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/atomic.h"
25 #include "qemu/option.h"
26 #include "qemu/config-file.h"
27 #include "sysemu/sysemu.h"
28 #include "sysemu/accel.h"
30 #include "hw/pci/msi.h"
31 #include "hw/s390x/adapter.h"
32 #include "exec/gdbstub.h"
33 #include "sysemu/kvm.h"
34 #include "qemu/bswap.h"
35 #include "exec/memory.h"
36 #include "exec/ram_addr.h"
37 #include "exec/address-spaces.h"
38 #include "qemu/event_notifier.h"
41 #include "hw/boards.h"
43 /* This check must be after config-host.h is included */
45 #include <sys/eventfd.h>
48 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
49 #define PAGE_SIZE TARGET_PAGE_SIZE
54 #define DPRINTF(fmt, ...) \
55 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
57 #define DPRINTF(fmt, ...) \
61 #define KVM_MSI_HASHTAB_SIZE 256
63 typedef struct KVMSlot
66 ram_addr_t memory_size
;
74 AccelState parent_obj
;
81 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
82 bool coalesced_flush_in_progress
;
83 int broken_set_mem_region
;
85 int robust_singlestep
;
87 #ifdef KVM_CAP_SET_GUEST_DEBUG
88 struct kvm_sw_breakpoint_head kvm_sw_breakpoints
;
94 /* The man page (and posix) say ioctl numbers are signed int, but
95 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
96 * unsigned, and treating them as signed here can break things */
97 unsigned irq_set_ioctl
;
98 unsigned int sigmask_len
;
99 #ifdef KVM_CAP_IRQ_ROUTING
100 struct kvm_irq_routing
*irq_routes
;
101 int nr_allocated_irq_routes
;
102 uint32_t *used_gsi_bitmap
;
103 unsigned int gsi_count
;
104 QTAILQ_HEAD(msi_hashtab
, KVMMSIRoute
) msi_hashtab
[KVM_MSI_HASHTAB_SIZE
];
109 #define TYPE_KVM_ACCEL ACCEL_CLASS_NAME("kvm")
111 #define KVM_STATE(obj) \
112 OBJECT_CHECK(KVMState, (obj), TYPE_KVM_ACCEL)
115 bool kvm_kernel_irqchip
;
116 bool kvm_async_interrupts_allowed
;
117 bool kvm_halt_in_kernel_allowed
;
118 bool kvm_eventfds_allowed
;
119 bool kvm_irqfds_allowed
;
120 bool kvm_resamplefds_allowed
;
121 bool kvm_msi_via_irqfd_allowed
;
122 bool kvm_gsi_routing_allowed
;
123 bool kvm_gsi_direct_mapping
;
125 bool kvm_readonly_mem_allowed
;
126 bool kvm_vm_attributes_allowed
;
128 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
129 KVM_CAP_INFO(USER_MEMORY
),
130 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
134 static KVMSlot
*kvm_get_free_slot(KVMState
*s
)
138 for (i
= 0; i
< s
->nr_slots
; i
++) {
139 if (s
->slots
[i
].memory_size
== 0) {
147 bool kvm_has_free_slot(MachineState
*ms
)
149 return kvm_get_free_slot(KVM_STATE(ms
->accelerator
));
152 static KVMSlot
*kvm_alloc_slot(KVMState
*s
)
154 KVMSlot
*slot
= kvm_get_free_slot(s
);
160 fprintf(stderr
, "%s: no free slot available\n", __func__
);
164 static KVMSlot
*kvm_lookup_matching_slot(KVMState
*s
,
170 for (i
= 0; i
< s
->nr_slots
; i
++) {
171 KVMSlot
*mem
= &s
->slots
[i
];
173 if (start_addr
== mem
->start_addr
&&
174 end_addr
== mem
->start_addr
+ mem
->memory_size
) {
183 * Find overlapping slot with lowest start address
185 static KVMSlot
*kvm_lookup_overlapping_slot(KVMState
*s
,
189 KVMSlot
*found
= NULL
;
192 for (i
= 0; i
< s
->nr_slots
; i
++) {
193 KVMSlot
*mem
= &s
->slots
[i
];
195 if (mem
->memory_size
== 0 ||
196 (found
&& found
->start_addr
< mem
->start_addr
)) {
200 if (end_addr
> mem
->start_addr
&&
201 start_addr
< mem
->start_addr
+ mem
->memory_size
) {
209 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
214 for (i
= 0; i
< s
->nr_slots
; i
++) {
215 KVMSlot
*mem
= &s
->slots
[i
];
217 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
218 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
226 static int kvm_set_user_memory_region(KVMState
*s
, KVMSlot
*slot
)
228 struct kvm_userspace_memory_region mem
;
230 mem
.slot
= slot
->slot
;
231 mem
.guest_phys_addr
= slot
->start_addr
;
232 mem
.userspace_addr
= (unsigned long)slot
->ram
;
233 mem
.flags
= slot
->flags
;
235 if (slot
->memory_size
&& mem
.flags
& KVM_MEM_READONLY
) {
236 /* Set the slot size to 0 before setting the slot to the desired
237 * value. This is needed based on KVM commit 75d61fbc. */
239 kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
241 mem
.memory_size
= slot
->memory_size
;
242 return kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
245 int kvm_init_vcpu(CPUState
*cpu
)
247 KVMState
*s
= kvm_state
;
251 DPRINTF("kvm_init_vcpu\n");
253 ret
= kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)kvm_arch_vcpu_id(cpu
));
255 DPRINTF("kvm_create_vcpu failed\n");
261 cpu
->kvm_vcpu_dirty
= true;
263 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
266 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
270 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
272 if (cpu
->kvm_run
== MAP_FAILED
) {
274 DPRINTF("mmap'ing vcpu state failed\n");
278 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
279 s
->coalesced_mmio_ring
=
280 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
283 ret
= kvm_arch_init_vcpu(cpu
);
289 * dirty pages logging control
292 static int kvm_mem_flags(MemoryRegion
*mr
)
294 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
297 if (memory_region_get_dirty_log_mask(mr
) != 0) {
298 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
300 if (readonly
&& kvm_readonly_mem_allowed
) {
301 flags
|= KVM_MEM_READONLY
;
306 static int kvm_slot_update_flags(KVMSlot
*mem
, MemoryRegion
*mr
)
308 KVMState
*s
= kvm_state
;
311 old_flags
= mem
->flags
;
312 mem
->flags
= kvm_mem_flags(mr
);
314 /* If nothing changed effectively, no need to issue ioctl */
315 if (mem
->flags
== old_flags
) {
319 return kvm_set_user_memory_region(s
, mem
);
322 static int kvm_section_update_flags(MemoryRegionSection
*section
)
324 KVMState
*s
= kvm_state
;
325 hwaddr phys_addr
= section
->offset_within_address_space
;
326 ram_addr_t size
= int128_get64(section
->size
);
327 KVMSlot
*mem
= kvm_lookup_matching_slot(s
, phys_addr
, phys_addr
+ size
);
332 return kvm_slot_update_flags(mem
, section
->mr
);
336 static void kvm_log_start(MemoryListener
*listener
,
337 MemoryRegionSection
*section
,
346 r
= kvm_section_update_flags(section
);
352 static void kvm_log_stop(MemoryListener
*listener
,
353 MemoryRegionSection
*section
,
362 r
= kvm_section_update_flags(section
);
368 /* get kvm's dirty pages bitmap and update qemu's */
369 static int kvm_get_dirty_pages_log_range(MemoryRegionSection
*section
,
370 unsigned long *bitmap
)
372 ram_addr_t start
= section
->offset_within_region
+ section
->mr
->ram_addr
;
373 ram_addr_t pages
= int128_get64(section
->size
) / getpagesize();
375 cpu_physical_memory_set_dirty_lebitmap(bitmap
, start
, pages
);
379 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
382 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
383 * This function updates qemu's dirty bitmap using
384 * memory_region_set_dirty(). This means all bits are set
387 * @start_add: start of logged region.
388 * @end_addr: end of logged region.
390 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection
*section
)
392 KVMState
*s
= kvm_state
;
393 unsigned long size
, allocated_size
= 0;
394 struct kvm_dirty_log d
= {};
397 hwaddr start_addr
= section
->offset_within_address_space
;
398 hwaddr end_addr
= start_addr
+ int128_get64(section
->size
);
400 d
.dirty_bitmap
= NULL
;
401 while (start_addr
< end_addr
) {
402 mem
= kvm_lookup_overlapping_slot(s
, start_addr
, end_addr
);
407 /* XXX bad kernel interface alert
408 * For dirty bitmap, kernel allocates array of size aligned to
409 * bits-per-long. But for case when the kernel is 64bits and
410 * the userspace is 32bits, userspace can't align to the same
411 * bits-per-long, since sizeof(long) is different between kernel
412 * and user space. This way, userspace will provide buffer which
413 * may be 4 bytes less than the kernel will use, resulting in
414 * userspace memory corruption (which is not detectable by valgrind
415 * too, in most cases).
416 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
417 * a hope that sizeof(long) wont become >8 any time soon.
419 size
= ALIGN(((mem
->memory_size
) >> TARGET_PAGE_BITS
),
420 /*HOST_LONG_BITS*/ 64) / 8;
421 if (!d
.dirty_bitmap
) {
422 d
.dirty_bitmap
= g_malloc(size
);
423 } else if (size
> allocated_size
) {
424 d
.dirty_bitmap
= g_realloc(d
.dirty_bitmap
, size
);
426 allocated_size
= size
;
427 memset(d
.dirty_bitmap
, 0, allocated_size
);
431 if (kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
) == -1) {
432 DPRINTF("ioctl failed %d\n", errno
);
437 kvm_get_dirty_pages_log_range(section
, d
.dirty_bitmap
);
438 start_addr
= mem
->start_addr
+ mem
->memory_size
;
440 g_free(d
.dirty_bitmap
);
445 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
446 MemoryRegionSection
*secion
,
447 hwaddr start
, hwaddr size
)
449 KVMState
*s
= kvm_state
;
451 if (s
->coalesced_mmio
) {
452 struct kvm_coalesced_mmio_zone zone
;
458 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
462 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
463 MemoryRegionSection
*secion
,
464 hwaddr start
, hwaddr size
)
466 KVMState
*s
= kvm_state
;
468 if (s
->coalesced_mmio
) {
469 struct kvm_coalesced_mmio_zone zone
;
475 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
479 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
483 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
491 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
495 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
497 /* VM wide version not implemented, use global one instead */
498 ret
= kvm_check_extension(s
, extension
);
504 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
506 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
507 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
508 * endianness, but the memory core hands them in target endianness.
509 * For example, PPC is always treated as big-endian even if running
510 * on KVM and on PPC64LE. Correct here.
524 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
525 bool assign
, uint32_t size
, bool datamatch
)
528 struct kvm_ioeventfd iofd
= {
529 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
536 if (!kvm_enabled()) {
541 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
544 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
547 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
556 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
557 bool assign
, uint32_t size
, bool datamatch
)
559 struct kvm_ioeventfd kick
= {
560 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
562 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
567 if (!kvm_enabled()) {
571 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
574 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
576 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
584 static int kvm_check_many_ioeventfds(void)
586 /* Userspace can use ioeventfd for io notification. This requires a host
587 * that supports eventfd(2) and an I/O thread; since eventfd does not
588 * support SIGIO it cannot interrupt the vcpu.
590 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
591 * can avoid creating too many ioeventfds.
593 #if defined(CONFIG_EVENTFD)
596 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
597 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
598 if (ioeventfds
[i
] < 0) {
601 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
603 close(ioeventfds
[i
]);
608 /* Decide whether many devices are supported or not */
609 ret
= i
== ARRAY_SIZE(ioeventfds
);
612 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
613 close(ioeventfds
[i
]);
621 static const KVMCapabilityInfo
*
622 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
625 if (!kvm_check_extension(s
, list
->value
)) {
633 static void kvm_set_phys_mem(MemoryRegionSection
*section
, bool add
)
635 KVMState
*s
= kvm_state
;
638 MemoryRegion
*mr
= section
->mr
;
639 bool writeable
= !mr
->readonly
&& !mr
->rom_device
;
640 hwaddr start_addr
= section
->offset_within_address_space
;
641 ram_addr_t size
= int128_get64(section
->size
);
645 /* kvm works in page size chunks, but the function may be called
646 with sub-page size and unaligned start address. Pad the start
647 address to next and truncate size to previous page boundary. */
648 delta
= (TARGET_PAGE_SIZE
- (start_addr
& ~TARGET_PAGE_MASK
));
649 delta
&= ~TARGET_PAGE_MASK
;
655 size
&= TARGET_PAGE_MASK
;
656 if (!size
|| (start_addr
& ~TARGET_PAGE_MASK
)) {
660 if (!memory_region_is_ram(mr
)) {
661 if (writeable
|| !kvm_readonly_mem_allowed
) {
663 } else if (!mr
->romd_mode
) {
664 /* If the memory device is not in romd_mode, then we actually want
665 * to remove the kvm memory slot so all accesses will trap. */
670 ram
= memory_region_get_ram_ptr(mr
) + section
->offset_within_region
+ delta
;
673 mem
= kvm_lookup_overlapping_slot(s
, start_addr
, start_addr
+ size
);
678 if (add
&& start_addr
>= mem
->start_addr
&&
679 (start_addr
+ size
<= mem
->start_addr
+ mem
->memory_size
) &&
680 (ram
- start_addr
== mem
->ram
- mem
->start_addr
)) {
681 /* The new slot fits into the existing one and comes with
682 * identical parameters - update flags and done. */
683 kvm_slot_update_flags(mem
, mr
);
689 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
690 kvm_physical_sync_dirty_bitmap(section
);
693 /* unregister the overlapping slot */
694 mem
->memory_size
= 0;
695 err
= kvm_set_user_memory_region(s
, mem
);
697 fprintf(stderr
, "%s: error unregistering overlapping slot: %s\n",
698 __func__
, strerror(-err
));
702 /* Workaround for older KVM versions: we can't join slots, even not by
703 * unregistering the previous ones and then registering the larger
704 * slot. We have to maintain the existing fragmentation. Sigh.
706 * This workaround assumes that the new slot starts at the same
707 * address as the first existing one. If not or if some overlapping
708 * slot comes around later, we will fail (not seen in practice so far)
709 * - and actually require a recent KVM version. */
710 if (s
->broken_set_mem_region
&&
711 old
.start_addr
== start_addr
&& old
.memory_size
< size
&& add
) {
712 mem
= kvm_alloc_slot(s
);
713 mem
->memory_size
= old
.memory_size
;
714 mem
->start_addr
= old
.start_addr
;
716 mem
->flags
= kvm_mem_flags(mr
);
718 err
= kvm_set_user_memory_region(s
, mem
);
720 fprintf(stderr
, "%s: error updating slot: %s\n", __func__
,
725 start_addr
+= old
.memory_size
;
726 ram
+= old
.memory_size
;
727 size
-= old
.memory_size
;
731 /* register prefix slot */
732 if (old
.start_addr
< start_addr
) {
733 mem
= kvm_alloc_slot(s
);
734 mem
->memory_size
= start_addr
- old
.start_addr
;
735 mem
->start_addr
= old
.start_addr
;
737 mem
->flags
= kvm_mem_flags(mr
);
739 err
= kvm_set_user_memory_region(s
, mem
);
741 fprintf(stderr
, "%s: error registering prefix slot: %s\n",
742 __func__
, strerror(-err
));
744 fprintf(stderr
, "%s: This is probably because your kernel's " \
745 "PAGE_SIZE is too big. Please try to use 4k " \
746 "PAGE_SIZE!\n", __func__
);
752 /* register suffix slot */
753 if (old
.start_addr
+ old
.memory_size
> start_addr
+ size
) {
754 ram_addr_t size_delta
;
756 mem
= kvm_alloc_slot(s
);
757 mem
->start_addr
= start_addr
+ size
;
758 size_delta
= mem
->start_addr
- old
.start_addr
;
759 mem
->memory_size
= old
.memory_size
- size_delta
;
760 mem
->ram
= old
.ram
+ size_delta
;
761 mem
->flags
= kvm_mem_flags(mr
);
763 err
= kvm_set_user_memory_region(s
, mem
);
765 fprintf(stderr
, "%s: error registering suffix slot: %s\n",
766 __func__
, strerror(-err
));
772 /* in case the KVM bug workaround already "consumed" the new slot */
779 mem
= kvm_alloc_slot(s
);
780 mem
->memory_size
= size
;
781 mem
->start_addr
= start_addr
;
783 mem
->flags
= kvm_mem_flags(mr
);
785 err
= kvm_set_user_memory_region(s
, mem
);
787 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
793 static void kvm_region_add(MemoryListener
*listener
,
794 MemoryRegionSection
*section
)
796 memory_region_ref(section
->mr
);
797 kvm_set_phys_mem(section
, true);
800 static void kvm_region_del(MemoryListener
*listener
,
801 MemoryRegionSection
*section
)
803 kvm_set_phys_mem(section
, false);
804 memory_region_unref(section
->mr
);
807 static void kvm_log_sync(MemoryListener
*listener
,
808 MemoryRegionSection
*section
)
812 r
= kvm_physical_sync_dirty_bitmap(section
);
818 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
819 MemoryRegionSection
*section
,
820 bool match_data
, uint64_t data
,
823 int fd
= event_notifier_get_fd(e
);
826 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
827 data
, true, int128_get64(section
->size
),
830 fprintf(stderr
, "%s: error adding ioeventfd: %s\n",
831 __func__
, strerror(-r
));
836 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
837 MemoryRegionSection
*section
,
838 bool match_data
, uint64_t data
,
841 int fd
= event_notifier_get_fd(e
);
844 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
845 data
, false, int128_get64(section
->size
),
852 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
853 MemoryRegionSection
*section
,
854 bool match_data
, uint64_t data
,
857 int fd
= event_notifier_get_fd(e
);
860 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
861 data
, true, int128_get64(section
->size
),
864 fprintf(stderr
, "%s: error adding ioeventfd: %s\n",
865 __func__
, strerror(-r
));
870 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
871 MemoryRegionSection
*section
,
872 bool match_data
, uint64_t data
,
876 int fd
= event_notifier_get_fd(e
);
879 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
880 data
, false, int128_get64(section
->size
),
887 static MemoryListener kvm_memory_listener
= {
888 .region_add
= kvm_region_add
,
889 .region_del
= kvm_region_del
,
890 .log_start
= kvm_log_start
,
891 .log_stop
= kvm_log_stop
,
892 .log_sync
= kvm_log_sync
,
893 .eventfd_add
= kvm_mem_ioeventfd_add
,
894 .eventfd_del
= kvm_mem_ioeventfd_del
,
895 .coalesced_mmio_add
= kvm_coalesce_mmio_region
,
896 .coalesced_mmio_del
= kvm_uncoalesce_mmio_region
,
900 static MemoryListener kvm_io_listener
= {
901 .eventfd_add
= kvm_io_ioeventfd_add
,
902 .eventfd_del
= kvm_io_ioeventfd_del
,
906 static void kvm_handle_interrupt(CPUState
*cpu
, int mask
)
908 cpu
->interrupt_request
|= mask
;
910 if (!qemu_cpu_is_self(cpu
)) {
915 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
917 struct kvm_irq_level event
;
920 assert(kvm_async_interrupts_enabled());
924 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
926 perror("kvm_set_irq");
930 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
933 #ifdef KVM_CAP_IRQ_ROUTING
934 typedef struct KVMMSIRoute
{
935 struct kvm_irq_routing_entry kroute
;
936 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
939 static void set_gsi(KVMState
*s
, unsigned int gsi
)
941 s
->used_gsi_bitmap
[gsi
/ 32] |= 1U << (gsi
% 32);
944 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
946 s
->used_gsi_bitmap
[gsi
/ 32] &= ~(1U << (gsi
% 32));
949 void kvm_init_irq_routing(KVMState
*s
)
953 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
955 unsigned int gsi_bits
, i
;
957 /* Round up so we can search ints using ffs */
958 gsi_bits
= ALIGN(gsi_count
, 32);
959 s
->used_gsi_bitmap
= g_malloc0(gsi_bits
/ 8);
960 s
->gsi_count
= gsi_count
;
962 /* Mark any over-allocated bits as already in use */
963 for (i
= gsi_count
; i
< gsi_bits
; i
++) {
968 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
969 s
->nr_allocated_irq_routes
= 0;
971 if (!s
->direct_msi
) {
972 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
973 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
977 kvm_arch_init_irq_routing(s
);
980 void kvm_irqchip_commit_routes(KVMState
*s
)
984 s
->irq_routes
->flags
= 0;
985 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
989 static void kvm_add_routing_entry(KVMState
*s
,
990 struct kvm_irq_routing_entry
*entry
)
992 struct kvm_irq_routing_entry
*new;
995 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
996 n
= s
->nr_allocated_irq_routes
* 2;
1000 size
= sizeof(struct kvm_irq_routing
);
1001 size
+= n
* sizeof(*new);
1002 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
1003 s
->nr_allocated_irq_routes
= n
;
1005 n
= s
->irq_routes
->nr
++;
1006 new = &s
->irq_routes
->entries
[n
];
1010 set_gsi(s
, entry
->gsi
);
1013 static int kvm_update_routing_entry(KVMState
*s
,
1014 struct kvm_irq_routing_entry
*new_entry
)
1016 struct kvm_irq_routing_entry
*entry
;
1019 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
1020 entry
= &s
->irq_routes
->entries
[n
];
1021 if (entry
->gsi
!= new_entry
->gsi
) {
1025 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1029 *entry
= *new_entry
;
1031 kvm_irqchip_commit_routes(s
);
1039 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1041 struct kvm_irq_routing_entry e
= {};
1043 assert(pin
< s
->gsi_count
);
1046 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1048 e
.u
.irqchip
.irqchip
= irqchip
;
1049 e
.u
.irqchip
.pin
= pin
;
1050 kvm_add_routing_entry(s
, &e
);
1053 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1055 struct kvm_irq_routing_entry
*e
;
1058 if (kvm_gsi_direct_mapping()) {
1062 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1063 e
= &s
->irq_routes
->entries
[i
];
1064 if (e
->gsi
== virq
) {
1065 s
->irq_routes
->nr
--;
1066 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1072 static unsigned int kvm_hash_msi(uint32_t data
)
1074 /* This is optimized for IA32 MSI layout. However, no other arch shall
1075 * repeat the mistake of not providing a direct MSI injection API. */
1079 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1081 KVMMSIRoute
*route
, *next
;
1084 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1085 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1086 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1087 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1093 static int kvm_irqchip_get_virq(KVMState
*s
)
1095 uint32_t *word
= s
->used_gsi_bitmap
;
1096 int max_words
= ALIGN(s
->gsi_count
, 32) / 32;
1100 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1101 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1102 * number can succeed even though a new route entry cannot be added.
1103 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1105 if (!s
->direct_msi
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1106 kvm_flush_dynamic_msi_routes(s
);
1109 /* Return the lowest unused GSI in the bitmap */
1110 for (i
= 0; i
< max_words
; i
++) {
1111 zeroes
= ctz32(~word
[i
]);
1116 return zeroes
+ i
* 32;
1122 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1124 unsigned int hash
= kvm_hash_msi(msg
.data
);
1127 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1128 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1129 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1130 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1137 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1142 if (s
->direct_msi
) {
1143 msi
.address_lo
= (uint32_t)msg
.address
;
1144 msi
.address_hi
= msg
.address
>> 32;
1145 msi
.data
= le32_to_cpu(msg
.data
);
1147 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1149 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1152 route
= kvm_lookup_msi_route(s
, msg
);
1156 virq
= kvm_irqchip_get_virq(s
);
1161 route
= g_malloc0(sizeof(KVMMSIRoute
));
1162 route
->kroute
.gsi
= virq
;
1163 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1164 route
->kroute
.flags
= 0;
1165 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1166 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1167 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1169 kvm_add_routing_entry(s
, &route
->kroute
);
1170 kvm_irqchip_commit_routes(s
);
1172 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1176 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1178 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
1181 int kvm_irqchip_add_msi_route(KVMState
*s
, MSIMessage msg
)
1183 struct kvm_irq_routing_entry kroute
= {};
1186 if (kvm_gsi_direct_mapping()) {
1187 return kvm_arch_msi_data_to_gsi(msg
.data
);
1190 if (!kvm_gsi_routing_enabled()) {
1194 virq
= kvm_irqchip_get_virq(s
);
1200 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1202 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1203 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1204 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1205 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
)) {
1206 kvm_irqchip_release_virq(s
, virq
);
1210 kvm_add_routing_entry(s
, &kroute
);
1211 kvm_irqchip_commit_routes(s
);
1216 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
1218 struct kvm_irq_routing_entry kroute
= {};
1220 if (kvm_gsi_direct_mapping()) {
1224 if (!kvm_irqchip_in_kernel()) {
1229 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1231 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1232 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1233 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1234 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
)) {
1238 return kvm_update_routing_entry(s
, &kroute
);
1241 static int kvm_irqchip_assign_irqfd(KVMState
*s
, int fd
, int rfd
, int virq
,
1244 struct kvm_irqfd irqfd
= {
1247 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
1251 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
1252 irqfd
.resamplefd
= rfd
;
1255 if (!kvm_irqfds_enabled()) {
1259 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
1262 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1264 struct kvm_irq_routing_entry kroute
= {};
1267 if (!kvm_gsi_routing_enabled()) {
1271 virq
= kvm_irqchip_get_virq(s
);
1277 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
1279 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
1280 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
1281 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
1282 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
1283 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
1285 kvm_add_routing_entry(s
, &kroute
);
1286 kvm_irqchip_commit_routes(s
);
1291 #else /* !KVM_CAP_IRQ_ROUTING */
1293 void kvm_init_irq_routing(KVMState
*s
)
1297 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1301 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1306 int kvm_irqchip_add_msi_route(KVMState
*s
, MSIMessage msg
)
1311 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1316 static int kvm_irqchip_assign_irqfd(KVMState
*s
, int fd
, int virq
, bool assign
)
1321 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
1325 #endif /* !KVM_CAP_IRQ_ROUTING */
1327 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1328 EventNotifier
*rn
, int virq
)
1330 return kvm_irqchip_assign_irqfd(s
, event_notifier_get_fd(n
),
1331 rn
? event_notifier_get_fd(rn
) : -1, virq
, true);
1334 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
, int virq
)
1336 return kvm_irqchip_assign_irqfd(s
, event_notifier_get_fd(n
), -1, virq
,
1340 static int kvm_irqchip_create(MachineState
*machine
, KVMState
*s
)
1344 if (!machine_kernel_irqchip_allowed(machine
) ||
1345 (!kvm_check_extension(s
, KVM_CAP_IRQCHIP
) &&
1346 (kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0) < 0))) {
1350 /* First probe and see if there's a arch-specific hook to create the
1351 * in-kernel irqchip for us */
1352 ret
= kvm_arch_irqchip_create(s
);
1355 } else if (ret
== 0) {
1356 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
1358 fprintf(stderr
, "Create kernel irqchip failed\n");
1363 kvm_kernel_irqchip
= true;
1364 /* If we have an in-kernel IRQ chip then we must have asynchronous
1365 * interrupt delivery (though the reverse is not necessarily true)
1367 kvm_async_interrupts_allowed
= true;
1368 kvm_halt_in_kernel_allowed
= true;
1370 kvm_init_irq_routing(s
);
1375 /* Find number of supported CPUs using the recommended
1376 * procedure from the kernel API documentation to cope with
1377 * older kernels that may be missing capabilities.
1379 static int kvm_recommended_vcpus(KVMState
*s
)
1381 int ret
= kvm_check_extension(s
, KVM_CAP_NR_VCPUS
);
1382 return (ret
) ? ret
: 4;
1385 static int kvm_max_vcpus(KVMState
*s
)
1387 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
1388 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
1391 static int kvm_init(MachineState
*ms
)
1393 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
1394 static const char upgrade_note
[] =
1395 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1396 "(see http://sourceforge.net/projects/kvm).\n";
1401 { "SMP", smp_cpus
},
1402 { "hotpluggable", max_cpus
},
1405 int soft_vcpus_limit
, hard_vcpus_limit
;
1407 const KVMCapabilityInfo
*missing_cap
;
1410 const char *kvm_type
;
1412 s
= KVM_STATE(ms
->accelerator
);
1415 * On systems where the kernel can support different base page
1416 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1417 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1418 * page size for the system though.
1420 assert(TARGET_PAGE_SIZE
<= getpagesize());
1425 #ifdef KVM_CAP_SET_GUEST_DEBUG
1426 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
1429 s
->fd
= qemu_open("/dev/kvm", O_RDWR
);
1431 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
1436 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
1437 if (ret
< KVM_API_VERSION
) {
1441 fprintf(stderr
, "kvm version too old\n");
1445 if (ret
> KVM_API_VERSION
) {
1447 fprintf(stderr
, "kvm version not supported\n");
1451 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
1453 /* If unspecified, use the default value */
1458 s
->slots
= g_malloc0(s
->nr_slots
* sizeof(KVMSlot
));
1460 for (i
= 0; i
< s
->nr_slots
; i
++) {
1461 s
->slots
[i
].slot
= i
;
1464 /* check the vcpu limits */
1465 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
1466 hard_vcpus_limit
= kvm_max_vcpus(s
);
1469 if (nc
->num
> soft_vcpus_limit
) {
1471 "Warning: Number of %s cpus requested (%d) exceeds "
1472 "the recommended cpus supported by KVM (%d)\n",
1473 nc
->name
, nc
->num
, soft_vcpus_limit
);
1475 if (nc
->num
> hard_vcpus_limit
) {
1476 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
1477 "the maximum cpus supported by KVM (%d)\n",
1478 nc
->name
, nc
->num
, hard_vcpus_limit
);
1485 kvm_type
= qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1487 type
= mc
->kvm_type(kvm_type
);
1488 } else if (kvm_type
) {
1490 fprintf(stderr
, "Invalid argument kvm-type=%s\n", kvm_type
);
1495 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
1496 } while (ret
== -EINTR
);
1499 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
1503 if (ret
== -EINVAL
) {
1505 "Host kernel setup problem detected. Please verify:\n");
1506 fprintf(stderr
, "- for kernels supporting the switch_amode or"
1507 " user_mode parameters, whether\n");
1509 " user space is running in primary address space\n");
1511 "- for kernels supporting the vm.allocate_pgste sysctl, "
1512 "whether it is enabled\n");
1519 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
1522 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
1526 fprintf(stderr
, "kvm does not support %s\n%s",
1527 missing_cap
->name
, upgrade_note
);
1531 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
1533 s
->broken_set_mem_region
= 1;
1534 ret
= kvm_check_extension(s
, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
);
1536 s
->broken_set_mem_region
= 0;
1539 #ifdef KVM_CAP_VCPU_EVENTS
1540 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
1543 s
->robust_singlestep
=
1544 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
1546 #ifdef KVM_CAP_DEBUGREGS
1547 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
1550 #ifdef KVM_CAP_XSAVE
1551 s
->xsave
= kvm_check_extension(s
, KVM_CAP_XSAVE
);
1555 s
->xcrs
= kvm_check_extension(s
, KVM_CAP_XCRS
);
1558 #ifdef KVM_CAP_PIT_STATE2
1559 s
->pit_state2
= kvm_check_extension(s
, KVM_CAP_PIT_STATE2
);
1562 #ifdef KVM_CAP_IRQ_ROUTING
1563 s
->direct_msi
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
1566 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
1568 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
1569 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
1570 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
1573 #ifdef KVM_CAP_READONLY_MEM
1574 kvm_readonly_mem_allowed
=
1575 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
1578 kvm_eventfds_allowed
=
1579 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
1581 kvm_irqfds_allowed
=
1582 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
1584 kvm_resamplefds_allowed
=
1585 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
1587 kvm_vm_attributes_allowed
=
1588 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
1590 ret
= kvm_arch_init(ms
, s
);
1595 ret
= kvm_irqchip_create(ms
, s
);
1601 memory_listener_register(&kvm_memory_listener
, &address_space_memory
);
1602 memory_listener_register(&kvm_io_listener
, &address_space_io
);
1604 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
1606 cpu_interrupt_handler
= kvm_handle_interrupt
;
1623 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
1625 s
->sigmask_len
= sigmask_len
;
1628 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
1629 int size
, uint32_t count
)
1632 uint8_t *ptr
= data
;
1634 for (i
= 0; i
< count
; i
++) {
1635 address_space_rw(&address_space_io
, port
, attrs
,
1637 direction
== KVM_EXIT_IO_OUT
);
1642 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
1644 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
1645 run
->internal
.suberror
);
1647 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
1650 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
1651 fprintf(stderr
, "extra data[%d]: %"PRIx64
"\n",
1652 i
, (uint64_t)run
->internal
.data
[i
]);
1655 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
1656 fprintf(stderr
, "emulation failure\n");
1657 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
1658 cpu_dump_state(cpu
, stderr
, fprintf
, CPU_DUMP_CODE
);
1659 return EXCP_INTERRUPT
;
1662 /* FIXME: Should trigger a qmp message to let management know
1663 * something went wrong.
1668 void kvm_flush_coalesced_mmio_buffer(void)
1670 KVMState
*s
= kvm_state
;
1672 if (s
->coalesced_flush_in_progress
) {
1676 s
->coalesced_flush_in_progress
= true;
1678 if (s
->coalesced_mmio_ring
) {
1679 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
1680 while (ring
->first
!= ring
->last
) {
1681 struct kvm_coalesced_mmio
*ent
;
1683 ent
= &ring
->coalesced_mmio
[ring
->first
];
1685 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
1687 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
1691 s
->coalesced_flush_in_progress
= false;
1694 static void do_kvm_cpu_synchronize_state(void *arg
)
1696 CPUState
*cpu
= arg
;
1698 if (!cpu
->kvm_vcpu_dirty
) {
1699 kvm_arch_get_registers(cpu
);
1700 cpu
->kvm_vcpu_dirty
= true;
1704 void kvm_cpu_synchronize_state(CPUState
*cpu
)
1706 if (!cpu
->kvm_vcpu_dirty
) {
1707 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, cpu
);
1711 static void do_kvm_cpu_synchronize_post_reset(void *arg
)
1713 CPUState
*cpu
= arg
;
1715 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
1716 cpu
->kvm_vcpu_dirty
= false;
1719 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
1721 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, cpu
);
1724 static void do_kvm_cpu_synchronize_post_init(void *arg
)
1726 CPUState
*cpu
= arg
;
1728 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
1729 cpu
->kvm_vcpu_dirty
= false;
1732 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
1734 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, cpu
);
1737 void kvm_cpu_clean_state(CPUState
*cpu
)
1739 cpu
->kvm_vcpu_dirty
= false;
1742 int kvm_cpu_exec(CPUState
*cpu
)
1744 struct kvm_run
*run
= cpu
->kvm_run
;
1747 DPRINTF("kvm_cpu_exec()\n");
1749 if (kvm_arch_process_async_events(cpu
)) {
1750 cpu
->exit_request
= 0;
1754 qemu_mutex_unlock_iothread();
1759 if (cpu
->kvm_vcpu_dirty
) {
1760 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
1761 cpu
->kvm_vcpu_dirty
= false;
1764 kvm_arch_pre_run(cpu
, run
);
1765 if (cpu
->exit_request
) {
1766 DPRINTF("interrupt exit requested\n");
1768 * KVM requires us to reenter the kernel after IO exits to complete
1769 * instruction emulation. This self-signal will ensure that we
1772 qemu_cpu_kick_self();
1775 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
1777 attrs
= kvm_arch_post_run(cpu
, run
);
1780 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
1781 DPRINTF("io window exit\n");
1782 ret
= EXCP_INTERRUPT
;
1785 fprintf(stderr
, "error: kvm run failed %s\n",
1786 strerror(-run_ret
));
1788 if (run_ret
== -EBUSY
) {
1790 "This is probably because your SMT is enabled.\n"
1791 "VCPU can only run on primary threads with all "
1792 "secondary threads offline.\n");
1799 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
1800 switch (run
->exit_reason
) {
1802 DPRINTF("handle_io\n");
1803 /* Called outside BQL */
1804 kvm_handle_io(run
->io
.port
, attrs
,
1805 (uint8_t *)run
+ run
->io
.data_offset
,
1812 DPRINTF("handle_mmio\n");
1813 /* Called outside BQL */
1814 address_space_rw(&address_space_memory
,
1815 run
->mmio
.phys_addr
, attrs
,
1818 run
->mmio
.is_write
);
1821 case KVM_EXIT_IRQ_WINDOW_OPEN
:
1822 DPRINTF("irq_window_open\n");
1823 ret
= EXCP_INTERRUPT
;
1825 case KVM_EXIT_SHUTDOWN
:
1826 DPRINTF("shutdown\n");
1827 qemu_system_reset_request();
1828 ret
= EXCP_INTERRUPT
;
1830 case KVM_EXIT_UNKNOWN
:
1831 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
1832 (uint64_t)run
->hw
.hardware_exit_reason
);
1835 case KVM_EXIT_INTERNAL_ERROR
:
1836 ret
= kvm_handle_internal_error(cpu
, run
);
1838 case KVM_EXIT_SYSTEM_EVENT
:
1839 switch (run
->system_event
.type
) {
1840 case KVM_SYSTEM_EVENT_SHUTDOWN
:
1841 qemu_system_shutdown_request();
1842 ret
= EXCP_INTERRUPT
;
1844 case KVM_SYSTEM_EVENT_RESET
:
1845 qemu_system_reset_request();
1846 ret
= EXCP_INTERRUPT
;
1849 DPRINTF("kvm_arch_handle_exit\n");
1850 ret
= kvm_arch_handle_exit(cpu
, run
);
1855 DPRINTF("kvm_arch_handle_exit\n");
1856 ret
= kvm_arch_handle_exit(cpu
, run
);
1861 qemu_mutex_lock_iothread();
1864 cpu_dump_state(cpu
, stderr
, fprintf
, CPU_DUMP_CODE
);
1865 vm_stop(RUN_STATE_INTERNAL_ERROR
);
1868 cpu
->exit_request
= 0;
1872 int kvm_ioctl(KVMState
*s
, int type
, ...)
1879 arg
= va_arg(ap
, void *);
1882 trace_kvm_ioctl(type
, arg
);
1883 ret
= ioctl(s
->fd
, type
, arg
);
1890 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
1897 arg
= va_arg(ap
, void *);
1900 trace_kvm_vm_ioctl(type
, arg
);
1901 ret
= ioctl(s
->vmfd
, type
, arg
);
1908 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
1915 arg
= va_arg(ap
, void *);
1918 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
1919 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
1926 int kvm_device_ioctl(int fd
, int type
, ...)
1933 arg
= va_arg(ap
, void *);
1936 trace_kvm_device_ioctl(fd
, type
, arg
);
1937 ret
= ioctl(fd
, type
, arg
);
1944 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
1947 struct kvm_device_attr attribute
= {
1952 if (!kvm_vm_attributes_allowed
) {
1956 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
1957 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
1961 int kvm_has_sync_mmu(void)
1963 return kvm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
1966 int kvm_has_vcpu_events(void)
1968 return kvm_state
->vcpu_events
;
1971 int kvm_has_robust_singlestep(void)
1973 return kvm_state
->robust_singlestep
;
1976 int kvm_has_debugregs(void)
1978 return kvm_state
->debugregs
;
1981 int kvm_has_xsave(void)
1983 return kvm_state
->xsave
;
1986 int kvm_has_xcrs(void)
1988 return kvm_state
->xcrs
;
1991 int kvm_has_pit_state2(void)
1993 return kvm_state
->pit_state2
;
1996 int kvm_has_many_ioeventfds(void)
1998 if (!kvm_enabled()) {
2001 return kvm_state
->many_ioeventfds
;
2004 int kvm_has_gsi_routing(void)
2006 #ifdef KVM_CAP_IRQ_ROUTING
2007 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
2013 int kvm_has_intx_set_mask(void)
2015 return kvm_state
->intx_set_mask
;
2018 void kvm_setup_guest_memory(void *start
, size_t size
)
2020 if (!kvm_has_sync_mmu()) {
2021 int ret
= qemu_madvise(start
, size
, QEMU_MADV_DONTFORK
);
2024 perror("qemu_madvise");
2026 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2032 #ifdef KVM_CAP_SET_GUEST_DEBUG
2033 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
2036 struct kvm_sw_breakpoint
*bp
;
2038 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
2046 int kvm_sw_breakpoints_active(CPUState
*cpu
)
2048 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
2051 struct kvm_set_guest_debug_data
{
2052 struct kvm_guest_debug dbg
;
2057 static void kvm_invoke_set_guest_debug(void *data
)
2059 struct kvm_set_guest_debug_data
*dbg_data
= data
;
2061 dbg_data
->err
= kvm_vcpu_ioctl(dbg_data
->cpu
, KVM_SET_GUEST_DEBUG
,
2065 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2067 struct kvm_set_guest_debug_data data
;
2069 data
.dbg
.control
= reinject_trap
;
2071 if (cpu
->singlestep_enabled
) {
2072 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
2074 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
2077 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
, &data
);
2081 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2082 target_ulong len
, int type
)
2084 struct kvm_sw_breakpoint
*bp
;
2087 if (type
== GDB_BREAKPOINT_SW
) {
2088 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2094 bp
= g_malloc(sizeof(struct kvm_sw_breakpoint
));
2097 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
2103 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2105 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
2112 err
= kvm_update_guest_debug(cpu
, 0);
2120 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2121 target_ulong len
, int type
)
2123 struct kvm_sw_breakpoint
*bp
;
2126 if (type
== GDB_BREAKPOINT_SW
) {
2127 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2132 if (bp
->use_count
> 1) {
2137 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
2142 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2145 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
2152 err
= kvm_update_guest_debug(cpu
, 0);
2160 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2162 struct kvm_sw_breakpoint
*bp
, *next
;
2163 KVMState
*s
= cpu
->kvm_state
;
2166 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
2167 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
2168 /* Try harder to find a CPU that currently sees the breakpoint. */
2169 CPU_FOREACH(tmpcpu
) {
2170 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
2175 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
2178 kvm_arch_remove_all_hw_breakpoints();
2181 kvm_update_guest_debug(cpu
, 0);
2185 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2187 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2192 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2193 target_ulong len
, int type
)
2198 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2199 target_ulong len
, int type
)
2204 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2207 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2209 int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
2211 KVMState
*s
= kvm_state
;
2212 struct kvm_signal_mask
*sigmask
;
2216 return kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, NULL
);
2219 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
2221 sigmask
->len
= s
->sigmask_len
;
2222 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
2223 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
2228 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
2230 return kvm_arch_on_sigbus_vcpu(cpu
, code
, addr
);
2233 int kvm_on_sigbus(int code
, void *addr
)
2235 return kvm_arch_on_sigbus(code
, addr
);
2238 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
2241 struct kvm_create_device create_dev
;
2243 create_dev
.type
= type
;
2245 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
2247 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
2251 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
2256 return test
? 0 : create_dev
.fd
;
2259 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
2261 struct kvm_one_reg reg
;
2265 reg
.addr
= (uintptr_t) source
;
2266 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
2268 trace_kvm_failed_reg_set(id
, strerror(r
));
2273 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
2275 struct kvm_one_reg reg
;
2279 reg
.addr
= (uintptr_t) target
;
2280 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
2282 trace_kvm_failed_reg_get(id
, strerror(r
));
2287 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
2289 AccelClass
*ac
= ACCEL_CLASS(oc
);
2291 ac
->init_machine
= kvm_init
;
2292 ac
->allowed
= &kvm_allowed
;
2295 static const TypeInfo kvm_accel_type
= {
2296 .name
= TYPE_KVM_ACCEL
,
2297 .parent
= TYPE_ACCEL
,
2298 .class_init
= kvm_accel_class_init
,
2299 .instance_size
= sizeof(KVMState
),
2302 static void kvm_type_init(void)
2304 type_register_static(&kvm_accel_type
);
2307 type_init(kvm_type_init
);