hw/arm/boot: Honour image size field in AArch64 Image format kernels
[qemu/ar7.git] / migration / postcopy-ram.c
blobe2aa57a7013543cca48a8001bd46d448a5bb7eeb
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 "sysemu/sysemu.h"
29 #include "sysemu/balloon.h"
30 #include "qemu/error-report.h"
31 #include "trace.h"
33 /* Arbitrary limit on size of each discard command,
34 * keeps them around ~200 bytes
36 #define MAX_DISCARDS_PER_COMMAND 12
38 struct PostcopyDiscardState {
39 const char *ramblock_name;
40 uint16_t cur_entry;
42 * Start and length of a discard range (bytes)
44 uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
45 uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
46 unsigned int nsentwords;
47 unsigned int nsentcmds;
50 static NotifierWithReturnList postcopy_notifier_list;
52 void postcopy_infrastructure_init(void)
54 notifier_with_return_list_init(&postcopy_notifier_list);
57 void postcopy_add_notifier(NotifierWithReturn *nn)
59 notifier_with_return_list_add(&postcopy_notifier_list, nn);
62 void postcopy_remove_notifier(NotifierWithReturn *n)
64 notifier_with_return_remove(n);
67 int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp)
69 struct PostcopyNotifyData pnd;
70 pnd.reason = reason;
71 pnd.errp = errp;
73 return notifier_with_return_list_notify(&postcopy_notifier_list,
74 &pnd);
77 /* Postcopy needs to detect accesses to pages that haven't yet been copied
78 * across, and efficiently map new pages in, the techniques for doing this
79 * are target OS specific.
81 #if defined(__linux__)
83 #include <poll.h>
84 #include <sys/ioctl.h>
85 #include <sys/syscall.h>
86 #include <asm/types.h> /* for __u64 */
87 #endif
89 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
90 #include <sys/eventfd.h>
91 #include <linux/userfaultfd.h>
93 typedef struct PostcopyBlocktimeContext {
94 /* time when page fault initiated per vCPU */
95 uint32_t *page_fault_vcpu_time;
96 /* page address per vCPU */
97 uintptr_t *vcpu_addr;
98 uint32_t total_blocktime;
99 /* blocktime per vCPU */
100 uint32_t *vcpu_blocktime;
101 /* point in time when last page fault was initiated */
102 uint32_t last_begin;
103 /* number of vCPU are suspended */
104 int smp_cpus_down;
105 uint64_t start_time;
108 * Handler for exit event, necessary for
109 * releasing whole blocktime_ctx
111 Notifier exit_notifier;
112 } PostcopyBlocktimeContext;
114 static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx)
116 g_free(ctx->page_fault_vcpu_time);
117 g_free(ctx->vcpu_addr);
118 g_free(ctx->vcpu_blocktime);
119 g_free(ctx);
122 static void migration_exit_cb(Notifier *n, void *data)
124 PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext,
125 exit_notifier);
126 destroy_blocktime_context(ctx);
129 static struct PostcopyBlocktimeContext *blocktime_context_new(void)
131 PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1);
132 ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus);
133 ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus);
134 ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus);
136 ctx->exit_notifier.notify = migration_exit_cb;
137 ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
138 qemu_add_exit_notifier(&ctx->exit_notifier);
139 return ctx;
142 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx)
144 uint32List *list = NULL, *entry = NULL;
145 int i;
147 for (i = smp_cpus - 1; i >= 0; i--) {
148 entry = g_new0(uint32List, 1);
149 entry->value = ctx->vcpu_blocktime[i];
150 entry->next = list;
151 list = entry;
154 return list;
158 * This function just populates MigrationInfo from postcopy's
159 * blocktime context. It will not populate MigrationInfo,
160 * unless postcopy-blocktime capability was set.
162 * @info: pointer to MigrationInfo to populate
164 void fill_destination_postcopy_migration_info(MigrationInfo *info)
166 MigrationIncomingState *mis = migration_incoming_get_current();
167 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
169 if (!bc) {
170 return;
173 info->has_postcopy_blocktime = true;
174 info->postcopy_blocktime = bc->total_blocktime;
175 info->has_postcopy_vcpu_blocktime = true;
176 info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc);
179 static uint32_t get_postcopy_total_blocktime(void)
181 MigrationIncomingState *mis = migration_incoming_get_current();
182 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
184 if (!bc) {
185 return 0;
188 return bc->total_blocktime;
192 * receive_ufd_features: check userfault fd features, to request only supported
193 * features in the future.
195 * Returns: true on success
197 * __NR_userfaultfd - should be checked before
198 * @features: out parameter will contain uffdio_api.features provided by kernel
199 * in case of success
201 static bool receive_ufd_features(uint64_t *features)
203 struct uffdio_api api_struct = {0};
204 int ufd;
205 bool ret = true;
207 /* if we are here __NR_userfaultfd should exists */
208 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
209 if (ufd == -1) {
210 error_report("%s: syscall __NR_userfaultfd failed: %s", __func__,
211 strerror(errno));
212 return false;
215 /* ask features */
216 api_struct.api = UFFD_API;
217 api_struct.features = 0;
218 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
219 error_report("%s: UFFDIO_API failed: %s", __func__,
220 strerror(errno));
221 ret = false;
222 goto release_ufd;
225 *features = api_struct.features;
227 release_ufd:
228 close(ufd);
229 return ret;
233 * request_ufd_features: this function should be called only once on a newly
234 * opened ufd, subsequent calls will lead to error.
236 * Returns: true on succes
238 * @ufd: fd obtained from userfaultfd syscall
239 * @features: bit mask see UFFD_API_FEATURES
241 static bool request_ufd_features(int ufd, uint64_t features)
243 struct uffdio_api api_struct = {0};
244 uint64_t ioctl_mask;
246 api_struct.api = UFFD_API;
247 api_struct.features = features;
248 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
249 error_report("%s failed: UFFDIO_API failed: %s", __func__,
250 strerror(errno));
251 return false;
254 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
255 (__u64)1 << _UFFDIO_UNREGISTER;
256 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
257 error_report("Missing userfault features: %" PRIx64,
258 (uint64_t)(~api_struct.ioctls & ioctl_mask));
259 return false;
262 return true;
265 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis)
267 uint64_t asked_features = 0;
268 static uint64_t supported_features;
271 * it's not possible to
272 * request UFFD_API twice per one fd
273 * userfault fd features is persistent
275 if (!supported_features) {
276 if (!receive_ufd_features(&supported_features)) {
277 error_report("%s failed", __func__);
278 return false;
282 #ifdef UFFD_FEATURE_THREAD_ID
283 if (migrate_postcopy_blocktime() && mis &&
284 UFFD_FEATURE_THREAD_ID & supported_features) {
285 /* kernel supports that feature */
286 /* don't create blocktime_context if it exists */
287 if (!mis->blocktime_ctx) {
288 mis->blocktime_ctx = blocktime_context_new();
291 asked_features |= UFFD_FEATURE_THREAD_ID;
293 #endif
296 * request features, even if asked_features is 0, due to
297 * kernel expects UFFD_API before UFFDIO_REGISTER, per
298 * userfault file descriptor
300 if (!request_ufd_features(ufd, asked_features)) {
301 error_report("%s failed: features %" PRIu64, __func__,
302 asked_features);
303 return false;
306 if (getpagesize() != ram_pagesize_summary()) {
307 bool have_hp = false;
308 /* We've got a huge page */
309 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
310 have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS;
311 #endif
312 if (!have_hp) {
313 error_report("Userfault on this host does not support huge pages");
314 return false;
317 return true;
320 /* Callback from postcopy_ram_supported_by_host block iterator.
322 static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque)
324 const char *block_name = qemu_ram_get_idstr(rb);
325 ram_addr_t length = qemu_ram_get_used_length(rb);
326 size_t pagesize = qemu_ram_pagesize(rb);
328 if (length % pagesize) {
329 error_report("Postcopy requires RAM blocks to be a page size multiple,"
330 " block %s is 0x" RAM_ADDR_FMT " bytes with a "
331 "page size of 0x%zx", block_name, length, pagesize);
332 return 1;
334 return 0;
338 * Note: This has the side effect of munlock'ing all of RAM, that's
339 * normally fine since if the postcopy succeeds it gets turned back on at the
340 * end.
342 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
344 long pagesize = getpagesize();
345 int ufd = -1;
346 bool ret = false; /* Error unless we change it */
347 void *testarea = NULL;
348 struct uffdio_register reg_struct;
349 struct uffdio_range range_struct;
350 uint64_t feature_mask;
351 Error *local_err = NULL;
353 if (qemu_target_page_size() > pagesize) {
354 error_report("Target page size bigger than host page size");
355 goto out;
358 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
359 if (ufd == -1) {
360 error_report("%s: userfaultfd not available: %s", __func__,
361 strerror(errno));
362 goto out;
365 /* Give devices a chance to object */
366 if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) {
367 error_report_err(local_err);
368 goto out;
371 /* Version and features check */
372 if (!ufd_check_and_apply(ufd, mis)) {
373 goto out;
376 /* We don't support postcopy with shared RAM yet */
377 if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) {
378 goto out;
382 * userfault and mlock don't go together; we'll put it back later if
383 * it was enabled.
385 if (munlockall()) {
386 error_report("%s: munlockall: %s", __func__, strerror(errno));
387 return -1;
391 * We need to check that the ops we need are supported on anon memory
392 * To do that we need to register a chunk and see the flags that
393 * are returned.
395 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
396 MAP_ANONYMOUS, -1, 0);
397 if (testarea == MAP_FAILED) {
398 error_report("%s: Failed to map test area: %s", __func__,
399 strerror(errno));
400 goto out;
402 g_assert(((size_t)testarea & (pagesize-1)) == 0);
404 reg_struct.range.start = (uintptr_t)testarea;
405 reg_struct.range.len = pagesize;
406 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
408 if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
409 error_report("%s userfault register: %s", __func__, strerror(errno));
410 goto out;
413 range_struct.start = (uintptr_t)testarea;
414 range_struct.len = pagesize;
415 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
416 error_report("%s userfault unregister: %s", __func__, strerror(errno));
417 goto out;
420 feature_mask = (__u64)1 << _UFFDIO_WAKE |
421 (__u64)1 << _UFFDIO_COPY |
422 (__u64)1 << _UFFDIO_ZEROPAGE;
423 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
424 error_report("Missing userfault map features: %" PRIx64,
425 (uint64_t)(~reg_struct.ioctls & feature_mask));
426 goto out;
429 /* Success! */
430 ret = true;
431 out:
432 if (testarea) {
433 munmap(testarea, pagesize);
435 if (ufd != -1) {
436 close(ufd);
438 return ret;
442 * Setup an area of RAM so that it *can* be used for postcopy later; this
443 * must be done right at the start prior to pre-copy.
444 * opaque should be the MIS.
446 static int init_range(RAMBlock *rb, void *opaque)
448 const char *block_name = qemu_ram_get_idstr(rb);
449 void *host_addr = qemu_ram_get_host_addr(rb);
450 ram_addr_t offset = qemu_ram_get_offset(rb);
451 ram_addr_t length = qemu_ram_get_used_length(rb);
452 trace_postcopy_init_range(block_name, host_addr, offset, length);
455 * We need the whole of RAM to be truly empty for postcopy, so things
456 * like ROMs and any data tables built during init must be zero'd
457 * - we're going to get the copy from the source anyway.
458 * (Precopy will just overwrite this data, so doesn't need the discard)
460 if (ram_discard_range(block_name, 0, length)) {
461 return -1;
464 return 0;
468 * At the end of migration, undo the effects of init_range
469 * opaque should be the MIS.
471 static int cleanup_range(RAMBlock *rb, void *opaque)
473 const char *block_name = qemu_ram_get_idstr(rb);
474 void *host_addr = qemu_ram_get_host_addr(rb);
475 ram_addr_t offset = qemu_ram_get_offset(rb);
476 ram_addr_t length = qemu_ram_get_used_length(rb);
477 MigrationIncomingState *mis = opaque;
478 struct uffdio_range range_struct;
479 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
482 * We turned off hugepage for the precopy stage with postcopy enabled
483 * we can turn it back on now.
485 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
488 * We can also turn off userfault now since we should have all the
489 * pages. It can be useful to leave it on to debug postcopy
490 * if you're not sure it's always getting every page.
492 range_struct.start = (uintptr_t)host_addr;
493 range_struct.len = length;
495 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
496 error_report("%s: userfault unregister %s", __func__, strerror(errno));
498 return -1;
501 return 0;
505 * Initialise postcopy-ram, setting the RAM to a state where we can go into
506 * postcopy later; must be called prior to any precopy.
507 * called from arch_init's similarly named ram_postcopy_incoming_init
509 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
511 if (foreach_not_ignored_block(init_range, NULL)) {
512 return -1;
515 return 0;
519 * Manage a single vote to the QEMU balloon inhibitor for all postcopy usage,
520 * last caller wins.
522 static void postcopy_balloon_inhibit(bool state)
524 static bool cur_state = false;
526 if (state != cur_state) {
527 qemu_balloon_inhibit(state);
528 cur_state = state;
533 * At the end of a migration where postcopy_ram_incoming_init was called.
535 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
537 trace_postcopy_ram_incoming_cleanup_entry();
539 if (mis->have_fault_thread) {
540 Error *local_err = NULL;
542 /* Let the fault thread quit */
543 atomic_set(&mis->fault_thread_quit, 1);
544 postcopy_fault_thread_notify(mis);
545 trace_postcopy_ram_incoming_cleanup_join();
546 qemu_thread_join(&mis->fault_thread);
548 if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
549 error_report_err(local_err);
550 return -1;
553 if (foreach_not_ignored_block(cleanup_range, mis)) {
554 return -1;
557 trace_postcopy_ram_incoming_cleanup_closeuf();
558 close(mis->userfault_fd);
559 close(mis->userfault_event_fd);
560 mis->have_fault_thread = false;
563 postcopy_balloon_inhibit(false);
565 if (enable_mlock) {
566 if (os_mlock() < 0) {
567 error_report("mlock: %s", strerror(errno));
569 * It doesn't feel right to fail at this point, we have a valid
570 * VM state.
575 postcopy_state_set(POSTCOPY_INCOMING_END);
577 if (mis->postcopy_tmp_page) {
578 munmap(mis->postcopy_tmp_page, mis->largest_page_size);
579 mis->postcopy_tmp_page = NULL;
581 if (mis->postcopy_tmp_zero_page) {
582 munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
583 mis->postcopy_tmp_zero_page = NULL;
585 trace_postcopy_ram_incoming_cleanup_blocktime(
586 get_postcopy_total_blocktime());
588 trace_postcopy_ram_incoming_cleanup_exit();
589 return 0;
593 * Disable huge pages on an area
595 static int nhp_range(RAMBlock *rb, void *opaque)
597 const char *block_name = qemu_ram_get_idstr(rb);
598 void *host_addr = qemu_ram_get_host_addr(rb);
599 ram_addr_t offset = qemu_ram_get_offset(rb);
600 ram_addr_t length = qemu_ram_get_used_length(rb);
601 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
604 * Before we do discards we need to ensure those discards really
605 * do delete areas of the page, even if THP thinks a hugepage would
606 * be a good idea, so force hugepages off.
608 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
610 return 0;
614 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
615 * however leaving it until after precopy means that most of the precopy
616 * data is still THPd
618 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
620 if (foreach_not_ignored_block(nhp_range, mis)) {
621 return -1;
624 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
626 return 0;
630 * Mark the given area of RAM as requiring notification to unwritten areas
631 * Used as a callback on foreach_not_ignored_block.
632 * host_addr: Base of area to mark
633 * offset: Offset in the whole ram arena
634 * length: Length of the section
635 * opaque: MigrationIncomingState pointer
636 * Returns 0 on success
638 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
640 MigrationIncomingState *mis = opaque;
641 struct uffdio_register reg_struct;
643 reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
644 reg_struct.range.len = qemu_ram_get_used_length(rb);
645 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
647 /* Now tell our userfault_fd that it's responsible for this area */
648 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
649 error_report("%s userfault register: %s", __func__, strerror(errno));
650 return -1;
652 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
653 error_report("%s userfault: Region doesn't support COPY", __func__);
654 return -1;
656 if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) {
657 qemu_ram_set_uf_zeroable(rb);
660 return 0;
663 int postcopy_wake_shared(struct PostCopyFD *pcfd,
664 uint64_t client_addr,
665 RAMBlock *rb)
667 size_t pagesize = qemu_ram_pagesize(rb);
668 struct uffdio_range range;
669 int ret;
670 trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
671 range.start = client_addr & ~(pagesize - 1);
672 range.len = pagesize;
673 ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
674 if (ret) {
675 error_report("%s: Failed to wake: %zx in %s (%s)",
676 __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
677 strerror(errno));
679 return ret;
683 * Callback from shared fault handlers to ask for a page,
684 * the page must be specified by a RAMBlock and an offset in that rb
685 * Note: Only for use by shared fault handlers (in fault thread)
687 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
688 uint64_t client_addr, uint64_t rb_offset)
690 size_t pagesize = qemu_ram_pagesize(rb);
691 uint64_t aligned_rbo = rb_offset & ~(pagesize - 1);
692 MigrationIncomingState *mis = migration_incoming_get_current();
694 trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
695 rb_offset);
696 if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
697 trace_postcopy_request_shared_page_present(pcfd->idstr,
698 qemu_ram_get_idstr(rb), rb_offset);
699 return postcopy_wake_shared(pcfd, client_addr, rb);
701 if (rb != mis->last_rb) {
702 mis->last_rb = rb;
703 migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
704 aligned_rbo, pagesize);
705 } else {
706 /* Save some space */
707 migrate_send_rp_req_pages(mis, NULL, aligned_rbo, pagesize);
709 return 0;
712 static int get_mem_fault_cpu_index(uint32_t pid)
714 CPUState *cpu_iter;
716 CPU_FOREACH(cpu_iter) {
717 if (cpu_iter->thread_id == pid) {
718 trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
719 return cpu_iter->cpu_index;
722 trace_get_mem_fault_cpu_index(-1, pid);
723 return -1;
726 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
728 int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
729 dc->start_time;
730 return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
734 * This function is being called when pagefault occurs. It
735 * tracks down vCPU blocking time.
737 * @addr: faulted host virtual address
738 * @ptid: faulted process thread id
739 * @rb: ramblock appropriate to addr
741 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
742 RAMBlock *rb)
744 int cpu, already_received;
745 MigrationIncomingState *mis = migration_incoming_get_current();
746 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
747 uint32_t low_time_offset;
749 if (!dc || ptid == 0) {
750 return;
752 cpu = get_mem_fault_cpu_index(ptid);
753 if (cpu < 0) {
754 return;
757 low_time_offset = get_low_time_offset(dc);
758 if (dc->vcpu_addr[cpu] == 0) {
759 atomic_inc(&dc->smp_cpus_down);
762 atomic_xchg(&dc->last_begin, low_time_offset);
763 atomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
764 atomic_xchg(&dc->vcpu_addr[cpu], addr);
766 /* check it here, not at the begining of the function,
767 * due to, check could accur early than bitmap_set in
768 * qemu_ufd_copy_ioctl */
769 already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
770 if (already_received) {
771 atomic_xchg(&dc->vcpu_addr[cpu], 0);
772 atomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
773 atomic_dec(&dc->smp_cpus_down);
775 trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
776 cpu, already_received);
780 * This function just provide calculated blocktime per cpu and trace it.
781 * Total blocktime is calculated in mark_postcopy_blocktime_end.
784 * Assume we have 3 CPU
786 * S1 E1 S1 E1
787 * -----***********------------xxx***************------------------------> CPU1
789 * S2 E2
790 * ------------****************xxx---------------------------------------> CPU2
792 * S3 E3
793 * ------------------------****xxx********-------------------------------> CPU3
795 * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
796 * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
797 * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
798 * it's a part of total blocktime.
799 * S1 - here is last_begin
800 * Legend of the picture is following:
801 * * - means blocktime per vCPU
802 * x - means overlapped blocktime (total blocktime)
804 * @addr: host virtual address
806 static void mark_postcopy_blocktime_end(uintptr_t addr)
808 MigrationIncomingState *mis = migration_incoming_get_current();
809 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
810 int i, affected_cpu = 0;
811 bool vcpu_total_blocktime = false;
812 uint32_t read_vcpu_time, low_time_offset;
814 if (!dc) {
815 return;
818 low_time_offset = get_low_time_offset(dc);
819 /* lookup cpu, to clear it,
820 * that algorithm looks straighforward, but it's not
821 * optimal, more optimal algorithm is keeping tree or hash
822 * where key is address value is a list of */
823 for (i = 0; i < smp_cpus; i++) {
824 uint32_t vcpu_blocktime = 0;
826 read_vcpu_time = atomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
827 if (atomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
828 read_vcpu_time == 0) {
829 continue;
831 atomic_xchg(&dc->vcpu_addr[i], 0);
832 vcpu_blocktime = low_time_offset - read_vcpu_time;
833 affected_cpu += 1;
834 /* we need to know is that mark_postcopy_end was due to
835 * faulted page, another possible case it's prefetched
836 * page and in that case we shouldn't be here */
837 if (!vcpu_total_blocktime &&
838 atomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
839 vcpu_total_blocktime = true;
841 /* continue cycle, due to one page could affect several vCPUs */
842 dc->vcpu_blocktime[i] += vcpu_blocktime;
845 atomic_sub(&dc->smp_cpus_down, affected_cpu);
846 if (vcpu_total_blocktime) {
847 dc->total_blocktime += low_time_offset - atomic_fetch_add(
848 &dc->last_begin, 0);
850 trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
851 affected_cpu);
854 static bool postcopy_pause_fault_thread(MigrationIncomingState *mis)
856 trace_postcopy_pause_fault_thread();
858 qemu_sem_wait(&mis->postcopy_pause_sem_fault);
860 trace_postcopy_pause_fault_thread_continued();
862 return true;
866 * Handle faults detected by the USERFAULT markings
868 static void *postcopy_ram_fault_thread(void *opaque)
870 MigrationIncomingState *mis = opaque;
871 struct uffd_msg msg;
872 int ret;
873 size_t index;
874 RAMBlock *rb = NULL;
876 trace_postcopy_ram_fault_thread_entry();
877 rcu_register_thread();
878 mis->last_rb = NULL; /* last RAMBlock we sent part of */
879 qemu_sem_post(&mis->fault_thread_sem);
881 struct pollfd *pfd;
882 size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
884 pfd = g_new0(struct pollfd, pfd_len);
886 pfd[0].fd = mis->userfault_fd;
887 pfd[0].events = POLLIN;
888 pfd[1].fd = mis->userfault_event_fd;
889 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
890 trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
891 for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
892 struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
893 struct PostCopyFD, index);
894 pfd[2 + index].fd = pcfd->fd;
895 pfd[2 + index].events = POLLIN;
896 trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
897 pcfd->fd);
900 while (true) {
901 ram_addr_t rb_offset;
902 int poll_result;
905 * We're mainly waiting for the kernel to give us a faulting HVA,
906 * however we can be told to quit via userfault_quit_fd which is
907 * an eventfd
910 poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
911 if (poll_result == -1) {
912 error_report("%s: userfault poll: %s", __func__, strerror(errno));
913 break;
916 if (!mis->to_src_file) {
918 * Possibly someone tells us that the return path is
919 * broken already using the event. We should hold until
920 * the channel is rebuilt.
922 if (postcopy_pause_fault_thread(mis)) {
923 mis->last_rb = NULL;
924 /* Continue to read the userfaultfd */
925 } else {
926 error_report("%s: paused but don't allow to continue",
927 __func__);
928 break;
932 if (pfd[1].revents) {
933 uint64_t tmp64 = 0;
935 /* Consume the signal */
936 if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
937 /* Nothing obviously nicer than posting this error. */
938 error_report("%s: read() failed", __func__);
941 if (atomic_read(&mis->fault_thread_quit)) {
942 trace_postcopy_ram_fault_thread_quit();
943 break;
947 if (pfd[0].revents) {
948 poll_result--;
949 ret = read(mis->userfault_fd, &msg, sizeof(msg));
950 if (ret != sizeof(msg)) {
951 if (errno == EAGAIN) {
953 * if a wake up happens on the other thread just after
954 * the poll, there is nothing to read.
956 continue;
958 if (ret < 0) {
959 error_report("%s: Failed to read full userfault "
960 "message: %s",
961 __func__, strerror(errno));
962 break;
963 } else {
964 error_report("%s: Read %d bytes from userfaultfd "
965 "expected %zd",
966 __func__, ret, sizeof(msg));
967 break; /* Lost alignment, don't know what we'd read next */
970 if (msg.event != UFFD_EVENT_PAGEFAULT) {
971 error_report("%s: Read unexpected event %ud from userfaultfd",
972 __func__, msg.event);
973 continue; /* It's not a page fault, shouldn't happen */
976 rb = qemu_ram_block_from_host(
977 (void *)(uintptr_t)msg.arg.pagefault.address,
978 true, &rb_offset);
979 if (!rb) {
980 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
981 PRIx64, (uint64_t)msg.arg.pagefault.address);
982 break;
985 rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
986 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
987 qemu_ram_get_idstr(rb),
988 rb_offset,
989 msg.arg.pagefault.feat.ptid);
990 mark_postcopy_blocktime_begin(
991 (uintptr_t)(msg.arg.pagefault.address),
992 msg.arg.pagefault.feat.ptid, rb);
994 retry:
996 * Send the request to the source - we want to request one
997 * of our host page sizes (which is >= TPS)
999 if (rb != mis->last_rb) {
1000 mis->last_rb = rb;
1001 ret = migrate_send_rp_req_pages(mis,
1002 qemu_ram_get_idstr(rb),
1003 rb_offset,
1004 qemu_ram_pagesize(rb));
1005 } else {
1006 /* Save some space */
1007 ret = migrate_send_rp_req_pages(mis,
1008 NULL,
1009 rb_offset,
1010 qemu_ram_pagesize(rb));
1013 if (ret) {
1014 /* May be network failure, try to wait for recovery */
1015 if (ret == -EIO && postcopy_pause_fault_thread(mis)) {
1016 /* We got reconnected somehow, try to continue */
1017 mis->last_rb = NULL;
1018 goto retry;
1019 } else {
1020 /* This is a unavoidable fault */
1021 error_report("%s: migrate_send_rp_req_pages() get %d",
1022 __func__, ret);
1023 break;
1028 /* Now handle any requests from external processes on shared memory */
1029 /* TODO: May need to handle devices deregistering during postcopy */
1030 for (index = 2; index < pfd_len && poll_result; index++) {
1031 if (pfd[index].revents) {
1032 struct PostCopyFD *pcfd =
1033 &g_array_index(mis->postcopy_remote_fds,
1034 struct PostCopyFD, index - 2);
1036 poll_result--;
1037 if (pfd[index].revents & POLLERR) {
1038 error_report("%s: POLLERR on poll %zd fd=%d",
1039 __func__, index, pcfd->fd);
1040 pfd[index].events = 0;
1041 continue;
1044 ret = read(pcfd->fd, &msg, sizeof(msg));
1045 if (ret != sizeof(msg)) {
1046 if (errno == EAGAIN) {
1048 * if a wake up happens on the other thread just after
1049 * the poll, there is nothing to read.
1051 continue;
1053 if (ret < 0) {
1054 error_report("%s: Failed to read full userfault "
1055 "message: %s (shared) revents=%d",
1056 __func__, strerror(errno),
1057 pfd[index].revents);
1058 /*TODO: Could just disable this sharer */
1059 break;
1060 } else {
1061 error_report("%s: Read %d bytes from userfaultfd "
1062 "expected %zd (shared)",
1063 __func__, ret, sizeof(msg));
1064 /*TODO: Could just disable this sharer */
1065 break; /*Lost alignment,don't know what we'd read next*/
1068 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1069 error_report("%s: Read unexpected event %ud "
1070 "from userfaultfd (shared)",
1071 __func__, msg.event);
1072 continue; /* It's not a page fault, shouldn't happen */
1074 /* Call the device handler registered with us */
1075 ret = pcfd->handler(pcfd, &msg);
1076 if (ret) {
1077 error_report("%s: Failed to resolve shared fault on %zd/%s",
1078 __func__, index, pcfd->idstr);
1079 /* TODO: Fail? Disable this sharer? */
1084 rcu_unregister_thread();
1085 trace_postcopy_ram_fault_thread_exit();
1086 g_free(pfd);
1087 return NULL;
1090 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
1092 /* Open the fd for the kernel to give us userfaults */
1093 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1094 if (mis->userfault_fd == -1) {
1095 error_report("%s: Failed to open userfault fd: %s", __func__,
1096 strerror(errno));
1097 return -1;
1101 * Although the host check already tested the API, we need to
1102 * do the check again as an ABI handshake on the new fd.
1104 if (!ufd_check_and_apply(mis->userfault_fd, mis)) {
1105 return -1;
1108 /* Now an eventfd we use to tell the fault-thread to quit */
1109 mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1110 if (mis->userfault_event_fd == -1) {
1111 error_report("%s: Opening userfault_event_fd: %s", __func__,
1112 strerror(errno));
1113 close(mis->userfault_fd);
1114 return -1;
1117 qemu_sem_init(&mis->fault_thread_sem, 0);
1118 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
1119 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
1120 qemu_sem_wait(&mis->fault_thread_sem);
1121 qemu_sem_destroy(&mis->fault_thread_sem);
1122 mis->have_fault_thread = true;
1124 /* Mark so that we get notified of accesses to unwritten areas */
1125 if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1126 error_report("ram_block_enable_notify failed");
1127 return -1;
1131 * Ballooning can mark pages as absent while we're postcopying
1132 * that would cause false userfaults.
1134 postcopy_balloon_inhibit(true);
1136 trace_postcopy_ram_enable_notify();
1138 return 0;
1141 static int qemu_ufd_copy_ioctl(int userfault_fd, void *host_addr,
1142 void *from_addr, uint64_t pagesize, RAMBlock *rb)
1144 int ret;
1145 if (from_addr) {
1146 struct uffdio_copy copy_struct;
1147 copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1148 copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1149 copy_struct.len = pagesize;
1150 copy_struct.mode = 0;
1151 ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1152 } else {
1153 struct uffdio_zeropage zero_struct;
1154 zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1155 zero_struct.range.len = pagesize;
1156 zero_struct.mode = 0;
1157 ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1159 if (!ret) {
1160 ramblock_recv_bitmap_set_range(rb, host_addr,
1161 pagesize / qemu_target_page_size());
1162 mark_postcopy_blocktime_end((uintptr_t)host_addr);
1165 return ret;
1168 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1170 int i;
1171 MigrationIncomingState *mis = migration_incoming_get_current();
1172 GArray *pcrfds = mis->postcopy_remote_fds;
1174 for (i = 0; i < pcrfds->len; i++) {
1175 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1176 int ret = cur->waker(cur, rb, offset);
1177 if (ret) {
1178 return ret;
1181 return 0;
1185 * Place a host page (from) at (host) atomically
1186 * returns 0 on success
1188 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1189 RAMBlock *rb)
1191 size_t pagesize = qemu_ram_pagesize(rb);
1193 /* copy also acks to the kernel waking the stalled thread up
1194 * TODO: We can inhibit that ack and only do it if it was requested
1195 * which would be slightly cheaper, but we'd have to be careful
1196 * of the order of updating our page state.
1198 if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, from, pagesize, rb)) {
1199 int e = errno;
1200 error_report("%s: %s copy host: %p from: %p (size: %zd)",
1201 __func__, strerror(e), host, from, pagesize);
1203 return -e;
1206 trace_postcopy_place_page(host);
1207 return postcopy_notify_shared_wake(rb,
1208 qemu_ram_block_host_offset(rb, host));
1212 * Place a zero page at (host) atomically
1213 * returns 0 on success
1215 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1216 RAMBlock *rb)
1218 size_t pagesize = qemu_ram_pagesize(rb);
1219 trace_postcopy_place_page_zero(host);
1221 /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1222 * but it's not available for everything (e.g. hugetlbpages)
1224 if (qemu_ram_is_uf_zeroable(rb)) {
1225 if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, NULL, pagesize, rb)) {
1226 int e = errno;
1227 error_report("%s: %s zero host: %p",
1228 __func__, strerror(e), host);
1230 return -e;
1232 return postcopy_notify_shared_wake(rb,
1233 qemu_ram_block_host_offset(rb,
1234 host));
1235 } else {
1236 /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */
1237 if (!mis->postcopy_tmp_zero_page) {
1238 mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1239 PROT_READ | PROT_WRITE,
1240 MAP_PRIVATE | MAP_ANONYMOUS,
1241 -1, 0);
1242 if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1243 int e = errno;
1244 mis->postcopy_tmp_zero_page = NULL;
1245 error_report("%s: %s mapping large zero page",
1246 __func__, strerror(e));
1247 return -e;
1249 memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1251 return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page,
1252 rb);
1257 * Returns a target page of memory that can be mapped at a later point in time
1258 * using postcopy_place_page
1259 * The same address is used repeatedly, postcopy_place_page just takes the
1260 * backing page away.
1261 * Returns: Pointer to allocated page
1264 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
1266 if (!mis->postcopy_tmp_page) {
1267 mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
1268 PROT_READ | PROT_WRITE, MAP_PRIVATE |
1269 MAP_ANONYMOUS, -1, 0);
1270 if (mis->postcopy_tmp_page == MAP_FAILED) {
1271 mis->postcopy_tmp_page = NULL;
1272 error_report("%s: %s", __func__, strerror(errno));
1273 return NULL;
1277 return mis->postcopy_tmp_page;
1280 #else
1281 /* No target OS support, stubs just fail */
1282 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1286 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
1288 error_report("%s: No OS support", __func__);
1289 return false;
1292 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1294 error_report("postcopy_ram_incoming_init: No OS support");
1295 return -1;
1298 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1300 assert(0);
1301 return -1;
1304 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1306 assert(0);
1307 return -1;
1310 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1311 uint64_t client_addr, uint64_t rb_offset)
1313 assert(0);
1314 return -1;
1317 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
1319 assert(0);
1320 return -1;
1323 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1324 RAMBlock *rb)
1326 assert(0);
1327 return -1;
1330 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1331 RAMBlock *rb)
1333 assert(0);
1334 return -1;
1337 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
1339 assert(0);
1340 return NULL;
1343 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1344 uint64_t client_addr,
1345 RAMBlock *rb)
1347 assert(0);
1348 return -1;
1350 #endif
1352 /* ------------------------------------------------------------------------- */
1354 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1356 uint64_t tmp64 = 1;
1359 * Wakeup the fault_thread. It's an eventfd that should currently
1360 * be at 0, we're going to increment it to 1
1362 if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1363 /* Not much we can do here, but may as well report it */
1364 error_report("%s: incrementing failed: %s", __func__,
1365 strerror(errno));
1370 * postcopy_discard_send_init: Called at the start of each RAMBlock before
1371 * asking to discard individual ranges.
1373 * @ms: The current migration state.
1374 * @offset: the bitmap offset of the named RAMBlock in the migration
1375 * bitmap.
1376 * @name: RAMBlock that discards will operate on.
1378 * returns: a new PDS.
1380 PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms,
1381 const char *name)
1383 PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState));
1385 if (res) {
1386 res->ramblock_name = name;
1389 return res;
1393 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1394 * discard. May send a discard message, may just leave it queued to
1395 * be sent later.
1397 * @ms: Current migration state.
1398 * @pds: Structure initialised by postcopy_discard_send_init().
1399 * @start,@length: a range of pages in the migration bitmap in the
1400 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1402 void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds,
1403 unsigned long start, unsigned long length)
1405 size_t tp_size = qemu_target_page_size();
1406 /* Convert to byte offsets within the RAM block */
1407 pds->start_list[pds->cur_entry] = start * tp_size;
1408 pds->length_list[pds->cur_entry] = length * tp_size;
1409 trace_postcopy_discard_send_range(pds->ramblock_name, start, length);
1410 pds->cur_entry++;
1411 pds->nsentwords++;
1413 if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) {
1414 /* Full set, ship it! */
1415 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1416 pds->ramblock_name,
1417 pds->cur_entry,
1418 pds->start_list,
1419 pds->length_list);
1420 pds->nsentcmds++;
1421 pds->cur_entry = 0;
1426 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1427 * bitmap code. Sends any outstanding discard messages, frees the PDS
1429 * @ms: Current migration state.
1430 * @pds: Structure initialised by postcopy_discard_send_init().
1432 void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds)
1434 /* Anything unsent? */
1435 if (pds->cur_entry) {
1436 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1437 pds->ramblock_name,
1438 pds->cur_entry,
1439 pds->start_list,
1440 pds->length_list);
1441 pds->nsentcmds++;
1444 trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords,
1445 pds->nsentcmds);
1447 g_free(pds);
1451 * Current state of incoming postcopy; note this is not part of
1452 * MigrationIncomingState since it's state is used during cleanup
1453 * at the end as MIS is being freed.
1455 static PostcopyState incoming_postcopy_state;
1457 PostcopyState postcopy_state_get(void)
1459 return atomic_mb_read(&incoming_postcopy_state);
1462 /* Set the state and return the old state */
1463 PostcopyState postcopy_state_set(PostcopyState new_state)
1465 return atomic_xchg(&incoming_postcopy_state, new_state);
1468 /* Register a handler for external shared memory postcopy
1469 * called on the destination.
1471 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1473 MigrationIncomingState *mis = migration_incoming_get_current();
1475 mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1476 *pcfd);
1479 /* Unregister a handler for external shared memory postcopy
1481 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1483 guint i;
1484 MigrationIncomingState *mis = migration_incoming_get_current();
1485 GArray *pcrfds = mis->postcopy_remote_fds;
1487 for (i = 0; i < pcrfds->len; i++) {
1488 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1489 if (cur->fd == pcfd->fd) {
1490 mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1491 return;