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.
23 #include "qemu-common.h"
24 #include "migration/migration.h"
25 #include "migration/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;
38 uint64_t offset; /* Bitmap entry for the 1st bit of this RAMBlock */
41 * Start and length of a discard range (bytes)
43 uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
44 uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
45 unsigned int nsentwords;
46 unsigned int nsentcmds;
49 /* Postcopy needs to detect accesses to pages that haven't yet been copied
50 * across, and efficiently map new pages in, the techniques for doing this
51 * are target OS specific.
53 #if defined(__linux__)
56 #include <sys/eventfd.h>
58 #include <sys/ioctl.h>
59 #include <sys/syscall.h>
60 #include <sys/types.h>
61 #include <asm/types.h> /* for __u64 */
64 #if defined(__linux__) && defined(__NR_userfaultfd)
65 #include <linux/userfaultfd.h>
67 static bool ufd_version_check(int ufd)
69 struct uffdio_api api_struct;
72 api_struct.api = UFFD_API;
73 api_struct.features = 0;
74 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
75 error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
80 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
81 (__u64)1 << _UFFDIO_UNREGISTER;
82 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
83 error_report("Missing userfault features: %" PRIx64,
84 (uint64_t)(~api_struct.ioctls & ioctl_mask));
92 * Note: This has the side effect of munlock'ing all of RAM, that's
93 * normally fine since if the postcopy succeeds it gets turned back on at the
96 bool postcopy_ram_supported_by_host(void)
98 long pagesize = getpagesize();
100 bool ret = false; /* Error unless we change it */
101 void *testarea = NULL;
102 struct uffdio_register reg_struct;
103 struct uffdio_range range_struct;
104 uint64_t feature_mask;
106 if ((1ul << qemu_target_page_bits()) > pagesize) {
107 error_report("Target page size bigger than host page size");
111 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
113 error_report("%s: userfaultfd not available: %s", __func__,
118 /* Version and features check */
119 if (!ufd_version_check(ufd)) {
124 * userfault and mlock don't go together; we'll put it back later if
128 error_report("%s: munlockall: %s", __func__, strerror(errno));
133 * We need to check that the ops we need are supported on anon memory
134 * To do that we need to register a chunk and see the flags that
137 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
138 MAP_ANONYMOUS, -1, 0);
139 if (testarea == MAP_FAILED) {
140 error_report("%s: Failed to map test area: %s", __func__,
144 g_assert(((size_t)testarea & (pagesize-1)) == 0);
146 reg_struct.range.start = (uintptr_t)testarea;
147 reg_struct.range.len = pagesize;
148 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
150 if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) {
151 error_report("%s userfault register: %s", __func__, strerror(errno));
155 range_struct.start = (uintptr_t)testarea;
156 range_struct.len = pagesize;
157 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
158 error_report("%s userfault unregister: %s", __func__, strerror(errno));
162 feature_mask = (__u64)1 << _UFFDIO_WAKE |
163 (__u64)1 << _UFFDIO_COPY |
164 (__u64)1 << _UFFDIO_ZEROPAGE;
165 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
166 error_report("Missing userfault map features: %" PRIx64,
167 (uint64_t)(~reg_struct.ioctls & feature_mask));
175 munmap(testarea, pagesize);
184 * postcopy_ram_discard_range: Discard a range of memory.
185 * We can assume that if we've been called postcopy_ram_hosttest returned true.
187 * @mis: Current incoming migration state.
188 * @start, @length: range of memory to discard.
190 * returns: 0 on success.
192 int postcopy_ram_discard_range(MigrationIncomingState *mis, uint8_t *start,
195 trace_postcopy_ram_discard_range(start, length);
196 if (madvise(start, length, MADV_DONTNEED)) {
197 error_report("%s MADV_DONTNEED: %s", __func__, strerror(errno));
205 * Setup an area of RAM so that it *can* be used for postcopy later; this
206 * must be done right at the start prior to pre-copy.
207 * opaque should be the MIS.
209 static int init_range(const char *block_name, void *host_addr,
210 ram_addr_t offset, ram_addr_t length, void *opaque)
212 MigrationIncomingState *mis = opaque;
214 trace_postcopy_init_range(block_name, host_addr, offset, length);
217 * We need the whole of RAM to be truly empty for postcopy, so things
218 * like ROMs and any data tables built during init must be zero'd
219 * - we're going to get the copy from the source anyway.
220 * (Precopy will just overwrite this data, so doesn't need the discard)
222 if (postcopy_ram_discard_range(mis, host_addr, length)) {
230 * At the end of migration, undo the effects of init_range
231 * opaque should be the MIS.
233 static int cleanup_range(const char *block_name, void *host_addr,
234 ram_addr_t offset, ram_addr_t length, void *opaque)
236 MigrationIncomingState *mis = opaque;
237 struct uffdio_range range_struct;
238 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
241 * We turned off hugepage for the precopy stage with postcopy enabled
242 * we can turn it back on now.
244 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
247 * We can also turn off userfault now since we should have all the
248 * pages. It can be useful to leave it on to debug postcopy
249 * if you're not sure it's always getting every page.
251 range_struct.start = (uintptr_t)host_addr;
252 range_struct.len = length;
254 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
255 error_report("%s: userfault unregister %s", __func__, strerror(errno));
264 * Initialise postcopy-ram, setting the RAM to a state where we can go into
265 * postcopy later; must be called prior to any precopy.
266 * called from arch_init's similarly named ram_postcopy_incoming_init
268 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
270 if (qemu_ram_foreach_block(init_range, mis)) {
278 * At the end of a migration where postcopy_ram_incoming_init was called.
280 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
282 trace_postcopy_ram_incoming_cleanup_entry();
284 if (mis->have_fault_thread) {
287 if (qemu_ram_foreach_block(cleanup_range, mis)) {
291 * Tell the fault_thread to exit, it's an eventfd that should
292 * currently be at 0, we're going to increment it to 1
295 if (write(mis->userfault_quit_fd, &tmp64, 8) == 8) {
296 trace_postcopy_ram_incoming_cleanup_join();
297 qemu_thread_join(&mis->fault_thread);
299 /* Not much we can do here, but may as well report it */
300 error_report("%s: incrementing userfault_quit_fd: %s", __func__,
303 trace_postcopy_ram_incoming_cleanup_closeuf();
304 close(mis->userfault_fd);
305 close(mis->userfault_quit_fd);
306 mis->have_fault_thread = false;
309 qemu_balloon_inhibit(false);
312 if (os_mlock() < 0) {
313 error_report("mlock: %s", strerror(errno));
315 * It doesn't feel right to fail at this point, we have a valid
321 postcopy_state_set(POSTCOPY_INCOMING_END);
322 migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
324 if (mis->postcopy_tmp_page) {
325 munmap(mis->postcopy_tmp_page, getpagesize());
326 mis->postcopy_tmp_page = NULL;
328 trace_postcopy_ram_incoming_cleanup_exit();
333 * Disable huge pages on an area
335 static int nhp_range(const char *block_name, void *host_addr,
336 ram_addr_t offset, ram_addr_t length, void *opaque)
338 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
341 * Before we do discards we need to ensure those discards really
342 * do delete areas of the page, even if THP thinks a hugepage would
343 * be a good idea, so force hugepages off.
345 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
351 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
352 * however leaving it until after precopy means that most of the precopy
355 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
357 if (qemu_ram_foreach_block(nhp_range, mis)) {
361 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
367 * Mark the given area of RAM as requiring notification to unwritten areas
368 * Used as a callback on qemu_ram_foreach_block.
369 * host_addr: Base of area to mark
370 * offset: Offset in the whole ram arena
371 * length: Length of the section
372 * opaque: MigrationIncomingState pointer
373 * Returns 0 on success
375 static int ram_block_enable_notify(const char *block_name, void *host_addr,
376 ram_addr_t offset, ram_addr_t length,
379 MigrationIncomingState *mis = opaque;
380 struct uffdio_register reg_struct;
382 reg_struct.range.start = (uintptr_t)host_addr;
383 reg_struct.range.len = length;
384 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
386 /* Now tell our userfault_fd that it's responsible for this area */
387 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) {
388 error_report("%s userfault register: %s", __func__, strerror(errno));
396 * Handle faults detected by the USERFAULT markings
398 static void *postcopy_ram_fault_thread(void *opaque)
400 MigrationIncomingState *mis = opaque;
403 size_t hostpagesize = getpagesize();
405 RAMBlock *last_rb = NULL; /* last RAMBlock we sent part of */
407 trace_postcopy_ram_fault_thread_entry();
408 qemu_sem_post(&mis->fault_thread_sem);
411 ram_addr_t rb_offset;
412 ram_addr_t in_raspace;
413 struct pollfd pfd[2];
416 * We're mainly waiting for the kernel to give us a faulting HVA,
417 * however we can be told to quit via userfault_quit_fd which is
420 pfd[0].fd = mis->userfault_fd;
421 pfd[0].events = POLLIN;
423 pfd[1].fd = mis->userfault_quit_fd;
424 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
427 if (poll(pfd, 2, -1 /* Wait forever */) == -1) {
428 error_report("%s: userfault poll: %s", __func__, strerror(errno));
432 if (pfd[1].revents) {
433 trace_postcopy_ram_fault_thread_quit();
437 ret = read(mis->userfault_fd, &msg, sizeof(msg));
438 if (ret != sizeof(msg)) {
439 if (errno == EAGAIN) {
441 * if a wake up happens on the other thread just after
442 * the poll, there is nothing to read.
447 error_report("%s: Failed to read full userfault message: %s",
448 __func__, strerror(errno));
451 error_report("%s: Read %d bytes from userfaultfd expected %zd",
452 __func__, ret, sizeof(msg));
453 break; /* Lost alignment, don't know what we'd read next */
456 if (msg.event != UFFD_EVENT_PAGEFAULT) {
457 error_report("%s: Read unexpected event %ud from userfaultfd",
458 __func__, msg.event);
459 continue; /* It's not a page fault, shouldn't happen */
462 rb = qemu_ram_block_from_host(
463 (void *)(uintptr_t)msg.arg.pagefault.address,
464 true, &in_raspace, &rb_offset);
466 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
467 PRIx64, (uint64_t)msg.arg.pagefault.address);
471 rb_offset &= ~(hostpagesize - 1);
472 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
473 qemu_ram_get_idstr(rb),
477 * Send the request to the source - we want to request one
478 * of our host page sizes (which is >= TPS)
482 migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
483 rb_offset, hostpagesize);
485 /* Save some space */
486 migrate_send_rp_req_pages(mis, NULL,
487 rb_offset, hostpagesize);
490 trace_postcopy_ram_fault_thread_exit();
494 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
496 /* Open the fd for the kernel to give us userfaults */
497 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
498 if (mis->userfault_fd == -1) {
499 error_report("%s: Failed to open userfault fd: %s", __func__,
505 * Although the host check already tested the API, we need to
506 * do the check again as an ABI handshake on the new fd.
508 if (!ufd_version_check(mis->userfault_fd)) {
512 /* Now an eventfd we use to tell the fault-thread to quit */
513 mis->userfault_quit_fd = eventfd(0, EFD_CLOEXEC);
514 if (mis->userfault_quit_fd == -1) {
515 error_report("%s: Opening userfault_quit_fd: %s", __func__,
517 close(mis->userfault_fd);
521 qemu_sem_init(&mis->fault_thread_sem, 0);
522 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
523 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
524 qemu_sem_wait(&mis->fault_thread_sem);
525 qemu_sem_destroy(&mis->fault_thread_sem);
526 mis->have_fault_thread = true;
528 /* Mark so that we get notified of accesses to unwritten areas */
529 if (qemu_ram_foreach_block(ram_block_enable_notify, mis)) {
534 * Ballooning can mark pages as absent while we're postcopying
535 * that would cause false userfaults.
537 qemu_balloon_inhibit(true);
539 trace_postcopy_ram_enable_notify();
545 * Place a host page (from) at (host) atomically
546 * returns 0 on success
548 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from)
550 struct uffdio_copy copy_struct;
552 copy_struct.dst = (uint64_t)(uintptr_t)host;
553 copy_struct.src = (uint64_t)(uintptr_t)from;
554 copy_struct.len = getpagesize();
555 copy_struct.mode = 0;
557 /* copy also acks to the kernel waking the stalled thread up
558 * TODO: We can inhibit that ack and only do it if it was requested
559 * which would be slightly cheaper, but we'd have to be careful
560 * of the order of updating our page state.
562 if (ioctl(mis->userfault_fd, UFFDIO_COPY, ©_struct)) {
564 error_report("%s: %s copy host: %p from: %p",
565 __func__, strerror(e), host, from);
570 trace_postcopy_place_page(host);
575 * Place a zero page at (host) atomically
576 * returns 0 on success
578 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host)
580 struct uffdio_zeropage zero_struct;
582 zero_struct.range.start = (uint64_t)(uintptr_t)host;
583 zero_struct.range.len = getpagesize();
584 zero_struct.mode = 0;
586 if (ioctl(mis->userfault_fd, UFFDIO_ZEROPAGE, &zero_struct)) {
588 error_report("%s: %s zero host: %p",
589 __func__, strerror(e), host);
594 trace_postcopy_place_page_zero(host);
599 * Returns a target page of memory that can be mapped at a later point in time
600 * using postcopy_place_page
601 * The same address is used repeatedly, postcopy_place_page just takes the
603 * Returns: Pointer to allocated page
606 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
608 if (!mis->postcopy_tmp_page) {
609 mis->postcopy_tmp_page = mmap(NULL, getpagesize(),
610 PROT_READ | PROT_WRITE, MAP_PRIVATE |
611 MAP_ANONYMOUS, -1, 0);
612 if (!mis->postcopy_tmp_page) {
613 error_report("%s: %s", __func__, strerror(errno));
618 return mis->postcopy_tmp_page;
622 /* No target OS support, stubs just fail */
623 bool postcopy_ram_supported_by_host(void)
625 error_report("%s: No OS support", __func__);
629 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
631 error_report("postcopy_ram_incoming_init: No OS support");
635 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
641 int postcopy_ram_discard_range(MigrationIncomingState *mis, uint8_t *start,
648 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
654 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
660 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from)
666 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host)
672 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
680 /* ------------------------------------------------------------------------- */
683 * postcopy_discard_send_init: Called at the start of each RAMBlock before
684 * asking to discard individual ranges.
686 * @ms: The current migration state.
687 * @offset: the bitmap offset of the named RAMBlock in the migration
689 * @name: RAMBlock that discards will operate on.
691 * returns: a new PDS.
693 PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms,
694 unsigned long offset,
697 PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState));
700 res->ramblock_name = name;
701 res->offset = offset;
708 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
709 * discard. May send a discard message, may just leave it queued to
712 * @ms: Current migration state.
713 * @pds: Structure initialised by postcopy_discard_send_init().
714 * @start,@length: a range of pages in the migration bitmap in the
715 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
717 void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds,
718 unsigned long start, unsigned long length)
720 size_t tp_bits = qemu_target_page_bits();
721 /* Convert to byte offsets within the RAM block */
722 pds->start_list[pds->cur_entry] = (start - pds->offset) << tp_bits;
723 pds->length_list[pds->cur_entry] = length << tp_bits;
724 trace_postcopy_discard_send_range(pds->ramblock_name, start, length);
728 if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) {
729 /* Full set, ship it! */
730 qemu_savevm_send_postcopy_ram_discard(ms->file, pds->ramblock_name,
740 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
741 * bitmap code. Sends any outstanding discard messages, frees the PDS
743 * @ms: Current migration state.
744 * @pds: Structure initialised by postcopy_discard_send_init().
746 void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds)
748 /* Anything unsent? */
749 if (pds->cur_entry) {
750 qemu_savevm_send_postcopy_ram_discard(ms->file, pds->ramblock_name,
757 trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords,