sm501: Ignore no-op blits
[qemu/ar7.git] / migration / postcopy-ram.c
bloba36402722b975bfd32a96d0d9c6ddc1f2af22f87
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 "sysemu/balloon.h"
31 #include "qemu/error-report.h"
32 #include "trace.h"
33 #include "hw/boards.h"
35 /* Arbitrary limit on size of each discard command,
36 * keeps them around ~200 bytes
38 #define MAX_DISCARDS_PER_COMMAND 12
40 struct PostcopyDiscardState {
41 const char *ramblock_name;
42 uint16_t cur_entry;
44 * Start and length of a discard range (bytes)
46 uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
47 uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
48 unsigned int nsentwords;
49 unsigned int nsentcmds;
52 static NotifierWithReturnList postcopy_notifier_list;
54 void postcopy_infrastructure_init(void)
56 notifier_with_return_list_init(&postcopy_notifier_list);
59 void postcopy_add_notifier(NotifierWithReturn *nn)
61 notifier_with_return_list_add(&postcopy_notifier_list, nn);
64 void postcopy_remove_notifier(NotifierWithReturn *n)
66 notifier_with_return_remove(n);
69 int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp)
71 struct PostcopyNotifyData pnd;
72 pnd.reason = reason;
73 pnd.errp = errp;
75 return notifier_with_return_list_notify(&postcopy_notifier_list,
76 &pnd);
79 /* Postcopy needs to detect accesses to pages that haven't yet been copied
80 * across, and efficiently map new pages in, the techniques for doing this
81 * are target OS specific.
83 #if defined(__linux__)
85 #include <poll.h>
86 #include <sys/ioctl.h>
87 #include <sys/syscall.h>
88 #include <asm/types.h> /* for __u64 */
89 #endif
91 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
92 #include <sys/eventfd.h>
93 #include <linux/userfaultfd.h>
95 typedef struct PostcopyBlocktimeContext {
96 /* time when page fault initiated per vCPU */
97 uint32_t *page_fault_vcpu_time;
98 /* page address per vCPU */
99 uintptr_t *vcpu_addr;
100 uint32_t total_blocktime;
101 /* blocktime per vCPU */
102 uint32_t *vcpu_blocktime;
103 /* point in time when last page fault was initiated */
104 uint32_t last_begin;
105 /* number of vCPU are suspended */
106 int smp_cpus_down;
107 uint64_t start_time;
110 * Handler for exit event, necessary for
111 * releasing whole blocktime_ctx
113 Notifier exit_notifier;
114 } PostcopyBlocktimeContext;
116 static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx)
118 g_free(ctx->page_fault_vcpu_time);
119 g_free(ctx->vcpu_addr);
120 g_free(ctx->vcpu_blocktime);
121 g_free(ctx);
124 static void migration_exit_cb(Notifier *n, void *data)
126 PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext,
127 exit_notifier);
128 destroy_blocktime_context(ctx);
131 static struct PostcopyBlocktimeContext *blocktime_context_new(void)
133 MachineState *ms = MACHINE(qdev_get_machine());
134 unsigned int smp_cpus = ms->smp.cpus;
135 PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1);
136 ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus);
137 ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus);
138 ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus);
140 ctx->exit_notifier.notify = migration_exit_cb;
141 ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
142 qemu_add_exit_notifier(&ctx->exit_notifier);
143 return ctx;
146 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx)
148 MachineState *ms = MACHINE(qdev_get_machine());
149 uint32List *list = NULL, *entry = NULL;
150 int i;
152 for (i = ms->smp.cpus - 1; i >= 0; i--) {
153 entry = g_new0(uint32List, 1);
154 entry->value = ctx->vcpu_blocktime[i];
155 entry->next = list;
156 list = entry;
159 return list;
163 * This function just populates MigrationInfo from postcopy's
164 * blocktime context. It will not populate MigrationInfo,
165 * unless postcopy-blocktime capability was set.
167 * @info: pointer to MigrationInfo to populate
169 void fill_destination_postcopy_migration_info(MigrationInfo *info)
171 MigrationIncomingState *mis = migration_incoming_get_current();
172 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
174 if (!bc) {
175 return;
178 info->has_postcopy_blocktime = true;
179 info->postcopy_blocktime = bc->total_blocktime;
180 info->has_postcopy_vcpu_blocktime = true;
181 info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc);
184 static uint32_t get_postcopy_total_blocktime(void)
186 MigrationIncomingState *mis = migration_incoming_get_current();
187 PostcopyBlocktimeContext *bc = mis->blocktime_ctx;
189 if (!bc) {
190 return 0;
193 return bc->total_blocktime;
197 * receive_ufd_features: check userfault fd features, to request only supported
198 * features in the future.
200 * Returns: true on success
202 * __NR_userfaultfd - should be checked before
203 * @features: out parameter will contain uffdio_api.features provided by kernel
204 * in case of success
206 static bool receive_ufd_features(uint64_t *features)
208 struct uffdio_api api_struct = {0};
209 int ufd;
210 bool ret = true;
212 /* if we are here __NR_userfaultfd should exists */
213 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
214 if (ufd == -1) {
215 error_report("%s: syscall __NR_userfaultfd failed: %s", __func__,
216 strerror(errno));
217 return false;
220 /* ask features */
221 api_struct.api = UFFD_API;
222 api_struct.features = 0;
223 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
224 error_report("%s: UFFDIO_API failed: %s", __func__,
225 strerror(errno));
226 ret = false;
227 goto release_ufd;
230 *features = api_struct.features;
232 release_ufd:
233 close(ufd);
234 return ret;
238 * request_ufd_features: this function should be called only once on a newly
239 * opened ufd, subsequent calls will lead to error.
241 * Returns: true on succes
243 * @ufd: fd obtained from userfaultfd syscall
244 * @features: bit mask see UFFD_API_FEATURES
246 static bool request_ufd_features(int ufd, uint64_t features)
248 struct uffdio_api api_struct = {0};
249 uint64_t ioctl_mask;
251 api_struct.api = UFFD_API;
252 api_struct.features = features;
253 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
254 error_report("%s failed: UFFDIO_API failed: %s", __func__,
255 strerror(errno));
256 return false;
259 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
260 (__u64)1 << _UFFDIO_UNREGISTER;
261 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
262 error_report("Missing userfault features: %" PRIx64,
263 (uint64_t)(~api_struct.ioctls & ioctl_mask));
264 return false;
267 return true;
270 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis)
272 uint64_t asked_features = 0;
273 static uint64_t supported_features;
276 * it's not possible to
277 * request UFFD_API twice per one fd
278 * userfault fd features is persistent
280 if (!supported_features) {
281 if (!receive_ufd_features(&supported_features)) {
282 error_report("%s failed", __func__);
283 return false;
287 #ifdef UFFD_FEATURE_THREAD_ID
288 if (migrate_postcopy_blocktime() && mis &&
289 UFFD_FEATURE_THREAD_ID & supported_features) {
290 /* kernel supports that feature */
291 /* don't create blocktime_context if it exists */
292 if (!mis->blocktime_ctx) {
293 mis->blocktime_ctx = blocktime_context_new();
296 asked_features |= UFFD_FEATURE_THREAD_ID;
298 #endif
301 * request features, even if asked_features is 0, due to
302 * kernel expects UFFD_API before UFFDIO_REGISTER, per
303 * userfault file descriptor
305 if (!request_ufd_features(ufd, asked_features)) {
306 error_report("%s failed: features %" PRIu64, __func__,
307 asked_features);
308 return false;
311 if (qemu_real_host_page_size != ram_pagesize_summary()) {
312 bool have_hp = false;
313 /* We've got a huge page */
314 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
315 have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS;
316 #endif
317 if (!have_hp) {
318 error_report("Userfault on this host does not support huge pages");
319 return false;
322 return true;
325 /* Callback from postcopy_ram_supported_by_host block iterator.
327 static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque)
329 const char *block_name = qemu_ram_get_idstr(rb);
330 ram_addr_t length = qemu_ram_get_used_length(rb);
331 size_t pagesize = qemu_ram_pagesize(rb);
333 if (length % pagesize) {
334 error_report("Postcopy requires RAM blocks to be a page size multiple,"
335 " block %s is 0x" RAM_ADDR_FMT " bytes with a "
336 "page size of 0x%zx", block_name, length, pagesize);
337 return 1;
339 return 0;
343 * Note: This has the side effect of munlock'ing all of RAM, that's
344 * normally fine since if the postcopy succeeds it gets turned back on at the
345 * end.
347 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
349 long pagesize = qemu_real_host_page_size;
350 int ufd = -1;
351 bool ret = false; /* Error unless we change it */
352 void *testarea = NULL;
353 struct uffdio_register reg_struct;
354 struct uffdio_range range_struct;
355 uint64_t feature_mask;
356 Error *local_err = NULL;
358 if (qemu_target_page_size() > pagesize) {
359 error_report("Target page size bigger than host page size");
360 goto out;
363 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
364 if (ufd == -1) {
365 error_report("%s: userfaultfd not available: %s", __func__,
366 strerror(errno));
367 goto out;
370 /* Give devices a chance to object */
371 if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) {
372 error_report_err(local_err);
373 goto out;
376 /* Version and features check */
377 if (!ufd_check_and_apply(ufd, mis)) {
378 goto out;
381 /* We don't support postcopy with shared RAM yet */
382 if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) {
383 goto out;
387 * userfault and mlock don't go together; we'll put it back later if
388 * it was enabled.
390 if (munlockall()) {
391 error_report("%s: munlockall: %s", __func__, strerror(errno));
392 return -1;
396 * We need to check that the ops we need are supported on anon memory
397 * To do that we need to register a chunk and see the flags that
398 * are returned.
400 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
401 MAP_ANONYMOUS, -1, 0);
402 if (testarea == MAP_FAILED) {
403 error_report("%s: Failed to map test area: %s", __func__,
404 strerror(errno));
405 goto out;
407 g_assert(((size_t)testarea & (pagesize-1)) == 0);
409 reg_struct.range.start = (uintptr_t)testarea;
410 reg_struct.range.len = pagesize;
411 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
413 if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
414 error_report("%s userfault register: %s", __func__, strerror(errno));
415 goto out;
418 range_struct.start = (uintptr_t)testarea;
419 range_struct.len = pagesize;
420 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
421 error_report("%s userfault unregister: %s", __func__, strerror(errno));
422 goto out;
425 feature_mask = (__u64)1 << _UFFDIO_WAKE |
426 (__u64)1 << _UFFDIO_COPY |
427 (__u64)1 << _UFFDIO_ZEROPAGE;
428 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
429 error_report("Missing userfault map features: %" PRIx64,
430 (uint64_t)(~reg_struct.ioctls & feature_mask));
431 goto out;
434 /* Success! */
435 ret = true;
436 out:
437 if (testarea) {
438 munmap(testarea, pagesize);
440 if (ufd != -1) {
441 close(ufd);
443 return ret;
447 * Setup an area of RAM so that it *can* be used for postcopy later; this
448 * must be done right at the start prior to pre-copy.
449 * opaque should be the MIS.
451 static int init_range(RAMBlock *rb, void *opaque)
453 const char *block_name = qemu_ram_get_idstr(rb);
454 void *host_addr = qemu_ram_get_host_addr(rb);
455 ram_addr_t offset = qemu_ram_get_offset(rb);
456 ram_addr_t length = qemu_ram_get_used_length(rb);
457 trace_postcopy_init_range(block_name, host_addr, offset, length);
460 * We need the whole of RAM to be truly empty for postcopy, so things
461 * like ROMs and any data tables built during init must be zero'd
462 * - we're going to get the copy from the source anyway.
463 * (Precopy will just overwrite this data, so doesn't need the discard)
465 if (ram_discard_range(block_name, 0, length)) {
466 return -1;
469 return 0;
473 * At the end of migration, undo the effects of init_range
474 * opaque should be the MIS.
476 static int cleanup_range(RAMBlock *rb, void *opaque)
478 const char *block_name = qemu_ram_get_idstr(rb);
479 void *host_addr = qemu_ram_get_host_addr(rb);
480 ram_addr_t offset = qemu_ram_get_offset(rb);
481 ram_addr_t length = qemu_ram_get_used_length(rb);
482 MigrationIncomingState *mis = opaque;
483 struct uffdio_range range_struct;
484 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
487 * We turned off hugepage for the precopy stage with postcopy enabled
488 * we can turn it back on now.
490 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
493 * We can also turn off userfault now since we should have all the
494 * pages. It can be useful to leave it on to debug postcopy
495 * if you're not sure it's always getting every page.
497 range_struct.start = (uintptr_t)host_addr;
498 range_struct.len = length;
500 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
501 error_report("%s: userfault unregister %s", __func__, strerror(errno));
503 return -1;
506 return 0;
510 * Initialise postcopy-ram, setting the RAM to a state where we can go into
511 * postcopy later; must be called prior to any precopy.
512 * called from arch_init's similarly named ram_postcopy_incoming_init
514 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
516 if (foreach_not_ignored_block(init_range, NULL)) {
517 return -1;
520 return 0;
524 * Manage a single vote to the QEMU balloon inhibitor for all postcopy usage,
525 * last caller wins.
527 static void postcopy_balloon_inhibit(bool state)
529 static bool cur_state = false;
531 if (state != cur_state) {
532 qemu_balloon_inhibit(state);
533 cur_state = state;
538 * At the end of a migration where postcopy_ram_incoming_init was called.
540 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
542 trace_postcopy_ram_incoming_cleanup_entry();
544 if (mis->have_fault_thread) {
545 Error *local_err = NULL;
547 /* Let the fault thread quit */
548 atomic_set(&mis->fault_thread_quit, 1);
549 postcopy_fault_thread_notify(mis);
550 trace_postcopy_ram_incoming_cleanup_join();
551 qemu_thread_join(&mis->fault_thread);
553 if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) {
554 error_report_err(local_err);
555 return -1;
558 if (foreach_not_ignored_block(cleanup_range, mis)) {
559 return -1;
562 trace_postcopy_ram_incoming_cleanup_closeuf();
563 close(mis->userfault_fd);
564 close(mis->userfault_event_fd);
565 mis->have_fault_thread = false;
568 postcopy_balloon_inhibit(false);
570 if (enable_mlock) {
571 if (os_mlock() < 0) {
572 error_report("mlock: %s", strerror(errno));
574 * It doesn't feel right to fail at this point, we have a valid
575 * VM state.
580 if (mis->postcopy_tmp_page) {
581 munmap(mis->postcopy_tmp_page, mis->largest_page_size);
582 mis->postcopy_tmp_page = NULL;
584 if (mis->postcopy_tmp_zero_page) {
585 munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
586 mis->postcopy_tmp_zero_page = NULL;
588 trace_postcopy_ram_incoming_cleanup_blocktime(
589 get_postcopy_total_blocktime());
591 trace_postcopy_ram_incoming_cleanup_exit();
592 return 0;
596 * Disable huge pages on an area
598 static int nhp_range(RAMBlock *rb, void *opaque)
600 const char *block_name = qemu_ram_get_idstr(rb);
601 void *host_addr = qemu_ram_get_host_addr(rb);
602 ram_addr_t offset = qemu_ram_get_offset(rb);
603 ram_addr_t length = qemu_ram_get_used_length(rb);
604 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
607 * Before we do discards we need to ensure those discards really
608 * do delete areas of the page, even if THP thinks a hugepage would
609 * be a good idea, so force hugepages off.
611 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
613 return 0;
617 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
618 * however leaving it until after precopy means that most of the precopy
619 * data is still THPd
621 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
623 if (foreach_not_ignored_block(nhp_range, mis)) {
624 return -1;
627 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
629 return 0;
633 * Mark the given area of RAM as requiring notification to unwritten areas
634 * Used as a callback on foreach_not_ignored_block.
635 * host_addr: Base of area to mark
636 * offset: Offset in the whole ram arena
637 * length: Length of the section
638 * opaque: MigrationIncomingState pointer
639 * Returns 0 on success
641 static int ram_block_enable_notify(RAMBlock *rb, void *opaque)
643 MigrationIncomingState *mis = opaque;
644 struct uffdio_register reg_struct;
646 reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
647 reg_struct.range.len = qemu_ram_get_used_length(rb);
648 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
650 /* Now tell our userfault_fd that it's responsible for this area */
651 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
652 error_report("%s userfault register: %s", __func__, strerror(errno));
653 return -1;
655 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
656 error_report("%s userfault: Region doesn't support COPY", __func__);
657 return -1;
659 if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) {
660 qemu_ram_set_uf_zeroable(rb);
663 return 0;
666 int postcopy_wake_shared(struct PostCopyFD *pcfd,
667 uint64_t client_addr,
668 RAMBlock *rb)
670 size_t pagesize = qemu_ram_pagesize(rb);
671 struct uffdio_range range;
672 int ret;
673 trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb));
674 range.start = client_addr & ~(pagesize - 1);
675 range.len = pagesize;
676 ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range);
677 if (ret) {
678 error_report("%s: Failed to wake: %zx in %s (%s)",
679 __func__, (size_t)client_addr, qemu_ram_get_idstr(rb),
680 strerror(errno));
682 return ret;
686 * Callback from shared fault handlers to ask for a page,
687 * the page must be specified by a RAMBlock and an offset in that rb
688 * Note: Only for use by shared fault handlers (in fault thread)
690 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
691 uint64_t client_addr, uint64_t rb_offset)
693 size_t pagesize = qemu_ram_pagesize(rb);
694 uint64_t aligned_rbo = rb_offset & ~(pagesize - 1);
695 MigrationIncomingState *mis = migration_incoming_get_current();
697 trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb),
698 rb_offset);
699 if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) {
700 trace_postcopy_request_shared_page_present(pcfd->idstr,
701 qemu_ram_get_idstr(rb), rb_offset);
702 return postcopy_wake_shared(pcfd, client_addr, rb);
704 if (rb != mis->last_rb) {
705 mis->last_rb = rb;
706 migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
707 aligned_rbo, pagesize);
708 } else {
709 /* Save some space */
710 migrate_send_rp_req_pages(mis, NULL, aligned_rbo, pagesize);
712 return 0;
715 static int get_mem_fault_cpu_index(uint32_t pid)
717 CPUState *cpu_iter;
719 CPU_FOREACH(cpu_iter) {
720 if (cpu_iter->thread_id == pid) {
721 trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid);
722 return cpu_iter->cpu_index;
725 trace_get_mem_fault_cpu_index(-1, pid);
726 return -1;
729 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc)
731 int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) -
732 dc->start_time;
733 return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX;
737 * This function is being called when pagefault occurs. It
738 * tracks down vCPU blocking time.
740 * @addr: faulted host virtual address
741 * @ptid: faulted process thread id
742 * @rb: ramblock appropriate to addr
744 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid,
745 RAMBlock *rb)
747 int cpu, already_received;
748 MigrationIncomingState *mis = migration_incoming_get_current();
749 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
750 uint32_t low_time_offset;
752 if (!dc || ptid == 0) {
753 return;
755 cpu = get_mem_fault_cpu_index(ptid);
756 if (cpu < 0) {
757 return;
760 low_time_offset = get_low_time_offset(dc);
761 if (dc->vcpu_addr[cpu] == 0) {
762 atomic_inc(&dc->smp_cpus_down);
765 atomic_xchg(&dc->last_begin, low_time_offset);
766 atomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset);
767 atomic_xchg(&dc->vcpu_addr[cpu], addr);
770 * check it here, not at the beginning of the function,
771 * due to, check could occur early than bitmap_set in
772 * qemu_ufd_copy_ioctl
774 already_received = ramblock_recv_bitmap_test(rb, (void *)addr);
775 if (already_received) {
776 atomic_xchg(&dc->vcpu_addr[cpu], 0);
777 atomic_xchg(&dc->page_fault_vcpu_time[cpu], 0);
778 atomic_dec(&dc->smp_cpus_down);
780 trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu],
781 cpu, already_received);
785 * This function just provide calculated blocktime per cpu and trace it.
786 * Total blocktime is calculated in mark_postcopy_blocktime_end.
789 * Assume we have 3 CPU
791 * S1 E1 S1 E1
792 * -----***********------------xxx***************------------------------> CPU1
794 * S2 E2
795 * ------------****************xxx---------------------------------------> CPU2
797 * S3 E3
798 * ------------------------****xxx********-------------------------------> CPU3
800 * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
801 * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3
802 * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 -
803 * it's a part of total blocktime.
804 * S1 - here is last_begin
805 * Legend of the picture is following:
806 * * - means blocktime per vCPU
807 * x - means overlapped blocktime (total blocktime)
809 * @addr: host virtual address
811 static void mark_postcopy_blocktime_end(uintptr_t addr)
813 MigrationIncomingState *mis = migration_incoming_get_current();
814 PostcopyBlocktimeContext *dc = mis->blocktime_ctx;
815 MachineState *ms = MACHINE(qdev_get_machine());
816 unsigned int smp_cpus = ms->smp.cpus;
817 int i, affected_cpu = 0;
818 bool vcpu_total_blocktime = false;
819 uint32_t read_vcpu_time, low_time_offset;
821 if (!dc) {
822 return;
825 low_time_offset = get_low_time_offset(dc);
826 /* lookup cpu, to clear it,
827 * that algorithm looks straighforward, but it's not
828 * optimal, more optimal algorithm is keeping tree or hash
829 * where key is address value is a list of */
830 for (i = 0; i < smp_cpus; i++) {
831 uint32_t vcpu_blocktime = 0;
833 read_vcpu_time = atomic_fetch_add(&dc->page_fault_vcpu_time[i], 0);
834 if (atomic_fetch_add(&dc->vcpu_addr[i], 0) != addr ||
835 read_vcpu_time == 0) {
836 continue;
838 atomic_xchg(&dc->vcpu_addr[i], 0);
839 vcpu_blocktime = low_time_offset - read_vcpu_time;
840 affected_cpu += 1;
841 /* we need to know is that mark_postcopy_end was due to
842 * faulted page, another possible case it's prefetched
843 * page and in that case we shouldn't be here */
844 if (!vcpu_total_blocktime &&
845 atomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) {
846 vcpu_total_blocktime = true;
848 /* continue cycle, due to one page could affect several vCPUs */
849 dc->vcpu_blocktime[i] += vcpu_blocktime;
852 atomic_sub(&dc->smp_cpus_down, affected_cpu);
853 if (vcpu_total_blocktime) {
854 dc->total_blocktime += low_time_offset - atomic_fetch_add(
855 &dc->last_begin, 0);
857 trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime,
858 affected_cpu);
861 static bool postcopy_pause_fault_thread(MigrationIncomingState *mis)
863 trace_postcopy_pause_fault_thread();
865 qemu_sem_wait(&mis->postcopy_pause_sem_fault);
867 trace_postcopy_pause_fault_thread_continued();
869 return true;
873 * Handle faults detected by the USERFAULT markings
875 static void *postcopy_ram_fault_thread(void *opaque)
877 MigrationIncomingState *mis = opaque;
878 struct uffd_msg msg;
879 int ret;
880 size_t index;
881 RAMBlock *rb = NULL;
883 trace_postcopy_ram_fault_thread_entry();
884 rcu_register_thread();
885 mis->last_rb = NULL; /* last RAMBlock we sent part of */
886 qemu_sem_post(&mis->fault_thread_sem);
888 struct pollfd *pfd;
889 size_t pfd_len = 2 + mis->postcopy_remote_fds->len;
891 pfd = g_new0(struct pollfd, pfd_len);
893 pfd[0].fd = mis->userfault_fd;
894 pfd[0].events = POLLIN;
895 pfd[1].fd = mis->userfault_event_fd;
896 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
897 trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd);
898 for (index = 0; index < mis->postcopy_remote_fds->len; index++) {
899 struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds,
900 struct PostCopyFD, index);
901 pfd[2 + index].fd = pcfd->fd;
902 pfd[2 + index].events = POLLIN;
903 trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr,
904 pcfd->fd);
907 while (true) {
908 ram_addr_t rb_offset;
909 int poll_result;
912 * We're mainly waiting for the kernel to give us a faulting HVA,
913 * however we can be told to quit via userfault_quit_fd which is
914 * an eventfd
917 poll_result = poll(pfd, pfd_len, -1 /* Wait forever */);
918 if (poll_result == -1) {
919 error_report("%s: userfault poll: %s", __func__, strerror(errno));
920 break;
923 if (!mis->to_src_file) {
925 * Possibly someone tells us that the return path is
926 * broken already using the event. We should hold until
927 * the channel is rebuilt.
929 if (postcopy_pause_fault_thread(mis)) {
930 mis->last_rb = NULL;
931 /* Continue to read the userfaultfd */
932 } else {
933 error_report("%s: paused but don't allow to continue",
934 __func__);
935 break;
939 if (pfd[1].revents) {
940 uint64_t tmp64 = 0;
942 /* Consume the signal */
943 if (read(mis->userfault_event_fd, &tmp64, 8) != 8) {
944 /* Nothing obviously nicer than posting this error. */
945 error_report("%s: read() failed", __func__);
948 if (atomic_read(&mis->fault_thread_quit)) {
949 trace_postcopy_ram_fault_thread_quit();
950 break;
954 if (pfd[0].revents) {
955 poll_result--;
956 ret = read(mis->userfault_fd, &msg, sizeof(msg));
957 if (ret != sizeof(msg)) {
958 if (errno == EAGAIN) {
960 * if a wake up happens on the other thread just after
961 * the poll, there is nothing to read.
963 continue;
965 if (ret < 0) {
966 error_report("%s: Failed to read full userfault "
967 "message: %s",
968 __func__, strerror(errno));
969 break;
970 } else {
971 error_report("%s: Read %d bytes from userfaultfd "
972 "expected %zd",
973 __func__, ret, sizeof(msg));
974 break; /* Lost alignment, don't know what we'd read next */
977 if (msg.event != UFFD_EVENT_PAGEFAULT) {
978 error_report("%s: Read unexpected event %ud from userfaultfd",
979 __func__, msg.event);
980 continue; /* It's not a page fault, shouldn't happen */
983 rb = qemu_ram_block_from_host(
984 (void *)(uintptr_t)msg.arg.pagefault.address,
985 true, &rb_offset);
986 if (!rb) {
987 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
988 PRIx64, (uint64_t)msg.arg.pagefault.address);
989 break;
992 rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
993 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
994 qemu_ram_get_idstr(rb),
995 rb_offset,
996 msg.arg.pagefault.feat.ptid);
997 mark_postcopy_blocktime_begin(
998 (uintptr_t)(msg.arg.pagefault.address),
999 msg.arg.pagefault.feat.ptid, rb);
1001 retry:
1003 * Send the request to the source - we want to request one
1004 * of our host page sizes (which is >= TPS)
1006 if (rb != mis->last_rb) {
1007 mis->last_rb = rb;
1008 ret = migrate_send_rp_req_pages(mis,
1009 qemu_ram_get_idstr(rb),
1010 rb_offset,
1011 qemu_ram_pagesize(rb));
1012 } else {
1013 /* Save some space */
1014 ret = migrate_send_rp_req_pages(mis,
1015 NULL,
1016 rb_offset,
1017 qemu_ram_pagesize(rb));
1020 if (ret) {
1021 /* May be network failure, try to wait for recovery */
1022 if (ret == -EIO && postcopy_pause_fault_thread(mis)) {
1023 /* We got reconnected somehow, try to continue */
1024 mis->last_rb = NULL;
1025 goto retry;
1026 } else {
1027 /* This is a unavoidable fault */
1028 error_report("%s: migrate_send_rp_req_pages() get %d",
1029 __func__, ret);
1030 break;
1035 /* Now handle any requests from external processes on shared memory */
1036 /* TODO: May need to handle devices deregistering during postcopy */
1037 for (index = 2; index < pfd_len && poll_result; index++) {
1038 if (pfd[index].revents) {
1039 struct PostCopyFD *pcfd =
1040 &g_array_index(mis->postcopy_remote_fds,
1041 struct PostCopyFD, index - 2);
1043 poll_result--;
1044 if (pfd[index].revents & POLLERR) {
1045 error_report("%s: POLLERR on poll %zd fd=%d",
1046 __func__, index, pcfd->fd);
1047 pfd[index].events = 0;
1048 continue;
1051 ret = read(pcfd->fd, &msg, sizeof(msg));
1052 if (ret != sizeof(msg)) {
1053 if (errno == EAGAIN) {
1055 * if a wake up happens on the other thread just after
1056 * the poll, there is nothing to read.
1058 continue;
1060 if (ret < 0) {
1061 error_report("%s: Failed to read full userfault "
1062 "message: %s (shared) revents=%d",
1063 __func__, strerror(errno),
1064 pfd[index].revents);
1065 /*TODO: Could just disable this sharer */
1066 break;
1067 } else {
1068 error_report("%s: Read %d bytes from userfaultfd "
1069 "expected %zd (shared)",
1070 __func__, ret, sizeof(msg));
1071 /*TODO: Could just disable this sharer */
1072 break; /*Lost alignment,don't know what we'd read next*/
1075 if (msg.event != UFFD_EVENT_PAGEFAULT) {
1076 error_report("%s: Read unexpected event %ud "
1077 "from userfaultfd (shared)",
1078 __func__, msg.event);
1079 continue; /* It's not a page fault, shouldn't happen */
1081 /* Call the device handler registered with us */
1082 ret = pcfd->handler(pcfd, &msg);
1083 if (ret) {
1084 error_report("%s: Failed to resolve shared fault on %zd/%s",
1085 __func__, index, pcfd->idstr);
1086 /* TODO: Fail? Disable this sharer? */
1091 rcu_unregister_thread();
1092 trace_postcopy_ram_fault_thread_exit();
1093 g_free(pfd);
1094 return NULL;
1097 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1099 /* Open the fd for the kernel to give us userfaults */
1100 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
1101 if (mis->userfault_fd == -1) {
1102 error_report("%s: Failed to open userfault fd: %s", __func__,
1103 strerror(errno));
1104 return -1;
1108 * Although the host check already tested the API, we need to
1109 * do the check again as an ABI handshake on the new fd.
1111 if (!ufd_check_and_apply(mis->userfault_fd, mis)) {
1112 return -1;
1115 /* Now an eventfd we use to tell the fault-thread to quit */
1116 mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC);
1117 if (mis->userfault_event_fd == -1) {
1118 error_report("%s: Opening userfault_event_fd: %s", __func__,
1119 strerror(errno));
1120 close(mis->userfault_fd);
1121 return -1;
1124 qemu_sem_init(&mis->fault_thread_sem, 0);
1125 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
1126 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
1127 qemu_sem_wait(&mis->fault_thread_sem);
1128 qemu_sem_destroy(&mis->fault_thread_sem);
1129 mis->have_fault_thread = true;
1131 /* Mark so that we get notified of accesses to unwritten areas */
1132 if (foreach_not_ignored_block(ram_block_enable_notify, mis)) {
1133 error_report("ram_block_enable_notify failed");
1134 return -1;
1137 mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
1138 PROT_READ | PROT_WRITE, MAP_PRIVATE |
1139 MAP_ANONYMOUS, -1, 0);
1140 if (mis->postcopy_tmp_page == MAP_FAILED) {
1141 mis->postcopy_tmp_page = NULL;
1142 error_report("%s: Failed to map postcopy_tmp_page %s",
1143 __func__, strerror(errno));
1144 return -1;
1148 * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages
1150 mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
1151 PROT_READ | PROT_WRITE,
1152 MAP_PRIVATE | MAP_ANONYMOUS,
1153 -1, 0);
1154 if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
1155 int e = errno;
1156 mis->postcopy_tmp_zero_page = NULL;
1157 error_report("%s: Failed to map large zero page %s",
1158 __func__, strerror(e));
1159 return -e;
1161 memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
1164 * Ballooning can mark pages as absent while we're postcopying
1165 * that would cause false userfaults.
1167 postcopy_balloon_inhibit(true);
1169 trace_postcopy_ram_enable_notify();
1171 return 0;
1174 static int qemu_ufd_copy_ioctl(int userfault_fd, void *host_addr,
1175 void *from_addr, uint64_t pagesize, RAMBlock *rb)
1177 int ret;
1178 if (from_addr) {
1179 struct uffdio_copy copy_struct;
1180 copy_struct.dst = (uint64_t)(uintptr_t)host_addr;
1181 copy_struct.src = (uint64_t)(uintptr_t)from_addr;
1182 copy_struct.len = pagesize;
1183 copy_struct.mode = 0;
1184 ret = ioctl(userfault_fd, UFFDIO_COPY, &copy_struct);
1185 } else {
1186 struct uffdio_zeropage zero_struct;
1187 zero_struct.range.start = (uint64_t)(uintptr_t)host_addr;
1188 zero_struct.range.len = pagesize;
1189 zero_struct.mode = 0;
1190 ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct);
1192 if (!ret) {
1193 ramblock_recv_bitmap_set_range(rb, host_addr,
1194 pagesize / qemu_target_page_size());
1195 mark_postcopy_blocktime_end((uintptr_t)host_addr);
1198 return ret;
1201 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset)
1203 int i;
1204 MigrationIncomingState *mis = migration_incoming_get_current();
1205 GArray *pcrfds = mis->postcopy_remote_fds;
1207 for (i = 0; i < pcrfds->len; i++) {
1208 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1209 int ret = cur->waker(cur, rb, offset);
1210 if (ret) {
1211 return ret;
1214 return 0;
1218 * Place a host page (from) at (host) atomically
1219 * returns 0 on success
1221 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1222 RAMBlock *rb)
1224 size_t pagesize = qemu_ram_pagesize(rb);
1226 /* copy also acks to the kernel waking the stalled thread up
1227 * TODO: We can inhibit that ack and only do it if it was requested
1228 * which would be slightly cheaper, but we'd have to be careful
1229 * of the order of updating our page state.
1231 if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, from, pagesize, rb)) {
1232 int e = errno;
1233 error_report("%s: %s copy host: %p from: %p (size: %zd)",
1234 __func__, strerror(e), host, from, pagesize);
1236 return -e;
1239 trace_postcopy_place_page(host);
1240 return postcopy_notify_shared_wake(rb,
1241 qemu_ram_block_host_offset(rb, host));
1245 * Place a zero page at (host) atomically
1246 * returns 0 on success
1248 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1249 RAMBlock *rb)
1251 size_t pagesize = qemu_ram_pagesize(rb);
1252 trace_postcopy_place_page_zero(host);
1254 /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE
1255 * but it's not available for everything (e.g. hugetlbpages)
1257 if (qemu_ram_is_uf_zeroable(rb)) {
1258 if (qemu_ufd_copy_ioctl(mis->userfault_fd, host, NULL, pagesize, rb)) {
1259 int e = errno;
1260 error_report("%s: %s zero host: %p",
1261 __func__, strerror(e), host);
1263 return -e;
1265 return postcopy_notify_shared_wake(rb,
1266 qemu_ram_block_host_offset(rb,
1267 host));
1268 } else {
1269 return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb);
1273 #else
1274 /* No target OS support, stubs just fail */
1275 void fill_destination_postcopy_migration_info(MigrationInfo *info)
1279 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis)
1281 error_report("%s: No OS support", __func__);
1282 return false;
1285 int postcopy_ram_incoming_init(MigrationIncomingState *mis)
1287 error_report("postcopy_ram_incoming_init: No OS support");
1288 return -1;
1291 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
1293 assert(0);
1294 return -1;
1297 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
1299 assert(0);
1300 return -1;
1303 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb,
1304 uint64_t client_addr, uint64_t rb_offset)
1306 assert(0);
1307 return -1;
1310 int postcopy_ram_incoming_setup(MigrationIncomingState *mis)
1312 assert(0);
1313 return -1;
1316 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
1317 RAMBlock *rb)
1319 assert(0);
1320 return -1;
1323 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
1324 RAMBlock *rb)
1326 assert(0);
1327 return -1;
1330 int postcopy_wake_shared(struct PostCopyFD *pcfd,
1331 uint64_t client_addr,
1332 RAMBlock *rb)
1334 assert(0);
1335 return -1;
1337 #endif
1339 /* ------------------------------------------------------------------------- */
1341 void postcopy_fault_thread_notify(MigrationIncomingState *mis)
1343 uint64_t tmp64 = 1;
1346 * Wakeup the fault_thread. It's an eventfd that should currently
1347 * be at 0, we're going to increment it to 1
1349 if (write(mis->userfault_event_fd, &tmp64, 8) != 8) {
1350 /* Not much we can do here, but may as well report it */
1351 error_report("%s: incrementing failed: %s", __func__,
1352 strerror(errno));
1357 * postcopy_discard_send_init: Called at the start of each RAMBlock before
1358 * asking to discard individual ranges.
1360 * @ms: The current migration state.
1361 * @offset: the bitmap offset of the named RAMBlock in the migration bitmap.
1362 * @name: RAMBlock that discards will operate on.
1364 static PostcopyDiscardState pds = {0};
1365 void postcopy_discard_send_init(MigrationState *ms, const char *name)
1367 pds.ramblock_name = name;
1368 pds.cur_entry = 0;
1369 pds.nsentwords = 0;
1370 pds.nsentcmds = 0;
1374 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
1375 * discard. May send a discard message, may just leave it queued to
1376 * be sent later.
1378 * @ms: Current migration state.
1379 * @start,@length: a range of pages in the migration bitmap in the
1380 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
1382 void postcopy_discard_send_range(MigrationState *ms, unsigned long start,
1383 unsigned long length)
1385 size_t tp_size = qemu_target_page_size();
1386 /* Convert to byte offsets within the RAM block */
1387 pds.start_list[pds.cur_entry] = start * tp_size;
1388 pds.length_list[pds.cur_entry] = length * tp_size;
1389 trace_postcopy_discard_send_range(pds.ramblock_name, start, length);
1390 pds.cur_entry++;
1391 pds.nsentwords++;
1393 if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) {
1394 /* Full set, ship it! */
1395 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
1396 pds.ramblock_name,
1397 pds.cur_entry,
1398 pds.start_list,
1399 pds.length_list);
1400 pds.nsentcmds++;
1401 pds.cur_entry = 0;
1406 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
1407 * bitmap code. Sends any outstanding discard messages, frees the PDS
1409 * @ms: Current migration state.
1411 void postcopy_discard_send_finish(MigrationState *ms)
1413 /* Anything unsent? */
1414 if (pds.cur_entry) {
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++;
1423 trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords,
1424 pds.nsentcmds);
1428 * Current state of incoming postcopy; note this is not part of
1429 * MigrationIncomingState since it's state is used during cleanup
1430 * at the end as MIS is being freed.
1432 static PostcopyState incoming_postcopy_state;
1434 PostcopyState postcopy_state_get(void)
1436 return atomic_mb_read(&incoming_postcopy_state);
1439 /* Set the state and return the old state */
1440 PostcopyState postcopy_state_set(PostcopyState new_state)
1442 return atomic_xchg(&incoming_postcopy_state, new_state);
1445 /* Register a handler for external shared memory postcopy
1446 * called on the destination.
1448 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd)
1450 MigrationIncomingState *mis = migration_incoming_get_current();
1452 mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds,
1453 *pcfd);
1456 /* Unregister a handler for external shared memory postcopy
1458 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd)
1460 guint i;
1461 MigrationIncomingState *mis = migration_incoming_get_current();
1462 GArray *pcrfds = mis->postcopy_remote_fds;
1464 for (i = 0; i < pcrfds->len; i++) {
1465 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i);
1466 if (cur->fd == pcfd->fd) {
1467 mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i);
1468 return;