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>
20 #include <linux/kvm.h>
22 #include "qemu/atomic.h"
23 #include "qemu/option.h"
24 #include "qemu/config-file.h"
25 #include "qemu/error-report.h"
26 #include "qapi/error.h"
27 #include "hw/pci/msi.h"
28 #include "hw/pci/msix.h"
29 #include "hw/s390x/adapter.h"
30 #include "exec/gdbstub.h"
31 #include "sysemu/kvm_int.h"
32 #include "sysemu/runstate.h"
33 #include "sysemu/cpus.h"
34 #include "sysemu/accel-blocker.h"
35 #include "qemu/bswap.h"
36 #include "exec/memory.h"
37 #include "exec/ram_addr.h"
38 #include "qemu/event_notifier.h"
39 #include "qemu/main-loop.h"
42 #include "qapi/visitor.h"
43 #include "qapi/qapi-types-common.h"
44 #include "qapi/qapi-visit-common.h"
45 #include "sysemu/reset.h"
46 #include "qemu/guest-random.h"
47 #include "sysemu/hw_accel.h"
49 #include "sysemu/dirtylimit.h"
50 #include "qemu/range.h"
52 #include "hw/boards.h"
53 #include "sysemu/stats.h"
55 /* This check must be after config-host.h is included */
57 #include <sys/eventfd.h>
60 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
61 * need to use the real host PAGE_SIZE, as that's what KVM will use.
66 #define PAGE_SIZE qemu_real_host_page_size()
68 #ifndef KVM_GUESTDBG_BLOCKIRQ
69 #define KVM_GUESTDBG_BLOCKIRQ 0
75 #define DPRINTF(fmt, ...) \
76 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
78 #define DPRINTF(fmt, ...) \
82 struct KVMParkedVcpu
{
83 unsigned long vcpu_id
;
85 QLIST_ENTRY(KVMParkedVcpu
) node
;
89 bool kvm_kernel_irqchip
;
90 bool kvm_split_irqchip
;
91 bool kvm_async_interrupts_allowed
;
92 bool kvm_halt_in_kernel_allowed
;
93 bool kvm_eventfds_allowed
;
94 bool kvm_irqfds_allowed
;
95 bool kvm_resamplefds_allowed
;
96 bool kvm_msi_via_irqfd_allowed
;
97 bool kvm_gsi_routing_allowed
;
98 bool kvm_gsi_direct_mapping
;
100 bool kvm_readonly_mem_allowed
;
101 bool kvm_vm_attributes_allowed
;
102 bool kvm_direct_msi_allowed
;
103 bool kvm_ioeventfd_any_length_allowed
;
104 bool kvm_msi_use_devid
;
105 bool kvm_has_guest_debug
;
106 static int kvm_sstep_flags
;
107 static bool kvm_immediate_exit
;
108 static hwaddr kvm_max_slot_size
= ~0;
110 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
111 KVM_CAP_INFO(USER_MEMORY
),
112 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
113 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS
),
117 static NotifierList kvm_irqchip_change_notifiers
=
118 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers
);
120 struct KVMResampleFd
{
122 EventNotifier
*resample_event
;
123 QLIST_ENTRY(KVMResampleFd
) node
;
125 typedef struct KVMResampleFd KVMResampleFd
;
128 * Only used with split irqchip where we need to do the resample fd
129 * kick for the kernel from userspace.
131 static QLIST_HEAD(, KVMResampleFd
) kvm_resample_fd_list
=
132 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list
);
134 static QemuMutex kml_slots_lock
;
136 #define kvm_slots_lock() qemu_mutex_lock(&kml_slots_lock)
137 #define kvm_slots_unlock() qemu_mutex_unlock(&kml_slots_lock)
139 static void kvm_slot_init_dirty_bitmap(KVMSlot
*mem
);
141 static inline void kvm_resample_fd_remove(int gsi
)
145 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
146 if (rfd
->gsi
== gsi
) {
147 QLIST_REMOVE(rfd
, node
);
154 static inline void kvm_resample_fd_insert(int gsi
, EventNotifier
*event
)
156 KVMResampleFd
*rfd
= g_new0(KVMResampleFd
, 1);
159 rfd
->resample_event
= event
;
161 QLIST_INSERT_HEAD(&kvm_resample_fd_list
, rfd
, node
);
164 void kvm_resample_fd_notify(int gsi
)
168 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
169 if (rfd
->gsi
== gsi
) {
170 event_notifier_set(rfd
->resample_event
);
171 trace_kvm_resample_fd_notify(gsi
);
177 int kvm_get_max_memslots(void)
179 KVMState
*s
= KVM_STATE(current_accel());
184 /* Called with KVMMemoryListener.slots_lock held */
185 static KVMSlot
*kvm_get_free_slot(KVMMemoryListener
*kml
)
187 KVMState
*s
= kvm_state
;
190 for (i
= 0; i
< s
->nr_slots
; i
++) {
191 if (kml
->slots
[i
].memory_size
== 0) {
192 return &kml
->slots
[i
];
199 bool kvm_has_free_slot(MachineState
*ms
)
201 KVMState
*s
= KVM_STATE(ms
->accelerator
);
203 KVMMemoryListener
*kml
= &s
->memory_listener
;
206 result
= !!kvm_get_free_slot(kml
);
212 /* Called with KVMMemoryListener.slots_lock held */
213 static KVMSlot
*kvm_alloc_slot(KVMMemoryListener
*kml
)
215 KVMSlot
*slot
= kvm_get_free_slot(kml
);
221 fprintf(stderr
, "%s: no free slot available\n", __func__
);
225 static KVMSlot
*kvm_lookup_matching_slot(KVMMemoryListener
*kml
,
229 KVMState
*s
= kvm_state
;
232 for (i
= 0; i
< s
->nr_slots
; i
++) {
233 KVMSlot
*mem
= &kml
->slots
[i
];
235 if (start_addr
== mem
->start_addr
&& size
== mem
->memory_size
) {
244 * Calculate and align the start address and the size of the section.
245 * Return the size. If the size is 0, the aligned section is empty.
247 static hwaddr
kvm_align_section(MemoryRegionSection
*section
,
250 hwaddr size
= int128_get64(section
->size
);
251 hwaddr delta
, aligned
;
253 /* kvm works in page size chunks, but the function may be called
254 with sub-page size and unaligned start address. Pad the start
255 address to next and truncate size to previous page boundary. */
256 aligned
= ROUND_UP(section
->offset_within_address_space
,
257 qemu_real_host_page_size());
258 delta
= aligned
- section
->offset_within_address_space
;
264 return (size
- delta
) & qemu_real_host_page_mask();
267 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
270 KVMMemoryListener
*kml
= &s
->memory_listener
;
274 for (i
= 0; i
< s
->nr_slots
; i
++) {
275 KVMSlot
*mem
= &kml
->slots
[i
];
277 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
278 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
288 static int kvm_set_user_memory_region(KVMMemoryListener
*kml
, KVMSlot
*slot
, bool new)
290 KVMState
*s
= kvm_state
;
291 struct kvm_userspace_memory_region mem
;
294 mem
.slot
= slot
->slot
| (kml
->as_id
<< 16);
295 mem
.guest_phys_addr
= slot
->start_addr
;
296 mem
.userspace_addr
= (unsigned long)slot
->ram
;
297 mem
.flags
= slot
->flags
;
299 if (slot
->memory_size
&& !new && (mem
.flags
^ slot
->old_flags
) & KVM_MEM_READONLY
) {
300 /* Set the slot size to 0 before setting the slot to the desired
301 * value. This is needed based on KVM commit 75d61fbc. */
303 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
308 mem
.memory_size
= slot
->memory_size
;
309 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
310 slot
->old_flags
= mem
.flags
;
312 trace_kvm_set_user_memory(mem
.slot
, mem
.flags
, mem
.guest_phys_addr
,
313 mem
.memory_size
, mem
.userspace_addr
, ret
);
315 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
316 " start=0x%" PRIx64
", size=0x%" PRIx64
": %s",
317 __func__
, mem
.slot
, slot
->start_addr
,
318 (uint64_t)mem
.memory_size
, strerror(errno
));
323 static int do_kvm_destroy_vcpu(CPUState
*cpu
)
325 KVMState
*s
= kvm_state
;
327 struct KVMParkedVcpu
*vcpu
= NULL
;
330 DPRINTF("kvm_destroy_vcpu\n");
332 ret
= kvm_arch_destroy_vcpu(cpu
);
337 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
340 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
344 ret
= munmap(cpu
->kvm_run
, mmap_size
);
349 if (cpu
->kvm_dirty_gfns
) {
350 ret
= munmap(cpu
->kvm_dirty_gfns
, s
->kvm_dirty_ring_bytes
);
356 vcpu
= g_malloc0(sizeof(*vcpu
));
357 vcpu
->vcpu_id
= kvm_arch_vcpu_id(cpu
);
358 vcpu
->kvm_fd
= cpu
->kvm_fd
;
359 QLIST_INSERT_HEAD(&kvm_state
->kvm_parked_vcpus
, vcpu
, node
);
364 void kvm_destroy_vcpu(CPUState
*cpu
)
366 if (do_kvm_destroy_vcpu(cpu
) < 0) {
367 error_report("kvm_destroy_vcpu failed");
372 static int kvm_get_vcpu(KVMState
*s
, unsigned long vcpu_id
)
374 struct KVMParkedVcpu
*cpu
;
376 QLIST_FOREACH(cpu
, &s
->kvm_parked_vcpus
, node
) {
377 if (cpu
->vcpu_id
== vcpu_id
) {
380 QLIST_REMOVE(cpu
, node
);
381 kvm_fd
= cpu
->kvm_fd
;
387 return kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)vcpu_id
);
390 int kvm_init_vcpu(CPUState
*cpu
, Error
**errp
)
392 KVMState
*s
= kvm_state
;
396 trace_kvm_init_vcpu(cpu
->cpu_index
, kvm_arch_vcpu_id(cpu
));
398 ret
= kvm_get_vcpu(s
, kvm_arch_vcpu_id(cpu
));
400 error_setg_errno(errp
, -ret
, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)",
401 kvm_arch_vcpu_id(cpu
));
407 cpu
->vcpu_dirty
= true;
408 cpu
->dirty_pages
= 0;
409 cpu
->throttle_us_per_full
= 0;
411 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
414 error_setg_errno(errp
, -mmap_size
,
415 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
419 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
421 if (cpu
->kvm_run
== MAP_FAILED
) {
423 error_setg_errno(errp
, ret
,
424 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
425 kvm_arch_vcpu_id(cpu
));
429 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
430 s
->coalesced_mmio_ring
=
431 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
434 if (s
->kvm_dirty_ring_size
) {
435 /* Use MAP_SHARED to share pages with the kernel */
436 cpu
->kvm_dirty_gfns
= mmap(NULL
, s
->kvm_dirty_ring_bytes
,
437 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
439 PAGE_SIZE
* KVM_DIRTY_LOG_PAGE_OFFSET
);
440 if (cpu
->kvm_dirty_gfns
== MAP_FAILED
) {
442 DPRINTF("mmap'ing vcpu dirty gfns failed: %d\n", ret
);
447 ret
= kvm_arch_init_vcpu(cpu
);
449 error_setg_errno(errp
, -ret
,
450 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
451 kvm_arch_vcpu_id(cpu
));
458 * dirty pages logging control
461 static int kvm_mem_flags(MemoryRegion
*mr
)
463 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
466 if (memory_region_get_dirty_log_mask(mr
) != 0) {
467 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
469 if (readonly
&& kvm_readonly_mem_allowed
) {
470 flags
|= KVM_MEM_READONLY
;
475 /* Called with KVMMemoryListener.slots_lock held */
476 static int kvm_slot_update_flags(KVMMemoryListener
*kml
, KVMSlot
*mem
,
479 mem
->flags
= kvm_mem_flags(mr
);
481 /* If nothing changed effectively, no need to issue ioctl */
482 if (mem
->flags
== mem
->old_flags
) {
486 kvm_slot_init_dirty_bitmap(mem
);
487 return kvm_set_user_memory_region(kml
, mem
, false);
490 static int kvm_section_update_flags(KVMMemoryListener
*kml
,
491 MemoryRegionSection
*section
)
493 hwaddr start_addr
, size
, slot_size
;
497 size
= kvm_align_section(section
, &start_addr
);
504 while (size
&& !ret
) {
505 slot_size
= MIN(kvm_max_slot_size
, size
);
506 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
508 /* We don't have a slot if we want to trap every access. */
512 ret
= kvm_slot_update_flags(kml
, mem
, section
->mr
);
513 start_addr
+= slot_size
;
522 static void kvm_log_start(MemoryListener
*listener
,
523 MemoryRegionSection
*section
,
526 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
533 r
= kvm_section_update_flags(kml
, section
);
539 static void kvm_log_stop(MemoryListener
*listener
,
540 MemoryRegionSection
*section
,
543 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
550 r
= kvm_section_update_flags(kml
, section
);
556 /* get kvm's dirty pages bitmap and update qemu's */
557 static void kvm_slot_sync_dirty_pages(KVMSlot
*slot
)
559 ram_addr_t start
= slot
->ram_start_offset
;
560 ram_addr_t pages
= slot
->memory_size
/ qemu_real_host_page_size();
562 cpu_physical_memory_set_dirty_lebitmap(slot
->dirty_bmap
, start
, pages
);
565 static void kvm_slot_reset_dirty_pages(KVMSlot
*slot
)
567 memset(slot
->dirty_bmap
, 0, slot
->dirty_bmap_size
);
570 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
572 /* Allocate the dirty bitmap for a slot */
573 static void kvm_slot_init_dirty_bitmap(KVMSlot
*mem
)
575 if (!(mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) || mem
->dirty_bmap
) {
580 * XXX bad kernel interface alert
581 * For dirty bitmap, kernel allocates array of size aligned to
582 * bits-per-long. But for case when the kernel is 64bits and
583 * the userspace is 32bits, userspace can't align to the same
584 * bits-per-long, since sizeof(long) is different between kernel
585 * and user space. This way, userspace will provide buffer which
586 * may be 4 bytes less than the kernel will use, resulting in
587 * userspace memory corruption (which is not detectable by valgrind
588 * too, in most cases).
589 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
590 * a hope that sizeof(long) won't become >8 any time soon.
592 * Note: the granule of kvm dirty log is qemu_real_host_page_size.
593 * And mem->memory_size is aligned to it (otherwise this mem can't
594 * be registered to KVM).
596 hwaddr bitmap_size
= ALIGN(mem
->memory_size
/ qemu_real_host_page_size(),
597 /*HOST_LONG_BITS*/ 64) / 8;
598 mem
->dirty_bmap
= g_malloc0(bitmap_size
);
599 mem
->dirty_bmap_size
= bitmap_size
;
603 * Sync dirty bitmap from kernel to KVMSlot.dirty_bmap, return true if
604 * succeeded, false otherwise
606 static bool kvm_slot_get_dirty_log(KVMState
*s
, KVMSlot
*slot
)
608 struct kvm_dirty_log d
= {};
611 d
.dirty_bitmap
= slot
->dirty_bmap
;
612 d
.slot
= slot
->slot
| (slot
->as_id
<< 16);
613 ret
= kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
);
615 if (ret
== -ENOENT
) {
616 /* kernel does not have dirty bitmap in this slot */
620 error_report_once("%s: KVM_GET_DIRTY_LOG failed with %d",
626 /* Should be with all slots_lock held for the address spaces. */
627 static void kvm_dirty_ring_mark_page(KVMState
*s
, uint32_t as_id
,
628 uint32_t slot_id
, uint64_t offset
)
630 KVMMemoryListener
*kml
;
633 if (as_id
>= s
->nr_as
) {
637 kml
= s
->as
[as_id
].ml
;
638 mem
= &kml
->slots
[slot_id
];
640 if (!mem
->memory_size
|| offset
>=
641 (mem
->memory_size
/ qemu_real_host_page_size())) {
645 set_bit(offset
, mem
->dirty_bmap
);
648 static bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn
*gfn
)
651 * Read the flags before the value. Pairs with barrier in
652 * KVM's kvm_dirty_ring_push() function.
654 return qatomic_load_acquire(&gfn
->flags
) == KVM_DIRTY_GFN_F_DIRTY
;
657 static void dirty_gfn_set_collected(struct kvm_dirty_gfn
*gfn
)
660 * Use a store-release so that the CPU that executes KVM_RESET_DIRTY_RINGS
661 * sees the full content of the ring:
664 * ------------------------------------------------------------------------------
666 * store-rel flags for gfn0
667 * load-acq flags for gfn0
668 * store-rel RESET for gfn0
670 * load-acq flags for gfn0
671 * check if flags have RESET
673 * The synchronization goes from CPU2 to CPU0 to CPU1.
675 qatomic_store_release(&gfn
->flags
, KVM_DIRTY_GFN_F_RESET
);
679 * Should be with all slots_lock held for the address spaces. It returns the
680 * dirty page we've collected on this dirty ring.
682 static uint32_t kvm_dirty_ring_reap_one(KVMState
*s
, CPUState
*cpu
)
684 struct kvm_dirty_gfn
*dirty_gfns
= cpu
->kvm_dirty_gfns
, *cur
;
685 uint32_t ring_size
= s
->kvm_dirty_ring_size
;
686 uint32_t count
= 0, fetch
= cpu
->kvm_fetch_index
;
688 assert(dirty_gfns
&& ring_size
);
689 trace_kvm_dirty_ring_reap_vcpu(cpu
->cpu_index
);
692 cur
= &dirty_gfns
[fetch
% ring_size
];
693 if (!dirty_gfn_is_dirtied(cur
)) {
696 kvm_dirty_ring_mark_page(s
, cur
->slot
>> 16, cur
->slot
& 0xffff,
698 dirty_gfn_set_collected(cur
);
699 trace_kvm_dirty_ring_page(cpu
->cpu_index
, fetch
, cur
->offset
);
703 cpu
->kvm_fetch_index
= fetch
;
704 cpu
->dirty_pages
+= count
;
709 /* Must be with slots_lock held */
710 static uint64_t kvm_dirty_ring_reap_locked(KVMState
*s
, CPUState
* cpu
)
719 total
= kvm_dirty_ring_reap_one(s
, cpu
);
722 total
+= kvm_dirty_ring_reap_one(s
, cpu
);
727 ret
= kvm_vm_ioctl(s
, KVM_RESET_DIRTY_RINGS
);
728 assert(ret
== total
);
731 stamp
= get_clock() - stamp
;
734 trace_kvm_dirty_ring_reap(total
, stamp
/ 1000);
741 * Currently for simplicity, we must hold BQL before calling this. We can
742 * consider to drop the BQL if we're clear with all the race conditions.
744 static uint64_t kvm_dirty_ring_reap(KVMState
*s
, CPUState
*cpu
)
749 * We need to lock all kvm slots for all address spaces here,
752 * (1) We need to mark dirty for dirty bitmaps in multiple slots
753 * and for tons of pages, so it's better to take the lock here
754 * once rather than once per page. And more importantly,
756 * (2) We must _NOT_ publish dirty bits to the other threads
757 * (e.g., the migration thread) via the kvm memory slot dirty
758 * bitmaps before correctly re-protect those dirtied pages.
759 * Otherwise we can have potential risk of data corruption if
760 * the page data is read in the other thread before we do
764 total
= kvm_dirty_ring_reap_locked(s
, cpu
);
770 static void do_kvm_cpu_synchronize_kick(CPUState
*cpu
, run_on_cpu_data arg
)
772 /* No need to do anything */
776 * Kick all vcpus out in a synchronized way. When returned, we
777 * guarantee that every vcpu has been kicked and at least returned to
780 static void kvm_cpu_synchronize_kick_all(void)
785 run_on_cpu(cpu
, do_kvm_cpu_synchronize_kick
, RUN_ON_CPU_NULL
);
790 * Flush all the existing dirty pages to the KVM slot buffers. When
791 * this call returns, we guarantee that all the touched dirty pages
792 * before calling this function have been put into the per-kvmslot
795 * This function must be called with BQL held.
797 static void kvm_dirty_ring_flush(void)
799 trace_kvm_dirty_ring_flush(0);
801 * The function needs to be serialized. Since this function
802 * should always be with BQL held, serialization is guaranteed.
803 * However, let's be sure of it.
805 assert(qemu_mutex_iothread_locked());
807 * First make sure to flush the hardware buffers by kicking all
808 * vcpus out in a synchronous way.
810 kvm_cpu_synchronize_kick_all();
811 kvm_dirty_ring_reap(kvm_state
, NULL
);
812 trace_kvm_dirty_ring_flush(1);
816 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
818 * This function will first try to fetch dirty bitmap from the kernel,
819 * and then updates qemu's dirty bitmap.
821 * NOTE: caller must be with kml->slots_lock held.
823 * @kml: the KVM memory listener object
824 * @section: the memory section to sync the dirty bitmap with
826 static void kvm_physical_sync_dirty_bitmap(KVMMemoryListener
*kml
,
827 MemoryRegionSection
*section
)
829 KVMState
*s
= kvm_state
;
831 hwaddr start_addr
, size
;
834 size
= kvm_align_section(section
, &start_addr
);
836 slot_size
= MIN(kvm_max_slot_size
, size
);
837 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
839 /* We don't have a slot if we want to trap every access. */
842 if (kvm_slot_get_dirty_log(s
, mem
)) {
843 kvm_slot_sync_dirty_pages(mem
);
845 start_addr
+= slot_size
;
850 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
851 #define KVM_CLEAR_LOG_SHIFT 6
852 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size() << KVM_CLEAR_LOG_SHIFT)
853 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN)
855 static int kvm_log_clear_one_slot(KVMSlot
*mem
, int as_id
, uint64_t start
,
858 KVMState
*s
= kvm_state
;
859 uint64_t end
, bmap_start
, start_delta
, bmap_npages
;
860 struct kvm_clear_dirty_log d
;
861 unsigned long *bmap_clear
= NULL
, psize
= qemu_real_host_page_size();
865 * We need to extend either the start or the size or both to
866 * satisfy the KVM interface requirement. Firstly, do the start
867 * page alignment on 64 host pages
869 bmap_start
= start
& KVM_CLEAR_LOG_MASK
;
870 start_delta
= start
- bmap_start
;
874 * The kernel interface has restriction on the size too, that either:
876 * (1) the size is 64 host pages aligned (just like the start), or
877 * (2) the size fills up until the end of the KVM memslot.
879 bmap_npages
= DIV_ROUND_UP(size
+ start_delta
, KVM_CLEAR_LOG_ALIGN
)
880 << KVM_CLEAR_LOG_SHIFT
;
881 end
= mem
->memory_size
/ psize
;
882 if (bmap_npages
> end
- bmap_start
) {
883 bmap_npages
= end
- bmap_start
;
885 start_delta
/= psize
;
888 * Prepare the bitmap to clear dirty bits. Here we must guarantee
889 * that we won't clear any unknown dirty bits otherwise we might
890 * accidentally clear some set bits which are not yet synced from
891 * the kernel into QEMU's bitmap, then we'll lose track of the
892 * guest modifications upon those pages (which can directly lead
893 * to guest data loss or panic after migration).
895 * Layout of the KVMSlot.dirty_bmap:
897 * |<-------- bmap_npages -----------..>|
900 * |----------------|-------------|------------------|------------|
903 * start bmap_start (start) end
904 * of memslot of memslot
906 * [1] bmap_npages can be aligned to either 64 pages or the end of slot
909 assert(bmap_start
% BITS_PER_LONG
== 0);
910 /* We should never do log_clear before log_sync */
911 assert(mem
->dirty_bmap
);
912 if (start_delta
|| bmap_npages
- size
/ psize
) {
913 /* Slow path - we need to manipulate a temp bitmap */
914 bmap_clear
= bitmap_new(bmap_npages
);
915 bitmap_copy_with_src_offset(bmap_clear
, mem
->dirty_bmap
,
916 bmap_start
, start_delta
+ size
/ psize
);
918 * We need to fill the holes at start because that was not
919 * specified by the caller and we extended the bitmap only for
922 bitmap_clear(bmap_clear
, 0, start_delta
);
923 d
.dirty_bitmap
= bmap_clear
;
926 * Fast path - both start and size align well with BITS_PER_LONG
927 * (or the end of memory slot)
929 d
.dirty_bitmap
= mem
->dirty_bmap
+ BIT_WORD(bmap_start
);
932 d
.first_page
= bmap_start
;
933 /* It should never overflow. If it happens, say something */
934 assert(bmap_npages
<= UINT32_MAX
);
935 d
.num_pages
= bmap_npages
;
936 d
.slot
= mem
->slot
| (as_id
<< 16);
938 ret
= kvm_vm_ioctl(s
, KVM_CLEAR_DIRTY_LOG
, &d
);
939 if (ret
< 0 && ret
!= -ENOENT
) {
940 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
941 "start=0x%"PRIx64
", size=0x%"PRIx32
", errno=%d",
942 __func__
, d
.slot
, (uint64_t)d
.first_page
,
943 (uint32_t)d
.num_pages
, ret
);
946 trace_kvm_clear_dirty_log(d
.slot
, d
.first_page
, d
.num_pages
);
950 * After we have updated the remote dirty bitmap, we update the
951 * cached bitmap as well for the memslot, then if another user
952 * clears the same region we know we shouldn't clear it again on
953 * the remote otherwise it's data loss as well.
955 bitmap_clear(mem
->dirty_bmap
, bmap_start
+ start_delta
,
957 /* This handles the NULL case well */
964 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
966 * NOTE: this will be a no-op if we haven't enabled manual dirty log
967 * protection in the host kernel because in that case this operation
968 * will be done within log_sync().
970 * @kml: the kvm memory listener
971 * @section: the memory range to clear dirty bitmap
973 static int kvm_physical_log_clear(KVMMemoryListener
*kml
,
974 MemoryRegionSection
*section
)
976 KVMState
*s
= kvm_state
;
977 uint64_t start
, size
, offset
, count
;
981 if (!s
->manual_dirty_log_protect
) {
982 /* No need to do explicit clear */
986 start
= section
->offset_within_address_space
;
987 size
= int128_get64(section
->size
);
990 /* Nothing more we can do... */
996 for (i
= 0; i
< s
->nr_slots
; i
++) {
997 mem
= &kml
->slots
[i
];
998 /* Discard slots that are empty or do not overlap the section */
999 if (!mem
->memory_size
||
1000 mem
->start_addr
> start
+ size
- 1 ||
1001 start
> mem
->start_addr
+ mem
->memory_size
- 1) {
1005 if (start
>= mem
->start_addr
) {
1006 /* The slot starts before section or is aligned to it. */
1007 offset
= start
- mem
->start_addr
;
1008 count
= MIN(mem
->memory_size
- offset
, size
);
1010 /* The slot starts after section. */
1012 count
= MIN(mem
->memory_size
, size
- (mem
->start_addr
- start
));
1014 ret
= kvm_log_clear_one_slot(mem
, kml
->as_id
, offset
, count
);
1025 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
1026 MemoryRegionSection
*secion
,
1027 hwaddr start
, hwaddr size
)
1029 KVMState
*s
= kvm_state
;
1031 if (s
->coalesced_mmio
) {
1032 struct kvm_coalesced_mmio_zone zone
;
1038 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
1042 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
1043 MemoryRegionSection
*secion
,
1044 hwaddr start
, hwaddr size
)
1046 KVMState
*s
= kvm_state
;
1048 if (s
->coalesced_mmio
) {
1049 struct kvm_coalesced_mmio_zone zone
;
1055 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
1059 static void kvm_coalesce_pio_add(MemoryListener
*listener
,
1060 MemoryRegionSection
*section
,
1061 hwaddr start
, hwaddr size
)
1063 KVMState
*s
= kvm_state
;
1065 if (s
->coalesced_pio
) {
1066 struct kvm_coalesced_mmio_zone zone
;
1072 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
1076 static void kvm_coalesce_pio_del(MemoryListener
*listener
,
1077 MemoryRegionSection
*section
,
1078 hwaddr start
, hwaddr size
)
1080 KVMState
*s
= kvm_state
;
1082 if (s
->coalesced_pio
) {
1083 struct kvm_coalesced_mmio_zone zone
;
1089 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
1093 static MemoryListener kvm_coalesced_pio_listener
= {
1094 .name
= "kvm-coalesced-pio",
1095 .coalesced_io_add
= kvm_coalesce_pio_add
,
1096 .coalesced_io_del
= kvm_coalesce_pio_del
,
1099 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
1103 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
1111 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
1115 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
1117 /* VM wide version not implemented, use global one instead */
1118 ret
= kvm_check_extension(s
, extension
);
1124 typedef struct HWPoisonPage
{
1125 ram_addr_t ram_addr
;
1126 QLIST_ENTRY(HWPoisonPage
) list
;
1129 static QLIST_HEAD(, HWPoisonPage
) hwpoison_page_list
=
1130 QLIST_HEAD_INITIALIZER(hwpoison_page_list
);
1132 static void kvm_unpoison_all(void *param
)
1134 HWPoisonPage
*page
, *next_page
;
1136 QLIST_FOREACH_SAFE(page
, &hwpoison_page_list
, list
, next_page
) {
1137 QLIST_REMOVE(page
, list
);
1138 qemu_ram_remap(page
->ram_addr
, TARGET_PAGE_SIZE
);
1143 void kvm_hwpoison_page_add(ram_addr_t ram_addr
)
1147 QLIST_FOREACH(page
, &hwpoison_page_list
, list
) {
1148 if (page
->ram_addr
== ram_addr
) {
1152 page
= g_new(HWPoisonPage
, 1);
1153 page
->ram_addr
= ram_addr
;
1154 QLIST_INSERT_HEAD(&hwpoison_page_list
, page
, list
);
1157 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
1159 #if HOST_BIG_ENDIAN != TARGET_BIG_ENDIAN
1160 /* The kernel expects ioeventfd values in HOST_BIG_ENDIAN
1161 * endianness, but the memory core hands them in target endianness.
1162 * For example, PPC is always treated as big-endian even if running
1163 * on KVM and on PPC64LE. Correct here.
1177 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
1178 bool assign
, uint32_t size
, bool datamatch
)
1181 struct kvm_ioeventfd iofd
= {
1182 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1189 trace_kvm_set_ioeventfd_mmio(fd
, (uint64_t)addr
, val
, assign
, size
,
1191 if (!kvm_enabled()) {
1196 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1199 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1202 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
1211 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
1212 bool assign
, uint32_t size
, bool datamatch
)
1214 struct kvm_ioeventfd kick
= {
1215 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1217 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
1222 trace_kvm_set_ioeventfd_pio(fd
, addr
, val
, assign
, size
, datamatch
);
1223 if (!kvm_enabled()) {
1227 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1230 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1232 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
1240 static int kvm_check_many_ioeventfds(void)
1242 /* Userspace can use ioeventfd for io notification. This requires a host
1243 * that supports eventfd(2) and an I/O thread; since eventfd does not
1244 * support SIGIO it cannot interrupt the vcpu.
1246 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
1247 * can avoid creating too many ioeventfds.
1249 #if defined(CONFIG_EVENTFD)
1252 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
1253 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
1254 if (ioeventfds
[i
] < 0) {
1257 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
1259 close(ioeventfds
[i
]);
1264 /* Decide whether many devices are supported or not */
1265 ret
= i
== ARRAY_SIZE(ioeventfds
);
1268 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
1269 close(ioeventfds
[i
]);
1277 static const KVMCapabilityInfo
*
1278 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
1280 while (list
->name
) {
1281 if (!kvm_check_extension(s
, list
->value
)) {
1289 void kvm_set_max_memslot_size(hwaddr max_slot_size
)
1292 ROUND_UP(max_slot_size
, qemu_real_host_page_size()) == max_slot_size
1294 kvm_max_slot_size
= max_slot_size
;
1297 /* Called with KVMMemoryListener.slots_lock held */
1298 static void kvm_set_phys_mem(KVMMemoryListener
*kml
,
1299 MemoryRegionSection
*section
, bool add
)
1303 MemoryRegion
*mr
= section
->mr
;
1304 bool writable
= !mr
->readonly
&& !mr
->rom_device
;
1305 hwaddr start_addr
, size
, slot_size
, mr_offset
;
1306 ram_addr_t ram_start_offset
;
1309 if (!memory_region_is_ram(mr
)) {
1310 if (writable
|| !kvm_readonly_mem_allowed
) {
1312 } else if (!mr
->romd_mode
) {
1313 /* If the memory device is not in romd_mode, then we actually want
1314 * to remove the kvm memory slot so all accesses will trap. */
1319 size
= kvm_align_section(section
, &start_addr
);
1324 /* The offset of the kvmslot within the memory region */
1325 mr_offset
= section
->offset_within_region
+ start_addr
-
1326 section
->offset_within_address_space
;
1328 /* use aligned delta to align the ram address and offset */
1329 ram
= memory_region_get_ram_ptr(mr
) + mr_offset
;
1330 ram_start_offset
= memory_region_get_ram_addr(mr
) + mr_offset
;
1334 slot_size
= MIN(kvm_max_slot_size
, size
);
1335 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
1339 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1341 * NOTE: We should be aware of the fact that here we're only
1342 * doing a best effort to sync dirty bits. No matter whether
1343 * we're using dirty log or dirty ring, we ignored two facts:
1345 * (1) dirty bits can reside in hardware buffers (PML)
1347 * (2) after we collected dirty bits here, pages can be dirtied
1348 * again before we do the final KVM_SET_USER_MEMORY_REGION to
1351 * Not easy. Let's cross the fingers until it's fixed.
1353 if (kvm_state
->kvm_dirty_ring_size
) {
1354 kvm_dirty_ring_reap_locked(kvm_state
, NULL
);
1356 kvm_slot_get_dirty_log(kvm_state
, mem
);
1358 kvm_slot_sync_dirty_pages(mem
);
1361 /* unregister the slot */
1362 g_free(mem
->dirty_bmap
);
1363 mem
->dirty_bmap
= NULL
;
1364 mem
->memory_size
= 0;
1366 err
= kvm_set_user_memory_region(kml
, mem
, false);
1368 fprintf(stderr
, "%s: error unregistering slot: %s\n",
1369 __func__
, strerror(-err
));
1372 start_addr
+= slot_size
;
1378 /* register the new slot */
1380 slot_size
= MIN(kvm_max_slot_size
, size
);
1381 mem
= kvm_alloc_slot(kml
);
1382 mem
->as_id
= kml
->as_id
;
1383 mem
->memory_size
= slot_size
;
1384 mem
->start_addr
= start_addr
;
1385 mem
->ram_start_offset
= ram_start_offset
;
1387 mem
->flags
= kvm_mem_flags(mr
);
1388 kvm_slot_init_dirty_bitmap(mem
);
1389 err
= kvm_set_user_memory_region(kml
, mem
, true);
1391 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
1395 start_addr
+= slot_size
;
1396 ram_start_offset
+= slot_size
;
1402 static void *kvm_dirty_ring_reaper_thread(void *data
)
1405 struct KVMDirtyRingReaper
*r
= &s
->reaper
;
1407 rcu_register_thread();
1409 trace_kvm_dirty_ring_reaper("init");
1412 r
->reaper_state
= KVM_DIRTY_RING_REAPER_WAIT
;
1413 trace_kvm_dirty_ring_reaper("wait");
1415 * TODO: provide a smarter timeout rather than a constant?
1419 /* keep sleeping so that dirtylimit not be interfered by reaper */
1420 if (dirtylimit_in_service()) {
1424 trace_kvm_dirty_ring_reaper("wakeup");
1425 r
->reaper_state
= KVM_DIRTY_RING_REAPER_REAPING
;
1427 qemu_mutex_lock_iothread();
1428 kvm_dirty_ring_reap(s
, NULL
);
1429 qemu_mutex_unlock_iothread();
1431 r
->reaper_iteration
++;
1434 trace_kvm_dirty_ring_reaper("exit");
1436 rcu_unregister_thread();
1441 static int kvm_dirty_ring_reaper_init(KVMState
*s
)
1443 struct KVMDirtyRingReaper
*r
= &s
->reaper
;
1445 qemu_thread_create(&r
->reaper_thr
, "kvm-reaper",
1446 kvm_dirty_ring_reaper_thread
,
1447 s
, QEMU_THREAD_JOINABLE
);
1452 static void kvm_region_add(MemoryListener
*listener
,
1453 MemoryRegionSection
*section
)
1455 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1456 KVMMemoryUpdate
*update
;
1458 update
= g_new0(KVMMemoryUpdate
, 1);
1459 update
->section
= *section
;
1461 QSIMPLEQ_INSERT_TAIL(&kml
->transaction_add
, update
, next
);
1464 static void kvm_region_del(MemoryListener
*listener
,
1465 MemoryRegionSection
*section
)
1467 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1468 KVMMemoryUpdate
*update
;
1470 update
= g_new0(KVMMemoryUpdate
, 1);
1471 update
->section
= *section
;
1473 QSIMPLEQ_INSERT_TAIL(&kml
->transaction_del
, update
, next
);
1476 static void kvm_region_commit(MemoryListener
*listener
)
1478 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
,
1480 KVMMemoryUpdate
*u1
, *u2
;
1481 bool need_inhibit
= false;
1483 if (QSIMPLEQ_EMPTY(&kml
->transaction_add
) &&
1484 QSIMPLEQ_EMPTY(&kml
->transaction_del
)) {
1489 * We have to be careful when regions to add overlap with ranges to remove.
1490 * We have to simulate atomic KVM memslot updates by making sure no ioctl()
1491 * is currently active.
1493 * The lists are order by addresses, so it's easy to find overlaps.
1495 u1
= QSIMPLEQ_FIRST(&kml
->transaction_del
);
1496 u2
= QSIMPLEQ_FIRST(&kml
->transaction_add
);
1500 range_init_nofail(&r1
, u1
->section
.offset_within_address_space
,
1501 int128_get64(u1
->section
.size
));
1502 range_init_nofail(&r2
, u2
->section
.offset_within_address_space
,
1503 int128_get64(u2
->section
.size
));
1505 if (range_overlaps_range(&r1
, &r2
)) {
1506 need_inhibit
= true;
1509 if (range_lob(&r1
) < range_lob(&r2
)) {
1510 u1
= QSIMPLEQ_NEXT(u1
, next
);
1512 u2
= QSIMPLEQ_NEXT(u2
, next
);
1518 accel_ioctl_inhibit_begin();
1521 /* Remove all memslots before adding the new ones. */
1522 while (!QSIMPLEQ_EMPTY(&kml
->transaction_del
)) {
1523 u1
= QSIMPLEQ_FIRST(&kml
->transaction_del
);
1524 QSIMPLEQ_REMOVE_HEAD(&kml
->transaction_del
, next
);
1526 kvm_set_phys_mem(kml
, &u1
->section
, false);
1527 memory_region_unref(u1
->section
.mr
);
1531 while (!QSIMPLEQ_EMPTY(&kml
->transaction_add
)) {
1532 u1
= QSIMPLEQ_FIRST(&kml
->transaction_add
);
1533 QSIMPLEQ_REMOVE_HEAD(&kml
->transaction_add
, next
);
1535 memory_region_ref(u1
->section
.mr
);
1536 kvm_set_phys_mem(kml
, &u1
->section
, true);
1542 accel_ioctl_inhibit_end();
1547 static void kvm_log_sync(MemoryListener
*listener
,
1548 MemoryRegionSection
*section
)
1550 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1553 kvm_physical_sync_dirty_bitmap(kml
, section
);
1557 static void kvm_log_sync_global(MemoryListener
*l
)
1559 KVMMemoryListener
*kml
= container_of(l
, KVMMemoryListener
, listener
);
1560 KVMState
*s
= kvm_state
;
1564 /* Flush all kernel dirty addresses into KVMSlot dirty bitmap */
1565 kvm_dirty_ring_flush();
1568 * TODO: make this faster when nr_slots is big while there are
1569 * only a few used slots (small VMs).
1572 for (i
= 0; i
< s
->nr_slots
; i
++) {
1573 mem
= &kml
->slots
[i
];
1574 if (mem
->memory_size
&& mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1575 kvm_slot_sync_dirty_pages(mem
);
1577 * This is not needed by KVM_GET_DIRTY_LOG because the
1578 * ioctl will unconditionally overwrite the whole region.
1579 * However kvm dirty ring has no such side effect.
1581 kvm_slot_reset_dirty_pages(mem
);
1587 static void kvm_log_clear(MemoryListener
*listener
,
1588 MemoryRegionSection
*section
)
1590 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1593 r
= kvm_physical_log_clear(kml
, section
);
1595 error_report_once("%s: kvm log clear failed: mr=%s "
1596 "offset=%"HWADDR_PRIx
" size=%"PRIx64
, __func__
,
1597 section
->mr
->name
, section
->offset_within_region
,
1598 int128_get64(section
->size
));
1603 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
1604 MemoryRegionSection
*section
,
1605 bool match_data
, uint64_t data
,
1608 int fd
= event_notifier_get_fd(e
);
1611 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1612 data
, true, int128_get64(section
->size
),
1615 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1616 __func__
, strerror(-r
), -r
);
1621 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
1622 MemoryRegionSection
*section
,
1623 bool match_data
, uint64_t data
,
1626 int fd
= event_notifier_get_fd(e
);
1629 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1630 data
, false, int128_get64(section
->size
),
1633 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1634 __func__
, strerror(-r
), -r
);
1639 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
1640 MemoryRegionSection
*section
,
1641 bool match_data
, uint64_t data
,
1644 int fd
= event_notifier_get_fd(e
);
1647 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1648 data
, true, int128_get64(section
->size
),
1651 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1652 __func__
, strerror(-r
), -r
);
1657 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
1658 MemoryRegionSection
*section
,
1659 bool match_data
, uint64_t data
,
1663 int fd
= event_notifier_get_fd(e
);
1666 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1667 data
, false, int128_get64(section
->size
),
1670 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1671 __func__
, strerror(-r
), -r
);
1676 void kvm_memory_listener_register(KVMState
*s
, KVMMemoryListener
*kml
,
1677 AddressSpace
*as
, int as_id
, const char *name
)
1681 kml
->slots
= g_new0(KVMSlot
, s
->nr_slots
);
1684 for (i
= 0; i
< s
->nr_slots
; i
++) {
1685 kml
->slots
[i
].slot
= i
;
1688 QSIMPLEQ_INIT(&kml
->transaction_add
);
1689 QSIMPLEQ_INIT(&kml
->transaction_del
);
1691 kml
->listener
.region_add
= kvm_region_add
;
1692 kml
->listener
.region_del
= kvm_region_del
;
1693 kml
->listener
.commit
= kvm_region_commit
;
1694 kml
->listener
.log_start
= kvm_log_start
;
1695 kml
->listener
.log_stop
= kvm_log_stop
;
1696 kml
->listener
.priority
= 10;
1697 kml
->listener
.name
= name
;
1699 if (s
->kvm_dirty_ring_size
) {
1700 kml
->listener
.log_sync_global
= kvm_log_sync_global
;
1702 kml
->listener
.log_sync
= kvm_log_sync
;
1703 kml
->listener
.log_clear
= kvm_log_clear
;
1706 memory_listener_register(&kml
->listener
, as
);
1708 for (i
= 0; i
< s
->nr_as
; ++i
) {
1717 static MemoryListener kvm_io_listener
= {
1719 .eventfd_add
= kvm_io_ioeventfd_add
,
1720 .eventfd_del
= kvm_io_ioeventfd_del
,
1724 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
1726 struct kvm_irq_level event
;
1729 assert(kvm_async_interrupts_enabled());
1731 event
.level
= level
;
1733 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
1735 perror("kvm_set_irq");
1739 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
1742 #ifdef KVM_CAP_IRQ_ROUTING
1743 typedef struct KVMMSIRoute
{
1744 struct kvm_irq_routing_entry kroute
;
1745 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
1748 static void set_gsi(KVMState
*s
, unsigned int gsi
)
1750 set_bit(gsi
, s
->used_gsi_bitmap
);
1753 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
1755 clear_bit(gsi
, s
->used_gsi_bitmap
);
1758 void kvm_init_irq_routing(KVMState
*s
)
1762 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
1763 if (gsi_count
> 0) {
1764 /* Round up so we can search ints using ffs */
1765 s
->used_gsi_bitmap
= bitmap_new(gsi_count
);
1766 s
->gsi_count
= gsi_count
;
1769 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
1770 s
->nr_allocated_irq_routes
= 0;
1772 if (!kvm_direct_msi_allowed
) {
1773 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
1774 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
1778 kvm_arch_init_irq_routing(s
);
1781 void kvm_irqchip_commit_routes(KVMState
*s
)
1785 if (kvm_gsi_direct_mapping()) {
1789 if (!kvm_gsi_routing_enabled()) {
1793 s
->irq_routes
->flags
= 0;
1794 trace_kvm_irqchip_commit_routes();
1795 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
1799 static void kvm_add_routing_entry(KVMState
*s
,
1800 struct kvm_irq_routing_entry
*entry
)
1802 struct kvm_irq_routing_entry
*new;
1805 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
1806 n
= s
->nr_allocated_irq_routes
* 2;
1810 size
= sizeof(struct kvm_irq_routing
);
1811 size
+= n
* sizeof(*new);
1812 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
1813 s
->nr_allocated_irq_routes
= n
;
1815 n
= s
->irq_routes
->nr
++;
1816 new = &s
->irq_routes
->entries
[n
];
1820 set_gsi(s
, entry
->gsi
);
1823 static int kvm_update_routing_entry(KVMState
*s
,
1824 struct kvm_irq_routing_entry
*new_entry
)
1826 struct kvm_irq_routing_entry
*entry
;
1829 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
1830 entry
= &s
->irq_routes
->entries
[n
];
1831 if (entry
->gsi
!= new_entry
->gsi
) {
1835 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1839 *entry
= *new_entry
;
1847 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1849 struct kvm_irq_routing_entry e
= {};
1851 assert(pin
< s
->gsi_count
);
1854 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1856 e
.u
.irqchip
.irqchip
= irqchip
;
1857 e
.u
.irqchip
.pin
= pin
;
1858 kvm_add_routing_entry(s
, &e
);
1861 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1863 struct kvm_irq_routing_entry
*e
;
1866 if (kvm_gsi_direct_mapping()) {
1870 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1871 e
= &s
->irq_routes
->entries
[i
];
1872 if (e
->gsi
== virq
) {
1873 s
->irq_routes
->nr
--;
1874 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1878 kvm_arch_release_virq_post(virq
);
1879 trace_kvm_irqchip_release_virq(virq
);
1882 void kvm_irqchip_add_change_notifier(Notifier
*n
)
1884 notifier_list_add(&kvm_irqchip_change_notifiers
, n
);
1887 void kvm_irqchip_remove_change_notifier(Notifier
*n
)
1892 void kvm_irqchip_change_notify(void)
1894 notifier_list_notify(&kvm_irqchip_change_notifiers
, NULL
);
1897 static unsigned int kvm_hash_msi(uint32_t data
)
1899 /* This is optimized for IA32 MSI layout. However, no other arch shall
1900 * repeat the mistake of not providing a direct MSI injection API. */
1904 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1906 KVMMSIRoute
*route
, *next
;
1909 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1910 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1911 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1912 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1918 static int kvm_irqchip_get_virq(KVMState
*s
)
1923 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1924 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1925 * number can succeed even though a new route entry cannot be added.
1926 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1928 if (!kvm_direct_msi_allowed
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1929 kvm_flush_dynamic_msi_routes(s
);
1932 /* Return the lowest unused GSI in the bitmap */
1933 next_virq
= find_first_zero_bit(s
->used_gsi_bitmap
, s
->gsi_count
);
1934 if (next_virq
>= s
->gsi_count
) {
1941 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1943 unsigned int hash
= kvm_hash_msi(msg
.data
);
1946 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1947 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1948 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1949 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1956 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1961 if (kvm_direct_msi_allowed
) {
1962 msi
.address_lo
= (uint32_t)msg
.address
;
1963 msi
.address_hi
= msg
.address
>> 32;
1964 msi
.data
= le32_to_cpu(msg
.data
);
1966 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1968 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1971 route
= kvm_lookup_msi_route(s
, msg
);
1975 virq
= kvm_irqchip_get_virq(s
);
1980 route
= g_new0(KVMMSIRoute
, 1);
1981 route
->kroute
.gsi
= virq
;
1982 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1983 route
->kroute
.flags
= 0;
1984 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1985 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1986 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1988 kvm_add_routing_entry(s
, &route
->kroute
);
1989 kvm_irqchip_commit_routes(s
);
1991 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1995 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1997 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
2000 int kvm_irqchip_add_msi_route(KVMRouteChange
*c
, int vector
, PCIDevice
*dev
)
2002 struct kvm_irq_routing_entry kroute
= {};
2005 MSIMessage msg
= {0, 0};
2007 if (pci_available
&& dev
) {
2008 msg
= pci_get_msi_message(dev
, vector
);
2011 if (kvm_gsi_direct_mapping()) {
2012 return kvm_arch_msi_data_to_gsi(msg
.data
);
2015 if (!kvm_gsi_routing_enabled()) {
2019 virq
= kvm_irqchip_get_virq(s
);
2025 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
2027 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
2028 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
2029 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
2030 if (pci_available
&& kvm_msi_devid_required()) {
2031 kroute
.flags
= KVM_MSI_VALID_DEVID
;
2032 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
2034 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
2035 kvm_irqchip_release_virq(s
, virq
);
2039 trace_kvm_irqchip_add_msi_route(dev
? dev
->name
: (char *)"N/A",
2042 kvm_add_routing_entry(s
, &kroute
);
2043 kvm_arch_add_msi_route_post(&kroute
, vector
, dev
);
2049 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
,
2052 struct kvm_irq_routing_entry kroute
= {};
2054 if (kvm_gsi_direct_mapping()) {
2058 if (!kvm_irqchip_in_kernel()) {
2063 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
2065 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
2066 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
2067 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
2068 if (pci_available
&& kvm_msi_devid_required()) {
2069 kroute
.flags
= KVM_MSI_VALID_DEVID
;
2070 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
2072 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
2076 trace_kvm_irqchip_update_msi_route(virq
);
2078 return kvm_update_routing_entry(s
, &kroute
);
2081 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
2082 EventNotifier
*resample
, int virq
,
2085 int fd
= event_notifier_get_fd(event
);
2086 int rfd
= resample
? event_notifier_get_fd(resample
) : -1;
2088 struct kvm_irqfd irqfd
= {
2091 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
2096 if (kvm_irqchip_is_split()) {
2098 * When the slow irqchip (e.g. IOAPIC) is in the
2099 * userspace, KVM kernel resamplefd will not work because
2100 * the EOI of the interrupt will be delivered to userspace
2101 * instead, so the KVM kernel resamplefd kick will be
2102 * skipped. The userspace here mimics what the kernel
2103 * provides with resamplefd, remember the resamplefd and
2104 * kick it when we receive EOI of this IRQ.
2106 * This is hackery because IOAPIC is mostly bypassed
2107 * (except EOI broadcasts) when irqfd is used. However
2108 * this can bring much performance back for split irqchip
2109 * with INTx IRQs (for VFIO, this gives 93% perf of the
2110 * full fast path, which is 46% perf boost comparing to
2111 * the INTx slow path).
2113 kvm_resample_fd_insert(virq
, resample
);
2115 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
2116 irqfd
.resamplefd
= rfd
;
2118 } else if (!assign
) {
2119 if (kvm_irqchip_is_split()) {
2120 kvm_resample_fd_remove(virq
);
2124 if (!kvm_irqfds_enabled()) {
2128 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
2131 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
2133 struct kvm_irq_routing_entry kroute
= {};
2136 if (!kvm_gsi_routing_enabled()) {
2140 virq
= kvm_irqchip_get_virq(s
);
2146 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
2148 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
2149 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
2150 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
2151 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
2152 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
2154 kvm_add_routing_entry(s
, &kroute
);
2159 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
2161 struct kvm_irq_routing_entry kroute
= {};
2164 if (!kvm_gsi_routing_enabled()) {
2167 if (!kvm_check_extension(s
, KVM_CAP_HYPERV_SYNIC
)) {
2170 virq
= kvm_irqchip_get_virq(s
);
2176 kroute
.type
= KVM_IRQ_ROUTING_HV_SINT
;
2178 kroute
.u
.hv_sint
.vcpu
= vcpu
;
2179 kroute
.u
.hv_sint
.sint
= sint
;
2181 kvm_add_routing_entry(s
, &kroute
);
2182 kvm_irqchip_commit_routes(s
);
2187 #else /* !KVM_CAP_IRQ_ROUTING */
2189 void kvm_init_irq_routing(KVMState
*s
)
2193 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
2197 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
2202 int kvm_irqchip_add_msi_route(KVMRouteChange
*c
, int vector
, PCIDevice
*dev
)
2207 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
2212 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
2217 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
2218 EventNotifier
*resample
, int virq
,
2224 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
2228 #endif /* !KVM_CAP_IRQ_ROUTING */
2230 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
2231 EventNotifier
*rn
, int virq
)
2233 return kvm_irqchip_assign_irqfd(s
, n
, rn
, virq
, true);
2236 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
2239 return kvm_irqchip_assign_irqfd(s
, n
, NULL
, virq
, false);
2242 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
2243 EventNotifier
*rn
, qemu_irq irq
)
2246 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
2251 return kvm_irqchip_add_irqfd_notifier_gsi(s
, n
, rn
, GPOINTER_TO_INT(gsi
));
2254 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
2258 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
2263 return kvm_irqchip_remove_irqfd_notifier_gsi(s
, n
, GPOINTER_TO_INT(gsi
));
2266 void kvm_irqchip_set_qemuirq_gsi(KVMState
*s
, qemu_irq irq
, int gsi
)
2268 g_hash_table_insert(s
->gsimap
, irq
, GINT_TO_POINTER(gsi
));
2271 static void kvm_irqchip_create(KVMState
*s
)
2275 assert(s
->kernel_irqchip_split
!= ON_OFF_AUTO_AUTO
);
2276 if (kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
2278 } else if (kvm_check_extension(s
, KVM_CAP_S390_IRQCHIP
)) {
2279 ret
= kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0);
2281 fprintf(stderr
, "Enable kernel irqchip failed: %s\n", strerror(-ret
));
2288 /* First probe and see if there's a arch-specific hook to create the
2289 * in-kernel irqchip for us */
2290 ret
= kvm_arch_irqchip_create(s
);
2292 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_ON
) {
2293 error_report("Split IRQ chip mode not supported.");
2296 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
2300 fprintf(stderr
, "Create kernel irqchip failed: %s\n", strerror(-ret
));
2304 kvm_kernel_irqchip
= true;
2305 /* If we have an in-kernel IRQ chip then we must have asynchronous
2306 * interrupt delivery (though the reverse is not necessarily true)
2308 kvm_async_interrupts_allowed
= true;
2309 kvm_halt_in_kernel_allowed
= true;
2311 kvm_init_irq_routing(s
);
2313 s
->gsimap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2316 /* Find number of supported CPUs using the recommended
2317 * procedure from the kernel API documentation to cope with
2318 * older kernels that may be missing capabilities.
2320 static int kvm_recommended_vcpus(KVMState
*s
)
2322 int ret
= kvm_vm_check_extension(s
, KVM_CAP_NR_VCPUS
);
2323 return (ret
) ? ret
: 4;
2326 static int kvm_max_vcpus(KVMState
*s
)
2328 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
2329 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
2332 static int kvm_max_vcpu_id(KVMState
*s
)
2334 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPU_ID
);
2335 return (ret
) ? ret
: kvm_max_vcpus(s
);
2338 bool kvm_vcpu_id_is_valid(int vcpu_id
)
2340 KVMState
*s
= KVM_STATE(current_accel());
2341 return vcpu_id
>= 0 && vcpu_id
< kvm_max_vcpu_id(s
);
2344 bool kvm_dirty_ring_enabled(void)
2346 return kvm_state
->kvm_dirty_ring_size
? true : false;
2349 static void query_stats_cb(StatsResultList
**result
, StatsTarget target
,
2350 strList
*names
, strList
*targets
, Error
**errp
);
2351 static void query_stats_schemas_cb(StatsSchemaList
**result
, Error
**errp
);
2353 uint32_t kvm_dirty_ring_size(void)
2355 return kvm_state
->kvm_dirty_ring_size
;
2358 static int kvm_init(MachineState
*ms
)
2360 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
2361 static const char upgrade_note
[] =
2362 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
2363 "(see http://sourceforge.net/projects/kvm).\n";
2368 { "SMP", ms
->smp
.cpus
},
2369 { "hotpluggable", ms
->smp
.max_cpus
},
2372 int soft_vcpus_limit
, hard_vcpus_limit
;
2374 const KVMCapabilityInfo
*missing_cap
;
2377 uint64_t dirty_log_manual_caps
;
2379 qemu_mutex_init(&kml_slots_lock
);
2381 s
= KVM_STATE(ms
->accelerator
);
2384 * On systems where the kernel can support different base page
2385 * sizes, host page size may be different from TARGET_PAGE_SIZE,
2386 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
2387 * page size for the system though.
2389 assert(TARGET_PAGE_SIZE
<= qemu_real_host_page_size());
2392 accel_blocker_init();
2394 #ifdef KVM_CAP_SET_GUEST_DEBUG
2395 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
2397 QLIST_INIT(&s
->kvm_parked_vcpus
);
2398 s
->fd
= qemu_open_old("/dev/kvm", O_RDWR
);
2400 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
2405 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
2406 if (ret
< KVM_API_VERSION
) {
2410 fprintf(stderr
, "kvm version too old\n");
2414 if (ret
> KVM_API_VERSION
) {
2416 fprintf(stderr
, "kvm version not supported\n");
2420 kvm_immediate_exit
= kvm_check_extension(s
, KVM_CAP_IMMEDIATE_EXIT
);
2421 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
2423 /* If unspecified, use the default value */
2428 s
->nr_as
= kvm_check_extension(s
, KVM_CAP_MULTI_ADDRESS_SPACE
);
2429 if (s
->nr_as
<= 1) {
2432 s
->as
= g_new0(struct KVMAs
, s
->nr_as
);
2434 if (object_property_find(OBJECT(current_machine
), "kvm-type")) {
2435 g_autofree
char *kvm_type
= object_property_get_str(OBJECT(current_machine
),
2438 type
= mc
->kvm_type(ms
, kvm_type
);
2439 } else if (mc
->kvm_type
) {
2440 type
= mc
->kvm_type(ms
, NULL
);
2444 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
2445 } while (ret
== -EINTR
);
2448 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
2452 if (ret
== -EINVAL
) {
2454 "Host kernel setup problem detected. Please verify:\n");
2455 fprintf(stderr
, "- for kernels supporting the switch_amode or"
2456 " user_mode parameters, whether\n");
2458 " user space is running in primary address space\n");
2460 "- for kernels supporting the vm.allocate_pgste sysctl, "
2461 "whether it is enabled\n");
2463 #elif defined(TARGET_PPC)
2464 if (ret
== -EINVAL
) {
2466 "PPC KVM module is not loaded. Try modprobe kvm_%s.\n",
2467 (type
== 2) ? "pr" : "hv");
2475 /* check the vcpu limits */
2476 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
2477 hard_vcpus_limit
= kvm_max_vcpus(s
);
2480 if (nc
->num
> soft_vcpus_limit
) {
2481 warn_report("Number of %s cpus requested (%d) exceeds "
2482 "the recommended cpus supported by KVM (%d)",
2483 nc
->name
, nc
->num
, soft_vcpus_limit
);
2485 if (nc
->num
> hard_vcpus_limit
) {
2486 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
2487 "the maximum cpus supported by KVM (%d)\n",
2488 nc
->name
, nc
->num
, hard_vcpus_limit
);
2495 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
2498 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
2502 fprintf(stderr
, "kvm does not support %s\n%s",
2503 missing_cap
->name
, upgrade_note
);
2507 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
2508 s
->coalesced_pio
= s
->coalesced_mmio
&&
2509 kvm_check_extension(s
, KVM_CAP_COALESCED_PIO
);
2512 * Enable KVM dirty ring if supported, otherwise fall back to
2513 * dirty logging mode
2515 if (s
->kvm_dirty_ring_size
> 0) {
2516 uint64_t ring_bytes
;
2518 ring_bytes
= s
->kvm_dirty_ring_size
* sizeof(struct kvm_dirty_gfn
);
2520 /* Read the max supported pages */
2521 ret
= kvm_vm_check_extension(s
, KVM_CAP_DIRTY_LOG_RING
);
2523 if (ring_bytes
> ret
) {
2524 error_report("KVM dirty ring size %" PRIu32
" too big "
2525 "(maximum is %ld). Please use a smaller value.",
2526 s
->kvm_dirty_ring_size
,
2527 (long)ret
/ sizeof(struct kvm_dirty_gfn
));
2532 ret
= kvm_vm_enable_cap(s
, KVM_CAP_DIRTY_LOG_RING
, 0, ring_bytes
);
2534 error_report("Enabling of KVM dirty ring failed: %s. "
2535 "Suggested minimum value is 1024.", strerror(-ret
));
2539 s
->kvm_dirty_ring_bytes
= ring_bytes
;
2541 warn_report("KVM dirty ring not available, using bitmap method");
2542 s
->kvm_dirty_ring_size
= 0;
2547 * KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 is not needed when dirty ring is
2548 * enabled. More importantly, KVM_DIRTY_LOG_INITIALLY_SET will assume no
2549 * page is wr-protected initially, which is against how kvm dirty ring is
2550 * usage - kvm dirty ring requires all pages are wr-protected at the very
2551 * beginning. Enabling this feature for dirty ring causes data corruption.
2553 * TODO: Without KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 and kvm clear dirty log,
2554 * we may expect a higher stall time when starting the migration. In the
2555 * future we can enable KVM_CLEAR_DIRTY_LOG to work with dirty ring too:
2556 * instead of clearing dirty bit, it can be a way to explicitly wr-protect
2559 if (!s
->kvm_dirty_ring_size
) {
2560 dirty_log_manual_caps
=
2561 kvm_check_extension(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
);
2562 dirty_log_manual_caps
&= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE
|
2563 KVM_DIRTY_LOG_INITIALLY_SET
);
2564 s
->manual_dirty_log_protect
= dirty_log_manual_caps
;
2565 if (dirty_log_manual_caps
) {
2566 ret
= kvm_vm_enable_cap(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
, 0,
2567 dirty_log_manual_caps
);
2569 warn_report("Trying to enable capability %"PRIu64
" of "
2570 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2571 "Falling back to the legacy mode. ",
2572 dirty_log_manual_caps
);
2573 s
->manual_dirty_log_protect
= 0;
2578 #ifdef KVM_CAP_VCPU_EVENTS
2579 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
2582 s
->robust_singlestep
=
2583 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
2585 #ifdef KVM_CAP_DEBUGREGS
2586 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
2589 s
->max_nested_state_len
= kvm_check_extension(s
, KVM_CAP_NESTED_STATE
);
2591 #ifdef KVM_CAP_IRQ_ROUTING
2592 kvm_direct_msi_allowed
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
2595 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
2597 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
2598 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
2599 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
2602 kvm_readonly_mem_allowed
=
2603 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
2605 kvm_eventfds_allowed
=
2606 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
2608 kvm_irqfds_allowed
=
2609 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
2611 kvm_resamplefds_allowed
=
2612 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
2614 kvm_vm_attributes_allowed
=
2615 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
2617 kvm_ioeventfd_any_length_allowed
=
2618 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD_ANY_LENGTH
) > 0);
2620 #ifdef KVM_CAP_SET_GUEST_DEBUG
2621 kvm_has_guest_debug
=
2622 (kvm_check_extension(s
, KVM_CAP_SET_GUEST_DEBUG
) > 0);
2625 kvm_sstep_flags
= 0;
2626 if (kvm_has_guest_debug
) {
2627 kvm_sstep_flags
= SSTEP_ENABLE
;
2629 #if defined KVM_CAP_SET_GUEST_DEBUG2
2630 int guest_debug_flags
=
2631 kvm_check_extension(s
, KVM_CAP_SET_GUEST_DEBUG2
);
2633 if (guest_debug_flags
& KVM_GUESTDBG_BLOCKIRQ
) {
2634 kvm_sstep_flags
|= SSTEP_NOIRQ
;
2641 ret
= kvm_arch_init(ms
, s
);
2646 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_AUTO
) {
2647 s
->kernel_irqchip_split
= mc
->default_kernel_irqchip_split
? ON_OFF_AUTO_ON
: ON_OFF_AUTO_OFF
;
2650 qemu_register_reset(kvm_unpoison_all
, NULL
);
2652 if (s
->kernel_irqchip_allowed
) {
2653 kvm_irqchip_create(s
);
2656 if (kvm_eventfds_allowed
) {
2657 s
->memory_listener
.listener
.eventfd_add
= kvm_mem_ioeventfd_add
;
2658 s
->memory_listener
.listener
.eventfd_del
= kvm_mem_ioeventfd_del
;
2660 s
->memory_listener
.listener
.coalesced_io_add
= kvm_coalesce_mmio_region
;
2661 s
->memory_listener
.listener
.coalesced_io_del
= kvm_uncoalesce_mmio_region
;
2663 kvm_memory_listener_register(s
, &s
->memory_listener
,
2664 &address_space_memory
, 0, "kvm-memory");
2665 if (kvm_eventfds_allowed
) {
2666 memory_listener_register(&kvm_io_listener
,
2669 memory_listener_register(&kvm_coalesced_pio_listener
,
2672 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
2674 s
->sync_mmu
= !!kvm_vm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
2676 ret
= ram_block_discard_disable(true);
2680 if (s
->kvm_dirty_ring_size
) {
2681 ret
= kvm_dirty_ring_reaper_init(s
);
2687 if (kvm_check_extension(kvm_state
, KVM_CAP_BINARY_STATS_FD
)) {
2688 add_stats_callbacks(STATS_PROVIDER_KVM
, query_stats_cb
,
2689 query_stats_schemas_cb
);
2702 g_free(s
->memory_listener
.slots
);
2707 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
2709 s
->sigmask_len
= sigmask_len
;
2712 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
2713 int size
, uint32_t count
)
2716 uint8_t *ptr
= data
;
2718 for (i
= 0; i
< count
; i
++) {
2719 address_space_rw(&address_space_io
, port
, attrs
,
2721 direction
== KVM_EXIT_IO_OUT
);
2726 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
2728 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
2729 run
->internal
.suberror
);
2731 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
2734 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
2735 fprintf(stderr
, "extra data[%d]: 0x%016"PRIx64
"\n",
2736 i
, (uint64_t)run
->internal
.data
[i
]);
2739 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
2740 fprintf(stderr
, "emulation failure\n");
2741 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
2742 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
2743 return EXCP_INTERRUPT
;
2746 /* FIXME: Should trigger a qmp message to let management know
2747 * something went wrong.
2752 void kvm_flush_coalesced_mmio_buffer(void)
2754 KVMState
*s
= kvm_state
;
2756 if (s
->coalesced_flush_in_progress
) {
2760 s
->coalesced_flush_in_progress
= true;
2762 if (s
->coalesced_mmio_ring
) {
2763 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
2764 while (ring
->first
!= ring
->last
) {
2765 struct kvm_coalesced_mmio
*ent
;
2767 ent
= &ring
->coalesced_mmio
[ring
->first
];
2769 if (ent
->pio
== 1) {
2770 address_space_write(&address_space_io
, ent
->phys_addr
,
2771 MEMTXATTRS_UNSPECIFIED
, ent
->data
,
2774 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
2777 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
2781 s
->coalesced_flush_in_progress
= false;
2784 bool kvm_cpu_check_are_resettable(void)
2786 return kvm_arch_cpu_check_are_resettable();
2789 static void do_kvm_cpu_synchronize_state(CPUState
*cpu
, run_on_cpu_data arg
)
2791 if (!cpu
->vcpu_dirty
) {
2792 kvm_arch_get_registers(cpu
);
2793 cpu
->vcpu_dirty
= true;
2797 void kvm_cpu_synchronize_state(CPUState
*cpu
)
2799 if (!cpu
->vcpu_dirty
) {
2800 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, RUN_ON_CPU_NULL
);
2804 static void do_kvm_cpu_synchronize_post_reset(CPUState
*cpu
, run_on_cpu_data arg
)
2806 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
2807 cpu
->vcpu_dirty
= false;
2810 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
2812 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, RUN_ON_CPU_NULL
);
2815 static void do_kvm_cpu_synchronize_post_init(CPUState
*cpu
, run_on_cpu_data arg
)
2817 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
2818 cpu
->vcpu_dirty
= false;
2821 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
2823 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, RUN_ON_CPU_NULL
);
2826 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
, run_on_cpu_data arg
)
2828 cpu
->vcpu_dirty
= true;
2831 void kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
)
2833 run_on_cpu(cpu
, do_kvm_cpu_synchronize_pre_loadvm
, RUN_ON_CPU_NULL
);
2836 #ifdef KVM_HAVE_MCE_INJECTION
2837 static __thread
void *pending_sigbus_addr
;
2838 static __thread
int pending_sigbus_code
;
2839 static __thread
bool have_sigbus_pending
;
2842 static void kvm_cpu_kick(CPUState
*cpu
)
2844 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 1);
2847 static void kvm_cpu_kick_self(void)
2849 if (kvm_immediate_exit
) {
2850 kvm_cpu_kick(current_cpu
);
2852 qemu_cpu_kick_self();
2856 static void kvm_eat_signals(CPUState
*cpu
)
2858 struct timespec ts
= { 0, 0 };
2864 if (kvm_immediate_exit
) {
2865 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 0);
2866 /* Write kvm_run->immediate_exit before the cpu->exit_request
2867 * write in kvm_cpu_exec.
2873 sigemptyset(&waitset
);
2874 sigaddset(&waitset
, SIG_IPI
);
2877 r
= sigtimedwait(&waitset
, &siginfo
, &ts
);
2878 if (r
== -1 && !(errno
== EAGAIN
|| errno
== EINTR
)) {
2879 perror("sigtimedwait");
2883 r
= sigpending(&chkset
);
2885 perror("sigpending");
2888 } while (sigismember(&chkset
, SIG_IPI
));
2891 int kvm_cpu_exec(CPUState
*cpu
)
2893 struct kvm_run
*run
= cpu
->kvm_run
;
2896 DPRINTF("kvm_cpu_exec()\n");
2898 if (kvm_arch_process_async_events(cpu
)) {
2899 qatomic_set(&cpu
->exit_request
, 0);
2903 qemu_mutex_unlock_iothread();
2904 cpu_exec_start(cpu
);
2909 if (cpu
->vcpu_dirty
) {
2910 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
2911 cpu
->vcpu_dirty
= false;
2914 kvm_arch_pre_run(cpu
, run
);
2915 if (qatomic_read(&cpu
->exit_request
)) {
2916 DPRINTF("interrupt exit requested\n");
2918 * KVM requires us to reenter the kernel after IO exits to complete
2919 * instruction emulation. This self-signal will ensure that we
2922 kvm_cpu_kick_self();
2925 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2926 * Matching barrier in kvm_eat_signals.
2930 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
2932 attrs
= kvm_arch_post_run(cpu
, run
);
2934 #ifdef KVM_HAVE_MCE_INJECTION
2935 if (unlikely(have_sigbus_pending
)) {
2936 qemu_mutex_lock_iothread();
2937 kvm_arch_on_sigbus_vcpu(cpu
, pending_sigbus_code
,
2938 pending_sigbus_addr
);
2939 have_sigbus_pending
= false;
2940 qemu_mutex_unlock_iothread();
2945 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
2946 DPRINTF("io window exit\n");
2947 kvm_eat_signals(cpu
);
2948 ret
= EXCP_INTERRUPT
;
2951 fprintf(stderr
, "error: kvm run failed %s\n",
2952 strerror(-run_ret
));
2954 if (run_ret
== -EBUSY
) {
2956 "This is probably because your SMT is enabled.\n"
2957 "VCPU can only run on primary threads with all "
2958 "secondary threads offline.\n");
2965 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
2966 switch (run
->exit_reason
) {
2968 DPRINTF("handle_io\n");
2969 /* Called outside BQL */
2970 kvm_handle_io(run
->io
.port
, attrs
,
2971 (uint8_t *)run
+ run
->io
.data_offset
,
2978 DPRINTF("handle_mmio\n");
2979 /* Called outside BQL */
2980 address_space_rw(&address_space_memory
,
2981 run
->mmio
.phys_addr
, attrs
,
2984 run
->mmio
.is_write
);
2987 case KVM_EXIT_IRQ_WINDOW_OPEN
:
2988 DPRINTF("irq_window_open\n");
2989 ret
= EXCP_INTERRUPT
;
2991 case KVM_EXIT_SHUTDOWN
:
2992 DPRINTF("shutdown\n");
2993 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
2994 ret
= EXCP_INTERRUPT
;
2996 case KVM_EXIT_UNKNOWN
:
2997 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
2998 (uint64_t)run
->hw
.hardware_exit_reason
);
3001 case KVM_EXIT_INTERNAL_ERROR
:
3002 ret
= kvm_handle_internal_error(cpu
, run
);
3004 case KVM_EXIT_DIRTY_RING_FULL
:
3006 * We shouldn't continue if the dirty ring of this vcpu is
3007 * still full. Got kicked by KVM_RESET_DIRTY_RINGS.
3009 trace_kvm_dirty_ring_full(cpu
->cpu_index
);
3010 qemu_mutex_lock_iothread();
3012 * We throttle vCPU by making it sleep once it exit from kernel
3013 * due to dirty ring full. In the dirtylimit scenario, reaping
3014 * all vCPUs after a single vCPU dirty ring get full result in
3015 * the miss of sleep, so just reap the ring-fulled vCPU.
3017 if (dirtylimit_in_service()) {
3018 kvm_dirty_ring_reap(kvm_state
, cpu
);
3020 kvm_dirty_ring_reap(kvm_state
, NULL
);
3022 qemu_mutex_unlock_iothread();
3023 dirtylimit_vcpu_execute(cpu
);
3026 case KVM_EXIT_SYSTEM_EVENT
:
3027 switch (run
->system_event
.type
) {
3028 case KVM_SYSTEM_EVENT_SHUTDOWN
:
3029 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN
);
3030 ret
= EXCP_INTERRUPT
;
3032 case KVM_SYSTEM_EVENT_RESET
:
3033 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
3034 ret
= EXCP_INTERRUPT
;
3036 case KVM_SYSTEM_EVENT_CRASH
:
3037 kvm_cpu_synchronize_state(cpu
);
3038 qemu_mutex_lock_iothread();
3039 qemu_system_guest_panicked(cpu_get_crash_info(cpu
));
3040 qemu_mutex_unlock_iothread();
3044 DPRINTF("kvm_arch_handle_exit\n");
3045 ret
= kvm_arch_handle_exit(cpu
, run
);
3050 DPRINTF("kvm_arch_handle_exit\n");
3051 ret
= kvm_arch_handle_exit(cpu
, run
);
3057 qemu_mutex_lock_iothread();
3060 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
3061 vm_stop(RUN_STATE_INTERNAL_ERROR
);
3064 qatomic_set(&cpu
->exit_request
, 0);
3068 int kvm_ioctl(KVMState
*s
, int type
, ...)
3075 arg
= va_arg(ap
, void *);
3078 trace_kvm_ioctl(type
, arg
);
3079 ret
= ioctl(s
->fd
, type
, arg
);
3086 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
3093 arg
= va_arg(ap
, void *);
3096 trace_kvm_vm_ioctl(type
, arg
);
3097 accel_ioctl_begin();
3098 ret
= ioctl(s
->vmfd
, type
, arg
);
3106 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
3113 arg
= va_arg(ap
, void *);
3116 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
3117 accel_cpu_ioctl_begin(cpu
);
3118 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
3119 accel_cpu_ioctl_end(cpu
);
3126 int kvm_device_ioctl(int fd
, int type
, ...)
3133 arg
= va_arg(ap
, void *);
3136 trace_kvm_device_ioctl(fd
, type
, arg
);
3137 accel_ioctl_begin();
3138 ret
= ioctl(fd
, type
, arg
);
3146 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
3149 struct kvm_device_attr attribute
= {
3154 if (!kvm_vm_attributes_allowed
) {
3158 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
3159 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
3163 int kvm_device_check_attr(int dev_fd
, uint32_t group
, uint64_t attr
)
3165 struct kvm_device_attr attribute
= {
3171 return kvm_device_ioctl(dev_fd
, KVM_HAS_DEVICE_ATTR
, &attribute
) ? 0 : 1;
3174 int kvm_device_access(int fd
, int group
, uint64_t attr
,
3175 void *val
, bool write
, Error
**errp
)
3177 struct kvm_device_attr kvmattr
;
3181 kvmattr
.group
= group
;
3182 kvmattr
.attr
= attr
;
3183 kvmattr
.addr
= (uintptr_t)val
;
3185 err
= kvm_device_ioctl(fd
,
3186 write
? KVM_SET_DEVICE_ATTR
: KVM_GET_DEVICE_ATTR
,
3189 error_setg_errno(errp
, -err
,
3190 "KVM_%s_DEVICE_ATTR failed: Group %d "
3191 "attr 0x%016" PRIx64
,
3192 write
? "SET" : "GET", group
, attr
);
3197 bool kvm_has_sync_mmu(void)
3199 return kvm_state
->sync_mmu
;
3202 int kvm_has_vcpu_events(void)
3204 return kvm_state
->vcpu_events
;
3207 int kvm_has_robust_singlestep(void)
3209 return kvm_state
->robust_singlestep
;
3212 int kvm_has_debugregs(void)
3214 return kvm_state
->debugregs
;
3217 int kvm_max_nested_state_length(void)
3219 return kvm_state
->max_nested_state_len
;
3222 int kvm_has_many_ioeventfds(void)
3224 if (!kvm_enabled()) {
3227 return kvm_state
->many_ioeventfds
;
3230 int kvm_has_gsi_routing(void)
3232 #ifdef KVM_CAP_IRQ_ROUTING
3233 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
3239 int kvm_has_intx_set_mask(void)
3241 return kvm_state
->intx_set_mask
;
3244 bool kvm_arm_supports_user_irq(void)
3246 return kvm_check_extension(kvm_state
, KVM_CAP_ARM_USER_IRQ
);
3249 #ifdef KVM_CAP_SET_GUEST_DEBUG
3250 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
3253 struct kvm_sw_breakpoint
*bp
;
3255 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
3263 int kvm_sw_breakpoints_active(CPUState
*cpu
)
3265 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
3268 struct kvm_set_guest_debug_data
{
3269 struct kvm_guest_debug dbg
;
3273 static void kvm_invoke_set_guest_debug(CPUState
*cpu
, run_on_cpu_data data
)
3275 struct kvm_set_guest_debug_data
*dbg_data
=
3276 (struct kvm_set_guest_debug_data
*) data
.host_ptr
;
3278 dbg_data
->err
= kvm_vcpu_ioctl(cpu
, KVM_SET_GUEST_DEBUG
,
3282 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
3284 struct kvm_set_guest_debug_data data
;
3286 data
.dbg
.control
= reinject_trap
;
3288 if (cpu
->singlestep_enabled
) {
3289 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
3291 if (cpu
->singlestep_enabled
& SSTEP_NOIRQ
) {
3292 data
.dbg
.control
|= KVM_GUESTDBG_BLOCKIRQ
;
3295 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
3297 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
,
3298 RUN_ON_CPU_HOST_PTR(&data
));
3302 bool kvm_supports_guest_debug(void)
3304 /* probed during kvm_init() */
3305 return kvm_has_guest_debug
;
3308 int kvm_insert_breakpoint(CPUState
*cpu
, int type
, vaddr addr
, vaddr len
)
3310 struct kvm_sw_breakpoint
*bp
;
3313 if (type
== GDB_BREAKPOINT_SW
) {
3314 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
3320 bp
= g_new(struct kvm_sw_breakpoint
, 1);
3323 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
3329 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
3331 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
3338 err
= kvm_update_guest_debug(cpu
, 0);
3346 int kvm_remove_breakpoint(CPUState
*cpu
, int type
, vaddr addr
, vaddr len
)
3348 struct kvm_sw_breakpoint
*bp
;
3351 if (type
== GDB_BREAKPOINT_SW
) {
3352 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
3357 if (bp
->use_count
> 1) {
3362 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
3367 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
3370 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
3377 err
= kvm_update_guest_debug(cpu
, 0);
3385 void kvm_remove_all_breakpoints(CPUState
*cpu
)
3387 struct kvm_sw_breakpoint
*bp
, *next
;
3388 KVMState
*s
= cpu
->kvm_state
;
3391 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
3392 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
3393 /* Try harder to find a CPU that currently sees the breakpoint. */
3394 CPU_FOREACH(tmpcpu
) {
3395 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
3400 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
3403 kvm_arch_remove_all_hw_breakpoints();
3406 kvm_update_guest_debug(cpu
, 0);
3410 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
3412 static int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
3414 KVMState
*s
= kvm_state
;
3415 struct kvm_signal_mask
*sigmask
;
3418 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
3420 sigmask
->len
= s
->sigmask_len
;
3421 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
3422 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
3428 static void kvm_ipi_signal(int sig
)
3431 assert(kvm_immediate_exit
);
3432 kvm_cpu_kick(current_cpu
);
3436 void kvm_init_cpu_signals(CPUState
*cpu
)
3440 struct sigaction sigact
;
3442 memset(&sigact
, 0, sizeof(sigact
));
3443 sigact
.sa_handler
= kvm_ipi_signal
;
3444 sigaction(SIG_IPI
, &sigact
, NULL
);
3446 pthread_sigmask(SIG_BLOCK
, NULL
, &set
);
3447 #if defined KVM_HAVE_MCE_INJECTION
3448 sigdelset(&set
, SIGBUS
);
3449 pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
3451 sigdelset(&set
, SIG_IPI
);
3452 if (kvm_immediate_exit
) {
3453 r
= pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
3455 r
= kvm_set_signal_mask(cpu
, &set
);
3458 fprintf(stderr
, "kvm_set_signal_mask: %s\n", strerror(-r
));
3463 /* Called asynchronously in VCPU thread. */
3464 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
3466 #ifdef KVM_HAVE_MCE_INJECTION
3467 if (have_sigbus_pending
) {
3470 have_sigbus_pending
= true;
3471 pending_sigbus_addr
= addr
;
3472 pending_sigbus_code
= code
;
3473 qatomic_set(&cpu
->exit_request
, 1);
3480 /* Called synchronously (via signalfd) in main thread. */
3481 int kvm_on_sigbus(int code
, void *addr
)
3483 #ifdef KVM_HAVE_MCE_INJECTION
3484 /* Action required MCE kills the process if SIGBUS is blocked. Because
3485 * that's what happens in the I/O thread, where we handle MCE via signalfd,
3486 * we can only get action optional here.
3488 assert(code
!= BUS_MCEERR_AR
);
3489 kvm_arch_on_sigbus_vcpu(first_cpu
, code
, addr
);
3496 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
3499 struct kvm_create_device create_dev
;
3501 create_dev
.type
= type
;
3503 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
3505 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
3509 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
3514 return test
? 0 : create_dev
.fd
;
3517 bool kvm_device_supported(int vmfd
, uint64_t type
)
3519 struct kvm_create_device create_dev
= {
3522 .flags
= KVM_CREATE_DEVICE_TEST
,
3525 if (ioctl(vmfd
, KVM_CHECK_EXTENSION
, KVM_CAP_DEVICE_CTRL
) <= 0) {
3529 return (ioctl(vmfd
, KVM_CREATE_DEVICE
, &create_dev
) >= 0);
3532 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
3534 struct kvm_one_reg reg
;
3538 reg
.addr
= (uintptr_t) source
;
3539 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
3541 trace_kvm_failed_reg_set(id
, strerror(-r
));
3546 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
3548 struct kvm_one_reg reg
;
3552 reg
.addr
= (uintptr_t) target
;
3553 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
3555 trace_kvm_failed_reg_get(id
, strerror(-r
));
3560 static bool kvm_accel_has_memory(MachineState
*ms
, AddressSpace
*as
,
3561 hwaddr start_addr
, hwaddr size
)
3563 KVMState
*kvm
= KVM_STATE(ms
->accelerator
);
3566 for (i
= 0; i
< kvm
->nr_as
; ++i
) {
3567 if (kvm
->as
[i
].as
== as
&& kvm
->as
[i
].ml
) {
3568 size
= MIN(kvm_max_slot_size
, size
);
3569 return NULL
!= kvm_lookup_matching_slot(kvm
->as
[i
].ml
,
3577 static void kvm_get_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3578 const char *name
, void *opaque
,
3581 KVMState
*s
= KVM_STATE(obj
);
3582 int64_t value
= s
->kvm_shadow_mem
;
3584 visit_type_int(v
, name
, &value
, errp
);
3587 static void kvm_set_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3588 const char *name
, void *opaque
,
3591 KVMState
*s
= KVM_STATE(obj
);
3595 error_setg(errp
, "Cannot set properties after the accelerator has been initialized");
3599 if (!visit_type_int(v
, name
, &value
, errp
)) {
3603 s
->kvm_shadow_mem
= value
;
3606 static void kvm_set_kernel_irqchip(Object
*obj
, Visitor
*v
,
3607 const char *name
, void *opaque
,
3610 KVMState
*s
= KVM_STATE(obj
);
3614 error_setg(errp
, "Cannot set properties after the accelerator has been initialized");
3618 if (!visit_type_OnOffSplit(v
, name
, &mode
, errp
)) {
3622 case ON_OFF_SPLIT_ON
:
3623 s
->kernel_irqchip_allowed
= true;
3624 s
->kernel_irqchip_required
= true;
3625 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3627 case ON_OFF_SPLIT_OFF
:
3628 s
->kernel_irqchip_allowed
= false;
3629 s
->kernel_irqchip_required
= false;
3630 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3632 case ON_OFF_SPLIT_SPLIT
:
3633 s
->kernel_irqchip_allowed
= true;
3634 s
->kernel_irqchip_required
= true;
3635 s
->kernel_irqchip_split
= ON_OFF_AUTO_ON
;
3638 /* The value was checked in visit_type_OnOffSplit() above. If
3639 * we get here, then something is wrong in QEMU.
3645 bool kvm_kernel_irqchip_allowed(void)
3647 return kvm_state
->kernel_irqchip_allowed
;
3650 bool kvm_kernel_irqchip_required(void)
3652 return kvm_state
->kernel_irqchip_required
;
3655 bool kvm_kernel_irqchip_split(void)
3657 return kvm_state
->kernel_irqchip_split
== ON_OFF_AUTO_ON
;
3660 static void kvm_get_dirty_ring_size(Object
*obj
, Visitor
*v
,
3661 const char *name
, void *opaque
,
3664 KVMState
*s
= KVM_STATE(obj
);
3665 uint32_t value
= s
->kvm_dirty_ring_size
;
3667 visit_type_uint32(v
, name
, &value
, errp
);
3670 static void kvm_set_dirty_ring_size(Object
*obj
, Visitor
*v
,
3671 const char *name
, void *opaque
,
3674 KVMState
*s
= KVM_STATE(obj
);
3678 error_setg(errp
, "Cannot set properties after the accelerator has been initialized");
3682 if (!visit_type_uint32(v
, name
, &value
, errp
)) {
3685 if (value
& (value
- 1)) {
3686 error_setg(errp
, "dirty-ring-size must be a power of two.");
3690 s
->kvm_dirty_ring_size
= value
;
3693 static void kvm_accel_instance_init(Object
*obj
)
3695 KVMState
*s
= KVM_STATE(obj
);
3699 s
->kvm_shadow_mem
= -1;
3700 s
->kernel_irqchip_allowed
= true;
3701 s
->kernel_irqchip_split
= ON_OFF_AUTO_AUTO
;
3702 /* KVM dirty ring is by default off */
3703 s
->kvm_dirty_ring_size
= 0;
3704 s
->notify_vmexit
= NOTIFY_VMEXIT_OPTION_RUN
;
3705 s
->notify_window
= 0;
3709 * kvm_gdbstub_sstep_flags():
3711 * Returns: SSTEP_* flags that KVM supports for guest debug. The
3712 * support is probed during kvm_init()
3714 static int kvm_gdbstub_sstep_flags(void)
3716 return kvm_sstep_flags
;
3719 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
3721 AccelClass
*ac
= ACCEL_CLASS(oc
);
3723 ac
->init_machine
= kvm_init
;
3724 ac
->has_memory
= kvm_accel_has_memory
;
3725 ac
->allowed
= &kvm_allowed
;
3726 ac
->gdbstub_supported_sstep_flags
= kvm_gdbstub_sstep_flags
;
3728 object_class_property_add(oc
, "kernel-irqchip", "on|off|split",
3729 NULL
, kvm_set_kernel_irqchip
,
3731 object_class_property_set_description(oc
, "kernel-irqchip",
3732 "Configure KVM in-kernel irqchip");
3734 object_class_property_add(oc
, "kvm-shadow-mem", "int",
3735 kvm_get_kvm_shadow_mem
, kvm_set_kvm_shadow_mem
,
3737 object_class_property_set_description(oc
, "kvm-shadow-mem",
3738 "KVM shadow MMU size");
3740 object_class_property_add(oc
, "dirty-ring-size", "uint32",
3741 kvm_get_dirty_ring_size
, kvm_set_dirty_ring_size
,
3743 object_class_property_set_description(oc
, "dirty-ring-size",
3744 "Size of KVM dirty page ring buffer (default: 0, i.e. use bitmap)");
3746 kvm_arch_accel_class_init(oc
);
3749 static const TypeInfo kvm_accel_type
= {
3750 .name
= TYPE_KVM_ACCEL
,
3751 .parent
= TYPE_ACCEL
,
3752 .instance_init
= kvm_accel_instance_init
,
3753 .class_init
= kvm_accel_class_init
,
3754 .instance_size
= sizeof(KVMState
),
3757 static void kvm_type_init(void)
3759 type_register_static(&kvm_accel_type
);
3762 type_init(kvm_type_init
);
3764 typedef struct StatsArgs
{
3765 union StatsResultsType
{
3766 StatsResultList
**stats
;
3767 StatsSchemaList
**schema
;
3773 static StatsList
*add_kvmstat_entry(struct kvm_stats_desc
*pdesc
,
3774 uint64_t *stats_data
,
3775 StatsList
*stats_list
,
3780 uint64List
*val_list
= NULL
;
3782 /* Only add stats that we understand. */
3783 switch (pdesc
->flags
& KVM_STATS_TYPE_MASK
) {
3784 case KVM_STATS_TYPE_CUMULATIVE
:
3785 case KVM_STATS_TYPE_INSTANT
:
3786 case KVM_STATS_TYPE_PEAK
:
3787 case KVM_STATS_TYPE_LINEAR_HIST
:
3788 case KVM_STATS_TYPE_LOG_HIST
:
3794 switch (pdesc
->flags
& KVM_STATS_UNIT_MASK
) {
3795 case KVM_STATS_UNIT_NONE
:
3796 case KVM_STATS_UNIT_BYTES
:
3797 case KVM_STATS_UNIT_CYCLES
:
3798 case KVM_STATS_UNIT_SECONDS
:
3799 case KVM_STATS_UNIT_BOOLEAN
:
3805 switch (pdesc
->flags
& KVM_STATS_BASE_MASK
) {
3806 case KVM_STATS_BASE_POW10
:
3807 case KVM_STATS_BASE_POW2
:
3813 /* Alloc and populate data list */
3814 stats
= g_new0(Stats
, 1);
3815 stats
->name
= g_strdup(pdesc
->name
);
3816 stats
->value
= g_new0(StatsValue
, 1);;
3818 if ((pdesc
->flags
& KVM_STATS_UNIT_MASK
) == KVM_STATS_UNIT_BOOLEAN
) {
3819 stats
->value
->u
.boolean
= *stats_data
;
3820 stats
->value
->type
= QTYPE_QBOOL
;
3821 } else if (pdesc
->size
== 1) {
3822 stats
->value
->u
.scalar
= *stats_data
;
3823 stats
->value
->type
= QTYPE_QNUM
;
3826 for (i
= 0; i
< pdesc
->size
; i
++) {
3827 QAPI_LIST_PREPEND(val_list
, stats_data
[i
]);
3829 stats
->value
->u
.list
= val_list
;
3830 stats
->value
->type
= QTYPE_QLIST
;
3833 QAPI_LIST_PREPEND(stats_list
, stats
);
3837 static StatsSchemaValueList
*add_kvmschema_entry(struct kvm_stats_desc
*pdesc
,
3838 StatsSchemaValueList
*list
,
3841 StatsSchemaValueList
*schema_entry
= g_new0(StatsSchemaValueList
, 1);
3842 schema_entry
->value
= g_new0(StatsSchemaValue
, 1);
3844 switch (pdesc
->flags
& KVM_STATS_TYPE_MASK
) {
3845 case KVM_STATS_TYPE_CUMULATIVE
:
3846 schema_entry
->value
->type
= STATS_TYPE_CUMULATIVE
;
3848 case KVM_STATS_TYPE_INSTANT
:
3849 schema_entry
->value
->type
= STATS_TYPE_INSTANT
;
3851 case KVM_STATS_TYPE_PEAK
:
3852 schema_entry
->value
->type
= STATS_TYPE_PEAK
;
3854 case KVM_STATS_TYPE_LINEAR_HIST
:
3855 schema_entry
->value
->type
= STATS_TYPE_LINEAR_HISTOGRAM
;
3856 schema_entry
->value
->bucket_size
= pdesc
->bucket_size
;
3857 schema_entry
->value
->has_bucket_size
= true;
3859 case KVM_STATS_TYPE_LOG_HIST
:
3860 schema_entry
->value
->type
= STATS_TYPE_LOG2_HISTOGRAM
;
3866 switch (pdesc
->flags
& KVM_STATS_UNIT_MASK
) {
3867 case KVM_STATS_UNIT_NONE
:
3869 case KVM_STATS_UNIT_BOOLEAN
:
3870 schema_entry
->value
->has_unit
= true;
3871 schema_entry
->value
->unit
= STATS_UNIT_BOOLEAN
;
3873 case KVM_STATS_UNIT_BYTES
:
3874 schema_entry
->value
->has_unit
= true;
3875 schema_entry
->value
->unit
= STATS_UNIT_BYTES
;
3877 case KVM_STATS_UNIT_CYCLES
:
3878 schema_entry
->value
->has_unit
= true;
3879 schema_entry
->value
->unit
= STATS_UNIT_CYCLES
;
3881 case KVM_STATS_UNIT_SECONDS
:
3882 schema_entry
->value
->has_unit
= true;
3883 schema_entry
->value
->unit
= STATS_UNIT_SECONDS
;
3889 schema_entry
->value
->exponent
= pdesc
->exponent
;
3890 if (pdesc
->exponent
) {
3891 switch (pdesc
->flags
& KVM_STATS_BASE_MASK
) {
3892 case KVM_STATS_BASE_POW10
:
3893 schema_entry
->value
->has_base
= true;
3894 schema_entry
->value
->base
= 10;
3896 case KVM_STATS_BASE_POW2
:
3897 schema_entry
->value
->has_base
= true;
3898 schema_entry
->value
->base
= 2;
3905 schema_entry
->value
->name
= g_strdup(pdesc
->name
);
3906 schema_entry
->next
= list
;
3907 return schema_entry
;
3909 g_free(schema_entry
->value
);
3910 g_free(schema_entry
);
3914 /* Cached stats descriptors */
3915 typedef struct StatsDescriptors
{
3916 const char *ident
; /* cache key, currently the StatsTarget */
3917 struct kvm_stats_desc
*kvm_stats_desc
;
3918 struct kvm_stats_header kvm_stats_header
;
3919 QTAILQ_ENTRY(StatsDescriptors
) next
;
3922 static QTAILQ_HEAD(, StatsDescriptors
) stats_descriptors
=
3923 QTAILQ_HEAD_INITIALIZER(stats_descriptors
);
3926 * Return the descriptors for 'target', that either have already been read
3927 * or are retrieved from 'stats_fd'.
3929 static StatsDescriptors
*find_stats_descriptors(StatsTarget target
, int stats_fd
,
3932 StatsDescriptors
*descriptors
;
3934 struct kvm_stats_desc
*kvm_stats_desc
;
3935 struct kvm_stats_header
*kvm_stats_header
;
3939 ident
= StatsTarget_str(target
);
3940 QTAILQ_FOREACH(descriptors
, &stats_descriptors
, next
) {
3941 if (g_str_equal(descriptors
->ident
, ident
)) {
3946 descriptors
= g_new0(StatsDescriptors
, 1);
3948 /* Read stats header */
3949 kvm_stats_header
= &descriptors
->kvm_stats_header
;
3950 ret
= read(stats_fd
, kvm_stats_header
, sizeof(*kvm_stats_header
));
3951 if (ret
!= sizeof(*kvm_stats_header
)) {
3952 error_setg(errp
, "KVM stats: failed to read stats header: "
3953 "expected %zu actual %zu",
3954 sizeof(*kvm_stats_header
), ret
);
3955 g_free(descriptors
);
3958 size_desc
= sizeof(*kvm_stats_desc
) + kvm_stats_header
->name_size
;
3960 /* Read stats descriptors */
3961 kvm_stats_desc
= g_malloc0_n(kvm_stats_header
->num_desc
, size_desc
);
3962 ret
= pread(stats_fd
, kvm_stats_desc
,
3963 size_desc
* kvm_stats_header
->num_desc
,
3964 kvm_stats_header
->desc_offset
);
3966 if (ret
!= size_desc
* kvm_stats_header
->num_desc
) {
3967 error_setg(errp
, "KVM stats: failed to read stats descriptors: "
3968 "expected %zu actual %zu",
3969 size_desc
* kvm_stats_header
->num_desc
, ret
);
3970 g_free(descriptors
);
3971 g_free(kvm_stats_desc
);
3974 descriptors
->kvm_stats_desc
= kvm_stats_desc
;
3975 descriptors
->ident
= ident
;
3976 QTAILQ_INSERT_TAIL(&stats_descriptors
, descriptors
, next
);
3980 static void query_stats(StatsResultList
**result
, StatsTarget target
,
3981 strList
*names
, int stats_fd
, Error
**errp
)
3983 struct kvm_stats_desc
*kvm_stats_desc
;
3984 struct kvm_stats_header
*kvm_stats_header
;
3985 StatsDescriptors
*descriptors
;
3986 g_autofree
uint64_t *stats_data
= NULL
;
3987 struct kvm_stats_desc
*pdesc
;
3988 StatsList
*stats_list
= NULL
;
3989 size_t size_desc
, size_data
= 0;
3993 descriptors
= find_stats_descriptors(target
, stats_fd
, errp
);
3998 kvm_stats_header
= &descriptors
->kvm_stats_header
;
3999 kvm_stats_desc
= descriptors
->kvm_stats_desc
;
4000 size_desc
= sizeof(*kvm_stats_desc
) + kvm_stats_header
->name_size
;
4002 /* Tally the total data size; read schema data */
4003 for (i
= 0; i
< kvm_stats_header
->num_desc
; ++i
) {
4004 pdesc
= (void *)kvm_stats_desc
+ i
* size_desc
;
4005 size_data
+= pdesc
->size
* sizeof(*stats_data
);
4008 stats_data
= g_malloc0(size_data
);
4009 ret
= pread(stats_fd
, stats_data
, size_data
, kvm_stats_header
->data_offset
);
4011 if (ret
!= size_data
) {
4012 error_setg(errp
, "KVM stats: failed to read data: "
4013 "expected %zu actual %zu", size_data
, ret
);
4017 for (i
= 0; i
< kvm_stats_header
->num_desc
; ++i
) {
4019 pdesc
= (void *)kvm_stats_desc
+ i
* size_desc
;
4021 /* Add entry to the list */
4022 stats
= (void *)stats_data
+ pdesc
->offset
;
4023 if (!apply_str_list_filter(pdesc
->name
, names
)) {
4026 stats_list
= add_kvmstat_entry(pdesc
, stats
, stats_list
, errp
);
4034 case STATS_TARGET_VM
:
4035 add_stats_entry(result
, STATS_PROVIDER_KVM
, NULL
, stats_list
);
4037 case STATS_TARGET_VCPU
:
4038 add_stats_entry(result
, STATS_PROVIDER_KVM
,
4039 current_cpu
->parent_obj
.canonical_path
,
4043 g_assert_not_reached();
4047 static void query_stats_schema(StatsSchemaList
**result
, StatsTarget target
,
4048 int stats_fd
, Error
**errp
)
4050 struct kvm_stats_desc
*kvm_stats_desc
;
4051 struct kvm_stats_header
*kvm_stats_header
;
4052 StatsDescriptors
*descriptors
;
4053 struct kvm_stats_desc
*pdesc
;
4054 StatsSchemaValueList
*stats_list
= NULL
;
4058 descriptors
= find_stats_descriptors(target
, stats_fd
, errp
);
4063 kvm_stats_header
= &descriptors
->kvm_stats_header
;
4064 kvm_stats_desc
= descriptors
->kvm_stats_desc
;
4065 size_desc
= sizeof(*kvm_stats_desc
) + kvm_stats_header
->name_size
;
4067 /* Tally the total data size; read schema data */
4068 for (i
= 0; i
< kvm_stats_header
->num_desc
; ++i
) {
4069 pdesc
= (void *)kvm_stats_desc
+ i
* size_desc
;
4070 stats_list
= add_kvmschema_entry(pdesc
, stats_list
, errp
);
4073 add_stats_schema(result
, STATS_PROVIDER_KVM
, target
, stats_list
);
4076 static void query_stats_vcpu(CPUState
*cpu
, run_on_cpu_data data
)
4078 StatsArgs
*kvm_stats_args
= (StatsArgs
*) data
.host_ptr
;
4079 int stats_fd
= kvm_vcpu_ioctl(cpu
, KVM_GET_STATS_FD
, NULL
);
4080 Error
*local_err
= NULL
;
4082 if (stats_fd
== -1) {
4083 error_setg_errno(&local_err
, errno
, "KVM stats: ioctl failed");
4084 error_propagate(kvm_stats_args
->errp
, local_err
);
4087 query_stats(kvm_stats_args
->result
.stats
, STATS_TARGET_VCPU
,
4088 kvm_stats_args
->names
, stats_fd
, kvm_stats_args
->errp
);
4092 static void query_stats_schema_vcpu(CPUState
*cpu
, run_on_cpu_data data
)
4094 StatsArgs
*kvm_stats_args
= (StatsArgs
*) data
.host_ptr
;
4095 int stats_fd
= kvm_vcpu_ioctl(cpu
, KVM_GET_STATS_FD
, NULL
);
4096 Error
*local_err
= NULL
;
4098 if (stats_fd
== -1) {
4099 error_setg_errno(&local_err
, errno
, "KVM stats: ioctl failed");
4100 error_propagate(kvm_stats_args
->errp
, local_err
);
4103 query_stats_schema(kvm_stats_args
->result
.schema
, STATS_TARGET_VCPU
, stats_fd
,
4104 kvm_stats_args
->errp
);
4108 static void query_stats_cb(StatsResultList
**result
, StatsTarget target
,
4109 strList
*names
, strList
*targets
, Error
**errp
)
4111 KVMState
*s
= kvm_state
;
4116 case STATS_TARGET_VM
:
4118 stats_fd
= kvm_vm_ioctl(s
, KVM_GET_STATS_FD
, NULL
);
4119 if (stats_fd
== -1) {
4120 error_setg_errno(errp
, errno
, "KVM stats: ioctl failed");
4123 query_stats(result
, target
, names
, stats_fd
, errp
);
4127 case STATS_TARGET_VCPU
:
4129 StatsArgs stats_args
;
4130 stats_args
.result
.stats
= result
;
4131 stats_args
.names
= names
;
4132 stats_args
.errp
= errp
;
4134 if (!apply_str_list_filter(cpu
->parent_obj
.canonical_path
, targets
)) {
4137 run_on_cpu(cpu
, query_stats_vcpu
, RUN_ON_CPU_HOST_PTR(&stats_args
));
4146 void query_stats_schemas_cb(StatsSchemaList
**result
, Error
**errp
)
4148 StatsArgs stats_args
;
4149 KVMState
*s
= kvm_state
;
4152 stats_fd
= kvm_vm_ioctl(s
, KVM_GET_STATS_FD
, NULL
);
4153 if (stats_fd
== -1) {
4154 error_setg_errno(errp
, errno
, "KVM stats: ioctl failed");
4157 query_stats_schema(result
, STATS_TARGET_VM
, stats_fd
, errp
);
4161 stats_args
.result
.schema
= result
;
4162 stats_args
.errp
= errp
;
4163 run_on_cpu(first_cpu
, query_stats_schema_vcpu
, RUN_ON_CPU_HOST_PTR(&stats_args
));