pc: fix auto_enable_numa_with_memhp/auto_enable_numa_with_memdev for the 5.0 machine
[qemu.git] / migration / postcopy-ram.c
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
19 #include "qemu/osdep.h"
20 #include "exec/target_page.h"
21 #include "migration.h"
22 #include "qemu-file.h"
23 #include "savevm.h"
24 #include "postcopy-ram.h"
25 #include "ram.h"
26 #include "qapi/error.h"
27 #include "qemu/notify.h"
28 #include "qemu/rcu.h"
29 #include "sysemu/sysemu.h"
30 #include "qemu/error-report.h"
31 #include "trace.h"
32 #include "hw/boards.h"
33
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;
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
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
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
84 #include <poll.h>
85 #include <sys/ioctl.h>
86 #include <sys/syscall.h>
87 #include <asm/types.h> /* for __u64 */
88 #endif
89
90 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
91 #include <sys/eventfd.h>
92 #include <linux/userfaultfd.h>
93
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 {
132 MachineState *ms = MACHINE(qdev_get_machine());
133 unsigned int smp_cpus = ms->smp.cpus;
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 }
144
145 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx)
146 {
147 MachineState *ms = MACHINE(qdev_get_machine());
148 uint32List *list = NULL, *entry = NULL;
149 int i;
150
151 for (i = ms->smp.cpus - 1; i >= 0; i--) {
152 entry = g_new0(uint32List, 1);
153 entry->value = ctx->vcpu_blocktime[i];
154 entry->next = list;
155 list = entry;
156 }
157
158 return list;
159 }
160
161 /*
162 * This function just populates MigrationInfo from postcopy's
163 * blocktime context. It will not populate MigrationInfo,
164 * unless postcopy-blocktime capability was set.
165 *
166 * @info: pointer to MigrationInfo to populate
167 */
168 void fill_destination_postcopy_migration_info(MigrationInfo *info)
169 {
170 MigrationIncomingState *mis = migration_incoming_get_current();
171 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
172
173 if (!bc) {
174 return;
175 }
176
177 info->has_postcopy_blocktime = true;
178 info->postcopy_blocktime = bc->total_blocktime;
179 info->has_postcopy_vcpu_blocktime = true;
180 info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc);
181 }
182
183 static uint32_t get_postcopy_total_blocktime(void)
184 {
185 MigrationIncomingState *mis = migration_incoming_get_current();
186 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
187
188 if (!bc) {
189 return 0;
190 }
191
192 return bc->total_blocktime;
193 }
194
195 /**
196 * receive_ufd_features: check userfault fd features, to request only supported
197 * features in the future.
198 *
199 * Returns: true on success
200 *
201 * __NR_userfaultfd - should be checked before
202 * @features: out parameter will contain uffdio_api.features provided by kernel
203 * in case of success
204 */
205 static bool receive_ufd_features(uint64_t *features)
206 {
207 struct uffdio_api api_struct = {0};
208 int ufd;
209 bool ret = true;
210
211 /* if we are here __NR_userfaultfd should exists */
212 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
213 if (ufd == -1) {
214 error_report("%s: syscall __NR_userfaultfd failed: %s", __func__,
215 strerror(errno));
216 return false;
217 }
218
219 /* ask features */
220 api_struct.api = UFFD_API;
221 api_struct.features = 0;
222 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
223 error_report("%s: UFFDIO_API failed: %s", __func__,
224 strerror(errno));
225 ret = false;
226 goto release_ufd;
227 }
228
229 *features = api_struct.features;
230
231 release_ufd:
232 close(ufd);
233 return ret;
234 }
235
236 /**
237 * request_ufd_features: this function should be called only once on a newly
238 * opened ufd, subsequent calls will lead to error.
239 *
240 * Returns: true on success
241 *
242 * @ufd: fd obtained from userfaultfd syscall
243 * @features: bit mask see UFFD_API_FEATURES
244 */
245 static bool request_ufd_features(int ufd, uint64_t features)
246 {
247 struct uffdio_api api_struct = {0};
248 uint64_t ioctl_mask;
249
250 api_struct.api = UFFD_API;
251 api_struct.features = features;
252 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
253 error_report("%s failed: UFFDIO_API failed: %s", __func__,
254 strerror(errno));
255 return false;
256 }
257
258 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
259 (__u64)1 << _UFFDIO_UNREGISTER;
260 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
261 error_report("Missing userfault features: %" PRIx64,
262 (uint64_t)(~api_struct.ioctls & ioctl_mask));
263 return false;
264 }
265
266 return true;
267 }
268
269 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis)
270 {
271 uint64_t asked_features = 0;
272 static uint64_t supported_features;
273
274 /*
275 * it's not possible to
276 * request UFFD_API twice per one fd
277 * userfault fd features is persistent
278 */
279 if (!supported_features) {
280 if (!receive_ufd_features(&supported_features)) {
281 error_report("%s failed", __func__);
282 return false;
283 }
284 }
285
286 #ifdef UFFD_FEATURE_THREAD_ID
287 if (migrate_postcopy_blocktime() && mis &&
288 UFFD_FEATURE_THREAD_ID & supported_features) {
289 /* kernel supports that feature */
290 /* don't create blocktime_context if it exists */
291 if (!mis->blocktime_ctx) {
292 mis->blocktime_ctx = blocktime_context_new();
293 }
294
295 asked_features |= UFFD_FEATURE_THREAD_ID;
296 }
297 #endif
298
299 /*
300 * request features, even if asked_features is 0, due to
301 * kernel expects UFFD_API before UFFDIO_REGISTER, per
302 * userfault file descriptor
303 */
304 if (!request_ufd_features(ufd, asked_features)) {
305 error_report("%s failed: features %" PRIu64, __func__,
306 asked_features);
307 return false;
308 }
309
310 if (qemu_real_host_page_size != ram_pagesize_summary()) {
311 bool have_hp = false;
312 /* We've got a huge page */
313 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
314 have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS;
315 #endif
316 if (!have_hp) {
317 error_report("Userfault on this host does not support huge pages");
318 return false;
319 }
320 }
321 return true;
322 }
323
324 /* Callback from postcopy_ram_supported_by_host block iterator.
325 */
326 static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque)
327 {
328 const char *block_name = qemu_ram_get_idstr(rb);
329 ram_addr_t length = qemu_ram_get_used_length(rb);
330 size_t pagesize = qemu_ram_pagesize(rb);
331
332 if (length % pagesize) {
333 error_report("Postcopy requires RAM blocks to be a page size multiple,"
334 " block %s is 0x" RAM_ADDR_FMT " bytes with a "
335 "page size of 0x%zx", block_name, length, pagesize);
336 return 1;
337 }
338 return 0;
339 }
340
341 /*
342 * Note: This has the side effect of munlock'ing all of RAM, that's
343 * normally fine since if the postcopy succeeds it gets turned back on at the
344 * end.
345 */
346 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
347 {
348 long pagesize = qemu_real_host_page_size;
349 int ufd = -1;
350 bool ret = false; /* Error unless we change it */
351 void *testarea = NULL;
352 struct uffdio_register reg_struct;
353 struct uffdio_range range_struct;
354 uint64_t feature_mask;
355 Error *local_err = NULL;
356
357 if (qemu_target_page_size() > pagesize) {
358 error_report("Target page size bigger than host page size");
359 goto out;
360 }
361
362 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
363 if (ufd == -1) {
364 error_report("%s: userfaultfd not available: %s", __func__,
365 strerror(errno));
366 goto out;
367 }
368
369 /* Give devices a chance to object */
370 if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) {
371 error_report_err(local_err);
372 goto out;
373 }
374
375 /* Version and features check */
376 if (!ufd_check_and_apply(ufd, mis)) {
377 goto out;
378 }
379
380 /* We don't support postcopy with shared RAM yet */
381 if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) {
382 goto out;
383 }
384
385 /*
386 * userfault and mlock don't go together; we'll put it back later if
387 * it was enabled.
388 */
389 if (munlockall()) {
390 error_report("%s: munlockall: %s", __func__, strerror(errno));
391 goto out;
392 }
393
394 /*
395 * We need to check that the ops we need are supported on anon memory
396 * To do that we need to register a chunk and see the flags that
397 * are returned.
398 */
399 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
400 MAP_ANONYMOUS, -1, 0);
401 if (testarea == MAP_FAILED) {
402 error_report("%s: Failed to map test area: %s", __func__,
403 strerror(errno));
404 goto out;
405 }
406 g_assert(((size_t)testarea & (pagesize-1)) == 0);
407
408 reg_struct.range.start = (uintptr_t)testarea;
409 reg_struct.range.len = pagesize;
410 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
411
412 if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
413 error_report("%s userfault register: %s", __func__, strerror(errno));
414 goto out;
415 }
416
417 range_struct.start = (uintptr_t)testarea;
418 range_struct.len = pagesize;
419 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
420 error_report("%s userfault unregister: %s", __func__, strerror(errno));
421 goto out;
422 }
423
424 feature_mask = (__u64)1 << _UFFDIO_WAKE |
425 (__u64)1 << _UFFDIO_COPY |
426 (__u64)1 << _UFFDIO_ZEROPAGE;
427 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
428 error_report("Missing userfault map features: %" PRIx64,
429 (uint64_t)(~reg_struct.ioctls & feature_mask));
430 goto out;
431 }
432
433 /* Success! */
434 ret = true;
435 out:
436 if (testarea) {
437 munmap(testarea, pagesize);
438 }
439 if (ufd != -1) {
440 close(ufd);
441 }
442 return ret;
443 }
444
445 /*
446 * Setup an area of RAM so that it *can* be used for postcopy later; this
447 * must be done right at the start prior to pre-copy.
448 * opaque should be the MIS.
449 */
450 static int init_range(RAMBlock *rb, void *opaque)
451 {
452 const char *block_name = qemu_ram_get_idstr(rb);
453 void *host_addr = qemu_ram_get_host_addr(rb);
454 ram_addr_t offset = qemu_ram_get_offset(rb);
455 ram_addr_t length = qemu_ram_get_used_length(rb);
456 trace_postcopy_init_range(block_name, host_addr, offset, length);
457
458 /*
459 * We need the whole of RAM to be truly empty for postcopy, so things
460 * like ROMs and any data tables built during init must be zero'd
461 * - we're going to get the copy from the source anyway.
462 * (Precopy will just overwrite this data, so doesn't need the discard)
463 */
464 if (ram_discard_range(block_name, 0, length)) {
465 return -1;
466 }
467
468 return 0;
469 }
470
471 /*
472 * At the end of migration, undo the effects of init_range
473 * opaque should be the MIS.
474 */
475 static int cleanup_range(RAMBlock *rb, void *opaque)
476 {
477 const char *block_name = qemu_ram_get_idstr(rb);
478 void *host_addr = qemu_ram_get_host_addr(rb);
479 ram_addr_t offset = qemu_ram_get_offset(rb);
480 ram_addr_t length = qemu_ram_get_used_length(rb);
481 MigrationIncomingState *mis = opaque;
482 struct uffdio_range range_struct;
483 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
484
485 /*
486 * We turned off hugepage for the precopy stage with postcopy enabled
487 * we can turn it back on now.
488 */
489 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
490
491 /*
492 * We can also turn off userfault now since we should have all the
493 * pages. It can be useful to leave it on to debug postcopy
494 * if you're not sure it's always getting every page.
495 */
496 range_struct.start = (uintptr_t)host_addr;
497 range_struct.len = length;
498
499 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
500 error_report("%s: userfault unregister %s", __func__, strerror(errno));
501
502 return -1;
503 }
504
505 return 0;
506 }
507
508 /*
509 * Initialise postcopy-ram, setting the RAM to a state where we can go into
510 * postcopy later; must be called prior to any precopy.
511 * called from arch_init's similarly named ram_postcopy_incoming_init
512 */
513 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
514 {
515 if (foreach_not_ignored_block(init_range, NULL)) {
516 return -1;
517 }
518
519 return 0;
520 }
521
522 /*
523 * At the end of a migration where postcopy_ram_incoming_init was called.
524 */
525 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
526 {
527 trace_postcopy_ram_incoming_cleanup_entry();
528
529 if (mis->have_fault_thread) {
530 Error *local_err = NULL;
531
532 /* Let the fault thread quit */
533 qatomic_set(&mis->fault_thread_quit, 1);
534 postcopy_fault_thread_notify(mis);
535 trace_postcopy_ram_incoming_cleanup_join();
536 qemu_thread_join(&mis->fault_thread);
537
538 if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
539 error_report_err(local_err);
540 return -1;
541 }
542
543 if (foreach_not_ignored_block(cleanup_range, mis)) {
544 return -1;
545 }
546
547 trace_postcopy_ram_incoming_cleanup_closeuf();
548 close(mis->userfault_fd);
549 close(mis->userfault_event_fd);
550 mis->have_fault_thread = false;
551 }
552
553 if (enable_mlock) {
554 if (os_mlock() < 0) {
555 error_report("mlock: %s", strerror(errno));
556 /*
557 * It doesn't feel right to fail at this point, we have a valid
558 * VM state.
559 */
560 }
561 }
562
563 if (mis->postcopy_tmp_page) {
564 munmap(mis->postcopy_tmp_page, mis->largest_page_size);
565 mis->postcopy_tmp_page = NULL;
566 }
567 if (mis->postcopy_tmp_zero_page) {
568 munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
569 mis->postcopy_tmp_zero_page = NULL;
570 }
571 trace_postcopy_ram_incoming_cleanup_blocktime(
572 get_postcopy_total_blocktime());
573
574 trace_postcopy_ram_incoming_cleanup_exit();
575 return 0;
576 }
577
578 /*
579 * Disable huge pages on an area
580 */
581 static int nhp_range(RAMBlock *rb, void *opaque)
582 {
583 const char *block_name = qemu_ram_get_idstr(rb);
584 void *host_addr = qemu_ram_get_host_addr(rb);
585 ram_addr_t offset = qemu_ram_get_offset(rb);
586 ram_addr_t length = qemu_ram_get_used_length(rb);
587 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
588
589 /*
590 * Before we do discards we need to ensure those discards really
591 * do delete areas of the page, even if THP thinks a hugepage would
592 * be a good idea, so force hugepages off.
593 */
594 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
595
596 return 0;
597 }
598
599 /*
600 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
601 * however leaving it until after precopy means that most of the precopy
602 * data is still THPd
603 */
604 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
605 {
606 if (foreach_not_ignored_block(nhp_range, mis)) {
607 return -1;
608 }
609
610 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
611
612 return 0;
613 }
614
615 /*
616 * Mark the given area of RAM as requiring notification to unwritten areas
617 * Used as a callback on foreach_not_ignored_block.
618 * host_addr: Base of area to mark
619 * offset: Offset in the whole ram arena
620 * length: Length of the section
621 * opaque: MigrationIncomingState pointer
622 * Returns 0 on success
623 */
624 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
625 {
626 MigrationIncomingState *mis = opaque;
627 struct uffdio_register reg_struct;
628
629 reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
630 reg_struct.range.len = qemu_ram_get_used_length(rb);
631 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
632
633 /* Now tell our userfault_fd that it's responsible for this area */
634 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
635 error_report("%s userfault register: %s", __func__, strerror(errno));
636 return -1;
637 }
638 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
639 error_report("%s userfault: Region doesn't support COPY", __func__);
640 return -1;
641 }
642 if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) {
643 qemu_ram_set_uf_zeroable(rb);
644 }
645
646 return 0;
647 }
648
649 int postcopy_wake_shared(struct PostCopyFD *pcfd,
650 uint64_t client_addr,
651 RAMBlock *rb)
652 {
653 size_t pagesize = qemu_ram_pagesize(rb);
654 struct uffdio_range range;
655 int ret;
656 trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
657 range.start = client_addr & ~(pagesize - 1);
658 range.len = pagesize;
659 ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
660 if (ret) {
661 error_report("%s: Failed to wake: %zx in %s (%s)",
662 __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
663 strerror(errno));
664 }
665 return ret;
666 }
667
668 /*
669 * Callback from shared fault handlers to ask for a page,
670 * the page must be specified by a RAMBlock and an offset in that rb
671 * Note: Only for use by shared fault handlers (in fault thread)
672 */
673 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
674 uint64_t client_addr, uint64_t rb_offset)
675 {
676 size_t pagesize = qemu_ram_pagesize(rb);
677 uint64_t aligned_rbo = rb_offset & ~(pagesize - 1);
678 MigrationIncomingState *mis = migration_incoming_get_current();
679
680 trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
681 rb_offset);
682 if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
683 trace_postcopy_request_shared_page_present(pcfd->idstr,
684 qemu_ram_get_idstr(rb), rb_offset);
685 return postcopy_wake_shared(pcfd, client_addr, rb);
686 }
687 migrate_send_rp_req_pages(mis, rb, aligned_rbo);
688 return 0;
689 }
690
691 static int get_mem_fault_cpu_index(uint32_t pid)
692 {
693 CPUState *cpu_iter;
694
695 CPU_FOREACH(cpu_iter) {
696 if (cpu_iter->thread_id == pid) {
697 trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
698 return cpu_iter->cpu_index;
699 }
700 }
701 trace_get_mem_fault_cpu_index(-1, pid);
702 return -1;
703 }
704
705 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
706 {
707 int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
708 dc->start_time;
709 return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
710 }
711
712 /*
713 * This function is being called when pagefault occurs. It
714 * tracks down vCPU blocking time.
715 *
716 * @addr: faulted host virtual address
717 * @ptid: faulted process thread id
718 * @rb: ramblock appropriate to addr
719 */
720 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
721 RAMBlock *rb)
722 {
723 int cpu, already_received;
724 MigrationIncomingState *mis = migration_incoming_get_current();
725 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
726 uint32_t low_time_offset;
727
728 if (!dc || ptid == 0) {
729 return;
730 }
731 cpu = get_mem_fault_cpu_index(ptid);
732 if (cpu < 0) {
733 return;
734 }
735
736 low_time_offset = get_low_time_offset(dc);
737 if (dc->vcpu_addr[cpu] == 0) {
738 qatomic_inc(&dc->smp_cpus_down);
739 }
740
741 qatomic_xchg(&dc->last_begin, low_time_offset);
742 qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
743 qatomic_xchg(&dc->vcpu_addr[cpu], addr);
744
745 /*
746 * check it here, not at the beginning of the function,
747 * due to, check could occur early than bitmap_set in
748 * qemu_ufd_copy_ioctl
749 */
750 already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
751 if (already_received) {
752 qatomic_xchg(&dc->vcpu_addr[cpu], 0);
753 qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
754 qatomic_dec(&dc->smp_cpus_down);
755 }
756 trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
757 cpu, already_received);
758 }
759
760 /*
761 * This function just provide calculated blocktime per cpu and trace it.
762 * Total blocktime is calculated in mark_postcopy_blocktime_end.
763 *
764 *
765 * Assume we have 3 CPU
766 *
767 * S1 E1 S1 E1
768 * -----***********------------xxx***************------------------------> CPU1
769 *
770 * S2 E2
771 * ------------****************xxx---------------------------------------> CPU2
772 *
773 * S3 E3
774 * ------------------------****xxx********-------------------------------> CPU3
775 *
776 * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
777 * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
778 * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
779 * it's a part of total blocktime.
780 * S1 - here is last_begin
781 * Legend of the picture is following:
782 * * - means blocktime per vCPU
783 * x - means overlapped blocktime (total blocktime)
784 *
785 * @addr: host virtual address
786 */
787 static void mark_postcopy_blocktime_end(uintptr_t addr)
788 {
789 MigrationIncomingState *mis = migration_incoming_get_current();
790 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
791 MachineState *ms = MACHINE(qdev_get_machine());
792 unsigned int smp_cpus = ms->smp.cpus;
793 int i, affected_cpu = 0;
794 bool vcpu_total_blocktime = false;
795 uint32_t read_vcpu_time, low_time_offset;
796
797 if (!dc) {
798 return;
799 }
800
801 low_time_offset = get_low_time_offset(dc);
802 /* lookup cpu, to clear it,
803 * that algorithm looks straightforward, but it's not
804 * optimal, more optimal algorithm is keeping tree or hash
805 * where key is address value is a list of */
806 for (i = 0; i < smp_cpus; i++) {
807 uint32_t vcpu_blocktime = 0;
808
809 read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
810 if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
811 read_vcpu_time == 0) {
812 continue;
813 }
814 qatomic_xchg(&dc->vcpu_addr[i], 0);
815 vcpu_blocktime = low_time_offset - read_vcpu_time;
816 affected_cpu += 1;
817 /* we need to know is that mark_postcopy_end was due to
818 * faulted page, another possible case it's prefetched
819 * page and in that case we shouldn't be here */
820 if (!vcpu_total_blocktime &&
821 qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
822 vcpu_total_blocktime = true;
823 }
824 /* continue cycle, due to one page could affect several vCPUs */
825 dc->vcpu_blocktime[i] += vcpu_blocktime;
826 }
827
828 qatomic_sub(&dc->smp_cpus_down, affected_cpu);
829 if (vcpu_total_blocktime) {
830 dc->total_blocktime += low_time_offset - qatomic_fetch_add(
831 &dc->last_begin, 0);
832 }
833 trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
834 affected_cpu);
835 }
836
837 static bool postcopy_pause_fault_thread(MigrationIncomingState *mis)
838 {
839 trace_postcopy_pause_fault_thread();
840
841 qemu_sem_wait(&mis->postcopy_pause_sem_fault);
842
843 trace_postcopy_pause_fault_thread_continued();
844
845 return true;
846 }
847
848 /*
849 * Handle faults detected by the USERFAULT markings
850 */
851 static void *postcopy_ram_fault_thread(void *opaque)
852 {
853 MigrationIncomingState *mis = opaque;
854 struct uffd_msg msg;
855 int ret;
856 size_t index;
857 RAMBlock *rb = NULL;
858
859 trace_postcopy_ram_fault_thread_entry();
860 rcu_register_thread();
861 mis->last_rb = NULL; /* last RAMBlock we sent part of */
862 qemu_sem_post(&mis->fault_thread_sem);
863
864 struct pollfd *pfd;
865 size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
866
867 pfd = g_new0(struct pollfd, pfd_len);
868
869 pfd[0].fd = mis->userfault_fd;
870 pfd[0].events = POLLIN;
871 pfd[1].fd = mis->userfault_event_fd;
872 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
873 trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
874 for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
875 struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
876 struct PostCopyFD, index);
877 pfd[2 + index].fd = pcfd->fd;
878 pfd[2 + index].events = POLLIN;
879 trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
880 pcfd->fd);
881 }
882
883 while (true) {
884 ram_addr_t rb_offset;
885 int poll_result;
886
887 /*
888 * We're mainly waiting for the kernel to give us a faulting HVA,
889 * however we can be told to quit via userfault_quit_fd which is
890 * an eventfd
891 */
892
893 poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
894 if (poll_result == -1) {
895 error_report("%s: userfault poll: %s", __func__, strerror(errno));
896 break;
897 }
898
899 if (!mis->to_src_file) {
900 /*
901 * Possibly someone tells us that the return path is
902 * broken already using the event. We should hold until
903 * the channel is rebuilt.
904 */
905 if (postcopy_pause_fault_thread(mis)) {
906 mis->last_rb = NULL;
907 /* Continue to read the userfaultfd */
908 } else {
909 error_report("%s: paused but don't allow to continue",
910 __func__);
911 break;
912 }
913 }
914
915 if (pfd[1].revents) {
916 uint64_t tmp64 = 0;
917
918 /* Consume the signal */
919 if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
920 /* Nothing obviously nicer than posting this error. */
921 error_report("%s: read() failed", __func__);
922 }
923
924 if (qatomic_read(&mis->fault_thread_quit)) {
925 trace_postcopy_ram_fault_thread_quit();
926 break;
927 }
928 }
929
930 if (pfd[0].revents) {
931 poll_result--;
932 ret = read(mis->userfault_fd, &msg, sizeof(msg));
933 if (ret != sizeof(msg)) {
934 if (errno == EAGAIN) {
935 /*
936 * if a wake up happens on the other thread just after
937 * the poll, there is nothing to read.
938 */
939 continue;
940 }
941 if (ret < 0) {
942 error_report("%s: Failed to read full userfault "
943 "message: %s",
944 __func__, strerror(errno));
945 break;
946 } else {
947 error_report("%s: Read %d bytes from userfaultfd "
948 "expected %zd",
949 __func__, ret, sizeof(msg));
950 break; /* Lost alignment, don't know what we'd read next */
951 }
952 }
953 if (msg.event != UFFD_EVENT_PAGEFAULT) {
954 error_report("%s: Read unexpected event %ud from userfaultfd",
955 __func__, msg.event);
956 continue; /* It's not a page fault, shouldn't happen */
957 }
958
959 rb = qemu_ram_block_from_host(
960 (void *)(uintptr_t)msg.arg.pagefault.address,
961 true, &rb_offset);
962 if (!rb) {
963 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
964 PRIx64, (uint64_t)msg.arg.pagefault.address);
965 break;
966 }
967
968 rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
969 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
970 qemu_ram_get_idstr(rb),
971 rb_offset,
972 msg.arg.pagefault.feat.ptid);
973 mark_postcopy_blocktime_begin(
974 (uintptr_t)(msg.arg.pagefault.address),
975 msg.arg.pagefault.feat.ptid, rb);
976
977 retry:
978 /*
979 * Send the request to the source - we want to request one
980 * of our host page sizes (which is >= TPS)
981 */
982 ret = migrate_send_rp_req_pages(mis, rb, rb_offset);
983 if (ret) {
984 /* May be network failure, try to wait for recovery */
985 if (ret == -EIO && postcopy_pause_fault_thread(mis)) {
986 /* We got reconnected somehow, try to continue */
987 mis->last_rb = NULL;
988 goto retry;
989 } else {
990 /* This is a unavoidable fault */
991 error_report("%s: migrate_send_rp_req_pages() get %d",
992 __func__, ret);
993 break;
994 }
995 }
996 }
997
998 /* Now handle any requests from external processes on shared memory */
999 /* TODO: May need to handle devices deregistering during postcopy */
1000 for (index = 2; index < pfd_len && poll_result; index++) {
1001 if (pfd[index].revents) {
1002 struct PostCopyFD *pcfd =
1003 &g_array_index(mis->postcopy_remote_fds,
1004 struct PostCopyFD, index - 2);
1005
1006 poll_result--;
1007 if (pfd[index].revents & POLLERR) {
1008 error_report("%s: POLLERR on poll %zd fd=%d",
1009 __func__, index, pcfd->fd);
1010 pfd[index].events = 0;
1011 continue;
1012 }
1013
1014 ret = read(pcfd->fd, &msg, sizeof(msg));
1015 if (ret != sizeof(msg)) {
1016 if (errno == EAGAIN) {
1017 /*
1018 * if a wake up happens on the other thread just after
1019 * the poll, there is nothing to read.
1020 */
1021 continue;
1022 }
1023 if (ret < 0) {
1024 error_report("%s: Failed to read full userfault "
1025 "message: %s (shared) revents=%d",
1026 __func__, strerror(errno),
1027 pfd[index].revents);
1028 /*TODO: Could just disable this sharer */
1029 break;
1030 } else {
1031 error_report("%s: Read %d bytes from userfaultfd "
1032 "expected %zd (shared)",
1033 __func__, ret, sizeof(msg));
1034 /*TODO: Could just disable this sharer */
1035 break; /*Lost alignment,don't know what we'd read next*/
1036 }
1037 }
1038 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1039 error_report("%s: Read unexpected event %ud "
1040 "from userfaultfd (shared)",
1041 __func__, msg.event);
1042 continue; /* It's not a page fault, shouldn't happen */
1043 }
1044 /* Call the device handler registered with us */
1045 ret = pcfd->handler(pcfd, &msg);
1046 if (ret) {
1047 error_report("%s: Failed to resolve shared fault on %zd/%s",
1048 __func__, index, pcfd->idstr);
1049 /* TODO: Fail? Disable this sharer? */
1050 }
1051 }
1052 }
1053 }
1054 rcu_unregister_thread();
1055 trace_postcopy_ram_fault_thread_exit();
1056 g_free(pfd);
1057 return NULL;
1058 }
1059
1060 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1061 {
1062 /* Open the fd for the kernel to give us userfaults */
1063 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1064 if (mis->userfault_fd == -1) {
1065 error_report("%s: Failed to open userfault fd: %s", __func__,
1066 strerror(errno));
1067 return -1;
1068 }
1069
1070 /*
1071 * Although the host check already tested the API, we need to
1072 * do the check again as an ABI handshake on the new fd.
1073 */
1074 if (!ufd_check_and_apply(mis->userfault_fd, mis)) {
1075 return -1;
1076 }
1077
1078 /* Now an eventfd we use to tell the fault-thread to quit */
1079 mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1080 if (mis->userfault_event_fd == -1) {
1081 error_report("%s: Opening userfault_event_fd: %s", __func__,
1082 strerror(errno));
1083 close(mis->userfault_fd);
1084 return -1;
1085 }
1086
1087 qemu_sem_init(&mis->fault_thread_sem, 0);
1088 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
1089 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
1090 qemu_sem_wait(&mis->fault_thread_sem);
1091 qemu_sem_destroy(&mis->fault_thread_sem);
1092 mis->have_fault_thread = true;
1093
1094 /* Mark so that we get notified of accesses to unwritten areas */
1095 if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1096 error_report("ram_block_enable_notify failed");
1097 return -1;
1098 }
1099
1100 mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
1101 PROT_READ | PROT_WRITE, MAP_PRIVATE |
1102 MAP_ANONYMOUS, -1, 0);
1103 if (mis->postcopy_tmp_page == MAP_FAILED) {
1104 mis->postcopy_tmp_page = NULL;
1105 error_report("%s: Failed to map postcopy_tmp_page %s",
1106 __func__, strerror(errno));
1107 return -1;
1108 }
1109
1110 /*
1111 * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages
1112 */
1113 mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1114 PROT_READ | PROT_WRITE,
1115 MAP_PRIVATE | MAP_ANONYMOUS,
1116 -1, 0);
1117 if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1118 int e = errno;
1119 mis->postcopy_tmp_zero_page = NULL;
1120 error_report("%s: Failed to map large zero page %s",
1121 __func__, strerror(e));
1122 return -e;
1123 }
1124 memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1125
1126 trace_postcopy_ram_enable_notify();
1127
1128 return 0;
1129 }
1130
1131 static int qemu_ufd_copy_ioctl(int userfault_fd, void *host_addr,
1132 void *from_addr, uint64_t pagesize, RAMBlock *rb)
1133 {
1134 int ret;
1135 if (from_addr) {
1136 struct uffdio_copy copy_struct;
1137 copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1138 copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1139 copy_struct.len = pagesize;
1140 copy_struct.mode = 0;
1141 ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1142 } else {
1143 struct uffdio_zeropage zero_struct;
1144 zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1145 zero_struct.range.len = pagesize;
1146 zero_struct.mode = 0;
1147 ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1148 }
1149 if (!ret) {
1150 ramblock_recv_bitmap_set_range(rb, host_addr,
1151 pagesize / qemu_target_page_size());
1152 mark_postcopy_blocktime_end((uintptr_t)host_addr);
1153
1154 }
1155 return ret;
1156 }
1157
1158 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1159 {
1160 int i;
1161 MigrationIncomingState *mis = migration_incoming_get_current();
1162 GArray *pcrfds = mis->postcopy_remote_fds;
1163
1164 for (i = 0; i < pcrfds->len; i++) {
1165 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1166 int ret = cur->waker(cur, rb, offset);
1167 if (ret) {
1168 return ret;
1169 }
1170 }
1171 return 0;
1172 }
1173
1174 /*
1175 * Place a host page (from) at (host) atomically
1176 * returns 0 on success
1177 */
1178 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1179 RAMBlock *rb)
1180 {
1181 size_t pagesize = qemu_ram_pagesize(rb);
1182
1183 /* copy also acks to the kernel waking the stalled thread up
1184 * TODO: We can inhibit that ack and only do it if it was requested
1185 * which would be slightly cheaper, but we'd have to be careful
1186 * of the order of updating our page state.
1187 */
1188 if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, from, pagesize, rb)) {
1189 int e = errno;
1190 error_report("%s: %s copy host: %p from: %p (size: %zd)",
1191 __func__, strerror(e), host, from, pagesize);
1192
1193 return -e;
1194 }
1195
1196 trace_postcopy_place_page(host);
1197 return postcopy_notify_shared_wake(rb,
1198 qemu_ram_block_host_offset(rb, host));
1199 }
1200
1201 /*
1202 * Place a zero page at (host) atomically
1203 * returns 0 on success
1204 */
1205 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1206 RAMBlock *rb)
1207 {
1208 size_t pagesize = qemu_ram_pagesize(rb);
1209 trace_postcopy_place_page_zero(host);
1210
1211 /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1212 * but it's not available for everything (e.g. hugetlbpages)
1213 */
1214 if (qemu_ram_is_uf_zeroable(rb)) {
1215 if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, NULL, pagesize, rb)) {
1216 int e = errno;
1217 error_report("%s: %s zero host: %p",
1218 __func__, strerror(e), host);
1219
1220 return -e;
1221 }
1222 return postcopy_notify_shared_wake(rb,
1223 qemu_ram_block_host_offset(rb,
1224 host));
1225 } else {
1226 return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb);
1227 }
1228 }
1229
1230 #else
1231 /* No target OS support, stubs just fail */
1232 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1233 {
1234 }
1235
1236 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
1237 {
1238 error_report("%s: No OS support", __func__);
1239 return false;
1240 }
1241
1242 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1243 {
1244 error_report("postcopy_ram_incoming_init: No OS support");
1245 return -1;
1246 }
1247
1248 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1249 {
1250 assert(0);
1251 return -1;
1252 }
1253
1254 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1255 {
1256 assert(0);
1257 return -1;
1258 }
1259
1260 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1261 uint64_t client_addr, uint64_t rb_offset)
1262 {
1263 assert(0);
1264 return -1;
1265 }
1266
1267 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1268 {
1269 assert(0);
1270 return -1;
1271 }
1272
1273 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1274 RAMBlock *rb)
1275 {
1276 assert(0);
1277 return -1;
1278 }
1279
1280 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1281 RAMBlock *rb)
1282 {
1283 assert(0);
1284 return -1;
1285 }
1286
1287 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1288 uint64_t client_addr,
1289 RAMBlock *rb)
1290 {
1291 assert(0);
1292 return -1;
1293 }
1294 #endif
1295
1296 /* ------------------------------------------------------------------------- */
1297
1298 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1299 {
1300 uint64_t tmp64 = 1;
1301
1302 /*
1303 * Wakeup the fault_thread. It's an eventfd that should currently
1304 * be at 0, we're going to increment it to 1
1305 */
1306 if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1307 /* Not much we can do here, but may as well report it */
1308 error_report("%s: incrementing failed: %s", __func__,
1309 strerror(errno));
1310 }
1311 }
1312
1313 /**
1314 * postcopy_discard_send_init: Called at the start of each RAMBlock before
1315 * asking to discard individual ranges.
1316 *
1317 * @ms: The current migration state.
1318 * @offset: the bitmap offset of the named RAMBlock in the migration bitmap.
1319 * @name: RAMBlock that discards will operate on.
1320 */
1321 static PostcopyDiscardState pds = {0};
1322 void postcopy_discard_send_init(MigrationState *ms, const char *name)
1323 {
1324 pds.ramblock_name = name;
1325 pds.cur_entry = 0;
1326 pds.nsentwords = 0;
1327 pds.nsentcmds = 0;
1328 }
1329
1330 /**
1331 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1332 * discard. May send a discard message, may just leave it queued to
1333 * be sent later.
1334 *
1335 * @ms: Current migration state.
1336 * @start,@length: a range of pages in the migration bitmap in the
1337 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1338 */
1339 void postcopy_discard_send_range(MigrationState *ms, unsigned long start,
1340 unsigned long length)
1341 {
1342 size_t tp_size = qemu_target_page_size();
1343 /* Convert to byte offsets within the RAM block */
1344 pds.start_list[pds.cur_entry] = start * tp_size;
1345 pds.length_list[pds.cur_entry] = length * tp_size;
1346 trace_postcopy_discard_send_range(pds.ramblock_name, start, length);
1347 pds.cur_entry++;
1348 pds.nsentwords++;
1349
1350 if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) {
1351 /* Full set, ship it! */
1352 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1353 pds.ramblock_name,
1354 pds.cur_entry,
1355 pds.start_list,
1356 pds.length_list);
1357 pds.nsentcmds++;
1358 pds.cur_entry = 0;
1359 }
1360 }
1361
1362 /**
1363 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1364 * bitmap code. Sends any outstanding discard messages, frees the PDS
1365 *
1366 * @ms: Current migration state.
1367 */
1368 void postcopy_discard_send_finish(MigrationState *ms)
1369 {
1370 /* Anything unsent? */
1371 if (pds.cur_entry) {
1372 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1373 pds.ramblock_name,
1374 pds.cur_entry,
1375 pds.start_list,
1376 pds.length_list);
1377 pds.nsentcmds++;
1378 }
1379
1380 trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords,
1381 pds.nsentcmds);
1382 }
1383
1384 /*
1385 * Current state of incoming postcopy; note this is not part of
1386 * MigrationIncomingState since it's state is used during cleanup
1387 * at the end as MIS is being freed.
1388 */
1389 static PostcopyState incoming_postcopy_state;
1390
1391 PostcopyState postcopy_state_get(void)
1392 {
1393 return qatomic_mb_read(&incoming_postcopy_state);
1394 }
1395
1396 /* Set the state and return the old state */
1397 PostcopyState postcopy_state_set(PostcopyState new_state)
1398 {
1399 return qatomic_xchg(&incoming_postcopy_state, new_state);
1400 }
1401
1402 /* Register a handler for external shared memory postcopy
1403 * called on the destination.
1404 */
1405 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1406 {
1407 MigrationIncomingState *mis = migration_incoming_get_current();
1408
1409 mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1410 *pcfd);
1411 }
1412
1413 /* Unregister a handler for external shared memory postcopy
1414 */
1415 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1416 {
1417 guint i;
1418 MigrationIncomingState *mis = migration_incoming_get_current();
1419 GArray *pcrfds = mis->postcopy_remote_fds;
1420
1421 for (i = 0; i < pcrfds->len; i++) {
1422 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1423 if (cur->fd == pcfd->fd) {
1424 mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1425 return;
1426 }
1427 }
1428 }