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"
22 #include "qemu-common.h"
23 #include "migration/migration.h"
24 #include "migration/postcopy-ram.h"
25 #include "sysemu/sysemu.h"
26 #include "sysemu/balloon.h"
27 #include "qemu/error-report.h"
30 /* Arbitrary limit on size of each discard command,
31 * keeps them around ~200 bytes
33 #define MAX_DISCARDS_PER_COMMAND 12
35 struct PostcopyDiscardState {
36 const char *ramblock_name;
37 uint64_t offset; /* Bitmap entry for the 1st bit of this RAMBlock */
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__)
56 #include <sys/ioctl.h>
57 #include <sys/syscall.h>
58 #include <asm/types.h> /* for __u64 */
61 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
62 #include <sys/eventfd.h>
63 #include <linux/userfaultfd.h>
65 static bool ufd_version_check(int ufd)
67 struct uffdio_api api_struct;
70 api_struct.api = UFFD_API;
71 api_struct.features = 0;
72 if (ioctl(ufd, UFFDIO_API, &api_struct)) {
73 error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
78 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
79 (__u64)1 << _UFFDIO_UNREGISTER;
80 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
81 error_report("Missing userfault features: %" PRIx64,
82 (uint64_t)(~api_struct.ioctls & ioctl_mask));
90 * Note: This has the side effect of munlock'ing all of RAM, that's
91 * normally fine since if the postcopy succeeds it gets turned back on at the
94 bool postcopy_ram_supported_by_host(void)
96 long pagesize = getpagesize();
98 bool ret = false; /* Error unless we change it */
99 void *testarea = NULL;
100 struct uffdio_register reg_struct;
101 struct uffdio_range range_struct;
102 uint64_t feature_mask;
104 if ((1ul << qemu_target_page_bits()) > pagesize) {
105 error_report("Target page size bigger than host page size");
109 ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
111 error_report("%s: userfaultfd not available: %s", __func__,
116 /* Version and features check */
117 if (!ufd_version_check(ufd)) {
122 * userfault and mlock don't go together; we'll put it back later if
126 error_report("%s: munlockall: %s", __func__, strerror(errno));
131 * We need to check that the ops we need are supported on anon memory
132 * To do that we need to register a chunk and see the flags that
135 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
136 MAP_ANONYMOUS, -1, 0);
137 if (testarea == MAP_FAILED) {
138 error_report("%s: Failed to map test area: %s", __func__,
142 g_assert(((size_t)testarea & (pagesize-1)) == 0);
144 reg_struct.range.start = (uintptr_t)testarea;
145 reg_struct.range.len = pagesize;
146 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
148 if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) {
149 error_report("%s userfault register: %s", __func__, strerror(errno));
153 range_struct.start = (uintptr_t)testarea;
154 range_struct.len = pagesize;
155 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
156 error_report("%s userfault unregister: %s", __func__, strerror(errno));
160 feature_mask = (__u64)1 << _UFFDIO_WAKE |
161 (__u64)1 << _UFFDIO_COPY |
162 (__u64)1 << _UFFDIO_ZEROPAGE;
163 if ((reg_struct.ioctls & feature_mask) != feature_mask) {
164 error_report("Missing userfault map features: %" PRIx64,
165 (uint64_t)(~reg_struct.ioctls & feature_mask));
173 munmap(testarea, pagesize);
182 * postcopy_ram_discard_range: Discard a range of memory.
183 * We can assume that if we've been called postcopy_ram_hosttest returned true.
185 * @mis: Current incoming migration state.
186 * @start, @length: range of memory to discard.
188 * returns: 0 on success.
190 int postcopy_ram_discard_range(MigrationIncomingState *mis, uint8_t *start,
193 trace_postcopy_ram_discard_range(start, length);
194 if (madvise(start, length, MADV_DONTNEED)) {
195 error_report("%s MADV_DONTNEED: %s", __func__, strerror(errno));
203 * Setup an area of RAM so that it *can* be used for postcopy later; this
204 * must be done right at the start prior to pre-copy.
205 * opaque should be the MIS.
207 static int init_range(const char *block_name, void *host_addr,
208 ram_addr_t offset, ram_addr_t length, void *opaque)
210 MigrationIncomingState *mis = opaque;
212 trace_postcopy_init_range(block_name, host_addr, offset, length);
215 * We need the whole of RAM to be truly empty for postcopy, so things
216 * like ROMs and any data tables built during init must be zero'd
217 * - we're going to get the copy from the source anyway.
218 * (Precopy will just overwrite this data, so doesn't need the discard)
220 if (postcopy_ram_discard_range(mis, host_addr, length)) {
228 * At the end of migration, undo the effects of init_range
229 * opaque should be the MIS.
231 static int cleanup_range(const char *block_name, void *host_addr,
232 ram_addr_t offset, ram_addr_t length, void *opaque)
234 MigrationIncomingState *mis = opaque;
235 struct uffdio_range range_struct;
236 trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
239 * We turned off hugepage for the precopy stage with postcopy enabled
240 * we can turn it back on now.
242 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
245 * We can also turn off userfault now since we should have all the
246 * pages. It can be useful to leave it on to debug postcopy
247 * if you're not sure it's always getting every page.
249 range_struct.start = (uintptr_t)host_addr;
250 range_struct.len = length;
252 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
253 error_report("%s: userfault unregister %s", __func__, strerror(errno));
262 * Initialise postcopy-ram, setting the RAM to a state where we can go into
263 * postcopy later; must be called prior to any precopy.
264 * called from arch_init's similarly named ram_postcopy_incoming_init
266 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
268 if (qemu_ram_foreach_block(init_range, mis)) {
276 * At the end of a migration where postcopy_ram_incoming_init was called.
278 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
280 trace_postcopy_ram_incoming_cleanup_entry();
282 if (mis->have_fault_thread) {
285 if (qemu_ram_foreach_block(cleanup_range, mis)) {
289 * Tell the fault_thread to exit, it's an eventfd that should
290 * currently be at 0, we're going to increment it to 1
293 if (write(mis->userfault_quit_fd, &tmp64, 8) == 8) {
294 trace_postcopy_ram_incoming_cleanup_join();
295 qemu_thread_join(&mis->fault_thread);
297 /* Not much we can do here, but may as well report it */
298 error_report("%s: incrementing userfault_quit_fd: %s", __func__,
301 trace_postcopy_ram_incoming_cleanup_closeuf();
302 close(mis->userfault_fd);
303 close(mis->userfault_quit_fd);
304 mis->have_fault_thread = false;
307 qemu_balloon_inhibit(false);
310 if (os_mlock() < 0) {
311 error_report("mlock: %s", strerror(errno));
313 * It doesn't feel right to fail at this point, we have a valid
319 postcopy_state_set(POSTCOPY_INCOMING_END);
320 migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
322 if (mis->postcopy_tmp_page) {
323 munmap(mis->postcopy_tmp_page, getpagesize());
324 mis->postcopy_tmp_page = NULL;
326 trace_postcopy_ram_incoming_cleanup_exit();
331 * Disable huge pages on an area
333 static int nhp_range(const char *block_name, void *host_addr,
334 ram_addr_t offset, ram_addr_t length, void *opaque)
336 trace_postcopy_nhp_range(block_name, host_addr, offset, length);
339 * Before we do discards we need to ensure those discards really
340 * do delete areas of the page, even if THP thinks a hugepage would
341 * be a good idea, so force hugepages off.
343 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
349 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
350 * however leaving it until after precopy means that most of the precopy
353 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
355 if (qemu_ram_foreach_block(nhp_range, mis)) {
359 postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
365 * Mark the given area of RAM as requiring notification to unwritten areas
366 * Used as a callback on qemu_ram_foreach_block.
367 * host_addr: Base of area to mark
368 * offset: Offset in the whole ram arena
369 * length: Length of the section
370 * opaque: MigrationIncomingState pointer
371 * Returns 0 on success
373 static int ram_block_enable_notify(const char *block_name, void *host_addr,
374 ram_addr_t offset, ram_addr_t length,
377 MigrationIncomingState *mis = opaque;
378 struct uffdio_register reg_struct;
380 reg_struct.range.start = (uintptr_t)host_addr;
381 reg_struct.range.len = length;
382 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
384 /* Now tell our userfault_fd that it's responsible for this area */
385 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) {
386 error_report("%s userfault register: %s", __func__, strerror(errno));
394 * Handle faults detected by the USERFAULT markings
396 static void *postcopy_ram_fault_thread(void *opaque)
398 MigrationIncomingState *mis = opaque;
401 size_t hostpagesize = getpagesize();
403 RAMBlock *last_rb = NULL; /* last RAMBlock we sent part of */
405 trace_postcopy_ram_fault_thread_entry();
406 qemu_sem_post(&mis->fault_thread_sem);
409 ram_addr_t rb_offset;
410 ram_addr_t in_raspace;
411 struct pollfd pfd[2];
414 * We're mainly waiting for the kernel to give us a faulting HVA,
415 * however we can be told to quit via userfault_quit_fd which is
418 pfd[0].fd = mis->userfault_fd;
419 pfd[0].events = POLLIN;
421 pfd[1].fd = mis->userfault_quit_fd;
422 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
425 if (poll(pfd, 2, -1 /* Wait forever */) == -1) {
426 error_report("%s: userfault poll: %s", __func__, strerror(errno));
430 if (pfd[1].revents) {
431 trace_postcopy_ram_fault_thread_quit();
435 ret = read(mis->userfault_fd, &msg, sizeof(msg));
436 if (ret != sizeof(msg)) {
437 if (errno == EAGAIN) {
439 * if a wake up happens on the other thread just after
440 * the poll, there is nothing to read.
445 error_report("%s: Failed to read full userfault message: %s",
446 __func__, strerror(errno));
449 error_report("%s: Read %d bytes from userfaultfd expected %zd",
450 __func__, ret, sizeof(msg));
451 break; /* Lost alignment, don't know what we'd read next */
454 if (msg.event != UFFD_EVENT_PAGEFAULT) {
455 error_report("%s: Read unexpected event %ud from userfaultfd",
456 __func__, msg.event);
457 continue; /* It's not a page fault, shouldn't happen */
460 rb = qemu_ram_block_from_host(
461 (void *)(uintptr_t)msg.arg.pagefault.address,
462 true, &in_raspace, &rb_offset);
464 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
465 PRIx64, (uint64_t)msg.arg.pagefault.address);
469 rb_offset &= ~(hostpagesize - 1);
470 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
471 qemu_ram_get_idstr(rb),
475 * Send the request to the source - we want to request one
476 * of our host page sizes (which is >= TPS)
480 migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
481 rb_offset, hostpagesize);
483 /* Save some space */
484 migrate_send_rp_req_pages(mis, NULL,
485 rb_offset, hostpagesize);
488 trace_postcopy_ram_fault_thread_exit();
492 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
494 /* Open the fd for the kernel to give us userfaults */
495 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
496 if (mis->userfault_fd == -1) {
497 error_report("%s: Failed to open userfault fd: %s", __func__,
503 * Although the host check already tested the API, we need to
504 * do the check again as an ABI handshake on the new fd.
506 if (!ufd_version_check(mis->userfault_fd)) {
510 /* Now an eventfd we use to tell the fault-thread to quit */
511 mis->userfault_quit_fd = eventfd(0, EFD_CLOEXEC);
512 if (mis->userfault_quit_fd == -1) {
513 error_report("%s: Opening userfault_quit_fd: %s", __func__,
515 close(mis->userfault_fd);
519 qemu_sem_init(&mis->fault_thread_sem, 0);
520 qemu_thread_create(&mis->fault_thread, "postcopy/fault",
521 postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
522 qemu_sem_wait(&mis->fault_thread_sem);
523 qemu_sem_destroy(&mis->fault_thread_sem);
524 mis->have_fault_thread = true;
526 /* Mark so that we get notified of accesses to unwritten areas */
527 if (qemu_ram_foreach_block(ram_block_enable_notify, mis)) {
532 * Ballooning can mark pages as absent while we're postcopying
533 * that would cause false userfaults.
535 qemu_balloon_inhibit(true);
537 trace_postcopy_ram_enable_notify();
543 * Place a host page (from) at (host) atomically
544 * returns 0 on success
546 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from)
548 struct uffdio_copy copy_struct;
550 copy_struct.dst = (uint64_t)(uintptr_t)host;
551 copy_struct.src = (uint64_t)(uintptr_t)from;
552 copy_struct.len = getpagesize();
553 copy_struct.mode = 0;
555 /* copy also acks to the kernel waking the stalled thread up
556 * TODO: We can inhibit that ack and only do it if it was requested
557 * which would be slightly cheaper, but we'd have to be careful
558 * of the order of updating our page state.
560 if (ioctl(mis->userfault_fd, UFFDIO_COPY, ©_struct)) {
562 error_report("%s: %s copy host: %p from: %p",
563 __func__, strerror(e), host, from);
568 trace_postcopy_place_page(host);
573 * Place a zero page at (host) atomically
574 * returns 0 on success
576 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host)
578 struct uffdio_zeropage zero_struct;
580 zero_struct.range.start = (uint64_t)(uintptr_t)host;
581 zero_struct.range.len = getpagesize();
582 zero_struct.mode = 0;
584 if (ioctl(mis->userfault_fd, UFFDIO_ZEROPAGE, &zero_struct)) {
586 error_report("%s: %s zero host: %p",
587 __func__, strerror(e), host);
592 trace_postcopy_place_page_zero(host);
597 * Returns a target page of memory that can be mapped at a later point in time
598 * using postcopy_place_page
599 * The same address is used repeatedly, postcopy_place_page just takes the
601 * Returns: Pointer to allocated page
604 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
606 if (!mis->postcopy_tmp_page) {
607 mis->postcopy_tmp_page = mmap(NULL, getpagesize(),
608 PROT_READ | PROT_WRITE, MAP_PRIVATE |
609 MAP_ANONYMOUS, -1, 0);
610 if (!mis->postcopy_tmp_page) {
611 error_report("%s: %s", __func__, strerror(errno));
616 return mis->postcopy_tmp_page;
620 /* No target OS support, stubs just fail */
621 bool postcopy_ram_supported_by_host(void)
623 error_report("%s: No OS support", __func__);
627 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
629 error_report("postcopy_ram_incoming_init: No OS support");
633 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
639 int postcopy_ram_discard_range(MigrationIncomingState *mis, uint8_t *start,
646 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
652 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
658 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from)
664 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host)
670 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
678 /* ------------------------------------------------------------------------- */
681 * postcopy_discard_send_init: Called at the start of each RAMBlock before
682 * asking to discard individual ranges.
684 * @ms: The current migration state.
685 * @offset: the bitmap offset of the named RAMBlock in the migration
687 * @name: RAMBlock that discards will operate on.
689 * returns: a new PDS.
691 PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms,
692 unsigned long offset,
695 PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState));
698 res->ramblock_name = name;
699 res->offset = offset;
706 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
707 * discard. May send a discard message, may just leave it queued to
710 * @ms: Current migration state.
711 * @pds: Structure initialised by postcopy_discard_send_init().
712 * @start,@length: a range of pages in the migration bitmap in the
713 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
715 void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds,
716 unsigned long start, unsigned long length)
718 size_t tp_bits = qemu_target_page_bits();
719 /* Convert to byte offsets within the RAM block */
720 pds->start_list[pds->cur_entry] = (start - pds->offset) << tp_bits;
721 pds->length_list[pds->cur_entry] = length << tp_bits;
722 trace_postcopy_discard_send_range(pds->ramblock_name, start, length);
726 if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) {
727 /* Full set, ship it! */
728 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
739 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
740 * bitmap code. Sends any outstanding discard messages, frees the PDS
742 * @ms: Current migration state.
743 * @pds: Structure initialised by postcopy_discard_send_init().
745 void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds)
747 /* Anything unsent? */
748 if (pds->cur_entry) {
749 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
757 trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords,