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
));
88 * Note: This has the side effect of munlock'ing all of RAM, that's
89 * normally fine since if the postcopy succeeds it gets turned back on at the
92 bool postcopy_ram_supported_by_host(void)
94 long pagesize
= getpagesize();
96 bool ret
= false; /* Error unless we change it */
97 void *testarea
= NULL
;
98 struct uffdio_register reg_struct
;
99 struct uffdio_range range_struct
;
100 uint64_t feature_mask
;
102 if ((1ul << qemu_target_page_bits()) > pagesize
) {
103 error_report("Target page size bigger than host page size");
107 ufd
= syscall(__NR_userfaultfd
, O_CLOEXEC
);
109 error_report("%s: userfaultfd not available: %s", __func__
,
114 /* Version and features check */
115 if (!ufd_version_check(ufd
)) {
118 /* TODO: Only allow huge pages if the kernel supports it */
121 * userfault and mlock don't go together; we'll put it back later if
125 error_report("%s: munlockall: %s", __func__
, strerror(errno
));
130 * We need to check that the ops we need are supported on anon memory
131 * To do that we need to register a chunk and see the flags that
134 testarea
= mmap(NULL
, pagesize
, PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
135 MAP_ANONYMOUS
, -1, 0);
136 if (testarea
== MAP_FAILED
) {
137 error_report("%s: Failed to map test area: %s", __func__
,
141 g_assert(((size_t)testarea
& (pagesize
-1)) == 0);
143 reg_struct
.range
.start
= (uintptr_t)testarea
;
144 reg_struct
.range
.len
= pagesize
;
145 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
147 if (ioctl(ufd
, UFFDIO_REGISTER
, ®_struct
)) {
148 error_report("%s userfault register: %s", __func__
, strerror(errno
));
152 range_struct
.start
= (uintptr_t)testarea
;
153 range_struct
.len
= pagesize
;
154 if (ioctl(ufd
, UFFDIO_UNREGISTER
, &range_struct
)) {
155 error_report("%s userfault unregister: %s", __func__
, strerror(errno
));
159 feature_mask
= (__u64
)1 << _UFFDIO_WAKE
|
160 (__u64
)1 << _UFFDIO_COPY
|
161 (__u64
)1 << _UFFDIO_ZEROPAGE
;
162 if ((reg_struct
.ioctls
& feature_mask
) != feature_mask
) {
163 error_report("Missing userfault map features: %" PRIx64
,
164 (uint64_t)(~reg_struct
.ioctls
& feature_mask
));
172 munmap(testarea
, pagesize
);
181 * Setup an area of RAM so that it *can* be used for postcopy later; this
182 * must be done right at the start prior to pre-copy.
183 * opaque should be the MIS.
185 static int init_range(const char *block_name
, void *host_addr
,
186 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
188 MigrationIncomingState
*mis
= opaque
;
190 trace_postcopy_init_range(block_name
, host_addr
, offset
, length
);
193 * We need the whole of RAM to be truly empty for postcopy, so things
194 * like ROMs and any data tables built during init must be zero'd
195 * - we're going to get the copy from the source anyway.
196 * (Precopy will just overwrite this data, so doesn't need the discard)
198 if (ram_discard_range(mis
, block_name
, 0, length
)) {
206 * At the end of migration, undo the effects of init_range
207 * opaque should be the MIS.
209 static int cleanup_range(const char *block_name
, void *host_addr
,
210 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
212 MigrationIncomingState
*mis
= opaque
;
213 struct uffdio_range range_struct
;
214 trace_postcopy_cleanup_range(block_name
, host_addr
, offset
, length
);
217 * We turned off hugepage for the precopy stage with postcopy enabled
218 * we can turn it back on now.
220 qemu_madvise(host_addr
, length
, QEMU_MADV_HUGEPAGE
);
223 * We can also turn off userfault now since we should have all the
224 * pages. It can be useful to leave it on to debug postcopy
225 * if you're not sure it's always getting every page.
227 range_struct
.start
= (uintptr_t)host_addr
;
228 range_struct
.len
= length
;
230 if (ioctl(mis
->userfault_fd
, UFFDIO_UNREGISTER
, &range_struct
)) {
231 error_report("%s: userfault unregister %s", __func__
, strerror(errno
));
240 * Initialise postcopy-ram, setting the RAM to a state where we can go into
241 * postcopy later; must be called prior to any precopy.
242 * called from arch_init's similarly named ram_postcopy_incoming_init
244 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
246 if (qemu_ram_foreach_block(init_range
, mis
)) {
254 * At the end of a migration where postcopy_ram_incoming_init was called.
256 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
258 trace_postcopy_ram_incoming_cleanup_entry();
260 if (mis
->have_fault_thread
) {
263 if (qemu_ram_foreach_block(cleanup_range
, mis
)) {
267 * Tell the fault_thread to exit, it's an eventfd that should
268 * currently be at 0, we're going to increment it to 1
271 if (write(mis
->userfault_quit_fd
, &tmp64
, 8) == 8) {
272 trace_postcopy_ram_incoming_cleanup_join();
273 qemu_thread_join(&mis
->fault_thread
);
275 /* Not much we can do here, but may as well report it */
276 error_report("%s: incrementing userfault_quit_fd: %s", __func__
,
279 trace_postcopy_ram_incoming_cleanup_closeuf();
280 close(mis
->userfault_fd
);
281 close(mis
->userfault_quit_fd
);
282 mis
->have_fault_thread
= false;
285 qemu_balloon_inhibit(false);
288 if (os_mlock() < 0) {
289 error_report("mlock: %s", strerror(errno
));
291 * It doesn't feel right to fail at this point, we have a valid
297 postcopy_state_set(POSTCOPY_INCOMING_END
);
298 migrate_send_rp_shut(mis
, qemu_file_get_error(mis
->from_src_file
) != 0);
300 if (mis
->postcopy_tmp_page
) {
301 munmap(mis
->postcopy_tmp_page
, mis
->largest_page_size
);
302 mis
->postcopy_tmp_page
= NULL
;
304 if (mis
->postcopy_tmp_zero_page
) {
305 munmap(mis
->postcopy_tmp_zero_page
, mis
->largest_page_size
);
306 mis
->postcopy_tmp_zero_page
= NULL
;
308 trace_postcopy_ram_incoming_cleanup_exit();
313 * Disable huge pages on an area
315 static int nhp_range(const char *block_name
, void *host_addr
,
316 ram_addr_t offset
, ram_addr_t length
, void *opaque
)
318 trace_postcopy_nhp_range(block_name
, host_addr
, offset
, length
);
321 * Before we do discards we need to ensure those discards really
322 * do delete areas of the page, even if THP thinks a hugepage would
323 * be a good idea, so force hugepages off.
325 qemu_madvise(host_addr
, length
, QEMU_MADV_NOHUGEPAGE
);
331 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
332 * however leaving it until after precopy means that most of the precopy
335 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
337 if (qemu_ram_foreach_block(nhp_range
, mis
)) {
341 postcopy_state_set(POSTCOPY_INCOMING_DISCARD
);
347 * Mark the given area of RAM as requiring notification to unwritten areas
348 * Used as a callback on qemu_ram_foreach_block.
349 * host_addr: Base of area to mark
350 * offset: Offset in the whole ram arena
351 * length: Length of the section
352 * opaque: MigrationIncomingState pointer
353 * Returns 0 on success
355 static int ram_block_enable_notify(const char *block_name
, void *host_addr
,
356 ram_addr_t offset
, ram_addr_t length
,
359 MigrationIncomingState
*mis
= opaque
;
360 struct uffdio_register reg_struct
;
362 reg_struct
.range
.start
= (uintptr_t)host_addr
;
363 reg_struct
.range
.len
= length
;
364 reg_struct
.mode
= UFFDIO_REGISTER_MODE_MISSING
;
366 /* Now tell our userfault_fd that it's responsible for this area */
367 if (ioctl(mis
->userfault_fd
, UFFDIO_REGISTER
, ®_struct
)) {
368 error_report("%s userfault register: %s", __func__
, strerror(errno
));
376 * Handle faults detected by the USERFAULT markings
378 static void *postcopy_ram_fault_thread(void *opaque
)
380 MigrationIncomingState
*mis
= opaque
;
384 RAMBlock
*last_rb
= NULL
; /* last RAMBlock we sent part of */
386 trace_postcopy_ram_fault_thread_entry();
387 qemu_sem_post(&mis
->fault_thread_sem
);
390 ram_addr_t rb_offset
;
391 struct pollfd pfd
[2];
394 * We're mainly waiting for the kernel to give us a faulting HVA,
395 * however we can be told to quit via userfault_quit_fd which is
398 pfd
[0].fd
= mis
->userfault_fd
;
399 pfd
[0].events
= POLLIN
;
401 pfd
[1].fd
= mis
->userfault_quit_fd
;
402 pfd
[1].events
= POLLIN
; /* Waiting for eventfd to go positive */
405 if (poll(pfd
, 2, -1 /* Wait forever */) == -1) {
406 error_report("%s: userfault poll: %s", __func__
, strerror(errno
));
410 if (pfd
[1].revents
) {
411 trace_postcopy_ram_fault_thread_quit();
415 ret
= read(mis
->userfault_fd
, &msg
, sizeof(msg
));
416 if (ret
!= sizeof(msg
)) {
417 if (errno
== EAGAIN
) {
419 * if a wake up happens on the other thread just after
420 * the poll, there is nothing to read.
425 error_report("%s: Failed to read full userfault message: %s",
426 __func__
, strerror(errno
));
429 error_report("%s: Read %d bytes from userfaultfd expected %zd",
430 __func__
, ret
, sizeof(msg
));
431 break; /* Lost alignment, don't know what we'd read next */
434 if (msg
.event
!= UFFD_EVENT_PAGEFAULT
) {
435 error_report("%s: Read unexpected event %ud from userfaultfd",
436 __func__
, msg
.event
);
437 continue; /* It's not a page fault, shouldn't happen */
440 rb
= qemu_ram_block_from_host(
441 (void *)(uintptr_t)msg
.arg
.pagefault
.address
,
444 error_report("postcopy_ram_fault_thread: Fault outside guest: %"
445 PRIx64
, (uint64_t)msg
.arg
.pagefault
.address
);
449 rb_offset
&= ~(qemu_ram_pagesize(rb
) - 1);
450 trace_postcopy_ram_fault_thread_request(msg
.arg
.pagefault
.address
,
451 qemu_ram_get_idstr(rb
),
455 * Send the request to the source - we want to request one
456 * of our host page sizes (which is >= TPS)
460 migrate_send_rp_req_pages(mis
, qemu_ram_get_idstr(rb
),
461 rb_offset
, qemu_ram_pagesize(rb
));
463 /* Save some space */
464 migrate_send_rp_req_pages(mis
, NULL
,
465 rb_offset
, qemu_ram_pagesize(rb
));
468 trace_postcopy_ram_fault_thread_exit();
472 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
474 /* Open the fd for the kernel to give us userfaults */
475 mis
->userfault_fd
= syscall(__NR_userfaultfd
, O_CLOEXEC
| O_NONBLOCK
);
476 if (mis
->userfault_fd
== -1) {
477 error_report("%s: Failed to open userfault fd: %s", __func__
,
483 * Although the host check already tested the API, we need to
484 * do the check again as an ABI handshake on the new fd.
486 if (!ufd_version_check(mis
->userfault_fd
)) {
490 /* Now an eventfd we use to tell the fault-thread to quit */
491 mis
->userfault_quit_fd
= eventfd(0, EFD_CLOEXEC
);
492 if (mis
->userfault_quit_fd
== -1) {
493 error_report("%s: Opening userfault_quit_fd: %s", __func__
,
495 close(mis
->userfault_fd
);
499 qemu_sem_init(&mis
->fault_thread_sem
, 0);
500 qemu_thread_create(&mis
->fault_thread
, "postcopy/fault",
501 postcopy_ram_fault_thread
, mis
, QEMU_THREAD_JOINABLE
);
502 qemu_sem_wait(&mis
->fault_thread_sem
);
503 qemu_sem_destroy(&mis
->fault_thread_sem
);
504 mis
->have_fault_thread
= true;
506 /* Mark so that we get notified of accesses to unwritten areas */
507 if (qemu_ram_foreach_block(ram_block_enable_notify
, mis
)) {
512 * Ballooning can mark pages as absent while we're postcopying
513 * that would cause false userfaults.
515 qemu_balloon_inhibit(true);
517 trace_postcopy_ram_enable_notify();
523 * Place a host page (from) at (host) atomically
524 * returns 0 on success
526 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
529 struct uffdio_copy copy_struct
;
531 copy_struct
.dst
= (uint64_t)(uintptr_t)host
;
532 copy_struct
.src
= (uint64_t)(uintptr_t)from
;
533 copy_struct
.len
= pagesize
;
534 copy_struct
.mode
= 0;
536 /* copy also acks to the kernel waking the stalled thread up
537 * TODO: We can inhibit that ack and only do it if it was requested
538 * which would be slightly cheaper, but we'd have to be careful
539 * of the order of updating our page state.
541 if (ioctl(mis
->userfault_fd
, UFFDIO_COPY
, ©_struct
)) {
543 error_report("%s: %s copy host: %p from: %p (size: %zd)",
544 __func__
, strerror(e
), host
, from
, pagesize
);
549 trace_postcopy_place_page(host
);
554 * Place a zero page at (host) atomically
555 * returns 0 on success
557 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
560 trace_postcopy_place_page_zero(host
);
562 if (pagesize
== getpagesize()) {
563 struct uffdio_zeropage zero_struct
;
564 zero_struct
.range
.start
= (uint64_t)(uintptr_t)host
;
565 zero_struct
.range
.len
= getpagesize();
566 zero_struct
.mode
= 0;
568 if (ioctl(mis
->userfault_fd
, UFFDIO_ZEROPAGE
, &zero_struct
)) {
570 error_report("%s: %s zero host: %p",
571 __func__
, strerror(e
), host
);
576 /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */
577 if (!mis
->postcopy_tmp_zero_page
) {
578 mis
->postcopy_tmp_zero_page
= mmap(NULL
, mis
->largest_page_size
,
579 PROT_READ
| PROT_WRITE
,
580 MAP_PRIVATE
| MAP_ANONYMOUS
,
582 if (mis
->postcopy_tmp_zero_page
== MAP_FAILED
) {
584 mis
->postcopy_tmp_zero_page
= NULL
;
585 error_report("%s: %s mapping large zero page",
586 __func__
, strerror(e
));
589 memset(mis
->postcopy_tmp_zero_page
, '\0', mis
->largest_page_size
);
591 return postcopy_place_page(mis
, host
, mis
->postcopy_tmp_zero_page
,
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
, mis
->largest_page_size
,
610 PROT_READ
| PROT_WRITE
, MAP_PRIVATE
|
611 MAP_ANONYMOUS
, -1, 0);
612 if (mis
->postcopy_tmp_page
== MAP_FAILED
) {
613 mis
->postcopy_tmp_page
= NULL
;
614 error_report("%s: %s", __func__
, strerror(errno
));
619 return mis
->postcopy_tmp_page
;
623 /* No target OS support, stubs just fail */
624 bool postcopy_ram_supported_by_host(void)
626 error_report("%s: No OS support", __func__
);
630 int postcopy_ram_incoming_init(MigrationIncomingState
*mis
, size_t ram_pages
)
632 error_report("postcopy_ram_incoming_init: No OS support");
636 int postcopy_ram_incoming_cleanup(MigrationIncomingState
*mis
)
642 int postcopy_ram_prepare_discard(MigrationIncomingState
*mis
)
648 int postcopy_ram_enable_notify(MigrationIncomingState
*mis
)
654 int postcopy_place_page(MigrationIncomingState
*mis
, void *host
, void *from
,
661 int postcopy_place_page_zero(MigrationIncomingState
*mis
, void *host
,
668 void *postcopy_get_tmp_page(MigrationIncomingState
*mis
)
676 /* ------------------------------------------------------------------------- */
679 * postcopy_discard_send_init: Called at the start of each RAMBlock before
680 * asking to discard individual ranges.
682 * @ms: The current migration state.
683 * @offset: the bitmap offset of the named RAMBlock in the migration
685 * @name: RAMBlock that discards will operate on.
687 * returns: a new PDS.
689 PostcopyDiscardState
*postcopy_discard_send_init(MigrationState
*ms
,
690 unsigned long offset
,
693 PostcopyDiscardState
*res
= g_malloc0(sizeof(PostcopyDiscardState
));
696 res
->ramblock_name
= name
;
697 res
->offset
= offset
;
704 * postcopy_discard_send_range: Called by the bitmap code for each chunk to
705 * discard. May send a discard message, may just leave it queued to
708 * @ms: Current migration state.
709 * @pds: Structure initialised by postcopy_discard_send_init().
710 * @start,@length: a range of pages in the migration bitmap in the
711 * RAM block passed to postcopy_discard_send_init() (length=1 is one page)
713 void postcopy_discard_send_range(MigrationState
*ms
, PostcopyDiscardState
*pds
,
714 unsigned long start
, unsigned long length
)
716 size_t tp_bits
= qemu_target_page_bits();
717 /* Convert to byte offsets within the RAM block */
718 pds
->start_list
[pds
->cur_entry
] = (start
- pds
->offset
) << tp_bits
;
719 pds
->length_list
[pds
->cur_entry
] = length
<< tp_bits
;
720 trace_postcopy_discard_send_range(pds
->ramblock_name
, start
, length
);
724 if (pds
->cur_entry
== MAX_DISCARDS_PER_COMMAND
) {
725 /* Full set, ship it! */
726 qemu_savevm_send_postcopy_ram_discard(ms
->to_dst_file
,
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
->to_dst_file
,
755 trace_postcopy_discard_send_finish(pds
->ramblock_name
, pds
->nsentwords
,