xref: /openbmc/qemu/migration/postcopy-ram.c (revision 64552b6b)
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 "sysemu/sysemu.h"
29 #include "sysemu/balloon.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 succes
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 (getpagesize() != 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 = getpagesize();
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         return -1;
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  * Manage a single vote to the QEMU balloon inhibitor for all postcopy usage,
524  * last caller wins.
525  */
526 static void postcopy_balloon_inhibit(bool state)
527 {
528     static bool cur_state = false;
529 
530     if (state != cur_state) {
531         qemu_balloon_inhibit(state);
532         cur_state = state;
533     }
534 }
535 
536 /*
537  * At the end of a migration where postcopy_ram_incoming_init was called.
538  */
539 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
540 {
541     trace_postcopy_ram_incoming_cleanup_entry();
542 
543     if (mis->have_fault_thread) {
544         Error *local_err = NULL;
545 
546         /* Let the fault thread quit */
547         atomic_set(&mis->fault_thread_quit, 1);
548         postcopy_fault_thread_notify(mis);
549         trace_postcopy_ram_incoming_cleanup_join();
550         qemu_thread_join(&mis->fault_thread);
551 
552         if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
553             error_report_err(local_err);
554             return -1;
555         }
556 
557         if (foreach_not_ignored_block(cleanup_range, mis)) {
558             return -1;
559         }
560 
561         trace_postcopy_ram_incoming_cleanup_closeuf();
562         close(mis->userfault_fd);
563         close(mis->userfault_event_fd);
564         mis->have_fault_thread = false;
565     }
566 
567     postcopy_balloon_inhibit(false);
568 
569     if (enable_mlock) {
570         if (os_mlock() < 0) {
571             error_report("mlock: %s", strerror(errno));
572             /*
573              * It doesn't feel right to fail at this point, we have a valid
574              * VM state.
575              */
576         }
577     }
578 
579     postcopy_state_set(POSTCOPY_INCOMING_END);
580 
581     if (mis->postcopy_tmp_page) {
582         munmap(mis->postcopy_tmp_page, mis->largest_page_size);
583         mis->postcopy_tmp_page = NULL;
584     }
585     if (mis->postcopy_tmp_zero_page) {
586         munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
587         mis->postcopy_tmp_zero_page = NULL;
588     }
589     trace_postcopy_ram_incoming_cleanup_blocktime(
590             get_postcopy_total_blocktime());
591 
592     trace_postcopy_ram_incoming_cleanup_exit();
593     return 0;
594 }
595 
596 /*
597  * Disable huge pages on an area
598  */
599 static int nhp_range(RAMBlock *rb, void *opaque)
600 {
601     const char *block_name = qemu_ram_get_idstr(rb);
602     void *host_addr = qemu_ram_get_host_addr(rb);
603     ram_addr_t offset = qemu_ram_get_offset(rb);
604     ram_addr_t length = qemu_ram_get_used_length(rb);
605     trace_postcopy_nhp_range(block_name, host_addr, offset, length);
606 
607     /*
608      * Before we do discards we need to ensure those discards really
609      * do delete areas of the page, even if THP thinks a hugepage would
610      * be a good idea, so force hugepages off.
611      */
612     qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
613 
614     return 0;
615 }
616 
617 /*
618  * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
619  * however leaving it until after precopy means that most of the precopy
620  * data is still THPd
621  */
622 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
623 {
624     if (foreach_not_ignored_block(nhp_range, mis)) {
625         return -1;
626     }
627 
628     postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
629 
630     return 0;
631 }
632 
633 /*
634  * Mark the given area of RAM as requiring notification to unwritten areas
635  * Used as a  callback on foreach_not_ignored_block.
636  *   host_addr: Base of area to mark
637  *   offset: Offset in the whole ram arena
638  *   length: Length of the section
639  *   opaque: MigrationIncomingState pointer
640  * Returns 0 on success
641  */
642 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
643 {
644     MigrationIncomingState *mis = opaque;
645     struct uffdio_register reg_struct;
646 
647     reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
648     reg_struct.range.len = qemu_ram_get_used_length(rb);
649     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
650 
651     /* Now tell our userfault_fd that it's responsible for this area */
652     if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
653         error_report("%s userfault register: %s", __func__, strerror(errno));
654         return -1;
655     }
656     if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
657         error_report("%s userfault: Region doesn't support COPY", __func__);
658         return -1;
659     }
660     if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) {
661         qemu_ram_set_uf_zeroable(rb);
662     }
663 
664     return 0;
665 }
666 
667 int postcopy_wake_shared(struct PostCopyFD *pcfd,
668                          uint64_t client_addr,
669                          RAMBlock *rb)
670 {
671     size_t pagesize = qemu_ram_pagesize(rb);
672     struct uffdio_range range;
673     int ret;
674     trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
675     range.start = client_addr & ~(pagesize - 1);
676     range.len = pagesize;
677     ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
678     if (ret) {
679         error_report("%s: Failed to wake: %zx in %s (%s)",
680                      __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
681                      strerror(errno));
682     }
683     return ret;
684 }
685 
686 /*
687  * Callback from shared fault handlers to ask for a page,
688  * the page must be specified by a RAMBlock and an offset in that rb
689  * Note: Only for use by shared fault handlers (in fault thread)
690  */
691 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
692                                  uint64_t client_addr, uint64_t rb_offset)
693 {
694     size_t pagesize = qemu_ram_pagesize(rb);
695     uint64_t aligned_rbo = rb_offset & ~(pagesize - 1);
696     MigrationIncomingState *mis = migration_incoming_get_current();
697 
698     trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
699                                        rb_offset);
700     if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
701         trace_postcopy_request_shared_page_present(pcfd->idstr,
702                                         qemu_ram_get_idstr(rb), rb_offset);
703         return postcopy_wake_shared(pcfd, client_addr, rb);
704     }
705     if (rb != mis->last_rb) {
706         mis->last_rb = rb;
707         migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
708                                   aligned_rbo, pagesize);
709     } else {
710         /* Save some space */
711         migrate_send_rp_req_pages(mis, NULL, aligned_rbo, pagesize);
712     }
713     return 0;
714 }
715 
716 static int get_mem_fault_cpu_index(uint32_t pid)
717 {
718     CPUState *cpu_iter;
719 
720     CPU_FOREACH(cpu_iter) {
721         if (cpu_iter->thread_id == pid) {
722             trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
723             return cpu_iter->cpu_index;
724         }
725     }
726     trace_get_mem_fault_cpu_index(-1, pid);
727     return -1;
728 }
729 
730 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
731 {
732     int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
733                                     dc->start_time;
734     return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
735 }
736 
737 /*
738  * This function is being called when pagefault occurs. It
739  * tracks down vCPU blocking time.
740  *
741  * @addr: faulted host virtual address
742  * @ptid: faulted process thread id
743  * @rb: ramblock appropriate to addr
744  */
745 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
746                                           RAMBlock *rb)
747 {
748     int cpu, already_received;
749     MigrationIncomingState *mis = migration_incoming_get_current();
750     PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
751     uint32_t low_time_offset;
752 
753     if (!dc || ptid == 0) {
754         return;
755     }
756     cpu = get_mem_fault_cpu_index(ptid);
757     if (cpu < 0) {
758         return;
759     }
760 
761     low_time_offset = get_low_time_offset(dc);
762     if (dc->vcpu_addr[cpu] == 0) {
763         atomic_inc(&dc->smp_cpus_down);
764     }
765 
766     atomic_xchg(&dc->last_begin, low_time_offset);
767     atomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
768     atomic_xchg(&dc->vcpu_addr[cpu], addr);
769 
770     /* check it here, not at the begining of the function,
771      * due to, check could accur early than bitmap_set in
772      * qemu_ufd_copy_ioctl */
773     already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
774     if (already_received) {
775         atomic_xchg(&dc->vcpu_addr[cpu], 0);
776         atomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
777         atomic_dec(&dc->smp_cpus_down);
778     }
779     trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
780                                         cpu, already_received);
781 }
782 
783 /*
784  *  This function just provide calculated blocktime per cpu and trace it.
785  *  Total blocktime is calculated in mark_postcopy_blocktime_end.
786  *
787  *
788  * Assume we have 3 CPU
789  *
790  *      S1        E1           S1               E1
791  * -----***********------------xxx***************------------------------> CPU1
792  *
793  *             S2                E2
794  * ------------****************xxx---------------------------------------> CPU2
795  *
796  *                         S3            E3
797  * ------------------------****xxx********-------------------------------> CPU3
798  *
799  * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
800  * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
801  * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
802  *            it's a part of total blocktime.
803  * S1 - here is last_begin
804  * Legend of the picture is following:
805  *              * - means blocktime per vCPU
806  *              x - means overlapped blocktime (total blocktime)
807  *
808  * @addr: host virtual address
809  */
810 static void mark_postcopy_blocktime_end(uintptr_t addr)
811 {
812     MigrationIncomingState *mis = migration_incoming_get_current();
813     PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
814     MachineState *ms = MACHINE(qdev_get_machine());
815     unsigned int smp_cpus = ms->smp.cpus;
816     int i, affected_cpu = 0;
817     bool vcpu_total_blocktime = false;
818     uint32_t read_vcpu_time, low_time_offset;
819 
820     if (!dc) {
821         return;
822     }
823 
824     low_time_offset = get_low_time_offset(dc);
825     /* lookup cpu, to clear it,
826      * that algorithm looks straighforward, but it's not
827      * optimal, more optimal algorithm is keeping tree or hash
828      * where key is address value is a list of  */
829     for (i = 0; i < smp_cpus; i++) {
830         uint32_t vcpu_blocktime = 0;
831 
832         read_vcpu_time = atomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
833         if (atomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
834             read_vcpu_time == 0) {
835             continue;
836         }
837         atomic_xchg(&dc->vcpu_addr[i], 0);
838         vcpu_blocktime = low_time_offset - read_vcpu_time;
839         affected_cpu += 1;
840         /* we need to know is that mark_postcopy_end was due to
841          * faulted page, another possible case it's prefetched
842          * page and in that case we shouldn't be here */
843         if (!vcpu_total_blocktime &&
844             atomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
845             vcpu_total_blocktime = true;
846         }
847         /* continue cycle, due to one page could affect several vCPUs */
848         dc->vcpu_blocktime[i] += vcpu_blocktime;
849     }
850 
851     atomic_sub(&dc->smp_cpus_down, affected_cpu);
852     if (vcpu_total_blocktime) {
853         dc->total_blocktime += low_time_offset - atomic_fetch_add(
854                 &dc->last_begin, 0);
855     }
856     trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
857                                       affected_cpu);
858 }
859 
860 static bool postcopy_pause_fault_thread(MigrationIncomingState *mis)
861 {
862     trace_postcopy_pause_fault_thread();
863 
864     qemu_sem_wait(&mis->postcopy_pause_sem_fault);
865 
866     trace_postcopy_pause_fault_thread_continued();
867 
868     return true;
869 }
870 
871 /*
872  * Handle faults detected by the USERFAULT markings
873  */
874 static void *postcopy_ram_fault_thread(void *opaque)
875 {
876     MigrationIncomingState *mis = opaque;
877     struct uffd_msg msg;
878     int ret;
879     size_t index;
880     RAMBlock *rb = NULL;
881 
882     trace_postcopy_ram_fault_thread_entry();
883     rcu_register_thread();
884     mis->last_rb = NULL; /* last RAMBlock we sent part of */
885     qemu_sem_post(&mis->fault_thread_sem);
886 
887     struct pollfd *pfd;
888     size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
889 
890     pfd = g_new0(struct pollfd, pfd_len);
891 
892     pfd[0].fd = mis->userfault_fd;
893     pfd[0].events = POLLIN;
894     pfd[1].fd = mis->userfault_event_fd;
895     pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
896     trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
897     for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
898         struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
899                                                  struct PostCopyFD, index);
900         pfd[2 + index].fd = pcfd->fd;
901         pfd[2 + index].events = POLLIN;
902         trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
903                                                   pcfd->fd);
904     }
905 
906     while (true) {
907         ram_addr_t rb_offset;
908         int poll_result;
909 
910         /*
911          * We're mainly waiting for the kernel to give us a faulting HVA,
912          * however we can be told to quit via userfault_quit_fd which is
913          * an eventfd
914          */
915 
916         poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
917         if (poll_result == -1) {
918             error_report("%s: userfault poll: %s", __func__, strerror(errno));
919             break;
920         }
921 
922         if (!mis->to_src_file) {
923             /*
924              * Possibly someone tells us that the return path is
925              * broken already using the event. We should hold until
926              * the channel is rebuilt.
927              */
928             if (postcopy_pause_fault_thread(mis)) {
929                 mis->last_rb = NULL;
930                 /* Continue to read the userfaultfd */
931             } else {
932                 error_report("%s: paused but don't allow to continue",
933                              __func__);
934                 break;
935             }
936         }
937 
938         if (pfd[1].revents) {
939             uint64_t tmp64 = 0;
940 
941             /* Consume the signal */
942             if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
943                 /* Nothing obviously nicer than posting this error. */
944                 error_report("%s: read() failed", __func__);
945             }
946 
947             if (atomic_read(&mis->fault_thread_quit)) {
948                 trace_postcopy_ram_fault_thread_quit();
949                 break;
950             }
951         }
952 
953         if (pfd[0].revents) {
954             poll_result--;
955             ret = read(mis->userfault_fd, &msg, sizeof(msg));
956             if (ret != sizeof(msg)) {
957                 if (errno == EAGAIN) {
958                     /*
959                      * if a wake up happens on the other thread just after
960                      * the poll, there is nothing to read.
961                      */
962                     continue;
963                 }
964                 if (ret < 0) {
965                     error_report("%s: Failed to read full userfault "
966                                  "message: %s",
967                                  __func__, strerror(errno));
968                     break;
969                 } else {
970                     error_report("%s: Read %d bytes from userfaultfd "
971                                  "expected %zd",
972                                  __func__, ret, sizeof(msg));
973                     break; /* Lost alignment, don't know what we'd read next */
974                 }
975             }
976             if (msg.event != UFFD_EVENT_PAGEFAULT) {
977                 error_report("%s: Read unexpected event %ud from userfaultfd",
978                              __func__, msg.event);
979                 continue; /* It's not a page fault, shouldn't happen */
980             }
981 
982             rb = qemu_ram_block_from_host(
983                      (void *)(uintptr_t)msg.arg.pagefault.address,
984                      true, &rb_offset);
985             if (!rb) {
986                 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
987                              PRIx64, (uint64_t)msg.arg.pagefault.address);
988                 break;
989             }
990 
991             rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
992             trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
993                                                 qemu_ram_get_idstr(rb),
994                                                 rb_offset,
995                                                 msg.arg.pagefault.feat.ptid);
996             mark_postcopy_blocktime_begin(
997                     (uintptr_t)(msg.arg.pagefault.address),
998                                 msg.arg.pagefault.feat.ptid, rb);
999 
1000 retry:
1001             /*
1002              * Send the request to the source - we want to request one
1003              * of our host page sizes (which is >= TPS)
1004              */
1005             if (rb != mis->last_rb) {
1006                 mis->last_rb = rb;
1007                 ret = migrate_send_rp_req_pages(mis,
1008                                                 qemu_ram_get_idstr(rb),
1009                                                 rb_offset,
1010                                                 qemu_ram_pagesize(rb));
1011             } else {
1012                 /* Save some space */
1013                 ret = migrate_send_rp_req_pages(mis,
1014                                                 NULL,
1015                                                 rb_offset,
1016                                                 qemu_ram_pagesize(rb));
1017             }
1018 
1019             if (ret) {
1020                 /* May be network failure, try to wait for recovery */
1021                 if (ret == -EIO && postcopy_pause_fault_thread(mis)) {
1022                     /* We got reconnected somehow, try to continue */
1023                     mis->last_rb = NULL;
1024                     goto retry;
1025                 } else {
1026                     /* This is a unavoidable fault */
1027                     error_report("%s: migrate_send_rp_req_pages() get %d",
1028                                  __func__, ret);
1029                     break;
1030                 }
1031             }
1032         }
1033 
1034         /* Now handle any requests from external processes on shared memory */
1035         /* TODO: May need to handle devices deregistering during postcopy */
1036         for (index = 2; index < pfd_len && poll_result; index++) {
1037             if (pfd[index].revents) {
1038                 struct PostCopyFD *pcfd =
1039                     &g_array_index(mis->postcopy_remote_fds,
1040                                    struct PostCopyFD, index - 2);
1041 
1042                 poll_result--;
1043                 if (pfd[index].revents & POLLERR) {
1044                     error_report("%s: POLLERR on poll %zd fd=%d",
1045                                  __func__, index, pcfd->fd);
1046                     pfd[index].events = 0;
1047                     continue;
1048                 }
1049 
1050                 ret = read(pcfd->fd, &msg, sizeof(msg));
1051                 if (ret != sizeof(msg)) {
1052                     if (errno == EAGAIN) {
1053                         /*
1054                          * if a wake up happens on the other thread just after
1055                          * the poll, there is nothing to read.
1056                          */
1057                         continue;
1058                     }
1059                     if (ret < 0) {
1060                         error_report("%s: Failed to read full userfault "
1061                                      "message: %s (shared) revents=%d",
1062                                      __func__, strerror(errno),
1063                                      pfd[index].revents);
1064                         /*TODO: Could just disable this sharer */
1065                         break;
1066                     } else {
1067                         error_report("%s: Read %d bytes from userfaultfd "
1068                                      "expected %zd (shared)",
1069                                      __func__, ret, sizeof(msg));
1070                         /*TODO: Could just disable this sharer */
1071                         break; /*Lost alignment,don't know what we'd read next*/
1072                     }
1073                 }
1074                 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1075                     error_report("%s: Read unexpected event %ud "
1076                                  "from userfaultfd (shared)",
1077                                  __func__, msg.event);
1078                     continue; /* It's not a page fault, shouldn't happen */
1079                 }
1080                 /* Call the device handler registered with us */
1081                 ret = pcfd->handler(pcfd, &msg);
1082                 if (ret) {
1083                     error_report("%s: Failed to resolve shared fault on %zd/%s",
1084                                  __func__, index, pcfd->idstr);
1085                     /* TODO: Fail? Disable this sharer? */
1086                 }
1087             }
1088         }
1089     }
1090     rcu_unregister_thread();
1091     trace_postcopy_ram_fault_thread_exit();
1092     g_free(pfd);
1093     return NULL;
1094 }
1095 
1096 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
1097 {
1098     /* Open the fd for the kernel to give us userfaults */
1099     mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1100     if (mis->userfault_fd == -1) {
1101         error_report("%s: Failed to open userfault fd: %s", __func__,
1102                      strerror(errno));
1103         return -1;
1104     }
1105 
1106     /*
1107      * Although the host check already tested the API, we need to
1108      * do the check again as an ABI handshake on the new fd.
1109      */
1110     if (!ufd_check_and_apply(mis->userfault_fd, mis)) {
1111         return -1;
1112     }
1113 
1114     /* Now an eventfd we use to tell the fault-thread to quit */
1115     mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1116     if (mis->userfault_event_fd == -1) {
1117         error_report("%s: Opening userfault_event_fd: %s", __func__,
1118                      strerror(errno));
1119         close(mis->userfault_fd);
1120         return -1;
1121     }
1122 
1123     qemu_sem_init(&mis->fault_thread_sem, 0);
1124     qemu_thread_create(&mis->fault_thread, "postcopy/fault",
1125                        postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
1126     qemu_sem_wait(&mis->fault_thread_sem);
1127     qemu_sem_destroy(&mis->fault_thread_sem);
1128     mis->have_fault_thread = true;
1129 
1130     /* Mark so that we get notified of accesses to unwritten areas */
1131     if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1132         error_report("ram_block_enable_notify failed");
1133         return -1;
1134     }
1135 
1136     /*
1137      * Ballooning can mark pages as absent while we're postcopying
1138      * that would cause false userfaults.
1139      */
1140     postcopy_balloon_inhibit(true);
1141 
1142     trace_postcopy_ram_enable_notify();
1143 
1144     return 0;
1145 }
1146 
1147 static int qemu_ufd_copy_ioctl(int userfault_fd, void *host_addr,
1148                                void *from_addr, uint64_t pagesize, RAMBlock *rb)
1149 {
1150     int ret;
1151     if (from_addr) {
1152         struct uffdio_copy copy_struct;
1153         copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1154         copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1155         copy_struct.len = pagesize;
1156         copy_struct.mode = 0;
1157         ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1158     } else {
1159         struct uffdio_zeropage zero_struct;
1160         zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1161         zero_struct.range.len = pagesize;
1162         zero_struct.mode = 0;
1163         ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1164     }
1165     if (!ret) {
1166         ramblock_recv_bitmap_set_range(rb, host_addr,
1167                                        pagesize / qemu_target_page_size());
1168         mark_postcopy_blocktime_end((uintptr_t)host_addr);
1169 
1170     }
1171     return ret;
1172 }
1173 
1174 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1175 {
1176     int i;
1177     MigrationIncomingState *mis = migration_incoming_get_current();
1178     GArray *pcrfds = mis->postcopy_remote_fds;
1179 
1180     for (i = 0; i < pcrfds->len; i++) {
1181         struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1182         int ret = cur->waker(cur, rb, offset);
1183         if (ret) {
1184             return ret;
1185         }
1186     }
1187     return 0;
1188 }
1189 
1190 /*
1191  * Place a host page (from) at (host) atomically
1192  * returns 0 on success
1193  */
1194 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1195                         RAMBlock *rb)
1196 {
1197     size_t pagesize = qemu_ram_pagesize(rb);
1198 
1199     /* copy also acks to the kernel waking the stalled thread up
1200      * TODO: We can inhibit that ack and only do it if it was requested
1201      * which would be slightly cheaper, but we'd have to be careful
1202      * of the order of updating our page state.
1203      */
1204     if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, from, pagesize, rb)) {
1205         int e = errno;
1206         error_report("%s: %s copy host: %p from: %p (size: %zd)",
1207                      __func__, strerror(e), host, from, pagesize);
1208 
1209         return -e;
1210     }
1211 
1212     trace_postcopy_place_page(host);
1213     return postcopy_notify_shared_wake(rb,
1214                                        qemu_ram_block_host_offset(rb, host));
1215 }
1216 
1217 /*
1218  * Place a zero page at (host) atomically
1219  * returns 0 on success
1220  */
1221 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1222                              RAMBlock *rb)
1223 {
1224     size_t pagesize = qemu_ram_pagesize(rb);
1225     trace_postcopy_place_page_zero(host);
1226 
1227     /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1228      * but it's not available for everything (e.g. hugetlbpages)
1229      */
1230     if (qemu_ram_is_uf_zeroable(rb)) {
1231         if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, NULL, pagesize, rb)) {
1232             int e = errno;
1233             error_report("%s: %s zero host: %p",
1234                          __func__, strerror(e), host);
1235 
1236             return -e;
1237         }
1238         return postcopy_notify_shared_wake(rb,
1239                                            qemu_ram_block_host_offset(rb,
1240                                                                       host));
1241     } else {
1242         /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */
1243         if (!mis->postcopy_tmp_zero_page) {
1244             mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1245                                                PROT_READ | PROT_WRITE,
1246                                                MAP_PRIVATE | MAP_ANONYMOUS,
1247                                                -1, 0);
1248             if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1249                 int e = errno;
1250                 mis->postcopy_tmp_zero_page = NULL;
1251                 error_report("%s: %s mapping large zero page",
1252                              __func__, strerror(e));
1253                 return -e;
1254             }
1255             memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1256         }
1257         return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page,
1258                                    rb);
1259     }
1260 }
1261 
1262 /*
1263  * Returns a target page of memory that can be mapped at a later point in time
1264  * using postcopy_place_page
1265  * The same address is used repeatedly, postcopy_place_page just takes the
1266  * backing page away.
1267  * Returns: Pointer to allocated page
1268  *
1269  */
1270 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
1271 {
1272     if (!mis->postcopy_tmp_page) {
1273         mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
1274                              PROT_READ | PROT_WRITE, MAP_PRIVATE |
1275                              MAP_ANONYMOUS, -1, 0);
1276         if (mis->postcopy_tmp_page == MAP_FAILED) {
1277             mis->postcopy_tmp_page = NULL;
1278             error_report("%s: %s", __func__, strerror(errno));
1279             return NULL;
1280         }
1281     }
1282 
1283     return mis->postcopy_tmp_page;
1284 }
1285 
1286 #else
1287 /* No target OS support, stubs just fail */
1288 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1289 {
1290 }
1291 
1292 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
1293 {
1294     error_report("%s: No OS support", __func__);
1295     return false;
1296 }
1297 
1298 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1299 {
1300     error_report("postcopy_ram_incoming_init: No OS support");
1301     return -1;
1302 }
1303 
1304 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1305 {
1306     assert(0);
1307     return -1;
1308 }
1309 
1310 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1311 {
1312     assert(0);
1313     return -1;
1314 }
1315 
1316 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1317                                  uint64_t client_addr, uint64_t rb_offset)
1318 {
1319     assert(0);
1320     return -1;
1321 }
1322 
1323 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
1324 {
1325     assert(0);
1326     return -1;
1327 }
1328 
1329 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1330                         RAMBlock *rb)
1331 {
1332     assert(0);
1333     return -1;
1334 }
1335 
1336 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1337                         RAMBlock *rb)
1338 {
1339     assert(0);
1340     return -1;
1341 }
1342 
1343 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
1344 {
1345     assert(0);
1346     return NULL;
1347 }
1348 
1349 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1350                          uint64_t client_addr,
1351                          RAMBlock *rb)
1352 {
1353     assert(0);
1354     return -1;
1355 }
1356 #endif
1357 
1358 /* ------------------------------------------------------------------------- */
1359 
1360 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1361 {
1362     uint64_t tmp64 = 1;
1363 
1364     /*
1365      * Wakeup the fault_thread.  It's an eventfd that should currently
1366      * be at 0, we're going to increment it to 1
1367      */
1368     if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1369         /* Not much we can do here, but may as well report it */
1370         error_report("%s: incrementing failed: %s", __func__,
1371                      strerror(errno));
1372     }
1373 }
1374 
1375 /**
1376  * postcopy_discard_send_init: Called at the start of each RAMBlock before
1377  *   asking to discard individual ranges.
1378  *
1379  * @ms: The current migration state.
1380  * @offset: the bitmap offset of the named RAMBlock in the migration
1381  *   bitmap.
1382  * @name: RAMBlock that discards will operate on.
1383  *
1384  * returns: a new PDS.
1385  */
1386 PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms,
1387                                                  const char *name)
1388 {
1389     PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState));
1390 
1391     if (res) {
1392         res->ramblock_name = name;
1393     }
1394 
1395     return res;
1396 }
1397 
1398 /**
1399  * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1400  *   discard. May send a discard message, may just leave it queued to
1401  *   be sent later.
1402  *
1403  * @ms: Current migration state.
1404  * @pds: Structure initialised by postcopy_discard_send_init().
1405  * @start,@length: a range of pages in the migration bitmap in the
1406  *   RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1407  */
1408 void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds,
1409                                 unsigned long start, unsigned long length)
1410 {
1411     size_t tp_size = qemu_target_page_size();
1412     /* Convert to byte offsets within the RAM block */
1413     pds->start_list[pds->cur_entry] = start  * tp_size;
1414     pds->length_list[pds->cur_entry] = length * tp_size;
1415     trace_postcopy_discard_send_range(pds->ramblock_name, start, length);
1416     pds->cur_entry++;
1417     pds->nsentwords++;
1418 
1419     if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) {
1420         /* Full set, ship it! */
1421         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1422                                               pds->ramblock_name,
1423                                               pds->cur_entry,
1424                                               pds->start_list,
1425                                               pds->length_list);
1426         pds->nsentcmds++;
1427         pds->cur_entry = 0;
1428     }
1429 }
1430 
1431 /**
1432  * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1433  * bitmap code. Sends any outstanding discard messages, frees the PDS
1434  *
1435  * @ms: Current migration state.
1436  * @pds: Structure initialised by postcopy_discard_send_init().
1437  */
1438 void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds)
1439 {
1440     /* Anything unsent? */
1441     if (pds->cur_entry) {
1442         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1443                                               pds->ramblock_name,
1444                                               pds->cur_entry,
1445                                               pds->start_list,
1446                                               pds->length_list);
1447         pds->nsentcmds++;
1448     }
1449 
1450     trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords,
1451                                        pds->nsentcmds);
1452 
1453     g_free(pds);
1454 }
1455 
1456 /*
1457  * Current state of incoming postcopy; note this is not part of
1458  * MigrationIncomingState since it's state is used during cleanup
1459  * at the end as MIS is being freed.
1460  */
1461 static PostcopyState incoming_postcopy_state;
1462 
1463 PostcopyState  postcopy_state_get(void)
1464 {
1465     return atomic_mb_read(&incoming_postcopy_state);
1466 }
1467 
1468 /* Set the state and return the old state */
1469 PostcopyState postcopy_state_set(PostcopyState new_state)
1470 {
1471     return atomic_xchg(&incoming_postcopy_state, new_state);
1472 }
1473 
1474 /* Register a handler for external shared memory postcopy
1475  * called on the destination.
1476  */
1477 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1478 {
1479     MigrationIncomingState *mis = migration_incoming_get_current();
1480 
1481     mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1482                                                   *pcfd);
1483 }
1484 
1485 /* Unregister a handler for external shared memory postcopy
1486  */
1487 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1488 {
1489     guint i;
1490     MigrationIncomingState *mis = migration_incoming_get_current();
1491     GArray *pcrfds = mis->postcopy_remote_fds;
1492 
1493     for (i = 0; i < pcrfds->len; i++) {
1494         struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1495         if (cur->fd == pcfd->fd) {
1496             mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1497             return;
1498         }
1499     }
1500 }
1501