4 * Copyright IBM, Corp. 2008
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
16 #include "qemu/osdep.h"
17 #include <sys/ioctl.h>
19 #include <linux/kvm.h>
21 #include "qemu-common.h"
22 #include "qemu/atomic.h"
23 #include "qemu/option.h"
24 #include "qemu/config-file.h"
25 #include "qemu/error-report.h"
26 #include "qapi/error.h"
28 #include "hw/pci/msi.h"
29 #include "hw/pci/msix.h"
30 #include "hw/s390x/adapter.h"
31 #include "exec/gdbstub.h"
32 #include "sysemu/kvm_int.h"
33 #include "sysemu/cpus.h"
34 #include "qemu/bswap.h"
35 #include "exec/memory.h"
36 #include "exec/ram_addr.h"
37 #include "exec/address-spaces.h"
38 #include "qemu/event_notifier.h"
41 #include "sysemu/sev.h"
42 #include "sysemu/balloon.h"
44 #include "hw/boards.h"
46 /* This check must be after config-host.h is included */
48 #include <sys/eventfd.h>
51 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
52 * need to use the real host PAGE_SIZE, as that's what KVM will use.
54 #define PAGE_SIZE getpagesize()
59 #define DPRINTF(fmt, ...) \
60 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
62 #define DPRINTF(fmt, ...) \
66 #define KVM_MSI_HASHTAB_SIZE 256
68 struct KVMParkedVcpu {
69 unsigned long vcpu_id;
71 QLIST_ENTRY(KVMParkedVcpu) node;
76 AccelState parent_obj;
83 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
84 bool coalesced_flush_in_progress;
86 int robust_singlestep;
88 #ifdef KVM_CAP_SET_GUEST_DEBUG
89 QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints;
94 /* The man page (and posix) say ioctl numbers are signed int, but
95 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
96 * unsigned, and treating them as signed here can break things */
97 unsigned irq_set_ioctl;
98 unsigned int sigmask_len;
100 #ifdef KVM_CAP_IRQ_ROUTING
101 struct kvm_irq_routing *irq_routes;
102 int nr_allocated_irq_routes;
103 unsigned long *used_gsi_bitmap;
104 unsigned int gsi_count;
105 QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
107 KVMMemoryListener memory_listener;
108 QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
110 /* memory encryption */
111 void *memcrypt_handle;
112 int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len);
116 bool kvm_kernel_irqchip;
117 bool kvm_split_irqchip;
118 bool kvm_async_interrupts_allowed;
119 bool kvm_halt_in_kernel_allowed;
120 bool kvm_eventfds_allowed;
121 bool kvm_irqfds_allowed;
122 bool kvm_resamplefds_allowed;
123 bool kvm_msi_via_irqfd_allowed;
124 bool kvm_gsi_routing_allowed;
125 bool kvm_gsi_direct_mapping;
127 bool kvm_readonly_mem_allowed;
128 bool kvm_vm_attributes_allowed;
129 bool kvm_direct_msi_allowed;
130 bool kvm_ioeventfd_any_length_allowed;
131 bool kvm_msi_use_devid;
132 static bool kvm_immediate_exit;
134 static const KVMCapabilityInfo kvm_required_capabilites[] = {
135 KVM_CAP_INFO(USER_MEMORY),
136 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
137 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
141 int kvm_get_max_memslots(void)
143 KVMState *s = KVM_STATE(current_machine->accelerator);
148 bool kvm_memcrypt_enabled(void)
150 if (kvm_state && kvm_state->memcrypt_handle) {
157 int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len)
159 if (kvm_state->memcrypt_handle &&
160 kvm_state->memcrypt_encrypt_data) {
161 return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle,
168 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
170 KVMState *s = kvm_state;
173 for (i = 0; i < s->nr_slots; i++) {
174 if (kml->slots[i].memory_size == 0) {
175 return &kml->slots[i];
182 bool kvm_has_free_slot(MachineState *ms)
184 KVMState *s = KVM_STATE(ms->accelerator);
186 return kvm_get_free_slot(&s->memory_listener);
189 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
191 KVMSlot *slot = kvm_get_free_slot(kml);
197 fprintf(stderr, "%s: no free slot available\n", __func__);
201 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
205 KVMState *s = kvm_state;
208 for (i = 0; i < s->nr_slots; i++) {
209 KVMSlot *mem = &kml->slots[i];
211 if (start_addr == mem->start_addr && size == mem->memory_size) {
220 * Calculate and align the start address and the size of the section.
221 * Return the size. If the size is 0, the aligned section is empty.
223 static hwaddr kvm_align_section(MemoryRegionSection *section,
226 hwaddr size = int128_get64(section->size);
227 hwaddr delta, aligned;
229 /* kvm works in page size chunks, but the function may be called
230 with sub-page size and unaligned start address. Pad the start
231 address to next and truncate size to previous page boundary. */
232 aligned = ROUND_UP(section->offset_within_address_space,
233 qemu_real_host_page_size);
234 delta = aligned - section->offset_within_address_space;
240 return (size - delta) & qemu_real_host_page_mask;
243 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
246 KVMMemoryListener *kml = &s->memory_listener;
249 for (i = 0; i < s->nr_slots; i++) {
250 KVMSlot *mem = &kml->slots[i];
252 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
253 *phys_addr = mem->start_addr + (ram - mem->ram);
261 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
263 KVMState *s = kvm_state;
264 struct kvm_userspace_memory_region mem;
267 mem.slot = slot->slot | (kml->as_id << 16);
268 mem.guest_phys_addr = slot->start_addr;
269 mem.userspace_addr = (unsigned long)slot->ram;
270 mem.flags = slot->flags;
272 if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
273 /* Set the slot size to 0 before setting the slot to the desired
274 * value. This is needed based on KVM commit 75d61fbc. */
276 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
278 mem.memory_size = slot->memory_size;
279 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
280 slot->old_flags = mem.flags;
281 trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
282 mem.memory_size, mem.userspace_addr, ret);
286 int kvm_destroy_vcpu(CPUState *cpu)
288 KVMState *s = kvm_state;
290 struct KVMParkedVcpu *vcpu = NULL;
293 DPRINTF("kvm_destroy_vcpu\n");
295 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
298 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
302 ret = munmap(cpu->kvm_run, mmap_size);
307 vcpu = g_malloc0(sizeof(*vcpu));
308 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
309 vcpu->kvm_fd = cpu->kvm_fd;
310 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
315 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
317 struct KVMParkedVcpu *cpu;
319 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
320 if (cpu->vcpu_id == vcpu_id) {
323 QLIST_REMOVE(cpu, node);
324 kvm_fd = cpu->kvm_fd;
330 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
333 int kvm_init_vcpu(CPUState *cpu)
335 KVMState *s = kvm_state;
339 DPRINTF("kvm_init_vcpu\n");
341 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
343 DPRINTF("kvm_create_vcpu failed\n");
349 cpu->vcpu_dirty = true;
351 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
354 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
358 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
360 if (cpu->kvm_run == MAP_FAILED) {
362 DPRINTF("mmap'ing vcpu state failed\n");
366 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
367 s->coalesced_mmio_ring =
368 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
371 ret = kvm_arch_init_vcpu(cpu);
377 * dirty pages logging control
380 static int kvm_mem_flags(MemoryRegion *mr)
382 bool readonly = mr->readonly || memory_region_is_romd(mr);
385 if (memory_region_get_dirty_log_mask(mr) != 0) {
386 flags |= KVM_MEM_LOG_DIRTY_PAGES;
388 if (readonly && kvm_readonly_mem_allowed) {
389 flags |= KVM_MEM_READONLY;
394 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
397 mem->flags = kvm_mem_flags(mr);
399 /* If nothing changed effectively, no need to issue ioctl */
400 if (mem->flags == mem->old_flags) {
404 return kvm_set_user_memory_region(kml, mem, false);
407 static int kvm_section_update_flags(KVMMemoryListener *kml,
408 MemoryRegionSection *section)
410 hwaddr start_addr, size;
413 size = kvm_align_section(section, &start_addr);
418 mem = kvm_lookup_matching_slot(kml, start_addr, size);
420 /* We don't have a slot if we want to trap every access. */
424 return kvm_slot_update_flags(kml, mem, section->mr);
427 static void kvm_log_start(MemoryListener *listener,
428 MemoryRegionSection *section,
431 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
438 r = kvm_section_update_flags(kml, section);
444 static void kvm_log_stop(MemoryListener *listener,
445 MemoryRegionSection *section,
448 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
455 r = kvm_section_update_flags(kml, section);
461 /* get kvm's dirty pages bitmap and update qemu's */
462 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
463 unsigned long *bitmap)
465 ram_addr_t start = section->offset_within_region +
466 memory_region_get_ram_addr(section->mr);
467 ram_addr_t pages = int128_get64(section->size) / getpagesize();
469 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
473 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
476 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
477 * This function updates qemu's dirty bitmap using
478 * memory_region_set_dirty(). This means all bits are set
481 * @start_add: start of logged region.
482 * @end_addr: end of logged region.
484 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
485 MemoryRegionSection *section)
487 KVMState *s = kvm_state;
488 struct kvm_dirty_log d = {};
490 hwaddr start_addr, size;
492 size = kvm_align_section(section, &start_addr);
494 mem = kvm_lookup_matching_slot(kml, start_addr, size);
496 /* We don't have a slot if we want to trap every access. */
500 /* XXX bad kernel interface alert
501 * For dirty bitmap, kernel allocates array of size aligned to
502 * bits-per-long. But for case when the kernel is 64bits and
503 * the userspace is 32bits, userspace can't align to the same
504 * bits-per-long, since sizeof(long) is different between kernel
505 * and user space. This way, userspace will provide buffer which
506 * may be 4 bytes less than the kernel will use, resulting in
507 * userspace memory corruption (which is not detectable by valgrind
508 * too, in most cases).
509 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
510 * a hope that sizeof(long) won't become >8 any time soon.
512 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
513 /*HOST_LONG_BITS*/ 64) / 8;
514 d.dirty_bitmap = g_malloc0(size);
516 d.slot = mem->slot | (kml->as_id << 16);
517 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
518 DPRINTF("ioctl failed %d\n", errno);
519 g_free(d.dirty_bitmap);
523 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
524 g_free(d.dirty_bitmap);
530 static void kvm_coalesce_mmio_region(MemoryListener *listener,
531 MemoryRegionSection *secion,
532 hwaddr start, hwaddr size)
534 KVMState *s = kvm_state;
536 if (s->coalesced_mmio) {
537 struct kvm_coalesced_mmio_zone zone;
543 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
547 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
548 MemoryRegionSection *secion,
549 hwaddr start, hwaddr size)
551 KVMState *s = kvm_state;
553 if (s->coalesced_mmio) {
554 struct kvm_coalesced_mmio_zone zone;
560 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
564 static void kvm_coalesce_pio_add(MemoryListener *listener,
565 MemoryRegionSection *section,
566 hwaddr start, hwaddr size)
568 KVMState *s = kvm_state;
570 if (s->coalesced_pio) {
571 struct kvm_coalesced_mmio_zone zone;
577 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
581 static void kvm_coalesce_pio_del(MemoryListener *listener,
582 MemoryRegionSection *section,
583 hwaddr start, hwaddr size)
585 KVMState *s = kvm_state;
587 if (s->coalesced_pio) {
588 struct kvm_coalesced_mmio_zone zone;
594 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
598 static MemoryListener kvm_coalesced_pio_listener = {
599 .coalesced_io_add = kvm_coalesce_pio_add,
600 .coalesced_io_del = kvm_coalesce_pio_del,
603 int kvm_check_extension(KVMState *s, unsigned int extension)
607 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
615 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
619 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
621 /* VM wide version not implemented, use global one instead */
622 ret = kvm_check_extension(s, extension);
628 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
630 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
631 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
632 * endianness, but the memory core hands them in target endianness.
633 * For example, PPC is always treated as big-endian even if running
634 * on KVM and on PPC64LE. Correct here.
648 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
649 bool assign, uint32_t size, bool datamatch)
652 struct kvm_ioeventfd iofd = {
653 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
660 trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size,
662 if (!kvm_enabled()) {
667 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
670 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
673 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
682 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
683 bool assign, uint32_t size, bool datamatch)
685 struct kvm_ioeventfd kick = {
686 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
688 .flags = KVM_IOEVENTFD_FLAG_PIO,
693 trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch);
694 if (!kvm_enabled()) {
698 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
701 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
703 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
711 static int kvm_check_many_ioeventfds(void)
713 /* Userspace can use ioeventfd for io notification. This requires a host
714 * that supports eventfd(2) and an I/O thread; since eventfd does not
715 * support SIGIO it cannot interrupt the vcpu.
717 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
718 * can avoid creating too many ioeventfds.
720 #if defined(CONFIG_EVENTFD)
723 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
724 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
725 if (ioeventfds[i] < 0) {
728 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
730 close(ioeventfds[i]);
735 /* Decide whether many devices are supported or not */
736 ret = i == ARRAY_SIZE(ioeventfds);
739 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
740 close(ioeventfds[i]);
748 static const KVMCapabilityInfo *
749 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
752 if (!kvm_check_extension(s, list->value)) {
760 static void kvm_set_phys_mem(KVMMemoryListener *kml,
761 MemoryRegionSection *section, bool add)
765 MemoryRegion *mr = section->mr;
766 bool writeable = !mr->readonly && !mr->rom_device;
767 hwaddr start_addr, size;
770 if (!memory_region_is_ram(mr)) {
771 if (writeable || !kvm_readonly_mem_allowed) {
773 } else if (!mr->romd_mode) {
774 /* If the memory device is not in romd_mode, then we actually want
775 * to remove the kvm memory slot so all accesses will trap. */
780 size = kvm_align_section(section, &start_addr);
785 /* use aligned delta to align the ram address */
786 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
787 (start_addr - section->offset_within_address_space);
790 mem = kvm_lookup_matching_slot(kml, start_addr, size);
794 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
795 kvm_physical_sync_dirty_bitmap(kml, section);
798 /* unregister the slot */
799 mem->memory_size = 0;
801 err = kvm_set_user_memory_region(kml, mem, false);
803 fprintf(stderr, "%s: error unregistering slot: %s\n",
804 __func__, strerror(-err));
810 /* register the new slot */
811 mem = kvm_alloc_slot(kml);
812 mem->memory_size = size;
813 mem->start_addr = start_addr;
815 mem->flags = kvm_mem_flags(mr);
817 err = kvm_set_user_memory_region(kml, mem, true);
819 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
825 static void kvm_region_add(MemoryListener *listener,
826 MemoryRegionSection *section)
828 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
830 memory_region_ref(section->mr);
831 kvm_set_phys_mem(kml, section, true);
834 static void kvm_region_del(MemoryListener *listener,
835 MemoryRegionSection *section)
837 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
839 kvm_set_phys_mem(kml, section, false);
840 memory_region_unref(section->mr);
843 static void kvm_log_sync(MemoryListener *listener,
844 MemoryRegionSection *section)
846 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
849 r = kvm_physical_sync_dirty_bitmap(kml, section);
855 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
856 MemoryRegionSection *section,
857 bool match_data, uint64_t data,
860 int fd = event_notifier_get_fd(e);
863 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
864 data, true, int128_get64(section->size),
867 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
868 __func__, strerror(-r));
873 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
874 MemoryRegionSection *section,
875 bool match_data, uint64_t data,
878 int fd = event_notifier_get_fd(e);
881 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
882 data, false, int128_get64(section->size),
889 static void kvm_io_ioeventfd_add(MemoryListener *listener,
890 MemoryRegionSection *section,
891 bool match_data, uint64_t data,
894 int fd = event_notifier_get_fd(e);
897 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
898 data, true, int128_get64(section->size),
901 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
902 __func__, strerror(-r));
907 static void kvm_io_ioeventfd_del(MemoryListener *listener,
908 MemoryRegionSection *section,
909 bool match_data, uint64_t data,
913 int fd = event_notifier_get_fd(e);
916 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
917 data, false, int128_get64(section->size),
924 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
925 AddressSpace *as, int as_id)
929 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
932 for (i = 0; i < s->nr_slots; i++) {
933 kml->slots[i].slot = i;
936 kml->listener.region_add = kvm_region_add;
937 kml->listener.region_del = kvm_region_del;
938 kml->listener.log_start = kvm_log_start;
939 kml->listener.log_stop = kvm_log_stop;
940 kml->listener.log_sync = kvm_log_sync;
941 kml->listener.priority = 10;
943 memory_listener_register(&kml->listener, as);
946 static MemoryListener kvm_io_listener = {
947 .eventfd_add = kvm_io_ioeventfd_add,
948 .eventfd_del = kvm_io_ioeventfd_del,
952 int kvm_set_irq(KVMState *s, int irq, int level)
954 struct kvm_irq_level event;
957 assert(kvm_async_interrupts_enabled());
961 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
963 perror("kvm_set_irq");
967 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
970 #ifdef KVM_CAP_IRQ_ROUTING
971 typedef struct KVMMSIRoute {
972 struct kvm_irq_routing_entry kroute;
973 QTAILQ_ENTRY(KVMMSIRoute) entry;
976 static void set_gsi(KVMState *s, unsigned int gsi)
978 set_bit(gsi, s->used_gsi_bitmap);
981 static void clear_gsi(KVMState *s, unsigned int gsi)
983 clear_bit(gsi, s->used_gsi_bitmap);
986 void kvm_init_irq_routing(KVMState *s)
990 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
992 /* Round up so we can search ints using ffs */
993 s->used_gsi_bitmap = bitmap_new(gsi_count);
994 s->gsi_count = gsi_count;
997 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
998 s->nr_allocated_irq_routes = 0;
1000 if (!kvm_direct_msi_allowed) {
1001 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1002 QTAILQ_INIT(&s->msi_hashtab[i]);
1006 kvm_arch_init_irq_routing(s);
1009 void kvm_irqchip_commit_routes(KVMState *s)
1013 if (kvm_gsi_direct_mapping()) {
1017 if (!kvm_gsi_routing_enabled()) {
1021 s->irq_routes->flags = 0;
1022 trace_kvm_irqchip_commit_routes();
1023 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1027 static void kvm_add_routing_entry(KVMState *s,
1028 struct kvm_irq_routing_entry *entry)
1030 struct kvm_irq_routing_entry *new;
1033 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1034 n = s->nr_allocated_irq_routes * 2;
1038 size = sizeof(struct kvm_irq_routing);
1039 size += n * sizeof(*new);
1040 s->irq_routes = g_realloc(s->irq_routes, size);
1041 s->nr_allocated_irq_routes = n;
1043 n = s->irq_routes->nr++;
1044 new = &s->irq_routes->entries[n];
1048 set_gsi(s, entry->gsi);
1051 static int kvm_update_routing_entry(KVMState *s,
1052 struct kvm_irq_routing_entry *new_entry)
1054 struct kvm_irq_routing_entry *entry;
1057 for (n = 0; n < s->irq_routes->nr; n++) {
1058 entry = &s->irq_routes->entries[n];
1059 if (entry->gsi != new_entry->gsi) {
1063 if(!memcmp(entry, new_entry, sizeof *entry)) {
1067 *entry = *new_entry;
1075 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1077 struct kvm_irq_routing_entry e = {};
1079 assert(pin < s->gsi_count);
1082 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1084 e.u.irqchip.irqchip = irqchip;
1085 e.u.irqchip.pin = pin;
1086 kvm_add_routing_entry(s, &e);
1089 void kvm_irqchip_release_virq(KVMState *s, int virq)
1091 struct kvm_irq_routing_entry *e;
1094 if (kvm_gsi_direct_mapping()) {
1098 for (i = 0; i < s->irq_routes->nr; i++) {
1099 e = &s->irq_routes->entries[i];
1100 if (e->gsi == virq) {
1101 s->irq_routes->nr--;
1102 *e = s->irq_routes->entries[s->irq_routes->nr];
1106 kvm_arch_release_virq_post(virq);
1107 trace_kvm_irqchip_release_virq(virq);
1110 static unsigned int kvm_hash_msi(uint32_t data)
1112 /* This is optimized for IA32 MSI layout. However, no other arch shall
1113 * repeat the mistake of not providing a direct MSI injection API. */
1117 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1119 KVMMSIRoute *route, *next;
1122 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1123 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1124 kvm_irqchip_release_virq(s, route->kroute.gsi);
1125 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1131 static int kvm_irqchip_get_virq(KVMState *s)
1136 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1137 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1138 * number can succeed even though a new route entry cannot be added.
1139 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1141 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1142 kvm_flush_dynamic_msi_routes(s);
1145 /* Return the lowest unused GSI in the bitmap */
1146 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1147 if (next_virq >= s->gsi_count) {
1154 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1156 unsigned int hash = kvm_hash_msi(msg.data);
1159 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1160 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1161 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1162 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1169 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1174 if (kvm_direct_msi_allowed) {
1175 msi.address_lo = (uint32_t)msg.address;
1176 msi.address_hi = msg.address >> 32;
1177 msi.data = le32_to_cpu(msg.data);
1179 memset(msi.pad, 0, sizeof(msi.pad));
1181 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1184 route = kvm_lookup_msi_route(s, msg);
1188 virq = kvm_irqchip_get_virq(s);
1193 route = g_malloc0(sizeof(KVMMSIRoute));
1194 route->kroute.gsi = virq;
1195 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1196 route->kroute.flags = 0;
1197 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1198 route->kroute.u.msi.address_hi = msg.address >> 32;
1199 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1201 kvm_add_routing_entry(s, &route->kroute);
1202 kvm_irqchip_commit_routes(s);
1204 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1208 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1210 return kvm_set_irq(s, route->kroute.gsi, 1);
1213 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1215 struct kvm_irq_routing_entry kroute = {};
1217 MSIMessage msg = {0, 0};
1219 if (pci_available && dev) {
1220 msg = pci_get_msi_message(dev, vector);
1223 if (kvm_gsi_direct_mapping()) {
1224 return kvm_arch_msi_data_to_gsi(msg.data);
1227 if (!kvm_gsi_routing_enabled()) {
1231 virq = kvm_irqchip_get_virq(s);
1237 kroute.type = KVM_IRQ_ROUTING_MSI;
1239 kroute.u.msi.address_lo = (uint32_t)msg.address;
1240 kroute.u.msi.address_hi = msg.address >> 32;
1241 kroute.u.msi.data = le32_to_cpu(msg.data);
1242 if (pci_available && kvm_msi_devid_required()) {
1243 kroute.flags = KVM_MSI_VALID_DEVID;
1244 kroute.u.msi.devid = pci_requester_id(dev);
1246 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1247 kvm_irqchip_release_virq(s, virq);
1251 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1254 kvm_add_routing_entry(s, &kroute);
1255 kvm_arch_add_msi_route_post(&kroute, vector, dev);
1256 kvm_irqchip_commit_routes(s);
1261 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1264 struct kvm_irq_routing_entry kroute = {};
1266 if (kvm_gsi_direct_mapping()) {
1270 if (!kvm_irqchip_in_kernel()) {
1275 kroute.type = KVM_IRQ_ROUTING_MSI;
1277 kroute.u.msi.address_lo = (uint32_t)msg.address;
1278 kroute.u.msi.address_hi = msg.address >> 32;
1279 kroute.u.msi.data = le32_to_cpu(msg.data);
1280 if (pci_available && kvm_msi_devid_required()) {
1281 kroute.flags = KVM_MSI_VALID_DEVID;
1282 kroute.u.msi.devid = pci_requester_id(dev);
1284 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1288 trace_kvm_irqchip_update_msi_route(virq);
1290 return kvm_update_routing_entry(s, &kroute);
1293 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1296 struct kvm_irqfd irqfd = {
1299 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1303 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1304 irqfd.resamplefd = rfd;
1307 if (!kvm_irqfds_enabled()) {
1311 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1314 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1316 struct kvm_irq_routing_entry kroute = {};
1319 if (!kvm_gsi_routing_enabled()) {
1323 virq = kvm_irqchip_get_virq(s);
1329 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1331 kroute.u.adapter.summary_addr = adapter->summary_addr;
1332 kroute.u.adapter.ind_addr = adapter->ind_addr;
1333 kroute.u.adapter.summary_offset = adapter->summary_offset;
1334 kroute.u.adapter.ind_offset = adapter->ind_offset;
1335 kroute.u.adapter.adapter_id = adapter->adapter_id;
1337 kvm_add_routing_entry(s, &kroute);
1342 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1344 struct kvm_irq_routing_entry kroute = {};
1347 if (!kvm_gsi_routing_enabled()) {
1350 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1353 virq = kvm_irqchip_get_virq(s);
1359 kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1361 kroute.u.hv_sint.vcpu = vcpu;
1362 kroute.u.hv_sint.sint = sint;
1364 kvm_add_routing_entry(s, &kroute);
1365 kvm_irqchip_commit_routes(s);
1370 #else /* !KVM_CAP_IRQ_ROUTING */
1372 void kvm_init_irq_routing(KVMState *s)
1376 void kvm_irqchip_release_virq(KVMState *s, int virq)
1380 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1385 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1390 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1395 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1400 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1405 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1409 #endif /* !KVM_CAP_IRQ_ROUTING */
1411 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1412 EventNotifier *rn, int virq)
1414 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1415 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1418 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1421 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1425 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1426 EventNotifier *rn, qemu_irq irq)
1429 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1434 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1437 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1441 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1446 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1449 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1451 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1454 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1458 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1460 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1461 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1463 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1470 /* First probe and see if there's a arch-specific hook to create the
1471 * in-kernel irqchip for us */
1472 ret = kvm_arch_irqchip_create(machine, s);
1474 if (machine_kernel_irqchip_split(machine)) {
1475 perror("Split IRQ chip mode not supported.");
1478 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1482 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1486 kvm_kernel_irqchip = true;
1487 /* If we have an in-kernel IRQ chip then we must have asynchronous
1488 * interrupt delivery (though the reverse is not necessarily true)
1490 kvm_async_interrupts_allowed = true;
1491 kvm_halt_in_kernel_allowed = true;
1493 kvm_init_irq_routing(s);
1495 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1498 /* Find number of supported CPUs using the recommended
1499 * procedure from the kernel API documentation to cope with
1500 * older kernels that may be missing capabilities.
1502 static int kvm_recommended_vcpus(KVMState *s)
1504 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1505 return (ret) ? ret : 4;
1508 static int kvm_max_vcpus(KVMState *s)
1510 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1511 return (ret) ? ret : kvm_recommended_vcpus(s);
1514 static int kvm_max_vcpu_id(KVMState *s)
1516 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1517 return (ret) ? ret : kvm_max_vcpus(s);
1520 bool kvm_vcpu_id_is_valid(int vcpu_id)
1522 KVMState *s = KVM_STATE(current_machine->accelerator);
1523 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1526 static int kvm_init(MachineState *ms)
1528 MachineClass *mc = MACHINE_GET_CLASS(ms);
1529 static const char upgrade_note[] =
1530 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1531 "(see http://sourceforge.net/projects/kvm).\n";
1536 { "SMP", smp_cpus },
1537 { "hotpluggable", max_cpus },
1540 int soft_vcpus_limit, hard_vcpus_limit;
1542 const KVMCapabilityInfo *missing_cap;
1545 const char *kvm_type;
1547 s = KVM_STATE(ms->accelerator);
1550 * On systems where the kernel can support different base page
1551 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1552 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1553 * page size for the system though.
1555 assert(TARGET_PAGE_SIZE <= getpagesize());
1559 #ifdef KVM_CAP_SET_GUEST_DEBUG
1560 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1562 QLIST_INIT(&s->kvm_parked_vcpus);
1564 s->fd = qemu_open("/dev/kvm", O_RDWR);
1566 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1571 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1572 if (ret < KVM_API_VERSION) {
1576 fprintf(stderr, "kvm version too old\n");
1580 if (ret > KVM_API_VERSION) {
1582 fprintf(stderr, "kvm version not supported\n");
1586 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1587 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1589 /* If unspecified, use the default value */
1594 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1596 type = mc->kvm_type(kvm_type);
1597 } else if (kvm_type) {
1599 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1604 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1605 } while (ret == -EINTR);
1608 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1612 if (ret == -EINVAL) {
1614 "Host kernel setup problem detected. Please verify:\n");
1615 fprintf(stderr, "- for kernels supporting the switch_amode or"
1616 " user_mode parameters, whether\n");
1618 " user space is running in primary address space\n");
1620 "- for kernels supporting the vm.allocate_pgste sysctl, "
1621 "whether it is enabled\n");
1629 /* check the vcpu limits */
1630 soft_vcpus_limit = kvm_recommended_vcpus(s);
1631 hard_vcpus_limit = kvm_max_vcpus(s);
1634 if (nc->num > soft_vcpus_limit) {
1635 warn_report("Number of %s cpus requested (%d) exceeds "
1636 "the recommended cpus supported by KVM (%d)",
1637 nc->name, nc->num, soft_vcpus_limit);
1639 if (nc->num > hard_vcpus_limit) {
1640 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1641 "the maximum cpus supported by KVM (%d)\n",
1642 nc->name, nc->num, hard_vcpus_limit);
1649 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1652 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1656 fprintf(stderr, "kvm does not support %s\n%s",
1657 missing_cap->name, upgrade_note);
1661 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1662 s->coalesced_pio = s->coalesced_mmio &&
1663 kvm_check_extension(s, KVM_CAP_COALESCED_PIO);
1665 #ifdef KVM_CAP_VCPU_EVENTS
1666 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1669 s->robust_singlestep =
1670 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1672 #ifdef KVM_CAP_DEBUGREGS
1673 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1676 #ifdef KVM_CAP_IRQ_ROUTING
1677 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1680 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1682 s->irq_set_ioctl = KVM_IRQ_LINE;
1683 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1684 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1687 kvm_readonly_mem_allowed =
1688 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1690 kvm_eventfds_allowed =
1691 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1693 kvm_irqfds_allowed =
1694 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1696 kvm_resamplefds_allowed =
1697 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1699 kvm_vm_attributes_allowed =
1700 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1702 kvm_ioeventfd_any_length_allowed =
1703 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1708 * if memory encryption object is specified then initialize the memory
1709 * encryption context.
1711 if (ms->memory_encryption) {
1712 kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption);
1713 if (!kvm_state->memcrypt_handle) {
1718 kvm_state->memcrypt_encrypt_data = sev_encrypt_data;
1721 ret = kvm_arch_init(ms, s);
1726 if (machine_kernel_irqchip_allowed(ms)) {
1727 kvm_irqchip_create(ms, s);
1730 if (kvm_eventfds_allowed) {
1731 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1732 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1734 s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region;
1735 s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region;
1737 kvm_memory_listener_register(s, &s->memory_listener,
1738 &address_space_memory, 0);
1739 memory_listener_register(&kvm_io_listener,
1741 memory_listener_register(&kvm_coalesced_pio_listener,
1744 s->many_ioeventfds = kvm_check_many_ioeventfds();
1746 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1748 qemu_balloon_inhibit(true);
1761 g_free(s->memory_listener.slots);
1766 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1768 s->sigmask_len = sigmask_len;
1771 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1772 int size, uint32_t count)
1775 uint8_t *ptr = data;
1777 for (i = 0; i < count; i++) {
1778 address_space_rw(&address_space_io, port, attrs,
1780 direction == KVM_EXIT_IO_OUT);
1785 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1787 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1788 run->internal.suberror);
1790 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1793 for (i = 0; i < run->internal.ndata; ++i) {
1794 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1795 i, (uint64_t)run->internal.data[i]);
1798 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1799 fprintf(stderr, "emulation failure\n");
1800 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1801 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1802 return EXCP_INTERRUPT;
1805 /* FIXME: Should trigger a qmp message to let management know
1806 * something went wrong.
1811 void kvm_flush_coalesced_mmio_buffer(void)
1813 KVMState *s = kvm_state;
1815 if (s->coalesced_flush_in_progress) {
1819 s->coalesced_flush_in_progress = true;
1821 if (s->coalesced_mmio_ring) {
1822 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1823 while (ring->first != ring->last) {
1824 struct kvm_coalesced_mmio *ent;
1826 ent = &ring->coalesced_mmio[ring->first];
1828 if (ent->pio == 1) {
1829 address_space_rw(&address_space_io, ent->phys_addr,
1830 MEMTXATTRS_UNSPECIFIED, ent->data,
1833 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1836 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1840 s->coalesced_flush_in_progress = false;
1843 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1845 if (!cpu->vcpu_dirty) {
1846 kvm_arch_get_registers(cpu);
1847 cpu->vcpu_dirty = true;
1851 void kvm_cpu_synchronize_state(CPUState *cpu)
1853 if (!cpu->vcpu_dirty) {
1854 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1858 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1860 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1861 cpu->vcpu_dirty = false;
1864 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1866 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1869 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1871 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1872 cpu->vcpu_dirty = false;
1875 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1877 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1880 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
1882 cpu->vcpu_dirty = true;
1885 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
1887 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1890 #ifdef KVM_HAVE_MCE_INJECTION
1891 static __thread void *pending_sigbus_addr;
1892 static __thread int pending_sigbus_code;
1893 static __thread bool have_sigbus_pending;
1896 static void kvm_cpu_kick(CPUState *cpu)
1898 atomic_set(&cpu->kvm_run->immediate_exit, 1);
1901 static void kvm_cpu_kick_self(void)
1903 if (kvm_immediate_exit) {
1904 kvm_cpu_kick(current_cpu);
1906 qemu_cpu_kick_self();
1910 static void kvm_eat_signals(CPUState *cpu)
1912 struct timespec ts = { 0, 0 };
1918 if (kvm_immediate_exit) {
1919 atomic_set(&cpu->kvm_run->immediate_exit, 0);
1920 /* Write kvm_run->immediate_exit before the cpu->exit_request
1921 * write in kvm_cpu_exec.
1927 sigemptyset(&waitset);
1928 sigaddset(&waitset, SIG_IPI);
1931 r = sigtimedwait(&waitset, &siginfo, &ts);
1932 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1933 perror("sigtimedwait");
1937 r = sigpending(&chkset);
1939 perror("sigpending");
1942 } while (sigismember(&chkset, SIG_IPI));
1945 int kvm_cpu_exec(CPUState *cpu)
1947 struct kvm_run *run = cpu->kvm_run;
1950 DPRINTF("kvm_cpu_exec()\n");
1952 if (kvm_arch_process_async_events(cpu)) {
1953 atomic_set(&cpu->exit_request, 0);
1957 qemu_mutex_unlock_iothread();
1958 cpu_exec_start(cpu);
1963 if (cpu->vcpu_dirty) {
1964 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1965 cpu->vcpu_dirty = false;
1968 kvm_arch_pre_run(cpu, run);
1969 if (atomic_read(&cpu->exit_request)) {
1970 DPRINTF("interrupt exit requested\n");
1972 * KVM requires us to reenter the kernel after IO exits to complete
1973 * instruction emulation. This self-signal will ensure that we
1976 kvm_cpu_kick_self();
1979 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1980 * Matching barrier in kvm_eat_signals.
1984 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1986 attrs = kvm_arch_post_run(cpu, run);
1988 #ifdef KVM_HAVE_MCE_INJECTION
1989 if (unlikely(have_sigbus_pending)) {
1990 qemu_mutex_lock_iothread();
1991 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
1992 pending_sigbus_addr);
1993 have_sigbus_pending = false;
1994 qemu_mutex_unlock_iothread();
1999 if (run_ret == -EINTR || run_ret == -EAGAIN) {
2000 DPRINTF("io window exit\n");
2001 kvm_eat_signals(cpu);
2002 ret = EXCP_INTERRUPT;
2005 fprintf(stderr, "error: kvm run failed %s\n",
2006 strerror(-run_ret));
2008 if (run_ret == -EBUSY) {
2010 "This is probably because your SMT is enabled.\n"
2011 "VCPU can only run on primary threads with all "
2012 "secondary threads offline.\n");
2019 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
2020 switch (run->exit_reason) {
2022 DPRINTF("handle_io\n");
2023 /* Called outside BQL */
2024 kvm_handle_io(run->io.port, attrs,
2025 (uint8_t *)run + run->io.data_offset,
2032 DPRINTF("handle_mmio\n");
2033 /* Called outside BQL */
2034 address_space_rw(&address_space_memory,
2035 run->mmio.phys_addr, attrs,
2038 run->mmio.is_write);
2041 case KVM_EXIT_IRQ_WINDOW_OPEN:
2042 DPRINTF("irq_window_open\n");
2043 ret = EXCP_INTERRUPT;
2045 case KVM_EXIT_SHUTDOWN:
2046 DPRINTF("shutdown\n");
2047 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2048 ret = EXCP_INTERRUPT;
2050 case KVM_EXIT_UNKNOWN:
2051 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
2052 (uint64_t)run->hw.hardware_exit_reason);
2055 case KVM_EXIT_INTERNAL_ERROR:
2056 ret = kvm_handle_internal_error(cpu, run);
2058 case KVM_EXIT_SYSTEM_EVENT:
2059 switch (run->system_event.type) {
2060 case KVM_SYSTEM_EVENT_SHUTDOWN:
2061 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2062 ret = EXCP_INTERRUPT;
2064 case KVM_SYSTEM_EVENT_RESET:
2065 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2066 ret = EXCP_INTERRUPT;
2068 case KVM_SYSTEM_EVENT_CRASH:
2069 kvm_cpu_synchronize_state(cpu);
2070 qemu_mutex_lock_iothread();
2071 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2072 qemu_mutex_unlock_iothread();
2076 DPRINTF("kvm_arch_handle_exit\n");
2077 ret = kvm_arch_handle_exit(cpu, run);
2082 DPRINTF("kvm_arch_handle_exit\n");
2083 ret = kvm_arch_handle_exit(cpu, run);
2089 qemu_mutex_lock_iothread();
2092 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
2093 vm_stop(RUN_STATE_INTERNAL_ERROR);
2096 atomic_set(&cpu->exit_request, 0);
2100 int kvm_ioctl(KVMState *s, int type, ...)
2107 arg = va_arg(ap, void *);
2110 trace_kvm_ioctl(type, arg);
2111 ret = ioctl(s->fd, type, arg);
2118 int kvm_vm_ioctl(KVMState *s, int type, ...)
2125 arg = va_arg(ap, void *);
2128 trace_kvm_vm_ioctl(type, arg);
2129 ret = ioctl(s->vmfd, type, arg);
2136 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2143 arg = va_arg(ap, void *);
2146 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2147 ret = ioctl(cpu->kvm_fd, type, arg);
2154 int kvm_device_ioctl(int fd, int type, ...)
2161 arg = va_arg(ap, void *);
2164 trace_kvm_device_ioctl(fd, type, arg);
2165 ret = ioctl(fd, type, arg);
2172 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2175 struct kvm_device_attr attribute = {
2180 if (!kvm_vm_attributes_allowed) {
2184 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2185 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2189 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2191 struct kvm_device_attr attribute = {
2197 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2200 int kvm_device_access(int fd, int group, uint64_t attr,
2201 void *val, bool write, Error **errp)
2203 struct kvm_device_attr kvmattr;
2207 kvmattr.group = group;
2208 kvmattr.attr = attr;
2209 kvmattr.addr = (uintptr_t)val;
2211 err = kvm_device_ioctl(fd,
2212 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2215 error_setg_errno(errp, -err,
2216 "KVM_%s_DEVICE_ATTR failed: Group %d "
2217 "attr 0x%016" PRIx64,
2218 write ? "SET" : "GET", group, attr);
2223 bool kvm_has_sync_mmu(void)
2225 return kvm_state->sync_mmu;
2228 int kvm_has_vcpu_events(void)
2230 return kvm_state->vcpu_events;
2233 int kvm_has_robust_singlestep(void)
2235 return kvm_state->robust_singlestep;
2238 int kvm_has_debugregs(void)
2240 return kvm_state->debugregs;
2243 int kvm_has_many_ioeventfds(void)
2245 if (!kvm_enabled()) {
2248 return kvm_state->many_ioeventfds;
2251 int kvm_has_gsi_routing(void)
2253 #ifdef KVM_CAP_IRQ_ROUTING
2254 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2260 int kvm_has_intx_set_mask(void)
2262 return kvm_state->intx_set_mask;
2265 bool kvm_arm_supports_user_irq(void)
2267 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2270 #ifdef KVM_CAP_SET_GUEST_DEBUG
2271 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2274 struct kvm_sw_breakpoint *bp;
2276 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2284 int kvm_sw_breakpoints_active(CPUState *cpu)
2286 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2289 struct kvm_set_guest_debug_data {
2290 struct kvm_guest_debug dbg;
2294 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2296 struct kvm_set_guest_debug_data *dbg_data =
2297 (struct kvm_set_guest_debug_data *) data.host_ptr;
2299 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2303 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2305 struct kvm_set_guest_debug_data data;
2307 data.dbg.control = reinject_trap;
2309 if (cpu->singlestep_enabled) {
2310 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2312 kvm_arch_update_guest_debug(cpu, &data.dbg);
2314 run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2315 RUN_ON_CPU_HOST_PTR(&data));
2319 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2320 target_ulong len, int type)
2322 struct kvm_sw_breakpoint *bp;
2325 if (type == GDB_BREAKPOINT_SW) {
2326 bp = kvm_find_sw_breakpoint(cpu, addr);
2332 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2335 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2341 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2343 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2350 err = kvm_update_guest_debug(cpu, 0);
2358 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2359 target_ulong len, int type)
2361 struct kvm_sw_breakpoint *bp;
2364 if (type == GDB_BREAKPOINT_SW) {
2365 bp = kvm_find_sw_breakpoint(cpu, addr);
2370 if (bp->use_count > 1) {
2375 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2380 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2383 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2390 err = kvm_update_guest_debug(cpu, 0);
2398 void kvm_remove_all_breakpoints(CPUState *cpu)
2400 struct kvm_sw_breakpoint *bp, *next;
2401 KVMState *s = cpu->kvm_state;
2404 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2405 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2406 /* Try harder to find a CPU that currently sees the breakpoint. */
2407 CPU_FOREACH(tmpcpu) {
2408 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2413 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2416 kvm_arch_remove_all_hw_breakpoints();
2419 kvm_update_guest_debug(cpu, 0);
2423 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2425 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2430 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2431 target_ulong len, int type)
2436 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2437 target_ulong len, int type)
2442 void kvm_remove_all_breakpoints(CPUState *cpu)
2445 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2447 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2449 KVMState *s = kvm_state;
2450 struct kvm_signal_mask *sigmask;
2453 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2455 sigmask->len = s->sigmask_len;
2456 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2457 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2463 static void kvm_ipi_signal(int sig)
2466 assert(kvm_immediate_exit);
2467 kvm_cpu_kick(current_cpu);
2471 void kvm_init_cpu_signals(CPUState *cpu)
2475 struct sigaction sigact;
2477 memset(&sigact, 0, sizeof(sigact));
2478 sigact.sa_handler = kvm_ipi_signal;
2479 sigaction(SIG_IPI, &sigact, NULL);
2481 pthread_sigmask(SIG_BLOCK, NULL, &set);
2482 #if defined KVM_HAVE_MCE_INJECTION
2483 sigdelset(&set, SIGBUS);
2484 pthread_sigmask(SIG_SETMASK, &set, NULL);
2486 sigdelset(&set, SIG_IPI);
2487 if (kvm_immediate_exit) {
2488 r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2490 r = kvm_set_signal_mask(cpu, &set);
2493 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2498 /* Called asynchronously in VCPU thread. */
2499 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2501 #ifdef KVM_HAVE_MCE_INJECTION
2502 if (have_sigbus_pending) {
2505 have_sigbus_pending = true;
2506 pending_sigbus_addr = addr;
2507 pending_sigbus_code = code;
2508 atomic_set(&cpu->exit_request, 1);
2515 /* Called synchronously (via signalfd) in main thread. */
2516 int kvm_on_sigbus(int code, void *addr)
2518 #ifdef KVM_HAVE_MCE_INJECTION
2519 /* Action required MCE kills the process if SIGBUS is blocked. Because
2520 * that's what happens in the I/O thread, where we handle MCE via signalfd,
2521 * we can only get action optional here.
2523 assert(code != BUS_MCEERR_AR);
2524 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2531 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2534 struct kvm_create_device create_dev;
2536 create_dev.type = type;
2538 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2540 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2544 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2549 return test ? 0 : create_dev.fd;
2552 bool kvm_device_supported(int vmfd, uint64_t type)
2554 struct kvm_create_device create_dev = {
2557 .flags = KVM_CREATE_DEVICE_TEST,
2560 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2564 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2567 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2569 struct kvm_one_reg reg;
2573 reg.addr = (uintptr_t) source;
2574 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®);
2576 trace_kvm_failed_reg_set(id, strerror(-r));
2581 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2583 struct kvm_one_reg reg;
2587 reg.addr = (uintptr_t) target;
2588 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®);
2590 trace_kvm_failed_reg_get(id, strerror(-r));
2595 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2597 AccelClass *ac = ACCEL_CLASS(oc);
2599 ac->init_machine = kvm_init;
2600 ac->allowed = &kvm_allowed;
2603 static const TypeInfo kvm_accel_type = {
2604 .name = TYPE_KVM_ACCEL,
2605 .parent = TYPE_ACCEL,
2606 .class_init = kvm_accel_class_init,
2607 .instance_size = sizeof(KVMState),
2610 static void kvm_type_init(void)
2612 type_register_static(&kvm_accel_type);
2615 type_init(kvm_type_init);