4 * Copyright IBM, Corp. 2008
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Glauber Costa <gcosta@redhat.com>
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.
16 #include <sys/types.h>
17 #include <sys/ioctl.h>
21 #include <linux/kvm.h>
23 #include "qemu-common.h"
24 #include "qemu-barrier.h"
32 /* This check must be after config-host.h is included */
34 #include <sys/eventfd.h>
37 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
38 #define PAGE_SIZE TARGET_PAGE_SIZE
43 #define DPRINTF(fmt, ...) \
44 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
46 #define DPRINTF(fmt, ...) \
50 typedef struct KVMSlot
52 target_phys_addr_t start_addr
;
53 ram_addr_t memory_size
;
59 typedef struct kvm_dirty_log KVMDirtyLog
;
67 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
68 bool coalesced_flush_in_progress
;
69 int broken_set_mem_region
;
72 int robust_singlestep
;
74 #ifdef KVM_CAP_SET_GUEST_DEBUG
75 struct kvm_sw_breakpoint_head kvm_sw_breakpoints
;
77 int irqchip_in_kernel
;
81 int irqchip_inject_ioctl
;
82 #ifdef KVM_CAP_IRQ_ROUTING
83 struct kvm_irq_routing
*irq_routes
;
84 int nr_allocated_irq_routes
;
85 uint32_t *used_gsi_bitmap
;
92 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
93 KVM_CAP_INFO(USER_MEMORY
),
94 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
98 static KVMSlot
*kvm_alloc_slot(KVMState
*s
)
102 for (i
= 0; i
< ARRAY_SIZE(s
->slots
); i
++) {
103 if (s
->slots
[i
].memory_size
== 0) {
108 fprintf(stderr
, "%s: no free slot available\n", __func__
);
112 static KVMSlot
*kvm_lookup_matching_slot(KVMState
*s
,
113 target_phys_addr_t start_addr
,
114 target_phys_addr_t end_addr
)
118 for (i
= 0; i
< ARRAY_SIZE(s
->slots
); i
++) {
119 KVMSlot
*mem
= &s
->slots
[i
];
121 if (start_addr
== mem
->start_addr
&&
122 end_addr
== mem
->start_addr
+ mem
->memory_size
) {
131 * Find overlapping slot with lowest start address
133 static KVMSlot
*kvm_lookup_overlapping_slot(KVMState
*s
,
134 target_phys_addr_t start_addr
,
135 target_phys_addr_t end_addr
)
137 KVMSlot
*found
= NULL
;
140 for (i
= 0; i
< ARRAY_SIZE(s
->slots
); i
++) {
141 KVMSlot
*mem
= &s
->slots
[i
];
143 if (mem
->memory_size
== 0 ||
144 (found
&& found
->start_addr
< mem
->start_addr
)) {
148 if (end_addr
> mem
->start_addr
&&
149 start_addr
< mem
->start_addr
+ mem
->memory_size
) {
157 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
158 target_phys_addr_t
*phys_addr
)
162 for (i
= 0; i
< ARRAY_SIZE(s
->slots
); i
++) {
163 KVMSlot
*mem
= &s
->slots
[i
];
165 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
166 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
174 static int kvm_set_user_memory_region(KVMState
*s
, KVMSlot
*slot
)
176 struct kvm_userspace_memory_region mem
;
178 mem
.slot
= slot
->slot
;
179 mem
.guest_phys_addr
= slot
->start_addr
;
180 mem
.memory_size
= slot
->memory_size
;
181 mem
.userspace_addr
= (unsigned long)slot
->ram
;
182 mem
.flags
= slot
->flags
;
183 if (s
->migration_log
) {
184 mem
.flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
186 return kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
189 static void kvm_reset_vcpu(void *opaque
)
191 CPUState
*env
= opaque
;
193 kvm_arch_reset_vcpu(env
);
196 int kvm_irqchip_in_kernel(void)
198 return kvm_state
->irqchip_in_kernel
;
201 int kvm_pit_in_kernel(void)
203 return kvm_state
->pit_in_kernel
;
206 int kvm_init_vcpu(CPUState
*env
)
208 KVMState
*s
= kvm_state
;
212 DPRINTF("kvm_init_vcpu\n");
214 ret
= kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, env
->cpu_index
);
216 DPRINTF("kvm_create_vcpu failed\n");
222 env
->kvm_vcpu_dirty
= 1;
224 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
227 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
231 env
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
233 if (env
->kvm_run
== MAP_FAILED
) {
235 DPRINTF("mmap'ing vcpu state failed\n");
239 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
240 s
->coalesced_mmio_ring
=
241 (void *)env
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
244 ret
= kvm_arch_init_vcpu(env
);
246 qemu_register_reset(kvm_reset_vcpu
, env
);
247 kvm_arch_reset_vcpu(env
);
254 * dirty pages logging control
257 static int kvm_mem_flags(KVMState
*s
, bool log_dirty
)
259 return log_dirty
? KVM_MEM_LOG_DIRTY_PAGES
: 0;
262 static int kvm_slot_dirty_pages_log_change(KVMSlot
*mem
, bool log_dirty
)
264 KVMState
*s
= kvm_state
;
265 int flags
, mask
= KVM_MEM_LOG_DIRTY_PAGES
;
268 old_flags
= mem
->flags
;
270 flags
= (mem
->flags
& ~mask
) | kvm_mem_flags(s
, log_dirty
);
273 /* If nothing changed effectively, no need to issue ioctl */
274 if (s
->migration_log
) {
275 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
278 if (flags
== old_flags
) {
282 return kvm_set_user_memory_region(s
, mem
);
285 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr
,
286 ram_addr_t size
, bool log_dirty
)
288 KVMState
*s
= kvm_state
;
289 KVMSlot
*mem
= kvm_lookup_matching_slot(s
, phys_addr
, phys_addr
+ size
);
292 fprintf(stderr
, "BUG: %s: invalid parameters " TARGET_FMT_plx
"-"
293 TARGET_FMT_plx
"\n", __func__
, phys_addr
,
294 (target_phys_addr_t
)(phys_addr
+ size
- 1));
297 return kvm_slot_dirty_pages_log_change(mem
, log_dirty
);
300 static void kvm_log_start(MemoryListener
*listener
,
301 MemoryRegionSection
*section
)
305 r
= kvm_dirty_pages_log_change(section
->offset_within_address_space
,
306 section
->size
, true);
312 static void kvm_log_stop(MemoryListener
*listener
,
313 MemoryRegionSection
*section
)
317 r
= kvm_dirty_pages_log_change(section
->offset_within_address_space
,
318 section
->size
, false);
324 static int kvm_set_migration_log(int enable
)
326 KVMState
*s
= kvm_state
;
330 s
->migration_log
= enable
;
332 for (i
= 0; i
< ARRAY_SIZE(s
->slots
); i
++) {
335 if (!mem
->memory_size
) {
338 if (!!(mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) == enable
) {
341 err
= kvm_set_user_memory_region(s
, mem
);
349 /* get kvm's dirty pages bitmap and update qemu's */
350 static int kvm_get_dirty_pages_log_range(MemoryRegionSection
*section
,
351 unsigned long *bitmap
)
354 unsigned long page_number
, c
;
355 target_phys_addr_t addr
, addr1
;
356 unsigned int len
= ((section
->size
/ TARGET_PAGE_SIZE
) + HOST_LONG_BITS
- 1) / HOST_LONG_BITS
;
359 * bitmap-traveling is faster than memory-traveling (for addr...)
360 * especially when most of the memory is not dirty.
362 for (i
= 0; i
< len
; i
++) {
363 if (bitmap
[i
] != 0) {
364 c
= leul_to_cpu(bitmap
[i
]);
368 page_number
= i
* HOST_LONG_BITS
+ j
;
369 addr1
= page_number
* TARGET_PAGE_SIZE
;
370 addr
= section
->offset_within_region
+ addr1
;
371 memory_region_set_dirty(section
->mr
, addr
);
378 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
381 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
382 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
383 * This means all bits are set to dirty.
385 * @start_add: start of logged region.
386 * @end_addr: end of logged region.
388 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection
*section
)
390 KVMState
*s
= kvm_state
;
391 unsigned long size
, allocated_size
= 0;
395 target_phys_addr_t start_addr
= section
->offset_within_address_space
;
396 target_phys_addr_t end_addr
= start_addr
+ section
->size
;
398 d
.dirty_bitmap
= NULL
;
399 while (start_addr
< end_addr
) {
400 mem
= kvm_lookup_overlapping_slot(s
, start_addr
, end_addr
);
405 /* XXX bad kernel interface alert
406 * For dirty bitmap, kernel allocates array of size aligned to
407 * bits-per-long. But for case when the kernel is 64bits and
408 * the userspace is 32bits, userspace can't align to the same
409 * bits-per-long, since sizeof(long) is different between kernel
410 * and user space. This way, userspace will provide buffer which
411 * may be 4 bytes less than the kernel will use, resulting in
412 * userspace memory corruption (which is not detectable by valgrind
413 * too, in most cases).
414 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
415 * a hope that sizeof(long) wont become >8 any time soon.
417 size
= ALIGN(((mem
->memory_size
) >> TARGET_PAGE_BITS
),
418 /*HOST_LONG_BITS*/ 64) / 8;
419 if (!d
.dirty_bitmap
) {
420 d
.dirty_bitmap
= g_malloc(size
);
421 } else if (size
> allocated_size
) {
422 d
.dirty_bitmap
= g_realloc(d
.dirty_bitmap
, size
);
424 allocated_size
= size
;
425 memset(d
.dirty_bitmap
, 0, allocated_size
);
429 if (kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
) == -1) {
430 DPRINTF("ioctl failed %d\n", errno
);
435 kvm_get_dirty_pages_log_range(section
, d
.dirty_bitmap
);
436 start_addr
= mem
->start_addr
+ mem
->memory_size
;
438 g_free(d
.dirty_bitmap
);
443 int kvm_coalesce_mmio_region(target_phys_addr_t start
, ram_addr_t size
)
446 KVMState
*s
= kvm_state
;
448 if (s
->coalesced_mmio
) {
449 struct kvm_coalesced_mmio_zone zone
;
454 ret
= kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
460 int kvm_uncoalesce_mmio_region(target_phys_addr_t start
, ram_addr_t size
)
463 KVMState
*s
= kvm_state
;
465 if (s
->coalesced_mmio
) {
466 struct kvm_coalesced_mmio_zone zone
;
471 ret
= kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
477 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
481 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
489 static int kvm_check_many_ioeventfds(void)
491 /* Userspace can use ioeventfd for io notification. This requires a host
492 * that supports eventfd(2) and an I/O thread; since eventfd does not
493 * support SIGIO it cannot interrupt the vcpu.
495 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
496 * can avoid creating too many ioeventfds.
498 #if defined(CONFIG_EVENTFD)
501 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
502 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
503 if (ioeventfds
[i
] < 0) {
506 ret
= kvm_set_ioeventfd_pio_word(ioeventfds
[i
], 0, i
, true);
508 close(ioeventfds
[i
]);
513 /* Decide whether many devices are supported or not */
514 ret
= i
== ARRAY_SIZE(ioeventfds
);
517 kvm_set_ioeventfd_pio_word(ioeventfds
[i
], 0, i
, false);
518 close(ioeventfds
[i
]);
526 static const KVMCapabilityInfo
*
527 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
530 if (!kvm_check_extension(s
, list
->value
)) {
538 static void kvm_set_phys_mem(MemoryRegionSection
*section
, bool add
)
540 KVMState
*s
= kvm_state
;
543 MemoryRegion
*mr
= section
->mr
;
544 bool log_dirty
= memory_region_is_logging(mr
);
545 target_phys_addr_t start_addr
= section
->offset_within_address_space
;
546 ram_addr_t size
= section
->size
;
549 /* kvm works in page size chunks, but the function may be called
550 with sub-page size and unaligned start address. */
551 size
= TARGET_PAGE_ALIGN(size
);
552 start_addr
= TARGET_PAGE_ALIGN(start_addr
);
554 if (!memory_region_is_ram(mr
)) {
558 ram
= memory_region_get_ram_ptr(mr
) + section
->offset_within_region
;
561 mem
= kvm_lookup_overlapping_slot(s
, start_addr
, start_addr
+ size
);
566 if (add
&& start_addr
>= mem
->start_addr
&&
567 (start_addr
+ size
<= mem
->start_addr
+ mem
->memory_size
) &&
568 (ram
- start_addr
== mem
->ram
- mem
->start_addr
)) {
569 /* The new slot fits into the existing one and comes with
570 * identical parameters - update flags and done. */
571 kvm_slot_dirty_pages_log_change(mem
, log_dirty
);
577 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
578 kvm_physical_sync_dirty_bitmap(section
);
581 /* unregister the overlapping slot */
582 mem
->memory_size
= 0;
583 err
= kvm_set_user_memory_region(s
, mem
);
585 fprintf(stderr
, "%s: error unregistering overlapping slot: %s\n",
586 __func__
, strerror(-err
));
590 /* Workaround for older KVM versions: we can't join slots, even not by
591 * unregistering the previous ones and then registering the larger
592 * slot. We have to maintain the existing fragmentation. Sigh.
594 * This workaround assumes that the new slot starts at the same
595 * address as the first existing one. If not or if some overlapping
596 * slot comes around later, we will fail (not seen in practice so far)
597 * - and actually require a recent KVM version. */
598 if (s
->broken_set_mem_region
&&
599 old
.start_addr
== start_addr
&& old
.memory_size
< size
&& add
) {
600 mem
= kvm_alloc_slot(s
);
601 mem
->memory_size
= old
.memory_size
;
602 mem
->start_addr
= old
.start_addr
;
604 mem
->flags
= kvm_mem_flags(s
, log_dirty
);
606 err
= kvm_set_user_memory_region(s
, mem
);
608 fprintf(stderr
, "%s: error updating slot: %s\n", __func__
,
613 start_addr
+= old
.memory_size
;
614 ram
+= old
.memory_size
;
615 size
-= old
.memory_size
;
619 /* register prefix slot */
620 if (old
.start_addr
< start_addr
) {
621 mem
= kvm_alloc_slot(s
);
622 mem
->memory_size
= start_addr
- old
.start_addr
;
623 mem
->start_addr
= old
.start_addr
;
625 mem
->flags
= kvm_mem_flags(s
, log_dirty
);
627 err
= kvm_set_user_memory_region(s
, mem
);
629 fprintf(stderr
, "%s: error registering prefix slot: %s\n",
630 __func__
, strerror(-err
));
632 fprintf(stderr
, "%s: This is probably because your kernel's " \
633 "PAGE_SIZE is too big. Please try to use 4k " \
634 "PAGE_SIZE!\n", __func__
);
640 /* register suffix slot */
641 if (old
.start_addr
+ old
.memory_size
> start_addr
+ size
) {
642 ram_addr_t size_delta
;
644 mem
= kvm_alloc_slot(s
);
645 mem
->start_addr
= start_addr
+ size
;
646 size_delta
= mem
->start_addr
- old
.start_addr
;
647 mem
->memory_size
= old
.memory_size
- size_delta
;
648 mem
->ram
= old
.ram
+ size_delta
;
649 mem
->flags
= kvm_mem_flags(s
, log_dirty
);
651 err
= kvm_set_user_memory_region(s
, mem
);
653 fprintf(stderr
, "%s: error registering suffix slot: %s\n",
654 __func__
, strerror(-err
));
660 /* in case the KVM bug workaround already "consumed" the new slot */
667 mem
= kvm_alloc_slot(s
);
668 mem
->memory_size
= size
;
669 mem
->start_addr
= start_addr
;
671 mem
->flags
= kvm_mem_flags(s
, log_dirty
);
673 err
= kvm_set_user_memory_region(s
, mem
);
675 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
681 static void kvm_region_add(MemoryListener
*listener
,
682 MemoryRegionSection
*section
)
684 kvm_set_phys_mem(section
, true);
687 static void kvm_region_del(MemoryListener
*listener
,
688 MemoryRegionSection
*section
)
690 kvm_set_phys_mem(section
, false);
693 static void kvm_log_sync(MemoryListener
*listener
,
694 MemoryRegionSection
*section
)
698 r
= kvm_physical_sync_dirty_bitmap(section
);
704 static void kvm_log_global_start(struct MemoryListener
*listener
)
708 r
= kvm_set_migration_log(1);
712 static void kvm_log_global_stop(struct MemoryListener
*listener
)
716 r
= kvm_set_migration_log(0);
720 static MemoryListener kvm_memory_listener
= {
721 .region_add
= kvm_region_add
,
722 .region_del
= kvm_region_del
,
723 .log_start
= kvm_log_start
,
724 .log_stop
= kvm_log_stop
,
725 .log_sync
= kvm_log_sync
,
726 .log_global_start
= kvm_log_global_start
,
727 .log_global_stop
= kvm_log_global_stop
,
730 static void kvm_handle_interrupt(CPUState
*env
, int mask
)
732 env
->interrupt_request
|= mask
;
734 if (!qemu_cpu_is_self(env
)) {
739 int kvm_irqchip_set_irq(KVMState
*s
, int irq
, int level
)
741 struct kvm_irq_level event
;
744 assert(s
->irqchip_in_kernel
);
748 ret
= kvm_vm_ioctl(s
, s
->irqchip_inject_ioctl
, &event
);
750 perror("kvm_set_irqchip_line");
754 return (s
->irqchip_inject_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
757 #ifdef KVM_CAP_IRQ_ROUTING
758 static void set_gsi(KVMState
*s
, unsigned int gsi
)
760 assert(gsi
< s
->max_gsi
);
762 s
->used_gsi_bitmap
[gsi
/ 32] |= 1U << (gsi
% 32);
765 static void kvm_init_irq_routing(KVMState
*s
)
769 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
);
771 unsigned int gsi_bits
, i
;
773 /* Round up so we can search ints using ffs */
774 gsi_bits
= (gsi_count
+ 31) / 32;
775 s
->used_gsi_bitmap
= g_malloc0(gsi_bits
/ 8);
776 s
->max_gsi
= gsi_bits
;
778 /* Mark any over-allocated bits as already in use */
779 for (i
= gsi_count
; i
< gsi_bits
; i
++) {
784 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
785 s
->nr_allocated_irq_routes
= 0;
787 kvm_arch_init_irq_routing(s
);
790 static void kvm_add_routing_entry(KVMState
*s
,
791 struct kvm_irq_routing_entry
*entry
)
793 struct kvm_irq_routing_entry
*new;
796 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
797 n
= s
->nr_allocated_irq_routes
* 2;
801 size
= sizeof(struct kvm_irq_routing
);
802 size
+= n
* sizeof(*new);
803 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
804 s
->nr_allocated_irq_routes
= n
;
806 n
= s
->irq_routes
->nr
++;
807 new = &s
->irq_routes
->entries
[n
];
808 memset(new, 0, sizeof(*new));
809 new->gsi
= entry
->gsi
;
810 new->type
= entry
->type
;
811 new->flags
= entry
->flags
;
814 set_gsi(s
, entry
->gsi
);
817 void kvm_irqchip_add_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
819 struct kvm_irq_routing_entry e
;
822 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
824 e
.u
.irqchip
.irqchip
= irqchip
;
825 e
.u
.irqchip
.pin
= pin
;
826 kvm_add_routing_entry(s
, &e
);
829 int kvm_irqchip_commit_routes(KVMState
*s
)
831 s
->irq_routes
->flags
= 0;
832 return kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
835 #else /* !KVM_CAP_IRQ_ROUTING */
837 static void kvm_init_irq_routing(KVMState
*s
)
840 #endif /* !KVM_CAP_IRQ_ROUTING */
842 static int kvm_irqchip_create(KVMState
*s
)
844 QemuOptsList
*list
= qemu_find_opts("machine");
847 if (QTAILQ_EMPTY(&list
->head
) ||
848 !qemu_opt_get_bool(QTAILQ_FIRST(&list
->head
),
849 "kernel_irqchip", false) ||
850 !kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
854 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
856 fprintf(stderr
, "Create kernel irqchip failed\n");
860 s
->irqchip_inject_ioctl
= KVM_IRQ_LINE
;
861 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
862 s
->irqchip_inject_ioctl
= KVM_IRQ_LINE_STATUS
;
864 s
->irqchip_in_kernel
= 1;
866 kvm_init_irq_routing(s
);
873 static const char upgrade_note
[] =
874 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
875 "(see http://sourceforge.net/projects/kvm).\n";
877 const KVMCapabilityInfo
*missing_cap
;
881 s
= g_malloc0(sizeof(KVMState
));
883 #ifdef KVM_CAP_SET_GUEST_DEBUG
884 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
886 for (i
= 0; i
< ARRAY_SIZE(s
->slots
); i
++) {
887 s
->slots
[i
].slot
= i
;
890 s
->fd
= qemu_open("/dev/kvm", O_RDWR
);
892 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
897 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
898 if (ret
< KVM_API_VERSION
) {
902 fprintf(stderr
, "kvm version too old\n");
906 if (ret
> KVM_API_VERSION
) {
908 fprintf(stderr
, "kvm version not supported\n");
912 s
->vmfd
= kvm_ioctl(s
, KVM_CREATE_VM
, 0);
915 fprintf(stderr
, "Please add the 'switch_amode' kernel parameter to "
916 "your host kernel command line\n");
922 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
925 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
929 fprintf(stderr
, "kvm does not support %s\n%s",
930 missing_cap
->name
, upgrade_note
);
934 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
936 s
->broken_set_mem_region
= 1;
937 ret
= kvm_check_extension(s
, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
);
939 s
->broken_set_mem_region
= 0;
942 #ifdef KVM_CAP_VCPU_EVENTS
943 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
946 s
->robust_singlestep
=
947 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
949 #ifdef KVM_CAP_DEBUGREGS
950 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
954 s
->xsave
= kvm_check_extension(s
, KVM_CAP_XSAVE
);
958 s
->xcrs
= kvm_check_extension(s
, KVM_CAP_XCRS
);
961 ret
= kvm_arch_init(s
);
966 ret
= kvm_irqchip_create(s
);
972 memory_listener_register(&kvm_memory_listener
);
974 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
976 cpu_interrupt_handler
= kvm_handle_interrupt
;
994 static void kvm_handle_io(uint16_t port
, void *data
, int direction
, int size
,
1000 for (i
= 0; i
< count
; i
++) {
1001 if (direction
== KVM_EXIT_IO_IN
) {
1004 stb_p(ptr
, cpu_inb(port
));
1007 stw_p(ptr
, cpu_inw(port
));
1010 stl_p(ptr
, cpu_inl(port
));
1016 cpu_outb(port
, ldub_p(ptr
));
1019 cpu_outw(port
, lduw_p(ptr
));
1022 cpu_outl(port
, ldl_p(ptr
));
1031 static int kvm_handle_internal_error(CPUState
*env
, struct kvm_run
*run
)
1033 fprintf(stderr
, "KVM internal error.");
1034 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
1037 fprintf(stderr
, " Suberror: %d\n", run
->internal
.suberror
);
1038 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
1039 fprintf(stderr
, "extra data[%d]: %"PRIx64
"\n",
1040 i
, (uint64_t)run
->internal
.data
[i
]);
1043 fprintf(stderr
, "\n");
1045 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
1046 fprintf(stderr
, "emulation failure\n");
1047 if (!kvm_arch_stop_on_emulation_error(env
)) {
1048 cpu_dump_state(env
, stderr
, fprintf
, CPU_DUMP_CODE
);
1049 return EXCP_INTERRUPT
;
1052 /* FIXME: Should trigger a qmp message to let management know
1053 * something went wrong.
1058 void kvm_flush_coalesced_mmio_buffer(void)
1060 KVMState
*s
= kvm_state
;
1062 if (s
->coalesced_flush_in_progress
) {
1066 s
->coalesced_flush_in_progress
= true;
1068 if (s
->coalesced_mmio_ring
) {
1069 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
1070 while (ring
->first
!= ring
->last
) {
1071 struct kvm_coalesced_mmio
*ent
;
1073 ent
= &ring
->coalesced_mmio
[ring
->first
];
1075 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
1077 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
1081 s
->coalesced_flush_in_progress
= false;
1084 static void do_kvm_cpu_synchronize_state(void *_env
)
1086 CPUState
*env
= _env
;
1088 if (!env
->kvm_vcpu_dirty
) {
1089 kvm_arch_get_registers(env
);
1090 env
->kvm_vcpu_dirty
= 1;
1094 void kvm_cpu_synchronize_state(CPUState
*env
)
1096 if (!env
->kvm_vcpu_dirty
) {
1097 run_on_cpu(env
, do_kvm_cpu_synchronize_state
, env
);
1101 void kvm_cpu_synchronize_post_reset(CPUState
*env
)
1103 kvm_arch_put_registers(env
, KVM_PUT_RESET_STATE
);
1104 env
->kvm_vcpu_dirty
= 0;
1107 void kvm_cpu_synchronize_post_init(CPUState
*env
)
1109 kvm_arch_put_registers(env
, KVM_PUT_FULL_STATE
);
1110 env
->kvm_vcpu_dirty
= 0;
1113 int kvm_cpu_exec(CPUState
*env
)
1115 struct kvm_run
*run
= env
->kvm_run
;
1118 DPRINTF("kvm_cpu_exec()\n");
1120 if (kvm_arch_process_async_events(env
)) {
1121 env
->exit_request
= 0;
1125 cpu_single_env
= env
;
1128 if (env
->kvm_vcpu_dirty
) {
1129 kvm_arch_put_registers(env
, KVM_PUT_RUNTIME_STATE
);
1130 env
->kvm_vcpu_dirty
= 0;
1133 kvm_arch_pre_run(env
, run
);
1134 if (env
->exit_request
) {
1135 DPRINTF("interrupt exit requested\n");
1137 * KVM requires us to reenter the kernel after IO exits to complete
1138 * instruction emulation. This self-signal will ensure that we
1141 qemu_cpu_kick_self();
1143 cpu_single_env
= NULL
;
1144 qemu_mutex_unlock_iothread();
1146 run_ret
= kvm_vcpu_ioctl(env
, KVM_RUN
, 0);
1148 qemu_mutex_lock_iothread();
1149 cpu_single_env
= env
;
1150 kvm_arch_post_run(env
, run
);
1152 kvm_flush_coalesced_mmio_buffer();
1155 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
1156 DPRINTF("io window exit\n");
1157 ret
= EXCP_INTERRUPT
;
1160 fprintf(stderr
, "error: kvm run failed %s\n",
1161 strerror(-run_ret
));
1165 switch (run
->exit_reason
) {
1167 DPRINTF("handle_io\n");
1168 kvm_handle_io(run
->io
.port
,
1169 (uint8_t *)run
+ run
->io
.data_offset
,
1176 DPRINTF("handle_mmio\n");
1177 cpu_physical_memory_rw(run
->mmio
.phys_addr
,
1180 run
->mmio
.is_write
);
1183 case KVM_EXIT_IRQ_WINDOW_OPEN
:
1184 DPRINTF("irq_window_open\n");
1185 ret
= EXCP_INTERRUPT
;
1187 case KVM_EXIT_SHUTDOWN
:
1188 DPRINTF("shutdown\n");
1189 qemu_system_reset_request();
1190 ret
= EXCP_INTERRUPT
;
1192 case KVM_EXIT_UNKNOWN
:
1193 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
1194 (uint64_t)run
->hw
.hardware_exit_reason
);
1197 case KVM_EXIT_INTERNAL_ERROR
:
1198 ret
= kvm_handle_internal_error(env
, run
);
1201 DPRINTF("kvm_arch_handle_exit\n");
1202 ret
= kvm_arch_handle_exit(env
, run
);
1208 cpu_dump_state(env
, stderr
, fprintf
, CPU_DUMP_CODE
);
1209 vm_stop(RUN_STATE_INTERNAL_ERROR
);
1212 env
->exit_request
= 0;
1213 cpu_single_env
= NULL
;
1217 int kvm_ioctl(KVMState
*s
, int type
, ...)
1224 arg
= va_arg(ap
, void *);
1227 ret
= ioctl(s
->fd
, type
, arg
);
1234 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
1241 arg
= va_arg(ap
, void *);
1244 ret
= ioctl(s
->vmfd
, type
, arg
);
1251 int kvm_vcpu_ioctl(CPUState
*env
, int type
, ...)
1258 arg
= va_arg(ap
, void *);
1261 ret
= ioctl(env
->kvm_fd
, type
, arg
);
1268 int kvm_has_sync_mmu(void)
1270 return kvm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
1273 int kvm_has_vcpu_events(void)
1275 return kvm_state
->vcpu_events
;
1278 int kvm_has_robust_singlestep(void)
1280 return kvm_state
->robust_singlestep
;
1283 int kvm_has_debugregs(void)
1285 return kvm_state
->debugregs
;
1288 int kvm_has_xsave(void)
1290 return kvm_state
->xsave
;
1293 int kvm_has_xcrs(void)
1295 return kvm_state
->xcrs
;
1298 int kvm_has_many_ioeventfds(void)
1300 if (!kvm_enabled()) {
1303 return kvm_state
->many_ioeventfds
;
1306 int kvm_has_gsi_routing(void)
1308 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
1311 int kvm_allows_irq0_override(void)
1313 return !kvm_enabled() || !kvm_irqchip_in_kernel() || kvm_has_gsi_routing();
1316 void kvm_setup_guest_memory(void *start
, size_t size
)
1318 if (!kvm_has_sync_mmu()) {
1319 int ret
= qemu_madvise(start
, size
, QEMU_MADV_DONTFORK
);
1322 perror("qemu_madvise");
1324 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1330 #ifdef KVM_CAP_SET_GUEST_DEBUG
1331 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*env
,
1334 struct kvm_sw_breakpoint
*bp
;
1336 QTAILQ_FOREACH(bp
, &env
->kvm_state
->kvm_sw_breakpoints
, entry
) {
1344 int kvm_sw_breakpoints_active(CPUState
*env
)
1346 return !QTAILQ_EMPTY(&env
->kvm_state
->kvm_sw_breakpoints
);
1349 struct kvm_set_guest_debug_data
{
1350 struct kvm_guest_debug dbg
;
1355 static void kvm_invoke_set_guest_debug(void *data
)
1357 struct kvm_set_guest_debug_data
*dbg_data
= data
;
1358 CPUState
*env
= dbg_data
->env
;
1360 dbg_data
->err
= kvm_vcpu_ioctl(env
, KVM_SET_GUEST_DEBUG
, &dbg_data
->dbg
);
1363 int kvm_update_guest_debug(CPUState
*env
, unsigned long reinject_trap
)
1365 struct kvm_set_guest_debug_data data
;
1367 data
.dbg
.control
= reinject_trap
;
1369 if (env
->singlestep_enabled
) {
1370 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
1372 kvm_arch_update_guest_debug(env
, &data
.dbg
);
1375 run_on_cpu(env
, kvm_invoke_set_guest_debug
, &data
);
1379 int kvm_insert_breakpoint(CPUState
*current_env
, target_ulong addr
,
1380 target_ulong len
, int type
)
1382 struct kvm_sw_breakpoint
*bp
;
1386 if (type
== GDB_BREAKPOINT_SW
) {
1387 bp
= kvm_find_sw_breakpoint(current_env
, addr
);
1393 bp
= g_malloc(sizeof(struct kvm_sw_breakpoint
));
1400 err
= kvm_arch_insert_sw_breakpoint(current_env
, bp
);
1406 QTAILQ_INSERT_HEAD(¤t_env
->kvm_state
->kvm_sw_breakpoints
,
1409 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
1415 for (env
= first_cpu
; env
!= NULL
; env
= env
->next_cpu
) {
1416 err
= kvm_update_guest_debug(env
, 0);
1424 int kvm_remove_breakpoint(CPUState
*current_env
, target_ulong addr
,
1425 target_ulong len
, int type
)
1427 struct kvm_sw_breakpoint
*bp
;
1431 if (type
== GDB_BREAKPOINT_SW
) {
1432 bp
= kvm_find_sw_breakpoint(current_env
, addr
);
1437 if (bp
->use_count
> 1) {
1442 err
= kvm_arch_remove_sw_breakpoint(current_env
, bp
);
1447 QTAILQ_REMOVE(¤t_env
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
1450 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
1456 for (env
= first_cpu
; env
!= NULL
; env
= env
->next_cpu
) {
1457 err
= kvm_update_guest_debug(env
, 0);
1465 void kvm_remove_all_breakpoints(CPUState
*current_env
)
1467 struct kvm_sw_breakpoint
*bp
, *next
;
1468 KVMState
*s
= current_env
->kvm_state
;
1471 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
1472 if (kvm_arch_remove_sw_breakpoint(current_env
, bp
) != 0) {
1473 /* Try harder to find a CPU that currently sees the breakpoint. */
1474 for (env
= first_cpu
; env
!= NULL
; env
= env
->next_cpu
) {
1475 if (kvm_arch_remove_sw_breakpoint(env
, bp
) == 0) {
1481 kvm_arch_remove_all_hw_breakpoints();
1483 for (env
= first_cpu
; env
!= NULL
; env
= env
->next_cpu
) {
1484 kvm_update_guest_debug(env
, 0);
1488 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1490 int kvm_update_guest_debug(CPUState
*env
, unsigned long reinject_trap
)
1495 int kvm_insert_breakpoint(CPUState
*current_env
, target_ulong addr
,
1496 target_ulong len
, int type
)
1501 int kvm_remove_breakpoint(CPUState
*current_env
, target_ulong addr
,
1502 target_ulong len
, int type
)
1507 void kvm_remove_all_breakpoints(CPUState
*current_env
)
1510 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1512 int kvm_set_signal_mask(CPUState
*env
, const sigset_t
*sigset
)
1514 struct kvm_signal_mask
*sigmask
;
1518 return kvm_vcpu_ioctl(env
, KVM_SET_SIGNAL_MASK
, NULL
);
1521 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
1524 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
1525 r
= kvm_vcpu_ioctl(env
, KVM_SET_SIGNAL_MASK
, sigmask
);
1531 int kvm_set_ioeventfd_mmio_long(int fd
, uint32_t addr
, uint32_t val
, bool assign
)
1534 struct kvm_ioeventfd iofd
;
1536 iofd
.datamatch
= val
;
1539 iofd
.flags
= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1542 if (!kvm_enabled()) {
1547 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1550 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
1559 int kvm_set_ioeventfd_pio_word(int fd
, uint16_t addr
, uint16_t val
, bool assign
)
1561 struct kvm_ioeventfd kick
= {
1565 .flags
= KVM_IOEVENTFD_FLAG_DATAMATCH
| KVM_IOEVENTFD_FLAG_PIO
,
1569 if (!kvm_enabled()) {
1573 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1575 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
1582 int kvm_on_sigbus_vcpu(CPUState
*env
, int code
, void *addr
)
1584 return kvm_arch_on_sigbus_vcpu(env
, code
, addr
);
1587 int kvm_on_sigbus(int code
, void *addr
)
1589 return kvm_arch_on_sigbus(code
, addr
);