]>
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__) | |
105 | ||
c4faeed2 | 106 | #include <poll.h> |
eb59db53 DDAG |
107 | #include <sys/ioctl.h> |
108 | #include <sys/syscall.h> | |
eb59db53 DDAG |
109 | #include <asm/types.h> /* for __u64 */ |
110 | #endif | |
111 | ||
d8b9d771 MF |
112 | #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD) |
113 | #include <sys/eventfd.h> | |
eb59db53 DDAG |
114 | #include <linux/userfaultfd.h> |
115 | ||
2a4c42f1 AP |
116 | typedef struct PostcopyBlocktimeContext { |
117 | /* time when page fault initiated per vCPU */ | |
118 | uint32_t *page_fault_vcpu_time; | |
119 | /* page address per vCPU */ | |
120 | uintptr_t *vcpu_addr; | |
121 | uint32_t total_blocktime; | |
122 | /* blocktime per vCPU */ | |
123 | uint32_t *vcpu_blocktime; | |
124 | /* point in time when last page fault was initiated */ | |
125 | uint32_t last_begin; | |
126 | /* number of vCPU are suspended */ | |
127 | int smp_cpus_down; | |
128 | uint64_t start_time; | |
129 | ||
130 | /* | |
131 | * Handler for exit event, necessary for | |
132 | * releasing whole blocktime_ctx | |
133 | */ | |
134 | Notifier exit_notifier; | |
135 | } PostcopyBlocktimeContext; | |
136 | ||
137 | static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx) | |
138 | { | |
139 | g_free(ctx->page_fault_vcpu_time); | |
140 | g_free(ctx->vcpu_addr); | |
141 | g_free(ctx->vcpu_blocktime); | |
142 | g_free(ctx); | |
143 | } | |
144 | ||
145 | static void migration_exit_cb(Notifier *n, void *data) | |
146 | { | |
147 | PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext, | |
148 | exit_notifier); | |
149 | destroy_blocktime_context(ctx); | |
150 | } | |
151 | ||
152 | static struct PostcopyBlocktimeContext *blocktime_context_new(void) | |
153 | { | |
5cc8767d LX |
154 | MachineState *ms = MACHINE(qdev_get_machine()); |
155 | unsigned int smp_cpus = ms->smp.cpus; | |
2a4c42f1 AP |
156 | PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1); |
157 | ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus); | |
158 | ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus); | |
159 | ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus); | |
160 | ||
161 | ctx->exit_notifier.notify = migration_exit_cb; | |
162 | ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
163 | qemu_add_exit_notifier(&ctx->exit_notifier); | |
164 | return ctx; | |
165 | } | |
ca6011c2 | 166 | |
65ace060 AP |
167 | static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx) |
168 | { | |
5cc8767d | 169 | MachineState *ms = MACHINE(qdev_get_machine()); |
54aa3de7 | 170 | uint32List *list = NULL; |
65ace060 AP |
171 | int i; |
172 | ||
5cc8767d | 173 | for (i = ms->smp.cpus - 1; i >= 0; i--) { |
54aa3de7 | 174 | QAPI_LIST_PREPEND(list, ctx->vcpu_blocktime[i]); |
65ace060 AP |
175 | } |
176 | ||
177 | return list; | |
178 | } | |
179 | ||
180 | /* | |
181 | * This function just populates MigrationInfo from postcopy's | |
182 | * blocktime context. It will not populate MigrationInfo, | |
183 | * unless postcopy-blocktime capability was set. | |
184 | * | |
185 | * @info: pointer to MigrationInfo to populate | |
186 | */ | |
187 | void fill_destination_postcopy_migration_info(MigrationInfo *info) | |
188 | { | |
189 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
190 | PostcopyBlocktimeContext *bc = mis->blocktime_ctx; | |
191 | ||
192 | if (!bc) { | |
193 | return; | |
194 | } | |
195 | ||
196 | info->has_postcopy_blocktime = true; | |
197 | info->postcopy_blocktime = bc->total_blocktime; | |
198 | info->has_postcopy_vcpu_blocktime = true; | |
199 | info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc); | |
200 | } | |
201 | ||
202 | static uint32_t get_postcopy_total_blocktime(void) | |
203 | { | |
204 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
205 | PostcopyBlocktimeContext *bc = mis->blocktime_ctx; | |
206 | ||
207 | if (!bc) { | |
208 | return 0; | |
209 | } | |
210 | ||
211 | return bc->total_blocktime; | |
212 | } | |
213 | ||
54ae0886 AP |
214 | /** |
215 | * receive_ufd_features: check userfault fd features, to request only supported | |
216 | * features in the future. | |
217 | * | |
218 | * Returns: true on success | |
219 | * | |
220 | * __NR_userfaultfd - should be checked before | |
221 | * @features: out parameter will contain uffdio_api.features provided by kernel | |
222 | * in case of success | |
223 | */ | |
224 | static bool receive_ufd_features(uint64_t *features) | |
eb59db53 | 225 | { |
54ae0886 AP |
226 | struct uffdio_api api_struct = {0}; |
227 | int ufd; | |
228 | bool ret = true; | |
229 | ||
d5890ea0 | 230 | ufd = uffd_open(O_CLOEXEC); |
54ae0886 | 231 | if (ufd == -1) { |
d5890ea0 | 232 | error_report("%s: uffd_open() failed: %s", __func__, strerror(errno)); |
54ae0886 AP |
233 | return false; |
234 | } | |
eb59db53 | 235 | |
54ae0886 | 236 | /* ask features */ |
eb59db53 DDAG |
237 | api_struct.api = UFFD_API; |
238 | api_struct.features = 0; | |
239 | if (ioctl(ufd, UFFDIO_API, &api_struct)) { | |
5553499f | 240 | error_report("%s: UFFDIO_API failed: %s", __func__, |
eb59db53 | 241 | strerror(errno)); |
54ae0886 AP |
242 | ret = false; |
243 | goto release_ufd; | |
244 | } | |
245 | ||
246 | *features = api_struct.features; | |
247 | ||
248 | release_ufd: | |
249 | close(ufd); | |
250 | return ret; | |
251 | } | |
252 | ||
253 | /** | |
254 | * request_ufd_features: this function should be called only once on a newly | |
255 | * opened ufd, subsequent calls will lead to error. | |
256 | * | |
3a4452d8 | 257 | * Returns: true on success |
54ae0886 AP |
258 | * |
259 | * @ufd: fd obtained from userfaultfd syscall | |
260 | * @features: bit mask see UFFD_API_FEATURES | |
261 | */ | |
262 | static bool request_ufd_features(int ufd, uint64_t features) | |
263 | { | |
264 | struct uffdio_api api_struct = {0}; | |
265 | uint64_t ioctl_mask; | |
266 | ||
267 | api_struct.api = UFFD_API; | |
268 | api_struct.features = features; | |
269 | if (ioctl(ufd, UFFDIO_API, &api_struct)) { | |
270 | error_report("%s failed: UFFDIO_API failed: %s", __func__, | |
271 | strerror(errno)); | |
eb59db53 DDAG |
272 | return false; |
273 | } | |
274 | ||
275 | ioctl_mask = (__u64)1 << _UFFDIO_REGISTER | | |
276 | (__u64)1 << _UFFDIO_UNREGISTER; | |
277 | if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) { | |
278 | error_report("Missing userfault features: %" PRIx64, | |
279 | (uint64_t)(~api_struct.ioctls & ioctl_mask)); | |
280 | return false; | |
281 | } | |
282 | ||
54ae0886 AP |
283 | return true; |
284 | } | |
285 | ||
74c38cf7 PX |
286 | static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis, |
287 | Error **errp) | |
54ae0886 AP |
288 | { |
289 | uint64_t asked_features = 0; | |
290 | static uint64_t supported_features; | |
291 | ||
74c38cf7 | 292 | ERRP_GUARD(); |
54ae0886 AP |
293 | /* |
294 | * it's not possible to | |
295 | * request UFFD_API twice per one fd | |
296 | * userfault fd features is persistent | |
297 | */ | |
298 | if (!supported_features) { | |
299 | if (!receive_ufd_features(&supported_features)) { | |
74c38cf7 | 300 | error_setg(errp, "Userfault feature detection failed"); |
54ae0886 AP |
301 | return false; |
302 | } | |
303 | } | |
304 | ||
2a4c42f1 | 305 | #ifdef UFFD_FEATURE_THREAD_ID |
2d1c37c6 | 306 | if (UFFD_FEATURE_THREAD_ID & supported_features) { |
2a4c42f1 | 307 | asked_features |= UFFD_FEATURE_THREAD_ID; |
2d1c37c6 PX |
308 | if (migrate_postcopy_blocktime()) { |
309 | if (!mis->blocktime_ctx) { | |
310 | mis->blocktime_ctx = blocktime_context_new(); | |
311 | } | |
312 | } | |
2a4c42f1 AP |
313 | } |
314 | #endif | |
315 | ||
54ae0886 AP |
316 | /* |
317 | * request features, even if asked_features is 0, due to | |
318 | * kernel expects UFFD_API before UFFDIO_REGISTER, per | |
319 | * userfault file descriptor | |
320 | */ | |
321 | if (!request_ufd_features(ufd, asked_features)) { | |
74c38cf7 | 322 | error_setg(errp, "Failed features %" PRIu64, asked_features); |
54ae0886 AP |
323 | return false; |
324 | } | |
325 | ||
8e3b0cbb | 326 | if (qemu_real_host_page_size() != ram_pagesize_summary()) { |
7e8cafb7 DDAG |
327 | bool have_hp = false; |
328 | /* We've got a huge page */ | |
329 | #ifdef UFFD_FEATURE_MISSING_HUGETLBFS | |
54ae0886 | 330 | have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS; |
7e8cafb7 DDAG |
331 | #endif |
332 | if (!have_hp) { | |
74c38cf7 PX |
333 | error_setg(errp, |
334 | "Userfault on this host does not support huge pages"); | |
7e8cafb7 DDAG |
335 | return false; |
336 | } | |
337 | } | |
eb59db53 DDAG |
338 | return true; |
339 | } | |
340 | ||
8679638b DDAG |
341 | /* Callback from postcopy_ram_supported_by_host block iterator. |
342 | */ | |
74c38cf7 | 343 | static int test_ramblock_postcopiable(RAMBlock *rb, Error **errp) |
8679638b | 344 | { |
754cb9c0 YK |
345 | const char *block_name = qemu_ram_get_idstr(rb); |
346 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
5d214a92 | 347 | size_t pagesize = qemu_ram_pagesize(rb); |
ae30b9b2 | 348 | QemuFsType fs; |
5d214a92 | 349 | |
5d214a92 | 350 | if (length % pagesize) { |
74c38cf7 PX |
351 | error_setg(errp, |
352 | "Postcopy requires RAM blocks to be a page size multiple," | |
353 | " block %s is 0x" RAM_ADDR_FMT " bytes with a " | |
354 | "page size of 0x%zx", block_name, length, pagesize); | |
5d214a92 DDAG |
355 | return 1; |
356 | } | |
ae30b9b2 PX |
357 | |
358 | if (rb->fd >= 0) { | |
359 | fs = qemu_fd_getfs(rb->fd); | |
360 | if (fs != QEMU_FS_TYPE_TMPFS && fs != QEMU_FS_TYPE_HUGETLBFS) { | |
74c38cf7 PX |
361 | error_setg(errp, |
362 | "Host backend files need to be TMPFS or HUGETLBFS only"); | |
ae30b9b2 PX |
363 | return 1; |
364 | } | |
365 | } | |
366 | ||
8679638b DDAG |
367 | return 0; |
368 | } | |
369 | ||
58b7c17e DDAG |
370 | /* |
371 | * Note: This has the side effect of munlock'ing all of RAM, that's | |
372 | * normally fine since if the postcopy succeeds it gets turned back on at the | |
373 | * end. | |
374 | */ | |
74c38cf7 | 375 | bool postcopy_ram_supported_by_host(MigrationIncomingState *mis, Error **errp) |
eb59db53 | 376 | { |
8e3b0cbb | 377 | long pagesize = qemu_real_host_page_size(); |
eb59db53 DDAG |
378 | int ufd = -1; |
379 | bool ret = false; /* Error unless we change it */ | |
380 | void *testarea = NULL; | |
381 | struct uffdio_register reg_struct; | |
382 | struct uffdio_range range_struct; | |
383 | uint64_t feature_mask; | |
ae30b9b2 | 384 | RAMBlock *block; |
eb59db53 | 385 | |
74c38cf7 | 386 | ERRP_GUARD(); |
20afaed9 | 387 | if (qemu_target_page_size() > pagesize) { |
74c38cf7 | 388 | error_setg(errp, "Target page size bigger than host page size"); |
eb59db53 DDAG |
389 | goto out; |
390 | } | |
391 | ||
d5890ea0 | 392 | ufd = uffd_open(O_CLOEXEC); |
eb59db53 | 393 | if (ufd == -1) { |
74c38cf7 | 394 | error_setg(errp, "Userfaultfd not available: %s", strerror(errno)); |
eb59db53 DDAG |
395 | goto out; |
396 | } | |
397 | ||
1693c64c | 398 | /* Give devices a chance to object */ |
74c38cf7 | 399 | if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, errp)) { |
1693c64c DDAG |
400 | goto out; |
401 | } | |
402 | ||
eb59db53 | 403 | /* Version and features check */ |
74c38cf7 | 404 | if (!ufd_check_and_apply(ufd, mis, errp)) { |
eb59db53 DDAG |
405 | goto out; |
406 | } | |
407 | ||
ae30b9b2 PX |
408 | /* |
409 | * We don't support postcopy with some type of ramblocks. | |
410 | * | |
f161c88a | 411 | * NOTE: we explicitly ignored migrate_ram_is_ignored() instead we checked |
ae30b9b2 PX |
412 | * all possible ramblocks. This is because this function can be called |
413 | * when creating the migration object, during the phase RAM_MIGRATABLE | |
414 | * is not even properly set for all the ramblocks. | |
415 | * | |
416 | * A side effect of this is we'll also check against RAM_SHARED | |
417 | * ramblocks even if migrate_ignore_shared() is set (in which case | |
418 | * we'll never migrate RAM_SHARED at all), but normally this shouldn't | |
419 | * affect in reality, or we can revisit. | |
420 | */ | |
421 | RAMBLOCK_FOREACH(block) { | |
74c38cf7 | 422 | if (test_ramblock_postcopiable(block, errp)) { |
ae30b9b2 PX |
423 | goto out; |
424 | } | |
8679638b DDAG |
425 | } |
426 | ||
58b7c17e DDAG |
427 | /* |
428 | * userfault and mlock don't go together; we'll put it back later if | |
429 | * it was enabled. | |
430 | */ | |
431 | if (munlockall()) { | |
74c38cf7 | 432 | error_setg(errp, "munlockall() failed: %s", strerror(errno)); |
617a32f5 | 433 | goto out; |
58b7c17e DDAG |
434 | } |
435 | ||
eb59db53 DDAG |
436 | /* |
437 | * We need to check that the ops we need are supported on anon memory | |
438 | * To do that we need to register a chunk and see the flags that | |
439 | * are returned. | |
440 | */ | |
441 | testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | | |
442 | MAP_ANONYMOUS, -1, 0); | |
443 | if (testarea == MAP_FAILED) { | |
74c38cf7 | 444 | error_setg(errp, "Failed to map test area: %s", strerror(errno)); |
eb59db53 DDAG |
445 | goto out; |
446 | } | |
7648297d | 447 | g_assert(QEMU_PTR_IS_ALIGNED(testarea, pagesize)); |
eb59db53 DDAG |
448 | |
449 | reg_struct.range.start = (uintptr_t)testarea; | |
450 | reg_struct.range.len = pagesize; | |
451 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; | |
452 | ||
453 | if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) { | |
74c38cf7 | 454 | error_setg(errp, "UFFDIO_REGISTER failed: %s", strerror(errno)); |
eb59db53 DDAG |
455 | goto out; |
456 | } | |
457 | ||
458 | range_struct.start = (uintptr_t)testarea; | |
459 | range_struct.len = pagesize; | |
460 | if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) { | |
74c38cf7 | 461 | error_setg(errp, "UFFDIO_UNREGISTER failed: %s", strerror(errno)); |
eb59db53 DDAG |
462 | goto out; |
463 | } | |
464 | ||
465 | feature_mask = (__u64)1 << _UFFDIO_WAKE | | |
466 | (__u64)1 << _UFFDIO_COPY | | |
467 | (__u64)1 << _UFFDIO_ZEROPAGE; | |
468 | if ((reg_struct.ioctls & feature_mask) != feature_mask) { | |
74c38cf7 PX |
469 | error_setg(errp, "Missing userfault map features: %" PRIx64, |
470 | (uint64_t)(~reg_struct.ioctls & feature_mask)); | |
eb59db53 DDAG |
471 | goto out; |
472 | } | |
473 | ||
474 | /* Success! */ | |
475 | ret = true; | |
476 | out: | |
477 | if (testarea) { | |
478 | munmap(testarea, pagesize); | |
479 | } | |
480 | if (ufd != -1) { | |
481 | close(ufd); | |
482 | } | |
483 | return ret; | |
484 | } | |
485 | ||
1caddf8a DDAG |
486 | /* |
487 | * Setup an area of RAM so that it *can* be used for postcopy later; this | |
488 | * must be done right at the start prior to pre-copy. | |
489 | * opaque should be the MIS. | |
490 | */ | |
754cb9c0 | 491 | static int init_range(RAMBlock *rb, void *opaque) |
1caddf8a | 492 | { |
754cb9c0 YK |
493 | const char *block_name = qemu_ram_get_idstr(rb); |
494 | void *host_addr = qemu_ram_get_host_addr(rb); | |
495 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
496 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
1caddf8a DDAG |
497 | trace_postcopy_init_range(block_name, host_addr, offset, length); |
498 | ||
898ba906 DH |
499 | /* |
500 | * Save the used_length before running the guest. In case we have to | |
501 | * resize RAM blocks when syncing RAM block sizes from the source during | |
502 | * precopy, we'll update it manually via the ram block notifier. | |
503 | */ | |
504 | rb->postcopy_length = length; | |
505 | ||
1caddf8a DDAG |
506 | /* |
507 | * We need the whole of RAM to be truly empty for postcopy, so things | |
508 | * like ROMs and any data tables built during init must be zero'd | |
509 | * - we're going to get the copy from the source anyway. | |
510 | * (Precopy will just overwrite this data, so doesn't need the discard) | |
511 | */ | |
aaa2064c | 512 | if (ram_discard_range(block_name, 0, length)) { |
1caddf8a DDAG |
513 | return -1; |
514 | } | |
515 | ||
516 | return 0; | |
517 | } | |
518 | ||
519 | /* | |
520 | * At the end of migration, undo the effects of init_range | |
521 | * opaque should be the MIS. | |
522 | */ | |
754cb9c0 | 523 | static int cleanup_range(RAMBlock *rb, void *opaque) |
1caddf8a | 524 | { |
754cb9c0 YK |
525 | const char *block_name = qemu_ram_get_idstr(rb); |
526 | void *host_addr = qemu_ram_get_host_addr(rb); | |
527 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
898ba906 | 528 | ram_addr_t length = rb->postcopy_length; |
1caddf8a DDAG |
529 | MigrationIncomingState *mis = opaque; |
530 | struct uffdio_range range_struct; | |
531 | trace_postcopy_cleanup_range(block_name, host_addr, offset, length); | |
532 | ||
533 | /* | |
534 | * We turned off hugepage for the precopy stage with postcopy enabled | |
535 | * we can turn it back on now. | |
536 | */ | |
1d741439 | 537 | qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE); |
1caddf8a DDAG |
538 | |
539 | /* | |
540 | * We can also turn off userfault now since we should have all the | |
541 | * pages. It can be useful to leave it on to debug postcopy | |
542 | * if you're not sure it's always getting every page. | |
543 | */ | |
544 | range_struct.start = (uintptr_t)host_addr; | |
545 | range_struct.len = length; | |
546 | ||
547 | if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) { | |
548 | error_report("%s: userfault unregister %s", __func__, strerror(errno)); | |
549 | ||
550 | return -1; | |
551 | } | |
552 | ||
553 | return 0; | |
554 | } | |
555 | ||
556 | /* | |
557 | * Initialise postcopy-ram, setting the RAM to a state where we can go into | |
558 | * postcopy later; must be called prior to any precopy. | |
559 | * called from arch_init's similarly named ram_postcopy_incoming_init | |
560 | */ | |
c136180c | 561 | int postcopy_ram_incoming_init(MigrationIncomingState *mis) |
1caddf8a | 562 | { |
fbd162e6 | 563 | if (foreach_not_ignored_block(init_range, NULL)) { |
1caddf8a DDAG |
564 | return -1; |
565 | } | |
566 | ||
567 | return 0; | |
568 | } | |
569 | ||
476ebf77 PX |
570 | static void postcopy_temp_pages_cleanup(MigrationIncomingState *mis) |
571 | { | |
77dadc3f PX |
572 | int i; |
573 | ||
574 | if (mis->postcopy_tmp_pages) { | |
575 | for (i = 0; i < mis->postcopy_channels; i++) { | |
576 | if (mis->postcopy_tmp_pages[i].tmp_huge_page) { | |
577 | munmap(mis->postcopy_tmp_pages[i].tmp_huge_page, | |
578 | mis->largest_page_size); | |
579 | mis->postcopy_tmp_pages[i].tmp_huge_page = NULL; | |
580 | } | |
581 | } | |
582 | g_free(mis->postcopy_tmp_pages); | |
583 | mis->postcopy_tmp_pages = NULL; | |
476ebf77 PX |
584 | } |
585 | ||
586 | if (mis->postcopy_tmp_zero_page) { | |
587 | munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size); | |
588 | mis->postcopy_tmp_zero_page = NULL; | |
589 | } | |
590 | } | |
591 | ||
1caddf8a DDAG |
592 | /* |
593 | * At the end of a migration where postcopy_ram_incoming_init was called. | |
594 | */ | |
595 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
596 | { | |
c4faeed2 DDAG |
597 | trace_postcopy_ram_incoming_cleanup_entry(); |
598 | ||
6621883f PX |
599 | if (mis->preempt_thread_status == PREEMPT_THREAD_CREATED) { |
600 | /* Notify the fast load thread to quit */ | |
601 | mis->preempt_thread_status = PREEMPT_THREAD_QUIT; | |
602 | if (mis->postcopy_qemufile_dst) { | |
603 | qemu_file_shutdown(mis->postcopy_qemufile_dst); | |
604 | } | |
36f62f11 | 605 | qemu_thread_join(&mis->postcopy_prio_thread); |
6621883f | 606 | mis->preempt_thread_status = PREEMPT_THREAD_NONE; |
36f62f11 PX |
607 | } |
608 | ||
c4faeed2 | 609 | if (mis->have_fault_thread) { |
46343570 DDAG |
610 | Error *local_err = NULL; |
611 | ||
55d0fe82 | 612 | /* Let the fault thread quit */ |
d73415a3 | 613 | qatomic_set(&mis->fault_thread_quit, 1); |
55d0fe82 IM |
614 | postcopy_fault_thread_notify(mis); |
615 | trace_postcopy_ram_incoming_cleanup_join(); | |
616 | qemu_thread_join(&mis->fault_thread); | |
617 | ||
46343570 DDAG |
618 | if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) { |
619 | error_report_err(local_err); | |
620 | return -1; | |
621 | } | |
622 | ||
fbd162e6 | 623 | if (foreach_not_ignored_block(cleanup_range, mis)) { |
c4faeed2 DDAG |
624 | return -1; |
625 | } | |
9ab7ef9b | 626 | |
c4faeed2 DDAG |
627 | trace_postcopy_ram_incoming_cleanup_closeuf(); |
628 | close(mis->userfault_fd); | |
64f615fe | 629 | close(mis->userfault_event_fd); |
c4faeed2 | 630 | mis->have_fault_thread = false; |
1caddf8a DDAG |
631 | } |
632 | ||
58b7c17e DDAG |
633 | if (enable_mlock) { |
634 | if (os_mlock() < 0) { | |
635 | error_report("mlock: %s", strerror(errno)); | |
636 | /* | |
637 | * It doesn't feel right to fail at this point, we have a valid | |
638 | * VM state. | |
639 | */ | |
640 | } | |
641 | } | |
642 | ||
476ebf77 PX |
643 | postcopy_temp_pages_cleanup(mis); |
644 | ||
65ace060 AP |
645 | trace_postcopy_ram_incoming_cleanup_blocktime( |
646 | get_postcopy_total_blocktime()); | |
647 | ||
c4faeed2 | 648 | trace_postcopy_ram_incoming_cleanup_exit(); |
1caddf8a DDAG |
649 | return 0; |
650 | } | |
651 | ||
f9527107 DDAG |
652 | /* |
653 | * Disable huge pages on an area | |
654 | */ | |
754cb9c0 | 655 | static int nhp_range(RAMBlock *rb, void *opaque) |
f9527107 | 656 | { |
754cb9c0 YK |
657 | const char *block_name = qemu_ram_get_idstr(rb); |
658 | void *host_addr = qemu_ram_get_host_addr(rb); | |
659 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
898ba906 | 660 | ram_addr_t length = rb->postcopy_length; |
f9527107 DDAG |
661 | trace_postcopy_nhp_range(block_name, host_addr, offset, length); |
662 | ||
663 | /* | |
664 | * Before we do discards we need to ensure those discards really | |
665 | * do delete areas of the page, even if THP thinks a hugepage would | |
666 | * be a good idea, so force hugepages off. | |
667 | */ | |
1d741439 | 668 | qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE); |
f9527107 DDAG |
669 | |
670 | return 0; | |
671 | } | |
672 | ||
673 | /* | |
674 | * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard | |
675 | * however leaving it until after precopy means that most of the precopy | |
676 | * data is still THPd | |
677 | */ | |
678 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) | |
679 | { | |
fbd162e6 | 680 | if (foreach_not_ignored_block(nhp_range, mis)) { |
f9527107 DDAG |
681 | return -1; |
682 | } | |
683 | ||
684 | postcopy_state_set(POSTCOPY_INCOMING_DISCARD); | |
685 | ||
686 | return 0; | |
687 | } | |
688 | ||
f0a227ad DDAG |
689 | /* |
690 | * Mark the given area of RAM as requiring notification to unwritten areas | |
fbd162e6 | 691 | * Used as a callback on foreach_not_ignored_block. |
f0a227ad DDAG |
692 | * host_addr: Base of area to mark |
693 | * offset: Offset in the whole ram arena | |
694 | * length: Length of the section | |
695 | * opaque: MigrationIncomingState pointer | |
696 | * Returns 0 on success | |
697 | */ | |
754cb9c0 | 698 | static int ram_block_enable_notify(RAMBlock *rb, void *opaque) |
f0a227ad DDAG |
699 | { |
700 | MigrationIncomingState *mis = opaque; | |
701 | struct uffdio_register reg_struct; | |
702 | ||
754cb9c0 | 703 | reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb); |
898ba906 | 704 | reg_struct.range.len = rb->postcopy_length; |
f0a227ad DDAG |
705 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; |
706 | ||
707 | /* Now tell our userfault_fd that it's responsible for this area */ | |
708 | if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) { | |
709 | error_report("%s userfault register: %s", __func__, strerror(errno)); | |
710 | return -1; | |
711 | } | |
665414ad DDAG |
712 | if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { |
713 | error_report("%s userfault: Region doesn't support COPY", __func__); | |
714 | return -1; | |
715 | } | |
2ce16640 | 716 | if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) { |
2ce16640 DDAG |
717 | qemu_ram_set_uf_zeroable(rb); |
718 | } | |
f0a227ad DDAG |
719 | |
720 | return 0; | |
721 | } | |
722 | ||
5efc3564 DDAG |
723 | int postcopy_wake_shared(struct PostCopyFD *pcfd, |
724 | uint64_t client_addr, | |
725 | RAMBlock *rb) | |
726 | { | |
727 | size_t pagesize = qemu_ram_pagesize(rb); | |
728 | struct uffdio_range range; | |
729 | int ret; | |
730 | trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb)); | |
7648297d | 731 | range.start = ROUND_DOWN(client_addr, pagesize); |
5efc3564 DDAG |
732 | range.len = pagesize; |
733 | ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range); | |
734 | if (ret) { | |
735 | error_report("%s: Failed to wake: %zx in %s (%s)", | |
736 | __func__, (size_t)client_addr, qemu_ram_get_idstr(rb), | |
737 | strerror(errno)); | |
738 | } | |
739 | return ret; | |
740 | } | |
741 | ||
9470c5e0 DH |
742 | static int postcopy_request_page(MigrationIncomingState *mis, RAMBlock *rb, |
743 | ram_addr_t start, uint64_t haddr) | |
744 | { | |
745 | void *aligned = (void *)(uintptr_t)ROUND_DOWN(haddr, qemu_ram_pagesize(rb)); | |
746 | ||
747 | /* | |
748 | * Discarded pages (via RamDiscardManager) are never migrated. On unlikely | |
749 | * access, place a zeropage, which will also set the relevant bits in the | |
750 | * recv_bitmap accordingly, so we won't try placing a zeropage twice. | |
751 | * | |
752 | * Checking a single bit is sufficient to handle pagesize > TPS as either | |
753 | * all relevant bits are set or not. | |
754 | */ | |
755 | assert(QEMU_IS_ALIGNED(start, qemu_ram_pagesize(rb))); | |
756 | if (ramblock_page_is_discarded(rb, start)) { | |
757 | bool received = ramblock_recv_bitmap_test_byte_offset(rb, start); | |
758 | ||
759 | return received ? 0 : postcopy_place_page_zero(mis, aligned, rb); | |
760 | } | |
761 | ||
762 | return migrate_send_rp_req_pages(mis, rb, start, haddr); | |
763 | } | |
764 | ||
096bf4c8 DDAG |
765 | /* |
766 | * Callback from shared fault handlers to ask for a page, | |
767 | * the page must be specified by a RAMBlock and an offset in that rb | |
768 | * Note: Only for use by shared fault handlers (in fault thread) | |
769 | */ | |
770 | int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, | |
771 | uint64_t client_addr, uint64_t rb_offset) | |
772 | { | |
7648297d | 773 | uint64_t aligned_rbo = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb)); |
096bf4c8 DDAG |
774 | MigrationIncomingState *mis = migration_incoming_get_current(); |
775 | ||
776 | trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb), | |
777 | rb_offset); | |
dedfb4b2 DDAG |
778 | if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) { |
779 | trace_postcopy_request_shared_page_present(pcfd->idstr, | |
780 | qemu_ram_get_idstr(rb), rb_offset); | |
781 | return postcopy_wake_shared(pcfd, client_addr, rb); | |
782 | } | |
9470c5e0 | 783 | postcopy_request_page(mis, rb, aligned_rbo, client_addr); |
096bf4c8 DDAG |
784 | return 0; |
785 | } | |
786 | ||
575b0b33 AP |
787 | static int get_mem_fault_cpu_index(uint32_t pid) |
788 | { | |
789 | CPUState *cpu_iter; | |
790 | ||
791 | CPU_FOREACH(cpu_iter) { | |
792 | if (cpu_iter->thread_id == pid) { | |
793 | trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid); | |
794 | return cpu_iter->cpu_index; | |
795 | } | |
796 | } | |
797 | trace_get_mem_fault_cpu_index(-1, pid); | |
798 | return -1; | |
799 | } | |
800 | ||
801 | static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc) | |
802 | { | |
803 | int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - | |
804 | dc->start_time; | |
805 | return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX; | |
806 | } | |
807 | ||
808 | /* | |
809 | * This function is being called when pagefault occurs. It | |
810 | * tracks down vCPU blocking time. | |
811 | * | |
812 | * @addr: faulted host virtual address | |
813 | * @ptid: faulted process thread id | |
814 | * @rb: ramblock appropriate to addr | |
815 | */ | |
816 | static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid, | |
817 | RAMBlock *rb) | |
818 | { | |
819 | int cpu, already_received; | |
820 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
821 | PostcopyBlocktimeContext *dc = mis->blocktime_ctx; | |
822 | uint32_t low_time_offset; | |
823 | ||
824 | if (!dc || ptid == 0) { | |
825 | return; | |
826 | } | |
827 | cpu = get_mem_fault_cpu_index(ptid); | |
828 | if (cpu < 0) { | |
829 | return; | |
830 | } | |
831 | ||
832 | low_time_offset = get_low_time_offset(dc); | |
833 | if (dc->vcpu_addr[cpu] == 0) { | |
d73415a3 | 834 | qatomic_inc(&dc->smp_cpus_down); |
575b0b33 AP |
835 | } |
836 | ||
d73415a3 SH |
837 | qatomic_xchg(&dc->last_begin, low_time_offset); |
838 | qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset); | |
839 | qatomic_xchg(&dc->vcpu_addr[cpu], addr); | |
575b0b33 | 840 | |
da1725d3 WY |
841 | /* |
842 | * check it here, not at the beginning of the function, | |
843 | * due to, check could occur early than bitmap_set in | |
844 | * qemu_ufd_copy_ioctl | |
845 | */ | |
575b0b33 AP |
846 | already_received = ramblock_recv_bitmap_test(rb, (void *)addr); |
847 | if (already_received) { | |
d73415a3 SH |
848 | qatomic_xchg(&dc->vcpu_addr[cpu], 0); |
849 | qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0); | |
850 | qatomic_dec(&dc->smp_cpus_down); | |
575b0b33 AP |
851 | } |
852 | trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu], | |
853 | cpu, already_received); | |
854 | } | |
855 | ||
856 | /* | |
857 | * This function just provide calculated blocktime per cpu and trace it. | |
858 | * Total blocktime is calculated in mark_postcopy_blocktime_end. | |
859 | * | |
860 | * | |
861 | * Assume we have 3 CPU | |
862 | * | |
863 | * S1 E1 S1 E1 | |
864 | * -----***********------------xxx***************------------------------> CPU1 | |
865 | * | |
866 | * S2 E2 | |
867 | * ------------****************xxx---------------------------------------> CPU2 | |
868 | * | |
869 | * S3 E3 | |
870 | * ------------------------****xxx********-------------------------------> CPU3 | |
871 | * | |
872 | * We have sequence S1,S2,E1,S3,S1,E2,E3,E1 | |
873 | * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3 | |
874 | * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 - | |
875 | * it's a part of total blocktime. | |
876 | * S1 - here is last_begin | |
877 | * Legend of the picture is following: | |
878 | * * - means blocktime per vCPU | |
879 | * x - means overlapped blocktime (total blocktime) | |
880 | * | |
881 | * @addr: host virtual address | |
882 | */ | |
883 | static void mark_postcopy_blocktime_end(uintptr_t addr) | |
884 | { | |
885 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
886 | PostcopyBlocktimeContext *dc = mis->blocktime_ctx; | |
5cc8767d LX |
887 | MachineState *ms = MACHINE(qdev_get_machine()); |
888 | unsigned int smp_cpus = ms->smp.cpus; | |
575b0b33 AP |
889 | int i, affected_cpu = 0; |
890 | bool vcpu_total_blocktime = false; | |
891 | uint32_t read_vcpu_time, low_time_offset; | |
892 | ||
893 | if (!dc) { | |
894 | return; | |
895 | } | |
896 | ||
897 | low_time_offset = get_low_time_offset(dc); | |
898 | /* lookup cpu, to clear it, | |
3a4452d8 | 899 | * that algorithm looks straightforward, but it's not |
575b0b33 AP |
900 | * optimal, more optimal algorithm is keeping tree or hash |
901 | * where key is address value is a list of */ | |
902 | for (i = 0; i < smp_cpus; i++) { | |
903 | uint32_t vcpu_blocktime = 0; | |
904 | ||
d73415a3 SH |
905 | read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0); |
906 | if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr || | |
575b0b33 AP |
907 | read_vcpu_time == 0) { |
908 | continue; | |
909 | } | |
d73415a3 | 910 | qatomic_xchg(&dc->vcpu_addr[i], 0); |
575b0b33 AP |
911 | vcpu_blocktime = low_time_offset - read_vcpu_time; |
912 | affected_cpu += 1; | |
913 | /* we need to know is that mark_postcopy_end was due to | |
914 | * faulted page, another possible case it's prefetched | |
915 | * page and in that case we shouldn't be here */ | |
916 | if (!vcpu_total_blocktime && | |
d73415a3 | 917 | qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) { |
575b0b33 AP |
918 | vcpu_total_blocktime = true; |
919 | } | |
920 | /* continue cycle, due to one page could affect several vCPUs */ | |
921 | dc->vcpu_blocktime[i] += vcpu_blocktime; | |
922 | } | |
923 | ||
d73415a3 | 924 | qatomic_sub(&dc->smp_cpus_down, affected_cpu); |
575b0b33 | 925 | if (vcpu_total_blocktime) { |
d73415a3 | 926 | dc->total_blocktime += low_time_offset - qatomic_fetch_add( |
575b0b33 AP |
927 | &dc->last_begin, 0); |
928 | } | |
929 | trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime, | |
930 | affected_cpu); | |
931 | } | |
932 | ||
27dd21b4 | 933 | static void postcopy_pause_fault_thread(MigrationIncomingState *mis) |
3a7804c3 PX |
934 | { |
935 | trace_postcopy_pause_fault_thread(); | |
3a7804c3 | 936 | qemu_sem_wait(&mis->postcopy_pause_sem_fault); |
3a7804c3 | 937 | trace_postcopy_pause_fault_thread_continued(); |
3a7804c3 PX |
938 | } |
939 | ||
f0a227ad DDAG |
940 | /* |
941 | * Handle faults detected by the USERFAULT markings | |
942 | */ | |
943 | static void *postcopy_ram_fault_thread(void *opaque) | |
944 | { | |
945 | MigrationIncomingState *mis = opaque; | |
c4faeed2 DDAG |
946 | struct uffd_msg msg; |
947 | int ret; | |
00fa4fc8 | 948 | size_t index; |
c4faeed2 | 949 | RAMBlock *rb = NULL; |
f0a227ad | 950 | |
c4faeed2 | 951 | trace_postcopy_ram_fault_thread_entry(); |
74637e6f | 952 | rcu_register_thread(); |
096bf4c8 | 953 | mis->last_rb = NULL; /* last RAMBlock we sent part of */ |
095c12a4 | 954 | qemu_sem_post(&mis->thread_sync_sem); |
f0a227ad | 955 | |
00fa4fc8 DDAG |
956 | struct pollfd *pfd; |
957 | size_t pfd_len = 2 + mis->postcopy_remote_fds->len; | |
958 | ||
959 | pfd = g_new0(struct pollfd, pfd_len); | |
960 | ||
961 | pfd[0].fd = mis->userfault_fd; | |
962 | pfd[0].events = POLLIN; | |
963 | pfd[1].fd = mis->userfault_event_fd; | |
964 | pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */ | |
965 | trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd); | |
966 | for (index = 0; index < mis->postcopy_remote_fds->len; index++) { | |
967 | struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds, | |
968 | struct PostCopyFD, index); | |
969 | pfd[2 + index].fd = pcfd->fd; | |
970 | pfd[2 + index].events = POLLIN; | |
971 | trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr, | |
972 | pcfd->fd); | |
973 | } | |
974 | ||
c4faeed2 DDAG |
975 | while (true) { |
976 | ram_addr_t rb_offset; | |
00fa4fc8 | 977 | int poll_result; |
c4faeed2 DDAG |
978 | |
979 | /* | |
980 | * We're mainly waiting for the kernel to give us a faulting HVA, | |
981 | * however we can be told to quit via userfault_quit_fd which is | |
982 | * an eventfd | |
983 | */ | |
00fa4fc8 DDAG |
984 | |
985 | poll_result = poll(pfd, pfd_len, -1 /* Wait forever */); | |
986 | if (poll_result == -1) { | |
c4faeed2 DDAG |
987 | error_report("%s: userfault poll: %s", __func__, strerror(errno)); |
988 | break; | |
989 | } | |
990 | ||
3a7804c3 PX |
991 | if (!mis->to_src_file) { |
992 | /* | |
993 | * Possibly someone tells us that the return path is | |
994 | * broken already using the event. We should hold until | |
995 | * the channel is rebuilt. | |
996 | */ | |
27dd21b4 | 997 | postcopy_pause_fault_thread(mis); |
3a7804c3 PX |
998 | } |
999 | ||
c4faeed2 | 1000 | if (pfd[1].revents) { |
64f615fe PX |
1001 | uint64_t tmp64 = 0; |
1002 | ||
1003 | /* Consume the signal */ | |
1004 | if (read(mis->userfault_event_fd, &tmp64, 8) != 8) { | |
1005 | /* Nothing obviously nicer than posting this error. */ | |
1006 | error_report("%s: read() failed", __func__); | |
1007 | } | |
1008 | ||
d73415a3 | 1009 | if (qatomic_read(&mis->fault_thread_quit)) { |
64f615fe PX |
1010 | trace_postcopy_ram_fault_thread_quit(); |
1011 | break; | |
1012 | } | |
c4faeed2 DDAG |
1013 | } |
1014 | ||
00fa4fc8 DDAG |
1015 | if (pfd[0].revents) { |
1016 | poll_result--; | |
1017 | ret = read(mis->userfault_fd, &msg, sizeof(msg)); | |
1018 | if (ret != sizeof(msg)) { | |
1019 | if (errno == EAGAIN) { | |
1020 | /* | |
1021 | * if a wake up happens on the other thread just after | |
1022 | * the poll, there is nothing to read. | |
1023 | */ | |
1024 | continue; | |
1025 | } | |
1026 | if (ret < 0) { | |
1027 | error_report("%s: Failed to read full userfault " | |
1028 | "message: %s", | |
1029 | __func__, strerror(errno)); | |
1030 | break; | |
1031 | } else { | |
1032 | error_report("%s: Read %d bytes from userfaultfd " | |
1033 | "expected %zd", | |
1034 | __func__, ret, sizeof(msg)); | |
1035 | break; /* Lost alignment, don't know what we'd read next */ | |
1036 | } | |
c4faeed2 | 1037 | } |
00fa4fc8 DDAG |
1038 | if (msg.event != UFFD_EVENT_PAGEFAULT) { |
1039 | error_report("%s: Read unexpected event %ud from userfaultfd", | |
1040 | __func__, msg.event); | |
1041 | continue; /* It's not a page fault, shouldn't happen */ | |
c4faeed2 | 1042 | } |
c4faeed2 | 1043 | |
00fa4fc8 DDAG |
1044 | rb = qemu_ram_block_from_host( |
1045 | (void *)(uintptr_t)msg.arg.pagefault.address, | |
1046 | true, &rb_offset); | |
1047 | if (!rb) { | |
1048 | error_report("postcopy_ram_fault_thread: Fault outside guest: %" | |
1049 | PRIx64, (uint64_t)msg.arg.pagefault.address); | |
1050 | break; | |
1051 | } | |
c4faeed2 | 1052 | |
7648297d | 1053 | rb_offset = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb)); |
00fa4fc8 | 1054 | trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, |
c4faeed2 | 1055 | qemu_ram_get_idstr(rb), |
575b0b33 AP |
1056 | rb_offset, |
1057 | msg.arg.pagefault.feat.ptid); | |
1058 | mark_postcopy_blocktime_begin( | |
1059 | (uintptr_t)(msg.arg.pagefault.address), | |
1060 | msg.arg.pagefault.feat.ptid, rb); | |
1061 | ||
3a7804c3 | 1062 | retry: |
00fa4fc8 DDAG |
1063 | /* |
1064 | * Send the request to the source - we want to request one | |
1065 | * of our host page sizes (which is >= TPS) | |
1066 | */ | |
9470c5e0 DH |
1067 | ret = postcopy_request_page(mis, rb, rb_offset, |
1068 | msg.arg.pagefault.address); | |
3a7804c3 PX |
1069 | if (ret) { |
1070 | /* May be network failure, try to wait for recovery */ | |
27dd21b4 PX |
1071 | postcopy_pause_fault_thread(mis); |
1072 | goto retry; | |
00fa4fc8 DDAG |
1073 | } |
1074 | } | |
c4faeed2 | 1075 | |
00fa4fc8 DDAG |
1076 | /* Now handle any requests from external processes on shared memory */ |
1077 | /* TODO: May need to handle devices deregistering during postcopy */ | |
1078 | for (index = 2; index < pfd_len && poll_result; index++) { | |
1079 | if (pfd[index].revents) { | |
1080 | struct PostCopyFD *pcfd = | |
1081 | &g_array_index(mis->postcopy_remote_fds, | |
1082 | struct PostCopyFD, index - 2); | |
1083 | ||
1084 | poll_result--; | |
1085 | if (pfd[index].revents & POLLERR) { | |
1086 | error_report("%s: POLLERR on poll %zd fd=%d", | |
1087 | __func__, index, pcfd->fd); | |
1088 | pfd[index].events = 0; | |
1089 | continue; | |
1090 | } | |
1091 | ||
1092 | ret = read(pcfd->fd, &msg, sizeof(msg)); | |
1093 | if (ret != sizeof(msg)) { | |
1094 | if (errno == EAGAIN) { | |
1095 | /* | |
1096 | * if a wake up happens on the other thread just after | |
1097 | * the poll, there is nothing to read. | |
1098 | */ | |
1099 | continue; | |
1100 | } | |
1101 | if (ret < 0) { | |
1102 | error_report("%s: Failed to read full userfault " | |
1103 | "message: %s (shared) revents=%d", | |
1104 | __func__, strerror(errno), | |
1105 | pfd[index].revents); | |
1106 | /*TODO: Could just disable this sharer */ | |
1107 | break; | |
1108 | } else { | |
1109 | error_report("%s: Read %d bytes from userfaultfd " | |
1110 | "expected %zd (shared)", | |
1111 | __func__, ret, sizeof(msg)); | |
1112 | /*TODO: Could just disable this sharer */ | |
1113 | break; /*Lost alignment,don't know what we'd read next*/ | |
1114 | } | |
1115 | } | |
1116 | if (msg.event != UFFD_EVENT_PAGEFAULT) { | |
1117 | error_report("%s: Read unexpected event %ud " | |
1118 | "from userfaultfd (shared)", | |
1119 | __func__, msg.event); | |
1120 | continue; /* It's not a page fault, shouldn't happen */ | |
1121 | } | |
1122 | /* Call the device handler registered with us */ | |
1123 | ret = pcfd->handler(pcfd, &msg); | |
1124 | if (ret) { | |
1125 | error_report("%s: Failed to resolve shared fault on %zd/%s", | |
1126 | __func__, index, pcfd->idstr); | |
1127 | /* TODO: Fail? Disable this sharer? */ | |
1128 | } | |
1129 | } | |
c4faeed2 DDAG |
1130 | } |
1131 | } | |
74637e6f | 1132 | rcu_unregister_thread(); |
c4faeed2 | 1133 | trace_postcopy_ram_fault_thread_exit(); |
fc6008f3 | 1134 | g_free(pfd); |
f0a227ad DDAG |
1135 | return NULL; |
1136 | } | |
1137 | ||
476ebf77 PX |
1138 | static int postcopy_temp_pages_setup(MigrationIncomingState *mis) |
1139 | { | |
77dadc3f PX |
1140 | PostcopyTmpPage *tmp_page; |
1141 | int err, i, channels; | |
1142 | void *temp_page; | |
1143 | ||
36f62f11 PX |
1144 | if (migrate_postcopy_preempt()) { |
1145 | /* If preemption enabled, need extra channel for urgent requests */ | |
1146 | mis->postcopy_channels = RAM_CHANNEL_MAX; | |
1147 | } else { | |
1148 | /* Both precopy/postcopy on the same channel */ | |
1149 | mis->postcopy_channels = 1; | |
1150 | } | |
77dadc3f PX |
1151 | |
1152 | channels = mis->postcopy_channels; | |
1153 | mis->postcopy_tmp_pages = g_malloc0_n(sizeof(PostcopyTmpPage), channels); | |
1154 | ||
1155 | for (i = 0; i < channels; i++) { | |
1156 | tmp_page = &mis->postcopy_tmp_pages[i]; | |
1157 | temp_page = mmap(NULL, mis->largest_page_size, PROT_READ | PROT_WRITE, | |
1158 | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); | |
1159 | if (temp_page == MAP_FAILED) { | |
1160 | err = errno; | |
1161 | error_report("%s: Failed to map postcopy_tmp_pages[%d]: %s", | |
1162 | __func__, i, strerror(err)); | |
1163 | /* Clean up will be done later */ | |
1164 | return -err; | |
1165 | } | |
1166 | tmp_page->tmp_huge_page = temp_page; | |
1167 | /* Initialize default states for each tmp page */ | |
1168 | postcopy_temp_page_reset(tmp_page); | |
476ebf77 PX |
1169 | } |
1170 | ||
1171 | /* | |
1172 | * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages | |
1173 | */ | |
1174 | mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size, | |
1175 | PROT_READ | PROT_WRITE, | |
1176 | MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); | |
1177 | if (mis->postcopy_tmp_zero_page == MAP_FAILED) { | |
1178 | err = errno; | |
1179 | mis->postcopy_tmp_zero_page = NULL; | |
1180 | error_report("%s: Failed to map large zero page %s", | |
1181 | __func__, strerror(err)); | |
1182 | return -err; | |
1183 | } | |
1184 | ||
1185 | memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size); | |
1186 | ||
1187 | return 0; | |
1188 | } | |
1189 | ||
2a7eb148 | 1190 | int postcopy_ram_incoming_setup(MigrationIncomingState *mis) |
f0a227ad | 1191 | { |
74c38cf7 PX |
1192 | Error *local_err = NULL; |
1193 | ||
c4faeed2 | 1194 | /* Open the fd for the kernel to give us userfaults */ |
d5890ea0 | 1195 | mis->userfault_fd = uffd_open(O_CLOEXEC | O_NONBLOCK); |
c4faeed2 DDAG |
1196 | if (mis->userfault_fd == -1) { |
1197 | error_report("%s: Failed to open userfault fd: %s", __func__, | |
1198 | strerror(errno)); | |
1199 | return -1; | |
1200 | } | |
1201 | ||
1202 | /* | |
1203 | * Although the host check already tested the API, we need to | |
1204 | * do the check again as an ABI handshake on the new fd. | |
1205 | */ | |
74c38cf7 PX |
1206 | if (!ufd_check_and_apply(mis->userfault_fd, mis, &local_err)) { |
1207 | error_report_err(local_err); | |
c4faeed2 DDAG |
1208 | return -1; |
1209 | } | |
1210 | ||
1211 | /* Now an eventfd we use to tell the fault-thread to quit */ | |
64f615fe PX |
1212 | mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC); |
1213 | if (mis->userfault_event_fd == -1) { | |
1214 | error_report("%s: Opening userfault_event_fd: %s", __func__, | |
c4faeed2 DDAG |
1215 | strerror(errno)); |
1216 | close(mis->userfault_fd); | |
1217 | return -1; | |
1218 | } | |
1219 | ||
36f62f11 | 1220 | postcopy_thread_create(mis, &mis->fault_thread, "fault-default", |
095c12a4 | 1221 | postcopy_ram_fault_thread, QEMU_THREAD_JOINABLE); |
c4faeed2 | 1222 | mis->have_fault_thread = true; |
f0a227ad DDAG |
1223 | |
1224 | /* Mark so that we get notified of accesses to unwritten areas */ | |
fbd162e6 | 1225 | if (foreach_not_ignored_block(ram_block_enable_notify, mis)) { |
91b02dc7 | 1226 | error_report("ram_block_enable_notify failed"); |
f0a227ad DDAG |
1227 | return -1; |
1228 | } | |
1229 | ||
476ebf77 PX |
1230 | if (postcopy_temp_pages_setup(mis)) { |
1231 | /* Error dumped in the sub-function */ | |
3414322a WY |
1232 | return -1; |
1233 | } | |
1234 | ||
36f62f11 PX |
1235 | if (migrate_postcopy_preempt()) { |
1236 | /* | |
1237 | * This thread needs to be created after the temp pages because | |
1238 | * it'll fetch RAM_CHANNEL_POSTCOPY PostcopyTmpPage immediately. | |
1239 | */ | |
1240 | postcopy_thread_create(mis, &mis->postcopy_prio_thread, "fault-fast", | |
1241 | postcopy_preempt_thread, QEMU_THREAD_JOINABLE); | |
6621883f | 1242 | mis->preempt_thread_status = PREEMPT_THREAD_CREATED; |
36f62f11 PX |
1243 | } |
1244 | ||
c4faeed2 DDAG |
1245 | trace_postcopy_ram_enable_notify(); |
1246 | ||
f0a227ad DDAG |
1247 | return 0; |
1248 | } | |
1249 | ||
eef621c4 | 1250 | static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr, |
f9494614 | 1251 | void *from_addr, uint64_t pagesize, RAMBlock *rb) |
727b9d7e | 1252 | { |
eef621c4 | 1253 | int userfault_fd = mis->userfault_fd; |
f9494614 | 1254 | int ret; |
eef621c4 | 1255 | |
727b9d7e AP |
1256 | if (from_addr) { |
1257 | struct uffdio_copy copy_struct; | |
1258 | copy_struct.dst = (uint64_t)(uintptr_t)host_addr; | |
1259 | copy_struct.src = (uint64_t)(uintptr_t)from_addr; | |
1260 | copy_struct.len = pagesize; | |
1261 | copy_struct.mode = 0; | |
f9494614 | 1262 | ret = ioctl(userfault_fd, UFFDIO_COPY, ©_struct); |
727b9d7e AP |
1263 | } else { |
1264 | struct uffdio_zeropage zero_struct; | |
1265 | zero_struct.range.start = (uint64_t)(uintptr_t)host_addr; | |
1266 | zero_struct.range.len = pagesize; | |
1267 | zero_struct.mode = 0; | |
f9494614 AP |
1268 | ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct); |
1269 | } | |
1270 | if (!ret) { | |
8f8bfffc | 1271 | qemu_mutex_lock(&mis->page_request_mutex); |
f9494614 AP |
1272 | ramblock_recv_bitmap_set_range(rb, host_addr, |
1273 | pagesize / qemu_target_page_size()); | |
8f8bfffc PX |
1274 | /* |
1275 | * If this page resolves a page fault for a previous recorded faulted | |
1276 | * address, take a special note to maintain the requested page list. | |
1277 | */ | |
1278 | if (g_tree_lookup(mis->page_requested, host_addr)) { | |
1279 | g_tree_remove(mis->page_requested, host_addr); | |
1280 | mis->page_requested_count--; | |
1281 | trace_postcopy_page_req_del(host_addr, mis->page_requested_count); | |
1282 | } | |
1283 | qemu_mutex_unlock(&mis->page_request_mutex); | |
575b0b33 | 1284 | mark_postcopy_blocktime_end((uintptr_t)host_addr); |
727b9d7e | 1285 | } |
f9494614 | 1286 | return ret; |
727b9d7e AP |
1287 | } |
1288 | ||
d488b349 DDAG |
1289 | int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset) |
1290 | { | |
1291 | int i; | |
1292 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1293 | GArray *pcrfds = mis->postcopy_remote_fds; | |
1294 | ||
1295 | for (i = 0; i < pcrfds->len; i++) { | |
1296 | struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); | |
1297 | int ret = cur->waker(cur, rb, offset); | |
1298 | if (ret) { | |
1299 | return ret; | |
1300 | } | |
1301 | } | |
1302 | return 0; | |
1303 | } | |
1304 | ||
696ed9a9 DDAG |
1305 | /* |
1306 | * Place a host page (from) at (host) atomically | |
1307 | * returns 0 on success | |
1308 | */ | |
df9ff5e1 | 1309 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
8be4620b | 1310 | RAMBlock *rb) |
696ed9a9 | 1311 | { |
8be4620b | 1312 | size_t pagesize = qemu_ram_pagesize(rb); |
696ed9a9 | 1313 | |
696ed9a9 DDAG |
1314 | /* copy also acks to the kernel waking the stalled thread up |
1315 | * TODO: We can inhibit that ack and only do it if it was requested | |
1316 | * which would be slightly cheaper, but we'd have to be careful | |
1317 | * of the order of updating our page state. | |
1318 | */ | |
eef621c4 | 1319 | if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) { |
696ed9a9 | 1320 | int e = errno; |
df9ff5e1 DDAG |
1321 | error_report("%s: %s copy host: %p from: %p (size: %zd)", |
1322 | __func__, strerror(e), host, from, pagesize); | |
696ed9a9 DDAG |
1323 | |
1324 | return -e; | |
1325 | } | |
1326 | ||
1327 | trace_postcopy_place_page(host); | |
dedfb4b2 DDAG |
1328 | return postcopy_notify_shared_wake(rb, |
1329 | qemu_ram_block_host_offset(rb, host)); | |
696ed9a9 DDAG |
1330 | } |
1331 | ||
1332 | /* | |
1333 | * Place a zero page at (host) atomically | |
1334 | * returns 0 on success | |
1335 | */ | |
df9ff5e1 | 1336 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
8be4620b | 1337 | RAMBlock *rb) |
696ed9a9 | 1338 | { |
2ce16640 | 1339 | size_t pagesize = qemu_ram_pagesize(rb); |
df9ff5e1 | 1340 | trace_postcopy_place_page_zero(host); |
696ed9a9 | 1341 | |
2ce16640 DDAG |
1342 | /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE |
1343 | * but it's not available for everything (e.g. hugetlbpages) | |
1344 | */ | |
1345 | if (qemu_ram_is_uf_zeroable(rb)) { | |
eef621c4 | 1346 | if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) { |
df9ff5e1 DDAG |
1347 | int e = errno; |
1348 | error_report("%s: %s zero host: %p", | |
1349 | __func__, strerror(e), host); | |
696ed9a9 | 1350 | |
df9ff5e1 DDAG |
1351 | return -e; |
1352 | } | |
dedfb4b2 DDAG |
1353 | return postcopy_notify_shared_wake(rb, |
1354 | qemu_ram_block_host_offset(rb, | |
1355 | host)); | |
df9ff5e1 | 1356 | } else { |
6629890d | 1357 | return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb); |
696ed9a9 | 1358 | } |
696ed9a9 DDAG |
1359 | } |
1360 | ||
eb59db53 DDAG |
1361 | #else |
1362 | /* No target OS support, stubs just fail */ | |
65ace060 AP |
1363 | void fill_destination_postcopy_migration_info(MigrationInfo *info) |
1364 | { | |
1365 | } | |
1366 | ||
74c38cf7 | 1367 | bool postcopy_ram_supported_by_host(MigrationIncomingState *mis, Error **errp) |
eb59db53 DDAG |
1368 | { |
1369 | error_report("%s: No OS support", __func__); | |
1370 | return false; | |
1371 | } | |
1372 | ||
c136180c | 1373 | int postcopy_ram_incoming_init(MigrationIncomingState *mis) |
1caddf8a DDAG |
1374 | { |
1375 | error_report("postcopy_ram_incoming_init: No OS support"); | |
1376 | return -1; | |
1377 | } | |
1378 | ||
1379 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
1380 | { | |
1381 | assert(0); | |
1382 | return -1; | |
1383 | } | |
1384 | ||
f9527107 DDAG |
1385 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) |
1386 | { | |
1387 | assert(0); | |
1388 | return -1; | |
1389 | } | |
1390 | ||
c188c539 MT |
1391 | int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, |
1392 | uint64_t client_addr, uint64_t rb_offset) | |
1393 | { | |
1394 | assert(0); | |
1395 | return -1; | |
1396 | } | |
1397 | ||
2a7eb148 | 1398 | int postcopy_ram_incoming_setup(MigrationIncomingState *mis) |
f0a227ad DDAG |
1399 | { |
1400 | assert(0); | |
1401 | return -1; | |
1402 | } | |
696ed9a9 | 1403 | |
df9ff5e1 | 1404 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
8be4620b | 1405 | RAMBlock *rb) |
696ed9a9 DDAG |
1406 | { |
1407 | assert(0); | |
1408 | return -1; | |
1409 | } | |
1410 | ||
df9ff5e1 | 1411 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
8be4620b | 1412 | RAMBlock *rb) |
696ed9a9 DDAG |
1413 | { |
1414 | assert(0); | |
1415 | return -1; | |
1416 | } | |
1417 | ||
5efc3564 DDAG |
1418 | int postcopy_wake_shared(struct PostCopyFD *pcfd, |
1419 | uint64_t client_addr, | |
1420 | RAMBlock *rb) | |
1421 | { | |
1422 | assert(0); | |
1423 | return -1; | |
1424 | } | |
eb59db53 DDAG |
1425 | #endif |
1426 | ||
e0b266f0 | 1427 | /* ------------------------------------------------------------------------- */ |
77dadc3f PX |
1428 | void postcopy_temp_page_reset(PostcopyTmpPage *tmp_page) |
1429 | { | |
1430 | tmp_page->target_pages = 0; | |
1431 | tmp_page->host_addr = NULL; | |
1432 | /* | |
1433 | * This is set to true when reset, and cleared as long as we received any | |
1434 | * of the non-zero small page within this huge page. | |
1435 | */ | |
1436 | tmp_page->all_zero = true; | |
1437 | } | |
e0b266f0 | 1438 | |
9ab7ef9b PX |
1439 | void postcopy_fault_thread_notify(MigrationIncomingState *mis) |
1440 | { | |
1441 | uint64_t tmp64 = 1; | |
1442 | ||
1443 | /* | |
1444 | * Wakeup the fault_thread. It's an eventfd that should currently | |
1445 | * be at 0, we're going to increment it to 1 | |
1446 | */ | |
1447 | if (write(mis->userfault_event_fd, &tmp64, 8) != 8) { | |
1448 | /* Not much we can do here, but may as well report it */ | |
1449 | error_report("%s: incrementing failed: %s", __func__, | |
1450 | strerror(errno)); | |
1451 | } | |
1452 | } | |
1453 | ||
e0b266f0 DDAG |
1454 | /** |
1455 | * postcopy_discard_send_init: Called at the start of each RAMBlock before | |
1456 | * asking to discard individual ranges. | |
1457 | * | |
1458 | * @ms: The current migration state. | |
810cf2bb | 1459 | * @offset: the bitmap offset of the named RAMBlock in the migration bitmap. |
e0b266f0 | 1460 | * @name: RAMBlock that discards will operate on. |
e0b266f0 | 1461 | */ |
810cf2bb WY |
1462 | static PostcopyDiscardState pds = {0}; |
1463 | void postcopy_discard_send_init(MigrationState *ms, const char *name) | |
e0b266f0 | 1464 | { |
810cf2bb WY |
1465 | pds.ramblock_name = name; |
1466 | pds.cur_entry = 0; | |
1467 | pds.nsentwords = 0; | |
1468 | pds.nsentcmds = 0; | |
e0b266f0 DDAG |
1469 | } |
1470 | ||
1471 | /** | |
1472 | * postcopy_discard_send_range: Called by the bitmap code for each chunk to | |
1473 | * discard. May send a discard message, may just leave it queued to | |
1474 | * be sent later. | |
1475 | * | |
1476 | * @ms: Current migration state. | |
e0b266f0 DDAG |
1477 | * @start,@length: a range of pages in the migration bitmap in the |
1478 | * RAM block passed to postcopy_discard_send_init() (length=1 is one page) | |
1479 | */ | |
810cf2bb WY |
1480 | void postcopy_discard_send_range(MigrationState *ms, unsigned long start, |
1481 | unsigned long length) | |
e0b266f0 | 1482 | { |
20afaed9 | 1483 | size_t tp_size = qemu_target_page_size(); |
e0b266f0 | 1484 | /* Convert to byte offsets within the RAM block */ |
810cf2bb WY |
1485 | pds.start_list[pds.cur_entry] = start * tp_size; |
1486 | pds.length_list[pds.cur_entry] = length * tp_size; | |
1487 | trace_postcopy_discard_send_range(pds.ramblock_name, start, length); | |
1488 | pds.cur_entry++; | |
1489 | pds.nsentwords++; | |
e0b266f0 | 1490 | |
810cf2bb | 1491 | if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) { |
e0b266f0 | 1492 | /* Full set, ship it! */ |
89a02a9f | 1493 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
810cf2bb WY |
1494 | pds.ramblock_name, |
1495 | pds.cur_entry, | |
1496 | pds.start_list, | |
1497 | pds.length_list); | |
1498 | pds.nsentcmds++; | |
1499 | pds.cur_entry = 0; | |
e0b266f0 DDAG |
1500 | } |
1501 | } | |
1502 | ||
1503 | /** | |
1504 | * postcopy_discard_send_finish: Called at the end of each RAMBlock by the | |
1505 | * bitmap code. Sends any outstanding discard messages, frees the PDS | |
1506 | * | |
1507 | * @ms: Current migration state. | |
e0b266f0 | 1508 | */ |
810cf2bb | 1509 | void postcopy_discard_send_finish(MigrationState *ms) |
e0b266f0 DDAG |
1510 | { |
1511 | /* Anything unsent? */ | |
810cf2bb | 1512 | if (pds.cur_entry) { |
89a02a9f | 1513 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
810cf2bb WY |
1514 | pds.ramblock_name, |
1515 | pds.cur_entry, | |
1516 | pds.start_list, | |
1517 | pds.length_list); | |
1518 | pds.nsentcmds++; | |
e0b266f0 DDAG |
1519 | } |
1520 | ||
810cf2bb WY |
1521 | trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords, |
1522 | pds.nsentcmds); | |
e0b266f0 | 1523 | } |
bac3b212 JQ |
1524 | |
1525 | /* | |
1526 | * Current state of incoming postcopy; note this is not part of | |
1527 | * MigrationIncomingState since it's state is used during cleanup | |
1528 | * at the end as MIS is being freed. | |
1529 | */ | |
1530 | static PostcopyState incoming_postcopy_state; | |
1531 | ||
1532 | PostcopyState postcopy_state_get(void) | |
1533 | { | |
4592eaf3 | 1534 | return qatomic_load_acquire(&incoming_postcopy_state); |
bac3b212 JQ |
1535 | } |
1536 | ||
1537 | /* Set the state and return the old state */ | |
1538 | PostcopyState postcopy_state_set(PostcopyState new_state) | |
1539 | { | |
d73415a3 | 1540 | return qatomic_xchg(&incoming_postcopy_state, new_state); |
bac3b212 | 1541 | } |
00fa4fc8 DDAG |
1542 | |
1543 | /* Register a handler for external shared memory postcopy | |
1544 | * called on the destination. | |
1545 | */ | |
1546 | void postcopy_register_shared_ufd(struct PostCopyFD *pcfd) | |
1547 | { | |
1548 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1549 | ||
1550 | mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds, | |
1551 | *pcfd); | |
1552 | } | |
1553 | ||
1554 | /* Unregister a handler for external shared memory postcopy | |
1555 | */ | |
1556 | void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd) | |
1557 | { | |
1558 | guint i; | |
1559 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1560 | GArray *pcrfds = mis->postcopy_remote_fds; | |
1561 | ||
56559980 JQ |
1562 | if (!pcrfds) { |
1563 | /* migration has already finished and freed the array */ | |
1564 | return; | |
1565 | } | |
00fa4fc8 DDAG |
1566 | for (i = 0; i < pcrfds->len; i++) { |
1567 | struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); | |
1568 | if (cur->fd == pcfd->fd) { | |
1569 | mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i); | |
1570 | return; | |
1571 | } | |
1572 | } | |
1573 | } | |
36f62f11 | 1574 | |
6720c2b3 | 1575 | void postcopy_preempt_new_channel(MigrationIncomingState *mis, QEMUFile *file) |
36f62f11 PX |
1576 | { |
1577 | /* | |
1578 | * The new loading channel has its own threads, so it needs to be | |
1579 | * blocked too. It's by default true, just be explicit. | |
1580 | */ | |
1581 | qemu_file_set_blocking(file, true); | |
1582 | mis->postcopy_qemufile_dst = file; | |
5655aab0 | 1583 | qemu_sem_post(&mis->postcopy_qemufile_dst_done); |
36f62f11 | 1584 | trace_postcopy_preempt_new_channel(); |
36f62f11 PX |
1585 | } |
1586 | ||
f0afaf6c PX |
1587 | /* |
1588 | * Setup the postcopy preempt channel with the IOC. If ERROR is specified, | |
1589 | * setup the error instead. This helper will free the ERROR if specified. | |
1590 | */ | |
d0edb8a1 | 1591 | static void |
f0afaf6c PX |
1592 | postcopy_preempt_send_channel_done(MigrationState *s, |
1593 | QIOChannel *ioc, Error *local_err) | |
36f62f11 | 1594 | { |
f0afaf6c | 1595 | if (local_err) { |
d0edb8a1 PX |
1596 | migrate_set_error(s, local_err); |
1597 | error_free(local_err); | |
1598 | } else { | |
1599 | migration_ioc_register_yank(ioc); | |
1600 | s->postcopy_qemufile_src = qemu_file_new_output(ioc); | |
1601 | trace_postcopy_preempt_new_channel(); | |
1602 | } | |
1603 | ||
1604 | /* | |
1605 | * Kick the waiter in all cases. The waiter should check upon | |
1606 | * postcopy_qemufile_src to know whether it failed or not. | |
1607 | */ | |
1608 | qemu_sem_post(&s->postcopy_qemufile_src_sem); | |
f0afaf6c PX |
1609 | } |
1610 | ||
1611 | static void | |
1612 | postcopy_preempt_tls_handshake(QIOTask *task, gpointer opaque) | |
1613 | { | |
1614 | g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task)); | |
1615 | MigrationState *s = opaque; | |
1616 | Error *local_err = NULL; | |
1617 | ||
1618 | qio_task_propagate_error(task, &local_err); | |
1619 | postcopy_preempt_send_channel_done(s, ioc, local_err); | |
1620 | } | |
1621 | ||
1622 | static void | |
1623 | postcopy_preempt_send_channel_new(QIOTask *task, gpointer opaque) | |
1624 | { | |
1625 | g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task)); | |
1626 | MigrationState *s = opaque; | |
1627 | QIOChannelTLS *tioc; | |
1628 | Error *local_err = NULL; | |
1629 | ||
1630 | if (qio_task_propagate_error(task, &local_err)) { | |
1631 | goto out; | |
1632 | } | |
1633 | ||
1634 | if (migrate_channel_requires_tls_upgrade(ioc)) { | |
0deb7e9b | 1635 | tioc = migration_tls_client_create(ioc, s->hostname, &local_err); |
f0afaf6c PX |
1636 | if (!tioc) { |
1637 | goto out; | |
1638 | } | |
1639 | trace_postcopy_preempt_tls_handshake(); | |
1640 | qio_channel_set_name(QIO_CHANNEL(tioc), "migration-tls-preempt"); | |
1641 | qio_channel_tls_handshake(tioc, postcopy_preempt_tls_handshake, | |
1642 | s, NULL, NULL); | |
1643 | /* Setup the channel until TLS handshake finished */ | |
1644 | return; | |
1645 | } | |
1646 | ||
1647 | out: | |
1648 | /* This handles both good and error cases */ | |
1649 | postcopy_preempt_send_channel_done(s, ioc, local_err); | |
d0edb8a1 | 1650 | } |
36f62f11 | 1651 | |
5655aab0 PX |
1652 | /* |
1653 | * This function will kick off an async task to establish the preempt | |
1654 | * channel, and wait until the connection setup completed. Returns 0 if | |
1655 | * channel established, -1 for error. | |
1656 | */ | |
1657 | int postcopy_preempt_establish_channel(MigrationState *s) | |
d0edb8a1 PX |
1658 | { |
1659 | /* If preempt not enabled, no need to wait */ | |
1660 | if (!migrate_postcopy_preempt()) { | |
1661 | return 0; | |
1662 | } | |
1663 | ||
06064a67 PX |
1664 | /* |
1665 | * Kick off async task to establish preempt channel. Only do so with | |
1666 | * 8.0+ machines, because 7.1/7.2 require the channel to be created in | |
1667 | * setup phase of migration (even if racy in an unreliable network). | |
1668 | */ | |
1669 | if (!s->preempt_pre_7_2) { | |
1670 | postcopy_preempt_setup(s); | |
1671 | } | |
5655aab0 | 1672 | |
d0edb8a1 PX |
1673 | /* |
1674 | * We need the postcopy preempt channel to be established before | |
1675 | * starting doing anything. | |
1676 | */ | |
1677 | qemu_sem_wait(&s->postcopy_qemufile_src_sem); | |
1678 | ||
1679 | return s->postcopy_qemufile_src ? 0 : -1; | |
1680 | } | |
1681 | ||
fc063a7b | 1682 | void postcopy_preempt_setup(MigrationState *s) |
d0edb8a1 | 1683 | { |
d0edb8a1 PX |
1684 | /* Kick an async task to connect */ |
1685 | socket_send_channel_create(postcopy_preempt_send_channel_new, s); | |
36f62f11 PX |
1686 | } |
1687 | ||
60bb3c58 PX |
1688 | static void postcopy_pause_ram_fast_load(MigrationIncomingState *mis) |
1689 | { | |
1690 | trace_postcopy_pause_fast_load(); | |
1691 | qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex); | |
1692 | qemu_sem_wait(&mis->postcopy_pause_sem_fast_load); | |
1693 | qemu_mutex_lock(&mis->postcopy_prio_thread_mutex); | |
1694 | trace_postcopy_pause_fast_load_continued(); | |
1695 | } | |
1696 | ||
6621883f PX |
1697 | static bool preempt_thread_should_run(MigrationIncomingState *mis) |
1698 | { | |
1699 | return mis->preempt_thread_status != PREEMPT_THREAD_QUIT; | |
1700 | } | |
1701 | ||
36f62f11 PX |
1702 | void *postcopy_preempt_thread(void *opaque) |
1703 | { | |
1704 | MigrationIncomingState *mis = opaque; | |
1705 | int ret; | |
1706 | ||
1707 | trace_postcopy_preempt_thread_entry(); | |
1708 | ||
1709 | rcu_register_thread(); | |
1710 | ||
1711 | qemu_sem_post(&mis->thread_sync_sem); | |
1712 | ||
a5d35dc7 PX |
1713 | /* |
1714 | * The preempt channel is established in asynchronous way. Wait | |
1715 | * for its completion. | |
1716 | */ | |
1717 | qemu_sem_wait(&mis->postcopy_qemufile_dst_done); | |
1718 | ||
36f62f11 | 1719 | /* Sending RAM_SAVE_FLAG_EOS to terminate this thread */ |
60bb3c58 | 1720 | qemu_mutex_lock(&mis->postcopy_prio_thread_mutex); |
6621883f | 1721 | while (preempt_thread_should_run(mis)) { |
60bb3c58 PX |
1722 | ret = ram_load_postcopy(mis->postcopy_qemufile_dst, |
1723 | RAM_CHANNEL_POSTCOPY); | |
1724 | /* If error happened, go into recovery routine */ | |
6621883f | 1725 | if (ret && preempt_thread_should_run(mis)) { |
60bb3c58 PX |
1726 | postcopy_pause_ram_fast_load(mis); |
1727 | } else { | |
1728 | /* We're done */ | |
1729 | break; | |
1730 | } | |
1731 | } | |
1732 | qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex); | |
36f62f11 PX |
1733 | |
1734 | rcu_unregister_thread(); | |
1735 | ||
1736 | trace_postcopy_preempt_thread_exit(); | |
1737 | ||
60bb3c58 | 1738 | return NULL; |
36f62f11 | 1739 | } |