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.
64 #define PAGE_SIZE qemu_real_host_page_size
69 #define DPRINTF(fmt, ...) \
70 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
72 #define DPRINTF(fmt, ...) \
76 #define KVM_MSI_HASHTAB_SIZE 256
78 struct KVMParkedVcpu
{
79 unsigned long vcpu_id
;
81 QLIST_ENTRY(KVMParkedVcpu
) node
;
86 AccelState parent_obj
;
93 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
94 bool coalesced_flush_in_progress
;
96 int robust_singlestep
;
98 #ifdef KVM_CAP_SET_GUEST_DEBUG
99 QTAILQ_HEAD(, kvm_sw_breakpoint
) kvm_sw_breakpoints
;
101 int max_nested_state_len
;
105 bool kernel_irqchip_allowed
;
106 bool kernel_irqchip_required
;
107 OnOffAuto kernel_irqchip_split
;
109 uint64_t manual_dirty_log_protect
;
110 /* The man page (and posix) say ioctl numbers are signed int, but
111 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
112 * unsigned, and treating them as signed here can break things */
113 unsigned irq_set_ioctl
;
114 unsigned int sigmask_len
;
116 #ifdef KVM_CAP_IRQ_ROUTING
117 struct kvm_irq_routing
*irq_routes
;
118 int nr_allocated_irq_routes
;
119 unsigned long *used_gsi_bitmap
;
120 unsigned int gsi_count
;
121 QTAILQ_HEAD(, KVMMSIRoute
) msi_hashtab
[KVM_MSI_HASHTAB_SIZE
];
123 KVMMemoryListener memory_listener
;
124 QLIST_HEAD(, KVMParkedVcpu
) kvm_parked_vcpus
;
126 /* For "info mtree -f" to tell if an MR is registered in KVM */
129 KVMMemoryListener
*ml
;
135 bool kvm_kernel_irqchip
;
136 bool kvm_split_irqchip
;
137 bool kvm_async_interrupts_allowed
;
138 bool kvm_halt_in_kernel_allowed
;
139 bool kvm_eventfds_allowed
;
140 bool kvm_irqfds_allowed
;
141 bool kvm_resamplefds_allowed
;
142 bool kvm_msi_via_irqfd_allowed
;
143 bool kvm_gsi_routing_allowed
;
144 bool kvm_gsi_direct_mapping
;
146 bool kvm_readonly_mem_allowed
;
147 bool kvm_vm_attributes_allowed
;
148 bool kvm_direct_msi_allowed
;
149 bool kvm_ioeventfd_any_length_allowed
;
150 bool kvm_msi_use_devid
;
151 static bool kvm_immediate_exit
;
152 static hwaddr kvm_max_slot_size
= ~0;
154 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
155 KVM_CAP_INFO(USER_MEMORY
),
156 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
157 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS
),
161 static NotifierList kvm_irqchip_change_notifiers
=
162 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers
);
164 struct KVMResampleFd
{
166 EventNotifier
*resample_event
;
167 QLIST_ENTRY(KVMResampleFd
) node
;
169 typedef struct KVMResampleFd KVMResampleFd
;
172 * Only used with split irqchip where we need to do the resample fd
173 * kick for the kernel from userspace.
175 static QLIST_HEAD(, KVMResampleFd
) kvm_resample_fd_list
=
176 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list
);
178 #define kvm_slots_lock(kml) qemu_mutex_lock(&(kml)->slots_lock)
179 #define kvm_slots_unlock(kml) qemu_mutex_unlock(&(kml)->slots_lock)
181 static inline void kvm_resample_fd_remove(int gsi
)
185 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
186 if (rfd
->gsi
== gsi
) {
187 QLIST_REMOVE(rfd
, node
);
194 static inline void kvm_resample_fd_insert(int gsi
, EventNotifier
*event
)
196 KVMResampleFd
*rfd
= g_new0(KVMResampleFd
, 1);
199 rfd
->resample_event
= event
;
201 QLIST_INSERT_HEAD(&kvm_resample_fd_list
, rfd
, node
);
204 void kvm_resample_fd_notify(int gsi
)
208 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
209 if (rfd
->gsi
== gsi
) {
210 event_notifier_set(rfd
->resample_event
);
211 trace_kvm_resample_fd_notify(gsi
);
217 int kvm_get_max_memslots(void)
219 KVMState
*s
= KVM_STATE(current_accel());
224 /* Called with KVMMemoryListener.slots_lock held */
225 static KVMSlot
*kvm_get_free_slot(KVMMemoryListener
*kml
)
227 KVMState
*s
= kvm_state
;
230 for (i
= 0; i
< s
->nr_slots
; i
++) {
231 if (kml
->slots
[i
].memory_size
== 0) {
232 return &kml
->slots
[i
];
239 bool kvm_has_free_slot(MachineState
*ms
)
241 KVMState
*s
= KVM_STATE(ms
->accelerator
);
243 KVMMemoryListener
*kml
= &s
->memory_listener
;
246 result
= !!kvm_get_free_slot(kml
);
247 kvm_slots_unlock(kml
);
252 /* Called with KVMMemoryListener.slots_lock held */
253 static KVMSlot
*kvm_alloc_slot(KVMMemoryListener
*kml
)
255 KVMSlot
*slot
= kvm_get_free_slot(kml
);
261 fprintf(stderr
, "%s: no free slot available\n", __func__
);
265 static KVMSlot
*kvm_lookup_matching_slot(KVMMemoryListener
*kml
,
269 KVMState
*s
= kvm_state
;
272 for (i
= 0; i
< s
->nr_slots
; i
++) {
273 KVMSlot
*mem
= &kml
->slots
[i
];
275 if (start_addr
== mem
->start_addr
&& size
== mem
->memory_size
) {
284 * Calculate and align the start address and the size of the section.
285 * Return the size. If the size is 0, the aligned section is empty.
287 static hwaddr
kvm_align_section(MemoryRegionSection
*section
,
290 hwaddr size
= int128_get64(section
->size
);
291 hwaddr delta
, aligned
;
293 /* kvm works in page size chunks, but the function may be called
294 with sub-page size and unaligned start address. Pad the start
295 address to next and truncate size to previous page boundary. */
296 aligned
= ROUND_UP(section
->offset_within_address_space
,
297 qemu_real_host_page_size
);
298 delta
= aligned
- section
->offset_within_address_space
;
304 return (size
- delta
) & qemu_real_host_page_mask
;
307 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
310 KVMMemoryListener
*kml
= &s
->memory_listener
;
314 for (i
= 0; i
< s
->nr_slots
; i
++) {
315 KVMSlot
*mem
= &kml
->slots
[i
];
317 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
318 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
323 kvm_slots_unlock(kml
);
328 static int kvm_set_user_memory_region(KVMMemoryListener
*kml
, KVMSlot
*slot
, bool new)
330 KVMState
*s
= kvm_state
;
331 struct kvm_userspace_memory_region mem
;
334 mem
.slot
= slot
->slot
| (kml
->as_id
<< 16);
335 mem
.guest_phys_addr
= slot
->start_addr
;
336 mem
.userspace_addr
= (unsigned long)slot
->ram
;
337 mem
.flags
= slot
->flags
;
339 if (slot
->memory_size
&& !new && (mem
.flags
^ slot
->old_flags
) & KVM_MEM_READONLY
) {
340 /* Set the slot size to 0 before setting the slot to the desired
341 * value. This is needed based on KVM commit 75d61fbc. */
343 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
348 mem
.memory_size
= slot
->memory_size
;
349 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
350 slot
->old_flags
= mem
.flags
;
352 trace_kvm_set_user_memory(mem
.slot
, mem
.flags
, mem
.guest_phys_addr
,
353 mem
.memory_size
, mem
.userspace_addr
, ret
);
355 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
356 " start=0x%" PRIx64
", size=0x%" PRIx64
": %s",
357 __func__
, mem
.slot
, slot
->start_addr
,
358 (uint64_t)mem
.memory_size
, strerror(errno
));
363 static int do_kvm_destroy_vcpu(CPUState
*cpu
)
365 KVMState
*s
= kvm_state
;
367 struct KVMParkedVcpu
*vcpu
= NULL
;
370 DPRINTF("kvm_destroy_vcpu\n");
372 ret
= kvm_arch_destroy_vcpu(cpu
);
377 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
380 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
384 ret
= munmap(cpu
->kvm_run
, mmap_size
);
389 vcpu
= g_malloc0(sizeof(*vcpu
));
390 vcpu
->vcpu_id
= kvm_arch_vcpu_id(cpu
);
391 vcpu
->kvm_fd
= cpu
->kvm_fd
;
392 QLIST_INSERT_HEAD(&kvm_state
->kvm_parked_vcpus
, vcpu
, node
);
397 void kvm_destroy_vcpu(CPUState
*cpu
)
399 if (do_kvm_destroy_vcpu(cpu
) < 0) {
400 error_report("kvm_destroy_vcpu failed");
405 static int kvm_get_vcpu(KVMState
*s
, unsigned long vcpu_id
)
407 struct KVMParkedVcpu
*cpu
;
409 QLIST_FOREACH(cpu
, &s
->kvm_parked_vcpus
, node
) {
410 if (cpu
->vcpu_id
== vcpu_id
) {
413 QLIST_REMOVE(cpu
, node
);
414 kvm_fd
= cpu
->kvm_fd
;
420 return kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)vcpu_id
);
423 int kvm_init_vcpu(CPUState
*cpu
, Error
**errp
)
425 KVMState
*s
= kvm_state
;
429 trace_kvm_init_vcpu(cpu
->cpu_index
, kvm_arch_vcpu_id(cpu
));
431 ret
= kvm_get_vcpu(s
, kvm_arch_vcpu_id(cpu
));
433 error_setg_errno(errp
, -ret
, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)",
434 kvm_arch_vcpu_id(cpu
));
440 cpu
->vcpu_dirty
= true;
442 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
445 error_setg_errno(errp
, -mmap_size
,
446 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
450 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
452 if (cpu
->kvm_run
== MAP_FAILED
) {
454 error_setg_errno(errp
, ret
,
455 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
456 kvm_arch_vcpu_id(cpu
));
460 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
461 s
->coalesced_mmio_ring
=
462 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
465 ret
= kvm_arch_init_vcpu(cpu
);
467 error_setg_errno(errp
, -ret
,
468 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
469 kvm_arch_vcpu_id(cpu
));
476 * dirty pages logging control
479 static int kvm_mem_flags(MemoryRegion
*mr
)
481 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
484 if (memory_region_get_dirty_log_mask(mr
) != 0) {
485 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
487 if (readonly
&& kvm_readonly_mem_allowed
) {
488 flags
|= KVM_MEM_READONLY
;
493 /* Called with KVMMemoryListener.slots_lock held */
494 static int kvm_slot_update_flags(KVMMemoryListener
*kml
, KVMSlot
*mem
,
497 mem
->flags
= kvm_mem_flags(mr
);
499 /* If nothing changed effectively, no need to issue ioctl */
500 if (mem
->flags
== mem
->old_flags
) {
504 return kvm_set_user_memory_region(kml
, mem
, false);
507 static int kvm_section_update_flags(KVMMemoryListener
*kml
,
508 MemoryRegionSection
*section
)
510 hwaddr start_addr
, size
, slot_size
;
514 size
= kvm_align_section(section
, &start_addr
);
521 while (size
&& !ret
) {
522 slot_size
= MIN(kvm_max_slot_size
, size
);
523 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
525 /* We don't have a slot if we want to trap every access. */
529 ret
= kvm_slot_update_flags(kml
, mem
, section
->mr
);
530 start_addr
+= slot_size
;
535 kvm_slots_unlock(kml
);
539 static void kvm_log_start(MemoryListener
*listener
,
540 MemoryRegionSection
*section
,
543 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
550 r
= kvm_section_update_flags(kml
, section
);
556 static void kvm_log_stop(MemoryListener
*listener
,
557 MemoryRegionSection
*section
,
560 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
567 r
= kvm_section_update_flags(kml
, section
);
573 /* get kvm's dirty pages bitmap and update qemu's */
574 static int kvm_get_dirty_pages_log_range(MemoryRegionSection
*section
,
575 unsigned long *bitmap
)
577 ram_addr_t start
= section
->offset_within_region
+
578 memory_region_get_ram_addr(section
->mr
);
579 ram_addr_t pages
= int128_get64(section
->size
) / qemu_real_host_page_size
;
581 cpu_physical_memory_set_dirty_lebitmap(bitmap
, start
, pages
);
585 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
587 /* Allocate the dirty bitmap for a slot */
588 static void kvm_memslot_init_dirty_bitmap(KVMSlot
*mem
)
591 * XXX bad kernel interface alert
592 * For dirty bitmap, kernel allocates array of size aligned to
593 * bits-per-long. But for case when the kernel is 64bits and
594 * the userspace is 32bits, userspace can't align to the same
595 * bits-per-long, since sizeof(long) is different between kernel
596 * and user space. This way, userspace will provide buffer which
597 * may be 4 bytes less than the kernel will use, resulting in
598 * userspace memory corruption (which is not detectable by valgrind
599 * too, in most cases).
600 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
601 * a hope that sizeof(long) won't become >8 any time soon.
603 hwaddr bitmap_size
= ALIGN(((mem
->memory_size
) >> TARGET_PAGE_BITS
),
604 /*HOST_LONG_BITS*/ 64) / 8;
605 mem
->dirty_bmap
= g_malloc0(bitmap_size
);
609 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
611 * This function will first try to fetch dirty bitmap from the kernel,
612 * and then updates qemu's dirty bitmap.
614 * NOTE: caller must be with kml->slots_lock held.
616 * @kml: the KVM memory listener object
617 * @section: the memory section to sync the dirty bitmap with
619 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener
*kml
,
620 MemoryRegionSection
*section
)
622 KVMState
*s
= kvm_state
;
623 struct kvm_dirty_log d
= {};
625 hwaddr start_addr
, size
;
626 hwaddr slot_size
, slot_offset
= 0;
629 size
= kvm_align_section(section
, &start_addr
);
631 MemoryRegionSection subsection
= *section
;
633 slot_size
= MIN(kvm_max_slot_size
, size
);
634 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
636 /* We don't have a slot if we want to trap every access. */
640 if (!mem
->dirty_bmap
) {
641 /* Allocate on the first log_sync, once and for all */
642 kvm_memslot_init_dirty_bitmap(mem
);
645 d
.dirty_bitmap
= mem
->dirty_bmap
;
646 d
.slot
= mem
->slot
| (kml
->as_id
<< 16);
647 if (kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
) == -1) {
648 DPRINTF("ioctl failed %d\n", errno
);
653 subsection
.offset_within_region
+= slot_offset
;
654 subsection
.size
= int128_make64(slot_size
);
655 kvm_get_dirty_pages_log_range(&subsection
, d
.dirty_bitmap
);
657 slot_offset
+= slot_size
;
658 start_addr
+= slot_size
;
665 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
666 #define KVM_CLEAR_LOG_SHIFT 6
667 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size << KVM_CLEAR_LOG_SHIFT)
668 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN)
670 static int kvm_log_clear_one_slot(KVMSlot
*mem
, int as_id
, uint64_t start
,
673 KVMState
*s
= kvm_state
;
674 uint64_t end
, bmap_start
, start_delta
, bmap_npages
;
675 struct kvm_clear_dirty_log d
;
676 unsigned long *bmap_clear
= NULL
, psize
= qemu_real_host_page_size
;
680 * We need to extend either the start or the size or both to
681 * satisfy the KVM interface requirement. Firstly, do the start
682 * page alignment on 64 host pages
684 bmap_start
= start
& KVM_CLEAR_LOG_MASK
;
685 start_delta
= start
- bmap_start
;
689 * The kernel interface has restriction on the size too, that either:
691 * (1) the size is 64 host pages aligned (just like the start), or
692 * (2) the size fills up until the end of the KVM memslot.
694 bmap_npages
= DIV_ROUND_UP(size
+ start_delta
, KVM_CLEAR_LOG_ALIGN
)
695 << KVM_CLEAR_LOG_SHIFT
;
696 end
= mem
->memory_size
/ psize
;
697 if (bmap_npages
> end
- bmap_start
) {
698 bmap_npages
= end
- bmap_start
;
700 start_delta
/= psize
;
703 * Prepare the bitmap to clear dirty bits. Here we must guarantee
704 * that we won't clear any unknown dirty bits otherwise we might
705 * accidentally clear some set bits which are not yet synced from
706 * the kernel into QEMU's bitmap, then we'll lose track of the
707 * guest modifications upon those pages (which can directly lead
708 * to guest data loss or panic after migration).
710 * Layout of the KVMSlot.dirty_bmap:
712 * |<-------- bmap_npages -----------..>|
715 * |----------------|-------------|------------------|------------|
718 * start bmap_start (start) end
719 * of memslot of memslot
721 * [1] bmap_npages can be aligned to either 64 pages or the end of slot
724 assert(bmap_start
% BITS_PER_LONG
== 0);
725 /* We should never do log_clear before log_sync */
726 assert(mem
->dirty_bmap
);
727 if (start_delta
|| bmap_npages
- size
/ psize
) {
728 /* Slow path - we need to manipulate a temp bitmap */
729 bmap_clear
= bitmap_new(bmap_npages
);
730 bitmap_copy_with_src_offset(bmap_clear
, mem
->dirty_bmap
,
731 bmap_start
, start_delta
+ size
/ psize
);
733 * We need to fill the holes at start because that was not
734 * specified by the caller and we extended the bitmap only for
737 bitmap_clear(bmap_clear
, 0, start_delta
);
738 d
.dirty_bitmap
= bmap_clear
;
741 * Fast path - both start and size align well with BITS_PER_LONG
742 * (or the end of memory slot)
744 d
.dirty_bitmap
= mem
->dirty_bmap
+ BIT_WORD(bmap_start
);
747 d
.first_page
= bmap_start
;
748 /* It should never overflow. If it happens, say something */
749 assert(bmap_npages
<= UINT32_MAX
);
750 d
.num_pages
= bmap_npages
;
751 d
.slot
= mem
->slot
| (as_id
<< 16);
753 if (kvm_vm_ioctl(s
, KVM_CLEAR_DIRTY_LOG
, &d
) == -1) {
755 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
756 "start=0x%"PRIx64
", size=0x%"PRIx32
", errno=%d",
757 __func__
, d
.slot
, (uint64_t)d
.first_page
,
758 (uint32_t)d
.num_pages
, ret
);
761 trace_kvm_clear_dirty_log(d
.slot
, d
.first_page
, d
.num_pages
);
765 * After we have updated the remote dirty bitmap, we update the
766 * cached bitmap as well for the memslot, then if another user
767 * clears the same region we know we shouldn't clear it again on
768 * the remote otherwise it's data loss as well.
770 bitmap_clear(mem
->dirty_bmap
, bmap_start
+ start_delta
,
772 /* This handles the NULL case well */
779 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
781 * NOTE: this will be a no-op if we haven't enabled manual dirty log
782 * protection in the host kernel because in that case this operation
783 * will be done within log_sync().
785 * @kml: the kvm memory listener
786 * @section: the memory range to clear dirty bitmap
788 static int kvm_physical_log_clear(KVMMemoryListener
*kml
,
789 MemoryRegionSection
*section
)
791 KVMState
*s
= kvm_state
;
792 uint64_t start
, size
, offset
, count
;
796 if (!s
->manual_dirty_log_protect
) {
797 /* No need to do explicit clear */
801 start
= section
->offset_within_address_space
;
802 size
= int128_get64(section
->size
);
805 /* Nothing more we can do... */
811 for (i
= 0; i
< s
->nr_slots
; i
++) {
812 mem
= &kml
->slots
[i
];
813 /* Discard slots that are empty or do not overlap the section */
814 if (!mem
->memory_size
||
815 mem
->start_addr
> start
+ size
- 1 ||
816 start
> mem
->start_addr
+ mem
->memory_size
- 1) {
820 if (start
>= mem
->start_addr
) {
821 /* The slot starts before section or is aligned to it. */
822 offset
= start
- mem
->start_addr
;
823 count
= MIN(mem
->memory_size
- offset
, size
);
825 /* The slot starts after section. */
827 count
= MIN(mem
->memory_size
, size
- (mem
->start_addr
- start
));
829 ret
= kvm_log_clear_one_slot(mem
, kml
->as_id
, offset
, count
);
835 kvm_slots_unlock(kml
);
840 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
841 MemoryRegionSection
*secion
,
842 hwaddr start
, hwaddr size
)
844 KVMState
*s
= kvm_state
;
846 if (s
->coalesced_mmio
) {
847 struct kvm_coalesced_mmio_zone zone
;
853 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
857 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
858 MemoryRegionSection
*secion
,
859 hwaddr start
, hwaddr size
)
861 KVMState
*s
= kvm_state
;
863 if (s
->coalesced_mmio
) {
864 struct kvm_coalesced_mmio_zone zone
;
870 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
874 static void kvm_coalesce_pio_add(MemoryListener
*listener
,
875 MemoryRegionSection
*section
,
876 hwaddr start
, hwaddr size
)
878 KVMState
*s
= kvm_state
;
880 if (s
->coalesced_pio
) {
881 struct kvm_coalesced_mmio_zone zone
;
887 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
891 static void kvm_coalesce_pio_del(MemoryListener
*listener
,
892 MemoryRegionSection
*section
,
893 hwaddr start
, hwaddr size
)
895 KVMState
*s
= kvm_state
;
897 if (s
->coalesced_pio
) {
898 struct kvm_coalesced_mmio_zone zone
;
904 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
908 static MemoryListener kvm_coalesced_pio_listener
= {
909 .coalesced_io_add
= kvm_coalesce_pio_add
,
910 .coalesced_io_del
= kvm_coalesce_pio_del
,
913 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
917 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
925 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
929 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
931 /* VM wide version not implemented, use global one instead */
932 ret
= kvm_check_extension(s
, extension
);
938 typedef struct HWPoisonPage
{
940 QLIST_ENTRY(HWPoisonPage
) list
;
943 static QLIST_HEAD(, HWPoisonPage
) hwpoison_page_list
=
944 QLIST_HEAD_INITIALIZER(hwpoison_page_list
);
946 static void kvm_unpoison_all(void *param
)
948 HWPoisonPage
*page
, *next_page
;
950 QLIST_FOREACH_SAFE(page
, &hwpoison_page_list
, list
, next_page
) {
951 QLIST_REMOVE(page
, list
);
952 qemu_ram_remap(page
->ram_addr
, TARGET_PAGE_SIZE
);
957 void kvm_hwpoison_page_add(ram_addr_t ram_addr
)
961 QLIST_FOREACH(page
, &hwpoison_page_list
, list
) {
962 if (page
->ram_addr
== ram_addr
) {
966 page
= g_new(HWPoisonPage
, 1);
967 page
->ram_addr
= ram_addr
;
968 QLIST_INSERT_HEAD(&hwpoison_page_list
, page
, list
);
971 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
973 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
974 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
975 * endianness, but the memory core hands them in target endianness.
976 * For example, PPC is always treated as big-endian even if running
977 * on KVM and on PPC64LE. Correct here.
991 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
992 bool assign
, uint32_t size
, bool datamatch
)
995 struct kvm_ioeventfd iofd
= {
996 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1003 trace_kvm_set_ioeventfd_mmio(fd
, (uint64_t)addr
, val
, assign
, size
,
1005 if (!kvm_enabled()) {
1010 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1013 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1016 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
1025 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
1026 bool assign
, uint32_t size
, bool datamatch
)
1028 struct kvm_ioeventfd kick
= {
1029 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1031 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
1036 trace_kvm_set_ioeventfd_pio(fd
, addr
, val
, assign
, size
, datamatch
);
1037 if (!kvm_enabled()) {
1041 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1044 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1046 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
1054 static int kvm_check_many_ioeventfds(void)
1056 /* Userspace can use ioeventfd for io notification. This requires a host
1057 * that supports eventfd(2) and an I/O thread; since eventfd does not
1058 * support SIGIO it cannot interrupt the vcpu.
1060 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
1061 * can avoid creating too many ioeventfds.
1063 #if defined(CONFIG_EVENTFD)
1066 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
1067 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
1068 if (ioeventfds
[i
] < 0) {
1071 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
1073 close(ioeventfds
[i
]);
1078 /* Decide whether many devices are supported or not */
1079 ret
= i
== ARRAY_SIZE(ioeventfds
);
1082 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
1083 close(ioeventfds
[i
]);
1091 static const KVMCapabilityInfo
*
1092 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
1094 while (list
->name
) {
1095 if (!kvm_check_extension(s
, list
->value
)) {
1103 void kvm_set_max_memslot_size(hwaddr max_slot_size
)
1106 ROUND_UP(max_slot_size
, qemu_real_host_page_size
) == max_slot_size
1108 kvm_max_slot_size
= max_slot_size
;
1111 static void kvm_set_phys_mem(KVMMemoryListener
*kml
,
1112 MemoryRegionSection
*section
, bool add
)
1116 MemoryRegion
*mr
= section
->mr
;
1117 bool writeable
= !mr
->readonly
&& !mr
->rom_device
;
1118 hwaddr start_addr
, size
, slot_size
;
1121 if (!memory_region_is_ram(mr
)) {
1122 if (writeable
|| !kvm_readonly_mem_allowed
) {
1124 } else if (!mr
->romd_mode
) {
1125 /* If the memory device is not in romd_mode, then we actually want
1126 * to remove the kvm memory slot so all accesses will trap. */
1131 size
= kvm_align_section(section
, &start_addr
);
1136 /* use aligned delta to align the ram address */
1137 ram
= memory_region_get_ram_ptr(mr
) + section
->offset_within_region
+
1138 (start_addr
- section
->offset_within_address_space
);
1140 kvm_slots_lock(kml
);
1144 slot_size
= MIN(kvm_max_slot_size
, size
);
1145 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
1149 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1150 kvm_physical_sync_dirty_bitmap(kml
, section
);
1153 /* unregister the slot */
1154 g_free(mem
->dirty_bmap
);
1155 mem
->dirty_bmap
= NULL
;
1156 mem
->memory_size
= 0;
1158 err
= kvm_set_user_memory_region(kml
, mem
, false);
1160 fprintf(stderr
, "%s: error unregistering slot: %s\n",
1161 __func__
, strerror(-err
));
1164 start_addr
+= slot_size
;
1170 /* register the new slot */
1172 slot_size
= MIN(kvm_max_slot_size
, size
);
1173 mem
= kvm_alloc_slot(kml
);
1174 mem
->memory_size
= slot_size
;
1175 mem
->start_addr
= start_addr
;
1177 mem
->flags
= kvm_mem_flags(mr
);
1179 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1181 * Reallocate the bmap; it means it doesn't disappear in
1182 * middle of a migrate.
1184 kvm_memslot_init_dirty_bitmap(mem
);
1186 err
= kvm_set_user_memory_region(kml
, mem
, true);
1188 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
1192 start_addr
+= slot_size
;
1198 kvm_slots_unlock(kml
);
1201 static void kvm_region_add(MemoryListener
*listener
,
1202 MemoryRegionSection
*section
)
1204 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1206 memory_region_ref(section
->mr
);
1207 kvm_set_phys_mem(kml
, section
, true);
1210 static void kvm_region_del(MemoryListener
*listener
,
1211 MemoryRegionSection
*section
)
1213 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1215 kvm_set_phys_mem(kml
, section
, false);
1216 memory_region_unref(section
->mr
);
1219 static void kvm_log_sync(MemoryListener
*listener
,
1220 MemoryRegionSection
*section
)
1222 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1225 kvm_slots_lock(kml
);
1226 r
= kvm_physical_sync_dirty_bitmap(kml
, section
);
1227 kvm_slots_unlock(kml
);
1233 static void kvm_log_clear(MemoryListener
*listener
,
1234 MemoryRegionSection
*section
)
1236 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1239 r
= kvm_physical_log_clear(kml
, section
);
1241 error_report_once("%s: kvm log clear failed: mr=%s "
1242 "offset=%"HWADDR_PRIx
" size=%"PRIx64
, __func__
,
1243 section
->mr
->name
, section
->offset_within_region
,
1244 int128_get64(section
->size
));
1249 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
1250 MemoryRegionSection
*section
,
1251 bool match_data
, uint64_t data
,
1254 int fd
= event_notifier_get_fd(e
);
1257 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1258 data
, true, int128_get64(section
->size
),
1261 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1262 __func__
, strerror(-r
), -r
);
1267 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
1268 MemoryRegionSection
*section
,
1269 bool match_data
, uint64_t data
,
1272 int fd
= event_notifier_get_fd(e
);
1275 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1276 data
, false, int128_get64(section
->size
),
1279 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1280 __func__
, strerror(-r
), -r
);
1285 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
1286 MemoryRegionSection
*section
,
1287 bool match_data
, uint64_t data
,
1290 int fd
= event_notifier_get_fd(e
);
1293 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1294 data
, true, int128_get64(section
->size
),
1297 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1298 __func__
, strerror(-r
), -r
);
1303 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
1304 MemoryRegionSection
*section
,
1305 bool match_data
, uint64_t data
,
1309 int fd
= event_notifier_get_fd(e
);
1312 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1313 data
, false, int128_get64(section
->size
),
1316 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1317 __func__
, strerror(-r
), -r
);
1322 void kvm_memory_listener_register(KVMState
*s
, KVMMemoryListener
*kml
,
1323 AddressSpace
*as
, int as_id
)
1327 qemu_mutex_init(&kml
->slots_lock
);
1328 kml
->slots
= g_malloc0(s
->nr_slots
* sizeof(KVMSlot
));
1331 for (i
= 0; i
< s
->nr_slots
; i
++) {
1332 kml
->slots
[i
].slot
= i
;
1335 kml
->listener
.region_add
= kvm_region_add
;
1336 kml
->listener
.region_del
= kvm_region_del
;
1337 kml
->listener
.log_start
= kvm_log_start
;
1338 kml
->listener
.log_stop
= kvm_log_stop
;
1339 kml
->listener
.log_sync
= kvm_log_sync
;
1340 kml
->listener
.log_clear
= kvm_log_clear
;
1341 kml
->listener
.priority
= 10;
1343 memory_listener_register(&kml
->listener
, as
);
1345 for (i
= 0; i
< s
->nr_as
; ++i
) {
1354 static MemoryListener kvm_io_listener
= {
1355 .eventfd_add
= kvm_io_ioeventfd_add
,
1356 .eventfd_del
= kvm_io_ioeventfd_del
,
1360 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
1362 struct kvm_irq_level event
;
1365 assert(kvm_async_interrupts_enabled());
1367 event
.level
= level
;
1369 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
1371 perror("kvm_set_irq");
1375 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
1378 #ifdef KVM_CAP_IRQ_ROUTING
1379 typedef struct KVMMSIRoute
{
1380 struct kvm_irq_routing_entry kroute
;
1381 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
1384 static void set_gsi(KVMState
*s
, unsigned int gsi
)
1386 set_bit(gsi
, s
->used_gsi_bitmap
);
1389 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
1391 clear_bit(gsi
, s
->used_gsi_bitmap
);
1394 void kvm_init_irq_routing(KVMState
*s
)
1398 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
1399 if (gsi_count
> 0) {
1400 /* Round up so we can search ints using ffs */
1401 s
->used_gsi_bitmap
= bitmap_new(gsi_count
);
1402 s
->gsi_count
= gsi_count
;
1405 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
1406 s
->nr_allocated_irq_routes
= 0;
1408 if (!kvm_direct_msi_allowed
) {
1409 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
1410 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
1414 kvm_arch_init_irq_routing(s
);
1417 void kvm_irqchip_commit_routes(KVMState
*s
)
1421 if (kvm_gsi_direct_mapping()) {
1425 if (!kvm_gsi_routing_enabled()) {
1429 s
->irq_routes
->flags
= 0;
1430 trace_kvm_irqchip_commit_routes();
1431 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
1435 static void kvm_add_routing_entry(KVMState
*s
,
1436 struct kvm_irq_routing_entry
*entry
)
1438 struct kvm_irq_routing_entry
*new;
1441 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
1442 n
= s
->nr_allocated_irq_routes
* 2;
1446 size
= sizeof(struct kvm_irq_routing
);
1447 size
+= n
* sizeof(*new);
1448 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
1449 s
->nr_allocated_irq_routes
= n
;
1451 n
= s
->irq_routes
->nr
++;
1452 new = &s
->irq_routes
->entries
[n
];
1456 set_gsi(s
, entry
->gsi
);
1459 static int kvm_update_routing_entry(KVMState
*s
,
1460 struct kvm_irq_routing_entry
*new_entry
)
1462 struct kvm_irq_routing_entry
*entry
;
1465 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
1466 entry
= &s
->irq_routes
->entries
[n
];
1467 if (entry
->gsi
!= new_entry
->gsi
) {
1471 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1475 *entry
= *new_entry
;
1483 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1485 struct kvm_irq_routing_entry e
= {};
1487 assert(pin
< s
->gsi_count
);
1490 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1492 e
.u
.irqchip
.irqchip
= irqchip
;
1493 e
.u
.irqchip
.pin
= pin
;
1494 kvm_add_routing_entry(s
, &e
);
1497 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1499 struct kvm_irq_routing_entry
*e
;
1502 if (kvm_gsi_direct_mapping()) {
1506 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1507 e
= &s
->irq_routes
->entries
[i
];
1508 if (e
->gsi
== virq
) {
1509 s
->irq_routes
->nr
--;
1510 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1514 kvm_arch_release_virq_post(virq
);
1515 trace_kvm_irqchip_release_virq(virq
);
1518 void kvm_irqchip_add_change_notifier(Notifier
*n
)
1520 notifier_list_add(&kvm_irqchip_change_notifiers
, n
);
1523 void kvm_irqchip_remove_change_notifier(Notifier
*n
)
1528 void kvm_irqchip_change_notify(void)
1530 notifier_list_notify(&kvm_irqchip_change_notifiers
, NULL
);
1533 static unsigned int kvm_hash_msi(uint32_t data
)
1535 /* This is optimized for IA32 MSI layout. However, no other arch shall
1536 * repeat the mistake of not providing a direct MSI injection API. */
1540 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1542 KVMMSIRoute
*route
, *next
;
1545 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1546 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1547 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1548 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1554 static int kvm_irqchip_get_virq(KVMState
*s
)
1559 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1560 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1561 * number can succeed even though a new route entry cannot be added.
1562 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1564 if (!kvm_direct_msi_allowed
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1565 kvm_flush_dynamic_msi_routes(s
);
1568 /* Return the lowest unused GSI in the bitmap */
1569 next_virq
= find_first_zero_bit(s
->used_gsi_bitmap
, s
->gsi_count
);
1570 if (next_virq
>= s
->gsi_count
) {
1577 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1579 unsigned int hash
= kvm_hash_msi(msg
.data
);
1582 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1583 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1584 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1585 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1592 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1597 if (kvm_direct_msi_allowed
) {
1598 msi
.address_lo
= (uint32_t)msg
.address
;
1599 msi
.address_hi
= msg
.address
>> 32;
1600 msi
.data
= le32_to_cpu(msg
.data
);
1602 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1604 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1607 route
= kvm_lookup_msi_route(s
, msg
);
1611 virq
= kvm_irqchip_get_virq(s
);
1616 route
= g_malloc0(sizeof(KVMMSIRoute
));
1617 route
->kroute
.gsi
= virq
;
1618 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1619 route
->kroute
.flags
= 0;
1620 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1621 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1622 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1624 kvm_add_routing_entry(s
, &route
->kroute
);
1625 kvm_irqchip_commit_routes(s
);
1627 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1631 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1633 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
1636 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1638 struct kvm_irq_routing_entry kroute
= {};
1640 MSIMessage msg
= {0, 0};
1642 if (pci_available
&& dev
) {
1643 msg
= pci_get_msi_message(dev
, vector
);
1646 if (kvm_gsi_direct_mapping()) {
1647 return kvm_arch_msi_data_to_gsi(msg
.data
);
1650 if (!kvm_gsi_routing_enabled()) {
1654 virq
= kvm_irqchip_get_virq(s
);
1660 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1662 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1663 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1664 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1665 if (pci_available
&& kvm_msi_devid_required()) {
1666 kroute
.flags
= KVM_MSI_VALID_DEVID
;
1667 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
1669 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1670 kvm_irqchip_release_virq(s
, virq
);
1674 trace_kvm_irqchip_add_msi_route(dev
? dev
->name
: (char *)"N/A",
1677 kvm_add_routing_entry(s
, &kroute
);
1678 kvm_arch_add_msi_route_post(&kroute
, vector
, dev
);
1679 kvm_irqchip_commit_routes(s
);
1684 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
,
1687 struct kvm_irq_routing_entry kroute
= {};
1689 if (kvm_gsi_direct_mapping()) {
1693 if (!kvm_irqchip_in_kernel()) {
1698 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1700 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1701 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1702 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1703 if (pci_available
&& kvm_msi_devid_required()) {
1704 kroute
.flags
= KVM_MSI_VALID_DEVID
;
1705 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
1707 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
1711 trace_kvm_irqchip_update_msi_route(virq
);
1713 return kvm_update_routing_entry(s
, &kroute
);
1716 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
1717 EventNotifier
*resample
, int virq
,
1720 int fd
= event_notifier_get_fd(event
);
1721 int rfd
= resample
? event_notifier_get_fd(resample
) : -1;
1723 struct kvm_irqfd irqfd
= {
1726 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
1731 if (kvm_irqchip_is_split()) {
1733 * When the slow irqchip (e.g. IOAPIC) is in the
1734 * userspace, KVM kernel resamplefd will not work because
1735 * the EOI of the interrupt will be delivered to userspace
1736 * instead, so the KVM kernel resamplefd kick will be
1737 * skipped. The userspace here mimics what the kernel
1738 * provides with resamplefd, remember the resamplefd and
1739 * kick it when we receive EOI of this IRQ.
1741 * This is hackery because IOAPIC is mostly bypassed
1742 * (except EOI broadcasts) when irqfd is used. However
1743 * this can bring much performance back for split irqchip
1744 * with INTx IRQs (for VFIO, this gives 93% perf of the
1745 * full fast path, which is 46% perf boost comparing to
1746 * the INTx slow path).
1748 kvm_resample_fd_insert(virq
, resample
);
1750 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
1751 irqfd
.resamplefd
= rfd
;
1753 } else if (!assign
) {
1754 if (kvm_irqchip_is_split()) {
1755 kvm_resample_fd_remove(virq
);
1759 if (!kvm_irqfds_enabled()) {
1763 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
1766 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1768 struct kvm_irq_routing_entry kroute
= {};
1771 if (!kvm_gsi_routing_enabled()) {
1775 virq
= kvm_irqchip_get_virq(s
);
1781 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
1783 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
1784 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
1785 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
1786 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
1787 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
1789 kvm_add_routing_entry(s
, &kroute
);
1794 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1796 struct kvm_irq_routing_entry kroute
= {};
1799 if (!kvm_gsi_routing_enabled()) {
1802 if (!kvm_check_extension(s
, KVM_CAP_HYPERV_SYNIC
)) {
1805 virq
= kvm_irqchip_get_virq(s
);
1811 kroute
.type
= KVM_IRQ_ROUTING_HV_SINT
;
1813 kroute
.u
.hv_sint
.vcpu
= vcpu
;
1814 kroute
.u
.hv_sint
.sint
= sint
;
1816 kvm_add_routing_entry(s
, &kroute
);
1817 kvm_irqchip_commit_routes(s
);
1822 #else /* !KVM_CAP_IRQ_ROUTING */
1824 void kvm_init_irq_routing(KVMState
*s
)
1828 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1832 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1837 int kvm_irqchip_add_msi_route(KVMState
*s
, int vector
, PCIDevice
*dev
)
1842 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
1847 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
1852 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
1853 EventNotifier
*resample
, int virq
,
1859 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
1863 #endif /* !KVM_CAP_IRQ_ROUTING */
1865 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1866 EventNotifier
*rn
, int virq
)
1868 return kvm_irqchip_assign_irqfd(s
, n
, rn
, virq
, true);
1871 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
1874 return kvm_irqchip_assign_irqfd(s
, n
, NULL
, virq
, false);
1877 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1878 EventNotifier
*rn
, qemu_irq irq
)
1881 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1886 return kvm_irqchip_add_irqfd_notifier_gsi(s
, n
, rn
, GPOINTER_TO_INT(gsi
));
1889 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
1893 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
1898 return kvm_irqchip_remove_irqfd_notifier_gsi(s
, n
, GPOINTER_TO_INT(gsi
));
1901 void kvm_irqchip_set_qemuirq_gsi(KVMState
*s
, qemu_irq irq
, int gsi
)
1903 g_hash_table_insert(s
->gsimap
, irq
, GINT_TO_POINTER(gsi
));
1906 static void kvm_irqchip_create(KVMState
*s
)
1910 assert(s
->kernel_irqchip_split
!= ON_OFF_AUTO_AUTO
);
1911 if (kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
1913 } else if (kvm_check_extension(s
, KVM_CAP_S390_IRQCHIP
)) {
1914 ret
= kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0);
1916 fprintf(stderr
, "Enable kernel irqchip failed: %s\n", strerror(-ret
));
1923 /* First probe and see if there's a arch-specific hook to create the
1924 * in-kernel irqchip for us */
1925 ret
= kvm_arch_irqchip_create(s
);
1927 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_ON
) {
1928 perror("Split IRQ chip mode not supported.");
1931 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
1935 fprintf(stderr
, "Create kernel irqchip failed: %s\n", strerror(-ret
));
1939 kvm_kernel_irqchip
= true;
1940 /* If we have an in-kernel IRQ chip then we must have asynchronous
1941 * interrupt delivery (though the reverse is not necessarily true)
1943 kvm_async_interrupts_allowed
= true;
1944 kvm_halt_in_kernel_allowed
= true;
1946 kvm_init_irq_routing(s
);
1948 s
->gsimap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
1951 /* Find number of supported CPUs using the recommended
1952 * procedure from the kernel API documentation to cope with
1953 * older kernels that may be missing capabilities.
1955 static int kvm_recommended_vcpus(KVMState
*s
)
1957 int ret
= kvm_vm_check_extension(s
, KVM_CAP_NR_VCPUS
);
1958 return (ret
) ? ret
: 4;
1961 static int kvm_max_vcpus(KVMState
*s
)
1963 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
1964 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
1967 static int kvm_max_vcpu_id(KVMState
*s
)
1969 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPU_ID
);
1970 return (ret
) ? ret
: kvm_max_vcpus(s
);
1973 bool kvm_vcpu_id_is_valid(int vcpu_id
)
1975 KVMState
*s
= KVM_STATE(current_accel());
1976 return vcpu_id
>= 0 && vcpu_id
< kvm_max_vcpu_id(s
);
1979 static int kvm_init(MachineState
*ms
)
1981 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
1982 static const char upgrade_note
[] =
1983 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1984 "(see http://sourceforge.net/projects/kvm).\n";
1989 { "SMP", ms
->smp
.cpus
},
1990 { "hotpluggable", ms
->smp
.max_cpus
},
1993 int soft_vcpus_limit
, hard_vcpus_limit
;
1995 const KVMCapabilityInfo
*missing_cap
;
1998 uint64_t dirty_log_manual_caps
;
2000 s
= KVM_STATE(ms
->accelerator
);
2003 * On systems where the kernel can support different base page
2004 * sizes, host page size may be different from TARGET_PAGE_SIZE,
2005 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
2006 * page size for the system though.
2008 assert(TARGET_PAGE_SIZE
<= qemu_real_host_page_size
);
2012 #ifdef KVM_CAP_SET_GUEST_DEBUG
2013 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
2015 QLIST_INIT(&s
->kvm_parked_vcpus
);
2017 s
->fd
= qemu_open_old("/dev/kvm", O_RDWR
);
2019 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
2024 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
2025 if (ret
< KVM_API_VERSION
) {
2029 fprintf(stderr
, "kvm version too old\n");
2033 if (ret
> KVM_API_VERSION
) {
2035 fprintf(stderr
, "kvm version not supported\n");
2039 kvm_immediate_exit
= kvm_check_extension(s
, KVM_CAP_IMMEDIATE_EXIT
);
2040 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
2042 /* If unspecified, use the default value */
2047 s
->nr_as
= kvm_check_extension(s
, KVM_CAP_MULTI_ADDRESS_SPACE
);
2048 if (s
->nr_as
<= 1) {
2051 s
->as
= g_new0(struct KVMAs
, s
->nr_as
);
2053 if (object_property_find(OBJECT(current_machine
), "kvm-type")) {
2054 g_autofree
char *kvm_type
= object_property_get_str(OBJECT(current_machine
),
2057 type
= mc
->kvm_type(ms
, kvm_type
);
2061 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
2062 } while (ret
== -EINTR
);
2065 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
2069 if (ret
== -EINVAL
) {
2071 "Host kernel setup problem detected. Please verify:\n");
2072 fprintf(stderr
, "- for kernels supporting the switch_amode or"
2073 " user_mode parameters, whether\n");
2075 " user space is running in primary address space\n");
2077 "- for kernels supporting the vm.allocate_pgste sysctl, "
2078 "whether it is enabled\n");
2086 /* check the vcpu limits */
2087 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
2088 hard_vcpus_limit
= kvm_max_vcpus(s
);
2091 if (nc
->num
> soft_vcpus_limit
) {
2092 warn_report("Number of %s cpus requested (%d) exceeds "
2093 "the recommended cpus supported by KVM (%d)",
2094 nc
->name
, nc
->num
, soft_vcpus_limit
);
2096 if (nc
->num
> hard_vcpus_limit
) {
2097 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
2098 "the maximum cpus supported by KVM (%d)\n",
2099 nc
->name
, nc
->num
, hard_vcpus_limit
);
2106 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
2109 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
2113 fprintf(stderr
, "kvm does not support %s\n%s",
2114 missing_cap
->name
, upgrade_note
);
2118 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
2119 s
->coalesced_pio
= s
->coalesced_mmio
&&
2120 kvm_check_extension(s
, KVM_CAP_COALESCED_PIO
);
2122 dirty_log_manual_caps
=
2123 kvm_check_extension(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
);
2124 dirty_log_manual_caps
&= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE
|
2125 KVM_DIRTY_LOG_INITIALLY_SET
);
2126 s
->manual_dirty_log_protect
= dirty_log_manual_caps
;
2127 if (dirty_log_manual_caps
) {
2128 ret
= kvm_vm_enable_cap(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
, 0,
2129 dirty_log_manual_caps
);
2131 warn_report("Trying to enable capability %"PRIu64
" of "
2132 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2133 "Falling back to the legacy mode. ",
2134 dirty_log_manual_caps
);
2135 s
->manual_dirty_log_protect
= 0;
2139 #ifdef KVM_CAP_VCPU_EVENTS
2140 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
2143 s
->robust_singlestep
=
2144 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
2146 #ifdef KVM_CAP_DEBUGREGS
2147 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
2150 s
->max_nested_state_len
= kvm_check_extension(s
, KVM_CAP_NESTED_STATE
);
2152 #ifdef KVM_CAP_IRQ_ROUTING
2153 kvm_direct_msi_allowed
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
2156 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
2158 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
2159 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
2160 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
2163 kvm_readonly_mem_allowed
=
2164 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
2166 kvm_eventfds_allowed
=
2167 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
2169 kvm_irqfds_allowed
=
2170 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
2172 kvm_resamplefds_allowed
=
2173 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
2175 kvm_vm_attributes_allowed
=
2176 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
2178 kvm_ioeventfd_any_length_allowed
=
2179 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD_ANY_LENGTH
) > 0);
2183 ret
= kvm_arch_init(ms
, s
);
2188 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_AUTO
) {
2189 s
->kernel_irqchip_split
= mc
->default_kernel_irqchip_split
? ON_OFF_AUTO_ON
: ON_OFF_AUTO_OFF
;
2192 qemu_register_reset(kvm_unpoison_all
, NULL
);
2194 if (s
->kernel_irqchip_allowed
) {
2195 kvm_irqchip_create(s
);
2198 if (kvm_eventfds_allowed
) {
2199 s
->memory_listener
.listener
.eventfd_add
= kvm_mem_ioeventfd_add
;
2200 s
->memory_listener
.listener
.eventfd_del
= kvm_mem_ioeventfd_del
;
2202 s
->memory_listener
.listener
.coalesced_io_add
= kvm_coalesce_mmio_region
;
2203 s
->memory_listener
.listener
.coalesced_io_del
= kvm_uncoalesce_mmio_region
;
2205 kvm_memory_listener_register(s
, &s
->memory_listener
,
2206 &address_space_memory
, 0);
2207 if (kvm_eventfds_allowed
) {
2208 memory_listener_register(&kvm_io_listener
,
2211 memory_listener_register(&kvm_coalesced_pio_listener
,
2214 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
2216 s
->sync_mmu
= !!kvm_vm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
2218 ret
= ram_block_discard_disable(true);
2231 g_free(s
->memory_listener
.slots
);
2236 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
2238 s
->sigmask_len
= sigmask_len
;
2241 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
2242 int size
, uint32_t count
)
2245 uint8_t *ptr
= data
;
2247 for (i
= 0; i
< count
; i
++) {
2248 address_space_rw(&address_space_io
, port
, attrs
,
2250 direction
== KVM_EXIT_IO_OUT
);
2255 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
2257 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
2258 run
->internal
.suberror
);
2260 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
2263 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
2264 fprintf(stderr
, "extra data[%d]: %"PRIx64
"\n",
2265 i
, (uint64_t)run
->internal
.data
[i
]);
2268 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
2269 fprintf(stderr
, "emulation failure\n");
2270 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
2271 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
2272 return EXCP_INTERRUPT
;
2275 /* FIXME: Should trigger a qmp message to let management know
2276 * something went wrong.
2281 void kvm_flush_coalesced_mmio_buffer(void)
2283 KVMState
*s
= kvm_state
;
2285 if (s
->coalesced_flush_in_progress
) {
2289 s
->coalesced_flush_in_progress
= true;
2291 if (s
->coalesced_mmio_ring
) {
2292 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
2293 while (ring
->first
!= ring
->last
) {
2294 struct kvm_coalesced_mmio
*ent
;
2296 ent
= &ring
->coalesced_mmio
[ring
->first
];
2298 if (ent
->pio
== 1) {
2299 address_space_write(&address_space_io
, ent
->phys_addr
,
2300 MEMTXATTRS_UNSPECIFIED
, ent
->data
,
2303 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
2306 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
2310 s
->coalesced_flush_in_progress
= false;
2313 static void do_kvm_cpu_synchronize_state(CPUState
*cpu
, run_on_cpu_data arg
)
2315 if (!cpu
->vcpu_dirty
) {
2316 kvm_arch_get_registers(cpu
);
2317 cpu
->vcpu_dirty
= true;
2321 void kvm_cpu_synchronize_state(CPUState
*cpu
)
2323 if (!cpu
->vcpu_dirty
) {
2324 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, RUN_ON_CPU_NULL
);
2328 static void do_kvm_cpu_synchronize_post_reset(CPUState
*cpu
, run_on_cpu_data arg
)
2330 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
2331 cpu
->vcpu_dirty
= false;
2334 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
2336 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, RUN_ON_CPU_NULL
);
2339 static void do_kvm_cpu_synchronize_post_init(CPUState
*cpu
, run_on_cpu_data arg
)
2341 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
2342 cpu
->vcpu_dirty
= false;
2345 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
2347 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, RUN_ON_CPU_NULL
);
2350 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
, run_on_cpu_data arg
)
2352 cpu
->vcpu_dirty
= true;
2355 void kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
)
2357 run_on_cpu(cpu
, do_kvm_cpu_synchronize_pre_loadvm
, RUN_ON_CPU_NULL
);
2360 #ifdef KVM_HAVE_MCE_INJECTION
2361 static __thread
void *pending_sigbus_addr
;
2362 static __thread
int pending_sigbus_code
;
2363 static __thread
bool have_sigbus_pending
;
2366 static void kvm_cpu_kick(CPUState
*cpu
)
2368 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 1);
2371 static void kvm_cpu_kick_self(void)
2373 if (kvm_immediate_exit
) {
2374 kvm_cpu_kick(current_cpu
);
2376 qemu_cpu_kick_self();
2380 static void kvm_eat_signals(CPUState
*cpu
)
2382 struct timespec ts
= { 0, 0 };
2388 if (kvm_immediate_exit
) {
2389 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 0);
2390 /* Write kvm_run->immediate_exit before the cpu->exit_request
2391 * write in kvm_cpu_exec.
2397 sigemptyset(&waitset
);
2398 sigaddset(&waitset
, SIG_IPI
);
2401 r
= sigtimedwait(&waitset
, &siginfo
, &ts
);
2402 if (r
== -1 && !(errno
== EAGAIN
|| errno
== EINTR
)) {
2403 perror("sigtimedwait");
2407 r
= sigpending(&chkset
);
2409 perror("sigpending");
2412 } while (sigismember(&chkset
, SIG_IPI
));
2415 int kvm_cpu_exec(CPUState
*cpu
)
2417 struct kvm_run
*run
= cpu
->kvm_run
;
2420 DPRINTF("kvm_cpu_exec()\n");
2422 if (kvm_arch_process_async_events(cpu
)) {
2423 qatomic_set(&cpu
->exit_request
, 0);
2427 qemu_mutex_unlock_iothread();
2428 cpu_exec_start(cpu
);
2433 if (cpu
->vcpu_dirty
) {
2434 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
2435 cpu
->vcpu_dirty
= false;
2438 kvm_arch_pre_run(cpu
, run
);
2439 if (qatomic_read(&cpu
->exit_request
)) {
2440 DPRINTF("interrupt exit requested\n");
2442 * KVM requires us to reenter the kernel after IO exits to complete
2443 * instruction emulation. This self-signal will ensure that we
2446 kvm_cpu_kick_self();
2449 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2450 * Matching barrier in kvm_eat_signals.
2454 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
2456 attrs
= kvm_arch_post_run(cpu
, run
);
2458 #ifdef KVM_HAVE_MCE_INJECTION
2459 if (unlikely(have_sigbus_pending
)) {
2460 qemu_mutex_lock_iothread();
2461 kvm_arch_on_sigbus_vcpu(cpu
, pending_sigbus_code
,
2462 pending_sigbus_addr
);
2463 have_sigbus_pending
= false;
2464 qemu_mutex_unlock_iothread();
2469 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
2470 DPRINTF("io window exit\n");
2471 kvm_eat_signals(cpu
);
2472 ret
= EXCP_INTERRUPT
;
2475 fprintf(stderr
, "error: kvm run failed %s\n",
2476 strerror(-run_ret
));
2478 if (run_ret
== -EBUSY
) {
2480 "This is probably because your SMT is enabled.\n"
2481 "VCPU can only run on primary threads with all "
2482 "secondary threads offline.\n");
2489 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
2490 switch (run
->exit_reason
) {
2492 DPRINTF("handle_io\n");
2493 /* Called outside BQL */
2494 kvm_handle_io(run
->io
.port
, attrs
,
2495 (uint8_t *)run
+ run
->io
.data_offset
,
2502 DPRINTF("handle_mmio\n");
2503 /* Called outside BQL */
2504 address_space_rw(&address_space_memory
,
2505 run
->mmio
.phys_addr
, attrs
,
2508 run
->mmio
.is_write
);
2511 case KVM_EXIT_IRQ_WINDOW_OPEN
:
2512 DPRINTF("irq_window_open\n");
2513 ret
= EXCP_INTERRUPT
;
2515 case KVM_EXIT_SHUTDOWN
:
2516 DPRINTF("shutdown\n");
2517 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
2518 ret
= EXCP_INTERRUPT
;
2520 case KVM_EXIT_UNKNOWN
:
2521 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
2522 (uint64_t)run
->hw
.hardware_exit_reason
);
2525 case KVM_EXIT_INTERNAL_ERROR
:
2526 ret
= kvm_handle_internal_error(cpu
, run
);
2528 case KVM_EXIT_SYSTEM_EVENT
:
2529 switch (run
->system_event
.type
) {
2530 case KVM_SYSTEM_EVENT_SHUTDOWN
:
2531 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN
);
2532 ret
= EXCP_INTERRUPT
;
2534 case KVM_SYSTEM_EVENT_RESET
:
2535 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
2536 ret
= EXCP_INTERRUPT
;
2538 case KVM_SYSTEM_EVENT_CRASH
:
2539 kvm_cpu_synchronize_state(cpu
);
2540 qemu_mutex_lock_iothread();
2541 qemu_system_guest_panicked(cpu_get_crash_info(cpu
));
2542 qemu_mutex_unlock_iothread();
2546 DPRINTF("kvm_arch_handle_exit\n");
2547 ret
= kvm_arch_handle_exit(cpu
, run
);
2552 DPRINTF("kvm_arch_handle_exit\n");
2553 ret
= kvm_arch_handle_exit(cpu
, run
);
2559 qemu_mutex_lock_iothread();
2562 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
2563 vm_stop(RUN_STATE_INTERNAL_ERROR
);
2566 qatomic_set(&cpu
->exit_request
, 0);
2570 int kvm_ioctl(KVMState
*s
, int type
, ...)
2577 arg
= va_arg(ap
, void *);
2580 trace_kvm_ioctl(type
, arg
);
2581 ret
= ioctl(s
->fd
, type
, arg
);
2588 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
2595 arg
= va_arg(ap
, void *);
2598 trace_kvm_vm_ioctl(type
, arg
);
2599 ret
= ioctl(s
->vmfd
, type
, arg
);
2606 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
2613 arg
= va_arg(ap
, void *);
2616 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
2617 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
2624 int kvm_device_ioctl(int fd
, int type
, ...)
2631 arg
= va_arg(ap
, void *);
2634 trace_kvm_device_ioctl(fd
, type
, arg
);
2635 ret
= ioctl(fd
, type
, arg
);
2642 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
2645 struct kvm_device_attr attribute
= {
2650 if (!kvm_vm_attributes_allowed
) {
2654 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
2655 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2659 int kvm_device_check_attr(int dev_fd
, uint32_t group
, uint64_t attr
)
2661 struct kvm_device_attr attribute
= {
2667 return kvm_device_ioctl(dev_fd
, KVM_HAS_DEVICE_ATTR
, &attribute
) ? 0 : 1;
2670 int kvm_device_access(int fd
, int group
, uint64_t attr
,
2671 void *val
, bool write
, Error
**errp
)
2673 struct kvm_device_attr kvmattr
;
2677 kvmattr
.group
= group
;
2678 kvmattr
.attr
= attr
;
2679 kvmattr
.addr
= (uintptr_t)val
;
2681 err
= kvm_device_ioctl(fd
,
2682 write
? KVM_SET_DEVICE_ATTR
: KVM_GET_DEVICE_ATTR
,
2685 error_setg_errno(errp
, -err
,
2686 "KVM_%s_DEVICE_ATTR failed: Group %d "
2687 "attr 0x%016" PRIx64
,
2688 write
? "SET" : "GET", group
, attr
);
2693 bool kvm_has_sync_mmu(void)
2695 return kvm_state
->sync_mmu
;
2698 int kvm_has_vcpu_events(void)
2700 return kvm_state
->vcpu_events
;
2703 int kvm_has_robust_singlestep(void)
2705 return kvm_state
->robust_singlestep
;
2708 int kvm_has_debugregs(void)
2710 return kvm_state
->debugregs
;
2713 int kvm_max_nested_state_length(void)
2715 return kvm_state
->max_nested_state_len
;
2718 int kvm_has_many_ioeventfds(void)
2720 if (!kvm_enabled()) {
2723 return kvm_state
->many_ioeventfds
;
2726 int kvm_has_gsi_routing(void)
2728 #ifdef KVM_CAP_IRQ_ROUTING
2729 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
2735 int kvm_has_intx_set_mask(void)
2737 return kvm_state
->intx_set_mask
;
2740 bool kvm_arm_supports_user_irq(void)
2742 return kvm_check_extension(kvm_state
, KVM_CAP_ARM_USER_IRQ
);
2745 #ifdef KVM_CAP_SET_GUEST_DEBUG
2746 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
2749 struct kvm_sw_breakpoint
*bp
;
2751 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
2759 int kvm_sw_breakpoints_active(CPUState
*cpu
)
2761 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
2764 struct kvm_set_guest_debug_data
{
2765 struct kvm_guest_debug dbg
;
2769 static void kvm_invoke_set_guest_debug(CPUState
*cpu
, run_on_cpu_data data
)
2771 struct kvm_set_guest_debug_data
*dbg_data
=
2772 (struct kvm_set_guest_debug_data
*) data
.host_ptr
;
2774 dbg_data
->err
= kvm_vcpu_ioctl(cpu
, KVM_SET_GUEST_DEBUG
,
2778 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2780 struct kvm_set_guest_debug_data data
;
2782 data
.dbg
.control
= reinject_trap
;
2784 if (cpu
->singlestep_enabled
) {
2785 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
2787 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
2789 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
,
2790 RUN_ON_CPU_HOST_PTR(&data
));
2794 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2795 target_ulong len
, int type
)
2797 struct kvm_sw_breakpoint
*bp
;
2800 if (type
== GDB_BREAKPOINT_SW
) {
2801 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2807 bp
= g_malloc(sizeof(struct kvm_sw_breakpoint
));
2810 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
2816 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2818 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
2825 err
= kvm_update_guest_debug(cpu
, 0);
2833 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2834 target_ulong len
, int type
)
2836 struct kvm_sw_breakpoint
*bp
;
2839 if (type
== GDB_BREAKPOINT_SW
) {
2840 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
2845 if (bp
->use_count
> 1) {
2850 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
2855 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
2858 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
2865 err
= kvm_update_guest_debug(cpu
, 0);
2873 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2875 struct kvm_sw_breakpoint
*bp
, *next
;
2876 KVMState
*s
= cpu
->kvm_state
;
2879 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
2880 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
2881 /* Try harder to find a CPU that currently sees the breakpoint. */
2882 CPU_FOREACH(tmpcpu
) {
2883 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
2888 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
2891 kvm_arch_remove_all_hw_breakpoints();
2894 kvm_update_guest_debug(cpu
, 0);
2898 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2900 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
2905 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
2906 target_ulong len
, int type
)
2911 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
2912 target_ulong len
, int type
)
2917 void kvm_remove_all_breakpoints(CPUState
*cpu
)
2920 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2922 static int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
2924 KVMState
*s
= kvm_state
;
2925 struct kvm_signal_mask
*sigmask
;
2928 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
2930 sigmask
->len
= s
->sigmask_len
;
2931 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
2932 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
2938 static void kvm_ipi_signal(int sig
)
2941 assert(kvm_immediate_exit
);
2942 kvm_cpu_kick(current_cpu
);
2946 void kvm_init_cpu_signals(CPUState
*cpu
)
2950 struct sigaction sigact
;
2952 memset(&sigact
, 0, sizeof(sigact
));
2953 sigact
.sa_handler
= kvm_ipi_signal
;
2954 sigaction(SIG_IPI
, &sigact
, NULL
);
2956 pthread_sigmask(SIG_BLOCK
, NULL
, &set
);
2957 #if defined KVM_HAVE_MCE_INJECTION
2958 sigdelset(&set
, SIGBUS
);
2959 pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
2961 sigdelset(&set
, SIG_IPI
);
2962 if (kvm_immediate_exit
) {
2963 r
= pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
2965 r
= kvm_set_signal_mask(cpu
, &set
);
2968 fprintf(stderr
, "kvm_set_signal_mask: %s\n", strerror(-r
));
2973 /* Called asynchronously in VCPU thread. */
2974 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
2976 #ifdef KVM_HAVE_MCE_INJECTION
2977 if (have_sigbus_pending
) {
2980 have_sigbus_pending
= true;
2981 pending_sigbus_addr
= addr
;
2982 pending_sigbus_code
= code
;
2983 qatomic_set(&cpu
->exit_request
, 1);
2990 /* Called synchronously (via signalfd) in main thread. */
2991 int kvm_on_sigbus(int code
, void *addr
)
2993 #ifdef KVM_HAVE_MCE_INJECTION
2994 /* Action required MCE kills the process if SIGBUS is blocked. Because
2995 * that's what happens in the I/O thread, where we handle MCE via signalfd,
2996 * we can only get action optional here.
2998 assert(code
!= BUS_MCEERR_AR
);
2999 kvm_arch_on_sigbus_vcpu(first_cpu
, code
, addr
);
3006 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
3009 struct kvm_create_device create_dev
;
3011 create_dev
.type
= type
;
3013 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
3015 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
3019 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
3024 return test
? 0 : create_dev
.fd
;
3027 bool kvm_device_supported(int vmfd
, uint64_t type
)
3029 struct kvm_create_device create_dev
= {
3032 .flags
= KVM_CREATE_DEVICE_TEST
,
3035 if (ioctl(vmfd
, KVM_CHECK_EXTENSION
, KVM_CAP_DEVICE_CTRL
) <= 0) {
3039 return (ioctl(vmfd
, KVM_CREATE_DEVICE
, &create_dev
) >= 0);
3042 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
3044 struct kvm_one_reg reg
;
3048 reg
.addr
= (uintptr_t) source
;
3049 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
3051 trace_kvm_failed_reg_set(id
, strerror(-r
));
3056 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
3058 struct kvm_one_reg reg
;
3062 reg
.addr
= (uintptr_t) target
;
3063 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
3065 trace_kvm_failed_reg_get(id
, strerror(-r
));
3070 static bool kvm_accel_has_memory(MachineState
*ms
, AddressSpace
*as
,
3071 hwaddr start_addr
, hwaddr size
)
3073 KVMState
*kvm
= KVM_STATE(ms
->accelerator
);
3076 for (i
= 0; i
< kvm
->nr_as
; ++i
) {
3077 if (kvm
->as
[i
].as
== as
&& kvm
->as
[i
].ml
) {
3078 size
= MIN(kvm_max_slot_size
, size
);
3079 return NULL
!= kvm_lookup_matching_slot(kvm
->as
[i
].ml
,
3087 static void kvm_get_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3088 const char *name
, void *opaque
,
3091 KVMState
*s
= KVM_STATE(obj
);
3092 int64_t value
= s
->kvm_shadow_mem
;
3094 visit_type_int(v
, name
, &value
, errp
);
3097 static void kvm_set_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3098 const char *name
, void *opaque
,
3101 KVMState
*s
= KVM_STATE(obj
);
3104 if (!visit_type_int(v
, name
, &value
, errp
)) {
3108 s
->kvm_shadow_mem
= value
;
3111 static void kvm_set_kernel_irqchip(Object
*obj
, Visitor
*v
,
3112 const char *name
, void *opaque
,
3115 KVMState
*s
= KVM_STATE(obj
);
3118 if (!visit_type_OnOffSplit(v
, name
, &mode
, errp
)) {
3122 case ON_OFF_SPLIT_ON
:
3123 s
->kernel_irqchip_allowed
= true;
3124 s
->kernel_irqchip_required
= true;
3125 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3127 case ON_OFF_SPLIT_OFF
:
3128 s
->kernel_irqchip_allowed
= false;
3129 s
->kernel_irqchip_required
= false;
3130 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3132 case ON_OFF_SPLIT_SPLIT
:
3133 s
->kernel_irqchip_allowed
= true;
3134 s
->kernel_irqchip_required
= true;
3135 s
->kernel_irqchip_split
= ON_OFF_AUTO_ON
;
3138 /* The value was checked in visit_type_OnOffSplit() above. If
3139 * we get here, then something is wrong in QEMU.
3145 bool kvm_kernel_irqchip_allowed(void)
3147 return kvm_state
->kernel_irqchip_allowed
;
3150 bool kvm_kernel_irqchip_required(void)
3152 return kvm_state
->kernel_irqchip_required
;
3155 bool kvm_kernel_irqchip_split(void)
3157 return kvm_state
->kernel_irqchip_split
== ON_OFF_AUTO_ON
;
3160 static void kvm_accel_instance_init(Object
*obj
)
3162 KVMState
*s
= KVM_STATE(obj
);
3164 s
->kvm_shadow_mem
= -1;
3165 s
->kernel_irqchip_allowed
= true;
3166 s
->kernel_irqchip_split
= ON_OFF_AUTO_AUTO
;
3169 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
3171 AccelClass
*ac
= ACCEL_CLASS(oc
);
3173 ac
->init_machine
= kvm_init
;
3174 ac
->has_memory
= kvm_accel_has_memory
;
3175 ac
->allowed
= &kvm_allowed
;
3177 object_class_property_add(oc
, "kernel-irqchip", "on|off|split",
3178 NULL
, kvm_set_kernel_irqchip
,
3180 object_class_property_set_description(oc
, "kernel-irqchip",
3181 "Configure KVM in-kernel irqchip");
3183 object_class_property_add(oc
, "kvm-shadow-mem", "int",
3184 kvm_get_kvm_shadow_mem
, kvm_set_kvm_shadow_mem
,
3186 object_class_property_set_description(oc
, "kvm-shadow-mem",
3187 "KVM shadow MMU size");
3190 static const TypeInfo kvm_accel_type
= {
3191 .name
= TYPE_KVM_ACCEL
,
3192 .parent
= TYPE_ACCEL
,
3193 .instance_init
= kvm_accel_instance_init
,
3194 .class_init
= kvm_accel_class_init
,
3195 .instance_size
= sizeof(KVMState
),
3198 static void kvm_type_init(void)
3200 type_register_static(&kvm_accel_type
);
3203 type_init(kvm_type_init
);