2 * Postcopy migration for RAM
4 * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates
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"
21 #include "qemu-common.h"
22 #include "migration/migration.h"
23 #include "migration/postcopy-ram.h"
24 #include "sysemu/sysemu.h"
25 #include "sysemu/balloon.h"
26 #include "qemu/error-report.h"
29 /* Arbitrary limit on size of each discard command,
30 * keeps them around ~200 bytes
32 #define MAX_DISCARDS_PER_COMMAND 12
34 struct PostcopyDiscardState
{
35 const char *ramblock_name
;
36 uint64_t offset
; /* Bitmap entry for the 1st bit of this RAMBlock */
39 * Start and length of a discard range (bytes)
41 uint64_t start_list
[MAX_DISCARDS_PER_COMMAND
];
42 uint64_t length_list
[MAX_DISCARDS_PER_COMMAND
];
43 unsigned int nsentwords
;
44 unsigned int nsentcmds
;
47 /* Postcopy needs to detect accesses to pages that haven't yet been copied
48 * across, and efficiently map new pages in, the techniques for doing this
49 * are target OS specific.
51 #if defined(__linux__)
54 #include <sys/ioctl.h>
55 #include <sys/syscall.h>
56 #include <asm/types.h> /* for __u64 */
59 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
60 #include <sys/eventfd.h>
61 #include <linux/userfaultfd.h>
63 static bool ufd_version_check(int ufd
)
65 struct uffdio_api api_struct
;
68 api_struct
.api
= UFFD_API
;
69 api_struct
.features
= 0;
70 if (ioctl(ufd
, UFFDIO_API
, &api_struct
)) {
71 error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
76 ioctl_mask
= (__u64
)1 << _UFFDIO_REGISTER
|
77 (__u64
)1 << _UFFDIO_UNREGISTER
;
78 if ((api_struct
.ioctls
& ioctl_mask
) != ioctl_mask
) {
79 error_report("Missing userfault features: %" PRIx64
,
80 (uint64_t)(~api_struct
.ioctls
& ioctl_mask
));
84 if (getpagesize() != ram_pagesize_summary()) {
86 /* We've got a huge page */
87 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
88 have_hp
= api_struct
.features
& UFFD_FEATURE_MISSING_HUGETLBFS
;
91 error_report("Userfault on this host does not support huge pages");
98 /* Callback from postcopy_ram_supported_by_host block iterator.
100 static int test_range_shared(const char *block_name
, void *host_addr
,
101 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
103 if (qemu_ram_is_shared(qemu_ram_block_by_name(block_name
))) {
104 error_report("Postcopy on shared RAM (%s) is not yet supported",
112 * Note: This has the side effect of munlock'ing all of RAM, that's
113 * normally fine since if the postcopy succeeds it gets turned back on at the
116 bool postcopy_ram_supported_by_host(void)
118 long pagesize
= getpagesize();
120 bool ret
= false; /* Error unless we change it */
121 void *testarea
= NULL
;
122 struct uffdio_register reg_struct
;
123 struct uffdio_range range_struct
;
124 uint64_t feature_mask
;
126 if ((1ul << qemu_target_page_bits()) > pagesize
) {
127 error_report("Target page size bigger than host page size");
131 ufd
= syscall(__NR_userfaultfd
, O_CLOEXEC
);
133 error_report("%s: userfaultfd not available: %s", __func__
,
138 /* Version and features check */
139 if (!ufd_version_check(ufd
)) {
143 /* We don't support postcopy with shared RAM yet */
144 if (qemu_ram_foreach_block(test_range_shared
, NULL
)) {
149 * userfault and mlock don't go together; we'll put it back later if
153 error_report("%s: munlockall: %s", __func__
, strerror(errno
));
158 * We need to check that the ops we need are supported on anon memory
159 * To do that we need to register a chunk and see the flags that
162 testarea
= mmap(NULL
, pagesize
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
163 MAP_ANONYMOUS
, -1, 0);
164 if (testarea
== MAP_FAILED
) {
165 error_report("%s: Failed to map test area: %s", __func__
,
169 g_assert(((size_t)testarea
& (pagesize
-1)) == 0);
171 reg_struct
.range
.start
= (uintptr_t)testarea
;
172 reg_struct
.range
.len
= pagesize
;
173 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
175 if (ioctl(ufd
, UFFDIO_REGISTER
, ®_struct
)) {
176 error_report("%s userfault register: %s", __func__
, strerror(errno
));
180 range_struct
.start
= (uintptr_t)testarea
;
181 range_struct
.len
= pagesize
;
182 if (ioctl(ufd
, UFFDIO_UNREGISTER
, &range_struct
)) {
183 error_report("%s userfault unregister: %s", __func__
, strerror(errno
));
187 feature_mask
= (__u64
)1 << _UFFDIO_WAKE
|
188 (__u64
)1 << _UFFDIO_COPY
|
189 (__u64
)1 << _UFFDIO_ZEROPAGE
;
190 if ((reg_struct
.ioctls
& feature_mask
) != feature_mask
) {
191 error_report("Missing userfault map features: %" PRIx64
,
192 (uint64_t)(~reg_struct
.ioctls
& feature_mask
));
200 munmap(testarea
, pagesize
);
209 * Setup an area of RAM so that it *can* be used for postcopy later; this
210 * must be done right at the start prior to pre-copy.
211 * opaque should be the MIS.
213 static int init_range(const char *block_name
, void *host_addr
,
214 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
216 MigrationIncomingState
*mis
= opaque
;
218 trace_postcopy_init_range(block_name
, host_addr
, offset
, length
);
221 * We need the whole of RAM to be truly empty for postcopy, so things
222 * like ROMs and any data tables built during init must be zero'd
223 * - we're going to get the copy from the source anyway.
224 * (Precopy will just overwrite this data, so doesn't need the discard)
226 if (ram_discard_range(mis
, block_name
, 0, length
)) {
234 * At the end of migration, undo the effects of init_range
235 * opaque should be the MIS.
237 static int cleanup_range(const char *block_name
, void *host_addr
,
238 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
240 MigrationIncomingState
*mis
= opaque
;
241 struct uffdio_range range_struct
;
242 trace_postcopy_cleanup_range(block_name
, host_addr
, offset
, length
);
245 * We turned off hugepage for the precopy stage with postcopy enabled
246 * we can turn it back on now.
248 qemu_madvise(host_addr
, length
, QEMU_MADV_HUGEPAGE
);
251 * We can also turn off userfault now since we should have all the
252 * pages. It can be useful to leave it on to debug postcopy
253 * if you're not sure it's always getting every page.
255 range_struct
.start
= (uintptr_t)host_addr
;
256 range_struct
.len
= length
;
258 if (ioctl(mis
->userfault_fd
, UFFDIO_UNREGISTER
, &range_struct
)) {
259 error_report("%s: userfault unregister %s", __func__
, strerror(errno
));
268 * Initialise postcopy-ram, setting the RAM to a state where we can go into
269 * postcopy later; must be called prior to any precopy.
270 * called from arch_init's similarly named ram_postcopy_incoming_init
272 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
274 if (qemu_ram_foreach_block(init_range
, mis
)) {
282 * At the end of a migration where postcopy_ram_incoming_init was called.
284 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
286 trace_postcopy_ram_incoming_cleanup_entry();
288 if (mis
->have_fault_thread
) {
291 if (qemu_ram_foreach_block(cleanup_range
, mis
)) {
295 * Tell the fault_thread to exit, it's an eventfd that should
296 * currently be at 0, we're going to increment it to 1
299 if (write(mis
->userfault_quit_fd
, &tmp64
, 8) == 8) {
300 trace_postcopy_ram_incoming_cleanup_join();
301 qemu_thread_join(&mis
->fault_thread
);
303 /* Not much we can do here, but may as well report it */
304 error_report("%s: incrementing userfault_quit_fd: %s", __func__
,
307 trace_postcopy_ram_incoming_cleanup_closeuf();
308 close(mis
->userfault_fd
);
309 close(mis
->userfault_quit_fd
);
310 mis
->have_fault_thread
= false;
313 qemu_balloon_inhibit(false);
316 if (os_mlock() < 0) {
317 error_report("mlock: %s", strerror(errno
));
319 * It doesn't feel right to fail at this point, we have a valid
325 postcopy_state_set(POSTCOPY_INCOMING_END
);
326 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
328 if (mis
->postcopy_tmp_page
) {
329 munmap(mis
->postcopy_tmp_page
, mis
->largest_page_size
);
330 mis
->postcopy_tmp_page
= NULL
;
332 if (mis
->postcopy_tmp_zero_page
) {
333 munmap(mis
->postcopy_tmp_zero_page
, mis
->largest_page_size
);
334 mis
->postcopy_tmp_zero_page
= NULL
;
336 trace_postcopy_ram_incoming_cleanup_exit();
341 * Disable huge pages on an area
343 static int nhp_range(const char *block_name
, void *host_addr
,
344 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
346 trace_postcopy_nhp_range(block_name
, host_addr
, offset
, length
);
349 * Before we do discards we need to ensure those discards really
350 * do delete areas of the page, even if THP thinks a hugepage would
351 * be a good idea, so force hugepages off.
353 qemu_madvise(host_addr
, length
, QEMU_MADV_NOHUGEPAGE
);
359 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
360 * however leaving it until after precopy means that most of the precopy
363 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
365 if (qemu_ram_foreach_block(nhp_range
, mis
)) {
369 postcopy_state_set(POSTCOPY_INCOMING_DISCARD
);
375 * Mark the given area of RAM as requiring notification to unwritten areas
376 * Used as a callback on qemu_ram_foreach_block.
377 * host_addr: Base of area to mark
378 * offset: Offset in the whole ram arena
379 * length: Length of the section
380 * opaque: MigrationIncomingState pointer
381 * Returns 0 on success
383 static int ram_block_enable_notify(const char *block_name
, void *host_addr
,
384 ram_addr_t offset
, ram_addr_t length
,
387 MigrationIncomingState
*mis
= opaque
;
388 struct uffdio_register reg_struct
;
390 reg_struct
.range
.start
= (uintptr_t)host_addr
;
391 reg_struct
.range
.len
= length
;
392 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
394 /* Now tell our userfault_fd that it's responsible for this area */
395 if (ioctl(mis
->userfault_fd
, UFFDIO_REGISTER
, ®_struct
)) {
396 error_report("%s userfault register: %s", __func__
, strerror(errno
));
399 if (!(reg_struct
.ioctls
& ((__u64
)1 << _UFFDIO_COPY
))) {
400 error_report("%s userfault: Region doesn't support COPY", __func__
);
408 * Handle faults detected by the USERFAULT markings
410 static void *postcopy_ram_fault_thread(void *opaque
)
412 MigrationIncomingState
*mis
= opaque
;
416 RAMBlock
*last_rb
= NULL
; /* last RAMBlock we sent part of */
418 trace_postcopy_ram_fault_thread_entry();
419 qemu_sem_post(&mis
->fault_thread_sem
);
422 ram_addr_t rb_offset
;
423 struct pollfd pfd
[2];
426 * We're mainly waiting for the kernel to give us a faulting HVA,
427 * however we can be told to quit via userfault_quit_fd which is
430 pfd
[0].fd
= mis
->userfault_fd
;
431 pfd
[0].events
= POLLIN
;
433 pfd
[1].fd
= mis
->userfault_quit_fd
;
434 pfd
[1].events
= POLLIN
; /* Waiting for eventfd to go positive */
437 if (poll(pfd
, 2, -1 /* Wait forever */) == -1) {
438 error_report("%s: userfault poll: %s", __func__
, strerror(errno
));
442 if (pfd
[1].revents
) {
443 trace_postcopy_ram_fault_thread_quit();
447 ret
= read(mis
->userfault_fd
, &msg
, sizeof(msg
));
448 if (ret
!= sizeof(msg
)) {
449 if (errno
== EAGAIN
) {
451 * if a wake up happens on the other thread just after
452 * the poll, there is nothing to read.
457 error_report("%s: Failed to read full userfault message: %s",
458 __func__
, strerror(errno
));
461 error_report("%s: Read %d bytes from userfaultfd expected %zd",
462 __func__
, ret
, sizeof(msg
));
463 break; /* Lost alignment, don't know what we'd read next */
466 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
) {
467 error_report("%s: Read unexpected event %ud from userfaultfd",
468 __func__
, msg
.event
);
469 continue; /* It's not a page fault, shouldn't happen */
472 rb
= qemu_ram_block_from_host(
473 (void *)(uintptr_t)msg
.arg
.pagefault
.address
,
476 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
477 PRIx64
, (uint64_t)msg
.arg
.pagefault
.address
);
481 rb_offset
&= ~(qemu_ram_pagesize(rb
) - 1);
482 trace_postcopy_ram_fault_thread_request(msg
.arg
.pagefault
.address
,
483 qemu_ram_get_idstr(rb
),
487 * Send the request to the source - we want to request one
488 * of our host page sizes (which is >= TPS)
492 migrate_send_rp_req_pages(mis
, qemu_ram_get_idstr(rb
),
493 rb_offset
, qemu_ram_pagesize(rb
));
495 /* Save some space */
496 migrate_send_rp_req_pages(mis
, NULL
,
497 rb_offset
, qemu_ram_pagesize(rb
));
500 trace_postcopy_ram_fault_thread_exit();
504 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
506 /* Open the fd for the kernel to give us userfaults */
507 mis
->userfault_fd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
508 if (mis
->userfault_fd
== -1) {
509 error_report("%s: Failed to open userfault fd: %s", __func__
,
515 * Although the host check already tested the API, we need to
516 * do the check again as an ABI handshake on the new fd.
518 if (!ufd_version_check(mis
->userfault_fd
)) {
522 /* Now an eventfd we use to tell the fault-thread to quit */
523 mis
->userfault_quit_fd
= eventfd(0, EFD_CLOEXEC
);
524 if (mis
->userfault_quit_fd
== -1) {
525 error_report("%s: Opening userfault_quit_fd: %s", __func__
,
527 close(mis
->userfault_fd
);
531 qemu_sem_init(&mis
->fault_thread_sem
, 0);
532 qemu_thread_create(&mis
->fault_thread
, "postcopy/fault",
533 postcopy_ram_fault_thread
, mis
, QEMU_THREAD_JOINABLE
);
534 qemu_sem_wait(&mis
->fault_thread_sem
);
535 qemu_sem_destroy(&mis
->fault_thread_sem
);
536 mis
->have_fault_thread
= true;
538 /* Mark so that we get notified of accesses to unwritten areas */
539 if (qemu_ram_foreach_block(ram_block_enable_notify
, mis
)) {
544 * Ballooning can mark pages as absent while we're postcopying
545 * that would cause false userfaults.
547 qemu_balloon_inhibit(true);
549 trace_postcopy_ram_enable_notify();
555 * Place a host page (from) at (host) atomically
556 * returns 0 on success
558 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
561 struct uffdio_copy copy_struct
;
563 copy_struct
.dst
= (uint64_t)(uintptr_t)host
;
564 copy_struct
.src
= (uint64_t)(uintptr_t)from
;
565 copy_struct
.len
= pagesize
;
566 copy_struct
.mode
= 0;
568 /* copy also acks to the kernel waking the stalled thread up
569 * TODO: We can inhibit that ack and only do it if it was requested
570 * which would be slightly cheaper, but we'd have to be careful
571 * of the order of updating our page state.
573 if (ioctl(mis
->userfault_fd
, UFFDIO_COPY
, ©_struct
)) {
575 error_report("%s: %s copy host: %p from: %p (size: %zd)",
576 __func__
, strerror(e
), host
, from
, pagesize
);
581 trace_postcopy_place_page(host
);
586 * Place a zero page at (host) atomically
587 * returns 0 on success
589 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
592 trace_postcopy_place_page_zero(host
);
594 if (pagesize
== getpagesize()) {
595 struct uffdio_zeropage zero_struct
;
596 zero_struct
.range
.start
= (uint64_t)(uintptr_t)host
;
597 zero_struct
.range
.len
= getpagesize();
598 zero_struct
.mode
= 0;
600 if (ioctl(mis
->userfault_fd
, UFFDIO_ZEROPAGE
, &zero_struct
)) {
602 error_report("%s: %s zero host: %p",
603 __func__
, strerror(e
), host
);
608 /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */
609 if (!mis
->postcopy_tmp_zero_page
) {
610 mis
->postcopy_tmp_zero_page
= mmap(NULL
, mis
->largest_page_size
,
611 PROT_READ
| PROT_WRITE
,
612 MAP_PRIVATE
| MAP_ANONYMOUS
,
614 if (mis
->postcopy_tmp_zero_page
== MAP_FAILED
) {
616 mis
->postcopy_tmp_zero_page
= NULL
;
617 error_report("%s: %s mapping large zero page",
618 __func__
, strerror(e
));
621 memset(mis
->postcopy_tmp_zero_page
, '\0', mis
->largest_page_size
);
623 return postcopy_place_page(mis
, host
, mis
->postcopy_tmp_zero_page
,
631 * Returns a target page of memory that can be mapped at a later point in time
632 * using postcopy_place_page
633 * The same address is used repeatedly, postcopy_place_page just takes the
635 * Returns: Pointer to allocated page
638 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
640 if (!mis
->postcopy_tmp_page
) {
641 mis
->postcopy_tmp_page
= mmap(NULL
, mis
->largest_page_size
,
642 PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
643 MAP_ANONYMOUS
, -1, 0);
644 if (mis
->postcopy_tmp_page
== MAP_FAILED
) {
645 mis
->postcopy_tmp_page
= NULL
;
646 error_report("%s: %s", __func__
, strerror(errno
));
651 return mis
->postcopy_tmp_page
;
655 /* No target OS support, stubs just fail */
656 bool postcopy_ram_supported_by_host(void)
658 error_report("%s: No OS support", __func__
);
662 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
664 error_report("postcopy_ram_incoming_init: No OS support");
668 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
674 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
680 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
686 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
693 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
700 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
708 /* ------------------------------------------------------------------------- */
711 * postcopy_discard_send_init: Called at the start of each RAMBlock before
712 * asking to discard individual ranges.
714 * @ms: The current migration state.
715 * @offset: the bitmap offset of the named RAMBlock in the migration
717 * @name: RAMBlock that discards will operate on.
719 * returns: a new PDS.
721 PostcopyDiscardState
*postcopy_discard_send_init(MigrationState
*ms
,
722 unsigned long offset
,
725 PostcopyDiscardState
*res
= g_malloc0(sizeof(PostcopyDiscardState
));
728 res
->ramblock_name
= name
;
729 res
->offset
= offset
;
736 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
737 * discard. May send a discard message, may just leave it queued to
740 * @ms: Current migration state.
741 * @pds: Structure initialised by postcopy_discard_send_init().
742 * @start,@length: a range of pages in the migration bitmap in the
743 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
745 void postcopy_discard_send_range(MigrationState
*ms
, PostcopyDiscardState
*pds
,
746 unsigned long start
, unsigned long length
)
748 size_t tp_bits
= qemu_target_page_bits();
749 /* Convert to byte offsets within the RAM block */
750 pds
->start_list
[pds
->cur_entry
] = (start
- pds
->offset
) << tp_bits
;
751 pds
->length_list
[pds
->cur_entry
] = length
<< tp_bits
;
752 trace_postcopy_discard_send_range(pds
->ramblock_name
, start
, length
);
756 if (pds
->cur_entry
== MAX_DISCARDS_PER_COMMAND
) {
757 /* Full set, ship it! */
758 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
769 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
770 * bitmap code. Sends any outstanding discard messages, frees the PDS
772 * @ms: Current migration state.
773 * @pds: Structure initialised by postcopy_discard_send_init().
775 void postcopy_discard_send_finish(MigrationState
*ms
, PostcopyDiscardState
*pds
)
777 /* Anything unsent? */
778 if (pds
->cur_entry
) {
779 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
787 trace_postcopy_discard_send_finish(pds
->ramblock_name
, pds
->nsentwords
,