]>
Commit | Line | Data |
---|---|---|
eb59db53 DDAG |
1 | /* |
2 | * Postcopy migration for RAM | |
3 | * | |
4 | * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates | |
5 | * | |
6 | * Authors: | |
7 | * Dave Gilbert <dgilbert@redhat.com> | |
8 | * | |
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. | |
11 | * | |
12 | */ | |
13 | ||
14 | /* | |
15 | * Postcopy is a migration technique where the execution flips from the | |
16 | * source to the destination before all the data has been copied. | |
17 | */ | |
18 | ||
1393a485 | 19 | #include "qemu/osdep.h" |
eb59db53 DDAG |
20 | |
21 | #include "qemu-common.h" | |
22 | #include "migration/migration.h" | |
23 | #include "migration/postcopy-ram.h" | |
24 | #include "sysemu/sysemu.h" | |
371ff5a3 | 25 | #include "sysemu/balloon.h" |
eb59db53 DDAG |
26 | #include "qemu/error-report.h" |
27 | #include "trace.h" | |
28 | ||
e0b266f0 DDAG |
29 | /* Arbitrary limit on size of each discard command, |
30 | * keeps them around ~200 bytes | |
31 | */ | |
32 | #define MAX_DISCARDS_PER_COMMAND 12 | |
33 | ||
34 | struct PostcopyDiscardState { | |
35 | const char *ramblock_name; | |
36 | uint64_t offset; /* Bitmap entry for the 1st bit of this RAMBlock */ | |
37 | uint16_t cur_entry; | |
38 | /* | |
39 | * Start and length of a discard range (bytes) | |
40 | */ | |
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; | |
45 | }; | |
46 | ||
eb59db53 DDAG |
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. | |
50 | */ | |
51 | #if defined(__linux__) | |
52 | ||
c4faeed2 | 53 | #include <poll.h> |
eb59db53 DDAG |
54 | #include <sys/ioctl.h> |
55 | #include <sys/syscall.h> | |
eb59db53 DDAG |
56 | #include <asm/types.h> /* for __u64 */ |
57 | #endif | |
58 | ||
d8b9d771 MF |
59 | #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD) |
60 | #include <sys/eventfd.h> | |
eb59db53 DDAG |
61 | #include <linux/userfaultfd.h> |
62 | ||
63 | static bool ufd_version_check(int ufd) | |
64 | { | |
65 | struct uffdio_api api_struct; | |
66 | uint64_t ioctl_mask; | |
67 | ||
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", | |
72 | strerror(errno)); | |
73 | return false; | |
74 | } | |
75 | ||
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)); | |
81 | return false; | |
82 | } | |
83 | ||
7e8cafb7 DDAG |
84 | if (getpagesize() != ram_pagesize_summary()) { |
85 | bool have_hp = false; | |
86 | /* We've got a huge page */ | |
87 | #ifdef UFFD_FEATURE_MISSING_HUGETLBFS | |
88 | have_hp = api_struct.features & UFFD_FEATURE_MISSING_HUGETLBFS; | |
89 | #endif | |
90 | if (!have_hp) { | |
91 | error_report("Userfault on this host does not support huge pages"); | |
92 | return false; | |
93 | } | |
94 | } | |
eb59db53 DDAG |
95 | return true; |
96 | } | |
97 | ||
58b7c17e DDAG |
98 | /* |
99 | * Note: This has the side effect of munlock'ing all of RAM, that's | |
100 | * normally fine since if the postcopy succeeds it gets turned back on at the | |
101 | * end. | |
102 | */ | |
eb59db53 DDAG |
103 | bool postcopy_ram_supported_by_host(void) |
104 | { | |
105 | long pagesize = getpagesize(); | |
106 | int ufd = -1; | |
107 | bool ret = false; /* Error unless we change it */ | |
108 | void *testarea = NULL; | |
109 | struct uffdio_register reg_struct; | |
110 | struct uffdio_range range_struct; | |
111 | uint64_t feature_mask; | |
112 | ||
113 | if ((1ul << qemu_target_page_bits()) > pagesize) { | |
114 | error_report("Target page size bigger than host page size"); | |
115 | goto out; | |
116 | } | |
117 | ||
118 | ufd = syscall(__NR_userfaultfd, O_CLOEXEC); | |
119 | if (ufd == -1) { | |
120 | error_report("%s: userfaultfd not available: %s", __func__, | |
121 | strerror(errno)); | |
122 | goto out; | |
123 | } | |
124 | ||
125 | /* Version and features check */ | |
126 | if (!ufd_version_check(ufd)) { | |
127 | goto out; | |
128 | } | |
129 | ||
58b7c17e DDAG |
130 | /* |
131 | * userfault and mlock don't go together; we'll put it back later if | |
132 | * it was enabled. | |
133 | */ | |
134 | if (munlockall()) { | |
135 | error_report("%s: munlockall: %s", __func__, strerror(errno)); | |
136 | return -1; | |
137 | } | |
138 | ||
eb59db53 DDAG |
139 | /* |
140 | * We need to check that the ops we need are supported on anon memory | |
141 | * To do that we need to register a chunk and see the flags that | |
142 | * are returned. | |
143 | */ | |
144 | testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | | |
145 | MAP_ANONYMOUS, -1, 0); | |
146 | if (testarea == MAP_FAILED) { | |
147 | error_report("%s: Failed to map test area: %s", __func__, | |
148 | strerror(errno)); | |
149 | goto out; | |
150 | } | |
151 | g_assert(((size_t)testarea & (pagesize-1)) == 0); | |
152 | ||
153 | reg_struct.range.start = (uintptr_t)testarea; | |
154 | reg_struct.range.len = pagesize; | |
155 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; | |
156 | ||
157 | if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) { | |
158 | error_report("%s userfault register: %s", __func__, strerror(errno)); | |
159 | goto out; | |
160 | } | |
161 | ||
162 | range_struct.start = (uintptr_t)testarea; | |
163 | range_struct.len = pagesize; | |
164 | if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) { | |
165 | error_report("%s userfault unregister: %s", __func__, strerror(errno)); | |
166 | goto out; | |
167 | } | |
168 | ||
169 | feature_mask = (__u64)1 << _UFFDIO_WAKE | | |
170 | (__u64)1 << _UFFDIO_COPY | | |
171 | (__u64)1 << _UFFDIO_ZEROPAGE; | |
172 | if ((reg_struct.ioctls & feature_mask) != feature_mask) { | |
173 | error_report("Missing userfault map features: %" PRIx64, | |
174 | (uint64_t)(~reg_struct.ioctls & feature_mask)); | |
175 | goto out; | |
176 | } | |
177 | ||
178 | /* Success! */ | |
179 | ret = true; | |
180 | out: | |
181 | if (testarea) { | |
182 | munmap(testarea, pagesize); | |
183 | } | |
184 | if (ufd != -1) { | |
185 | close(ufd); | |
186 | } | |
187 | return ret; | |
188 | } | |
189 | ||
1caddf8a DDAG |
190 | /* |
191 | * Setup an area of RAM so that it *can* be used for postcopy later; this | |
192 | * must be done right at the start prior to pre-copy. | |
193 | * opaque should be the MIS. | |
194 | */ | |
195 | static int init_range(const char *block_name, void *host_addr, | |
196 | ram_addr_t offset, ram_addr_t length, void *opaque) | |
197 | { | |
198 | MigrationIncomingState *mis = opaque; | |
199 | ||
200 | trace_postcopy_init_range(block_name, host_addr, offset, length); | |
201 | ||
202 | /* | |
203 | * We need the whole of RAM to be truly empty for postcopy, so things | |
204 | * like ROMs and any data tables built during init must be zero'd | |
205 | * - we're going to get the copy from the source anyway. | |
206 | * (Precopy will just overwrite this data, so doesn't need the discard) | |
207 | */ | |
d3a5038c | 208 | if (ram_discard_range(mis, block_name, 0, length)) { |
1caddf8a DDAG |
209 | return -1; |
210 | } | |
211 | ||
212 | return 0; | |
213 | } | |
214 | ||
215 | /* | |
216 | * At the end of migration, undo the effects of init_range | |
217 | * opaque should be the MIS. | |
218 | */ | |
219 | static int cleanup_range(const char *block_name, void *host_addr, | |
220 | ram_addr_t offset, ram_addr_t length, void *opaque) | |
221 | { | |
222 | MigrationIncomingState *mis = opaque; | |
223 | struct uffdio_range range_struct; | |
224 | trace_postcopy_cleanup_range(block_name, host_addr, offset, length); | |
225 | ||
226 | /* | |
227 | * We turned off hugepage for the precopy stage with postcopy enabled | |
228 | * we can turn it back on now. | |
229 | */ | |
1d741439 | 230 | qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE); |
1caddf8a DDAG |
231 | |
232 | /* | |
233 | * We can also turn off userfault now since we should have all the | |
234 | * pages. It can be useful to leave it on to debug postcopy | |
235 | * if you're not sure it's always getting every page. | |
236 | */ | |
237 | range_struct.start = (uintptr_t)host_addr; | |
238 | range_struct.len = length; | |
239 | ||
240 | if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) { | |
241 | error_report("%s: userfault unregister %s", __func__, strerror(errno)); | |
242 | ||
243 | return -1; | |
244 | } | |
245 | ||
246 | return 0; | |
247 | } | |
248 | ||
249 | /* | |
250 | * Initialise postcopy-ram, setting the RAM to a state where we can go into | |
251 | * postcopy later; must be called prior to any precopy. | |
252 | * called from arch_init's similarly named ram_postcopy_incoming_init | |
253 | */ | |
254 | int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages) | |
255 | { | |
256 | if (qemu_ram_foreach_block(init_range, mis)) { | |
257 | return -1; | |
258 | } | |
259 | ||
260 | return 0; | |
261 | } | |
262 | ||
263 | /* | |
264 | * At the end of a migration where postcopy_ram_incoming_init was called. | |
265 | */ | |
266 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
267 | { | |
c4faeed2 DDAG |
268 | trace_postcopy_ram_incoming_cleanup_entry(); |
269 | ||
270 | if (mis->have_fault_thread) { | |
271 | uint64_t tmp64; | |
272 | ||
273 | if (qemu_ram_foreach_block(cleanup_range, mis)) { | |
274 | return -1; | |
275 | } | |
276 | /* | |
277 | * Tell the fault_thread to exit, it's an eventfd that should | |
278 | * currently be at 0, we're going to increment it to 1 | |
279 | */ | |
280 | tmp64 = 1; | |
281 | if (write(mis->userfault_quit_fd, &tmp64, 8) == 8) { | |
282 | trace_postcopy_ram_incoming_cleanup_join(); | |
283 | qemu_thread_join(&mis->fault_thread); | |
284 | } else { | |
285 | /* Not much we can do here, but may as well report it */ | |
286 | error_report("%s: incrementing userfault_quit_fd: %s", __func__, | |
287 | strerror(errno)); | |
288 | } | |
289 | trace_postcopy_ram_incoming_cleanup_closeuf(); | |
290 | close(mis->userfault_fd); | |
291 | close(mis->userfault_quit_fd); | |
292 | mis->have_fault_thread = false; | |
1caddf8a DDAG |
293 | } |
294 | ||
371ff5a3 DDAG |
295 | qemu_balloon_inhibit(false); |
296 | ||
58b7c17e DDAG |
297 | if (enable_mlock) { |
298 | if (os_mlock() < 0) { | |
299 | error_report("mlock: %s", strerror(errno)); | |
300 | /* | |
301 | * It doesn't feel right to fail at this point, we have a valid | |
302 | * VM state. | |
303 | */ | |
304 | } | |
305 | } | |
306 | ||
c4faeed2 DDAG |
307 | postcopy_state_set(POSTCOPY_INCOMING_END); |
308 | migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0); | |
309 | ||
696ed9a9 | 310 | if (mis->postcopy_tmp_page) { |
df9ff5e1 | 311 | munmap(mis->postcopy_tmp_page, mis->largest_page_size); |
696ed9a9 DDAG |
312 | mis->postcopy_tmp_page = NULL; |
313 | } | |
41d84210 DDAG |
314 | if (mis->postcopy_tmp_zero_page) { |
315 | munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size); | |
316 | mis->postcopy_tmp_zero_page = NULL; | |
317 | } | |
c4faeed2 | 318 | trace_postcopy_ram_incoming_cleanup_exit(); |
1caddf8a DDAG |
319 | return 0; |
320 | } | |
321 | ||
f9527107 DDAG |
322 | /* |
323 | * Disable huge pages on an area | |
324 | */ | |
325 | static int nhp_range(const char *block_name, void *host_addr, | |
326 | ram_addr_t offset, ram_addr_t length, void *opaque) | |
327 | { | |
328 | trace_postcopy_nhp_range(block_name, host_addr, offset, length); | |
329 | ||
330 | /* | |
331 | * Before we do discards we need to ensure those discards really | |
332 | * do delete areas of the page, even if THP thinks a hugepage would | |
333 | * be a good idea, so force hugepages off. | |
334 | */ | |
1d741439 | 335 | qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE); |
f9527107 DDAG |
336 | |
337 | return 0; | |
338 | } | |
339 | ||
340 | /* | |
341 | * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard | |
342 | * however leaving it until after precopy means that most of the precopy | |
343 | * data is still THPd | |
344 | */ | |
345 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) | |
346 | { | |
347 | if (qemu_ram_foreach_block(nhp_range, mis)) { | |
348 | return -1; | |
349 | } | |
350 | ||
351 | postcopy_state_set(POSTCOPY_INCOMING_DISCARD); | |
352 | ||
353 | return 0; | |
354 | } | |
355 | ||
f0a227ad DDAG |
356 | /* |
357 | * Mark the given area of RAM as requiring notification to unwritten areas | |
358 | * Used as a callback on qemu_ram_foreach_block. | |
359 | * host_addr: Base of area to mark | |
360 | * offset: Offset in the whole ram arena | |
361 | * length: Length of the section | |
362 | * opaque: MigrationIncomingState pointer | |
363 | * Returns 0 on success | |
364 | */ | |
365 | static int ram_block_enable_notify(const char *block_name, void *host_addr, | |
366 | ram_addr_t offset, ram_addr_t length, | |
367 | void *opaque) | |
368 | { | |
369 | MigrationIncomingState *mis = opaque; | |
370 | struct uffdio_register reg_struct; | |
371 | ||
372 | reg_struct.range.start = (uintptr_t)host_addr; | |
373 | reg_struct.range.len = length; | |
374 | reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; | |
375 | ||
376 | /* Now tell our userfault_fd that it's responsible for this area */ | |
377 | if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) { | |
378 | error_report("%s userfault register: %s", __func__, strerror(errno)); | |
379 | return -1; | |
380 | } | |
665414ad DDAG |
381 | if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { |
382 | error_report("%s userfault: Region doesn't support COPY", __func__); | |
383 | return -1; | |
384 | } | |
f0a227ad DDAG |
385 | |
386 | return 0; | |
387 | } | |
388 | ||
389 | /* | |
390 | * Handle faults detected by the USERFAULT markings | |
391 | */ | |
392 | static void *postcopy_ram_fault_thread(void *opaque) | |
393 | { | |
394 | MigrationIncomingState *mis = opaque; | |
c4faeed2 DDAG |
395 | struct uffd_msg msg; |
396 | int ret; | |
c4faeed2 DDAG |
397 | RAMBlock *rb = NULL; |
398 | RAMBlock *last_rb = NULL; /* last RAMBlock we sent part of */ | |
f0a227ad | 399 | |
c4faeed2 | 400 | trace_postcopy_ram_fault_thread_entry(); |
f0a227ad | 401 | qemu_sem_post(&mis->fault_thread_sem); |
f0a227ad | 402 | |
c4faeed2 DDAG |
403 | while (true) { |
404 | ram_addr_t rb_offset; | |
c4faeed2 DDAG |
405 | struct pollfd pfd[2]; |
406 | ||
407 | /* | |
408 | * We're mainly waiting for the kernel to give us a faulting HVA, | |
409 | * however we can be told to quit via userfault_quit_fd which is | |
410 | * an eventfd | |
411 | */ | |
412 | pfd[0].fd = mis->userfault_fd; | |
413 | pfd[0].events = POLLIN; | |
414 | pfd[0].revents = 0; | |
415 | pfd[1].fd = mis->userfault_quit_fd; | |
416 | pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */ | |
417 | pfd[1].revents = 0; | |
418 | ||
419 | if (poll(pfd, 2, -1 /* Wait forever */) == -1) { | |
420 | error_report("%s: userfault poll: %s", __func__, strerror(errno)); | |
421 | break; | |
422 | } | |
423 | ||
424 | if (pfd[1].revents) { | |
425 | trace_postcopy_ram_fault_thread_quit(); | |
426 | break; | |
427 | } | |
428 | ||
429 | ret = read(mis->userfault_fd, &msg, sizeof(msg)); | |
430 | if (ret != sizeof(msg)) { | |
431 | if (errno == EAGAIN) { | |
432 | /* | |
433 | * if a wake up happens on the other thread just after | |
434 | * the poll, there is nothing to read. | |
435 | */ | |
436 | continue; | |
437 | } | |
438 | if (ret < 0) { | |
439 | error_report("%s: Failed to read full userfault message: %s", | |
440 | __func__, strerror(errno)); | |
441 | break; | |
442 | } else { | |
443 | error_report("%s: Read %d bytes from userfaultfd expected %zd", | |
444 | __func__, ret, sizeof(msg)); | |
445 | break; /* Lost alignment, don't know what we'd read next */ | |
446 | } | |
447 | } | |
448 | if (msg.event != UFFD_EVENT_PAGEFAULT) { | |
449 | error_report("%s: Read unexpected event %ud from userfaultfd", | |
450 | __func__, msg.event); | |
451 | continue; /* It's not a page fault, shouldn't happen */ | |
452 | } | |
453 | ||
454 | rb = qemu_ram_block_from_host( | |
455 | (void *)(uintptr_t)msg.arg.pagefault.address, | |
f615f396 | 456 | true, &rb_offset); |
c4faeed2 DDAG |
457 | if (!rb) { |
458 | error_report("postcopy_ram_fault_thread: Fault outside guest: %" | |
459 | PRIx64, (uint64_t)msg.arg.pagefault.address); | |
460 | break; | |
461 | } | |
462 | ||
332847f0 | 463 | rb_offset &= ~(qemu_ram_pagesize(rb) - 1); |
c4faeed2 DDAG |
464 | trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, |
465 | qemu_ram_get_idstr(rb), | |
466 | rb_offset); | |
467 | ||
468 | /* | |
469 | * Send the request to the source - we want to request one | |
470 | * of our host page sizes (which is >= TPS) | |
471 | */ | |
472 | if (rb != last_rb) { | |
473 | last_rb = rb; | |
474 | migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb), | |
332847f0 | 475 | rb_offset, qemu_ram_pagesize(rb)); |
c4faeed2 DDAG |
476 | } else { |
477 | /* Save some space */ | |
478 | migrate_send_rp_req_pages(mis, NULL, | |
332847f0 | 479 | rb_offset, qemu_ram_pagesize(rb)); |
c4faeed2 DDAG |
480 | } |
481 | } | |
482 | trace_postcopy_ram_fault_thread_exit(); | |
f0a227ad DDAG |
483 | return NULL; |
484 | } | |
485 | ||
486 | int postcopy_ram_enable_notify(MigrationIncomingState *mis) | |
487 | { | |
c4faeed2 DDAG |
488 | /* Open the fd for the kernel to give us userfaults */ |
489 | mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK); | |
490 | if (mis->userfault_fd == -1) { | |
491 | error_report("%s: Failed to open userfault fd: %s", __func__, | |
492 | strerror(errno)); | |
493 | return -1; | |
494 | } | |
495 | ||
496 | /* | |
497 | * Although the host check already tested the API, we need to | |
498 | * do the check again as an ABI handshake on the new fd. | |
499 | */ | |
500 | if (!ufd_version_check(mis->userfault_fd)) { | |
501 | return -1; | |
502 | } | |
503 | ||
504 | /* Now an eventfd we use to tell the fault-thread to quit */ | |
505 | mis->userfault_quit_fd = eventfd(0, EFD_CLOEXEC); | |
506 | if (mis->userfault_quit_fd == -1) { | |
507 | error_report("%s: Opening userfault_quit_fd: %s", __func__, | |
508 | strerror(errno)); | |
509 | close(mis->userfault_fd); | |
510 | return -1; | |
511 | } | |
512 | ||
f0a227ad DDAG |
513 | qemu_sem_init(&mis->fault_thread_sem, 0); |
514 | qemu_thread_create(&mis->fault_thread, "postcopy/fault", | |
515 | postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE); | |
516 | qemu_sem_wait(&mis->fault_thread_sem); | |
517 | qemu_sem_destroy(&mis->fault_thread_sem); | |
c4faeed2 | 518 | mis->have_fault_thread = true; |
f0a227ad DDAG |
519 | |
520 | /* Mark so that we get notified of accesses to unwritten areas */ | |
521 | if (qemu_ram_foreach_block(ram_block_enable_notify, mis)) { | |
522 | return -1; | |
523 | } | |
524 | ||
371ff5a3 DDAG |
525 | /* |
526 | * Ballooning can mark pages as absent while we're postcopying | |
527 | * that would cause false userfaults. | |
528 | */ | |
529 | qemu_balloon_inhibit(true); | |
530 | ||
c4faeed2 DDAG |
531 | trace_postcopy_ram_enable_notify(); |
532 | ||
f0a227ad DDAG |
533 | return 0; |
534 | } | |
535 | ||
696ed9a9 DDAG |
536 | /* |
537 | * Place a host page (from) at (host) atomically | |
538 | * returns 0 on success | |
539 | */ | |
df9ff5e1 DDAG |
540 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
541 | size_t pagesize) | |
696ed9a9 DDAG |
542 | { |
543 | struct uffdio_copy copy_struct; | |
544 | ||
545 | copy_struct.dst = (uint64_t)(uintptr_t)host; | |
546 | copy_struct.src = (uint64_t)(uintptr_t)from; | |
df9ff5e1 | 547 | copy_struct.len = pagesize; |
696ed9a9 DDAG |
548 | copy_struct.mode = 0; |
549 | ||
550 | /* copy also acks to the kernel waking the stalled thread up | |
551 | * TODO: We can inhibit that ack and only do it if it was requested | |
552 | * which would be slightly cheaper, but we'd have to be careful | |
553 | * of the order of updating our page state. | |
554 | */ | |
555 | if (ioctl(mis->userfault_fd, UFFDIO_COPY, ©_struct)) { | |
556 | int e = errno; | |
df9ff5e1 DDAG |
557 | error_report("%s: %s copy host: %p from: %p (size: %zd)", |
558 | __func__, strerror(e), host, from, pagesize); | |
696ed9a9 DDAG |
559 | |
560 | return -e; | |
561 | } | |
562 | ||
563 | trace_postcopy_place_page(host); | |
564 | return 0; | |
565 | } | |
566 | ||
567 | /* | |
568 | * Place a zero page at (host) atomically | |
569 | * returns 0 on success | |
570 | */ | |
df9ff5e1 DDAG |
571 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
572 | size_t pagesize) | |
696ed9a9 | 573 | { |
df9ff5e1 | 574 | trace_postcopy_place_page_zero(host); |
696ed9a9 | 575 | |
df9ff5e1 DDAG |
576 | if (pagesize == getpagesize()) { |
577 | struct uffdio_zeropage zero_struct; | |
578 | zero_struct.range.start = (uint64_t)(uintptr_t)host; | |
579 | zero_struct.range.len = getpagesize(); | |
580 | zero_struct.mode = 0; | |
696ed9a9 | 581 | |
df9ff5e1 DDAG |
582 | if (ioctl(mis->userfault_fd, UFFDIO_ZEROPAGE, &zero_struct)) { |
583 | int e = errno; | |
584 | error_report("%s: %s zero host: %p", | |
585 | __func__, strerror(e), host); | |
696ed9a9 | 586 | |
df9ff5e1 DDAG |
587 | return -e; |
588 | } | |
589 | } else { | |
41d84210 DDAG |
590 | /* The kernel can't use UFFDIO_ZEROPAGE for hugepages */ |
591 | if (!mis->postcopy_tmp_zero_page) { | |
592 | mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size, | |
593 | PROT_READ | PROT_WRITE, | |
594 | MAP_PRIVATE | MAP_ANONYMOUS, | |
595 | -1, 0); | |
596 | if (mis->postcopy_tmp_zero_page == MAP_FAILED) { | |
597 | int e = errno; | |
598 | mis->postcopy_tmp_zero_page = NULL; | |
599 | error_report("%s: %s mapping large zero page", | |
600 | __func__, strerror(e)); | |
601 | return -e; | |
602 | } | |
603 | memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size); | |
604 | } | |
605 | return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, | |
606 | pagesize); | |
696ed9a9 DDAG |
607 | } |
608 | ||
696ed9a9 DDAG |
609 | return 0; |
610 | } | |
611 | ||
612 | /* | |
613 | * Returns a target page of memory that can be mapped at a later point in time | |
614 | * using postcopy_place_page | |
615 | * The same address is used repeatedly, postcopy_place_page just takes the | |
616 | * backing page away. | |
617 | * Returns: Pointer to allocated page | |
618 | * | |
619 | */ | |
620 | void *postcopy_get_tmp_page(MigrationIncomingState *mis) | |
621 | { | |
622 | if (!mis->postcopy_tmp_page) { | |
df9ff5e1 | 623 | mis->postcopy_tmp_page = mmap(NULL, mis->largest_page_size, |
696ed9a9 DDAG |
624 | PROT_READ | PROT_WRITE, MAP_PRIVATE | |
625 | MAP_ANONYMOUS, -1, 0); | |
0e8b3cdf EY |
626 | if (mis->postcopy_tmp_page == MAP_FAILED) { |
627 | mis->postcopy_tmp_page = NULL; | |
696ed9a9 DDAG |
628 | error_report("%s: %s", __func__, strerror(errno)); |
629 | return NULL; | |
630 | } | |
631 | } | |
632 | ||
633 | return mis->postcopy_tmp_page; | |
634 | } | |
635 | ||
eb59db53 DDAG |
636 | #else |
637 | /* No target OS support, stubs just fail */ | |
638 | bool postcopy_ram_supported_by_host(void) | |
639 | { | |
640 | error_report("%s: No OS support", __func__); | |
641 | return false; | |
642 | } | |
643 | ||
1caddf8a DDAG |
644 | int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages) |
645 | { | |
646 | error_report("postcopy_ram_incoming_init: No OS support"); | |
647 | return -1; | |
648 | } | |
649 | ||
650 | int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) | |
651 | { | |
652 | assert(0); | |
653 | return -1; | |
654 | } | |
655 | ||
f9527107 DDAG |
656 | int postcopy_ram_prepare_discard(MigrationIncomingState *mis) |
657 | { | |
658 | assert(0); | |
659 | return -1; | |
660 | } | |
661 | ||
f0a227ad DDAG |
662 | int postcopy_ram_enable_notify(MigrationIncomingState *mis) |
663 | { | |
664 | assert(0); | |
665 | return -1; | |
666 | } | |
696ed9a9 | 667 | |
df9ff5e1 DDAG |
668 | int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, |
669 | size_t pagesize) | |
696ed9a9 DDAG |
670 | { |
671 | assert(0); | |
672 | return -1; | |
673 | } | |
674 | ||
df9ff5e1 DDAG |
675 | int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, |
676 | size_t pagesize) | |
696ed9a9 DDAG |
677 | { |
678 | assert(0); | |
679 | return -1; | |
680 | } | |
681 | ||
682 | void *postcopy_get_tmp_page(MigrationIncomingState *mis) | |
683 | { | |
684 | assert(0); | |
685 | return NULL; | |
686 | } | |
687 | ||
eb59db53 DDAG |
688 | #endif |
689 | ||
e0b266f0 DDAG |
690 | /* ------------------------------------------------------------------------- */ |
691 | ||
692 | /** | |
693 | * postcopy_discard_send_init: Called at the start of each RAMBlock before | |
694 | * asking to discard individual ranges. | |
695 | * | |
696 | * @ms: The current migration state. | |
697 | * @offset: the bitmap offset of the named RAMBlock in the migration | |
698 | * bitmap. | |
699 | * @name: RAMBlock that discards will operate on. | |
700 | * | |
701 | * returns: a new PDS. | |
702 | */ | |
703 | PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms, | |
704 | unsigned long offset, | |
705 | const char *name) | |
706 | { | |
707 | PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState)); | |
708 | ||
709 | if (res) { | |
710 | res->ramblock_name = name; | |
711 | res->offset = offset; | |
712 | } | |
713 | ||
714 | return res; | |
715 | } | |
716 | ||
717 | /** | |
718 | * postcopy_discard_send_range: Called by the bitmap code for each chunk to | |
719 | * discard. May send a discard message, may just leave it queued to | |
720 | * be sent later. | |
721 | * | |
722 | * @ms: Current migration state. | |
723 | * @pds: Structure initialised by postcopy_discard_send_init(). | |
724 | * @start,@length: a range of pages in the migration bitmap in the | |
725 | * RAM block passed to postcopy_discard_send_init() (length=1 is one page) | |
726 | */ | |
727 | void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds, | |
728 | unsigned long start, unsigned long length) | |
729 | { | |
730 | size_t tp_bits = qemu_target_page_bits(); | |
731 | /* Convert to byte offsets within the RAM block */ | |
732 | pds->start_list[pds->cur_entry] = (start - pds->offset) << tp_bits; | |
733 | pds->length_list[pds->cur_entry] = length << tp_bits; | |
734 | trace_postcopy_discard_send_range(pds->ramblock_name, start, length); | |
735 | pds->cur_entry++; | |
736 | pds->nsentwords++; | |
737 | ||
738 | if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) { | |
739 | /* Full set, ship it! */ | |
89a02a9f HZ |
740 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
741 | pds->ramblock_name, | |
e0b266f0 DDAG |
742 | pds->cur_entry, |
743 | pds->start_list, | |
744 | pds->length_list); | |
745 | pds->nsentcmds++; | |
746 | pds->cur_entry = 0; | |
747 | } | |
748 | } | |
749 | ||
750 | /** | |
751 | * postcopy_discard_send_finish: Called at the end of each RAMBlock by the | |
752 | * bitmap code. Sends any outstanding discard messages, frees the PDS | |
753 | * | |
754 | * @ms: Current migration state. | |
755 | * @pds: Structure initialised by postcopy_discard_send_init(). | |
756 | */ | |
757 | void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds) | |
758 | { | |
759 | /* Anything unsent? */ | |
760 | if (pds->cur_entry) { | |
89a02a9f HZ |
761 | qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, |
762 | pds->ramblock_name, | |
e0b266f0 DDAG |
763 | pds->cur_entry, |
764 | pds->start_list, | |
765 | pds->length_list); | |
766 | pds->nsentcmds++; | |
767 | } | |
768 | ||
769 | trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords, | |
770 | pds->nsentcmds); | |
771 | ||
772 | g_free(pds); | |
773 | } |