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.
23 #include "qemu-common.h"
24 #include "migration/migration.h"
25 #include "migration/postcopy-ram.h"
26 #include "sysemu/sysemu.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__)
55 #include <sys/eventfd.h>
57 #include <sys/ioctl.h>
58 #include <sys/syscall.h>
59 #include <sys/types.h>
60 #include <asm/types.h> /* for __u64 */
63 #if defined(__linux__) && defined(__NR_userfaultfd)
64 #include <linux/userfaultfd.h>
66 static bool ufd_version_check(int ufd
)
68 struct uffdio_api api_struct
;
71 api_struct
.api
= UFFD_API
;
72 api_struct
.features
= 0;
73 if (ioctl(ufd
, UFFDIO_API
, &api_struct
)) {
74 error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
79 ioctl_mask
= (__u64
)1 << _UFFDIO_REGISTER
|
80 (__u64
)1 << _UFFDIO_UNREGISTER
;
81 if ((api_struct
.ioctls
& ioctl_mask
) != ioctl_mask
) {
82 error_report("Missing userfault features: %" PRIx64
,
83 (uint64_t)(~api_struct
.ioctls
& ioctl_mask
));
91 * Note: This has the side effect of munlock'ing all of RAM, that's
92 * normally fine since if the postcopy succeeds it gets turned back on at the
95 bool postcopy_ram_supported_by_host(void)
97 long pagesize
= getpagesize();
99 bool ret
= false; /* Error unless we change it */
100 void *testarea
= NULL
;
101 struct uffdio_register reg_struct
;
102 struct uffdio_range range_struct
;
103 uint64_t feature_mask
;
105 if ((1ul << qemu_target_page_bits()) > pagesize
) {
106 error_report("Target page size bigger than host page size");
110 ufd
= syscall(__NR_userfaultfd
, O_CLOEXEC
);
112 error_report("%s: userfaultfd not available: %s", __func__
,
117 /* Version and features check */
118 if (!ufd_version_check(ufd
)) {
123 * userfault and mlock don't go together; we'll put it back later if
127 error_report("%s: munlockall: %s", __func__
, strerror(errno
));
132 * We need to check that the ops we need are supported on anon memory
133 * To do that we need to register a chunk and see the flags that
136 testarea
= mmap(NULL
, pagesize
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
137 MAP_ANONYMOUS
, -1, 0);
138 if (testarea
== MAP_FAILED
) {
139 error_report("%s: Failed to map test area: %s", __func__
,
143 g_assert(((size_t)testarea
& (pagesize
-1)) == 0);
145 reg_struct
.range
.start
= (uintptr_t)testarea
;
146 reg_struct
.range
.len
= pagesize
;
147 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
149 if (ioctl(ufd
, UFFDIO_REGISTER
, ®_struct
)) {
150 error_report("%s userfault register: %s", __func__
, strerror(errno
));
154 range_struct
.start
= (uintptr_t)testarea
;
155 range_struct
.len
= pagesize
;
156 if (ioctl(ufd
, UFFDIO_UNREGISTER
, &range_struct
)) {
157 error_report("%s userfault unregister: %s", __func__
, strerror(errno
));
161 feature_mask
= (__u64
)1 << _UFFDIO_WAKE
|
162 (__u64
)1 << _UFFDIO_COPY
|
163 (__u64
)1 << _UFFDIO_ZEROPAGE
;
164 if ((reg_struct
.ioctls
& feature_mask
) != feature_mask
) {
165 error_report("Missing userfault map features: %" PRIx64
,
166 (uint64_t)(~reg_struct
.ioctls
& feature_mask
));
174 munmap(testarea
, pagesize
);
183 * postcopy_ram_discard_range: Discard a range of memory.
184 * We can assume that if we've been called postcopy_ram_hosttest returned true.
186 * @mis: Current incoming migration state.
187 * @start, @length: range of memory to discard.
189 * returns: 0 on success.
191 int postcopy_ram_discard_range(MigrationIncomingState
*mis
, uint8_t *start
,
194 trace_postcopy_ram_discard_range(start
, length
);
195 if (madvise(start
, length
, MADV_DONTNEED
)) {
196 error_report("%s MADV_DONTNEED: %s", __func__
, strerror(errno
));
204 * Setup an area of RAM so that it *can* be used for postcopy later; this
205 * must be done right at the start prior to pre-copy.
206 * opaque should be the MIS.
208 static int init_range(const char *block_name
, void *host_addr
,
209 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
211 MigrationIncomingState
*mis
= opaque
;
213 trace_postcopy_init_range(block_name
, host_addr
, offset
, length
);
216 * We need the whole of RAM to be truly empty for postcopy, so things
217 * like ROMs and any data tables built during init must be zero'd
218 * - we're going to get the copy from the source anyway.
219 * (Precopy will just overwrite this data, so doesn't need the discard)
221 if (postcopy_ram_discard_range(mis
, host_addr
, length
)) {
229 * At the end of migration, undo the effects of init_range
230 * opaque should be the MIS.
232 static int cleanup_range(const char *block_name
, void *host_addr
,
233 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
235 MigrationIncomingState
*mis
= opaque
;
236 struct uffdio_range range_struct
;
237 trace_postcopy_cleanup_range(block_name
, host_addr
, offset
, length
);
240 * We turned off hugepage for the precopy stage with postcopy enabled
241 * we can turn it back on now.
243 if (qemu_madvise(host_addr
, length
, QEMU_MADV_HUGEPAGE
)) {
244 error_report("%s HUGEPAGE: %s", __func__
, strerror(errno
));
249 * We can also turn off userfault now since we should have all the
250 * pages. It can be useful to leave it on to debug postcopy
251 * if you're not sure it's always getting every page.
253 range_struct
.start
= (uintptr_t)host_addr
;
254 range_struct
.len
= length
;
256 if (ioctl(mis
->userfault_fd
, UFFDIO_UNREGISTER
, &range_struct
)) {
257 error_report("%s: userfault unregister %s", __func__
, strerror(errno
));
266 * Initialise postcopy-ram, setting the RAM to a state where we can go into
267 * postcopy later; must be called prior to any precopy.
268 * called from arch_init's similarly named ram_postcopy_incoming_init
270 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
272 if (qemu_ram_foreach_block(init_range
, mis
)) {
280 * At the end of a migration where postcopy_ram_incoming_init was called.
282 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
284 trace_postcopy_ram_incoming_cleanup_entry();
286 if (mis
->have_fault_thread
) {
289 if (qemu_ram_foreach_block(cleanup_range
, mis
)) {
293 * Tell the fault_thread to exit, it's an eventfd that should
294 * currently be at 0, we're going to increment it to 1
297 if (write(mis
->userfault_quit_fd
, &tmp64
, 8) == 8) {
298 trace_postcopy_ram_incoming_cleanup_join();
299 qemu_thread_join(&mis
->fault_thread
);
301 /* Not much we can do here, but may as well report it */
302 error_report("%s: incrementing userfault_quit_fd: %s", __func__
,
305 trace_postcopy_ram_incoming_cleanup_closeuf();
306 close(mis
->userfault_fd
);
307 close(mis
->userfault_quit_fd
);
308 mis
->have_fault_thread
= 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 if (qemu_madvise(host_addr
, length
, QEMU_MADV_NOHUGEPAGE
)) {
346 error_report("%s: NOHUGEPAGE: %s", __func__
, strerror(errno
));
354 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
355 * however leaving it until after precopy means that most of the precopy
358 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
360 if (qemu_ram_foreach_block(nhp_range
, mis
)) {
364 postcopy_state_set(POSTCOPY_INCOMING_DISCARD
);
370 * Mark the given area of RAM as requiring notification to unwritten areas
371 * Used as a callback on qemu_ram_foreach_block.
372 * host_addr: Base of area to mark
373 * offset: Offset in the whole ram arena
374 * length: Length of the section
375 * opaque: MigrationIncomingState pointer
376 * Returns 0 on success
378 static int ram_block_enable_notify(const char *block_name
, void *host_addr
,
379 ram_addr_t offset
, ram_addr_t length
,
382 MigrationIncomingState
*mis
= opaque
;
383 struct uffdio_register reg_struct
;
385 reg_struct
.range
.start
= (uintptr_t)host_addr
;
386 reg_struct
.range
.len
= length
;
387 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
389 /* Now tell our userfault_fd that it's responsible for this area */
390 if (ioctl(mis
->userfault_fd
, UFFDIO_REGISTER
, ®_struct
)) {
391 error_report("%s userfault register: %s", __func__
, strerror(errno
));
399 * Handle faults detected by the USERFAULT markings
401 static void *postcopy_ram_fault_thread(void *opaque
)
403 MigrationIncomingState
*mis
= opaque
;
406 size_t hostpagesize
= getpagesize();
408 RAMBlock
*last_rb
= NULL
; /* last RAMBlock we sent part of */
410 trace_postcopy_ram_fault_thread_entry();
411 qemu_sem_post(&mis
->fault_thread_sem
);
414 ram_addr_t rb_offset
;
415 ram_addr_t in_raspace
;
416 struct pollfd pfd
[2];
419 * We're mainly waiting for the kernel to give us a faulting HVA,
420 * however we can be told to quit via userfault_quit_fd which is
423 pfd
[0].fd
= mis
->userfault_fd
;
424 pfd
[0].events
= POLLIN
;
426 pfd
[1].fd
= mis
->userfault_quit_fd
;
427 pfd
[1].events
= POLLIN
; /* Waiting for eventfd to go positive */
430 if (poll(pfd
, 2, -1 /* Wait forever */) == -1) {
431 error_report("%s: userfault poll: %s", __func__
, strerror(errno
));
435 if (pfd
[1].revents
) {
436 trace_postcopy_ram_fault_thread_quit();
440 ret
= read(mis
->userfault_fd
, &msg
, sizeof(msg
));
441 if (ret
!= sizeof(msg
)) {
442 if (errno
== EAGAIN
) {
444 * if a wake up happens on the other thread just after
445 * the poll, there is nothing to read.
450 error_report("%s: Failed to read full userfault message: %s",
451 __func__
, strerror(errno
));
454 error_report("%s: Read %d bytes from userfaultfd expected %zd",
455 __func__
, ret
, sizeof(msg
));
456 break; /* Lost alignment, don't know what we'd read next */
459 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
) {
460 error_report("%s: Read unexpected event %ud from userfaultfd",
461 __func__
, msg
.event
);
462 continue; /* It's not a page fault, shouldn't happen */
465 rb
= qemu_ram_block_from_host(
466 (void *)(uintptr_t)msg
.arg
.pagefault
.address
,
467 true, &in_raspace
, &rb_offset
);
469 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
470 PRIx64
, (uint64_t)msg
.arg
.pagefault
.address
);
474 rb_offset
&= ~(hostpagesize
- 1);
475 trace_postcopy_ram_fault_thread_request(msg
.arg
.pagefault
.address
,
476 qemu_ram_get_idstr(rb
),
480 * Send the request to the source - we want to request one
481 * of our host page sizes (which is >= TPS)
485 migrate_send_rp_req_pages(mis
, qemu_ram_get_idstr(rb
),
486 rb_offset
, hostpagesize
);
488 /* Save some space */
489 migrate_send_rp_req_pages(mis
, NULL
,
490 rb_offset
, hostpagesize
);
493 trace_postcopy_ram_fault_thread_exit();
497 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
499 /* Open the fd for the kernel to give us userfaults */
500 mis
->userfault_fd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
501 if (mis
->userfault_fd
== -1) {
502 error_report("%s: Failed to open userfault fd: %s", __func__
,
508 * Although the host check already tested the API, we need to
509 * do the check again as an ABI handshake on the new fd.
511 if (!ufd_version_check(mis
->userfault_fd
)) {
515 /* Now an eventfd we use to tell the fault-thread to quit */
516 mis
->userfault_quit_fd
= eventfd(0, EFD_CLOEXEC
);
517 if (mis
->userfault_quit_fd
== -1) {
518 error_report("%s: Opening userfault_quit_fd: %s", __func__
,
520 close(mis
->userfault_fd
);
524 qemu_sem_init(&mis
->fault_thread_sem
, 0);
525 qemu_thread_create(&mis
->fault_thread
, "postcopy/fault",
526 postcopy_ram_fault_thread
, mis
, QEMU_THREAD_JOINABLE
);
527 qemu_sem_wait(&mis
->fault_thread_sem
);
528 qemu_sem_destroy(&mis
->fault_thread_sem
);
529 mis
->have_fault_thread
= true;
531 /* Mark so that we get notified of accesses to unwritten areas */
532 if (qemu_ram_foreach_block(ram_block_enable_notify
, mis
)) {
536 trace_postcopy_ram_enable_notify();
542 * Place a host page (from) at (host) atomically
543 * returns 0 on success
545 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
)
547 struct uffdio_copy copy_struct
;
549 copy_struct
.dst
= (uint64_t)(uintptr_t)host
;
550 copy_struct
.src
= (uint64_t)(uintptr_t)from
;
551 copy_struct
.len
= getpagesize();
552 copy_struct
.mode
= 0;
554 /* copy also acks to the kernel waking the stalled thread up
555 * TODO: We can inhibit that ack and only do it if it was requested
556 * which would be slightly cheaper, but we'd have to be careful
557 * of the order of updating our page state.
559 if (ioctl(mis
->userfault_fd
, UFFDIO_COPY
, ©_struct
)) {
561 error_report("%s: %s copy host: %p from: %p",
562 __func__
, strerror(e
), host
, from
);
567 trace_postcopy_place_page(host
);
572 * Place a zero page at (host) atomically
573 * returns 0 on success
575 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
)
577 struct uffdio_zeropage zero_struct
;
579 zero_struct
.range
.start
= (uint64_t)(uintptr_t)host
;
580 zero_struct
.range
.len
= getpagesize();
581 zero_struct
.mode
= 0;
583 if (ioctl(mis
->userfault_fd
, UFFDIO_ZEROPAGE
, &zero_struct
)) {
585 error_report("%s: %s zero host: %p",
586 __func__
, strerror(e
), host
);
591 trace_postcopy_place_page_zero(host
);
596 * Returns a target page of memory that can be mapped at a later point in time
597 * using postcopy_place_page
598 * The same address is used repeatedly, postcopy_place_page just takes the
600 * Returns: Pointer to allocated page
603 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
605 if (!mis
->postcopy_tmp_page
) {
606 mis
->postcopy_tmp_page
= mmap(NULL
, getpagesize(),
607 PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
608 MAP_ANONYMOUS
, -1, 0);
609 if (!mis
->postcopy_tmp_page
) {
610 error_report("%s: %s", __func__
, strerror(errno
));
615 return mis
->postcopy_tmp_page
;
619 /* No target OS support, stubs just fail */
620 bool postcopy_ram_supported_by_host(void)
622 error_report("%s: No OS support", __func__
);
626 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
628 error_report("postcopy_ram_incoming_init: No OS support");
632 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
638 int postcopy_ram_discard_range(MigrationIncomingState
*mis
, uint8_t *start
,
645 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
651 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
657 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
)
663 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
)
669 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
677 /* ------------------------------------------------------------------------- */
680 * postcopy_discard_send_init: Called at the start of each RAMBlock before
681 * asking to discard individual ranges.
683 * @ms: The current migration state.
684 * @offset: the bitmap offset of the named RAMBlock in the migration
686 * @name: RAMBlock that discards will operate on.
688 * returns: a new PDS.
690 PostcopyDiscardState
*postcopy_discard_send_init(MigrationState
*ms
,
691 unsigned long offset
,
694 PostcopyDiscardState
*res
= g_malloc0(sizeof(PostcopyDiscardState
));
697 res
->ramblock_name
= name
;
698 res
->offset
= offset
;
705 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
706 * discard. May send a discard message, may just leave it queued to
709 * @ms: Current migration state.
710 * @pds: Structure initialised by postcopy_discard_send_init().
711 * @start,@length: a range of pages in the migration bitmap in the
712 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
714 void postcopy_discard_send_range(MigrationState
*ms
, PostcopyDiscardState
*pds
,
715 unsigned long start
, unsigned long length
)
717 size_t tp_bits
= qemu_target_page_bits();
718 /* Convert to byte offsets within the RAM block */
719 pds
->start_list
[pds
->cur_entry
] = (start
- pds
->offset
) << tp_bits
;
720 pds
->length_list
[pds
->cur_entry
] = length
<< tp_bits
;
721 trace_postcopy_discard_send_range(pds
->ramblock_name
, start
, length
);
725 if (pds
->cur_entry
== MAX_DISCARDS_PER_COMMAND
) {
726 /* Full set, ship it! */
727 qemu_savevm_send_postcopy_ram_discard(ms
->file
, pds
->ramblock_name
,
737 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
738 * bitmap code. Sends any outstanding discard messages, frees the PDS
740 * @ms: Current migration state.
741 * @pds: Structure initialised by postcopy_discard_send_init().
743 void postcopy_discard_send_finish(MigrationState
*ms
, PostcopyDiscardState
*pds
)
745 /* Anything unsent? */
746 if (pds
->cur_entry
) {
747 qemu_savevm_send_postcopy_ram_discard(ms
->file
, pds
->ramblock_name
,
754 trace_postcopy_discard_send_finish(pds
->ramblock_name
, pds
->nsentwords
,