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/atomic.h"
22 #include "qemu/option.h"
23 #include "qemu/config-file.h"
24 #include "qemu/error-report.h"
25 #include "qapi/error.h"
26 #include "hw/pci/msi.h"
27 #include "hw/pci/msix.h"
28 #include "hw/s390x/adapter.h"
29 #include "exec/gdbstub.h"
30 #include "sysemu/kvm_int.h"
31 #include "sysemu/runstate.h"
32 #include "sysemu/cpus.h"
33 #include "sysemu/sysemu.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"
39 #include "qemu/main-loop.h"
42 #include "sysemu/sev.h"
43 #include "qapi/visitor.h"
44 #include "qapi/qapi-types-common.h"
45 #include "qapi/qapi-visit-common.h"
46 #include "sysemu/reset.h"
47 #include "qemu/guest-random.h"
48 #include "sysemu/hw_accel.h"
51 #include "hw/boards.h"
53 /* This check must be after config-host.h is included */
55 #include <sys/eventfd.h>
58 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
59 * need to use the real host PAGE_SIZE, as that's what KVM will use.
61 #define PAGE_SIZE qemu_real_host_page_size
66 #define DPRINTF(fmt, ...) \
67 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
69 #define DPRINTF(fmt, ...) \
73 #define KVM_MSI_HASHTAB_SIZE 256
75 struct KVMParkedVcpu
{
76 unsigned long vcpu_id
;
78 QLIST_ENTRY(KVMParkedVcpu
) node
;
83 AccelState parent_obj
;
90 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
91 bool coalesced_flush_in_progress
;
93 int robust_singlestep
;
95 #ifdef KVM_CAP_SET_GUEST_DEBUG
96 QTAILQ_HEAD(, kvm_sw_breakpoint
) kvm_sw_breakpoints
;
98 int max_nested_state_len
;
102 bool kernel_irqchip_allowed
;
103 bool kernel_irqchip_required
;
104 OnOffAuto kernel_irqchip_split
;
106 uint64_t manual_dirty_log_protect
;
107 /* The man page (and posix) say ioctl numbers are signed int, but
108 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
109 * unsigned, and treating them as signed here can break things */
110 unsigned irq_set_ioctl
;
111 unsigned int sigmask_len
;
113 #ifdef KVM_CAP_IRQ_ROUTING
114 struct kvm_irq_routing
*irq_routes
;
115 int nr_allocated_irq_routes
;
116 unsigned long *used_gsi_bitmap
;
117 unsigned int gsi_count
;
118 QTAILQ_HEAD(, KVMMSIRoute
) msi_hashtab
[KVM_MSI_HASHTAB_SIZE
];
120 KVMMemoryListener memory_listener
;
121 QLIST_HEAD(, KVMParkedVcpu
) kvm_parked_vcpus
;
123 /* memory encryption */
124 void *memcrypt_handle
;
125 int (*memcrypt_encrypt_data
)(void *handle
, uint8_t *ptr
, uint64_t len
);
127 /* For "info mtree -f" to tell if an MR is registered in KVM */
130 KVMMemoryListener
*ml
;
136 bool kvm_kernel_irqchip
;
137 bool kvm_split_irqchip
;
138 bool kvm_async_interrupts_allowed
;
139 bool kvm_halt_in_kernel_allowed
;
140 bool kvm_eventfds_allowed
;
141 bool kvm_irqfds_allowed
;
142 bool kvm_resamplefds_allowed
;
143 bool kvm_msi_via_irqfd_allowed
;
144 bool kvm_gsi_routing_allowed
;
145 bool kvm_gsi_direct_mapping
;
147 bool kvm_readonly_mem_allowed
;
148 bool kvm_vm_attributes_allowed
;
149 bool kvm_direct_msi_allowed
;
150 bool kvm_ioeventfd_any_length_allowed
;
151 bool kvm_msi_use_devid
;
152 static bool kvm_immediate_exit
;
153 static hwaddr kvm_max_slot_size
= ~0;
155 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
156 KVM_CAP_INFO(USER_MEMORY
),
157 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
158 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS
),
162 static NotifierList kvm_irqchip_change_notifiers
=
163 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers
);
165 struct KVMResampleFd
{
167 EventNotifier
*resample_event
;
168 QLIST_ENTRY(KVMResampleFd
) node
;
170 typedef struct KVMResampleFd KVMResampleFd
;
173 * Only used with split irqchip where we need to do the resample fd
174 * kick for the kernel from userspace.
176 static QLIST_HEAD(, KVMResampleFd
) kvm_resample_fd_list
=
177 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list
);
179 #define kvm_slots_lock(kml) qemu_mutex_lock(&(kml)->slots_lock)
180 #define kvm_slots_unlock(kml) qemu_mutex_unlock(&(kml)->slots_lock)
182 static inline void kvm_resample_fd_remove(int gsi
)
186 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
187 if (rfd
->gsi
== gsi
) {
188 QLIST_REMOVE(rfd
, node
);
195 static inline void kvm_resample_fd_insert(int gsi
, EventNotifier
*event
)
197 KVMResampleFd
*rfd
= g_new0(KVMResampleFd
, 1);
200 rfd
->resample_event
= event
;
202 QLIST_INSERT_HEAD(&kvm_resample_fd_list
, rfd
, node
);
205 void kvm_resample_fd_notify(int gsi
)
209 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
210 if (rfd
->gsi
== gsi
) {
211 event_notifier_set(rfd
->resample_event
);
212 trace_kvm_resample_fd_notify(gsi
);
218 int kvm_get_max_memslots(void)
220 KVMState
*s
= KVM_STATE(current_accel());
225 bool kvm_memcrypt_enabled(void)
227 if (kvm_state
&& kvm_state
->memcrypt_handle
) {
234 int kvm_memcrypt_encrypt_data(uint8_t *ptr
, uint64_t len
)
236 if (kvm_state
->memcrypt_handle
&&
237 kvm_state
->memcrypt_encrypt_data
) {
238 return kvm_state
->memcrypt_encrypt_data(kvm_state
->memcrypt_handle
,
245 /* Called with KVMMemoryListener.slots_lock held */
246 static KVMSlot
*kvm_get_free_slot(KVMMemoryListener
*kml
)
248 KVMState
*s
= kvm_state
;
251 for (i
= 0; i
< s
->nr_slots
; i
++) {
252 if (kml
->slots
[i
].memory_size
== 0) {
253 return &kml
->slots
[i
];
260 bool kvm_has_free_slot(MachineState
*ms
)
262 KVMState
*s
= KVM_STATE(ms
->accelerator
);
264 KVMMemoryListener
*kml
= &s
->memory_listener
;
267 result
= !!kvm_get_free_slot(kml
);
268 kvm_slots_unlock(kml
);
273 /* Called with KVMMemoryListener.slots_lock held */
274 static KVMSlot
*kvm_alloc_slot(KVMMemoryListener
*kml
)
276 KVMSlot
*slot
= kvm_get_free_slot(kml
);
282 fprintf(stderr
, "%s: no free slot available\n", __func__
);
286 static KVMSlot
*kvm_lookup_matching_slot(KVMMemoryListener
*kml
,
290 KVMState
*s
= kvm_state
;
293 for (i
= 0; i
< s
->nr_slots
; i
++) {
294 KVMSlot
*mem
= &kml
->slots
[i
];
296 if (start_addr
== mem
->start_addr
&& size
== mem
->memory_size
) {
305 * Calculate and align the start address and the size of the section.
306 * Return the size. If the size is 0, the aligned section is empty.
308 static hwaddr
kvm_align_section(MemoryRegionSection
*section
,
311 hwaddr size
= int128_get64(section
->size
);
312 hwaddr delta
, aligned
;
314 /* kvm works in page size chunks, but the function may be called
315 with sub-page size and unaligned start address. Pad the start
316 address to next and truncate size to previous page boundary. */
317 aligned
= ROUND_UP(section
->offset_within_address_space
,
318 qemu_real_host_page_size
);
319 delta
= aligned
- section
->offset_within_address_space
;
325 return (size
- delta
) & qemu_real_host_page_mask
;
328 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
331 KVMMemoryListener
*kml
= &s
->memory_listener
;
335 for (i
= 0; i
< s
->nr_slots
; i
++) {
336 KVMSlot
*mem
= &kml
->slots
[i
];
338 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
339 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
344 kvm_slots_unlock(kml
);
349 static int kvm_set_user_memory_region(KVMMemoryListener
*kml
, KVMSlot
*slot
, bool new)
351 KVMState
*s
= kvm_state
;
352 struct kvm_userspace_memory_region mem
;
355 mem
.slot
= slot
->slot
| (kml
->as_id
<< 16);
356 mem
.guest_phys_addr
= slot
->start_addr
;
357 mem
.userspace_addr
= (unsigned long)slot
->ram
;
358 mem
.flags
= slot
->flags
;
360 if (slot
->memory_size
&& !new && (mem
.flags
^ slot
->old_flags
) & KVM_MEM_READONLY
) {
361 /* Set the slot size to 0 before setting the slot to the desired
362 * value. This is needed based on KVM commit 75d61fbc. */
364 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
369 mem
.memory_size
= slot
->memory_size
;
370 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
371 slot
->old_flags
= mem
.flags
;
373 trace_kvm_set_user_memory(mem
.slot
, mem
.flags
, mem
.guest_phys_addr
,
374 mem
.memory_size
, mem
.userspace_addr
, ret
);
376 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
377 " start=0x%" PRIx64
", size=0x%" PRIx64
": %s",
378 __func__
, mem
.slot
, slot
->start_addr
,
379 (uint64_t)mem
.memory_size
, strerror(errno
));
384 static int do_kvm_destroy_vcpu(CPUState
*cpu
)
386 KVMState
*s
= kvm_state
;
388 struct KVMParkedVcpu
*vcpu
= NULL
;
391 DPRINTF("kvm_destroy_vcpu\n");
393 ret
= kvm_arch_destroy_vcpu(cpu
);
398 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
401 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
405 ret
= munmap(cpu
->kvm_run
, mmap_size
);
410 vcpu
= g_malloc0(sizeof(*vcpu
));
411 vcpu
->vcpu_id
= kvm_arch_vcpu_id(cpu
);
412 vcpu
->kvm_fd
= cpu
->kvm_fd
;
413 QLIST_INSERT_HEAD(&kvm_state
->kvm_parked_vcpus
, vcpu
, node
);
418 void kvm_destroy_vcpu(CPUState
*cpu
)
420 if (do_kvm_destroy_vcpu(cpu
) < 0) {
421 error_report("kvm_destroy_vcpu failed");
426 static int kvm_get_vcpu(KVMState
*s
, unsigned long vcpu_id
)
428 struct KVMParkedVcpu
*cpu
;
430 QLIST_FOREACH(cpu
, &s
->kvm_parked_vcpus
, node
) {
431 if (cpu
->vcpu_id
== vcpu_id
) {
434 QLIST_REMOVE(cpu
, node
);
435 kvm_fd
= cpu
->kvm_fd
;
441 return kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)vcpu_id
);
444 int kvm_init_vcpu(CPUState
*cpu
, Error
**errp
)
446 KVMState
*s
= kvm_state
;
450 trace_kvm_init_vcpu(cpu
->cpu_index
, kvm_arch_vcpu_id(cpu
));
452 ret
= kvm_get_vcpu(s
, kvm_arch_vcpu_id(cpu
));
454 error_setg_errno(errp
, -ret
, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)",
455 kvm_arch_vcpu_id(cpu
));
461 cpu
->vcpu_dirty
= true;
463 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
466 error_setg_errno(errp
, -mmap_size
,
467 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
471 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
473 if (cpu
->kvm_run
== MAP_FAILED
) {
475 error_setg_errno(errp
, ret
,
476 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
477 kvm_arch_vcpu_id(cpu
));
481 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
482 s
->coalesced_mmio_ring
=
483 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
486 ret
= kvm_arch_init_vcpu(cpu
);
488 error_setg_errno(errp
, -ret
,
489 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
490 kvm_arch_vcpu_id(cpu
));
497 * dirty pages logging control
500 static int kvm_mem_flags(MemoryRegion
*mr
)
502 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
505 if (memory_region_get_dirty_log_mask(mr
) != 0) {
506 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
508 if (readonly
&& kvm_readonly_mem_allowed
) {
509 flags
|= KVM_MEM_READONLY
;
514 /* Called with KVMMemoryListener.slots_lock held */
515 static int kvm_slot_update_flags(KVMMemoryListener
*kml
, KVMSlot
*mem
,
518 mem
->flags
= kvm_mem_flags(mr
);
520 /* If nothing changed effectively, no need to issue ioctl */
521 if (mem
->flags
== mem
->old_flags
) {
525 return kvm_set_user_memory_region(kml
, mem
, false);
528 static int kvm_section_update_flags(KVMMemoryListener
*kml
,
529 MemoryRegionSection
*section
)
531 hwaddr start_addr
, size
, slot_size
;
535 size
= kvm_align_section(section
, &start_addr
);
542 while (size
&& !ret
) {
543 slot_size
= MIN(kvm_max_slot_size
, size
);
544 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
546 /* We don't have a slot if we want to trap every access. */
550 ret
= kvm_slot_update_flags(kml
, mem
, section
->mr
);
551 start_addr
+= slot_size
;
556 kvm_slots_unlock(kml
);
560 static void kvm_log_start(MemoryListener
*listener
,
561 MemoryRegionSection
*section
,
564 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
571 r
= kvm_section_update_flags(kml
, section
);
577 static void kvm_log_stop(MemoryListener
*listener
,
578 MemoryRegionSection
*section
,
581 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
588 r
= kvm_section_update_flags(kml
, section
);
594 /* get kvm's dirty pages bitmap and update qemu's */
595 static int kvm_get_dirty_pages_log_range(MemoryRegionSection
*section
,
596 unsigned long *bitmap
)
598 ram_addr_t start
= section
->offset_within_region
+
599 memory_region_get_ram_addr(section
->mr
);
600 ram_addr_t pages
= int128_get64(section
->size
) / qemu_real_host_page_size
;
602 cpu_physical_memory_set_dirty_lebitmap(bitmap
, start
, pages
);
606 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
608 /* Allocate the dirty bitmap for a slot */
609 static void kvm_memslot_init_dirty_bitmap(KVMSlot
*mem
)
612 * XXX bad kernel interface alert
613 * For dirty bitmap, kernel allocates array of size aligned to
614 * bits-per-long. But for case when the kernel is 64bits and
615 * the userspace is 32bits, userspace can't align to the same
616 * bits-per-long, since sizeof(long) is different between kernel
617 * and user space. This way, userspace will provide buffer which
618 * may be 4 bytes less than the kernel will use, resulting in
619 * userspace memory corruption (which is not detectable by valgrind
620 * too, in most cases).
621 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
622 * a hope that sizeof(long) won't become >8 any time soon.
624 hwaddr bitmap_size
= ALIGN(((mem
->memory_size
) >> TARGET_PAGE_BITS
),
625 /*HOST_LONG_BITS*/ 64) / 8;
626 mem
->dirty_bmap
= g_malloc0(bitmap_size
);
630 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
632 * This function will first try to fetch dirty bitmap from the kernel,
633 * and then updates qemu's dirty bitmap.
635 * NOTE: caller must be with kml->slots_lock held.
637 * @kml: the KVM memory listener object
638 * @section: the memory section to sync the dirty bitmap with
640 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener
*kml
,
641 MemoryRegionSection
*section
)
643 KVMState
*s
= kvm_state
;
644 struct kvm_dirty_log d
= {};
646 hwaddr start_addr
, size
;
647 hwaddr slot_size
, slot_offset
= 0;
650 size
= kvm_align_section(section
, &start_addr
);
652 MemoryRegionSection subsection
= *section
;
654 slot_size
= MIN(kvm_max_slot_size
, size
);
655 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
657 /* We don't have a slot if we want to trap every access. */
661 if (!mem
->dirty_bmap
) {
662 /* Allocate on the first log_sync, once and for all */
663 kvm_memslot_init_dirty_bitmap(mem
);
666 d
.dirty_bitmap
= mem
->dirty_bmap
;
667 d
.slot
= mem
->slot
| (kml
->as_id
<< 16);
668 if (kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
) == -1) {
669 DPRINTF("ioctl failed %d\n", errno
);
674 subsection
.offset_within_region
+= slot_offset
;
675 subsection
.size
= int128_make64(slot_size
);
676 kvm_get_dirty_pages_log_range(&subsection
, d
.dirty_bitmap
);
678 slot_offset
+= slot_size
;
679 start_addr
+= slot_size
;
686 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
687 #define KVM_CLEAR_LOG_SHIFT 6
688 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT)
689 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN)
691 static int kvm_log_clear_one_slot(KVMSlot
*mem
, int as_id
, uint64_t start
,
694 KVMState
*s
= kvm_state
;
695 uint64_t end
, bmap_start
, start_delta
, bmap_npages
;
696 struct kvm_clear_dirty_log d
;
697 unsigned long *bmap_clear
= NULL
, psize
= qemu_real_host_page_size
;
701 * We need to extend either the start or the size or both to
702 * satisfy the KVM interface requirement. Firstly, do the start
703 * page alignment on 64 host pages
705 bmap_start
= start
& KVM_CLEAR_LOG_MASK
;
706 start_delta
= start
- bmap_start
;
710 * The kernel interface has restriction on the size too, that either:
712 * (1) the size is 64 host pages aligned (just like the start), or
713 * (2) the size fills up until the end of the KVM memslot.
715 bmap_npages
= DIV_ROUND_UP(size
+ start_delta
, KVM_CLEAR_LOG_ALIGN
)
716 << KVM_CLEAR_LOG_SHIFT
;
717 end
= mem
->memory_size
/ psize
;
718 if (bmap_npages
> end
- bmap_start
) {
719 bmap_npages
= end
- bmap_start
;
721 start_delta
/= psize
;
724 * Prepare the bitmap to clear dirty bits. Here we must guarantee
725 * that we won't clear any unknown dirty bits otherwise we might
726 * accidentally clear some set bits which are not yet synced from
727 * the kernel into QEMU's bitmap, then we'll lose track of the
728 * guest modifications upon those pages (which can directly lead
729 * to guest data loss or panic after migration).
731 * Layout of the KVMSlot.dirty_bmap:
733 * |<-------- bmap_npages -----------..>|
736 * |----------------|-------------|------------------|------------|
739 * start bmap_start (start) end
740 * of memslot of memslot
742 * [1] bmap_npages can be aligned to either 64 pages or the end of slot
745 assert(bmap_start
% BITS_PER_LONG
== 0);
746 /* We should never do log_clear before log_sync */
747 assert(mem
->dirty_bmap
);
748 if (start_delta
|| bmap_npages
- size
/ psize
) {
749 /* Slow path - we need to manipulate a temp bitmap */
750 bmap_clear
= bitmap_new(bmap_npages
);
751 bitmap_copy_with_src_offset(bmap_clear
, mem
->dirty_bmap
,
752 bmap_start
, start_delta
+ size
/ psize
);
754 * We need to fill the holes at start because that was not
755 * specified by the caller and we extended the bitmap only for
758 bitmap_clear(bmap_clear
, 0, start_delta
);
759 d
.dirty_bitmap
= bmap_clear
;
762 * Fast path - both start and size align well with BITS_PER_LONG
763 * (or the end of memory slot)
765 d
.dirty_bitmap
= mem
->dirty_bmap
+ BIT_WORD(bmap_start
);
768 d
.first_page
= bmap_start
;
769 /* It should never overflow. If it happens, say something */
770 assert(bmap_npages
<= UINT32_MAX
);
771 d
.num_pages
= bmap_npages
;
772 d
.slot
= mem
->slot
| (as_id
<< 16);
774 if (kvm_vm_ioctl(s
, KVM_CLEAR_DIRTY_LOG
, &d
) == -1) {
776 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
777 "start=0x%"PRIx64
", size=0x%"PRIx32
", errno=%d",
778 __func__
, d
.slot
, (uint64_t)d
.first_page
,
779 (uint32_t)d
.num_pages
, ret
);
782 trace_kvm_clear_dirty_log(d
.slot
, d
.first_page
, d
.num_pages
);
786 * After we have updated the remote dirty bitmap, we update the
787 * cached bitmap as well for the memslot, then if another user
788 * clears the same region we know we shouldn't clear it again on
789 * the remote otherwise it's data loss as well.
791 bitmap_clear(mem
->dirty_bmap
, bmap_start
+ start_delta
,
793 /* This handles the NULL case well */
800 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
802 * NOTE: this will be a no-op if we haven't enabled manual dirty log
803 * protection in the host kernel because in that case this operation
804 * will be done within log_sync().
806 * @kml: the kvm memory listener
807 * @section: the memory range to clear dirty bitmap
809 static int kvm_physical_log_clear(KVMMemoryListener
*kml
,
810 MemoryRegionSection
*section
)
812 KVMState
*s
= kvm_state
;
813 uint64_t start
, size
, offset
, count
;
817 if (!s
->manual_dirty_log_protect
) {
818 /* No need to do explicit clear */
822 start
= section
->offset_within_address_space
;
823 size
= int128_get64(section
->size
);
826 /* Nothing more we can do... */
832 for (i
= 0; i
< s
->nr_slots
; i
++) {
833 mem
= &kml
->slots
[i
];
834 /* Discard slots that are empty or do not overlap the section */
835 if (!mem
->memory_size
||
836 mem
->start_addr
> start
+ size
- 1 ||
837 start
> mem
->start_addr
+ mem
->memory_size
- 1) {
841 if (start
>= mem
->start_addr
) {
842 /* The slot starts before section or is aligned to it. */
843 offset
= start
- mem
->start_addr
;
844 count
= MIN(mem
->memory_size
- offset
, size
);
846 /* The slot starts after section. */
848 count
= MIN(mem
->memory_size
, size
- (mem
->start_addr
- start
));
850 ret
= kvm_log_clear_one_slot(mem
, kml
->as_id
, offset
, count
);
856 kvm_slots_unlock(kml
);
861 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
862 MemoryRegionSection
*secion
,
863 hwaddr start
, hwaddr size
)
865 KVMState
*s
= kvm_state
;
867 if (s
->coalesced_mmio
) {
868 struct kvm_coalesced_mmio_zone zone
;
874 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
878 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
879 MemoryRegionSection
*secion
,
880 hwaddr start
, hwaddr size
)
882 KVMState
*s
= kvm_state
;
884 if (s
->coalesced_mmio
) {
885 struct kvm_coalesced_mmio_zone zone
;
891 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
895 static void kvm_coalesce_pio_add(MemoryListener
*listener
,
896 MemoryRegionSection
*section
,
897 hwaddr start
, hwaddr size
)
899 KVMState
*s
= kvm_state
;
901 if (s
->coalesced_pio
) {
902 struct kvm_coalesced_mmio_zone zone
;
908 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
912 static void kvm_coalesce_pio_del(MemoryListener
*listener
,
913 MemoryRegionSection
*section
,
914 hwaddr start
, hwaddr size
)
916 KVMState
*s
= kvm_state
;
918 if (s
->coalesced_pio
) {
919 struct kvm_coalesced_mmio_zone zone
;
925 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
929 static MemoryListener kvm_coalesced_pio_listener
= {
930 .coalesced_io_add
= kvm_coalesce_pio_add
,
931 .coalesced_io_del
= kvm_coalesce_pio_del
,
934 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
938 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
946 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
950 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
952 /* VM wide version not implemented, use global one instead */
953 ret
= kvm_check_extension(s
, extension
);
959 typedef struct HWPoisonPage
{
961 QLIST_ENTRY(HWPoisonPage
) list
;
964 static QLIST_HEAD(, HWPoisonPage
) hwpoison_page_list
=
965 QLIST_HEAD_INITIALIZER(hwpoison_page_list
);
967 static void kvm_unpoison_all(void *param
)
969 HWPoisonPage
*page
, *next_page
;
971 QLIST_FOREACH_SAFE(page
, &hwpoison_page_list
, list
, next_page
) {
972 QLIST_REMOVE(page
, list
);
973 qemu_ram_remap(page
->ram_addr
, TARGET_PAGE_SIZE
);
978 void kvm_hwpoison_page_add(ram_addr_t ram_addr
)
982 QLIST_FOREACH(page
, &hwpoison_page_list
, list
) {
983 if (page
->ram_addr
== ram_addr
) {
987 page
= g_new(HWPoisonPage
, 1);
988 page
->ram_addr
= ram_addr
;
989 QLIST_INSERT_HEAD(&hwpoison_page_list
, page
, list
);
992 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
994 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
995 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
996 * endianness, but the memory core hands them in target endianness.
997 * For example, PPC is always treated as big-endian even if running
998 * on KVM and on PPC64LE. Correct here.
1012 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
1013 bool assign
, uint32_t size
, bool datamatch
)
1016 struct kvm_ioeventfd iofd
= {
1017 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1024 trace_kvm_set_ioeventfd_mmio(fd
, (uint64_t)addr
, val
, assign
, size
,
1026 if (!kvm_enabled()) {
1031 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1034 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1037 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
1046 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
1047 bool assign
, uint32_t size
, bool datamatch
)
1049 struct kvm_ioeventfd kick
= {
1050 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1052 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
1057 trace_kvm_set_ioeventfd_pio(fd
, addr
, val
, assign
, size
, datamatch
);
1058 if (!kvm_enabled()) {
1062 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1065 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1067 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
1075 static int kvm_check_many_ioeventfds(void)
1077 /* Userspace can use ioeventfd for io notification. This requires a host
1078 * that supports eventfd(2) and an I/O thread; since eventfd does not
1079 * support SIGIO it cannot interrupt the vcpu.
1081 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
1082 * can avoid creating too many ioeventfds.
1084 #if defined(CONFIG_EVENTFD)
1087 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
1088 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
1089 if (ioeventfds
[i
] < 0) {
1092 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
1094 close(ioeventfds
[i
]);
1099 /* Decide whether many devices are supported or not */
1100 ret
= i
== ARRAY_SIZE(ioeventfds
);
1103 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
1104 close(ioeventfds
[i
]);
1112 static const KVMCapabilityInfo
*
1113 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
1115 while (list
->name
) {
1116 if (!kvm_check_extension(s
, list
->value
)) {
1124 void kvm_set_max_memslot_size(hwaddr max_slot_size
)
1127 ROUND_UP(max_slot_size
, qemu_real_host_page_size
) == max_slot_size
1129 kvm_max_slot_size
= max_slot_size
;
1132 static void kvm_set_phys_mem(KVMMemoryListener
*kml
,
1133 MemoryRegionSection
*section
, bool add
)
1137 MemoryRegion
*mr
= section
->mr
;
1138 bool writeable
= !mr
->readonly
&& !mr
->rom_device
;
1139 hwaddr start_addr
, size
, slot_size
;
1142 if (!memory_region_is_ram(mr
)) {
1143 if (writeable
|| !kvm_readonly_mem_allowed
) {
1145 } else if (!mr
->romd_mode
) {
1146 /* If the memory device is not in romd_mode, then we actually want
1147 * to remove the kvm memory slot so all accesses will trap. */
1152 size
= kvm_align_section(section
, &start_addr
);
1157 /* use aligned delta to align the ram address */
1158 ram
= memory_region_get_ram_ptr(mr
) + section
->offset_within_region
+
1159 (start_addr
- section
->offset_within_address_space
);
1161 kvm_slots_lock(kml
);
1165 slot_size
= MIN(kvm_max_slot_size
, size
);
1166 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
1170 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1171 kvm_physical_sync_dirty_bitmap(kml
, section
);
1174 /* unregister the slot */
1175 g_free(mem
->dirty_bmap
);
1176 mem
->dirty_bmap
= NULL
;
1177 mem
->memory_size
= 0;
1179 err
= kvm_set_user_memory_region(kml
, mem
, false);
1181 fprintf(stderr
, "%s: error unregistering slot: %s\n",
1182 __func__
, strerror(-err
));
1185 start_addr
+= slot_size
;
1191 /* register the new slot */
1193 slot_size
= MIN(kvm_max_slot_size
, size
);
1194 mem
= kvm_alloc_slot(kml
);
1195 mem
->memory_size
= slot_size
;
1196 mem
->start_addr
= start_addr
;
1198 mem
->flags
= kvm_mem_flags(mr
);
1200 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1202 * Reallocate the bmap; it means it doesn't disappear in
1203 * middle of a migrate.
1205 kvm_memslot_init_dirty_bitmap(mem
);
1207 err
= kvm_set_user_memory_region(kml
, mem
, true);
1209 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
1213 start_addr
+= slot_size
;
1219 kvm_slots_unlock(kml
);
1222 static void kvm_region_add(MemoryListener
*listener
,
1223 MemoryRegionSection
*section
)
1225 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1227 memory_region_ref(section
->mr
);
1228 kvm_set_phys_mem(kml
, section
, true);
1231 static void kvm_region_del(MemoryListener
*listener
,
1232 MemoryRegionSection
*section
)
1234 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1236 kvm_set_phys_mem(kml
, section
, false);
1237 memory_region_unref(section
->mr
);
1240 static void kvm_log_sync(MemoryListener
*listener
,
1241 MemoryRegionSection
*section
)
1243 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1246 kvm_slots_lock(kml
);
1247 r
= kvm_physical_sync_dirty_bitmap(kml
, section
);
1248 kvm_slots_unlock(kml
);
1254 static void kvm_log_clear(MemoryListener
*listener
,
1255 MemoryRegionSection
*section
)
1257 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1260 r
= kvm_physical_log_clear(kml
, section
);
1262 error_report_once("%s: kvm log clear failed: mr=%s "
1263 "offset=%"HWADDR_PRIx
" size=%"PRIx64
, __func__
,
1264 section
->mr
->name
, section
->offset_within_region
,
1265 int128_get64(section
->size
));
1270 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
1271 MemoryRegionSection
*section
,
1272 bool match_data
, uint64_t data
,
1275 int fd
= event_notifier_get_fd(e
);
1278 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1279 data
, true, int128_get64(section
->size
),
1282 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1283 __func__
, strerror(-r
), -r
);
1288 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
1289 MemoryRegionSection
*section
,
1290 bool match_data
, uint64_t data
,
1293 int fd
= event_notifier_get_fd(e
);
1296 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1297 data
, false, int128_get64(section
->size
),
1300 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1301 __func__
, strerror(-r
), -r
);
1306 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
1307 MemoryRegionSection
*section
,
1308 bool match_data
, uint64_t data
,
1311 int fd
= event_notifier_get_fd(e
);
1314 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1315 data
, true, int128_get64(section
->size
),
1318 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1319 __func__
, strerror(-r
), -r
);
1324 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
1325 MemoryRegionSection
*section
,
1326 bool match_data
, uint64_t data
,
1330 int fd
= event_notifier_get_fd(e
);
1333 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1334 data
, false, int128_get64(section
->size
),
1337 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1338 __func__
, strerror(-r
), -r
);
1343 void kvm_memory_listener_register(KVMState
*s
, KVMMemoryListener
*kml
,
1344 AddressSpace
*as
, int as_id
)
1348 qemu_mutex_init(&kml
->slots_lock
);
1349 kml
->slots
= g_malloc0(s
->nr_slots
* sizeof(KVMSlot
));
1352 for (i
= 0; i
< s
->nr_slots
; i
++) {
1353 kml
->slots
[i
].slot
= i
;
1356 kml
->listener
.region_add
= kvm_region_add
;
1357 kml
->listener
.region_del
= kvm_region_del
;
1358 kml
->listener
.log_start
= kvm_log_start
;
1359 kml
->listener
.log_stop
= kvm_log_stop
;
1360 kml
->listener
.log_sync
= kvm_log_sync
;
1361 kml
->listener
.log_clear
= kvm_log_clear
;
1362 kml
->listener
.priority
= 10;
1364 memory_listener_register(&kml
->listener
, as
);
1366 for (i
= 0; i
< s
->nr_as
; ++i
) {
1375 static MemoryListener kvm_io_listener
= {
1376 .eventfd_add
= kvm_io_ioeventfd_add
,
1377 .eventfd_del
= kvm_io_ioeventfd_del
,
1381 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
1383 struct kvm_irq_level event
;
1386 assert(kvm_async_interrupts_enabled());
1388 event
.level
= level
;
1390 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
1392 perror("kvm_set_irq");
1396 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
1399 #ifdef KVM_CAP_IRQ_ROUTING
1400 typedef struct KVMMSIRoute
{
1401 struct kvm_irq_routing_entry kroute
;
1402 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
1405 static void set_gsi(KVMState
*s
, unsigned int gsi
)
1407 set_bit(gsi
, s
->used_gsi_bitmap
);
1410 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
1412 clear_bit(gsi
, s
->used_gsi_bitmap
);
1415 void kvm_init_irq_routing(KVMState
*s
)
1419 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
1420 if (gsi_count
> 0) {
1421 /* Round up so we can search ints using ffs */
1422 s
->used_gsi_bitmap
= bitmap_new(gsi_count
);
1423 s
->gsi_count
= gsi_count
;
1426 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
1427 s
->nr_allocated_irq_routes
= 0;
1429 if (!kvm_direct_msi_allowed
) {
1430 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
1431 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
1435 kvm_arch_init_irq_routing(s
);
1438 void kvm_irqchip_commit_routes(KVMState
*s
)
1442 if (kvm_gsi_direct_mapping()) {
1446 if (!kvm_gsi_routing_enabled()) {
1450 s
->irq_routes
->flags
= 0;
1451 trace_kvm_irqchip_commit_routes();
1452 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
1456 static void kvm_add_routing_entry(KVMState
*s
,
1457 struct kvm_irq_routing_entry
*entry
)
1459 struct kvm_irq_routing_entry
*new;
1462 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
1463 n
= s
->nr_allocated_irq_routes
* 2;
1467 size
= sizeof(struct kvm_irq_routing
);
1468 size
+= n
* sizeof(*new);
1469 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
1470 s
->nr_allocated_irq_routes
= n
;
1472 n
= s
->irq_routes
->nr
++;
1473 new = &s
->irq_routes
->entries
[n
];
1477 set_gsi(s
, entry
->gsi
);
1480 static int kvm_update_routing_entry(KVMState
*s
,
1481 struct kvm_irq_routing_entry
*new_entry
)
1483 struct kvm_irq_routing_entry
*entry
;
1486 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
1487 entry
= &s
->irq_routes
->entries
[n
];
1488 if (entry
->gsi
!= new_entry
->gsi
) {
1492 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1496 *entry
= *new_entry
;
1504 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1506 struct kvm_irq_routing_entry e
= {};
1508 assert(pin
< s
->gsi_count
);
1511 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1513 e
.u
.irqchip
.irqchip
= irqchip
;
1514 e
.u
.irqchip
.pin
= pin
;
1515 kvm_add_routing_entry(s
, &e
);
1518 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1520 struct kvm_irq_routing_entry
*e
;
1523 if (kvm_gsi_direct_mapping()) {
1527 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1528 e
= &s
->irq_routes
->entries
[i
];
1529 if (e
->gsi
== virq
) {
1530 s
->irq_routes
->nr
--;
1531 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1535 kvm_arch_release_virq_post(virq
);
1536 trace_kvm_irqchip_release_virq(virq
);
1539 void kvm_irqchip_add_change_notifier(Notifier
*n
)
1541 notifier_list_add(&kvm_irqchip_change_notifiers
, n
);
1544 void kvm_irqchip_remove_change_notifier(Notifier
*n
)
1549 void kvm_irqchip_change_notify(void)
1551 notifier_list_notify(&kvm_irqchip_change_notifiers
, NULL
);
1554 static unsigned int kvm_hash_msi(uint32_t data
)
1556 /* This is optimized for IA32 MSI layout. However, no other arch shall
1557 * repeat the mistake of not providing a direct MSI injection API. */
1561 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1563 KVMMSIRoute
*route
, *next
;
1566 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1567 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1568 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1569 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1575 static int kvm_irqchip_get_virq(KVMState
*s
)
1580 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1581 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1582 * number can succeed even though a new route entry cannot be added.
1583 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1585 if (!kvm_direct_msi_allowed
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1586 kvm_flush_dynamic_msi_routes(s
);
1589 /* Return the lowest unused GSI in the bitmap */
1590 next_virq
= find_first_zero_bit(s
->used_gsi_bitmap
, s
->gsi_count
);
1591 if (next_virq
>= s
->gsi_count
) {
1598 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1600 unsigned int hash
= kvm_hash_msi(msg
.data
);
1603 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1604 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1605 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1606 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1613 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1618 if (kvm_direct_msi_allowed
) {
1619 msi
.address_lo
= (uint32_t)msg
.address
;
1620 msi
.address_hi
= msg
.address
>> 32;
1621 msi
.data
= le32_to_cpu(msg
.data
);
1623 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1625 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1628 route
= kvm_lookup_msi_route(s
, msg
);
1632 virq
= kvm_irqchip_get_virq(s
);
1637 route
= g_malloc0(sizeof(KVMMSIRoute
));
1638 route
->kroute
.gsi
= virq
;
1639 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1640 route
->kroute
.flags
= 0;
1641 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1642 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1643 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1645 kvm_add_routing_entry(s
, &route
->kroute
);
1646 kvm_irqchip_commit_routes(s
);
1648 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1652 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1654 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
1657 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1659 struct kvm_irq_routing_entry kroute
= {};
1661 MSIMessage msg
= {0, 0};
1663 if (pci_available
&& dev
) {
1664 msg
= pci_get_msi_message(dev
, vector
);
1667 if (kvm_gsi_direct_mapping()) {
1668 return kvm_arch_msi_data_to_gsi(msg
.data
);
1671 if (!kvm_gsi_routing_enabled()) {
1675 virq
= kvm_irqchip_get_virq(s
);
1681 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1683 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1684 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1685 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1686 if (pci_available
&& kvm_msi_devid_required()) {
1687 kroute
.flags
= KVM_MSI_VALID_DEVID
;
1688 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
1690 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1691 kvm_irqchip_release_virq(s
, virq
);
1695 trace_kvm_irqchip_add_msi_route(dev
? dev
->name
: (char *)"N/A",
1698 kvm_add_routing_entry(s
, &kroute
);
1699 kvm_arch_add_msi_route_post(&kroute
, vector
, dev
);
1700 kvm_irqchip_commit_routes(s
);
1705 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
,
1708 struct kvm_irq_routing_entry kroute
= {};
1710 if (kvm_gsi_direct_mapping()) {
1714 if (!kvm_irqchip_in_kernel()) {
1719 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1721 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1722 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1723 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1724 if (pci_available
&& kvm_msi_devid_required()) {
1725 kroute
.flags
= KVM_MSI_VALID_DEVID
;
1726 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
1728 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1732 trace_kvm_irqchip_update_msi_route(virq
);
1734 return kvm_update_routing_entry(s
, &kroute
);
1737 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
1738 EventNotifier
*resample
, int virq
,
1741 int fd
= event_notifier_get_fd(event
);
1742 int rfd
= resample
? event_notifier_get_fd(resample
) : -1;
1744 struct kvm_irqfd irqfd
= {
1747 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
1752 if (kvm_irqchip_is_split()) {
1754 * When the slow irqchip (e.g. IOAPIC) is in the
1755 * userspace, KVM kernel resamplefd will not work because
1756 * the EOI of the interrupt will be delivered to userspace
1757 * instead, so the KVM kernel resamplefd kick will be
1758 * skipped. The userspace here mimics what the kernel
1759 * provides with resamplefd, remember the resamplefd and
1760 * kick it when we receive EOI of this IRQ.
1762 * This is hackery because IOAPIC is mostly bypassed
1763 * (except EOI broadcasts) when irqfd is used. However
1764 * this can bring much performance back for split irqchip
1765 * with INTx IRQs (for VFIO, this gives 93% perf of the
1766 * full fast path, which is 46% perf boost comparing to
1767 * the INTx slow path).
1769 kvm_resample_fd_insert(virq
, resample
);
1771 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
1772 irqfd
.resamplefd
= rfd
;
1774 } else if (!assign
) {
1775 if (kvm_irqchip_is_split()) {
1776 kvm_resample_fd_remove(virq
);
1780 if (!kvm_irqfds_enabled()) {
1784 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
1787 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1789 struct kvm_irq_routing_entry kroute
= {};
1792 if (!kvm_gsi_routing_enabled()) {
1796 virq
= kvm_irqchip_get_virq(s
);
1802 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
1804 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
1805 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
1806 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
1807 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
1808 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
1810 kvm_add_routing_entry(s
, &kroute
);
1815 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1817 struct kvm_irq_routing_entry kroute
= {};
1820 if (!kvm_gsi_routing_enabled()) {
1823 if (!kvm_check_extension(s
, KVM_CAP_HYPERV_SYNIC
)) {
1826 virq
= kvm_irqchip_get_virq(s
);
1832 kroute
.type
= KVM_IRQ_ROUTING_HV_SINT
;
1834 kroute
.u
.hv_sint
.vcpu
= vcpu
;
1835 kroute
.u
.hv_sint
.sint
= sint
;
1837 kvm_add_routing_entry(s
, &kroute
);
1838 kvm_irqchip_commit_routes(s
);
1843 #else /* !KVM_CAP_IRQ_ROUTING */
1845 void kvm_init_irq_routing(KVMState
*s
)
1849 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1853 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1858 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1863 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1868 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1873 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
1874 EventNotifier
*resample
, int virq
,
1880 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
1884 #endif /* !KVM_CAP_IRQ_ROUTING */
1886 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1887 EventNotifier
*rn
, int virq
)
1889 return kvm_irqchip_assign_irqfd(s
, n
, rn
, virq
, true);
1892 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1895 return kvm_irqchip_assign_irqfd(s
, n
, NULL
, virq
, false);
1898 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1899 EventNotifier
*rn
, qemu_irq irq
)
1902 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1907 return kvm_irqchip_add_irqfd_notifier_gsi(s
, n
, rn
, GPOINTER_TO_INT(gsi
));
1910 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1914 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1919 return kvm_irqchip_remove_irqfd_notifier_gsi(s
, n
, GPOINTER_TO_INT(gsi
));
1922 void kvm_irqchip_set_qemuirq_gsi(KVMState
*s
, qemu_irq irq
, int gsi
)
1924 g_hash_table_insert(s
->gsimap
, irq
, GINT_TO_POINTER(gsi
));
1927 static void kvm_irqchip_create(KVMState
*s
)
1931 assert(s
->kernel_irqchip_split
!= ON_OFF_AUTO_AUTO
);
1932 if (kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
1934 } else if (kvm_check_extension(s
, KVM_CAP_S390_IRQCHIP
)) {
1935 ret
= kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0);
1937 fprintf(stderr
, "Enable kernel irqchip failed: %s\n", strerror(-ret
));
1944 /* First probe and see if there's a arch-specific hook to create the
1945 * in-kernel irqchip for us */
1946 ret
= kvm_arch_irqchip_create(s
);
1948 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_ON
) {
1949 perror("Split IRQ chip mode not supported.");
1952 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
1956 fprintf(stderr
, "Create kernel irqchip failed: %s\n", strerror(-ret
));
1960 kvm_kernel_irqchip
= true;
1961 /* If we have an in-kernel IRQ chip then we must have asynchronous
1962 * interrupt delivery (though the reverse is not necessarily true)
1964 kvm_async_interrupts_allowed
= true;
1965 kvm_halt_in_kernel_allowed
= true;
1967 kvm_init_irq_routing(s
);
1969 s
->gsimap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
1972 /* Find number of supported CPUs using the recommended
1973 * procedure from the kernel API documentation to cope with
1974 * older kernels that may be missing capabilities.
1976 static int kvm_recommended_vcpus(KVMState
*s
)
1978 int ret
= kvm_vm_check_extension(s
, KVM_CAP_NR_VCPUS
);
1979 return (ret
) ? ret
: 4;
1982 static int kvm_max_vcpus(KVMState
*s
)
1984 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
1985 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
1988 static int kvm_max_vcpu_id(KVMState
*s
)
1990 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPU_ID
);
1991 return (ret
) ? ret
: kvm_max_vcpus(s
);
1994 bool kvm_vcpu_id_is_valid(int vcpu_id
)
1996 KVMState
*s
= KVM_STATE(current_accel());
1997 return vcpu_id
>= 0 && vcpu_id
< kvm_max_vcpu_id(s
);
2000 static int kvm_init(MachineState
*ms
)
2002 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
2003 static const char upgrade_note
[] =
2004 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
2005 "(see http://sourceforge.net/projects/kvm).\n";
2010 { "SMP", ms
->smp
.cpus
},
2011 { "hotpluggable", ms
->smp
.max_cpus
},
2014 int soft_vcpus_limit
, hard_vcpus_limit
;
2016 const KVMCapabilityInfo
*missing_cap
;
2019 uint64_t dirty_log_manual_caps
;
2021 s
= KVM_STATE(ms
->accelerator
);
2024 * On systems where the kernel can support different base page
2025 * sizes, host page size may be different from TARGET_PAGE_SIZE,
2026 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
2027 * page size for the system though.
2029 assert(TARGET_PAGE_SIZE
<= qemu_real_host_page_size
);
2033 #ifdef KVM_CAP_SET_GUEST_DEBUG
2034 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
2036 QLIST_INIT(&s
->kvm_parked_vcpus
);
2038 s
->fd
= qemu_open_old("/dev/kvm", O_RDWR
);
2040 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
2045 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
2046 if (ret
< KVM_API_VERSION
) {
2050 fprintf(stderr
, "kvm version too old\n");
2054 if (ret
> KVM_API_VERSION
) {
2056 fprintf(stderr
, "kvm version not supported\n");
2060 kvm_immediate_exit
= kvm_check_extension(s
, KVM_CAP_IMMEDIATE_EXIT
);
2061 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
2063 /* If unspecified, use the default value */
2068 s
->nr_as
= kvm_check_extension(s
, KVM_CAP_MULTI_ADDRESS_SPACE
);
2069 if (s
->nr_as
<= 1) {
2072 s
->as
= g_new0(struct KVMAs
, s
->nr_as
);
2074 if (object_property_find(OBJECT(current_machine
), "kvm-type")) {
2075 g_autofree
char *kvm_type
= object_property_get_str(OBJECT(current_machine
),
2078 type
= mc
->kvm_type(ms
, kvm_type
);
2082 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
2083 } while (ret
== -EINTR
);
2086 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
2090 if (ret
== -EINVAL
) {
2092 "Host kernel setup problem detected. Please verify:\n");
2093 fprintf(stderr
, "- for kernels supporting the switch_amode or"
2094 " user_mode parameters, whether\n");
2096 " user space is running in primary address space\n");
2098 "- for kernels supporting the vm.allocate_pgste sysctl, "
2099 "whether it is enabled\n");
2107 /* check the vcpu limits */
2108 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
2109 hard_vcpus_limit
= kvm_max_vcpus(s
);
2112 if (nc
->num
> soft_vcpus_limit
) {
2113 warn_report("Number of %s cpus requested (%d) exceeds "
2114 "the recommended cpus supported by KVM (%d)",
2115 nc
->name
, nc
->num
, soft_vcpus_limit
);
2117 if (nc
->num
> hard_vcpus_limit
) {
2118 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
2119 "the maximum cpus supported by KVM (%d)\n",
2120 nc
->name
, nc
->num
, hard_vcpus_limit
);
2127 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
2130 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
2134 fprintf(stderr
, "kvm does not support %s\n%s",
2135 missing_cap
->name
, upgrade_note
);
2139 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
2140 s
->coalesced_pio
= s
->coalesced_mmio
&&
2141 kvm_check_extension(s
, KVM_CAP_COALESCED_PIO
);
2143 dirty_log_manual_caps
=
2144 kvm_check_extension(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
);
2145 dirty_log_manual_caps
&= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE
|
2146 KVM_DIRTY_LOG_INITIALLY_SET
);
2147 s
->manual_dirty_log_protect
= dirty_log_manual_caps
;
2148 if (dirty_log_manual_caps
) {
2149 ret
= kvm_vm_enable_cap(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
, 0,
2150 dirty_log_manual_caps
);
2152 warn_report("Trying to enable capability %"PRIu64
" of "
2153 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2154 "Falling back to the legacy mode. ",
2155 dirty_log_manual_caps
);
2156 s
->manual_dirty_log_protect
= 0;
2160 #ifdef KVM_CAP_VCPU_EVENTS
2161 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
2164 s
->robust_singlestep
=
2165 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
2167 #ifdef KVM_CAP_DEBUGREGS
2168 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
2171 s
->max_nested_state_len
= kvm_check_extension(s
, KVM_CAP_NESTED_STATE
);
2173 #ifdef KVM_CAP_IRQ_ROUTING
2174 kvm_direct_msi_allowed
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
2177 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
2179 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
2180 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
2181 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
2184 kvm_readonly_mem_allowed
=
2185 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
2187 kvm_eventfds_allowed
=
2188 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
2190 kvm_irqfds_allowed
=
2191 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
2193 kvm_resamplefds_allowed
=
2194 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
2196 kvm_vm_attributes_allowed
=
2197 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
2199 kvm_ioeventfd_any_length_allowed
=
2200 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD_ANY_LENGTH
) > 0);
2205 * if memory encryption object is specified then initialize the memory
2206 * encryption context.
2208 if (ms
->memory_encryption
) {
2209 kvm_state
->memcrypt_handle
= sev_guest_init(ms
->memory_encryption
);
2210 if (!kvm_state
->memcrypt_handle
) {
2215 kvm_state
->memcrypt_encrypt_data
= sev_encrypt_data
;
2218 ret
= kvm_arch_init(ms
, s
);
2223 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_AUTO
) {
2224 s
->kernel_irqchip_split
= mc
->default_kernel_irqchip_split
? ON_OFF_AUTO_ON
: ON_OFF_AUTO_OFF
;
2227 qemu_register_reset(kvm_unpoison_all
, NULL
);
2229 if (s
->kernel_irqchip_allowed
) {
2230 kvm_irqchip_create(s
);
2233 if (kvm_eventfds_allowed
) {
2234 s
->memory_listener
.listener
.eventfd_add
= kvm_mem_ioeventfd_add
;
2235 s
->memory_listener
.listener
.eventfd_del
= kvm_mem_ioeventfd_del
;
2237 s
->memory_listener
.listener
.coalesced_io_add
= kvm_coalesce_mmio_region
;
2238 s
->memory_listener
.listener
.coalesced_io_del
= kvm_uncoalesce_mmio_region
;
2240 kvm_memory_listener_register(s
, &s
->memory_listener
,
2241 &address_space_memory
, 0);
2242 if (kvm_eventfds_allowed
) {
2243 memory_listener_register(&kvm_io_listener
,
2246 memory_listener_register(&kvm_coalesced_pio_listener
,
2249 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
2251 s
->sync_mmu
= !!kvm_vm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
2253 ret
= ram_block_discard_disable(true);
2257 cpus_register_accel(&kvm_cpus
);
2268 g_free(s
->memory_listener
.slots
);
2273 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
2275 s
->sigmask_len
= sigmask_len
;
2278 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
2279 int size
, uint32_t count
)
2282 uint8_t *ptr
= data
;
2284 for (i
= 0; i
< count
; i
++) {
2285 address_space_rw(&address_space_io
, port
, attrs
,
2287 direction
== KVM_EXIT_IO_OUT
);
2292 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
2294 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
2295 run
->internal
.suberror
);
2297 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
2300 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
2301 fprintf(stderr
, "extra data[%d]: %"PRIx64
"\n",
2302 i
, (uint64_t)run
->internal
.data
[i
]);
2305 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
2306 fprintf(stderr
, "emulation failure\n");
2307 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
2308 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
2309 return EXCP_INTERRUPT
;
2312 /* FIXME: Should trigger a qmp message to let management know
2313 * something went wrong.
2318 void kvm_flush_coalesced_mmio_buffer(void)
2320 KVMState
*s
= kvm_state
;
2322 if (s
->coalesced_flush_in_progress
) {
2326 s
->coalesced_flush_in_progress
= true;
2328 if (s
->coalesced_mmio_ring
) {
2329 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
2330 while (ring
->first
!= ring
->last
) {
2331 struct kvm_coalesced_mmio
*ent
;
2333 ent
= &ring
->coalesced_mmio
[ring
->first
];
2335 if (ent
->pio
== 1) {
2336 address_space_write(&address_space_io
, ent
->phys_addr
,
2337 MEMTXATTRS_UNSPECIFIED
, ent
->data
,
2340 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
2343 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
2347 s
->coalesced_flush_in_progress
= false;
2350 static void do_kvm_cpu_synchronize_state(CPUState
*cpu
, run_on_cpu_data arg
)
2352 if (!cpu
->vcpu_dirty
) {
2353 kvm_arch_get_registers(cpu
);
2354 cpu
->vcpu_dirty
= true;
2358 void kvm_cpu_synchronize_state(CPUState
*cpu
)
2360 if (!cpu
->vcpu_dirty
) {
2361 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, RUN_ON_CPU_NULL
);
2365 static void do_kvm_cpu_synchronize_post_reset(CPUState
*cpu
, run_on_cpu_data arg
)
2367 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
2368 cpu
->vcpu_dirty
= false;
2371 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
2373 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, RUN_ON_CPU_NULL
);
2376 static void do_kvm_cpu_synchronize_post_init(CPUState
*cpu
, run_on_cpu_data arg
)
2378 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
2379 cpu
->vcpu_dirty
= false;
2382 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
2384 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, RUN_ON_CPU_NULL
);
2387 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
, run_on_cpu_data arg
)
2389 cpu
->vcpu_dirty
= true;
2392 void kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
)
2394 run_on_cpu(cpu
, do_kvm_cpu_synchronize_pre_loadvm
, RUN_ON_CPU_NULL
);
2397 #ifdef KVM_HAVE_MCE_INJECTION
2398 static __thread
void *pending_sigbus_addr
;
2399 static __thread
int pending_sigbus_code
;
2400 static __thread
bool have_sigbus_pending
;
2403 static void kvm_cpu_kick(CPUState
*cpu
)
2405 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 1);
2408 static void kvm_cpu_kick_self(void)
2410 if (kvm_immediate_exit
) {
2411 kvm_cpu_kick(current_cpu
);
2413 qemu_cpu_kick_self();
2417 static void kvm_eat_signals(CPUState
*cpu
)
2419 struct timespec ts
= { 0, 0 };
2425 if (kvm_immediate_exit
) {
2426 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 0);
2427 /* Write kvm_run->immediate_exit before the cpu->exit_request
2428 * write in kvm_cpu_exec.
2434 sigemptyset(&waitset
);
2435 sigaddset(&waitset
, SIG_IPI
);
2438 r
= sigtimedwait(&waitset
, &siginfo
, &ts
);
2439 if (r
== -1 && !(errno
== EAGAIN
|| errno
== EINTR
)) {
2440 perror("sigtimedwait");
2444 r
= sigpending(&chkset
);
2446 perror("sigpending");
2449 } while (sigismember(&chkset
, SIG_IPI
));
2452 int kvm_cpu_exec(CPUState
*cpu
)
2454 struct kvm_run
*run
= cpu
->kvm_run
;
2457 DPRINTF("kvm_cpu_exec()\n");
2459 if (kvm_arch_process_async_events(cpu
)) {
2460 qatomic_set(&cpu
->exit_request
, 0);
2464 qemu_mutex_unlock_iothread();
2465 cpu_exec_start(cpu
);
2470 if (cpu
->vcpu_dirty
) {
2471 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
2472 cpu
->vcpu_dirty
= false;
2475 kvm_arch_pre_run(cpu
, run
);
2476 if (qatomic_read(&cpu
->exit_request
)) {
2477 DPRINTF("interrupt exit requested\n");
2479 * KVM requires us to reenter the kernel after IO exits to complete
2480 * instruction emulation. This self-signal will ensure that we
2483 kvm_cpu_kick_self();
2486 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2487 * Matching barrier in kvm_eat_signals.
2491 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
2493 attrs
= kvm_arch_post_run(cpu
, run
);
2495 #ifdef KVM_HAVE_MCE_INJECTION
2496 if (unlikely(have_sigbus_pending
)) {
2497 qemu_mutex_lock_iothread();
2498 kvm_arch_on_sigbus_vcpu(cpu
, pending_sigbus_code
,
2499 pending_sigbus_addr
);
2500 have_sigbus_pending
= false;
2501 qemu_mutex_unlock_iothread();
2506 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
2507 DPRINTF("io window exit\n");
2508 kvm_eat_signals(cpu
);
2509 ret
= EXCP_INTERRUPT
;
2512 fprintf(stderr
, "error: kvm run failed %s\n",
2513 strerror(-run_ret
));
2515 if (run_ret
== -EBUSY
) {
2517 "This is probably because your SMT is enabled.\n"
2518 "VCPU can only run on primary threads with all "
2519 "secondary threads offline.\n");
2526 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
2527 switch (run
->exit_reason
) {
2529 DPRINTF("handle_io\n");
2530 /* Called outside BQL */
2531 kvm_handle_io(run
->io
.port
, attrs
,
2532 (uint8_t *)run
+ run
->io
.data_offset
,
2539 DPRINTF("handle_mmio\n");
2540 /* Called outside BQL */
2541 address_space_rw(&address_space_memory
,
2542 run
->mmio
.phys_addr
, attrs
,
2545 run
->mmio
.is_write
);
2548 case KVM_EXIT_IRQ_WINDOW_OPEN
:
2549 DPRINTF("irq_window_open\n");
2550 ret
= EXCP_INTERRUPT
;
2552 case KVM_EXIT_SHUTDOWN
:
2553 DPRINTF("shutdown\n");
2554 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
2555 ret
= EXCP_INTERRUPT
;
2557 case KVM_EXIT_UNKNOWN
:
2558 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
2559 (uint64_t)run
->hw
.hardware_exit_reason
);
2562 case KVM_EXIT_INTERNAL_ERROR
:
2563 ret
= kvm_handle_internal_error(cpu
, run
);
2565 case KVM_EXIT_SYSTEM_EVENT
:
2566 switch (run
->system_event
.type
) {
2567 case KVM_SYSTEM_EVENT_SHUTDOWN
:
2568 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN
);
2569 ret
= EXCP_INTERRUPT
;
2571 case KVM_SYSTEM_EVENT_RESET
:
2572 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
2573 ret
= EXCP_INTERRUPT
;
2575 case KVM_SYSTEM_EVENT_CRASH
:
2576 kvm_cpu_synchronize_state(cpu
);
2577 qemu_mutex_lock_iothread();
2578 qemu_system_guest_panicked(cpu_get_crash_info(cpu
));
2579 qemu_mutex_unlock_iothread();
2583 DPRINTF("kvm_arch_handle_exit\n");
2584 ret
= kvm_arch_handle_exit(cpu
, run
);
2589 DPRINTF("kvm_arch_handle_exit\n");
2590 ret
= kvm_arch_handle_exit(cpu
, run
);
2596 qemu_mutex_lock_iothread();
2599 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
2600 vm_stop(RUN_STATE_INTERNAL_ERROR
);
2603 qatomic_set(&cpu
->exit_request
, 0);
2607 int kvm_ioctl(KVMState
*s
, int type
, ...)
2614 arg
= va_arg(ap
, void *);
2617 trace_kvm_ioctl(type
, arg
);
2618 ret
= ioctl(s
->fd
, type
, arg
);
2625 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
2632 arg
= va_arg(ap
, void *);
2635 trace_kvm_vm_ioctl(type
, arg
);
2636 ret
= ioctl(s
->vmfd
, type
, arg
);
2643 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
2650 arg
= va_arg(ap
, void *);
2653 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
2654 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
2661 int kvm_device_ioctl(int fd
, int type
, ...)
2668 arg
= va_arg(ap
, void *);
2671 trace_kvm_device_ioctl(fd
, type
, arg
);
2672 ret
= ioctl(fd
, type
, arg
);
2679 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
2682 struct kvm_device_attr attribute
= {
2687 if (!kvm_vm_attributes_allowed
) {
2691 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
2692 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2696 int kvm_device_check_attr(int dev_fd
, uint32_t group
, uint64_t attr
)
2698 struct kvm_device_attr attribute
= {
2704 return kvm_device_ioctl(dev_fd
, KVM_HAS_DEVICE_ATTR
, &attribute
) ? 0 : 1;
2707 int kvm_device_access(int fd
, int group
, uint64_t attr
,
2708 void *val
, bool write
, Error
**errp
)
2710 struct kvm_device_attr kvmattr
;
2714 kvmattr
.group
= group
;
2715 kvmattr
.attr
= attr
;
2716 kvmattr
.addr
= (uintptr_t)val
;
2718 err
= kvm_device_ioctl(fd
,
2719 write
? KVM_SET_DEVICE_ATTR
: KVM_GET_DEVICE_ATTR
,
2722 error_setg_errno(errp
, -err
,
2723 "KVM_%s_DEVICE_ATTR failed: Group %d "
2724 "attr 0x%016" PRIx64
,
2725 write
? "SET" : "GET", group
, attr
);
2730 bool kvm_has_sync_mmu(void)
2732 return kvm_state
->sync_mmu
;
2735 int kvm_has_vcpu_events(void)
2737 return kvm_state
->vcpu_events
;
2740 int kvm_has_robust_singlestep(void)
2742 return kvm_state
->robust_singlestep
;
2745 int kvm_has_debugregs(void)
2747 return kvm_state
->debugregs
;
2750 int kvm_max_nested_state_length(void)
2752 return kvm_state
->max_nested_state_len
;
2755 int kvm_has_many_ioeventfds(void)
2757 if (!kvm_enabled()) {
2760 return kvm_state
->many_ioeventfds
;
2763 int kvm_has_gsi_routing(void)
2765 #ifdef KVM_CAP_IRQ_ROUTING
2766 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
2772 int kvm_has_intx_set_mask(void)
2774 return kvm_state
->intx_set_mask
;
2777 bool kvm_arm_supports_user_irq(void)
2779 return kvm_check_extension(kvm_state
, KVM_CAP_ARM_USER_IRQ
);
2782 #ifdef KVM_CAP_SET_GUEST_DEBUG
2783 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
2786 struct kvm_sw_breakpoint
*bp
;
2788 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
2796 int kvm_sw_breakpoints_active(CPUState
*cpu
)
2798 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
2801 struct kvm_set_guest_debug_data
{
2802 struct kvm_guest_debug dbg
;
2806 static void kvm_invoke_set_guest_debug(CPUState
*cpu
, run_on_cpu_data data
)
2808 struct kvm_set_guest_debug_data
*dbg_data
=
2809 (struct kvm_set_guest_debug_data
*) data
.host_ptr
;
2811 dbg_data
->err
= kvm_vcpu_ioctl(cpu
, KVM_SET_GUEST_DEBUG
,
2815 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2817 struct kvm_set_guest_debug_data data
;
2819 data
.dbg
.control
= reinject_trap
;
2821 if (cpu
->singlestep_enabled
) {
2822 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
2824 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
2826 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
,
2827 RUN_ON_CPU_HOST_PTR(&data
));
2831 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2832 target_ulong len
, int type
)
2834 struct kvm_sw_breakpoint
*bp
;
2837 if (type
== GDB_BREAKPOINT_SW
) {
2838 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2844 bp
= g_malloc(sizeof(struct kvm_sw_breakpoint
));
2847 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
2853 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2855 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
2862 err
= kvm_update_guest_debug(cpu
, 0);
2870 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2871 target_ulong len
, int type
)
2873 struct kvm_sw_breakpoint
*bp
;
2876 if (type
== GDB_BREAKPOINT_SW
) {
2877 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2882 if (bp
->use_count
> 1) {
2887 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
2892 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2895 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
2902 err
= kvm_update_guest_debug(cpu
, 0);
2910 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2912 struct kvm_sw_breakpoint
*bp
, *next
;
2913 KVMState
*s
= cpu
->kvm_state
;
2916 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
2917 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
2918 /* Try harder to find a CPU that currently sees the breakpoint. */
2919 CPU_FOREACH(tmpcpu
) {
2920 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
2925 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
2928 kvm_arch_remove_all_hw_breakpoints();
2931 kvm_update_guest_debug(cpu
, 0);
2935 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2937 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2942 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2943 target_ulong len
, int type
)
2948 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2949 target_ulong len
, int type
)
2954 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2957 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2959 static int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
2961 KVMState
*s
= kvm_state
;
2962 struct kvm_signal_mask
*sigmask
;
2965 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
2967 sigmask
->len
= s
->sigmask_len
;
2968 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
2969 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
2975 static void kvm_ipi_signal(int sig
)
2978 assert(kvm_immediate_exit
);
2979 kvm_cpu_kick(current_cpu
);
2983 void kvm_init_cpu_signals(CPUState
*cpu
)
2987 struct sigaction sigact
;
2989 memset(&sigact
, 0, sizeof(sigact
));
2990 sigact
.sa_handler
= kvm_ipi_signal
;
2991 sigaction(SIG_IPI
, &sigact
, NULL
);
2993 pthread_sigmask(SIG_BLOCK
, NULL
, &set
);
2994 #if defined KVM_HAVE_MCE_INJECTION
2995 sigdelset(&set
, SIGBUS
);
2996 pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
2998 sigdelset(&set
, SIG_IPI
);
2999 if (kvm_immediate_exit
) {
3000 r
= pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
3002 r
= kvm_set_signal_mask(cpu
, &set
);
3005 fprintf(stderr
, "kvm_set_signal_mask: %s\n", strerror(-r
));
3010 /* Called asynchronously in VCPU thread. */
3011 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
3013 #ifdef KVM_HAVE_MCE_INJECTION
3014 if (have_sigbus_pending
) {
3017 have_sigbus_pending
= true;
3018 pending_sigbus_addr
= addr
;
3019 pending_sigbus_code
= code
;
3020 qatomic_set(&cpu
->exit_request
, 1);
3027 /* Called synchronously (via signalfd) in main thread. */
3028 int kvm_on_sigbus(int code
, void *addr
)
3030 #ifdef KVM_HAVE_MCE_INJECTION
3031 /* Action required MCE kills the process if SIGBUS is blocked. Because
3032 * that's what happens in the I/O thread, where we handle MCE via signalfd,
3033 * we can only get action optional here.
3035 assert(code
!= BUS_MCEERR_AR
);
3036 kvm_arch_on_sigbus_vcpu(first_cpu
, code
, addr
);
3043 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
3046 struct kvm_create_device create_dev
;
3048 create_dev
.type
= type
;
3050 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
3052 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
3056 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
3061 return test
? 0 : create_dev
.fd
;
3064 bool kvm_device_supported(int vmfd
, uint64_t type
)
3066 struct kvm_create_device create_dev
= {
3069 .flags
= KVM_CREATE_DEVICE_TEST
,
3072 if (ioctl(vmfd
, KVM_CHECK_EXTENSION
, KVM_CAP_DEVICE_CTRL
) <= 0) {
3076 return (ioctl(vmfd
, KVM_CREATE_DEVICE
, &create_dev
) >= 0);
3079 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
3081 struct kvm_one_reg reg
;
3085 reg
.addr
= (uintptr_t) source
;
3086 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
3088 trace_kvm_failed_reg_set(id
, strerror(-r
));
3093 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
3095 struct kvm_one_reg reg
;
3099 reg
.addr
= (uintptr_t) target
;
3100 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
3102 trace_kvm_failed_reg_get(id
, strerror(-r
));
3107 static bool kvm_accel_has_memory(MachineState
*ms
, AddressSpace
*as
,
3108 hwaddr start_addr
, hwaddr size
)
3110 KVMState
*kvm
= KVM_STATE(ms
->accelerator
);
3113 for (i
= 0; i
< kvm
->nr_as
; ++i
) {
3114 if (kvm
->as
[i
].as
== as
&& kvm
->as
[i
].ml
) {
3115 size
= MIN(kvm_max_slot_size
, size
);
3116 return NULL
!= kvm_lookup_matching_slot(kvm
->as
[i
].ml
,
3124 static void kvm_get_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3125 const char *name
, void *opaque
,
3128 KVMState
*s
= KVM_STATE(obj
);
3129 int64_t value
= s
->kvm_shadow_mem
;
3131 visit_type_int(v
, name
, &value
, errp
);
3134 static void kvm_set_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3135 const char *name
, void *opaque
,
3138 KVMState
*s
= KVM_STATE(obj
);
3141 if (!visit_type_int(v
, name
, &value
, errp
)) {
3145 s
->kvm_shadow_mem
= value
;
3148 static void kvm_set_kernel_irqchip(Object
*obj
, Visitor
*v
,
3149 const char *name
, void *opaque
,
3152 KVMState
*s
= KVM_STATE(obj
);
3155 if (!visit_type_OnOffSplit(v
, name
, &mode
, errp
)) {
3159 case ON_OFF_SPLIT_ON
:
3160 s
->kernel_irqchip_allowed
= true;
3161 s
->kernel_irqchip_required
= true;
3162 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3164 case ON_OFF_SPLIT_OFF
:
3165 s
->kernel_irqchip_allowed
= false;
3166 s
->kernel_irqchip_required
= false;
3167 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3169 case ON_OFF_SPLIT_SPLIT
:
3170 s
->kernel_irqchip_allowed
= true;
3171 s
->kernel_irqchip_required
= true;
3172 s
->kernel_irqchip_split
= ON_OFF_AUTO_ON
;
3175 /* The value was checked in visit_type_OnOffSplit() above. If
3176 * we get here, then something is wrong in QEMU.
3182 bool kvm_kernel_irqchip_allowed(void)
3184 return kvm_state
->kernel_irqchip_allowed
;
3187 bool kvm_kernel_irqchip_required(void)
3189 return kvm_state
->kernel_irqchip_required
;
3192 bool kvm_kernel_irqchip_split(void)
3194 return kvm_state
->kernel_irqchip_split
== ON_OFF_AUTO_ON
;
3197 static void kvm_accel_instance_init(Object
*obj
)
3199 KVMState
*s
= KVM_STATE(obj
);
3201 s
->kvm_shadow_mem
= -1;
3202 s
->kernel_irqchip_allowed
= true;
3203 s
->kernel_irqchip_split
= ON_OFF_AUTO_AUTO
;
3206 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
3208 AccelClass
*ac
= ACCEL_CLASS(oc
);
3210 ac
->init_machine
= kvm_init
;
3211 ac
->has_memory
= kvm_accel_has_memory
;
3212 ac
->allowed
= &kvm_allowed
;
3214 object_class_property_add(oc
, "kernel-irqchip", "on|off|split",
3215 NULL
, kvm_set_kernel_irqchip
,
3217 object_class_property_set_description(oc
, "kernel-irqchip",
3218 "Configure KVM in-kernel irqchip");
3220 object_class_property_add(oc
, "kvm-shadow-mem", "int",
3221 kvm_get_kvm_shadow_mem
, kvm_set_kvm_shadow_mem
,
3223 object_class_property_set_description(oc
, "kvm-shadow-mem",
3224 "KVM shadow MMU size");
3227 static const TypeInfo kvm_accel_type
= {
3228 .name
= TYPE_KVM_ACCEL
,
3229 .parent
= TYPE_ACCEL
,
3230 .instance_init
= kvm_accel_instance_init
,
3231 .class_init
= kvm_accel_class_init
,
3232 .instance_size
= sizeof(KVMState
),
3235 static void kvm_type_init(void)
3237 type_register_static(&kvm_accel_type
);
3240 type_init(kvm_type_init
);