xref: /openbmc/qemu/migration/postcopy-ram.c (revision b519e2e982f7f5368c5976ca44d9ca7bbb7a3378)
1 /*
2  * Postcopy migration for RAM
3  *
4  * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates
5  *
6  * Authors:
7  *  Dave Gilbert  <dgilbert@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
10  * See the COPYING file in the top-level directory.
11  *
12  */
13 
14 /*
15  * Postcopy is a migration technique where the execution flips from the
16  * source to the destination before all the data has been copied.
17  */
18 
19 #include "qemu/osdep.h"
20 #include "exec/target_page.h"
21 #include "migration.h"
22 #include "qemu-file.h"
23 #include "savevm.h"
24 #include "postcopy-ram.h"
25 #include "ram.h"
26 #include "qapi/error.h"
27 #include "qemu/notify.h"
28 #include "qemu/rcu.h"
29 #include "sysemu/sysemu.h"
30 #include "qemu/error-report.h"
31 #include "trace.h"
32 #include "hw/boards.h"
33 
34 /* Arbitrary limit on size of each discard command,
35  * keeps them around ~200 bytes
36  */
37 #define MAX_DISCARDS_PER_COMMAND 12
38 
39 struct PostcopyDiscardState {
40     const char *ramblock_name;
41     uint16_t cur_entry;
42     /*
43      * Start and length of a discard range (bytes)
44      */
45     uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
46     uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
47     unsigned int nsentwords;
48     unsigned int nsentcmds;
49 };
50 
51 static NotifierWithReturnList postcopy_notifier_list;
52 
53 void postcopy_infrastructure_init(void)
54 {
55     notifier_with_return_list_init(&postcopy_notifier_list);
56 }
57 
58 void postcopy_add_notifier(NotifierWithReturn *nn)
59 {
60     notifier_with_return_list_add(&postcopy_notifier_list, nn);
61 }
62 
63 void postcopy_remove_notifier(NotifierWithReturn *n)
64 {
65     notifier_with_return_remove(n);
66 }
67 
68 int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp)
69 {
70     struct PostcopyNotifyData pnd;
71     pnd.reason = reason;
72     pnd.errp = errp;
73 
74     return notifier_with_return_list_notify(&postcopy_notifier_list,
75                                             &pnd);
76 }
77 
78 /* Postcopy needs to detect accesses to pages that haven't yet been copied
79  * across, and efficiently map new pages in, the techniques for doing this
80  * are target OS specific.
81  */
82 #if defined(__linux__)
83 
84 #include <poll.h>
85 #include <sys/ioctl.h>
86 #include <sys/syscall.h>
87 #include <asm/types.h> /* for __u64 */
88 #endif
89 
90 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
91 #include <sys/eventfd.h>
92 #include <linux/userfaultfd.h>
93 
94 typedef struct PostcopyBlocktimeContext {
95     /* time when page fault initiated per vCPU */
96     uint32_t *page_fault_vcpu_time;
97     /* page address per vCPU */
98     uintptr_t *vcpu_addr;
99     uint32_t total_blocktime;
100     /* blocktime per vCPU */
101     uint32_t *vcpu_blocktime;
102     /* point in time when last page fault was initiated */
103     uint32_t last_begin;
104     /* number of vCPU are suspended */
105     int smp_cpus_down;
106     uint64_t start_time;
107 
108     /*
109      * Handler for exit event, necessary for
110      * releasing whole blocktime_ctx
111      */
112     Notifier exit_notifier;
113 } PostcopyBlocktimeContext;
114 
115 static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx)
116 {
117     g_free(ctx->page_fault_vcpu_time);
118     g_free(ctx->vcpu_addr);
119     g_free(ctx->vcpu_blocktime);
120     g_free(ctx);
121 }
122 
123 static void migration_exit_cb(Notifier *n, void *data)
124 {
125     PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext,
126                                                  exit_notifier);
127     destroy_blocktime_context(ctx);
128 }
129 
130 static struct PostcopyBlocktimeContext *blocktime_context_new(void)
131 {
132     MachineState *ms = MACHINE(qdev_get_machine());
133     unsigned int smp_cpus = ms->smp.cpus;
134     PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1);
135     ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus);
136     ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus);
137     ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus);
138 
139     ctx->exit_notifier.notify = migration_exit_cb;
140     ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
141     qemu_add_exit_notifier(&ctx->exit_notifier);
142     return ctx;
143 }
144 
145 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx)
146 {
147     MachineState *ms = MACHINE(qdev_get_machine());
148     uint32List *list = NULL, *entry = NULL;
149     int i;
150 
151     for (i = ms->smp.cpus - 1; i >= 0; i--) {
152         entry = g_new0(uint32List, 1);
153         entry->value = ctx->vcpu_blocktime[i];
154         entry->next = list;
155         list = entry;
156     }
157 
158     return list;
159 }
160 
161 /*
162  * This function just populates MigrationInfo from postcopy's
163  * blocktime context. It will not populate MigrationInfo,
164  * unless postcopy-blocktime capability was set.
165  *
166  * @info: pointer to MigrationInfo to populate
167  */
168 void fill_destination_postcopy_migration_info(MigrationInfo *info)
169 {
170     MigrationIncomingState *mis = migration_incoming_get_current();
171     PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
172 
173     if (!bc) {
174         return;
175     }
176 
177     info->has_postcopy_blocktime = true;
178     info->postcopy_blocktime = bc->total_blocktime;
179     info->has_postcopy_vcpu_blocktime = true;
180     info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc);
181 }
182 
183 static uint32_t get_postcopy_total_blocktime(void)
184 {
185     MigrationIncomingState *mis = migration_incoming_get_current();
186     PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
187 
188     if (!bc) {
189         return 0;
190     }
191 
192     return bc->total_blocktime;
193 }
194 
195 /**
196  * receive_ufd_features: check userfault fd features, to request only supported
197  * features in the future.
198  *
199  * Returns: true on success
200  *
201  * __NR_userfaultfd - should be checked before
202  *  @features: out parameter will contain uffdio_api.features provided by kernel
203  *              in case of success
204  */
205 static bool receive_ufd_features(uint64_t *features)
206 {
207     struct uffdio_api api_struct = {0};
208     int ufd;
209     bool ret = true;
210 
211     /* if we are here __NR_userfaultfd should exists */
212     ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
213     if (ufd == -1) {
214         error_report("%s: syscall __NR_userfaultfd failed: %s", __func__,
215                      strerror(errno));
216         return false;
217     }
218 
219     /* ask features */
220     api_struct.api = UFFD_API;
221     api_struct.features = 0;
222     if (ioctl(ufd, UFFDIO_API, &api_struct)) {
223         error_report("%s: UFFDIO_API failed: %s", __func__,
224                      strerror(errno));
225         ret = false;
226         goto release_ufd;
227     }
228 
229     *features = api_struct.features;
230 
231 release_ufd:
232     close(ufd);
233     return ret;
234 }
235 
236 /**
237  * request_ufd_features: this function should be called only once on a newly
238  * opened ufd, subsequent calls will lead to error.
239  *
240  * Returns: true on success
241  *
242  * @ufd: fd obtained from userfaultfd syscall
243  * @features: bit mask see UFFD_API_FEATURES
244  */
245 static bool request_ufd_features(int ufd, uint64_t features)
246 {
247     struct uffdio_api api_struct = {0};
248     uint64_t ioctl_mask;
249 
250     api_struct.api = UFFD_API;
251     api_struct.features = features;
252     if (ioctl(ufd, UFFDIO_API, &api_struct)) {
253         error_report("%s failed: UFFDIO_API failed: %s", __func__,
254                      strerror(errno));
255         return false;
256     }
257 
258     ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
259                  (__u64)1 << _UFFDIO_UNREGISTER;
260     if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
261         error_report("Missing userfault features: %" PRIx64,
262                      (uint64_t)(~api_struct.ioctls & ioctl_mask));
263         return false;
264     }
265 
266     return true;
267 }
268 
269 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis)
270 {
271     uint64_t asked_features = 0;
272     static uint64_t supported_features;
273 
274     /*
275      * it's not possible to
276      * request UFFD_API twice per one fd
277      * userfault fd features is persistent
278      */
279     if (!supported_features) {
280         if (!receive_ufd_features(&supported_features)) {
281             error_report("%s failed", __func__);
282             return false;
283         }
284     }
285 
286 #ifdef UFFD_FEATURE_THREAD_ID
287     if (migrate_postcopy_blocktime() && mis &&
288         UFFD_FEATURE_THREAD_ID & supported_features) {
289         /* kernel supports that feature */
290         /* don't create blocktime_context if it exists */
291         if (!mis->blocktime_ctx) {
292             mis->blocktime_ctx = blocktime_context_new();
293         }
294 
295         asked_features |= UFFD_FEATURE_THREAD_ID;
296     }
297 #endif
298 
299     /*
300      * request features, even if asked_features is 0, due to
301      * kernel expects UFFD_API before UFFDIO_REGISTER, per
302      * userfault file descriptor
303      */
304     if (!request_ufd_features(ufd, asked_features)) {
305         error_report("%s failed: features %" PRIu64, __func__,
306                      asked_features);
307         return false;
308     }
309 
310     if (qemu_real_host_page_size != ram_pagesize_summary()) {
311         bool have_hp = false;
312         /* We've got a huge page */
313 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
314         have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS;
315 #endif
316         if (!have_hp) {
317             error_report("Userfault on this host does not support huge pages");
318             return false;
319         }
320     }
321     return true;
322 }
323 
324 /* Callback from postcopy_ram_supported_by_host block iterator.
325  */
326 static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque)
327 {
328     const char *block_name = qemu_ram_get_idstr(rb);
329     ram_addr_t length = qemu_ram_get_used_length(rb);
330     size_t pagesize = qemu_ram_pagesize(rb);
331 
332     if (length % pagesize) {
333         error_report("Postcopy requires RAM blocks to be a page size multiple,"
334                      " block %s is 0x" RAM_ADDR_FMT " bytes with a "
335                      "page size of 0x%zx", block_name, length, pagesize);
336         return 1;
337     }
338     return 0;
339 }
340 
341 /*
342  * Note: This has the side effect of munlock'ing all of RAM, that's
343  * normally fine since if the postcopy succeeds it gets turned back on at the
344  * end.
345  */
346 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
347 {
348     long pagesize = qemu_real_host_page_size;
349     int ufd = -1;
350     bool ret = false; /* Error unless we change it */
351     void *testarea = NULL;
352     struct uffdio_register reg_struct;
353     struct uffdio_range range_struct;
354     uint64_t feature_mask;
355     Error *local_err = NULL;
356 
357     if (qemu_target_page_size() > pagesize) {
358         error_report("Target page size bigger than host page size");
359         goto out;
360     }
361 
362     ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
363     if (ufd == -1) {
364         error_report("%s: userfaultfd not available: %s", __func__,
365                      strerror(errno));
366         goto out;
367     }
368 
369     /* Give devices a chance to object */
370     if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) {
371         error_report_err(local_err);
372         goto out;
373     }
374 
375     /* Version and features check */
376     if (!ufd_check_and_apply(ufd, mis)) {
377         goto out;
378     }
379 
380     /* We don't support postcopy with shared RAM yet */
381     if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) {
382         goto out;
383     }
384 
385     /*
386      * userfault and mlock don't go together; we'll put it back later if
387      * it was enabled.
388      */
389     if (munlockall()) {
390         error_report("%s: munlockall: %s", __func__,  strerror(errno));
391         goto out;
392     }
393 
394     /*
395      *  We need to check that the ops we need are supported on anon memory
396      *  To do that we need to register a chunk and see the flags that
397      *  are returned.
398      */
399     testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
400                                     MAP_ANONYMOUS, -1, 0);
401     if (testarea == MAP_FAILED) {
402         error_report("%s: Failed to map test area: %s", __func__,
403                      strerror(errno));
404         goto out;
405     }
406     g_assert(((size_t)testarea & (pagesize - 1)) == 0);
407 
408     reg_struct.range.start = (uintptr_t)testarea;
409     reg_struct.range.len = pagesize;
410     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
411 
412     if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
413         error_report("%s userfault register: %s", __func__, strerror(errno));
414         goto out;
415     }
416 
417     range_struct.start = (uintptr_t)testarea;
418     range_struct.len = pagesize;
419     if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
420         error_report("%s userfault unregister: %s", __func__, strerror(errno));
421         goto out;
422     }
423 
424     feature_mask = (__u64)1 << _UFFDIO_WAKE |
425                    (__u64)1 << _UFFDIO_COPY |
426                    (__u64)1 << _UFFDIO_ZEROPAGE;
427     if ((reg_struct.ioctls & feature_mask) != feature_mask) {
428         error_report("Missing userfault map features: %" PRIx64,
429                      (uint64_t)(~reg_struct.ioctls & feature_mask));
430         goto out;
431     }
432 
433     /* Success! */
434     ret = true;
435 out:
436     if (testarea) {
437         munmap(testarea, pagesize);
438     }
439     if (ufd != -1) {
440         close(ufd);
441     }
442     return ret;
443 }
444 
445 /*
446  * Setup an area of RAM so that it *can* be used for postcopy later; this
447  * must be done right at the start prior to pre-copy.
448  * opaque should be the MIS.
449  */
450 static int init_range(RAMBlock *rb, void *opaque)
451 {
452     const char *block_name = qemu_ram_get_idstr(rb);
453     void *host_addr = qemu_ram_get_host_addr(rb);
454     ram_addr_t offset = qemu_ram_get_offset(rb);
455     ram_addr_t length = qemu_ram_get_used_length(rb);
456     trace_postcopy_init_range(block_name, host_addr, offset, length);
457 
458     /*
459      * We need the whole of RAM to be truly empty for postcopy, so things
460      * like ROMs and any data tables built during init must be zero'd
461      * - we're going to get the copy from the source anyway.
462      * (Precopy will just overwrite this data, so doesn't need the discard)
463      */
464     if (ram_discard_range(block_name, 0, length)) {
465         return -1;
466     }
467 
468     return 0;
469 }
470 
471 /*
472  * At the end of migration, undo the effects of init_range
473  * opaque should be the MIS.
474  */
475 static int cleanup_range(RAMBlock *rb, void *opaque)
476 {
477     const char *block_name = qemu_ram_get_idstr(rb);
478     void *host_addr = qemu_ram_get_host_addr(rb);
479     ram_addr_t offset = qemu_ram_get_offset(rb);
480     ram_addr_t length = qemu_ram_get_used_length(rb);
481     MigrationIncomingState *mis = opaque;
482     struct uffdio_range range_struct;
483     trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
484 
485     /*
486      * We turned off hugepage for the precopy stage with postcopy enabled
487      * we can turn it back on now.
488      */
489     qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
490 
491     /*
492      * We can also turn off userfault now since we should have all the
493      * pages.   It can be useful to leave it on to debug postcopy
494      * if you're not sure it's always getting every page.
495      */
496     range_struct.start = (uintptr_t)host_addr;
497     range_struct.len = length;
498 
499     if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
500         error_report("%s: userfault unregister %s", __func__, strerror(errno));
501 
502         return -1;
503     }
504 
505     return 0;
506 }
507 
508 /*
509  * Initialise postcopy-ram, setting the RAM to a state where we can go into
510  * postcopy later; must be called prior to any precopy.
511  * called from arch_init's similarly named ram_postcopy_incoming_init
512  */
513 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
514 {
515     if (foreach_not_ignored_block(init_range, NULL)) {
516         return -1;
517     }
518 
519     return 0;
520 }
521 
522 /*
523  * At the end of a migration where postcopy_ram_incoming_init was called.
524  */
525 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
526 {
527     trace_postcopy_ram_incoming_cleanup_entry();
528 
529     if (mis->have_fault_thread) {
530         Error *local_err = NULL;
531 
532         /* Let the fault thread quit */
533         qatomic_set(&mis->fault_thread_quit, 1);
534         postcopy_fault_thread_notify(mis);
535         trace_postcopy_ram_incoming_cleanup_join();
536         qemu_thread_join(&mis->fault_thread);
537 
538         if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
539             error_report_err(local_err);
540             return -1;
541         }
542 
543         if (foreach_not_ignored_block(cleanup_range, mis)) {
544             return -1;
545         }
546 
547         trace_postcopy_ram_incoming_cleanup_closeuf();
548         close(mis->userfault_fd);
549         close(mis->userfault_event_fd);
550         mis->have_fault_thread = false;
551     }
552 
553     if (enable_mlock) {
554         if (os_mlock() < 0) {
555             error_report("mlock: %s", strerror(errno));
556             /*
557              * It doesn't feel right to fail at this point, we have a valid
558              * VM state.
559              */
560         }
561     }
562 
563     if (mis->postcopy_tmp_page) {
564         munmap(mis->postcopy_tmp_page, mis->largest_page_size);
565         mis->postcopy_tmp_page = NULL;
566     }
567     if (mis->postcopy_tmp_zero_page) {
568         munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
569         mis->postcopy_tmp_zero_page = NULL;
570     }
571     trace_postcopy_ram_incoming_cleanup_blocktime(
572             get_postcopy_total_blocktime());
573 
574     trace_postcopy_ram_incoming_cleanup_exit();
575     return 0;
576 }
577 
578 /*
579  * Disable huge pages on an area
580  */
581 static int nhp_range(RAMBlock *rb, void *opaque)
582 {
583     const char *block_name = qemu_ram_get_idstr(rb);
584     void *host_addr = qemu_ram_get_host_addr(rb);
585     ram_addr_t offset = qemu_ram_get_offset(rb);
586     ram_addr_t length = qemu_ram_get_used_length(rb);
587     trace_postcopy_nhp_range(block_name, host_addr, offset, length);
588 
589     /*
590      * Before we do discards we need to ensure those discards really
591      * do delete areas of the page, even if THP thinks a hugepage would
592      * be a good idea, so force hugepages off.
593      */
594     qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
595 
596     return 0;
597 }
598 
599 /*
600  * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
601  * however leaving it until after precopy means that most of the precopy
602  * data is still THPd
603  */
604 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
605 {
606     if (foreach_not_ignored_block(nhp_range, mis)) {
607         return -1;
608     }
609 
610     postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
611 
612     return 0;
613 }
614 
615 /*
616  * Mark the given area of RAM as requiring notification to unwritten areas
617  * Used as a  callback on foreach_not_ignored_block.
618  *   host_addr: Base of area to mark
619  *   offset: Offset in the whole ram arena
620  *   length: Length of the section
621  *   opaque: MigrationIncomingState pointer
622  * Returns 0 on success
623  */
624 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
625 {
626     MigrationIncomingState *mis = opaque;
627     struct uffdio_register reg_struct;
628 
629     reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
630     reg_struct.range.len = qemu_ram_get_used_length(rb);
631     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
632 
633     /* Now tell our userfault_fd that it's responsible for this area */
634     if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
635         error_report("%s userfault register: %s", __func__, strerror(errno));
636         return -1;
637     }
638     if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
639         error_report("%s userfault: Region doesn't support COPY", __func__);
640         return -1;
641     }
642     if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) {
643         qemu_ram_set_uf_zeroable(rb);
644     }
645 
646     return 0;
647 }
648 
649 int postcopy_wake_shared(struct PostCopyFD *pcfd,
650                          uint64_t client_addr,
651                          RAMBlock *rb)
652 {
653     size_t pagesize = qemu_ram_pagesize(rb);
654     struct uffdio_range range;
655     int ret;
656     trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
657     range.start = client_addr & ~(pagesize - 1);
658     range.len = pagesize;
659     ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
660     if (ret) {
661         error_report("%s: Failed to wake: %zx in %s (%s)",
662                      __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
663                      strerror(errno));
664     }
665     return ret;
666 }
667 
668 /*
669  * Callback from shared fault handlers to ask for a page,
670  * the page must be specified by a RAMBlock and an offset in that rb
671  * Note: Only for use by shared fault handlers (in fault thread)
672  */
673 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
674                                  uint64_t client_addr, uint64_t rb_offset)
675 {
676     size_t pagesize = qemu_ram_pagesize(rb);
677     uint64_t aligned_rbo = rb_offset & ~(pagesize - 1);
678     MigrationIncomingState *mis = migration_incoming_get_current();
679 
680     trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
681                                        rb_offset);
682     if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
683         trace_postcopy_request_shared_page_present(pcfd->idstr,
684                                         qemu_ram_get_idstr(rb), rb_offset);
685         return postcopy_wake_shared(pcfd, client_addr, rb);
686     }
687     migrate_send_rp_req_pages(mis, rb, aligned_rbo, client_addr);
688     return 0;
689 }
690 
691 static int get_mem_fault_cpu_index(uint32_t pid)
692 {
693     CPUState *cpu_iter;
694 
695     CPU_FOREACH(cpu_iter) {
696         if (cpu_iter->thread_id == pid) {
697             trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
698             return cpu_iter->cpu_index;
699         }
700     }
701     trace_get_mem_fault_cpu_index(-1, pid);
702     return -1;
703 }
704 
705 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
706 {
707     int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
708                                     dc->start_time;
709     return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
710 }
711 
712 /*
713  * This function is being called when pagefault occurs. It
714  * tracks down vCPU blocking time.
715  *
716  * @addr: faulted host virtual address
717  * @ptid: faulted process thread id
718  * @rb: ramblock appropriate to addr
719  */
720 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
721                                           RAMBlock *rb)
722 {
723     int cpu, already_received;
724     MigrationIncomingState *mis = migration_incoming_get_current();
725     PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
726     uint32_t low_time_offset;
727 
728     if (!dc || ptid == 0) {
729         return;
730     }
731     cpu = get_mem_fault_cpu_index(ptid);
732     if (cpu < 0) {
733         return;
734     }
735 
736     low_time_offset = get_low_time_offset(dc);
737     if (dc->vcpu_addr[cpu] == 0) {
738         qatomic_inc(&dc->smp_cpus_down);
739     }
740 
741     qatomic_xchg(&dc->last_begin, low_time_offset);
742     qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
743     qatomic_xchg(&dc->vcpu_addr[cpu], addr);
744 
745     /*
746      * check it here, not at the beginning of the function,
747      * due to, check could occur early than bitmap_set in
748      * qemu_ufd_copy_ioctl
749      */
750     already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
751     if (already_received) {
752         qatomic_xchg(&dc->vcpu_addr[cpu], 0);
753         qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
754         qatomic_dec(&dc->smp_cpus_down);
755     }
756     trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
757                                         cpu, already_received);
758 }
759 
760 /*
761  *  This function just provide calculated blocktime per cpu and trace it.
762  *  Total blocktime is calculated in mark_postcopy_blocktime_end.
763  *
764  *
765  * Assume we have 3 CPU
766  *
767  *      S1        E1           S1               E1
768  * -----***********------------xxx***************------------------------> CPU1
769  *
770  *             S2                E2
771  * ------------****************xxx---------------------------------------> CPU2
772  *
773  *                         S3            E3
774  * ------------------------****xxx********-------------------------------> CPU3
775  *
776  * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
777  * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
778  * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
779  *            it's a part of total blocktime.
780  * S1 - here is last_begin
781  * Legend of the picture is following:
782  *              * - means blocktime per vCPU
783  *              x - means overlapped blocktime (total blocktime)
784  *
785  * @addr: host virtual address
786  */
787 static void mark_postcopy_blocktime_end(uintptr_t addr)
788 {
789     MigrationIncomingState *mis = migration_incoming_get_current();
790     PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
791     MachineState *ms = MACHINE(qdev_get_machine());
792     unsigned int smp_cpus = ms->smp.cpus;
793     int i, affected_cpu = 0;
794     bool vcpu_total_blocktime = false;
795     uint32_t read_vcpu_time, low_time_offset;
796 
797     if (!dc) {
798         return;
799     }
800 
801     low_time_offset = get_low_time_offset(dc);
802     /* lookup cpu, to clear it,
803      * that algorithm looks straightforward, but it's not
804      * optimal, more optimal algorithm is keeping tree or hash
805      * where key is address value is a list of  */
806     for (i = 0; i < smp_cpus; i++) {
807         uint32_t vcpu_blocktime = 0;
808 
809         read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
810         if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
811             read_vcpu_time == 0) {
812             continue;
813         }
814         qatomic_xchg(&dc->vcpu_addr[i], 0);
815         vcpu_blocktime = low_time_offset - read_vcpu_time;
816         affected_cpu += 1;
817         /* we need to know is that mark_postcopy_end was due to
818          * faulted page, another possible case it's prefetched
819          * page and in that case we shouldn't be here */
820         if (!vcpu_total_blocktime &&
821             qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
822             vcpu_total_blocktime = true;
823         }
824         /* continue cycle, due to one page could affect several vCPUs */
825         dc->vcpu_blocktime[i] += vcpu_blocktime;
826     }
827 
828     qatomic_sub(&dc->smp_cpus_down, affected_cpu);
829     if (vcpu_total_blocktime) {
830         dc->total_blocktime += low_time_offset - qatomic_fetch_add(
831                 &dc->last_begin, 0);
832     }
833     trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
834                                       affected_cpu);
835 }
836 
837 static bool postcopy_pause_fault_thread(MigrationIncomingState *mis)
838 {
839     trace_postcopy_pause_fault_thread();
840 
841     qemu_sem_wait(&mis->postcopy_pause_sem_fault);
842 
843     trace_postcopy_pause_fault_thread_continued();
844 
845     return true;
846 }
847 
848 /*
849  * Handle faults detected by the USERFAULT markings
850  */
851 static void *postcopy_ram_fault_thread(void *opaque)
852 {
853     MigrationIncomingState *mis = opaque;
854     struct uffd_msg msg;
855     int ret;
856     size_t index;
857     RAMBlock *rb = NULL;
858 
859     trace_postcopy_ram_fault_thread_entry();
860     rcu_register_thread();
861     mis->last_rb = NULL; /* last RAMBlock we sent part of */
862     qemu_sem_post(&mis->fault_thread_sem);
863 
864     struct pollfd *pfd;
865     size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
866 
867     pfd = g_new0(struct pollfd, pfd_len);
868 
869     pfd[0].fd = mis->userfault_fd;
870     pfd[0].events = POLLIN;
871     pfd[1].fd = mis->userfault_event_fd;
872     pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
873     trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
874     for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
875         struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
876                                                  struct PostCopyFD, index);
877         pfd[2 + index].fd = pcfd->fd;
878         pfd[2 + index].events = POLLIN;
879         trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
880                                                   pcfd->fd);
881     }
882 
883     while (true) {
884         ram_addr_t rb_offset;
885         int poll_result;
886 
887         /*
888          * We're mainly waiting for the kernel to give us a faulting HVA,
889          * however we can be told to quit via userfault_quit_fd which is
890          * an eventfd
891          */
892 
893         poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
894         if (poll_result == -1) {
895             error_report("%s: userfault poll: %s", __func__, strerror(errno));
896             break;
897         }
898 
899         if (!mis->to_src_file) {
900             /*
901              * Possibly someone tells us that the return path is
902              * broken already using the event. We should hold until
903              * the channel is rebuilt.
904              */
905             if (postcopy_pause_fault_thread(mis)) {
906                 mis->last_rb = NULL;
907                 /* Continue to read the userfaultfd */
908             } else {
909                 error_report("%s: paused but don't allow to continue",
910                              __func__);
911                 break;
912             }
913         }
914 
915         if (pfd[1].revents) {
916             uint64_t tmp64 = 0;
917 
918             /* Consume the signal */
919             if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
920                 /* Nothing obviously nicer than posting this error. */
921                 error_report("%s: read() failed", __func__);
922             }
923 
924             if (qatomic_read(&mis->fault_thread_quit)) {
925                 trace_postcopy_ram_fault_thread_quit();
926                 break;
927             }
928         }
929 
930         if (pfd[0].revents) {
931             poll_result--;
932             ret = read(mis->userfault_fd, &msg, sizeof(msg));
933             if (ret != sizeof(msg)) {
934                 if (errno == EAGAIN) {
935                     /*
936                      * if a wake up happens on the other thread just after
937                      * the poll, there is nothing to read.
938                      */
939                     continue;
940                 }
941                 if (ret < 0) {
942                     error_report("%s: Failed to read full userfault "
943                                  "message: %s",
944                                  __func__, strerror(errno));
945                     break;
946                 } else {
947                     error_report("%s: Read %d bytes from userfaultfd "
948                                  "expected %zd",
949                                  __func__, ret, sizeof(msg));
950                     break; /* Lost alignment, don't know what we'd read next */
951                 }
952             }
953             if (msg.event != UFFD_EVENT_PAGEFAULT) {
954                 error_report("%s: Read unexpected event %ud from userfaultfd",
955                              __func__, msg.event);
956                 continue; /* It's not a page fault, shouldn't happen */
957             }
958 
959             rb = qemu_ram_block_from_host(
960                      (void *)(uintptr_t)msg.arg.pagefault.address,
961                      true, &rb_offset);
962             if (!rb) {
963                 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
964                              PRIx64, (uint64_t)msg.arg.pagefault.address);
965                 break;
966             }
967 
968             rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
969             trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
970                                                 qemu_ram_get_idstr(rb),
971                                                 rb_offset,
972                                                 msg.arg.pagefault.feat.ptid);
973             mark_postcopy_blocktime_begin(
974                     (uintptr_t)(msg.arg.pagefault.address),
975                                 msg.arg.pagefault.feat.ptid, rb);
976 
977 retry:
978             /*
979              * Send the request to the source - we want to request one
980              * of our host page sizes (which is >= TPS)
981              */
982             ret = migrate_send_rp_req_pages(mis, rb, rb_offset,
983                                             msg.arg.pagefault.address);
984             if (ret) {
985                 /* May be network failure, try to wait for recovery */
986                 if (ret == -EIO && postcopy_pause_fault_thread(mis)) {
987                     /* We got reconnected somehow, try to continue */
988                     mis->last_rb = NULL;
989                     goto retry;
990                 } else {
991                     /* This is a unavoidable fault */
992                     error_report("%s: migrate_send_rp_req_pages() get %d",
993                                  __func__, ret);
994                     break;
995                 }
996             }
997         }
998 
999         /* Now handle any requests from external processes on shared memory */
1000         /* TODO: May need to handle devices deregistering during postcopy */
1001         for (index = 2; index < pfd_len && poll_result; index++) {
1002             if (pfd[index].revents) {
1003                 struct PostCopyFD *pcfd =
1004                     &g_array_index(mis->postcopy_remote_fds,
1005                                    struct PostCopyFD, index - 2);
1006 
1007                 poll_result--;
1008                 if (pfd[index].revents & POLLERR) {
1009                     error_report("%s: POLLERR on poll %zd fd=%d",
1010                                  __func__, index, pcfd->fd);
1011                     pfd[index].events = 0;
1012                     continue;
1013                 }
1014 
1015                 ret = read(pcfd->fd, &msg, sizeof(msg));
1016                 if (ret != sizeof(msg)) {
1017                     if (errno == EAGAIN) {
1018                         /*
1019                          * if a wake up happens on the other thread just after
1020                          * the poll, there is nothing to read.
1021                          */
1022                         continue;
1023                     }
1024                     if (ret < 0) {
1025                         error_report("%s: Failed to read full userfault "
1026                                      "message: %s (shared) revents=%d",
1027                                      __func__, strerror(errno),
1028                                      pfd[index].revents);
1029                         /*TODO: Could just disable this sharer */
1030                         break;
1031                     } else {
1032                         error_report("%s: Read %d bytes from userfaultfd "
1033                                      "expected %zd (shared)",
1034                                      __func__, ret, sizeof(msg));
1035                         /*TODO: Could just disable this sharer */
1036                         break; /*Lost alignment,don't know what we'd read next*/
1037                     }
1038                 }
1039                 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1040                     error_report("%s: Read unexpected event %ud "
1041                                  "from userfaultfd (shared)",
1042                                  __func__, msg.event);
1043                     continue; /* It's not a page fault, shouldn't happen */
1044                 }
1045                 /* Call the device handler registered with us */
1046                 ret = pcfd->handler(pcfd, &msg);
1047                 if (ret) {
1048                     error_report("%s: Failed to resolve shared fault on %zd/%s",
1049                                  __func__, index, pcfd->idstr);
1050                     /* TODO: Fail? Disable this sharer? */
1051                 }
1052             }
1053         }
1054     }
1055     rcu_unregister_thread();
1056     trace_postcopy_ram_fault_thread_exit();
1057     g_free(pfd);
1058     return NULL;
1059 }
1060 
1061 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1062 {
1063     /* Open the fd for the kernel to give us userfaults */
1064     mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1065     if (mis->userfault_fd == -1) {
1066         error_report("%s: Failed to open userfault fd: %s", __func__,
1067                      strerror(errno));
1068         return -1;
1069     }
1070 
1071     /*
1072      * Although the host check already tested the API, we need to
1073      * do the check again as an ABI handshake on the new fd.
1074      */
1075     if (!ufd_check_and_apply(mis->userfault_fd, mis)) {
1076         return -1;
1077     }
1078 
1079     /* Now an eventfd we use to tell the fault-thread to quit */
1080     mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1081     if (mis->userfault_event_fd == -1) {
1082         error_report("%s: Opening userfault_event_fd: %s", __func__,
1083                      strerror(errno));
1084         close(mis->userfault_fd);
1085         return -1;
1086     }
1087 
1088     qemu_sem_init(&mis->fault_thread_sem, 0);
1089     qemu_thread_create(&mis->fault_thread, "postcopy/fault",
1090                        postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
1091     qemu_sem_wait(&mis->fault_thread_sem);
1092     qemu_sem_destroy(&mis->fault_thread_sem);
1093     mis->have_fault_thread = true;
1094 
1095     /* Mark so that we get notified of accesses to unwritten areas */
1096     if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1097         error_report("ram_block_enable_notify failed");
1098         return -1;
1099     }
1100 
1101     mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
1102                                   PROT_READ | PROT_WRITE, MAP_PRIVATE |
1103                                   MAP_ANONYMOUS, -1, 0);
1104     if (mis->postcopy_tmp_page == MAP_FAILED) {
1105         mis->postcopy_tmp_page = NULL;
1106         error_report("%s: Failed to map postcopy_tmp_page %s",
1107                      __func__, strerror(errno));
1108         return -1;
1109     }
1110 
1111     /*
1112      * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages
1113      */
1114     mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1115                                        PROT_READ | PROT_WRITE,
1116                                        MAP_PRIVATE | MAP_ANONYMOUS,
1117                                        -1, 0);
1118     if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1119         int e = errno;
1120         mis->postcopy_tmp_zero_page = NULL;
1121         error_report("%s: Failed to map large zero page %s",
1122                      __func__, strerror(e));
1123         return -e;
1124     }
1125     memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1126 
1127     trace_postcopy_ram_enable_notify();
1128 
1129     return 0;
1130 }
1131 
1132 static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr,
1133                                void *from_addr, uint64_t pagesize, RAMBlock *rb)
1134 {
1135     int userfault_fd = mis->userfault_fd;
1136     int ret;
1137 
1138     if (from_addr) {
1139         struct uffdio_copy copy_struct;
1140         copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1141         copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1142         copy_struct.len = pagesize;
1143         copy_struct.mode = 0;
1144         ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1145     } else {
1146         struct uffdio_zeropage zero_struct;
1147         zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1148         zero_struct.range.len = pagesize;
1149         zero_struct.mode = 0;
1150         ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1151     }
1152     if (!ret) {
1153         qemu_mutex_lock(&mis->page_request_mutex);
1154         ramblock_recv_bitmap_set_range(rb, host_addr,
1155                                        pagesize / qemu_target_page_size());
1156         /*
1157          * If this page resolves a page fault for a previous recorded faulted
1158          * address, take a special note to maintain the requested page list.
1159          */
1160         if (g_tree_lookup(mis->page_requested, host_addr)) {
1161             g_tree_remove(mis->page_requested, host_addr);
1162             mis->page_requested_count--;
1163             trace_postcopy_page_req_del(host_addr, mis->page_requested_count);
1164         }
1165         qemu_mutex_unlock(&mis->page_request_mutex);
1166         mark_postcopy_blocktime_end((uintptr_t)host_addr);
1167     }
1168     return ret;
1169 }
1170 
1171 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1172 {
1173     int i;
1174     MigrationIncomingState *mis = migration_incoming_get_current();
1175     GArray *pcrfds = mis->postcopy_remote_fds;
1176 
1177     for (i = 0; i < pcrfds->len; i++) {
1178         struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1179         int ret = cur->waker(cur, rb, offset);
1180         if (ret) {
1181             return ret;
1182         }
1183     }
1184     return 0;
1185 }
1186 
1187 /*
1188  * Place a host page (from) at (host) atomically
1189  * returns 0 on success
1190  */
1191 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1192                         RAMBlock *rb)
1193 {
1194     size_t pagesize = qemu_ram_pagesize(rb);
1195 
1196     /* copy also acks to the kernel waking the stalled thread up
1197      * TODO: We can inhibit that ack and only do it if it was requested
1198      * which would be slightly cheaper, but we'd have to be careful
1199      * of the order of updating our page state.
1200      */
1201     if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) {
1202         int e = errno;
1203         error_report("%s: %s copy host: %p from: %p (size: %zd)",
1204                      __func__, strerror(e), host, from, pagesize);
1205 
1206         return -e;
1207     }
1208 
1209     trace_postcopy_place_page(host);
1210     return postcopy_notify_shared_wake(rb,
1211                                        qemu_ram_block_host_offset(rb, host));
1212 }
1213 
1214 /*
1215  * Place a zero page at (host) atomically
1216  * returns 0 on success
1217  */
1218 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1219                              RAMBlock *rb)
1220 {
1221     size_t pagesize = qemu_ram_pagesize(rb);
1222     trace_postcopy_place_page_zero(host);
1223 
1224     /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1225      * but it's not available for everything (e.g. hugetlbpages)
1226      */
1227     if (qemu_ram_is_uf_zeroable(rb)) {
1228         if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) {
1229             int e = errno;
1230             error_report("%s: %s zero host: %p",
1231                          __func__, strerror(e), host);
1232 
1233             return -e;
1234         }
1235         return postcopy_notify_shared_wake(rb,
1236                                            qemu_ram_block_host_offset(rb,
1237                                                                       host));
1238     } else {
1239         return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb);
1240     }
1241 }
1242 
1243 #else
1244 /* No target OS support, stubs just fail */
1245 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1246 {
1247 }
1248 
1249 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
1250 {
1251     error_report("%s: No OS support", __func__);
1252     return false;
1253 }
1254 
1255 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1256 {
1257     error_report("postcopy_ram_incoming_init: No OS support");
1258     return -1;
1259 }
1260 
1261 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1262 {
1263     assert(0);
1264     return -1;
1265 }
1266 
1267 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1268 {
1269     assert(0);
1270     return -1;
1271 }
1272 
1273 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1274                                  uint64_t client_addr, uint64_t rb_offset)
1275 {
1276     assert(0);
1277     return -1;
1278 }
1279 
1280 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1281 {
1282     assert(0);
1283     return -1;
1284 }
1285 
1286 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1287                         RAMBlock *rb)
1288 {
1289     assert(0);
1290     return -1;
1291 }
1292 
1293 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1294                         RAMBlock *rb)
1295 {
1296     assert(0);
1297     return -1;
1298 }
1299 
1300 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1301                          uint64_t client_addr,
1302                          RAMBlock *rb)
1303 {
1304     assert(0);
1305     return -1;
1306 }
1307 #endif
1308 
1309 /* ------------------------------------------------------------------------- */
1310 
1311 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1312 {
1313     uint64_t tmp64 = 1;
1314 
1315     /*
1316      * Wakeup the fault_thread.  It's an eventfd that should currently
1317      * be at 0, we're going to increment it to 1
1318      */
1319     if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1320         /* Not much we can do here, but may as well report it */
1321         error_report("%s: incrementing failed: %s", __func__,
1322                      strerror(errno));
1323     }
1324 }
1325 
1326 /**
1327  * postcopy_discard_send_init: Called at the start of each RAMBlock before
1328  *   asking to discard individual ranges.
1329  *
1330  * @ms: The current migration state.
1331  * @offset: the bitmap offset of the named RAMBlock in the migration bitmap.
1332  * @name: RAMBlock that discards will operate on.
1333  */
1334 static PostcopyDiscardState pds = {0};
1335 void postcopy_discard_send_init(MigrationState *ms, const char *name)
1336 {
1337     pds.ramblock_name = name;
1338     pds.cur_entry = 0;
1339     pds.nsentwords = 0;
1340     pds.nsentcmds = 0;
1341 }
1342 
1343 /**
1344  * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1345  *   discard. May send a discard message, may just leave it queued to
1346  *   be sent later.
1347  *
1348  * @ms: Current migration state.
1349  * @start,@length: a range of pages in the migration bitmap in the
1350  *   RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1351  */
1352 void postcopy_discard_send_range(MigrationState *ms, unsigned long start,
1353                                  unsigned long length)
1354 {
1355     size_t tp_size = qemu_target_page_size();
1356     /* Convert to byte offsets within the RAM block */
1357     pds.start_list[pds.cur_entry] = start  * tp_size;
1358     pds.length_list[pds.cur_entry] = length * tp_size;
1359     trace_postcopy_discard_send_range(pds.ramblock_name, start, length);
1360     pds.cur_entry++;
1361     pds.nsentwords++;
1362 
1363     if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) {
1364         /* Full set, ship it! */
1365         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1366                                               pds.ramblock_name,
1367                                               pds.cur_entry,
1368                                               pds.start_list,
1369                                               pds.length_list);
1370         pds.nsentcmds++;
1371         pds.cur_entry = 0;
1372     }
1373 }
1374 
1375 /**
1376  * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1377  * bitmap code. Sends any outstanding discard messages, frees the PDS
1378  *
1379  * @ms: Current migration state.
1380  */
1381 void postcopy_discard_send_finish(MigrationState *ms)
1382 {
1383     /* Anything unsent? */
1384     if (pds.cur_entry) {
1385         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1386                                               pds.ramblock_name,
1387                                               pds.cur_entry,
1388                                               pds.start_list,
1389                                               pds.length_list);
1390         pds.nsentcmds++;
1391     }
1392 
1393     trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords,
1394                                        pds.nsentcmds);
1395 }
1396 
1397 /*
1398  * Current state of incoming postcopy; note this is not part of
1399  * MigrationIncomingState since it's state is used during cleanup
1400  * at the end as MIS is being freed.
1401  */
1402 static PostcopyState incoming_postcopy_state;
1403 
1404 PostcopyState  postcopy_state_get(void)
1405 {
1406     return qatomic_mb_read(&incoming_postcopy_state);
1407 }
1408 
1409 /* Set the state and return the old state */
1410 PostcopyState postcopy_state_set(PostcopyState new_state)
1411 {
1412     return qatomic_xchg(&incoming_postcopy_state, new_state);
1413 }
1414 
1415 /* Register a handler for external shared memory postcopy
1416  * called on the destination.
1417  */
1418 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1419 {
1420     MigrationIncomingState *mis = migration_incoming_get_current();
1421 
1422     mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1423                                                   *pcfd);
1424 }
1425 
1426 /* Unregister a handler for external shared memory postcopy
1427  */
1428 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1429 {
1430     guint i;
1431     MigrationIncomingState *mis = migration_incoming_get_current();
1432     GArray *pcrfds = mis->postcopy_remote_fds;
1433 
1434     for (i = 0; i < pcrfds->len; i++) {
1435         struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1436         if (cur->fd == pcfd->fd) {
1437             mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1438             return;
1439         }
1440     }
1441 }
1442