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/atomic.h"
22 #include "qemu/option.h"
23 #include "qemu/config-file.h"
24 #include "qemu/error-report.h"
25 #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/cpus.h"
33 #include "qemu/bswap.h"
34 #include "exec/memory.h"
35 #include "exec/ram_addr.h"
36 #include "exec/address-spaces.h"
37 #include "qemu/event_notifier.h"
40 #include "sysemu/sev.h"
41 #include "sysemu/balloon.h"
43 #include "hw/boards.h"
45 /* This check must be after config-host.h is included */
47 #include <sys/eventfd.h>
50 /* KVM uses PAGE_SIZE in its definition of KVM_COALESCED_MMIO_MAX. We
51 * need to use the real host PAGE_SIZE, as that's what KVM will use.
53 #define PAGE_SIZE getpagesize()
58 #define DPRINTF(fmt, ...) \
59 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
61 #define DPRINTF(fmt, ...) \
65 #define KVM_MSI_HASHTAB_SIZE 256
67 struct KVMParkedVcpu {
68 unsigned long vcpu_id;
70 QLIST_ENTRY(KVMParkedVcpu) node;
75 AccelState parent_obj;
82 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
83 bool coalesced_flush_in_progress;
85 int robust_singlestep;
87 #ifdef KVM_CAP_SET_GUEST_DEBUG
88 QTAILQ_HEAD(, kvm_sw_breakpoint) kvm_sw_breakpoints;
93 /* The man page (and posix) say ioctl numbers are signed int, but
94 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
95 * unsigned, and treating them as signed here can break things */
96 unsigned irq_set_ioctl;
97 unsigned int sigmask_len;
99 #ifdef KVM_CAP_IRQ_ROUTING
100 struct kvm_irq_routing *irq_routes;
101 int nr_allocated_irq_routes;
102 unsigned long *used_gsi_bitmap;
103 unsigned int gsi_count;
104 QTAILQ_HEAD(, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
106 KVMMemoryListener memory_listener;
107 QLIST_HEAD(, KVMParkedVcpu) kvm_parked_vcpus;
109 /* memory encryption */
110 void *memcrypt_handle;
111 int (*memcrypt_encrypt_data)(void *handle, uint8_t *ptr, uint64_t len);
115 bool kvm_kernel_irqchip;
116 bool kvm_split_irqchip;
117 bool kvm_async_interrupts_allowed;
118 bool kvm_halt_in_kernel_allowed;
119 bool kvm_eventfds_allowed;
120 bool kvm_irqfds_allowed;
121 bool kvm_resamplefds_allowed;
122 bool kvm_msi_via_irqfd_allowed;
123 bool kvm_gsi_routing_allowed;
124 bool kvm_gsi_direct_mapping;
126 bool kvm_readonly_mem_allowed;
127 bool kvm_vm_attributes_allowed;
128 bool kvm_direct_msi_allowed;
129 bool kvm_ioeventfd_any_length_allowed;
130 bool kvm_msi_use_devid;
131 static bool kvm_immediate_exit;
133 static const KVMCapabilityInfo kvm_required_capabilites[] = {
134 KVM_CAP_INFO(USER_MEMORY),
135 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
136 KVM_CAP_INFO(JOIN_MEMORY_REGIONS_WORKS),
140 int kvm_get_max_memslots(void)
142 KVMState *s = KVM_STATE(current_machine->accelerator);
147 bool kvm_memcrypt_enabled(void)
149 if (kvm_state && kvm_state->memcrypt_handle) {
156 int kvm_memcrypt_encrypt_data(uint8_t *ptr, uint64_t len)
158 if (kvm_state->memcrypt_handle &&
159 kvm_state->memcrypt_encrypt_data) {
160 return kvm_state->memcrypt_encrypt_data(kvm_state->memcrypt_handle,
167 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
169 KVMState *s = kvm_state;
172 for (i = 0; i < s->nr_slots; i++) {
173 if (kml->slots[i].memory_size == 0) {
174 return &kml->slots[i];
181 bool kvm_has_free_slot(MachineState *ms)
183 KVMState *s = KVM_STATE(ms->accelerator);
185 return kvm_get_free_slot(&s->memory_listener);
188 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
190 KVMSlot *slot = kvm_get_free_slot(kml);
196 fprintf(stderr, "%s: no free slot available\n", __func__);
200 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
204 KVMState *s = kvm_state;
207 for (i = 0; i < s->nr_slots; i++) {
208 KVMSlot *mem = &kml->slots[i];
210 if (start_addr == mem->start_addr && size == mem->memory_size) {
219 * Calculate and align the start address and the size of the section.
220 * Return the size. If the size is 0, the aligned section is empty.
222 static hwaddr kvm_align_section(MemoryRegionSection *section,
225 hwaddr size = int128_get64(section->size);
226 hwaddr delta, aligned;
228 /* kvm works in page size chunks, but the function may be called
229 with sub-page size and unaligned start address. Pad the start
230 address to next and truncate size to previous page boundary. */
231 aligned = ROUND_UP(section->offset_within_address_space,
232 qemu_real_host_page_size);
233 delta = aligned - section->offset_within_address_space;
239 return (size - delta) & qemu_real_host_page_mask;
242 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
245 KVMMemoryListener *kml = &s->memory_listener;
248 for (i = 0; i < s->nr_slots; i++) {
249 KVMSlot *mem = &kml->slots[i];
251 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
252 *phys_addr = mem->start_addr + (ram - mem->ram);
260 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot, bool new)
262 KVMState *s = kvm_state;
263 struct kvm_userspace_memory_region mem;
266 mem.slot = slot->slot | (kml->as_id << 16);
267 mem.guest_phys_addr = slot->start_addr;
268 mem.userspace_addr = (unsigned long)slot->ram;
269 mem.flags = slot->flags;
271 if (slot->memory_size && !new && (mem.flags ^ slot->old_flags) & KVM_MEM_READONLY) {
272 /* Set the slot size to 0 before setting the slot to the desired
273 * value. This is needed based on KVM commit 75d61fbc. */
275 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
277 mem.memory_size = slot->memory_size;
278 ret = kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
279 slot->old_flags = mem.flags;
280 trace_kvm_set_user_memory(mem.slot, mem.flags, mem.guest_phys_addr,
281 mem.memory_size, mem.userspace_addr, ret);
285 int kvm_destroy_vcpu(CPUState *cpu)
287 KVMState *s = kvm_state;
289 struct KVMParkedVcpu *vcpu = NULL;
292 DPRINTF("kvm_destroy_vcpu\n");
294 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
297 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
301 ret = munmap(cpu->kvm_run, mmap_size);
306 vcpu = g_malloc0(sizeof(*vcpu));
307 vcpu->vcpu_id = kvm_arch_vcpu_id(cpu);
308 vcpu->kvm_fd = cpu->kvm_fd;
309 QLIST_INSERT_HEAD(&kvm_state->kvm_parked_vcpus, vcpu, node);
314 static int kvm_get_vcpu(KVMState *s, unsigned long vcpu_id)
316 struct KVMParkedVcpu *cpu;
318 QLIST_FOREACH(cpu, &s->kvm_parked_vcpus, node) {
319 if (cpu->vcpu_id == vcpu_id) {
322 QLIST_REMOVE(cpu, node);
323 kvm_fd = cpu->kvm_fd;
329 return kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)vcpu_id);
332 int kvm_init_vcpu(CPUState *cpu)
334 KVMState *s = kvm_state;
338 DPRINTF("kvm_init_vcpu\n");
340 ret = kvm_get_vcpu(s, kvm_arch_vcpu_id(cpu));
342 DPRINTF("kvm_create_vcpu failed\n");
348 cpu->vcpu_dirty = true;
350 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
353 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
357 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
359 if (cpu->kvm_run == MAP_FAILED) {
361 DPRINTF("mmap'ing vcpu state failed\n");
365 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
366 s->coalesced_mmio_ring =
367 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
370 ret = kvm_arch_init_vcpu(cpu);
376 * dirty pages logging control
379 static int kvm_mem_flags(MemoryRegion *mr)
381 bool readonly = mr->readonly || memory_region_is_romd(mr);
384 if (memory_region_get_dirty_log_mask(mr) != 0) {
385 flags |= KVM_MEM_LOG_DIRTY_PAGES;
387 if (readonly && kvm_readonly_mem_allowed) {
388 flags |= KVM_MEM_READONLY;
393 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
396 mem->flags = kvm_mem_flags(mr);
398 /* If nothing changed effectively, no need to issue ioctl */
399 if (mem->flags == mem->old_flags) {
403 return kvm_set_user_memory_region(kml, mem, false);
406 static int kvm_section_update_flags(KVMMemoryListener *kml,
407 MemoryRegionSection *section)
409 hwaddr start_addr, size;
412 size = kvm_align_section(section, &start_addr);
417 mem = kvm_lookup_matching_slot(kml, start_addr, size);
419 /* We don't have a slot if we want to trap every access. */
423 return kvm_slot_update_flags(kml, mem, section->mr);
426 static void kvm_log_start(MemoryListener *listener,
427 MemoryRegionSection *section,
430 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
437 r = kvm_section_update_flags(kml, section);
443 static void kvm_log_stop(MemoryListener *listener,
444 MemoryRegionSection *section,
447 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
454 r = kvm_section_update_flags(kml, section);
460 /* get kvm's dirty pages bitmap and update qemu's */
461 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
462 unsigned long *bitmap)
464 ram_addr_t start = section->offset_within_region +
465 memory_region_get_ram_addr(section->mr);
466 ram_addr_t pages = int128_get64(section->size) / getpagesize();
468 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
472 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
475 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
476 * This function updates qemu's dirty bitmap using
477 * memory_region_set_dirty(). This means all bits are set
480 * @start_add: start of logged region.
481 * @end_addr: end of logged region.
483 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
484 MemoryRegionSection *section)
486 KVMState *s = kvm_state;
487 struct kvm_dirty_log d = {};
489 hwaddr start_addr, size;
491 size = kvm_align_section(section, &start_addr);
493 mem = kvm_lookup_matching_slot(kml, start_addr, size);
495 /* We don't have a slot if we want to trap every access. */
499 /* XXX bad kernel interface alert
500 * For dirty bitmap, kernel allocates array of size aligned to
501 * bits-per-long. But for case when the kernel is 64bits and
502 * the userspace is 32bits, userspace can't align to the same
503 * bits-per-long, since sizeof(long) is different between kernel
504 * and user space. This way, userspace will provide buffer which
505 * may be 4 bytes less than the kernel will use, resulting in
506 * userspace memory corruption (which is not detectable by valgrind
507 * too, in most cases).
508 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
509 * a hope that sizeof(long) won't become >8 any time soon.
511 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
512 /*HOST_LONG_BITS*/ 64) / 8;
513 d.dirty_bitmap = g_malloc0(size);
515 d.slot = mem->slot | (kml->as_id << 16);
516 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
517 DPRINTF("ioctl failed %d\n", errno);
518 g_free(d.dirty_bitmap);
522 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
523 g_free(d.dirty_bitmap);
529 static void kvm_coalesce_mmio_region(MemoryListener *listener,
530 MemoryRegionSection *secion,
531 hwaddr start, hwaddr size)
533 KVMState *s = kvm_state;
535 if (s->coalesced_mmio) {
536 struct kvm_coalesced_mmio_zone zone;
542 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
546 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
547 MemoryRegionSection *secion,
548 hwaddr start, hwaddr size)
550 KVMState *s = kvm_state;
552 if (s->coalesced_mmio) {
553 struct kvm_coalesced_mmio_zone zone;
559 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
563 static void kvm_coalesce_pio_add(MemoryListener *listener,
564 MemoryRegionSection *section,
565 hwaddr start, hwaddr size)
567 KVMState *s = kvm_state;
569 if (s->coalesced_pio) {
570 struct kvm_coalesced_mmio_zone zone;
576 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
580 static void kvm_coalesce_pio_del(MemoryListener *listener,
581 MemoryRegionSection *section,
582 hwaddr start, hwaddr size)
584 KVMState *s = kvm_state;
586 if (s->coalesced_pio) {
587 struct kvm_coalesced_mmio_zone zone;
593 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
597 static MemoryListener kvm_coalesced_pio_listener = {
598 .coalesced_io_add = kvm_coalesce_pio_add,
599 .coalesced_io_del = kvm_coalesce_pio_del,
602 int kvm_check_extension(KVMState *s, unsigned int extension)
606 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
614 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
618 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
620 /* VM wide version not implemented, use global one instead */
621 ret = kvm_check_extension(s, extension);
627 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
629 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
630 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
631 * endianness, but the memory core hands them in target endianness.
632 * For example, PPC is always treated as big-endian even if running
633 * on KVM and on PPC64LE. Correct here.
647 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
648 bool assign, uint32_t size, bool datamatch)
651 struct kvm_ioeventfd iofd = {
652 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
659 trace_kvm_set_ioeventfd_mmio(fd, (uint64_t)addr, val, assign, size,
661 if (!kvm_enabled()) {
666 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
669 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
672 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
681 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
682 bool assign, uint32_t size, bool datamatch)
684 struct kvm_ioeventfd kick = {
685 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
687 .flags = KVM_IOEVENTFD_FLAG_PIO,
692 trace_kvm_set_ioeventfd_pio(fd, addr, val, assign, size, datamatch);
693 if (!kvm_enabled()) {
697 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
700 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
702 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
710 static int kvm_check_many_ioeventfds(void)
712 /* Userspace can use ioeventfd for io notification. This requires a host
713 * that supports eventfd(2) and an I/O thread; since eventfd does not
714 * support SIGIO it cannot interrupt the vcpu.
716 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
717 * can avoid creating too many ioeventfds.
719 #if defined(CONFIG_EVENTFD)
722 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
723 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
724 if (ioeventfds[i] < 0) {
727 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
729 close(ioeventfds[i]);
734 /* Decide whether many devices are supported or not */
735 ret = i == ARRAY_SIZE(ioeventfds);
738 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
739 close(ioeventfds[i]);
747 static const KVMCapabilityInfo *
748 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
751 if (!kvm_check_extension(s, list->value)) {
759 static void kvm_set_phys_mem(KVMMemoryListener *kml,
760 MemoryRegionSection *section, bool add)
764 MemoryRegion *mr = section->mr;
765 bool writeable = !mr->readonly && !mr->rom_device;
766 hwaddr start_addr, size;
769 if (!memory_region_is_ram(mr)) {
770 if (writeable || !kvm_readonly_mem_allowed) {
772 } else if (!mr->romd_mode) {
773 /* If the memory device is not in romd_mode, then we actually want
774 * to remove the kvm memory slot so all accesses will trap. */
779 size = kvm_align_section(section, &start_addr);
784 /* use aligned delta to align the ram address */
785 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region +
786 (start_addr - section->offset_within_address_space);
789 mem = kvm_lookup_matching_slot(kml, start_addr, size);
793 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
794 kvm_physical_sync_dirty_bitmap(kml, section);
797 /* unregister the slot */
798 mem->memory_size = 0;
800 err = kvm_set_user_memory_region(kml, mem, false);
802 fprintf(stderr, "%s: error unregistering slot: %s\n",
803 __func__, strerror(-err));
809 /* register the new slot */
810 mem = kvm_alloc_slot(kml);
811 mem->memory_size = size;
812 mem->start_addr = start_addr;
814 mem->flags = kvm_mem_flags(mr);
816 err = kvm_set_user_memory_region(kml, mem, true);
818 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
824 static void kvm_region_add(MemoryListener *listener,
825 MemoryRegionSection *section)
827 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
829 memory_region_ref(section->mr);
830 kvm_set_phys_mem(kml, section, true);
833 static void kvm_region_del(MemoryListener *listener,
834 MemoryRegionSection *section)
836 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
838 kvm_set_phys_mem(kml, section, false);
839 memory_region_unref(section->mr);
842 static void kvm_log_sync(MemoryListener *listener,
843 MemoryRegionSection *section)
845 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
848 r = kvm_physical_sync_dirty_bitmap(kml, section);
854 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
855 MemoryRegionSection *section,
856 bool match_data, uint64_t data,
859 int fd = event_notifier_get_fd(e);
862 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
863 data, true, int128_get64(section->size),
866 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
867 __func__, strerror(-r), -r);
872 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
873 MemoryRegionSection *section,
874 bool match_data, uint64_t data,
877 int fd = event_notifier_get_fd(e);
880 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
881 data, false, int128_get64(section->size),
884 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
885 __func__, strerror(-r), -r);
890 static void kvm_io_ioeventfd_add(MemoryListener *listener,
891 MemoryRegionSection *section,
892 bool match_data, uint64_t data,
895 int fd = event_notifier_get_fd(e);
898 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
899 data, true, int128_get64(section->size),
902 fprintf(stderr, "%s: error adding ioeventfd: %s (%d)\n",
903 __func__, strerror(-r), -r);
908 static void kvm_io_ioeventfd_del(MemoryListener *listener,
909 MemoryRegionSection *section,
910 bool match_data, uint64_t data,
914 int fd = event_notifier_get_fd(e);
917 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
918 data, false, int128_get64(section->size),
921 fprintf(stderr, "%s: error deleting ioeventfd: %s (%d)\n",
922 __func__, strerror(-r), -r);
927 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
928 AddressSpace *as, int as_id)
932 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
935 for (i = 0; i < s->nr_slots; i++) {
936 kml->slots[i].slot = i;
939 kml->listener.region_add = kvm_region_add;
940 kml->listener.region_del = kvm_region_del;
941 kml->listener.log_start = kvm_log_start;
942 kml->listener.log_stop = kvm_log_stop;
943 kml->listener.log_sync = kvm_log_sync;
944 kml->listener.priority = 10;
946 memory_listener_register(&kml->listener, as);
949 static MemoryListener kvm_io_listener = {
950 .eventfd_add = kvm_io_ioeventfd_add,
951 .eventfd_del = kvm_io_ioeventfd_del,
955 int kvm_set_irq(KVMState *s, int irq, int level)
957 struct kvm_irq_level event;
960 assert(kvm_async_interrupts_enabled());
964 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
966 perror("kvm_set_irq");
970 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
973 #ifdef KVM_CAP_IRQ_ROUTING
974 typedef struct KVMMSIRoute {
975 struct kvm_irq_routing_entry kroute;
976 QTAILQ_ENTRY(KVMMSIRoute) entry;
979 static void set_gsi(KVMState *s, unsigned int gsi)
981 set_bit(gsi, s->used_gsi_bitmap);
984 static void clear_gsi(KVMState *s, unsigned int gsi)
986 clear_bit(gsi, s->used_gsi_bitmap);
989 void kvm_init_irq_routing(KVMState *s)
993 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
995 /* Round up so we can search ints using ffs */
996 s->used_gsi_bitmap = bitmap_new(gsi_count);
997 s->gsi_count = gsi_count;
1000 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1001 s->nr_allocated_irq_routes = 0;
1003 if (!kvm_direct_msi_allowed) {
1004 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1005 QTAILQ_INIT(&s->msi_hashtab[i]);
1009 kvm_arch_init_irq_routing(s);
1012 void kvm_irqchip_commit_routes(KVMState *s)
1016 if (kvm_gsi_direct_mapping()) {
1020 if (!kvm_gsi_routing_enabled()) {
1024 s->irq_routes->flags = 0;
1025 trace_kvm_irqchip_commit_routes();
1026 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1030 static void kvm_add_routing_entry(KVMState *s,
1031 struct kvm_irq_routing_entry *entry)
1033 struct kvm_irq_routing_entry *new;
1036 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1037 n = s->nr_allocated_irq_routes * 2;
1041 size = sizeof(struct kvm_irq_routing);
1042 size += n * sizeof(*new);
1043 s->irq_routes = g_realloc(s->irq_routes, size);
1044 s->nr_allocated_irq_routes = n;
1046 n = s->irq_routes->nr++;
1047 new = &s->irq_routes->entries[n];
1051 set_gsi(s, entry->gsi);
1054 static int kvm_update_routing_entry(KVMState *s,
1055 struct kvm_irq_routing_entry *new_entry)
1057 struct kvm_irq_routing_entry *entry;
1060 for (n = 0; n < s->irq_routes->nr; n++) {
1061 entry = &s->irq_routes->entries[n];
1062 if (entry->gsi != new_entry->gsi) {
1066 if(!memcmp(entry, new_entry, sizeof *entry)) {
1070 *entry = *new_entry;
1078 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1080 struct kvm_irq_routing_entry e = {};
1082 assert(pin < s->gsi_count);
1085 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1087 e.u.irqchip.irqchip = irqchip;
1088 e.u.irqchip.pin = pin;
1089 kvm_add_routing_entry(s, &e);
1092 void kvm_irqchip_release_virq(KVMState *s, int virq)
1094 struct kvm_irq_routing_entry *e;
1097 if (kvm_gsi_direct_mapping()) {
1101 for (i = 0; i < s->irq_routes->nr; i++) {
1102 e = &s->irq_routes->entries[i];
1103 if (e->gsi == virq) {
1104 s->irq_routes->nr--;
1105 *e = s->irq_routes->entries[s->irq_routes->nr];
1109 kvm_arch_release_virq_post(virq);
1110 trace_kvm_irqchip_release_virq(virq);
1113 static unsigned int kvm_hash_msi(uint32_t data)
1115 /* This is optimized for IA32 MSI layout. However, no other arch shall
1116 * repeat the mistake of not providing a direct MSI injection API. */
1120 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1122 KVMMSIRoute *route, *next;
1125 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1126 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1127 kvm_irqchip_release_virq(s, route->kroute.gsi);
1128 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1134 static int kvm_irqchip_get_virq(KVMState *s)
1139 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1140 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1141 * number can succeed even though a new route entry cannot be added.
1142 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1144 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1145 kvm_flush_dynamic_msi_routes(s);
1148 /* Return the lowest unused GSI in the bitmap */
1149 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1150 if (next_virq >= s->gsi_count) {
1157 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1159 unsigned int hash = kvm_hash_msi(msg.data);
1162 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1163 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1164 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1165 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1172 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1177 if (kvm_direct_msi_allowed) {
1178 msi.address_lo = (uint32_t)msg.address;
1179 msi.address_hi = msg.address >> 32;
1180 msi.data = le32_to_cpu(msg.data);
1182 memset(msi.pad, 0, sizeof(msi.pad));
1184 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1187 route = kvm_lookup_msi_route(s, msg);
1191 virq = kvm_irqchip_get_virq(s);
1196 route = g_malloc0(sizeof(KVMMSIRoute));
1197 route->kroute.gsi = virq;
1198 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1199 route->kroute.flags = 0;
1200 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1201 route->kroute.u.msi.address_hi = msg.address >> 32;
1202 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1204 kvm_add_routing_entry(s, &route->kroute);
1205 kvm_irqchip_commit_routes(s);
1207 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1211 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1213 return kvm_set_irq(s, route->kroute.gsi, 1);
1216 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1218 struct kvm_irq_routing_entry kroute = {};
1220 MSIMessage msg = {0, 0};
1222 if (pci_available && dev) {
1223 msg = pci_get_msi_message(dev, vector);
1226 if (kvm_gsi_direct_mapping()) {
1227 return kvm_arch_msi_data_to_gsi(msg.data);
1230 if (!kvm_gsi_routing_enabled()) {
1234 virq = kvm_irqchip_get_virq(s);
1240 kroute.type = KVM_IRQ_ROUTING_MSI;
1242 kroute.u.msi.address_lo = (uint32_t)msg.address;
1243 kroute.u.msi.address_hi = msg.address >> 32;
1244 kroute.u.msi.data = le32_to_cpu(msg.data);
1245 if (pci_available && kvm_msi_devid_required()) {
1246 kroute.flags = KVM_MSI_VALID_DEVID;
1247 kroute.u.msi.devid = pci_requester_id(dev);
1249 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1250 kvm_irqchip_release_virq(s, virq);
1254 trace_kvm_irqchip_add_msi_route(dev ? dev->name : (char *)"N/A",
1257 kvm_add_routing_entry(s, &kroute);
1258 kvm_arch_add_msi_route_post(&kroute, vector, dev);
1259 kvm_irqchip_commit_routes(s);
1264 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1267 struct kvm_irq_routing_entry kroute = {};
1269 if (kvm_gsi_direct_mapping()) {
1273 if (!kvm_irqchip_in_kernel()) {
1278 kroute.type = KVM_IRQ_ROUTING_MSI;
1280 kroute.u.msi.address_lo = (uint32_t)msg.address;
1281 kroute.u.msi.address_hi = msg.address >> 32;
1282 kroute.u.msi.data = le32_to_cpu(msg.data);
1283 if (pci_available && kvm_msi_devid_required()) {
1284 kroute.flags = KVM_MSI_VALID_DEVID;
1285 kroute.u.msi.devid = pci_requester_id(dev);
1287 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1291 trace_kvm_irqchip_update_msi_route(virq);
1293 return kvm_update_routing_entry(s, &kroute);
1296 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1299 struct kvm_irqfd irqfd = {
1302 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1306 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1307 irqfd.resamplefd = rfd;
1310 if (!kvm_irqfds_enabled()) {
1314 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1317 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1319 struct kvm_irq_routing_entry kroute = {};
1322 if (!kvm_gsi_routing_enabled()) {
1326 virq = kvm_irqchip_get_virq(s);
1332 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1334 kroute.u.adapter.summary_addr = adapter->summary_addr;
1335 kroute.u.adapter.ind_addr = adapter->ind_addr;
1336 kroute.u.adapter.summary_offset = adapter->summary_offset;
1337 kroute.u.adapter.ind_offset = adapter->ind_offset;
1338 kroute.u.adapter.adapter_id = adapter->adapter_id;
1340 kvm_add_routing_entry(s, &kroute);
1345 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1347 struct kvm_irq_routing_entry kroute = {};
1350 if (!kvm_gsi_routing_enabled()) {
1353 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1356 virq = kvm_irqchip_get_virq(s);
1362 kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1364 kroute.u.hv_sint.vcpu = vcpu;
1365 kroute.u.hv_sint.sint = sint;
1367 kvm_add_routing_entry(s, &kroute);
1368 kvm_irqchip_commit_routes(s);
1373 #else /* !KVM_CAP_IRQ_ROUTING */
1375 void kvm_init_irq_routing(KVMState *s)
1379 void kvm_irqchip_release_virq(KVMState *s, int virq)
1383 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1388 int kvm_irqchip_add_msi_route(KVMState *s, int vector, PCIDevice *dev)
1393 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1398 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1403 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1408 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1412 #endif /* !KVM_CAP_IRQ_ROUTING */
1414 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1415 EventNotifier *rn, int virq)
1417 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1418 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1421 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1424 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1428 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1429 EventNotifier *rn, qemu_irq irq)
1432 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1437 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1440 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1444 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1449 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1452 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1454 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1457 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1461 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1463 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1464 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1466 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1473 /* First probe and see if there's a arch-specific hook to create the
1474 * in-kernel irqchip for us */
1475 ret = kvm_arch_irqchip_create(machine, s);
1477 if (machine_kernel_irqchip_split(machine)) {
1478 perror("Split IRQ chip mode not supported.");
1481 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1485 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1489 kvm_kernel_irqchip = true;
1490 /* If we have an in-kernel IRQ chip then we must have asynchronous
1491 * interrupt delivery (though the reverse is not necessarily true)
1493 kvm_async_interrupts_allowed = true;
1494 kvm_halt_in_kernel_allowed = true;
1496 kvm_init_irq_routing(s);
1498 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1501 /* Find number of supported CPUs using the recommended
1502 * procedure from the kernel API documentation to cope with
1503 * older kernels that may be missing capabilities.
1505 static int kvm_recommended_vcpus(KVMState *s)
1507 int ret = kvm_vm_check_extension(s, KVM_CAP_NR_VCPUS);
1508 return (ret) ? ret : 4;
1511 static int kvm_max_vcpus(KVMState *s)
1513 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1514 return (ret) ? ret : kvm_recommended_vcpus(s);
1517 static int kvm_max_vcpu_id(KVMState *s)
1519 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPU_ID);
1520 return (ret) ? ret : kvm_max_vcpus(s);
1523 bool kvm_vcpu_id_is_valid(int vcpu_id)
1525 KVMState *s = KVM_STATE(current_machine->accelerator);
1526 return vcpu_id >= 0 && vcpu_id < kvm_max_vcpu_id(s);
1529 static int kvm_init(MachineState *ms)
1531 MachineClass *mc = MACHINE_GET_CLASS(ms);
1532 static const char upgrade_note[] =
1533 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1534 "(see http://sourceforge.net/projects/kvm).\n";
1539 { "SMP", smp_cpus },
1540 { "hotpluggable", max_cpus },
1543 int soft_vcpus_limit, hard_vcpus_limit;
1545 const KVMCapabilityInfo *missing_cap;
1548 const char *kvm_type;
1550 s = KVM_STATE(ms->accelerator);
1553 * On systems where the kernel can support different base page
1554 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1555 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1556 * page size for the system though.
1558 assert(TARGET_PAGE_SIZE <= getpagesize());
1562 #ifdef KVM_CAP_SET_GUEST_DEBUG
1563 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1565 QLIST_INIT(&s->kvm_parked_vcpus);
1567 s->fd = qemu_open("/dev/kvm", O_RDWR);
1569 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1574 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1575 if (ret < KVM_API_VERSION) {
1579 fprintf(stderr, "kvm version too old\n");
1583 if (ret > KVM_API_VERSION) {
1585 fprintf(stderr, "kvm version not supported\n");
1589 kvm_immediate_exit = kvm_check_extension(s, KVM_CAP_IMMEDIATE_EXIT);
1590 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1592 /* If unspecified, use the default value */
1597 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1599 type = mc->kvm_type(ms, kvm_type);
1600 } else if (kvm_type) {
1602 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1607 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1608 } while (ret == -EINTR);
1611 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1615 if (ret == -EINVAL) {
1617 "Host kernel setup problem detected. Please verify:\n");
1618 fprintf(stderr, "- for kernels supporting the switch_amode or"
1619 " user_mode parameters, whether\n");
1621 " user space is running in primary address space\n");
1623 "- for kernels supporting the vm.allocate_pgste sysctl, "
1624 "whether it is enabled\n");
1632 /* check the vcpu limits */
1633 soft_vcpus_limit = kvm_recommended_vcpus(s);
1634 hard_vcpus_limit = kvm_max_vcpus(s);
1637 if (nc->num > soft_vcpus_limit) {
1638 warn_report("Number of %s cpus requested (%d) exceeds "
1639 "the recommended cpus supported by KVM (%d)",
1640 nc->name, nc->num, soft_vcpus_limit);
1642 if (nc->num > hard_vcpus_limit) {
1643 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1644 "the maximum cpus supported by KVM (%d)\n",
1645 nc->name, nc->num, hard_vcpus_limit);
1652 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1655 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1659 fprintf(stderr, "kvm does not support %s\n%s",
1660 missing_cap->name, upgrade_note);
1664 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1665 s->coalesced_pio = s->coalesced_mmio &&
1666 kvm_check_extension(s, KVM_CAP_COALESCED_PIO);
1668 #ifdef KVM_CAP_VCPU_EVENTS
1669 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1672 s->robust_singlestep =
1673 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1675 #ifdef KVM_CAP_DEBUGREGS
1676 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1679 #ifdef KVM_CAP_IRQ_ROUTING
1680 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1683 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1685 s->irq_set_ioctl = KVM_IRQ_LINE;
1686 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1687 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1690 kvm_readonly_mem_allowed =
1691 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1693 kvm_eventfds_allowed =
1694 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1696 kvm_irqfds_allowed =
1697 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1699 kvm_resamplefds_allowed =
1700 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1702 kvm_vm_attributes_allowed =
1703 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1705 kvm_ioeventfd_any_length_allowed =
1706 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1711 * if memory encryption object is specified then initialize the memory
1712 * encryption context.
1714 if (ms->memory_encryption) {
1715 kvm_state->memcrypt_handle = sev_guest_init(ms->memory_encryption);
1716 if (!kvm_state->memcrypt_handle) {
1721 kvm_state->memcrypt_encrypt_data = sev_encrypt_data;
1724 ret = kvm_arch_init(ms, s);
1729 if (machine_kernel_irqchip_allowed(ms)) {
1730 kvm_irqchip_create(ms, s);
1733 if (kvm_eventfds_allowed) {
1734 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1735 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1737 s->memory_listener.listener.coalesced_io_add = kvm_coalesce_mmio_region;
1738 s->memory_listener.listener.coalesced_io_del = kvm_uncoalesce_mmio_region;
1740 kvm_memory_listener_register(s, &s->memory_listener,
1741 &address_space_memory, 0);
1742 memory_listener_register(&kvm_io_listener,
1744 memory_listener_register(&kvm_coalesced_pio_listener,
1747 s->many_ioeventfds = kvm_check_many_ioeventfds();
1749 s->sync_mmu = !!kvm_vm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1751 qemu_balloon_inhibit(true);
1764 g_free(s->memory_listener.slots);
1769 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1771 s->sigmask_len = sigmask_len;
1774 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1775 int size, uint32_t count)
1778 uint8_t *ptr = data;
1780 for (i = 0; i < count; i++) {
1781 address_space_rw(&address_space_io, port, attrs,
1783 direction == KVM_EXIT_IO_OUT);
1788 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1790 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1791 run->internal.suberror);
1793 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1796 for (i = 0; i < run->internal.ndata; ++i) {
1797 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1798 i, (uint64_t)run->internal.data[i]);
1801 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1802 fprintf(stderr, "emulation failure\n");
1803 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1804 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
1805 return EXCP_INTERRUPT;
1808 /* FIXME: Should trigger a qmp message to let management know
1809 * something went wrong.
1814 void kvm_flush_coalesced_mmio_buffer(void)
1816 KVMState *s = kvm_state;
1818 if (s->coalesced_flush_in_progress) {
1822 s->coalesced_flush_in_progress = true;
1824 if (s->coalesced_mmio_ring) {
1825 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1826 while (ring->first != ring->last) {
1827 struct kvm_coalesced_mmio *ent;
1829 ent = &ring->coalesced_mmio[ring->first];
1831 if (ent->pio == 1) {
1832 address_space_rw(&address_space_io, ent->phys_addr,
1833 MEMTXATTRS_UNSPECIFIED, ent->data,
1836 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1839 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1843 s->coalesced_flush_in_progress = false;
1846 static void do_kvm_cpu_synchronize_state(CPUState *cpu, run_on_cpu_data arg)
1848 if (!cpu->vcpu_dirty) {
1849 kvm_arch_get_registers(cpu);
1850 cpu->vcpu_dirty = true;
1854 void kvm_cpu_synchronize_state(CPUState *cpu)
1856 if (!cpu->vcpu_dirty) {
1857 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, RUN_ON_CPU_NULL);
1861 static void do_kvm_cpu_synchronize_post_reset(CPUState *cpu, run_on_cpu_data arg)
1863 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1864 cpu->vcpu_dirty = false;
1867 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1869 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, RUN_ON_CPU_NULL);
1872 static void do_kvm_cpu_synchronize_post_init(CPUState *cpu, run_on_cpu_data arg)
1874 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1875 cpu->vcpu_dirty = false;
1878 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1880 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, RUN_ON_CPU_NULL);
1883 static void do_kvm_cpu_synchronize_pre_loadvm(CPUState *cpu, run_on_cpu_data arg)
1885 cpu->vcpu_dirty = true;
1888 void kvm_cpu_synchronize_pre_loadvm(CPUState *cpu)
1890 run_on_cpu(cpu, do_kvm_cpu_synchronize_pre_loadvm, RUN_ON_CPU_NULL);
1893 #ifdef KVM_HAVE_MCE_INJECTION
1894 static __thread void *pending_sigbus_addr;
1895 static __thread int pending_sigbus_code;
1896 static __thread bool have_sigbus_pending;
1899 static void kvm_cpu_kick(CPUState *cpu)
1901 atomic_set(&cpu->kvm_run->immediate_exit, 1);
1904 static void kvm_cpu_kick_self(void)
1906 if (kvm_immediate_exit) {
1907 kvm_cpu_kick(current_cpu);
1909 qemu_cpu_kick_self();
1913 static void kvm_eat_signals(CPUState *cpu)
1915 struct timespec ts = { 0, 0 };
1921 if (kvm_immediate_exit) {
1922 atomic_set(&cpu->kvm_run->immediate_exit, 0);
1923 /* Write kvm_run->immediate_exit before the cpu->exit_request
1924 * write in kvm_cpu_exec.
1930 sigemptyset(&waitset);
1931 sigaddset(&waitset, SIG_IPI);
1934 r = sigtimedwait(&waitset, &siginfo, &ts);
1935 if (r == -1 && !(errno == EAGAIN || errno == EINTR)) {
1936 perror("sigtimedwait");
1940 r = sigpending(&chkset);
1942 perror("sigpending");
1945 } while (sigismember(&chkset, SIG_IPI));
1948 int kvm_cpu_exec(CPUState *cpu)
1950 struct kvm_run *run = cpu->kvm_run;
1953 DPRINTF("kvm_cpu_exec()\n");
1955 if (kvm_arch_process_async_events(cpu)) {
1956 atomic_set(&cpu->exit_request, 0);
1960 qemu_mutex_unlock_iothread();
1961 cpu_exec_start(cpu);
1966 if (cpu->vcpu_dirty) {
1967 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1968 cpu->vcpu_dirty = false;
1971 kvm_arch_pre_run(cpu, run);
1972 if (atomic_read(&cpu->exit_request)) {
1973 DPRINTF("interrupt exit requested\n");
1975 * KVM requires us to reenter the kernel after IO exits to complete
1976 * instruction emulation. This self-signal will ensure that we
1979 kvm_cpu_kick_self();
1982 /* Read cpu->exit_request before KVM_RUN reads run->immediate_exit.
1983 * Matching barrier in kvm_eat_signals.
1987 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1989 attrs = kvm_arch_post_run(cpu, run);
1991 #ifdef KVM_HAVE_MCE_INJECTION
1992 if (unlikely(have_sigbus_pending)) {
1993 qemu_mutex_lock_iothread();
1994 kvm_arch_on_sigbus_vcpu(cpu, pending_sigbus_code,
1995 pending_sigbus_addr);
1996 have_sigbus_pending = false;
1997 qemu_mutex_unlock_iothread();
2002 if (run_ret == -EINTR || run_ret == -EAGAIN) {
2003 DPRINTF("io window exit\n");
2004 kvm_eat_signals(cpu);
2005 ret = EXCP_INTERRUPT;
2008 fprintf(stderr, "error: kvm run failed %s\n",
2009 strerror(-run_ret));
2011 if (run_ret == -EBUSY) {
2013 "This is probably because your SMT is enabled.\n"
2014 "VCPU can only run on primary threads with all "
2015 "secondary threads offline.\n");
2022 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
2023 switch (run->exit_reason) {
2025 DPRINTF("handle_io\n");
2026 /* Called outside BQL */
2027 kvm_handle_io(run->io.port, attrs,
2028 (uint8_t *)run + run->io.data_offset,
2035 DPRINTF("handle_mmio\n");
2036 /* Called outside BQL */
2037 address_space_rw(&address_space_memory,
2038 run->mmio.phys_addr, attrs,
2041 run->mmio.is_write);
2044 case KVM_EXIT_IRQ_WINDOW_OPEN:
2045 DPRINTF("irq_window_open\n");
2046 ret = EXCP_INTERRUPT;
2048 case KVM_EXIT_SHUTDOWN:
2049 DPRINTF("shutdown\n");
2050 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2051 ret = EXCP_INTERRUPT;
2053 case KVM_EXIT_UNKNOWN:
2054 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
2055 (uint64_t)run->hw.hardware_exit_reason);
2058 case KVM_EXIT_INTERNAL_ERROR:
2059 ret = kvm_handle_internal_error(cpu, run);
2061 case KVM_EXIT_SYSTEM_EVENT:
2062 switch (run->system_event.type) {
2063 case KVM_SYSTEM_EVENT_SHUTDOWN:
2064 qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
2065 ret = EXCP_INTERRUPT;
2067 case KVM_SYSTEM_EVENT_RESET:
2068 qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
2069 ret = EXCP_INTERRUPT;
2071 case KVM_SYSTEM_EVENT_CRASH:
2072 kvm_cpu_synchronize_state(cpu);
2073 qemu_mutex_lock_iothread();
2074 qemu_system_guest_panicked(cpu_get_crash_info(cpu));
2075 qemu_mutex_unlock_iothread();
2079 DPRINTF("kvm_arch_handle_exit\n");
2080 ret = kvm_arch_handle_exit(cpu, run);
2085 DPRINTF("kvm_arch_handle_exit\n");
2086 ret = kvm_arch_handle_exit(cpu, run);
2092 qemu_mutex_lock_iothread();
2095 cpu_dump_state(cpu, stderr, CPU_DUMP_CODE);
2096 vm_stop(RUN_STATE_INTERNAL_ERROR);
2099 atomic_set(&cpu->exit_request, 0);
2103 int kvm_ioctl(KVMState *s, int type, ...)
2110 arg = va_arg(ap, void *);
2113 trace_kvm_ioctl(type, arg);
2114 ret = ioctl(s->fd, type, arg);
2121 int kvm_vm_ioctl(KVMState *s, int type, ...)
2128 arg = va_arg(ap, void *);
2131 trace_kvm_vm_ioctl(type, arg);
2132 ret = ioctl(s->vmfd, type, arg);
2139 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
2146 arg = va_arg(ap, void *);
2149 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
2150 ret = ioctl(cpu->kvm_fd, type, arg);
2157 int kvm_device_ioctl(int fd, int type, ...)
2164 arg = va_arg(ap, void *);
2167 trace_kvm_device_ioctl(fd, type, arg);
2168 ret = ioctl(fd, type, arg);
2175 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2178 struct kvm_device_attr attribute = {
2183 if (!kvm_vm_attributes_allowed) {
2187 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2188 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2192 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2194 struct kvm_device_attr attribute = {
2200 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2203 int kvm_device_access(int fd, int group, uint64_t attr,
2204 void *val, bool write, Error **errp)
2206 struct kvm_device_attr kvmattr;
2210 kvmattr.group = group;
2211 kvmattr.attr = attr;
2212 kvmattr.addr = (uintptr_t)val;
2214 err = kvm_device_ioctl(fd,
2215 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2218 error_setg_errno(errp, -err,
2219 "KVM_%s_DEVICE_ATTR failed: Group %d "
2220 "attr 0x%016" PRIx64,
2221 write ? "SET" : "GET", group, attr);
2226 bool kvm_has_sync_mmu(void)
2228 return kvm_state->sync_mmu;
2231 int kvm_has_vcpu_events(void)
2233 return kvm_state->vcpu_events;
2236 int kvm_has_robust_singlestep(void)
2238 return kvm_state->robust_singlestep;
2241 int kvm_has_debugregs(void)
2243 return kvm_state->debugregs;
2246 int kvm_has_many_ioeventfds(void)
2248 if (!kvm_enabled()) {
2251 return kvm_state->many_ioeventfds;
2254 int kvm_has_gsi_routing(void)
2256 #ifdef KVM_CAP_IRQ_ROUTING
2257 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2263 int kvm_has_intx_set_mask(void)
2265 return kvm_state->intx_set_mask;
2268 bool kvm_arm_supports_user_irq(void)
2270 return kvm_check_extension(kvm_state, KVM_CAP_ARM_USER_IRQ);
2273 #ifdef KVM_CAP_SET_GUEST_DEBUG
2274 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2277 struct kvm_sw_breakpoint *bp;
2279 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2287 int kvm_sw_breakpoints_active(CPUState *cpu)
2289 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2292 struct kvm_set_guest_debug_data {
2293 struct kvm_guest_debug dbg;
2297 static void kvm_invoke_set_guest_debug(CPUState *cpu, run_on_cpu_data data)
2299 struct kvm_set_guest_debug_data *dbg_data =
2300 (struct kvm_set_guest_debug_data *) data.host_ptr;
2302 dbg_data->err = kvm_vcpu_ioctl(cpu, KVM_SET_GUEST_DEBUG,
2306 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2308 struct kvm_set_guest_debug_data data;
2310 data.dbg.control = reinject_trap;
2312 if (cpu->singlestep_enabled) {
2313 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2315 kvm_arch_update_guest_debug(cpu, &data.dbg);
2317 run_on_cpu(cpu, kvm_invoke_set_guest_debug,
2318 RUN_ON_CPU_HOST_PTR(&data));
2322 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2323 target_ulong len, int type)
2325 struct kvm_sw_breakpoint *bp;
2328 if (type == GDB_BREAKPOINT_SW) {
2329 bp = kvm_find_sw_breakpoint(cpu, addr);
2335 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2338 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2344 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2346 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2353 err = kvm_update_guest_debug(cpu, 0);
2361 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2362 target_ulong len, int type)
2364 struct kvm_sw_breakpoint *bp;
2367 if (type == GDB_BREAKPOINT_SW) {
2368 bp = kvm_find_sw_breakpoint(cpu, addr);
2373 if (bp->use_count > 1) {
2378 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2383 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2386 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2393 err = kvm_update_guest_debug(cpu, 0);
2401 void kvm_remove_all_breakpoints(CPUState *cpu)
2403 struct kvm_sw_breakpoint *bp, *next;
2404 KVMState *s = cpu->kvm_state;
2407 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2408 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2409 /* Try harder to find a CPU that currently sees the breakpoint. */
2410 CPU_FOREACH(tmpcpu) {
2411 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2416 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2419 kvm_arch_remove_all_hw_breakpoints();
2422 kvm_update_guest_debug(cpu, 0);
2426 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2428 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2433 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2434 target_ulong len, int type)
2439 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2440 target_ulong len, int type)
2445 void kvm_remove_all_breakpoints(CPUState *cpu)
2448 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2450 static int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2452 KVMState *s = kvm_state;
2453 struct kvm_signal_mask *sigmask;
2456 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2458 sigmask->len = s->sigmask_len;
2459 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2460 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2466 static void kvm_ipi_signal(int sig)
2469 assert(kvm_immediate_exit);
2470 kvm_cpu_kick(current_cpu);
2474 void kvm_init_cpu_signals(CPUState *cpu)
2478 struct sigaction sigact;
2480 memset(&sigact, 0, sizeof(sigact));
2481 sigact.sa_handler = kvm_ipi_signal;
2482 sigaction(SIG_IPI, &sigact, NULL);
2484 pthread_sigmask(SIG_BLOCK, NULL, &set);
2485 #if defined KVM_HAVE_MCE_INJECTION
2486 sigdelset(&set, SIGBUS);
2487 pthread_sigmask(SIG_SETMASK, &set, NULL);
2489 sigdelset(&set, SIG_IPI);
2490 if (kvm_immediate_exit) {
2491 r = pthread_sigmask(SIG_SETMASK, &set, NULL);
2493 r = kvm_set_signal_mask(cpu, &set);
2496 fprintf(stderr, "kvm_set_signal_mask: %s\n", strerror(-r));
2501 /* Called asynchronously in VCPU thread. */
2502 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2504 #ifdef KVM_HAVE_MCE_INJECTION
2505 if (have_sigbus_pending) {
2508 have_sigbus_pending = true;
2509 pending_sigbus_addr = addr;
2510 pending_sigbus_code = code;
2511 atomic_set(&cpu->exit_request, 1);
2518 /* Called synchronously (via signalfd) in main thread. */
2519 int kvm_on_sigbus(int code, void *addr)
2521 #ifdef KVM_HAVE_MCE_INJECTION
2522 /* Action required MCE kills the process if SIGBUS is blocked. Because
2523 * that's what happens in the I/O thread, where we handle MCE via signalfd,
2524 * we can only get action optional here.
2526 assert(code != BUS_MCEERR_AR);
2527 kvm_arch_on_sigbus_vcpu(first_cpu, code, addr);
2534 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2537 struct kvm_create_device create_dev;
2539 create_dev.type = type;
2541 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2543 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2547 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2552 return test ? 0 : create_dev.fd;
2555 bool kvm_device_supported(int vmfd, uint64_t type)
2557 struct kvm_create_device create_dev = {
2560 .flags = KVM_CREATE_DEVICE_TEST,
2563 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2567 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2570 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2572 struct kvm_one_reg reg;
2576 reg.addr = (uintptr_t) source;
2577 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®);
2579 trace_kvm_failed_reg_set(id, strerror(-r));
2584 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2586 struct kvm_one_reg reg;
2590 reg.addr = (uintptr_t) target;
2591 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®);
2593 trace_kvm_failed_reg_get(id, strerror(-r));
2598 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2600 AccelClass *ac = ACCEL_CLASS(oc);
2602 ac->init_machine = kvm_init;
2603 ac->allowed = &kvm_allowed;
2606 static const TypeInfo kvm_accel_type = {
2607 .name = TYPE_KVM_ACCEL,
2608 .parent = TYPE_ACCEL,
2609 .class_init = kvm_accel_class_init,
2610 .instance_size = sizeof(KVMState),
2613 static void kvm_type_init(void)
2615 type_register_static(&kvm_accel_type);
2618 type_init(kvm_type_init);