2 * Postcopy migration for RAM
4 * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates
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"
24 #include "postcopy-ram.h"
26 #include "sysemu/sysemu.h"
27 #include "sysemu/balloon.h"
28 #include "qemu/error-report.h"
31 /* Arbitrary limit on size of each discard command,
32 * keeps them around ~200 bytes
34 #define MAX_DISCARDS_PER_COMMAND 12
36 struct PostcopyDiscardState {
37 const char *ramblock_name;
40 * Start and length of a discard range (bytes)
42 uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
43 uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
44 unsigned int nsentwords;
45 unsigned int nsentcmds;
48 /* Postcopy needs to detect accesses to pages that haven't yet been copied
49 * across, and efficiently map new pages in, the techniques for doing this
50 * are target OS specific.
52 #if defined(__linux__)
55 #include <sys/ioctl.h>
56 #include <sys/syscall.h>
57 #include <asm/types.h> /* for __u64 */
60 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
61 #include <sys/eventfd.h>
62 #include <linux/userfaultfd.h>
64 static bool ufd_version_check(int ufd)
66 struct uffdio_api api_struct;
69 api_struct.api = UFFD_API;
70 api_struct.features = 0;
71 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
72 error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
77 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
78 (__u64)1 << _UFFDIO_UNREGISTER;
79 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
80 error_report("Missing userfault features: %" PRIx64,
81 (uint64_t)(~api_struct.ioctls & ioctl_mask));
85 if (getpagesize() != ram_pagesize_summary()) {
87 /* We've got a huge page */
88 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
89 have_hp = api_struct.features & UFFD_FEATURE_MISSING_HUGETLBFS;
92 error_report("Userfault on this host does not support huge pages");
99 /* Callback from postcopy_ram_supported_by_host block iterator.
101 static int test_ramblock_postcopiable(const char *block_name, void *host_addr,
102 ram_addr_t offset, ram_addr_t length, void *opaque)
104 RAMBlock *rb = qemu_ram_block_by_name(block_name);
105 size_t pagesize = qemu_ram_pagesize(rb);
107 if (qemu_ram_is_shared(rb)) {
108 error_report("Postcopy on shared RAM (%s) is not yet supported",
113 if (length % pagesize) {
114 error_report("Postcopy requires RAM blocks to be a page size multiple,"
115 " block %s is 0x" RAM_ADDR_FMT " bytes with a "
116 "page size of 0x%zx", block_name, length, pagesize);
123 * Note: This has the side effect of munlock'ing all of RAM, that's
124 * normally fine since if the postcopy succeeds it gets turned back on at the
127 bool postcopy_ram_supported_by_host(void)
129 long pagesize = getpagesize();
131 bool ret = false; /* Error unless we change it */
132 void *testarea = NULL;
133 struct uffdio_register reg_struct;
134 struct uffdio_range range_struct;
135 uint64_t feature_mask;
137 if (qemu_target_page_size() > pagesize) {
138 error_report("Target page size bigger than host page size");
142 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
144 error_report("%s: userfaultfd not available: %s", __func__,
149 /* Version and features check */
150 if (!ufd_version_check(ufd)) {
154 /* We don't support postcopy with shared RAM yet */
155 if (qemu_ram_foreach_block(test_ramblock_postcopiable, NULL)) {
160 * userfault and mlock don't go together; we'll put it back later if
164 error_report("%s: munlockall: %s", __func__, strerror(errno));
169 * We need to check that the ops we need are supported on anon memory
170 * To do that we need to register a chunk and see the flags that
173 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
174 MAP_ANONYMOUS, -1, 0);
175 if (testarea == MAP_FAILED) {
176 error_report("%s: Failed to map test area: %s", __func__,
180 g_assert(((size_t)testarea & (pagesize-1)) == 0);
182 reg_struct.range.start = (uintptr_t)testarea;
183 reg_struct.range.len = pagesize;
184 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
186 if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) {
187 error_report("%s userfault register: %s", __func__, strerror(errno));
191 range_struct.start = (uintptr_t)testarea;
192 range_struct.len = pagesize;
193 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
194 error_report("%s userfault unregister: %s", __func__, strerror(errno));
198 feature_mask = (__u64)1 << _UFFDIO_WAKE |
199 (__u64)1 << _UFFDIO_COPY |
200 (__u64)1 << _UFFDIO_ZEROPAGE;
201 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
202 error_report("Missing userfault map features: %" PRIx64,
203 (uint64_t)(~reg_struct.ioctls & feature_mask));
211 munmap(testarea, pagesize);
220 * Setup an area of RAM so that it *can* be used for postcopy later; this
221 * must be done right at the start prior to pre-copy.
222 * opaque should be the MIS.
224 static int init_range(const char *block_name, void *host_addr,
225 ram_addr_t offset, ram_addr_t length, void *opaque)
227 trace_postcopy_init_range(block_name, host_addr, offset, length);
230 * We need the whole of RAM to be truly empty for postcopy, so things
231 * like ROMs and any data tables built during init must be zero'd
232 * - we're going to get the copy from the source anyway.
233 * (Precopy will just overwrite this data, so doesn't need the discard)
235 if (ram_discard_range(block_name, 0, length)) {
243 * At the end of migration, undo the effects of init_range
244 * opaque should be the MIS.
246 static int cleanup_range(const char *block_name, void *host_addr,
247 ram_addr_t offset, ram_addr_t length, void *opaque)
249 MigrationIncomingState *mis = opaque;
250 struct uffdio_range range_struct;
251 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
254 * We turned off hugepage for the precopy stage with postcopy enabled
255 * we can turn it back on now.
257 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
260 * We can also turn off userfault now since we should have all the
261 * pages. It can be useful to leave it on to debug postcopy
262 * if you're not sure it's always getting every page.
264 range_struct.start = (uintptr_t)host_addr;
265 range_struct.len = length;
267 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
268 error_report("%s: userfault unregister %s", __func__, strerror(errno));
277 * Initialise postcopy-ram, setting the RAM to a state where we can go into
278 * postcopy later; must be called prior to any precopy.
279 * called from arch_init's similarly named ram_postcopy_incoming_init
281 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
283 if (qemu_ram_foreach_block(init_range, NULL)) {
291 * At the end of a migration where postcopy_ram_incoming_init was called.
293 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
295 trace_postcopy_ram_incoming_cleanup_entry();
297 if (mis->have_fault_thread) {
300 if (qemu_ram_foreach_block(cleanup_range, mis)) {
304 * Tell the fault_thread to exit, it's an eventfd that should
305 * currently be at 0, we're going to increment it to 1
308 if (write(mis->userfault_quit_fd, &tmp64, 8) == 8) {
309 trace_postcopy_ram_incoming_cleanup_join();
310 qemu_thread_join(&mis->fault_thread);
312 /* Not much we can do here, but may as well report it */
313 error_report("%s: incrementing userfault_quit_fd: %s", __func__,
316 trace_postcopy_ram_incoming_cleanup_closeuf();
317 close(mis->userfault_fd);
318 close(mis->userfault_quit_fd);
319 mis->have_fault_thread = false;
322 qemu_balloon_inhibit(false);
325 if (os_mlock() < 0) {
326 error_report("mlock: %s", strerror(errno));
328 * It doesn't feel right to fail at this point, we have a valid
334 postcopy_state_set(POSTCOPY_INCOMING_END);
336 if (mis->postcopy_tmp_page) {
337 munmap(mis->postcopy_tmp_page, mis->largest_page_size);
338 mis->postcopy_tmp_page = NULL;
340 if (mis->postcopy_tmp_zero_page) {
341 munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size);
342 mis->postcopy_tmp_zero_page = NULL;
344 trace_postcopy_ram_incoming_cleanup_exit();
349 * Disable huge pages on an area
351 static int nhp_range(const char *block_name, void *host_addr,
352 ram_addr_t offset, ram_addr_t length, void *opaque)
354 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
357 * Before we do discards we need to ensure those discards really
358 * do delete areas of the page, even if THP thinks a hugepage would
359 * be a good idea, so force hugepages off.
361 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
367 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
368 * however leaving it until after precopy means that most of the precopy
371 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
373 if (qemu_ram_foreach_block(nhp_range, mis)) {
377 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
383 * Mark the given area of RAM as requiring notification to unwritten areas
384 * Used as a callback on qemu_ram_foreach_block.
385 * host_addr: Base of area to mark
386 * offset: Offset in the whole ram arena
387 * length: Length of the section
388 * opaque: MigrationIncomingState pointer
389 * Returns 0 on success
391 static int ram_block_enable_notify(const char *block_name, void *host_addr,
392 ram_addr_t offset, ram_addr_t length,
395 MigrationIncomingState *mis = opaque;
396 struct uffdio_register reg_struct;
398 reg_struct.range.start = (uintptr_t)host_addr;
399 reg_struct.range.len = length;
400 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
402 /* Now tell our userfault_fd that it's responsible for this area */
403 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) {
404 error_report("%s userfault register: %s", __func__, strerror(errno));
407 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) {
408 error_report("%s userfault: Region doesn't support COPY", __func__);
416 * Handle faults detected by the USERFAULT markings
418 static void *postcopy_ram_fault_thread(void *opaque)
420 MigrationIncomingState *mis = opaque;
424 RAMBlock *last_rb = NULL; /* last RAMBlock we sent part of */
426 trace_postcopy_ram_fault_thread_entry();
427 qemu_sem_post(&mis->fault_thread_sem);
430 ram_addr_t rb_offset;
431 struct pollfd pfd[2];
434 * We're mainly waiting for the kernel to give us a faulting HVA,
435 * however we can be told to quit via userfault_quit_fd which is
438 pfd[0].fd = mis->userfault_fd;
439 pfd[0].events = POLLIN;
441 pfd[1].fd = mis->userfault_quit_fd;
442 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
445 if (poll(pfd, 2, -1 /* Wait forever */) == -1) {
446 error_report("%s: userfault poll: %s", __func__, strerror(errno));
450 if (pfd[1].revents) {
451 trace_postcopy_ram_fault_thread_quit();
455 ret = read(mis->userfault_fd, &msg, sizeof(msg));
456 if (ret != sizeof(msg)) {
457 if (errno == EAGAIN) {
459 * if a wake up happens on the other thread just after
460 * the poll, there is nothing to read.
465 error_report("%s: Failed to read full userfault message: %s",
466 __func__, strerror(errno));
469 error_report("%s: Read %d bytes from userfaultfd expected %zd",
470 __func__, ret, sizeof(msg));
471 break; /* Lost alignment, don't know what we'd read next */
474 if (msg.event != UFFD_EVENT_PAGEFAULT) {
475 error_report("%s: Read unexpected event %ud from userfaultfd",
476 __func__, msg.event);
477 continue; /* It's not a page fault, shouldn't happen */
480 rb = qemu_ram_block_from_host(
481 (void *)(uintptr_t)msg.arg.pagefault.address,
484 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
485 PRIx64, (uint64_t)msg.arg.pagefault.address);
489 rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
490 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
491 qemu_ram_get_idstr(rb),
495 * Send the request to the source - we want to request one
496 * of our host page sizes (which is >= TPS)
500 migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
501 rb_offset, qemu_ram_pagesize(rb));
503 /* Save some space */
504 migrate_send_rp_req_pages(mis, NULL,
505 rb_offset, qemu_ram_pagesize(rb));
508 trace_postcopy_ram_fault_thread_exit();
512 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
514 /* Open the fd for the kernel to give us userfaults */
515 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
516 if (mis->userfault_fd == -1) {
517 error_report("%s: Failed to open userfault fd: %s", __func__,
523 * Although the host check already tested the API, we need to
524 * do the check again as an ABI handshake on the new fd.
526 if (!ufd_version_check(mis->userfault_fd)) {
530 /* Now an eventfd we use to tell the fault-thread to quit */
531 mis->userfault_quit_fd = eventfd(0, EFD_CLOEXEC);
532 if (mis->userfault_quit_fd == -1) {
533 error_report("%s: Opening userfault_quit_fd: %s", __func__,
535 close(mis->userfault_fd);
539 qemu_sem_init(&mis->fault_thread_sem, 0);
540 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
541 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
542 qemu_sem_wait(&mis->fault_thread_sem);
543 qemu_sem_destroy(&mis->fault_thread_sem);
544 mis->have_fault_thread = true;
546 /* Mark so that we get notified of accesses to unwritten areas */
547 if (qemu_ram_foreach_block(ram_block_enable_notify, mis)) {
552 * Ballooning can mark pages as absent while we're postcopying
553 * that would cause false userfaults.
555 qemu_balloon_inhibit(true);
557 trace_postcopy_ram_enable_notify();
563 * Place a host page (from) at (host) atomically
564 * returns 0 on success
566 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
569 struct uffdio_copy copy_struct;
571 copy_struct.dst = (uint64_t)(uintptr_t)host;
572 copy_struct.src = (uint64_t)(uintptr_t)from;
573 copy_struct.len = pagesize;
574 copy_struct.mode = 0;
576 /* copy also acks to the kernel waking the stalled thread up
577 * TODO: We can inhibit that ack and only do it if it was requested
578 * which would be slightly cheaper, but we'd have to be careful
579 * of the order of updating our page state.
581 if (ioctl(mis->userfault_fd, UFFDIO_COPY, ©_struct)) {
583 error_report("%s: %s copy host: %p from: %p (size: %zd)",
584 __func__, strerror(e), host, from, pagesize);
589 trace_postcopy_place_page(host);
594 * Place a zero page at (host) atomically
595 * returns 0 on success
597 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
600 trace_postcopy_place_page_zero(host);
602 if (pagesize == getpagesize()) {
603 struct uffdio_zeropage zero_struct;
604 zero_struct.range.start = (uint64_t)(uintptr_t)host;
605 zero_struct.range.len = getpagesize();
606 zero_struct.mode = 0;
608 if (ioctl(mis->userfault_fd, UFFDIO_ZEROPAGE, &zero_struct)) {
610 error_report("%s: %s zero host: %p",
611 __func__, strerror(e), host);
616 /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */
617 if (!mis->postcopy_tmp_zero_page) {
618 mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size,
619 PROT_READ | PROT_WRITE,
620 MAP_PRIVATE | MAP_ANONYMOUS,
622 if (mis->postcopy_tmp_zero_page == MAP_FAILED) {
624 mis->postcopy_tmp_zero_page = NULL;
625 error_report("%s: %s mapping large zero page",
626 __func__, strerror(e));
629 memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size);
631 return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page,
639 * Returns a target page of memory that can be mapped at a later point in time
640 * using postcopy_place_page
641 * The same address is used repeatedly, postcopy_place_page just takes the
643 * Returns: Pointer to allocated page
646 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
648 if (!mis->postcopy_tmp_page) {
649 mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size,
650 PROT_READ | PROT_WRITE, MAP_PRIVATE |
651 MAP_ANONYMOUS, -1, 0);
652 if (mis->postcopy_tmp_page == MAP_FAILED) {
653 mis->postcopy_tmp_page = NULL;
654 error_report("%s: %s", __func__, strerror(errno));
659 return mis->postcopy_tmp_page;
663 /* No target OS support, stubs just fail */
664 bool postcopy_ram_supported_by_host(void)
666 error_report("%s: No OS support", __func__);
670 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
672 error_report("postcopy_ram_incoming_init: No OS support");
676 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
682 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
688 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
694 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from,
701 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host,
708 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
716 /* ------------------------------------------------------------------------- */
719 * postcopy_discard_send_init: Called at the start of each RAMBlock before
720 * asking to discard individual ranges.
722 * @ms: The current migration state.
723 * @offset: the bitmap offset of the named RAMBlock in the migration
725 * @name: RAMBlock that discards will operate on.
727 * returns: a new PDS.
729 PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms,
732 PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState));
735 res->ramblock_name = name;
742 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
743 * discard. May send a discard message, may just leave it queued to
746 * @ms: Current migration state.
747 * @pds: Structure initialised by postcopy_discard_send_init().
748 * @start,@length: a range of pages in the migration bitmap in the
749 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
751 void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds,
752 unsigned long start, unsigned long length)
754 size_t tp_size = qemu_target_page_size();
755 /* Convert to byte offsets within the RAM block */
756 pds->start_list[pds->cur_entry] = start * tp_size;
757 pds->length_list[pds->cur_entry] = length * tp_size;
758 trace_postcopy_discard_send_range(pds->ramblock_name, start, length);
762 if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) {
763 /* Full set, ship it! */
764 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
775 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
776 * bitmap code. Sends any outstanding discard messages, frees the PDS
778 * @ms: Current migration state.
779 * @pds: Structure initialised by postcopy_discard_send_init().
781 void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds)
783 /* Anything unsent? */
784 if (pds->cur_entry) {
785 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
793 trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords,
800 * Current state of incoming postcopy; note this is not part of
801 * MigrationIncomingState since it's state is used during cleanup
802 * at the end as MIS is being freed.
804 static PostcopyState incoming_postcopy_state;
806 PostcopyState postcopy_state_get(void)
808 return atomic_mb_read(&incoming_postcopy_state);
811 /* Set the state and return the old state */
812 PostcopyState postcopy_state_set(PostcopyState new_state)
814 return atomic_xchg(&incoming_postcopy_state, new_state);