Merge remote-tracking branch 'remotes/kevin/tags/for-upstream' into staging
[qemu.git] / kvm-all.c
1 /*
2 * QEMU KVM support
3 *
4 * Copyright IBM, Corp. 2008
5 * Red Hat, Inc. 2008
6 *
7 * Authors:
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Glauber Costa <gcosta@redhat.com>
10 *
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
13 *
14 */
15
16 #include <sys/types.h>
17 #include <sys/ioctl.h>
18 #include <sys/mman.h>
19 #include <stdarg.h>
20
21 #include <linux/kvm.h>
22
23 #include "qemu-common.h"
24 #include "qemu/atomic.h"
25 #include "qemu/option.h"
26 #include "qemu/config-file.h"
27 #include "sysemu/sysemu.h"
28 #include "hw/hw.h"
29 #include "hw/pci/msi.h"
30 #include "exec/gdbstub.h"
31 #include "sysemu/kvm.h"
32 #include "qemu/bswap.h"
33 #include "exec/memory.h"
34 #include "exec/ram_addr.h"
35 #include "exec/address-spaces.h"
36 #include "qemu/event_notifier.h"
37 #include "trace.h"
38
39 #include "hw/boards.h"
40
41 /* This check must be after config-host.h is included */
42 #ifdef CONFIG_EVENTFD
43 #include <sys/eventfd.h>
44 #endif
45
46 #ifdef CONFIG_VALGRIND_H
47 #include <valgrind/memcheck.h>
48 #endif
49
50 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
51 #define PAGE_SIZE TARGET_PAGE_SIZE
52
53 //#define DEBUG_KVM
54
55 #ifdef DEBUG_KVM
56 #define DPRINTF(fmt, ...) \
57 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
58 #else
59 #define DPRINTF(fmt, ...) \
60 do { } while (0)
61 #endif
62
63 #define KVM_MSI_HASHTAB_SIZE 256
64
65 typedef struct KVMSlot
66 {
67 hwaddr start_addr;
68 ram_addr_t memory_size;
69 void *ram;
70 int slot;
71 int flags;
72 } KVMSlot;
73
74 typedef struct kvm_dirty_log KVMDirtyLog;
75
76 struct KVMState
77 {
78 KVMSlot *slots;
79 int nr_slots;
80 int fd;
81 int vmfd;
82 int coalesced_mmio;
83 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
84 bool coalesced_flush_in_progress;
85 int broken_set_mem_region;
86 int migration_log;
87 int vcpu_events;
88 int robust_singlestep;
89 int debugregs;
90 #ifdef KVM_CAP_SET_GUEST_DEBUG
91 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
92 #endif
93 int pit_state2;
94 int xsave, xcrs;
95 int many_ioeventfds;
96 int intx_set_mask;
97 /* The man page (and posix) say ioctl numbers are signed int, but
98 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
99 * unsigned, and treating them as signed here can break things */
100 unsigned irq_set_ioctl;
101 #ifdef KVM_CAP_IRQ_ROUTING
102 struct kvm_irq_routing *irq_routes;
103 int nr_allocated_irq_routes;
104 uint32_t *used_gsi_bitmap;
105 unsigned int gsi_count;
106 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
107 bool direct_msi;
108 #endif
109 };
110
111 KVMState *kvm_state;
112 bool kvm_kernel_irqchip;
113 bool kvm_async_interrupts_allowed;
114 bool kvm_halt_in_kernel_allowed;
115 bool kvm_irqfds_allowed;
116 bool kvm_msi_via_irqfd_allowed;
117 bool kvm_gsi_routing_allowed;
118 bool kvm_gsi_direct_mapping;
119 bool kvm_allowed;
120 bool kvm_readonly_mem_allowed;
121
122 static const KVMCapabilityInfo kvm_required_capabilites[] = {
123 KVM_CAP_INFO(USER_MEMORY),
124 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
125 KVM_CAP_LAST_INFO
126 };
127
128 static KVMSlot *kvm_alloc_slot(KVMState *s)
129 {
130 int i;
131
132 for (i = 0; i < s->nr_slots; i++) {
133 if (s->slots[i].memory_size == 0) {
134 return &s->slots[i];
135 }
136 }
137
138 fprintf(stderr, "%s: no free slot available\n", __func__);
139 abort();
140 }
141
142 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
143 hwaddr start_addr,
144 hwaddr end_addr)
145 {
146 int i;
147
148 for (i = 0; i < s->nr_slots; i++) {
149 KVMSlot *mem = &s->slots[i];
150
151 if (start_addr == mem->start_addr &&
152 end_addr == mem->start_addr + mem->memory_size) {
153 return mem;
154 }
155 }
156
157 return NULL;
158 }
159
160 /*
161 * Find overlapping slot with lowest start address
162 */
163 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
164 hwaddr start_addr,
165 hwaddr end_addr)
166 {
167 KVMSlot *found = NULL;
168 int i;
169
170 for (i = 0; i < s->nr_slots; i++) {
171 KVMSlot *mem = &s->slots[i];
172
173 if (mem->memory_size == 0 ||
174 (found && found->start_addr < mem->start_addr)) {
175 continue;
176 }
177
178 if (end_addr > mem->start_addr &&
179 start_addr < mem->start_addr + mem->memory_size) {
180 found = mem;
181 }
182 }
183
184 return found;
185 }
186
187 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
188 hwaddr *phys_addr)
189 {
190 int i;
191
192 for (i = 0; i < s->nr_slots; i++) {
193 KVMSlot *mem = &s->slots[i];
194
195 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
196 *phys_addr = mem->start_addr + (ram - mem->ram);
197 return 1;
198 }
199 }
200
201 return 0;
202 }
203
204 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
205 {
206 struct kvm_userspace_memory_region mem;
207
208 mem.slot = slot->slot;
209 mem.guest_phys_addr = slot->start_addr;
210 mem.userspace_addr = (unsigned long)slot->ram;
211 mem.flags = slot->flags;
212 if (s->migration_log) {
213 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
214 }
215
216 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
217 /* Set the slot size to 0 before setting the slot to the desired
218 * value. This is needed based on KVM commit 75d61fbc. */
219 mem.memory_size = 0;
220 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
221 }
222 mem.memory_size = slot->memory_size;
223 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
224 }
225
226 static void kvm_reset_vcpu(void *opaque)
227 {
228 CPUState *cpu = opaque;
229
230 kvm_arch_reset_vcpu(cpu);
231 }
232
233 int kvm_init_vcpu(CPUState *cpu)
234 {
235 KVMState *s = kvm_state;
236 long mmap_size;
237 int ret;
238
239 DPRINTF("kvm_init_vcpu\n");
240
241 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
242 if (ret < 0) {
243 DPRINTF("kvm_create_vcpu failed\n");
244 goto err;
245 }
246
247 cpu->kvm_fd = ret;
248 cpu->kvm_state = s;
249 cpu->kvm_vcpu_dirty = true;
250
251 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
252 if (mmap_size < 0) {
253 ret = mmap_size;
254 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
255 goto err;
256 }
257
258 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
259 cpu->kvm_fd, 0);
260 if (cpu->kvm_run == MAP_FAILED) {
261 ret = -errno;
262 DPRINTF("mmap'ing vcpu state failed\n");
263 goto err;
264 }
265
266 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
267 s->coalesced_mmio_ring =
268 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
269 }
270
271 ret = kvm_arch_init_vcpu(cpu);
272 if (ret == 0) {
273 qemu_register_reset(kvm_reset_vcpu, cpu);
274 kvm_arch_reset_vcpu(cpu);
275 }
276 err:
277 return ret;
278 }
279
280 /*
281 * dirty pages logging control
282 */
283
284 static int kvm_mem_flags(KVMState *s, bool log_dirty, bool readonly)
285 {
286 int flags = 0;
287 flags = log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
288 if (readonly && kvm_readonly_mem_allowed) {
289 flags |= KVM_MEM_READONLY;
290 }
291 return flags;
292 }
293
294 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
295 {
296 KVMState *s = kvm_state;
297 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
298 int old_flags;
299
300 old_flags = mem->flags;
301
302 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
303 mem->flags = flags;
304
305 /* If nothing changed effectively, no need to issue ioctl */
306 if (s->migration_log) {
307 flags |= KVM_MEM_LOG_DIRTY_PAGES;
308 }
309
310 if (flags == old_flags) {
311 return 0;
312 }
313
314 return kvm_set_user_memory_region(s, mem);
315 }
316
317 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
318 ram_addr_t size, bool log_dirty)
319 {
320 KVMState *s = kvm_state;
321 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
322
323 if (mem == NULL) {
324 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
325 TARGET_FMT_plx "\n", __func__, phys_addr,
326 (hwaddr)(phys_addr + size - 1));
327 return -EINVAL;
328 }
329 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
330 }
331
332 static void kvm_log_start(MemoryListener *listener,
333 MemoryRegionSection *section)
334 {
335 int r;
336
337 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
338 int128_get64(section->size), true);
339 if (r < 0) {
340 abort();
341 }
342 }
343
344 static void kvm_log_stop(MemoryListener *listener,
345 MemoryRegionSection *section)
346 {
347 int r;
348
349 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
350 int128_get64(section->size), false);
351 if (r < 0) {
352 abort();
353 }
354 }
355
356 static int kvm_set_migration_log(int enable)
357 {
358 KVMState *s = kvm_state;
359 KVMSlot *mem;
360 int i, err;
361
362 s->migration_log = enable;
363
364 for (i = 0; i < s->nr_slots; i++) {
365 mem = &s->slots[i];
366
367 if (!mem->memory_size) {
368 continue;
369 }
370 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
371 continue;
372 }
373 err = kvm_set_user_memory_region(s, mem);
374 if (err) {
375 return err;
376 }
377 }
378 return 0;
379 }
380
381 /* get kvm's dirty pages bitmap and update qemu's */
382 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
383 unsigned long *bitmap)
384 {
385 ram_addr_t start = section->offset_within_region + section->mr->ram_addr;
386 ram_addr_t pages = int128_get64(section->size) / getpagesize();
387
388 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
389 return 0;
390 }
391
392 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
393
394 /**
395 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
396 * This function updates qemu's dirty bitmap using
397 * memory_region_set_dirty(). This means all bits are set
398 * to dirty.
399 *
400 * @start_add: start of logged region.
401 * @end_addr: end of logged region.
402 */
403 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
404 {
405 KVMState *s = kvm_state;
406 unsigned long size, allocated_size = 0;
407 KVMDirtyLog d;
408 KVMSlot *mem;
409 int ret = 0;
410 hwaddr start_addr = section->offset_within_address_space;
411 hwaddr end_addr = start_addr + int128_get64(section->size);
412
413 d.dirty_bitmap = NULL;
414 while (start_addr < end_addr) {
415 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
416 if (mem == NULL) {
417 break;
418 }
419
420 /* XXX bad kernel interface alert
421 * For dirty bitmap, kernel allocates array of size aligned to
422 * bits-per-long. But for case when the kernel is 64bits and
423 * the userspace is 32bits, userspace can't align to the same
424 * bits-per-long, since sizeof(long) is different between kernel
425 * and user space. This way, userspace will provide buffer which
426 * may be 4 bytes less than the kernel will use, resulting in
427 * userspace memory corruption (which is not detectable by valgrind
428 * too, in most cases).
429 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
430 * a hope that sizeof(long) wont become >8 any time soon.
431 */
432 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
433 /*HOST_LONG_BITS*/ 64) / 8;
434 if (!d.dirty_bitmap) {
435 d.dirty_bitmap = g_malloc(size);
436 } else if (size > allocated_size) {
437 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
438 }
439 allocated_size = size;
440 memset(d.dirty_bitmap, 0, allocated_size);
441
442 d.slot = mem->slot;
443
444 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
445 DPRINTF("ioctl failed %d\n", errno);
446 ret = -1;
447 break;
448 }
449
450 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
451 start_addr = mem->start_addr + mem->memory_size;
452 }
453 g_free(d.dirty_bitmap);
454
455 return ret;
456 }
457
458 static void kvm_coalesce_mmio_region(MemoryListener *listener,
459 MemoryRegionSection *secion,
460 hwaddr start, hwaddr size)
461 {
462 KVMState *s = kvm_state;
463
464 if (s->coalesced_mmio) {
465 struct kvm_coalesced_mmio_zone zone;
466
467 zone.addr = start;
468 zone.size = size;
469 zone.pad = 0;
470
471 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
472 }
473 }
474
475 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
476 MemoryRegionSection *secion,
477 hwaddr start, hwaddr size)
478 {
479 KVMState *s = kvm_state;
480
481 if (s->coalesced_mmio) {
482 struct kvm_coalesced_mmio_zone zone;
483
484 zone.addr = start;
485 zone.size = size;
486 zone.pad = 0;
487
488 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
489 }
490 }
491
492 int kvm_check_extension(KVMState *s, unsigned int extension)
493 {
494 int ret;
495
496 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
497 if (ret < 0) {
498 ret = 0;
499 }
500
501 return ret;
502 }
503
504 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
505 bool assign, uint32_t size, bool datamatch)
506 {
507 int ret;
508 struct kvm_ioeventfd iofd;
509
510 iofd.datamatch = datamatch ? val : 0;
511 iofd.addr = addr;
512 iofd.len = size;
513 iofd.flags = 0;
514 iofd.fd = fd;
515
516 if (!kvm_enabled()) {
517 return -ENOSYS;
518 }
519
520 if (datamatch) {
521 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
522 }
523 if (!assign) {
524 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
525 }
526
527 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
528
529 if (ret < 0) {
530 return -errno;
531 }
532
533 return 0;
534 }
535
536 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
537 bool assign, uint32_t size, bool datamatch)
538 {
539 struct kvm_ioeventfd kick = {
540 .datamatch = datamatch ? val : 0,
541 .addr = addr,
542 .flags = KVM_IOEVENTFD_FLAG_PIO,
543 .len = size,
544 .fd = fd,
545 };
546 int r;
547 if (!kvm_enabled()) {
548 return -ENOSYS;
549 }
550 if (datamatch) {
551 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
552 }
553 if (!assign) {
554 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
555 }
556 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
557 if (r < 0) {
558 return r;
559 }
560 return 0;
561 }
562
563
564 static int kvm_check_many_ioeventfds(void)
565 {
566 /* Userspace can use ioeventfd for io notification. This requires a host
567 * that supports eventfd(2) and an I/O thread; since eventfd does not
568 * support SIGIO it cannot interrupt the vcpu.
569 *
570 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
571 * can avoid creating too many ioeventfds.
572 */
573 #if defined(CONFIG_EVENTFD)
574 int ioeventfds[7];
575 int i, ret = 0;
576 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
577 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
578 if (ioeventfds[i] < 0) {
579 break;
580 }
581 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
582 if (ret < 0) {
583 close(ioeventfds[i]);
584 break;
585 }
586 }
587
588 /* Decide whether many devices are supported or not */
589 ret = i == ARRAY_SIZE(ioeventfds);
590
591 while (i-- > 0) {
592 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
593 close(ioeventfds[i]);
594 }
595 return ret;
596 #else
597 return 0;
598 #endif
599 }
600
601 static const KVMCapabilityInfo *
602 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
603 {
604 while (list->name) {
605 if (!kvm_check_extension(s, list->value)) {
606 return list;
607 }
608 list++;
609 }
610 return NULL;
611 }
612
613 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
614 {
615 KVMState *s = kvm_state;
616 KVMSlot *mem, old;
617 int err;
618 MemoryRegion *mr = section->mr;
619 bool log_dirty = memory_region_is_logging(mr);
620 bool writeable = !mr->readonly && !mr->rom_device;
621 bool readonly_flag = mr->readonly || memory_region_is_romd(mr);
622 hwaddr start_addr = section->offset_within_address_space;
623 ram_addr_t size = int128_get64(section->size);
624 void *ram = NULL;
625 unsigned delta;
626
627 /* kvm works in page size chunks, but the function may be called
628 with sub-page size and unaligned start address. */
629 delta = TARGET_PAGE_ALIGN(size) - size;
630 if (delta > size) {
631 return;
632 }
633 start_addr += delta;
634 size -= delta;
635 size &= TARGET_PAGE_MASK;
636 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
637 return;
638 }
639
640 if (!memory_region_is_ram(mr)) {
641 if (writeable || !kvm_readonly_mem_allowed) {
642 return;
643 } else if (!mr->romd_mode) {
644 /* If the memory device is not in romd_mode, then we actually want
645 * to remove the kvm memory slot so all accesses will trap. */
646 add = false;
647 }
648 }
649
650 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
651
652 while (1) {
653 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
654 if (!mem) {
655 break;
656 }
657
658 if (add && start_addr >= mem->start_addr &&
659 (start_addr + size <= mem->start_addr + mem->memory_size) &&
660 (ram - start_addr == mem->ram - mem->start_addr)) {
661 /* The new slot fits into the existing one and comes with
662 * identical parameters - update flags and done. */
663 kvm_slot_dirty_pages_log_change(mem, log_dirty);
664 return;
665 }
666
667 old = *mem;
668
669 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
670 kvm_physical_sync_dirty_bitmap(section);
671 }
672
673 /* unregister the overlapping slot */
674 mem->memory_size = 0;
675 err = kvm_set_user_memory_region(s, mem);
676 if (err) {
677 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
678 __func__, strerror(-err));
679 abort();
680 }
681
682 /* Workaround for older KVM versions: we can't join slots, even not by
683 * unregistering the previous ones and then registering the larger
684 * slot. We have to maintain the existing fragmentation. Sigh.
685 *
686 * This workaround assumes that the new slot starts at the same
687 * address as the first existing one. If not or if some overlapping
688 * slot comes around later, we will fail (not seen in practice so far)
689 * - and actually require a recent KVM version. */
690 if (s->broken_set_mem_region &&
691 old.start_addr == start_addr && old.memory_size < size && add) {
692 mem = kvm_alloc_slot(s);
693 mem->memory_size = old.memory_size;
694 mem->start_addr = old.start_addr;
695 mem->ram = old.ram;
696 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
697
698 err = kvm_set_user_memory_region(s, mem);
699 if (err) {
700 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
701 strerror(-err));
702 abort();
703 }
704
705 start_addr += old.memory_size;
706 ram += old.memory_size;
707 size -= old.memory_size;
708 continue;
709 }
710
711 /* register prefix slot */
712 if (old.start_addr < start_addr) {
713 mem = kvm_alloc_slot(s);
714 mem->memory_size = start_addr - old.start_addr;
715 mem->start_addr = old.start_addr;
716 mem->ram = old.ram;
717 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
718
719 err = kvm_set_user_memory_region(s, mem);
720 if (err) {
721 fprintf(stderr, "%s: error registering prefix slot: %s\n",
722 __func__, strerror(-err));
723 #ifdef TARGET_PPC
724 fprintf(stderr, "%s: This is probably because your kernel's " \
725 "PAGE_SIZE is too big. Please try to use 4k " \
726 "PAGE_SIZE!\n", __func__);
727 #endif
728 abort();
729 }
730 }
731
732 /* register suffix slot */
733 if (old.start_addr + old.memory_size > start_addr + size) {
734 ram_addr_t size_delta;
735
736 mem = kvm_alloc_slot(s);
737 mem->start_addr = start_addr + size;
738 size_delta = mem->start_addr - old.start_addr;
739 mem->memory_size = old.memory_size - size_delta;
740 mem->ram = old.ram + size_delta;
741 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
742
743 err = kvm_set_user_memory_region(s, mem);
744 if (err) {
745 fprintf(stderr, "%s: error registering suffix slot: %s\n",
746 __func__, strerror(-err));
747 abort();
748 }
749 }
750 }
751
752 /* in case the KVM bug workaround already "consumed" the new slot */
753 if (!size) {
754 return;
755 }
756 if (!add) {
757 return;
758 }
759 mem = kvm_alloc_slot(s);
760 mem->memory_size = size;
761 mem->start_addr = start_addr;
762 mem->ram = ram;
763 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
764
765 err = kvm_set_user_memory_region(s, mem);
766 if (err) {
767 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
768 strerror(-err));
769 abort();
770 }
771 }
772
773 static void kvm_region_add(MemoryListener *listener,
774 MemoryRegionSection *section)
775 {
776 memory_region_ref(section->mr);
777 kvm_set_phys_mem(section, true);
778 }
779
780 static void kvm_region_del(MemoryListener *listener,
781 MemoryRegionSection *section)
782 {
783 kvm_set_phys_mem(section, false);
784 memory_region_unref(section->mr);
785 }
786
787 static void kvm_log_sync(MemoryListener *listener,
788 MemoryRegionSection *section)
789 {
790 int r;
791
792 r = kvm_physical_sync_dirty_bitmap(section);
793 if (r < 0) {
794 abort();
795 }
796 }
797
798 static void kvm_log_global_start(struct MemoryListener *listener)
799 {
800 int r;
801
802 r = kvm_set_migration_log(1);
803 assert(r >= 0);
804 }
805
806 static void kvm_log_global_stop(struct MemoryListener *listener)
807 {
808 int r;
809
810 r = kvm_set_migration_log(0);
811 assert(r >= 0);
812 }
813
814 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
815 MemoryRegionSection *section,
816 bool match_data, uint64_t data,
817 EventNotifier *e)
818 {
819 int fd = event_notifier_get_fd(e);
820 int r;
821
822 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
823 data, true, int128_get64(section->size),
824 match_data);
825 if (r < 0) {
826 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
827 __func__, strerror(-r));
828 abort();
829 }
830 }
831
832 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
833 MemoryRegionSection *section,
834 bool match_data, uint64_t data,
835 EventNotifier *e)
836 {
837 int fd = event_notifier_get_fd(e);
838 int r;
839
840 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
841 data, false, int128_get64(section->size),
842 match_data);
843 if (r < 0) {
844 abort();
845 }
846 }
847
848 static void kvm_io_ioeventfd_add(MemoryListener *listener,
849 MemoryRegionSection *section,
850 bool match_data, uint64_t data,
851 EventNotifier *e)
852 {
853 int fd = event_notifier_get_fd(e);
854 int r;
855
856 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
857 data, true, int128_get64(section->size),
858 match_data);
859 if (r < 0) {
860 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
861 __func__, strerror(-r));
862 abort();
863 }
864 }
865
866 static void kvm_io_ioeventfd_del(MemoryListener *listener,
867 MemoryRegionSection *section,
868 bool match_data, uint64_t data,
869 EventNotifier *e)
870
871 {
872 int fd = event_notifier_get_fd(e);
873 int r;
874
875 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
876 data, false, int128_get64(section->size),
877 match_data);
878 if (r < 0) {
879 abort();
880 }
881 }
882
883 static MemoryListener kvm_memory_listener = {
884 .region_add = kvm_region_add,
885 .region_del = kvm_region_del,
886 .log_start = kvm_log_start,
887 .log_stop = kvm_log_stop,
888 .log_sync = kvm_log_sync,
889 .log_global_start = kvm_log_global_start,
890 .log_global_stop = kvm_log_global_stop,
891 .eventfd_add = kvm_mem_ioeventfd_add,
892 .eventfd_del = kvm_mem_ioeventfd_del,
893 .coalesced_mmio_add = kvm_coalesce_mmio_region,
894 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
895 .priority = 10,
896 };
897
898 static MemoryListener kvm_io_listener = {
899 .eventfd_add = kvm_io_ioeventfd_add,
900 .eventfd_del = kvm_io_ioeventfd_del,
901 .priority = 10,
902 };
903
904 static void kvm_handle_interrupt(CPUState *cpu, int mask)
905 {
906 cpu->interrupt_request |= mask;
907
908 if (!qemu_cpu_is_self(cpu)) {
909 qemu_cpu_kick(cpu);
910 }
911 }
912
913 int kvm_set_irq(KVMState *s, int irq, int level)
914 {
915 struct kvm_irq_level event;
916 int ret;
917
918 assert(kvm_async_interrupts_enabled());
919
920 event.level = level;
921 event.irq = irq;
922 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
923 if (ret < 0) {
924 perror("kvm_set_irq");
925 abort();
926 }
927
928 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
929 }
930
931 #ifdef KVM_CAP_IRQ_ROUTING
932 typedef struct KVMMSIRoute {
933 struct kvm_irq_routing_entry kroute;
934 QTAILQ_ENTRY(KVMMSIRoute) entry;
935 } KVMMSIRoute;
936
937 static void set_gsi(KVMState *s, unsigned int gsi)
938 {
939 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
940 }
941
942 static void clear_gsi(KVMState *s, unsigned int gsi)
943 {
944 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
945 }
946
947 void kvm_init_irq_routing(KVMState *s)
948 {
949 int gsi_count, i;
950
951 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
952 if (gsi_count > 0) {
953 unsigned int gsi_bits, i;
954
955 /* Round up so we can search ints using ffs */
956 gsi_bits = ALIGN(gsi_count, 32);
957 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
958 s->gsi_count = gsi_count;
959
960 /* Mark any over-allocated bits as already in use */
961 for (i = gsi_count; i < gsi_bits; i++) {
962 set_gsi(s, i);
963 }
964 }
965
966 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
967 s->nr_allocated_irq_routes = 0;
968
969 if (!s->direct_msi) {
970 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
971 QTAILQ_INIT(&s->msi_hashtab[i]);
972 }
973 }
974
975 kvm_arch_init_irq_routing(s);
976 }
977
978 void kvm_irqchip_commit_routes(KVMState *s)
979 {
980 int ret;
981
982 s->irq_routes->flags = 0;
983 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
984 assert(ret == 0);
985 }
986
987 static void kvm_add_routing_entry(KVMState *s,
988 struct kvm_irq_routing_entry *entry)
989 {
990 struct kvm_irq_routing_entry *new;
991 int n, size;
992
993 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
994 n = s->nr_allocated_irq_routes * 2;
995 if (n < 64) {
996 n = 64;
997 }
998 size = sizeof(struct kvm_irq_routing);
999 size += n * sizeof(*new);
1000 s->irq_routes = g_realloc(s->irq_routes, size);
1001 s->nr_allocated_irq_routes = n;
1002 }
1003 n = s->irq_routes->nr++;
1004 new = &s->irq_routes->entries[n];
1005
1006 *new = *entry;
1007
1008 set_gsi(s, entry->gsi);
1009 }
1010
1011 static int kvm_update_routing_entry(KVMState *s,
1012 struct kvm_irq_routing_entry *new_entry)
1013 {
1014 struct kvm_irq_routing_entry *entry;
1015 int n;
1016
1017 for (n = 0; n < s->irq_routes->nr; n++) {
1018 entry = &s->irq_routes->entries[n];
1019 if (entry->gsi != new_entry->gsi) {
1020 continue;
1021 }
1022
1023 if(!memcmp(entry, new_entry, sizeof *entry)) {
1024 return 0;
1025 }
1026
1027 *entry = *new_entry;
1028
1029 kvm_irqchip_commit_routes(s);
1030
1031 return 0;
1032 }
1033
1034 return -ESRCH;
1035 }
1036
1037 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1038 {
1039 struct kvm_irq_routing_entry e = {};
1040
1041 assert(pin < s->gsi_count);
1042
1043 e.gsi = irq;
1044 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1045 e.flags = 0;
1046 e.u.irqchip.irqchip = irqchip;
1047 e.u.irqchip.pin = pin;
1048 kvm_add_routing_entry(s, &e);
1049 }
1050
1051 void kvm_irqchip_release_virq(KVMState *s, int virq)
1052 {
1053 struct kvm_irq_routing_entry *e;
1054 int i;
1055
1056 if (kvm_gsi_direct_mapping()) {
1057 return;
1058 }
1059
1060 for (i = 0; i < s->irq_routes->nr; i++) {
1061 e = &s->irq_routes->entries[i];
1062 if (e->gsi == virq) {
1063 s->irq_routes->nr--;
1064 *e = s->irq_routes->entries[s->irq_routes->nr];
1065 }
1066 }
1067 clear_gsi(s, virq);
1068 }
1069
1070 static unsigned int kvm_hash_msi(uint32_t data)
1071 {
1072 /* This is optimized for IA32 MSI layout. However, no other arch shall
1073 * repeat the mistake of not providing a direct MSI injection API. */
1074 return data & 0xff;
1075 }
1076
1077 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1078 {
1079 KVMMSIRoute *route, *next;
1080 unsigned int hash;
1081
1082 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1083 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1084 kvm_irqchip_release_virq(s, route->kroute.gsi);
1085 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1086 g_free(route);
1087 }
1088 }
1089 }
1090
1091 static int kvm_irqchip_get_virq(KVMState *s)
1092 {
1093 uint32_t *word = s->used_gsi_bitmap;
1094 int max_words = ALIGN(s->gsi_count, 32) / 32;
1095 int i, bit;
1096 bool retry = true;
1097
1098 again:
1099 /* Return the lowest unused GSI in the bitmap */
1100 for (i = 0; i < max_words; i++) {
1101 bit = ffs(~word[i]);
1102 if (!bit) {
1103 continue;
1104 }
1105
1106 return bit - 1 + i * 32;
1107 }
1108 if (!s->direct_msi && retry) {
1109 retry = false;
1110 kvm_flush_dynamic_msi_routes(s);
1111 goto again;
1112 }
1113 return -ENOSPC;
1114
1115 }
1116
1117 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1118 {
1119 unsigned int hash = kvm_hash_msi(msg.data);
1120 KVMMSIRoute *route;
1121
1122 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1123 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1124 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1125 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1126 return route;
1127 }
1128 }
1129 return NULL;
1130 }
1131
1132 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1133 {
1134 struct kvm_msi msi;
1135 KVMMSIRoute *route;
1136
1137 if (s->direct_msi) {
1138 msi.address_lo = (uint32_t)msg.address;
1139 msi.address_hi = msg.address >> 32;
1140 msi.data = le32_to_cpu(msg.data);
1141 msi.flags = 0;
1142 memset(msi.pad, 0, sizeof(msi.pad));
1143
1144 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1145 }
1146
1147 route = kvm_lookup_msi_route(s, msg);
1148 if (!route) {
1149 int virq;
1150
1151 virq = kvm_irqchip_get_virq(s);
1152 if (virq < 0) {
1153 return virq;
1154 }
1155
1156 route = g_malloc0(sizeof(KVMMSIRoute));
1157 route->kroute.gsi = virq;
1158 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1159 route->kroute.flags = 0;
1160 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1161 route->kroute.u.msi.address_hi = msg.address >> 32;
1162 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1163
1164 kvm_add_routing_entry(s, &route->kroute);
1165 kvm_irqchip_commit_routes(s);
1166
1167 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1168 entry);
1169 }
1170
1171 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1172
1173 return kvm_set_irq(s, route->kroute.gsi, 1);
1174 }
1175
1176 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1177 {
1178 struct kvm_irq_routing_entry kroute = {};
1179 int virq;
1180
1181 if (kvm_gsi_direct_mapping()) {
1182 return msg.data & 0xffff;
1183 }
1184
1185 if (!kvm_gsi_routing_enabled()) {
1186 return -ENOSYS;
1187 }
1188
1189 virq = kvm_irqchip_get_virq(s);
1190 if (virq < 0) {
1191 return virq;
1192 }
1193
1194 kroute.gsi = virq;
1195 kroute.type = KVM_IRQ_ROUTING_MSI;
1196 kroute.flags = 0;
1197 kroute.u.msi.address_lo = (uint32_t)msg.address;
1198 kroute.u.msi.address_hi = msg.address >> 32;
1199 kroute.u.msi.data = le32_to_cpu(msg.data);
1200
1201 kvm_add_routing_entry(s, &kroute);
1202 kvm_irqchip_commit_routes(s);
1203
1204 return virq;
1205 }
1206
1207 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1208 {
1209 struct kvm_irq_routing_entry kroute = {};
1210
1211 if (kvm_gsi_direct_mapping()) {
1212 return 0;
1213 }
1214
1215 if (!kvm_irqchip_in_kernel()) {
1216 return -ENOSYS;
1217 }
1218
1219 kroute.gsi = virq;
1220 kroute.type = KVM_IRQ_ROUTING_MSI;
1221 kroute.flags = 0;
1222 kroute.u.msi.address_lo = (uint32_t)msg.address;
1223 kroute.u.msi.address_hi = msg.address >> 32;
1224 kroute.u.msi.data = le32_to_cpu(msg.data);
1225
1226 return kvm_update_routing_entry(s, &kroute);
1227 }
1228
1229 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1230 bool assign)
1231 {
1232 struct kvm_irqfd irqfd = {
1233 .fd = fd,
1234 .gsi = virq,
1235 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1236 };
1237
1238 if (rfd != -1) {
1239 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1240 irqfd.resamplefd = rfd;
1241 }
1242
1243 if (!kvm_irqfds_enabled()) {
1244 return -ENOSYS;
1245 }
1246
1247 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1248 }
1249
1250 #else /* !KVM_CAP_IRQ_ROUTING */
1251
1252 void kvm_init_irq_routing(KVMState *s)
1253 {
1254 }
1255
1256 void kvm_irqchip_release_virq(KVMState *s, int virq)
1257 {
1258 }
1259
1260 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1261 {
1262 abort();
1263 }
1264
1265 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1266 {
1267 return -ENOSYS;
1268 }
1269
1270 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1271 {
1272 abort();
1273 }
1274
1275 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1276 {
1277 return -ENOSYS;
1278 }
1279 #endif /* !KVM_CAP_IRQ_ROUTING */
1280
1281 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1282 EventNotifier *rn, int virq)
1283 {
1284 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1285 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1286 }
1287
1288 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1289 {
1290 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1291 false);
1292 }
1293
1294 static int kvm_irqchip_create(KVMState *s)
1295 {
1296 int ret;
1297
1298 if (!qemu_opt_get_bool(qemu_get_machine_opts(), "kernel_irqchip", true) ||
1299 !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1300 return 0;
1301 }
1302
1303 /* First probe and see if there's a arch-specific hook to create the
1304 * in-kernel irqchip for us */
1305 ret = kvm_arch_irqchip_create(s);
1306 if (ret < 0) {
1307 return ret;
1308 } else if (ret == 0) {
1309 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1310 if (ret < 0) {
1311 fprintf(stderr, "Create kernel irqchip failed\n");
1312 return ret;
1313 }
1314 }
1315
1316 kvm_kernel_irqchip = true;
1317 /* If we have an in-kernel IRQ chip then we must have asynchronous
1318 * interrupt delivery (though the reverse is not necessarily true)
1319 */
1320 kvm_async_interrupts_allowed = true;
1321 kvm_halt_in_kernel_allowed = true;
1322
1323 kvm_init_irq_routing(s);
1324
1325 return 0;
1326 }
1327
1328 /* Find number of supported CPUs using the recommended
1329 * procedure from the kernel API documentation to cope with
1330 * older kernels that may be missing capabilities.
1331 */
1332 static int kvm_recommended_vcpus(KVMState *s)
1333 {
1334 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1335 return (ret) ? ret : 4;
1336 }
1337
1338 static int kvm_max_vcpus(KVMState *s)
1339 {
1340 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1341 return (ret) ? ret : kvm_recommended_vcpus(s);
1342 }
1343
1344 int kvm_init(QEMUMachine *machine)
1345 {
1346 static const char upgrade_note[] =
1347 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1348 "(see http://sourceforge.net/projects/kvm).\n";
1349 struct {
1350 const char *name;
1351 int num;
1352 } num_cpus[] = {
1353 { "SMP", smp_cpus },
1354 { "hotpluggable", max_cpus },
1355 { NULL, }
1356 }, *nc = num_cpus;
1357 int soft_vcpus_limit, hard_vcpus_limit;
1358 KVMState *s;
1359 const KVMCapabilityInfo *missing_cap;
1360 int ret;
1361 int i, type = 0;
1362 const char *kvm_type;
1363
1364 s = g_malloc0(sizeof(KVMState));
1365
1366 /*
1367 * On systems where the kernel can support different base page
1368 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1369 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1370 * page size for the system though.
1371 */
1372 assert(TARGET_PAGE_SIZE <= getpagesize());
1373 page_size_init();
1374
1375 #ifdef KVM_CAP_SET_GUEST_DEBUG
1376 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1377 #endif
1378 s->vmfd = -1;
1379 s->fd = qemu_open("/dev/kvm", O_RDWR);
1380 if (s->fd == -1) {
1381 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1382 ret = -errno;
1383 goto err;
1384 }
1385
1386 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1387 if (ret < KVM_API_VERSION) {
1388 if (ret > 0) {
1389 ret = -EINVAL;
1390 }
1391 fprintf(stderr, "kvm version too old\n");
1392 goto err;
1393 }
1394
1395 if (ret > KVM_API_VERSION) {
1396 ret = -EINVAL;
1397 fprintf(stderr, "kvm version not supported\n");
1398 goto err;
1399 }
1400
1401 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1402
1403 /* If unspecified, use the default value */
1404 if (!s->nr_slots) {
1405 s->nr_slots = 32;
1406 }
1407
1408 s->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1409
1410 for (i = 0; i < s->nr_slots; i++) {
1411 s->slots[i].slot = i;
1412 }
1413
1414 /* check the vcpu limits */
1415 soft_vcpus_limit = kvm_recommended_vcpus(s);
1416 hard_vcpus_limit = kvm_max_vcpus(s);
1417
1418 while (nc->name) {
1419 if (nc->num > soft_vcpus_limit) {
1420 fprintf(stderr,
1421 "Warning: Number of %s cpus requested (%d) exceeds "
1422 "the recommended cpus supported by KVM (%d)\n",
1423 nc->name, nc->num, soft_vcpus_limit);
1424
1425 if (nc->num > hard_vcpus_limit) {
1426 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1427 "the maximum cpus supported by KVM (%d)\n",
1428 nc->name, nc->num, hard_vcpus_limit);
1429 exit(1);
1430 }
1431 }
1432 nc++;
1433 }
1434
1435 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1436 if (machine->kvm_type) {
1437 type = machine->kvm_type(kvm_type);
1438 } else if (kvm_type) {
1439 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1440 goto err;
1441 }
1442
1443 do {
1444 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1445 } while (ret == -EINTR);
1446
1447 if (ret < 0) {
1448 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1449 strerror(-ret));
1450
1451 #ifdef TARGET_S390X
1452 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1453 "your host kernel command line\n");
1454 #endif
1455 goto err;
1456 }
1457
1458 s->vmfd = ret;
1459 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1460 if (!missing_cap) {
1461 missing_cap =
1462 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1463 }
1464 if (missing_cap) {
1465 ret = -EINVAL;
1466 fprintf(stderr, "kvm does not support %s\n%s",
1467 missing_cap->name, upgrade_note);
1468 goto err;
1469 }
1470
1471 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1472
1473 s->broken_set_mem_region = 1;
1474 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1475 if (ret > 0) {
1476 s->broken_set_mem_region = 0;
1477 }
1478
1479 #ifdef KVM_CAP_VCPU_EVENTS
1480 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1481 #endif
1482
1483 s->robust_singlestep =
1484 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1485
1486 #ifdef KVM_CAP_DEBUGREGS
1487 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1488 #endif
1489
1490 #ifdef KVM_CAP_XSAVE
1491 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1492 #endif
1493
1494 #ifdef KVM_CAP_XCRS
1495 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1496 #endif
1497
1498 #ifdef KVM_CAP_PIT_STATE2
1499 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1500 #endif
1501
1502 #ifdef KVM_CAP_IRQ_ROUTING
1503 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1504 #endif
1505
1506 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1507
1508 s->irq_set_ioctl = KVM_IRQ_LINE;
1509 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1510 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1511 }
1512
1513 #ifdef KVM_CAP_READONLY_MEM
1514 kvm_readonly_mem_allowed =
1515 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1516 #endif
1517
1518 ret = kvm_arch_init(s);
1519 if (ret < 0) {
1520 goto err;
1521 }
1522
1523 ret = kvm_irqchip_create(s);
1524 if (ret < 0) {
1525 goto err;
1526 }
1527
1528 kvm_state = s;
1529 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1530 memory_listener_register(&kvm_io_listener, &address_space_io);
1531
1532 s->many_ioeventfds = kvm_check_many_ioeventfds();
1533
1534 cpu_interrupt_handler = kvm_handle_interrupt;
1535
1536 return 0;
1537
1538 err:
1539 if (s->vmfd >= 0) {
1540 close(s->vmfd);
1541 }
1542 if (s->fd != -1) {
1543 close(s->fd);
1544 }
1545 g_free(s->slots);
1546 g_free(s);
1547
1548 return ret;
1549 }
1550
1551 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1552 uint32_t count)
1553 {
1554 int i;
1555 uint8_t *ptr = data;
1556
1557 for (i = 0; i < count; i++) {
1558 address_space_rw(&address_space_io, port, ptr, size,
1559 direction == KVM_EXIT_IO_OUT);
1560 ptr += size;
1561 }
1562 }
1563
1564 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1565 {
1566 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1567 run->internal.suberror);
1568
1569 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1570 int i;
1571
1572 for (i = 0; i < run->internal.ndata; ++i) {
1573 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1574 i, (uint64_t)run->internal.data[i]);
1575 }
1576 }
1577 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1578 fprintf(stderr, "emulation failure\n");
1579 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1580 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1581 return EXCP_INTERRUPT;
1582 }
1583 }
1584 /* FIXME: Should trigger a qmp message to let management know
1585 * something went wrong.
1586 */
1587 return -1;
1588 }
1589
1590 void kvm_flush_coalesced_mmio_buffer(void)
1591 {
1592 KVMState *s = kvm_state;
1593
1594 if (s->coalesced_flush_in_progress) {
1595 return;
1596 }
1597
1598 s->coalesced_flush_in_progress = true;
1599
1600 if (s->coalesced_mmio_ring) {
1601 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1602 while (ring->first != ring->last) {
1603 struct kvm_coalesced_mmio *ent;
1604
1605 ent = &ring->coalesced_mmio[ring->first];
1606
1607 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1608 smp_wmb();
1609 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1610 }
1611 }
1612
1613 s->coalesced_flush_in_progress = false;
1614 }
1615
1616 static void do_kvm_cpu_synchronize_state(void *arg)
1617 {
1618 CPUState *cpu = arg;
1619
1620 if (!cpu->kvm_vcpu_dirty) {
1621 kvm_arch_get_registers(cpu);
1622 cpu->kvm_vcpu_dirty = true;
1623 }
1624 }
1625
1626 void kvm_cpu_synchronize_state(CPUState *cpu)
1627 {
1628 if (!cpu->kvm_vcpu_dirty) {
1629 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1630 }
1631 }
1632
1633 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1634 {
1635 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1636 cpu->kvm_vcpu_dirty = false;
1637 }
1638
1639 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1640 {
1641 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1642 cpu->kvm_vcpu_dirty = false;
1643 }
1644
1645 int kvm_cpu_exec(CPUState *cpu)
1646 {
1647 struct kvm_run *run = cpu->kvm_run;
1648 int ret, run_ret;
1649
1650 DPRINTF("kvm_cpu_exec()\n");
1651
1652 if (kvm_arch_process_async_events(cpu)) {
1653 cpu->exit_request = 0;
1654 return EXCP_HLT;
1655 }
1656
1657 do {
1658 if (cpu->kvm_vcpu_dirty) {
1659 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1660 cpu->kvm_vcpu_dirty = false;
1661 }
1662
1663 kvm_arch_pre_run(cpu, run);
1664 if (cpu->exit_request) {
1665 DPRINTF("interrupt exit requested\n");
1666 /*
1667 * KVM requires us to reenter the kernel after IO exits to complete
1668 * instruction emulation. This self-signal will ensure that we
1669 * leave ASAP again.
1670 */
1671 qemu_cpu_kick_self();
1672 }
1673 qemu_mutex_unlock_iothread();
1674
1675 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1676
1677 qemu_mutex_lock_iothread();
1678 kvm_arch_post_run(cpu, run);
1679
1680 if (run_ret < 0) {
1681 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1682 DPRINTF("io window exit\n");
1683 ret = EXCP_INTERRUPT;
1684 break;
1685 }
1686 fprintf(stderr, "error: kvm run failed %s\n",
1687 strerror(-run_ret));
1688 abort();
1689 }
1690
1691 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1692 switch (run->exit_reason) {
1693 case KVM_EXIT_IO:
1694 DPRINTF("handle_io\n");
1695 kvm_handle_io(run->io.port,
1696 (uint8_t *)run + run->io.data_offset,
1697 run->io.direction,
1698 run->io.size,
1699 run->io.count);
1700 ret = 0;
1701 break;
1702 case KVM_EXIT_MMIO:
1703 DPRINTF("handle_mmio\n");
1704 cpu_physical_memory_rw(run->mmio.phys_addr,
1705 run->mmio.data,
1706 run->mmio.len,
1707 run->mmio.is_write);
1708 ret = 0;
1709 break;
1710 case KVM_EXIT_IRQ_WINDOW_OPEN:
1711 DPRINTF("irq_window_open\n");
1712 ret = EXCP_INTERRUPT;
1713 break;
1714 case KVM_EXIT_SHUTDOWN:
1715 DPRINTF("shutdown\n");
1716 qemu_system_reset_request();
1717 ret = EXCP_INTERRUPT;
1718 break;
1719 case KVM_EXIT_UNKNOWN:
1720 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1721 (uint64_t)run->hw.hardware_exit_reason);
1722 ret = -1;
1723 break;
1724 case KVM_EXIT_INTERNAL_ERROR:
1725 ret = kvm_handle_internal_error(cpu, run);
1726 break;
1727 default:
1728 DPRINTF("kvm_arch_handle_exit\n");
1729 ret = kvm_arch_handle_exit(cpu, run);
1730 break;
1731 }
1732 } while (ret == 0);
1733
1734 if (ret < 0) {
1735 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1736 vm_stop(RUN_STATE_INTERNAL_ERROR);
1737 }
1738
1739 cpu->exit_request = 0;
1740 return ret;
1741 }
1742
1743 int kvm_ioctl(KVMState *s, int type, ...)
1744 {
1745 int ret;
1746 void *arg;
1747 va_list ap;
1748
1749 va_start(ap, type);
1750 arg = va_arg(ap, void *);
1751 va_end(ap);
1752
1753 trace_kvm_ioctl(type, arg);
1754 ret = ioctl(s->fd, type, arg);
1755 if (ret == -1) {
1756 ret = -errno;
1757 }
1758 return ret;
1759 }
1760
1761 int kvm_vm_ioctl(KVMState *s, int type, ...)
1762 {
1763 int ret;
1764 void *arg;
1765 va_list ap;
1766
1767 va_start(ap, type);
1768 arg = va_arg(ap, void *);
1769 va_end(ap);
1770
1771 trace_kvm_vm_ioctl(type, arg);
1772 ret = ioctl(s->vmfd, type, arg);
1773 if (ret == -1) {
1774 ret = -errno;
1775 }
1776 return ret;
1777 }
1778
1779 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1780 {
1781 int ret;
1782 void *arg;
1783 va_list ap;
1784
1785 va_start(ap, type);
1786 arg = va_arg(ap, void *);
1787 va_end(ap);
1788
1789 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1790 ret = ioctl(cpu->kvm_fd, type, arg);
1791 if (ret == -1) {
1792 ret = -errno;
1793 }
1794 return ret;
1795 }
1796
1797 int kvm_device_ioctl(int fd, int type, ...)
1798 {
1799 int ret;
1800 void *arg;
1801 va_list ap;
1802
1803 va_start(ap, type);
1804 arg = va_arg(ap, void *);
1805 va_end(ap);
1806
1807 trace_kvm_device_ioctl(fd, type, arg);
1808 ret = ioctl(fd, type, arg);
1809 if (ret == -1) {
1810 ret = -errno;
1811 }
1812 return ret;
1813 }
1814
1815 int kvm_has_sync_mmu(void)
1816 {
1817 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1818 }
1819
1820 int kvm_has_vcpu_events(void)
1821 {
1822 return kvm_state->vcpu_events;
1823 }
1824
1825 int kvm_has_robust_singlestep(void)
1826 {
1827 return kvm_state->robust_singlestep;
1828 }
1829
1830 int kvm_has_debugregs(void)
1831 {
1832 return kvm_state->debugregs;
1833 }
1834
1835 int kvm_has_xsave(void)
1836 {
1837 return kvm_state->xsave;
1838 }
1839
1840 int kvm_has_xcrs(void)
1841 {
1842 return kvm_state->xcrs;
1843 }
1844
1845 int kvm_has_pit_state2(void)
1846 {
1847 return kvm_state->pit_state2;
1848 }
1849
1850 int kvm_has_many_ioeventfds(void)
1851 {
1852 if (!kvm_enabled()) {
1853 return 0;
1854 }
1855 return kvm_state->many_ioeventfds;
1856 }
1857
1858 int kvm_has_gsi_routing(void)
1859 {
1860 #ifdef KVM_CAP_IRQ_ROUTING
1861 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1862 #else
1863 return false;
1864 #endif
1865 }
1866
1867 int kvm_has_intx_set_mask(void)
1868 {
1869 return kvm_state->intx_set_mask;
1870 }
1871
1872 void kvm_setup_guest_memory(void *start, size_t size)
1873 {
1874 #ifdef CONFIG_VALGRIND_H
1875 VALGRIND_MAKE_MEM_DEFINED(start, size);
1876 #endif
1877 if (!kvm_has_sync_mmu()) {
1878 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1879
1880 if (ret) {
1881 perror("qemu_madvise");
1882 fprintf(stderr,
1883 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1884 exit(1);
1885 }
1886 }
1887 }
1888
1889 #ifdef KVM_CAP_SET_GUEST_DEBUG
1890 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
1891 target_ulong pc)
1892 {
1893 struct kvm_sw_breakpoint *bp;
1894
1895 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
1896 if (bp->pc == pc) {
1897 return bp;
1898 }
1899 }
1900 return NULL;
1901 }
1902
1903 int kvm_sw_breakpoints_active(CPUState *cpu)
1904 {
1905 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
1906 }
1907
1908 struct kvm_set_guest_debug_data {
1909 struct kvm_guest_debug dbg;
1910 CPUState *cpu;
1911 int err;
1912 };
1913
1914 static void kvm_invoke_set_guest_debug(void *data)
1915 {
1916 struct kvm_set_guest_debug_data *dbg_data = data;
1917
1918 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
1919 &dbg_data->dbg);
1920 }
1921
1922 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
1923 {
1924 struct kvm_set_guest_debug_data data;
1925
1926 data.dbg.control = reinject_trap;
1927
1928 if (cpu->singlestep_enabled) {
1929 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1930 }
1931 kvm_arch_update_guest_debug(cpu, &data.dbg);
1932 data.cpu = cpu;
1933
1934 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
1935 return data.err;
1936 }
1937
1938 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
1939 target_ulong len, int type)
1940 {
1941 struct kvm_sw_breakpoint *bp;
1942 int err;
1943
1944 if (type == GDB_BREAKPOINT_SW) {
1945 bp = kvm_find_sw_breakpoint(cpu, addr);
1946 if (bp) {
1947 bp->use_count++;
1948 return 0;
1949 }
1950
1951 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1952 if (!bp) {
1953 return -ENOMEM;
1954 }
1955
1956 bp->pc = addr;
1957 bp->use_count = 1;
1958 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
1959 if (err) {
1960 g_free(bp);
1961 return err;
1962 }
1963
1964 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
1965 } else {
1966 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1967 if (err) {
1968 return err;
1969 }
1970 }
1971
1972 CPU_FOREACH(cpu) {
1973 err = kvm_update_guest_debug(cpu, 0);
1974 if (err) {
1975 return err;
1976 }
1977 }
1978 return 0;
1979 }
1980
1981 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
1982 target_ulong len, int type)
1983 {
1984 struct kvm_sw_breakpoint *bp;
1985 int err;
1986
1987 if (type == GDB_BREAKPOINT_SW) {
1988 bp = kvm_find_sw_breakpoint(cpu, addr);
1989 if (!bp) {
1990 return -ENOENT;
1991 }
1992
1993 if (bp->use_count > 1) {
1994 bp->use_count--;
1995 return 0;
1996 }
1997
1998 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
1999 if (err) {
2000 return err;
2001 }
2002
2003 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2004 g_free(bp);
2005 } else {
2006 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2007 if (err) {
2008 return err;
2009 }
2010 }
2011
2012 CPU_FOREACH(cpu) {
2013 err = kvm_update_guest_debug(cpu, 0);
2014 if (err) {
2015 return err;
2016 }
2017 }
2018 return 0;
2019 }
2020
2021 void kvm_remove_all_breakpoints(CPUState *cpu)
2022 {
2023 struct kvm_sw_breakpoint *bp, *next;
2024 KVMState *s = cpu->kvm_state;
2025
2026 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2027 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2028 /* Try harder to find a CPU that currently sees the breakpoint. */
2029 CPU_FOREACH(cpu) {
2030 if (kvm_arch_remove_sw_breakpoint(cpu, bp) == 0) {
2031 break;
2032 }
2033 }
2034 }
2035 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2036 g_free(bp);
2037 }
2038 kvm_arch_remove_all_hw_breakpoints();
2039
2040 CPU_FOREACH(cpu) {
2041 kvm_update_guest_debug(cpu, 0);
2042 }
2043 }
2044
2045 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2046
2047 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2048 {
2049 return -EINVAL;
2050 }
2051
2052 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2053 target_ulong len, int type)
2054 {
2055 return -EINVAL;
2056 }
2057
2058 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2059 target_ulong len, int type)
2060 {
2061 return -EINVAL;
2062 }
2063
2064 void kvm_remove_all_breakpoints(CPUState *cpu)
2065 {
2066 }
2067 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2068
2069 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2070 {
2071 struct kvm_signal_mask *sigmask;
2072 int r;
2073
2074 if (!sigset) {
2075 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2076 }
2077
2078 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2079
2080 sigmask->len = 8;
2081 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2082 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2083 g_free(sigmask);
2084
2085 return r;
2086 }
2087 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2088 {
2089 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2090 }
2091
2092 int kvm_on_sigbus(int code, void *addr)
2093 {
2094 return kvm_arch_on_sigbus(code, addr);
2095 }
2096
2097 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2098 {
2099 int ret;
2100 struct kvm_create_device create_dev;
2101
2102 create_dev.type = type;
2103 create_dev.fd = -1;
2104 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2105
2106 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2107 return -ENOTSUP;
2108 }
2109
2110 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2111 if (ret) {
2112 return ret;
2113 }
2114
2115 return test ? 0 : create_dev.fd;
2116 }