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 "qemu/bswap.h"
35 #include "exec/memory.h"
36 #include "exec/ram_addr.h"
37 #include "qemu/event_notifier.h"
38 #include "qemu/main-loop.h"
41 #include "qapi/visitor.h"
42 #include "qapi/qapi-types-common.h"
43 #include "qapi/qapi-visit-common.h"
44 #include "sysemu/reset.h"
45 #include "qemu/guest-random.h"
46 #include "sysemu/hw_accel.h"
48 #include "sysemu/dirtylimit.h"
50 #include "hw/boards.h"
51 #include "monitor/stats.h"
53 /* This check must be after config-host.h is included */
55 #include <sys/eventfd.h>
58 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
59 * need to use the real host PAGE_SIZE, as that's what KVM will use.
64 #define PAGE_SIZE qemu_real_host_page_size()
66 #ifndef KVM_GUESTDBG_BLOCKIRQ
67 #define KVM_GUESTDBG_BLOCKIRQ 0
73 #define DPRINTF(fmt, ...) \
74 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
76 #define DPRINTF(fmt, ...) \
80 #define KVM_MSI_HASHTAB_SIZE 256
82 struct KVMParkedVcpu
{
83 unsigned long vcpu_id
;
85 QLIST_ENTRY(KVMParkedVcpu
) node
;
88 enum KVMDirtyRingReaperState
{
89 KVM_DIRTY_RING_REAPER_NONE
= 0,
90 /* The reaper is sleeping */
91 KVM_DIRTY_RING_REAPER_WAIT
,
92 /* The reaper is reaping for dirty pages */
93 KVM_DIRTY_RING_REAPER_REAPING
,
97 * KVM reaper instance, responsible for collecting the KVM dirty bits
100 struct KVMDirtyRingReaper
{
101 /* The reaper thread */
102 QemuThread reaper_thr
;
103 volatile uint64_t reaper_iteration
; /* iteration number of reaper thr */
104 volatile enum KVMDirtyRingReaperState reaper_state
; /* reap thr state */
109 AccelState parent_obj
;
116 struct kvm_coalesced_mmio_ring
*coalesced_mmio_ring
;
117 bool coalesced_flush_in_progress
;
119 int robust_singlestep
;
121 #ifdef KVM_CAP_SET_GUEST_DEBUG
122 QTAILQ_HEAD(, kvm_sw_breakpoint
) kvm_sw_breakpoints
;
124 int max_nested_state_len
;
128 bool kernel_irqchip_allowed
;
129 bool kernel_irqchip_required
;
130 OnOffAuto kernel_irqchip_split
;
132 uint64_t manual_dirty_log_protect
;
133 /* The man page (and posix) say ioctl numbers are signed int, but
134 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
135 * unsigned, and treating them as signed here can break things */
136 unsigned irq_set_ioctl
;
137 unsigned int sigmask_len
;
139 #ifdef KVM_CAP_IRQ_ROUTING
140 struct kvm_irq_routing
*irq_routes
;
141 int nr_allocated_irq_routes
;
142 unsigned long *used_gsi_bitmap
;
143 unsigned int gsi_count
;
144 QTAILQ_HEAD(, KVMMSIRoute
) msi_hashtab
[KVM_MSI_HASHTAB_SIZE
];
146 KVMMemoryListener memory_listener
;
147 QLIST_HEAD(, KVMParkedVcpu
) kvm_parked_vcpus
;
149 /* For "info mtree -f" to tell if an MR is registered in KVM */
152 KVMMemoryListener
*ml
;
155 uint64_t kvm_dirty_ring_bytes
; /* Size of the per-vcpu dirty ring */
156 uint32_t kvm_dirty_ring_size
; /* Number of dirty GFNs per ring */
157 struct KVMDirtyRingReaper reaper
;
161 bool kvm_kernel_irqchip
;
162 bool kvm_split_irqchip
;
163 bool kvm_async_interrupts_allowed
;
164 bool kvm_halt_in_kernel_allowed
;
165 bool kvm_eventfds_allowed
;
166 bool kvm_irqfds_allowed
;
167 bool kvm_resamplefds_allowed
;
168 bool kvm_msi_via_irqfd_allowed
;
169 bool kvm_gsi_routing_allowed
;
170 bool kvm_gsi_direct_mapping
;
172 bool kvm_readonly_mem_allowed
;
173 bool kvm_vm_attributes_allowed
;
174 bool kvm_direct_msi_allowed
;
175 bool kvm_ioeventfd_any_length_allowed
;
176 bool kvm_msi_use_devid
;
177 bool kvm_has_guest_debug
;
179 static bool kvm_immediate_exit
;
180 static hwaddr kvm_max_slot_size
= ~0;
182 static const KVMCapabilityInfo kvm_required_capabilites
[] = {
183 KVM_CAP_INFO(USER_MEMORY
),
184 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS
),
185 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS
),
189 static NotifierList kvm_irqchip_change_notifiers
=
190 NOTIFIER_LIST_INITIALIZER(kvm_irqchip_change_notifiers
);
192 struct KVMResampleFd
{
194 EventNotifier
*resample_event
;
195 QLIST_ENTRY(KVMResampleFd
) node
;
197 typedef struct KVMResampleFd KVMResampleFd
;
200 * Only used with split irqchip where we need to do the resample fd
201 * kick for the kernel from userspace.
203 static QLIST_HEAD(, KVMResampleFd
) kvm_resample_fd_list
=
204 QLIST_HEAD_INITIALIZER(kvm_resample_fd_list
);
206 static QemuMutex kml_slots_lock
;
208 #define kvm_slots_lock() qemu_mutex_lock(&kml_slots_lock)
209 #define kvm_slots_unlock() qemu_mutex_unlock(&kml_slots_lock)
211 static void kvm_slot_init_dirty_bitmap(KVMSlot
*mem
);
213 static inline void kvm_resample_fd_remove(int gsi
)
217 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
218 if (rfd
->gsi
== gsi
) {
219 QLIST_REMOVE(rfd
, node
);
226 static inline void kvm_resample_fd_insert(int gsi
, EventNotifier
*event
)
228 KVMResampleFd
*rfd
= g_new0(KVMResampleFd
, 1);
231 rfd
->resample_event
= event
;
233 QLIST_INSERT_HEAD(&kvm_resample_fd_list
, rfd
, node
);
236 void kvm_resample_fd_notify(int gsi
)
240 QLIST_FOREACH(rfd
, &kvm_resample_fd_list
, node
) {
241 if (rfd
->gsi
== gsi
) {
242 event_notifier_set(rfd
->resample_event
);
243 trace_kvm_resample_fd_notify(gsi
);
249 int kvm_get_max_memslots(void)
251 KVMState
*s
= KVM_STATE(current_accel());
256 /* Called with KVMMemoryListener.slots_lock held */
257 static KVMSlot
*kvm_get_free_slot(KVMMemoryListener
*kml
)
259 KVMState
*s
= kvm_state
;
262 for (i
= 0; i
< s
->nr_slots
; i
++) {
263 if (kml
->slots
[i
].memory_size
== 0) {
264 return &kml
->slots
[i
];
271 bool kvm_has_free_slot(MachineState
*ms
)
273 KVMState
*s
= KVM_STATE(ms
->accelerator
);
275 KVMMemoryListener
*kml
= &s
->memory_listener
;
278 result
= !!kvm_get_free_slot(kml
);
284 /* Called with KVMMemoryListener.slots_lock held */
285 static KVMSlot
*kvm_alloc_slot(KVMMemoryListener
*kml
)
287 KVMSlot
*slot
= kvm_get_free_slot(kml
);
293 fprintf(stderr
, "%s: no free slot available\n", __func__
);
297 static KVMSlot
*kvm_lookup_matching_slot(KVMMemoryListener
*kml
,
301 KVMState
*s
= kvm_state
;
304 for (i
= 0; i
< s
->nr_slots
; i
++) {
305 KVMSlot
*mem
= &kml
->slots
[i
];
307 if (start_addr
== mem
->start_addr
&& size
== mem
->memory_size
) {
316 * Calculate and align the start address and the size of the section.
317 * Return the size. If the size is 0, the aligned section is empty.
319 static hwaddr
kvm_align_section(MemoryRegionSection
*section
,
322 hwaddr size
= int128_get64(section
->size
);
323 hwaddr delta
, aligned
;
325 /* kvm works in page size chunks, but the function may be called
326 with sub-page size and unaligned start address. Pad the start
327 address to next and truncate size to previous page boundary. */
328 aligned
= ROUND_UP(section
->offset_within_address_space
,
329 qemu_real_host_page_size());
330 delta
= aligned
- section
->offset_within_address_space
;
336 return (size
- delta
) & qemu_real_host_page_mask();
339 int kvm_physical_memory_addr_from_host(KVMState
*s
, void *ram
,
342 KVMMemoryListener
*kml
= &s
->memory_listener
;
346 for (i
= 0; i
< s
->nr_slots
; i
++) {
347 KVMSlot
*mem
= &kml
->slots
[i
];
349 if (ram
>= mem
->ram
&& ram
< mem
->ram
+ mem
->memory_size
) {
350 *phys_addr
= mem
->start_addr
+ (ram
- mem
->ram
);
360 static int kvm_set_user_memory_region(KVMMemoryListener
*kml
, KVMSlot
*slot
, bool new)
362 KVMState
*s
= kvm_state
;
363 struct kvm_userspace_memory_region mem
;
366 mem
.slot
= slot
->slot
| (kml
->as_id
<< 16);
367 mem
.guest_phys_addr
= slot
->start_addr
;
368 mem
.userspace_addr
= (unsigned long)slot
->ram
;
369 mem
.flags
= slot
->flags
;
371 if (slot
->memory_size
&& !new && (mem
.flags
^ slot
->old_flags
) & KVM_MEM_READONLY
) {
372 /* Set the slot size to 0 before setting the slot to the desired
373 * value. This is needed based on KVM commit 75d61fbc. */
375 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
380 mem
.memory_size
= slot
->memory_size
;
381 ret
= kvm_vm_ioctl(s
, KVM_SET_USER_MEMORY_REGION
, &mem
);
382 slot
->old_flags
= mem
.flags
;
384 trace_kvm_set_user_memory(mem
.slot
, mem
.flags
, mem
.guest_phys_addr
,
385 mem
.memory_size
, mem
.userspace_addr
, ret
);
387 error_report("%s: KVM_SET_USER_MEMORY_REGION failed, slot=%d,"
388 " start=0x%" PRIx64
", size=0x%" PRIx64
": %s",
389 __func__
, mem
.slot
, slot
->start_addr
,
390 (uint64_t)mem
.memory_size
, strerror(errno
));
395 static int do_kvm_destroy_vcpu(CPUState
*cpu
)
397 KVMState
*s
= kvm_state
;
399 struct KVMParkedVcpu
*vcpu
= NULL
;
402 DPRINTF("kvm_destroy_vcpu\n");
404 ret
= kvm_arch_destroy_vcpu(cpu
);
409 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
412 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
416 ret
= munmap(cpu
->kvm_run
, mmap_size
);
421 if (cpu
->kvm_dirty_gfns
) {
422 ret
= munmap(cpu
->kvm_dirty_gfns
, s
->kvm_dirty_ring_bytes
);
428 vcpu
= g_malloc0(sizeof(*vcpu
));
429 vcpu
->vcpu_id
= kvm_arch_vcpu_id(cpu
);
430 vcpu
->kvm_fd
= cpu
->kvm_fd
;
431 QLIST_INSERT_HEAD(&kvm_state
->kvm_parked_vcpus
, vcpu
, node
);
436 void kvm_destroy_vcpu(CPUState
*cpu
)
438 if (do_kvm_destroy_vcpu(cpu
) < 0) {
439 error_report("kvm_destroy_vcpu failed");
444 static int kvm_get_vcpu(KVMState
*s
, unsigned long vcpu_id
)
446 struct KVMParkedVcpu
*cpu
;
448 QLIST_FOREACH(cpu
, &s
->kvm_parked_vcpus
, node
) {
449 if (cpu
->vcpu_id
== vcpu_id
) {
452 QLIST_REMOVE(cpu
, node
);
453 kvm_fd
= cpu
->kvm_fd
;
459 return kvm_vm_ioctl(s
, KVM_CREATE_VCPU
, (void *)vcpu_id
);
462 int kvm_init_vcpu(CPUState
*cpu
, Error
**errp
)
464 KVMState
*s
= kvm_state
;
468 trace_kvm_init_vcpu(cpu
->cpu_index
, kvm_arch_vcpu_id(cpu
));
470 ret
= kvm_get_vcpu(s
, kvm_arch_vcpu_id(cpu
));
472 error_setg_errno(errp
, -ret
, "kvm_init_vcpu: kvm_get_vcpu failed (%lu)",
473 kvm_arch_vcpu_id(cpu
));
479 cpu
->vcpu_dirty
= true;
480 cpu
->dirty_pages
= 0;
481 cpu
->throttle_us_per_full
= 0;
483 mmap_size
= kvm_ioctl(s
, KVM_GET_VCPU_MMAP_SIZE
, 0);
486 error_setg_errno(errp
, -mmap_size
,
487 "kvm_init_vcpu: KVM_GET_VCPU_MMAP_SIZE failed");
491 cpu
->kvm_run
= mmap(NULL
, mmap_size
, PROT_READ
| PROT_WRITE
, MAP_SHARED
,
493 if (cpu
->kvm_run
== MAP_FAILED
) {
495 error_setg_errno(errp
, ret
,
496 "kvm_init_vcpu: mmap'ing vcpu state failed (%lu)",
497 kvm_arch_vcpu_id(cpu
));
501 if (s
->coalesced_mmio
&& !s
->coalesced_mmio_ring
) {
502 s
->coalesced_mmio_ring
=
503 (void *)cpu
->kvm_run
+ s
->coalesced_mmio
* PAGE_SIZE
;
506 if (s
->kvm_dirty_ring_size
) {
507 /* Use MAP_SHARED to share pages with the kernel */
508 cpu
->kvm_dirty_gfns
= mmap(NULL
, s
->kvm_dirty_ring_bytes
,
509 PROT_READ
| PROT_WRITE
, MAP_SHARED
,
511 PAGE_SIZE
* KVM_DIRTY_LOG_PAGE_OFFSET
);
512 if (cpu
->kvm_dirty_gfns
== MAP_FAILED
) {
514 DPRINTF("mmap'ing vcpu dirty gfns failed: %d\n", ret
);
519 ret
= kvm_arch_init_vcpu(cpu
);
521 error_setg_errno(errp
, -ret
,
522 "kvm_init_vcpu: kvm_arch_init_vcpu failed (%lu)",
523 kvm_arch_vcpu_id(cpu
));
530 * dirty pages logging control
533 static int kvm_mem_flags(MemoryRegion
*mr
)
535 bool readonly
= mr
->readonly
|| memory_region_is_romd(mr
);
538 if (memory_region_get_dirty_log_mask(mr
) != 0) {
539 flags
|= KVM_MEM_LOG_DIRTY_PAGES
;
541 if (readonly
&& kvm_readonly_mem_allowed
) {
542 flags
|= KVM_MEM_READONLY
;
547 /* Called with KVMMemoryListener.slots_lock held */
548 static int kvm_slot_update_flags(KVMMemoryListener
*kml
, KVMSlot
*mem
,
551 mem
->flags
= kvm_mem_flags(mr
);
553 /* If nothing changed effectively, no need to issue ioctl */
554 if (mem
->flags
== mem
->old_flags
) {
558 kvm_slot_init_dirty_bitmap(mem
);
559 return kvm_set_user_memory_region(kml
, mem
, false);
562 static int kvm_section_update_flags(KVMMemoryListener
*kml
,
563 MemoryRegionSection
*section
)
565 hwaddr start_addr
, size
, slot_size
;
569 size
= kvm_align_section(section
, &start_addr
);
576 while (size
&& !ret
) {
577 slot_size
= MIN(kvm_max_slot_size
, size
);
578 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
580 /* We don't have a slot if we want to trap every access. */
584 ret
= kvm_slot_update_flags(kml
, mem
, section
->mr
);
585 start_addr
+= slot_size
;
594 static void kvm_log_start(MemoryListener
*listener
,
595 MemoryRegionSection
*section
,
598 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
605 r
= kvm_section_update_flags(kml
, section
);
611 static void kvm_log_stop(MemoryListener
*listener
,
612 MemoryRegionSection
*section
,
615 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
622 r
= kvm_section_update_flags(kml
, section
);
628 /* get kvm's dirty pages bitmap and update qemu's */
629 static void kvm_slot_sync_dirty_pages(KVMSlot
*slot
)
631 ram_addr_t start
= slot
->ram_start_offset
;
632 ram_addr_t pages
= slot
->memory_size
/ qemu_real_host_page_size();
634 cpu_physical_memory_set_dirty_lebitmap(slot
->dirty_bmap
, start
, pages
);
637 static void kvm_slot_reset_dirty_pages(KVMSlot
*slot
)
639 memset(slot
->dirty_bmap
, 0, slot
->dirty_bmap_size
);
642 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
644 /* Allocate the dirty bitmap for a slot */
645 static void kvm_slot_init_dirty_bitmap(KVMSlot
*mem
)
647 if (!(mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) || mem
->dirty_bmap
) {
652 * XXX bad kernel interface alert
653 * For dirty bitmap, kernel allocates array of size aligned to
654 * bits-per-long. But for case when the kernel is 64bits and
655 * the userspace is 32bits, userspace can't align to the same
656 * bits-per-long, since sizeof(long) is different between kernel
657 * and user space. This way, userspace will provide buffer which
658 * may be 4 bytes less than the kernel will use, resulting in
659 * userspace memory corruption (which is not detectable by valgrind
660 * too, in most cases).
661 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
662 * a hope that sizeof(long) won't become >8 any time soon.
664 * Note: the granule of kvm dirty log is qemu_real_host_page_size.
665 * And mem->memory_size is aligned to it (otherwise this mem can't
666 * be registered to KVM).
668 hwaddr bitmap_size
= ALIGN(mem
->memory_size
/ qemu_real_host_page_size(),
669 /*HOST_LONG_BITS*/ 64) / 8;
670 mem
->dirty_bmap
= g_malloc0(bitmap_size
);
671 mem
->dirty_bmap_size
= bitmap_size
;
675 * Sync dirty bitmap from kernel to KVMSlot.dirty_bmap, return true if
676 * succeeded, false otherwise
678 static bool kvm_slot_get_dirty_log(KVMState
*s
, KVMSlot
*slot
)
680 struct kvm_dirty_log d
= {};
683 d
.dirty_bitmap
= slot
->dirty_bmap
;
684 d
.slot
= slot
->slot
| (slot
->as_id
<< 16);
685 ret
= kvm_vm_ioctl(s
, KVM_GET_DIRTY_LOG
, &d
);
687 if (ret
== -ENOENT
) {
688 /* kernel does not have dirty bitmap in this slot */
692 error_report_once("%s: KVM_GET_DIRTY_LOG failed with %d",
698 /* Should be with all slots_lock held for the address spaces. */
699 static void kvm_dirty_ring_mark_page(KVMState
*s
, uint32_t as_id
,
700 uint32_t slot_id
, uint64_t offset
)
702 KVMMemoryListener
*kml
;
705 if (as_id
>= s
->nr_as
) {
709 kml
= s
->as
[as_id
].ml
;
710 mem
= &kml
->slots
[slot_id
];
712 if (!mem
->memory_size
|| offset
>=
713 (mem
->memory_size
/ qemu_real_host_page_size())) {
717 set_bit(offset
, mem
->dirty_bmap
);
720 static bool dirty_gfn_is_dirtied(struct kvm_dirty_gfn
*gfn
)
722 return gfn
->flags
== KVM_DIRTY_GFN_F_DIRTY
;
725 static void dirty_gfn_set_collected(struct kvm_dirty_gfn
*gfn
)
727 gfn
->flags
= KVM_DIRTY_GFN_F_RESET
;
731 * Should be with all slots_lock held for the address spaces. It returns the
732 * dirty page we've collected on this dirty ring.
734 static uint32_t kvm_dirty_ring_reap_one(KVMState
*s
, CPUState
*cpu
)
736 struct kvm_dirty_gfn
*dirty_gfns
= cpu
->kvm_dirty_gfns
, *cur
;
737 uint32_t ring_size
= s
->kvm_dirty_ring_size
;
738 uint32_t count
= 0, fetch
= cpu
->kvm_fetch_index
;
740 assert(dirty_gfns
&& ring_size
);
741 trace_kvm_dirty_ring_reap_vcpu(cpu
->cpu_index
);
744 cur
= &dirty_gfns
[fetch
% ring_size
];
745 if (!dirty_gfn_is_dirtied(cur
)) {
748 kvm_dirty_ring_mark_page(s
, cur
->slot
>> 16, cur
->slot
& 0xffff,
750 dirty_gfn_set_collected(cur
);
751 trace_kvm_dirty_ring_page(cpu
->cpu_index
, fetch
, cur
->offset
);
755 cpu
->kvm_fetch_index
= fetch
;
756 cpu
->dirty_pages
+= count
;
761 /* Must be with slots_lock held */
762 static uint64_t kvm_dirty_ring_reap_locked(KVMState
*s
, CPUState
* cpu
)
771 total
= kvm_dirty_ring_reap_one(s
, cpu
);
774 total
+= kvm_dirty_ring_reap_one(s
, cpu
);
779 ret
= kvm_vm_ioctl(s
, KVM_RESET_DIRTY_RINGS
);
780 assert(ret
== total
);
783 stamp
= get_clock() - stamp
;
786 trace_kvm_dirty_ring_reap(total
, stamp
/ 1000);
793 * Currently for simplicity, we must hold BQL before calling this. We can
794 * consider to drop the BQL if we're clear with all the race conditions.
796 static uint64_t kvm_dirty_ring_reap(KVMState
*s
, CPUState
*cpu
)
801 * We need to lock all kvm slots for all address spaces here,
804 * (1) We need to mark dirty for dirty bitmaps in multiple slots
805 * and for tons of pages, so it's better to take the lock here
806 * once rather than once per page. And more importantly,
808 * (2) We must _NOT_ publish dirty bits to the other threads
809 * (e.g., the migration thread) via the kvm memory slot dirty
810 * bitmaps before correctly re-protect those dirtied pages.
811 * Otherwise we can have potential risk of data corruption if
812 * the page data is read in the other thread before we do
816 total
= kvm_dirty_ring_reap_locked(s
, cpu
);
822 static void do_kvm_cpu_synchronize_kick(CPUState
*cpu
, run_on_cpu_data arg
)
824 /* No need to do anything */
828 * Kick all vcpus out in a synchronized way. When returned, we
829 * guarantee that every vcpu has been kicked and at least returned to
832 static void kvm_cpu_synchronize_kick_all(void)
837 run_on_cpu(cpu
, do_kvm_cpu_synchronize_kick
, RUN_ON_CPU_NULL
);
842 * Flush all the existing dirty pages to the KVM slot buffers. When
843 * this call returns, we guarantee that all the touched dirty pages
844 * before calling this function have been put into the per-kvmslot
847 * This function must be called with BQL held.
849 static void kvm_dirty_ring_flush(void)
851 trace_kvm_dirty_ring_flush(0);
853 * The function needs to be serialized. Since this function
854 * should always be with BQL held, serialization is guaranteed.
855 * However, let's be sure of it.
857 assert(qemu_mutex_iothread_locked());
859 * First make sure to flush the hardware buffers by kicking all
860 * vcpus out in a synchronous way.
862 kvm_cpu_synchronize_kick_all();
863 kvm_dirty_ring_reap(kvm_state
, NULL
);
864 trace_kvm_dirty_ring_flush(1);
868 * kvm_physical_sync_dirty_bitmap - Sync dirty bitmap from kernel space
870 * This function will first try to fetch dirty bitmap from the kernel,
871 * and then updates qemu's dirty bitmap.
873 * NOTE: caller must be with kml->slots_lock held.
875 * @kml: the KVM memory listener object
876 * @section: the memory section to sync the dirty bitmap with
878 static void kvm_physical_sync_dirty_bitmap(KVMMemoryListener
*kml
,
879 MemoryRegionSection
*section
)
881 KVMState
*s
= kvm_state
;
883 hwaddr start_addr
, size
;
886 size
= kvm_align_section(section
, &start_addr
);
888 slot_size
= MIN(kvm_max_slot_size
, size
);
889 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
891 /* We don't have a slot if we want to trap every access. */
894 if (kvm_slot_get_dirty_log(s
, mem
)) {
895 kvm_slot_sync_dirty_pages(mem
);
897 start_addr
+= slot_size
;
902 /* Alignment requirement for KVM_CLEAR_DIRTY_LOG - 64 pages */
903 #define KVM_CLEAR_LOG_SHIFT 6
904 #define KVM_CLEAR_LOG_ALIGN (qemu_real_host_page_size() << KVM_CLEAR_LOG_SHIFT)
905 #define KVM_CLEAR_LOG_MASK (-KVM_CLEAR_LOG_ALIGN)
907 static int kvm_log_clear_one_slot(KVMSlot
*mem
, int as_id
, uint64_t start
,
910 KVMState
*s
= kvm_state
;
911 uint64_t end
, bmap_start
, start_delta
, bmap_npages
;
912 struct kvm_clear_dirty_log d
;
913 unsigned long *bmap_clear
= NULL
, psize
= qemu_real_host_page_size();
917 * We need to extend either the start or the size or both to
918 * satisfy the KVM interface requirement. Firstly, do the start
919 * page alignment on 64 host pages
921 bmap_start
= start
& KVM_CLEAR_LOG_MASK
;
922 start_delta
= start
- bmap_start
;
926 * The kernel interface has restriction on the size too, that either:
928 * (1) the size is 64 host pages aligned (just like the start), or
929 * (2) the size fills up until the end of the KVM memslot.
931 bmap_npages
= DIV_ROUND_UP(size
+ start_delta
, KVM_CLEAR_LOG_ALIGN
)
932 << KVM_CLEAR_LOG_SHIFT
;
933 end
= mem
->memory_size
/ psize
;
934 if (bmap_npages
> end
- bmap_start
) {
935 bmap_npages
= end
- bmap_start
;
937 start_delta
/= psize
;
940 * Prepare the bitmap to clear dirty bits. Here we must guarantee
941 * that we won't clear any unknown dirty bits otherwise we might
942 * accidentally clear some set bits which are not yet synced from
943 * the kernel into QEMU's bitmap, then we'll lose track of the
944 * guest modifications upon those pages (which can directly lead
945 * to guest data loss or panic after migration).
947 * Layout of the KVMSlot.dirty_bmap:
949 * |<-------- bmap_npages -----------..>|
952 * |----------------|-------------|------------------|------------|
955 * start bmap_start (start) end
956 * of memslot of memslot
958 * [1] bmap_npages can be aligned to either 64 pages or the end of slot
961 assert(bmap_start
% BITS_PER_LONG
== 0);
962 /* We should never do log_clear before log_sync */
963 assert(mem
->dirty_bmap
);
964 if (start_delta
|| bmap_npages
- size
/ psize
) {
965 /* Slow path - we need to manipulate a temp bitmap */
966 bmap_clear
= bitmap_new(bmap_npages
);
967 bitmap_copy_with_src_offset(bmap_clear
, mem
->dirty_bmap
,
968 bmap_start
, start_delta
+ size
/ psize
);
970 * We need to fill the holes at start because that was not
971 * specified by the caller and we extended the bitmap only for
974 bitmap_clear(bmap_clear
, 0, start_delta
);
975 d
.dirty_bitmap
= bmap_clear
;
978 * Fast path - both start and size align well with BITS_PER_LONG
979 * (or the end of memory slot)
981 d
.dirty_bitmap
= mem
->dirty_bmap
+ BIT_WORD(bmap_start
);
984 d
.first_page
= bmap_start
;
985 /* It should never overflow. If it happens, say something */
986 assert(bmap_npages
<= UINT32_MAX
);
987 d
.num_pages
= bmap_npages
;
988 d
.slot
= mem
->slot
| (as_id
<< 16);
990 ret
= kvm_vm_ioctl(s
, KVM_CLEAR_DIRTY_LOG
, &d
);
991 if (ret
< 0 && ret
!= -ENOENT
) {
992 error_report("%s: KVM_CLEAR_DIRTY_LOG failed, slot=%d, "
993 "start=0x%"PRIx64
", size=0x%"PRIx32
", errno=%d",
994 __func__
, d
.slot
, (uint64_t)d
.first_page
,
995 (uint32_t)d
.num_pages
, ret
);
998 trace_kvm_clear_dirty_log(d
.slot
, d
.first_page
, d
.num_pages
);
1002 * After we have updated the remote dirty bitmap, we update the
1003 * cached bitmap as well for the memslot, then if another user
1004 * clears the same region we know we shouldn't clear it again on
1005 * the remote otherwise it's data loss as well.
1007 bitmap_clear(mem
->dirty_bmap
, bmap_start
+ start_delta
,
1009 /* This handles the NULL case well */
1016 * kvm_physical_log_clear - Clear the kernel's dirty bitmap for range
1018 * NOTE: this will be a no-op if we haven't enabled manual dirty log
1019 * protection in the host kernel because in that case this operation
1020 * will be done within log_sync().
1022 * @kml: the kvm memory listener
1023 * @section: the memory range to clear dirty bitmap
1025 static int kvm_physical_log_clear(KVMMemoryListener
*kml
,
1026 MemoryRegionSection
*section
)
1028 KVMState
*s
= kvm_state
;
1029 uint64_t start
, size
, offset
, count
;
1033 if (!s
->manual_dirty_log_protect
) {
1034 /* No need to do explicit clear */
1038 start
= section
->offset_within_address_space
;
1039 size
= int128_get64(section
->size
);
1042 /* Nothing more we can do... */
1048 for (i
= 0; i
< s
->nr_slots
; i
++) {
1049 mem
= &kml
->slots
[i
];
1050 /* Discard slots that are empty or do not overlap the section */
1051 if (!mem
->memory_size
||
1052 mem
->start_addr
> start
+ size
- 1 ||
1053 start
> mem
->start_addr
+ mem
->memory_size
- 1) {
1057 if (start
>= mem
->start_addr
) {
1058 /* The slot starts before section or is aligned to it. */
1059 offset
= start
- mem
->start_addr
;
1060 count
= MIN(mem
->memory_size
- offset
, size
);
1062 /* The slot starts after section. */
1064 count
= MIN(mem
->memory_size
, size
- (mem
->start_addr
- start
));
1066 ret
= kvm_log_clear_one_slot(mem
, kml
->as_id
, offset
, count
);
1077 static void kvm_coalesce_mmio_region(MemoryListener
*listener
,
1078 MemoryRegionSection
*secion
,
1079 hwaddr start
, hwaddr size
)
1081 KVMState
*s
= kvm_state
;
1083 if (s
->coalesced_mmio
) {
1084 struct kvm_coalesced_mmio_zone zone
;
1090 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
1094 static void kvm_uncoalesce_mmio_region(MemoryListener
*listener
,
1095 MemoryRegionSection
*secion
,
1096 hwaddr start
, hwaddr size
)
1098 KVMState
*s
= kvm_state
;
1100 if (s
->coalesced_mmio
) {
1101 struct kvm_coalesced_mmio_zone zone
;
1107 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
1111 static void kvm_coalesce_pio_add(MemoryListener
*listener
,
1112 MemoryRegionSection
*section
,
1113 hwaddr start
, hwaddr size
)
1115 KVMState
*s
= kvm_state
;
1117 if (s
->coalesced_pio
) {
1118 struct kvm_coalesced_mmio_zone zone
;
1124 (void)kvm_vm_ioctl(s
, KVM_REGISTER_COALESCED_MMIO
, &zone
);
1128 static void kvm_coalesce_pio_del(MemoryListener
*listener
,
1129 MemoryRegionSection
*section
,
1130 hwaddr start
, hwaddr size
)
1132 KVMState
*s
= kvm_state
;
1134 if (s
->coalesced_pio
) {
1135 struct kvm_coalesced_mmio_zone zone
;
1141 (void)kvm_vm_ioctl(s
, KVM_UNREGISTER_COALESCED_MMIO
, &zone
);
1145 static MemoryListener kvm_coalesced_pio_listener
= {
1146 .name
= "kvm-coalesced-pio",
1147 .coalesced_io_add
= kvm_coalesce_pio_add
,
1148 .coalesced_io_del
= kvm_coalesce_pio_del
,
1151 int kvm_check_extension(KVMState
*s
, unsigned int extension
)
1155 ret
= kvm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
1163 int kvm_vm_check_extension(KVMState
*s
, unsigned int extension
)
1167 ret
= kvm_vm_ioctl(s
, KVM_CHECK_EXTENSION
, extension
);
1169 /* VM wide version not implemented, use global one instead */
1170 ret
= kvm_check_extension(s
, extension
);
1176 typedef struct HWPoisonPage
{
1177 ram_addr_t ram_addr
;
1178 QLIST_ENTRY(HWPoisonPage
) list
;
1181 static QLIST_HEAD(, HWPoisonPage
) hwpoison_page_list
=
1182 QLIST_HEAD_INITIALIZER(hwpoison_page_list
);
1184 static void kvm_unpoison_all(void *param
)
1186 HWPoisonPage
*page
, *next_page
;
1188 QLIST_FOREACH_SAFE(page
, &hwpoison_page_list
, list
, next_page
) {
1189 QLIST_REMOVE(page
, list
);
1190 qemu_ram_remap(page
->ram_addr
, TARGET_PAGE_SIZE
);
1195 void kvm_hwpoison_page_add(ram_addr_t ram_addr
)
1199 QLIST_FOREACH(page
, &hwpoison_page_list
, list
) {
1200 if (page
->ram_addr
== ram_addr
) {
1204 page
= g_new(HWPoisonPage
, 1);
1205 page
->ram_addr
= ram_addr
;
1206 QLIST_INSERT_HEAD(&hwpoison_page_list
, page
, list
);
1209 static uint32_t adjust_ioeventfd_endianness(uint32_t val
, uint32_t size
)
1211 #if HOST_BIG_ENDIAN != TARGET_BIG_ENDIAN
1212 /* The kernel expects ioeventfd values in HOST_BIG_ENDIAN
1213 * endianness, but the memory core hands them in target endianness.
1214 * For example, PPC is always treated as big-endian even if running
1215 * on KVM and on PPC64LE. Correct here.
1229 static int kvm_set_ioeventfd_mmio(int fd
, hwaddr addr
, uint32_t val
,
1230 bool assign
, uint32_t size
, bool datamatch
)
1233 struct kvm_ioeventfd iofd
= {
1234 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1241 trace_kvm_set_ioeventfd_mmio(fd
, (uint64_t)addr
, val
, assign
, size
,
1243 if (!kvm_enabled()) {
1248 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1251 iofd
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1254 ret
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &iofd
);
1263 static int kvm_set_ioeventfd_pio(int fd
, uint16_t addr
, uint16_t val
,
1264 bool assign
, uint32_t size
, bool datamatch
)
1266 struct kvm_ioeventfd kick
= {
1267 .datamatch
= datamatch
? adjust_ioeventfd_endianness(val
, size
) : 0,
1269 .flags
= KVM_IOEVENTFD_FLAG_PIO
,
1274 trace_kvm_set_ioeventfd_pio(fd
, addr
, val
, assign
, size
, datamatch
);
1275 if (!kvm_enabled()) {
1279 kick
.flags
|= KVM_IOEVENTFD_FLAG_DATAMATCH
;
1282 kick
.flags
|= KVM_IOEVENTFD_FLAG_DEASSIGN
;
1284 r
= kvm_vm_ioctl(kvm_state
, KVM_IOEVENTFD
, &kick
);
1292 static int kvm_check_many_ioeventfds(void)
1294 /* Userspace can use ioeventfd for io notification. This requires a host
1295 * that supports eventfd(2) and an I/O thread; since eventfd does not
1296 * support SIGIO it cannot interrupt the vcpu.
1298 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
1299 * can avoid creating too many ioeventfds.
1301 #if defined(CONFIG_EVENTFD)
1304 for (i
= 0; i
< ARRAY_SIZE(ioeventfds
); i
++) {
1305 ioeventfds
[i
] = eventfd(0, EFD_CLOEXEC
);
1306 if (ioeventfds
[i
] < 0) {
1309 ret
= kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, true, 2, true);
1311 close(ioeventfds
[i
]);
1316 /* Decide whether many devices are supported or not */
1317 ret
= i
== ARRAY_SIZE(ioeventfds
);
1320 kvm_set_ioeventfd_pio(ioeventfds
[i
], 0, i
, false, 2, true);
1321 close(ioeventfds
[i
]);
1329 static const KVMCapabilityInfo
*
1330 kvm_check_extension_list(KVMState
*s
, const KVMCapabilityInfo
*list
)
1332 while (list
->name
) {
1333 if (!kvm_check_extension(s
, list
->value
)) {
1341 void kvm_set_max_memslot_size(hwaddr max_slot_size
)
1344 ROUND_UP(max_slot_size
, qemu_real_host_page_size()) == max_slot_size
1346 kvm_max_slot_size
= max_slot_size
;
1349 static void kvm_set_phys_mem(KVMMemoryListener
*kml
,
1350 MemoryRegionSection
*section
, bool add
)
1354 MemoryRegion
*mr
= section
->mr
;
1355 bool writable
= !mr
->readonly
&& !mr
->rom_device
;
1356 hwaddr start_addr
, size
, slot_size
, mr_offset
;
1357 ram_addr_t ram_start_offset
;
1360 if (!memory_region_is_ram(mr
)) {
1361 if (writable
|| !kvm_readonly_mem_allowed
) {
1363 } else if (!mr
->romd_mode
) {
1364 /* If the memory device is not in romd_mode, then we actually want
1365 * to remove the kvm memory slot so all accesses will trap. */
1370 size
= kvm_align_section(section
, &start_addr
);
1375 /* The offset of the kvmslot within the memory region */
1376 mr_offset
= section
->offset_within_region
+ start_addr
-
1377 section
->offset_within_address_space
;
1379 /* use aligned delta to align the ram address and offset */
1380 ram
= memory_region_get_ram_ptr(mr
) + mr_offset
;
1381 ram_start_offset
= memory_region_get_ram_addr(mr
) + mr_offset
;
1387 slot_size
= MIN(kvm_max_slot_size
, size
);
1388 mem
= kvm_lookup_matching_slot(kml
, start_addr
, slot_size
);
1392 if (mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1394 * NOTE: We should be aware of the fact that here we're only
1395 * doing a best effort to sync dirty bits. No matter whether
1396 * we're using dirty log or dirty ring, we ignored two facts:
1398 * (1) dirty bits can reside in hardware buffers (PML)
1400 * (2) after we collected dirty bits here, pages can be dirtied
1401 * again before we do the final KVM_SET_USER_MEMORY_REGION to
1404 * Not easy. Let's cross the fingers until it's fixed.
1406 if (kvm_state
->kvm_dirty_ring_size
) {
1407 kvm_dirty_ring_reap_locked(kvm_state
, NULL
);
1409 kvm_slot_get_dirty_log(kvm_state
, mem
);
1411 kvm_slot_sync_dirty_pages(mem
);
1414 /* unregister the slot */
1415 g_free(mem
->dirty_bmap
);
1416 mem
->dirty_bmap
= NULL
;
1417 mem
->memory_size
= 0;
1419 err
= kvm_set_user_memory_region(kml
, mem
, false);
1421 fprintf(stderr
, "%s: error unregistering slot: %s\n",
1422 __func__
, strerror(-err
));
1425 start_addr
+= slot_size
;
1431 /* register the new slot */
1433 slot_size
= MIN(kvm_max_slot_size
, size
);
1434 mem
= kvm_alloc_slot(kml
);
1435 mem
->as_id
= kml
->as_id
;
1436 mem
->memory_size
= slot_size
;
1437 mem
->start_addr
= start_addr
;
1438 mem
->ram_start_offset
= ram_start_offset
;
1440 mem
->flags
= kvm_mem_flags(mr
);
1441 kvm_slot_init_dirty_bitmap(mem
);
1442 err
= kvm_set_user_memory_region(kml
, mem
, true);
1444 fprintf(stderr
, "%s: error registering slot: %s\n", __func__
,
1448 start_addr
+= slot_size
;
1449 ram_start_offset
+= slot_size
;
1458 static void *kvm_dirty_ring_reaper_thread(void *data
)
1461 struct KVMDirtyRingReaper
*r
= &s
->reaper
;
1463 rcu_register_thread();
1465 trace_kvm_dirty_ring_reaper("init");
1468 r
->reaper_state
= KVM_DIRTY_RING_REAPER_WAIT
;
1469 trace_kvm_dirty_ring_reaper("wait");
1471 * TODO: provide a smarter timeout rather than a constant?
1475 /* keep sleeping so that dirtylimit not be interfered by reaper */
1476 if (dirtylimit_in_service()) {
1480 trace_kvm_dirty_ring_reaper("wakeup");
1481 r
->reaper_state
= KVM_DIRTY_RING_REAPER_REAPING
;
1483 qemu_mutex_lock_iothread();
1484 kvm_dirty_ring_reap(s
, NULL
);
1485 qemu_mutex_unlock_iothread();
1487 r
->reaper_iteration
++;
1490 trace_kvm_dirty_ring_reaper("exit");
1492 rcu_unregister_thread();
1497 static int kvm_dirty_ring_reaper_init(KVMState
*s
)
1499 struct KVMDirtyRingReaper
*r
= &s
->reaper
;
1501 qemu_thread_create(&r
->reaper_thr
, "kvm-reaper",
1502 kvm_dirty_ring_reaper_thread
,
1503 s
, QEMU_THREAD_JOINABLE
);
1508 static void kvm_region_add(MemoryListener
*listener
,
1509 MemoryRegionSection
*section
)
1511 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1513 memory_region_ref(section
->mr
);
1514 kvm_set_phys_mem(kml
, section
, true);
1517 static void kvm_region_del(MemoryListener
*listener
,
1518 MemoryRegionSection
*section
)
1520 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1522 kvm_set_phys_mem(kml
, section
, false);
1523 memory_region_unref(section
->mr
);
1526 static void kvm_log_sync(MemoryListener
*listener
,
1527 MemoryRegionSection
*section
)
1529 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1532 kvm_physical_sync_dirty_bitmap(kml
, section
);
1536 static void kvm_log_sync_global(MemoryListener
*l
)
1538 KVMMemoryListener
*kml
= container_of(l
, KVMMemoryListener
, listener
);
1539 KVMState
*s
= kvm_state
;
1543 /* Flush all kernel dirty addresses into KVMSlot dirty bitmap */
1544 kvm_dirty_ring_flush();
1547 * TODO: make this faster when nr_slots is big while there are
1548 * only a few used slots (small VMs).
1551 for (i
= 0; i
< s
->nr_slots
; i
++) {
1552 mem
= &kml
->slots
[i
];
1553 if (mem
->memory_size
&& mem
->flags
& KVM_MEM_LOG_DIRTY_PAGES
) {
1554 kvm_slot_sync_dirty_pages(mem
);
1556 * This is not needed by KVM_GET_DIRTY_LOG because the
1557 * ioctl will unconditionally overwrite the whole region.
1558 * However kvm dirty ring has no such side effect.
1560 kvm_slot_reset_dirty_pages(mem
);
1566 static void kvm_log_clear(MemoryListener
*listener
,
1567 MemoryRegionSection
*section
)
1569 KVMMemoryListener
*kml
= container_of(listener
, KVMMemoryListener
, listener
);
1572 r
= kvm_physical_log_clear(kml
, section
);
1574 error_report_once("%s: kvm log clear failed: mr=%s "
1575 "offset=%"HWADDR_PRIx
" size=%"PRIx64
, __func__
,
1576 section
->mr
->name
, section
->offset_within_region
,
1577 int128_get64(section
->size
));
1582 static void kvm_mem_ioeventfd_add(MemoryListener
*listener
,
1583 MemoryRegionSection
*section
,
1584 bool match_data
, uint64_t data
,
1587 int fd
= event_notifier_get_fd(e
);
1590 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1591 data
, true, int128_get64(section
->size
),
1594 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1595 __func__
, strerror(-r
), -r
);
1600 static void kvm_mem_ioeventfd_del(MemoryListener
*listener
,
1601 MemoryRegionSection
*section
,
1602 bool match_data
, uint64_t data
,
1605 int fd
= event_notifier_get_fd(e
);
1608 r
= kvm_set_ioeventfd_mmio(fd
, section
->offset_within_address_space
,
1609 data
, false, int128_get64(section
->size
),
1612 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1613 __func__
, strerror(-r
), -r
);
1618 static void kvm_io_ioeventfd_add(MemoryListener
*listener
,
1619 MemoryRegionSection
*section
,
1620 bool match_data
, uint64_t data
,
1623 int fd
= event_notifier_get_fd(e
);
1626 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1627 data
, true, int128_get64(section
->size
),
1630 fprintf(stderr
, "%s: error adding ioeventfd: %s (%d)\n",
1631 __func__
, strerror(-r
), -r
);
1636 static void kvm_io_ioeventfd_del(MemoryListener
*listener
,
1637 MemoryRegionSection
*section
,
1638 bool match_data
, uint64_t data
,
1642 int fd
= event_notifier_get_fd(e
);
1645 r
= kvm_set_ioeventfd_pio(fd
, section
->offset_within_address_space
,
1646 data
, false, int128_get64(section
->size
),
1649 fprintf(stderr
, "%s: error deleting ioeventfd: %s (%d)\n",
1650 __func__
, strerror(-r
), -r
);
1655 void kvm_memory_listener_register(KVMState
*s
, KVMMemoryListener
*kml
,
1656 AddressSpace
*as
, int as_id
, const char *name
)
1660 kml
->slots
= g_new0(KVMSlot
, s
->nr_slots
);
1663 for (i
= 0; i
< s
->nr_slots
; i
++) {
1664 kml
->slots
[i
].slot
= i
;
1667 kml
->listener
.region_add
= kvm_region_add
;
1668 kml
->listener
.region_del
= kvm_region_del
;
1669 kml
->listener
.log_start
= kvm_log_start
;
1670 kml
->listener
.log_stop
= kvm_log_stop
;
1671 kml
->listener
.priority
= 10;
1672 kml
->listener
.name
= name
;
1674 if (s
->kvm_dirty_ring_size
) {
1675 kml
->listener
.log_sync_global
= kvm_log_sync_global
;
1677 kml
->listener
.log_sync
= kvm_log_sync
;
1678 kml
->listener
.log_clear
= kvm_log_clear
;
1681 memory_listener_register(&kml
->listener
, as
);
1683 for (i
= 0; i
< s
->nr_as
; ++i
) {
1692 static MemoryListener kvm_io_listener
= {
1694 .eventfd_add
= kvm_io_ioeventfd_add
,
1695 .eventfd_del
= kvm_io_ioeventfd_del
,
1699 int kvm_set_irq(KVMState
*s
, int irq
, int level
)
1701 struct kvm_irq_level event
;
1704 assert(kvm_async_interrupts_enabled());
1706 event
.level
= level
;
1708 ret
= kvm_vm_ioctl(s
, s
->irq_set_ioctl
, &event
);
1710 perror("kvm_set_irq");
1714 return (s
->irq_set_ioctl
== KVM_IRQ_LINE
) ? 1 : event
.status
;
1717 #ifdef KVM_CAP_IRQ_ROUTING
1718 typedef struct KVMMSIRoute
{
1719 struct kvm_irq_routing_entry kroute
;
1720 QTAILQ_ENTRY(KVMMSIRoute
) entry
;
1723 static void set_gsi(KVMState
*s
, unsigned int gsi
)
1725 set_bit(gsi
, s
->used_gsi_bitmap
);
1728 static void clear_gsi(KVMState
*s
, unsigned int gsi
)
1730 clear_bit(gsi
, s
->used_gsi_bitmap
);
1733 void kvm_init_irq_routing(KVMState
*s
)
1737 gsi_count
= kvm_check_extension(s
, KVM_CAP_IRQ_ROUTING
) - 1;
1738 if (gsi_count
> 0) {
1739 /* Round up so we can search ints using ffs */
1740 s
->used_gsi_bitmap
= bitmap_new(gsi_count
);
1741 s
->gsi_count
= gsi_count
;
1744 s
->irq_routes
= g_malloc0(sizeof(*s
->irq_routes
));
1745 s
->nr_allocated_irq_routes
= 0;
1747 if (!kvm_direct_msi_allowed
) {
1748 for (i
= 0; i
< KVM_MSI_HASHTAB_SIZE
; i
++) {
1749 QTAILQ_INIT(&s
->msi_hashtab
[i
]);
1753 kvm_arch_init_irq_routing(s
);
1756 void kvm_irqchip_commit_routes(KVMState
*s
)
1760 if (kvm_gsi_direct_mapping()) {
1764 if (!kvm_gsi_routing_enabled()) {
1768 s
->irq_routes
->flags
= 0;
1769 trace_kvm_irqchip_commit_routes();
1770 ret
= kvm_vm_ioctl(s
, KVM_SET_GSI_ROUTING
, s
->irq_routes
);
1774 static void kvm_add_routing_entry(KVMState
*s
,
1775 struct kvm_irq_routing_entry
*entry
)
1777 struct kvm_irq_routing_entry
*new;
1780 if (s
->irq_routes
->nr
== s
->nr_allocated_irq_routes
) {
1781 n
= s
->nr_allocated_irq_routes
* 2;
1785 size
= sizeof(struct kvm_irq_routing
);
1786 size
+= n
* sizeof(*new);
1787 s
->irq_routes
= g_realloc(s
->irq_routes
, size
);
1788 s
->nr_allocated_irq_routes
= n
;
1790 n
= s
->irq_routes
->nr
++;
1791 new = &s
->irq_routes
->entries
[n
];
1795 set_gsi(s
, entry
->gsi
);
1798 static int kvm_update_routing_entry(KVMState
*s
,
1799 struct kvm_irq_routing_entry
*new_entry
)
1801 struct kvm_irq_routing_entry
*entry
;
1804 for (n
= 0; n
< s
->irq_routes
->nr
; n
++) {
1805 entry
= &s
->irq_routes
->entries
[n
];
1806 if (entry
->gsi
!= new_entry
->gsi
) {
1810 if(!memcmp(entry
, new_entry
, sizeof *entry
)) {
1814 *entry
= *new_entry
;
1822 void kvm_irqchip_add_irq_route(KVMState
*s
, int irq
, int irqchip
, int pin
)
1824 struct kvm_irq_routing_entry e
= {};
1826 assert(pin
< s
->gsi_count
);
1829 e
.type
= KVM_IRQ_ROUTING_IRQCHIP
;
1831 e
.u
.irqchip
.irqchip
= irqchip
;
1832 e
.u
.irqchip
.pin
= pin
;
1833 kvm_add_routing_entry(s
, &e
);
1836 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
1838 struct kvm_irq_routing_entry
*e
;
1841 if (kvm_gsi_direct_mapping()) {
1845 for (i
= 0; i
< s
->irq_routes
->nr
; i
++) {
1846 e
= &s
->irq_routes
->entries
[i
];
1847 if (e
->gsi
== virq
) {
1848 s
->irq_routes
->nr
--;
1849 *e
= s
->irq_routes
->entries
[s
->irq_routes
->nr
];
1853 kvm_arch_release_virq_post(virq
);
1854 trace_kvm_irqchip_release_virq(virq
);
1857 void kvm_irqchip_add_change_notifier(Notifier
*n
)
1859 notifier_list_add(&kvm_irqchip_change_notifiers
, n
);
1862 void kvm_irqchip_remove_change_notifier(Notifier
*n
)
1867 void kvm_irqchip_change_notify(void)
1869 notifier_list_notify(&kvm_irqchip_change_notifiers
, NULL
);
1872 static unsigned int kvm_hash_msi(uint32_t data
)
1874 /* This is optimized for IA32 MSI layout. However, no other arch shall
1875 * repeat the mistake of not providing a direct MSI injection API. */
1879 static void kvm_flush_dynamic_msi_routes(KVMState
*s
)
1881 KVMMSIRoute
*route
, *next
;
1884 for (hash
= 0; hash
< KVM_MSI_HASHTAB_SIZE
; hash
++) {
1885 QTAILQ_FOREACH_SAFE(route
, &s
->msi_hashtab
[hash
], entry
, next
) {
1886 kvm_irqchip_release_virq(s
, route
->kroute
.gsi
);
1887 QTAILQ_REMOVE(&s
->msi_hashtab
[hash
], route
, entry
);
1893 static int kvm_irqchip_get_virq(KVMState
*s
)
1898 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1899 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1900 * number can succeed even though a new route entry cannot be added.
1901 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1903 if (!kvm_direct_msi_allowed
&& s
->irq_routes
->nr
== s
->gsi_count
) {
1904 kvm_flush_dynamic_msi_routes(s
);
1907 /* Return the lowest unused GSI in the bitmap */
1908 next_virq
= find_first_zero_bit(s
->used_gsi_bitmap
, s
->gsi_count
);
1909 if (next_virq
>= s
->gsi_count
) {
1916 static KVMMSIRoute
*kvm_lookup_msi_route(KVMState
*s
, MSIMessage msg
)
1918 unsigned int hash
= kvm_hash_msi(msg
.data
);
1921 QTAILQ_FOREACH(route
, &s
->msi_hashtab
[hash
], entry
) {
1922 if (route
->kroute
.u
.msi
.address_lo
== (uint32_t)msg
.address
&&
1923 route
->kroute
.u
.msi
.address_hi
== (msg
.address
>> 32) &&
1924 route
->kroute
.u
.msi
.data
== le32_to_cpu(msg
.data
)) {
1931 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
1936 if (kvm_direct_msi_allowed
) {
1937 msi
.address_lo
= (uint32_t)msg
.address
;
1938 msi
.address_hi
= msg
.address
>> 32;
1939 msi
.data
= le32_to_cpu(msg
.data
);
1941 memset(msi
.pad
, 0, sizeof(msi
.pad
));
1943 return kvm_vm_ioctl(s
, KVM_SIGNAL_MSI
, &msi
);
1946 route
= kvm_lookup_msi_route(s
, msg
);
1950 virq
= kvm_irqchip_get_virq(s
);
1955 route
= g_new0(KVMMSIRoute
, 1);
1956 route
->kroute
.gsi
= virq
;
1957 route
->kroute
.type
= KVM_IRQ_ROUTING_MSI
;
1958 route
->kroute
.flags
= 0;
1959 route
->kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
1960 route
->kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
1961 route
->kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
1963 kvm_add_routing_entry(s
, &route
->kroute
);
1964 kvm_irqchip_commit_routes(s
);
1966 QTAILQ_INSERT_TAIL(&s
->msi_hashtab
[kvm_hash_msi(msg
.data
)], route
,
1970 assert(route
->kroute
.type
== KVM_IRQ_ROUTING_MSI
);
1972 return kvm_set_irq(s
, route
->kroute
.gsi
, 1);
1975 int kvm_irqchip_add_msi_route(KVMRouteChange
*c
, int vector
, PCIDevice
*dev
)
1977 struct kvm_irq_routing_entry kroute
= {};
1980 MSIMessage msg
= {0, 0};
1982 if (pci_available
&& dev
) {
1983 msg
= pci_get_msi_message(dev
, vector
);
1986 if (kvm_gsi_direct_mapping()) {
1987 return kvm_arch_msi_data_to_gsi(msg
.data
);
1990 if (!kvm_gsi_routing_enabled()) {
1994 virq
= kvm_irqchip_get_virq(s
);
2000 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
2002 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
2003 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
2004 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
2005 if (pci_available
&& kvm_msi_devid_required()) {
2006 kroute
.flags
= KVM_MSI_VALID_DEVID
;
2007 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
2009 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
2010 kvm_irqchip_release_virq(s
, virq
);
2014 trace_kvm_irqchip_add_msi_route(dev
? dev
->name
: (char *)"N/A",
2017 kvm_add_routing_entry(s
, &kroute
);
2018 kvm_arch_add_msi_route_post(&kroute
, vector
, dev
);
2024 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
,
2027 struct kvm_irq_routing_entry kroute
= {};
2029 if (kvm_gsi_direct_mapping()) {
2033 if (!kvm_irqchip_in_kernel()) {
2038 kroute
.type
= KVM_IRQ_ROUTING_MSI
;
2040 kroute
.u
.msi
.address_lo
= (uint32_t)msg
.address
;
2041 kroute
.u
.msi
.address_hi
= msg
.address
>> 32;
2042 kroute
.u
.msi
.data
= le32_to_cpu(msg
.data
);
2043 if (pci_available
&& kvm_msi_devid_required()) {
2044 kroute
.flags
= KVM_MSI_VALID_DEVID
;
2045 kroute
.u
.msi
.devid
= pci_requester_id(dev
);
2047 if (kvm_arch_fixup_msi_route(&kroute
, msg
.address
, msg
.data
, dev
)) {
2051 trace_kvm_irqchip_update_msi_route(virq
);
2053 return kvm_update_routing_entry(s
, &kroute
);
2056 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
2057 EventNotifier
*resample
, int virq
,
2060 int fd
= event_notifier_get_fd(event
);
2061 int rfd
= resample
? event_notifier_get_fd(resample
) : -1;
2063 struct kvm_irqfd irqfd
= {
2066 .flags
= assign
? 0 : KVM_IRQFD_FLAG_DEASSIGN
,
2071 if (kvm_irqchip_is_split()) {
2073 * When the slow irqchip (e.g. IOAPIC) is in the
2074 * userspace, KVM kernel resamplefd will not work because
2075 * the EOI of the interrupt will be delivered to userspace
2076 * instead, so the KVM kernel resamplefd kick will be
2077 * skipped. The userspace here mimics what the kernel
2078 * provides with resamplefd, remember the resamplefd and
2079 * kick it when we receive EOI of this IRQ.
2081 * This is hackery because IOAPIC is mostly bypassed
2082 * (except EOI broadcasts) when irqfd is used. However
2083 * this can bring much performance back for split irqchip
2084 * with INTx IRQs (for VFIO, this gives 93% perf of the
2085 * full fast path, which is 46% perf boost comparing to
2086 * the INTx slow path).
2088 kvm_resample_fd_insert(virq
, resample
);
2090 irqfd
.flags
|= KVM_IRQFD_FLAG_RESAMPLE
;
2091 irqfd
.resamplefd
= rfd
;
2093 } else if (!assign
) {
2094 if (kvm_irqchip_is_split()) {
2095 kvm_resample_fd_remove(virq
);
2099 if (!kvm_irqfds_enabled()) {
2103 return kvm_vm_ioctl(s
, KVM_IRQFD
, &irqfd
);
2106 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
2108 struct kvm_irq_routing_entry kroute
= {};
2111 if (!kvm_gsi_routing_enabled()) {
2115 virq
= kvm_irqchip_get_virq(s
);
2121 kroute
.type
= KVM_IRQ_ROUTING_S390_ADAPTER
;
2123 kroute
.u
.adapter
.summary_addr
= adapter
->summary_addr
;
2124 kroute
.u
.adapter
.ind_addr
= adapter
->ind_addr
;
2125 kroute
.u
.adapter
.summary_offset
= adapter
->summary_offset
;
2126 kroute
.u
.adapter
.ind_offset
= adapter
->ind_offset
;
2127 kroute
.u
.adapter
.adapter_id
= adapter
->adapter_id
;
2129 kvm_add_routing_entry(s
, &kroute
);
2134 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
2136 struct kvm_irq_routing_entry kroute
= {};
2139 if (!kvm_gsi_routing_enabled()) {
2142 if (!kvm_check_extension(s
, KVM_CAP_HYPERV_SYNIC
)) {
2145 virq
= kvm_irqchip_get_virq(s
);
2151 kroute
.type
= KVM_IRQ_ROUTING_HV_SINT
;
2153 kroute
.u
.hv_sint
.vcpu
= vcpu
;
2154 kroute
.u
.hv_sint
.sint
= sint
;
2156 kvm_add_routing_entry(s
, &kroute
);
2157 kvm_irqchip_commit_routes(s
);
2162 #else /* !KVM_CAP_IRQ_ROUTING */
2164 void kvm_init_irq_routing(KVMState
*s
)
2168 void kvm_irqchip_release_virq(KVMState
*s
, int virq
)
2172 int kvm_irqchip_send_msi(KVMState
*s
, MSIMessage msg
)
2177 int kvm_irqchip_add_msi_route(KVMRouteChange
*c
, int vector
, PCIDevice
*dev
)
2182 int kvm_irqchip_add_adapter_route(KVMState
*s
, AdapterInfo
*adapter
)
2187 int kvm_irqchip_add_hv_sint_route(KVMState
*s
, uint32_t vcpu
, uint32_t sint
)
2192 static int kvm_irqchip_assign_irqfd(KVMState
*s
, EventNotifier
*event
,
2193 EventNotifier
*resample
, int virq
,
2199 int kvm_irqchip_update_msi_route(KVMState
*s
, int virq
, MSIMessage msg
)
2203 #endif /* !KVM_CAP_IRQ_ROUTING */
2205 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
2206 EventNotifier
*rn
, int virq
)
2208 return kvm_irqchip_assign_irqfd(s
, n
, rn
, virq
, true);
2211 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState
*s
, EventNotifier
*n
,
2214 return kvm_irqchip_assign_irqfd(s
, n
, NULL
, virq
, false);
2217 int kvm_irqchip_add_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
2218 EventNotifier
*rn
, qemu_irq irq
)
2221 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
2226 return kvm_irqchip_add_irqfd_notifier_gsi(s
, n
, rn
, GPOINTER_TO_INT(gsi
));
2229 int kvm_irqchip_remove_irqfd_notifier(KVMState
*s
, EventNotifier
*n
,
2233 gboolean found
= g_hash_table_lookup_extended(s
->gsimap
, irq
, &key
, &gsi
);
2238 return kvm_irqchip_remove_irqfd_notifier_gsi(s
, n
, GPOINTER_TO_INT(gsi
));
2241 void kvm_irqchip_set_qemuirq_gsi(KVMState
*s
, qemu_irq irq
, int gsi
)
2243 g_hash_table_insert(s
->gsimap
, irq
, GINT_TO_POINTER(gsi
));
2246 static void kvm_irqchip_create(KVMState
*s
)
2250 assert(s
->kernel_irqchip_split
!= ON_OFF_AUTO_AUTO
);
2251 if (kvm_check_extension(s
, KVM_CAP_IRQCHIP
)) {
2253 } else if (kvm_check_extension(s
, KVM_CAP_S390_IRQCHIP
)) {
2254 ret
= kvm_vm_enable_cap(s
, KVM_CAP_S390_IRQCHIP
, 0);
2256 fprintf(stderr
, "Enable kernel irqchip failed: %s\n", strerror(-ret
));
2263 /* First probe and see if there's a arch-specific hook to create the
2264 * in-kernel irqchip for us */
2265 ret
= kvm_arch_irqchip_create(s
);
2267 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_ON
) {
2268 error_report("Split IRQ chip mode not supported.");
2271 ret
= kvm_vm_ioctl(s
, KVM_CREATE_IRQCHIP
);
2275 fprintf(stderr
, "Create kernel irqchip failed: %s\n", strerror(-ret
));
2279 kvm_kernel_irqchip
= true;
2280 /* If we have an in-kernel IRQ chip then we must have asynchronous
2281 * interrupt delivery (though the reverse is not necessarily true)
2283 kvm_async_interrupts_allowed
= true;
2284 kvm_halt_in_kernel_allowed
= true;
2286 kvm_init_irq_routing(s
);
2288 s
->gsimap
= g_hash_table_new(g_direct_hash
, g_direct_equal
);
2291 /* Find number of supported CPUs using the recommended
2292 * procedure from the kernel API documentation to cope with
2293 * older kernels that may be missing capabilities.
2295 static int kvm_recommended_vcpus(KVMState
*s
)
2297 int ret
= kvm_vm_check_extension(s
, KVM_CAP_NR_VCPUS
);
2298 return (ret
) ? ret
: 4;
2301 static int kvm_max_vcpus(KVMState
*s
)
2303 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPUS
);
2304 return (ret
) ? ret
: kvm_recommended_vcpus(s
);
2307 static int kvm_max_vcpu_id(KVMState
*s
)
2309 int ret
= kvm_check_extension(s
, KVM_CAP_MAX_VCPU_ID
);
2310 return (ret
) ? ret
: kvm_max_vcpus(s
);
2313 bool kvm_vcpu_id_is_valid(int vcpu_id
)
2315 KVMState
*s
= KVM_STATE(current_accel());
2316 return vcpu_id
>= 0 && vcpu_id
< kvm_max_vcpu_id(s
);
2319 bool kvm_dirty_ring_enabled(void)
2321 return kvm_state
->kvm_dirty_ring_size
? true : false;
2324 static void query_stats_cb(StatsResultList
**result
, StatsTarget target
,
2325 strList
*names
, strList
*targets
, Error
**errp
);
2326 static void query_stats_schemas_cb(StatsSchemaList
**result
, Error
**errp
);
2328 uint32_t kvm_dirty_ring_size(void)
2330 return kvm_state
->kvm_dirty_ring_size
;
2333 static int kvm_init(MachineState
*ms
)
2335 MachineClass
*mc
= MACHINE_GET_CLASS(ms
);
2336 static const char upgrade_note
[] =
2337 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
2338 "(see http://sourceforge.net/projects/kvm).\n";
2343 { "SMP", ms
->smp
.cpus
},
2344 { "hotpluggable", ms
->smp
.max_cpus
},
2347 int soft_vcpus_limit
, hard_vcpus_limit
;
2349 const KVMCapabilityInfo
*missing_cap
;
2352 uint64_t dirty_log_manual_caps
;
2354 qemu_mutex_init(&kml_slots_lock
);
2356 s
= KVM_STATE(ms
->accelerator
);
2359 * On systems where the kernel can support different base page
2360 * sizes, host page size may be different from TARGET_PAGE_SIZE,
2361 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
2362 * page size for the system though.
2364 assert(TARGET_PAGE_SIZE
<= qemu_real_host_page_size());
2368 #ifdef KVM_CAP_SET_GUEST_DEBUG
2369 QTAILQ_INIT(&s
->kvm_sw_breakpoints
);
2371 QLIST_INIT(&s
->kvm_parked_vcpus
);
2372 s
->fd
= qemu_open_old("/dev/kvm", O_RDWR
);
2374 fprintf(stderr
, "Could not access KVM kernel module: %m\n");
2379 ret
= kvm_ioctl(s
, KVM_GET_API_VERSION
, 0);
2380 if (ret
< KVM_API_VERSION
) {
2384 fprintf(stderr
, "kvm version too old\n");
2388 if (ret
> KVM_API_VERSION
) {
2390 fprintf(stderr
, "kvm version not supported\n");
2394 kvm_immediate_exit
= kvm_check_extension(s
, KVM_CAP_IMMEDIATE_EXIT
);
2395 s
->nr_slots
= kvm_check_extension(s
, KVM_CAP_NR_MEMSLOTS
);
2397 /* If unspecified, use the default value */
2402 s
->nr_as
= kvm_check_extension(s
, KVM_CAP_MULTI_ADDRESS_SPACE
);
2403 if (s
->nr_as
<= 1) {
2406 s
->as
= g_new0(struct KVMAs
, s
->nr_as
);
2408 if (object_property_find(OBJECT(current_machine
), "kvm-type")) {
2409 g_autofree
char *kvm_type
= object_property_get_str(OBJECT(current_machine
),
2412 type
= mc
->kvm_type(ms
, kvm_type
);
2413 } else if (mc
->kvm_type
) {
2414 type
= mc
->kvm_type(ms
, NULL
);
2418 ret
= kvm_ioctl(s
, KVM_CREATE_VM
, type
);
2419 } while (ret
== -EINTR
);
2422 fprintf(stderr
, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret
,
2426 if (ret
== -EINVAL
) {
2428 "Host kernel setup problem detected. Please verify:\n");
2429 fprintf(stderr
, "- for kernels supporting the switch_amode or"
2430 " user_mode parameters, whether\n");
2432 " user space is running in primary address space\n");
2434 "- for kernels supporting the vm.allocate_pgste sysctl, "
2435 "whether it is enabled\n");
2437 #elif defined(TARGET_PPC)
2438 if (ret
== -EINVAL
) {
2440 "PPC KVM module is not loaded. Try modprobe kvm_%s.\n",
2441 (type
== 2) ? "pr" : "hv");
2449 /* check the vcpu limits */
2450 soft_vcpus_limit
= kvm_recommended_vcpus(s
);
2451 hard_vcpus_limit
= kvm_max_vcpus(s
);
2454 if (nc
->num
> soft_vcpus_limit
) {
2455 warn_report("Number of %s cpus requested (%d) exceeds "
2456 "the recommended cpus supported by KVM (%d)",
2457 nc
->name
, nc
->num
, soft_vcpus_limit
);
2459 if (nc
->num
> hard_vcpus_limit
) {
2460 fprintf(stderr
, "Number of %s cpus requested (%d) exceeds "
2461 "the maximum cpus supported by KVM (%d)\n",
2462 nc
->name
, nc
->num
, hard_vcpus_limit
);
2469 missing_cap
= kvm_check_extension_list(s
, kvm_required_capabilites
);
2472 kvm_check_extension_list(s
, kvm_arch_required_capabilities
);
2476 fprintf(stderr
, "kvm does not support %s\n%s",
2477 missing_cap
->name
, upgrade_note
);
2481 s
->coalesced_mmio
= kvm_check_extension(s
, KVM_CAP_COALESCED_MMIO
);
2482 s
->coalesced_pio
= s
->coalesced_mmio
&&
2483 kvm_check_extension(s
, KVM_CAP_COALESCED_PIO
);
2486 * Enable KVM dirty ring if supported, otherwise fall back to
2487 * dirty logging mode
2489 if (s
->kvm_dirty_ring_size
> 0) {
2490 uint64_t ring_bytes
;
2492 ring_bytes
= s
->kvm_dirty_ring_size
* sizeof(struct kvm_dirty_gfn
);
2494 /* Read the max supported pages */
2495 ret
= kvm_vm_check_extension(s
, KVM_CAP_DIRTY_LOG_RING
);
2497 if (ring_bytes
> ret
) {
2498 error_report("KVM dirty ring size %" PRIu32
" too big "
2499 "(maximum is %ld). Please use a smaller value.",
2500 s
->kvm_dirty_ring_size
,
2501 (long)ret
/ sizeof(struct kvm_dirty_gfn
));
2506 ret
= kvm_vm_enable_cap(s
, KVM_CAP_DIRTY_LOG_RING
, 0, ring_bytes
);
2508 error_report("Enabling of KVM dirty ring failed: %s. "
2509 "Suggested minimum value is 1024.", strerror(-ret
));
2513 s
->kvm_dirty_ring_bytes
= ring_bytes
;
2515 warn_report("KVM dirty ring not available, using bitmap method");
2516 s
->kvm_dirty_ring_size
= 0;
2521 * KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 is not needed when dirty ring is
2522 * enabled. More importantly, KVM_DIRTY_LOG_INITIALLY_SET will assume no
2523 * page is wr-protected initially, which is against how kvm dirty ring is
2524 * usage - kvm dirty ring requires all pages are wr-protected at the very
2525 * beginning. Enabling this feature for dirty ring causes data corruption.
2527 * TODO: Without KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 and kvm clear dirty log,
2528 * we may expect a higher stall time when starting the migration. In the
2529 * future we can enable KVM_CLEAR_DIRTY_LOG to work with dirty ring too:
2530 * instead of clearing dirty bit, it can be a way to explicitly wr-protect
2533 if (!s
->kvm_dirty_ring_size
) {
2534 dirty_log_manual_caps
=
2535 kvm_check_extension(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
);
2536 dirty_log_manual_caps
&= (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE
|
2537 KVM_DIRTY_LOG_INITIALLY_SET
);
2538 s
->manual_dirty_log_protect
= dirty_log_manual_caps
;
2539 if (dirty_log_manual_caps
) {
2540 ret
= kvm_vm_enable_cap(s
, KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2
, 0,
2541 dirty_log_manual_caps
);
2543 warn_report("Trying to enable capability %"PRIu64
" of "
2544 "KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 but failed. "
2545 "Falling back to the legacy mode. ",
2546 dirty_log_manual_caps
);
2547 s
->manual_dirty_log_protect
= 0;
2552 #ifdef KVM_CAP_VCPU_EVENTS
2553 s
->vcpu_events
= kvm_check_extension(s
, KVM_CAP_VCPU_EVENTS
);
2556 s
->robust_singlestep
=
2557 kvm_check_extension(s
, KVM_CAP_X86_ROBUST_SINGLESTEP
);
2559 #ifdef KVM_CAP_DEBUGREGS
2560 s
->debugregs
= kvm_check_extension(s
, KVM_CAP_DEBUGREGS
);
2563 s
->max_nested_state_len
= kvm_check_extension(s
, KVM_CAP_NESTED_STATE
);
2565 #ifdef KVM_CAP_IRQ_ROUTING
2566 kvm_direct_msi_allowed
= (kvm_check_extension(s
, KVM_CAP_SIGNAL_MSI
) > 0);
2569 s
->intx_set_mask
= kvm_check_extension(s
, KVM_CAP_PCI_2_3
);
2571 s
->irq_set_ioctl
= KVM_IRQ_LINE
;
2572 if (kvm_check_extension(s
, KVM_CAP_IRQ_INJECT_STATUS
)) {
2573 s
->irq_set_ioctl
= KVM_IRQ_LINE_STATUS
;
2576 kvm_readonly_mem_allowed
=
2577 (kvm_check_extension(s
, KVM_CAP_READONLY_MEM
) > 0);
2579 kvm_eventfds_allowed
=
2580 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD
) > 0);
2582 kvm_irqfds_allowed
=
2583 (kvm_check_extension(s
, KVM_CAP_IRQFD
) > 0);
2585 kvm_resamplefds_allowed
=
2586 (kvm_check_extension(s
, KVM_CAP_IRQFD_RESAMPLE
) > 0);
2588 kvm_vm_attributes_allowed
=
2589 (kvm_check_extension(s
, KVM_CAP_VM_ATTRIBUTES
) > 0);
2591 kvm_ioeventfd_any_length_allowed
=
2592 (kvm_check_extension(s
, KVM_CAP_IOEVENTFD_ANY_LENGTH
) > 0);
2594 #ifdef KVM_CAP_SET_GUEST_DEBUG
2595 kvm_has_guest_debug
=
2596 (kvm_check_extension(s
, KVM_CAP_SET_GUEST_DEBUG
) > 0);
2599 kvm_sstep_flags
= 0;
2600 if (kvm_has_guest_debug
) {
2601 kvm_sstep_flags
= SSTEP_ENABLE
;
2603 #if defined KVM_CAP_SET_GUEST_DEBUG2
2604 int guest_debug_flags
=
2605 kvm_check_extension(s
, KVM_CAP_SET_GUEST_DEBUG2
);
2607 if (guest_debug_flags
& KVM_GUESTDBG_BLOCKIRQ
) {
2608 kvm_sstep_flags
|= SSTEP_NOIRQ
;
2615 ret
= kvm_arch_init(ms
, s
);
2620 if (s
->kernel_irqchip_split
== ON_OFF_AUTO_AUTO
) {
2621 s
->kernel_irqchip_split
= mc
->default_kernel_irqchip_split
? ON_OFF_AUTO_ON
: ON_OFF_AUTO_OFF
;
2624 qemu_register_reset(kvm_unpoison_all
, NULL
);
2626 if (s
->kernel_irqchip_allowed
) {
2627 kvm_irqchip_create(s
);
2630 if (kvm_eventfds_allowed
) {
2631 s
->memory_listener
.listener
.eventfd_add
= kvm_mem_ioeventfd_add
;
2632 s
->memory_listener
.listener
.eventfd_del
= kvm_mem_ioeventfd_del
;
2634 s
->memory_listener
.listener
.coalesced_io_add
= kvm_coalesce_mmio_region
;
2635 s
->memory_listener
.listener
.coalesced_io_del
= kvm_uncoalesce_mmio_region
;
2637 kvm_memory_listener_register(s
, &s
->memory_listener
,
2638 &address_space_memory
, 0, "kvm-memory");
2639 if (kvm_eventfds_allowed
) {
2640 memory_listener_register(&kvm_io_listener
,
2643 memory_listener_register(&kvm_coalesced_pio_listener
,
2646 s
->many_ioeventfds
= kvm_check_many_ioeventfds();
2648 s
->sync_mmu
= !!kvm_vm_check_extension(kvm_state
, KVM_CAP_SYNC_MMU
);
2650 ret
= ram_block_discard_disable(true);
2654 if (s
->kvm_dirty_ring_size
) {
2655 ret
= kvm_dirty_ring_reaper_init(s
);
2661 if (kvm_check_extension(kvm_state
, KVM_CAP_BINARY_STATS_FD
)) {
2662 add_stats_callbacks(STATS_PROVIDER_KVM
, query_stats_cb
,
2663 query_stats_schemas_cb
);
2676 g_free(s
->memory_listener
.slots
);
2681 void kvm_set_sigmask_len(KVMState
*s
, unsigned int sigmask_len
)
2683 s
->sigmask_len
= sigmask_len
;
2686 static void kvm_handle_io(uint16_t port
, MemTxAttrs attrs
, void *data
, int direction
,
2687 int size
, uint32_t count
)
2690 uint8_t *ptr
= data
;
2692 for (i
= 0; i
< count
; i
++) {
2693 address_space_rw(&address_space_io
, port
, attrs
,
2695 direction
== KVM_EXIT_IO_OUT
);
2700 static int kvm_handle_internal_error(CPUState
*cpu
, struct kvm_run
*run
)
2702 fprintf(stderr
, "KVM internal error. Suberror: %d\n",
2703 run
->internal
.suberror
);
2705 if (kvm_check_extension(kvm_state
, KVM_CAP_INTERNAL_ERROR_DATA
)) {
2708 for (i
= 0; i
< run
->internal
.ndata
; ++i
) {
2709 fprintf(stderr
, "extra data[%d]: 0x%016"PRIx64
"\n",
2710 i
, (uint64_t)run
->internal
.data
[i
]);
2713 if (run
->internal
.suberror
== KVM_INTERNAL_ERROR_EMULATION
) {
2714 fprintf(stderr
, "emulation failure\n");
2715 if (!kvm_arch_stop_on_emulation_error(cpu
)) {
2716 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
2717 return EXCP_INTERRUPT
;
2720 /* FIXME: Should trigger a qmp message to let management know
2721 * something went wrong.
2726 void kvm_flush_coalesced_mmio_buffer(void)
2728 KVMState
*s
= kvm_state
;
2730 if (s
->coalesced_flush_in_progress
) {
2734 s
->coalesced_flush_in_progress
= true;
2736 if (s
->coalesced_mmio_ring
) {
2737 struct kvm_coalesced_mmio_ring
*ring
= s
->coalesced_mmio_ring
;
2738 while (ring
->first
!= ring
->last
) {
2739 struct kvm_coalesced_mmio
*ent
;
2741 ent
= &ring
->coalesced_mmio
[ring
->first
];
2743 if (ent
->pio
== 1) {
2744 address_space_write(&address_space_io
, ent
->phys_addr
,
2745 MEMTXATTRS_UNSPECIFIED
, ent
->data
,
2748 cpu_physical_memory_write(ent
->phys_addr
, ent
->data
, ent
->len
);
2751 ring
->first
= (ring
->first
+ 1) % KVM_COALESCED_MMIO_MAX
;
2755 s
->coalesced_flush_in_progress
= false;
2758 bool kvm_cpu_check_are_resettable(void)
2760 return kvm_arch_cpu_check_are_resettable();
2763 static void do_kvm_cpu_synchronize_state(CPUState
*cpu
, run_on_cpu_data arg
)
2765 if (!cpu
->vcpu_dirty
) {
2766 kvm_arch_get_registers(cpu
);
2767 cpu
->vcpu_dirty
= true;
2771 void kvm_cpu_synchronize_state(CPUState
*cpu
)
2773 if (!cpu
->vcpu_dirty
) {
2774 run_on_cpu(cpu
, do_kvm_cpu_synchronize_state
, RUN_ON_CPU_NULL
);
2778 static void do_kvm_cpu_synchronize_post_reset(CPUState
*cpu
, run_on_cpu_data arg
)
2780 kvm_arch_put_registers(cpu
, KVM_PUT_RESET_STATE
);
2781 cpu
->vcpu_dirty
= false;
2784 void kvm_cpu_synchronize_post_reset(CPUState
*cpu
)
2786 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_reset
, RUN_ON_CPU_NULL
);
2789 static void do_kvm_cpu_synchronize_post_init(CPUState
*cpu
, run_on_cpu_data arg
)
2791 kvm_arch_put_registers(cpu
, KVM_PUT_FULL_STATE
);
2792 cpu
->vcpu_dirty
= false;
2795 void kvm_cpu_synchronize_post_init(CPUState
*cpu
)
2797 run_on_cpu(cpu
, do_kvm_cpu_synchronize_post_init
, RUN_ON_CPU_NULL
);
2800 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
, run_on_cpu_data arg
)
2802 cpu
->vcpu_dirty
= true;
2805 void kvm_cpu_synchronize_pre_loadvm(CPUState
*cpu
)
2807 run_on_cpu(cpu
, do_kvm_cpu_synchronize_pre_loadvm
, RUN_ON_CPU_NULL
);
2810 #ifdef KVM_HAVE_MCE_INJECTION
2811 static __thread
void *pending_sigbus_addr
;
2812 static __thread
int pending_sigbus_code
;
2813 static __thread
bool have_sigbus_pending
;
2816 static void kvm_cpu_kick(CPUState
*cpu
)
2818 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 1);
2821 static void kvm_cpu_kick_self(void)
2823 if (kvm_immediate_exit
) {
2824 kvm_cpu_kick(current_cpu
);
2826 qemu_cpu_kick_self();
2830 static void kvm_eat_signals(CPUState
*cpu
)
2832 struct timespec ts
= { 0, 0 };
2838 if (kvm_immediate_exit
) {
2839 qatomic_set(&cpu
->kvm_run
->immediate_exit
, 0);
2840 /* Write kvm_run->immediate_exit before the cpu->exit_request
2841 * write in kvm_cpu_exec.
2847 sigemptyset(&waitset
);
2848 sigaddset(&waitset
, SIG_IPI
);
2851 r
= sigtimedwait(&waitset
, &siginfo
, &ts
);
2852 if (r
== -1 && !(errno
== EAGAIN
|| errno
== EINTR
)) {
2853 perror("sigtimedwait");
2857 r
= sigpending(&chkset
);
2859 perror("sigpending");
2862 } while (sigismember(&chkset
, SIG_IPI
));
2865 int kvm_cpu_exec(CPUState
*cpu
)
2867 struct kvm_run
*run
= cpu
->kvm_run
;
2870 DPRINTF("kvm_cpu_exec()\n");
2872 if (kvm_arch_process_async_events(cpu
)) {
2873 qatomic_set(&cpu
->exit_request
, 0);
2877 qemu_mutex_unlock_iothread();
2878 cpu_exec_start(cpu
);
2883 if (cpu
->vcpu_dirty
) {
2884 kvm_arch_put_registers(cpu
, KVM_PUT_RUNTIME_STATE
);
2885 cpu
->vcpu_dirty
= false;
2888 kvm_arch_pre_run(cpu
, run
);
2889 if (qatomic_read(&cpu
->exit_request
)) {
2890 DPRINTF("interrupt exit requested\n");
2892 * KVM requires us to reenter the kernel after IO exits to complete
2893 * instruction emulation. This self-signal will ensure that we
2896 kvm_cpu_kick_self();
2899 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
2900 * Matching barrier in kvm_eat_signals.
2904 run_ret
= kvm_vcpu_ioctl(cpu
, KVM_RUN
, 0);
2906 attrs
= kvm_arch_post_run(cpu
, run
);
2908 #ifdef KVM_HAVE_MCE_INJECTION
2909 if (unlikely(have_sigbus_pending
)) {
2910 qemu_mutex_lock_iothread();
2911 kvm_arch_on_sigbus_vcpu(cpu
, pending_sigbus_code
,
2912 pending_sigbus_addr
);
2913 have_sigbus_pending
= false;
2914 qemu_mutex_unlock_iothread();
2919 if (run_ret
== -EINTR
|| run_ret
== -EAGAIN
) {
2920 DPRINTF("io window exit\n");
2921 kvm_eat_signals(cpu
);
2922 ret
= EXCP_INTERRUPT
;
2925 fprintf(stderr
, "error: kvm run failed %s\n",
2926 strerror(-run_ret
));
2928 if (run_ret
== -EBUSY
) {
2930 "This is probably because your SMT is enabled.\n"
2931 "VCPU can only run on primary threads with all "
2932 "secondary threads offline.\n");
2939 trace_kvm_run_exit(cpu
->cpu_index
, run
->exit_reason
);
2940 switch (run
->exit_reason
) {
2942 DPRINTF("handle_io\n");
2943 /* Called outside BQL */
2944 kvm_handle_io(run
->io
.port
, attrs
,
2945 (uint8_t *)run
+ run
->io
.data_offset
,
2952 DPRINTF("handle_mmio\n");
2953 /* Called outside BQL */
2954 address_space_rw(&address_space_memory
,
2955 run
->mmio
.phys_addr
, attrs
,
2958 run
->mmio
.is_write
);
2961 case KVM_EXIT_IRQ_WINDOW_OPEN
:
2962 DPRINTF("irq_window_open\n");
2963 ret
= EXCP_INTERRUPT
;
2965 case KVM_EXIT_SHUTDOWN
:
2966 DPRINTF("shutdown\n");
2967 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
2968 ret
= EXCP_INTERRUPT
;
2970 case KVM_EXIT_UNKNOWN
:
2971 fprintf(stderr
, "KVM: unknown exit, hardware reason %" PRIx64
"\n",
2972 (uint64_t)run
->hw
.hardware_exit_reason
);
2975 case KVM_EXIT_INTERNAL_ERROR
:
2976 ret
= kvm_handle_internal_error(cpu
, run
);
2978 case KVM_EXIT_DIRTY_RING_FULL
:
2980 * We shouldn't continue if the dirty ring of this vcpu is
2981 * still full. Got kicked by KVM_RESET_DIRTY_RINGS.
2983 trace_kvm_dirty_ring_full(cpu
->cpu_index
);
2984 qemu_mutex_lock_iothread();
2986 * We throttle vCPU by making it sleep once it exit from kernel
2987 * due to dirty ring full. In the dirtylimit scenario, reaping
2988 * all vCPUs after a single vCPU dirty ring get full result in
2989 * the miss of sleep, so just reap the ring-fulled vCPU.
2991 if (dirtylimit_in_service()) {
2992 kvm_dirty_ring_reap(kvm_state
, cpu
);
2994 kvm_dirty_ring_reap(kvm_state
, NULL
);
2996 qemu_mutex_unlock_iothread();
2997 dirtylimit_vcpu_execute(cpu
);
3000 case KVM_EXIT_SYSTEM_EVENT
:
3001 switch (run
->system_event
.type
) {
3002 case KVM_SYSTEM_EVENT_SHUTDOWN
:
3003 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN
);
3004 ret
= EXCP_INTERRUPT
;
3006 case KVM_SYSTEM_EVENT_RESET
:
3007 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET
);
3008 ret
= EXCP_INTERRUPT
;
3010 case KVM_SYSTEM_EVENT_CRASH
:
3011 kvm_cpu_synchronize_state(cpu
);
3012 qemu_mutex_lock_iothread();
3013 qemu_system_guest_panicked(cpu_get_crash_info(cpu
));
3014 qemu_mutex_unlock_iothread();
3018 DPRINTF("kvm_arch_handle_exit\n");
3019 ret
= kvm_arch_handle_exit(cpu
, run
);
3024 DPRINTF("kvm_arch_handle_exit\n");
3025 ret
= kvm_arch_handle_exit(cpu
, run
);
3031 qemu_mutex_lock_iothread();
3034 cpu_dump_state(cpu
, stderr
, CPU_DUMP_CODE
);
3035 vm_stop(RUN_STATE_INTERNAL_ERROR
);
3038 qatomic_set(&cpu
->exit_request
, 0);
3042 int kvm_ioctl(KVMState
*s
, int type
, ...)
3049 arg
= va_arg(ap
, void *);
3052 trace_kvm_ioctl(type
, arg
);
3053 ret
= ioctl(s
->fd
, type
, arg
);
3060 int kvm_vm_ioctl(KVMState
*s
, int type
, ...)
3067 arg
= va_arg(ap
, void *);
3070 trace_kvm_vm_ioctl(type
, arg
);
3071 ret
= ioctl(s
->vmfd
, type
, arg
);
3078 int kvm_vcpu_ioctl(CPUState
*cpu
, int type
, ...)
3085 arg
= va_arg(ap
, void *);
3088 trace_kvm_vcpu_ioctl(cpu
->cpu_index
, type
, arg
);
3089 ret
= ioctl(cpu
->kvm_fd
, type
, arg
);
3096 int kvm_device_ioctl(int fd
, int type
, ...)
3103 arg
= va_arg(ap
, void *);
3106 trace_kvm_device_ioctl(fd
, type
, arg
);
3107 ret
= ioctl(fd
, type
, arg
);
3114 int kvm_vm_check_attr(KVMState
*s
, uint32_t group
, uint64_t attr
)
3117 struct kvm_device_attr attribute
= {
3122 if (!kvm_vm_attributes_allowed
) {
3126 ret
= kvm_vm_ioctl(s
, KVM_HAS_DEVICE_ATTR
, &attribute
);
3127 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
3131 int kvm_device_check_attr(int dev_fd
, uint32_t group
, uint64_t attr
)
3133 struct kvm_device_attr attribute
= {
3139 return kvm_device_ioctl(dev_fd
, KVM_HAS_DEVICE_ATTR
, &attribute
) ? 0 : 1;
3142 int kvm_device_access(int fd
, int group
, uint64_t attr
,
3143 void *val
, bool write
, Error
**errp
)
3145 struct kvm_device_attr kvmattr
;
3149 kvmattr
.group
= group
;
3150 kvmattr
.attr
= attr
;
3151 kvmattr
.addr
= (uintptr_t)val
;
3153 err
= kvm_device_ioctl(fd
,
3154 write
? KVM_SET_DEVICE_ATTR
: KVM_GET_DEVICE_ATTR
,
3157 error_setg_errno(errp
, -err
,
3158 "KVM_%s_DEVICE_ATTR failed: Group %d "
3159 "attr 0x%016" PRIx64
,
3160 write
? "SET" : "GET", group
, attr
);
3165 bool kvm_has_sync_mmu(void)
3167 return kvm_state
->sync_mmu
;
3170 int kvm_has_vcpu_events(void)
3172 return kvm_state
->vcpu_events
;
3175 int kvm_has_robust_singlestep(void)
3177 return kvm_state
->robust_singlestep
;
3180 int kvm_has_debugregs(void)
3182 return kvm_state
->debugregs
;
3185 int kvm_max_nested_state_length(void)
3187 return kvm_state
->max_nested_state_len
;
3190 int kvm_has_many_ioeventfds(void)
3192 if (!kvm_enabled()) {
3195 return kvm_state
->many_ioeventfds
;
3198 int kvm_has_gsi_routing(void)
3200 #ifdef KVM_CAP_IRQ_ROUTING
3201 return kvm_check_extension(kvm_state
, KVM_CAP_IRQ_ROUTING
);
3207 int kvm_has_intx_set_mask(void)
3209 return kvm_state
->intx_set_mask
;
3212 bool kvm_arm_supports_user_irq(void)
3214 return kvm_check_extension(kvm_state
, KVM_CAP_ARM_USER_IRQ
);
3217 #ifdef KVM_CAP_SET_GUEST_DEBUG
3218 struct kvm_sw_breakpoint
*kvm_find_sw_breakpoint(CPUState
*cpu
,
3221 struct kvm_sw_breakpoint
*bp
;
3223 QTAILQ_FOREACH(bp
, &cpu
->kvm_state
->kvm_sw_breakpoints
, entry
) {
3231 int kvm_sw_breakpoints_active(CPUState
*cpu
)
3233 return !QTAILQ_EMPTY(&cpu
->kvm_state
->kvm_sw_breakpoints
);
3236 struct kvm_set_guest_debug_data
{
3237 struct kvm_guest_debug dbg
;
3241 static void kvm_invoke_set_guest_debug(CPUState
*cpu
, run_on_cpu_data data
)
3243 struct kvm_set_guest_debug_data
*dbg_data
=
3244 (struct kvm_set_guest_debug_data
*) data
.host_ptr
;
3246 dbg_data
->err
= kvm_vcpu_ioctl(cpu
, KVM_SET_GUEST_DEBUG
,
3250 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
3252 struct kvm_set_guest_debug_data data
;
3254 data
.dbg
.control
= reinject_trap
;
3256 if (cpu
->singlestep_enabled
) {
3257 data
.dbg
.control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_SINGLESTEP
;
3259 if (cpu
->singlestep_enabled
& SSTEP_NOIRQ
) {
3260 data
.dbg
.control
|= KVM_GUESTDBG_BLOCKIRQ
;
3263 kvm_arch_update_guest_debug(cpu
, &data
.dbg
);
3265 run_on_cpu(cpu
, kvm_invoke_set_guest_debug
,
3266 RUN_ON_CPU_HOST_PTR(&data
));
3270 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
3271 target_ulong len
, int type
)
3273 struct kvm_sw_breakpoint
*bp
;
3276 if (type
== GDB_BREAKPOINT_SW
) {
3277 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
3283 bp
= g_new(struct kvm_sw_breakpoint
, 1);
3286 err
= kvm_arch_insert_sw_breakpoint(cpu
, bp
);
3292 QTAILQ_INSERT_HEAD(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
3294 err
= kvm_arch_insert_hw_breakpoint(addr
, len
, type
);
3301 err
= kvm_update_guest_debug(cpu
, 0);
3309 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
3310 target_ulong len
, int type
)
3312 struct kvm_sw_breakpoint
*bp
;
3315 if (type
== GDB_BREAKPOINT_SW
) {
3316 bp
= kvm_find_sw_breakpoint(cpu
, addr
);
3321 if (bp
->use_count
> 1) {
3326 err
= kvm_arch_remove_sw_breakpoint(cpu
, bp
);
3331 QTAILQ_REMOVE(&cpu
->kvm_state
->kvm_sw_breakpoints
, bp
, entry
);
3334 err
= kvm_arch_remove_hw_breakpoint(addr
, len
, type
);
3341 err
= kvm_update_guest_debug(cpu
, 0);
3349 void kvm_remove_all_breakpoints(CPUState
*cpu
)
3351 struct kvm_sw_breakpoint
*bp
, *next
;
3352 KVMState
*s
= cpu
->kvm_state
;
3355 QTAILQ_FOREACH_SAFE(bp
, &s
->kvm_sw_breakpoints
, entry
, next
) {
3356 if (kvm_arch_remove_sw_breakpoint(cpu
, bp
) != 0) {
3357 /* Try harder to find a CPU that currently sees the breakpoint. */
3358 CPU_FOREACH(tmpcpu
) {
3359 if (kvm_arch_remove_sw_breakpoint(tmpcpu
, bp
) == 0) {
3364 QTAILQ_REMOVE(&s
->kvm_sw_breakpoints
, bp
, entry
);
3367 kvm_arch_remove_all_hw_breakpoints();
3370 kvm_update_guest_debug(cpu
, 0);
3374 #else /* !KVM_CAP_SET_GUEST_DEBUG */
3376 int kvm_update_guest_debug(CPUState
*cpu
, unsigned long reinject_trap
)
3381 int kvm_insert_breakpoint(CPUState
*cpu
, target_ulong addr
,
3382 target_ulong len
, int type
)
3387 int kvm_remove_breakpoint(CPUState
*cpu
, target_ulong addr
,
3388 target_ulong len
, int type
)
3393 void kvm_remove_all_breakpoints(CPUState
*cpu
)
3396 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
3398 static int kvm_set_signal_mask(CPUState
*cpu
, const sigset_t
*sigset
)
3400 KVMState
*s
= kvm_state
;
3401 struct kvm_signal_mask
*sigmask
;
3404 sigmask
= g_malloc(sizeof(*sigmask
) + sizeof(*sigset
));
3406 sigmask
->len
= s
->sigmask_len
;
3407 memcpy(sigmask
->sigset
, sigset
, sizeof(*sigset
));
3408 r
= kvm_vcpu_ioctl(cpu
, KVM_SET_SIGNAL_MASK
, sigmask
);
3414 static void kvm_ipi_signal(int sig
)
3417 assert(kvm_immediate_exit
);
3418 kvm_cpu_kick(current_cpu
);
3422 void kvm_init_cpu_signals(CPUState
*cpu
)
3426 struct sigaction sigact
;
3428 memset(&sigact
, 0, sizeof(sigact
));
3429 sigact
.sa_handler
= kvm_ipi_signal
;
3430 sigaction(SIG_IPI
, &sigact
, NULL
);
3432 pthread_sigmask(SIG_BLOCK
, NULL
, &set
);
3433 #if defined KVM_HAVE_MCE_INJECTION
3434 sigdelset(&set
, SIGBUS
);
3435 pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
3437 sigdelset(&set
, SIG_IPI
);
3438 if (kvm_immediate_exit
) {
3439 r
= pthread_sigmask(SIG_SETMASK
, &set
, NULL
);
3441 r
= kvm_set_signal_mask(cpu
, &set
);
3444 fprintf(stderr
, "kvm_set_signal_mask: %s\n", strerror(-r
));
3449 /* Called asynchronously in VCPU thread. */
3450 int kvm_on_sigbus_vcpu(CPUState
*cpu
, int code
, void *addr
)
3452 #ifdef KVM_HAVE_MCE_INJECTION
3453 if (have_sigbus_pending
) {
3456 have_sigbus_pending
= true;
3457 pending_sigbus_addr
= addr
;
3458 pending_sigbus_code
= code
;
3459 qatomic_set(&cpu
->exit_request
, 1);
3466 /* Called synchronously (via signalfd) in main thread. */
3467 int kvm_on_sigbus(int code
, void *addr
)
3469 #ifdef KVM_HAVE_MCE_INJECTION
3470 /* Action required MCE kills the process if SIGBUS is blocked. Because
3471 * that's what happens in the I/O thread, where we handle MCE via signalfd,
3472 * we can only get action optional here.
3474 assert(code
!= BUS_MCEERR_AR
);
3475 kvm_arch_on_sigbus_vcpu(first_cpu
, code
, addr
);
3482 int kvm_create_device(KVMState
*s
, uint64_t type
, bool test
)
3485 struct kvm_create_device create_dev
;
3487 create_dev
.type
= type
;
3489 create_dev
.flags
= test
? KVM_CREATE_DEVICE_TEST
: 0;
3491 if (!kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
)) {
3495 ret
= kvm_vm_ioctl(s
, KVM_CREATE_DEVICE
, &create_dev
);
3500 return test
? 0 : create_dev
.fd
;
3503 bool kvm_device_supported(int vmfd
, uint64_t type
)
3505 struct kvm_create_device create_dev
= {
3508 .flags
= KVM_CREATE_DEVICE_TEST
,
3511 if (ioctl(vmfd
, KVM_CHECK_EXTENSION
, KVM_CAP_DEVICE_CTRL
) <= 0) {
3515 return (ioctl(vmfd
, KVM_CREATE_DEVICE
, &create_dev
) >= 0);
3518 int kvm_set_one_reg(CPUState
*cs
, uint64_t id
, void *source
)
3520 struct kvm_one_reg reg
;
3524 reg
.addr
= (uintptr_t) source
;
3525 r
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, ®
);
3527 trace_kvm_failed_reg_set(id
, strerror(-r
));
3532 int kvm_get_one_reg(CPUState
*cs
, uint64_t id
, void *target
)
3534 struct kvm_one_reg reg
;
3538 reg
.addr
= (uintptr_t) target
;
3539 r
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, ®
);
3541 trace_kvm_failed_reg_get(id
, strerror(-r
));
3546 static bool kvm_accel_has_memory(MachineState
*ms
, AddressSpace
*as
,
3547 hwaddr start_addr
, hwaddr size
)
3549 KVMState
*kvm
= KVM_STATE(ms
->accelerator
);
3552 for (i
= 0; i
< kvm
->nr_as
; ++i
) {
3553 if (kvm
->as
[i
].as
== as
&& kvm
->as
[i
].ml
) {
3554 size
= MIN(kvm_max_slot_size
, size
);
3555 return NULL
!= kvm_lookup_matching_slot(kvm
->as
[i
].ml
,
3563 static void kvm_get_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3564 const char *name
, void *opaque
,
3567 KVMState
*s
= KVM_STATE(obj
);
3568 int64_t value
= s
->kvm_shadow_mem
;
3570 visit_type_int(v
, name
, &value
, errp
);
3573 static void kvm_set_kvm_shadow_mem(Object
*obj
, Visitor
*v
,
3574 const char *name
, void *opaque
,
3577 KVMState
*s
= KVM_STATE(obj
);
3581 error_setg(errp
, "Cannot set properties after the accelerator has been initialized");
3585 if (!visit_type_int(v
, name
, &value
, errp
)) {
3589 s
->kvm_shadow_mem
= value
;
3592 static void kvm_set_kernel_irqchip(Object
*obj
, Visitor
*v
,
3593 const char *name
, void *opaque
,
3596 KVMState
*s
= KVM_STATE(obj
);
3600 error_setg(errp
, "Cannot set properties after the accelerator has been initialized");
3604 if (!visit_type_OnOffSplit(v
, name
, &mode
, errp
)) {
3608 case ON_OFF_SPLIT_ON
:
3609 s
->kernel_irqchip_allowed
= true;
3610 s
->kernel_irqchip_required
= true;
3611 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3613 case ON_OFF_SPLIT_OFF
:
3614 s
->kernel_irqchip_allowed
= false;
3615 s
->kernel_irqchip_required
= false;
3616 s
->kernel_irqchip_split
= ON_OFF_AUTO_OFF
;
3618 case ON_OFF_SPLIT_SPLIT
:
3619 s
->kernel_irqchip_allowed
= true;
3620 s
->kernel_irqchip_required
= true;
3621 s
->kernel_irqchip_split
= ON_OFF_AUTO_ON
;
3624 /* The value was checked in visit_type_OnOffSplit() above. If
3625 * we get here, then something is wrong in QEMU.
3631 bool kvm_kernel_irqchip_allowed(void)
3633 return kvm_state
->kernel_irqchip_allowed
;
3636 bool kvm_kernel_irqchip_required(void)
3638 return kvm_state
->kernel_irqchip_required
;
3641 bool kvm_kernel_irqchip_split(void)
3643 return kvm_state
->kernel_irqchip_split
== ON_OFF_AUTO_ON
;
3646 static void kvm_get_dirty_ring_size(Object
*obj
, Visitor
*v
,
3647 const char *name
, void *opaque
,
3650 KVMState
*s
= KVM_STATE(obj
);
3651 uint32_t value
= s
->kvm_dirty_ring_size
;
3653 visit_type_uint32(v
, name
, &value
, errp
);
3656 static void kvm_set_dirty_ring_size(Object
*obj
, Visitor
*v
,
3657 const char *name
, void *opaque
,
3660 KVMState
*s
= KVM_STATE(obj
);
3661 Error
*error
= NULL
;
3665 error_setg(errp
, "Cannot set properties after the accelerator has been initialized");
3669 visit_type_uint32(v
, name
, &value
, &error
);
3671 error_propagate(errp
, error
);
3674 if (value
& (value
- 1)) {
3675 error_setg(errp
, "dirty-ring-size must be a power of two.");
3679 s
->kvm_dirty_ring_size
= value
;
3682 static void kvm_accel_instance_init(Object
*obj
)
3684 KVMState
*s
= KVM_STATE(obj
);
3688 s
->kvm_shadow_mem
= -1;
3689 s
->kernel_irqchip_allowed
= true;
3690 s
->kernel_irqchip_split
= ON_OFF_AUTO_AUTO
;
3691 /* KVM dirty ring is by default off */
3692 s
->kvm_dirty_ring_size
= 0;
3695 static void kvm_accel_class_init(ObjectClass
*oc
, void *data
)
3697 AccelClass
*ac
= ACCEL_CLASS(oc
);
3699 ac
->init_machine
= kvm_init
;
3700 ac
->has_memory
= kvm_accel_has_memory
;
3701 ac
->allowed
= &kvm_allowed
;
3703 object_class_property_add(oc
, "kernel-irqchip", "on|off|split",
3704 NULL
, kvm_set_kernel_irqchip
,
3706 object_class_property_set_description(oc
, "kernel-irqchip",
3707 "Configure KVM in-kernel irqchip");
3709 object_class_property_add(oc
, "kvm-shadow-mem", "int",
3710 kvm_get_kvm_shadow_mem
, kvm_set_kvm_shadow_mem
,
3712 object_class_property_set_description(oc
, "kvm-shadow-mem",
3713 "KVM shadow MMU size");
3715 object_class_property_add(oc
, "dirty-ring-size", "uint32",
3716 kvm_get_dirty_ring_size
, kvm_set_dirty_ring_size
,
3718 object_class_property_set_description(oc
, "dirty-ring-size",
3719 "Size of KVM dirty page ring buffer (default: 0, i.e. use bitmap)");
3722 static const TypeInfo kvm_accel_type
= {
3723 .name
= TYPE_KVM_ACCEL
,
3724 .parent
= TYPE_ACCEL
,
3725 .instance_init
= kvm_accel_instance_init
,
3726 .class_init
= kvm_accel_class_init
,
3727 .instance_size
= sizeof(KVMState
),
3730 static void kvm_type_init(void)
3732 type_register_static(&kvm_accel_type
);
3735 type_init(kvm_type_init
);
3737 typedef struct StatsArgs
{
3738 union StatsResultsType
{
3739 StatsResultList
**stats
;
3740 StatsSchemaList
**schema
;
3746 static StatsList
*add_kvmstat_entry(struct kvm_stats_desc
*pdesc
,
3747 uint64_t *stats_data
,
3748 StatsList
*stats_list
,
3753 uint64List
*val_list
= NULL
;
3755 /* Only add stats that we understand. */
3756 switch (pdesc
->flags
& KVM_STATS_TYPE_MASK
) {
3757 case KVM_STATS_TYPE_CUMULATIVE
:
3758 case KVM_STATS_TYPE_INSTANT
:
3759 case KVM_STATS_TYPE_PEAK
:
3760 case KVM_STATS_TYPE_LINEAR_HIST
:
3761 case KVM_STATS_TYPE_LOG_HIST
:
3767 switch (pdesc
->flags
& KVM_STATS_UNIT_MASK
) {
3768 case KVM_STATS_UNIT_NONE
:
3769 case KVM_STATS_UNIT_BYTES
:
3770 case KVM_STATS_UNIT_CYCLES
:
3771 case KVM_STATS_UNIT_SECONDS
:
3772 case KVM_STATS_UNIT_BOOLEAN
:
3778 switch (pdesc
->flags
& KVM_STATS_BASE_MASK
) {
3779 case KVM_STATS_BASE_POW10
:
3780 case KVM_STATS_BASE_POW2
:
3786 /* Alloc and populate data list */
3787 stats
= g_new0(Stats
, 1);
3788 stats
->name
= g_strdup(pdesc
->name
);
3789 stats
->value
= g_new0(StatsValue
, 1);;
3791 if ((pdesc
->flags
& KVM_STATS_UNIT_MASK
) == KVM_STATS_UNIT_BOOLEAN
) {
3792 stats
->value
->u
.boolean
= *stats_data
;
3793 stats
->value
->type
= QTYPE_QBOOL
;
3794 } else if (pdesc
->size
== 1) {
3795 stats
->value
->u
.scalar
= *stats_data
;
3796 stats
->value
->type
= QTYPE_QNUM
;
3799 for (i
= 0; i
< pdesc
->size
; i
++) {
3800 QAPI_LIST_PREPEND(val_list
, stats_data
[i
]);
3802 stats
->value
->u
.list
= val_list
;
3803 stats
->value
->type
= QTYPE_QLIST
;
3806 QAPI_LIST_PREPEND(stats_list
, stats
);
3810 static StatsSchemaValueList
*add_kvmschema_entry(struct kvm_stats_desc
*pdesc
,
3811 StatsSchemaValueList
*list
,
3814 StatsSchemaValueList
*schema_entry
= g_new0(StatsSchemaValueList
, 1);
3815 schema_entry
->value
= g_new0(StatsSchemaValue
, 1);
3817 switch (pdesc
->flags
& KVM_STATS_TYPE_MASK
) {
3818 case KVM_STATS_TYPE_CUMULATIVE
:
3819 schema_entry
->value
->type
= STATS_TYPE_CUMULATIVE
;
3821 case KVM_STATS_TYPE_INSTANT
:
3822 schema_entry
->value
->type
= STATS_TYPE_INSTANT
;
3824 case KVM_STATS_TYPE_PEAK
:
3825 schema_entry
->value
->type
= STATS_TYPE_PEAK
;
3827 case KVM_STATS_TYPE_LINEAR_HIST
:
3828 schema_entry
->value
->type
= STATS_TYPE_LINEAR_HISTOGRAM
;
3829 schema_entry
->value
->bucket_size
= pdesc
->bucket_size
;
3830 schema_entry
->value
->has_bucket_size
= true;
3832 case KVM_STATS_TYPE_LOG_HIST
:
3833 schema_entry
->value
->type
= STATS_TYPE_LOG2_HISTOGRAM
;
3839 switch (pdesc
->flags
& KVM_STATS_UNIT_MASK
) {
3840 case KVM_STATS_UNIT_NONE
:
3842 case KVM_STATS_UNIT_BOOLEAN
:
3843 schema_entry
->value
->has_unit
= true;
3844 schema_entry
->value
->unit
= STATS_UNIT_BOOLEAN
;
3846 case KVM_STATS_UNIT_BYTES
:
3847 schema_entry
->value
->has_unit
= true;
3848 schema_entry
->value
->unit
= STATS_UNIT_BYTES
;
3850 case KVM_STATS_UNIT_CYCLES
:
3851 schema_entry
->value
->has_unit
= true;
3852 schema_entry
->value
->unit
= STATS_UNIT_CYCLES
;
3854 case KVM_STATS_UNIT_SECONDS
:
3855 schema_entry
->value
->has_unit
= true;
3856 schema_entry
->value
->unit
= STATS_UNIT_SECONDS
;
3862 schema_entry
->value
->exponent
= pdesc
->exponent
;
3863 if (pdesc
->exponent
) {
3864 switch (pdesc
->flags
& KVM_STATS_BASE_MASK
) {
3865 case KVM_STATS_BASE_POW10
:
3866 schema_entry
->value
->has_base
= true;
3867 schema_entry
->value
->base
= 10;
3869 case KVM_STATS_BASE_POW2
:
3870 schema_entry
->value
->has_base
= true;
3871 schema_entry
->value
->base
= 2;
3878 schema_entry
->value
->name
= g_strdup(pdesc
->name
);
3879 schema_entry
->next
= list
;
3880 return schema_entry
;
3882 g_free(schema_entry
->value
);
3883 g_free(schema_entry
);
3887 /* Cached stats descriptors */
3888 typedef struct StatsDescriptors
{
3889 const char *ident
; /* cache key, currently the StatsTarget */
3890 struct kvm_stats_desc
*kvm_stats_desc
;
3891 struct kvm_stats_header
*kvm_stats_header
;
3892 QTAILQ_ENTRY(StatsDescriptors
) next
;
3895 static QTAILQ_HEAD(, StatsDescriptors
) stats_descriptors
=
3896 QTAILQ_HEAD_INITIALIZER(stats_descriptors
);
3899 * Return the descriptors for 'target', that either have already been read
3900 * or are retrieved from 'stats_fd'.
3902 static StatsDescriptors
*find_stats_descriptors(StatsTarget target
, int stats_fd
,
3905 StatsDescriptors
*descriptors
;
3907 struct kvm_stats_desc
*kvm_stats_desc
;
3908 struct kvm_stats_header
*kvm_stats_header
;
3912 ident
= StatsTarget_str(target
);
3913 QTAILQ_FOREACH(descriptors
, &stats_descriptors
, next
) {
3914 if (g_str_equal(descriptors
->ident
, ident
)) {
3919 descriptors
= g_new0(StatsDescriptors
, 1);
3921 /* Read stats header */
3922 kvm_stats_header
= g_malloc(sizeof(*kvm_stats_header
));
3923 ret
= read(stats_fd
, kvm_stats_header
, sizeof(*kvm_stats_header
));
3924 if (ret
!= sizeof(*kvm_stats_header
)) {
3925 error_setg(errp
, "KVM stats: failed to read stats header: "
3926 "expected %zu actual %zu",
3927 sizeof(*kvm_stats_header
), ret
);
3928 g_free(descriptors
);
3931 size_desc
= sizeof(*kvm_stats_desc
) + kvm_stats_header
->name_size
;
3933 /* Read stats descriptors */
3934 kvm_stats_desc
= g_malloc0_n(kvm_stats_header
->num_desc
, size_desc
);
3935 ret
= pread(stats_fd
, kvm_stats_desc
,
3936 size_desc
* kvm_stats_header
->num_desc
,
3937 kvm_stats_header
->desc_offset
);
3939 if (ret
!= size_desc
* kvm_stats_header
->num_desc
) {
3940 error_setg(errp
, "KVM stats: failed to read stats descriptors: "
3941 "expected %zu actual %zu",
3942 size_desc
* kvm_stats_header
->num_desc
, ret
);
3943 g_free(descriptors
);
3944 g_free(kvm_stats_desc
);
3947 descriptors
->kvm_stats_header
= kvm_stats_header
;
3948 descriptors
->kvm_stats_desc
= kvm_stats_desc
;
3949 descriptors
->ident
= ident
;
3950 QTAILQ_INSERT_TAIL(&stats_descriptors
, descriptors
, next
);
3954 static void query_stats(StatsResultList
**result
, StatsTarget target
,
3955 strList
*names
, int stats_fd
, Error
**errp
)
3957 struct kvm_stats_desc
*kvm_stats_desc
;
3958 struct kvm_stats_header
*kvm_stats_header
;
3959 StatsDescriptors
*descriptors
;
3960 g_autofree
uint64_t *stats_data
= NULL
;
3961 struct kvm_stats_desc
*pdesc
;
3962 StatsList
*stats_list
= NULL
;
3963 size_t size_desc
, size_data
= 0;
3967 descriptors
= find_stats_descriptors(target
, stats_fd
, errp
);
3972 kvm_stats_header
= descriptors
->kvm_stats_header
;
3973 kvm_stats_desc
= descriptors
->kvm_stats_desc
;
3974 size_desc
= sizeof(*kvm_stats_desc
) + kvm_stats_header
->name_size
;
3976 /* Tally the total data size; read schema data */
3977 for (i
= 0; i
< kvm_stats_header
->num_desc
; ++i
) {
3978 pdesc
= (void *)kvm_stats_desc
+ i
* size_desc
;
3979 size_data
+= pdesc
->size
* sizeof(*stats_data
);
3982 stats_data
= g_malloc0(size_data
);
3983 ret
= pread(stats_fd
, stats_data
, size_data
, kvm_stats_header
->data_offset
);
3985 if (ret
!= size_data
) {
3986 error_setg(errp
, "KVM stats: failed to read data: "
3987 "expected %zu actual %zu", size_data
, ret
);
3991 for (i
= 0; i
< kvm_stats_header
->num_desc
; ++i
) {
3993 pdesc
= (void *)kvm_stats_desc
+ i
* size_desc
;
3995 /* Add entry to the list */
3996 stats
= (void *)stats_data
+ pdesc
->offset
;
3997 if (!apply_str_list_filter(pdesc
->name
, names
)) {
4000 stats_list
= add_kvmstat_entry(pdesc
, stats
, stats_list
, errp
);
4008 case STATS_TARGET_VM
:
4009 add_stats_entry(result
, STATS_PROVIDER_KVM
, NULL
, stats_list
);
4011 case STATS_TARGET_VCPU
:
4012 add_stats_entry(result
, STATS_PROVIDER_KVM
,
4013 current_cpu
->parent_obj
.canonical_path
,
4017 g_assert_not_reached();
4021 static void query_stats_schema(StatsSchemaList
**result
, StatsTarget target
,
4022 int stats_fd
, Error
**errp
)
4024 struct kvm_stats_desc
*kvm_stats_desc
;
4025 struct kvm_stats_header
*kvm_stats_header
;
4026 StatsDescriptors
*descriptors
;
4027 struct kvm_stats_desc
*pdesc
;
4028 StatsSchemaValueList
*stats_list
= NULL
;
4032 descriptors
= find_stats_descriptors(target
, stats_fd
, errp
);
4037 kvm_stats_header
= descriptors
->kvm_stats_header
;
4038 kvm_stats_desc
= descriptors
->kvm_stats_desc
;
4039 size_desc
= sizeof(*kvm_stats_desc
) + kvm_stats_header
->name_size
;
4041 /* Tally the total data size; read schema data */
4042 for (i
= 0; i
< kvm_stats_header
->num_desc
; ++i
) {
4043 pdesc
= (void *)kvm_stats_desc
+ i
* size_desc
;
4044 stats_list
= add_kvmschema_entry(pdesc
, stats_list
, errp
);
4047 add_stats_schema(result
, STATS_PROVIDER_KVM
, target
, stats_list
);
4050 static void query_stats_vcpu(CPUState
*cpu
, run_on_cpu_data data
)
4052 StatsArgs
*kvm_stats_args
= (StatsArgs
*) data
.host_ptr
;
4053 int stats_fd
= kvm_vcpu_ioctl(cpu
, KVM_GET_STATS_FD
, NULL
);
4054 Error
*local_err
= NULL
;
4056 if (stats_fd
== -1) {
4057 error_setg_errno(&local_err
, errno
, "KVM stats: ioctl failed");
4058 error_propagate(kvm_stats_args
->errp
, local_err
);
4061 query_stats(kvm_stats_args
->result
.stats
, STATS_TARGET_VCPU
,
4062 kvm_stats_args
->names
, stats_fd
, kvm_stats_args
->errp
);
4066 static void query_stats_schema_vcpu(CPUState
*cpu
, run_on_cpu_data data
)
4068 StatsArgs
*kvm_stats_args
= (StatsArgs
*) data
.host_ptr
;
4069 int stats_fd
= kvm_vcpu_ioctl(cpu
, KVM_GET_STATS_FD
, NULL
);
4070 Error
*local_err
= NULL
;
4072 if (stats_fd
== -1) {
4073 error_setg_errno(&local_err
, errno
, "KVM stats: ioctl failed");
4074 error_propagate(kvm_stats_args
->errp
, local_err
);
4077 query_stats_schema(kvm_stats_args
->result
.schema
, STATS_TARGET_VCPU
, stats_fd
,
4078 kvm_stats_args
->errp
);
4082 static void query_stats_cb(StatsResultList
**result
, StatsTarget target
,
4083 strList
*names
, strList
*targets
, Error
**errp
)
4085 KVMState
*s
= kvm_state
;
4090 case STATS_TARGET_VM
:
4092 stats_fd
= kvm_vm_ioctl(s
, KVM_GET_STATS_FD
, NULL
);
4093 if (stats_fd
== -1) {
4094 error_setg_errno(errp
, errno
, "KVM stats: ioctl failed");
4097 query_stats(result
, target
, names
, stats_fd
, errp
);
4101 case STATS_TARGET_VCPU
:
4103 StatsArgs stats_args
;
4104 stats_args
.result
.stats
= result
;
4105 stats_args
.names
= names
;
4106 stats_args
.errp
= errp
;
4108 if (!apply_str_list_filter(cpu
->parent_obj
.canonical_path
, targets
)) {
4111 run_on_cpu(cpu
, query_stats_vcpu
, RUN_ON_CPU_HOST_PTR(&stats_args
));
4120 void query_stats_schemas_cb(StatsSchemaList
**result
, Error
**errp
)
4122 StatsArgs stats_args
;
4123 KVMState
*s
= kvm_state
;
4126 stats_fd
= kvm_vm_ioctl(s
, KVM_GET_STATS_FD
, NULL
);
4127 if (stats_fd
== -1) {
4128 error_setg_errno(errp
, errno
, "KVM stats: ioctl failed");
4131 query_stats_schema(result
, STATS_TARGET_VM
, stats_fd
, errp
);
4134 stats_args
.result
.schema
= result
;
4135 stats_args
.errp
= errp
;
4136 run_on_cpu(first_cpu
, query_stats_schema_vcpu
, RUN_ON_CPU_HOST_PTR(&stats_args
));