MAINTAINERS: Add include/exec/gen-icount.h to 'Main Loop' section
[qemu/ar7.git] / migration / postcopy-ram.c
blobab482adef10b84ea45045bb05412b74a3930768f
1 /*
2 * Postcopy migration for RAM
4 * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates
6 * Authors:
7 * Dave Gilbert <dgilbert@redhat.com>
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.
15 * Postcopy is a migration technique where the execution flips from the
16 * source to the destination before all the data has been copied.
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"
34 /* Arbitrary limit on size of each discard command,
35 * keeps them around ~200 bytes
37 #define MAX_DISCARDS_PER_COMMAND 12
39 struct PostcopyDiscardState {
40 const char *ramblock_name;
41 uint16_t cur_entry;
43 * Start and length of a discard range (bytes)
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;
51 static NotifierWithReturnList postcopy_notifier_list;
53 void postcopy_infrastructure_init(void)
55 notifier_with_return_list_init(&postcopy_notifier_list);
58 void postcopy_add_notifier(NotifierWithReturn *nn)
60 notifier_with_return_list_add(&postcopy_notifier_list, nn);
63 void postcopy_remove_notifier(NotifierWithReturn *n)
65 notifier_with_return_remove(n);
68 int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp)
70 struct PostcopyNotifyData pnd;
71 pnd.reason = reason;
72 pnd.errp = errp;
74 return notifier_with_return_list_notify(&postcopy_notifier_list,
75 &pnd);
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.
82 #if defined(__linux__)
84 #include <poll.h>
85 #include <sys/ioctl.h>
86 #include <sys/syscall.h>
87 #include <asm/types.h> /* for __u64 */
88 #endif
90 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
91 #include <sys/eventfd.h>
92 #include <linux/userfaultfd.h>
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;
109 * Handler for exit event, necessary for
110 * releasing whole blocktime_ctx
112 Notifier exit_notifier;
113 } PostcopyBlocktimeContext;
115 static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx)
117 g_free(ctx->page_fault_vcpu_time);
118 g_free(ctx->vcpu_addr);
119 g_free(ctx->vcpu_blocktime);
120 g_free(ctx);
123 static void migration_exit_cb(Notifier *n, void *data)
125 PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext,
126 exit_notifier);
127 destroy_blocktime_context(ctx);
130 static struct PostcopyBlocktimeContext *blocktime_context_new(void)
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);
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;
145 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx)
147 MachineState *ms = MACHINE(qdev_get_machine());
148 uint32List *list = NULL;
149 int i;
151 for (i = ms->smp.cpus - 1; i >= 0; i--) {
152 QAPI_LIST_PREPEND(list, ctx->vcpu_blocktime[i]);
155 return list;
159 * This function just populates MigrationInfo from postcopy's
160 * blocktime context. It will not populate MigrationInfo,
161 * unless postcopy-blocktime capability was set.
163 * @info: pointer to MigrationInfo to populate
165 void fill_destination_postcopy_migration_info(MigrationInfo *info)
167 MigrationIncomingState *mis = migration_incoming_get_current();
168 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
170 if (!bc) {
171 return;
174 info->has_postcopy_blocktime = true;
175 info->postcopy_blocktime = bc->total_blocktime;
176 info->has_postcopy_vcpu_blocktime = true;
177 info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc);
180 static uint32_t get_postcopy_total_blocktime(void)
182 MigrationIncomingState *mis = migration_incoming_get_current();
183 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
185 if (!bc) {
186 return 0;
189 return bc->total_blocktime;
193 * receive_ufd_features: check userfault fd features, to request only supported
194 * features in the future.
196 * Returns: true on success
198 * __NR_userfaultfd - should be checked before
199 * @features: out parameter will contain uffdio_api.features provided by kernel
200 * in case of success
202 static bool receive_ufd_features(uint64_t *features)
204 struct uffdio_api api_struct = {0};
205 int ufd;
206 bool ret = true;
208 /* if we are here __NR_userfaultfd should exists */
209 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
210 if (ufd == -1) {
211 error_report("%s: syscall __NR_userfaultfd failed: %s", __func__,
212 strerror(errno));
213 return false;
216 /* ask features */
217 api_struct.api = UFFD_API;
218 api_struct.features = 0;
219 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
220 error_report("%s: UFFDIO_API failed: %s", __func__,
221 strerror(errno));
222 ret = false;
223 goto release_ufd;
226 *features = api_struct.features;
228 release_ufd:
229 close(ufd);
230 return ret;
234 * request_ufd_features: this function should be called only once on a newly
235 * opened ufd, subsequent calls will lead to error.
237 * Returns: true on success
239 * @ufd: fd obtained from userfaultfd syscall
240 * @features: bit mask see UFFD_API_FEATURES
242 static bool request_ufd_features(int ufd, uint64_t features)
244 struct uffdio_api api_struct = {0};
245 uint64_t ioctl_mask;
247 api_struct.api = UFFD_API;
248 api_struct.features = features;
249 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
250 error_report("%s failed: UFFDIO_API failed: %s", __func__,
251 strerror(errno));
252 return false;
255 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
256 (__u64)1 << _UFFDIO_UNREGISTER;
257 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
258 error_report("Missing userfault features: %" PRIx64,
259 (uint64_t)(~api_struct.ioctls & ioctl_mask));
260 return false;
263 return true;
266 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis)
268 uint64_t asked_features = 0;
269 static uint64_t supported_features;
272 * it's not possible to
273 * request UFFD_API twice per one fd
274 * userfault fd features is persistent
276 if (!supported_features) {
277 if (!receive_ufd_features(&supported_features)) {
278 error_report("%s failed", __func__);
279 return false;
283 #ifdef UFFD_FEATURE_THREAD_ID
284 if (migrate_postcopy_blocktime() && mis &&
285 UFFD_FEATURE_THREAD_ID & supported_features) {
286 /* kernel supports that feature */
287 /* don't create blocktime_context if it exists */
288 if (!mis->blocktime_ctx) {
289 mis->blocktime_ctx = blocktime_context_new();
292 asked_features |= UFFD_FEATURE_THREAD_ID;
294 #endif
297 * request features, even if asked_features is 0, due to
298 * kernel expects UFFD_API before UFFDIO_REGISTER, per
299 * userfault file descriptor
301 if (!request_ufd_features(ufd, asked_features)) {
302 error_report("%s failed: features %" PRIu64, __func__,
303 asked_features);
304 return false;
307 if (qemu_real_host_page_size != ram_pagesize_summary()) {
308 bool have_hp = false;
309 /* We've got a huge page */
310 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
311 have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS;
312 #endif
313 if (!have_hp) {
314 error_report("Userfault on this host does not support huge pages");
315 return false;
318 return true;
321 /* Callback from postcopy_ram_supported_by_host block iterator.
323 static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque)
325 const char *block_name = qemu_ram_get_idstr(rb);
326 ram_addr_t length = qemu_ram_get_used_length(rb);
327 size_t pagesize = qemu_ram_pagesize(rb);
329 if (length % pagesize) {
330 error_report("Postcopy requires RAM blocks to be a page size multiple,"
331 " block %s is 0x" RAM_ADDR_FMT " bytes with a "
332 "page size of 0x%zx", block_name, length, pagesize);
333 return 1;
335 return 0;
339 * Note: This has the side effect of munlock'ing all of RAM, that's
340 * normally fine since if the postcopy succeeds it gets turned back on at the
341 * end.
343 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
345 long pagesize = qemu_real_host_page_size;
346 int ufd = -1;
347 bool ret = false; /* Error unless we change it */
348 void *testarea = NULL;
349 struct uffdio_register reg_struct;
350 struct uffdio_range range_struct;
351 uint64_t feature_mask;
352 Error *local_err = NULL;
354 if (qemu_target_page_size() > pagesize) {
355 error_report("Target page size bigger than host page size");
356 goto out;
359 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
360 if (ufd == -1) {
361 error_report("%s: userfaultfd not available: %s", __func__,
362 strerror(errno));
363 goto out;
366 /* Give devices a chance to object */
367 if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) {
368 error_report_err(local_err);
369 goto out;
372 /* Version and features check */
373 if (!ufd_check_and_apply(ufd, mis)) {
374 goto out;
377 /* We don't support postcopy with shared RAM yet */
378 if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) {
379 goto out;
383 * userfault and mlock don't go together; we'll put it back later if
384 * it was enabled.
386 if (munlockall()) {
387 error_report("%s: munlockall: %s", __func__, strerror(errno));
388 goto out;
392 * We need to check that the ops we need are supported on anon memory
393 * To do that we need to register a chunk and see the flags that
394 * are returned.
396 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
397 MAP_ANONYMOUS, -1, 0);
398 if (testarea == MAP_FAILED) {
399 error_report("%s: Failed to map test area: %s", __func__,
400 strerror(errno));
401 goto out;
403 g_assert(((size_t)testarea & (pagesize - 1)) == 0);
405 reg_struct.range.start = (uintptr_t)testarea;
406 reg_struct.range.len = pagesize;
407 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
409 if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
410 error_report("%s userfault register: %s", __func__, strerror(errno));
411 goto out;
414 range_struct.start = (uintptr_t)testarea;
415 range_struct.len = pagesize;
416 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
417 error_report("%s userfault unregister: %s", __func__, strerror(errno));
418 goto out;
421 feature_mask = (__u64)1 << _UFFDIO_WAKE |
422 (__u64)1 << _UFFDIO_COPY |
423 (__u64)1 << _UFFDIO_ZEROPAGE;
424 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
425 error_report("Missing userfault map features: %" PRIx64,
426 (uint64_t)(~reg_struct.ioctls & feature_mask));
427 goto out;
430 /* Success! */
431 ret = true;
432 out:
433 if (testarea) {
434 munmap(testarea, pagesize);
436 if (ufd != -1) {
437 close(ufd);
439 return ret;
443 * Setup an area of RAM so that it *can* be used for postcopy later; this
444 * must be done right at the start prior to pre-copy.
445 * opaque should be the MIS.
447 static int init_range(RAMBlock *rb, void *opaque)
449 const char *block_name = qemu_ram_get_idstr(rb);
450 void *host_addr = qemu_ram_get_host_addr(rb);
451 ram_addr_t offset = qemu_ram_get_offset(rb);
452 ram_addr_t length = qemu_ram_get_used_length(rb);
453 trace_postcopy_init_range(block_name, host_addr, offset, length);
456 * We need the whole of RAM to be truly empty for postcopy, so things
457 * like ROMs and any data tables built during init must be zero'd
458 * - we're going to get the copy from the source anyway.
459 * (Precopy will just overwrite this data, so doesn't need the discard)
461 if (ram_discard_range(block_name, 0, length)) {
462 return -1;
465 return 0;
469 * At the end of migration, undo the effects of init_range
470 * opaque should be the MIS.
472 static int cleanup_range(RAMBlock *rb, void *opaque)
474 const char *block_name = qemu_ram_get_idstr(rb);
475 void *host_addr = qemu_ram_get_host_addr(rb);
476 ram_addr_t offset = qemu_ram_get_offset(rb);
477 ram_addr_t length = qemu_ram_get_used_length(rb);
478 MigrationIncomingState *mis = opaque;
479 struct uffdio_range range_struct;
480 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
483 * We turned off hugepage for the precopy stage with postcopy enabled
484 * we can turn it back on now.
486 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
489 * We can also turn off userfault now since we should have all the
490 * pages. It can be useful to leave it on to debug postcopy
491 * if you're not sure it's always getting every page.
493 range_struct.start = (uintptr_t)host_addr;
494 range_struct.len = length;
496 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
497 error_report("%s: userfault unregister %s", __func__, strerror(errno));
499 return -1;
502 return 0;
506 * Initialise postcopy-ram, setting the RAM to a state where we can go into
507 * postcopy later; must be called prior to any precopy.
508 * called from arch_init's similarly named ram_postcopy_incoming_init
510 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
512 if (foreach_not_ignored_block(init_range, NULL)) {
513 return -1;
516 return 0;
520 * At the end of a migration where postcopy_ram_incoming_init was called.
522 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
524 trace_postcopy_ram_incoming_cleanup_entry();
526 if (mis->have_fault_thread) {
527 Error *local_err = NULL;
529 /* Let the fault thread quit */
530 qatomic_set(&mis->fault_thread_quit, 1);
531 postcopy_fault_thread_notify(mis);
532 trace_postcopy_ram_incoming_cleanup_join();
533 qemu_thread_join(&mis->fault_thread);
535 if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
536 error_report_err(local_err);
537 return -1;
540 if (foreach_not_ignored_block(cleanup_range, mis)) {
541 return -1;
544 trace_postcopy_ram_incoming_cleanup_closeuf();
545 close(mis->userfault_fd);
546 close(mis->userfault_event_fd);
547 mis->have_fault_thread = false;
550 if (enable_mlock) {
551 if (os_mlock() < 0) {
552 error_report("mlock: %s", strerror(errno));
554 * It doesn't feel right to fail at this point, we have a valid
555 * VM state.
560 if (mis->postcopy_tmp_page) {
561 munmap(mis->postcopy_tmp_page, mis->largest_page_size);
562 mis->postcopy_tmp_page = NULL;
564 if (mis->postcopy_tmp_zero_page) {
565 munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
566 mis->postcopy_tmp_zero_page = NULL;
568 trace_postcopy_ram_incoming_cleanup_blocktime(
569 get_postcopy_total_blocktime());
571 trace_postcopy_ram_incoming_cleanup_exit();
572 return 0;
576 * Disable huge pages on an area
578 static int nhp_range(RAMBlock *rb, void *opaque)
580 const char *block_name = qemu_ram_get_idstr(rb);
581 void *host_addr = qemu_ram_get_host_addr(rb);
582 ram_addr_t offset = qemu_ram_get_offset(rb);
583 ram_addr_t length = qemu_ram_get_used_length(rb);
584 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
587 * Before we do discards we need to ensure those discards really
588 * do delete areas of the page, even if THP thinks a hugepage would
589 * be a good idea, so force hugepages off.
591 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
593 return 0;
597 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
598 * however leaving it until after precopy means that most of the precopy
599 * data is still THPd
601 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
603 if (foreach_not_ignored_block(nhp_range, mis)) {
604 return -1;
607 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
609 return 0;
613 * Mark the given area of RAM as requiring notification to unwritten areas
614 * Used as a callback on foreach_not_ignored_block.
615 * host_addr: Base of area to mark
616 * offset: Offset in the whole ram arena
617 * length: Length of the section
618 * opaque: MigrationIncomingState pointer
619 * Returns 0 on success
621 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
623 MigrationIncomingState *mis = opaque;
624 struct uffdio_register reg_struct;
626 reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
627 reg_struct.range.len = qemu_ram_get_used_length(rb);
628 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
630 /* Now tell our userfault_fd that it's responsible for this area */
631 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
632 error_report("%s userfault register: %s", __func__, strerror(errno));
633 return -1;
635 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
636 error_report("%s userfault: Region doesn't support COPY", __func__);
637 return -1;
639 if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) {
640 qemu_ram_set_uf_zeroable(rb);
643 return 0;
646 int postcopy_wake_shared(struct PostCopyFD *pcfd,
647 uint64_t client_addr,
648 RAMBlock *rb)
650 size_t pagesize = qemu_ram_pagesize(rb);
651 struct uffdio_range range;
652 int ret;
653 trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
654 range.start = client_addr & ~(pagesize - 1);
655 range.len = pagesize;
656 ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
657 if (ret) {
658 error_report("%s: Failed to wake: %zx in %s (%s)",
659 __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
660 strerror(errno));
662 return ret;
666 * Callback from shared fault handlers to ask for a page,
667 * the page must be specified by a RAMBlock and an offset in that rb
668 * Note: Only for use by shared fault handlers (in fault thread)
670 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
671 uint64_t client_addr, uint64_t rb_offset)
673 size_t pagesize = qemu_ram_pagesize(rb);
674 uint64_t aligned_rbo = rb_offset & ~(pagesize - 1);
675 MigrationIncomingState *mis = migration_incoming_get_current();
677 trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
678 rb_offset);
679 if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
680 trace_postcopy_request_shared_page_present(pcfd->idstr,
681 qemu_ram_get_idstr(rb), rb_offset);
682 return postcopy_wake_shared(pcfd, client_addr, rb);
684 migrate_send_rp_req_pages(mis, rb, aligned_rbo, client_addr);
685 return 0;
688 static int get_mem_fault_cpu_index(uint32_t pid)
690 CPUState *cpu_iter;
692 CPU_FOREACH(cpu_iter) {
693 if (cpu_iter->thread_id == pid) {
694 trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
695 return cpu_iter->cpu_index;
698 trace_get_mem_fault_cpu_index(-1, pid);
699 return -1;
702 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
704 int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
705 dc->start_time;
706 return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
710 * This function is being called when pagefault occurs. It
711 * tracks down vCPU blocking time.
713 * @addr: faulted host virtual address
714 * @ptid: faulted process thread id
715 * @rb: ramblock appropriate to addr
717 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
718 RAMBlock *rb)
720 int cpu, already_received;
721 MigrationIncomingState *mis = migration_incoming_get_current();
722 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
723 uint32_t low_time_offset;
725 if (!dc || ptid == 0) {
726 return;
728 cpu = get_mem_fault_cpu_index(ptid);
729 if (cpu < 0) {
730 return;
733 low_time_offset = get_low_time_offset(dc);
734 if (dc->vcpu_addr[cpu] == 0) {
735 qatomic_inc(&dc->smp_cpus_down);
738 qatomic_xchg(&dc->last_begin, low_time_offset);
739 qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
740 qatomic_xchg(&dc->vcpu_addr[cpu], addr);
743 * check it here, not at the beginning of the function,
744 * due to, check could occur early than bitmap_set in
745 * qemu_ufd_copy_ioctl
747 already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
748 if (already_received) {
749 qatomic_xchg(&dc->vcpu_addr[cpu], 0);
750 qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
751 qatomic_dec(&dc->smp_cpus_down);
753 trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
754 cpu, already_received);
758 * This function just provide calculated blocktime per cpu and trace it.
759 * Total blocktime is calculated in mark_postcopy_blocktime_end.
762 * Assume we have 3 CPU
764 * S1 E1 S1 E1
765 * -----***********------------xxx***************------------------------> CPU1
767 * S2 E2
768 * ------------****************xxx---------------------------------------> CPU2
770 * S3 E3
771 * ------------------------****xxx********-------------------------------> CPU3
773 * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
774 * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
775 * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
776 * it's a part of total blocktime.
777 * S1 - here is last_begin
778 * Legend of the picture is following:
779 * * - means blocktime per vCPU
780 * x - means overlapped blocktime (total blocktime)
782 * @addr: host virtual address
784 static void mark_postcopy_blocktime_end(uintptr_t addr)
786 MigrationIncomingState *mis = migration_incoming_get_current();
787 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
788 MachineState *ms = MACHINE(qdev_get_machine());
789 unsigned int smp_cpus = ms->smp.cpus;
790 int i, affected_cpu = 0;
791 bool vcpu_total_blocktime = false;
792 uint32_t read_vcpu_time, low_time_offset;
794 if (!dc) {
795 return;
798 low_time_offset = get_low_time_offset(dc);
799 /* lookup cpu, to clear it,
800 * that algorithm looks straightforward, but it's not
801 * optimal, more optimal algorithm is keeping tree or hash
802 * where key is address value is a list of */
803 for (i = 0; i < smp_cpus; i++) {
804 uint32_t vcpu_blocktime = 0;
806 read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
807 if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
808 read_vcpu_time == 0) {
809 continue;
811 qatomic_xchg(&dc->vcpu_addr[i], 0);
812 vcpu_blocktime = low_time_offset - read_vcpu_time;
813 affected_cpu += 1;
814 /* we need to know is that mark_postcopy_end was due to
815 * faulted page, another possible case it's prefetched
816 * page and in that case we shouldn't be here */
817 if (!vcpu_total_blocktime &&
818 qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
819 vcpu_total_blocktime = true;
821 /* continue cycle, due to one page could affect several vCPUs */
822 dc->vcpu_blocktime[i] += vcpu_blocktime;
825 qatomic_sub(&dc->smp_cpus_down, affected_cpu);
826 if (vcpu_total_blocktime) {
827 dc->total_blocktime += low_time_offset - qatomic_fetch_add(
828 &dc->last_begin, 0);
830 trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
831 affected_cpu);
834 static bool postcopy_pause_fault_thread(MigrationIncomingState *mis)
836 trace_postcopy_pause_fault_thread();
838 qemu_sem_wait(&mis->postcopy_pause_sem_fault);
840 trace_postcopy_pause_fault_thread_continued();
842 return true;
846 * Handle faults detected by the USERFAULT markings
848 static void *postcopy_ram_fault_thread(void *opaque)
850 MigrationIncomingState *mis = opaque;
851 struct uffd_msg msg;
852 int ret;
853 size_t index;
854 RAMBlock *rb = NULL;
856 trace_postcopy_ram_fault_thread_entry();
857 rcu_register_thread();
858 mis->last_rb = NULL; /* last RAMBlock we sent part of */
859 qemu_sem_post(&mis->fault_thread_sem);
861 struct pollfd *pfd;
862 size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
864 pfd = g_new0(struct pollfd, pfd_len);
866 pfd[0].fd = mis->userfault_fd;
867 pfd[0].events = POLLIN;
868 pfd[1].fd = mis->userfault_event_fd;
869 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
870 trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
871 for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
872 struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
873 struct PostCopyFD, index);
874 pfd[2 + index].fd = pcfd->fd;
875 pfd[2 + index].events = POLLIN;
876 trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
877 pcfd->fd);
880 while (true) {
881 ram_addr_t rb_offset;
882 int poll_result;
885 * We're mainly waiting for the kernel to give us a faulting HVA,
886 * however we can be told to quit via userfault_quit_fd which is
887 * an eventfd
890 poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
891 if (poll_result == -1) {
892 error_report("%s: userfault poll: %s", __func__, strerror(errno));
893 break;
896 if (!mis->to_src_file) {
898 * Possibly someone tells us that the return path is
899 * broken already using the event. We should hold until
900 * the channel is rebuilt.
902 if (postcopy_pause_fault_thread(mis)) {
903 /* Continue to read the userfaultfd */
904 } else {
905 error_report("%s: paused but don't allow to continue",
906 __func__);
907 break;
911 if (pfd[1].revents) {
912 uint64_t tmp64 = 0;
914 /* Consume the signal */
915 if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
916 /* Nothing obviously nicer than posting this error. */
917 error_report("%s: read() failed", __func__);
920 if (qatomic_read(&mis->fault_thread_quit)) {
921 trace_postcopy_ram_fault_thread_quit();
922 break;
926 if (pfd[0].revents) {
927 poll_result--;
928 ret = read(mis->userfault_fd, &msg, sizeof(msg));
929 if (ret != sizeof(msg)) {
930 if (errno == EAGAIN) {
932 * if a wake up happens on the other thread just after
933 * the poll, there is nothing to read.
935 continue;
937 if (ret < 0) {
938 error_report("%s: Failed to read full userfault "
939 "message: %s",
940 __func__, strerror(errno));
941 break;
942 } else {
943 error_report("%s: Read %d bytes from userfaultfd "
944 "expected %zd",
945 __func__, ret, sizeof(msg));
946 break; /* Lost alignment, don't know what we'd read next */
949 if (msg.event != UFFD_EVENT_PAGEFAULT) {
950 error_report("%s: Read unexpected event %ud from userfaultfd",
951 __func__, msg.event);
952 continue; /* It's not a page fault, shouldn't happen */
955 rb = qemu_ram_block_from_host(
956 (void *)(uintptr_t)msg.arg.pagefault.address,
957 true, &rb_offset);
958 if (!rb) {
959 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
960 PRIx64, (uint64_t)msg.arg.pagefault.address);
961 break;
964 rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
965 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
966 qemu_ram_get_idstr(rb),
967 rb_offset,
968 msg.arg.pagefault.feat.ptid);
969 mark_postcopy_blocktime_begin(
970 (uintptr_t)(msg.arg.pagefault.address),
971 msg.arg.pagefault.feat.ptid, rb);
973 retry:
975 * Send the request to the source - we want to request one
976 * of our host page sizes (which is >= TPS)
978 ret = migrate_send_rp_req_pages(mis, rb, rb_offset,
979 msg.arg.pagefault.address);
980 if (ret) {
981 /* May be network failure, try to wait for recovery */
982 if (ret == -EIO && postcopy_pause_fault_thread(mis)) {
983 /* We got reconnected somehow, try to continue */
984 goto retry;
985 } else {
986 /* This is a unavoidable fault */
987 error_report("%s: migrate_send_rp_req_pages() get %d",
988 __func__, ret);
989 break;
994 /* Now handle any requests from external processes on shared memory */
995 /* TODO: May need to handle devices deregistering during postcopy */
996 for (index = 2; index < pfd_len && poll_result; index++) {
997 if (pfd[index].revents) {
998 struct PostCopyFD *pcfd =
999 &g_array_index(mis->postcopy_remote_fds,
1000 struct PostCopyFD, index - 2);
1002 poll_result--;
1003 if (pfd[index].revents & POLLERR) {
1004 error_report("%s: POLLERR on poll %zd fd=%d",
1005 __func__, index, pcfd->fd);
1006 pfd[index].events = 0;
1007 continue;
1010 ret = read(pcfd->fd, &msg, sizeof(msg));
1011 if (ret != sizeof(msg)) {
1012 if (errno == EAGAIN) {
1014 * if a wake up happens on the other thread just after
1015 * the poll, there is nothing to read.
1017 continue;
1019 if (ret < 0) {
1020 error_report("%s: Failed to read full userfault "
1021 "message: %s (shared) revents=%d",
1022 __func__, strerror(errno),
1023 pfd[index].revents);
1024 /*TODO: Could just disable this sharer */
1025 break;
1026 } else {
1027 error_report("%s: Read %d bytes from userfaultfd "
1028 "expected %zd (shared)",
1029 __func__, ret, sizeof(msg));
1030 /*TODO: Could just disable this sharer */
1031 break; /*Lost alignment,don't know what we'd read next*/
1034 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1035 error_report("%s: Read unexpected event %ud "
1036 "from userfaultfd (shared)",
1037 __func__, msg.event);
1038 continue; /* It's not a page fault, shouldn't happen */
1040 /* Call the device handler registered with us */
1041 ret = pcfd->handler(pcfd, &msg);
1042 if (ret) {
1043 error_report("%s: Failed to resolve shared fault on %zd/%s",
1044 __func__, index, pcfd->idstr);
1045 /* TODO: Fail? Disable this sharer? */
1050 rcu_unregister_thread();
1051 trace_postcopy_ram_fault_thread_exit();
1052 g_free(pfd);
1053 return NULL;
1056 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1058 /* Open the fd for the kernel to give us userfaults */
1059 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1060 if (mis->userfault_fd == -1) {
1061 error_report("%s: Failed to open userfault fd: %s", __func__,
1062 strerror(errno));
1063 return -1;
1067 * Although the host check already tested the API, we need to
1068 * do the check again as an ABI handshake on the new fd.
1070 if (!ufd_check_and_apply(mis->userfault_fd, mis)) {
1071 return -1;
1074 /* Now an eventfd we use to tell the fault-thread to quit */
1075 mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1076 if (mis->userfault_event_fd == -1) {
1077 error_report("%s: Opening userfault_event_fd: %s", __func__,
1078 strerror(errno));
1079 close(mis->userfault_fd);
1080 return -1;
1083 qemu_sem_init(&mis->fault_thread_sem, 0);
1084 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
1085 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
1086 qemu_sem_wait(&mis->fault_thread_sem);
1087 qemu_sem_destroy(&mis->fault_thread_sem);
1088 mis->have_fault_thread = true;
1090 /* Mark so that we get notified of accesses to unwritten areas */
1091 if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1092 error_report("ram_block_enable_notify failed");
1093 return -1;
1096 mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
1097 PROT_READ | PROT_WRITE, MAP_PRIVATE |
1098 MAP_ANONYMOUS, -1, 0);
1099 if (mis->postcopy_tmp_page == MAP_FAILED) {
1100 mis->postcopy_tmp_page = NULL;
1101 error_report("%s: Failed to map postcopy_tmp_page %s",
1102 __func__, strerror(errno));
1103 return -1;
1107 * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages
1109 mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1110 PROT_READ | PROT_WRITE,
1111 MAP_PRIVATE | MAP_ANONYMOUS,
1112 -1, 0);
1113 if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1114 int e = errno;
1115 mis->postcopy_tmp_zero_page = NULL;
1116 error_report("%s: Failed to map large zero page %s",
1117 __func__, strerror(e));
1118 return -e;
1120 memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1122 trace_postcopy_ram_enable_notify();
1124 return 0;
1127 static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr,
1128 void *from_addr, uint64_t pagesize, RAMBlock *rb)
1130 int userfault_fd = mis->userfault_fd;
1131 int ret;
1133 if (from_addr) {
1134 struct uffdio_copy copy_struct;
1135 copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1136 copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1137 copy_struct.len = pagesize;
1138 copy_struct.mode = 0;
1139 ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1140 } else {
1141 struct uffdio_zeropage zero_struct;
1142 zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1143 zero_struct.range.len = pagesize;
1144 zero_struct.mode = 0;
1145 ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1147 if (!ret) {
1148 qemu_mutex_lock(&mis->page_request_mutex);
1149 ramblock_recv_bitmap_set_range(rb, host_addr,
1150 pagesize / qemu_target_page_size());
1152 * If this page resolves a page fault for a previous recorded faulted
1153 * address, take a special note to maintain the requested page list.
1155 if (g_tree_lookup(mis->page_requested, host_addr)) {
1156 g_tree_remove(mis->page_requested, host_addr);
1157 mis->page_requested_count--;
1158 trace_postcopy_page_req_del(host_addr, mis->page_requested_count);
1160 qemu_mutex_unlock(&mis->page_request_mutex);
1161 mark_postcopy_blocktime_end((uintptr_t)host_addr);
1163 return ret;
1166 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1168 int i;
1169 MigrationIncomingState *mis = migration_incoming_get_current();
1170 GArray *pcrfds = mis->postcopy_remote_fds;
1172 for (i = 0; i < pcrfds->len; i++) {
1173 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1174 int ret = cur->waker(cur, rb, offset);
1175 if (ret) {
1176 return ret;
1179 return 0;
1183 * Place a host page (from) at (host) atomically
1184 * returns 0 on success
1186 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1187 RAMBlock *rb)
1189 size_t pagesize = qemu_ram_pagesize(rb);
1191 /* copy also acks to the kernel waking the stalled thread up
1192 * TODO: We can inhibit that ack and only do it if it was requested
1193 * which would be slightly cheaper, but we'd have to be careful
1194 * of the order of updating our page state.
1196 if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) {
1197 int e = errno;
1198 error_report("%s: %s copy host: %p from: %p (size: %zd)",
1199 __func__, strerror(e), host, from, pagesize);
1201 return -e;
1204 trace_postcopy_place_page(host);
1205 return postcopy_notify_shared_wake(rb,
1206 qemu_ram_block_host_offset(rb, host));
1210 * Place a zero page at (host) atomically
1211 * returns 0 on success
1213 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1214 RAMBlock *rb)
1216 size_t pagesize = qemu_ram_pagesize(rb);
1217 trace_postcopy_place_page_zero(host);
1219 /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1220 * but it's not available for everything (e.g. hugetlbpages)
1222 if (qemu_ram_is_uf_zeroable(rb)) {
1223 if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) {
1224 int e = errno;
1225 error_report("%s: %s zero host: %p",
1226 __func__, strerror(e), host);
1228 return -e;
1230 return postcopy_notify_shared_wake(rb,
1231 qemu_ram_block_host_offset(rb,
1232 host));
1233 } else {
1234 return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb);
1238 #else
1239 /* No target OS support, stubs just fail */
1240 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1244 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
1246 error_report("%s: No OS support", __func__);
1247 return false;
1250 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1252 error_report("postcopy_ram_incoming_init: No OS support");
1253 return -1;
1256 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1258 assert(0);
1259 return -1;
1262 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1264 assert(0);
1265 return -1;
1268 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1269 uint64_t client_addr, uint64_t rb_offset)
1271 assert(0);
1272 return -1;
1275 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1277 assert(0);
1278 return -1;
1281 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1282 RAMBlock *rb)
1284 assert(0);
1285 return -1;
1288 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1289 RAMBlock *rb)
1291 assert(0);
1292 return -1;
1295 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1296 uint64_t client_addr,
1297 RAMBlock *rb)
1299 assert(0);
1300 return -1;
1302 #endif
1304 /* ------------------------------------------------------------------------- */
1306 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1308 uint64_t tmp64 = 1;
1311 * Wakeup the fault_thread. It's an eventfd that should currently
1312 * be at 0, we're going to increment it to 1
1314 if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1315 /* Not much we can do here, but may as well report it */
1316 error_report("%s: incrementing failed: %s", __func__,
1317 strerror(errno));
1322 * postcopy_discard_send_init: Called at the start of each RAMBlock before
1323 * asking to discard individual ranges.
1325 * @ms: The current migration state.
1326 * @offset: the bitmap offset of the named RAMBlock in the migration bitmap.
1327 * @name: RAMBlock that discards will operate on.
1329 static PostcopyDiscardState pds = {0};
1330 void postcopy_discard_send_init(MigrationState *ms, const char *name)
1332 pds.ramblock_name = name;
1333 pds.cur_entry = 0;
1334 pds.nsentwords = 0;
1335 pds.nsentcmds = 0;
1339 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1340 * discard. May send a discard message, may just leave it queued to
1341 * be sent later.
1343 * @ms: Current migration state.
1344 * @start,@length: a range of pages in the migration bitmap in the
1345 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1347 void postcopy_discard_send_range(MigrationState *ms, unsigned long start,
1348 unsigned long length)
1350 size_t tp_size = qemu_target_page_size();
1351 /* Convert to byte offsets within the RAM block */
1352 pds.start_list[pds.cur_entry] = start * tp_size;
1353 pds.length_list[pds.cur_entry] = length * tp_size;
1354 trace_postcopy_discard_send_range(pds.ramblock_name, start, length);
1355 pds.cur_entry++;
1356 pds.nsentwords++;
1358 if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) {
1359 /* Full set, ship it! */
1360 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1361 pds.ramblock_name,
1362 pds.cur_entry,
1363 pds.start_list,
1364 pds.length_list);
1365 pds.nsentcmds++;
1366 pds.cur_entry = 0;
1371 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1372 * bitmap code. Sends any outstanding discard messages, frees the PDS
1374 * @ms: Current migration state.
1376 void postcopy_discard_send_finish(MigrationState *ms)
1378 /* Anything unsent? */
1379 if (pds.cur_entry) {
1380 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1381 pds.ramblock_name,
1382 pds.cur_entry,
1383 pds.start_list,
1384 pds.length_list);
1385 pds.nsentcmds++;
1388 trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords,
1389 pds.nsentcmds);
1393 * Current state of incoming postcopy; note this is not part of
1394 * MigrationIncomingState since it's state is used during cleanup
1395 * at the end as MIS is being freed.
1397 static PostcopyState incoming_postcopy_state;
1399 PostcopyState postcopy_state_get(void)
1401 return qatomic_mb_read(&incoming_postcopy_state);
1404 /* Set the state and return the old state */
1405 PostcopyState postcopy_state_set(PostcopyState new_state)
1407 return qatomic_xchg(&incoming_postcopy_state, new_state);
1410 /* Register a handler for external shared memory postcopy
1411 * called on the destination.
1413 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1415 MigrationIncomingState *mis = migration_incoming_get_current();
1417 mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1418 *pcfd);
1421 /* Unregister a handler for external shared memory postcopy
1423 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1425 guint i;
1426 MigrationIncomingState *mis = migration_incoming_get_current();
1427 GArray *pcrfds = mis->postcopy_remote_fds;
1429 for (i = 0; i < pcrfds->len; i++) {
1430 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1431 if (cur->fd == pcfd->fd) {
1432 mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1433 return;