xref: /openbmc/qemu/migration/postcopy-ram.c (revision 05caa062)
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 "qemu/madvise.h"
21 #include "exec/target_page.h"
22 #include "migration.h"
23 #include "qemu-file.h"
24 #include "savevm.h"
25 #include "postcopy-ram.h"
26 #include "ram.h"
27 #include "qapi/error.h"
28 #include "qemu/notify.h"
29 #include "qemu/rcu.h"
30 #include "sysemu/sysemu.h"
31 #include "qemu/error-report.h"
32 #include "trace.h"
33 #include "hw/boards.h"
34 #include "exec/ramblock.h"
35 #include "socket.h"
36 #include "yank_functions.h"
37 #include "tls.h"
38 #include "qemu/userfaultfd.h"
39 #include "qemu/mmap-alloc.h"
40 #include "options.h"
41 
42 /* Arbitrary limit on size of each discard command,
43  * keeps them around ~200 bytes
44  */
45 #define MAX_DISCARDS_PER_COMMAND 12
46 
47 typedef struct PostcopyDiscardState {
48     const char *ramblock_name;
49     uint16_t cur_entry;
50     /*
51      * Start and length of a discard range (bytes)
52      */
53     uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
54     uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
55     unsigned int nsentwords;
56     unsigned int nsentcmds;
57 } PostcopyDiscardState;
58 
59 static NotifierWithReturnList postcopy_notifier_list;
60 
61 void postcopy_infrastructure_init(void)
62 {
63     notifier_with_return_list_init(&postcopy_notifier_list);
64 }
65 
66 void postcopy_add_notifier(NotifierWithReturn *nn)
67 {
68     notifier_with_return_list_add(&postcopy_notifier_list, nn);
69 }
70 
71 void postcopy_remove_notifier(NotifierWithReturn *n)
72 {
73     notifier_with_return_remove(n);
74 }
75 
76 int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp)
77 {
78     struct PostcopyNotifyData pnd;
79     pnd.reason = reason;
80 
81     return notifier_with_return_list_notify(&postcopy_notifier_list,
82                                             &pnd, errp);
83 }
84 
85 /*
86  * NOTE: this routine is not thread safe, we can't call it concurrently. But it
87  * should be good enough for migration's purposes.
88  */
89 void postcopy_thread_create(MigrationIncomingState *mis,
90                             QemuThread *thread, const char *name,
91                             void *(*fn)(void *), int joinable)
92 {
93     qemu_sem_init(&mis->thread_sync_sem, 0);
94     qemu_thread_create(thread, name, fn, mis, joinable);
95     qemu_sem_wait(&mis->thread_sync_sem);
96     qemu_sem_destroy(&mis->thread_sync_sem);
97 }
98 
99 /* Postcopy needs to detect accesses to pages that haven't yet been copied
100  * across, and efficiently map new pages in, the techniques for doing this
101  * are target OS specific.
102  */
103 #if defined(__linux__)
104 #include <poll.h>
105 #include <sys/ioctl.h>
106 #include <sys/syscall.h>
107 #endif
108 
109 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
110 #include <sys/eventfd.h>
111 #include <linux/userfaultfd.h>
112 
113 typedef struct PostcopyBlocktimeContext {
114     /* time when page fault initiated per vCPU */
115     uint32_t *page_fault_vcpu_time;
116     /* page address per vCPU */
117     uintptr_t *vcpu_addr;
118     uint32_t total_blocktime;
119     /* blocktime per vCPU */
120     uint32_t *vcpu_blocktime;
121     /* point in time when last page fault was initiated */
122     uint32_t last_begin;
123     /* number of vCPU are suspended */
124     int smp_cpus_down;
125     uint64_t start_time;
126 
127     /*
128      * Handler for exit event, necessary for
129      * releasing whole blocktime_ctx
130      */
131     Notifier exit_notifier;
132 } PostcopyBlocktimeContext;
133 
134 static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx)
135 {
136     g_free(ctx->page_fault_vcpu_time);
137     g_free(ctx->vcpu_addr);
138     g_free(ctx->vcpu_blocktime);
139     g_free(ctx);
140 }
141 
142 static void migration_exit_cb(Notifier *n, void *data)
143 {
144     PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext,
145                                                  exit_notifier);
146     destroy_blocktime_context(ctx);
147 }
148 
149 static struct PostcopyBlocktimeContext *blocktime_context_new(void)
150 {
151     MachineState *ms = MACHINE(qdev_get_machine());
152     unsigned int smp_cpus = ms->smp.cpus;
153     PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1);
154     ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus);
155     ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus);
156     ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus);
157 
158     ctx->exit_notifier.notify = migration_exit_cb;
159     ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
160     qemu_add_exit_notifier(&ctx->exit_notifier);
161     return ctx;
162 }
163 
164 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx)
165 {
166     MachineState *ms = MACHINE(qdev_get_machine());
167     uint32List *list = NULL;
168     int i;
169 
170     for (i = ms->smp.cpus - 1; i >= 0; i--) {
171         QAPI_LIST_PREPEND(list, ctx->vcpu_blocktime[i]);
172     }
173 
174     return list;
175 }
176 
177 /*
178  * This function just populates MigrationInfo from postcopy's
179  * blocktime context. It will not populate MigrationInfo,
180  * unless postcopy-blocktime capability was set.
181  *
182  * @info: pointer to MigrationInfo to populate
183  */
184 void fill_destination_postcopy_migration_info(MigrationInfo *info)
185 {
186     MigrationIncomingState *mis = migration_incoming_get_current();
187     PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
188 
189     if (!bc) {
190         return;
191     }
192 
193     info->has_postcopy_blocktime = true;
194     info->postcopy_blocktime = bc->total_blocktime;
195     info->has_postcopy_vcpu_blocktime = true;
196     info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc);
197 }
198 
199 static uint32_t get_postcopy_total_blocktime(void)
200 {
201     MigrationIncomingState *mis = migration_incoming_get_current();
202     PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
203 
204     if (!bc) {
205         return 0;
206     }
207 
208     return bc->total_blocktime;
209 }
210 
211 /**
212  * receive_ufd_features: check userfault fd features, to request only supported
213  * features in the future.
214  *
215  * Returns: true on success
216  *
217  * __NR_userfaultfd - should be checked before
218  *  @features: out parameter will contain uffdio_api.features provided by kernel
219  *              in case of success
220  */
221 static bool receive_ufd_features(uint64_t *features)
222 {
223     struct uffdio_api api_struct = {0};
224     int ufd;
225     bool ret = true;
226 
227     ufd = uffd_open(O_CLOEXEC);
228     if (ufd == -1) {
229         error_report("%s: uffd_open() failed: %s", __func__, strerror(errno));
230         return false;
231     }
232 
233     /* ask features */
234     api_struct.api = UFFD_API;
235     api_struct.features = 0;
236     if (ioctl(ufd, UFFDIO_API, &api_struct)) {
237         error_report("%s: UFFDIO_API failed: %s", __func__,
238                      strerror(errno));
239         ret = false;
240         goto release_ufd;
241     }
242 
243     *features = api_struct.features;
244 
245 release_ufd:
246     close(ufd);
247     return ret;
248 }
249 
250 /**
251  * request_ufd_features: this function should be called only once on a newly
252  * opened ufd, subsequent calls will lead to error.
253  *
254  * Returns: true on success
255  *
256  * @ufd: fd obtained from userfaultfd syscall
257  * @features: bit mask see UFFD_API_FEATURES
258  */
259 static bool request_ufd_features(int ufd, uint64_t features)
260 {
261     struct uffdio_api api_struct = {0};
262     uint64_t ioctl_mask;
263 
264     api_struct.api = UFFD_API;
265     api_struct.features = features;
266     if (ioctl(ufd, UFFDIO_API, &api_struct)) {
267         error_report("%s failed: UFFDIO_API failed: %s", __func__,
268                      strerror(errno));
269         return false;
270     }
271 
272     ioctl_mask = 1ULL << _UFFDIO_REGISTER |
273                  1ULL << _UFFDIO_UNREGISTER;
274     if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
275         error_report("Missing userfault features: %" PRIx64,
276                      (uint64_t)(~api_struct.ioctls & ioctl_mask));
277         return false;
278     }
279 
280     return true;
281 }
282 
283 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis,
284                                 Error **errp)
285 {
286     ERRP_GUARD();
287     uint64_t asked_features = 0;
288     static uint64_t supported_features;
289 
290     /*
291      * it's not possible to
292      * request UFFD_API twice per one fd
293      * userfault fd features is persistent
294      */
295     if (!supported_features) {
296         if (!receive_ufd_features(&supported_features)) {
297             error_setg(errp, "Userfault feature detection failed");
298             return false;
299         }
300     }
301 
302 #ifdef UFFD_FEATURE_THREAD_ID
303     if (UFFD_FEATURE_THREAD_ID & supported_features) {
304         asked_features |= UFFD_FEATURE_THREAD_ID;
305         if (migrate_postcopy_blocktime()) {
306             if (!mis->blocktime_ctx) {
307                 mis->blocktime_ctx = blocktime_context_new();
308             }
309         }
310     }
311 #endif
312 
313     /*
314      * request features, even if asked_features is 0, due to
315      * kernel expects UFFD_API before UFFDIO_REGISTER, per
316      * userfault file descriptor
317      */
318     if (!request_ufd_features(ufd, asked_features)) {
319         error_setg(errp, "Failed features %" PRIu64, asked_features);
320         return false;
321     }
322 
323     if (qemu_real_host_page_size() != ram_pagesize_summary()) {
324         bool have_hp = false;
325         /* We've got a huge page */
326 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
327         have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS;
328 #endif
329         if (!have_hp) {
330             error_setg(errp,
331                        "Userfault on this host does not support huge pages");
332             return false;
333         }
334     }
335     return true;
336 }
337 
338 /* Callback from postcopy_ram_supported_by_host block iterator.
339  */
340 static int test_ramblock_postcopiable(RAMBlock *rb, Error **errp)
341 {
342     const char *block_name = qemu_ram_get_idstr(rb);
343     ram_addr_t length = qemu_ram_get_used_length(rb);
344     size_t pagesize = qemu_ram_pagesize(rb);
345     QemuFsType fs;
346 
347     if (length % pagesize) {
348         error_setg(errp,
349                    "Postcopy requires RAM blocks to be a page size multiple,"
350                    " block %s is 0x" RAM_ADDR_FMT " bytes with a "
351                    "page size of 0x%zx", block_name, length, pagesize);
352         return 1;
353     }
354 
355     if (rb->fd >= 0) {
356         fs = qemu_fd_getfs(rb->fd);
357         if (fs != QEMU_FS_TYPE_TMPFS && fs != QEMU_FS_TYPE_HUGETLBFS) {
358             error_setg(errp,
359                        "Host backend files need to be TMPFS or HUGETLBFS only");
360             return 1;
361         }
362     }
363 
364     return 0;
365 }
366 
367 /*
368  * Note: This has the side effect of munlock'ing all of RAM, that's
369  * normally fine since if the postcopy succeeds it gets turned back on at the
370  * end.
371  */
372 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis, Error **errp)
373 {
374     ERRP_GUARD();
375     long pagesize = qemu_real_host_page_size();
376     int ufd = -1;
377     bool ret = false; /* Error unless we change it */
378     void *testarea = NULL;
379     struct uffdio_register reg_struct;
380     struct uffdio_range range_struct;
381     uint64_t feature_mask;
382     RAMBlock *block;
383 
384     if (qemu_target_page_size() > pagesize) {
385         error_setg(errp, "Target page size bigger than host page size");
386         goto out;
387     }
388 
389     ufd = uffd_open(O_CLOEXEC);
390     if (ufd == -1) {
391         error_setg(errp, "Userfaultfd not available: %s", strerror(errno));
392         goto out;
393     }
394 
395     /* Give devices a chance to object */
396     if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, errp)) {
397         goto out;
398     }
399 
400     /* Version and features check */
401     if (!ufd_check_and_apply(ufd, mis, errp)) {
402         goto out;
403     }
404 
405     /*
406      * We don't support postcopy with some type of ramblocks.
407      *
408      * NOTE: we explicitly ignored migrate_ram_is_ignored() instead we checked
409      * all possible ramblocks.  This is because this function can be called
410      * when creating the migration object, during the phase RAM_MIGRATABLE
411      * is not even properly set for all the ramblocks.
412      *
413      * A side effect of this is we'll also check against RAM_SHARED
414      * ramblocks even if migrate_ignore_shared() is set (in which case
415      * we'll never migrate RAM_SHARED at all), but normally this shouldn't
416      * affect in reality, or we can revisit.
417      */
418     RAMBLOCK_FOREACH(block) {
419         if (test_ramblock_postcopiable(block, errp)) {
420             goto out;
421         }
422     }
423 
424     /*
425      * userfault and mlock don't go together; we'll put it back later if
426      * it was enabled.
427      */
428     if (munlockall()) {
429         error_setg(errp, "munlockall() failed: %s", strerror(errno));
430         goto out;
431     }
432 
433     /*
434      *  We need to check that the ops we need are supported on anon memory
435      *  To do that we need to register a chunk and see the flags that
436      *  are returned.
437      */
438     testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
439                                     MAP_ANONYMOUS, -1, 0);
440     if (testarea == MAP_FAILED) {
441         error_setg(errp, "Failed to map test area: %s", strerror(errno));
442         goto out;
443     }
444     g_assert(QEMU_PTR_IS_ALIGNED(testarea, pagesize));
445 
446     reg_struct.range.start = (uintptr_t)testarea;
447     reg_struct.range.len = pagesize;
448     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
449 
450     if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
451         error_setg(errp, "UFFDIO_REGISTER failed: %s", strerror(errno));
452         goto out;
453     }
454 
455     range_struct.start = (uintptr_t)testarea;
456     range_struct.len = pagesize;
457     if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
458         error_setg(errp, "UFFDIO_UNREGISTER failed: %s", strerror(errno));
459         goto out;
460     }
461 
462     feature_mask = 1ULL << _UFFDIO_WAKE |
463                    1ULL << _UFFDIO_COPY |
464                    1ULL << _UFFDIO_ZEROPAGE;
465     if ((reg_struct.ioctls & feature_mask) != feature_mask) {
466         error_setg(errp, "Missing userfault map features: %" PRIx64,
467                    (uint64_t)(~reg_struct.ioctls & feature_mask));
468         goto out;
469     }
470 
471     /* Success! */
472     ret = true;
473 out:
474     if (testarea) {
475         munmap(testarea, pagesize);
476     }
477     if (ufd != -1) {
478         close(ufd);
479     }
480     return ret;
481 }
482 
483 /*
484  * Setup an area of RAM so that it *can* be used for postcopy later; this
485  * must be done right at the start prior to pre-copy.
486  * opaque should be the MIS.
487  */
488 static int init_range(RAMBlock *rb, void *opaque)
489 {
490     const char *block_name = qemu_ram_get_idstr(rb);
491     void *host_addr = qemu_ram_get_host_addr(rb);
492     ram_addr_t offset = qemu_ram_get_offset(rb);
493     ram_addr_t length = qemu_ram_get_used_length(rb);
494     trace_postcopy_init_range(block_name, host_addr, offset, length);
495 
496     /*
497      * Save the used_length before running the guest. In case we have to
498      * resize RAM blocks when syncing RAM block sizes from the source during
499      * precopy, we'll update it manually via the ram block notifier.
500      */
501     rb->postcopy_length = length;
502 
503     /*
504      * We need the whole of RAM to be truly empty for postcopy, so things
505      * like ROMs and any data tables built during init must be zero'd
506      * - we're going to get the copy from the source anyway.
507      * (Precopy will just overwrite this data, so doesn't need the discard)
508      */
509     if (ram_discard_range(block_name, 0, length)) {
510         return -1;
511     }
512 
513     return 0;
514 }
515 
516 /*
517  * At the end of migration, undo the effects of init_range
518  * opaque should be the MIS.
519  */
520 static int cleanup_range(RAMBlock *rb, void *opaque)
521 {
522     const char *block_name = qemu_ram_get_idstr(rb);
523     void *host_addr = qemu_ram_get_host_addr(rb);
524     ram_addr_t offset = qemu_ram_get_offset(rb);
525     ram_addr_t length = rb->postcopy_length;
526     MigrationIncomingState *mis = opaque;
527     struct uffdio_range range_struct;
528     trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
529 
530     /*
531      * We turned off hugepage for the precopy stage with postcopy enabled
532      * we can turn it back on now.
533      */
534     qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
535 
536     /*
537      * We can also turn off userfault now since we should have all the
538      * pages.   It can be useful to leave it on to debug postcopy
539      * if you're not sure it's always getting every page.
540      */
541     range_struct.start = (uintptr_t)host_addr;
542     range_struct.len = length;
543 
544     if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
545         error_report("%s: userfault unregister %s", __func__, strerror(errno));
546 
547         return -1;
548     }
549 
550     return 0;
551 }
552 
553 /*
554  * Initialise postcopy-ram, setting the RAM to a state where we can go into
555  * postcopy later; must be called prior to any precopy.
556  * called from arch_init's similarly named ram_postcopy_incoming_init
557  */
558 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
559 {
560     if (foreach_not_ignored_block(init_range, NULL)) {
561         return -1;
562     }
563 
564     return 0;
565 }
566 
567 static void postcopy_temp_pages_cleanup(MigrationIncomingState *mis)
568 {
569     int i;
570 
571     if (mis->postcopy_tmp_pages) {
572         for (i = 0; i < mis->postcopy_channels; i++) {
573             if (mis->postcopy_tmp_pages[i].tmp_huge_page) {
574                 munmap(mis->postcopy_tmp_pages[i].tmp_huge_page,
575                        mis->largest_page_size);
576                 mis->postcopy_tmp_pages[i].tmp_huge_page = NULL;
577             }
578         }
579         g_free(mis->postcopy_tmp_pages);
580         mis->postcopy_tmp_pages = NULL;
581     }
582 
583     if (mis->postcopy_tmp_zero_page) {
584         munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
585         mis->postcopy_tmp_zero_page = NULL;
586     }
587 }
588 
589 /*
590  * At the end of a migration where postcopy_ram_incoming_init was called.
591  */
592 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
593 {
594     trace_postcopy_ram_incoming_cleanup_entry();
595 
596     if (mis->preempt_thread_status == PREEMPT_THREAD_CREATED) {
597         /* Notify the fast load thread to quit */
598         mis->preempt_thread_status = PREEMPT_THREAD_QUIT;
599         /*
600          * Update preempt_thread_status before reading count.  Note: mutex
601          * lock only provide ACQUIRE semantic, and it doesn't stops this
602          * write to be reordered after reading the count.
603          */
604         smp_mb();
605         /*
606          * It's possible that the preempt thread is still handling the last
607          * pages to arrive which were requested by guest page faults.
608          * Making sure nothing is left behind by waiting on the condvar if
609          * that unlikely case happened.
610          */
611         WITH_QEMU_LOCK_GUARD(&mis->page_request_mutex) {
612             if (qatomic_read(&mis->page_requested_count)) {
613                 /*
614                  * It is guaranteed to receive a signal later, because the
615                  * count>0 now, so it's destined to be decreased to zero
616                  * very soon by the preempt thread.
617                  */
618                 qemu_cond_wait(&mis->page_request_cond,
619                                &mis->page_request_mutex);
620             }
621         }
622         /* Notify the fast load thread to quit */
623         if (mis->postcopy_qemufile_dst) {
624             qemu_file_shutdown(mis->postcopy_qemufile_dst);
625         }
626         qemu_thread_join(&mis->postcopy_prio_thread);
627         mis->preempt_thread_status = PREEMPT_THREAD_NONE;
628     }
629 
630     if (mis->have_fault_thread) {
631         Error *local_err = NULL;
632 
633         /* Let the fault thread quit */
634         qatomic_set(&mis->fault_thread_quit, 1);
635         postcopy_fault_thread_notify(mis);
636         trace_postcopy_ram_incoming_cleanup_join();
637         qemu_thread_join(&mis->fault_thread);
638 
639         if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
640             error_report_err(local_err);
641             return -1;
642         }
643 
644         if (foreach_not_ignored_block(cleanup_range, mis)) {
645             return -1;
646         }
647 
648         trace_postcopy_ram_incoming_cleanup_closeuf();
649         close(mis->userfault_fd);
650         close(mis->userfault_event_fd);
651         mis->have_fault_thread = false;
652     }
653 
654     if (enable_mlock) {
655         if (os_mlock() < 0) {
656             error_report("mlock: %s", strerror(errno));
657             /*
658              * It doesn't feel right to fail at this point, we have a valid
659              * VM state.
660              */
661         }
662     }
663 
664     postcopy_temp_pages_cleanup(mis);
665 
666     trace_postcopy_ram_incoming_cleanup_blocktime(
667             get_postcopy_total_blocktime());
668 
669     trace_postcopy_ram_incoming_cleanup_exit();
670     return 0;
671 }
672 
673 /*
674  * Disable huge pages on an area
675  */
676 static int nhp_range(RAMBlock *rb, void *opaque)
677 {
678     const char *block_name = qemu_ram_get_idstr(rb);
679     void *host_addr = qemu_ram_get_host_addr(rb);
680     ram_addr_t offset = qemu_ram_get_offset(rb);
681     ram_addr_t length = rb->postcopy_length;
682     trace_postcopy_nhp_range(block_name, host_addr, offset, length);
683 
684     /*
685      * Before we do discards we need to ensure those discards really
686      * do delete areas of the page, even if THP thinks a hugepage would
687      * be a good idea, so force hugepages off.
688      */
689     qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
690 
691     return 0;
692 }
693 
694 /*
695  * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
696  * however leaving it until after precopy means that most of the precopy
697  * data is still THPd
698  */
699 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
700 {
701     if (foreach_not_ignored_block(nhp_range, mis)) {
702         return -1;
703     }
704 
705     postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
706 
707     return 0;
708 }
709 
710 /*
711  * Mark the given area of RAM as requiring notification to unwritten areas
712  * Used as a  callback on foreach_not_ignored_block.
713  *   host_addr: Base of area to mark
714  *   offset: Offset in the whole ram arena
715  *   length: Length of the section
716  *   opaque: MigrationIncomingState pointer
717  * Returns 0 on success
718  */
719 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
720 {
721     MigrationIncomingState *mis = opaque;
722     struct uffdio_register reg_struct;
723 
724     reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
725     reg_struct.range.len = rb->postcopy_length;
726     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
727 
728     /* Now tell our userfault_fd that it's responsible for this area */
729     if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
730         error_report("%s userfault register: %s", __func__, strerror(errno));
731         return -1;
732     }
733     if (!(reg_struct.ioctls & (1ULL << _UFFDIO_COPY))) {
734         error_report("%s userfault: Region doesn't support COPY", __func__);
735         return -1;
736     }
737     if (reg_struct.ioctls & (1ULL << _UFFDIO_ZEROPAGE)) {
738         qemu_ram_set_uf_zeroable(rb);
739     }
740 
741     return 0;
742 }
743 
744 int postcopy_wake_shared(struct PostCopyFD *pcfd,
745                          uint64_t client_addr,
746                          RAMBlock *rb)
747 {
748     size_t pagesize = qemu_ram_pagesize(rb);
749     struct uffdio_range range;
750     int ret;
751     trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
752     range.start = ROUND_DOWN(client_addr, pagesize);
753     range.len = pagesize;
754     ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
755     if (ret) {
756         error_report("%s: Failed to wake: %zx in %s (%s)",
757                      __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
758                      strerror(errno));
759     }
760     return ret;
761 }
762 
763 static int postcopy_request_page(MigrationIncomingState *mis, RAMBlock *rb,
764                                  ram_addr_t start, uint64_t haddr)
765 {
766     void *aligned = (void *)(uintptr_t)ROUND_DOWN(haddr, qemu_ram_pagesize(rb));
767 
768     /*
769      * Discarded pages (via RamDiscardManager) are never migrated. On unlikely
770      * access, place a zeropage, which will also set the relevant bits in the
771      * recv_bitmap accordingly, so we won't try placing a zeropage twice.
772      *
773      * Checking a single bit is sufficient to handle pagesize > TPS as either
774      * all relevant bits are set or not.
775      */
776     assert(QEMU_IS_ALIGNED(start, qemu_ram_pagesize(rb)));
777     if (ramblock_page_is_discarded(rb, start)) {
778         bool received = ramblock_recv_bitmap_test_byte_offset(rb, start);
779 
780         return received ? 0 : postcopy_place_page_zero(mis, aligned, rb);
781     }
782 
783     return migrate_send_rp_req_pages(mis, rb, start, haddr);
784 }
785 
786 /*
787  * Callback from shared fault handlers to ask for a page,
788  * the page must be specified by a RAMBlock and an offset in that rb
789  * Note: Only for use by shared fault handlers (in fault thread)
790  */
791 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
792                                  uint64_t client_addr, uint64_t rb_offset)
793 {
794     uint64_t aligned_rbo = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb));
795     MigrationIncomingState *mis = migration_incoming_get_current();
796 
797     trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
798                                        rb_offset);
799     if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
800         trace_postcopy_request_shared_page_present(pcfd->idstr,
801                                         qemu_ram_get_idstr(rb), rb_offset);
802         return postcopy_wake_shared(pcfd, client_addr, rb);
803     }
804     postcopy_request_page(mis, rb, aligned_rbo, client_addr);
805     return 0;
806 }
807 
808 static int get_mem_fault_cpu_index(uint32_t pid)
809 {
810     CPUState *cpu_iter;
811 
812     CPU_FOREACH(cpu_iter) {
813         if (cpu_iter->thread_id == pid) {
814             trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
815             return cpu_iter->cpu_index;
816         }
817     }
818     trace_get_mem_fault_cpu_index(-1, pid);
819     return -1;
820 }
821 
822 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
823 {
824     int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
825                                     dc->start_time;
826     return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
827 }
828 
829 /*
830  * This function is being called when pagefault occurs. It
831  * tracks down vCPU blocking time.
832  *
833  * @addr: faulted host virtual address
834  * @ptid: faulted process thread id
835  * @rb: ramblock appropriate to addr
836  */
837 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
838                                           RAMBlock *rb)
839 {
840     int cpu, already_received;
841     MigrationIncomingState *mis = migration_incoming_get_current();
842     PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
843     uint32_t low_time_offset;
844 
845     if (!dc || ptid == 0) {
846         return;
847     }
848     cpu = get_mem_fault_cpu_index(ptid);
849     if (cpu < 0) {
850         return;
851     }
852 
853     low_time_offset = get_low_time_offset(dc);
854     if (dc->vcpu_addr[cpu] == 0) {
855         qatomic_inc(&dc->smp_cpus_down);
856     }
857 
858     qatomic_xchg(&dc->last_begin, low_time_offset);
859     qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
860     qatomic_xchg(&dc->vcpu_addr[cpu], addr);
861 
862     /*
863      * check it here, not at the beginning of the function,
864      * due to, check could occur early than bitmap_set in
865      * qemu_ufd_copy_ioctl
866      */
867     already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
868     if (already_received) {
869         qatomic_xchg(&dc->vcpu_addr[cpu], 0);
870         qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
871         qatomic_dec(&dc->smp_cpus_down);
872     }
873     trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
874                                         cpu, already_received);
875 }
876 
877 /*
878  *  This function just provide calculated blocktime per cpu and trace it.
879  *  Total blocktime is calculated in mark_postcopy_blocktime_end.
880  *
881  *
882  * Assume we have 3 CPU
883  *
884  *      S1        E1           S1               E1
885  * -----***********------------xxx***************------------------------> CPU1
886  *
887  *             S2                E2
888  * ------------****************xxx---------------------------------------> CPU2
889  *
890  *                         S3            E3
891  * ------------------------****xxx********-------------------------------> CPU3
892  *
893  * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
894  * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
895  * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
896  *            it's a part of total blocktime.
897  * S1 - here is last_begin
898  * Legend of the picture is following:
899  *              * - means blocktime per vCPU
900  *              x - means overlapped blocktime (total blocktime)
901  *
902  * @addr: host virtual address
903  */
904 static void mark_postcopy_blocktime_end(uintptr_t addr)
905 {
906     MigrationIncomingState *mis = migration_incoming_get_current();
907     PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
908     MachineState *ms = MACHINE(qdev_get_machine());
909     unsigned int smp_cpus = ms->smp.cpus;
910     int i, affected_cpu = 0;
911     bool vcpu_total_blocktime = false;
912     uint32_t read_vcpu_time, low_time_offset;
913 
914     if (!dc) {
915         return;
916     }
917 
918     low_time_offset = get_low_time_offset(dc);
919     /* lookup cpu, to clear it,
920      * that algorithm looks straightforward, but it's not
921      * optimal, more optimal algorithm is keeping tree or hash
922      * where key is address value is a list of  */
923     for (i = 0; i < smp_cpus; i++) {
924         uint32_t vcpu_blocktime = 0;
925 
926         read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
927         if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
928             read_vcpu_time == 0) {
929             continue;
930         }
931         qatomic_xchg(&dc->vcpu_addr[i], 0);
932         vcpu_blocktime = low_time_offset - read_vcpu_time;
933         affected_cpu += 1;
934         /* we need to know is that mark_postcopy_end was due to
935          * faulted page, another possible case it's prefetched
936          * page and in that case we shouldn't be here */
937         if (!vcpu_total_blocktime &&
938             qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
939             vcpu_total_blocktime = true;
940         }
941         /* continue cycle, due to one page could affect several vCPUs */
942         dc->vcpu_blocktime[i] += vcpu_blocktime;
943     }
944 
945     qatomic_sub(&dc->smp_cpus_down, affected_cpu);
946     if (vcpu_total_blocktime) {
947         dc->total_blocktime += low_time_offset - qatomic_fetch_add(
948                 &dc->last_begin, 0);
949     }
950     trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
951                                       affected_cpu);
952 }
953 
954 static void postcopy_pause_fault_thread(MigrationIncomingState *mis)
955 {
956     trace_postcopy_pause_fault_thread();
957     qemu_sem_wait(&mis->postcopy_pause_sem_fault);
958     trace_postcopy_pause_fault_thread_continued();
959 }
960 
961 /*
962  * Handle faults detected by the USERFAULT markings
963  */
964 static void *postcopy_ram_fault_thread(void *opaque)
965 {
966     MigrationIncomingState *mis = opaque;
967     struct uffd_msg msg;
968     int ret;
969     size_t index;
970     RAMBlock *rb = NULL;
971 
972     trace_postcopy_ram_fault_thread_entry();
973     rcu_register_thread();
974     mis->last_rb = NULL; /* last RAMBlock we sent part of */
975     qemu_sem_post(&mis->thread_sync_sem);
976 
977     struct pollfd *pfd;
978     size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
979 
980     pfd = g_new0(struct pollfd, pfd_len);
981 
982     pfd[0].fd = mis->userfault_fd;
983     pfd[0].events = POLLIN;
984     pfd[1].fd = mis->userfault_event_fd;
985     pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
986     trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
987     for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
988         struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
989                                                  struct PostCopyFD, index);
990         pfd[2 + index].fd = pcfd->fd;
991         pfd[2 + index].events = POLLIN;
992         trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
993                                                   pcfd->fd);
994     }
995 
996     while (true) {
997         ram_addr_t rb_offset;
998         int poll_result;
999 
1000         /*
1001          * We're mainly waiting for the kernel to give us a faulting HVA,
1002          * however we can be told to quit via userfault_quit_fd which is
1003          * an eventfd
1004          */
1005 
1006         poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
1007         if (poll_result == -1) {
1008             error_report("%s: userfault poll: %s", __func__, strerror(errno));
1009             break;
1010         }
1011 
1012         if (!mis->to_src_file) {
1013             /*
1014              * Possibly someone tells us that the return path is
1015              * broken already using the event. We should hold until
1016              * the channel is rebuilt.
1017              */
1018             postcopy_pause_fault_thread(mis);
1019         }
1020 
1021         if (pfd[1].revents) {
1022             uint64_t tmp64 = 0;
1023 
1024             /* Consume the signal */
1025             if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
1026                 /* Nothing obviously nicer than posting this error. */
1027                 error_report("%s: read() failed", __func__);
1028             }
1029 
1030             if (qatomic_read(&mis->fault_thread_quit)) {
1031                 trace_postcopy_ram_fault_thread_quit();
1032                 break;
1033             }
1034         }
1035 
1036         if (pfd[0].revents) {
1037             poll_result--;
1038             ret = read(mis->userfault_fd, &msg, sizeof(msg));
1039             if (ret != sizeof(msg)) {
1040                 if (errno == EAGAIN) {
1041                     /*
1042                      * if a wake up happens on the other thread just after
1043                      * the poll, there is nothing to read.
1044                      */
1045                     continue;
1046                 }
1047                 if (ret < 0) {
1048                     error_report("%s: Failed to read full userfault "
1049                                  "message: %s",
1050                                  __func__, strerror(errno));
1051                     break;
1052                 } else {
1053                     error_report("%s: Read %d bytes from userfaultfd "
1054                                  "expected %zd",
1055                                  __func__, ret, sizeof(msg));
1056                     break; /* Lost alignment, don't know what we'd read next */
1057                 }
1058             }
1059             if (msg.event != UFFD_EVENT_PAGEFAULT) {
1060                 error_report("%s: Read unexpected event %ud from userfaultfd",
1061                              __func__, msg.event);
1062                 continue; /* It's not a page fault, shouldn't happen */
1063             }
1064 
1065             rb = qemu_ram_block_from_host(
1066                      (void *)(uintptr_t)msg.arg.pagefault.address,
1067                      true, &rb_offset);
1068             if (!rb) {
1069                 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
1070                              PRIx64, (uint64_t)msg.arg.pagefault.address);
1071                 break;
1072             }
1073 
1074             rb_offset = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb));
1075             trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
1076                                                 qemu_ram_get_idstr(rb),
1077                                                 rb_offset,
1078                                                 msg.arg.pagefault.feat.ptid);
1079             mark_postcopy_blocktime_begin(
1080                     (uintptr_t)(msg.arg.pagefault.address),
1081                                 msg.arg.pagefault.feat.ptid, rb);
1082 
1083 retry:
1084             /*
1085              * Send the request to the source - we want to request one
1086              * of our host page sizes (which is >= TPS)
1087              */
1088             ret = postcopy_request_page(mis, rb, rb_offset,
1089                                         msg.arg.pagefault.address);
1090             if (ret) {
1091                 /* May be network failure, try to wait for recovery */
1092                 postcopy_pause_fault_thread(mis);
1093                 goto retry;
1094             }
1095         }
1096 
1097         /* Now handle any requests from external processes on shared memory */
1098         /* TODO: May need to handle devices deregistering during postcopy */
1099         for (index = 2; index < pfd_len && poll_result; index++) {
1100             if (pfd[index].revents) {
1101                 struct PostCopyFD *pcfd =
1102                     &g_array_index(mis->postcopy_remote_fds,
1103                                    struct PostCopyFD, index - 2);
1104 
1105                 poll_result--;
1106                 if (pfd[index].revents & POLLERR) {
1107                     error_report("%s: POLLERR on poll %zd fd=%d",
1108                                  __func__, index, pcfd->fd);
1109                     pfd[index].events = 0;
1110                     continue;
1111                 }
1112 
1113                 ret = read(pcfd->fd, &msg, sizeof(msg));
1114                 if (ret != sizeof(msg)) {
1115                     if (errno == EAGAIN) {
1116                         /*
1117                          * if a wake up happens on the other thread just after
1118                          * the poll, there is nothing to read.
1119                          */
1120                         continue;
1121                     }
1122                     if (ret < 0) {
1123                         error_report("%s: Failed to read full userfault "
1124                                      "message: %s (shared) revents=%d",
1125                                      __func__, strerror(errno),
1126                                      pfd[index].revents);
1127                         /*TODO: Could just disable this sharer */
1128                         break;
1129                     } else {
1130                         error_report("%s: Read %d bytes from userfaultfd "
1131                                      "expected %zd (shared)",
1132                                      __func__, ret, sizeof(msg));
1133                         /*TODO: Could just disable this sharer */
1134                         break; /*Lost alignment,don't know what we'd read next*/
1135                     }
1136                 }
1137                 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1138                     error_report("%s: Read unexpected event %ud "
1139                                  "from userfaultfd (shared)",
1140                                  __func__, msg.event);
1141                     continue; /* It's not a page fault, shouldn't happen */
1142                 }
1143                 /* Call the device handler registered with us */
1144                 ret = pcfd->handler(pcfd, &msg);
1145                 if (ret) {
1146                     error_report("%s: Failed to resolve shared fault on %zd/%s",
1147                                  __func__, index, pcfd->idstr);
1148                     /* TODO: Fail? Disable this sharer? */
1149                 }
1150             }
1151         }
1152     }
1153     rcu_unregister_thread();
1154     trace_postcopy_ram_fault_thread_exit();
1155     g_free(pfd);
1156     return NULL;
1157 }
1158 
1159 static int postcopy_temp_pages_setup(MigrationIncomingState *mis)
1160 {
1161     PostcopyTmpPage *tmp_page;
1162     int err, i, channels;
1163     void *temp_page;
1164 
1165     if (migrate_postcopy_preempt()) {
1166         /* If preemption enabled, need extra channel for urgent requests */
1167         mis->postcopy_channels = RAM_CHANNEL_MAX;
1168     } else {
1169         /* Both precopy/postcopy on the same channel */
1170         mis->postcopy_channels = 1;
1171     }
1172 
1173     channels = mis->postcopy_channels;
1174     mis->postcopy_tmp_pages = g_malloc0_n(sizeof(PostcopyTmpPage), channels);
1175 
1176     for (i = 0; i < channels; i++) {
1177         tmp_page = &mis->postcopy_tmp_pages[i];
1178         temp_page = mmap(NULL, mis->largest_page_size, PROT_READ | PROT_WRITE,
1179                          MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1180         if (temp_page == MAP_FAILED) {
1181             err = errno;
1182             error_report("%s: Failed to map postcopy_tmp_pages[%d]: %s",
1183                          __func__, i, strerror(err));
1184             /* Clean up will be done later */
1185             return -err;
1186         }
1187         tmp_page->tmp_huge_page = temp_page;
1188         /* Initialize default states for each tmp page */
1189         postcopy_temp_page_reset(tmp_page);
1190     }
1191 
1192     /*
1193      * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages
1194      */
1195     mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1196                                        PROT_READ | PROT_WRITE,
1197                                        MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
1198     if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1199         err = errno;
1200         mis->postcopy_tmp_zero_page = NULL;
1201         error_report("%s: Failed to map large zero page %s",
1202                      __func__, strerror(err));
1203         return -err;
1204     }
1205 
1206     memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1207 
1208     return 0;
1209 }
1210 
1211 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1212 {
1213     Error *local_err = NULL;
1214 
1215     /* Open the fd for the kernel to give us userfaults */
1216     mis->userfault_fd = uffd_open(O_CLOEXEC | O_NONBLOCK);
1217     if (mis->userfault_fd == -1) {
1218         error_report("%s: Failed to open userfault fd: %s", __func__,
1219                      strerror(errno));
1220         return -1;
1221     }
1222 
1223     /*
1224      * Although the host check already tested the API, we need to
1225      * do the check again as an ABI handshake on the new fd.
1226      */
1227     if (!ufd_check_and_apply(mis->userfault_fd, mis, &local_err)) {
1228         error_report_err(local_err);
1229         return -1;
1230     }
1231 
1232     /* Now an eventfd we use to tell the fault-thread to quit */
1233     mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1234     if (mis->userfault_event_fd == -1) {
1235         error_report("%s: Opening userfault_event_fd: %s", __func__,
1236                      strerror(errno));
1237         close(mis->userfault_fd);
1238         return -1;
1239     }
1240 
1241     postcopy_thread_create(mis, &mis->fault_thread, "mig/dst/fault",
1242                            postcopy_ram_fault_thread, QEMU_THREAD_JOINABLE);
1243     mis->have_fault_thread = true;
1244 
1245     /* Mark so that we get notified of accesses to unwritten areas */
1246     if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1247         error_report("ram_block_enable_notify failed");
1248         return -1;
1249     }
1250 
1251     if (postcopy_temp_pages_setup(mis)) {
1252         /* Error dumped in the sub-function */
1253         return -1;
1254     }
1255 
1256     if (migrate_postcopy_preempt()) {
1257         /*
1258          * This thread needs to be created after the temp pages because
1259          * it'll fetch RAM_CHANNEL_POSTCOPY PostcopyTmpPage immediately.
1260          */
1261         postcopy_thread_create(mis, &mis->postcopy_prio_thread, "mig/dst/preempt",
1262                                postcopy_preempt_thread, QEMU_THREAD_JOINABLE);
1263         mis->preempt_thread_status = PREEMPT_THREAD_CREATED;
1264     }
1265 
1266     trace_postcopy_ram_enable_notify();
1267 
1268     return 0;
1269 }
1270 
1271 static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr,
1272                                void *from_addr, uint64_t pagesize, RAMBlock *rb)
1273 {
1274     int userfault_fd = mis->userfault_fd;
1275     int ret;
1276 
1277     if (from_addr) {
1278         struct uffdio_copy copy_struct;
1279         copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1280         copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1281         copy_struct.len = pagesize;
1282         copy_struct.mode = 0;
1283         ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1284     } else {
1285         struct uffdio_zeropage zero_struct;
1286         zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1287         zero_struct.range.len = pagesize;
1288         zero_struct.mode = 0;
1289         ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1290     }
1291     if (!ret) {
1292         qemu_mutex_lock(&mis->page_request_mutex);
1293         ramblock_recv_bitmap_set_range(rb, host_addr,
1294                                        pagesize / qemu_target_page_size());
1295         /*
1296          * If this page resolves a page fault for a previous recorded faulted
1297          * address, take a special note to maintain the requested page list.
1298          */
1299         if (g_tree_lookup(mis->page_requested, host_addr)) {
1300             g_tree_remove(mis->page_requested, host_addr);
1301             int left_pages = qatomic_dec_fetch(&mis->page_requested_count);
1302 
1303             trace_postcopy_page_req_del(host_addr, mis->page_requested_count);
1304             /* Order the update of count and read of preempt status */
1305             smp_mb();
1306             if (mis->preempt_thread_status == PREEMPT_THREAD_QUIT &&
1307                 left_pages == 0) {
1308                 /*
1309                  * This probably means the main thread is waiting for us.
1310                  * Notify that we've finished receiving the last requested
1311                  * page.
1312                  */
1313                 qemu_cond_signal(&mis->page_request_cond);
1314             }
1315         }
1316         qemu_mutex_unlock(&mis->page_request_mutex);
1317         mark_postcopy_blocktime_end((uintptr_t)host_addr);
1318     }
1319     return ret;
1320 }
1321 
1322 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1323 {
1324     int i;
1325     MigrationIncomingState *mis = migration_incoming_get_current();
1326     GArray *pcrfds = mis->postcopy_remote_fds;
1327 
1328     for (i = 0; i < pcrfds->len; i++) {
1329         struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1330         int ret = cur->waker(cur, rb, offset);
1331         if (ret) {
1332             return ret;
1333         }
1334     }
1335     return 0;
1336 }
1337 
1338 /*
1339  * Place a host page (from) at (host) atomically
1340  * returns 0 on success
1341  */
1342 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1343                         RAMBlock *rb)
1344 {
1345     size_t pagesize = qemu_ram_pagesize(rb);
1346 
1347     /* copy also acks to the kernel waking the stalled thread up
1348      * TODO: We can inhibit that ack and only do it if it was requested
1349      * which would be slightly cheaper, but we'd have to be careful
1350      * of the order of updating our page state.
1351      */
1352     if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) {
1353         int e = errno;
1354         error_report("%s: %s copy host: %p from: %p (size: %zd)",
1355                      __func__, strerror(e), host, from, pagesize);
1356 
1357         return -e;
1358     }
1359 
1360     trace_postcopy_place_page(host);
1361     return postcopy_notify_shared_wake(rb,
1362                                        qemu_ram_block_host_offset(rb, host));
1363 }
1364 
1365 /*
1366  * Place a zero page at (host) atomically
1367  * returns 0 on success
1368  */
1369 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1370                              RAMBlock *rb)
1371 {
1372     size_t pagesize = qemu_ram_pagesize(rb);
1373     trace_postcopy_place_page_zero(host);
1374 
1375     /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1376      * but it's not available for everything (e.g. hugetlbpages)
1377      */
1378     if (qemu_ram_is_uf_zeroable(rb)) {
1379         if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) {
1380             int e = errno;
1381             error_report("%s: %s zero host: %p",
1382                          __func__, strerror(e), host);
1383 
1384             return -e;
1385         }
1386         return postcopy_notify_shared_wake(rb,
1387                                            qemu_ram_block_host_offset(rb,
1388                                                                       host));
1389     } else {
1390         return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb);
1391     }
1392 }
1393 
1394 #else
1395 /* No target OS support, stubs just fail */
1396 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1397 {
1398 }
1399 
1400 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis, Error **errp)
1401 {
1402     error_report("%s: No OS support", __func__);
1403     return false;
1404 }
1405 
1406 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1407 {
1408     error_report("postcopy_ram_incoming_init: No OS support");
1409     return -1;
1410 }
1411 
1412 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1413 {
1414     assert(0);
1415     return -1;
1416 }
1417 
1418 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1419 {
1420     assert(0);
1421     return -1;
1422 }
1423 
1424 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1425                                  uint64_t client_addr, uint64_t rb_offset)
1426 {
1427     assert(0);
1428     return -1;
1429 }
1430 
1431 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1432 {
1433     assert(0);
1434     return -1;
1435 }
1436 
1437 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1438                         RAMBlock *rb)
1439 {
1440     assert(0);
1441     return -1;
1442 }
1443 
1444 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1445                         RAMBlock *rb)
1446 {
1447     assert(0);
1448     return -1;
1449 }
1450 
1451 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1452                          uint64_t client_addr,
1453                          RAMBlock *rb)
1454 {
1455     assert(0);
1456     return -1;
1457 }
1458 #endif
1459 
1460 /* ------------------------------------------------------------------------- */
1461 void postcopy_temp_page_reset(PostcopyTmpPage *tmp_page)
1462 {
1463     tmp_page->target_pages = 0;
1464     tmp_page->host_addr = NULL;
1465     /*
1466      * This is set to true when reset, and cleared as long as we received any
1467      * of the non-zero small page within this huge page.
1468      */
1469     tmp_page->all_zero = true;
1470 }
1471 
1472 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1473 {
1474     uint64_t tmp64 = 1;
1475 
1476     /*
1477      * Wakeup the fault_thread.  It's an eventfd that should currently
1478      * be at 0, we're going to increment it to 1
1479      */
1480     if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1481         /* Not much we can do here, but may as well report it */
1482         error_report("%s: incrementing failed: %s", __func__,
1483                      strerror(errno));
1484     }
1485 }
1486 
1487 /**
1488  * postcopy_discard_send_init: Called at the start of each RAMBlock before
1489  *   asking to discard individual ranges.
1490  *
1491  * @ms: The current migration state.
1492  * @offset: the bitmap offset of the named RAMBlock in the migration bitmap.
1493  * @name: RAMBlock that discards will operate on.
1494  */
1495 static PostcopyDiscardState pds = {0};
1496 void postcopy_discard_send_init(MigrationState *ms, const char *name)
1497 {
1498     pds.ramblock_name = name;
1499     pds.cur_entry = 0;
1500     pds.nsentwords = 0;
1501     pds.nsentcmds = 0;
1502 }
1503 
1504 /**
1505  * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1506  *   discard. May send a discard message, may just leave it queued to
1507  *   be sent later.
1508  *
1509  * @ms: Current migration state.
1510  * @start,@length: a range of pages in the migration bitmap in the
1511  *   RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1512  */
1513 void postcopy_discard_send_range(MigrationState *ms, unsigned long start,
1514                                  unsigned long length)
1515 {
1516     size_t tp_size = qemu_target_page_size();
1517     /* Convert to byte offsets within the RAM block */
1518     pds.start_list[pds.cur_entry] = start  * tp_size;
1519     pds.length_list[pds.cur_entry] = length * tp_size;
1520     trace_postcopy_discard_send_range(pds.ramblock_name, start, length);
1521     pds.cur_entry++;
1522     pds.nsentwords++;
1523 
1524     if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) {
1525         /* Full set, ship it! */
1526         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1527                                               pds.ramblock_name,
1528                                               pds.cur_entry,
1529                                               pds.start_list,
1530                                               pds.length_list);
1531         pds.nsentcmds++;
1532         pds.cur_entry = 0;
1533     }
1534 }
1535 
1536 /**
1537  * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1538  * bitmap code. Sends any outstanding discard messages, frees the PDS
1539  *
1540  * @ms: Current migration state.
1541  */
1542 void postcopy_discard_send_finish(MigrationState *ms)
1543 {
1544     /* Anything unsent? */
1545     if (pds.cur_entry) {
1546         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1547                                               pds.ramblock_name,
1548                                               pds.cur_entry,
1549                                               pds.start_list,
1550                                               pds.length_list);
1551         pds.nsentcmds++;
1552     }
1553 
1554     trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords,
1555                                        pds.nsentcmds);
1556 }
1557 
1558 /*
1559  * Current state of incoming postcopy; note this is not part of
1560  * MigrationIncomingState since it's state is used during cleanup
1561  * at the end as MIS is being freed.
1562  */
1563 static PostcopyState incoming_postcopy_state;
1564 
1565 PostcopyState  postcopy_state_get(void)
1566 {
1567     return qatomic_load_acquire(&incoming_postcopy_state);
1568 }
1569 
1570 /* Set the state and return the old state */
1571 PostcopyState postcopy_state_set(PostcopyState new_state)
1572 {
1573     return qatomic_xchg(&incoming_postcopy_state, new_state);
1574 }
1575 
1576 /* Register a handler for external shared memory postcopy
1577  * called on the destination.
1578  */
1579 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1580 {
1581     MigrationIncomingState *mis = migration_incoming_get_current();
1582 
1583     mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1584                                                   *pcfd);
1585 }
1586 
1587 /* Unregister a handler for external shared memory postcopy
1588  */
1589 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1590 {
1591     guint i;
1592     MigrationIncomingState *mis = migration_incoming_get_current();
1593     GArray *pcrfds = mis->postcopy_remote_fds;
1594 
1595     if (!pcrfds) {
1596         /* migration has already finished and freed the array */
1597         return;
1598     }
1599     for (i = 0; i < pcrfds->len; i++) {
1600         struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1601         if (cur->fd == pcfd->fd) {
1602             mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1603             return;
1604         }
1605     }
1606 }
1607 
1608 void postcopy_preempt_new_channel(MigrationIncomingState *mis, QEMUFile *file)
1609 {
1610     /*
1611      * The new loading channel has its own threads, so it needs to be
1612      * blocked too.  It's by default true, just be explicit.
1613      */
1614     qemu_file_set_blocking(file, true);
1615     mis->postcopy_qemufile_dst = file;
1616     qemu_sem_post(&mis->postcopy_qemufile_dst_done);
1617     trace_postcopy_preempt_new_channel();
1618 }
1619 
1620 /*
1621  * Setup the postcopy preempt channel with the IOC.  If ERROR is specified,
1622  * setup the error instead.  This helper will free the ERROR if specified.
1623  */
1624 static void
1625 postcopy_preempt_send_channel_done(MigrationState *s,
1626                                    QIOChannel *ioc, Error *local_err)
1627 {
1628     if (local_err) {
1629         migrate_set_error(s, local_err);
1630         error_free(local_err);
1631     } else {
1632         migration_ioc_register_yank(ioc);
1633         s->postcopy_qemufile_src = qemu_file_new_output(ioc);
1634         trace_postcopy_preempt_new_channel();
1635     }
1636 
1637     /*
1638      * Kick the waiter in all cases.  The waiter should check upon
1639      * postcopy_qemufile_src to know whether it failed or not.
1640      */
1641     qemu_sem_post(&s->postcopy_qemufile_src_sem);
1642 }
1643 
1644 static void
1645 postcopy_preempt_tls_handshake(QIOTask *task, gpointer opaque)
1646 {
1647     g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task));
1648     MigrationState *s = opaque;
1649     Error *local_err = NULL;
1650 
1651     qio_task_propagate_error(task, &local_err);
1652     postcopy_preempt_send_channel_done(s, ioc, local_err);
1653 }
1654 
1655 static void
1656 postcopy_preempt_send_channel_new(QIOTask *task, gpointer opaque)
1657 {
1658     g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task));
1659     MigrationState *s = opaque;
1660     QIOChannelTLS *tioc;
1661     Error *local_err = NULL;
1662 
1663     if (qio_task_propagate_error(task, &local_err)) {
1664         goto out;
1665     }
1666 
1667     if (migrate_channel_requires_tls_upgrade(ioc)) {
1668         tioc = migration_tls_client_create(ioc, s->hostname, &local_err);
1669         if (!tioc) {
1670             goto out;
1671         }
1672         trace_postcopy_preempt_tls_handshake();
1673         qio_channel_set_name(QIO_CHANNEL(tioc), "migration-tls-preempt");
1674         qio_channel_tls_handshake(tioc, postcopy_preempt_tls_handshake,
1675                                   s, NULL, NULL);
1676         /* Setup the channel until TLS handshake finished */
1677         return;
1678     }
1679 
1680 out:
1681     /* This handles both good and error cases */
1682     postcopy_preempt_send_channel_done(s, ioc, local_err);
1683 }
1684 
1685 /*
1686  * This function will kick off an async task to establish the preempt
1687  * channel, and wait until the connection setup completed.  Returns 0 if
1688  * channel established, -1 for error.
1689  */
1690 int postcopy_preempt_establish_channel(MigrationState *s)
1691 {
1692     /* If preempt not enabled, no need to wait */
1693     if (!migrate_postcopy_preempt()) {
1694         return 0;
1695     }
1696 
1697     /*
1698      * Kick off async task to establish preempt channel.  Only do so with
1699      * 8.0+ machines, because 7.1/7.2 require the channel to be created in
1700      * setup phase of migration (even if racy in an unreliable network).
1701      */
1702     if (!s->preempt_pre_7_2) {
1703         postcopy_preempt_setup(s);
1704     }
1705 
1706     /*
1707      * We need the postcopy preempt channel to be established before
1708      * starting doing anything.
1709      */
1710     qemu_sem_wait(&s->postcopy_qemufile_src_sem);
1711 
1712     return s->postcopy_qemufile_src ? 0 : -1;
1713 }
1714 
1715 void postcopy_preempt_setup(MigrationState *s)
1716 {
1717     /* Kick an async task to connect */
1718     socket_send_channel_create(postcopy_preempt_send_channel_new, s);
1719 }
1720 
1721 static void postcopy_pause_ram_fast_load(MigrationIncomingState *mis)
1722 {
1723     trace_postcopy_pause_fast_load();
1724     qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex);
1725     qemu_sem_wait(&mis->postcopy_pause_sem_fast_load);
1726     qemu_mutex_lock(&mis->postcopy_prio_thread_mutex);
1727     trace_postcopy_pause_fast_load_continued();
1728 }
1729 
1730 static bool preempt_thread_should_run(MigrationIncomingState *mis)
1731 {
1732     return mis->preempt_thread_status != PREEMPT_THREAD_QUIT;
1733 }
1734 
1735 void *postcopy_preempt_thread(void *opaque)
1736 {
1737     MigrationIncomingState *mis = opaque;
1738     int ret;
1739 
1740     trace_postcopy_preempt_thread_entry();
1741 
1742     rcu_register_thread();
1743 
1744     qemu_sem_post(&mis->thread_sync_sem);
1745 
1746     /*
1747      * The preempt channel is established in asynchronous way.  Wait
1748      * for its completion.
1749      */
1750     qemu_sem_wait(&mis->postcopy_qemufile_dst_done);
1751 
1752     /* Sending RAM_SAVE_FLAG_EOS to terminate this thread */
1753     qemu_mutex_lock(&mis->postcopy_prio_thread_mutex);
1754     while (preempt_thread_should_run(mis)) {
1755         ret = ram_load_postcopy(mis->postcopy_qemufile_dst,
1756                                 RAM_CHANNEL_POSTCOPY);
1757         /* If error happened, go into recovery routine */
1758         if (ret && preempt_thread_should_run(mis)) {
1759             postcopy_pause_ram_fast_load(mis);
1760         } else {
1761             /* We're done */
1762             break;
1763         }
1764     }
1765     qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex);
1766 
1767     rcu_unregister_thread();
1768 
1769     trace_postcopy_preempt_thread_exit();
1770 
1771     return NULL;
1772 }
1773 
1774 bool postcopy_is_paused(MigrationStatus status)
1775 {
1776     return status == MIGRATION_STATUS_POSTCOPY_PAUSED ||
1777         status == MIGRATION_STATUS_POSTCOPY_RECOVER_SETUP;
1778 }
1779