]>
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" |
51180423 | 20 | #include "exec/target_page.h" |
6666c96a | 21 | #include "migration.h" |
08a0aee1 | 22 | #include "qemu-file.h" |
20a519a0 | 23 | #include "savevm.h" |
be07b0ac | 24 | #include "postcopy-ram.h" |
7b1e1a22 | 25 | #include "ram.h" |
1693c64c DDAG |
26 | #include "qapi/error.h" |
27 | #include "qemu/notify.h" | |
d4842052 | 28 | #include "qemu/rcu.h" |
eb59db53 DDAG |
29 | #include "sysemu/sysemu.h" |
30 | #include "qemu/error-report.h" | |
31 | #include "trace.h" | |
5cc8767d | 32 | #include "hw/boards.h" |
eb59db53 | 33 | |
e0b266f0 DDAG |
34 | /* Arbitrary limit on size of each discard command, |
35 | * keeps them around ~200 bytes | |
36 | */ | |
37 | #define MAX_DISCARDS_PER_COMMAND 12 | |
38 | ||
39 | struct PostcopyDiscardState { | |
40 | const char *ramblock_name; | |
e0b266f0 DDAG |
41 | uint16_t cur_entry; |
42 | /* | |
43 | * Start and length of a discard range (bytes) | |
44 | */ | |
45 | uint64_t start_list[MAX_DISCARDS_PER_COMMAND]; | |
46 | uint64_t length_list[MAX_DISCARDS_PER_COMMAND]; | |
47 | unsigned int nsentwords; | |
48 | unsigned int nsentcmds; | |
49 | }; | |
50 | ||
1693c64c DDAG |
51 | static NotifierWithReturnList postcopy_notifier_list; |
52 | ||
53 | void postcopy_infrastructure_init(void) | |
54 | { | |
55 | notifier_with_return_list_init(&postcopy_notifier_list); | |
56 | } | |
57 | ||
58 | void postcopy_add_notifier(NotifierWithReturn *nn) | |
59 | { | |
60 | notifier_with_return_list_add(&postcopy_notifier_list, nn); | |
61 | } | |
62 | ||
63 | void postcopy_remove_notifier(NotifierWithReturn *n) | |
64 | { | |
65 | notifier_with_return_remove(n); | |
66 | } | |
67 | ||
68 | int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp) | |
69 | { | |
70 | struct PostcopyNotifyData pnd; | |
71 | pnd.reason = reason; | |
72 | pnd.errp = errp; | |
73 | ||
74 | return notifier_with_return_list_notify(&postcopy_notifier_list, | |
75 | &pnd); | |
76 | } | |
77 | ||
eb59db53 DDAG |
78 | /* Postcopy needs to detect accesses to pages that haven't yet been copied |
79 | * across, and efficiently map new pages in, the techniques for doing this | |
80 | * are target OS specific. | |
81 | */ | |
82 | #if defined(__linux__) | |
83 | ||
c4faeed2 | 84 | #include <poll.h> |
eb59db53 DDAG |
85 | #include <sys/ioctl.h> |
86 | #include <sys/syscall.h> | |
eb59db53 DDAG |
87 | #include <asm/types.h> /* for __u64 */ |
88 | #endif | |
89 | ||
d8b9d771 MF |
90 | #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD) |
91 | #include <sys/eventfd.h> | |
eb59db53 DDAG |
92 | #include <linux/userfaultfd.h> |
93 | ||
2a4c42f1 AP |
94 | typedef struct PostcopyBlocktimeContext { |
95 | /* time when page fault initiated per vCPU */ | |
96 | uint32_t *page_fault_vcpu_time; | |
97 | /* page address per vCPU */ | |
98 | uintptr_t *vcpu_addr; | |
99 | uint32_t total_blocktime; | |
100 | /* blocktime per vCPU */ | |
101 | uint32_t *vcpu_blocktime; | |
102 | /* point in time when last page fault was initiated */ | |
103 | uint32_t last_begin; | |
104 | /* number of vCPU are suspended */ | |
105 | int smp_cpus_down; | |
106 | uint64_t start_time; | |
107 | ||
108 | /* | |
109 | * Handler for exit event, necessary for | |
110 | * releasing whole blocktime_ctx | |
111 | */ | |
112 | Notifier exit_notifier; | |
113 | } PostcopyBlocktimeContext; | |
114 | ||
115 | static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx) | |
116 | { | |
117 | g_free(ctx->page_fault_vcpu_time); | |
118 | g_free(ctx->vcpu_addr); | |
119 | g_free(ctx->vcpu_blocktime); | |
120 | g_free(ctx); | |
121 | } | |
122 | ||
123 | static void migration_exit_cb(Notifier *n, void *data) | |
124 | { | |
125 | PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext, | |
126 | exit_notifier); | |
127 | destroy_blocktime_context(ctx); | |
128 | } | |
129 | ||
130 | static struct PostcopyBlocktimeContext *blocktime_context_new(void) | |
131 | { | |
5cc8767d LX |
132 | MachineState *ms = MACHINE(qdev_get_machine()); |
133 | unsigned int smp_cpus = ms->smp.cpus; | |
2a4c42f1 AP |
134 | PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1); |
135 | ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus); | |
136 | ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus); | |
137 | ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus); | |
138 | ||
139 | ctx->exit_notifier.notify = migration_exit_cb; | |
140 | ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); | |
141 | qemu_add_exit_notifier(&ctx->exit_notifier); | |
142 | return ctx; | |
143 | } | |
ca6011c2 | 144 | |
65ace060 AP |
145 | static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx) |
146 | { | |
5cc8767d | 147 | MachineState *ms = MACHINE(qdev_get_machine()); |
54aa3de7 | 148 | uint32List *list = NULL; |
65ace060 AP |
149 | int i; |
150 | ||
5cc8767d | 151 | for (i = ms->smp.cpus - 1; i >= 0; i--) { |
54aa3de7 | 152 | QAPI_LIST_PREPEND(list, ctx->vcpu_blocktime[i]); |
65ace060 AP |
153 | } |
154 | ||
155 | return list; | |
156 | } | |
157 | ||
158 | /* | |
159 | * This function just populates MigrationInfo from postcopy's | |
160 | * blocktime context. It will not populate MigrationInfo, | |
161 | * unless postcopy-blocktime capability was set. | |
162 | * | |
163 | * @info: pointer to MigrationInfo to populate | |
164 | */ | |
165 | void fill_destination_postcopy_migration_info(MigrationInfo *info) | |
166 | { | |
167 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
168 | PostcopyBlocktimeContext *bc = mis->blocktime_ctx; | |
169 | ||
170 | if (!bc) { | |
171 | return; | |
172 | } | |
173 | ||
174 | info->has_postcopy_blocktime = true; | |
175 | info->postcopy_blocktime = bc->total_blocktime; | |
176 | info->has_postcopy_vcpu_blocktime = true; | |
177 | info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc); | |
178 | } | |
179 | ||
180 | static uint32_t get_postcopy_total_blocktime(void) | |
181 | { | |
182 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
183 | PostcopyBlocktimeContext *bc = mis->blocktime_ctx; | |
184 | ||
185 | if (!bc) { | |
186 | return 0; | |
187 | } | |
188 | ||
189 | return bc->total_blocktime; | |
190 | } | |
191 | ||
54ae0886 AP |
192 | /** |
193 | * receive_ufd_features: check userfault fd features, to request only supported | |
194 | * features in the future. | |
195 | * | |
196 | * Returns: true on success | |
197 | * | |
198 | * __NR_userfaultfd - should be checked before | |
199 | * @features: out parameter will contain uffdio_api.features provided by kernel | |
200 | * in case of success | |
201 | */ | |
202 | static bool receive_ufd_features(uint64_t *features) | |
eb59db53 | 203 | { |
54ae0886 AP |
204 | struct uffdio_api api_struct = {0}; |
205 | int ufd; | |
206 | bool ret = true; | |
207 | ||
208 | /* if we are here __NR_userfaultfd should exists */ | |
209 | ufd = syscall(__NR_userfaultfd, O_CLOEXEC); | |
210 | if (ufd == -1) { | |
211 | error_report("%s: syscall __NR_userfaultfd failed: %s", __func__, | |
212 | strerror(errno)); | |
213 | return false; | |
214 | } | |
eb59db53 | 215 | |
54ae0886 | 216 | /* ask features */ |
eb59db53 DDAG |
217 | api_struct.api = UFFD_API; |
218 | api_struct.features = 0; | |
219 | if (ioctl(ufd, UFFDIO_API, &api_struct)) { | |
5553499f | 220 | error_report("%s: UFFDIO_API failed: %s", __func__, |
eb59db53 | 221 | strerror(errno)); |
54ae0886 AP |
222 | ret = false; |
223 | goto release_ufd; | |
224 | } | |
225 | ||
226 | *features = api_struct.features; | |
227 | ||
228 | release_ufd: | |
229 | close(ufd); | |
230 | return ret; | |
231 | } | |
232 | ||
233 | /** | |
234 | * request_ufd_features: this function should be called only once on a newly | |
235 | * opened ufd, subsequent calls will lead to error. | |
236 | * | |
3a4452d8 | 237 | * Returns: true on success |
54ae0886 AP |
238 | * |
239 | * @ufd: fd obtained from userfaultfd syscall | |
240 | * @features: bit mask see UFFD_API_FEATURES | |
241 | */ | |
242 | static bool request_ufd_features(int ufd, uint64_t features) | |
243 | { | |
244 | struct uffdio_api api_struct = {0}; | |
245 | uint64_t ioctl_mask; | |
246 | ||
247 | api_struct.api = UFFD_API; | |
248 | api_struct.features = features; | |
249 | if (ioctl(ufd, UFFDIO_API, &api_struct)) { | |
250 | error_report("%s failed: UFFDIO_API failed: %s", __func__, | |
251 | strerror(errno)); | |
eb59db53 DDAG |
252 | return false; |
253 | } | |
254 | ||
255 | ioctl_mask = (__u64)1 << _UFFDIO_REGISTER | | |
256 | (__u64)1 << _UFFDIO_UNREGISTER; | |
257 | if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) { | |
258 | error_report("Missing userfault features: %" PRIx64, | |
259 | (uint64_t)(~api_struct.ioctls & ioctl_mask)); | |
260 | return false; | |
261 | } | |
262 | ||
54ae0886 AP |
263 | return true; |
264 | } | |
265 | ||
266 | static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis) | |
267 | { | |
268 | uint64_t asked_features = 0; | |
269 | static uint64_t supported_features; | |
270 | ||
271 | /* | |
272 | * it's not possible to | |
273 | * request UFFD_API twice per one fd | |
274 | * userfault fd features is persistent | |
275 | */ | |
276 | if (!supported_features) { | |
277 | if (!receive_ufd_features(&supported_features)) { | |
278 | error_report("%s failed", __func__); | |
279 | return false; | |
280 | } | |
281 | } | |
282 | ||
2a4c42f1 AP |
283 | #ifdef UFFD_FEATURE_THREAD_ID |
284 | if (migrate_postcopy_blocktime() && mis && | |
285 | UFFD_FEATURE_THREAD_ID & supported_features) { | |
286 | /* kernel supports that feature */ | |
287 | /* don't create blocktime_context if it exists */ | |
288 | if (!mis->blocktime_ctx) { | |
289 | mis->blocktime_ctx = blocktime_context_new(); | |
290 | } | |
291 | ||
292 | asked_features |= UFFD_FEATURE_THREAD_ID; | |
293 | } | |
294 | #endif | |
295 | ||
54ae0886 AP |
296 | /* |
297 | * request features, even if asked_features is 0, due to | |
298 | * kernel expects UFFD_API before UFFDIO_REGISTER, per | |
299 | * userfault file descriptor | |
300 | */ | |
301 | if (!request_ufd_features(ufd, asked_features)) { | |
302 | error_report("%s failed: features %" PRIu64, __func__, | |
303 | asked_features); | |
304 | return false; | |
305 | } | |
306 | ||
038adc2f | 307 | if (qemu_real_host_page_size != ram_pagesize_summary()) { |
7e8cafb7 DDAG |
308 | bool have_hp = false; |
309 | /* We've got a huge page */ | |
310 | #ifdef UFFD_FEATURE_MISSING_HUGETLBFS | |
54ae0886 | 311 | have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS; |
7e8cafb7 DDAG |
312 | #endif |
313 | if (!have_hp) { | |
314 | error_report("Userfault on this host does not support huge pages"); | |
315 | return false; | |
316 | } | |
317 | } | |
eb59db53 DDAG |
318 | return true; |
319 | } | |
320 | ||
8679638b DDAG |
321 | /* Callback from postcopy_ram_supported_by_host block iterator. |
322 | */ | |
754cb9c0 | 323 | static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque) |
8679638b | 324 | { |
754cb9c0 YK |
325 | const char *block_name = qemu_ram_get_idstr(rb); |
326 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
5d214a92 DDAG |
327 | size_t pagesize = qemu_ram_pagesize(rb); |
328 | ||
5d214a92 DDAG |
329 | if (length % pagesize) { |
330 | error_report("Postcopy requires RAM blocks to be a page size multiple," | |
331 | " block %s is 0x" RAM_ADDR_FMT " bytes with a " | |
332 | "page size of 0x%zx", block_name, length, pagesize); | |
333 | return 1; | |
334 | } | |
8679638b DDAG |
335 | return 0; |
336 | } | |
337 | ||
58b7c17e DDAG |
338 | /* |
339 | * Note: This has the side effect of munlock'ing all of RAM, that's | |
340 | * normally fine since if the postcopy succeeds it gets turned back on at the | |
341 | * end. | |
342 | */ | |
d7651f15 | 343 | bool postcopy_ram_supported_by_host(MigrationIncomingState *mis) |
eb59db53 | 344 | { |
038adc2f | 345 | long pagesize = qemu_real_host_page_size; |
eb59db53 DDAG |
346 | int ufd = -1; |
347 | bool ret = false; /* Error unless we change it */ | |
348 | void *testarea = NULL; | |
349 | struct uffdio_register reg_struct; | |
350 | struct uffdio_range range_struct; | |
351 | uint64_t feature_mask; | |
1693c64c | 352 | Error *local_err = NULL; |
eb59db53 | 353 | |
20afaed9 | 354 | if (qemu_target_page_size() > pagesize) { |
eb59db53 DDAG |
355 | error_report("Target page size bigger than host page size"); |
356 | goto out; | |
357 | } | |
358 | ||
359 | ufd = syscall(__NR_userfaultfd, O_CLOEXEC); | |
360 | if (ufd == -1) { | |
361 | error_report("%s: userfaultfd not available: %s", __func__, | |
362 | strerror(errno)); | |
363 | goto out; | |
364 | } | |
365 | ||
1693c64c DDAG |
366 | /* Give devices a chance to object */ |
367 | if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) { | |
368 | error_report_err(local_err); | |
369 | goto out; | |
370 | } | |
371 | ||
eb59db53 | 372 | /* Version and features check */ |
54ae0886 | 373 | if (!ufd_check_and_apply(ufd, mis)) { |
eb59db53 DDAG |
374 | goto out; |
375 | } | |
376 | ||
8679638b | 377 | /* We don't support postcopy with shared RAM yet */ |
fbd162e6 | 378 | if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) { |
8679638b DDAG |
379 | goto out; |
380 | } | |
381 | ||
58b7c17e DDAG |
382 | /* |
383 | * userfault and mlock don't go together; we'll put it back later if | |
384 | * it was enabled. | |
385 | */ | |
386 | if (munlockall()) { | |
387 | error_report("%s: munlockall: %s", __func__, strerror(errno)); | |
617a32f5 | 388 | goto out; |
58b7c17e DDAG |
389 | } |
390 | ||
eb59db53 DDAG |
391 | /* |
392 | * We need to check that the ops we need are supported on anon memory | |
393 | * To do that we need to register a chunk and see the flags that | |
394 | * are returned. | |
395 | */ | |
396 | testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | | |
397 | MAP_ANONYMOUS, -1, 0); | |
398 | if (testarea == MAP_FAILED) { | |
399 | error_report("%s: Failed to map test area: %s", __func__, | |
400 | strerror(errno)); | |
401 | goto out; | |
402 | } | |
395cb450 | 403 | g_assert(((size_t)testarea & (pagesize - 1)) == 0); |
eb59db53 DDAG |
404 | |
405 | reg_struct.range.start = (uintptr_t)testarea; | |
406 | reg_struct.range.len = pagesize; | |
407 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; | |
408 | ||
409 | if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) { | |
410 | error_report("%s userfault register: %s", __func__, strerror(errno)); | |
411 | goto out; | |
412 | } | |
413 | ||
414 | range_struct.start = (uintptr_t)testarea; | |
415 | range_struct.len = pagesize; | |
416 | if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) { | |
417 | error_report("%s userfault unregister: %s", __func__, strerror(errno)); | |
418 | goto out; | |
419 | } | |
420 | ||
421 | feature_mask = (__u64)1 << _UFFDIO_WAKE | | |
422 | (__u64)1 << _UFFDIO_COPY | | |
423 | (__u64)1 << _UFFDIO_ZEROPAGE; | |
424 | if ((reg_struct.ioctls & feature_mask) != feature_mask) { | |
425 | error_report("Missing userfault map features: %" PRIx64, | |
426 | (uint64_t)(~reg_struct.ioctls & feature_mask)); | |
427 | goto out; | |
428 | } | |
429 | ||
430 | /* Success! */ | |
431 | ret = true; | |
432 | out: | |
433 | if (testarea) { | |
434 | munmap(testarea, pagesize); | |
435 | } | |
436 | if (ufd != -1) { | |
437 | close(ufd); | |
438 | } | |
439 | return ret; | |
440 | } | |
441 | ||
1caddf8a DDAG |
442 | /* |
443 | * Setup an area of RAM so that it *can* be used for postcopy later; this | |
444 | * must be done right at the start prior to pre-copy. | |
445 | * opaque should be the MIS. | |
446 | */ | |
754cb9c0 | 447 | static int init_range(RAMBlock *rb, void *opaque) |
1caddf8a | 448 | { |
754cb9c0 YK |
449 | const char *block_name = qemu_ram_get_idstr(rb); |
450 | void *host_addr = qemu_ram_get_host_addr(rb); | |
451 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
452 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
1caddf8a DDAG |
453 | trace_postcopy_init_range(block_name, host_addr, offset, length); |
454 | ||
455 | /* | |
456 | * We need the whole of RAM to be truly empty for postcopy, so things | |
457 | * like ROMs and any data tables built during init must be zero'd | |
458 | * - we're going to get the copy from the source anyway. | |
459 | * (Precopy will just overwrite this data, so doesn't need the discard) | |
460 | */ | |
aaa2064c | 461 | if (ram_discard_range(block_name, 0, length)) { |
1caddf8a DDAG |
462 | return -1; |
463 | } | |
464 | ||
465 | return 0; | |
466 | } | |
467 | ||
468 | /* | |
469 | * At the end of migration, undo the effects of init_range | |
470 | * opaque should be the MIS. | |
471 | */ | |
754cb9c0 | 472 | static int cleanup_range(RAMBlock *rb, void *opaque) |
1caddf8a | 473 | { |
754cb9c0 YK |
474 | const char *block_name = qemu_ram_get_idstr(rb); |
475 | void *host_addr = qemu_ram_get_host_addr(rb); | |
476 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
477 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
1caddf8a DDAG |
478 | MigrationIncomingState *mis = opaque; |
479 | struct uffdio_range range_struct; | |
480 | trace_postcopy_cleanup_range(block_name, host_addr, offset, length); | |
481 | ||
482 | /* | |
483 | * We turned off hugepage for the precopy stage with postcopy enabled | |
484 | * we can turn it back on now. | |
485 | */ | |
1d741439 | 486 | qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE); |
1caddf8a DDAG |
487 | |
488 | /* | |
489 | * We can also turn off userfault now since we should have all the | |
490 | * pages. It can be useful to leave it on to debug postcopy | |
491 | * if you're not sure it's always getting every page. | |
492 | */ | |
493 | range_struct.start = (uintptr_t)host_addr; | |
494 | range_struct.len = length; | |
495 | ||
496 | if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) { | |
497 | error_report("%s: userfault unregister %s", __func__, strerror(errno)); | |
498 | ||
499 | return -1; | |
500 | } | |
501 | ||
502 | return 0; | |
503 | } | |
504 | ||
505 | /* | |
506 | * Initialise postcopy-ram, setting the RAM to a state where we can go into | |
507 | * postcopy later; must be called prior to any precopy. | |
508 | * called from arch_init's similarly named ram_postcopy_incoming_init | |
509 | */ | |
c136180c | 510 | int postcopy_ram_incoming_init(MigrationIncomingState *mis) |
1caddf8a | 511 | { |
fbd162e6 | 512 | if (foreach_not_ignored_block(init_range, NULL)) { |
1caddf8a DDAG |
513 | return -1; |
514 | } | |
515 | ||
516 | return 0; | |
517 | } | |
518 | ||
519 | /* | |
520 | * At the end of a migration where postcopy_ram_incoming_init was called. | |
521 | */ | |
522 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
523 | { | |
c4faeed2 DDAG |
524 | trace_postcopy_ram_incoming_cleanup_entry(); |
525 | ||
526 | if (mis->have_fault_thread) { | |
46343570 DDAG |
527 | Error *local_err = NULL; |
528 | ||
55d0fe82 | 529 | /* Let the fault thread quit */ |
d73415a3 | 530 | qatomic_set(&mis->fault_thread_quit, 1); |
55d0fe82 IM |
531 | postcopy_fault_thread_notify(mis); |
532 | trace_postcopy_ram_incoming_cleanup_join(); | |
533 | qemu_thread_join(&mis->fault_thread); | |
534 | ||
46343570 DDAG |
535 | if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) { |
536 | error_report_err(local_err); | |
537 | return -1; | |
538 | } | |
539 | ||
fbd162e6 | 540 | if (foreach_not_ignored_block(cleanup_range, mis)) { |
c4faeed2 DDAG |
541 | return -1; |
542 | } | |
9ab7ef9b | 543 | |
c4faeed2 DDAG |
544 | trace_postcopy_ram_incoming_cleanup_closeuf(); |
545 | close(mis->userfault_fd); | |
64f615fe | 546 | close(mis->userfault_event_fd); |
c4faeed2 | 547 | mis->have_fault_thread = false; |
1caddf8a DDAG |
548 | } |
549 | ||
58b7c17e DDAG |
550 | if (enable_mlock) { |
551 | if (os_mlock() < 0) { | |
552 | error_report("mlock: %s", strerror(errno)); | |
553 | /* | |
554 | * It doesn't feel right to fail at this point, we have a valid | |
555 | * VM state. | |
556 | */ | |
557 | } | |
558 | } | |
559 | ||
696ed9a9 | 560 | if (mis->postcopy_tmp_page) { |
df9ff5e1 | 561 | munmap(mis->postcopy_tmp_page, mis->largest_page_size); |
696ed9a9 DDAG |
562 | mis->postcopy_tmp_page = NULL; |
563 | } | |
41d84210 DDAG |
564 | if (mis->postcopy_tmp_zero_page) { |
565 | munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size); | |
566 | mis->postcopy_tmp_zero_page = NULL; | |
567 | } | |
65ace060 AP |
568 | trace_postcopy_ram_incoming_cleanup_blocktime( |
569 | get_postcopy_total_blocktime()); | |
570 | ||
c4faeed2 | 571 | trace_postcopy_ram_incoming_cleanup_exit(); |
1caddf8a DDAG |
572 | return 0; |
573 | } | |
574 | ||
f9527107 DDAG |
575 | /* |
576 | * Disable huge pages on an area | |
577 | */ | |
754cb9c0 | 578 | static int nhp_range(RAMBlock *rb, void *opaque) |
f9527107 | 579 | { |
754cb9c0 YK |
580 | const char *block_name = qemu_ram_get_idstr(rb); |
581 | void *host_addr = qemu_ram_get_host_addr(rb); | |
582 | ram_addr_t offset = qemu_ram_get_offset(rb); | |
583 | ram_addr_t length = qemu_ram_get_used_length(rb); | |
f9527107 DDAG |
584 | trace_postcopy_nhp_range(block_name, host_addr, offset, length); |
585 | ||
586 | /* | |
587 | * Before we do discards we need to ensure those discards really | |
588 | * do delete areas of the page, even if THP thinks a hugepage would | |
589 | * be a good idea, so force hugepages off. | |
590 | */ | |
1d741439 | 591 | qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE); |
f9527107 DDAG |
592 | |
593 | return 0; | |
594 | } | |
595 | ||
596 | /* | |
597 | * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard | |
598 | * however leaving it until after precopy means that most of the precopy | |
599 | * data is still THPd | |
600 | */ | |
601 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) | |
602 | { | |
fbd162e6 | 603 | if (foreach_not_ignored_block(nhp_range, mis)) { |
f9527107 DDAG |
604 | return -1; |
605 | } | |
606 | ||
607 | postcopy_state_set(POSTCOPY_INCOMING_DISCARD); | |
608 | ||
609 | return 0; | |
610 | } | |
611 | ||
f0a227ad DDAG |
612 | /* |
613 | * Mark the given area of RAM as requiring notification to unwritten areas | |
fbd162e6 | 614 | * Used as a callback on foreach_not_ignored_block. |
f0a227ad DDAG |
615 | * host_addr: Base of area to mark |
616 | * offset: Offset in the whole ram arena | |
617 | * length: Length of the section | |
618 | * opaque: MigrationIncomingState pointer | |
619 | * Returns 0 on success | |
620 | */ | |
754cb9c0 | 621 | static int ram_block_enable_notify(RAMBlock *rb, void *opaque) |
f0a227ad DDAG |
622 | { |
623 | MigrationIncomingState *mis = opaque; | |
624 | struct uffdio_register reg_struct; | |
625 | ||
754cb9c0 YK |
626 | reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb); |
627 | reg_struct.range.len = qemu_ram_get_used_length(rb); | |
f0a227ad DDAG |
628 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; |
629 | ||
630 | /* Now tell our userfault_fd that it's responsible for this area */ | |
631 | if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) { | |
632 | error_report("%s userfault register: %s", __func__, strerror(errno)); | |
633 | return -1; | |
634 | } | |
665414ad DDAG |
635 | if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { |
636 | error_report("%s userfault: Region doesn't support COPY", __func__); | |
637 | return -1; | |
638 | } | |
2ce16640 | 639 | if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) { |
2ce16640 DDAG |
640 | qemu_ram_set_uf_zeroable(rb); |
641 | } | |
f0a227ad DDAG |
642 | |
643 | return 0; | |
644 | } | |
645 | ||
5efc3564 DDAG |
646 | int postcopy_wake_shared(struct PostCopyFD *pcfd, |
647 | uint64_t client_addr, | |
648 | RAMBlock *rb) | |
649 | { | |
650 | size_t pagesize = qemu_ram_pagesize(rb); | |
651 | struct uffdio_range range; | |
652 | int ret; | |
653 | trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb)); | |
654 | range.start = client_addr & ~(pagesize - 1); | |
655 | range.len = pagesize; | |
656 | ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range); | |
657 | if (ret) { | |
658 | error_report("%s: Failed to wake: %zx in %s (%s)", | |
659 | __func__, (size_t)client_addr, qemu_ram_get_idstr(rb), | |
660 | strerror(errno)); | |
661 | } | |
662 | return ret; | |
663 | } | |
664 | ||
096bf4c8 DDAG |
665 | /* |
666 | * Callback from shared fault handlers to ask for a page, | |
667 | * the page must be specified by a RAMBlock and an offset in that rb | |
668 | * Note: Only for use by shared fault handlers (in fault thread) | |
669 | */ | |
670 | int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, | |
671 | uint64_t client_addr, uint64_t rb_offset) | |
672 | { | |
673 | size_t pagesize = qemu_ram_pagesize(rb); | |
674 | uint64_t aligned_rbo = rb_offset & ~(pagesize - 1); | |
675 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
676 | ||
677 | trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb), | |
678 | rb_offset); | |
dedfb4b2 DDAG |
679 | if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) { |
680 | trace_postcopy_request_shared_page_present(pcfd->idstr, | |
681 | qemu_ram_get_idstr(rb), rb_offset); | |
682 | return postcopy_wake_shared(pcfd, client_addr, rb); | |
683 | } | |
8f8bfffc | 684 | migrate_send_rp_req_pages(mis, rb, aligned_rbo, client_addr); |
096bf4c8 DDAG |
685 | return 0; |
686 | } | |
687 | ||
575b0b33 AP |
688 | static int get_mem_fault_cpu_index(uint32_t pid) |
689 | { | |
690 | CPUState *cpu_iter; | |
691 | ||
692 | CPU_FOREACH(cpu_iter) { | |
693 | if (cpu_iter->thread_id == pid) { | |
694 | trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid); | |
695 | return cpu_iter->cpu_index; | |
696 | } | |
697 | } | |
698 | trace_get_mem_fault_cpu_index(-1, pid); | |
699 | return -1; | |
700 | } | |
701 | ||
702 | static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc) | |
703 | { | |
704 | int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - | |
705 | dc->start_time; | |
706 | return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX; | |
707 | } | |
708 | ||
709 | /* | |
710 | * This function is being called when pagefault occurs. It | |
711 | * tracks down vCPU blocking time. | |
712 | * | |
713 | * @addr: faulted host virtual address | |
714 | * @ptid: faulted process thread id | |
715 | * @rb: ramblock appropriate to addr | |
716 | */ | |
717 | static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid, | |
718 | RAMBlock *rb) | |
719 | { | |
720 | int cpu, already_received; | |
721 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
722 | PostcopyBlocktimeContext *dc = mis->blocktime_ctx; | |
723 | uint32_t low_time_offset; | |
724 | ||
725 | if (!dc || ptid == 0) { | |
726 | return; | |
727 | } | |
728 | cpu = get_mem_fault_cpu_index(ptid); | |
729 | if (cpu < 0) { | |
730 | return; | |
731 | } | |
732 | ||
733 | low_time_offset = get_low_time_offset(dc); | |
734 | if (dc->vcpu_addr[cpu] == 0) { | |
d73415a3 | 735 | qatomic_inc(&dc->smp_cpus_down); |
575b0b33 AP |
736 | } |
737 | ||
d73415a3 SH |
738 | qatomic_xchg(&dc->last_begin, low_time_offset); |
739 | qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset); | |
740 | qatomic_xchg(&dc->vcpu_addr[cpu], addr); | |
575b0b33 | 741 | |
da1725d3 WY |
742 | /* |
743 | * check it here, not at the beginning of the function, | |
744 | * due to, check could occur early than bitmap_set in | |
745 | * qemu_ufd_copy_ioctl | |
746 | */ | |
575b0b33 AP |
747 | already_received = ramblock_recv_bitmap_test(rb, (void *)addr); |
748 | if (already_received) { | |
d73415a3 SH |
749 | qatomic_xchg(&dc->vcpu_addr[cpu], 0); |
750 | qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0); | |
751 | qatomic_dec(&dc->smp_cpus_down); | |
575b0b33 AP |
752 | } |
753 | trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu], | |
754 | cpu, already_received); | |
755 | } | |
756 | ||
757 | /* | |
758 | * This function just provide calculated blocktime per cpu and trace it. | |
759 | * Total blocktime is calculated in mark_postcopy_blocktime_end. | |
760 | * | |
761 | * | |
762 | * Assume we have 3 CPU | |
763 | * | |
764 | * S1 E1 S1 E1 | |
765 | * -----***********------------xxx***************------------------------> CPU1 | |
766 | * | |
767 | * S2 E2 | |
768 | * ------------****************xxx---------------------------------------> CPU2 | |
769 | * | |
770 | * S3 E3 | |
771 | * ------------------------****xxx********-------------------------------> CPU3 | |
772 | * | |
773 | * We have sequence S1,S2,E1,S3,S1,E2,E3,E1 | |
774 | * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3 | |
775 | * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 - | |
776 | * it's a part of total blocktime. | |
777 | * S1 - here is last_begin | |
778 | * Legend of the picture is following: | |
779 | * * - means blocktime per vCPU | |
780 | * x - means overlapped blocktime (total blocktime) | |
781 | * | |
782 | * @addr: host virtual address | |
783 | */ | |
784 | static void mark_postcopy_blocktime_end(uintptr_t addr) | |
785 | { | |
786 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
787 | PostcopyBlocktimeContext *dc = mis->blocktime_ctx; | |
5cc8767d LX |
788 | MachineState *ms = MACHINE(qdev_get_machine()); |
789 | unsigned int smp_cpus = ms->smp.cpus; | |
575b0b33 AP |
790 | int i, affected_cpu = 0; |
791 | bool vcpu_total_blocktime = false; | |
792 | uint32_t read_vcpu_time, low_time_offset; | |
793 | ||
794 | if (!dc) { | |
795 | return; | |
796 | } | |
797 | ||
798 | low_time_offset = get_low_time_offset(dc); | |
799 | /* lookup cpu, to clear it, | |
3a4452d8 | 800 | * that algorithm looks straightforward, but it's not |
575b0b33 AP |
801 | * optimal, more optimal algorithm is keeping tree or hash |
802 | * where key is address value is a list of */ | |
803 | for (i = 0; i < smp_cpus; i++) { | |
804 | uint32_t vcpu_blocktime = 0; | |
805 | ||
d73415a3 SH |
806 | read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0); |
807 | if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr || | |
575b0b33 AP |
808 | read_vcpu_time == 0) { |
809 | continue; | |
810 | } | |
d73415a3 | 811 | qatomic_xchg(&dc->vcpu_addr[i], 0); |
575b0b33 AP |
812 | vcpu_blocktime = low_time_offset - read_vcpu_time; |
813 | affected_cpu += 1; | |
814 | /* we need to know is that mark_postcopy_end was due to | |
815 | * faulted page, another possible case it's prefetched | |
816 | * page and in that case we shouldn't be here */ | |
817 | if (!vcpu_total_blocktime && | |
d73415a3 | 818 | qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) { |
575b0b33 AP |
819 | vcpu_total_blocktime = true; |
820 | } | |
821 | /* continue cycle, due to one page could affect several vCPUs */ | |
822 | dc->vcpu_blocktime[i] += vcpu_blocktime; | |
823 | } | |
824 | ||
d73415a3 | 825 | qatomic_sub(&dc->smp_cpus_down, affected_cpu); |
575b0b33 | 826 | if (vcpu_total_blocktime) { |
d73415a3 | 827 | dc->total_blocktime += low_time_offset - qatomic_fetch_add( |
575b0b33 AP |
828 | &dc->last_begin, 0); |
829 | } | |
830 | trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime, | |
831 | affected_cpu); | |
832 | } | |
833 | ||
3a7804c3 PX |
834 | static bool postcopy_pause_fault_thread(MigrationIncomingState *mis) |
835 | { | |
836 | trace_postcopy_pause_fault_thread(); | |
837 | ||
838 | qemu_sem_wait(&mis->postcopy_pause_sem_fault); | |
839 | ||
840 | trace_postcopy_pause_fault_thread_continued(); | |
841 | ||
842 | return true; | |
843 | } | |
844 | ||
f0a227ad DDAG |
845 | /* |
846 | * Handle faults detected by the USERFAULT markings | |
847 | */ | |
848 | static void *postcopy_ram_fault_thread(void *opaque) | |
849 | { | |
850 | MigrationIncomingState *mis = opaque; | |
c4faeed2 DDAG |
851 | struct uffd_msg msg; |
852 | int ret; | |
00fa4fc8 | 853 | size_t index; |
c4faeed2 | 854 | RAMBlock *rb = NULL; |
f0a227ad | 855 | |
c4faeed2 | 856 | trace_postcopy_ram_fault_thread_entry(); |
74637e6f | 857 | rcu_register_thread(); |
096bf4c8 | 858 | mis->last_rb = NULL; /* last RAMBlock we sent part of */ |
f0a227ad | 859 | qemu_sem_post(&mis->fault_thread_sem); |
f0a227ad | 860 | |
00fa4fc8 DDAG |
861 | struct pollfd *pfd; |
862 | size_t pfd_len = 2 + mis->postcopy_remote_fds->len; | |
863 | ||
864 | pfd = g_new0(struct pollfd, pfd_len); | |
865 | ||
866 | pfd[0].fd = mis->userfault_fd; | |
867 | pfd[0].events = POLLIN; | |
868 | pfd[1].fd = mis->userfault_event_fd; | |
869 | pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */ | |
870 | trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd); | |
871 | for (index = 0; index < mis->postcopy_remote_fds->len; index++) { | |
872 | struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds, | |
873 | struct PostCopyFD, index); | |
874 | pfd[2 + index].fd = pcfd->fd; | |
875 | pfd[2 + index].events = POLLIN; | |
876 | trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr, | |
877 | pcfd->fd); | |
878 | } | |
879 | ||
c4faeed2 DDAG |
880 | while (true) { |
881 | ram_addr_t rb_offset; | |
00fa4fc8 | 882 | int poll_result; |
c4faeed2 DDAG |
883 | |
884 | /* | |
885 | * We're mainly waiting for the kernel to give us a faulting HVA, | |
886 | * however we can be told to quit via userfault_quit_fd which is | |
887 | * an eventfd | |
888 | */ | |
00fa4fc8 DDAG |
889 | |
890 | poll_result = poll(pfd, pfd_len, -1 /* Wait forever */); | |
891 | if (poll_result == -1) { | |
c4faeed2 DDAG |
892 | error_report("%s: userfault poll: %s", __func__, strerror(errno)); |
893 | break; | |
894 | } | |
895 | ||
3a7804c3 PX |
896 | if (!mis->to_src_file) { |
897 | /* | |
898 | * Possibly someone tells us that the return path is | |
899 | * broken already using the event. We should hold until | |
900 | * the channel is rebuilt. | |
901 | */ | |
902 | if (postcopy_pause_fault_thread(mis)) { | |
3a7804c3 PX |
903 | /* Continue to read the userfaultfd */ |
904 | } else { | |
905 | error_report("%s: paused but don't allow to continue", | |
906 | __func__); | |
907 | break; | |
908 | } | |
909 | } | |
910 | ||
c4faeed2 | 911 | if (pfd[1].revents) { |
64f615fe PX |
912 | uint64_t tmp64 = 0; |
913 | ||
914 | /* Consume the signal */ | |
915 | if (read(mis->userfault_event_fd, &tmp64, 8) != 8) { | |
916 | /* Nothing obviously nicer than posting this error. */ | |
917 | error_report("%s: read() failed", __func__); | |
918 | } | |
919 | ||
d73415a3 | 920 | if (qatomic_read(&mis->fault_thread_quit)) { |
64f615fe PX |
921 | trace_postcopy_ram_fault_thread_quit(); |
922 | break; | |
923 | } | |
c4faeed2 DDAG |
924 | } |
925 | ||
00fa4fc8 DDAG |
926 | if (pfd[0].revents) { |
927 | poll_result--; | |
928 | ret = read(mis->userfault_fd, &msg, sizeof(msg)); | |
929 | if (ret != sizeof(msg)) { | |
930 | if (errno == EAGAIN) { | |
931 | /* | |
932 | * if a wake up happens on the other thread just after | |
933 | * the poll, there is nothing to read. | |
934 | */ | |
935 | continue; | |
936 | } | |
937 | if (ret < 0) { | |
938 | error_report("%s: Failed to read full userfault " | |
939 | "message: %s", | |
940 | __func__, strerror(errno)); | |
941 | break; | |
942 | } else { | |
943 | error_report("%s: Read %d bytes from userfaultfd " | |
944 | "expected %zd", | |
945 | __func__, ret, sizeof(msg)); | |
946 | break; /* Lost alignment, don't know what we'd read next */ | |
947 | } | |
c4faeed2 | 948 | } |
00fa4fc8 DDAG |
949 | if (msg.event != UFFD_EVENT_PAGEFAULT) { |
950 | error_report("%s: Read unexpected event %ud from userfaultfd", | |
951 | __func__, msg.event); | |
952 | continue; /* It's not a page fault, shouldn't happen */ | |
c4faeed2 | 953 | } |
c4faeed2 | 954 | |
00fa4fc8 DDAG |
955 | rb = qemu_ram_block_from_host( |
956 | (void *)(uintptr_t)msg.arg.pagefault.address, | |
957 | true, &rb_offset); | |
958 | if (!rb) { | |
959 | error_report("postcopy_ram_fault_thread: Fault outside guest: %" | |
960 | PRIx64, (uint64_t)msg.arg.pagefault.address); | |
961 | break; | |
962 | } | |
c4faeed2 | 963 | |
00fa4fc8 DDAG |
964 | rb_offset &= ~(qemu_ram_pagesize(rb) - 1); |
965 | trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, | |
c4faeed2 | 966 | qemu_ram_get_idstr(rb), |
575b0b33 AP |
967 | rb_offset, |
968 | msg.arg.pagefault.feat.ptid); | |
969 | mark_postcopy_blocktime_begin( | |
970 | (uintptr_t)(msg.arg.pagefault.address), | |
971 | msg.arg.pagefault.feat.ptid, rb); | |
972 | ||
3a7804c3 | 973 | retry: |
00fa4fc8 DDAG |
974 | /* |
975 | * Send the request to the source - we want to request one | |
976 | * of our host page sizes (which is >= TPS) | |
977 | */ | |
8f8bfffc PX |
978 | ret = migrate_send_rp_req_pages(mis, rb, rb_offset, |
979 | msg.arg.pagefault.address); | |
3a7804c3 PX |
980 | if (ret) { |
981 | /* May be network failure, try to wait for recovery */ | |
982 | if (ret == -EIO && postcopy_pause_fault_thread(mis)) { | |
983 | /* We got reconnected somehow, try to continue */ | |
3a7804c3 PX |
984 | goto retry; |
985 | } else { | |
986 | /* This is a unavoidable fault */ | |
987 | error_report("%s: migrate_send_rp_req_pages() get %d", | |
988 | __func__, ret); | |
989 | break; | |
990 | } | |
00fa4fc8 DDAG |
991 | } |
992 | } | |
c4faeed2 | 993 | |
00fa4fc8 DDAG |
994 | /* Now handle any requests from external processes on shared memory */ |
995 | /* TODO: May need to handle devices deregistering during postcopy */ | |
996 | for (index = 2; index < pfd_len && poll_result; index++) { | |
997 | if (pfd[index].revents) { | |
998 | struct PostCopyFD *pcfd = | |
999 | &g_array_index(mis->postcopy_remote_fds, | |
1000 | struct PostCopyFD, index - 2); | |
1001 | ||
1002 | poll_result--; | |
1003 | if (pfd[index].revents & POLLERR) { | |
1004 | error_report("%s: POLLERR on poll %zd fd=%d", | |
1005 | __func__, index, pcfd->fd); | |
1006 | pfd[index].events = 0; | |
1007 | continue; | |
1008 | } | |
1009 | ||
1010 | ret = read(pcfd->fd, &msg, sizeof(msg)); | |
1011 | if (ret != sizeof(msg)) { | |
1012 | if (errno == EAGAIN) { | |
1013 | /* | |
1014 | * if a wake up happens on the other thread just after | |
1015 | * the poll, there is nothing to read. | |
1016 | */ | |
1017 | continue; | |
1018 | } | |
1019 | if (ret < 0) { | |
1020 | error_report("%s: Failed to read full userfault " | |
1021 | "message: %s (shared) revents=%d", | |
1022 | __func__, strerror(errno), | |
1023 | pfd[index].revents); | |
1024 | /*TODO: Could just disable this sharer */ | |
1025 | break; | |
1026 | } else { | |
1027 | error_report("%s: Read %d bytes from userfaultfd " | |
1028 | "expected %zd (shared)", | |
1029 | __func__, ret, sizeof(msg)); | |
1030 | /*TODO: Could just disable this sharer */ | |
1031 | break; /*Lost alignment,don't know what we'd read next*/ | |
1032 | } | |
1033 | } | |
1034 | if (msg.event != UFFD_EVENT_PAGEFAULT) { | |
1035 | error_report("%s: Read unexpected event %ud " | |
1036 | "from userfaultfd (shared)", | |
1037 | __func__, msg.event); | |
1038 | continue; /* It's not a page fault, shouldn't happen */ | |
1039 | } | |
1040 | /* Call the device handler registered with us */ | |
1041 | ret = pcfd->handler(pcfd, &msg); | |
1042 | if (ret) { | |
1043 | error_report("%s: Failed to resolve shared fault on %zd/%s", | |
1044 | __func__, index, pcfd->idstr); | |
1045 | /* TODO: Fail? Disable this sharer? */ | |
1046 | } | |
1047 | } | |
c4faeed2 DDAG |
1048 | } |
1049 | } | |
74637e6f | 1050 | rcu_unregister_thread(); |
c4faeed2 | 1051 | trace_postcopy_ram_fault_thread_exit(); |
fc6008f3 | 1052 | g_free(pfd); |
f0a227ad DDAG |
1053 | return NULL; |
1054 | } | |
1055 | ||
2a7eb148 | 1056 | int postcopy_ram_incoming_setup(MigrationIncomingState *mis) |
f0a227ad | 1057 | { |
c4faeed2 DDAG |
1058 | /* Open the fd for the kernel to give us userfaults */ |
1059 | mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK); | |
1060 | if (mis->userfault_fd == -1) { | |
1061 | error_report("%s: Failed to open userfault fd: %s", __func__, | |
1062 | strerror(errno)); | |
1063 | return -1; | |
1064 | } | |
1065 | ||
1066 | /* | |
1067 | * Although the host check already tested the API, we need to | |
1068 | * do the check again as an ABI handshake on the new fd. | |
1069 | */ | |
54ae0886 | 1070 | if (!ufd_check_and_apply(mis->userfault_fd, mis)) { |
c4faeed2 DDAG |
1071 | return -1; |
1072 | } | |
1073 | ||
1074 | /* Now an eventfd we use to tell the fault-thread to quit */ | |
64f615fe PX |
1075 | mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC); |
1076 | if (mis->userfault_event_fd == -1) { | |
1077 | error_report("%s: Opening userfault_event_fd: %s", __func__, | |
c4faeed2 DDAG |
1078 | strerror(errno)); |
1079 | close(mis->userfault_fd); | |
1080 | return -1; | |
1081 | } | |
1082 | ||
f0a227ad DDAG |
1083 | qemu_sem_init(&mis->fault_thread_sem, 0); |
1084 | qemu_thread_create(&mis->fault_thread, "postcopy/fault", | |
1085 | postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE); | |
1086 | qemu_sem_wait(&mis->fault_thread_sem); | |
1087 | qemu_sem_destroy(&mis->fault_thread_sem); | |
c4faeed2 | 1088 | mis->have_fault_thread = true; |
f0a227ad DDAG |
1089 | |
1090 | /* Mark so that we get notified of accesses to unwritten areas */ | |
fbd162e6 | 1091 | if (foreach_not_ignored_block(ram_block_enable_notify, mis)) { |
91b02dc7 | 1092 | error_report("ram_block_enable_notify failed"); |
f0a227ad DDAG |
1093 | return -1; |
1094 | } | |
1095 | ||
3414322a WY |
1096 | mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size, |
1097 | PROT_READ | PROT_WRITE, MAP_PRIVATE | | |
1098 | MAP_ANONYMOUS, -1, 0); | |
1099 | if (mis->postcopy_tmp_page == MAP_FAILED) { | |
1100 | mis->postcopy_tmp_page = NULL; | |
1101 | error_report("%s: Failed to map postcopy_tmp_page %s", | |
1102 | __func__, strerror(errno)); | |
1103 | return -1; | |
1104 | } | |
1105 | ||
6629890d WY |
1106 | /* |
1107 | * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages | |
1108 | */ | |
1109 | mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size, | |
1110 | PROT_READ | PROT_WRITE, | |
1111 | MAP_PRIVATE | MAP_ANONYMOUS, | |
1112 | -1, 0); | |
1113 | if (mis->postcopy_tmp_zero_page == MAP_FAILED) { | |
1114 | int e = errno; | |
1115 | mis->postcopy_tmp_zero_page = NULL; | |
1116 | error_report("%s: Failed to map large zero page %s", | |
1117 | __func__, strerror(e)); | |
1118 | return -e; | |
1119 | } | |
1120 | memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size); | |
1121 | ||
c4faeed2 DDAG |
1122 | trace_postcopy_ram_enable_notify(); |
1123 | ||
f0a227ad DDAG |
1124 | return 0; |
1125 | } | |
1126 | ||
eef621c4 | 1127 | static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr, |
f9494614 | 1128 | void *from_addr, uint64_t pagesize, RAMBlock *rb) |
727b9d7e | 1129 | { |
eef621c4 | 1130 | int userfault_fd = mis->userfault_fd; |
f9494614 | 1131 | int ret; |
eef621c4 | 1132 | |
727b9d7e AP |
1133 | if (from_addr) { |
1134 | struct uffdio_copy copy_struct; | |
1135 | copy_struct.dst = (uint64_t)(uintptr_t)host_addr; | |
1136 | copy_struct.src = (uint64_t)(uintptr_t)from_addr; | |
1137 | copy_struct.len = pagesize; | |
1138 | copy_struct.mode = 0; | |
f9494614 | 1139 | ret = ioctl(userfault_fd, UFFDIO_COPY, ©_struct); |
727b9d7e AP |
1140 | } else { |
1141 | struct uffdio_zeropage zero_struct; | |
1142 | zero_struct.range.start = (uint64_t)(uintptr_t)host_addr; | |
1143 | zero_struct.range.len = pagesize; | |
1144 | zero_struct.mode = 0; | |
f9494614 AP |
1145 | ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct); |
1146 | } | |
1147 | if (!ret) { | |
8f8bfffc | 1148 | qemu_mutex_lock(&mis->page_request_mutex); |
f9494614 AP |
1149 | ramblock_recv_bitmap_set_range(rb, host_addr, |
1150 | pagesize / qemu_target_page_size()); | |
8f8bfffc PX |
1151 | /* |
1152 | * If this page resolves a page fault for a previous recorded faulted | |
1153 | * address, take a special note to maintain the requested page list. | |
1154 | */ | |
1155 | if (g_tree_lookup(mis->page_requested, host_addr)) { | |
1156 | g_tree_remove(mis->page_requested, host_addr); | |
1157 | mis->page_requested_count--; | |
1158 | trace_postcopy_page_req_del(host_addr, mis->page_requested_count); | |
1159 | } | |
1160 | qemu_mutex_unlock(&mis->page_request_mutex); | |
575b0b33 | 1161 | mark_postcopy_blocktime_end((uintptr_t)host_addr); |
727b9d7e | 1162 | } |
f9494614 | 1163 | return ret; |
727b9d7e AP |
1164 | } |
1165 | ||
d488b349 DDAG |
1166 | int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset) |
1167 | { | |
1168 | int i; | |
1169 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1170 | GArray *pcrfds = mis->postcopy_remote_fds; | |
1171 | ||
1172 | for (i = 0; i < pcrfds->len; i++) { | |
1173 | struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); | |
1174 | int ret = cur->waker(cur, rb, offset); | |
1175 | if (ret) { | |
1176 | return ret; | |
1177 | } | |
1178 | } | |
1179 | return 0; | |
1180 | } | |
1181 | ||
696ed9a9 DDAG |
1182 | /* |
1183 | * Place a host page (from) at (host) atomically | |
1184 | * returns 0 on success | |
1185 | */ | |
df9ff5e1 | 1186 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
8be4620b | 1187 | RAMBlock *rb) |
696ed9a9 | 1188 | { |
8be4620b | 1189 | size_t pagesize = qemu_ram_pagesize(rb); |
696ed9a9 | 1190 | |
696ed9a9 DDAG |
1191 | /* copy also acks to the kernel waking the stalled thread up |
1192 | * TODO: We can inhibit that ack and only do it if it was requested | |
1193 | * which would be slightly cheaper, but we'd have to be careful | |
1194 | * of the order of updating our page state. | |
1195 | */ | |
eef621c4 | 1196 | if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) { |
696ed9a9 | 1197 | int e = errno; |
df9ff5e1 DDAG |
1198 | error_report("%s: %s copy host: %p from: %p (size: %zd)", |
1199 | __func__, strerror(e), host, from, pagesize); | |
696ed9a9 DDAG |
1200 | |
1201 | return -e; | |
1202 | } | |
1203 | ||
1204 | trace_postcopy_place_page(host); | |
dedfb4b2 DDAG |
1205 | return postcopy_notify_shared_wake(rb, |
1206 | qemu_ram_block_host_offset(rb, host)); | |
696ed9a9 DDAG |
1207 | } |
1208 | ||
1209 | /* | |
1210 | * Place a zero page at (host) atomically | |
1211 | * returns 0 on success | |
1212 | */ | |
df9ff5e1 | 1213 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
8be4620b | 1214 | RAMBlock *rb) |
696ed9a9 | 1215 | { |
2ce16640 | 1216 | size_t pagesize = qemu_ram_pagesize(rb); |
df9ff5e1 | 1217 | trace_postcopy_place_page_zero(host); |
696ed9a9 | 1218 | |
2ce16640 DDAG |
1219 | /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE |
1220 | * but it's not available for everything (e.g. hugetlbpages) | |
1221 | */ | |
1222 | if (qemu_ram_is_uf_zeroable(rb)) { | |
eef621c4 | 1223 | if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) { |
df9ff5e1 DDAG |
1224 | int e = errno; |
1225 | error_report("%s: %s zero host: %p", | |
1226 | __func__, strerror(e), host); | |
696ed9a9 | 1227 | |
df9ff5e1 DDAG |
1228 | return -e; |
1229 | } | |
dedfb4b2 DDAG |
1230 | return postcopy_notify_shared_wake(rb, |
1231 | qemu_ram_block_host_offset(rb, | |
1232 | host)); | |
df9ff5e1 | 1233 | } else { |
6629890d | 1234 | return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb); |
696ed9a9 | 1235 | } |
696ed9a9 DDAG |
1236 | } |
1237 | ||
eb59db53 DDAG |
1238 | #else |
1239 | /* No target OS support, stubs just fail */ | |
65ace060 AP |
1240 | void fill_destination_postcopy_migration_info(MigrationInfo *info) |
1241 | { | |
1242 | } | |
1243 | ||
d7651f15 | 1244 | bool postcopy_ram_supported_by_host(MigrationIncomingState *mis) |
eb59db53 DDAG |
1245 | { |
1246 | error_report("%s: No OS support", __func__); | |
1247 | return false; | |
1248 | } | |
1249 | ||
c136180c | 1250 | int postcopy_ram_incoming_init(MigrationIncomingState *mis) |
1caddf8a DDAG |
1251 | { |
1252 | error_report("postcopy_ram_incoming_init: No OS support"); | |
1253 | return -1; | |
1254 | } | |
1255 | ||
1256 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
1257 | { | |
1258 | assert(0); | |
1259 | return -1; | |
1260 | } | |
1261 | ||
f9527107 DDAG |
1262 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) |
1263 | { | |
1264 | assert(0); | |
1265 | return -1; | |
1266 | } | |
1267 | ||
c188c539 MT |
1268 | int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, |
1269 | uint64_t client_addr, uint64_t rb_offset) | |
1270 | { | |
1271 | assert(0); | |
1272 | return -1; | |
1273 | } | |
1274 | ||
2a7eb148 | 1275 | int postcopy_ram_incoming_setup(MigrationIncomingState *mis) |
f0a227ad DDAG |
1276 | { |
1277 | assert(0); | |
1278 | return -1; | |
1279 | } | |
696ed9a9 | 1280 | |
df9ff5e1 | 1281 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
8be4620b | 1282 | RAMBlock *rb) |
696ed9a9 DDAG |
1283 | { |
1284 | assert(0); | |
1285 | return -1; | |
1286 | } | |
1287 | ||
df9ff5e1 | 1288 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
8be4620b | 1289 | RAMBlock *rb) |
696ed9a9 DDAG |
1290 | { |
1291 | assert(0); | |
1292 | return -1; | |
1293 | } | |
1294 | ||
5efc3564 DDAG |
1295 | int postcopy_wake_shared(struct PostCopyFD *pcfd, |
1296 | uint64_t client_addr, | |
1297 | RAMBlock *rb) | |
1298 | { | |
1299 | assert(0); | |
1300 | return -1; | |
1301 | } | |
eb59db53 DDAG |
1302 | #endif |
1303 | ||
e0b266f0 DDAG |
1304 | /* ------------------------------------------------------------------------- */ |
1305 | ||
9ab7ef9b PX |
1306 | void postcopy_fault_thread_notify(MigrationIncomingState *mis) |
1307 | { | |
1308 | uint64_t tmp64 = 1; | |
1309 | ||
1310 | /* | |
1311 | * Wakeup the fault_thread. It's an eventfd that should currently | |
1312 | * be at 0, we're going to increment it to 1 | |
1313 | */ | |
1314 | if (write(mis->userfault_event_fd, &tmp64, 8) != 8) { | |
1315 | /* Not much we can do here, but may as well report it */ | |
1316 | error_report("%s: incrementing failed: %s", __func__, | |
1317 | strerror(errno)); | |
1318 | } | |
1319 | } | |
1320 | ||
e0b266f0 DDAG |
1321 | /** |
1322 | * postcopy_discard_send_init: Called at the start of each RAMBlock before | |
1323 | * asking to discard individual ranges. | |
1324 | * | |
1325 | * @ms: The current migration state. | |
810cf2bb | 1326 | * @offset: the bitmap offset of the named RAMBlock in the migration bitmap. |
e0b266f0 | 1327 | * @name: RAMBlock that discards will operate on. |
e0b266f0 | 1328 | */ |
810cf2bb WY |
1329 | static PostcopyDiscardState pds = {0}; |
1330 | void postcopy_discard_send_init(MigrationState *ms, const char *name) | |
e0b266f0 | 1331 | { |
810cf2bb WY |
1332 | pds.ramblock_name = name; |
1333 | pds.cur_entry = 0; | |
1334 | pds.nsentwords = 0; | |
1335 | pds.nsentcmds = 0; | |
e0b266f0 DDAG |
1336 | } |
1337 | ||
1338 | /** | |
1339 | * postcopy_discard_send_range: Called by the bitmap code for each chunk to | |
1340 | * discard. May send a discard message, may just leave it queued to | |
1341 | * be sent later. | |
1342 | * | |
1343 | * @ms: Current migration state. | |
e0b266f0 DDAG |
1344 | * @start,@length: a range of pages in the migration bitmap in the |
1345 | * RAM block passed to postcopy_discard_send_init() (length=1 is one page) | |
1346 | */ | |
810cf2bb WY |
1347 | void postcopy_discard_send_range(MigrationState *ms, unsigned long start, |
1348 | unsigned long length) | |
e0b266f0 | 1349 | { |
20afaed9 | 1350 | size_t tp_size = qemu_target_page_size(); |
e0b266f0 | 1351 | /* Convert to byte offsets within the RAM block */ |
810cf2bb WY |
1352 | pds.start_list[pds.cur_entry] = start * tp_size; |
1353 | pds.length_list[pds.cur_entry] = length * tp_size; | |
1354 | trace_postcopy_discard_send_range(pds.ramblock_name, start, length); | |
1355 | pds.cur_entry++; | |
1356 | pds.nsentwords++; | |
e0b266f0 | 1357 | |
810cf2bb | 1358 | if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) { |
e0b266f0 | 1359 | /* Full set, ship it! */ |
89a02a9f | 1360 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
810cf2bb WY |
1361 | pds.ramblock_name, |
1362 | pds.cur_entry, | |
1363 | pds.start_list, | |
1364 | pds.length_list); | |
1365 | pds.nsentcmds++; | |
1366 | pds.cur_entry = 0; | |
e0b266f0 DDAG |
1367 | } |
1368 | } | |
1369 | ||
1370 | /** | |
1371 | * postcopy_discard_send_finish: Called at the end of each RAMBlock by the | |
1372 | * bitmap code. Sends any outstanding discard messages, frees the PDS | |
1373 | * | |
1374 | * @ms: Current migration state. | |
e0b266f0 | 1375 | */ |
810cf2bb | 1376 | void postcopy_discard_send_finish(MigrationState *ms) |
e0b266f0 DDAG |
1377 | { |
1378 | /* Anything unsent? */ | |
810cf2bb | 1379 | if (pds.cur_entry) { |
89a02a9f | 1380 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
810cf2bb WY |
1381 | pds.ramblock_name, |
1382 | pds.cur_entry, | |
1383 | pds.start_list, | |
1384 | pds.length_list); | |
1385 | pds.nsentcmds++; | |
e0b266f0 DDAG |
1386 | } |
1387 | ||
810cf2bb WY |
1388 | trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords, |
1389 | pds.nsentcmds); | |
e0b266f0 | 1390 | } |
bac3b212 JQ |
1391 | |
1392 | /* | |
1393 | * Current state of incoming postcopy; note this is not part of | |
1394 | * MigrationIncomingState since it's state is used during cleanup | |
1395 | * at the end as MIS is being freed. | |
1396 | */ | |
1397 | static PostcopyState incoming_postcopy_state; | |
1398 | ||
1399 | PostcopyState postcopy_state_get(void) | |
1400 | { | |
d73415a3 | 1401 | return qatomic_mb_read(&incoming_postcopy_state); |
bac3b212 JQ |
1402 | } |
1403 | ||
1404 | /* Set the state and return the old state */ | |
1405 | PostcopyState postcopy_state_set(PostcopyState new_state) | |
1406 | { | |
d73415a3 | 1407 | return qatomic_xchg(&incoming_postcopy_state, new_state); |
bac3b212 | 1408 | } |
00fa4fc8 DDAG |
1409 | |
1410 | /* Register a handler for external shared memory postcopy | |
1411 | * called on the destination. | |
1412 | */ | |
1413 | void postcopy_register_shared_ufd(struct PostCopyFD *pcfd) | |
1414 | { | |
1415 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1416 | ||
1417 | mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds, | |
1418 | *pcfd); | |
1419 | } | |
1420 | ||
1421 | /* Unregister a handler for external shared memory postcopy | |
1422 | */ | |
1423 | void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd) | |
1424 | { | |
1425 | guint i; | |
1426 | MigrationIncomingState *mis = migration_incoming_get_current(); | |
1427 | GArray *pcrfds = mis->postcopy_remote_fds; | |
1428 | ||
1429 | for (i = 0; i < pcrfds->len; i++) { | |
1430 | struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); | |
1431 | if (cur->fd == pcfd->fd) { | |
1432 | mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i); | |
1433 | return; | |
1434 | } | |
1435 | } | |
1436 | } |