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 "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
;
38 * Start and length of a discard range (bytes)
40 uint64_t start_list
[MAX_DISCARDS_PER_COMMAND
];
41 uint64_t length_list
[MAX_DISCARDS_PER_COMMAND
];
42 unsigned int nsentwords
;
43 unsigned int nsentcmds
;
46 /* Postcopy needs to detect accesses to pages that haven't yet been copied
47 * across, and efficiently map new pages in, the techniques for doing this
48 * are target OS specific.
50 #if defined(__linux__)
53 #include <sys/ioctl.h>
54 #include <sys/syscall.h>
55 #include <asm/types.h> /* for __u64 */
58 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
59 #include <sys/eventfd.h>
60 #include <linux/userfaultfd.h>
62 static bool ufd_version_check(int ufd
)
64 struct uffdio_api api_struct
;
67 api_struct
.api
= UFFD_API
;
68 api_struct
.features
= 0;
69 if (ioctl(ufd
, UFFDIO_API
, &api_struct
)) {
70 error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
75 ioctl_mask
= (__u64
)1 << _UFFDIO_REGISTER
|
76 (__u64
)1 << _UFFDIO_UNREGISTER
;
77 if ((api_struct
.ioctls
& ioctl_mask
) != ioctl_mask
) {
78 error_report("Missing userfault features: %" PRIx64
,
79 (uint64_t)(~api_struct
.ioctls
& ioctl_mask
));
83 if (getpagesize() != ram_pagesize_summary()) {
85 /* We've got a huge page */
86 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS
87 have_hp
= api_struct
.features
& UFFD_FEATURE_MISSING_HUGETLBFS
;
90 error_report("Userfault on this host does not support huge pages");
97 /* Callback from postcopy_ram_supported_by_host block iterator.
99 static int test_range_shared(const char *block_name
, void *host_addr
,
100 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
102 if (qemu_ram_is_shared(qemu_ram_block_by_name(block_name
))) {
103 error_report("Postcopy on shared RAM (%s) is not yet supported",
111 * Note: This has the side effect of munlock'ing all of RAM, that's
112 * normally fine since if the postcopy succeeds it gets turned back on at the
115 bool postcopy_ram_supported_by_host(void)
117 long pagesize
= getpagesize();
119 bool ret
= false; /* Error unless we change it */
120 void *testarea
= NULL
;
121 struct uffdio_register reg_struct
;
122 struct uffdio_range range_struct
;
123 uint64_t feature_mask
;
125 if (qemu_target_page_size() > pagesize
) {
126 error_report("Target page size bigger than host page size");
130 ufd
= syscall(__NR_userfaultfd
, O_CLOEXEC
);
132 error_report("%s: userfaultfd not available: %s", __func__
,
137 /* Version and features check */
138 if (!ufd_version_check(ufd
)) {
142 /* We don't support postcopy with shared RAM yet */
143 if (qemu_ram_foreach_block(test_range_shared
, NULL
)) {
148 * userfault and mlock don't go together; we'll put it back later if
152 error_report("%s: munlockall: %s", __func__
, strerror(errno
));
157 * We need to check that the ops we need are supported on anon memory
158 * To do that we need to register a chunk and see the flags that
161 testarea
= mmap(NULL
, pagesize
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
162 MAP_ANONYMOUS
, -1, 0);
163 if (testarea
== MAP_FAILED
) {
164 error_report("%s: Failed to map test area: %s", __func__
,
168 g_assert(((size_t)testarea
& (pagesize
-1)) == 0);
170 reg_struct
.range
.start
= (uintptr_t)testarea
;
171 reg_struct
.range
.len
= pagesize
;
172 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
174 if (ioctl(ufd
, UFFDIO_REGISTER
, ®_struct
)) {
175 error_report("%s userfault register: %s", __func__
, strerror(errno
));
179 range_struct
.start
= (uintptr_t)testarea
;
180 range_struct
.len
= pagesize
;
181 if (ioctl(ufd
, UFFDIO_UNREGISTER
, &range_struct
)) {
182 error_report("%s userfault unregister: %s", __func__
, strerror(errno
));
186 feature_mask
= (__u64
)1 << _UFFDIO_WAKE
|
187 (__u64
)1 << _UFFDIO_COPY
|
188 (__u64
)1 << _UFFDIO_ZEROPAGE
;
189 if ((reg_struct
.ioctls
& feature_mask
) != feature_mask
) {
190 error_report("Missing userfault map features: %" PRIx64
,
191 (uint64_t)(~reg_struct
.ioctls
& feature_mask
));
199 munmap(testarea
, pagesize
);
208 * Setup an area of RAM so that it *can* be used for postcopy later; this
209 * must be done right at the start prior to pre-copy.
210 * opaque should be the MIS.
212 static int init_range(const char *block_name
, void *host_addr
,
213 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
215 trace_postcopy_init_range(block_name
, host_addr
, offset
, length
);
218 * We need the whole of RAM to be truly empty for postcopy, so things
219 * like ROMs and any data tables built during init must be zero'd
220 * - we're going to get the copy from the source anyway.
221 * (Precopy will just overwrite this data, so doesn't need the discard)
223 if (ram_discard_range(block_name
, 0, length
)) {
231 * At the end of migration, undo the effects of init_range
232 * opaque should be the MIS.
234 static int cleanup_range(const char *block_name
, void *host_addr
,
235 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
237 MigrationIncomingState
*mis
= opaque
;
238 struct uffdio_range range_struct
;
239 trace_postcopy_cleanup_range(block_name
, host_addr
, offset
, length
);
242 * We turned off hugepage for the precopy stage with postcopy enabled
243 * we can turn it back on now.
245 qemu_madvise(host_addr
, length
, QEMU_MADV_HUGEPAGE
);
248 * We can also turn off userfault now since we should have all the
249 * pages. It can be useful to leave it on to debug postcopy
250 * if you're not sure it's always getting every page.
252 range_struct
.start
= (uintptr_t)host_addr
;
253 range_struct
.len
= length
;
255 if (ioctl(mis
->userfault_fd
, UFFDIO_UNREGISTER
, &range_struct
)) {
256 error_report("%s: userfault unregister %s", __func__
, strerror(errno
));
265 * Initialise postcopy-ram, setting the RAM to a state where we can go into
266 * postcopy later; must be called prior to any precopy.
267 * called from arch_init's similarly named ram_postcopy_incoming_init
269 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
271 if (qemu_ram_foreach_block(init_range
, NULL
)) {
279 * At the end of a migration where postcopy_ram_incoming_init was called.
281 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
283 trace_postcopy_ram_incoming_cleanup_entry();
285 if (mis
->have_fault_thread
) {
288 if (qemu_ram_foreach_block(cleanup_range
, mis
)) {
292 * Tell the fault_thread to exit, it's an eventfd that should
293 * currently be at 0, we're going to increment it to 1
296 if (write(mis
->userfault_quit_fd
, &tmp64
, 8) == 8) {
297 trace_postcopy_ram_incoming_cleanup_join();
298 qemu_thread_join(&mis
->fault_thread
);
300 /* Not much we can do here, but may as well report it */
301 error_report("%s: incrementing userfault_quit_fd: %s", __func__
,
304 trace_postcopy_ram_incoming_cleanup_closeuf();
305 close(mis
->userfault_fd
);
306 close(mis
->userfault_quit_fd
);
307 mis
->have_fault_thread
= false;
310 qemu_balloon_inhibit(false);
313 if (os_mlock() < 0) {
314 error_report("mlock: %s", strerror(errno
));
316 * It doesn't feel right to fail at this point, we have a valid
322 postcopy_state_set(POSTCOPY_INCOMING_END
);
323 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
325 if (mis
->postcopy_tmp_page
) {
326 munmap(mis
->postcopy_tmp_page
, mis
->largest_page_size
);
327 mis
->postcopy_tmp_page
= NULL
;
329 if (mis
->postcopy_tmp_zero_page
) {
330 munmap(mis
->postcopy_tmp_zero_page
, mis
->largest_page_size
);
331 mis
->postcopy_tmp_zero_page
= NULL
;
333 trace_postcopy_ram_incoming_cleanup_exit();
338 * Disable huge pages on an area
340 static int nhp_range(const char *block_name
, void *host_addr
,
341 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
343 trace_postcopy_nhp_range(block_name
, host_addr
, offset
, length
);
346 * Before we do discards we need to ensure those discards really
347 * do delete areas of the page, even if THP thinks a hugepage would
348 * be a good idea, so force hugepages off.
350 qemu_madvise(host_addr
, length
, QEMU_MADV_NOHUGEPAGE
);
356 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
357 * however leaving it until after precopy means that most of the precopy
360 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
362 if (qemu_ram_foreach_block(nhp_range
, mis
)) {
366 postcopy_state_set(POSTCOPY_INCOMING_DISCARD
);
372 * Mark the given area of RAM as requiring notification to unwritten areas
373 * Used as a callback on qemu_ram_foreach_block.
374 * host_addr: Base of area to mark
375 * offset: Offset in the whole ram arena
376 * length: Length of the section
377 * opaque: MigrationIncomingState pointer
378 * Returns 0 on success
380 static int ram_block_enable_notify(const char *block_name
, void *host_addr
,
381 ram_addr_t offset
, ram_addr_t length
,
384 MigrationIncomingState
*mis
= opaque
;
385 struct uffdio_register reg_struct
;
387 reg_struct
.range
.start
= (uintptr_t)host_addr
;
388 reg_struct
.range
.len
= length
;
389 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
391 /* Now tell our userfault_fd that it's responsible for this area */
392 if (ioctl(mis
->userfault_fd
, UFFDIO_REGISTER
, ®_struct
)) {
393 error_report("%s userfault register: %s", __func__
, strerror(errno
));
396 if (!(reg_struct
.ioctls
& ((__u64
)1 << _UFFDIO_COPY
))) {
397 error_report("%s userfault: Region doesn't support COPY", __func__
);
405 * Handle faults detected by the USERFAULT markings
407 static void *postcopy_ram_fault_thread(void *opaque
)
409 MigrationIncomingState
*mis
= opaque
;
413 RAMBlock
*last_rb
= NULL
; /* last RAMBlock we sent part of */
415 trace_postcopy_ram_fault_thread_entry();
416 qemu_sem_post(&mis
->fault_thread_sem
);
419 ram_addr_t rb_offset
;
420 struct pollfd pfd
[2];
423 * We're mainly waiting for the kernel to give us a faulting HVA,
424 * however we can be told to quit via userfault_quit_fd which is
427 pfd
[0].fd
= mis
->userfault_fd
;
428 pfd
[0].events
= POLLIN
;
430 pfd
[1].fd
= mis
->userfault_quit_fd
;
431 pfd
[1].events
= POLLIN
; /* Waiting for eventfd to go positive */
434 if (poll(pfd
, 2, -1 /* Wait forever */) == -1) {
435 error_report("%s: userfault poll: %s", __func__
, strerror(errno
));
439 if (pfd
[1].revents
) {
440 trace_postcopy_ram_fault_thread_quit();
444 ret
= read(mis
->userfault_fd
, &msg
, sizeof(msg
));
445 if (ret
!= sizeof(msg
)) {
446 if (errno
== EAGAIN
) {
448 * if a wake up happens on the other thread just after
449 * the poll, there is nothing to read.
454 error_report("%s: Failed to read full userfault message: %s",
455 __func__
, strerror(errno
));
458 error_report("%s: Read %d bytes from userfaultfd expected %zd",
459 __func__
, ret
, sizeof(msg
));
460 break; /* Lost alignment, don't know what we'd read next */
463 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
) {
464 error_report("%s: Read unexpected event %ud from userfaultfd",
465 __func__
, msg
.event
);
466 continue; /* It's not a page fault, shouldn't happen */
469 rb
= qemu_ram_block_from_host(
470 (void *)(uintptr_t)msg
.arg
.pagefault
.address
,
473 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
474 PRIx64
, (uint64_t)msg
.arg
.pagefault
.address
);
478 rb_offset
&= ~(qemu_ram_pagesize(rb
) - 1);
479 trace_postcopy_ram_fault_thread_request(msg
.arg
.pagefault
.address
,
480 qemu_ram_get_idstr(rb
),
484 * Send the request to the source - we want to request one
485 * of our host page sizes (which is >= TPS)
489 migrate_send_rp_req_pages(mis
, qemu_ram_get_idstr(rb
),
490 rb_offset
, qemu_ram_pagesize(rb
));
492 /* Save some space */
493 migrate_send_rp_req_pages(mis
, NULL
,
494 rb_offset
, qemu_ram_pagesize(rb
));
497 trace_postcopy_ram_fault_thread_exit();
501 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
503 /* Open the fd for the kernel to give us userfaults */
504 mis
->userfault_fd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
505 if (mis
->userfault_fd
== -1) {
506 error_report("%s: Failed to open userfault fd: %s", __func__
,
512 * Although the host check already tested the API, we need to
513 * do the check again as an ABI handshake on the new fd.
515 if (!ufd_version_check(mis
->userfault_fd
)) {
519 /* Now an eventfd we use to tell the fault-thread to quit */
520 mis
->userfault_quit_fd
= eventfd(0, EFD_CLOEXEC
);
521 if (mis
->userfault_quit_fd
== -1) {
522 error_report("%s: Opening userfault_quit_fd: %s", __func__
,
524 close(mis
->userfault_fd
);
528 qemu_sem_init(&mis
->fault_thread_sem
, 0);
529 qemu_thread_create(&mis
->fault_thread
, "postcopy/fault",
530 postcopy_ram_fault_thread
, mis
, QEMU_THREAD_JOINABLE
);
531 qemu_sem_wait(&mis
->fault_thread_sem
);
532 qemu_sem_destroy(&mis
->fault_thread_sem
);
533 mis
->have_fault_thread
= true;
535 /* Mark so that we get notified of accesses to unwritten areas */
536 if (qemu_ram_foreach_block(ram_block_enable_notify
, mis
)) {
541 * Ballooning can mark pages as absent while we're postcopying
542 * that would cause false userfaults.
544 qemu_balloon_inhibit(true);
546 trace_postcopy_ram_enable_notify();
552 * Place a host page (from) at (host) atomically
553 * returns 0 on success
555 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
558 struct uffdio_copy copy_struct
;
560 copy_struct
.dst
= (uint64_t)(uintptr_t)host
;
561 copy_struct
.src
= (uint64_t)(uintptr_t)from
;
562 copy_struct
.len
= pagesize
;
563 copy_struct
.mode
= 0;
565 /* copy also acks to the kernel waking the stalled thread up
566 * TODO: We can inhibit that ack and only do it if it was requested
567 * which would be slightly cheaper, but we'd have to be careful
568 * of the order of updating our page state.
570 if (ioctl(mis
->userfault_fd
, UFFDIO_COPY
, ©_struct
)) {
572 error_report("%s: %s copy host: %p from: %p (size: %zd)",
573 __func__
, strerror(e
), host
, from
, pagesize
);
578 trace_postcopy_place_page(host
);
583 * Place a zero page at (host) atomically
584 * returns 0 on success
586 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
589 trace_postcopy_place_page_zero(host
);
591 if (pagesize
== getpagesize()) {
592 struct uffdio_zeropage zero_struct
;
593 zero_struct
.range
.start
= (uint64_t)(uintptr_t)host
;
594 zero_struct
.range
.len
= getpagesize();
595 zero_struct
.mode
= 0;
597 if (ioctl(mis
->userfault_fd
, UFFDIO_ZEROPAGE
, &zero_struct
)) {
599 error_report("%s: %s zero host: %p",
600 __func__
, strerror(e
), host
);
605 /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */
606 if (!mis
->postcopy_tmp_zero_page
) {
607 mis
->postcopy_tmp_zero_page
= mmap(NULL
, mis
->largest_page_size
,
608 PROT_READ
| PROT_WRITE
,
609 MAP_PRIVATE
| MAP_ANONYMOUS
,
611 if (mis
->postcopy_tmp_zero_page
== MAP_FAILED
) {
613 mis
->postcopy_tmp_zero_page
= NULL
;
614 error_report("%s: %s mapping large zero page",
615 __func__
, strerror(e
));
618 memset(mis
->postcopy_tmp_zero_page
, '\0', mis
->largest_page_size
);
620 return postcopy_place_page(mis
, host
, mis
->postcopy_tmp_zero_page
,
628 * Returns a target page of memory that can be mapped at a later point in time
629 * using postcopy_place_page
630 * The same address is used repeatedly, postcopy_place_page just takes the
632 * Returns: Pointer to allocated page
635 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
637 if (!mis
->postcopy_tmp_page
) {
638 mis
->postcopy_tmp_page
= mmap(NULL
, mis
->largest_page_size
,
639 PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
640 MAP_ANONYMOUS
, -1, 0);
641 if (mis
->postcopy_tmp_page
== MAP_FAILED
) {
642 mis
->postcopy_tmp_page
= NULL
;
643 error_report("%s: %s", __func__
, strerror(errno
));
648 return mis
->postcopy_tmp_page
;
652 /* No target OS support, stubs just fail */
653 bool postcopy_ram_supported_by_host(void)
655 error_report("%s: No OS support", __func__
);
659 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
661 error_report("postcopy_ram_incoming_init: No OS support");
665 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
671 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
677 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
683 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
690 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
697 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
705 /* ------------------------------------------------------------------------- */
708 * postcopy_discard_send_init: Called at the start of each RAMBlock before
709 * asking to discard individual ranges.
711 * @ms: The current migration state.
712 * @offset: the bitmap offset of the named RAMBlock in the migration
714 * @name: RAMBlock that discards will operate on.
716 * returns: a new PDS.
718 PostcopyDiscardState
*postcopy_discard_send_init(MigrationState
*ms
,
721 PostcopyDiscardState
*res
= g_malloc0(sizeof(PostcopyDiscardState
));
724 res
->ramblock_name
= name
;
731 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
732 * discard. May send a discard message, may just leave it queued to
735 * @ms: Current migration state.
736 * @pds: Structure initialised by postcopy_discard_send_init().
737 * @start,@length: a range of pages in the migration bitmap in the
738 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
740 void postcopy_discard_send_range(MigrationState
*ms
, PostcopyDiscardState
*pds
,
741 unsigned long start
, unsigned long length
)
743 size_t tp_size
= qemu_target_page_size();
744 /* Convert to byte offsets within the RAM block */
745 pds
->start_list
[pds
->cur_entry
] = start
* tp_size
;
746 pds
->length_list
[pds
->cur_entry
] = length
* tp_size
;
747 trace_postcopy_discard_send_range(pds
->ramblock_name
, start
, length
);
751 if (pds
->cur_entry
== MAX_DISCARDS_PER_COMMAND
) {
752 /* Full set, ship it! */
753 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
764 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
765 * bitmap code. Sends any outstanding discard messages, frees the PDS
767 * @ms: Current migration state.
768 * @pds: Structure initialised by postcopy_discard_send_init().
770 void postcopy_discard_send_finish(MigrationState
*ms
, PostcopyDiscardState
*pds
)
772 /* Anything unsent? */
773 if (pds
->cur_entry
) {
774 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
782 trace_postcopy_discard_send_finish(pds
->ramblock_name
, pds
->nsentwords
,
789 * Current state of incoming postcopy; note this is not part of
790 * MigrationIncomingState since it's state is used during cleanup
791 * at the end as MIS is being freed.
793 static PostcopyState incoming_postcopy_state
;
795 PostcopyState
postcopy_state_get(void)
797 return atomic_mb_read(&incoming_postcopy_state
);
800 /* Set the state and return the old state */
801 PostcopyState
postcopy_state_set(PostcopyState new_state
)
803 return atomic_xchg(&incoming_postcopy_state
, new_state
);