]>
Commit | Line | Data |
---|---|---|
eb59db53 DDAG |
1 | /* |
2 | * Postcopy migration for RAM | |
3 | * | |
4 | * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates | |
5 | * | |
6 | * Authors: | |
7 | * Dave Gilbert <dgilbert@redhat.com> | |
8 | * | |
9 | * This work is licensed under the terms of the GNU GPL, version 2 or later. | |
10 | * See the COPYING file in the top-level directory. | |
11 | * | |
12 | */ | |
13 | ||
14 | /* | |
15 | * Postcopy is a migration technique where the execution flips from the | |
16 | * source to the destination before all the data has been copied. | |
17 | */ | |
18 | ||
1393a485 | 19 | #include "qemu/osdep.h" |
b85ea5fa | 20 | #include "qemu/madvise.h" |
51180423 | 21 | #include "exec/target_page.h" |
6666c96a | 22 | #include "migration.h" |
08a0aee1 | 23 | #include "qemu-file.h" |
20a519a0 | 24 | #include "savevm.h" |
be07b0ac | 25 | #include "postcopy-ram.h" |
7b1e1a22 | 26 | #include "ram.h" |
1693c64c DDAG |
27 | #include "qapi/error.h" |
28 | #include "qemu/notify.h" | |
d4842052 | 29 | #include "qemu/rcu.h" |
eb59db53 DDAG |
30 | #include "sysemu/sysemu.h" |
31 | #include "qemu/error-report.h" | |
32 | #include "trace.h" | |
5cc8767d | 33 | #include "hw/boards.h" |
898ba906 | 34 | #include "exec/ramblock.h" |
36f62f11 | 35 | #include "socket.h" |
36f62f11 | 36 | #include "yank_functions.h" |
f0afaf6c | 37 | #include "tls.h" |
d5890ea0 | 38 | #include "qemu/userfaultfd.h" |
ae30b9b2 | 39 | #include "qemu/mmap-alloc.h" |
1f0776f1 | 40 | #include "options.h" |
eb59db53 | 41 | |
e0b266f0 DDAG |
42 | /* Arbitrary limit on size of each discard command, |
43 | * keeps them around ~200 bytes | |
44 | */ | |
45 | #define MAX_DISCARDS_PER_COMMAND 12 | |
46 | ||
47 | struct PostcopyDiscardState { | |
48 | const char *ramblock_name; | |
e0b266f0 DDAG |
49 | uint16_t cur_entry; |
50 | /* | |
51 | * Start and length of a discard range (bytes) | |
52 | */ | |
53 | uint64_t start_list[MAX_DISCARDS_PER_COMMAND]; | |
54 | uint64_t length_list[MAX_DISCARDS_PER_COMMAND]; | |
55 | unsigned int nsentwords; | |
56 | unsigned int nsentcmds; | |
57 | }; | |
58 | ||
1693c64c DDAG |
59 | static NotifierWithReturnList postcopy_notifier_list; |
60 | ||
61 | void postcopy_infrastructure_init(void) | |
62 | { | |
63 | notifier_with_return_list_init(&postcopy_notifier_list); | |
64 | } | |
65 | ||
66 | void postcopy_add_notifier(NotifierWithReturn *nn) | |
67 | { | |
68 | notifier_with_return_list_add(&postcopy_notifier_list, nn); | |
69 | } | |
70 | ||
71 | void postcopy_remove_notifier(NotifierWithReturn *n) | |
72 | { | |
73 | notifier_with_return_remove(n); | |
74 | } | |
75 | ||
76 | int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp) | |
77 | { | |
78 | struct PostcopyNotifyData pnd; | |
79 | pnd.reason = reason; | |
80 | pnd.errp = errp; | |
81 | ||
82 | return notifier_with_return_list_notify(&postcopy_notifier_list, | |
83 | &pnd); | |
84 | } | |
85 | ||
095c12a4 PX |
86 | /* |
87 | * NOTE: this routine is not thread safe, we can't call it concurrently. But it | |
88 | * should be good enough for migration's purposes. | |
89 | */ | |
90 | void postcopy_thread_create(MigrationIncomingState *mis, | |
91 | QemuThread *thread, const char *name, | |
92 | void *(*fn)(void *), int joinable) | |
93 | { | |
94 | qemu_sem_init(&mis->thread_sync_sem, 0); | |
95 | qemu_thread_create(thread, name, fn, mis, joinable); | |
96 | qemu_sem_wait(&mis->thread_sync_sem); | |
97 | qemu_sem_destroy(&mis->thread_sync_sem); | |
98 | } | |
99 | ||
eb59db53 DDAG |
100 | /* Postcopy needs to detect accesses to pages that haven't yet been copied |
101 | * across, and efficiently map new pages in, the techniques for doing this | |
102 | * are target OS specific. | |
103 | */ | |
104 | #if defined(__linux__) | |
c4faeed2 | 105 | #include <poll.h> |
eb59db53 DDAG |
106 | #include <sys/ioctl.h> |
107 | #include <sys/syscall.h> | |
eb59db53 DDAG |
108 | #endif |
109 | ||
d8b9d771 MF |
110 | #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD) |
111 | #include <sys/eventfd.h> | |
eb59db53 DDAG |
112 | #include <linux/userfaultfd.h> |
113 | ||
2a4c42f1 AP |
114 | typedef struct PostcopyBlocktimeContext { |
115 | /* time when page fault initiated per vCPU */ | |
116 | uint32_t *page_fault_vcpu_time; | |
117 | /* page address per vCPU */ | |
118 | uintptr_t *vcpu_addr; | |
119 | uint32_t total_blocktime; | |
120 | /* blocktime per vCPU */ | |
121 | uint32_t *vcpu_blocktime; | |
122 | /* point in time when last page fault was initiated */ | |
123 | uint32_t last_begin; | |
124 | /* number of vCPU are suspended */ | |
125 | int smp_cpus_down; | |
126 | uint64_t start_time; | |
127 | ||
128 | /* | |
129 | * Handler for exit event, necessary for | |
130 | * releasing whole blocktime_ctx | |
131 | */ | |
132 | Notifier exit_notifier; | |
133 | } PostcopyBlocktimeContext; | |
134 | ||
135 | static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx) | |
136 | { | |
137 | g_free(ctx->page_fault_vcpu_time); | |
138 | g_free(ctx->vcpu_addr); | |
139 | g_free(ctx->vcpu_blocktime); | |
140 | g_free(ctx); | |
141 | } | |
142 | ||
143 | static void migration_exit_cb(Notifier *n, void *data) | |
144 | { | |
145 | PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext, | |
146 | exit_notifier); | |
147 | destroy_blocktime_context(ctx); | |
148 | } | |
149 | ||
150 | static struct PostcopyBlocktimeContext *blocktime_context_new(void) | |
151 | { | |
5cc8767d LX |
152 | MachineState *ms = MACHINE(qdev_get_machine()); |
153 | unsigned int smp_cpus = ms->smp.cpus; | |
2a4c42f1 AP |
154 | PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1); |
155 | ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus); | |
156 | ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus); | |
157 | ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus); | |
158 | ||
159 | ctx->exit_notifier.notify = migration_exit_cb; | |
160 | ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
161 | qemu_add_exit_notifier(&ctx->exit_notifier); | |
162 | return ctx; | |
163 | } | |
ca6011c2 | 164 | |
65ace060 AP |
165 | static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx) |
166 | { | |
5cc8767d | 167 | MachineState *ms = MACHINE(qdev_get_machine()); |
54aa3de7 | 168 | uint32List *list = NULL; |
65ace060 AP |
169 | int i; |
170 | ||
5cc8767d | 171 | for (i = ms->smp.cpus - 1; i >= 0; i--) { |
54aa3de7 | 172 | QAPI_LIST_PREPEND(list, ctx->vcpu_blocktime[i]); |
65ace060 AP |
173 | } |
174 | ||
175 | return list; | |
176 | } | |
177 | ||
178 | /* | |
179 | * This function just populates MigrationInfo from postcopy's | |
180 | * blocktime context. It will not populate MigrationInfo, | |
181 | * unless postcopy-blocktime capability was set. | |
182 | * | |
183 | * @info: pointer to MigrationInfo to populate | |
184 | */ | |
185 | void fill_destination_postcopy_migration_info(MigrationInfo *info) | |
186 | { | |
187 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
188 | PostcopyBlocktimeContext *bc = mis->blocktime_ctx; | |
189 | ||
190 | if (!bc) { | |
191 | return; | |
192 | } | |
193 | ||
194 | info->has_postcopy_blocktime = true; | |
195 | info->postcopy_blocktime = bc->total_blocktime; | |
196 | info->has_postcopy_vcpu_blocktime = true; | |
197 | info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc); | |
198 | } | |
199 | ||
200 | static uint32_t get_postcopy_total_blocktime(void) | |
201 | { | |
202 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
203 | PostcopyBlocktimeContext *bc = mis->blocktime_ctx; | |
204 | ||
205 | if (!bc) { | |
206 | return 0; | |
207 | } | |
208 | ||
209 | return bc->total_blocktime; | |
210 | } | |
211 | ||
54ae0886 AP |
212 | /** |
213 | * receive_ufd_features: check userfault fd features, to request only supported | |
214 | * features in the future. | |
215 | * | |
216 | * Returns: true on success | |
217 | * | |
218 | * __NR_userfaultfd - should be checked before | |
219 | * @features: out parameter will contain uffdio_api.features provided by kernel | |
220 | * in case of success | |
221 | */ | |
222 | static bool receive_ufd_features(uint64_t *features) | |
eb59db53 | 223 | { |
54ae0886 AP |
224 | struct uffdio_api api_struct = {0}; |
225 | int ufd; | |
226 | bool ret = true; | |
227 | ||
d5890ea0 | 228 | ufd = uffd_open(O_CLOEXEC); |
54ae0886 | 229 | if (ufd == -1) { |
d5890ea0 | 230 | error_report("%s: uffd_open() failed: %s", __func__, strerror(errno)); |
54ae0886 AP |
231 | return false; |
232 | } | |
eb59db53 | 233 | |
54ae0886 | 234 | /* ask features */ |
eb59db53 DDAG |
235 | api_struct.api = UFFD_API; |
236 | api_struct.features = 0; | |
237 | if (ioctl(ufd, UFFDIO_API, &api_struct)) { | |
5553499f | 238 | error_report("%s: UFFDIO_API failed: %s", __func__, |
eb59db53 | 239 | strerror(errno)); |
54ae0886 AP |
240 | ret = false; |
241 | goto release_ufd; | |
242 | } | |
243 | ||
244 | *features = api_struct.features; | |
245 | ||
246 | release_ufd: | |
247 | close(ufd); | |
248 | return ret; | |
249 | } | |
250 | ||
251 | /** | |
252 | * request_ufd_features: this function should be called only once on a newly | |
253 | * opened ufd, subsequent calls will lead to error. | |
254 | * | |
3a4452d8 | 255 | * Returns: true on success |
54ae0886 AP |
256 | * |
257 | * @ufd: fd obtained from userfaultfd syscall | |
258 | * @features: bit mask see UFFD_API_FEATURES | |
259 | */ | |
260 | static bool request_ufd_features(int ufd, uint64_t features) | |
261 | { | |
262 | struct uffdio_api api_struct = {0}; | |
263 | uint64_t ioctl_mask; | |
264 | ||
265 | api_struct.api = UFFD_API; | |
266 | api_struct.features = features; | |
267 | if (ioctl(ufd, UFFDIO_API, &api_struct)) { | |
268 | error_report("%s failed: UFFDIO_API failed: %s", __func__, | |
269 | strerror(errno)); | |
eb59db53 DDAG |
270 | return false; |
271 | } | |
272 | ||
73b49878 PB |
273 | ioctl_mask = 1ULL << _UFFDIO_REGISTER | |
274 | 1ULL << _UFFDIO_UNREGISTER; | |
eb59db53 DDAG |
275 | if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) { |
276 | error_report("Missing userfault features: %" PRIx64, | |
277 | (uint64_t)(~api_struct.ioctls & ioctl_mask)); | |
278 | return false; | |
279 | } | |
280 | ||
54ae0886 AP |
281 | return true; |
282 | } | |
283 | ||
74c38cf7 PX |
284 | static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis, |
285 | Error **errp) | |
54ae0886 AP |
286 | { |
287 | uint64_t asked_features = 0; | |
288 | static uint64_t supported_features; | |
289 | ||
74c38cf7 | 290 | ERRP_GUARD(); |
54ae0886 AP |
291 | /* |
292 | * it's not possible to | |
293 | * request UFFD_API twice per one fd | |
294 | * userfault fd features is persistent | |
295 | */ | |
296 | if (!supported_features) { | |
297 | if (!receive_ufd_features(&supported_features)) { | |
74c38cf7 | 298 | error_setg(errp, "Userfault feature detection failed"); |
54ae0886 AP |
299 | return false; |
300 | } | |
301 | } | |
302 | ||
2a4c42f1 | 303 | #ifdef UFFD_FEATURE_THREAD_ID |
2d1c37c6 | 304 | if (UFFD_FEATURE_THREAD_ID & supported_features) { |
2a4c42f1 | 305 | asked_features |= UFFD_FEATURE_THREAD_ID; |
2d1c37c6 PX |
306 | if (migrate_postcopy_blocktime()) { |
307 | if (!mis->blocktime_ctx) { | |
308 | mis->blocktime_ctx = blocktime_context_new(); | |
309 | } | |
310 | } | |
2a4c42f1 AP |
311 | } |
312 | #endif | |
313 | ||
54ae0886 AP |
314 | /* |
315 | * request features, even if asked_features is 0, due to | |
316 | * kernel expects UFFD_API before UFFDIO_REGISTER, per | |
317 | * userfault file descriptor | |
318 | */ | |
319 | if (!request_ufd_features(ufd, asked_features)) { | |
74c38cf7 | 320 | error_setg(errp, "Failed features %" PRIu64, asked_features); |
54ae0886 AP |
321 | return false; |
322 | } | |
323 | ||
8e3b0cbb | 324 | if (qemu_real_host_page_size() != ram_pagesize_summary()) { |
7e8cafb7 DDAG |
325 | bool have_hp = false; |
326 | /* We've got a huge page */ | |
327 | #ifdef UFFD_FEATURE_MISSING_HUGETLBFS | |
54ae0886 | 328 | have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS; |
7e8cafb7 DDAG |
329 | #endif |
330 | if (!have_hp) { | |
74c38cf7 PX |
331 | error_setg(errp, |
332 | "Userfault on this host does not support huge pages"); | |
7e8cafb7 DDAG |
333 | return false; |
334 | } | |
335 | } | |
eb59db53 DDAG |
336 | return true; |
337 | } | |
338 | ||
8679638b DDAG |
339 | /* Callback from postcopy_ram_supported_by_host block iterator. |
340 | */ | |
74c38cf7 | 341 | static int test_ramblock_postcopiable(RAMBlock *rb, Error **errp) |
8679638b | 342 | { |
754cb9c0 YK |
343 | const char *block_name = qemu_ram_get_idstr(rb); |
344 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
5d214a92 | 345 | size_t pagesize = qemu_ram_pagesize(rb); |
ae30b9b2 | 346 | QemuFsType fs; |
5d214a92 | 347 | |
5d214a92 | 348 | if (length % pagesize) { |
74c38cf7 PX |
349 | error_setg(errp, |
350 | "Postcopy requires RAM blocks to be a page size multiple," | |
351 | " block %s is 0x" RAM_ADDR_FMT " bytes with a " | |
352 | "page size of 0x%zx", block_name, length, pagesize); | |
5d214a92 DDAG |
353 | return 1; |
354 | } | |
ae30b9b2 PX |
355 | |
356 | if (rb->fd >= 0) { | |
357 | fs = qemu_fd_getfs(rb->fd); | |
358 | if (fs != QEMU_FS_TYPE_TMPFS && fs != QEMU_FS_TYPE_HUGETLBFS) { | |
74c38cf7 PX |
359 | error_setg(errp, |
360 | "Host backend files need to be TMPFS or HUGETLBFS only"); | |
ae30b9b2 PX |
361 | return 1; |
362 | } | |
363 | } | |
364 | ||
8679638b DDAG |
365 | return 0; |
366 | } | |
367 | ||
58b7c17e DDAG |
368 | /* |
369 | * Note: This has the side effect of munlock'ing all of RAM, that's | |
370 | * normally fine since if the postcopy succeeds it gets turned back on at the | |
371 | * end. | |
372 | */ | |
74c38cf7 | 373 | bool postcopy_ram_supported_by_host(MigrationIncomingState *mis, Error **errp) |
eb59db53 | 374 | { |
8e3b0cbb | 375 | long pagesize = qemu_real_host_page_size(); |
eb59db53 DDAG |
376 | int ufd = -1; |
377 | bool ret = false; /* Error unless we change it */ | |
378 | void *testarea = NULL; | |
379 | struct uffdio_register reg_struct; | |
380 | struct uffdio_range range_struct; | |
381 | uint64_t feature_mask; | |
ae30b9b2 | 382 | RAMBlock *block; |
eb59db53 | 383 | |
74c38cf7 | 384 | ERRP_GUARD(); |
20afaed9 | 385 | if (qemu_target_page_size() > pagesize) { |
74c38cf7 | 386 | error_setg(errp, "Target page size bigger than host page size"); |
eb59db53 DDAG |
387 | goto out; |
388 | } | |
389 | ||
d5890ea0 | 390 | ufd = uffd_open(O_CLOEXEC); |
eb59db53 | 391 | if (ufd == -1) { |
74c38cf7 | 392 | error_setg(errp, "Userfaultfd not available: %s", strerror(errno)); |
eb59db53 DDAG |
393 | goto out; |
394 | } | |
395 | ||
1693c64c | 396 | /* Give devices a chance to object */ |
74c38cf7 | 397 | if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, errp)) { |
1693c64c DDAG |
398 | goto out; |
399 | } | |
400 | ||
eb59db53 | 401 | /* Version and features check */ |
74c38cf7 | 402 | if (!ufd_check_and_apply(ufd, mis, errp)) { |
eb59db53 DDAG |
403 | goto out; |
404 | } | |
405 | ||
ae30b9b2 PX |
406 | /* |
407 | * We don't support postcopy with some type of ramblocks. | |
408 | * | |
f161c88a | 409 | * NOTE: we explicitly ignored migrate_ram_is_ignored() instead we checked |
ae30b9b2 PX |
410 | * all possible ramblocks. This is because this function can be called |
411 | * when creating the migration object, during the phase RAM_MIGRATABLE | |
412 | * is not even properly set for all the ramblocks. | |
413 | * | |
414 | * A side effect of this is we'll also check against RAM_SHARED | |
415 | * ramblocks even if migrate_ignore_shared() is set (in which case | |
416 | * we'll never migrate RAM_SHARED at all), but normally this shouldn't | |
417 | * affect in reality, or we can revisit. | |
418 | */ | |
419 | RAMBLOCK_FOREACH(block) { | |
74c38cf7 | 420 | if (test_ramblock_postcopiable(block, errp)) { |
ae30b9b2 PX |
421 | goto out; |
422 | } | |
8679638b DDAG |
423 | } |
424 | ||
58b7c17e DDAG |
425 | /* |
426 | * userfault and mlock don't go together; we'll put it back later if | |
427 | * it was enabled. | |
428 | */ | |
429 | if (munlockall()) { | |
74c38cf7 | 430 | error_setg(errp, "munlockall() failed: %s", strerror(errno)); |
617a32f5 | 431 | goto out; |
58b7c17e DDAG |
432 | } |
433 | ||
eb59db53 DDAG |
434 | /* |
435 | * We need to check that the ops we need are supported on anon memory | |
436 | * To do that we need to register a chunk and see the flags that | |
437 | * are returned. | |
438 | */ | |
439 | testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | | |
440 | MAP_ANONYMOUS, -1, 0); | |
441 | if (testarea == MAP_FAILED) { | |
74c38cf7 | 442 | error_setg(errp, "Failed to map test area: %s", strerror(errno)); |
eb59db53 DDAG |
443 | goto out; |
444 | } | |
7648297d | 445 | g_assert(QEMU_PTR_IS_ALIGNED(testarea, pagesize)); |
eb59db53 DDAG |
446 | |
447 | reg_struct.range.start = (uintptr_t)testarea; | |
448 | reg_struct.range.len = pagesize; | |
449 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; | |
450 | ||
451 | if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) { | |
74c38cf7 | 452 | error_setg(errp, "UFFDIO_REGISTER failed: %s", strerror(errno)); |
eb59db53 DDAG |
453 | goto out; |
454 | } | |
455 | ||
456 | range_struct.start = (uintptr_t)testarea; | |
457 | range_struct.len = pagesize; | |
458 | if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) { | |
74c38cf7 | 459 | error_setg(errp, "UFFDIO_UNREGISTER failed: %s", strerror(errno)); |
eb59db53 DDAG |
460 | goto out; |
461 | } | |
462 | ||
73b49878 PB |
463 | feature_mask = 1ULL << _UFFDIO_WAKE | |
464 | 1ULL << _UFFDIO_COPY | | |
465 | 1ULL << _UFFDIO_ZEROPAGE; | |
eb59db53 | 466 | if ((reg_struct.ioctls & feature_mask) != feature_mask) { |
74c38cf7 PX |
467 | error_setg(errp, "Missing userfault map features: %" PRIx64, |
468 | (uint64_t)(~reg_struct.ioctls & feature_mask)); | |
eb59db53 DDAG |
469 | goto out; |
470 | } | |
471 | ||
472 | /* Success! */ | |
473 | ret = true; | |
474 | out: | |
475 | if (testarea) { | |
476 | munmap(testarea, pagesize); | |
477 | } | |
478 | if (ufd != -1) { | |
479 | close(ufd); | |
480 | } | |
481 | return ret; | |
482 | } | |
483 | ||
1caddf8a DDAG |
484 | /* |
485 | * Setup an area of RAM so that it *can* be used for postcopy later; this | |
486 | * must be done right at the start prior to pre-copy. | |
487 | * opaque should be the MIS. | |
488 | */ | |
754cb9c0 | 489 | static int init_range(RAMBlock *rb, void *opaque) |
1caddf8a | 490 | { |
754cb9c0 YK |
491 | const char *block_name = qemu_ram_get_idstr(rb); |
492 | void *host_addr = qemu_ram_get_host_addr(rb); | |
493 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
494 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
1caddf8a DDAG |
495 | trace_postcopy_init_range(block_name, host_addr, offset, length); |
496 | ||
898ba906 DH |
497 | /* |
498 | * Save the used_length before running the guest. In case we have to | |
499 | * resize RAM blocks when syncing RAM block sizes from the source during | |
500 | * precopy, we'll update it manually via the ram block notifier. | |
501 | */ | |
502 | rb->postcopy_length = length; | |
503 | ||
1caddf8a DDAG |
504 | /* |
505 | * We need the whole of RAM to be truly empty for postcopy, so things | |
506 | * like ROMs and any data tables built during init must be zero'd | |
507 | * - we're going to get the copy from the source anyway. | |
508 | * (Precopy will just overwrite this data, so doesn't need the discard) | |
509 | */ | |
aaa2064c | 510 | if (ram_discard_range(block_name, 0, length)) { |
1caddf8a DDAG |
511 | return -1; |
512 | } | |
513 | ||
514 | return 0; | |
515 | } | |
516 | ||
517 | /* | |
518 | * At the end of migration, undo the effects of init_range | |
519 | * opaque should be the MIS. | |
520 | */ | |
754cb9c0 | 521 | static int cleanup_range(RAMBlock *rb, void *opaque) |
1caddf8a | 522 | { |
754cb9c0 YK |
523 | const char *block_name = qemu_ram_get_idstr(rb); |
524 | void *host_addr = qemu_ram_get_host_addr(rb); | |
525 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
898ba906 | 526 | ram_addr_t length = rb->postcopy_length; |
1caddf8a DDAG |
527 | MigrationIncomingState *mis = opaque; |
528 | struct uffdio_range range_struct; | |
529 | trace_postcopy_cleanup_range(block_name, host_addr, offset, length); | |
530 | ||
531 | /* | |
532 | * We turned off hugepage for the precopy stage with postcopy enabled | |
533 | * we can turn it back on now. | |
534 | */ | |
1d741439 | 535 | qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE); |
1caddf8a DDAG |
536 | |
537 | /* | |
538 | * We can also turn off userfault now since we should have all the | |
539 | * pages. It can be useful to leave it on to debug postcopy | |
540 | * if you're not sure it's always getting every page. | |
541 | */ | |
542 | range_struct.start = (uintptr_t)host_addr; | |
543 | range_struct.len = length; | |
544 | ||
545 | if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) { | |
546 | error_report("%s: userfault unregister %s", __func__, strerror(errno)); | |
547 | ||
548 | return -1; | |
549 | } | |
550 | ||
551 | return 0; | |
552 | } | |
553 | ||
554 | /* | |
555 | * Initialise postcopy-ram, setting the RAM to a state where we can go into | |
556 | * postcopy later; must be called prior to any precopy. | |
557 | * called from arch_init's similarly named ram_postcopy_incoming_init | |
558 | */ | |
c136180c | 559 | int postcopy_ram_incoming_init(MigrationIncomingState *mis) |
1caddf8a | 560 | { |
fbd162e6 | 561 | if (foreach_not_ignored_block(init_range, NULL)) { |
1caddf8a DDAG |
562 | return -1; |
563 | } | |
564 | ||
565 | return 0; | |
566 | } | |
567 | ||
476ebf77 PX |
568 | static void postcopy_temp_pages_cleanup(MigrationIncomingState *mis) |
569 | { | |
77dadc3f PX |
570 | int i; |
571 | ||
572 | if (mis->postcopy_tmp_pages) { | |
573 | for (i = 0; i < mis->postcopy_channels; i++) { | |
574 | if (mis->postcopy_tmp_pages[i].tmp_huge_page) { | |
575 | munmap(mis->postcopy_tmp_pages[i].tmp_huge_page, | |
576 | mis->largest_page_size); | |
577 | mis->postcopy_tmp_pages[i].tmp_huge_page = NULL; | |
578 | } | |
579 | } | |
580 | g_free(mis->postcopy_tmp_pages); | |
581 | mis->postcopy_tmp_pages = NULL; | |
476ebf77 PX |
582 | } |
583 | ||
584 | if (mis->postcopy_tmp_zero_page) { | |
585 | munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size); | |
586 | mis->postcopy_tmp_zero_page = NULL; | |
587 | } | |
588 | } | |
589 | ||
1caddf8a DDAG |
590 | /* |
591 | * At the end of a migration where postcopy_ram_incoming_init was called. | |
592 | */ | |
593 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
594 | { | |
c4faeed2 DDAG |
595 | trace_postcopy_ram_incoming_cleanup_entry(); |
596 | ||
6621883f PX |
597 | if (mis->preempt_thread_status == PREEMPT_THREAD_CREATED) { |
598 | /* Notify the fast load thread to quit */ | |
599 | mis->preempt_thread_status = PREEMPT_THREAD_QUIT; | |
cf02f29e PX |
600 | /* |
601 | * Update preempt_thread_status before reading count. Note: mutex | |
602 | * lock only provide ACQUIRE semantic, and it doesn't stops this | |
603 | * write to be reordered after reading the count. | |
604 | */ | |
605 | smp_mb(); | |
606 | /* | |
607 | * It's possible that the preempt thread is still handling the last | |
608 | * pages to arrive which were requested by guest page faults. | |
609 | * Making sure nothing is left behind by waiting on the condvar if | |
610 | * that unlikely case happened. | |
611 | */ | |
612 | WITH_QEMU_LOCK_GUARD(&mis->page_request_mutex) { | |
613 | if (qatomic_read(&mis->page_requested_count)) { | |
614 | /* | |
615 | * It is guaranteed to receive a signal later, because the | |
616 | * count>0 now, so it's destined to be decreased to zero | |
617 | * very soon by the preempt thread. | |
618 | */ | |
619 | qemu_cond_wait(&mis->page_request_cond, | |
620 | &mis->page_request_mutex); | |
621 | } | |
622 | } | |
623 | /* Notify the fast load thread to quit */ | |
6621883f PX |
624 | if (mis->postcopy_qemufile_dst) { |
625 | qemu_file_shutdown(mis->postcopy_qemufile_dst); | |
626 | } | |
36f62f11 | 627 | qemu_thread_join(&mis->postcopy_prio_thread); |
6621883f | 628 | mis->preempt_thread_status = PREEMPT_THREAD_NONE; |
36f62f11 PX |
629 | } |
630 | ||
c4faeed2 | 631 | if (mis->have_fault_thread) { |
46343570 DDAG |
632 | Error *local_err = NULL; |
633 | ||
55d0fe82 | 634 | /* Let the fault thread quit */ |
d73415a3 | 635 | qatomic_set(&mis->fault_thread_quit, 1); |
55d0fe82 IM |
636 | postcopy_fault_thread_notify(mis); |
637 | trace_postcopy_ram_incoming_cleanup_join(); | |
638 | qemu_thread_join(&mis->fault_thread); | |
639 | ||
46343570 DDAG |
640 | if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) { |
641 | error_report_err(local_err); | |
642 | return -1; | |
643 | } | |
644 | ||
fbd162e6 | 645 | if (foreach_not_ignored_block(cleanup_range, mis)) { |
c4faeed2 DDAG |
646 | return -1; |
647 | } | |
9ab7ef9b | 648 | |
c4faeed2 DDAG |
649 | trace_postcopy_ram_incoming_cleanup_closeuf(); |
650 | close(mis->userfault_fd); | |
64f615fe | 651 | close(mis->userfault_event_fd); |
c4faeed2 | 652 | mis->have_fault_thread = false; |
1caddf8a DDAG |
653 | } |
654 | ||
58b7c17e DDAG |
655 | if (enable_mlock) { |
656 | if (os_mlock() < 0) { | |
657 | error_report("mlock: %s", strerror(errno)); | |
658 | /* | |
659 | * It doesn't feel right to fail at this point, we have a valid | |
660 | * VM state. | |
661 | */ | |
662 | } | |
663 | } | |
664 | ||
476ebf77 PX |
665 | postcopy_temp_pages_cleanup(mis); |
666 | ||
65ace060 AP |
667 | trace_postcopy_ram_incoming_cleanup_blocktime( |
668 | get_postcopy_total_blocktime()); | |
669 | ||
c4faeed2 | 670 | trace_postcopy_ram_incoming_cleanup_exit(); |
1caddf8a DDAG |
671 | return 0; |
672 | } | |
673 | ||
f9527107 DDAG |
674 | /* |
675 | * Disable huge pages on an area | |
676 | */ | |
754cb9c0 | 677 | static int nhp_range(RAMBlock *rb, void *opaque) |
f9527107 | 678 | { |
754cb9c0 YK |
679 | const char *block_name = qemu_ram_get_idstr(rb); |
680 | void *host_addr = qemu_ram_get_host_addr(rb); | |
681 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
898ba906 | 682 | ram_addr_t length = rb->postcopy_length; |
f9527107 DDAG |
683 | trace_postcopy_nhp_range(block_name, host_addr, offset, length); |
684 | ||
685 | /* | |
686 | * Before we do discards we need to ensure those discards really | |
687 | * do delete areas of the page, even if THP thinks a hugepage would | |
688 | * be a good idea, so force hugepages off. | |
689 | */ | |
1d741439 | 690 | qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE); |
f9527107 DDAG |
691 | |
692 | return 0; | |
693 | } | |
694 | ||
695 | /* | |
696 | * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard | |
697 | * however leaving it until after precopy means that most of the precopy | |
698 | * data is still THPd | |
699 | */ | |
700 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) | |
701 | { | |
fbd162e6 | 702 | if (foreach_not_ignored_block(nhp_range, mis)) { |
f9527107 DDAG |
703 | return -1; |
704 | } | |
705 | ||
706 | postcopy_state_set(POSTCOPY_INCOMING_DISCARD); | |
707 | ||
708 | return 0; | |
709 | } | |
710 | ||
f0a227ad DDAG |
711 | /* |
712 | * Mark the given area of RAM as requiring notification to unwritten areas | |
fbd162e6 | 713 | * Used as a callback on foreach_not_ignored_block. |
f0a227ad DDAG |
714 | * host_addr: Base of area to mark |
715 | * offset: Offset in the whole ram arena | |
716 | * length: Length of the section | |
717 | * opaque: MigrationIncomingState pointer | |
718 | * Returns 0 on success | |
719 | */ | |
754cb9c0 | 720 | static int ram_block_enable_notify(RAMBlock *rb, void *opaque) |
f0a227ad DDAG |
721 | { |
722 | MigrationIncomingState *mis = opaque; | |
723 | struct uffdio_register reg_struct; | |
724 | ||
754cb9c0 | 725 | reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb); |
898ba906 | 726 | reg_struct.range.len = rb->postcopy_length; |
f0a227ad DDAG |
727 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; |
728 | ||
729 | /* Now tell our userfault_fd that it's responsible for this area */ | |
730 | if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) { | |
731 | error_report("%s userfault register: %s", __func__, strerror(errno)); | |
732 | return -1; | |
733 | } | |
73b49878 | 734 | if (!(reg_struct.ioctls & (1ULL << _UFFDIO_COPY))) { |
665414ad DDAG |
735 | error_report("%s userfault: Region doesn't support COPY", __func__); |
736 | return -1; | |
737 | } | |
73b49878 | 738 | if (reg_struct.ioctls & (1ULL << _UFFDIO_ZEROPAGE)) { |
2ce16640 DDAG |
739 | qemu_ram_set_uf_zeroable(rb); |
740 | } | |
f0a227ad DDAG |
741 | |
742 | return 0; | |
743 | } | |
744 | ||
5efc3564 DDAG |
745 | int postcopy_wake_shared(struct PostCopyFD *pcfd, |
746 | uint64_t client_addr, | |
747 | RAMBlock *rb) | |
748 | { | |
749 | size_t pagesize = qemu_ram_pagesize(rb); | |
750 | struct uffdio_range range; | |
751 | int ret; | |
752 | trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb)); | |
7648297d | 753 | range.start = ROUND_DOWN(client_addr, pagesize); |
5efc3564 DDAG |
754 | range.len = pagesize; |
755 | ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range); | |
756 | if (ret) { | |
757 | error_report("%s: Failed to wake: %zx in %s (%s)", | |
758 | __func__, (size_t)client_addr, qemu_ram_get_idstr(rb), | |
759 | strerror(errno)); | |
760 | } | |
761 | return ret; | |
762 | } | |
763 | ||
9470c5e0 DH |
764 | static int postcopy_request_page(MigrationIncomingState *mis, RAMBlock *rb, |
765 | ram_addr_t start, uint64_t haddr) | |
766 | { | |
767 | void *aligned = (void *)(uintptr_t)ROUND_DOWN(haddr, qemu_ram_pagesize(rb)); | |
768 | ||
769 | /* | |
770 | * Discarded pages (via RamDiscardManager) are never migrated. On unlikely | |
771 | * access, place a zeropage, which will also set the relevant bits in the | |
772 | * recv_bitmap accordingly, so we won't try placing a zeropage twice. | |
773 | * | |
774 | * Checking a single bit is sufficient to handle pagesize > TPS as either | |
775 | * all relevant bits are set or not. | |
776 | */ | |
777 | assert(QEMU_IS_ALIGNED(start, qemu_ram_pagesize(rb))); | |
778 | if (ramblock_page_is_discarded(rb, start)) { | |
779 | bool received = ramblock_recv_bitmap_test_byte_offset(rb, start); | |
780 | ||
781 | return received ? 0 : postcopy_place_page_zero(mis, aligned, rb); | |
782 | } | |
783 | ||
784 | return migrate_send_rp_req_pages(mis, rb, start, haddr); | |
785 | } | |
786 | ||
096bf4c8 DDAG |
787 | /* |
788 | * Callback from shared fault handlers to ask for a page, | |
789 | * the page must be specified by a RAMBlock and an offset in that rb | |
790 | * Note: Only for use by shared fault handlers (in fault thread) | |
791 | */ | |
792 | int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, | |
793 | uint64_t client_addr, uint64_t rb_offset) | |
794 | { | |
7648297d | 795 | uint64_t aligned_rbo = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb)); |
096bf4c8 DDAG |
796 | MigrationIncomingState *mis = migration_incoming_get_current(); |
797 | ||
798 | trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb), | |
799 | rb_offset); | |
dedfb4b2 DDAG |
800 | if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) { |
801 | trace_postcopy_request_shared_page_present(pcfd->idstr, | |
802 | qemu_ram_get_idstr(rb), rb_offset); | |
803 | return postcopy_wake_shared(pcfd, client_addr, rb); | |
804 | } | |
9470c5e0 | 805 | postcopy_request_page(mis, rb, aligned_rbo, client_addr); |
096bf4c8 DDAG |
806 | return 0; |
807 | } | |
808 | ||
575b0b33 AP |
809 | static int get_mem_fault_cpu_index(uint32_t pid) |
810 | { | |
811 | CPUState *cpu_iter; | |
812 | ||
813 | CPU_FOREACH(cpu_iter) { | |
814 | if (cpu_iter->thread_id == pid) { | |
815 | trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid); | |
816 | return cpu_iter->cpu_index; | |
817 | } | |
818 | } | |
819 | trace_get_mem_fault_cpu_index(-1, pid); | |
820 | return -1; | |
821 | } | |
822 | ||
823 | static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc) | |
824 | { | |
825 | int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - | |
826 | dc->start_time; | |
827 | return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX; | |
828 | } | |
829 | ||
830 | /* | |
831 | * This function is being called when pagefault occurs. It | |
832 | * tracks down vCPU blocking time. | |
833 | * | |
834 | * @addr: faulted host virtual address | |
835 | * @ptid: faulted process thread id | |
836 | * @rb: ramblock appropriate to addr | |
837 | */ | |
838 | static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid, | |
839 | RAMBlock *rb) | |
840 | { | |
841 | int cpu, already_received; | |
842 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
843 | PostcopyBlocktimeContext *dc = mis->blocktime_ctx; | |
844 | uint32_t low_time_offset; | |
845 | ||
846 | if (!dc || ptid == 0) { | |
847 | return; | |
848 | } | |
849 | cpu = get_mem_fault_cpu_index(ptid); | |
850 | if (cpu < 0) { | |
851 | return; | |
852 | } | |
853 | ||
854 | low_time_offset = get_low_time_offset(dc); | |
855 | if (dc->vcpu_addr[cpu] == 0) { | |
d73415a3 | 856 | qatomic_inc(&dc->smp_cpus_down); |
575b0b33 AP |
857 | } |
858 | ||
d73415a3 SH |
859 | qatomic_xchg(&dc->last_begin, low_time_offset); |
860 | qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset); | |
861 | qatomic_xchg(&dc->vcpu_addr[cpu], addr); | |
575b0b33 | 862 | |
da1725d3 WY |
863 | /* |
864 | * check it here, not at the beginning of the function, | |
865 | * due to, check could occur early than bitmap_set in | |
866 | * qemu_ufd_copy_ioctl | |
867 | */ | |
575b0b33 AP |
868 | already_received = ramblock_recv_bitmap_test(rb, (void *)addr); |
869 | if (already_received) { | |
d73415a3 SH |
870 | qatomic_xchg(&dc->vcpu_addr[cpu], 0); |
871 | qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0); | |
872 | qatomic_dec(&dc->smp_cpus_down); | |
575b0b33 AP |
873 | } |
874 | trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu], | |
875 | cpu, already_received); | |
876 | } | |
877 | ||
878 | /* | |
879 | * This function just provide calculated blocktime per cpu and trace it. | |
880 | * Total blocktime is calculated in mark_postcopy_blocktime_end. | |
881 | * | |
882 | * | |
883 | * Assume we have 3 CPU | |
884 | * | |
885 | * S1 E1 S1 E1 | |
886 | * -----***********------------xxx***************------------------------> CPU1 | |
887 | * | |
888 | * S2 E2 | |
889 | * ------------****************xxx---------------------------------------> CPU2 | |
890 | * | |
891 | * S3 E3 | |
892 | * ------------------------****xxx********-------------------------------> CPU3 | |
893 | * | |
894 | * We have sequence S1,S2,E1,S3,S1,E2,E3,E1 | |
895 | * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3 | |
896 | * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 - | |
897 | * it's a part of total blocktime. | |
898 | * S1 - here is last_begin | |
899 | * Legend of the picture is following: | |
900 | * * - means blocktime per vCPU | |
901 | * x - means overlapped blocktime (total blocktime) | |
902 | * | |
903 | * @addr: host virtual address | |
904 | */ | |
905 | static void mark_postcopy_blocktime_end(uintptr_t addr) | |
906 | { | |
907 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
908 | PostcopyBlocktimeContext *dc = mis->blocktime_ctx; | |
5cc8767d LX |
909 | MachineState *ms = MACHINE(qdev_get_machine()); |
910 | unsigned int smp_cpus = ms->smp.cpus; | |
575b0b33 AP |
911 | int i, affected_cpu = 0; |
912 | bool vcpu_total_blocktime = false; | |
913 | uint32_t read_vcpu_time, low_time_offset; | |
914 | ||
915 | if (!dc) { | |
916 | return; | |
917 | } | |
918 | ||
919 | low_time_offset = get_low_time_offset(dc); | |
920 | /* lookup cpu, to clear it, | |
3a4452d8 | 921 | * that algorithm looks straightforward, but it's not |
575b0b33 AP |
922 | * optimal, more optimal algorithm is keeping tree or hash |
923 | * where key is address value is a list of */ | |
924 | for (i = 0; i < smp_cpus; i++) { | |
925 | uint32_t vcpu_blocktime = 0; | |
926 | ||
d73415a3 SH |
927 | read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0); |
928 | if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr || | |
575b0b33 AP |
929 | read_vcpu_time == 0) { |
930 | continue; | |
931 | } | |
d73415a3 | 932 | qatomic_xchg(&dc->vcpu_addr[i], 0); |
575b0b33 AP |
933 | vcpu_blocktime = low_time_offset - read_vcpu_time; |
934 | affected_cpu += 1; | |
935 | /* we need to know is that mark_postcopy_end was due to | |
936 | * faulted page, another possible case it's prefetched | |
937 | * page and in that case we shouldn't be here */ | |
938 | if (!vcpu_total_blocktime && | |
d73415a3 | 939 | qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) { |
575b0b33 AP |
940 | vcpu_total_blocktime = true; |
941 | } | |
942 | /* continue cycle, due to one page could affect several vCPUs */ | |
943 | dc->vcpu_blocktime[i] += vcpu_blocktime; | |
944 | } | |
945 | ||
d73415a3 | 946 | qatomic_sub(&dc->smp_cpus_down, affected_cpu); |
575b0b33 | 947 | if (vcpu_total_blocktime) { |
d73415a3 | 948 | dc->total_blocktime += low_time_offset - qatomic_fetch_add( |
575b0b33 AP |
949 | &dc->last_begin, 0); |
950 | } | |
951 | trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime, | |
952 | affected_cpu); | |
953 | } | |
954 | ||
27dd21b4 | 955 | static void postcopy_pause_fault_thread(MigrationIncomingState *mis) |
3a7804c3 PX |
956 | { |
957 | trace_postcopy_pause_fault_thread(); | |
3a7804c3 | 958 | qemu_sem_wait(&mis->postcopy_pause_sem_fault); |
3a7804c3 | 959 | trace_postcopy_pause_fault_thread_continued(); |
3a7804c3 PX |
960 | } |
961 | ||
f0a227ad DDAG |
962 | /* |
963 | * Handle faults detected by the USERFAULT markings | |
964 | */ | |
965 | static void *postcopy_ram_fault_thread(void *opaque) | |
966 | { | |
967 | MigrationIncomingState *mis = opaque; | |
c4faeed2 DDAG |
968 | struct uffd_msg msg; |
969 | int ret; | |
00fa4fc8 | 970 | size_t index; |
c4faeed2 | 971 | RAMBlock *rb = NULL; |
f0a227ad | 972 | |
c4faeed2 | 973 | trace_postcopy_ram_fault_thread_entry(); |
74637e6f | 974 | rcu_register_thread(); |
096bf4c8 | 975 | mis->last_rb = NULL; /* last RAMBlock we sent part of */ |
095c12a4 | 976 | qemu_sem_post(&mis->thread_sync_sem); |
f0a227ad | 977 | |
00fa4fc8 DDAG |
978 | struct pollfd *pfd; |
979 | size_t pfd_len = 2 + mis->postcopy_remote_fds->len; | |
980 | ||
981 | pfd = g_new0(struct pollfd, pfd_len); | |
982 | ||
983 | pfd[0].fd = mis->userfault_fd; | |
984 | pfd[0].events = POLLIN; | |
985 | pfd[1].fd = mis->userfault_event_fd; | |
986 | pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */ | |
987 | trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd); | |
988 | for (index = 0; index < mis->postcopy_remote_fds->len; index++) { | |
989 | struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds, | |
990 | struct PostCopyFD, index); | |
991 | pfd[2 + index].fd = pcfd->fd; | |
992 | pfd[2 + index].events = POLLIN; | |
993 | trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr, | |
994 | pcfd->fd); | |
995 | } | |
996 | ||
c4faeed2 DDAG |
997 | while (true) { |
998 | ram_addr_t rb_offset; | |
00fa4fc8 | 999 | int poll_result; |
c4faeed2 DDAG |
1000 | |
1001 | /* | |
1002 | * We're mainly waiting for the kernel to give us a faulting HVA, | |
1003 | * however we can be told to quit via userfault_quit_fd which is | |
1004 | * an eventfd | |
1005 | */ | |
00fa4fc8 DDAG |
1006 | |
1007 | poll_result = poll(pfd, pfd_len, -1 /* Wait forever */); | |
1008 | if (poll_result == -1) { | |
c4faeed2 DDAG |
1009 | error_report("%s: userfault poll: %s", __func__, strerror(errno)); |
1010 | break; | |
1011 | } | |
1012 | ||
3a7804c3 PX |
1013 | if (!mis->to_src_file) { |
1014 | /* | |
1015 | * Possibly someone tells us that the return path is | |
1016 | * broken already using the event. We should hold until | |
1017 | * the channel is rebuilt. | |
1018 | */ | |
27dd21b4 | 1019 | postcopy_pause_fault_thread(mis); |
3a7804c3 PX |
1020 | } |
1021 | ||
c4faeed2 | 1022 | if (pfd[1].revents) { |
64f615fe PX |
1023 | uint64_t tmp64 = 0; |
1024 | ||
1025 | /* Consume the signal */ | |
1026 | if (read(mis->userfault_event_fd, &tmp64, 8) != 8) { | |
1027 | /* Nothing obviously nicer than posting this error. */ | |
1028 | error_report("%s: read() failed", __func__); | |
1029 | } | |
1030 | ||
d73415a3 | 1031 | if (qatomic_read(&mis->fault_thread_quit)) { |
64f615fe PX |
1032 | trace_postcopy_ram_fault_thread_quit(); |
1033 | break; | |
1034 | } | |
c4faeed2 DDAG |
1035 | } |
1036 | ||
00fa4fc8 DDAG |
1037 | if (pfd[0].revents) { |
1038 | poll_result--; | |
1039 | ret = read(mis->userfault_fd, &msg, sizeof(msg)); | |
1040 | if (ret != sizeof(msg)) { | |
1041 | if (errno == EAGAIN) { | |
1042 | /* | |
1043 | * if a wake up happens on the other thread just after | |
1044 | * the poll, there is nothing to read. | |
1045 | */ | |
1046 | continue; | |
1047 | } | |
1048 | if (ret < 0) { | |
1049 | error_report("%s: Failed to read full userfault " | |
1050 | "message: %s", | |
1051 | __func__, strerror(errno)); | |
1052 | break; | |
1053 | } else { | |
1054 | error_report("%s: Read %d bytes from userfaultfd " | |
1055 | "expected %zd", | |
1056 | __func__, ret, sizeof(msg)); | |
1057 | break; /* Lost alignment, don't know what we'd read next */ | |
1058 | } | |
c4faeed2 | 1059 | } |
00fa4fc8 DDAG |
1060 | if (msg.event != UFFD_EVENT_PAGEFAULT) { |
1061 | error_report("%s: Read unexpected event %ud from userfaultfd", | |
1062 | __func__, msg.event); | |
1063 | continue; /* It's not a page fault, shouldn't happen */ | |
c4faeed2 | 1064 | } |
c4faeed2 | 1065 | |
00fa4fc8 DDAG |
1066 | rb = qemu_ram_block_from_host( |
1067 | (void *)(uintptr_t)msg.arg.pagefault.address, | |
1068 | true, &rb_offset); | |
1069 | if (!rb) { | |
1070 | error_report("postcopy_ram_fault_thread: Fault outside guest: %" | |
1071 | PRIx64, (uint64_t)msg.arg.pagefault.address); | |
1072 | break; | |
1073 | } | |
c4faeed2 | 1074 | |
7648297d | 1075 | rb_offset = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb)); |
00fa4fc8 | 1076 | trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, |
c4faeed2 | 1077 | qemu_ram_get_idstr(rb), |
575b0b33 AP |
1078 | rb_offset, |
1079 | msg.arg.pagefault.feat.ptid); | |
1080 | mark_postcopy_blocktime_begin( | |
1081 | (uintptr_t)(msg.arg.pagefault.address), | |
1082 | msg.arg.pagefault.feat.ptid, rb); | |
1083 | ||
3a7804c3 | 1084 | retry: |
00fa4fc8 DDAG |
1085 | /* |
1086 | * Send the request to the source - we want to request one | |
1087 | * of our host page sizes (which is >= TPS) | |
1088 | */ | |
9470c5e0 DH |
1089 | ret = postcopy_request_page(mis, rb, rb_offset, |
1090 | msg.arg.pagefault.address); | |
3a7804c3 PX |
1091 | if (ret) { |
1092 | /* May be network failure, try to wait for recovery */ | |
27dd21b4 PX |
1093 | postcopy_pause_fault_thread(mis); |
1094 | goto retry; | |
00fa4fc8 DDAG |
1095 | } |
1096 | } | |
c4faeed2 | 1097 | |
00fa4fc8 DDAG |
1098 | /* Now handle any requests from external processes on shared memory */ |
1099 | /* TODO: May need to handle devices deregistering during postcopy */ | |
1100 | for (index = 2; index < pfd_len && poll_result; index++) { | |
1101 | if (pfd[index].revents) { | |
1102 | struct PostCopyFD *pcfd = | |
1103 | &g_array_index(mis->postcopy_remote_fds, | |
1104 | struct PostCopyFD, index - 2); | |
1105 | ||
1106 | poll_result--; | |
1107 | if (pfd[index].revents & POLLERR) { | |
1108 | error_report("%s: POLLERR on poll %zd fd=%d", | |
1109 | __func__, index, pcfd->fd); | |
1110 | pfd[index].events = 0; | |
1111 | continue; | |
1112 | } | |
1113 | ||
1114 | ret = read(pcfd->fd, &msg, sizeof(msg)); | |
1115 | if (ret != sizeof(msg)) { | |
1116 | if (errno == EAGAIN) { | |
1117 | /* | |
1118 | * if a wake up happens on the other thread just after | |
1119 | * the poll, there is nothing to read. | |
1120 | */ | |
1121 | continue; | |
1122 | } | |
1123 | if (ret < 0) { | |
1124 | error_report("%s: Failed to read full userfault " | |
1125 | "message: %s (shared) revents=%d", | |
1126 | __func__, strerror(errno), | |
1127 | pfd[index].revents); | |
1128 | /*TODO: Could just disable this sharer */ | |
1129 | break; | |
1130 | } else { | |
1131 | error_report("%s: Read %d bytes from userfaultfd " | |
1132 | "expected %zd (shared)", | |
1133 | __func__, ret, sizeof(msg)); | |
1134 | /*TODO: Could just disable this sharer */ | |
1135 | break; /*Lost alignment,don't know what we'd read next*/ | |
1136 | } | |
1137 | } | |
1138 | if (msg.event != UFFD_EVENT_PAGEFAULT) { | |
1139 | error_report("%s: Read unexpected event %ud " | |
1140 | "from userfaultfd (shared)", | |
1141 | __func__, msg.event); | |
1142 | continue; /* It's not a page fault, shouldn't happen */ | |
1143 | } | |
1144 | /* Call the device handler registered with us */ | |
1145 | ret = pcfd->handler(pcfd, &msg); | |
1146 | if (ret) { | |
1147 | error_report("%s: Failed to resolve shared fault on %zd/%s", | |
1148 | __func__, index, pcfd->idstr); | |
1149 | /* TODO: Fail? Disable this sharer? */ | |
1150 | } | |
1151 | } | |
c4faeed2 DDAG |
1152 | } |
1153 | } | |
74637e6f | 1154 | rcu_unregister_thread(); |
c4faeed2 | 1155 | trace_postcopy_ram_fault_thread_exit(); |
fc6008f3 | 1156 | g_free(pfd); |
f0a227ad DDAG |
1157 | return NULL; |
1158 | } | |
1159 | ||
476ebf77 PX |
1160 | static int postcopy_temp_pages_setup(MigrationIncomingState *mis) |
1161 | { | |
77dadc3f PX |
1162 | PostcopyTmpPage *tmp_page; |
1163 | int err, i, channels; | |
1164 | void *temp_page; | |
1165 | ||
36f62f11 PX |
1166 | if (migrate_postcopy_preempt()) { |
1167 | /* If preemption enabled, need extra channel for urgent requests */ | |
1168 | mis->postcopy_channels = RAM_CHANNEL_MAX; | |
1169 | } else { | |
1170 | /* Both precopy/postcopy on the same channel */ | |
1171 | mis->postcopy_channels = 1; | |
1172 | } | |
77dadc3f PX |
1173 | |
1174 | channels = mis->postcopy_channels; | |
1175 | mis->postcopy_tmp_pages = g_malloc0_n(sizeof(PostcopyTmpPage), channels); | |
1176 | ||
1177 | for (i = 0; i < channels; i++) { | |
1178 | tmp_page = &mis->postcopy_tmp_pages[i]; | |
1179 | temp_page = mmap(NULL, mis->largest_page_size, PROT_READ | PROT_WRITE, | |
1180 | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); | |
1181 | if (temp_page == MAP_FAILED) { | |
1182 | err = errno; | |
1183 | error_report("%s: Failed to map postcopy_tmp_pages[%d]: %s", | |
1184 | __func__, i, strerror(err)); | |
1185 | /* Clean up will be done later */ | |
1186 | return -err; | |
1187 | } | |
1188 | tmp_page->tmp_huge_page = temp_page; | |
1189 | /* Initialize default states for each tmp page */ | |
1190 | postcopy_temp_page_reset(tmp_page); | |
476ebf77 PX |
1191 | } |
1192 | ||
1193 | /* | |
1194 | * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages | |
1195 | */ | |
1196 | mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size, | |
1197 | PROT_READ | PROT_WRITE, | |
1198 | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); | |
1199 | if (mis->postcopy_tmp_zero_page == MAP_FAILED) { | |
1200 | err = errno; | |
1201 | mis->postcopy_tmp_zero_page = NULL; | |
1202 | error_report("%s: Failed to map large zero page %s", | |
1203 | __func__, strerror(err)); | |
1204 | return -err; | |
1205 | } | |
1206 | ||
1207 | memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size); | |
1208 | ||
1209 | return 0; | |
1210 | } | |
1211 | ||
2a7eb148 | 1212 | int postcopy_ram_incoming_setup(MigrationIncomingState *mis) |
f0a227ad | 1213 | { |
74c38cf7 PX |
1214 | Error *local_err = NULL; |
1215 | ||
c4faeed2 | 1216 | /* Open the fd for the kernel to give us userfaults */ |
d5890ea0 | 1217 | mis->userfault_fd = uffd_open(O_CLOEXEC | O_NONBLOCK); |
c4faeed2 DDAG |
1218 | if (mis->userfault_fd == -1) { |
1219 | error_report("%s: Failed to open userfault fd: %s", __func__, | |
1220 | strerror(errno)); | |
1221 | return -1; | |
1222 | } | |
1223 | ||
1224 | /* | |
1225 | * Although the host check already tested the API, we need to | |
1226 | * do the check again as an ABI handshake on the new fd. | |
1227 | */ | |
74c38cf7 PX |
1228 | if (!ufd_check_and_apply(mis->userfault_fd, mis, &local_err)) { |
1229 | error_report_err(local_err); | |
c4faeed2 DDAG |
1230 | return -1; |
1231 | } | |
1232 | ||
1233 | /* Now an eventfd we use to tell the fault-thread to quit */ | |
64f615fe PX |
1234 | mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC); |
1235 | if (mis->userfault_event_fd == -1) { | |
1236 | error_report("%s: Opening userfault_event_fd: %s", __func__, | |
c4faeed2 DDAG |
1237 | strerror(errno)); |
1238 | close(mis->userfault_fd); | |
1239 | return -1; | |
1240 | } | |
1241 | ||
36f62f11 | 1242 | postcopy_thread_create(mis, &mis->fault_thread, "fault-default", |
095c12a4 | 1243 | postcopy_ram_fault_thread, QEMU_THREAD_JOINABLE); |
c4faeed2 | 1244 | mis->have_fault_thread = true; |
f0a227ad DDAG |
1245 | |
1246 | /* Mark so that we get notified of accesses to unwritten areas */ | |
fbd162e6 | 1247 | if (foreach_not_ignored_block(ram_block_enable_notify, mis)) { |
91b02dc7 | 1248 | error_report("ram_block_enable_notify failed"); |
f0a227ad DDAG |
1249 | return -1; |
1250 | } | |
1251 | ||
476ebf77 PX |
1252 | if (postcopy_temp_pages_setup(mis)) { |
1253 | /* Error dumped in the sub-function */ | |
3414322a WY |
1254 | return -1; |
1255 | } | |
1256 | ||
36f62f11 PX |
1257 | if (migrate_postcopy_preempt()) { |
1258 | /* | |
1259 | * This thread needs to be created after the temp pages because | |
1260 | * it'll fetch RAM_CHANNEL_POSTCOPY PostcopyTmpPage immediately. | |
1261 | */ | |
1262 | postcopy_thread_create(mis, &mis->postcopy_prio_thread, "fault-fast", | |
1263 | postcopy_preempt_thread, QEMU_THREAD_JOINABLE); | |
6621883f | 1264 | mis->preempt_thread_status = PREEMPT_THREAD_CREATED; |
36f62f11 PX |
1265 | } |
1266 | ||
c4faeed2 DDAG |
1267 | trace_postcopy_ram_enable_notify(); |
1268 | ||
f0a227ad DDAG |
1269 | return 0; |
1270 | } | |
1271 | ||
eef621c4 | 1272 | static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr, |
f9494614 | 1273 | void *from_addr, uint64_t pagesize, RAMBlock *rb) |
727b9d7e | 1274 | { |
eef621c4 | 1275 | int userfault_fd = mis->userfault_fd; |
f9494614 | 1276 | int ret; |
eef621c4 | 1277 | |
727b9d7e AP |
1278 | if (from_addr) { |
1279 | struct uffdio_copy copy_struct; | |
1280 | copy_struct.dst = (uint64_t)(uintptr_t)host_addr; | |
1281 | copy_struct.src = (uint64_t)(uintptr_t)from_addr; | |
1282 | copy_struct.len = pagesize; | |
1283 | copy_struct.mode = 0; | |
f9494614 | 1284 | ret = ioctl(userfault_fd, UFFDIO_COPY, ©_struct); |
727b9d7e AP |
1285 | } else { |
1286 | struct uffdio_zeropage zero_struct; | |
1287 | zero_struct.range.start = (uint64_t)(uintptr_t)host_addr; | |
1288 | zero_struct.range.len = pagesize; | |
1289 | zero_struct.mode = 0; | |
f9494614 AP |
1290 | ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct); |
1291 | } | |
1292 | if (!ret) { | |
8f8bfffc | 1293 | qemu_mutex_lock(&mis->page_request_mutex); |
f9494614 AP |
1294 | ramblock_recv_bitmap_set_range(rb, host_addr, |
1295 | pagesize / qemu_target_page_size()); | |
8f8bfffc PX |
1296 | /* |
1297 | * If this page resolves a page fault for a previous recorded faulted | |
1298 | * address, take a special note to maintain the requested page list. | |
1299 | */ | |
1300 | if (g_tree_lookup(mis->page_requested, host_addr)) { | |
1301 | g_tree_remove(mis->page_requested, host_addr); | |
cf02f29e PX |
1302 | int left_pages = qatomic_dec_fetch(&mis->page_requested_count); |
1303 | ||
8f8bfffc | 1304 | trace_postcopy_page_req_del(host_addr, mis->page_requested_count); |
cf02f29e PX |
1305 | /* Order the update of count and read of preempt status */ |
1306 | smp_mb(); | |
1307 | if (mis->preempt_thread_status == PREEMPT_THREAD_QUIT && | |
1308 | left_pages == 0) { | |
1309 | /* | |
1310 | * This probably means the main thread is waiting for us. | |
1311 | * Notify that we've finished receiving the last requested | |
1312 | * page. | |
1313 | */ | |
1314 | qemu_cond_signal(&mis->page_request_cond); | |
1315 | } | |
8f8bfffc PX |
1316 | } |
1317 | qemu_mutex_unlock(&mis->page_request_mutex); | |
575b0b33 | 1318 | mark_postcopy_blocktime_end((uintptr_t)host_addr); |
727b9d7e | 1319 | } |
f9494614 | 1320 | return ret; |
727b9d7e AP |
1321 | } |
1322 | ||
d488b349 DDAG |
1323 | int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset) |
1324 | { | |
1325 | int i; | |
1326 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1327 | GArray *pcrfds = mis->postcopy_remote_fds; | |
1328 | ||
1329 | for (i = 0; i < pcrfds->len; i++) { | |
1330 | struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); | |
1331 | int ret = cur->waker(cur, rb, offset); | |
1332 | if (ret) { | |
1333 | return ret; | |
1334 | } | |
1335 | } | |
1336 | return 0; | |
1337 | } | |
1338 | ||
696ed9a9 DDAG |
1339 | /* |
1340 | * Place a host page (from) at (host) atomically | |
1341 | * returns 0 on success | |
1342 | */ | |
df9ff5e1 | 1343 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
8be4620b | 1344 | RAMBlock *rb) |
696ed9a9 | 1345 | { |
8be4620b | 1346 | size_t pagesize = qemu_ram_pagesize(rb); |
696ed9a9 | 1347 | |
696ed9a9 DDAG |
1348 | /* copy also acks to the kernel waking the stalled thread up |
1349 | * TODO: We can inhibit that ack and only do it if it was requested | |
1350 | * which would be slightly cheaper, but we'd have to be careful | |
1351 | * of the order of updating our page state. | |
1352 | */ | |
eef621c4 | 1353 | if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) { |
696ed9a9 | 1354 | int e = errno; |
df9ff5e1 DDAG |
1355 | error_report("%s: %s copy host: %p from: %p (size: %zd)", |
1356 | __func__, strerror(e), host, from, pagesize); | |
696ed9a9 DDAG |
1357 | |
1358 | return -e; | |
1359 | } | |
1360 | ||
1361 | trace_postcopy_place_page(host); | |
dedfb4b2 DDAG |
1362 | return postcopy_notify_shared_wake(rb, |
1363 | qemu_ram_block_host_offset(rb, host)); | |
696ed9a9 DDAG |
1364 | } |
1365 | ||
1366 | /* | |
1367 | * Place a zero page at (host) atomically | |
1368 | * returns 0 on success | |
1369 | */ | |
df9ff5e1 | 1370 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
8be4620b | 1371 | RAMBlock *rb) |
696ed9a9 | 1372 | { |
2ce16640 | 1373 | size_t pagesize = qemu_ram_pagesize(rb); |
df9ff5e1 | 1374 | trace_postcopy_place_page_zero(host); |
696ed9a9 | 1375 | |
2ce16640 DDAG |
1376 | /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE |
1377 | * but it's not available for everything (e.g. hugetlbpages) | |
1378 | */ | |
1379 | if (qemu_ram_is_uf_zeroable(rb)) { | |
eef621c4 | 1380 | if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) { |
df9ff5e1 DDAG |
1381 | int e = errno; |
1382 | error_report("%s: %s zero host: %p", | |
1383 | __func__, strerror(e), host); | |
696ed9a9 | 1384 | |
df9ff5e1 DDAG |
1385 | return -e; |
1386 | } | |
dedfb4b2 DDAG |
1387 | return postcopy_notify_shared_wake(rb, |
1388 | qemu_ram_block_host_offset(rb, | |
1389 | host)); | |
df9ff5e1 | 1390 | } else { |
6629890d | 1391 | return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb); |
696ed9a9 | 1392 | } |
696ed9a9 DDAG |
1393 | } |
1394 | ||
eb59db53 DDAG |
1395 | #else |
1396 | /* No target OS support, stubs just fail */ | |
65ace060 AP |
1397 | void fill_destination_postcopy_migration_info(MigrationInfo *info) |
1398 | { | |
1399 | } | |
1400 | ||
74c38cf7 | 1401 | bool postcopy_ram_supported_by_host(MigrationIncomingState *mis, Error **errp) |
eb59db53 DDAG |
1402 | { |
1403 | error_report("%s: No OS support", __func__); | |
1404 | return false; | |
1405 | } | |
1406 | ||
c136180c | 1407 | int postcopy_ram_incoming_init(MigrationIncomingState *mis) |
1caddf8a DDAG |
1408 | { |
1409 | error_report("postcopy_ram_incoming_init: No OS support"); | |
1410 | return -1; | |
1411 | } | |
1412 | ||
1413 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
1414 | { | |
1415 | assert(0); | |
1416 | return -1; | |
1417 | } | |
1418 | ||
f9527107 DDAG |
1419 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) |
1420 | { | |
1421 | assert(0); | |
1422 | return -1; | |
1423 | } | |
1424 | ||
c188c539 MT |
1425 | int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, |
1426 | uint64_t client_addr, uint64_t rb_offset) | |
1427 | { | |
1428 | assert(0); | |
1429 | return -1; | |
1430 | } | |
1431 | ||
2a7eb148 | 1432 | int postcopy_ram_incoming_setup(MigrationIncomingState *mis) |
f0a227ad DDAG |
1433 | { |
1434 | assert(0); | |
1435 | return -1; | |
1436 | } | |
696ed9a9 | 1437 | |
df9ff5e1 | 1438 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
8be4620b | 1439 | RAMBlock *rb) |
696ed9a9 DDAG |
1440 | { |
1441 | assert(0); | |
1442 | return -1; | |
1443 | } | |
1444 | ||
df9ff5e1 | 1445 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
8be4620b | 1446 | RAMBlock *rb) |
696ed9a9 DDAG |
1447 | { |
1448 | assert(0); | |
1449 | return -1; | |
1450 | } | |
1451 | ||
5efc3564 DDAG |
1452 | int postcopy_wake_shared(struct PostCopyFD *pcfd, |
1453 | uint64_t client_addr, | |
1454 | RAMBlock *rb) | |
1455 | { | |
1456 | assert(0); | |
1457 | return -1; | |
1458 | } | |
eb59db53 DDAG |
1459 | #endif |
1460 | ||
e0b266f0 | 1461 | /* ------------------------------------------------------------------------- */ |
77dadc3f PX |
1462 | void postcopy_temp_page_reset(PostcopyTmpPage *tmp_page) |
1463 | { | |
1464 | tmp_page->target_pages = 0; | |
1465 | tmp_page->host_addr = NULL; | |
1466 | /* | |
1467 | * This is set to true when reset, and cleared as long as we received any | |
1468 | * of the non-zero small page within this huge page. | |
1469 | */ | |
1470 | tmp_page->all_zero = true; | |
1471 | } | |
e0b266f0 | 1472 | |
9ab7ef9b PX |
1473 | void postcopy_fault_thread_notify(MigrationIncomingState *mis) |
1474 | { | |
1475 | uint64_t tmp64 = 1; | |
1476 | ||
1477 | /* | |
1478 | * Wakeup the fault_thread. It's an eventfd that should currently | |
1479 | * be at 0, we're going to increment it to 1 | |
1480 | */ | |
1481 | if (write(mis->userfault_event_fd, &tmp64, 8) != 8) { | |
1482 | /* Not much we can do here, but may as well report it */ | |
1483 | error_report("%s: incrementing failed: %s", __func__, | |
1484 | strerror(errno)); | |
1485 | } | |
1486 | } | |
1487 | ||
e0b266f0 DDAG |
1488 | /** |
1489 | * postcopy_discard_send_init: Called at the start of each RAMBlock before | |
1490 | * asking to discard individual ranges. | |
1491 | * | |
1492 | * @ms: The current migration state. | |
810cf2bb | 1493 | * @offset: the bitmap offset of the named RAMBlock in the migration bitmap. |
e0b266f0 | 1494 | * @name: RAMBlock that discards will operate on. |
e0b266f0 | 1495 | */ |
810cf2bb WY |
1496 | static PostcopyDiscardState pds = {0}; |
1497 | void postcopy_discard_send_init(MigrationState *ms, const char *name) | |
e0b266f0 | 1498 | { |
810cf2bb WY |
1499 | pds.ramblock_name = name; |
1500 | pds.cur_entry = 0; | |
1501 | pds.nsentwords = 0; | |
1502 | pds.nsentcmds = 0; | |
e0b266f0 DDAG |
1503 | } |
1504 | ||
1505 | /** | |
1506 | * postcopy_discard_send_range: Called by the bitmap code for each chunk to | |
1507 | * discard. May send a discard message, may just leave it queued to | |
1508 | * be sent later. | |
1509 | * | |
1510 | * @ms: Current migration state. | |
e0b266f0 DDAG |
1511 | * @start,@length: a range of pages in the migration bitmap in the |
1512 | * RAM block passed to postcopy_discard_send_init() (length=1 is one page) | |
1513 | */ | |
810cf2bb WY |
1514 | void postcopy_discard_send_range(MigrationState *ms, unsigned long start, |
1515 | unsigned long length) | |
e0b266f0 | 1516 | { |
20afaed9 | 1517 | size_t tp_size = qemu_target_page_size(); |
e0b266f0 | 1518 | /* Convert to byte offsets within the RAM block */ |
810cf2bb WY |
1519 | pds.start_list[pds.cur_entry] = start * tp_size; |
1520 | pds.length_list[pds.cur_entry] = length * tp_size; | |
1521 | trace_postcopy_discard_send_range(pds.ramblock_name, start, length); | |
1522 | pds.cur_entry++; | |
1523 | pds.nsentwords++; | |
e0b266f0 | 1524 | |
810cf2bb | 1525 | if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) { |
e0b266f0 | 1526 | /* Full set, ship it! */ |
89a02a9f | 1527 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
810cf2bb WY |
1528 | pds.ramblock_name, |
1529 | pds.cur_entry, | |
1530 | pds.start_list, | |
1531 | pds.length_list); | |
1532 | pds.nsentcmds++; | |
1533 | pds.cur_entry = 0; | |
e0b266f0 DDAG |
1534 | } |
1535 | } | |
1536 | ||
1537 | /** | |
1538 | * postcopy_discard_send_finish: Called at the end of each RAMBlock by the | |
1539 | * bitmap code. Sends any outstanding discard messages, frees the PDS | |
1540 | * | |
1541 | * @ms: Current migration state. | |
e0b266f0 | 1542 | */ |
810cf2bb | 1543 | void postcopy_discard_send_finish(MigrationState *ms) |
e0b266f0 DDAG |
1544 | { |
1545 | /* Anything unsent? */ | |
810cf2bb | 1546 | if (pds.cur_entry) { |
89a02a9f | 1547 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
810cf2bb WY |
1548 | pds.ramblock_name, |
1549 | pds.cur_entry, | |
1550 | pds.start_list, | |
1551 | pds.length_list); | |
1552 | pds.nsentcmds++; | |
e0b266f0 DDAG |
1553 | } |
1554 | ||
810cf2bb WY |
1555 | trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords, |
1556 | pds.nsentcmds); | |
e0b266f0 | 1557 | } |
bac3b212 JQ |
1558 | |
1559 | /* | |
1560 | * Current state of incoming postcopy; note this is not part of | |
1561 | * MigrationIncomingState since it's state is used during cleanup | |
1562 | * at the end as MIS is being freed. | |
1563 | */ | |
1564 | static PostcopyState incoming_postcopy_state; | |
1565 | ||
1566 | PostcopyState postcopy_state_get(void) | |
1567 | { | |
4592eaf3 | 1568 | return qatomic_load_acquire(&incoming_postcopy_state); |
bac3b212 JQ |
1569 | } |
1570 | ||
1571 | /* Set the state and return the old state */ | |
1572 | PostcopyState postcopy_state_set(PostcopyState new_state) | |
1573 | { | |
d73415a3 | 1574 | return qatomic_xchg(&incoming_postcopy_state, new_state); |
bac3b212 | 1575 | } |
00fa4fc8 DDAG |
1576 | |
1577 | /* Register a handler for external shared memory postcopy | |
1578 | * called on the destination. | |
1579 | */ | |
1580 | void postcopy_register_shared_ufd(struct PostCopyFD *pcfd) | |
1581 | { | |
1582 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1583 | ||
1584 | mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds, | |
1585 | *pcfd); | |
1586 | } | |
1587 | ||
1588 | /* Unregister a handler for external shared memory postcopy | |
1589 | */ | |
1590 | void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd) | |
1591 | { | |
1592 | guint i; | |
1593 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1594 | GArray *pcrfds = mis->postcopy_remote_fds; | |
1595 | ||
56559980 JQ |
1596 | if (!pcrfds) { |
1597 | /* migration has already finished and freed the array */ | |
1598 | return; | |
1599 | } | |
00fa4fc8 DDAG |
1600 | for (i = 0; i < pcrfds->len; i++) { |
1601 | struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); | |
1602 | if (cur->fd == pcfd->fd) { | |
1603 | mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i); | |
1604 | return; | |
1605 | } | |
1606 | } | |
1607 | } | |
36f62f11 | 1608 | |
6720c2b3 | 1609 | void postcopy_preempt_new_channel(MigrationIncomingState *mis, QEMUFile *file) |
36f62f11 PX |
1610 | { |
1611 | /* | |
1612 | * The new loading channel has its own threads, so it needs to be | |
1613 | * blocked too. It's by default true, just be explicit. | |
1614 | */ | |
1615 | qemu_file_set_blocking(file, true); | |
1616 | mis->postcopy_qemufile_dst = file; | |
5655aab0 | 1617 | qemu_sem_post(&mis->postcopy_qemufile_dst_done); |
36f62f11 | 1618 | trace_postcopy_preempt_new_channel(); |
36f62f11 PX |
1619 | } |
1620 | ||
f0afaf6c PX |
1621 | /* |
1622 | * Setup the postcopy preempt channel with the IOC. If ERROR is specified, | |
1623 | * setup the error instead. This helper will free the ERROR if specified. | |
1624 | */ | |
d0edb8a1 | 1625 | static void |
f0afaf6c PX |
1626 | postcopy_preempt_send_channel_done(MigrationState *s, |
1627 | QIOChannel *ioc, Error *local_err) | |
36f62f11 | 1628 | { |
f0afaf6c | 1629 | if (local_err) { |
d0edb8a1 PX |
1630 | migrate_set_error(s, local_err); |
1631 | error_free(local_err); | |
1632 | } else { | |
1633 | migration_ioc_register_yank(ioc); | |
1634 | s->postcopy_qemufile_src = qemu_file_new_output(ioc); | |
1635 | trace_postcopy_preempt_new_channel(); | |
1636 | } | |
1637 | ||
1638 | /* | |
1639 | * Kick the waiter in all cases. The waiter should check upon | |
1640 | * postcopy_qemufile_src to know whether it failed or not. | |
1641 | */ | |
1642 | qemu_sem_post(&s->postcopy_qemufile_src_sem); | |
f0afaf6c PX |
1643 | } |
1644 | ||
1645 | static void | |
1646 | postcopy_preempt_tls_handshake(QIOTask *task, gpointer opaque) | |
1647 | { | |
1648 | g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task)); | |
1649 | MigrationState *s = opaque; | |
1650 | Error *local_err = NULL; | |
1651 | ||
1652 | qio_task_propagate_error(task, &local_err); | |
1653 | postcopy_preempt_send_channel_done(s, ioc, local_err); | |
1654 | } | |
1655 | ||
1656 | static void | |
1657 | postcopy_preempt_send_channel_new(QIOTask *task, gpointer opaque) | |
1658 | { | |
1659 | g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task)); | |
1660 | MigrationState *s = opaque; | |
1661 | QIOChannelTLS *tioc; | |
1662 | Error *local_err = NULL; | |
1663 | ||
1664 | if (qio_task_propagate_error(task, &local_err)) { | |
1665 | goto out; | |
1666 | } | |
1667 | ||
1668 | if (migrate_channel_requires_tls_upgrade(ioc)) { | |
0deb7e9b | 1669 | tioc = migration_tls_client_create(ioc, s->hostname, &local_err); |
f0afaf6c PX |
1670 | if (!tioc) { |
1671 | goto out; | |
1672 | } | |
1673 | trace_postcopy_preempt_tls_handshake(); | |
1674 | qio_channel_set_name(QIO_CHANNEL(tioc), "migration-tls-preempt"); | |
1675 | qio_channel_tls_handshake(tioc, postcopy_preempt_tls_handshake, | |
1676 | s, NULL, NULL); | |
1677 | /* Setup the channel until TLS handshake finished */ | |
1678 | return; | |
1679 | } | |
1680 | ||
1681 | out: | |
1682 | /* This handles both good and error cases */ | |
1683 | postcopy_preempt_send_channel_done(s, ioc, local_err); | |
d0edb8a1 | 1684 | } |
36f62f11 | 1685 | |
5655aab0 PX |
1686 | /* |
1687 | * This function will kick off an async task to establish the preempt | |
1688 | * channel, and wait until the connection setup completed. Returns 0 if | |
1689 | * channel established, -1 for error. | |
1690 | */ | |
1691 | int postcopy_preempt_establish_channel(MigrationState *s) | |
d0edb8a1 PX |
1692 | { |
1693 | /* If preempt not enabled, no need to wait */ | |
1694 | if (!migrate_postcopy_preempt()) { | |
1695 | return 0; | |
1696 | } | |
1697 | ||
06064a67 PX |
1698 | /* |
1699 | * Kick off async task to establish preempt channel. Only do so with | |
1700 | * 8.0+ machines, because 7.1/7.2 require the channel to be created in | |
1701 | * setup phase of migration (even if racy in an unreliable network). | |
1702 | */ | |
1703 | if (!s->preempt_pre_7_2) { | |
1704 | postcopy_preempt_setup(s); | |
1705 | } | |
5655aab0 | 1706 | |
d0edb8a1 PX |
1707 | /* |
1708 | * We need the postcopy preempt channel to be established before | |
1709 | * starting doing anything. | |
1710 | */ | |
1711 | qemu_sem_wait(&s->postcopy_qemufile_src_sem); | |
1712 | ||
1713 | return s->postcopy_qemufile_src ? 0 : -1; | |
1714 | } | |
1715 | ||
fc063a7b | 1716 | void postcopy_preempt_setup(MigrationState *s) |
d0edb8a1 | 1717 | { |
d0edb8a1 PX |
1718 | /* Kick an async task to connect */ |
1719 | socket_send_channel_create(postcopy_preempt_send_channel_new, s); | |
36f62f11 PX |
1720 | } |
1721 | ||
60bb3c58 PX |
1722 | static void postcopy_pause_ram_fast_load(MigrationIncomingState *mis) |
1723 | { | |
1724 | trace_postcopy_pause_fast_load(); | |
1725 | qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex); | |
1726 | qemu_sem_wait(&mis->postcopy_pause_sem_fast_load); | |
1727 | qemu_mutex_lock(&mis->postcopy_prio_thread_mutex); | |
1728 | trace_postcopy_pause_fast_load_continued(); | |
1729 | } | |
1730 | ||
6621883f PX |
1731 | static bool preempt_thread_should_run(MigrationIncomingState *mis) |
1732 | { | |
1733 | return mis->preempt_thread_status != PREEMPT_THREAD_QUIT; | |
1734 | } | |
1735 | ||
36f62f11 PX |
1736 | void *postcopy_preempt_thread(void *opaque) |
1737 | { | |
1738 | MigrationIncomingState *mis = opaque; | |
1739 | int ret; | |
1740 | ||
1741 | trace_postcopy_preempt_thread_entry(); | |
1742 | ||
1743 | rcu_register_thread(); | |
1744 | ||
1745 | qemu_sem_post(&mis->thread_sync_sem); | |
1746 | ||
a5d35dc7 PX |
1747 | /* |
1748 | * The preempt channel is established in asynchronous way. Wait | |
1749 | * for its completion. | |
1750 | */ | |
1751 | qemu_sem_wait(&mis->postcopy_qemufile_dst_done); | |
1752 | ||
36f62f11 | 1753 | /* Sending RAM_SAVE_FLAG_EOS to terminate this thread */ |
60bb3c58 | 1754 | qemu_mutex_lock(&mis->postcopy_prio_thread_mutex); |
6621883f | 1755 | while (preempt_thread_should_run(mis)) { |
60bb3c58 PX |
1756 | ret = ram_load_postcopy(mis->postcopy_qemufile_dst, |
1757 | RAM_CHANNEL_POSTCOPY); | |
1758 | /* If error happened, go into recovery routine */ | |
6621883f | 1759 | if (ret && preempt_thread_should_run(mis)) { |
60bb3c58 PX |
1760 | postcopy_pause_ram_fast_load(mis); |
1761 | } else { | |
1762 | /* We're done */ | |
1763 | break; | |
1764 | } | |
1765 | } | |
1766 | qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex); | |
36f62f11 PX |
1767 | |
1768 | rcu_unregister_thread(); | |
1769 | ||
1770 | trace_postcopy_preempt_thread_exit(); | |
1771 | ||
60bb3c58 | 1772 | return NULL; |
36f62f11 | 1773 | } |