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"
27 #include "hw/pci/msi.h"
28 #include "hw/pci/msix.h"
29 #include "hw/s390x/adapter.h"
30 #include "exec/gdbstub.h"
31 #include "sysemu/kvm_int.h"
32 #include "qemu/bswap.h"
33 #include "exec/memory.h"
34 #include "exec/ram_addr.h"
35 #include "exec/address-spaces.h"
36 #include "qemu/event_notifier.h"
40 #include "hw/boards.h"
42 /* This check must be after config-host.h is included */
44 #include <sys/eventfd.h>
47 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
48 * need to use the real host PAGE_SIZE, as that's what KVM will use.
50 #define PAGE_SIZE getpagesize()
55 #define DPRINTF(fmt, ...) \
56 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
58 #define DPRINTF(fmt, ...) \
62 #define KVM_MSI_HASHTAB_SIZE 256
64 struct KVMParkedVcpu
{
65 unsigned long vcpu_id
;
67 QLIST_ENTRY(KVMParkedVcpu
) node
;
72 AccelState parent_obj
;
78 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
79 bool coalesced_flush_in_progress
;
80 int broken_set_mem_region
;
82 int robust_singlestep
;
84 #ifdef KVM_CAP_SET_GUEST_DEBUG
85 struct kvm_sw_breakpoint_head kvm_sw_breakpoints
;
89 /* The man page (and posix) say ioctl numbers are signed int, but
90 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
91 * unsigned, and treating them as signed here can break things */
92 unsigned irq_set_ioctl
;
93 unsigned int sigmask_len
;
95 #ifdef KVM_CAP_IRQ_ROUTING
96 struct kvm_irq_routing
*irq_routes
;
97 int nr_allocated_irq_routes
;
98 unsigned long *used_gsi_bitmap
;
99 unsigned int gsi_count
;
100 QTAILQ_HEAD(msi_hashtab
, KVMMSIRoute
) msi_hashtab
[KVM_MSI_HASHTAB_SIZE
];
102 KVMMemoryListener memory_listener
;
103 QLIST_HEAD(, KVMParkedVcpu
) kvm_parked_vcpus
;
107 bool kvm_kernel_irqchip
;
108 bool kvm_split_irqchip
;
109 bool kvm_async_interrupts_allowed
;
110 bool kvm_halt_in_kernel_allowed
;
111 bool kvm_eventfds_allowed
;
112 bool kvm_irqfds_allowed
;
113 bool kvm_resamplefds_allowed
;
114 bool kvm_msi_via_irqfd_allowed
;
115 bool kvm_gsi_routing_allowed
;
116 bool kvm_gsi_direct_mapping
;
118 bool kvm_readonly_mem_allowed
;
119 bool kvm_vm_attributes_allowed
;
120 bool kvm_direct_msi_allowed
;
121 bool kvm_ioeventfd_any_length_allowed
;
123 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
124 KVM_CAP_INFO(USER_MEMORY
),
125 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
129 int kvm_get_max_memslots(void)
131 KVMState
*s
= KVM_STATE(current_machine
->accelerator
);
136 static KVMSlot
*kvm_get_free_slot(KVMMemoryListener
*kml
)
138 KVMState
*s
= kvm_state
;
141 for (i
= 0; i
< s
->nr_slots
; i
++) {
142 if (kml
->slots
[i
].memory_size
== 0) {
143 return &kml
->slots
[i
];
150 bool kvm_has_free_slot(MachineState
*ms
)
152 KVMState
*s
= KVM_STATE(ms
->accelerator
);
154 return kvm_get_free_slot(&s
->memory_listener
);
157 static KVMSlot
*kvm_alloc_slot(KVMMemoryListener
*kml
)
159 KVMSlot
*slot
= kvm_get_free_slot(kml
);
165 fprintf(stderr
, "%s: no free slot available\n", __func__
);
169 static KVMSlot
*kvm_lookup_matching_slot(KVMMemoryListener
*kml
,
173 KVMState
*s
= kvm_state
;
176 for (i
= 0; i
< s
->nr_slots
; i
++) {
177 KVMSlot
*mem
= &kml
->slots
[i
];
179 if (start_addr
== mem
->start_addr
&&
180 end_addr
== mem
->start_addr
+ mem
->memory_size
) {
189 * Find overlapping slot with lowest start address
191 static KVMSlot
*kvm_lookup_overlapping_slot(KVMMemoryListener
*kml
,
195 KVMState
*s
= kvm_state
;
196 KVMSlot
*found
= NULL
;
199 for (i
= 0; i
< s
->nr_slots
; i
++) {
200 KVMSlot
*mem
= &kml
->slots
[i
];
202 if (mem
->memory_size
== 0 ||
203 (found
&& found
->start_addr
< mem
->start_addr
)) {
207 if (end_addr
> mem
->start_addr
&&
208 start_addr
< mem
->start_addr
+ mem
->memory_size
) {
216 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
219 KVMMemoryListener
*kml
= &s
->memory_listener
;
222 for (i
= 0; i
< s
->nr_slots
; i
++) {
223 KVMSlot
*mem
= &kml
->slots
[i
];
225 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
226 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
234 static int kvm_set_user_memory_region(KVMMemoryListener
*kml
, KVMSlot
*slot
)
236 KVMState
*s
= kvm_state
;
237 struct kvm_userspace_memory_region mem
;
239 mem
.slot
= slot
->slot
| (kml
->as_id
<< 16);
240 mem
.guest_phys_addr
= slot
->start_addr
;
241 mem
.userspace_addr
= (unsigned long)slot
->ram
;
242 mem
.flags
= slot
->flags
;
244 if (slot
->memory_size
&& mem
.flags
& KVM_MEM_READONLY
) {
245 /* Set the slot size to 0 before setting the slot to the desired
246 * value. This is needed based on KVM commit 75d61fbc. */
248 kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
250 mem
.memory_size
= slot
->memory_size
;
251 return kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
254 int kvm_destroy_vcpu(CPUState
*cpu
)
256 KVMState
*s
= kvm_state
;
258 struct KVMParkedVcpu
*vcpu
= NULL
;
261 DPRINTF("kvm_destroy_vcpu\n");
263 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
266 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
270 ret
= munmap(cpu
->kvm_run
, mmap_size
);
275 vcpu
= g_malloc0(sizeof(*vcpu
));
276 vcpu
->vcpu_id
= kvm_arch_vcpu_id(cpu
);
277 vcpu
->kvm_fd
= cpu
->kvm_fd
;
278 QLIST_INSERT_HEAD(&kvm_state
->kvm_parked_vcpus
, vcpu
, node
);
283 static int kvm_get_vcpu(KVMState
*s
, unsigned long vcpu_id
)
285 struct KVMParkedVcpu
*cpu
;
287 QLIST_FOREACH(cpu
, &s
->kvm_parked_vcpus
, node
) {
288 if (cpu
->vcpu_id
== vcpu_id
) {
291 QLIST_REMOVE(cpu
, node
);
292 kvm_fd
= cpu
->kvm_fd
;
298 return kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)vcpu_id
);
301 int kvm_init_vcpu(CPUState
*cpu
)
303 KVMState
*s
= kvm_state
;
307 DPRINTF("kvm_init_vcpu\n");
309 ret
= kvm_get_vcpu(s
, kvm_arch_vcpu_id(cpu
));
311 DPRINTF("kvm_create_vcpu failed\n");
317 cpu
->kvm_vcpu_dirty
= true;
319 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
322 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
326 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
328 if (cpu
->kvm_run
== MAP_FAILED
) {
330 DPRINTF("mmap'ing vcpu state failed\n");
334 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
335 s
->coalesced_mmio_ring
=
336 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
339 ret
= kvm_arch_init_vcpu(cpu
);
345 * dirty pages logging control
348 static int kvm_mem_flags(MemoryRegion
*mr
)
350 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
353 if (memory_region_get_dirty_log_mask(mr
) != 0) {
354 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
356 if (readonly
&& kvm_readonly_mem_allowed
) {
357 flags
|= KVM_MEM_READONLY
;
362 static int kvm_slot_update_flags(KVMMemoryListener
*kml
, KVMSlot
*mem
,
367 old_flags
= mem
->flags
;
368 mem
->flags
= kvm_mem_flags(mr
);
370 /* If nothing changed effectively, no need to issue ioctl */
371 if (mem
->flags
== old_flags
) {
375 return kvm_set_user_memory_region(kml
, mem
);
378 static int kvm_section_update_flags(KVMMemoryListener
*kml
,
379 MemoryRegionSection
*section
)
381 hwaddr phys_addr
= section
->offset_within_address_space
;
382 ram_addr_t size
= int128_get64(section
->size
);
383 KVMSlot
*mem
= kvm_lookup_matching_slot(kml
, phys_addr
, phys_addr
+ size
);
388 return kvm_slot_update_flags(kml
, mem
, section
->mr
);
392 static void kvm_log_start(MemoryListener
*listener
,
393 MemoryRegionSection
*section
,
396 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
403 r
= kvm_section_update_flags(kml
, section
);
409 static void kvm_log_stop(MemoryListener
*listener
,
410 MemoryRegionSection
*section
,
413 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
420 r
= kvm_section_update_flags(kml
, section
);
426 /* get kvm's dirty pages bitmap and update qemu's */
427 static int kvm_get_dirty_pages_log_range(MemoryRegionSection
*section
,
428 unsigned long *bitmap
)
430 ram_addr_t start
= section
->offset_within_region
+
431 memory_region_get_ram_addr(section
->mr
);
432 ram_addr_t pages
= int128_get64(section
->size
) / getpagesize();
434 cpu_physical_memory_set_dirty_lebitmap(bitmap
, start
, pages
);
438 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
441 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
442 * This function updates qemu's dirty bitmap using
443 * memory_region_set_dirty(). This means all bits are set
446 * @start_add: start of logged region.
447 * @end_addr: end of logged region.
449 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener
*kml
,
450 MemoryRegionSection
*section
)
452 KVMState
*s
= kvm_state
;
453 unsigned long size
, allocated_size
= 0;
454 struct kvm_dirty_log d
= {};
457 hwaddr start_addr
= section
->offset_within_address_space
;
458 hwaddr end_addr
= start_addr
+ int128_get64(section
->size
);
460 d
.dirty_bitmap
= NULL
;
461 while (start_addr
< end_addr
) {
462 mem
= kvm_lookup_overlapping_slot(kml
, start_addr
, end_addr
);
467 /* XXX bad kernel interface alert
468 * For dirty bitmap, kernel allocates array of size aligned to
469 * bits-per-long. But for case when the kernel is 64bits and
470 * the userspace is 32bits, userspace can't align to the same
471 * bits-per-long, since sizeof(long) is different between kernel
472 * and user space. This way, userspace will provide buffer which
473 * may be 4 bytes less than the kernel will use, resulting in
474 * userspace memory corruption (which is not detectable by valgrind
475 * too, in most cases).
476 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
477 * a hope that sizeof(long) won't become >8 any time soon.
479 size
= ALIGN(((mem
->memory_size
) >> TARGET_PAGE_BITS
),
480 /*HOST_LONG_BITS*/ 64) / 8;
481 if (!d
.dirty_bitmap
) {
482 d
.dirty_bitmap
= g_malloc(size
);
483 } else if (size
> allocated_size
) {
484 d
.dirty_bitmap
= g_realloc(d
.dirty_bitmap
, size
);
486 allocated_size
= size
;
487 memset(d
.dirty_bitmap
, 0, allocated_size
);
489 d
.slot
= mem
->slot
| (kml
->as_id
<< 16);
490 if (kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
) == -1) {
491 DPRINTF("ioctl failed %d\n", errno
);
496 kvm_get_dirty_pages_log_range(section
, d
.dirty_bitmap
);
497 start_addr
= mem
->start_addr
+ mem
->memory_size
;
499 g_free(d
.dirty_bitmap
);
504 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
505 MemoryRegionSection
*secion
,
506 hwaddr start
, hwaddr size
)
508 KVMState
*s
= kvm_state
;
510 if (s
->coalesced_mmio
) {
511 struct kvm_coalesced_mmio_zone zone
;
517 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
521 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
522 MemoryRegionSection
*secion
,
523 hwaddr start
, hwaddr size
)
525 KVMState
*s
= kvm_state
;
527 if (s
->coalesced_mmio
) {
528 struct kvm_coalesced_mmio_zone zone
;
534 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
538 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
542 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
550 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
554 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
556 /* VM wide version not implemented, use global one instead */
557 ret
= kvm_check_extension(s
, extension
);
563 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
565 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
566 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
567 * endianness, but the memory core hands them in target endianness.
568 * For example, PPC is always treated as big-endian even if running
569 * on KVM and on PPC64LE. Correct here.
583 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
584 bool assign
, uint32_t size
, bool datamatch
)
587 struct kvm_ioeventfd iofd
= {
588 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
595 if (!kvm_enabled()) {
600 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
603 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
606 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
615 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
616 bool assign
, uint32_t size
, bool datamatch
)
618 struct kvm_ioeventfd kick
= {
619 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
621 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
626 if (!kvm_enabled()) {
630 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
633 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
635 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
643 static int kvm_check_many_ioeventfds(void)
645 /* Userspace can use ioeventfd for io notification. This requires a host
646 * that supports eventfd(2) and an I/O thread; since eventfd does not
647 * support SIGIO it cannot interrupt the vcpu.
649 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
650 * can avoid creating too many ioeventfds.
652 #if defined(CONFIG_EVENTFD)
655 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
656 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
657 if (ioeventfds
[i
] < 0) {
660 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
662 close(ioeventfds
[i
]);
667 /* Decide whether many devices are supported or not */
668 ret
= i
== ARRAY_SIZE(ioeventfds
);
671 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
672 close(ioeventfds
[i
]);
680 static const KVMCapabilityInfo
*
681 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
684 if (!kvm_check_extension(s
, list
->value
)) {
692 static void kvm_set_phys_mem(KVMMemoryListener
*kml
,
693 MemoryRegionSection
*section
, bool add
)
695 KVMState
*s
= kvm_state
;
698 MemoryRegion
*mr
= section
->mr
;
699 bool writeable
= !mr
->readonly
&& !mr
->rom_device
;
700 hwaddr start_addr
= section
->offset_within_address_space
;
701 ram_addr_t size
= int128_get64(section
->size
);
705 /* kvm works in page size chunks, but the function may be called
706 with sub-page size and unaligned start address. Pad the start
707 address to next and truncate size to previous page boundary. */
708 delta
= qemu_real_host_page_size
- (start_addr
& ~qemu_real_host_page_mask
);
709 delta
&= ~qemu_real_host_page_mask
;
715 size
&= qemu_real_host_page_mask
;
716 if (!size
|| (start_addr
& ~qemu_real_host_page_mask
)) {
720 if (!memory_region_is_ram(mr
)) {
721 if (writeable
|| !kvm_readonly_mem_allowed
) {
723 } else if (!mr
->romd_mode
) {
724 /* If the memory device is not in romd_mode, then we actually want
725 * to remove the kvm memory slot so all accesses will trap. */
730 ram
= memory_region_get_ram_ptr(mr
) + section
->offset_within_region
+ delta
;
733 mem
= kvm_lookup_overlapping_slot(kml
, start_addr
, start_addr
+ size
);
738 if (add
&& start_addr
>= mem
->start_addr
&&
739 (start_addr
+ size
<= mem
->start_addr
+ mem
->memory_size
) &&
740 (ram
- start_addr
== mem
->ram
- mem
->start_addr
)) {
741 /* The new slot fits into the existing one and comes with
742 * identical parameters - update flags and done. */
743 kvm_slot_update_flags(kml
, mem
, mr
);
749 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
750 kvm_physical_sync_dirty_bitmap(kml
, section
);
753 /* unregister the overlapping slot */
754 mem
->memory_size
= 0;
755 err
= kvm_set_user_memory_region(kml
, mem
);
757 fprintf(stderr
, "%s: error unregistering overlapping slot: %s\n",
758 __func__
, strerror(-err
));
762 /* Workaround for older KVM versions: we can't join slots, even not by
763 * unregistering the previous ones and then registering the larger
764 * slot. We have to maintain the existing fragmentation. Sigh.
766 * This workaround assumes that the new slot starts at the same
767 * address as the first existing one. If not or if some overlapping
768 * slot comes around later, we will fail (not seen in practice so far)
769 * - and actually require a recent KVM version. */
770 if (s
->broken_set_mem_region
&&
771 old
.start_addr
== start_addr
&& old
.memory_size
< size
&& add
) {
772 mem
= kvm_alloc_slot(kml
);
773 mem
->memory_size
= old
.memory_size
;
774 mem
->start_addr
= old
.start_addr
;
776 mem
->flags
= kvm_mem_flags(mr
);
778 err
= kvm_set_user_memory_region(kml
, mem
);
780 fprintf(stderr
, "%s: error updating slot: %s\n", __func__
,
785 start_addr
+= old
.memory_size
;
786 ram
+= old
.memory_size
;
787 size
-= old
.memory_size
;
791 /* register prefix slot */
792 if (old
.start_addr
< start_addr
) {
793 mem
= kvm_alloc_slot(kml
);
794 mem
->memory_size
= start_addr
- old
.start_addr
;
795 mem
->start_addr
= old
.start_addr
;
797 mem
->flags
= kvm_mem_flags(mr
);
799 err
= kvm_set_user_memory_region(kml
, mem
);
801 fprintf(stderr
, "%s: error registering prefix slot: %s\n",
802 __func__
, strerror(-err
));
804 fprintf(stderr
, "%s: This is probably because your kernel's " \
805 "PAGE_SIZE is too big. Please try to use 4k " \
806 "PAGE_SIZE!\n", __func__
);
812 /* register suffix slot */
813 if (old
.start_addr
+ old
.memory_size
> start_addr
+ size
) {
814 ram_addr_t size_delta
;
816 mem
= kvm_alloc_slot(kml
);
817 mem
->start_addr
= start_addr
+ size
;
818 size_delta
= mem
->start_addr
- old
.start_addr
;
819 mem
->memory_size
= old
.memory_size
- size_delta
;
820 mem
->ram
= old
.ram
+ size_delta
;
821 mem
->flags
= kvm_mem_flags(mr
);
823 err
= kvm_set_user_memory_region(kml
, mem
);
825 fprintf(stderr
, "%s: error registering suffix slot: %s\n",
826 __func__
, strerror(-err
));
832 /* in case the KVM bug workaround already "consumed" the new slot */
839 mem
= kvm_alloc_slot(kml
);
840 mem
->memory_size
= size
;
841 mem
->start_addr
= start_addr
;
843 mem
->flags
= kvm_mem_flags(mr
);
845 err
= kvm_set_user_memory_region(kml
, mem
);
847 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
853 static void kvm_region_add(MemoryListener
*listener
,
854 MemoryRegionSection
*section
)
856 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
858 memory_region_ref(section
->mr
);
859 kvm_set_phys_mem(kml
, section
, true);
862 static void kvm_region_del(MemoryListener
*listener
,
863 MemoryRegionSection
*section
)
865 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
867 kvm_set_phys_mem(kml
, section
, false);
868 memory_region_unref(section
->mr
);
871 static void kvm_log_sync(MemoryListener
*listener
,
872 MemoryRegionSection
*section
)
874 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
877 r
= kvm_physical_sync_dirty_bitmap(kml
, section
);
883 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
884 MemoryRegionSection
*section
,
885 bool match_data
, uint64_t data
,
888 int fd
= event_notifier_get_fd(e
);
891 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
892 data
, true, int128_get64(section
->size
),
895 fprintf(stderr
, "%s: error adding ioeventfd: %s\n",
896 __func__
, strerror(-r
));
901 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
902 MemoryRegionSection
*section
,
903 bool match_data
, uint64_t data
,
906 int fd
= event_notifier_get_fd(e
);
909 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
910 data
, false, int128_get64(section
->size
),
917 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
918 MemoryRegionSection
*section
,
919 bool match_data
, uint64_t data
,
922 int fd
= event_notifier_get_fd(e
);
925 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
926 data
, true, int128_get64(section
->size
),
929 fprintf(stderr
, "%s: error adding ioeventfd: %s\n",
930 __func__
, strerror(-r
));
935 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
936 MemoryRegionSection
*section
,
937 bool match_data
, uint64_t data
,
941 int fd
= event_notifier_get_fd(e
);
944 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
945 data
, false, int128_get64(section
->size
),
952 void kvm_memory_listener_register(KVMState
*s
, KVMMemoryListener
*kml
,
953 AddressSpace
*as
, int as_id
)
957 kml
->slots
= g_malloc0(s
->nr_slots
* sizeof(KVMSlot
));
960 for (i
= 0; i
< s
->nr_slots
; i
++) {
961 kml
->slots
[i
].slot
= i
;
964 kml
->listener
.region_add
= kvm_region_add
;
965 kml
->listener
.region_del
= kvm_region_del
;
966 kml
->listener
.log_start
= kvm_log_start
;
967 kml
->listener
.log_stop
= kvm_log_stop
;
968 kml
->listener
.log_sync
= kvm_log_sync
;
969 kml
->listener
.priority
= 10;
971 memory_listener_register(&kml
->listener
, as
);
974 static MemoryListener kvm_io_listener
= {
975 .eventfd_add
= kvm_io_ioeventfd_add
,
976 .eventfd_del
= kvm_io_ioeventfd_del
,
980 static void kvm_handle_interrupt(CPUState
*cpu
, int mask
)
982 cpu
->interrupt_request
|= mask
;
984 if (!qemu_cpu_is_self(cpu
)) {
989 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
991 struct kvm_irq_level event
;
994 assert(kvm_async_interrupts_enabled());
998 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
1000 perror("kvm_set_irq");
1004 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
1007 #ifdef KVM_CAP_IRQ_ROUTING
1008 typedef struct KVMMSIRoute
{
1009 struct kvm_irq_routing_entry kroute
;
1010 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
1013 static void set_gsi(KVMState
*s
, unsigned int gsi
)
1015 set_bit(gsi
, s
->used_gsi_bitmap
);
1018 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
1020 clear_bit(gsi
, s
->used_gsi_bitmap
);
1023 void kvm_init_irq_routing(KVMState
*s
)
1027 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
1028 if (gsi_count
> 0) {
1029 /* Round up so we can search ints using ffs */
1030 s
->used_gsi_bitmap
= bitmap_new(gsi_count
);
1031 s
->gsi_count
= gsi_count
;
1034 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
1035 s
->nr_allocated_irq_routes
= 0;
1037 if (!kvm_direct_msi_allowed
) {
1038 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
1039 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
1043 kvm_arch_init_irq_routing(s
);
1046 void kvm_irqchip_commit_routes(KVMState
*s
)
1050 if (kvm_gsi_direct_mapping()) {
1054 if (!kvm_gsi_routing_enabled()) {
1058 s
->irq_routes
->flags
= 0;
1059 trace_kvm_irqchip_commit_routes();
1060 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
1064 static void kvm_add_routing_entry(KVMState
*s
,
1065 struct kvm_irq_routing_entry
*entry
)
1067 struct kvm_irq_routing_entry
*new;
1070 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
1071 n
= s
->nr_allocated_irq_routes
* 2;
1075 size
= sizeof(struct kvm_irq_routing
);
1076 size
+= n
* sizeof(*new);
1077 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
1078 s
->nr_allocated_irq_routes
= n
;
1080 n
= s
->irq_routes
->nr
++;
1081 new = &s
->irq_routes
->entries
[n
];
1085 set_gsi(s
, entry
->gsi
);
1088 static int kvm_update_routing_entry(KVMState
*s
,
1089 struct kvm_irq_routing_entry
*new_entry
)
1091 struct kvm_irq_routing_entry
*entry
;
1094 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
1095 entry
= &s
->irq_routes
->entries
[n
];
1096 if (entry
->gsi
!= new_entry
->gsi
) {
1100 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1104 *entry
= *new_entry
;
1112 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1114 struct kvm_irq_routing_entry e
= {};
1116 assert(pin
< s
->gsi_count
);
1119 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1121 e
.u
.irqchip
.irqchip
= irqchip
;
1122 e
.u
.irqchip
.pin
= pin
;
1123 kvm_add_routing_entry(s
, &e
);
1126 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1128 struct kvm_irq_routing_entry
*e
;
1131 if (kvm_gsi_direct_mapping()) {
1135 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1136 e
= &s
->irq_routes
->entries
[i
];
1137 if (e
->gsi
== virq
) {
1138 s
->irq_routes
->nr
--;
1139 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1143 kvm_arch_release_virq_post(virq
);
1146 static unsigned int kvm_hash_msi(uint32_t data
)
1148 /* This is optimized for IA32 MSI layout. However, no other arch shall
1149 * repeat the mistake of not providing a direct MSI injection API. */
1153 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1155 KVMMSIRoute
*route
, *next
;
1158 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1159 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1160 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1161 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1167 static int kvm_irqchip_get_virq(KVMState
*s
)
1172 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1173 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1174 * number can succeed even though a new route entry cannot be added.
1175 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1177 if (!kvm_direct_msi_allowed
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1178 kvm_flush_dynamic_msi_routes(s
);
1181 /* Return the lowest unused GSI in the bitmap */
1182 next_virq
= find_first_zero_bit(s
->used_gsi_bitmap
, s
->gsi_count
);
1183 if (next_virq
>= s
->gsi_count
) {
1190 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1192 unsigned int hash
= kvm_hash_msi(msg
.data
);
1195 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1196 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1197 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1198 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1205 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1210 if (kvm_direct_msi_allowed
) {
1211 msi
.address_lo
= (uint32_t)msg
.address
;
1212 msi
.address_hi
= msg
.address
>> 32;
1213 msi
.data
= le32_to_cpu(msg
.data
);
1215 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1217 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1220 route
= kvm_lookup_msi_route(s
, msg
);
1224 virq
= kvm_irqchip_get_virq(s
);
1229 route
= g_malloc0(sizeof(KVMMSIRoute
));
1230 route
->kroute
.gsi
= virq
;
1231 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1232 route
->kroute
.flags
= 0;
1233 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1234 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1235 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1237 kvm_add_routing_entry(s
, &route
->kroute
);
1238 kvm_irqchip_commit_routes(s
);
1240 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1244 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1246 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
1249 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1251 struct kvm_irq_routing_entry kroute
= {};
1253 MSIMessage msg
= {0, 0};
1256 msg
= pci_get_msi_message(dev
, vector
);
1259 if (kvm_gsi_direct_mapping()) {
1260 return kvm_arch_msi_data_to_gsi(msg
.data
);
1263 if (!kvm_gsi_routing_enabled()) {
1267 virq
= kvm_irqchip_get_virq(s
);
1273 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1275 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1276 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1277 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1278 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1279 kvm_irqchip_release_virq(s
, virq
);
1283 trace_kvm_irqchip_add_msi_route(virq
);
1285 kvm_add_routing_entry(s
, &kroute
);
1286 kvm_arch_add_msi_route_post(&kroute
, vector
, dev
);
1287 kvm_irqchip_commit_routes(s
);
1292 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
,
1295 struct kvm_irq_routing_entry kroute
= {};
1297 if (kvm_gsi_direct_mapping()) {
1301 if (!kvm_irqchip_in_kernel()) {
1306 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1308 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1309 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1310 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1311 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1315 trace_kvm_irqchip_update_msi_route(virq
);
1317 return kvm_update_routing_entry(s
, &kroute
);
1320 static int kvm_irqchip_assign_irqfd(KVMState
*s
, int fd
, int rfd
, int virq
,
1323 struct kvm_irqfd irqfd
= {
1326 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
1330 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
1331 irqfd
.resamplefd
= rfd
;
1334 if (!kvm_irqfds_enabled()) {
1338 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
1341 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1343 struct kvm_irq_routing_entry kroute
= {};
1346 if (!kvm_gsi_routing_enabled()) {
1350 virq
= kvm_irqchip_get_virq(s
);
1356 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
1358 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
1359 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
1360 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
1361 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
1362 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
1364 kvm_add_routing_entry(s
, &kroute
);
1369 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1371 struct kvm_irq_routing_entry kroute
= {};
1374 if (!kvm_gsi_routing_enabled()) {
1377 if (!kvm_check_extension(s
, KVM_CAP_HYPERV_SYNIC
)) {
1380 virq
= kvm_irqchip_get_virq(s
);
1386 kroute
.type
= KVM_IRQ_ROUTING_HV_SINT
;
1388 kroute
.u
.hv_sint
.vcpu
= vcpu
;
1389 kroute
.u
.hv_sint
.sint
= sint
;
1391 kvm_add_routing_entry(s
, &kroute
);
1392 kvm_irqchip_commit_routes(s
);
1397 #else /* !KVM_CAP_IRQ_ROUTING */
1399 void kvm_init_irq_routing(KVMState
*s
)
1403 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1407 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1412 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1417 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1422 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1427 static int kvm_irqchip_assign_irqfd(KVMState
*s
, int fd
, int virq
, bool assign
)
1432 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
1436 #endif /* !KVM_CAP_IRQ_ROUTING */
1438 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1439 EventNotifier
*rn
, int virq
)
1441 return kvm_irqchip_assign_irqfd(s
, event_notifier_get_fd(n
),
1442 rn
? event_notifier_get_fd(rn
) : -1, virq
, true);
1445 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1448 return kvm_irqchip_assign_irqfd(s
, event_notifier_get_fd(n
), -1, virq
,
1452 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1453 EventNotifier
*rn
, qemu_irq irq
)
1456 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1461 return kvm_irqchip_add_irqfd_notifier_gsi(s
, n
, rn
, GPOINTER_TO_INT(gsi
));
1464 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1468 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1473 return kvm_irqchip_remove_irqfd_notifier_gsi(s
, n
, GPOINTER_TO_INT(gsi
));
1476 void kvm_irqchip_set_qemuirq_gsi(KVMState
*s
, qemu_irq irq
, int gsi
)
1478 g_hash_table_insert(s
->gsimap
, irq
, GINT_TO_POINTER(gsi
));
1481 static void kvm_irqchip_create(MachineState
*machine
, KVMState
*s
)
1485 if (kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
1487 } else if (kvm_check_extension(s
, KVM_CAP_S390_IRQCHIP
)) {
1488 ret
= kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0);
1490 fprintf(stderr
, "Enable kernel irqchip failed: %s\n", strerror(-ret
));
1497 /* First probe and see if there's a arch-specific hook to create the
1498 * in-kernel irqchip for us */
1499 ret
= kvm_arch_irqchip_create(machine
, s
);
1501 if (machine_kernel_irqchip_split(machine
)) {
1502 perror("Split IRQ chip mode not supported.");
1505 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
1509 fprintf(stderr
, "Create kernel irqchip failed: %s\n", strerror(-ret
));
1513 kvm_kernel_irqchip
= true;
1514 /* If we have an in-kernel IRQ chip then we must have asynchronous
1515 * interrupt delivery (though the reverse is not necessarily true)
1517 kvm_async_interrupts_allowed
= true;
1518 kvm_halt_in_kernel_allowed
= true;
1520 kvm_init_irq_routing(s
);
1522 s
->gsimap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
1525 /* Find number of supported CPUs using the recommended
1526 * procedure from the kernel API documentation to cope with
1527 * older kernels that may be missing capabilities.
1529 static int kvm_recommended_vcpus(KVMState
*s
)
1531 int ret
= kvm_check_extension(s
, KVM_CAP_NR_VCPUS
);
1532 return (ret
) ? ret
: 4;
1535 static int kvm_max_vcpus(KVMState
*s
)
1537 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
1538 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
1541 static int kvm_max_vcpu_id(KVMState
*s
)
1543 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPU_ID
);
1544 return (ret
) ? ret
: kvm_max_vcpus(s
);
1547 bool kvm_vcpu_id_is_valid(int vcpu_id
)
1549 KVMState
*s
= KVM_STATE(current_machine
->accelerator
);
1550 return vcpu_id
>= 0 && vcpu_id
< kvm_max_vcpu_id(s
);
1553 static int kvm_init(MachineState
*ms
)
1555 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
1556 static const char upgrade_note
[] =
1557 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1558 "(see http://sourceforge.net/projects/kvm).\n";
1563 { "SMP", smp_cpus
},
1564 { "hotpluggable", max_cpus
},
1567 int soft_vcpus_limit
, hard_vcpus_limit
;
1569 const KVMCapabilityInfo
*missing_cap
;
1572 const char *kvm_type
;
1574 s
= KVM_STATE(ms
->accelerator
);
1577 * On systems where the kernel can support different base page
1578 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1579 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1580 * page size for the system though.
1582 assert(TARGET_PAGE_SIZE
<= getpagesize());
1586 #ifdef KVM_CAP_SET_GUEST_DEBUG
1587 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
1589 QLIST_INIT(&s
->kvm_parked_vcpus
);
1591 s
->fd
= qemu_open("/dev/kvm", O_RDWR
);
1593 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
1598 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
1599 if (ret
< KVM_API_VERSION
) {
1603 fprintf(stderr
, "kvm version too old\n");
1607 if (ret
> KVM_API_VERSION
) {
1609 fprintf(stderr
, "kvm version not supported\n");
1613 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
1615 /* If unspecified, use the default value */
1620 /* check the vcpu limits */
1621 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
1622 hard_vcpus_limit
= kvm_max_vcpus(s
);
1625 if (nc
->num
> soft_vcpus_limit
) {
1627 "Warning: Number of %s cpus requested (%d) exceeds "
1628 "the recommended cpus supported by KVM (%d)\n",
1629 nc
->name
, nc
->num
, soft_vcpus_limit
);
1631 if (nc
->num
> hard_vcpus_limit
) {
1632 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
1633 "the maximum cpus supported by KVM (%d)\n",
1634 nc
->name
, nc
->num
, hard_vcpus_limit
);
1641 kvm_type
= qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1643 type
= mc
->kvm_type(kvm_type
);
1644 } else if (kvm_type
) {
1646 fprintf(stderr
, "Invalid argument kvm-type=%s\n", kvm_type
);
1651 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
1652 } while (ret
== -EINTR
);
1655 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
1659 if (ret
== -EINVAL
) {
1661 "Host kernel setup problem detected. Please verify:\n");
1662 fprintf(stderr
, "- for kernels supporting the switch_amode or"
1663 " user_mode parameters, whether\n");
1665 " user space is running in primary address space\n");
1667 "- for kernels supporting the vm.allocate_pgste sysctl, "
1668 "whether it is enabled\n");
1675 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
1678 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
1682 fprintf(stderr
, "kvm does not support %s\n%s",
1683 missing_cap
->name
, upgrade_note
);
1687 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
1689 s
->broken_set_mem_region
= 1;
1690 ret
= kvm_check_extension(s
, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
);
1692 s
->broken_set_mem_region
= 0;
1695 #ifdef KVM_CAP_VCPU_EVENTS
1696 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
1699 s
->robust_singlestep
=
1700 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
1702 #ifdef KVM_CAP_DEBUGREGS
1703 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
1706 #ifdef KVM_CAP_IRQ_ROUTING
1707 kvm_direct_msi_allowed
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
1710 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
1712 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
1713 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
1714 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
1717 #ifdef KVM_CAP_READONLY_MEM
1718 kvm_readonly_mem_allowed
=
1719 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
1722 kvm_eventfds_allowed
=
1723 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
1725 kvm_irqfds_allowed
=
1726 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
1728 kvm_resamplefds_allowed
=
1729 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
1731 kvm_vm_attributes_allowed
=
1732 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
1734 kvm_ioeventfd_any_length_allowed
=
1735 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD_ANY_LENGTH
) > 0);
1737 ret
= kvm_arch_init(ms
, s
);
1742 if (machine_kernel_irqchip_allowed(ms
)) {
1743 kvm_irqchip_create(ms
, s
);
1748 if (kvm_eventfds_allowed
) {
1749 s
->memory_listener
.listener
.eventfd_add
= kvm_mem_ioeventfd_add
;
1750 s
->memory_listener
.listener
.eventfd_del
= kvm_mem_ioeventfd_del
;
1752 s
->memory_listener
.listener
.coalesced_mmio_add
= kvm_coalesce_mmio_region
;
1753 s
->memory_listener
.listener
.coalesced_mmio_del
= kvm_uncoalesce_mmio_region
;
1755 kvm_memory_listener_register(s
, &s
->memory_listener
,
1756 &address_space_memory
, 0);
1757 memory_listener_register(&kvm_io_listener
,
1760 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
1762 cpu_interrupt_handler
= kvm_handle_interrupt
;
1774 g_free(s
->memory_listener
.slots
);
1779 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
1781 s
->sigmask_len
= sigmask_len
;
1784 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
1785 int size
, uint32_t count
)
1788 uint8_t *ptr
= data
;
1790 for (i
= 0; i
< count
; i
++) {
1791 address_space_rw(&address_space_io
, port
, attrs
,
1793 direction
== KVM_EXIT_IO_OUT
);
1798 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
1800 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
1801 run
->internal
.suberror
);
1803 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
1806 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
1807 fprintf(stderr
, "extra data[%d]: %"PRIx64
"\n",
1808 i
, (uint64_t)run
->internal
.data
[i
]);
1811 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
1812 fprintf(stderr
, "emulation failure\n");
1813 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
1814 cpu_dump_state(cpu
, stderr
, fprintf
, CPU_DUMP_CODE
);
1815 return EXCP_INTERRUPT
;
1818 /* FIXME: Should trigger a qmp message to let management know
1819 * something went wrong.
1824 void kvm_flush_coalesced_mmio_buffer(void)
1826 KVMState
*s
= kvm_state
;
1828 if (s
->coalesced_flush_in_progress
) {
1832 s
->coalesced_flush_in_progress
= true;
1834 if (s
->coalesced_mmio_ring
) {
1835 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
1836 while (ring
->first
!= ring
->last
) {
1837 struct kvm_coalesced_mmio
*ent
;
1839 ent
= &ring
->coalesced_mmio
[ring
->first
];
1841 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
1843 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
1847 s
->coalesced_flush_in_progress
= false;
1850 static void do_kvm_cpu_synchronize_state(void *arg
)
1852 CPUState
*cpu
= arg
;
1854 if (!cpu
->kvm_vcpu_dirty
) {
1855 kvm_arch_get_registers(cpu
);
1856 cpu
->kvm_vcpu_dirty
= true;
1860 void kvm_cpu_synchronize_state(CPUState
*cpu
)
1862 if (!cpu
->kvm_vcpu_dirty
) {
1863 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, cpu
);
1867 static void do_kvm_cpu_synchronize_post_reset(void *arg
)
1869 CPUState
*cpu
= arg
;
1871 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
1872 cpu
->kvm_vcpu_dirty
= false;
1875 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
1877 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, cpu
);
1880 static void do_kvm_cpu_synchronize_post_init(void *arg
)
1882 CPUState
*cpu
= arg
;
1884 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
1885 cpu
->kvm_vcpu_dirty
= false;
1888 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
1890 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, cpu
);
1893 int kvm_cpu_exec(CPUState
*cpu
)
1895 struct kvm_run
*run
= cpu
->kvm_run
;
1898 DPRINTF("kvm_cpu_exec()\n");
1900 if (kvm_arch_process_async_events(cpu
)) {
1901 cpu
->exit_request
= 0;
1905 qemu_mutex_unlock_iothread();
1910 if (cpu
->kvm_vcpu_dirty
) {
1911 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
1912 cpu
->kvm_vcpu_dirty
= false;
1915 kvm_arch_pre_run(cpu
, run
);
1916 if (cpu
->exit_request
) {
1917 DPRINTF("interrupt exit requested\n");
1919 * KVM requires us to reenter the kernel after IO exits to complete
1920 * instruction emulation. This self-signal will ensure that we
1923 qemu_cpu_kick_self();
1926 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
1928 attrs
= kvm_arch_post_run(cpu
, run
);
1931 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
1932 DPRINTF("io window exit\n");
1933 ret
= EXCP_INTERRUPT
;
1936 fprintf(stderr
, "error: kvm run failed %s\n",
1937 strerror(-run_ret
));
1939 if (run_ret
== -EBUSY
) {
1941 "This is probably because your SMT is enabled.\n"
1942 "VCPU can only run on primary threads with all "
1943 "secondary threads offline.\n");
1950 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
1951 switch (run
->exit_reason
) {
1953 DPRINTF("handle_io\n");
1954 /* Called outside BQL */
1955 kvm_handle_io(run
->io
.port
, attrs
,
1956 (uint8_t *)run
+ run
->io
.data_offset
,
1963 DPRINTF("handle_mmio\n");
1964 /* Called outside BQL */
1965 address_space_rw(&address_space_memory
,
1966 run
->mmio
.phys_addr
, attrs
,
1969 run
->mmio
.is_write
);
1972 case KVM_EXIT_IRQ_WINDOW_OPEN
:
1973 DPRINTF("irq_window_open\n");
1974 ret
= EXCP_INTERRUPT
;
1976 case KVM_EXIT_SHUTDOWN
:
1977 DPRINTF("shutdown\n");
1978 qemu_system_reset_request();
1979 ret
= EXCP_INTERRUPT
;
1981 case KVM_EXIT_UNKNOWN
:
1982 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
1983 (uint64_t)run
->hw
.hardware_exit_reason
);
1986 case KVM_EXIT_INTERNAL_ERROR
:
1987 ret
= kvm_handle_internal_error(cpu
, run
);
1989 case KVM_EXIT_SYSTEM_EVENT
:
1990 switch (run
->system_event
.type
) {
1991 case KVM_SYSTEM_EVENT_SHUTDOWN
:
1992 qemu_system_shutdown_request();
1993 ret
= EXCP_INTERRUPT
;
1995 case KVM_SYSTEM_EVENT_RESET
:
1996 qemu_system_reset_request();
1997 ret
= EXCP_INTERRUPT
;
1999 case KVM_SYSTEM_EVENT_CRASH
:
2000 qemu_mutex_lock_iothread();
2001 qemu_system_guest_panicked();
2002 qemu_mutex_unlock_iothread();
2006 DPRINTF("kvm_arch_handle_exit\n");
2007 ret
= kvm_arch_handle_exit(cpu
, run
);
2012 DPRINTF("kvm_arch_handle_exit\n");
2013 ret
= kvm_arch_handle_exit(cpu
, run
);
2018 qemu_mutex_lock_iothread();
2021 cpu_dump_state(cpu
, stderr
, fprintf
, CPU_DUMP_CODE
);
2022 vm_stop(RUN_STATE_INTERNAL_ERROR
);
2025 cpu
->exit_request
= 0;
2029 int kvm_ioctl(KVMState
*s
, int type
, ...)
2036 arg
= va_arg(ap
, void *);
2039 trace_kvm_ioctl(type
, arg
);
2040 ret
= ioctl(s
->fd
, type
, arg
);
2047 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
2054 arg
= va_arg(ap
, void *);
2057 trace_kvm_vm_ioctl(type
, arg
);
2058 ret
= ioctl(s
->vmfd
, type
, arg
);
2065 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
2072 arg
= va_arg(ap
, void *);
2075 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
2076 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
2083 int kvm_device_ioctl(int fd
, int type
, ...)
2090 arg
= va_arg(ap
, void *);
2093 trace_kvm_device_ioctl(fd
, type
, arg
);
2094 ret
= ioctl(fd
, type
, arg
);
2101 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
2104 struct kvm_device_attr attribute
= {
2109 if (!kvm_vm_attributes_allowed
) {
2113 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
2114 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2118 int kvm_device_check_attr(int dev_fd
, uint32_t group
, uint64_t attr
)
2120 struct kvm_device_attr attribute
= {
2126 return kvm_device_ioctl(dev_fd
, KVM_HAS_DEVICE_ATTR
, &attribute
) ? 0 : 1;
2129 void kvm_device_access(int fd
, int group
, uint64_t attr
,
2130 void *val
, bool write
)
2132 struct kvm_device_attr kvmattr
;
2136 kvmattr
.group
= group
;
2137 kvmattr
.attr
= attr
;
2138 kvmattr
.addr
= (uintptr_t)val
;
2140 err
= kvm_device_ioctl(fd
,
2141 write
? KVM_SET_DEVICE_ATTR
: KVM_GET_DEVICE_ATTR
,
2144 error_report("KVM_%s_DEVICE_ATTR failed: %s",
2145 write
? "SET" : "GET", strerror(-err
));
2146 error_printf("Group %d attr 0x%016" PRIx64
"\n", group
, attr
);
2151 int kvm_has_sync_mmu(void)
2153 return kvm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
2156 int kvm_has_vcpu_events(void)
2158 return kvm_state
->vcpu_events
;
2161 int kvm_has_robust_singlestep(void)
2163 return kvm_state
->robust_singlestep
;
2166 int kvm_has_debugregs(void)
2168 return kvm_state
->debugregs
;
2171 int kvm_has_many_ioeventfds(void)
2173 if (!kvm_enabled()) {
2176 return kvm_state
->many_ioeventfds
;
2179 int kvm_has_gsi_routing(void)
2181 #ifdef KVM_CAP_IRQ_ROUTING
2182 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
2188 int kvm_has_intx_set_mask(void)
2190 return kvm_state
->intx_set_mask
;
2193 void kvm_setup_guest_memory(void *start
, size_t size
)
2195 if (!kvm_has_sync_mmu()) {
2196 int ret
= qemu_madvise(start
, size
, QEMU_MADV_DONTFORK
);
2199 perror("qemu_madvise");
2201 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2207 #ifdef KVM_CAP_SET_GUEST_DEBUG
2208 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
2211 struct kvm_sw_breakpoint
*bp
;
2213 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
2221 int kvm_sw_breakpoints_active(CPUState
*cpu
)
2223 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
2226 struct kvm_set_guest_debug_data
{
2227 struct kvm_guest_debug dbg
;
2232 static void kvm_invoke_set_guest_debug(void *data
)
2234 struct kvm_set_guest_debug_data
*dbg_data
= data
;
2236 dbg_data
->err
= kvm_vcpu_ioctl(dbg_data
->cpu
, KVM_SET_GUEST_DEBUG
,
2240 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2242 struct kvm_set_guest_debug_data data
;
2244 data
.dbg
.control
= reinject_trap
;
2246 if (cpu
->singlestep_enabled
) {
2247 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
2249 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
2252 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
, &data
);
2256 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2257 target_ulong len
, int type
)
2259 struct kvm_sw_breakpoint
*bp
;
2262 if (type
== GDB_BREAKPOINT_SW
) {
2263 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2269 bp
= g_malloc(sizeof(struct kvm_sw_breakpoint
));
2272 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
2278 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2280 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
2287 err
= kvm_update_guest_debug(cpu
, 0);
2295 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2296 target_ulong len
, int type
)
2298 struct kvm_sw_breakpoint
*bp
;
2301 if (type
== GDB_BREAKPOINT_SW
) {
2302 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2307 if (bp
->use_count
> 1) {
2312 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
2317 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2320 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
2327 err
= kvm_update_guest_debug(cpu
, 0);
2335 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2337 struct kvm_sw_breakpoint
*bp
, *next
;
2338 KVMState
*s
= cpu
->kvm_state
;
2341 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
2342 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
2343 /* Try harder to find a CPU that currently sees the breakpoint. */
2344 CPU_FOREACH(tmpcpu
) {
2345 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
2350 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
2353 kvm_arch_remove_all_hw_breakpoints();
2356 kvm_update_guest_debug(cpu
, 0);
2360 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2362 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2367 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2368 target_ulong len
, int type
)
2373 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2374 target_ulong len
, int type
)
2379 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2382 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2384 int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
2386 KVMState
*s
= kvm_state
;
2387 struct kvm_signal_mask
*sigmask
;
2391 return kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, NULL
);
2394 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
2396 sigmask
->len
= s
->sigmask_len
;
2397 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
2398 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
2403 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
2405 return kvm_arch_on_sigbus_vcpu(cpu
, code
, addr
);
2408 int kvm_on_sigbus(int code
, void *addr
)
2410 return kvm_arch_on_sigbus(code
, addr
);
2413 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
2416 struct kvm_create_device create_dev
;
2418 create_dev
.type
= type
;
2420 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
2422 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
2426 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
2431 return test
? 0 : create_dev
.fd
;
2434 bool kvm_device_supported(int vmfd
, uint64_t type
)
2436 struct kvm_create_device create_dev
= {
2439 .flags
= KVM_CREATE_DEVICE_TEST
,
2442 if (ioctl(vmfd
, KVM_CHECK_EXTENSION
, KVM_CAP_DEVICE_CTRL
) <= 0) {
2446 return (ioctl(vmfd
, KVM_CREATE_DEVICE
, &create_dev
) >= 0);
2449 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
2451 struct kvm_one_reg reg
;
2455 reg
.addr
= (uintptr_t) source
;
2456 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
2458 trace_kvm_failed_reg_set(id
, strerror(-r
));
2463 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
2465 struct kvm_one_reg reg
;
2469 reg
.addr
= (uintptr_t) target
;
2470 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
2472 trace_kvm_failed_reg_get(id
, strerror(-r
));
2477 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
2479 AccelClass
*ac
= ACCEL_CLASS(oc
);
2481 ac
->init_machine
= kvm_init
;
2482 ac
->allowed
= &kvm_allowed
;
2485 static const TypeInfo kvm_accel_type
= {
2486 .name
= TYPE_KVM_ACCEL
,
2487 .parent
= TYPE_ACCEL
,
2488 .class_init
= kvm_accel_class_init
,
2489 .instance_size
= sizeof(KVMState
),
2492 static void kvm_type_init(void)
2494 type_register_static(&kvm_accel_type
);
2497 type_init(kvm_type_init
);