4 * Copyright IBM, Corp. 2008
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Glauber Costa <gcosta@redhat.com>
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
16 #include <sys/types.h>
17 #include <sys/ioctl.h>
21 #include <linux/kvm.h>
23 #include "qemu-common.h"
24 #include "qemu/atomic.h"
25 #include "qemu/option.h"
26 #include "qemu/config-file.h"
28 #include "hw/pci/msi.h"
29 #include "hw/s390x/adapter.h"
30 #include "exec/gdbstub.h"
31 #include "sysemu/kvm_int.h"
32 #include "qemu/bswap.h"
33 #include "exec/memory.h"
34 #include "exec/ram_addr.h"
35 #include "exec/address-spaces.h"
36 #include "qemu/event_notifier.h"
40 #include "hw/boards.h"
42 /* This check must be after config-host.h is included */
44 #include <sys/eventfd.h>
47 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
48 #define PAGE_SIZE TARGET_PAGE_SIZE
53 #define DPRINTF(fmt, ...) \
54 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
56 #define DPRINTF(fmt, ...) \
60 #define KVM_MSI_HASHTAB_SIZE 256
64 AccelState parent_obj;
70 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
71 bool coalesced_flush_in_progress;
72 int broken_set_mem_region;
74 int robust_singlestep;
76 #ifdef KVM_CAP_SET_GUEST_DEBUG
77 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
83 /* The man page (and posix) say ioctl numbers are signed int, but
84 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
85 * unsigned, and treating them as signed here can break things */
86 unsigned irq_set_ioctl;
87 unsigned int sigmask_len;
89 #ifdef KVM_CAP_IRQ_ROUTING
90 struct kvm_irq_routing *irq_routes;
91 int nr_allocated_irq_routes;
92 uint32_t *used_gsi_bitmap;
93 unsigned int gsi_count;
94 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
97 KVMMemoryListener memory_listener;
101 bool kvm_kernel_irqchip;
102 bool kvm_async_interrupts_allowed;
103 bool kvm_halt_in_kernel_allowed;
104 bool kvm_eventfds_allowed;
105 bool kvm_irqfds_allowed;
106 bool kvm_resamplefds_allowed;
107 bool kvm_msi_via_irqfd_allowed;
108 bool kvm_gsi_routing_allowed;
109 bool kvm_gsi_direct_mapping;
111 bool kvm_readonly_mem_allowed;
112 bool kvm_vm_attributes_allowed;
114 static const KVMCapabilityInfo kvm_required_capabilites[] = {
115 KVM_CAP_INFO(USER_MEMORY),
116 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
120 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
122 KVMState *s = kvm_state;
125 for (i = 0; i < s->nr_slots; i++) {
126 if (kml->slots[i].memory_size == 0) {
127 return &kml->slots[i];
134 bool kvm_has_free_slot(MachineState *ms)
136 KVMState *s = KVM_STATE(ms->accelerator);
138 return kvm_get_free_slot(&s->memory_listener);
141 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
143 KVMSlot *slot = kvm_get_free_slot(kml);
149 fprintf(stderr, "%s: no free slot available\n", __func__);
153 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
157 KVMState *s = kvm_state;
160 for (i = 0; i < s->nr_slots; i++) {
161 KVMSlot *mem = &kml->slots[i];
163 if (start_addr == mem->start_addr &&
164 end_addr == mem->start_addr + mem->memory_size) {
173 * Find overlapping slot with lowest start address
175 static KVMSlot *kvm_lookup_overlapping_slot(KVMMemoryListener *kml,
179 KVMState *s = kvm_state;
180 KVMSlot *found = NULL;
183 for (i = 0; i < s->nr_slots; i++) {
184 KVMSlot *mem = &kml->slots[i];
186 if (mem->memory_size == 0 ||
187 (found && found->start_addr < mem->start_addr)) {
191 if (end_addr > mem->start_addr &&
192 start_addr < mem->start_addr + mem->memory_size) {
200 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
203 KVMMemoryListener *kml = &s->memory_listener;
206 for (i = 0; i < s->nr_slots; i++) {
207 KVMSlot *mem = &kml->slots[i];
209 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
210 *phys_addr = mem->start_addr + (ram - mem->ram);
218 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot)
220 KVMState *s = kvm_state;
221 struct kvm_userspace_memory_region mem;
223 mem.slot = slot->slot | (kml->as_id << 16);
224 mem.guest_phys_addr = slot->start_addr;
225 mem.userspace_addr = (unsigned long)slot->ram;
226 mem.flags = slot->flags;
228 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
229 /* Set the slot size to 0 before setting the slot to the desired
230 * value. This is needed based on KVM commit 75d61fbc. */
232 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
234 mem.memory_size = slot->memory_size;
235 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
238 int kvm_init_vcpu(CPUState *cpu)
240 KVMState *s = kvm_state;
244 DPRINTF("kvm_init_vcpu\n");
246 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
248 DPRINTF("kvm_create_vcpu failed\n");
254 cpu->kvm_vcpu_dirty = true;
256 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
259 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
263 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
265 if (cpu->kvm_run == MAP_FAILED) {
267 DPRINTF("mmap'ing vcpu state failed\n");
271 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
272 s->coalesced_mmio_ring =
273 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
276 ret = kvm_arch_init_vcpu(cpu);
282 * dirty pages logging control
285 static int kvm_mem_flags(MemoryRegion *mr)
287 bool readonly = mr->readonly || memory_region_is_romd(mr);
290 if (memory_region_get_dirty_log_mask(mr) != 0) {
291 flags |= KVM_MEM_LOG_DIRTY_PAGES;
293 if (readonly && kvm_readonly_mem_allowed) {
294 flags |= KVM_MEM_READONLY;
299 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
304 old_flags = mem->flags;
305 mem->flags = kvm_mem_flags(mr);
307 /* If nothing changed effectively, no need to issue ioctl */
308 if (mem->flags == old_flags) {
312 return kvm_set_user_memory_region(kml, mem);
315 static int kvm_section_update_flags(KVMMemoryListener *kml,
316 MemoryRegionSection *section)
318 hwaddr phys_addr = section->offset_within_address_space;
319 ram_addr_t size = int128_get64(section->size);
320 KVMSlot *mem = kvm_lookup_matching_slot(kml, phys_addr, phys_addr + size);
325 return kvm_slot_update_flags(kml, mem, section->mr);
329 static void kvm_log_start(MemoryListener *listener,
330 MemoryRegionSection *section,
333 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
340 r = kvm_section_update_flags(kml, section);
346 static void kvm_log_stop(MemoryListener *listener,
347 MemoryRegionSection *section,
350 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
357 r = kvm_section_update_flags(kml, section);
363 /* get kvm's dirty pages bitmap and update qemu's */
364 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
365 unsigned long *bitmap)
367 ram_addr_t start = section->offset_within_region + section->mr->ram_addr;
368 ram_addr_t pages = int128_get64(section->size) / getpagesize();
370 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
374 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
377 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
378 * This function updates qemu's dirty bitmap using
379 * memory_region_set_dirty(). This means all bits are set
382 * @start_add: start of logged region.
383 * @end_addr: end of logged region.
385 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
386 MemoryRegionSection *section)
388 KVMState *s = kvm_state;
389 unsigned long size, allocated_size = 0;
390 struct kvm_dirty_log d = {};
393 hwaddr start_addr = section->offset_within_address_space;
394 hwaddr end_addr = start_addr + int128_get64(section->size);
396 d.dirty_bitmap = NULL;
397 while (start_addr < end_addr) {
398 mem = kvm_lookup_overlapping_slot(kml, start_addr, end_addr);
403 /* XXX bad kernel interface alert
404 * For dirty bitmap, kernel allocates array of size aligned to
405 * bits-per-long. But for case when the kernel is 64bits and
406 * the userspace is 32bits, userspace can't align to the same
407 * bits-per-long, since sizeof(long) is different between kernel
408 * and user space. This way, userspace will provide buffer which
409 * may be 4 bytes less than the kernel will use, resulting in
410 * userspace memory corruption (which is not detectable by valgrind
411 * too, in most cases).
412 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
413 * a hope that sizeof(long) wont become >8 any time soon.
415 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
416 /*HOST_LONG_BITS*/ 64) / 8;
417 if (!d.dirty_bitmap) {
418 d.dirty_bitmap = g_malloc(size);
419 } else if (size > allocated_size) {
420 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
422 allocated_size = size;
423 memset(d.dirty_bitmap, 0, allocated_size);
425 d.slot = mem->slot | (kml->as_id << 16);
426 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
427 DPRINTF("ioctl failed %d\n", errno);
432 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
433 start_addr = mem->start_addr + mem->memory_size;
435 g_free(d.dirty_bitmap);
440 static void kvm_coalesce_mmio_region(MemoryListener *listener,
441 MemoryRegionSection *secion,
442 hwaddr start, hwaddr size)
444 KVMState *s = kvm_state;
446 if (s->coalesced_mmio) {
447 struct kvm_coalesced_mmio_zone zone;
453 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
457 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
458 MemoryRegionSection *secion,
459 hwaddr start, hwaddr size)
461 KVMState *s = kvm_state;
463 if (s->coalesced_mmio) {
464 struct kvm_coalesced_mmio_zone zone;
470 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
474 int kvm_check_extension(KVMState *s, unsigned int extension)
478 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
486 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
490 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
492 /* VM wide version not implemented, use global one instead */
493 ret = kvm_check_extension(s, extension);
499 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
501 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
502 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
503 * endianness, but the memory core hands them in target endianness.
504 * For example, PPC is always treated as big-endian even if running
505 * on KVM and on PPC64LE. Correct here.
519 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
520 bool assign, uint32_t size, bool datamatch)
523 struct kvm_ioeventfd iofd = {
524 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
531 if (!kvm_enabled()) {
536 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
539 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
542 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
551 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
552 bool assign, uint32_t size, bool datamatch)
554 struct kvm_ioeventfd kick = {
555 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
557 .flags = KVM_IOEVENTFD_FLAG_PIO,
562 if (!kvm_enabled()) {
566 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
569 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
571 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
579 static int kvm_check_many_ioeventfds(void)
581 /* Userspace can use ioeventfd for io notification. This requires a host
582 * that supports eventfd(2) and an I/O thread; since eventfd does not
583 * support SIGIO it cannot interrupt the vcpu.
585 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
586 * can avoid creating too many ioeventfds.
588 #if defined(CONFIG_EVENTFD)
591 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
592 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
593 if (ioeventfds[i] < 0) {
596 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
598 close(ioeventfds[i]);
603 /* Decide whether many devices are supported or not */
604 ret = i == ARRAY_SIZE(ioeventfds);
607 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
608 close(ioeventfds[i]);
616 static const KVMCapabilityInfo *
617 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
620 if (!kvm_check_extension(s, list->value)) {
628 static void kvm_set_phys_mem(KVMMemoryListener *kml,
629 MemoryRegionSection *section, bool add)
631 KVMState *s = kvm_state;
634 MemoryRegion *mr = section->mr;
635 bool writeable = !mr->readonly && !mr->rom_device;
636 hwaddr start_addr = section->offset_within_address_space;
637 ram_addr_t size = int128_get64(section->size);
641 /* kvm works in page size chunks, but the function may be called
642 with sub-page size and unaligned start address. Pad the start
643 address to next and truncate size to previous page boundary. */
644 delta = (TARGET_PAGE_SIZE - (start_addr & ~TARGET_PAGE_MASK));
645 delta &= ~TARGET_PAGE_MASK;
651 size &= TARGET_PAGE_MASK;
652 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
656 if (!memory_region_is_ram(mr)) {
657 if (writeable || !kvm_readonly_mem_allowed) {
659 } else if (!mr->romd_mode) {
660 /* If the memory device is not in romd_mode, then we actually want
661 * to remove the kvm memory slot so all accesses will trap. */
666 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
669 mem = kvm_lookup_overlapping_slot(kml, start_addr, start_addr + size);
674 if (add && start_addr >= mem->start_addr &&
675 (start_addr + size <= mem->start_addr + mem->memory_size) &&
676 (ram - start_addr == mem->ram - mem->start_addr)) {
677 /* The new slot fits into the existing one and comes with
678 * identical parameters - update flags and done. */
679 kvm_slot_update_flags(kml, mem, mr);
685 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
686 kvm_physical_sync_dirty_bitmap(kml, section);
689 /* unregister the overlapping slot */
690 mem->memory_size = 0;
691 err = kvm_set_user_memory_region(kml, mem);
693 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
694 __func__, strerror(-err));
698 /* Workaround for older KVM versions: we can't join slots, even not by
699 * unregistering the previous ones and then registering the larger
700 * slot. We have to maintain the existing fragmentation. Sigh.
702 * This workaround assumes that the new slot starts at the same
703 * address as the first existing one. If not or if some overlapping
704 * slot comes around later, we will fail (not seen in practice so far)
705 * - and actually require a recent KVM version. */
706 if (s->broken_set_mem_region &&
707 old.start_addr == start_addr && old.memory_size < size && add) {
708 mem = kvm_alloc_slot(kml);
709 mem->memory_size = old.memory_size;
710 mem->start_addr = old.start_addr;
712 mem->flags = kvm_mem_flags(mr);
714 err = kvm_set_user_memory_region(kml, mem);
716 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
721 start_addr += old.memory_size;
722 ram += old.memory_size;
723 size -= old.memory_size;
727 /* register prefix slot */
728 if (old.start_addr < start_addr) {
729 mem = kvm_alloc_slot(kml);
730 mem->memory_size = start_addr - old.start_addr;
731 mem->start_addr = old.start_addr;
733 mem->flags = kvm_mem_flags(mr);
735 err = kvm_set_user_memory_region(kml, mem);
737 fprintf(stderr, "%s: error registering prefix slot: %s\n",
738 __func__, strerror(-err));
740 fprintf(stderr, "%s: This is probably because your kernel's " \
741 "PAGE_SIZE is too big. Please try to use 4k " \
742 "PAGE_SIZE!\n", __func__);
748 /* register suffix slot */
749 if (old.start_addr + old.memory_size > start_addr + size) {
750 ram_addr_t size_delta;
752 mem = kvm_alloc_slot(kml);
753 mem->start_addr = start_addr + size;
754 size_delta = mem->start_addr - old.start_addr;
755 mem->memory_size = old.memory_size - size_delta;
756 mem->ram = old.ram + size_delta;
757 mem->flags = kvm_mem_flags(mr);
759 err = kvm_set_user_memory_region(kml, mem);
761 fprintf(stderr, "%s: error registering suffix slot: %s\n",
762 __func__, strerror(-err));
768 /* in case the KVM bug workaround already "consumed" the new slot */
775 mem = kvm_alloc_slot(kml);
776 mem->memory_size = size;
777 mem->start_addr = start_addr;
779 mem->flags = kvm_mem_flags(mr);
781 err = kvm_set_user_memory_region(kml, mem);
783 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
789 static void kvm_region_add(MemoryListener *listener,
790 MemoryRegionSection *section)
792 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
794 memory_region_ref(section->mr);
795 kvm_set_phys_mem(kml, section, true);
798 static void kvm_region_del(MemoryListener *listener,
799 MemoryRegionSection *section)
801 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
803 kvm_set_phys_mem(kml, section, false);
804 memory_region_unref(section->mr);
807 static void kvm_log_sync(MemoryListener *listener,
808 MemoryRegionSection *section)
810 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
813 r = kvm_physical_sync_dirty_bitmap(kml, section);
819 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
820 MemoryRegionSection *section,
821 bool match_data, uint64_t data,
824 int fd = event_notifier_get_fd(e);
827 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
828 data, true, int128_get64(section->size),
831 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
832 __func__, strerror(-r));
837 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
838 MemoryRegionSection *section,
839 bool match_data, uint64_t data,
842 int fd = event_notifier_get_fd(e);
845 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
846 data, false, int128_get64(section->size),
853 static void kvm_io_ioeventfd_add(MemoryListener *listener,
854 MemoryRegionSection *section,
855 bool match_data, uint64_t data,
858 int fd = event_notifier_get_fd(e);
861 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
862 data, true, int128_get64(section->size),
865 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
866 __func__, strerror(-r));
871 static void kvm_io_ioeventfd_del(MemoryListener *listener,
872 MemoryRegionSection *section,
873 bool match_data, uint64_t data,
877 int fd = event_notifier_get_fd(e);
880 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
881 data, false, int128_get64(section->size),
888 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
889 AddressSpace *as, int as_id)
893 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
896 for (i = 0; i < s->nr_slots; i++) {
897 kml->slots[i].slot = i;
900 kml->listener.region_add = kvm_region_add;
901 kml->listener.region_del = kvm_region_del;
902 kml->listener.log_start = kvm_log_start;
903 kml->listener.log_stop = kvm_log_stop;
904 kml->listener.log_sync = kvm_log_sync;
905 kml->listener.priority = 10;
907 memory_listener_register(&kml->listener, as);
910 static MemoryListener kvm_io_listener = {
911 .eventfd_add = kvm_io_ioeventfd_add,
912 .eventfd_del = kvm_io_ioeventfd_del,
916 static void kvm_handle_interrupt(CPUState *cpu, int mask)
918 cpu->interrupt_request |= mask;
920 if (!qemu_cpu_is_self(cpu)) {
925 int kvm_set_irq(KVMState *s, int irq, int level)
927 struct kvm_irq_level event;
930 assert(kvm_async_interrupts_enabled());
934 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
936 perror("kvm_set_irq");
940 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
943 #ifdef KVM_CAP_IRQ_ROUTING
944 typedef struct KVMMSIRoute {
945 struct kvm_irq_routing_entry kroute;
946 QTAILQ_ENTRY(KVMMSIRoute) entry;
949 static void set_gsi(KVMState *s, unsigned int gsi)
951 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
954 static void clear_gsi(KVMState *s, unsigned int gsi)
956 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
959 void kvm_init_irq_routing(KVMState *s)
963 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
965 unsigned int gsi_bits, i;
967 /* Round up so we can search ints using ffs */
968 gsi_bits = ALIGN(gsi_count, 32);
969 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
970 s->gsi_count = gsi_count;
972 /* Mark any over-allocated bits as already in use */
973 for (i = gsi_count; i < gsi_bits; i++) {
978 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
979 s->nr_allocated_irq_routes = 0;
981 if (!s->direct_msi) {
982 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
983 QTAILQ_INIT(&s->msi_hashtab[i]);
987 kvm_arch_init_irq_routing(s);
990 void kvm_irqchip_commit_routes(KVMState *s)
994 s->irq_routes->flags = 0;
995 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
999 static void kvm_add_routing_entry(KVMState *s,
1000 struct kvm_irq_routing_entry *entry)
1002 struct kvm_irq_routing_entry *new;
1005 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1006 n = s->nr_allocated_irq_routes * 2;
1010 size = sizeof(struct kvm_irq_routing);
1011 size += n * sizeof(*new);
1012 s->irq_routes = g_realloc(s->irq_routes, size);
1013 s->nr_allocated_irq_routes = n;
1015 n = s->irq_routes->nr++;
1016 new = &s->irq_routes->entries[n];
1020 set_gsi(s, entry->gsi);
1023 static int kvm_update_routing_entry(KVMState *s,
1024 struct kvm_irq_routing_entry *new_entry)
1026 struct kvm_irq_routing_entry *entry;
1029 for (n = 0; n < s->irq_routes->nr; n++) {
1030 entry = &s->irq_routes->entries[n];
1031 if (entry->gsi != new_entry->gsi) {
1035 if(!memcmp(entry, new_entry, sizeof *entry)) {
1039 *entry = *new_entry;
1041 kvm_irqchip_commit_routes(s);
1049 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1051 struct kvm_irq_routing_entry e = {};
1053 assert(pin < s->gsi_count);
1056 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1058 e.u.irqchip.irqchip = irqchip;
1059 e.u.irqchip.pin = pin;
1060 kvm_add_routing_entry(s, &e);
1063 void kvm_irqchip_release_virq(KVMState *s, int virq)
1065 struct kvm_irq_routing_entry *e;
1068 if (kvm_gsi_direct_mapping()) {
1072 for (i = 0; i < s->irq_routes->nr; i++) {
1073 e = &s->irq_routes->entries[i];
1074 if (e->gsi == virq) {
1075 s->irq_routes->nr--;
1076 *e = s->irq_routes->entries[s->irq_routes->nr];
1082 static unsigned int kvm_hash_msi(uint32_t data)
1084 /* This is optimized for IA32 MSI layout. However, no other arch shall
1085 * repeat the mistake of not providing a direct MSI injection API. */
1089 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1091 KVMMSIRoute *route, *next;
1094 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1095 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1096 kvm_irqchip_release_virq(s, route->kroute.gsi);
1097 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1103 static int kvm_irqchip_get_virq(KVMState *s)
1105 uint32_t *word = s->used_gsi_bitmap;
1106 int max_words = ALIGN(s->gsi_count, 32) / 32;
1110 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1111 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1112 * number can succeed even though a new route entry cannot be added.
1113 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1115 if (!s->direct_msi && s->irq_routes->nr == s->gsi_count) {
1116 kvm_flush_dynamic_msi_routes(s);
1119 /* Return the lowest unused GSI in the bitmap */
1120 for (i = 0; i < max_words; i++) {
1121 zeroes = ctz32(~word[i]);
1126 return zeroes + i * 32;
1132 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1134 unsigned int hash = kvm_hash_msi(msg.data);
1137 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1138 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1139 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1140 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1147 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1152 if (s->direct_msi) {
1153 msi.address_lo = (uint32_t)msg.address;
1154 msi.address_hi = msg.address >> 32;
1155 msi.data = le32_to_cpu(msg.data);
1157 memset(msi.pad, 0, sizeof(msi.pad));
1159 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1162 route = kvm_lookup_msi_route(s, msg);
1166 virq = kvm_irqchip_get_virq(s);
1171 route = g_malloc0(sizeof(KVMMSIRoute));
1172 route->kroute.gsi = virq;
1173 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1174 route->kroute.flags = 0;
1175 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1176 route->kroute.u.msi.address_hi = msg.address >> 32;
1177 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1179 kvm_add_routing_entry(s, &route->kroute);
1180 kvm_irqchip_commit_routes(s);
1182 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1186 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1188 return kvm_set_irq(s, route->kroute.gsi, 1);
1191 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1193 struct kvm_irq_routing_entry kroute = {};
1196 if (kvm_gsi_direct_mapping()) {
1197 return kvm_arch_msi_data_to_gsi(msg.data);
1200 if (!kvm_gsi_routing_enabled()) {
1204 virq = kvm_irqchip_get_virq(s);
1210 kroute.type = KVM_IRQ_ROUTING_MSI;
1212 kroute.u.msi.address_lo = (uint32_t)msg.address;
1213 kroute.u.msi.address_hi = msg.address >> 32;
1214 kroute.u.msi.data = le32_to_cpu(msg.data);
1215 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1216 kvm_irqchip_release_virq(s, virq);
1220 kvm_add_routing_entry(s, &kroute);
1221 kvm_irqchip_commit_routes(s);
1226 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1228 struct kvm_irq_routing_entry kroute = {};
1230 if (kvm_gsi_direct_mapping()) {
1234 if (!kvm_irqchip_in_kernel()) {
1239 kroute.type = KVM_IRQ_ROUTING_MSI;
1241 kroute.u.msi.address_lo = (uint32_t)msg.address;
1242 kroute.u.msi.address_hi = msg.address >> 32;
1243 kroute.u.msi.data = le32_to_cpu(msg.data);
1244 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1248 return kvm_update_routing_entry(s, &kroute);
1251 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1254 struct kvm_irqfd irqfd = {
1257 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1261 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1262 irqfd.resamplefd = rfd;
1265 if (!kvm_irqfds_enabled()) {
1269 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1272 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1274 struct kvm_irq_routing_entry kroute = {};
1277 if (!kvm_gsi_routing_enabled()) {
1281 virq = kvm_irqchip_get_virq(s);
1287 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1289 kroute.u.adapter.summary_addr = adapter->summary_addr;
1290 kroute.u.adapter.ind_addr = adapter->ind_addr;
1291 kroute.u.adapter.summary_offset = adapter->summary_offset;
1292 kroute.u.adapter.ind_offset = adapter->ind_offset;
1293 kroute.u.adapter.adapter_id = adapter->adapter_id;
1295 kvm_add_routing_entry(s, &kroute);
1300 #else /* !KVM_CAP_IRQ_ROUTING */
1302 void kvm_init_irq_routing(KVMState *s)
1306 void kvm_irqchip_release_virq(KVMState *s, int virq)
1310 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1315 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1320 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1325 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1330 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1334 #endif /* !KVM_CAP_IRQ_ROUTING */
1336 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1337 EventNotifier *rn, int virq)
1339 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1340 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1343 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1346 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1350 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1351 EventNotifier *rn, qemu_irq irq)
1354 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1359 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1362 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1366 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1371 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1374 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1376 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1379 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1383 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1385 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1386 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1388 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1395 /* First probe and see if there's a arch-specific hook to create the
1396 * in-kernel irqchip for us */
1397 ret = kvm_arch_irqchip_create(s);
1399 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1402 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1406 kvm_kernel_irqchip = true;
1407 /* If we have an in-kernel IRQ chip then we must have asynchronous
1408 * interrupt delivery (though the reverse is not necessarily true)
1410 kvm_async_interrupts_allowed = true;
1411 kvm_halt_in_kernel_allowed = true;
1413 kvm_init_irq_routing(s);
1415 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1418 /* Find number of supported CPUs using the recommended
1419 * procedure from the kernel API documentation to cope with
1420 * older kernels that may be missing capabilities.
1422 static int kvm_recommended_vcpus(KVMState *s)
1424 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1425 return (ret) ? ret : 4;
1428 static int kvm_max_vcpus(KVMState *s)
1430 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1431 return (ret) ? ret : kvm_recommended_vcpus(s);
1434 static int kvm_init(MachineState *ms)
1436 MachineClass *mc = MACHINE_GET_CLASS(ms);
1437 static const char upgrade_note[] =
1438 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1439 "(see http://sourceforge.net/projects/kvm).\n";
1444 { "SMP", smp_cpus },
1445 { "hotpluggable", max_cpus },
1448 int soft_vcpus_limit, hard_vcpus_limit;
1450 const KVMCapabilityInfo *missing_cap;
1453 const char *kvm_type;
1455 s = KVM_STATE(ms->accelerator);
1458 * On systems where the kernel can support different base page
1459 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1460 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1461 * page size for the system though.
1463 assert(TARGET_PAGE_SIZE <= getpagesize());
1468 #ifdef KVM_CAP_SET_GUEST_DEBUG
1469 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1472 s->fd = qemu_open("/dev/kvm", O_RDWR);
1474 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1479 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1480 if (ret < KVM_API_VERSION) {
1484 fprintf(stderr, "kvm version too old\n");
1488 if (ret > KVM_API_VERSION) {
1490 fprintf(stderr, "kvm version not supported\n");
1494 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1496 /* If unspecified, use the default value */
1501 /* check the vcpu limits */
1502 soft_vcpus_limit = kvm_recommended_vcpus(s);
1503 hard_vcpus_limit = kvm_max_vcpus(s);
1506 if (nc->num > soft_vcpus_limit) {
1508 "Warning: Number of %s cpus requested (%d) exceeds "
1509 "the recommended cpus supported by KVM (%d)\n",
1510 nc->name, nc->num, soft_vcpus_limit);
1512 if (nc->num > hard_vcpus_limit) {
1513 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1514 "the maximum cpus supported by KVM (%d)\n",
1515 nc->name, nc->num, hard_vcpus_limit);
1522 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1524 type = mc->kvm_type(kvm_type);
1525 } else if (kvm_type) {
1527 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1532 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1533 } while (ret == -EINTR);
1536 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1540 if (ret == -EINVAL) {
1542 "Host kernel setup problem detected. Please verify:\n");
1543 fprintf(stderr, "- for kernels supporting the switch_amode or"
1544 " user_mode parameters, whether\n");
1546 " user space is running in primary address space\n");
1548 "- for kernels supporting the vm.allocate_pgste sysctl, "
1549 "whether it is enabled\n");
1556 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1559 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1563 fprintf(stderr, "kvm does not support %s\n%s",
1564 missing_cap->name, upgrade_note);
1568 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1570 s->broken_set_mem_region = 1;
1571 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1573 s->broken_set_mem_region = 0;
1576 #ifdef KVM_CAP_VCPU_EVENTS
1577 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1580 s->robust_singlestep =
1581 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1583 #ifdef KVM_CAP_DEBUGREGS
1584 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1587 #ifdef KVM_CAP_XSAVE
1588 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1592 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1595 #ifdef KVM_CAP_PIT_STATE2
1596 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1599 #ifdef KVM_CAP_IRQ_ROUTING
1600 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1603 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1605 s->irq_set_ioctl = KVM_IRQ_LINE;
1606 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1607 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1610 #ifdef KVM_CAP_READONLY_MEM
1611 kvm_readonly_mem_allowed =
1612 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1615 kvm_eventfds_allowed =
1616 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1618 kvm_irqfds_allowed =
1619 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1621 kvm_resamplefds_allowed =
1622 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1624 kvm_vm_attributes_allowed =
1625 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1627 ret = kvm_arch_init(ms, s);
1632 if (machine_kernel_irqchip_allowed(ms)) {
1633 kvm_irqchip_create(ms, s);
1638 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1639 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1640 s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1641 s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1643 kvm_memory_listener_register(s, &s->memory_listener,
1644 &address_space_memory, 0);
1645 memory_listener_register(&kvm_io_listener,
1648 s->many_ioeventfds = kvm_check_many_ioeventfds();
1650 cpu_interrupt_handler = kvm_handle_interrupt;
1662 g_free(s->memory_listener.slots);
1667 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1669 s->sigmask_len = sigmask_len;
1672 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1673 int size, uint32_t count)
1676 uint8_t *ptr = data;
1678 for (i = 0; i < count; i++) {
1679 address_space_rw(&address_space_io, port, attrs,
1681 direction == KVM_EXIT_IO_OUT);
1686 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1688 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1689 run->internal.suberror);
1691 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1694 for (i = 0; i < run->internal.ndata; ++i) {
1695 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1696 i, (uint64_t)run->internal.data[i]);
1699 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1700 fprintf(stderr, "emulation failure\n");
1701 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1702 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1703 return EXCP_INTERRUPT;
1706 /* FIXME: Should trigger a qmp message to let management know
1707 * something went wrong.
1712 void kvm_flush_coalesced_mmio_buffer(void)
1714 KVMState *s = kvm_state;
1716 if (s->coalesced_flush_in_progress) {
1720 s->coalesced_flush_in_progress = true;
1722 if (s->coalesced_mmio_ring) {
1723 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1724 while (ring->first != ring->last) {
1725 struct kvm_coalesced_mmio *ent;
1727 ent = &ring->coalesced_mmio[ring->first];
1729 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1731 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1735 s->coalesced_flush_in_progress = false;
1738 static void do_kvm_cpu_synchronize_state(void *arg)
1740 CPUState *cpu = arg;
1742 if (!cpu->kvm_vcpu_dirty) {
1743 kvm_arch_get_registers(cpu);
1744 cpu->kvm_vcpu_dirty = true;
1748 void kvm_cpu_synchronize_state(CPUState *cpu)
1750 if (!cpu->kvm_vcpu_dirty) {
1751 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1755 static void do_kvm_cpu_synchronize_post_reset(void *arg)
1757 CPUState *cpu = arg;
1759 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1760 cpu->kvm_vcpu_dirty = false;
1763 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1765 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, cpu);
1768 static void do_kvm_cpu_synchronize_post_init(void *arg)
1770 CPUState *cpu = arg;
1772 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1773 cpu->kvm_vcpu_dirty = false;
1776 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1778 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, cpu);
1781 void kvm_cpu_clean_state(CPUState *cpu)
1783 cpu->kvm_vcpu_dirty = false;
1786 int kvm_cpu_exec(CPUState *cpu)
1788 struct kvm_run *run = cpu->kvm_run;
1791 DPRINTF("kvm_cpu_exec()\n");
1793 if (kvm_arch_process_async_events(cpu)) {
1794 cpu->exit_request = 0;
1798 qemu_mutex_unlock_iothread();
1803 if (cpu->kvm_vcpu_dirty) {
1804 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1805 cpu->kvm_vcpu_dirty = false;
1808 kvm_arch_pre_run(cpu, run);
1809 if (cpu->exit_request) {
1810 DPRINTF("interrupt exit requested\n");
1812 * KVM requires us to reenter the kernel after IO exits to complete
1813 * instruction emulation. This self-signal will ensure that we
1816 qemu_cpu_kick_self();
1819 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1821 attrs = kvm_arch_post_run(cpu, run);
1824 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1825 DPRINTF("io window exit\n");
1826 ret = EXCP_INTERRUPT;
1829 fprintf(stderr, "error: kvm run failed %s\n",
1830 strerror(-run_ret));
1832 if (run_ret == -EBUSY) {
1834 "This is probably because your SMT is enabled.\n"
1835 "VCPU can only run on primary threads with all "
1836 "secondary threads offline.\n");
1843 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1844 switch (run->exit_reason) {
1846 DPRINTF("handle_io\n");
1847 /* Called outside BQL */
1848 kvm_handle_io(run->io.port, attrs,
1849 (uint8_t *)run + run->io.data_offset,
1856 DPRINTF("handle_mmio\n");
1857 /* Called outside BQL */
1858 address_space_rw(&address_space_memory,
1859 run->mmio.phys_addr, attrs,
1862 run->mmio.is_write);
1865 case KVM_EXIT_IRQ_WINDOW_OPEN:
1866 DPRINTF("irq_window_open\n");
1867 ret = EXCP_INTERRUPT;
1869 case KVM_EXIT_SHUTDOWN:
1870 DPRINTF("shutdown\n");
1871 qemu_system_reset_request();
1872 ret = EXCP_INTERRUPT;
1874 case KVM_EXIT_UNKNOWN:
1875 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1876 (uint64_t)run->hw.hardware_exit_reason);
1879 case KVM_EXIT_INTERNAL_ERROR:
1880 ret = kvm_handle_internal_error(cpu, run);
1882 case KVM_EXIT_SYSTEM_EVENT:
1883 switch (run->system_event.type) {
1884 case KVM_SYSTEM_EVENT_SHUTDOWN:
1885 qemu_system_shutdown_request();
1886 ret = EXCP_INTERRUPT;
1888 case KVM_SYSTEM_EVENT_RESET:
1889 qemu_system_reset_request();
1890 ret = EXCP_INTERRUPT;
1892 case KVM_SYSTEM_EVENT_CRASH:
1893 qemu_mutex_lock_iothread();
1894 qemu_system_guest_panicked();
1895 qemu_mutex_unlock_iothread();
1899 DPRINTF("kvm_arch_handle_exit\n");
1900 ret = kvm_arch_handle_exit(cpu, run);
1905 DPRINTF("kvm_arch_handle_exit\n");
1906 ret = kvm_arch_handle_exit(cpu, run);
1911 qemu_mutex_lock_iothread();
1914 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1915 vm_stop(RUN_STATE_INTERNAL_ERROR);
1918 cpu->exit_request = 0;
1922 int kvm_ioctl(KVMState *s, int type, ...)
1929 arg = va_arg(ap, void *);
1932 trace_kvm_ioctl(type, arg);
1933 ret = ioctl(s->fd, type, arg);
1940 int kvm_vm_ioctl(KVMState *s, int type, ...)
1947 arg = va_arg(ap, void *);
1950 trace_kvm_vm_ioctl(type, arg);
1951 ret = ioctl(s->vmfd, type, arg);
1958 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1965 arg = va_arg(ap, void *);
1968 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1969 ret = ioctl(cpu->kvm_fd, type, arg);
1976 int kvm_device_ioctl(int fd, int type, ...)
1983 arg = va_arg(ap, void *);
1986 trace_kvm_device_ioctl(fd, type, arg);
1987 ret = ioctl(fd, type, arg);
1994 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
1997 struct kvm_device_attr attribute = {
2002 if (!kvm_vm_attributes_allowed) {
2006 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2007 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2011 int kvm_has_sync_mmu(void)
2013 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2016 int kvm_has_vcpu_events(void)
2018 return kvm_state->vcpu_events;
2021 int kvm_has_robust_singlestep(void)
2023 return kvm_state->robust_singlestep;
2026 int kvm_has_debugregs(void)
2028 return kvm_state->debugregs;
2031 int kvm_has_xsave(void)
2033 return kvm_state->xsave;
2036 int kvm_has_xcrs(void)
2038 return kvm_state->xcrs;
2041 int kvm_has_pit_state2(void)
2043 return kvm_state->pit_state2;
2046 int kvm_has_many_ioeventfds(void)
2048 if (!kvm_enabled()) {
2051 return kvm_state->many_ioeventfds;
2054 int kvm_has_gsi_routing(void)
2056 #ifdef KVM_CAP_IRQ_ROUTING
2057 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2063 int kvm_has_intx_set_mask(void)
2065 return kvm_state->intx_set_mask;
2068 void kvm_setup_guest_memory(void *start, size_t size)
2070 if (!kvm_has_sync_mmu()) {
2071 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
2074 perror("qemu_madvise");
2076 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2082 #ifdef KVM_CAP_SET_GUEST_DEBUG
2083 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2086 struct kvm_sw_breakpoint *bp;
2088 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2096 int kvm_sw_breakpoints_active(CPUState *cpu)
2098 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2101 struct kvm_set_guest_debug_data {
2102 struct kvm_guest_debug dbg;
2107 static void kvm_invoke_set_guest_debug(void *data)
2109 struct kvm_set_guest_debug_data *dbg_data = data;
2111 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
2115 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2117 struct kvm_set_guest_debug_data data;
2119 data.dbg.control = reinject_trap;
2121 if (cpu->singlestep_enabled) {
2122 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2124 kvm_arch_update_guest_debug(cpu, &data.dbg);
2127 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
2131 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2132 target_ulong len, int type)
2134 struct kvm_sw_breakpoint *bp;
2137 if (type == GDB_BREAKPOINT_SW) {
2138 bp = kvm_find_sw_breakpoint(cpu, addr);
2144 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2147 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2153 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2155 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2162 err = kvm_update_guest_debug(cpu, 0);
2170 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2171 target_ulong len, int type)
2173 struct kvm_sw_breakpoint *bp;
2176 if (type == GDB_BREAKPOINT_SW) {
2177 bp = kvm_find_sw_breakpoint(cpu, addr);
2182 if (bp->use_count > 1) {
2187 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2192 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2195 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2202 err = kvm_update_guest_debug(cpu, 0);
2210 void kvm_remove_all_breakpoints(CPUState *cpu)
2212 struct kvm_sw_breakpoint *bp, *next;
2213 KVMState *s = cpu->kvm_state;
2216 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2217 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2218 /* Try harder to find a CPU that currently sees the breakpoint. */
2219 CPU_FOREACH(tmpcpu) {
2220 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2225 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2228 kvm_arch_remove_all_hw_breakpoints();
2231 kvm_update_guest_debug(cpu, 0);
2235 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2237 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2242 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2243 target_ulong len, int type)
2248 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2249 target_ulong len, int type)
2254 void kvm_remove_all_breakpoints(CPUState *cpu)
2257 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2259 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2261 KVMState *s = kvm_state;
2262 struct kvm_signal_mask *sigmask;
2266 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2269 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2271 sigmask->len = s->sigmask_len;
2272 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2273 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2278 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2280 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2283 int kvm_on_sigbus(int code, void *addr)
2285 return kvm_arch_on_sigbus(code, addr);
2288 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2291 struct kvm_create_device create_dev;
2293 create_dev.type = type;
2295 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2297 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2301 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2306 return test ? 0 : create_dev.fd;
2309 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2311 struct kvm_one_reg reg;
2315 reg.addr = (uintptr_t) source;
2316 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®);
2318 trace_kvm_failed_reg_set(id, strerror(r));
2323 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2325 struct kvm_one_reg reg;
2329 reg.addr = (uintptr_t) target;
2330 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®);
2332 trace_kvm_failed_reg_get(id, strerror(r));
2337 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2339 AccelClass *ac = ACCEL_CLASS(oc);
2341 ac->init_machine = kvm_init;
2342 ac->allowed = &kvm_allowed;
2345 static const TypeInfo kvm_accel_type = {
2346 .name = TYPE_KVM_ACCEL,
2347 .parent = TYPE_ACCEL,
2348 .class_init = kvm_accel_class_init,
2349 .instance_size = sizeof(KVMState),
2352 static void kvm_type_init(void)
2354 type_register_static(&kvm_accel_type);
2357 type_init(kvm_type_init);