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 "qemu/osdep.h"
17 #include <sys/ioctl.h>
19 #include <linux/kvm.h>
21 #include "qemu-common.h"
22 #include "qemu/atomic.h"
23 #include "qemu/option.h"
24 #include "qemu/config-file.h"
25 #include "qemu/error-report.h"
26 #include "qapi/error.h"
28 #include "hw/pci/msi.h"
29 #include "hw/pci/msix.h"
30 #include "hw/s390x/adapter.h"
31 #include "exec/gdbstub.h"
32 #include "sysemu/kvm_int.h"
33 #include "sysemu/cpus.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"
42 #include "hw/boards.h"
44 /* This check must be after config-host.h is included */
46 #include <sys/eventfd.h>
49 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
50 * need to use the real host PAGE_SIZE, as that's what KVM will use.
52 #define PAGE_SIZE getpagesize()
57 #define DPRINTF(fmt, ...) \
58 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
60 #define DPRINTF(fmt, ...) \
64 #define KVM_MSI_HASHTAB_SIZE 256
66 struct KVMParkedVcpu
{
67 unsigned long vcpu_id
;
69 QLIST_ENTRY(KVMParkedVcpu
) node
;
74 AccelState parent_obj
;
80 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
81 bool coalesced_flush_in_progress
;
83 int robust_singlestep
;
85 #ifdef KVM_CAP_SET_GUEST_DEBUG
86 struct kvm_sw_breakpoint_head kvm_sw_breakpoints
;
91 /* The man page (and posix) say ioctl numbers are signed int, but
92 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
93 * unsigned, and treating them as signed here can break things */
94 unsigned irq_set_ioctl
;
95 unsigned int sigmask_len
;
97 #ifdef KVM_CAP_IRQ_ROUTING
98 struct kvm_irq_routing
*irq_routes
;
99 int nr_allocated_irq_routes
;
100 unsigned long *used_gsi_bitmap
;
101 unsigned int gsi_count
;
102 QTAILQ_HEAD(msi_hashtab
, KVMMSIRoute
) msi_hashtab
[KVM_MSI_HASHTAB_SIZE
];
104 KVMMemoryListener memory_listener
;
105 QLIST_HEAD(, KVMParkedVcpu
) kvm_parked_vcpus
;
109 bool kvm_kernel_irqchip
;
110 bool kvm_split_irqchip
;
111 bool kvm_async_interrupts_allowed
;
112 bool kvm_halt_in_kernel_allowed
;
113 bool kvm_eventfds_allowed
;
114 bool kvm_irqfds_allowed
;
115 bool kvm_resamplefds_allowed
;
116 bool kvm_msi_via_irqfd_allowed
;
117 bool kvm_gsi_routing_allowed
;
118 bool kvm_gsi_direct_mapping
;
120 bool kvm_readonly_mem_allowed
;
121 bool kvm_vm_attributes_allowed
;
122 bool kvm_direct_msi_allowed
;
123 bool kvm_ioeventfd_any_length_allowed
;
124 bool kvm_msi_use_devid
;
125 static bool kvm_immediate_exit
;
127 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
128 KVM_CAP_INFO(USER_MEMORY
),
129 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
130 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS
),
134 int kvm_get_max_memslots(void)
136 KVMState
*s
= KVM_STATE(current_machine
->accelerator
);
141 static KVMSlot
*kvm_get_free_slot(KVMMemoryListener
*kml
)
143 KVMState
*s
= kvm_state
;
146 for (i
= 0; i
< s
->nr_slots
; i
++) {
147 if (kml
->slots
[i
].memory_size
== 0) {
148 return &kml
->slots
[i
];
155 bool kvm_has_free_slot(MachineState
*ms
)
157 KVMState
*s
= KVM_STATE(ms
->accelerator
);
159 return kvm_get_free_slot(&s
->memory_listener
);
162 static KVMSlot
*kvm_alloc_slot(KVMMemoryListener
*kml
)
164 KVMSlot
*slot
= kvm_get_free_slot(kml
);
170 fprintf(stderr
, "%s: no free slot available\n", __func__
);
174 static KVMSlot
*kvm_lookup_matching_slot(KVMMemoryListener
*kml
,
178 KVMState
*s
= kvm_state
;
181 for (i
= 0; i
< s
->nr_slots
; i
++) {
182 KVMSlot
*mem
= &kml
->slots
[i
];
184 if (start_addr
== mem
->start_addr
&& size
== mem
->memory_size
) {
193 * Calculate and align the start address and the size of the section.
194 * Return the size. If the size is 0, the aligned section is empty.
196 static hwaddr
kvm_align_section(MemoryRegionSection
*section
,
199 hwaddr size
= int128_get64(section
->size
);
202 *start
= section
->offset_within_address_space
;
204 /* kvm works in page size chunks, but the function may be called
205 with sub-page size and unaligned start address. Pad the start
206 address to next and truncate size to previous page boundary. */
207 delta
= qemu_real_host_page_size
- (*start
& ~qemu_real_host_page_mask
);
208 delta
&= ~qemu_real_host_page_mask
;
214 size
&= qemu_real_host_page_mask
;
215 if (*start
& ~qemu_real_host_page_mask
) {
222 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
225 KVMMemoryListener
*kml
= &s
->memory_listener
;
228 for (i
= 0; i
< s
->nr_slots
; i
++) {
229 KVMSlot
*mem
= &kml
->slots
[i
];
231 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
232 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
240 static int kvm_set_user_memory_region(KVMMemoryListener
*kml
, KVMSlot
*slot
)
242 KVMState
*s
= kvm_state
;
243 struct kvm_userspace_memory_region mem
;
245 mem
.slot
= slot
->slot
| (kml
->as_id
<< 16);
246 mem
.guest_phys_addr
= slot
->start_addr
;
247 mem
.userspace_addr
= (unsigned long)slot
->ram
;
248 mem
.flags
= slot
->flags
;
250 if (slot
->memory_size
&& mem
.flags
& KVM_MEM_READONLY
) {
251 /* Set the slot size to 0 before setting the slot to the desired
252 * value. This is needed based on KVM commit 75d61fbc. */
254 kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
256 mem
.memory_size
= slot
->memory_size
;
257 return kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
260 int kvm_destroy_vcpu(CPUState
*cpu
)
262 KVMState
*s
= kvm_state
;
264 struct KVMParkedVcpu
*vcpu
= NULL
;
267 DPRINTF("kvm_destroy_vcpu\n");
269 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
272 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
276 ret
= munmap(cpu
->kvm_run
, mmap_size
);
281 vcpu
= g_malloc0(sizeof(*vcpu
));
282 vcpu
->vcpu_id
= kvm_arch_vcpu_id(cpu
);
283 vcpu
->kvm_fd
= cpu
->kvm_fd
;
284 QLIST_INSERT_HEAD(&kvm_state
->kvm_parked_vcpus
, vcpu
, node
);
289 static int kvm_get_vcpu(KVMState
*s
, unsigned long vcpu_id
)
291 struct KVMParkedVcpu
*cpu
;
293 QLIST_FOREACH(cpu
, &s
->kvm_parked_vcpus
, node
) {
294 if (cpu
->vcpu_id
== vcpu_id
) {
297 QLIST_REMOVE(cpu
, node
);
298 kvm_fd
= cpu
->kvm_fd
;
304 return kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)vcpu_id
);
307 int kvm_init_vcpu(CPUState
*cpu
)
309 KVMState
*s
= kvm_state
;
313 DPRINTF("kvm_init_vcpu\n");
315 ret
= kvm_get_vcpu(s
, kvm_arch_vcpu_id(cpu
));
317 DPRINTF("kvm_create_vcpu failed\n");
323 cpu
->vcpu_dirty
= true;
325 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
328 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
332 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
334 if (cpu
->kvm_run
== MAP_FAILED
) {
336 DPRINTF("mmap'ing vcpu state failed\n");
340 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
341 s
->coalesced_mmio_ring
=
342 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
345 ret
= kvm_arch_init_vcpu(cpu
);
351 * dirty pages logging control
354 static int kvm_mem_flags(MemoryRegion
*mr
)
356 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
359 if (memory_region_get_dirty_log_mask(mr
) != 0) {
360 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
362 if (readonly
&& kvm_readonly_mem_allowed
) {
363 flags
|= KVM_MEM_READONLY
;
368 static int kvm_slot_update_flags(KVMMemoryListener
*kml
, KVMSlot
*mem
,
373 old_flags
= mem
->flags
;
374 mem
->flags
= kvm_mem_flags(mr
);
376 /* If nothing changed effectively, no need to issue ioctl */
377 if (mem
->flags
== old_flags
) {
381 return kvm_set_user_memory_region(kml
, mem
);
384 static int kvm_section_update_flags(KVMMemoryListener
*kml
,
385 MemoryRegionSection
*section
)
387 hwaddr start_addr
, size
;
390 size
= kvm_align_section(section
, &start_addr
);
395 mem
= kvm_lookup_matching_slot(kml
, start_addr
, size
);
397 fprintf(stderr
, "%s: error finding slot\n", __func__
);
401 return kvm_slot_update_flags(kml
, mem
, section
->mr
);
404 static void kvm_log_start(MemoryListener
*listener
,
405 MemoryRegionSection
*section
,
408 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
415 r
= kvm_section_update_flags(kml
, section
);
421 static void kvm_log_stop(MemoryListener
*listener
,
422 MemoryRegionSection
*section
,
425 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
432 r
= kvm_section_update_flags(kml
, section
);
438 /* get kvm's dirty pages bitmap and update qemu's */
439 static int kvm_get_dirty_pages_log_range(MemoryRegionSection
*section
,
440 unsigned long *bitmap
)
442 ram_addr_t start
= section
->offset_within_region
+
443 memory_region_get_ram_addr(section
->mr
);
444 ram_addr_t pages
= int128_get64(section
->size
) / getpagesize();
446 cpu_physical_memory_set_dirty_lebitmap(bitmap
, start
, pages
);
450 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
453 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
454 * This function updates qemu's dirty bitmap using
455 * memory_region_set_dirty(). This means all bits are set
458 * @start_add: start of logged region.
459 * @end_addr: end of logged region.
461 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener
*kml
,
462 MemoryRegionSection
*section
)
464 KVMState
*s
= kvm_state
;
465 struct kvm_dirty_log d
= {};
467 hwaddr start_addr
, size
;
469 size
= kvm_align_section(section
, &start_addr
);
471 mem
= kvm_lookup_matching_slot(kml
, start_addr
, size
);
473 fprintf(stderr
, "%s: error finding slot\n", __func__
);
477 /* XXX bad kernel interface alert
478 * For dirty bitmap, kernel allocates array of size aligned to
479 * bits-per-long. But for case when the kernel is 64bits and
480 * the userspace is 32bits, userspace can't align to the same
481 * bits-per-long, since sizeof(long) is different between kernel
482 * and user space. This way, userspace will provide buffer which
483 * may be 4 bytes less than the kernel will use, resulting in
484 * userspace memory corruption (which is not detectable by valgrind
485 * too, in most cases).
486 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
487 * a hope that sizeof(long) won't become >8 any time soon.
489 size
= ALIGN(((mem
->memory_size
) >> TARGET_PAGE_BITS
),
490 /*HOST_LONG_BITS*/ 64) / 8;
491 d
.dirty_bitmap
= g_malloc0(size
);
493 d
.slot
= mem
->slot
| (kml
->as_id
<< 16);
494 if (kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
) == -1) {
495 DPRINTF("ioctl failed %d\n", errno
);
496 g_free(d
.dirty_bitmap
);
500 kvm_get_dirty_pages_log_range(section
, d
.dirty_bitmap
);
501 g_free(d
.dirty_bitmap
);
507 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
508 MemoryRegionSection
*secion
,
509 hwaddr start
, hwaddr size
)
511 KVMState
*s
= kvm_state
;
513 if (s
->coalesced_mmio
) {
514 struct kvm_coalesced_mmio_zone zone
;
520 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
524 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
525 MemoryRegionSection
*secion
,
526 hwaddr start
, hwaddr size
)
528 KVMState
*s
= kvm_state
;
530 if (s
->coalesced_mmio
) {
531 struct kvm_coalesced_mmio_zone zone
;
537 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
541 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
545 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
553 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
557 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
559 /* VM wide version not implemented, use global one instead */
560 ret
= kvm_check_extension(s
, extension
);
566 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
568 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
569 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
570 * endianness, but the memory core hands them in target endianness.
571 * For example, PPC is always treated as big-endian even if running
572 * on KVM and on PPC64LE. Correct here.
586 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
587 bool assign
, uint32_t size
, bool datamatch
)
590 struct kvm_ioeventfd iofd
= {
591 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
598 if (!kvm_enabled()) {
603 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
606 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
609 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
618 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
619 bool assign
, uint32_t size
, bool datamatch
)
621 struct kvm_ioeventfd kick
= {
622 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
624 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
629 if (!kvm_enabled()) {
633 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
636 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
638 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
646 static int kvm_check_many_ioeventfds(void)
648 /* Userspace can use ioeventfd for io notification. This requires a host
649 * that supports eventfd(2) and an I/O thread; since eventfd does not
650 * support SIGIO it cannot interrupt the vcpu.
652 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
653 * can avoid creating too many ioeventfds.
655 #if defined(CONFIG_EVENTFD)
658 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
659 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
660 if (ioeventfds
[i
] < 0) {
663 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
665 close(ioeventfds
[i
]);
670 /* Decide whether many devices are supported or not */
671 ret
= i
== ARRAY_SIZE(ioeventfds
);
674 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
675 close(ioeventfds
[i
]);
683 static const KVMCapabilityInfo
*
684 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
687 if (!kvm_check_extension(s
, list
->value
)) {
695 static void kvm_set_phys_mem(KVMMemoryListener
*kml
,
696 MemoryRegionSection
*section
, bool add
)
700 MemoryRegion
*mr
= section
->mr
;
701 bool writeable
= !mr
->readonly
&& !mr
->rom_device
;
702 hwaddr start_addr
, size
;
705 if (!memory_region_is_ram(mr
)) {
706 if (writeable
|| !kvm_readonly_mem_allowed
) {
708 } else if (!mr
->romd_mode
) {
709 /* If the memory device is not in romd_mode, then we actually want
710 * to remove the kvm memory slot so all accesses will trap. */
715 size
= kvm_align_section(section
, &start_addr
);
720 ram
= memory_region_get_ram_ptr(mr
) + section
->offset_within_region
+
721 (section
->offset_within_address_space
- start_addr
);
723 mem
= kvm_lookup_matching_slot(kml
, start_addr
, size
);
728 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
729 kvm_physical_sync_dirty_bitmap(kml
, section
);
732 /* unregister the slot */
733 mem
->memory_size
= 0;
734 err
= kvm_set_user_memory_region(kml
, mem
);
736 fprintf(stderr
, "%s: error unregistering overlapping slot: %s\n",
737 __func__
, strerror(-err
));
744 /* update the slot */
745 kvm_slot_update_flags(kml
, mem
, mr
);
749 /* register the new slot */
750 mem
= kvm_alloc_slot(kml
);
751 mem
->memory_size
= size
;
752 mem
->start_addr
= start_addr
;
754 mem
->flags
= kvm_mem_flags(mr
);
756 err
= kvm_set_user_memory_region(kml
, mem
);
758 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
764 static void kvm_region_add(MemoryListener
*listener
,
765 MemoryRegionSection
*section
)
767 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
769 memory_region_ref(section
->mr
);
770 kvm_set_phys_mem(kml
, section
, true);
773 static void kvm_region_del(MemoryListener
*listener
,
774 MemoryRegionSection
*section
)
776 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
778 kvm_set_phys_mem(kml
, section
, false);
779 memory_region_unref(section
->mr
);
782 static void kvm_log_sync(MemoryListener
*listener
,
783 MemoryRegionSection
*section
)
785 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
788 r
= kvm_physical_sync_dirty_bitmap(kml
, section
);
794 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
795 MemoryRegionSection
*section
,
796 bool match_data
, uint64_t data
,
799 int fd
= event_notifier_get_fd(e
);
802 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
803 data
, true, int128_get64(section
->size
),
806 fprintf(stderr
, "%s: error adding ioeventfd: %s\n",
807 __func__
, strerror(-r
));
812 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
813 MemoryRegionSection
*section
,
814 bool match_data
, uint64_t data
,
817 int fd
= event_notifier_get_fd(e
);
820 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
821 data
, false, int128_get64(section
->size
),
828 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
829 MemoryRegionSection
*section
,
830 bool match_data
, uint64_t data
,
833 int fd
= event_notifier_get_fd(e
);
836 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
837 data
, true, int128_get64(section
->size
),
840 fprintf(stderr
, "%s: error adding ioeventfd: %s\n",
841 __func__
, strerror(-r
));
846 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
847 MemoryRegionSection
*section
,
848 bool match_data
, uint64_t data
,
852 int fd
= event_notifier_get_fd(e
);
855 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
856 data
, false, int128_get64(section
->size
),
863 void kvm_memory_listener_register(KVMState
*s
, KVMMemoryListener
*kml
,
864 AddressSpace
*as
, int as_id
)
868 kml
->slots
= g_malloc0(s
->nr_slots
* sizeof(KVMSlot
));
871 for (i
= 0; i
< s
->nr_slots
; i
++) {
872 kml
->slots
[i
].slot
= i
;
875 kml
->listener
.region_add
= kvm_region_add
;
876 kml
->listener
.region_del
= kvm_region_del
;
877 kml
->listener
.log_start
= kvm_log_start
;
878 kml
->listener
.log_stop
= kvm_log_stop
;
879 kml
->listener
.log_sync
= kvm_log_sync
;
880 kml
->listener
.priority
= 10;
882 memory_listener_register(&kml
->listener
, as
);
885 static MemoryListener kvm_io_listener
= {
886 .eventfd_add
= kvm_io_ioeventfd_add
,
887 .eventfd_del
= kvm_io_ioeventfd_del
,
891 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
893 struct kvm_irq_level event
;
896 assert(kvm_async_interrupts_enabled());
900 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
902 perror("kvm_set_irq");
906 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
909 #ifdef KVM_CAP_IRQ_ROUTING
910 typedef struct KVMMSIRoute
{
911 struct kvm_irq_routing_entry kroute
;
912 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
915 static void set_gsi(KVMState
*s
, unsigned int gsi
)
917 set_bit(gsi
, s
->used_gsi_bitmap
);
920 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
922 clear_bit(gsi
, s
->used_gsi_bitmap
);
925 void kvm_init_irq_routing(KVMState
*s
)
929 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
931 /* Round up so we can search ints using ffs */
932 s
->used_gsi_bitmap
= bitmap_new(gsi_count
);
933 s
->gsi_count
= gsi_count
;
936 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
937 s
->nr_allocated_irq_routes
= 0;
939 if (!kvm_direct_msi_allowed
) {
940 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
941 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
945 kvm_arch_init_irq_routing(s
);
948 void kvm_irqchip_commit_routes(KVMState
*s
)
952 if (kvm_gsi_direct_mapping()) {
956 if (!kvm_gsi_routing_enabled()) {
960 s
->irq_routes
->flags
= 0;
961 trace_kvm_irqchip_commit_routes();
962 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
966 static void kvm_add_routing_entry(KVMState
*s
,
967 struct kvm_irq_routing_entry
*entry
)
969 struct kvm_irq_routing_entry
*new;
972 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
973 n
= s
->nr_allocated_irq_routes
* 2;
977 size
= sizeof(struct kvm_irq_routing
);
978 size
+= n
* sizeof(*new);
979 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
980 s
->nr_allocated_irq_routes
= n
;
982 n
= s
->irq_routes
->nr
++;
983 new = &s
->irq_routes
->entries
[n
];
987 set_gsi(s
, entry
->gsi
);
990 static int kvm_update_routing_entry(KVMState
*s
,
991 struct kvm_irq_routing_entry
*new_entry
)
993 struct kvm_irq_routing_entry
*entry
;
996 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
997 entry
= &s
->irq_routes
->entries
[n
];
998 if (entry
->gsi
!= new_entry
->gsi
) {
1002 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1006 *entry
= *new_entry
;
1014 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1016 struct kvm_irq_routing_entry e
= {};
1018 assert(pin
< s
->gsi_count
);
1021 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1023 e
.u
.irqchip
.irqchip
= irqchip
;
1024 e
.u
.irqchip
.pin
= pin
;
1025 kvm_add_routing_entry(s
, &e
);
1028 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1030 struct kvm_irq_routing_entry
*e
;
1033 if (kvm_gsi_direct_mapping()) {
1037 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1038 e
= &s
->irq_routes
->entries
[i
];
1039 if (e
->gsi
== virq
) {
1040 s
->irq_routes
->nr
--;
1041 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1045 kvm_arch_release_virq_post(virq
);
1046 trace_kvm_irqchip_release_virq(virq
);
1049 static unsigned int kvm_hash_msi(uint32_t data
)
1051 /* This is optimized for IA32 MSI layout. However, no other arch shall
1052 * repeat the mistake of not providing a direct MSI injection API. */
1056 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1058 KVMMSIRoute
*route
, *next
;
1061 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1062 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1063 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1064 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1070 static int kvm_irqchip_get_virq(KVMState
*s
)
1075 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1076 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1077 * number can succeed even though a new route entry cannot be added.
1078 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1080 if (!kvm_direct_msi_allowed
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1081 kvm_flush_dynamic_msi_routes(s
);
1084 /* Return the lowest unused GSI in the bitmap */
1085 next_virq
= find_first_zero_bit(s
->used_gsi_bitmap
, s
->gsi_count
);
1086 if (next_virq
>= s
->gsi_count
) {
1093 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1095 unsigned int hash
= kvm_hash_msi(msg
.data
);
1098 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1099 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1100 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1101 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1108 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1113 if (kvm_direct_msi_allowed
) {
1114 msi
.address_lo
= (uint32_t)msg
.address
;
1115 msi
.address_hi
= msg
.address
>> 32;
1116 msi
.data
= le32_to_cpu(msg
.data
);
1118 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1120 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1123 route
= kvm_lookup_msi_route(s
, msg
);
1127 virq
= kvm_irqchip_get_virq(s
);
1132 route
= g_malloc0(sizeof(KVMMSIRoute
));
1133 route
->kroute
.gsi
= virq
;
1134 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1135 route
->kroute
.flags
= 0;
1136 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1137 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1138 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1140 kvm_add_routing_entry(s
, &route
->kroute
);
1141 kvm_irqchip_commit_routes(s
);
1143 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1147 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1149 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
1152 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1154 struct kvm_irq_routing_entry kroute
= {};
1156 MSIMessage msg
= {0, 0};
1158 if (pci_available
&& dev
) {
1159 msg
= pci_get_msi_message(dev
, vector
);
1162 if (kvm_gsi_direct_mapping()) {
1163 return kvm_arch_msi_data_to_gsi(msg
.data
);
1166 if (!kvm_gsi_routing_enabled()) {
1170 virq
= kvm_irqchip_get_virq(s
);
1176 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1178 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1179 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1180 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1181 if (pci_available
&& kvm_msi_devid_required()) {
1182 kroute
.flags
= KVM_MSI_VALID_DEVID
;
1183 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
1185 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1186 kvm_irqchip_release_virq(s
, virq
);
1190 trace_kvm_irqchip_add_msi_route(dev
? dev
->name
: (char *)"N/A",
1193 kvm_add_routing_entry(s
, &kroute
);
1194 kvm_arch_add_msi_route_post(&kroute
, vector
, dev
);
1195 kvm_irqchip_commit_routes(s
);
1200 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
,
1203 struct kvm_irq_routing_entry kroute
= {};
1205 if (kvm_gsi_direct_mapping()) {
1209 if (!kvm_irqchip_in_kernel()) {
1214 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1216 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1217 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1218 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1219 if (pci_available
&& kvm_msi_devid_required()) {
1220 kroute
.flags
= KVM_MSI_VALID_DEVID
;
1221 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
1223 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1227 trace_kvm_irqchip_update_msi_route(virq
);
1229 return kvm_update_routing_entry(s
, &kroute
);
1232 static int kvm_irqchip_assign_irqfd(KVMState
*s
, int fd
, int rfd
, int virq
,
1235 struct kvm_irqfd irqfd
= {
1238 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
1242 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
1243 irqfd
.resamplefd
= rfd
;
1246 if (!kvm_irqfds_enabled()) {
1250 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
1253 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1255 struct kvm_irq_routing_entry kroute
= {};
1258 if (!kvm_gsi_routing_enabled()) {
1262 virq
= kvm_irqchip_get_virq(s
);
1268 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
1270 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
1271 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
1272 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
1273 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
1274 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
1276 kvm_add_routing_entry(s
, &kroute
);
1281 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1283 struct kvm_irq_routing_entry kroute
= {};
1286 if (!kvm_gsi_routing_enabled()) {
1289 if (!kvm_check_extension(s
, KVM_CAP_HYPERV_SYNIC
)) {
1292 virq
= kvm_irqchip_get_virq(s
);
1298 kroute
.type
= KVM_IRQ_ROUTING_HV_SINT
;
1300 kroute
.u
.hv_sint
.vcpu
= vcpu
;
1301 kroute
.u
.hv_sint
.sint
= sint
;
1303 kvm_add_routing_entry(s
, &kroute
);
1304 kvm_irqchip_commit_routes(s
);
1309 #else /* !KVM_CAP_IRQ_ROUTING */
1311 void kvm_init_irq_routing(KVMState
*s
)
1315 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1319 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1324 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1329 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1334 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1339 static int kvm_irqchip_assign_irqfd(KVMState
*s
, int fd
, int virq
, bool assign
)
1344 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
1348 #endif /* !KVM_CAP_IRQ_ROUTING */
1350 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1351 EventNotifier
*rn
, int virq
)
1353 return kvm_irqchip_assign_irqfd(s
, event_notifier_get_fd(n
),
1354 rn
? event_notifier_get_fd(rn
) : -1, virq
, true);
1357 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1360 return kvm_irqchip_assign_irqfd(s
, event_notifier_get_fd(n
), -1, virq
,
1364 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1365 EventNotifier
*rn
, qemu_irq irq
)
1368 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1373 return kvm_irqchip_add_irqfd_notifier_gsi(s
, n
, rn
, GPOINTER_TO_INT(gsi
));
1376 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1380 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1385 return kvm_irqchip_remove_irqfd_notifier_gsi(s
, n
, GPOINTER_TO_INT(gsi
));
1388 void kvm_irqchip_set_qemuirq_gsi(KVMState
*s
, qemu_irq irq
, int gsi
)
1390 g_hash_table_insert(s
->gsimap
, irq
, GINT_TO_POINTER(gsi
));
1393 static void kvm_irqchip_create(MachineState
*machine
, KVMState
*s
)
1397 if (kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
1399 } else if (kvm_check_extension(s
, KVM_CAP_S390_IRQCHIP
)) {
1400 ret
= kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0);
1402 fprintf(stderr
, "Enable kernel irqchip failed: %s\n", strerror(-ret
));
1409 /* First probe and see if there's a arch-specific hook to create the
1410 * in-kernel irqchip for us */
1411 ret
= kvm_arch_irqchip_create(machine
, s
);
1413 if (machine_kernel_irqchip_split(machine
)) {
1414 perror("Split IRQ chip mode not supported.");
1417 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
1421 fprintf(stderr
, "Create kernel irqchip failed: %s\n", strerror(-ret
));
1425 kvm_kernel_irqchip
= true;
1426 /* If we have an in-kernel IRQ chip then we must have asynchronous
1427 * interrupt delivery (though the reverse is not necessarily true)
1429 kvm_async_interrupts_allowed
= true;
1430 kvm_halt_in_kernel_allowed
= true;
1432 kvm_init_irq_routing(s
);
1434 s
->gsimap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
1437 /* Find number of supported CPUs using the recommended
1438 * procedure from the kernel API documentation to cope with
1439 * older kernels that may be missing capabilities.
1441 static int kvm_recommended_vcpus(KVMState
*s
)
1443 int ret
= kvm_vm_check_extension(s
, KVM_CAP_NR_VCPUS
);
1444 return (ret
) ? ret
: 4;
1447 static int kvm_max_vcpus(KVMState
*s
)
1449 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
1450 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
1453 static int kvm_max_vcpu_id(KVMState
*s
)
1455 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPU_ID
);
1456 return (ret
) ? ret
: kvm_max_vcpus(s
);
1459 bool kvm_vcpu_id_is_valid(int vcpu_id
)
1461 KVMState
*s
= KVM_STATE(current_machine
->accelerator
);
1462 return vcpu_id
>= 0 && vcpu_id
< kvm_max_vcpu_id(s
);
1465 static int kvm_init(MachineState
*ms
)
1467 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
1468 static const char upgrade_note
[] =
1469 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1470 "(see http://sourceforge.net/projects/kvm).\n";
1475 { "SMP", smp_cpus
},
1476 { "hotpluggable", max_cpus
},
1479 int soft_vcpus_limit
, hard_vcpus_limit
;
1481 const KVMCapabilityInfo
*missing_cap
;
1484 const char *kvm_type
;
1486 s
= KVM_STATE(ms
->accelerator
);
1489 * On systems where the kernel can support different base page
1490 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1491 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1492 * page size for the system though.
1494 assert(TARGET_PAGE_SIZE
<= getpagesize());
1498 #ifdef KVM_CAP_SET_GUEST_DEBUG
1499 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
1501 QLIST_INIT(&s
->kvm_parked_vcpus
);
1503 s
->fd
= qemu_open("/dev/kvm", O_RDWR
);
1505 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
1510 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
1511 if (ret
< KVM_API_VERSION
) {
1515 fprintf(stderr
, "kvm version too old\n");
1519 if (ret
> KVM_API_VERSION
) {
1521 fprintf(stderr
, "kvm version not supported\n");
1525 kvm_immediate_exit
= kvm_check_extension(s
, KVM_CAP_IMMEDIATE_EXIT
);
1526 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
1528 /* If unspecified, use the default value */
1533 kvm_type
= qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1535 type
= mc
->kvm_type(kvm_type
);
1536 } else if (kvm_type
) {
1538 fprintf(stderr
, "Invalid argument kvm-type=%s\n", kvm_type
);
1543 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
1544 } while (ret
== -EINTR
);
1547 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
1551 if (ret
== -EINVAL
) {
1553 "Host kernel setup problem detected. Please verify:\n");
1554 fprintf(stderr
, "- for kernels supporting the switch_amode or"
1555 " user_mode parameters, whether\n");
1557 " user space is running in primary address space\n");
1559 "- for kernels supporting the vm.allocate_pgste sysctl, "
1560 "whether it is enabled\n");
1568 /* check the vcpu limits */
1569 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
1570 hard_vcpus_limit
= kvm_max_vcpus(s
);
1573 if (nc
->num
> soft_vcpus_limit
) {
1574 warn_report("Number of %s cpus requested (%d) exceeds "
1575 "the recommended cpus supported by KVM (%d)",
1576 nc
->name
, nc
->num
, soft_vcpus_limit
);
1578 if (nc
->num
> hard_vcpus_limit
) {
1579 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
1580 "the maximum cpus supported by KVM (%d)\n",
1581 nc
->name
, nc
->num
, hard_vcpus_limit
);
1588 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
1591 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
1595 fprintf(stderr
, "kvm does not support %s\n%s",
1596 missing_cap
->name
, upgrade_note
);
1600 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
1602 #ifdef KVM_CAP_VCPU_EVENTS
1603 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
1606 s
->robust_singlestep
=
1607 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
1609 #ifdef KVM_CAP_DEBUGREGS
1610 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
1613 #ifdef KVM_CAP_IRQ_ROUTING
1614 kvm_direct_msi_allowed
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
1617 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
1619 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
1620 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
1621 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
1624 #ifdef KVM_CAP_READONLY_MEM
1625 kvm_readonly_mem_allowed
=
1626 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
1629 kvm_eventfds_allowed
=
1630 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
1632 kvm_irqfds_allowed
=
1633 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
1635 kvm_resamplefds_allowed
=
1636 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
1638 kvm_vm_attributes_allowed
=
1639 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
1641 kvm_ioeventfd_any_length_allowed
=
1642 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD_ANY_LENGTH
) > 0);
1646 ret
= kvm_arch_init(ms
, s
);
1651 if (machine_kernel_irqchip_allowed(ms
)) {
1652 kvm_irqchip_create(ms
, s
);
1655 if (kvm_eventfds_allowed
) {
1656 s
->memory_listener
.listener
.eventfd_add
= kvm_mem_ioeventfd_add
;
1657 s
->memory_listener
.listener
.eventfd_del
= kvm_mem_ioeventfd_del
;
1659 s
->memory_listener
.listener
.coalesced_mmio_add
= kvm_coalesce_mmio_region
;
1660 s
->memory_listener
.listener
.coalesced_mmio_del
= kvm_uncoalesce_mmio_region
;
1662 kvm_memory_listener_register(s
, &s
->memory_listener
,
1663 &address_space_memory
, 0);
1664 memory_listener_register(&kvm_io_listener
,
1667 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
1669 s
->sync_mmu
= !!kvm_vm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
1681 g_free(s
->memory_listener
.slots
);
1686 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
1688 s
->sigmask_len
= sigmask_len
;
1691 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
1692 int size
, uint32_t count
)
1695 uint8_t *ptr
= data
;
1697 for (i
= 0; i
< count
; i
++) {
1698 address_space_rw(&address_space_io
, port
, attrs
,
1700 direction
== KVM_EXIT_IO_OUT
);
1705 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
1707 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
1708 run
->internal
.suberror
);
1710 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
1713 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
1714 fprintf(stderr
, "extra data[%d]: %"PRIx64
"\n",
1715 i
, (uint64_t)run
->internal
.data
[i
]);
1718 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
1719 fprintf(stderr
, "emulation failure\n");
1720 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
1721 cpu_dump_state(cpu
, stderr
, fprintf
, CPU_DUMP_CODE
);
1722 return EXCP_INTERRUPT
;
1725 /* FIXME: Should trigger a qmp message to let management know
1726 * something went wrong.
1731 void kvm_flush_coalesced_mmio_buffer(void)
1733 KVMState
*s
= kvm_state
;
1735 if (s
->coalesced_flush_in_progress
) {
1739 s
->coalesced_flush_in_progress
= true;
1741 if (s
->coalesced_mmio_ring
) {
1742 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
1743 while (ring
->first
!= ring
->last
) {
1744 struct kvm_coalesced_mmio
*ent
;
1746 ent
= &ring
->coalesced_mmio
[ring
->first
];
1748 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
1750 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
1754 s
->coalesced_flush_in_progress
= false;
1757 static void do_kvm_cpu_synchronize_state(CPUState
*cpu
, run_on_cpu_data arg
)
1759 if (!cpu
->vcpu_dirty
) {
1760 kvm_arch_get_registers(cpu
);
1761 cpu
->vcpu_dirty
= true;
1765 void kvm_cpu_synchronize_state(CPUState
*cpu
)
1767 if (!cpu
->vcpu_dirty
) {
1768 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, RUN_ON_CPU_NULL
);
1772 static void do_kvm_cpu_synchronize_post_reset(CPUState
*cpu
, run_on_cpu_data arg
)
1774 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
1775 cpu
->vcpu_dirty
= false;
1778 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
1780 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, RUN_ON_CPU_NULL
);
1783 static void do_kvm_cpu_synchronize_post_init(CPUState
*cpu
, run_on_cpu_data arg
)
1785 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
1786 cpu
->vcpu_dirty
= false;
1789 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
1791 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, RUN_ON_CPU_NULL
);
1794 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
, run_on_cpu_data arg
)
1796 cpu
->vcpu_dirty
= true;
1799 void kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
)
1801 run_on_cpu(cpu
, do_kvm_cpu_synchronize_pre_loadvm
, RUN_ON_CPU_NULL
);
1804 #ifdef KVM_HAVE_MCE_INJECTION
1805 static __thread
void *pending_sigbus_addr
;
1806 static __thread
int pending_sigbus_code
;
1807 static __thread
bool have_sigbus_pending
;
1810 static void kvm_cpu_kick(CPUState
*cpu
)
1812 atomic_set(&cpu
->kvm_run
->immediate_exit
, 1);
1815 static void kvm_cpu_kick_self(void)
1817 if (kvm_immediate_exit
) {
1818 kvm_cpu_kick(current_cpu
);
1820 qemu_cpu_kick_self();
1824 static void kvm_eat_signals(CPUState
*cpu
)
1826 struct timespec ts
= { 0, 0 };
1832 if (kvm_immediate_exit
) {
1833 atomic_set(&cpu
->kvm_run
->immediate_exit
, 0);
1834 /* Write kvm_run->immediate_exit before the cpu->exit_request
1835 * write in kvm_cpu_exec.
1841 sigemptyset(&waitset
);
1842 sigaddset(&waitset
, SIG_IPI
);
1845 r
= sigtimedwait(&waitset
, &siginfo
, &ts
);
1846 if (r
== -1 && !(errno
== EAGAIN
|| errno
== EINTR
)) {
1847 perror("sigtimedwait");
1851 r
= sigpending(&chkset
);
1853 perror("sigpending");
1856 } while (sigismember(&chkset
, SIG_IPI
));
1859 int kvm_cpu_exec(CPUState
*cpu
)
1861 struct kvm_run
*run
= cpu
->kvm_run
;
1864 DPRINTF("kvm_cpu_exec()\n");
1866 if (kvm_arch_process_async_events(cpu
)) {
1867 atomic_set(&cpu
->exit_request
, 0);
1871 qemu_mutex_unlock_iothread();
1872 cpu_exec_start(cpu
);
1877 if (cpu
->vcpu_dirty
) {
1878 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
1879 cpu
->vcpu_dirty
= false;
1882 kvm_arch_pre_run(cpu
, run
);
1883 if (atomic_read(&cpu
->exit_request
)) {
1884 DPRINTF("interrupt exit requested\n");
1886 * KVM requires us to reenter the kernel after IO exits to complete
1887 * instruction emulation. This self-signal will ensure that we
1890 kvm_cpu_kick_self();
1893 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1894 * Matching barrier in kvm_eat_signals.
1898 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
1900 attrs
= kvm_arch_post_run(cpu
, run
);
1902 #ifdef KVM_HAVE_MCE_INJECTION
1903 if (unlikely(have_sigbus_pending
)) {
1904 qemu_mutex_lock_iothread();
1905 kvm_arch_on_sigbus_vcpu(cpu
, pending_sigbus_code
,
1906 pending_sigbus_addr
);
1907 have_sigbus_pending
= false;
1908 qemu_mutex_unlock_iothread();
1913 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
1914 DPRINTF("io window exit\n");
1915 kvm_eat_signals(cpu
);
1916 ret
= EXCP_INTERRUPT
;
1919 fprintf(stderr
, "error: kvm run failed %s\n",
1920 strerror(-run_ret
));
1922 if (run_ret
== -EBUSY
) {
1924 "This is probably because your SMT is enabled.\n"
1925 "VCPU can only run on primary threads with all "
1926 "secondary threads offline.\n");
1933 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
1934 switch (run
->exit_reason
) {
1936 DPRINTF("handle_io\n");
1937 /* Called outside BQL */
1938 kvm_handle_io(run
->io
.port
, attrs
,
1939 (uint8_t *)run
+ run
->io
.data_offset
,
1946 DPRINTF("handle_mmio\n");
1947 /* Called outside BQL */
1948 address_space_rw(&address_space_memory
,
1949 run
->mmio
.phys_addr
, attrs
,
1952 run
->mmio
.is_write
);
1955 case KVM_EXIT_IRQ_WINDOW_OPEN
:
1956 DPRINTF("irq_window_open\n");
1957 ret
= EXCP_INTERRUPT
;
1959 case KVM_EXIT_SHUTDOWN
:
1960 DPRINTF("shutdown\n");
1961 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
1962 ret
= EXCP_INTERRUPT
;
1964 case KVM_EXIT_UNKNOWN
:
1965 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
1966 (uint64_t)run
->hw
.hardware_exit_reason
);
1969 case KVM_EXIT_INTERNAL_ERROR
:
1970 ret
= kvm_handle_internal_error(cpu
, run
);
1972 case KVM_EXIT_SYSTEM_EVENT
:
1973 switch (run
->system_event
.type
) {
1974 case KVM_SYSTEM_EVENT_SHUTDOWN
:
1975 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN
);
1976 ret
= EXCP_INTERRUPT
;
1978 case KVM_SYSTEM_EVENT_RESET
:
1979 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
1980 ret
= EXCP_INTERRUPT
;
1982 case KVM_SYSTEM_EVENT_CRASH
:
1983 kvm_cpu_synchronize_state(cpu
);
1984 qemu_mutex_lock_iothread();
1985 qemu_system_guest_panicked(cpu_get_crash_info(cpu
));
1986 qemu_mutex_unlock_iothread();
1990 DPRINTF("kvm_arch_handle_exit\n");
1991 ret
= kvm_arch_handle_exit(cpu
, run
);
1996 DPRINTF("kvm_arch_handle_exit\n");
1997 ret
= kvm_arch_handle_exit(cpu
, run
);
2003 qemu_mutex_lock_iothread();
2006 cpu_dump_state(cpu
, stderr
, fprintf
, CPU_DUMP_CODE
);
2007 vm_stop(RUN_STATE_INTERNAL_ERROR
);
2010 atomic_set(&cpu
->exit_request
, 0);
2014 int kvm_ioctl(KVMState
*s
, int type
, ...)
2021 arg
= va_arg(ap
, void *);
2024 trace_kvm_ioctl(type
, arg
);
2025 ret
= ioctl(s
->fd
, type
, arg
);
2032 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
2039 arg
= va_arg(ap
, void *);
2042 trace_kvm_vm_ioctl(type
, arg
);
2043 ret
= ioctl(s
->vmfd
, type
, arg
);
2050 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
2057 arg
= va_arg(ap
, void *);
2060 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
2061 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
2068 int kvm_device_ioctl(int fd
, int type
, ...)
2075 arg
= va_arg(ap
, void *);
2078 trace_kvm_device_ioctl(fd
, type
, arg
);
2079 ret
= ioctl(fd
, type
, arg
);
2086 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
2089 struct kvm_device_attr attribute
= {
2094 if (!kvm_vm_attributes_allowed
) {
2098 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
2099 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2103 int kvm_device_check_attr(int dev_fd
, uint32_t group
, uint64_t attr
)
2105 struct kvm_device_attr attribute
= {
2111 return kvm_device_ioctl(dev_fd
, KVM_HAS_DEVICE_ATTR
, &attribute
) ? 0 : 1;
2114 int kvm_device_access(int fd
, int group
, uint64_t attr
,
2115 void *val
, bool write
, Error
**errp
)
2117 struct kvm_device_attr kvmattr
;
2121 kvmattr
.group
= group
;
2122 kvmattr
.attr
= attr
;
2123 kvmattr
.addr
= (uintptr_t)val
;
2125 err
= kvm_device_ioctl(fd
,
2126 write
? KVM_SET_DEVICE_ATTR
: KVM_GET_DEVICE_ATTR
,
2129 error_setg_errno(errp
, -err
,
2130 "KVM_%s_DEVICE_ATTR failed: Group %d "
2131 "attr 0x%016" PRIx64
,
2132 write
? "SET" : "GET", group
, attr
);
2137 bool kvm_has_sync_mmu(void)
2139 return kvm_state
->sync_mmu
;
2142 int kvm_has_vcpu_events(void)
2144 return kvm_state
->vcpu_events
;
2147 int kvm_has_robust_singlestep(void)
2149 return kvm_state
->robust_singlestep
;
2152 int kvm_has_debugregs(void)
2154 return kvm_state
->debugregs
;
2157 int kvm_has_many_ioeventfds(void)
2159 if (!kvm_enabled()) {
2162 return kvm_state
->many_ioeventfds
;
2165 int kvm_has_gsi_routing(void)
2167 #ifdef KVM_CAP_IRQ_ROUTING
2168 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
2174 int kvm_has_intx_set_mask(void)
2176 return kvm_state
->intx_set_mask
;
2179 bool kvm_arm_supports_user_irq(void)
2181 return kvm_check_extension(kvm_state
, KVM_CAP_ARM_USER_IRQ
);
2184 #ifdef KVM_CAP_SET_GUEST_DEBUG
2185 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
2188 struct kvm_sw_breakpoint
*bp
;
2190 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
2198 int kvm_sw_breakpoints_active(CPUState
*cpu
)
2200 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
2203 struct kvm_set_guest_debug_data
{
2204 struct kvm_guest_debug dbg
;
2208 static void kvm_invoke_set_guest_debug(CPUState
*cpu
, run_on_cpu_data data
)
2210 struct kvm_set_guest_debug_data
*dbg_data
=
2211 (struct kvm_set_guest_debug_data
*) data
.host_ptr
;
2213 dbg_data
->err
= kvm_vcpu_ioctl(cpu
, KVM_SET_GUEST_DEBUG
,
2217 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2219 struct kvm_set_guest_debug_data data
;
2221 data
.dbg
.control
= reinject_trap
;
2223 if (cpu
->singlestep_enabled
) {
2224 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
2226 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
2228 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
,
2229 RUN_ON_CPU_HOST_PTR(&data
));
2233 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2234 target_ulong len
, int type
)
2236 struct kvm_sw_breakpoint
*bp
;
2239 if (type
== GDB_BREAKPOINT_SW
) {
2240 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2246 bp
= g_malloc(sizeof(struct kvm_sw_breakpoint
));
2249 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
2255 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2257 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
2264 err
= kvm_update_guest_debug(cpu
, 0);
2272 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2273 target_ulong len
, int type
)
2275 struct kvm_sw_breakpoint
*bp
;
2278 if (type
== GDB_BREAKPOINT_SW
) {
2279 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2284 if (bp
->use_count
> 1) {
2289 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
2294 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2297 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
2304 err
= kvm_update_guest_debug(cpu
, 0);
2312 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2314 struct kvm_sw_breakpoint
*bp
, *next
;
2315 KVMState
*s
= cpu
->kvm_state
;
2318 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
2319 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
2320 /* Try harder to find a CPU that currently sees the breakpoint. */
2321 CPU_FOREACH(tmpcpu
) {
2322 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
2327 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
2330 kvm_arch_remove_all_hw_breakpoints();
2333 kvm_update_guest_debug(cpu
, 0);
2337 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2339 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2344 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2345 target_ulong len
, int type
)
2350 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2351 target_ulong len
, int type
)
2356 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2359 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2361 static int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
2363 KVMState
*s
= kvm_state
;
2364 struct kvm_signal_mask
*sigmask
;
2367 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
2369 sigmask
->len
= s
->sigmask_len
;
2370 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
2371 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
2377 static void kvm_ipi_signal(int sig
)
2380 assert(kvm_immediate_exit
);
2381 kvm_cpu_kick(current_cpu
);
2385 void kvm_init_cpu_signals(CPUState
*cpu
)
2389 struct sigaction sigact
;
2391 memset(&sigact
, 0, sizeof(sigact
));
2392 sigact
.sa_handler
= kvm_ipi_signal
;
2393 sigaction(SIG_IPI
, &sigact
, NULL
);
2395 pthread_sigmask(SIG_BLOCK
, NULL
, &set
);
2396 #if defined KVM_HAVE_MCE_INJECTION
2397 sigdelset(&set
, SIGBUS
);
2398 pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
2400 sigdelset(&set
, SIG_IPI
);
2401 if (kvm_immediate_exit
) {
2402 r
= pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
2404 r
= kvm_set_signal_mask(cpu
, &set
);
2407 fprintf(stderr
, "kvm_set_signal_mask: %s\n", strerror(-r
));
2412 /* Called asynchronously in VCPU thread. */
2413 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
2415 #ifdef KVM_HAVE_MCE_INJECTION
2416 if (have_sigbus_pending
) {
2419 have_sigbus_pending
= true;
2420 pending_sigbus_addr
= addr
;
2421 pending_sigbus_code
= code
;
2422 atomic_set(&cpu
->exit_request
, 1);
2429 /* Called synchronously (via signalfd) in main thread. */
2430 int kvm_on_sigbus(int code
, void *addr
)
2432 #ifdef KVM_HAVE_MCE_INJECTION
2433 /* Action required MCE kills the process if SIGBUS is blocked. Because
2434 * that's what happens in the I/O thread, where we handle MCE via signalfd,
2435 * we can only get action optional here.
2437 assert(code
!= BUS_MCEERR_AR
);
2438 kvm_arch_on_sigbus_vcpu(first_cpu
, code
, addr
);
2445 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
2448 struct kvm_create_device create_dev
;
2450 create_dev
.type
= type
;
2452 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
2454 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
2458 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
2463 return test
? 0 : create_dev
.fd
;
2466 bool kvm_device_supported(int vmfd
, uint64_t type
)
2468 struct kvm_create_device create_dev
= {
2471 .flags
= KVM_CREATE_DEVICE_TEST
,
2474 if (ioctl(vmfd
, KVM_CHECK_EXTENSION
, KVM_CAP_DEVICE_CTRL
) <= 0) {
2478 return (ioctl(vmfd
, KVM_CREATE_DEVICE
, &create_dev
) >= 0);
2481 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
2483 struct kvm_one_reg reg
;
2487 reg
.addr
= (uintptr_t) source
;
2488 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
2490 trace_kvm_failed_reg_set(id
, strerror(-r
));
2495 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
2497 struct kvm_one_reg reg
;
2501 reg
.addr
= (uintptr_t) target
;
2502 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
2504 trace_kvm_failed_reg_get(id
, strerror(-r
));
2509 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
2511 AccelClass
*ac
= ACCEL_CLASS(oc
);
2513 ac
->init_machine
= kvm_init
;
2514 ac
->allowed
= &kvm_allowed
;
2517 static const TypeInfo kvm_accel_type
= {
2518 .name
= TYPE_KVM_ACCEL
,
2519 .parent
= TYPE_ACCEL
,
2520 .class_init
= kvm_accel_class_init
,
2521 .instance_size
= sizeof(KVMState
),
2524 static void kvm_type_init(void)
2526 type_register_static(&kvm_accel_type
);
2529 type_init(kvm_type_init
);