3e3b | 2044 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 2045 | /* Cleanup flags and remove reference */ |
82ea3e3b AG |
2046 | block->flags &= ~RAM_UF_WRITEPROTECT; |
2047 | memory_region_unref(block->mr); | |
278e2f55 AG |
2048 | } |
2049 | ||
2050 | uffd_close_fd(uffd_fd); | |
2051 | rs->uffdio_fd = -1; | |
2052 | return -1; | |
2053 | } | |
2054 | ||
2055 | /** | |
2056 | * ram_write_tracking_stop: stop UFFD-WP memory tracking and remove protection | |
2057 | */ | |
2058 | void ram_write_tracking_stop(void) | |
2059 | { | |
2060 | RAMState *rs = ram_state; | |
82ea3e3b | 2061 | RAMBlock *block; |
278e2f55 AG |
2062 | |
2063 | RCU_READ_LOCK_GUARD(); | |
2064 | ||
82ea3e3b AG |
2065 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
2066 | if ((block->flags & RAM_UF_WRITEPROTECT) == 0) { | |
278e2f55 AG |
2067 | continue; |
2068 | } | |
82ea3e3b | 2069 | uffd_unregister_memory(rs->uffdio_fd, block->host, block->max_length); |
278e2f55 | 2070 | |
82ea3e3b AG |
2071 | trace_ram_write_tracking_ramblock_stop(block->idstr, block->page_size, |
2072 | block->host, block->max_length); | |
278e2f55 AG |
2073 | |
2074 | /* Cleanup flags and remove reference */ | |
82ea3e3b AG |
2075 | block->flags &= ~RAM_UF_WRITEPROTECT; |
2076 | memory_region_unref(block->mr); | |
278e2f55 AG |
2077 | } |
2078 | ||
2079 | /* Finally close UFFD file descriptor */ | |
2080 | uffd_close_fd(rs->uffdio_fd); | |
2081 | rs->uffdio_fd = -1; | |
2082 | } | |
2083 | ||
2084 | #else | |
2085 | /* No target OS support, stubs just fail or ignore */ | |
2086 | ||
2087 | static RAMBlock *poll_fault_page(RAMState *rs, ram_addr_t *offset) | |
2088 | { | |
2089 | (void) rs; | |
2090 | (void) offset; | |
2091 | ||
2092 | return NULL; | |
2093 | } | |
2094 | ||
2095 | static int ram_save_release_protection(RAMState *rs, PageSearchStatus *pss, | |
2096 | unsigned long start_page) | |
2097 | { | |
2098 | (void) rs; | |
2099 | (void) pss; | |
2100 | (void) start_page; | |
2101 | ||
2102 | return 0; | |
2103 | } | |
2104 | ||
2105 | bool ram_write_tracking_available(void) | |
2106 | { | |
2107 | return false; | |
2108 | } | |
2109 | ||
2110 | bool ram_write_tracking_compatible(void) | |
2111 | { | |
2112 | assert(0); | |
2113 | return false; | |
2114 | } | |
2115 | ||
2116 | int ram_write_tracking_start(void) | |
2117 | { | |
2118 | assert(0); | |
2119 | return -1; | |
2120 | } | |
2121 | ||
2122 | void ram_write_tracking_stop(void) | |
2123 | { | |
2124 | assert(0); | |
2125 | } | |
2126 | #endif /* defined(__linux__) */ | |
2127 | ||
3d0684b2 | 2128 | /** |
ff1543af | 2129 | * get_queued_page: unqueue a page from the postcopy requests |
3d0684b2 JQ |
2130 | * |
2131 | * Skips pages that are already sent (!dirty) | |
a82d593b | 2132 | * |
a5f7b1a6 | 2133 | * Returns true if a queued page is found |
a82d593b | 2134 | * |
6f37bb8b | 2135 | * @rs: current RAM state |
3d0684b2 | 2136 | * @pss: data about the state of the current dirty page scan |
a82d593b | 2137 | */ |
f20e2865 | 2138 | static bool get_queued_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b DDAG |
2139 | { |
2140 | RAMBlock *block; | |
2141 | ram_addr_t offset; | |
777f53c7 TH |
2142 | bool dirty; |
2143 | ||
2144 | do { | |
2145 | block = unqueue_page(rs, &offset); | |
2146 | /* | |
2147 | * We're sending this page, and since it's postcopy nothing else | |
2148 | * will dirty it, and we must make sure it doesn't get sent again | |
2149 | * even if this queue request was received after the background | |
2150 | * search already sent it. | |
2151 | */ | |
2152 | if (block) { | |
2153 | unsigned long page; | |
2154 | ||
2155 | page = offset >> TARGET_PAGE_BITS; | |
2156 | dirty = test_bit(page, block->bmap); | |
2157 | if (!dirty) { | |
2158 | trace_get_queued_page_not_dirty(block->idstr, (uint64_t)offset, | |
2159 | page); | |
2160 | } else { | |
2161 | trace_get_queued_page(block->idstr, (uint64_t)offset, page); | |
2162 | } | |
2163 | } | |
a82d593b | 2164 | |
777f53c7 | 2165 | } while (block && !dirty); |
a82d593b | 2166 | |
b062106d | 2167 | if (!block) { |
278e2f55 AG |
2168 | /* |
2169 | * Poll write faults too if background snapshot is enabled; that's | |
2170 | * when we have vcpus got blocked by the write protected pages. | |
2171 | */ | |
2172 | block = poll_fault_page(rs, &offset); | |
2173 | } | |
2174 | ||
a82d593b | 2175 | if (block) { |
a82d593b DDAG |
2176 | /* |
2177 | * We want the background search to continue from the queued page | |
2178 | * since the guest is likely to want other pages near to the page | |
2179 | * it just requested. | |
2180 | */ | |
2181 | pss->block = block; | |
a935e30f | 2182 | pss->page = offset >> TARGET_PAGE_BITS; |
422314e7 WY |
2183 | |
2184 | /* | |
2185 | * This unqueued page would break the "one round" check, even is | |
2186 | * really rare. | |
2187 | */ | |
2188 | pss->complete_round = false; | |
a82d593b DDAG |
2189 | } |
2190 | ||
2191 | return !!block; | |
2192 | } | |
2193 | ||
6c595cde | 2194 | /** |
5e58f968 JQ |
2195 | * migration_page_queue_free: drop any remaining pages in the ram |
2196 | * request queue | |
6c595cde | 2197 | * |
3d0684b2 JQ |
2198 | * It should be empty at the end anyway, but in error cases there may |
2199 | * be some left. in case that there is any page left, we drop it. | |
2200 | * | |
6c595cde | 2201 | */ |
83c13382 | 2202 | static void migration_page_queue_free(RAMState *rs) |
6c595cde | 2203 | { |
ec481c6c | 2204 | struct RAMSrcPageRequest *mspr, *next_mspr; |
6c595cde DDAG |
2205 | /* This queue generally should be empty - but in the case of a failed |
2206 | * migration might have some droppings in. | |
2207 | */ | |
89ac5a1d | 2208 | RCU_READ_LOCK_GUARD(); |
ec481c6c | 2209 | QSIMPLEQ_FOREACH_SAFE(mspr, &rs->src_page_requests, next_req, next_mspr) { |
6c595cde | 2210 | memory_region_unref(mspr->rb->mr); |
ec481c6c | 2211 | QSIMPLEQ_REMOVE_HEAD(&rs->src_page_requests, next_req); |
6c595cde DDAG |
2212 | g_free(mspr); |
2213 | } | |
6c595cde DDAG |
2214 | } |
2215 | ||
2216 | /** | |
3d0684b2 JQ |
2217 | * ram_save_queue_pages: queue the page for transmission |
2218 | * | |
2219 | * A request from postcopy destination for example. | |
2220 | * | |
2221 | * Returns zero on success or negative on error | |
2222 | * | |
3d0684b2 JQ |
2223 | * @rbname: Name of the RAMBLock of the request. NULL means the |
2224 | * same that last one. | |
2225 | * @start: starting address from the start of the RAMBlock | |
2226 | * @len: length (in bytes) to send | |
6c595cde | 2227 | */ |
96506894 | 2228 | int ram_save_queue_pages(const char *rbname, ram_addr_t start, ram_addr_t len) |
6c595cde DDAG |
2229 | { |
2230 | RAMBlock *ramblock; | |
53518d94 | 2231 | RAMState *rs = ram_state; |
6c595cde | 2232 | |
aff3f660 | 2233 | stat64_add(&mig_stats.postcopy_requests, 1); |
89ac5a1d DDAG |
2234 | RCU_READ_LOCK_GUARD(); |
2235 | ||
6c595cde DDAG |
2236 | if (!rbname) { |
2237 | /* Reuse last RAMBlock */ | |
68a098f3 | 2238 | ramblock = rs->last_req_rb; |
6c595cde DDAG |
2239 | |
2240 | if (!ramblock) { | |
2241 | /* | |
2242 | * Shouldn't happen, we can't reuse the last RAMBlock if | |
2243 | * it's the 1st request. | |
2244 | */ | |
2245 | error_report("ram_save_queue_pages no previous block"); | |
03acb4e9 | 2246 | return -1; |
6c595cde DDAG |
2247 | } |
2248 | } else { | |
2249 | ramblock = qemu_ram_block_by_name(rbname); | |
2250 | ||
2251 | if (!ramblock) { | |
2252 | /* We shouldn't be asked for a non-existent RAMBlock */ | |
2253 | error_report("ram_save_queue_pages no block '%s'", rbname); | |
03acb4e9 | 2254 | return -1; |
6c595cde | 2255 | } |
68a098f3 | 2256 | rs->last_req_rb = ramblock; |
6c595cde DDAG |
2257 | } |
2258 | trace_ram_save_queue_pages(ramblock->idstr, start, len); | |
542147f4 | 2259 | if (!offset_in_ramblock(ramblock, start + len - 1)) { |
9458ad6b JQ |
2260 | error_report("%s request overrun start=" RAM_ADDR_FMT " len=" |
2261 | RAM_ADDR_FMT " blocklen=" RAM_ADDR_FMT, | |
6c595cde | 2262 | __func__, start, len, ramblock->used_length); |
03acb4e9 | 2263 | return -1; |
6c595cde DDAG |
2264 | } |
2265 | ||
93589827 PX |
2266 | /* |
2267 | * When with postcopy preempt, we send back the page directly in the | |
2268 | * rp-return thread. | |
2269 | */ | |
2270 | if (postcopy_preempt_active()) { | |
2271 | ram_addr_t page_start = start >> TARGET_PAGE_BITS; | |
2272 | size_t page_size = qemu_ram_pagesize(ramblock); | |
2273 | PageSearchStatus *pss = &ram_state->pss[RAM_CHANNEL_POSTCOPY]; | |
2274 | int ret = 0; | |
2275 | ||
2276 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2277 | ||
2278 | pss_init(pss, ramblock, page_start); | |
2279 | /* | |
2280 | * Always use the preempt channel, and make sure it's there. It's | |
2281 | * safe to access without lock, because when rp-thread is running | |
2282 | * we should be the only one who operates on the qemufile | |
2283 | */ | |
2284 | pss->pss_channel = migrate_get_current()->postcopy_qemufile_src; | |
93589827 PX |
2285 | assert(pss->pss_channel); |
2286 | ||
2287 | /* | |
2288 | * It must be either one or multiple of host page size. Just | |
2289 | * assert; if something wrong we're mostly split brain anyway. | |
2290 | */ | |
2291 | assert(len % page_size == 0); | |
2292 | while (len) { | |
2293 | if (ram_save_host_page_urgent(pss)) { | |
2294 | error_report("%s: ram_save_host_page_urgent() failed: " | |
2295 | "ramblock=%s, start_addr=0x"RAM_ADDR_FMT, | |
2296 | __func__, ramblock->idstr, start); | |
2297 | ret = -1; | |
2298 | break; | |
2299 | } | |
2300 | /* | |
2301 | * NOTE: after ram_save_host_page_urgent() succeeded, pss->page | |
2302 | * will automatically be moved and point to the next host page | |
2303 | * we're going to send, so no need to update here. | |
2304 | * | |
2305 | * Normally QEMU never sends >1 host page in requests, so | |
2306 | * logically we don't even need that as the loop should only | |
2307 | * run once, but just to be consistent. | |
2308 | */ | |
2309 | len -= page_size; | |
2310 | }; | |
2311 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
2312 | ||
2313 | return ret; | |
2314 | } | |
2315 | ||
ec481c6c | 2316 | struct RAMSrcPageRequest *new_entry = |
b21e2380 | 2317 | g_new0(struct RAMSrcPageRequest, 1); |
6c595cde DDAG |
2318 | new_entry->rb = ramblock; |
2319 | new_entry->offset = start; | |
2320 | new_entry->len = len; | |
2321 | ||
2322 | memory_region_ref(ramblock->mr); | |
ec481c6c JQ |
2323 | qemu_mutex_lock(&rs->src_page_req_mutex); |
2324 | QSIMPLEQ_INSERT_TAIL(&rs->src_page_requests, new_entry, next_req); | |
e03a34f8 | 2325 | migration_make_urgent_request(); |
ec481c6c | 2326 | qemu_mutex_unlock(&rs->src_page_req_mutex); |
6c595cde DDAG |
2327 | |
2328 | return 0; | |
6c595cde DDAG |
2329 | } |
2330 | ||
d7400a34 XG |
2331 | static bool save_page_use_compression(RAMState *rs) |
2332 | { | |
a7a94d14 | 2333 | if (!migrate_compress()) { |
d7400a34 XG |
2334 | return false; |
2335 | } | |
2336 | ||
2337 | /* | |
1a373522 DH |
2338 | * If xbzrle is enabled (e.g., after first round of migration), stop |
2339 | * using the data compression. In theory, xbzrle can do better than | |
2340 | * compression. | |
d7400a34 | 2341 | */ |
f3095cc8 | 2342 | if (rs->xbzrle_started) { |
1a373522 | 2343 | return false; |
d7400a34 XG |
2344 | } |
2345 | ||
1a373522 | 2346 | return true; |
d7400a34 XG |
2347 | } |
2348 | ||
5e5fdcff XG |
2349 | /* |
2350 | * try to compress the page before posting it out, return true if the page | |
2351 | * has been properly handled by compression, otherwise needs other | |
2352 | * paths to handle it | |
2353 | */ | |
ec6f3ab9 PX |
2354 | static bool save_compress_page(RAMState *rs, PageSearchStatus *pss, |
2355 | RAMBlock *block, ram_addr_t offset) | |
5e5fdcff XG |
2356 | { |
2357 | if (!save_page_use_compression(rs)) { | |
2358 | return false; | |
2359 | } | |
2360 | ||
2361 | /* | |
2362 | * When starting the process of a new block, the first page of | |
2363 | * the block should be sent out before other pages in the same | |
2364 | * block, and all the pages in last block should have been sent | |
2365 | * out, keeping this order is important, because the 'cont' flag | |
2366 | * is used to avoid resending the block name. | |
2367 | * | |
2368 | * We post the fist page as normal page as compression will take | |
2369 | * much CPU resource. | |
2370 | */ | |
ec6f3ab9 | 2371 | if (block != pss->last_sent_block) { |
ef4f5f5d | 2372 | ram_flush_compressed_data(rs); |
5e5fdcff XG |
2373 | return false; |
2374 | } | |
2375 | ||
ef4f5f5d | 2376 | if (compress_page_with_multi_thread(block, offset, send_queued_data) > 0) { |
5e5fdcff XG |
2377 | return true; |
2378 | } | |
2379 | ||
76e03000 | 2380 | compression_counters.busy++; |
5e5fdcff XG |
2381 | return false; |
2382 | } | |
2383 | ||
a82d593b | 2384 | /** |
4010ba38 | 2385 | * ram_save_target_page_legacy: save one target page |
a82d593b | 2386 | * |
3d0684b2 | 2387 | * Returns the number of pages written |
a82d593b | 2388 | * |
6f37bb8b | 2389 | * @rs: current RAM state |
3d0684b2 | 2390 | * @pss: data about the page we want to send |
a82d593b | 2391 | */ |
4010ba38 | 2392 | static int ram_save_target_page_legacy(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2393 | { |
a8ec91f9 | 2394 | RAMBlock *block = pss->block; |
8bba004c | 2395 | ram_addr_t offset = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; |
a8ec91f9 XG |
2396 | int res; |
2397 | ||
61717ea9 | 2398 | if (control_save_page(pss, block, offset, &res)) { |
a8ec91f9 XG |
2399 | return res; |
2400 | } | |
2401 | ||
ec6f3ab9 | 2402 | if (save_compress_page(rs, pss, block, offset)) { |
5e5fdcff | 2403 | return 1; |
d7400a34 XG |
2404 | } |
2405 | ||
37502df3 | 2406 | res = save_zero_page(pss, pss->pss_channel, block, offset); |
d7400a34 XG |
2407 | if (res > 0) { |
2408 | /* Must let xbzrle know, otherwise a previous (now 0'd) cached | |
2409 | * page would be stale | |
2410 | */ | |
f3095cc8 | 2411 | if (rs->xbzrle_started) { |
d7400a34 XG |
2412 | XBZRLE_cache_lock(); |
2413 | xbzrle_cache_zero_page(rs, block->offset + offset); | |
2414 | XBZRLE_cache_unlock(); | |
2415 | } | |
d7400a34 XG |
2416 | return res; |
2417 | } | |
2418 | ||
da3f56cb | 2419 | /* |
6f39c90b PX |
2420 | * Do not use multifd in postcopy as one whole host page should be |
2421 | * placed. Meanwhile postcopy requires atomic update of pages, so even | |
2422 | * if host page size == guest page size the dest guest during run may | |
2423 | * still see partially copied pages which is data corruption. | |
da3f56cb | 2424 | */ |
51b07548 | 2425 | if (migrate_multifd() && !migration_in_postcopy()) { |
61717ea9 | 2426 | return ram_save_multifd_page(pss->pss_channel, block, offset); |
a82d593b DDAG |
2427 | } |
2428 | ||
05931ec5 | 2429 | return ram_save_page(rs, pss); |
a82d593b DDAG |
2430 | } |
2431 | ||
d9e474ea PX |
2432 | /* Should be called before sending a host page */ |
2433 | static void pss_host_page_prepare(PageSearchStatus *pss) | |
2434 | { | |
2435 | /* How many guest pages are there in one host page? */ | |
2436 | size_t guest_pfns = qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
2437 | ||
2438 | pss->host_page_sending = true; | |
301d7ffe PX |
2439 | if (guest_pfns <= 1) { |
2440 | /* | |
2441 | * This covers both when guest psize == host psize, or when guest | |
2442 | * has larger psize than the host (guest_pfns==0). | |
2443 | * | |
2444 | * For the latter, we always send one whole guest page per | |
2445 | * iteration of the host page (example: an Alpha VM on x86 host | |
2446 | * will have guest psize 8K while host psize 4K). | |
2447 | */ | |
2448 | pss->host_page_start = pss->page; | |
2449 | pss->host_page_end = pss->page + 1; | |
2450 | } else { | |
2451 | /* | |
2452 | * The host page spans over multiple guest pages, we send them | |
2453 | * within the same host page iteration. | |
2454 | */ | |
2455 | pss->host_page_start = ROUND_DOWN(pss->page, guest_pfns); | |
2456 | pss->host_page_end = ROUND_UP(pss->page + 1, guest_pfns); | |
2457 | } | |
d9e474ea PX |
2458 | } |
2459 | ||
2460 | /* | |
2461 | * Whether the page pointed by PSS is within the host page being sent. | |
2462 | * Must be called after a previous pss_host_page_prepare(). | |
2463 | */ | |
2464 | static bool pss_within_range(PageSearchStatus *pss) | |
2465 | { | |
2466 | ram_addr_t ram_addr; | |
2467 | ||
2468 | assert(pss->host_page_sending); | |
2469 | ||
2470 | /* Over host-page boundary? */ | |
2471 | if (pss->page >= pss->host_page_end) { | |
2472 | return false; | |
2473 | } | |
2474 | ||
2475 | ram_addr = ((ram_addr_t)pss->page) << TARGET_PAGE_BITS; | |
2476 | ||
2477 | return offset_in_ramblock(pss->block, ram_addr); | |
2478 | } | |
2479 | ||
2480 | static void pss_host_page_finish(PageSearchStatus *pss) | |
2481 | { | |
2482 | pss->host_page_sending = false; | |
2483 | /* This is not needed, but just to reset it */ | |
2484 | pss->host_page_start = pss->host_page_end = 0; | |
2485 | } | |
2486 | ||
93589827 PX |
2487 | /* |
2488 | * Send an urgent host page specified by `pss'. Need to be called with | |
2489 | * bitmap_mutex held. | |
2490 | * | |
2491 | * Returns 0 if save host page succeeded, false otherwise. | |
2492 | */ | |
2493 | static int ram_save_host_page_urgent(PageSearchStatus *pss) | |
2494 | { | |
2495 | bool page_dirty, sent = false; | |
2496 | RAMState *rs = ram_state; | |
2497 | int ret = 0; | |
2498 | ||
2499 | trace_postcopy_preempt_send_host_page(pss->block->idstr, pss->page); | |
2500 | pss_host_page_prepare(pss); | |
2501 | ||
2502 | /* | |
2503 | * If precopy is sending the same page, let it be done in precopy, or | |
2504 | * we could send the same page in two channels and none of them will | |
2505 | * receive the whole page. | |
2506 | */ | |
2507 | if (pss_overlap(pss, &ram_state->pss[RAM_CHANNEL_PRECOPY])) { | |
2508 | trace_postcopy_preempt_hit(pss->block->idstr, | |
2509 | pss->page << TARGET_PAGE_BITS); | |
2510 | return 0; | |
2511 | } | |
2512 | ||
2513 | do { | |
2514 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); | |
2515 | ||
2516 | if (page_dirty) { | |
2517 | /* Be strict to return code; it must be 1, or what else? */ | |
4010ba38 | 2518 | if (migration_ops->ram_save_target_page(rs, pss) != 1) { |
93589827 PX |
2519 | error_report_once("%s: ram_save_target_page failed", __func__); |
2520 | ret = -1; | |
2521 | goto out; | |
2522 | } | |
2523 | sent = true; | |
2524 | } | |
2525 | pss_find_next_dirty(pss); | |
2526 | } while (pss_within_range(pss)); | |
2527 | out: | |
2528 | pss_host_page_finish(pss); | |
2529 | /* For urgent requests, flush immediately if sent */ | |
2530 | if (sent) { | |
2531 | qemu_fflush(pss->pss_channel); | |
2532 | } | |
2533 | return ret; | |
2534 | } | |
2535 | ||
a82d593b | 2536 | /** |
3d0684b2 | 2537 | * ram_save_host_page: save a whole host page |
a82d593b | 2538 | * |
3d0684b2 JQ |
2539 | * Starting at *offset send pages up to the end of the current host |
2540 | * page. It's valid for the initial offset to point into the middle of | |
2541 | * a host page in which case the remainder of the hostpage is sent. | |
2542 | * Only dirty target pages are sent. Note that the host page size may | |
2543 | * be a huge page for this block. | |
f3321554 | 2544 | * |
1eb3fc0a DDAG |
2545 | * The saving stops at the boundary of the used_length of the block |
2546 | * if the RAMBlock isn't a multiple of the host page size. | |
a82d593b | 2547 | * |
f3321554 PX |
2548 | * The caller must be with ram_state.bitmap_mutex held to call this |
2549 | * function. Note that this function can temporarily release the lock, but | |
2550 | * when the function is returned it'll make sure the lock is still held. | |
2551 | * | |
3d0684b2 JQ |
2552 | * Returns the number of pages written or negative on error |
2553 | * | |
6f37bb8b | 2554 | * @rs: current RAM state |
3d0684b2 | 2555 | * @pss: data about the page we want to send |
a82d593b | 2556 | */ |
05931ec5 | 2557 | static int ram_save_host_page(RAMState *rs, PageSearchStatus *pss) |
a82d593b | 2558 | { |
f3321554 | 2559 | bool page_dirty, preempt_active = postcopy_preempt_active(); |
a82d593b | 2560 | int tmppages, pages = 0; |
a935e30f JQ |
2561 | size_t pagesize_bits = |
2562 | qemu_ram_pagesize(pss->block) >> TARGET_PAGE_BITS; | |
278e2f55 AG |
2563 | unsigned long start_page = pss->page; |
2564 | int res; | |
4c011c37 | 2565 | |
fbd162e6 | 2566 | if (ramblock_is_ignored(pss->block)) { |
b895de50 CLG |
2567 | error_report("block %s should not be migrated !", pss->block->idstr); |
2568 | return 0; | |
2569 | } | |
2570 | ||
d9e474ea PX |
2571 | /* Update host page boundary information */ |
2572 | pss_host_page_prepare(pss); | |
2573 | ||
a82d593b | 2574 | do { |
f3321554 | 2575 | page_dirty = migration_bitmap_clear_dirty(rs, pss->block, pss->page); |
a82d593b | 2576 | |
f3321554 PX |
2577 | /* Check the pages is dirty and if it is send it */ |
2578 | if (page_dirty) { | |
ba1b7c81 | 2579 | /* |
f3321554 PX |
2580 | * Properly yield the lock only in postcopy preempt mode |
2581 | * because both migration thread and rp-return thread can | |
2582 | * operate on the bitmaps. | |
ba1b7c81 | 2583 | */ |
f3321554 PX |
2584 | if (preempt_active) { |
2585 | qemu_mutex_unlock(&rs->bitmap_mutex); | |
ba1b7c81 | 2586 | } |
4010ba38 | 2587 | tmppages = migration_ops->ram_save_target_page(rs, pss); |
f3321554 PX |
2588 | if (tmppages >= 0) { |
2589 | pages += tmppages; | |
2590 | /* | |
2591 | * Allow rate limiting to happen in the middle of huge pages if | |
2592 | * something is sent in the current iteration. | |
2593 | */ | |
2594 | if (pagesize_bits > 1 && tmppages > 0) { | |
2595 | migration_rate_limit(); | |
2596 | } | |
2597 | } | |
2598 | if (preempt_active) { | |
2599 | qemu_mutex_lock(&rs->bitmap_mutex); | |
2600 | } | |
2601 | } else { | |
2602 | tmppages = 0; | |
23feba90 | 2603 | } |
f3321554 PX |
2604 | |
2605 | if (tmppages < 0) { | |
d9e474ea | 2606 | pss_host_page_finish(pss); |
f3321554 PX |
2607 | return tmppages; |
2608 | } | |
2609 | ||
d9e474ea PX |
2610 | pss_find_next_dirty(pss); |
2611 | } while (pss_within_range(pss)); | |
2612 | ||
2613 | pss_host_page_finish(pss); | |
278e2f55 AG |
2614 | |
2615 | res = ram_save_release_protection(rs, pss, start_page); | |
2616 | return (res < 0 ? res : pages); | |
a82d593b | 2617 | } |
6c595cde | 2618 | |
56e93d26 | 2619 | /** |
3d0684b2 | 2620 | * ram_find_and_save_block: finds a dirty page and sends it to f |
56e93d26 JQ |
2621 | * |
2622 | * Called within an RCU critical section. | |
2623 | * | |
e8f3735f XG |
2624 | * Returns the number of pages written where zero means no dirty pages, |
2625 | * or negative on error | |
56e93d26 | 2626 | * |
6f37bb8b | 2627 | * @rs: current RAM state |
a82d593b DDAG |
2628 | * |
2629 | * On systems where host-page-size > target-page-size it will send all the | |
2630 | * pages in a host page that are dirty. | |
56e93d26 | 2631 | */ |
05931ec5 | 2632 | static int ram_find_and_save_block(RAMState *rs) |
56e93d26 | 2633 | { |
f1668764 | 2634 | PageSearchStatus *pss = &rs->pss[RAM_CHANNEL_PRECOPY]; |
56e93d26 | 2635 | int pages = 0; |
56e93d26 | 2636 | |
0827b9e9 | 2637 | /* No dirty page as there is zero RAM */ |
8d80e195 | 2638 | if (!rs->ram_bytes_total) { |
0827b9e9 AA |
2639 | return pages; |
2640 | } | |
2641 | ||
4934a5dd PX |
2642 | /* |
2643 | * Always keep last_seen_block/last_page valid during this procedure, | |
2644 | * because find_dirty_block() relies on these values (e.g., we compare | |
2645 | * last_seen_block with pss.block to see whether we searched all the | |
2646 | * ramblocks) to detect the completion of migration. Having NULL value | |
2647 | * of last_seen_block can conditionally cause below loop to run forever. | |
2648 | */ | |
2649 | if (!rs->last_seen_block) { | |
2650 | rs->last_seen_block = QLIST_FIRST_RCU(&ram_list.blocks); | |
2651 | rs->last_page = 0; | |
2652 | } | |
2653 | ||
f1668764 | 2654 | pss_init(pss, rs->last_seen_block, rs->last_page); |
b8fb8cb7 | 2655 | |
31e2ac74 | 2656 | while (true){ |
51efd36f | 2657 | if (!get_queued_page(rs, pss)) { |
b062106d | 2658 | /* priority queue empty, so just search for something dirty */ |
31e2ac74 JQ |
2659 | int res = find_dirty_block(rs, pss); |
2660 | if (res != PAGE_DIRTY_FOUND) { | |
2661 | if (res == PAGE_ALL_CLEAN) { | |
51efd36f | 2662 | break; |
31e2ac74 JQ |
2663 | } else if (res == PAGE_TRY_AGAIN) { |
2664 | continue; | |
294e5a40 JQ |
2665 | } else if (res < 0) { |
2666 | pages = res; | |
2667 | break; | |
51efd36f JQ |
2668 | } |
2669 | } | |
56e93d26 | 2670 | } |
51efd36f | 2671 | pages = ram_save_host_page(rs, pss); |
31e2ac74 JQ |
2672 | if (pages) { |
2673 | break; | |
2674 | } | |
2675 | } | |
56e93d26 | 2676 | |
f1668764 PX |
2677 | rs->last_seen_block = pss->block; |
2678 | rs->last_page = pss->page; | |
56e93d26 JQ |
2679 | |
2680 | return pages; | |
2681 | } | |
2682 | ||
8008a272 | 2683 | static uint64_t ram_bytes_total_with_ignored(void) |
56e93d26 JQ |
2684 | { |
2685 | RAMBlock *block; | |
2686 | uint64_t total = 0; | |
2687 | ||
89ac5a1d DDAG |
2688 | RCU_READ_LOCK_GUARD(); |
2689 | ||
8008a272 JQ |
2690 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
2691 | total += block->used_length; | |
99e15582 | 2692 | } |
56e93d26 JQ |
2693 | return total; |
2694 | } | |
2695 | ||
fbd162e6 YK |
2696 | uint64_t ram_bytes_total(void) |
2697 | { | |
8008a272 JQ |
2698 | RAMBlock *block; |
2699 | uint64_t total = 0; | |
2700 | ||
2701 | RCU_READ_LOCK_GUARD(); | |
2702 | ||
2703 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
2704 | total += block->used_length; | |
2705 | } | |
2706 | return total; | |
fbd162e6 YK |
2707 | } |
2708 | ||
f265e0e4 | 2709 | static void xbzrle_load_setup(void) |
56e93d26 | 2710 | { |
f265e0e4 | 2711 | XBZRLE.decoded_buf = g_malloc(TARGET_PAGE_SIZE); |
56e93d26 JQ |
2712 | } |
2713 | ||
f265e0e4 JQ |
2714 | static void xbzrle_load_cleanup(void) |
2715 | { | |
2716 | g_free(XBZRLE.decoded_buf); | |
2717 | XBZRLE.decoded_buf = NULL; | |
2718 | } | |
2719 | ||
7d7c96be PX |
2720 | static void ram_state_cleanup(RAMState **rsp) |
2721 | { | |
b9ccaf6d DDAG |
2722 | if (*rsp) { |
2723 | migration_page_queue_free(*rsp); | |
2724 | qemu_mutex_destroy(&(*rsp)->bitmap_mutex); | |
2725 | qemu_mutex_destroy(&(*rsp)->src_page_req_mutex); | |
2726 | g_free(*rsp); | |
2727 | *rsp = NULL; | |
2728 | } | |
7d7c96be PX |
2729 | } |
2730 | ||
84593a08 PX |
2731 | static void xbzrle_cleanup(void) |
2732 | { | |
2733 | XBZRLE_cache_lock(); | |
2734 | if (XBZRLE.cache) { | |
2735 | cache_fini(XBZRLE.cache); | |
2736 | g_free(XBZRLE.encoded_buf); | |
2737 | g_free(XBZRLE.current_buf); | |
2738 | g_free(XBZRLE.zero_target_page); | |
2739 | XBZRLE.cache = NULL; | |
2740 | XBZRLE.encoded_buf = NULL; | |
2741 | XBZRLE.current_buf = NULL; | |
2742 | XBZRLE.zero_target_page = NULL; | |
2743 | } | |
2744 | XBZRLE_cache_unlock(); | |
2745 | } | |
2746 | ||
f265e0e4 | 2747 | static void ram_save_cleanup(void *opaque) |
56e93d26 | 2748 | { |
53518d94 | 2749 | RAMState **rsp = opaque; |
6b6712ef | 2750 | RAMBlock *block; |
eb859c53 | 2751 | |
278e2f55 AG |
2752 | /* We don't use dirty log with background snapshots */ |
2753 | if (!migrate_background_snapshot()) { | |
2754 | /* caller have hold iothread lock or is in a bh, so there is | |
2755 | * no writing race against the migration bitmap | |
2756 | */ | |
63b41db4 HH |
2757 | if (global_dirty_tracking & GLOBAL_DIRTY_MIGRATION) { |
2758 | /* | |
2759 | * do not stop dirty log without starting it, since | |
2760 | * memory_global_dirty_log_stop will assert that | |
2761 | * memory_global_dirty_log_start/stop used in pairs | |
2762 | */ | |
2763 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); | |
2764 | } | |
278e2f55 | 2765 | } |
6b6712ef | 2766 | |
fbd162e6 | 2767 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
002cad6b PX |
2768 | g_free(block->clear_bmap); |
2769 | block->clear_bmap = NULL; | |
6b6712ef JQ |
2770 | g_free(block->bmap); |
2771 | block->bmap = NULL; | |
56e93d26 JQ |
2772 | } |
2773 | ||
84593a08 | 2774 | xbzrle_cleanup(); |
f0afa331 | 2775 | compress_threads_save_cleanup(); |
7d7c96be | 2776 | ram_state_cleanup(rsp); |
4010ba38 JQ |
2777 | g_free(migration_ops); |
2778 | migration_ops = NULL; | |
56e93d26 JQ |
2779 | } |
2780 | ||
6f37bb8b | 2781 | static void ram_state_reset(RAMState *rs) |
56e93d26 | 2782 | { |
ec6f3ab9 PX |
2783 | int i; |
2784 | ||
2785 | for (i = 0; i < RAM_CHANNEL_MAX; i++) { | |
2786 | rs->pss[i].last_sent_block = NULL; | |
2787 | } | |
2788 | ||
6f37bb8b | 2789 | rs->last_seen_block = NULL; |
269ace29 | 2790 | rs->last_page = 0; |
6f37bb8b | 2791 | rs->last_version = ram_list.version; |
f3095cc8 | 2792 | rs->xbzrle_started = false; |
56e93d26 JQ |
2793 | } |
2794 | ||
2795 | #define MAX_WAIT 50 /* ms, half buffered_file limit */ | |
2796 | ||
e0b266f0 DDAG |
2797 | /* **** functions for postcopy ***** */ |
2798 | ||
ced1c616 PB |
2799 | void ram_postcopy_migrated_memory_release(MigrationState *ms) |
2800 | { | |
2801 | struct RAMBlock *block; | |
ced1c616 | 2802 | |
fbd162e6 | 2803 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
6b6712ef JQ |
2804 | unsigned long *bitmap = block->bmap; |
2805 | unsigned long range = block->used_length >> TARGET_PAGE_BITS; | |
2806 | unsigned long run_start = find_next_zero_bit(bitmap, range, 0); | |
ced1c616 PB |
2807 | |
2808 | while (run_start < range) { | |
2809 | unsigned long run_end = find_next_bit(bitmap, range, run_start + 1); | |
8bba004c AR |
2810 | ram_discard_range(block->idstr, |
2811 | ((ram_addr_t)run_start) << TARGET_PAGE_BITS, | |
2812 | ((ram_addr_t)(run_end - run_start)) | |
2813 | << TARGET_PAGE_BITS); | |
ced1c616 PB |
2814 | run_start = find_next_zero_bit(bitmap, range, run_end + 1); |
2815 | } | |
2816 | } | |
2817 | } | |
2818 | ||
3d0684b2 JQ |
2819 | /** |
2820 | * postcopy_send_discard_bm_ram: discard a RAMBlock | |
2821 | * | |
e0b266f0 | 2822 | * Callback from postcopy_each_ram_send_discard for each RAMBlock |
3d0684b2 JQ |
2823 | * |
2824 | * @ms: current migration state | |
89dab31b | 2825 | * @block: RAMBlock to discard |
e0b266f0 | 2826 | */ |
9e7d1223 | 2827 | static void postcopy_send_discard_bm_ram(MigrationState *ms, RAMBlock *block) |
e0b266f0 | 2828 | { |
6b6712ef | 2829 | unsigned long end = block->used_length >> TARGET_PAGE_BITS; |
e0b266f0 | 2830 | unsigned long current; |
1e7cf8c3 | 2831 | unsigned long *bitmap = block->bmap; |
e0b266f0 | 2832 | |
6b6712ef | 2833 | for (current = 0; current < end; ) { |
1e7cf8c3 | 2834 | unsigned long one = find_next_bit(bitmap, end, current); |
33a5cb62 | 2835 | unsigned long zero, discard_length; |
e0b266f0 | 2836 | |
33a5cb62 WY |
2837 | if (one >= end) { |
2838 | break; | |
2839 | } | |
e0b266f0 | 2840 | |
1e7cf8c3 | 2841 | zero = find_next_zero_bit(bitmap, end, one + 1); |
33a5cb62 WY |
2842 | |
2843 | if (zero >= end) { | |
2844 | discard_length = end - one; | |
e0b266f0 | 2845 | } else { |
33a5cb62 WY |
2846 | discard_length = zero - one; |
2847 | } | |
810cf2bb | 2848 | postcopy_discard_send_range(ms, one, discard_length); |
33a5cb62 | 2849 | current = one + discard_length; |
e0b266f0 | 2850 | } |
e0b266f0 DDAG |
2851 | } |
2852 | ||
f30c2e5b PX |
2853 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block); |
2854 | ||
3d0684b2 JQ |
2855 | /** |
2856 | * postcopy_each_ram_send_discard: discard all RAMBlocks | |
2857 | * | |
e0b266f0 DDAG |
2858 | * Utility for the outgoing postcopy code. |
2859 | * Calls postcopy_send_discard_bm_ram for each RAMBlock | |
2860 | * passing it bitmap indexes and name. | |
e0b266f0 DDAG |
2861 | * (qemu_ram_foreach_block ends up passing unscaled lengths |
2862 | * which would mean postcopy code would have to deal with target page) | |
3d0684b2 JQ |
2863 | * |
2864 | * @ms: current migration state | |
e0b266f0 | 2865 | */ |
739fcc1b | 2866 | static void postcopy_each_ram_send_discard(MigrationState *ms) |
e0b266f0 DDAG |
2867 | { |
2868 | struct RAMBlock *block; | |
e0b266f0 | 2869 | |
fbd162e6 | 2870 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
810cf2bb | 2871 | postcopy_discard_send_init(ms, block->idstr); |
e0b266f0 | 2872 | |
f30c2e5b PX |
2873 | /* |
2874 | * Deal with TPS != HPS and huge pages. It discard any partially sent | |
2875 | * host-page size chunks, mark any partially dirty host-page size | |
2876 | * chunks as all dirty. In this case the host-page is the host-page | |
2877 | * for the particular RAMBlock, i.e. it might be a huge page. | |
2878 | */ | |
2879 | postcopy_chunk_hostpages_pass(ms, block); | |
2880 | ||
e0b266f0 DDAG |
2881 | /* |
2882 | * Postcopy sends chunks of bitmap over the wire, but it | |
2883 | * just needs indexes at this point, avoids it having | |
2884 | * target page specific code. | |
2885 | */ | |
739fcc1b | 2886 | postcopy_send_discard_bm_ram(ms, block); |
810cf2bb | 2887 | postcopy_discard_send_finish(ms); |
e0b266f0 | 2888 | } |
e0b266f0 DDAG |
2889 | } |
2890 | ||
3d0684b2 | 2891 | /** |
8324ef86 | 2892 | * postcopy_chunk_hostpages_pass: canonicalize bitmap in hostpages |
3d0684b2 JQ |
2893 | * |
2894 | * Helper for postcopy_chunk_hostpages; it's called twice to | |
2895 | * canonicalize the two bitmaps, that are similar, but one is | |
2896 | * inverted. | |
99e314eb | 2897 | * |
3d0684b2 JQ |
2898 | * Postcopy requires that all target pages in a hostpage are dirty or |
2899 | * clean, not a mix. This function canonicalizes the bitmaps. | |
99e314eb | 2900 | * |
3d0684b2 | 2901 | * @ms: current migration state |
3d0684b2 | 2902 | * @block: block that contains the page we want to canonicalize |
99e314eb | 2903 | */ |
1e7cf8c3 | 2904 | static void postcopy_chunk_hostpages_pass(MigrationState *ms, RAMBlock *block) |
99e314eb | 2905 | { |
53518d94 | 2906 | RAMState *rs = ram_state; |
6b6712ef | 2907 | unsigned long *bitmap = block->bmap; |
29c59172 | 2908 | unsigned int host_ratio = block->page_size / TARGET_PAGE_SIZE; |
6b6712ef | 2909 | unsigned long pages = block->used_length >> TARGET_PAGE_BITS; |
99e314eb DDAG |
2910 | unsigned long run_start; |
2911 | ||
29c59172 DDAG |
2912 | if (block->page_size == TARGET_PAGE_SIZE) { |
2913 | /* Easy case - TPS==HPS for a non-huge page RAMBlock */ | |
2914 | return; | |
2915 | } | |
2916 | ||
1e7cf8c3 WY |
2917 | /* Find a dirty page */ |
2918 | run_start = find_next_bit(bitmap, pages, 0); | |
99e314eb | 2919 | |
6b6712ef | 2920 | while (run_start < pages) { |
99e314eb DDAG |
2921 | |
2922 | /* | |
2923 | * If the start of this run of pages is in the middle of a host | |
2924 | * page, then we need to fixup this host page. | |
2925 | */ | |
9dec3cc3 | 2926 | if (QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2927 | /* Find the end of this run */ |
1e7cf8c3 | 2928 | run_start = find_next_zero_bit(bitmap, pages, run_start + 1); |
99e314eb DDAG |
2929 | /* |
2930 | * If the end isn't at the start of a host page, then the | |
2931 | * run doesn't finish at the end of a host page | |
2932 | * and we need to discard. | |
2933 | */ | |
99e314eb DDAG |
2934 | } |
2935 | ||
9dec3cc3 | 2936 | if (!QEMU_IS_ALIGNED(run_start, host_ratio)) { |
99e314eb | 2937 | unsigned long page; |
dad45ab2 WY |
2938 | unsigned long fixup_start_addr = QEMU_ALIGN_DOWN(run_start, |
2939 | host_ratio); | |
2940 | run_start = QEMU_ALIGN_UP(run_start, host_ratio); | |
99e314eb | 2941 | |
99e314eb DDAG |
2942 | /* Clean up the bitmap */ |
2943 | for (page = fixup_start_addr; | |
2944 | page < fixup_start_addr + host_ratio; page++) { | |
99e314eb DDAG |
2945 | /* |
2946 | * Remark them as dirty, updating the count for any pages | |
2947 | * that weren't previously dirty. | |
2948 | */ | |
0d8ec885 | 2949 | rs->migration_dirty_pages += !test_and_set_bit(page, bitmap); |
99e314eb DDAG |
2950 | } |
2951 | } | |
2952 | ||
1e7cf8c3 WY |
2953 | /* Find the next dirty page for the next iteration */ |
2954 | run_start = find_next_bit(bitmap, pages, run_start); | |
99e314eb DDAG |
2955 | } |
2956 | } | |
2957 | ||
3d0684b2 JQ |
2958 | /** |
2959 | * ram_postcopy_send_discard_bitmap: transmit the discard bitmap | |
2960 | * | |
e0b266f0 DDAG |
2961 | * Transmit the set of pages to be discarded after precopy to the target |
2962 | * these are pages that: | |
2963 | * a) Have been previously transmitted but are now dirty again | |
2964 | * b) Pages that have never been transmitted, this ensures that | |
2965 | * any pages on the destination that have been mapped by background | |
2966 | * tasks get discarded (transparent huge pages is the specific concern) | |
2967 | * Hopefully this is pretty sparse | |
3d0684b2 JQ |
2968 | * |
2969 | * @ms: current migration state | |
e0b266f0 | 2970 | */ |
739fcc1b | 2971 | void ram_postcopy_send_discard_bitmap(MigrationState *ms) |
e0b266f0 | 2972 | { |
53518d94 | 2973 | RAMState *rs = ram_state; |
e0b266f0 | 2974 | |
89ac5a1d | 2975 | RCU_READ_LOCK_GUARD(); |
e0b266f0 DDAG |
2976 | |
2977 | /* This should be our last sync, the src is now paused */ | |
eb859c53 | 2978 | migration_bitmap_sync(rs); |
e0b266f0 | 2979 | |
6b6712ef | 2980 | /* Easiest way to make sure we don't resume in the middle of a host-page */ |
ec6f3ab9 | 2981 | rs->pss[RAM_CHANNEL_PRECOPY].last_sent_block = NULL; |
6b6712ef | 2982 | rs->last_seen_block = NULL; |
6b6712ef | 2983 | rs->last_page = 0; |
e0b266f0 | 2984 | |
739fcc1b | 2985 | postcopy_each_ram_send_discard(ms); |
e0b266f0 | 2986 | |
739fcc1b | 2987 | trace_ram_postcopy_send_discard_bitmap(); |
e0b266f0 DDAG |
2988 | } |
2989 | ||
3d0684b2 JQ |
2990 | /** |
2991 | * ram_discard_range: discard dirtied pages at the beginning of postcopy | |
e0b266f0 | 2992 | * |
3d0684b2 | 2993 | * Returns zero on success |
e0b266f0 | 2994 | * |
36449157 JQ |
2995 | * @rbname: name of the RAMBlock of the request. NULL means the |
2996 | * same that last one. | |
3d0684b2 JQ |
2997 | * @start: RAMBlock starting page |
2998 | * @length: RAMBlock size | |
e0b266f0 | 2999 | */ |
aaa2064c | 3000 | int ram_discard_range(const char *rbname, uint64_t start, size_t length) |
e0b266f0 | 3001 | { |
36449157 | 3002 | trace_ram_discard_range(rbname, start, length); |
d3a5038c | 3003 | |
89ac5a1d | 3004 | RCU_READ_LOCK_GUARD(); |
36449157 | 3005 | RAMBlock *rb = qemu_ram_block_by_name(rbname); |
e0b266f0 DDAG |
3006 | |
3007 | if (!rb) { | |
36449157 | 3008 | error_report("ram_discard_range: Failed to find block '%s'", rbname); |
03acb4e9 | 3009 | return -1; |
e0b266f0 DDAG |
3010 | } |
3011 | ||
814bb08f PX |
3012 | /* |
3013 | * On source VM, we don't need to update the received bitmap since | |
3014 | * we don't even have one. | |
3015 | */ | |
3016 | if (rb->receivedmap) { | |
3017 | bitmap_clear(rb->receivedmap, start >> qemu_target_page_bits(), | |
3018 | length >> qemu_target_page_bits()); | |
3019 | } | |
3020 | ||
03acb4e9 | 3021 | return ram_block_discard_range(rb, start, length); |
e0b266f0 DDAG |
3022 | } |
3023 | ||
84593a08 PX |
3024 | /* |
3025 | * For every allocation, we will try not to crash the VM if the | |
3026 | * allocation failed. | |
3027 | */ | |
3028 | static int xbzrle_init(void) | |
3029 | { | |
3030 | Error *local_err = NULL; | |
3031 | ||
87dca0c9 | 3032 | if (!migrate_xbzrle()) { |
84593a08 PX |
3033 | return 0; |
3034 | } | |
3035 | ||
3036 | XBZRLE_cache_lock(); | |
3037 | ||
3038 | XBZRLE.zero_target_page = g_try_malloc0(TARGET_PAGE_SIZE); | |
3039 | if (!XBZRLE.zero_target_page) { | |
3040 | error_report("%s: Error allocating zero page", __func__); | |
3041 | goto err_out; | |
3042 | } | |
3043 | ||
3044 | XBZRLE.cache = cache_init(migrate_xbzrle_cache_size(), | |
3045 | TARGET_PAGE_SIZE, &local_err); | |
3046 | if (!XBZRLE.cache) { | |
3047 | error_report_err(local_err); | |
3048 | goto free_zero_page; | |
3049 | } | |
3050 | ||
3051 | XBZRLE.encoded_buf = g_try_malloc0(TARGET_PAGE_SIZE); | |
3052 | if (!XBZRLE.encoded_buf) { | |
3053 | error_report("%s: Error allocating encoded_buf", __func__); | |
3054 | goto free_cache; | |
3055 | } | |
3056 | ||
3057 | XBZRLE.current_buf = g_try_malloc(TARGET_PAGE_SIZE); | |
3058 | if (!XBZRLE.current_buf) { | |
3059 | error_report("%s: Error allocating current_buf", __func__); | |
3060 | goto free_encoded_buf; | |
3061 | } | |
3062 | ||
3063 | /* We are all good */ | |
3064 | XBZRLE_cache_unlock(); | |
3065 | return 0; | |
3066 | ||
3067 | free_encoded_buf: | |
3068 | g_free(XBZRLE.encoded_buf); | |
3069 | XBZRLE.encoded_buf = NULL; | |
3070 | free_cache: | |
3071 | cache_fini(XBZRLE.cache); | |
3072 | XBZRLE.cache = NULL; | |
3073 | free_zero_page: | |
3074 | g_free(XBZRLE.zero_target_page); | |
3075 | XBZRLE.zero_target_page = NULL; | |
3076 | err_out: | |
3077 | XBZRLE_cache_unlock(); | |
3078 | return -ENOMEM; | |
3079 | } | |
3080 | ||
53518d94 | 3081 | static int ram_state_init(RAMState **rsp) |
56e93d26 | 3082 | { |
7d00ee6a PX |
3083 | *rsp = g_try_new0(RAMState, 1); |
3084 | ||
3085 | if (!*rsp) { | |
3086 | error_report("%s: Init ramstate fail", __func__); | |
3087 | return -1; | |
3088 | } | |
53518d94 JQ |
3089 | |
3090 | qemu_mutex_init(&(*rsp)->bitmap_mutex); | |
3091 | qemu_mutex_init(&(*rsp)->src_page_req_mutex); | |
3092 | QSIMPLEQ_INIT(&(*rsp)->src_page_requests); | |
8d80e195 | 3093 | (*rsp)->ram_bytes_total = ram_bytes_total(); |
56e93d26 | 3094 | |
7d00ee6a | 3095 | /* |
40c4d4a8 IR |
3096 | * Count the total number of pages used by ram blocks not including any |
3097 | * gaps due to alignment or unplugs. | |
03158519 | 3098 | * This must match with the initial values of dirty bitmap. |
7d00ee6a | 3099 | */ |
8d80e195 | 3100 | (*rsp)->migration_dirty_pages = (*rsp)->ram_bytes_total >> TARGET_PAGE_BITS; |
7d00ee6a PX |
3101 | ram_state_reset(*rsp); |
3102 | ||
3103 | return 0; | |
3104 | } | |
3105 | ||
d6eff5d7 | 3106 | static void ram_list_init_bitmaps(void) |
7d00ee6a | 3107 | { |
002cad6b | 3108 | MigrationState *ms = migrate_get_current(); |
d6eff5d7 PX |
3109 | RAMBlock *block; |
3110 | unsigned long pages; | |
002cad6b | 3111 | uint8_t shift; |
56e93d26 | 3112 | |
0827b9e9 AA |
3113 | /* Skip setting bitmap if there is no RAM */ |
3114 | if (ram_bytes_total()) { | |
002cad6b PX |
3115 | shift = ms->clear_bitmap_shift; |
3116 | if (shift > CLEAR_BITMAP_SHIFT_MAX) { | |
3117 | error_report("clear_bitmap_shift (%u) too big, using " | |
3118 | "max value (%u)", shift, CLEAR_BITMAP_SHIFT_MAX); | |
3119 | shift = CLEAR_BITMAP_SHIFT_MAX; | |
3120 | } else if (shift < CLEAR_BITMAP_SHIFT_MIN) { | |
3121 | error_report("clear_bitmap_shift (%u) too small, using " | |
3122 | "min value (%u)", shift, CLEAR_BITMAP_SHIFT_MIN); | |
3123 | shift = CLEAR_BITMAP_SHIFT_MIN; | |
3124 | } | |
3125 | ||
fbd162e6 | 3126 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
d6eff5d7 | 3127 | pages = block->max_length >> TARGET_PAGE_BITS; |
03158519 WY |
3128 | /* |
3129 | * The initial dirty bitmap for migration must be set with all | |
3130 | * ones to make sure we'll migrate every guest RAM page to | |
3131 | * destination. | |
40c4d4a8 IR |
3132 | * Here we set RAMBlock.bmap all to 1 because when rebegin a |
3133 | * new migration after a failed migration, ram_list. | |
3134 | * dirty_memory[DIRTY_MEMORY_MIGRATION] don't include the whole | |
3135 | * guest memory. | |
03158519 | 3136 | */ |
6b6712ef | 3137 | block->bmap = bitmap_new(pages); |
40c4d4a8 | 3138 | bitmap_set(block->bmap, 0, pages); |
002cad6b PX |
3139 | block->clear_bmap_shift = shift; |
3140 | block->clear_bmap = bitmap_new(clear_bmap_size(pages, shift)); | |
0827b9e9 | 3141 | } |
f3f491fc | 3142 | } |
d6eff5d7 PX |
3143 | } |
3144 | ||
be39b4cd DH |
3145 | static void migration_bitmap_clear_discarded_pages(RAMState *rs) |
3146 | { | |
3147 | unsigned long pages; | |
3148 | RAMBlock *rb; | |
3149 | ||
3150 | RCU_READ_LOCK_GUARD(); | |
3151 | ||
3152 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { | |
3153 | pages = ramblock_dirty_bitmap_clear_discarded_pages(rb); | |
3154 | rs->migration_dirty_pages -= pages; | |
3155 | } | |
3156 | } | |
3157 | ||
d6eff5d7 PX |
3158 | static void ram_init_bitmaps(RAMState *rs) |
3159 | { | |
3160 | /* For memory_global_dirty_log_start below. */ | |
3161 | qemu_mutex_lock_iothread(); | |
3162 | qemu_mutex_lock_ramlist(); | |
f3f491fc | 3163 | |
89ac5a1d DDAG |
3164 | WITH_RCU_READ_LOCK_GUARD() { |
3165 | ram_list_init_bitmaps(); | |
278e2f55 AG |
3166 | /* We don't use dirty log with background snapshots */ |
3167 | if (!migrate_background_snapshot()) { | |
63b41db4 | 3168 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
278e2f55 AG |
3169 | migration_bitmap_sync_precopy(rs); |
3170 | } | |
89ac5a1d | 3171 | } |
56e93d26 | 3172 | qemu_mutex_unlock_ramlist(); |
49877834 | 3173 | qemu_mutex_unlock_iothread(); |
be39b4cd DH |
3174 | |
3175 | /* | |
3176 | * After an eventual first bitmap sync, fixup the initial bitmap | |
3177 | * containing all 1s to exclude any discarded pages from migration. | |
3178 | */ | |
3179 | migration_bitmap_clear_discarded_pages(rs); | |
d6eff5d7 PX |
3180 | } |
3181 | ||
3182 | static int ram_init_all(RAMState **rsp) | |
3183 | { | |
3184 | if (ram_state_init(rsp)) { | |
3185 | return -1; | |
3186 | } | |
3187 | ||
3188 | if (xbzrle_init()) { | |
3189 | ram_state_cleanup(rsp); | |
3190 | return -1; | |
3191 | } | |
3192 | ||
3193 | ram_init_bitmaps(*rsp); | |
a91246c9 HZ |
3194 | |
3195 | return 0; | |
3196 | } | |
3197 | ||
08614f34 PX |
3198 | static void ram_state_resume_prepare(RAMState *rs, QEMUFile *out) |
3199 | { | |
3200 | RAMBlock *block; | |
3201 | uint64_t pages = 0; | |
3202 | ||
3203 | /* | |
3204 | * Postcopy is not using xbzrle/compression, so no need for that. | |
3205 | * Also, since source are already halted, we don't need to care | |
3206 | * about dirty page logging as well. | |
3207 | */ | |
3208 | ||
fbd162e6 | 3209 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
08614f34 PX |
3210 | pages += bitmap_count_one(block->bmap, |
3211 | block->used_length >> TARGET_PAGE_BITS); | |
3212 | } | |
3213 | ||
3214 | /* This may not be aligned with current bitmaps. Recalculate. */ | |
3215 | rs->migration_dirty_pages = pages; | |
3216 | ||
1a373522 | 3217 | ram_state_reset(rs); |
08614f34 PX |
3218 | |
3219 | /* Update RAMState cache of output QEMUFile */ | |
7f401b80 | 3220 | rs->pss[RAM_CHANNEL_PRECOPY].pss_channel = out; |
08614f34 PX |
3221 | |
3222 | trace_ram_state_resume_prepare(pages); | |
3223 | } | |
3224 | ||
6bcb05fc WW |
3225 | /* |
3226 | * This function clears bits of the free pages reported by the caller from the | |
3227 | * migration dirty bitmap. @addr is the host address corresponding to the | |
3228 | * start of the continuous guest free pages, and @len is the total bytes of | |
3229 | * those pages. | |
3230 | */ | |
3231 | void qemu_guest_free_page_hint(void *addr, size_t len) | |
3232 | { | |
3233 | RAMBlock *block; | |
3234 | ram_addr_t offset; | |
3235 | size_t used_len, start, npages; | |
3236 | MigrationState *s = migrate_get_current(); | |
3237 | ||
3238 | /* This function is currently expected to be used during live migration */ | |
3239 | if (!migration_is_setup_or_active(s->state)) { | |
3240 | return; | |
3241 | } | |
3242 | ||
3243 | for (; len > 0; len -= used_len, addr += used_len) { | |
3244 | block = qemu_ram_block_from_host(addr, false, &offset); | |
3245 | if (unlikely(!block || offset >= block->used_length)) { | |
3246 | /* | |
3247 | * The implementation might not support RAMBlock resize during | |
3248 | * live migration, but it could happen in theory with future | |
3249 | * updates. So we add a check here to capture that case. | |
3250 | */ | |
3251 | error_report_once("%s unexpected error", __func__); | |
3252 | return; | |
3253 | } | |
3254 | ||
3255 | if (len <= block->used_length - offset) { | |
3256 | used_len = len; | |
3257 | } else { | |
3258 | used_len = block->used_length - offset; | |
3259 | } | |
3260 | ||
3261 | start = offset >> TARGET_PAGE_BITS; | |
3262 | npages = used_len >> TARGET_PAGE_BITS; | |
3263 | ||
3264 | qemu_mutex_lock(&ram_state->bitmap_mutex); | |
3143577d WW |
3265 | /* |
3266 | * The skipped free pages are equavalent to be sent from clear_bmap's | |
3267 | * perspective, so clear the bits from the memory region bitmap which | |
3268 | * are initially set. Otherwise those skipped pages will be sent in | |
3269 | * the next round after syncing from the memory region bitmap. | |
3270 | */ | |
1230a25f | 3271 | migration_clear_memory_region_dirty_bitmap_range(block, start, npages); |
6bcb05fc WW |
3272 | ram_state->migration_dirty_pages -= |
3273 | bitmap_count_one_with_offset(block->bmap, start, npages); | |
3274 | bitmap_clear(block->bmap, start, npages); | |
3275 | qemu_mutex_unlock(&ram_state->bitmap_mutex); | |
3276 | } | |
3277 | } | |
3278 | ||
3d0684b2 JQ |
3279 | /* |
3280 | * Each of ram_save_setup, ram_save_iterate and ram_save_complete has | |
a91246c9 HZ |
3281 | * long-running RCU critical section. When rcu-reclaims in the code |
3282 | * start to become numerous it will be necessary to reduce the | |
3283 | * granularity of these critical sections. | |
3284 | */ | |
3285 | ||
3d0684b2 JQ |
3286 | /** |
3287 | * ram_save_setup: Setup RAM for migration | |
3288 | * | |
3289 | * Returns zero to indicate success and negative for error | |
3290 | * | |
3291 | * @f: QEMUFile where to send the data | |
3292 | * @opaque: RAMState pointer | |
3293 | */ | |
a91246c9 HZ |
3294 | static int ram_save_setup(QEMUFile *f, void *opaque) |
3295 | { | |
53518d94 | 3296 | RAMState **rsp = opaque; |
a91246c9 | 3297 | RAMBlock *block; |
33d70973 | 3298 | int ret; |
a91246c9 | 3299 | |
dcaf446e XG |
3300 | if (compress_threads_save_setup()) { |
3301 | return -1; | |
3302 | } | |
3303 | ||
a91246c9 HZ |
3304 | /* migration has already setup the bitmap, reuse it. */ |
3305 | if (!migration_in_colo_state()) { | |
7d00ee6a | 3306 | if (ram_init_all(rsp) != 0) { |
dcaf446e | 3307 | compress_threads_save_cleanup(); |
a91246c9 | 3308 | return -1; |
53518d94 | 3309 | } |
a91246c9 | 3310 | } |
7f401b80 | 3311 | (*rsp)->pss[RAM_CHANNEL_PRECOPY].pss_channel = f; |
a91246c9 | 3312 | |
0e6ebd48 | 3313 | WITH_RCU_READ_LOCK_GUARD() { |
8008a272 JQ |
3314 | qemu_put_be64(f, ram_bytes_total_with_ignored() |
3315 | | RAM_SAVE_FLAG_MEM_SIZE); | |
56e93d26 | 3316 | |
0e6ebd48 DDAG |
3317 | RAMBLOCK_FOREACH_MIGRATABLE(block) { |
3318 | qemu_put_byte(f, strlen(block->idstr)); | |
3319 | qemu_put_buffer(f, (uint8_t *)block->idstr, strlen(block->idstr)); | |
3320 | qemu_put_be64(f, block->used_length); | |
3321 | if (migrate_postcopy_ram() && block->page_size != | |
3322 | qemu_host_page_size) { | |
3323 | qemu_put_be64(f, block->page_size); | |
3324 | } | |
3325 | if (migrate_ignore_shared()) { | |
3326 | qemu_put_be64(f, block->mr->addr); | |
3327 | } | |
fbd162e6 | 3328 | } |
56e93d26 JQ |
3329 | } |
3330 | ||
56e93d26 JQ |
3331 | ram_control_before_iterate(f, RAM_CONTROL_SETUP); |
3332 | ram_control_after_iterate(f, RAM_CONTROL_SETUP); | |
3333 | ||
4010ba38 JQ |
3334 | migration_ops = g_malloc0(sizeof(MigrationOps)); |
3335 | migration_ops->ram_save_target_page = ram_save_target_page_legacy; | |
8ebb6ecc | 3336 | ret = multifd_send_sync_main(f); |
33d70973 LB |
3337 | if (ret < 0) { |
3338 | return ret; | |
3339 | } | |
3340 | ||
294e5a40 JQ |
3341 | if (!migrate_multifd_flush_after_each_section()) { |
3342 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); | |
3343 | } | |
3344 | ||
56e93d26 | 3345 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
35374cbd | 3346 | qemu_fflush(f); |
56e93d26 JQ |
3347 | |
3348 | return 0; | |
3349 | } | |
3350 | ||
3d0684b2 JQ |
3351 | /** |
3352 | * ram_save_iterate: iterative stage for migration | |
3353 | * | |
3354 | * Returns zero to indicate success and negative for error | |
3355 | * | |
3356 | * @f: QEMUFile where to send the data | |
3357 | * @opaque: RAMState pointer | |
3358 | */ | |
56e93d26 JQ |
3359 | static int ram_save_iterate(QEMUFile *f, void *opaque) |
3360 | { | |
53518d94 JQ |
3361 | RAMState **temp = opaque; |
3362 | RAMState *rs = *temp; | |
3d4095b2 | 3363 | int ret = 0; |
56e93d26 JQ |
3364 | int i; |
3365 | int64_t t0; | |
5c90308f | 3366 | int done = 0; |
56e93d26 | 3367 | |
b2557345 PL |
3368 | if (blk_mig_bulk_active()) { |
3369 | /* Avoid transferring ram during bulk phase of block migration as | |
3370 | * the bulk phase will usually take a long time and transferring | |
3371 | * ram updates during that time is pointless. */ | |
3372 | goto out; | |
3373 | } | |
3374 | ||
63268c49 PX |
3375 | /* |
3376 | * We'll take this lock a little bit long, but it's okay for two reasons. | |
3377 | * Firstly, the only possible other thread to take it is who calls | |
3378 | * qemu_guest_free_page_hint(), which should be rare; secondly, see | |
3379 | * MAX_WAIT (if curious, further see commit 4508bd9ed8053ce) below, which | |
3380 | * guarantees that we'll at least released it in a regular basis. | |
3381 | */ | |
3382 | qemu_mutex_lock(&rs->bitmap_mutex); | |
89ac5a1d DDAG |
3383 | WITH_RCU_READ_LOCK_GUARD() { |
3384 | if (ram_list.version != rs->last_version) { | |
3385 | ram_state_reset(rs); | |
3386 | } | |
56e93d26 | 3387 | |
89ac5a1d DDAG |
3388 | /* Read version before ram_list.blocks */ |
3389 | smp_rmb(); | |
56e93d26 | 3390 | |
89ac5a1d | 3391 | ram_control_before_iterate(f, RAM_CONTROL_ROUND); |
56e93d26 | 3392 | |
89ac5a1d DDAG |
3393 | t0 = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); |
3394 | i = 0; | |
3395 | while ((ret = qemu_file_rate_limit(f)) == 0 || | |
a1fe28df | 3396 | postcopy_has_request(rs)) { |
89ac5a1d | 3397 | int pages; |
e03a34f8 | 3398 | |
89ac5a1d DDAG |
3399 | if (qemu_file_get_error(f)) { |
3400 | break; | |
3401 | } | |
e8f3735f | 3402 | |
05931ec5 | 3403 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3404 | /* no more pages to sent */ |
3405 | if (pages == 0) { | |
3406 | done = 1; | |
3407 | break; | |
3408 | } | |
e8f3735f | 3409 | |
89ac5a1d DDAG |
3410 | if (pages < 0) { |
3411 | qemu_file_set_error(f, pages); | |
56e93d26 JQ |
3412 | break; |
3413 | } | |
89ac5a1d DDAG |
3414 | |
3415 | rs->target_page_count += pages; | |
3416 | ||
644acf99 WY |
3417 | /* |
3418 | * During postcopy, it is necessary to make sure one whole host | |
3419 | * page is sent in one chunk. | |
3420 | */ | |
3421 | if (migrate_postcopy_ram()) { | |
ef4f5f5d | 3422 | ram_flush_compressed_data(rs); |
644acf99 WY |
3423 | } |
3424 | ||
89ac5a1d DDAG |
3425 | /* |
3426 | * we want to check in the 1st loop, just in case it was the 1st | |
3427 | * time and we had to sync the dirty bitmap. | |
3428 | * qemu_clock_get_ns() is a bit expensive, so we only check each | |
3429 | * some iterations | |
3430 | */ | |
3431 | if ((i & 63) == 0) { | |
3432 | uint64_t t1 = (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - t0) / | |
3433 | 1000000; | |
3434 | if (t1 > MAX_WAIT) { | |
3435 | trace_ram_save_iterate_big_wait(t1, i); | |
3436 | break; | |
3437 | } | |
3438 | } | |
3439 | i++; | |
56e93d26 | 3440 | } |
56e93d26 | 3441 | } |
63268c49 | 3442 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 JQ |
3443 | |
3444 | /* | |
3445 | * Must occur before EOS (or any QEMUFile operation) | |
3446 | * because of RDMA protocol. | |
3447 | */ | |
3448 | ram_control_after_iterate(f, RAM_CONTROL_ROUND); | |
3449 | ||
b2557345 | 3450 | out: |
b69a0227 JQ |
3451 | if (ret >= 0 |
3452 | && migration_is_setup_or_active(migrate_get_current()->state)) { | |
b05292c2 JQ |
3453 | if (migrate_multifd_flush_after_each_section()) { |
3454 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); | |
3455 | if (ret < 0) { | |
3456 | return ret; | |
3457 | } | |
33d70973 LB |
3458 | } |
3459 | ||
3d4095b2 JQ |
3460 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3461 | qemu_fflush(f); | |
4c2d0f6d | 3462 | ram_transferred_add(8); |
56e93d26 | 3463 | |
3d4095b2 JQ |
3464 | ret = qemu_file_get_error(f); |
3465 | } | |
56e93d26 JQ |
3466 | if (ret < 0) { |
3467 | return ret; | |
3468 | } | |
3469 | ||
5c90308f | 3470 | return done; |
56e93d26 JQ |
3471 | } |
3472 | ||
3d0684b2 JQ |
3473 | /** |
3474 | * ram_save_complete: function called to send the remaining amount of ram | |
3475 | * | |
e8f3735f | 3476 | * Returns zero to indicate success or negative on error |
3d0684b2 JQ |
3477 | * |
3478 | * Called with iothread lock | |
3479 | * | |
3480 | * @f: QEMUFile where to send the data | |
3481 | * @opaque: RAMState pointer | |
3482 | */ | |
56e93d26 JQ |
3483 | static int ram_save_complete(QEMUFile *f, void *opaque) |
3484 | { | |
53518d94 JQ |
3485 | RAMState **temp = opaque; |
3486 | RAMState *rs = *temp; | |
e8f3735f | 3487 | int ret = 0; |
6f37bb8b | 3488 | |
05931ec5 JQ |
3489 | rs->last_stage = !migration_in_colo_state(); |
3490 | ||
89ac5a1d DDAG |
3491 | WITH_RCU_READ_LOCK_GUARD() { |
3492 | if (!migration_in_postcopy()) { | |
3493 | migration_bitmap_sync_precopy(rs); | |
3494 | } | |
56e93d26 | 3495 | |
89ac5a1d | 3496 | ram_control_before_iterate(f, RAM_CONTROL_FINISH); |
56e93d26 | 3497 | |
89ac5a1d | 3498 | /* try transferring iterative blocks of memory */ |
56e93d26 | 3499 | |
89ac5a1d | 3500 | /* flush all remaining blocks regardless of rate limiting */ |
c13221b5 | 3501 | qemu_mutex_lock(&rs->bitmap_mutex); |
89ac5a1d DDAG |
3502 | while (true) { |
3503 | int pages; | |
56e93d26 | 3504 | |
05931ec5 | 3505 | pages = ram_find_and_save_block(rs); |
89ac5a1d DDAG |
3506 | /* no more blocks to sent */ |
3507 | if (pages == 0) { | |
3508 | break; | |
3509 | } | |
3510 | if (pages < 0) { | |
3511 | ret = pages; | |
3512 | break; | |
3513 | } | |
e8f3735f | 3514 | } |
c13221b5 | 3515 | qemu_mutex_unlock(&rs->bitmap_mutex); |
56e93d26 | 3516 | |
ef4f5f5d | 3517 | ram_flush_compressed_data(rs); |
89ac5a1d DDAG |
3518 | ram_control_after_iterate(f, RAM_CONTROL_FINISH); |
3519 | } | |
d09a6fde | 3520 | |
33d70973 LB |
3521 | if (ret < 0) { |
3522 | return ret; | |
3d4095b2 | 3523 | } |
56e93d26 | 3524 | |
7f401b80 | 3525 | ret = multifd_send_sync_main(rs->pss[RAM_CHANNEL_PRECOPY].pss_channel); |
33d70973 LB |
3526 | if (ret < 0) { |
3527 | return ret; | |
3528 | } | |
3529 | ||
294e5a40 JQ |
3530 | if (!migrate_multifd_flush_after_each_section()) { |
3531 | qemu_put_be64(f, RAM_SAVE_FLAG_MULTIFD_FLUSH); | |
3532 | } | |
33d70973 LB |
3533 | qemu_put_be64(f, RAM_SAVE_FLAG_EOS); |
3534 | qemu_fflush(f); | |
3535 | ||
3536 | return 0; | |
56e93d26 JQ |
3537 | } |
3538 | ||
24beea4e JQ |
3539 | static void ram_state_pending_estimate(void *opaque, uint64_t *must_precopy, |
3540 | uint64_t *can_postcopy) | |
56e93d26 | 3541 | { |
53518d94 JQ |
3542 | RAMState **temp = opaque; |
3543 | RAMState *rs = *temp; | |
56e93d26 | 3544 | |
c8df4a7a | 3545 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3546 | |
c8df4a7a JQ |
3547 | if (migrate_postcopy_ram()) { |
3548 | /* We can do postcopy, and all the data is postcopiable */ | |
24beea4e | 3549 | *can_postcopy += remaining_size; |
c8df4a7a | 3550 | } else { |
24beea4e | 3551 | *must_precopy += remaining_size; |
c8df4a7a JQ |
3552 | } |
3553 | } | |
3554 | ||
24beea4e JQ |
3555 | static void ram_state_pending_exact(void *opaque, uint64_t *must_precopy, |
3556 | uint64_t *can_postcopy) | |
c8df4a7a | 3557 | { |
28ef5339 | 3558 | MigrationState *s = migrate_get_current(); |
c8df4a7a JQ |
3559 | RAMState **temp = opaque; |
3560 | RAMState *rs = *temp; | |
3561 | ||
3562 | uint64_t remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; | |
3563 | ||
28ef5339 | 3564 | if (!migration_in_postcopy() && remaining_size < s->threshold_size) { |
56e93d26 | 3565 | qemu_mutex_lock_iothread(); |
89ac5a1d DDAG |
3566 | WITH_RCU_READ_LOCK_GUARD() { |
3567 | migration_bitmap_sync_precopy(rs); | |
3568 | } | |
56e93d26 | 3569 | qemu_mutex_unlock_iothread(); |
9edabd4d | 3570 | remaining_size = rs->migration_dirty_pages * TARGET_PAGE_SIZE; |
56e93d26 | 3571 | } |
c31b098f | 3572 | |
86e1167e VSO |
3573 | if (migrate_postcopy_ram()) { |
3574 | /* We can do postcopy, and all the data is postcopiable */ | |
24beea4e | 3575 | *can_postcopy += remaining_size; |
86e1167e | 3576 | } else { |
24beea4e | 3577 | *must_precopy += remaining_size; |
86e1167e | 3578 | } |
56e93d26 JQ |
3579 | } |
3580 | ||
3581 | static int load_xbzrle(QEMUFile *f, ram_addr_t addr, void *host) | |
3582 | { | |
3583 | unsigned int xh_len; | |
3584 | int xh_flags; | |
063e760a | 3585 | uint8_t *loaded_data; |
56e93d26 | 3586 | |
56e93d26 JQ |
3587 | /* extract RLE header */ |
3588 | xh_flags = qemu_get_byte(f); | |
3589 | xh_len = qemu_get_be16(f); | |
3590 | ||
3591 | if (xh_flags != ENCODING_FLAG_XBZRLE) { | |
3592 | error_report("Failed to load XBZRLE page - wrong compression!"); | |
3593 | return -1; | |
3594 | } | |
3595 | ||
3596 | if (xh_len > TARGET_PAGE_SIZE) { | |
3597 | error_report("Failed to load XBZRLE page - len overflow!"); | |
3598 | return -1; | |
3599 | } | |
f265e0e4 | 3600 | loaded_data = XBZRLE.decoded_buf; |
56e93d26 | 3601 | /* load data and decode */ |
f265e0e4 | 3602 | /* it can change loaded_data to point to an internal buffer */ |
063e760a | 3603 | qemu_get_buffer_in_place(f, &loaded_data, xh_len); |
56e93d26 JQ |
3604 | |
3605 | /* decode RLE */ | |
063e760a | 3606 | if (xbzrle_decode_buffer(loaded_data, xh_len, host, |
56e93d26 JQ |
3607 | TARGET_PAGE_SIZE) == -1) { |
3608 | error_report("Failed to load XBZRLE page - decode error!"); | |
3609 | return -1; | |
3610 | } | |
3611 | ||
3612 | return 0; | |
3613 | } | |
3614 | ||
3d0684b2 JQ |
3615 | /** |
3616 | * ram_block_from_stream: read a RAMBlock id from the migration stream | |
3617 | * | |
3618 | * Must be called from within a rcu critical section. | |
3619 | * | |
56e93d26 | 3620 | * Returns a pointer from within the RCU-protected ram_list. |
a7180877 | 3621 | * |
755e8d7c | 3622 | * @mis: the migration incoming state pointer |
3d0684b2 JQ |
3623 | * @f: QEMUFile where to read the data from |
3624 | * @flags: Page flags (mostly to see if it's a continuation of previous block) | |
c01b16ed | 3625 | * @channel: the channel we're using |
a7180877 | 3626 | */ |
755e8d7c | 3627 | static inline RAMBlock *ram_block_from_stream(MigrationIncomingState *mis, |
c01b16ed PX |
3628 | QEMUFile *f, int flags, |
3629 | int channel) | |
56e93d26 | 3630 | { |
c01b16ed | 3631 | RAMBlock *block = mis->last_recv_block[channel]; |
56e93d26 JQ |
3632 | char id[256]; |
3633 | uint8_t len; | |
3634 | ||
3635 | if (flags & RAM_SAVE_FLAG_CONTINUE) { | |
4c4bad48 | 3636 | if (!block) { |
56e93d26 JQ |
3637 | error_report("Ack, bad migration stream!"); |
3638 | return NULL; | |
3639 | } | |
4c4bad48 | 3640 | return block; |
56e93d26 JQ |
3641 | } |
3642 | ||
3643 | len = qemu_get_byte(f); | |
3644 | qemu_get_buffer(f, (uint8_t *)id, len); | |
3645 | id[len] = 0; | |
3646 | ||
e3dd7493 | 3647 | block = qemu_ram_block_by_name(id); |
4c4bad48 HZ |
3648 | if (!block) { |
3649 | error_report("Can't find block %s", id); | |
3650 | return NULL; | |
56e93d26 JQ |
3651 | } |
3652 | ||
fbd162e6 | 3653 | if (ramblock_is_ignored(block)) { |
b895de50 CLG |
3654 | error_report("block %s should not be migrated !", id); |
3655 | return NULL; | |
3656 | } | |
3657 | ||
c01b16ed | 3658 | mis->last_recv_block[channel] = block; |
755e8d7c | 3659 | |
4c4bad48 HZ |
3660 | return block; |
3661 | } | |
3662 | ||
3663 | static inline void *host_from_ram_block_offset(RAMBlock *block, | |
3664 | ram_addr_t offset) | |
3665 | { | |
3666 | if (!offset_in_ramblock(block, offset)) { | |
3667 | return NULL; | |
3668 | } | |
3669 | ||
3670 | return block->host + offset; | |
56e93d26 JQ |
3671 | } |
3672 | ||
6a23f639 DH |
3673 | static void *host_page_from_ram_block_offset(RAMBlock *block, |
3674 | ram_addr_t offset) | |
3675 | { | |
3676 | /* Note: Explicitly no check against offset_in_ramblock(). */ | |
3677 | return (void *)QEMU_ALIGN_DOWN((uintptr_t)(block->host + offset), | |
3678 | block->page_size); | |
3679 | } | |
3680 | ||
3681 | static ram_addr_t host_page_offset_from_ram_block_offset(RAMBlock *block, | |
3682 | ram_addr_t offset) | |
3683 | { | |
3684 | return ((uintptr_t)block->host + offset) & (block->page_size - 1); | |
3685 | } | |
3686 | ||
13af18f2 | 3687 | static inline void *colo_cache_from_block_offset(RAMBlock *block, |
8af66371 | 3688 | ram_addr_t offset, bool record_bitmap) |
13af18f2 ZC |
3689 | { |
3690 | if (!offset_in_ramblock(block, offset)) { | |
3691 | return NULL; | |
3692 | } | |
3693 | if (!block->colo_cache) { | |
3694 | error_report("%s: colo_cache is NULL in block :%s", | |
3695 | __func__, block->idstr); | |
3696 | return NULL; | |
3697 | } | |
7d9acafa ZC |
3698 | |
3699 | /* | |
3700 | * During colo checkpoint, we need bitmap of these migrated pages. | |
3701 | * It help us to decide which pages in ram cache should be flushed | |
3702 | * into VM's RAM later. | |
3703 | */ | |
8af66371 HZ |
3704 | if (record_bitmap && |
3705 | !test_and_set_bit(offset >> TARGET_PAGE_BITS, block->bmap)) { | |
7d9acafa ZC |
3706 | ram_state->migration_dirty_pages++; |
3707 | } | |
13af18f2 ZC |
3708 | return block->colo_cache + offset; |
3709 | } | |
3710 | ||
3d0684b2 JQ |
3711 | /** |
3712 | * ram_handle_compressed: handle the zero page case | |
3713 | * | |
56e93d26 JQ |
3714 | * If a page (or a whole RDMA chunk) has been |
3715 | * determined to be zero, then zap it. | |
3d0684b2 JQ |
3716 | * |
3717 | * @host: host address for the zero page | |
3718 | * @ch: what the page is filled from. We only support zero | |
3719 | * @size: size of the zero page | |
56e93d26 JQ |
3720 | */ |
3721 | void ram_handle_compressed(void *host, uint8_t ch, uint64_t size) | |
3722 | { | |
bad452a7 | 3723 | if (ch != 0 || !buffer_is_zero(host, size)) { |
56e93d26 JQ |
3724 | memset(host, ch, size); |
3725 | } | |
3726 | } | |
3727 | ||
797ca154 XG |
3728 | /* return the size after decompression, or negative value on error */ |
3729 | static int | |
3730 | qemu_uncompress_data(z_stream *stream, uint8_t *dest, size_t dest_len, | |
3731 | const uint8_t *source, size_t source_len) | |
3732 | { | |
3733 | int err; | |
3734 | ||
3735 | err = inflateReset(stream); | |
3736 | if (err != Z_OK) { | |
3737 | return -1; | |
3738 | } | |
3739 | ||
3740 | stream->avail_in = source_len; | |
3741 | stream->next_in = (uint8_t *)source; | |
3742 | stream->avail_out = dest_len; | |
3743 | stream->next_out = dest; | |
3744 | ||
3745 | err = inflate(stream, Z_NO_FLUSH); | |
3746 | if (err != Z_STREAM_END) { | |
3747 | return -1; | |
3748 | } | |
3749 | ||
3750 | return stream->total_out; | |
3751 | } | |
3752 | ||
56e93d26 JQ |
3753 | static void *do_data_decompress(void *opaque) |
3754 | { | |
3755 | DecompressParam *param = opaque; | |
3756 | unsigned long pagesize; | |
33d151f4 | 3757 | uint8_t *des; |
34ab9e97 | 3758 | int len, ret; |
56e93d26 | 3759 | |
33d151f4 | 3760 | qemu_mutex_lock(¶m->mutex); |
90e56fb4 | 3761 | while (!param->quit) { |
33d151f4 LL |
3762 | if (param->des) { |
3763 | des = param->des; | |
3764 | len = param->len; | |
3765 | param->des = 0; | |
3766 | qemu_mutex_unlock(¶m->mutex); | |
3767 | ||
56e93d26 | 3768 | pagesize = TARGET_PAGE_SIZE; |
34ab9e97 XG |
3769 | |
3770 | ret = qemu_uncompress_data(¶m->stream, des, pagesize, | |
3771 | param->compbuf, len); | |
f548222c | 3772 | if (ret < 0 && migrate_get_current()->decompress_error_check) { |
34ab9e97 XG |
3773 | error_report("decompress data failed"); |
3774 | qemu_file_set_error(decomp_file, ret); | |
3775 | } | |
73a8912b | 3776 | |
33d151f4 LL |
3777 | qemu_mutex_lock(&decomp_done_lock); |
3778 | param->done = true; | |
3779 | qemu_cond_signal(&decomp_done_cond); | |
3780 | qemu_mutex_unlock(&decomp_done_lock); | |
3781 | ||
3782 | qemu_mutex_lock(¶m->mutex); | |
3783 | } else { | |
3784 | qemu_cond_wait(¶m->cond, ¶m->mutex); | |
3785 | } | |
56e93d26 | 3786 | } |
33d151f4 | 3787 | qemu_mutex_unlock(¶m->mutex); |
56e93d26 JQ |
3788 | |
3789 | return NULL; | |
3790 | } | |
3791 | ||
34ab9e97 | 3792 | static int wait_for_decompress_done(void) |
5533b2e9 LL |
3793 | { |
3794 | int idx, thread_count; | |
3795 | ||
a7a94d14 | 3796 | if (!migrate_compress()) { |
34ab9e97 | 3797 | return 0; |
5533b2e9 LL |
3798 | } |
3799 | ||
3800 | thread_count = migrate_decompress_threads(); | |
3801 | qemu_mutex_lock(&decomp_done_lock); | |
3802 | for (idx = 0; idx < thread_count; idx++) { | |
3803 | while (!decomp_param[idx].done) { | |
3804 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
3805 | } | |
3806 | } | |
3807 | qemu_mutex_unlock(&decomp_done_lock); | |
34ab9e97 | 3808 | return qemu_file_get_error(decomp_file); |
5533b2e9 LL |
3809 | } |
3810 | ||
f0afa331 | 3811 | static void compress_threads_load_cleanup(void) |
56e93d26 JQ |
3812 | { |
3813 | int i, thread_count; | |
3814 | ||
a7a94d14 | 3815 | if (!migrate_compress()) { |
3416ab5b JQ |
3816 | return; |
3817 | } | |
56e93d26 JQ |
3818 | thread_count = migrate_decompress_threads(); |
3819 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3820 | /* |
3821 | * we use it as a indicator which shows if the thread is | |
3822 | * properly init'd or not | |
3823 | */ | |
3824 | if (!decomp_param[i].compbuf) { | |
3825 | break; | |
3826 | } | |
3827 | ||
56e93d26 | 3828 | qemu_mutex_lock(&decomp_param[i].mutex); |
90e56fb4 | 3829 | decomp_param[i].quit = true; |
56e93d26 JQ |
3830 | qemu_cond_signal(&decomp_param[i].cond); |
3831 | qemu_mutex_unlock(&decomp_param[i].mutex); | |
3832 | } | |
3833 | for (i = 0; i < thread_count; i++) { | |
797ca154 XG |
3834 | if (!decomp_param[i].compbuf) { |
3835 | break; | |
3836 | } | |
3837 | ||
56e93d26 JQ |
3838 | qemu_thread_join(decompress_threads + i); |
3839 | qemu_mutex_destroy(&decomp_param[i].mutex); | |
3840 | qemu_cond_destroy(&decomp_param[i].cond); | |
797ca154 | 3841 | inflateEnd(&decomp_param[i].stream); |
56e93d26 | 3842 | g_free(decomp_param[i].compbuf); |
797ca154 | 3843 | decomp_param[i].compbuf = NULL; |
56e93d26 JQ |
3844 | } |
3845 | g_free(decompress_threads); | |
3846 | g_free(decomp_param); | |
56e93d26 JQ |
3847 | decompress_threads = NULL; |
3848 | decomp_param = NULL; | |
34ab9e97 | 3849 | decomp_file = NULL; |
56e93d26 JQ |
3850 | } |
3851 | ||
34ab9e97 | 3852 | static int compress_threads_load_setup(QEMUFile *f) |
797ca154 XG |
3853 | { |
3854 | int i, thread_count; | |
3855 | ||
a7a94d14 | 3856 | if (!migrate_compress()) { |
797ca154 XG |
3857 | return 0; |
3858 | } | |
3859 | ||
3860 | thread_count = migrate_decompress_threads(); | |
3861 | decompress_threads = g_new0(QemuThread, thread_count); | |
3862 | decomp_param = g_new0(DecompressParam, thread_count); | |
3863 | qemu_mutex_init(&decomp_done_lock); | |
3864 | qemu_cond_init(&decomp_done_cond); | |
34ab9e97 | 3865 | decomp_file = f; |
797ca154 XG |
3866 | for (i = 0; i < thread_count; i++) { |
3867 | if (inflateInit(&decomp_param[i].stream) != Z_OK) { | |
3868 | goto exit; | |
3869 | } | |
3870 | ||
3871 | decomp_param[i].compbuf = g_malloc0(compressBound(TARGET_PAGE_SIZE)); | |
3872 | qemu_mutex_init(&decomp_param[i].mutex); | |
3873 | qemu_cond_init(&decomp_param[i].cond); | |
3874 | decomp_param[i].done = true; | |
3875 | decomp_param[i].quit = false; | |
3876 | qemu_thread_create(decompress_threads + i, "decompress", | |
3877 | do_data_decompress, decomp_param + i, | |
3878 | QEMU_THREAD_JOINABLE); | |
3879 | } | |
3880 | return 0; | |
3881 | exit: | |
3882 | compress_threads_load_cleanup(); | |
3883 | return -1; | |
3884 | } | |
3885 | ||
c1bc6626 | 3886 | static void decompress_data_with_multi_threads(QEMUFile *f, |
56e93d26 JQ |
3887 | void *host, int len) |
3888 | { | |
3889 | int idx, thread_count; | |
3890 | ||
3891 | thread_count = migrate_decompress_threads(); | |
37396950 | 3892 | QEMU_LOCK_GUARD(&decomp_done_lock); |
56e93d26 JQ |
3893 | while (true) { |
3894 | for (idx = 0; idx < thread_count; idx++) { | |
73a8912b | 3895 | if (decomp_param[idx].done) { |
33d151f4 LL |
3896 | decomp_param[idx].done = false; |
3897 | qemu_mutex_lock(&decomp_param[idx].mutex); | |
c1bc6626 | 3898 | qemu_get_buffer(f, decomp_param[idx].compbuf, len); |
56e93d26 JQ |
3899 | decomp_param[idx].des = host; |
3900 | decomp_param[idx].len = len; | |
33d151f4 LL |
3901 | qemu_cond_signal(&decomp_param[idx].cond); |
3902 | qemu_mutex_unlock(&decomp_param[idx].mutex); | |
56e93d26 JQ |
3903 | break; |
3904 | } | |
3905 | } | |
3906 | if (idx < thread_count) { | |
3907 | break; | |
73a8912b LL |
3908 | } else { |
3909 | qemu_cond_wait(&decomp_done_cond, &decomp_done_lock); | |
56e93d26 JQ |
3910 | } |
3911 | } | |
3912 | } | |
3913 | ||
b70cb3b4 RL |
3914 | static void colo_init_ram_state(void) |
3915 | { | |
3916 | ram_state_init(&ram_state); | |
b70cb3b4 RL |
3917 | } |
3918 | ||
13af18f2 ZC |
3919 | /* |
3920 | * colo cache: this is for secondary VM, we cache the whole | |
3921 | * memory of the secondary VM, it is need to hold the global lock | |
3922 | * to call this helper. | |
3923 | */ | |
3924 | int colo_init_ram_cache(void) | |
3925 | { | |
3926 | RAMBlock *block; | |
3927 | ||
44901b5a PB |
3928 | WITH_RCU_READ_LOCK_GUARD() { |
3929 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3930 | block->colo_cache = qemu_anon_ram_alloc(block->used_length, | |
8dbe22c6 | 3931 | NULL, false, false); |
44901b5a PB |
3932 | if (!block->colo_cache) { |
3933 | error_report("%s: Can't alloc memory for COLO cache of block %s," | |
3934 | "size 0x" RAM_ADDR_FMT, __func__, block->idstr, | |
3935 | block->used_length); | |
3936 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3937 | if (block->colo_cache) { | |
3938 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
3939 | block->colo_cache = NULL; | |
3940 | } | |
89ac5a1d | 3941 | } |
44901b5a | 3942 | return -errno; |
89ac5a1d | 3943 | } |
e5fdf920 LS |
3944 | if (!machine_dump_guest_core(current_machine)) { |
3945 | qemu_madvise(block->colo_cache, block->used_length, | |
3946 | QEMU_MADV_DONTDUMP); | |
3947 | } | |
13af18f2 | 3948 | } |
13af18f2 | 3949 | } |
44901b5a | 3950 | |
7d9acafa ZC |
3951 | /* |
3952 | * Record the dirty pages that sent by PVM, we use this dirty bitmap together | |
3953 | * with to decide which page in cache should be flushed into SVM's RAM. Here | |
3954 | * we use the same name 'ram_bitmap' as for migration. | |
3955 | */ | |
3956 | if (ram_bytes_total()) { | |
3957 | RAMBlock *block; | |
3958 | ||
fbd162e6 | 3959 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa | 3960 | unsigned long pages = block->max_length >> TARGET_PAGE_BITS; |
7d9acafa | 3961 | block->bmap = bitmap_new(pages); |
7d9acafa ZC |
3962 | } |
3963 | } | |
7d9acafa | 3964 | |
b70cb3b4 | 3965 | colo_init_ram_state(); |
13af18f2 | 3966 | return 0; |
13af18f2 ZC |
3967 | } |
3968 | ||
0393031a HZ |
3969 | /* TODO: duplicated with ram_init_bitmaps */ |
3970 | void colo_incoming_start_dirty_log(void) | |
3971 | { | |
3972 | RAMBlock *block = NULL; | |
3973 | /* For memory_global_dirty_log_start below. */ | |
3974 | qemu_mutex_lock_iothread(); | |
3975 | qemu_mutex_lock_ramlist(); | |
3976 | ||
3977 | memory_global_dirty_log_sync(); | |
3978 | WITH_RCU_READ_LOCK_GUARD() { | |
3979 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
3980 | ramblock_sync_dirty_bitmap(ram_state, block); | |
3981 | /* Discard this dirty bitmap record */ | |
3982 | bitmap_zero(block->bmap, block->max_length >> TARGET_PAGE_BITS); | |
3983 | } | |
63b41db4 | 3984 | memory_global_dirty_log_start(GLOBAL_DIRTY_MIGRATION); |
0393031a HZ |
3985 | } |
3986 | ram_state->migration_dirty_pages = 0; | |
3987 | qemu_mutex_unlock_ramlist(); | |
3988 | qemu_mutex_unlock_iothread(); | |
3989 | } | |
3990 | ||
13af18f2 ZC |
3991 | /* It is need to hold the global lock to call this helper */ |
3992 | void colo_release_ram_cache(void) | |
3993 | { | |
3994 | RAMBlock *block; | |
3995 | ||
63b41db4 | 3996 | memory_global_dirty_log_stop(GLOBAL_DIRTY_MIGRATION); |
fbd162e6 | 3997 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
7d9acafa ZC |
3998 | g_free(block->bmap); |
3999 | block->bmap = NULL; | |
4000 | } | |
4001 | ||
89ac5a1d DDAG |
4002 | WITH_RCU_READ_LOCK_GUARD() { |
4003 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
4004 | if (block->colo_cache) { | |
4005 | qemu_anon_ram_free(block->colo_cache, block->used_length); | |
4006 | block->colo_cache = NULL; | |
4007 | } | |
13af18f2 ZC |
4008 | } |
4009 | } | |
0393031a | 4010 | ram_state_cleanup(&ram_state); |
13af18f2 ZC |
4011 | } |
4012 | ||
f265e0e4 JQ |
4013 | /** |
4014 | * ram_load_setup: Setup RAM for migration incoming side | |
4015 | * | |
4016 | * Returns zero to indicate success and negative for error | |
4017 | * | |
4018 | * @f: QEMUFile where to receive the data | |
4019 | * @opaque: RAMState pointer | |
4020 | */ | |
4021 | static int ram_load_setup(QEMUFile *f, void *opaque) | |
4022 | { | |
34ab9e97 | 4023 | if (compress_threads_load_setup(f)) { |
797ca154 XG |
4024 | return -1; |
4025 | } | |
4026 | ||
f265e0e4 | 4027 | xbzrle_load_setup(); |
f9494614 | 4028 | ramblock_recv_map_init(); |
13af18f2 | 4029 | |
f265e0e4 JQ |
4030 | return 0; |
4031 | } | |
4032 | ||
4033 | static int ram_load_cleanup(void *opaque) | |
4034 | { | |
f9494614 | 4035 | RAMBlock *rb; |
56eb90af | 4036 | |
fbd162e6 | 4037 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
bd108a44 | 4038 | qemu_ram_block_writeback(rb); |
56eb90af JH |
4039 | } |
4040 | ||
f265e0e4 | 4041 | xbzrle_load_cleanup(); |
f0afa331 | 4042 | compress_threads_load_cleanup(); |
f9494614 | 4043 | |
fbd162e6 | 4044 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
f9494614 AP |
4045 | g_free(rb->receivedmap); |
4046 | rb->receivedmap = NULL; | |
4047 | } | |
13af18f2 | 4048 | |
f265e0e4 JQ |
4049 | return 0; |
4050 | } | |
4051 | ||
3d0684b2 JQ |
4052 | /** |
4053 | * ram_postcopy_incoming_init: allocate postcopy data structures | |
4054 | * | |
4055 | * Returns 0 for success and negative if there was one error | |
4056 | * | |
4057 | * @mis: current migration incoming state | |
4058 | * | |
4059 | * Allocate data structures etc needed by incoming migration with | |
4060 | * postcopy-ram. postcopy-ram's similarly names | |
4061 | * postcopy_ram_incoming_init does the work. | |
1caddf8a DDAG |
4062 | */ |
4063 | int ram_postcopy_incoming_init(MigrationIncomingState *mis) | |
4064 | { | |
c136180c | 4065 | return postcopy_ram_incoming_init(mis); |
1caddf8a DDAG |
4066 | } |
4067 | ||
3d0684b2 JQ |
4068 | /** |
4069 | * ram_load_postcopy: load a page in postcopy case | |
4070 | * | |
4071 | * Returns 0 for success or -errno in case of error | |
4072 | * | |
a7180877 DDAG |
4073 | * Called in postcopy mode by ram_load(). |
4074 | * rcu_read_lock is taken prior to this being called. | |
3d0684b2 JQ |
4075 | * |
4076 | * @f: QEMUFile where to send the data | |
36f62f11 | 4077 | * @channel: the channel to use for loading |
a7180877 | 4078 | */ |
36f62f11 | 4079 | int ram_load_postcopy(QEMUFile *f, int channel) |
a7180877 DDAG |
4080 | { |
4081 | int flags = 0, ret = 0; | |
4082 | bool place_needed = false; | |
1aa83678 | 4083 | bool matches_target_page_size = false; |
a7180877 | 4084 | MigrationIncomingState *mis = migration_incoming_get_current(); |
36f62f11 | 4085 | PostcopyTmpPage *tmp_page = &mis->postcopy_tmp_pages[channel]; |
a7180877 DDAG |
4086 | |
4087 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { | |
4088 | ram_addr_t addr; | |
a7180877 DDAG |
4089 | void *page_buffer = NULL; |
4090 | void *place_source = NULL; | |
df9ff5e1 | 4091 | RAMBlock *block = NULL; |
a7180877 | 4092 | uint8_t ch; |
644acf99 | 4093 | int len; |
a7180877 DDAG |
4094 | |
4095 | addr = qemu_get_be64(f); | |
7a9ddfbf PX |
4096 | |
4097 | /* | |
4098 | * If qemu file error, we should stop here, and then "addr" | |
4099 | * may be invalid | |
4100 | */ | |
4101 | ret = qemu_file_get_error(f); | |
4102 | if (ret) { | |
4103 | break; | |
4104 | } | |
4105 | ||
a7180877 DDAG |
4106 | flags = addr & ~TARGET_PAGE_MASK; |
4107 | addr &= TARGET_PAGE_MASK; | |
4108 | ||
36f62f11 | 4109 | trace_ram_load_postcopy_loop(channel, (uint64_t)addr, flags); |
644acf99 WY |
4110 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
4111 | RAM_SAVE_FLAG_COMPRESS_PAGE)) { | |
c01b16ed | 4112 | block = ram_block_from_stream(mis, f, flags, channel); |
6a23f639 DH |
4113 | if (!block) { |
4114 | ret = -EINVAL; | |
4115 | break; | |
4116 | } | |
4c4bad48 | 4117 | |
898ba906 DH |
4118 | /* |
4119 | * Relying on used_length is racy and can result in false positives. | |
4120 | * We might place pages beyond used_length in case RAM was shrunk | |
4121 | * while in postcopy, which is fine - trying to place via | |
4122 | * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault. | |
4123 | */ | |
4124 | if (!block->host || addr >= block->postcopy_length) { | |
a7180877 DDAG |
4125 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); |
4126 | ret = -EINVAL; | |
4127 | break; | |
4128 | } | |
77dadc3f | 4129 | tmp_page->target_pages++; |
1aa83678 | 4130 | matches_target_page_size = block->page_size == TARGET_PAGE_SIZE; |
a7180877 | 4131 | /* |
28abd200 DDAG |
4132 | * Postcopy requires that we place whole host pages atomically; |
4133 | * these may be huge pages for RAMBlocks that are backed by | |
4134 | * hugetlbfs. | |
a7180877 DDAG |
4135 | * To make it atomic, the data is read into a temporary page |
4136 | * that's moved into place later. | |
4137 | * The migration protocol uses, possibly smaller, target-pages | |
4138 | * however the source ensures it always sends all the components | |
91ba442f | 4139 | * of a host page in one chunk. |
a7180877 | 4140 | */ |
77dadc3f | 4141 | page_buffer = tmp_page->tmp_huge_page + |
6a23f639 DH |
4142 | host_page_offset_from_ram_block_offset(block, addr); |
4143 | /* If all TP are zero then we can optimise the place */ | |
77dadc3f PX |
4144 | if (tmp_page->target_pages == 1) { |
4145 | tmp_page->host_addr = | |
4146 | host_page_from_ram_block_offset(block, addr); | |
4147 | } else if (tmp_page->host_addr != | |
4148 | host_page_from_ram_block_offset(block, addr)) { | |
c53b7ddc | 4149 | /* not the 1st TP within the HP */ |
36f62f11 | 4150 | error_report("Non-same host page detected on channel %d: " |
cfc7dc8a PX |
4151 | "Target host page %p, received host page %p " |
4152 | "(rb %s offset 0x"RAM_ADDR_FMT" target_pages %d)", | |
36f62f11 | 4153 | channel, tmp_page->host_addr, |
cfc7dc8a PX |
4154 | host_page_from_ram_block_offset(block, addr), |
4155 | block->idstr, addr, tmp_page->target_pages); | |
6a23f639 DH |
4156 | ret = -EINVAL; |
4157 | break; | |
a7180877 DDAG |
4158 | } |
4159 | ||
4160 | /* | |
4161 | * If it's the last part of a host page then we place the host | |
4162 | * page | |
4163 | */ | |
77dadc3f PX |
4164 | if (tmp_page->target_pages == |
4165 | (block->page_size / TARGET_PAGE_SIZE)) { | |
4cbb3c63 | 4166 | place_needed = true; |
4cbb3c63 | 4167 | } |
77dadc3f | 4168 | place_source = tmp_page->tmp_huge_page; |
a7180877 DDAG |
4169 | } |
4170 | ||
4171 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { | |
bb890ed5 | 4172 | case RAM_SAVE_FLAG_ZERO: |
a7180877 | 4173 | ch = qemu_get_byte(f); |
2e36bc1b WY |
4174 | /* |
4175 | * Can skip to set page_buffer when | |
4176 | * this is a zero page and (block->page_size == TARGET_PAGE_SIZE). | |
4177 | */ | |
4178 | if (ch || !matches_target_page_size) { | |
4179 | memset(page_buffer, ch, TARGET_PAGE_SIZE); | |
4180 | } | |
a7180877 | 4181 | if (ch) { |
77dadc3f | 4182 | tmp_page->all_zero = false; |
a7180877 DDAG |
4183 | } |
4184 | break; | |
4185 | ||
4186 | case RAM_SAVE_FLAG_PAGE: | |
77dadc3f | 4187 | tmp_page->all_zero = false; |
1aa83678 PX |
4188 | if (!matches_target_page_size) { |
4189 | /* For huge pages, we always use temporary buffer */ | |
a7180877 DDAG |
4190 | qemu_get_buffer(f, page_buffer, TARGET_PAGE_SIZE); |
4191 | } else { | |
1aa83678 PX |
4192 | /* |
4193 | * For small pages that matches target page size, we | |
4194 | * avoid the qemu_file copy. Instead we directly use | |
4195 | * the buffer of QEMUFile to place the page. Note: we | |
4196 | * cannot do any QEMUFile operation before using that | |
4197 | * buffer to make sure the buffer is valid when | |
4198 | * placing the page. | |
a7180877 DDAG |
4199 | */ |
4200 | qemu_get_buffer_in_place(f, (uint8_t **)&place_source, | |
4201 | TARGET_PAGE_SIZE); | |
4202 | } | |
4203 | break; | |
644acf99 | 4204 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
77dadc3f | 4205 | tmp_page->all_zero = false; |
644acf99 WY |
4206 | len = qemu_get_be32(f); |
4207 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4208 | error_report("Invalid compressed data length: %d", len); | |
4209 | ret = -EINVAL; | |
4210 | break; | |
4211 | } | |
4212 | decompress_data_with_multi_threads(f, page_buffer, len); | |
4213 | break; | |
294e5a40 JQ |
4214 | case RAM_SAVE_FLAG_MULTIFD_FLUSH: |
4215 | multifd_recv_sync_main(); | |
4216 | break; | |
a7180877 DDAG |
4217 | case RAM_SAVE_FLAG_EOS: |
4218 | /* normal exit */ | |
b05292c2 JQ |
4219 | if (migrate_multifd_flush_after_each_section()) { |
4220 | multifd_recv_sync_main(); | |
4221 | } | |
a7180877 DDAG |
4222 | break; |
4223 | default: | |
29fccade | 4224 | error_report("Unknown combination of migration flags: 0x%x" |
a7180877 DDAG |
4225 | " (postcopy mode)", flags); |
4226 | ret = -EINVAL; | |
7a9ddfbf PX |
4227 | break; |
4228 | } | |
4229 | ||
644acf99 WY |
4230 | /* Got the whole host page, wait for decompress before placing. */ |
4231 | if (place_needed) { | |
4232 | ret |= wait_for_decompress_done(); | |
4233 | } | |
4234 | ||
7a9ddfbf PX |
4235 | /* Detect for any possible file errors */ |
4236 | if (!ret && qemu_file_get_error(f)) { | |
4237 | ret = qemu_file_get_error(f); | |
a7180877 DDAG |
4238 | } |
4239 | ||
7a9ddfbf | 4240 | if (!ret && place_needed) { |
77dadc3f PX |
4241 | if (tmp_page->all_zero) { |
4242 | ret = postcopy_place_page_zero(mis, tmp_page->host_addr, block); | |
a7180877 | 4243 | } else { |
77dadc3f PX |
4244 | ret = postcopy_place_page(mis, tmp_page->host_addr, |
4245 | place_source, block); | |
a7180877 | 4246 | } |
ddf35bdf | 4247 | place_needed = false; |
77dadc3f | 4248 | postcopy_temp_page_reset(tmp_page); |
a7180877 | 4249 | } |
a7180877 DDAG |
4250 | } |
4251 | ||
4252 | return ret; | |
4253 | } | |
4254 | ||
acab30b8 DHB |
4255 | static bool postcopy_is_running(void) |
4256 | { | |
4257 | PostcopyState ps = postcopy_state_get(); | |
4258 | return ps >= POSTCOPY_INCOMING_LISTENING && ps < POSTCOPY_INCOMING_END; | |
4259 | } | |
4260 | ||
e6f4aa18 ZC |
4261 | /* |
4262 | * Flush content of RAM cache into SVM's memory. | |
4263 | * Only flush the pages that be dirtied by PVM or SVM or both. | |
4264 | */ | |
24fa16f8 | 4265 | void colo_flush_ram_cache(void) |
e6f4aa18 ZC |
4266 | { |
4267 | RAMBlock *block = NULL; | |
4268 | void *dst_host; | |
4269 | void *src_host; | |
4270 | unsigned long offset = 0; | |
4271 | ||
d1955d22 | 4272 | memory_global_dirty_log_sync(); |
89ac5a1d DDAG |
4273 | WITH_RCU_READ_LOCK_GUARD() { |
4274 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { | |
4275 | ramblock_sync_dirty_bitmap(ram_state, block); | |
4276 | } | |
d1955d22 | 4277 | } |
d1955d22 | 4278 | |
e6f4aa18 | 4279 | trace_colo_flush_ram_cache_begin(ram_state->migration_dirty_pages); |
89ac5a1d DDAG |
4280 | WITH_RCU_READ_LOCK_GUARD() { |
4281 | block = QLIST_FIRST_RCU(&ram_list.blocks); | |
e6f4aa18 | 4282 | |
89ac5a1d | 4283 | while (block) { |
a6a83cef | 4284 | unsigned long num = 0; |
e6f4aa18 | 4285 | |
a6a83cef | 4286 | offset = colo_bitmap_find_dirty(ram_state, block, offset, &num); |
542147f4 DH |
4287 | if (!offset_in_ramblock(block, |
4288 | ((ram_addr_t)offset) << TARGET_PAGE_BITS)) { | |
89ac5a1d | 4289 | offset = 0; |
a6a83cef | 4290 | num = 0; |
89ac5a1d DDAG |
4291 | block = QLIST_NEXT_RCU(block, next); |
4292 | } else { | |
a6a83cef RL |
4293 | unsigned long i = 0; |
4294 | ||
4295 | for (i = 0; i < num; i++) { | |
4296 | migration_bitmap_clear_dirty(ram_state, block, offset + i); | |
4297 | } | |
8bba004c AR |
4298 | dst_host = block->host |
4299 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
4300 | src_host = block->colo_cache | |
4301 | + (((ram_addr_t)offset) << TARGET_PAGE_BITS); | |
a6a83cef RL |
4302 | memcpy(dst_host, src_host, TARGET_PAGE_SIZE * num); |
4303 | offset += num; | |
89ac5a1d | 4304 | } |
e6f4aa18 ZC |
4305 | } |
4306 | } | |
e6f4aa18 ZC |
4307 | trace_colo_flush_ram_cache_end(); |
4308 | } | |
4309 | ||
10da4a36 WY |
4310 | /** |
4311 | * ram_load_precopy: load pages in precopy case | |
4312 | * | |
4313 | * Returns 0 for success or -errno in case of error | |
4314 | * | |
4315 | * Called in precopy mode by ram_load(). | |
4316 | * rcu_read_lock is taken prior to this being called. | |
4317 | * | |
4318 | * @f: QEMUFile where to send the data | |
4319 | */ | |
4320 | static int ram_load_precopy(QEMUFile *f) | |
56e93d26 | 4321 | { |
755e8d7c | 4322 | MigrationIncomingState *mis = migration_incoming_get_current(); |
e65cec5e | 4323 | int flags = 0, ret = 0, invalid_flags = 0, len = 0, i = 0; |
ef08fb38 | 4324 | /* ADVISE is earlier, it shows the source has the postcopy capability on */ |
80fe315c | 4325 | bool postcopy_advised = migration_incoming_postcopy_advised(); |
a7a94d14 | 4326 | if (!migrate_compress()) { |
edc60127 JQ |
4327 | invalid_flags |= RAM_SAVE_FLAG_COMPRESS_PAGE; |
4328 | } | |
a7180877 | 4329 | |
10da4a36 | 4330 | while (!ret && !(flags & RAM_SAVE_FLAG_EOS)) { |
56e93d26 | 4331 | ram_addr_t addr, total_ram_bytes; |
0393031a | 4332 | void *host = NULL, *host_bak = NULL; |
56e93d26 JQ |
4333 | uint8_t ch; |
4334 | ||
e65cec5e YK |
4335 | /* |
4336 | * Yield periodically to let main loop run, but an iteration of | |
4337 | * the main loop is expensive, so do it each some iterations | |
4338 | */ | |
4339 | if ((i & 32767) == 0 && qemu_in_coroutine()) { | |
4340 | aio_co_schedule(qemu_get_current_aio_context(), | |
4341 | qemu_coroutine_self()); | |
4342 | qemu_coroutine_yield(); | |
4343 | } | |
4344 | i++; | |
4345 | ||
56e93d26 JQ |
4346 | addr = qemu_get_be64(f); |
4347 | flags = addr & ~TARGET_PAGE_MASK; | |
4348 | addr &= TARGET_PAGE_MASK; | |
4349 | ||
edc60127 JQ |
4350 | if (flags & invalid_flags) { |
4351 | if (flags & invalid_flags & RAM_SAVE_FLAG_COMPRESS_PAGE) { | |
4352 | error_report("Received an unexpected compressed page"); | |
4353 | } | |
4354 | ||
4355 | ret = -EINVAL; | |
4356 | break; | |
4357 | } | |
4358 | ||
bb890ed5 | 4359 | if (flags & (RAM_SAVE_FLAG_ZERO | RAM_SAVE_FLAG_PAGE | |
a776aa15 | 4360 | RAM_SAVE_FLAG_COMPRESS_PAGE | RAM_SAVE_FLAG_XBZRLE)) { |
c01b16ed PX |
4361 | RAMBlock *block = ram_block_from_stream(mis, f, flags, |
4362 | RAM_CHANNEL_PRECOPY); | |
4c4bad48 | 4363 | |
0393031a | 4364 | host = host_from_ram_block_offset(block, addr); |
13af18f2 | 4365 | /* |
0393031a HZ |
4366 | * After going into COLO stage, we should not load the page |
4367 | * into SVM's memory directly, we put them into colo_cache firstly. | |
4368 | * NOTE: We need to keep a copy of SVM's ram in colo_cache. | |
4369 | * Previously, we copied all these memory in preparing stage of COLO | |
4370 | * while we need to stop VM, which is a time-consuming process. | |
4371 | * Here we optimize it by a trick, back-up every page while in | |
4372 | * migration process while COLO is enabled, though it affects the | |
4373 | * speed of the migration, but it obviously reduce the downtime of | |
4374 | * back-up all SVM'S memory in COLO preparing stage. | |
13af18f2 | 4375 | */ |
0393031a HZ |
4376 | if (migration_incoming_colo_enabled()) { |
4377 | if (migration_incoming_in_colo_state()) { | |
4378 | /* In COLO stage, put all pages into cache temporarily */ | |
8af66371 | 4379 | host = colo_cache_from_block_offset(block, addr, true); |
0393031a HZ |
4380 | } else { |
4381 | /* | |
4382 | * In migration stage but before COLO stage, | |
4383 | * Put all pages into both cache and SVM's memory. | |
4384 | */ | |
8af66371 | 4385 | host_bak = colo_cache_from_block_offset(block, addr, false); |
0393031a | 4386 | } |
13af18f2 | 4387 | } |
a776aa15 DDAG |
4388 | if (!host) { |
4389 | error_report("Illegal RAM offset " RAM_ADDR_FMT, addr); | |
4390 | ret = -EINVAL; | |
4391 | break; | |
4392 | } | |
13af18f2 ZC |
4393 | if (!migration_incoming_in_colo_state()) { |
4394 | ramblock_recv_bitmap_set(block, host); | |
4395 | } | |
4396 | ||
1db9d8e5 | 4397 | trace_ram_load_loop(block->idstr, (uint64_t)addr, flags, host); |
a776aa15 DDAG |
4398 | } |
4399 | ||
56e93d26 JQ |
4400 | switch (flags & ~RAM_SAVE_FLAG_CONTINUE) { |
4401 | case RAM_SAVE_FLAG_MEM_SIZE: | |
4402 | /* Synchronize RAM block list */ | |
4403 | total_ram_bytes = addr; | |
4404 | while (!ret && total_ram_bytes) { | |
4405 | RAMBlock *block; | |
56e93d26 JQ |
4406 | char id[256]; |
4407 | ram_addr_t length; | |
4408 | ||
4409 | len = qemu_get_byte(f); | |
4410 | qemu_get_buffer(f, (uint8_t *)id, len); | |
4411 | id[len] = 0; | |
4412 | length = qemu_get_be64(f); | |
4413 | ||
e3dd7493 | 4414 | block = qemu_ram_block_by_name(id); |
b895de50 CLG |
4415 | if (block && !qemu_ram_is_migratable(block)) { |
4416 | error_report("block %s should not be migrated !", id); | |
4417 | ret = -EINVAL; | |
4418 | } else if (block) { | |
e3dd7493 DDAG |
4419 | if (length != block->used_length) { |
4420 | Error *local_err = NULL; | |
56e93d26 | 4421 | |
fa53a0e5 | 4422 | ret = qemu_ram_resize(block, length, |
e3dd7493 DDAG |
4423 | &local_err); |
4424 | if (local_err) { | |
4425 | error_report_err(local_err); | |
56e93d26 | 4426 | } |
56e93d26 | 4427 | } |
ef08fb38 | 4428 | /* For postcopy we need to check hugepage sizes match */ |
e846b746 | 4429 | if (postcopy_advised && migrate_postcopy_ram() && |
ef08fb38 DDAG |
4430 | block->page_size != qemu_host_page_size) { |
4431 | uint64_t remote_page_size = qemu_get_be64(f); | |
4432 | if (remote_page_size != block->page_size) { | |
4433 | error_report("Mismatched RAM page size %s " | |
4434 | "(local) %zd != %" PRId64, | |
4435 | id, block->page_size, | |
4436 | remote_page_size); | |
4437 | ret = -EINVAL; | |
4438 | } | |
4439 | } | |
fbd162e6 YK |
4440 | if (migrate_ignore_shared()) { |
4441 | hwaddr addr = qemu_get_be64(f); | |
fbd162e6 YK |
4442 | if (ramblock_is_ignored(block) && |
4443 | block->mr->addr != addr) { | |
4444 | error_report("Mismatched GPAs for block %s " | |
4445 | "%" PRId64 "!= %" PRId64, | |
4446 | id, (uint64_t)addr, | |
4447 | (uint64_t)block->mr->addr); | |
4448 | ret = -EINVAL; | |
4449 | } | |
4450 | } | |
e3dd7493 DDAG |
4451 | ram_control_load_hook(f, RAM_CONTROL_BLOCK_REG, |
4452 | block->idstr); | |
4453 | } else { | |
56e93d26 JQ |
4454 | error_report("Unknown ramblock \"%s\", cannot " |
4455 | "accept migration", id); | |
4456 | ret = -EINVAL; | |
4457 | } | |
4458 | ||
4459 | total_ram_bytes -= length; | |
4460 | } | |
4461 | break; | |
a776aa15 | 4462 | |
bb890ed5 | 4463 | case RAM_SAVE_FLAG_ZERO: |
56e93d26 JQ |
4464 | ch = qemu_get_byte(f); |
4465 | ram_handle_compressed(host, ch, TARGET_PAGE_SIZE); | |
4466 | break; | |
a776aa15 | 4467 | |
56e93d26 | 4468 | case RAM_SAVE_FLAG_PAGE: |
56e93d26 JQ |
4469 | qemu_get_buffer(f, host, TARGET_PAGE_SIZE); |
4470 | break; | |
56e93d26 | 4471 | |
a776aa15 | 4472 | case RAM_SAVE_FLAG_COMPRESS_PAGE: |
56e93d26 JQ |
4473 | len = qemu_get_be32(f); |
4474 | if (len < 0 || len > compressBound(TARGET_PAGE_SIZE)) { | |
4475 | error_report("Invalid compressed data length: %d", len); | |
4476 | ret = -EINVAL; | |
4477 | break; | |
4478 | } | |
c1bc6626 | 4479 | decompress_data_with_multi_threads(f, host, len); |
56e93d26 | 4480 | break; |
a776aa15 | 4481 | |
56e93d26 | 4482 | case RAM_SAVE_FLAG_XBZRLE: |
56e93d26 JQ |
4483 | if (load_xbzrle(f, addr, host) < 0) { |
4484 | error_report("Failed to decompress XBZRLE page at " | |
4485 | RAM_ADDR_FMT, addr); | |
4486 | ret = -EINVAL; | |
4487 | break; | |
4488 | } | |
4489 | break; | |
294e5a40 JQ |
4490 | case RAM_SAVE_FLAG_MULTIFD_FLUSH: |
4491 | multifd_recv_sync_main(); | |
4492 | break; | |
56e93d26 JQ |
4493 | case RAM_SAVE_FLAG_EOS: |
4494 | /* normal exit */ | |
b05292c2 JQ |
4495 | if (migrate_multifd_flush_after_each_section()) { |
4496 | multifd_recv_sync_main(); | |
4497 | } | |
56e93d26 | 4498 | break; |
5f1e7540 JQ |
4499 | case RAM_SAVE_FLAG_HOOK: |
4500 | ram_control_load_hook(f, RAM_CONTROL_HOOK, NULL); | |
4501 | break; | |
56e93d26 | 4502 | default: |
5f1e7540 JQ |
4503 | error_report("Unknown combination of migration flags: 0x%x", flags); |
4504 | ret = -EINVAL; | |
56e93d26 JQ |
4505 | } |
4506 | if (!ret) { | |
4507 | ret = qemu_file_get_error(f); | |
4508 | } | |
0393031a HZ |
4509 | if (!ret && host_bak) { |
4510 | memcpy(host_bak, host, TARGET_PAGE_SIZE); | |
4511 | } | |
56e93d26 JQ |
4512 | } |
4513 | ||
ca1a6b70 | 4514 | ret |= wait_for_decompress_done(); |
10da4a36 WY |
4515 | return ret; |
4516 | } | |
4517 | ||
4518 | static int ram_load(QEMUFile *f, void *opaque, int version_id) | |
4519 | { | |
4520 | int ret = 0; | |
4521 | static uint64_t seq_iter; | |
4522 | /* | |
4523 | * If system is running in postcopy mode, page inserts to host memory must | |
4524 | * be atomic | |
4525 | */ | |
4526 | bool postcopy_running = postcopy_is_running(); | |
4527 | ||
4528 | seq_iter++; | |
4529 | ||
4530 | if (version_id != 4) { | |
4531 | return -EINVAL; | |
4532 | } | |
4533 | ||
4534 | /* | |
4535 | * This RCU critical section can be very long running. | |
4536 | * When RCU reclaims in the code start to become numerous, | |
4537 | * it will be necessary to reduce the granularity of this | |
4538 | * critical section. | |
4539 | */ | |
89ac5a1d DDAG |
4540 | WITH_RCU_READ_LOCK_GUARD() { |
4541 | if (postcopy_running) { | |
36f62f11 PX |
4542 | /* |
4543 | * Note! Here RAM_CHANNEL_PRECOPY is the precopy channel of | |
4544 | * postcopy migration, we have another RAM_CHANNEL_POSTCOPY to | |
4545 | * service fast page faults. | |
4546 | */ | |
4547 | ret = ram_load_postcopy(f, RAM_CHANNEL_PRECOPY); | |
89ac5a1d DDAG |
4548 | } else { |
4549 | ret = ram_load_precopy(f); | |
4550 | } | |
10da4a36 | 4551 | } |
55c4446b | 4552 | trace_ram_load_complete(ret, seq_iter); |
e6f4aa18 | 4553 | |
56e93d26 JQ |
4554 | return ret; |
4555 | } | |
4556 | ||
c6467627 VSO |
4557 | static bool ram_has_postcopy(void *opaque) |
4558 | { | |
469dd51b | 4559 | RAMBlock *rb; |
fbd162e6 | 4560 | RAMBLOCK_FOREACH_NOT_IGNORED(rb) { |
469dd51b JH |
4561 | if (ramblock_is_pmem(rb)) { |
4562 | info_report("Block: %s, host: %p is a nvdimm memory, postcopy" | |
4563 | "is not supported now!", rb->idstr, rb->host); | |
4564 | return false; | |
4565 | } | |
4566 | } | |
4567 | ||
c6467627 VSO |
4568 | return migrate_postcopy_ram(); |
4569 | } | |
4570 | ||
edd090c7 PX |
4571 | /* Sync all the dirty bitmap with destination VM. */ |
4572 | static int ram_dirty_bitmap_sync_all(MigrationState *s, RAMState *rs) | |
4573 | { | |
4574 | RAMBlock *block; | |
4575 | QEMUFile *file = s->to_dst_file; | |
4576 | int ramblock_count = 0; | |
4577 | ||
4578 | trace_ram_dirty_bitmap_sync_start(); | |
4579 | ||
fbd162e6 | 4580 | RAMBLOCK_FOREACH_NOT_IGNORED(block) { |
edd090c7 PX |
4581 | qemu_savevm_send_recv_bitmap(file, block->idstr); |
4582 | trace_ram_dirty_bitmap_request(block->idstr); | |
4583 | ramblock_count++; | |
4584 | } | |
4585 | ||
4586 | trace_ram_dirty_bitmap_sync_wait(); | |
4587 | ||
4588 | /* Wait until all the ramblocks' dirty bitmap synced */ | |
4589 | while (ramblock_count--) { | |
4590 | qemu_sem_wait(&s->rp_state.rp_sem); | |
4591 | } | |
4592 | ||
4593 | trace_ram_dirty_bitmap_sync_complete(); | |
4594 | ||
4595 | return 0; | |
4596 | } | |
4597 | ||
4598 | static void ram_dirty_bitmap_reload_notify(MigrationState *s) | |
4599 | { | |
4600 | qemu_sem_post(&s->rp_state.rp_sem); | |
4601 | } | |
4602 | ||
a335debb PX |
4603 | /* |
4604 | * Read the received bitmap, revert it as the initial dirty bitmap. | |
4605 | * This is only used when the postcopy migration is paused but wants | |
4606 | * to resume from a middle point. | |
4607 | */ | |
4608 | int ram_dirty_bitmap_reload(MigrationState *s, RAMBlock *block) | |
4609 | { | |
4610 | int ret = -EINVAL; | |
43044ac0 | 4611 | /* from_dst_file is always valid because we're within rp_thread */ |
a335debb PX |
4612 | QEMUFile *file = s->rp_state.from_dst_file; |
4613 | unsigned long *le_bitmap, nbits = block->used_length >> TARGET_PAGE_BITS; | |
a725ef9f | 4614 | uint64_t local_size = DIV_ROUND_UP(nbits, 8); |
a335debb PX |
4615 | uint64_t size, end_mark; |
4616 | ||
4617 | trace_ram_dirty_bitmap_reload_begin(block->idstr); | |
4618 | ||
4619 | if (s->state != MIGRATION_STATUS_POSTCOPY_RECOVER) { | |
4620 | error_report("%s: incorrect state %s", __func__, | |
4621 | MigrationStatus_str(s->state)); | |
4622 | return -EINVAL; | |
4623 | } | |
4624 | ||
4625 | /* | |
4626 | * Note: see comments in ramblock_recv_bitmap_send() on why we | |
3a4452d8 | 4627 | * need the endianness conversion, and the paddings. |
a335debb PX |
4628 | */ |
4629 | local_size = ROUND_UP(local_size, 8); | |
4630 | ||
4631 | /* Add paddings */ | |
4632 | le_bitmap = bitmap_new(nbits + BITS_PER_LONG); | |
4633 | ||
4634 | size = qemu_get_be64(file); | |
4635 | ||
4636 | /* The size of the bitmap should match with our ramblock */ | |
4637 | if (size != local_size) { | |
4638 | error_report("%s: ramblock '%s' bitmap size mismatch " | |
4639 | "(0x%"PRIx64" != 0x%"PRIx64")", __func__, | |
4640 | block->idstr, size, local_size); | |
4641 | ret = -EINVAL; | |
4642 | goto out; | |
4643 | } | |
4644 | ||
4645 | size = qemu_get_buffer(file, (uint8_t *)le_bitmap, local_size); | |
4646 | end_mark = qemu_get_be64(file); | |
4647 | ||
4648 | ret = qemu_file_get_error(file); | |
4649 | if (ret || size != local_size) { | |
4650 | error_report("%s: read bitmap failed for ramblock '%s': %d" | |
4651 | " (size 0x%"PRIx64", got: 0x%"PRIx64")", | |
4652 | __func__, block->idstr, ret, local_size, size); | |
4653 | ret = -EIO; | |
4654 | goto out; | |
4655 | } | |
4656 | ||
4657 | if (end_mark != RAMBLOCK_RECV_BITMAP_ENDING) { | |
af3bbbe9 | 4658 | error_report("%s: ramblock '%s' end mark incorrect: 0x%"PRIx64, |
a335debb PX |
4659 | __func__, block->idstr, end_mark); |
4660 | ret = -EINVAL; | |
4661 | goto out; | |
4662 | } | |
4663 | ||
4664 | /* | |
3a4452d8 | 4665 | * Endianness conversion. We are during postcopy (though paused). |
a335debb PX |
4666 | * The dirty bitmap won't change. We can directly modify it. |
4667 | */ | |
4668 | bitmap_from_le(block->bmap, le_bitmap, nbits); | |
4669 | ||
4670 | /* | |
4671 | * What we received is "received bitmap". Revert it as the initial | |
4672 | * dirty bitmap for this ramblock. | |
4673 | */ | |
4674 | bitmap_complement(block->bmap, block->bmap, nbits); | |
4675 | ||
be39b4cd DH |
4676 | /* Clear dirty bits of discarded ranges that we don't want to migrate. */ |
4677 | ramblock_dirty_bitmap_clear_discarded_pages(block); | |
4678 | ||
4679 | /* We'll recalculate migration_dirty_pages in ram_state_resume_prepare(). */ | |
a335debb PX |
4680 | trace_ram_dirty_bitmap_reload_complete(block->idstr); |
4681 | ||
edd090c7 PX |
4682 | /* |
4683 | * We succeeded to sync bitmap for current ramblock. If this is | |
4684 | * the last one to sync, we need to notify the main send thread. | |
4685 | */ | |
4686 | ram_dirty_bitmap_reload_notify(s); | |
4687 | ||
a335debb PX |
4688 | ret = 0; |
4689 | out: | |
bf269906 | 4690 | g_free(le_bitmap); |
a335debb PX |
4691 | return ret; |
4692 | } | |
4693 | ||
edd090c7 PX |
4694 | static int ram_resume_prepare(MigrationState *s, void *opaque) |
4695 | { | |
4696 | RAMState *rs = *(RAMState **)opaque; | |
08614f34 | 4697 | int ret; |
edd090c7 | 4698 | |
08614f34 PX |
4699 | ret = ram_dirty_bitmap_sync_all(s, rs); |
4700 | if (ret) { | |
4701 | return ret; | |
4702 | } | |
4703 | ||
4704 | ram_state_resume_prepare(rs, s->to_dst_file); | |
4705 | ||
4706 | return 0; | |
edd090c7 PX |
4707 | } |
4708 | ||
36f62f11 PX |
4709 | void postcopy_preempt_shutdown_file(MigrationState *s) |
4710 | { | |
4711 | qemu_put_be64(s->postcopy_qemufile_src, RAM_SAVE_FLAG_EOS); | |
4712 | qemu_fflush(s->postcopy_qemufile_src); | |
4713 | } | |
4714 | ||
56e93d26 | 4715 | static SaveVMHandlers savevm_ram_handlers = { |
9907e842 | 4716 | .save_setup = ram_save_setup, |
56e93d26 | 4717 | .save_live_iterate = ram_save_iterate, |
763c906b | 4718 | .save_live_complete_postcopy = ram_save_complete, |
a3e06c3d | 4719 | .save_live_complete_precopy = ram_save_complete, |
c6467627 | 4720 | .has_postcopy = ram_has_postcopy, |
c8df4a7a JQ |
4721 | .state_pending_exact = ram_state_pending_exact, |
4722 | .state_pending_estimate = ram_state_pending_estimate, | |
56e93d26 | 4723 | .load_state = ram_load, |
f265e0e4 JQ |
4724 | .save_cleanup = ram_save_cleanup, |
4725 | .load_setup = ram_load_setup, | |
4726 | .load_cleanup = ram_load_cleanup, | |
edd090c7 | 4727 | .resume_prepare = ram_resume_prepare, |
56e93d26 JQ |
4728 | }; |
4729 | ||
c7c0e724 DH |
4730 | static void ram_mig_ram_block_resized(RAMBlockNotifier *n, void *host, |
4731 | size_t old_size, size_t new_size) | |
4732 | { | |
cc61c703 | 4733 | PostcopyState ps = postcopy_state_get(); |
c7c0e724 DH |
4734 | ram_addr_t offset; |
4735 | RAMBlock *rb = qemu_ram_block_from_host(host, false, &offset); | |
4736 | Error *err = NULL; | |
4737 | ||
4738 | if (ramblock_is_ignored(rb)) { | |
4739 | return; | |
4740 | } | |
4741 | ||
4742 | if (!migration_is_idle()) { | |
4743 | /* | |
4744 | * Precopy code on the source cannot deal with the size of RAM blocks | |
4745 | * changing at random points in time - especially after sending the | |
4746 | * RAM block sizes in the migration stream, they must no longer change. | |
4747 | * Abort and indicate a proper reason. | |
4748 | */ | |
4749 | error_setg(&err, "RAM block '%s' resized during precopy.", rb->idstr); | |
458fecca | 4750 | migration_cancel(err); |
c7c0e724 | 4751 | error_free(err); |
c7c0e724 | 4752 | } |
cc61c703 DH |
4753 | |
4754 | switch (ps) { | |
4755 | case POSTCOPY_INCOMING_ADVISE: | |
4756 | /* | |
4757 | * Update what ram_postcopy_incoming_init()->init_range() does at the | |
4758 | * time postcopy was advised. Syncing RAM blocks with the source will | |
4759 | * result in RAM resizes. | |
4760 | */ | |
4761 | if (old_size < new_size) { | |
4762 | if (ram_discard_range(rb->idstr, old_size, new_size - old_size)) { | |
4763 | error_report("RAM block '%s' discard of resized RAM failed", | |
4764 | rb->idstr); | |
4765 | } | |
4766 | } | |
898ba906 | 4767 | rb->postcopy_length = new_size; |
cc61c703 DH |
4768 | break; |
4769 | case POSTCOPY_INCOMING_NONE: | |
4770 | case POSTCOPY_INCOMING_RUNNING: | |
4771 | case POSTCOPY_INCOMING_END: | |
4772 | /* | |
4773 | * Once our guest is running, postcopy does no longer care about | |
4774 | * resizes. When growing, the new memory was not available on the | |
4775 | * source, no handler needed. | |
4776 | */ | |
4777 | break; | |
4778 | default: | |
4779 | error_report("RAM block '%s' resized during postcopy state: %d", | |
4780 | rb->idstr, ps); | |
4781 | exit(-1); | |
4782 | } | |
c7c0e724 DH |
4783 | } |
4784 | ||
4785 | static RAMBlockNotifier ram_mig_ram_notifier = { | |
4786 | .ram_block_resized = ram_mig_ram_block_resized, | |
4787 | }; | |
4788 | ||
56e93d26 JQ |
4789 | void ram_mig_init(void) |
4790 | { | |
4791 | qemu_mutex_init(&XBZRLE.lock); | |
ce62df53 | 4792 | register_savevm_live("ram", 0, 4, &savevm_ram_handlers, &ram_state); |
c7c0e724 | 4793 | ram_block_notifier_add(&ram_mig_ram_notifier); |
56e93d26 | 4794 | } |