4 * Copyright IBM, Corp. 2008
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Glauber Costa <gcosta@redhat.com>
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
16 #include "qemu/osdep.h"
17 #include <sys/ioctl.h>
20 #include <linux/kvm.h>
22 #include "qemu-common.h"
23 #include "qemu/atomic.h"
24 #include "qemu/option.h"
25 #include "qemu/config-file.h"
26 #include "qemu/error-report.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 KVM_COALESCED_MMIO_MAX. We
48 * need to use the real host PAGE_SIZE, as that's what KVM will use.
50 #define PAGE_SIZE getpagesize()
55 #define DPRINTF(fmt, ...) \
56 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
58 #define DPRINTF(fmt, ...) \
62 #define KVM_MSI_HASHTAB_SIZE 256
66 AccelState parent_obj;
72 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
73 bool coalesced_flush_in_progress;
74 int broken_set_mem_region;
76 int robust_singlestep;
78 #ifdef KVM_CAP_SET_GUEST_DEBUG
79 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 unsigned long *used_gsi_bitmap;
93 unsigned int gsi_count;
94 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
96 KVMMemoryListener memory_listener;
100 bool kvm_kernel_irqchip;
101 bool kvm_split_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;
113 bool kvm_direct_msi_allowed;
114 bool kvm_ioeventfd_any_length_allowed;
116 static const KVMCapabilityInfo kvm_required_capabilites[] = {
117 KVM_CAP_INFO(USER_MEMORY),
118 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
122 static KVMSlot *kvm_get_free_slot(KVMMemoryListener *kml)
124 KVMState *s = kvm_state;
127 for (i = 0; i < s->nr_slots; i++) {
128 if (kml->slots[i].memory_size == 0) {
129 return &kml->slots[i];
136 bool kvm_has_free_slot(MachineState *ms)
138 KVMState *s = KVM_STATE(ms->accelerator);
140 return kvm_get_free_slot(&s->memory_listener);
143 static KVMSlot *kvm_alloc_slot(KVMMemoryListener *kml)
145 KVMSlot *slot = kvm_get_free_slot(kml);
151 fprintf(stderr, "%s: no free slot available\n", __func__);
155 static KVMSlot *kvm_lookup_matching_slot(KVMMemoryListener *kml,
159 KVMState *s = kvm_state;
162 for (i = 0; i < s->nr_slots; i++) {
163 KVMSlot *mem = &kml->slots[i];
165 if (start_addr == mem->start_addr &&
166 end_addr == mem->start_addr + mem->memory_size) {
175 * Find overlapping slot with lowest start address
177 static KVMSlot *kvm_lookup_overlapping_slot(KVMMemoryListener *kml,
181 KVMState *s = kvm_state;
182 KVMSlot *found = NULL;
185 for (i = 0; i < s->nr_slots; i++) {
186 KVMSlot *mem = &kml->slots[i];
188 if (mem->memory_size == 0 ||
189 (found && found->start_addr < mem->start_addr)) {
193 if (end_addr > mem->start_addr &&
194 start_addr < mem->start_addr + mem->memory_size) {
202 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
205 KVMMemoryListener *kml = &s->memory_listener;
208 for (i = 0; i < s->nr_slots; i++) {
209 KVMSlot *mem = &kml->slots[i];
211 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
212 *phys_addr = mem->start_addr + (ram - mem->ram);
220 static int kvm_set_user_memory_region(KVMMemoryListener *kml, KVMSlot *slot)
222 KVMState *s = kvm_state;
223 struct kvm_userspace_memory_region mem;
225 mem.slot = slot->slot | (kml->as_id << 16);
226 mem.guest_phys_addr = slot->start_addr;
227 mem.userspace_addr = (unsigned long)slot->ram;
228 mem.flags = slot->flags;
230 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
231 /* Set the slot size to 0 before setting the slot to the desired
232 * value. This is needed based on KVM commit 75d61fbc. */
234 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
236 mem.memory_size = slot->memory_size;
237 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
240 int kvm_init_vcpu(CPUState *cpu)
242 KVMState *s = kvm_state;
246 DPRINTF("kvm_init_vcpu\n");
248 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
250 DPRINTF("kvm_create_vcpu failed\n");
256 cpu->kvm_vcpu_dirty = true;
258 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
261 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
265 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
267 if (cpu->kvm_run == MAP_FAILED) {
269 DPRINTF("mmap'ing vcpu state failed\n");
273 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
274 s->coalesced_mmio_ring =
275 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
278 ret = kvm_arch_init_vcpu(cpu);
284 * dirty pages logging control
287 static int kvm_mem_flags(MemoryRegion *mr)
289 bool readonly = mr->readonly || memory_region_is_romd(mr);
292 if (memory_region_get_dirty_log_mask(mr) != 0) {
293 flags |= KVM_MEM_LOG_DIRTY_PAGES;
295 if (readonly && kvm_readonly_mem_allowed) {
296 flags |= KVM_MEM_READONLY;
301 static int kvm_slot_update_flags(KVMMemoryListener *kml, KVMSlot *mem,
306 old_flags = mem->flags;
307 mem->flags = kvm_mem_flags(mr);
309 /* If nothing changed effectively, no need to issue ioctl */
310 if (mem->flags == old_flags) {
314 return kvm_set_user_memory_region(kml, mem);
317 static int kvm_section_update_flags(KVMMemoryListener *kml,
318 MemoryRegionSection *section)
320 hwaddr phys_addr = section->offset_within_address_space;
321 ram_addr_t size = int128_get64(section->size);
322 KVMSlot *mem = kvm_lookup_matching_slot(kml, phys_addr, phys_addr + size);
327 return kvm_slot_update_flags(kml, mem, section->mr);
331 static void kvm_log_start(MemoryListener *listener,
332 MemoryRegionSection *section,
335 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
342 r = kvm_section_update_flags(kml, section);
348 static void kvm_log_stop(MemoryListener *listener,
349 MemoryRegionSection *section,
352 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
359 r = kvm_section_update_flags(kml, section);
365 /* get kvm's dirty pages bitmap and update qemu's */
366 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
367 unsigned long *bitmap)
369 ram_addr_t start = section->offset_within_region +
370 memory_region_get_ram_addr(section->mr);
371 ram_addr_t pages = int128_get64(section->size) / getpagesize();
373 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
377 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
380 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
381 * This function updates qemu's dirty bitmap using
382 * memory_region_set_dirty(). This means all bits are set
385 * @start_add: start of logged region.
386 * @end_addr: end of logged region.
388 static int kvm_physical_sync_dirty_bitmap(KVMMemoryListener *kml,
389 MemoryRegionSection *section)
391 KVMState *s = kvm_state;
392 unsigned long size, allocated_size = 0;
393 struct kvm_dirty_log d = {};
396 hwaddr start_addr = section->offset_within_address_space;
397 hwaddr end_addr = start_addr + int128_get64(section->size);
399 d.dirty_bitmap = NULL;
400 while (start_addr < end_addr) {
401 mem = kvm_lookup_overlapping_slot(kml, start_addr, end_addr);
406 /* XXX bad kernel interface alert
407 * For dirty bitmap, kernel allocates array of size aligned to
408 * bits-per-long. But for case when the kernel is 64bits and
409 * the userspace is 32bits, userspace can't align to the same
410 * bits-per-long, since sizeof(long) is different between kernel
411 * and user space. This way, userspace will provide buffer which
412 * may be 4 bytes less than the kernel will use, resulting in
413 * userspace memory corruption (which is not detectable by valgrind
414 * too, in most cases).
415 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
416 * a hope that sizeof(long) wont become >8 any time soon.
418 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
419 /*HOST_LONG_BITS*/ 64) / 8;
420 if (!d.dirty_bitmap) {
421 d.dirty_bitmap = g_malloc(size);
422 } else if (size > allocated_size) {
423 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
425 allocated_size = size;
426 memset(d.dirty_bitmap, 0, allocated_size);
428 d.slot = mem->slot | (kml->as_id << 16);
429 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
430 DPRINTF("ioctl failed %d\n", errno);
435 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
436 start_addr = mem->start_addr + mem->memory_size;
438 g_free(d.dirty_bitmap);
443 static void kvm_coalesce_mmio_region(MemoryListener *listener,
444 MemoryRegionSection *secion,
445 hwaddr start, hwaddr size)
447 KVMState *s = kvm_state;
449 if (s->coalesced_mmio) {
450 struct kvm_coalesced_mmio_zone zone;
456 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
460 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
461 MemoryRegionSection *secion,
462 hwaddr start, hwaddr size)
464 KVMState *s = kvm_state;
466 if (s->coalesced_mmio) {
467 struct kvm_coalesced_mmio_zone zone;
473 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
477 int kvm_check_extension(KVMState *s, unsigned int extension)
481 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
489 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
493 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
495 /* VM wide version not implemented, use global one instead */
496 ret = kvm_check_extension(s, extension);
502 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
504 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
505 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
506 * endianness, but the memory core hands them in target endianness.
507 * For example, PPC is always treated as big-endian even if running
508 * on KVM and on PPC64LE. Correct here.
522 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
523 bool assign, uint32_t size, bool datamatch)
526 struct kvm_ioeventfd iofd = {
527 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
534 if (!kvm_enabled()) {
539 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
542 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
545 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
554 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
555 bool assign, uint32_t size, bool datamatch)
557 struct kvm_ioeventfd kick = {
558 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
560 .flags = KVM_IOEVENTFD_FLAG_PIO,
565 if (!kvm_enabled()) {
569 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
572 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
574 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
582 static int kvm_check_many_ioeventfds(void)
584 /* Userspace can use ioeventfd for io notification. This requires a host
585 * that supports eventfd(2) and an I/O thread; since eventfd does not
586 * support SIGIO it cannot interrupt the vcpu.
588 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
589 * can avoid creating too many ioeventfds.
591 #if defined(CONFIG_EVENTFD)
594 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
595 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
596 if (ioeventfds[i] < 0) {
599 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
601 close(ioeventfds[i]);
606 /* Decide whether many devices are supported or not */
607 ret = i == ARRAY_SIZE(ioeventfds);
610 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
611 close(ioeventfds[i]);
619 static const KVMCapabilityInfo *
620 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
623 if (!kvm_check_extension(s, list->value)) {
631 static void kvm_set_phys_mem(KVMMemoryListener *kml,
632 MemoryRegionSection *section, bool add)
634 KVMState *s = kvm_state;
637 MemoryRegion *mr = section->mr;
638 bool writeable = !mr->readonly && !mr->rom_device;
639 hwaddr start_addr = section->offset_within_address_space;
640 ram_addr_t size = int128_get64(section->size);
644 /* kvm works in page size chunks, but the function may be called
645 with sub-page size and unaligned start address. Pad the start
646 address to next and truncate size to previous page boundary. */
647 delta = qemu_real_host_page_size - (start_addr & ~qemu_real_host_page_mask);
648 delta &= ~qemu_real_host_page_mask;
654 size &= qemu_real_host_page_mask;
655 if (!size || (start_addr & ~qemu_real_host_page_mask)) {
659 if (!memory_region_is_ram(mr)) {
660 if (writeable || !kvm_readonly_mem_allowed) {
662 } else if (!mr->romd_mode) {
663 /* If the memory device is not in romd_mode, then we actually want
664 * to remove the kvm memory slot so all accesses will trap. */
669 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
672 mem = kvm_lookup_overlapping_slot(kml, start_addr, start_addr + size);
677 if (add && start_addr >= mem->start_addr &&
678 (start_addr + size <= mem->start_addr + mem->memory_size) &&
679 (ram - start_addr == mem->ram - mem->start_addr)) {
680 /* The new slot fits into the existing one and comes with
681 * identical parameters - update flags and done. */
682 kvm_slot_update_flags(kml, mem, mr);
688 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
689 kvm_physical_sync_dirty_bitmap(kml, section);
692 /* unregister the overlapping slot */
693 mem->memory_size = 0;
694 err = kvm_set_user_memory_region(kml, mem);
696 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
697 __func__, strerror(-err));
701 /* Workaround for older KVM versions: we can't join slots, even not by
702 * unregistering the previous ones and then registering the larger
703 * slot. We have to maintain the existing fragmentation. Sigh.
705 * This workaround assumes that the new slot starts at the same
706 * address as the first existing one. If not or if some overlapping
707 * slot comes around later, we will fail (not seen in practice so far)
708 * - and actually require a recent KVM version. */
709 if (s->broken_set_mem_region &&
710 old.start_addr == start_addr && old.memory_size < size && add) {
711 mem = kvm_alloc_slot(kml);
712 mem->memory_size = old.memory_size;
713 mem->start_addr = old.start_addr;
715 mem->flags = kvm_mem_flags(mr);
717 err = kvm_set_user_memory_region(kml, mem);
719 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
724 start_addr += old.memory_size;
725 ram += old.memory_size;
726 size -= old.memory_size;
730 /* register prefix slot */
731 if (old.start_addr < start_addr) {
732 mem = kvm_alloc_slot(kml);
733 mem->memory_size = start_addr - old.start_addr;
734 mem->start_addr = old.start_addr;
736 mem->flags = kvm_mem_flags(mr);
738 err = kvm_set_user_memory_region(kml, mem);
740 fprintf(stderr, "%s: error registering prefix slot: %s\n",
741 __func__, strerror(-err));
743 fprintf(stderr, "%s: This is probably because your kernel's " \
744 "PAGE_SIZE is too big. Please try to use 4k " \
745 "PAGE_SIZE!\n", __func__);
751 /* register suffix slot */
752 if (old.start_addr + old.memory_size > start_addr + size) {
753 ram_addr_t size_delta;
755 mem = kvm_alloc_slot(kml);
756 mem->start_addr = start_addr + size;
757 size_delta = mem->start_addr - old.start_addr;
758 mem->memory_size = old.memory_size - size_delta;
759 mem->ram = old.ram + size_delta;
760 mem->flags = kvm_mem_flags(mr);
762 err = kvm_set_user_memory_region(kml, mem);
764 fprintf(stderr, "%s: error registering suffix slot: %s\n",
765 __func__, strerror(-err));
771 /* in case the KVM bug workaround already "consumed" the new slot */
778 mem = kvm_alloc_slot(kml);
779 mem->memory_size = size;
780 mem->start_addr = start_addr;
782 mem->flags = kvm_mem_flags(mr);
784 err = kvm_set_user_memory_region(kml, mem);
786 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
792 static void kvm_region_add(MemoryListener *listener,
793 MemoryRegionSection *section)
795 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
797 memory_region_ref(section->mr);
798 kvm_set_phys_mem(kml, section, true);
801 static void kvm_region_del(MemoryListener *listener,
802 MemoryRegionSection *section)
804 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
806 kvm_set_phys_mem(kml, section, false);
807 memory_region_unref(section->mr);
810 static void kvm_log_sync(MemoryListener *listener,
811 MemoryRegionSection *section)
813 KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
816 r = kvm_physical_sync_dirty_bitmap(kml, section);
822 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
823 MemoryRegionSection *section,
824 bool match_data, uint64_t data,
827 int fd = event_notifier_get_fd(e);
830 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
831 data, true, int128_get64(section->size),
834 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
835 __func__, strerror(-r));
840 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
841 MemoryRegionSection *section,
842 bool match_data, uint64_t data,
845 int fd = event_notifier_get_fd(e);
848 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
849 data, false, int128_get64(section->size),
856 static void kvm_io_ioeventfd_add(MemoryListener *listener,
857 MemoryRegionSection *section,
858 bool match_data, uint64_t data,
861 int fd = event_notifier_get_fd(e);
864 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
865 data, true, int128_get64(section->size),
868 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
869 __func__, strerror(-r));
874 static void kvm_io_ioeventfd_del(MemoryListener *listener,
875 MemoryRegionSection *section,
876 bool match_data, uint64_t data,
880 int fd = event_notifier_get_fd(e);
883 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
884 data, false, int128_get64(section->size),
891 void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
892 AddressSpace *as, int as_id)
896 kml->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
899 for (i = 0; i < s->nr_slots; i++) {
900 kml->slots[i].slot = i;
903 kml->listener.region_add = kvm_region_add;
904 kml->listener.region_del = kvm_region_del;
905 kml->listener.log_start = kvm_log_start;
906 kml->listener.log_stop = kvm_log_stop;
907 kml->listener.log_sync = kvm_log_sync;
908 kml->listener.priority = 10;
910 memory_listener_register(&kml->listener, as);
913 static MemoryListener kvm_io_listener = {
914 .eventfd_add = kvm_io_ioeventfd_add,
915 .eventfd_del = kvm_io_ioeventfd_del,
919 static void kvm_handle_interrupt(CPUState *cpu, int mask)
921 cpu->interrupt_request |= mask;
923 if (!qemu_cpu_is_self(cpu)) {
928 int kvm_set_irq(KVMState *s, int irq, int level)
930 struct kvm_irq_level event;
933 assert(kvm_async_interrupts_enabled());
937 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
939 perror("kvm_set_irq");
943 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
946 #ifdef KVM_CAP_IRQ_ROUTING
947 typedef struct KVMMSIRoute {
948 struct kvm_irq_routing_entry kroute;
949 QTAILQ_ENTRY(KVMMSIRoute) entry;
952 static void set_gsi(KVMState *s, unsigned int gsi)
954 set_bit(gsi, s->used_gsi_bitmap);
957 static void clear_gsi(KVMState *s, unsigned int gsi)
959 clear_bit(gsi, s->used_gsi_bitmap);
962 void kvm_init_irq_routing(KVMState *s)
966 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
968 /* Round up so we can search ints using ffs */
969 s->used_gsi_bitmap = bitmap_new(gsi_count);
970 s->gsi_count = gsi_count;
973 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
974 s->nr_allocated_irq_routes = 0;
976 if (!kvm_direct_msi_allowed) {
977 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
978 QTAILQ_INIT(&s->msi_hashtab[i]);
982 kvm_arch_init_irq_routing(s);
985 void kvm_irqchip_commit_routes(KVMState *s)
989 s->irq_routes->flags = 0;
990 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
994 static void kvm_add_routing_entry(KVMState *s,
995 struct kvm_irq_routing_entry *entry)
997 struct kvm_irq_routing_entry *new;
1000 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1001 n = s->nr_allocated_irq_routes * 2;
1005 size = sizeof(struct kvm_irq_routing);
1006 size += n * sizeof(*new);
1007 s->irq_routes = g_realloc(s->irq_routes, size);
1008 s->nr_allocated_irq_routes = n;
1010 n = s->irq_routes->nr++;
1011 new = &s->irq_routes->entries[n];
1015 set_gsi(s, entry->gsi);
1018 static int kvm_update_routing_entry(KVMState *s,
1019 struct kvm_irq_routing_entry *new_entry)
1021 struct kvm_irq_routing_entry *entry;
1024 for (n = 0; n < s->irq_routes->nr; n++) {
1025 entry = &s->irq_routes->entries[n];
1026 if (entry->gsi != new_entry->gsi) {
1030 if(!memcmp(entry, new_entry, sizeof *entry)) {
1034 *entry = *new_entry;
1036 kvm_irqchip_commit_routes(s);
1044 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1046 struct kvm_irq_routing_entry e = {};
1048 assert(pin < s->gsi_count);
1051 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1053 e.u.irqchip.irqchip = irqchip;
1054 e.u.irqchip.pin = pin;
1055 kvm_add_routing_entry(s, &e);
1058 void kvm_irqchip_release_virq(KVMState *s, int virq)
1060 struct kvm_irq_routing_entry *e;
1063 if (kvm_gsi_direct_mapping()) {
1067 for (i = 0; i < s->irq_routes->nr; i++) {
1068 e = &s->irq_routes->entries[i];
1069 if (e->gsi == virq) {
1070 s->irq_routes->nr--;
1071 *e = s->irq_routes->entries[s->irq_routes->nr];
1077 static unsigned int kvm_hash_msi(uint32_t data)
1079 /* This is optimized for IA32 MSI layout. However, no other arch shall
1080 * repeat the mistake of not providing a direct MSI injection API. */
1084 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1086 KVMMSIRoute *route, *next;
1089 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1090 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1091 kvm_irqchip_release_virq(s, route->kroute.gsi);
1092 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1098 static int kvm_irqchip_get_virq(KVMState *s)
1103 * PIC and IOAPIC share the first 16 GSI numbers, thus the available
1104 * GSI numbers are more than the number of IRQ route. Allocating a GSI
1105 * number can succeed even though a new route entry cannot be added.
1106 * When this happens, flush dynamic MSI entries to free IRQ route entries.
1108 if (!kvm_direct_msi_allowed && s->irq_routes->nr == s->gsi_count) {
1109 kvm_flush_dynamic_msi_routes(s);
1112 /* Return the lowest unused GSI in the bitmap */
1113 next_virq = find_first_zero_bit(s->used_gsi_bitmap, s->gsi_count);
1114 if (next_virq >= s->gsi_count) {
1121 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1123 unsigned int hash = kvm_hash_msi(msg.data);
1126 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1127 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1128 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1129 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1136 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1141 if (kvm_direct_msi_allowed) {
1142 msi.address_lo = (uint32_t)msg.address;
1143 msi.address_hi = msg.address >> 32;
1144 msi.data = le32_to_cpu(msg.data);
1146 memset(msi.pad, 0, sizeof(msi.pad));
1148 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1151 route = kvm_lookup_msi_route(s, msg);
1155 virq = kvm_irqchip_get_virq(s);
1160 route = g_malloc0(sizeof(KVMMSIRoute));
1161 route->kroute.gsi = virq;
1162 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1163 route->kroute.flags = 0;
1164 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1165 route->kroute.u.msi.address_hi = msg.address >> 32;
1166 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1168 kvm_add_routing_entry(s, &route->kroute);
1169 kvm_irqchip_commit_routes(s);
1171 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1175 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1177 return kvm_set_irq(s, route->kroute.gsi, 1);
1180 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg, PCIDevice *dev)
1182 struct kvm_irq_routing_entry kroute = {};
1185 if (kvm_gsi_direct_mapping()) {
1186 return kvm_arch_msi_data_to_gsi(msg.data);
1189 if (!kvm_gsi_routing_enabled()) {
1193 virq = kvm_irqchip_get_virq(s);
1199 kroute.type = KVM_IRQ_ROUTING_MSI;
1201 kroute.u.msi.address_lo = (uint32_t)msg.address;
1202 kroute.u.msi.address_hi = msg.address >> 32;
1203 kroute.u.msi.data = le32_to_cpu(msg.data);
1204 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1205 kvm_irqchip_release_virq(s, virq);
1209 kvm_add_routing_entry(s, &kroute);
1210 kvm_irqchip_commit_routes(s);
1215 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg,
1218 struct kvm_irq_routing_entry kroute = {};
1220 if (kvm_gsi_direct_mapping()) {
1224 if (!kvm_irqchip_in_kernel()) {
1229 kroute.type = KVM_IRQ_ROUTING_MSI;
1231 kroute.u.msi.address_lo = (uint32_t)msg.address;
1232 kroute.u.msi.address_hi = msg.address >> 32;
1233 kroute.u.msi.data = le32_to_cpu(msg.data);
1234 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data, dev)) {
1238 return kvm_update_routing_entry(s, &kroute);
1241 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1244 struct kvm_irqfd irqfd = {
1247 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1251 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1252 irqfd.resamplefd = rfd;
1255 if (!kvm_irqfds_enabled()) {
1259 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1262 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1264 struct kvm_irq_routing_entry kroute = {};
1267 if (!kvm_gsi_routing_enabled()) {
1271 virq = kvm_irqchip_get_virq(s);
1277 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1279 kroute.u.adapter.summary_addr = adapter->summary_addr;
1280 kroute.u.adapter.ind_addr = adapter->ind_addr;
1281 kroute.u.adapter.summary_offset = adapter->summary_offset;
1282 kroute.u.adapter.ind_offset = adapter->ind_offset;
1283 kroute.u.adapter.adapter_id = adapter->adapter_id;
1285 kvm_add_routing_entry(s, &kroute);
1290 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1292 struct kvm_irq_routing_entry kroute = {};
1295 if (!kvm_gsi_routing_enabled()) {
1298 if (!kvm_check_extension(s, KVM_CAP_HYPERV_SYNIC)) {
1301 virq = kvm_irqchip_get_virq(s);
1307 kroute.type = KVM_IRQ_ROUTING_HV_SINT;
1309 kroute.u.hv_sint.vcpu = vcpu;
1310 kroute.u.hv_sint.sint = sint;
1312 kvm_add_routing_entry(s, &kroute);
1313 kvm_irqchip_commit_routes(s);
1318 #else /* !KVM_CAP_IRQ_ROUTING */
1320 void kvm_init_irq_routing(KVMState *s)
1324 void kvm_irqchip_release_virq(KVMState *s, int virq)
1328 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1333 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1338 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1343 int kvm_irqchip_add_hv_sint_route(KVMState *s, uint32_t vcpu, uint32_t sint)
1348 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1353 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1357 #endif /* !KVM_CAP_IRQ_ROUTING */
1359 int kvm_irqchip_add_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1360 EventNotifier *rn, int virq)
1362 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1363 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1366 int kvm_irqchip_remove_irqfd_notifier_gsi(KVMState *s, EventNotifier *n,
1369 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1373 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1374 EventNotifier *rn, qemu_irq irq)
1377 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1382 return kvm_irqchip_add_irqfd_notifier_gsi(s, n, rn, GPOINTER_TO_INT(gsi));
1385 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n,
1389 gboolean found = g_hash_table_lookup_extended(s->gsimap, irq, &key, &gsi);
1394 return kvm_irqchip_remove_irqfd_notifier_gsi(s, n, GPOINTER_TO_INT(gsi));
1397 void kvm_irqchip_set_qemuirq_gsi(KVMState *s, qemu_irq irq, int gsi)
1399 g_hash_table_insert(s->gsimap, irq, GINT_TO_POINTER(gsi));
1402 static void kvm_irqchip_create(MachineState *machine, KVMState *s)
1406 if (kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1408 } else if (kvm_check_extension(s, KVM_CAP_S390_IRQCHIP)) {
1409 ret = kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0);
1411 fprintf(stderr, "Enable kernel irqchip failed: %s\n", strerror(-ret));
1418 /* First probe and see if there's a arch-specific hook to create the
1419 * in-kernel irqchip for us */
1420 ret = kvm_arch_irqchip_create(machine, s);
1422 if (machine_kernel_irqchip_split(machine)) {
1423 perror("Split IRQ chip mode not supported.");
1426 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1430 fprintf(stderr, "Create kernel irqchip failed: %s\n", strerror(-ret));
1434 kvm_kernel_irqchip = true;
1435 /* If we have an in-kernel IRQ chip then we must have asynchronous
1436 * interrupt delivery (though the reverse is not necessarily true)
1438 kvm_async_interrupts_allowed = true;
1439 kvm_halt_in_kernel_allowed = true;
1441 kvm_init_irq_routing(s);
1443 s->gsimap = g_hash_table_new(g_direct_hash, g_direct_equal);
1446 /* Find number of supported CPUs using the recommended
1447 * procedure from the kernel API documentation to cope with
1448 * older kernels that may be missing capabilities.
1450 static int kvm_recommended_vcpus(KVMState *s)
1452 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1453 return (ret) ? ret : 4;
1456 static int kvm_max_vcpus(KVMState *s)
1458 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1459 return (ret) ? ret : kvm_recommended_vcpus(s);
1462 static int kvm_init(MachineState *ms)
1464 MachineClass *mc = MACHINE_GET_CLASS(ms);
1465 static const char upgrade_note[] =
1466 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1467 "(see http://sourceforge.net/projects/kvm).\n";
1472 { "SMP", smp_cpus },
1473 { "hotpluggable", max_cpus },
1476 int soft_vcpus_limit, hard_vcpus_limit;
1478 const KVMCapabilityInfo *missing_cap;
1481 const char *kvm_type;
1483 s = KVM_STATE(ms->accelerator);
1486 * On systems where the kernel can support different base page
1487 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1488 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1489 * page size for the system though.
1491 assert(TARGET_PAGE_SIZE <= getpagesize());
1495 #ifdef KVM_CAP_SET_GUEST_DEBUG
1496 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1499 s->fd = qemu_open("/dev/kvm", O_RDWR);
1501 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1506 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1507 if (ret < KVM_API_VERSION) {
1511 fprintf(stderr, "kvm version too old\n");
1515 if (ret > KVM_API_VERSION) {
1517 fprintf(stderr, "kvm version not supported\n");
1521 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1523 /* If unspecified, use the default value */
1528 /* check the vcpu limits */
1529 soft_vcpus_limit = kvm_recommended_vcpus(s);
1530 hard_vcpus_limit = kvm_max_vcpus(s);
1533 if (nc->num > soft_vcpus_limit) {
1535 "Warning: Number of %s cpus requested (%d) exceeds "
1536 "the recommended cpus supported by KVM (%d)\n",
1537 nc->name, nc->num, soft_vcpus_limit);
1539 if (nc->num > hard_vcpus_limit) {
1540 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1541 "the maximum cpus supported by KVM (%d)\n",
1542 nc->name, nc->num, hard_vcpus_limit);
1549 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1551 type = mc->kvm_type(kvm_type);
1552 } else if (kvm_type) {
1554 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1559 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1560 } while (ret == -EINTR);
1563 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1567 if (ret == -EINVAL) {
1569 "Host kernel setup problem detected. Please verify:\n");
1570 fprintf(stderr, "- for kernels supporting the switch_amode or"
1571 " user_mode parameters, whether\n");
1573 " user space is running in primary address space\n");
1575 "- for kernels supporting the vm.allocate_pgste sysctl, "
1576 "whether it is enabled\n");
1583 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1586 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1590 fprintf(stderr, "kvm does not support %s\n%s",
1591 missing_cap->name, upgrade_note);
1595 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1597 s->broken_set_mem_region = 1;
1598 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1600 s->broken_set_mem_region = 0;
1603 #ifdef KVM_CAP_VCPU_EVENTS
1604 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1607 s->robust_singlestep =
1608 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1610 #ifdef KVM_CAP_DEBUGREGS
1611 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1614 #ifdef KVM_CAP_IRQ_ROUTING
1615 kvm_direct_msi_allowed = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1618 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1620 s->irq_set_ioctl = KVM_IRQ_LINE;
1621 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1622 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1625 #ifdef KVM_CAP_READONLY_MEM
1626 kvm_readonly_mem_allowed =
1627 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1630 kvm_eventfds_allowed =
1631 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1633 kvm_irqfds_allowed =
1634 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1636 kvm_resamplefds_allowed =
1637 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1639 kvm_vm_attributes_allowed =
1640 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1642 kvm_ioeventfd_any_length_allowed =
1643 (kvm_check_extension(s, KVM_CAP_IOEVENTFD_ANY_LENGTH) > 0);
1645 ret = kvm_arch_init(ms, s);
1650 if (machine_kernel_irqchip_allowed(ms)) {
1651 kvm_irqchip_create(ms, s);
1656 if (kvm_eventfds_allowed) {
1657 s->memory_listener.listener.eventfd_add = kvm_mem_ioeventfd_add;
1658 s->memory_listener.listener.eventfd_del = kvm_mem_ioeventfd_del;
1660 s->memory_listener.listener.coalesced_mmio_add = kvm_coalesce_mmio_region;
1661 s->memory_listener.listener.coalesced_mmio_del = kvm_uncoalesce_mmio_region;
1663 kvm_memory_listener_register(s, &s->memory_listener,
1664 &address_space_memory, 0);
1665 memory_listener_register(&kvm_io_listener,
1668 s->many_ioeventfds = kvm_check_many_ioeventfds();
1670 cpu_interrupt_handler = kvm_handle_interrupt;
1682 g_free(s->memory_listener.slots);
1687 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1689 s->sigmask_len = sigmask_len;
1692 static void kvm_handle_io(uint16_t port, MemTxAttrs attrs, void *data, int direction,
1693 int size, uint32_t count)
1696 uint8_t *ptr = data;
1698 for (i = 0; i < count; i++) {
1699 address_space_rw(&address_space_io, port, attrs,
1701 direction == KVM_EXIT_IO_OUT);
1706 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1708 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1709 run->internal.suberror);
1711 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1714 for (i = 0; i < run->internal.ndata; ++i) {
1715 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1716 i, (uint64_t)run->internal.data[i]);
1719 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1720 fprintf(stderr, "emulation failure\n");
1721 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1722 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1723 return EXCP_INTERRUPT;
1726 /* FIXME: Should trigger a qmp message to let management know
1727 * something went wrong.
1732 void kvm_flush_coalesced_mmio_buffer(void)
1734 KVMState *s = kvm_state;
1736 if (s->coalesced_flush_in_progress) {
1740 s->coalesced_flush_in_progress = true;
1742 if (s->coalesced_mmio_ring) {
1743 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1744 while (ring->first != ring->last) {
1745 struct kvm_coalesced_mmio *ent;
1747 ent = &ring->coalesced_mmio[ring->first];
1749 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1751 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1755 s->coalesced_flush_in_progress = false;
1758 static void do_kvm_cpu_synchronize_state(void *arg)
1760 CPUState *cpu = arg;
1762 if (!cpu->kvm_vcpu_dirty) {
1763 kvm_arch_get_registers(cpu);
1764 cpu->kvm_vcpu_dirty = true;
1768 void kvm_cpu_synchronize_state(CPUState *cpu)
1770 if (!cpu->kvm_vcpu_dirty) {
1771 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1775 static void do_kvm_cpu_synchronize_post_reset(void *arg)
1777 CPUState *cpu = arg;
1779 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1780 cpu->kvm_vcpu_dirty = false;
1783 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1785 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, cpu);
1788 static void do_kvm_cpu_synchronize_post_init(void *arg)
1790 CPUState *cpu = arg;
1792 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1793 cpu->kvm_vcpu_dirty = false;
1796 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1798 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, cpu);
1801 int kvm_cpu_exec(CPUState *cpu)
1803 struct kvm_run *run = cpu->kvm_run;
1806 DPRINTF("kvm_cpu_exec()\n");
1808 if (kvm_arch_process_async_events(cpu)) {
1809 cpu->exit_request = 0;
1813 qemu_mutex_unlock_iothread();
1818 if (cpu->kvm_vcpu_dirty) {
1819 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1820 cpu->kvm_vcpu_dirty = false;
1823 kvm_arch_pre_run(cpu, run);
1824 if (cpu->exit_request) {
1825 DPRINTF("interrupt exit requested\n");
1827 * KVM requires us to reenter the kernel after IO exits to complete
1828 * instruction emulation. This self-signal will ensure that we
1831 qemu_cpu_kick_self();
1834 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1836 attrs = kvm_arch_post_run(cpu, run);
1839 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1840 DPRINTF("io window exit\n");
1841 ret = EXCP_INTERRUPT;
1844 fprintf(stderr, "error: kvm run failed %s\n",
1845 strerror(-run_ret));
1847 if (run_ret == -EBUSY) {
1849 "This is probably because your SMT is enabled.\n"
1850 "VCPU can only run on primary threads with all "
1851 "secondary threads offline.\n");
1858 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1859 switch (run->exit_reason) {
1861 DPRINTF("handle_io\n");
1862 /* Called outside BQL */
1863 kvm_handle_io(run->io.port, attrs,
1864 (uint8_t *)run + run->io.data_offset,
1871 DPRINTF("handle_mmio\n");
1872 /* Called outside BQL */
1873 address_space_rw(&address_space_memory,
1874 run->mmio.phys_addr, attrs,
1877 run->mmio.is_write);
1880 case KVM_EXIT_IRQ_WINDOW_OPEN:
1881 DPRINTF("irq_window_open\n");
1882 ret = EXCP_INTERRUPT;
1884 case KVM_EXIT_SHUTDOWN:
1885 DPRINTF("shutdown\n");
1886 qemu_system_reset_request();
1887 ret = EXCP_INTERRUPT;
1889 case KVM_EXIT_UNKNOWN:
1890 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1891 (uint64_t)run->hw.hardware_exit_reason);
1894 case KVM_EXIT_INTERNAL_ERROR:
1895 ret = kvm_handle_internal_error(cpu, run);
1897 case KVM_EXIT_SYSTEM_EVENT:
1898 switch (run->system_event.type) {
1899 case KVM_SYSTEM_EVENT_SHUTDOWN:
1900 qemu_system_shutdown_request();
1901 ret = EXCP_INTERRUPT;
1903 case KVM_SYSTEM_EVENT_RESET:
1904 qemu_system_reset_request();
1905 ret = EXCP_INTERRUPT;
1907 case KVM_SYSTEM_EVENT_CRASH:
1908 qemu_mutex_lock_iothread();
1909 qemu_system_guest_panicked();
1910 qemu_mutex_unlock_iothread();
1914 DPRINTF("kvm_arch_handle_exit\n");
1915 ret = kvm_arch_handle_exit(cpu, run);
1920 DPRINTF("kvm_arch_handle_exit\n");
1921 ret = kvm_arch_handle_exit(cpu, run);
1926 qemu_mutex_lock_iothread();
1929 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1930 vm_stop(RUN_STATE_INTERNAL_ERROR);
1933 cpu->exit_request = 0;
1937 int kvm_ioctl(KVMState *s, int type, ...)
1944 arg = va_arg(ap, void *);
1947 trace_kvm_ioctl(type, arg);
1948 ret = ioctl(s->fd, type, arg);
1955 int kvm_vm_ioctl(KVMState *s, int type, ...)
1962 arg = va_arg(ap, void *);
1965 trace_kvm_vm_ioctl(type, arg);
1966 ret = ioctl(s->vmfd, type, arg);
1973 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1980 arg = va_arg(ap, void *);
1983 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1984 ret = ioctl(cpu->kvm_fd, type, arg);
1991 int kvm_device_ioctl(int fd, int type, ...)
1998 arg = va_arg(ap, void *);
2001 trace_kvm_device_ioctl(fd, type, arg);
2002 ret = ioctl(fd, type, arg);
2009 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
2012 struct kvm_device_attr attribute = {
2017 if (!kvm_vm_attributes_allowed) {
2021 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
2022 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
2026 int kvm_device_check_attr(int dev_fd, uint32_t group, uint64_t attr)
2028 struct kvm_device_attr attribute = {
2034 return kvm_device_ioctl(dev_fd, KVM_HAS_DEVICE_ATTR, &attribute) ? 0 : 1;
2037 void kvm_device_access(int fd, int group, uint64_t attr,
2038 void *val, bool write)
2040 struct kvm_device_attr kvmattr;
2044 kvmattr.group = group;
2045 kvmattr.attr = attr;
2046 kvmattr.addr = (uintptr_t)val;
2048 err = kvm_device_ioctl(fd,
2049 write ? KVM_SET_DEVICE_ATTR : KVM_GET_DEVICE_ATTR,
2052 error_report("KVM_%s_DEVICE_ATTR failed: %s",
2053 write ? "SET" : "GET", strerror(-err));
2054 error_printf("Group %d attr 0x%016" PRIx64, group, attr);
2059 int kvm_has_sync_mmu(void)
2061 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
2064 int kvm_has_vcpu_events(void)
2066 return kvm_state->vcpu_events;
2069 int kvm_has_robust_singlestep(void)
2071 return kvm_state->robust_singlestep;
2074 int kvm_has_debugregs(void)
2076 return kvm_state->debugregs;
2079 int kvm_has_many_ioeventfds(void)
2081 if (!kvm_enabled()) {
2084 return kvm_state->many_ioeventfds;
2087 int kvm_has_gsi_routing(void)
2089 #ifdef KVM_CAP_IRQ_ROUTING
2090 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2096 int kvm_has_intx_set_mask(void)
2098 return kvm_state->intx_set_mask;
2101 void kvm_setup_guest_memory(void *start, size_t size)
2103 if (!kvm_has_sync_mmu()) {
2104 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
2107 perror("qemu_madvise");
2109 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2115 #ifdef KVM_CAP_SET_GUEST_DEBUG
2116 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2119 struct kvm_sw_breakpoint *bp;
2121 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2129 int kvm_sw_breakpoints_active(CPUState *cpu)
2131 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2134 struct kvm_set_guest_debug_data {
2135 struct kvm_guest_debug dbg;
2140 static void kvm_invoke_set_guest_debug(void *data)
2142 struct kvm_set_guest_debug_data *dbg_data = data;
2144 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
2148 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2150 struct kvm_set_guest_debug_data data;
2152 data.dbg.control = reinject_trap;
2154 if (cpu->singlestep_enabled) {
2155 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2157 kvm_arch_update_guest_debug(cpu, &data.dbg);
2160 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
2164 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2165 target_ulong len, int type)
2167 struct kvm_sw_breakpoint *bp;
2170 if (type == GDB_BREAKPOINT_SW) {
2171 bp = kvm_find_sw_breakpoint(cpu, addr);
2177 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2180 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2186 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2188 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2195 err = kvm_update_guest_debug(cpu, 0);
2203 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2204 target_ulong len, int type)
2206 struct kvm_sw_breakpoint *bp;
2209 if (type == GDB_BREAKPOINT_SW) {
2210 bp = kvm_find_sw_breakpoint(cpu, addr);
2215 if (bp->use_count > 1) {
2220 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2225 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2228 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2235 err = kvm_update_guest_debug(cpu, 0);
2243 void kvm_remove_all_breakpoints(CPUState *cpu)
2245 struct kvm_sw_breakpoint *bp, *next;
2246 KVMState *s = cpu->kvm_state;
2249 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2250 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2251 /* Try harder to find a CPU that currently sees the breakpoint. */
2252 CPU_FOREACH(tmpcpu) {
2253 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2258 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2261 kvm_arch_remove_all_hw_breakpoints();
2264 kvm_update_guest_debug(cpu, 0);
2268 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2270 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2275 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2276 target_ulong len, int type)
2281 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2282 target_ulong len, int type)
2287 void kvm_remove_all_breakpoints(CPUState *cpu)
2290 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2292 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2294 KVMState *s = kvm_state;
2295 struct kvm_signal_mask *sigmask;
2299 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2302 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2304 sigmask->len = s->sigmask_len;
2305 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2306 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2311 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2313 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2316 int kvm_on_sigbus(int code, void *addr)
2318 return kvm_arch_on_sigbus(code, addr);
2321 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2324 struct kvm_create_device create_dev;
2326 create_dev.type = type;
2328 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2330 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2334 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2339 return test ? 0 : create_dev.fd;
2342 bool kvm_device_supported(int vmfd, uint64_t type)
2344 struct kvm_create_device create_dev = {
2347 .flags = KVM_CREATE_DEVICE_TEST,
2350 if (ioctl(vmfd, KVM_CHECK_EXTENSION, KVM_CAP_DEVICE_CTRL) <= 0) {
2354 return (ioctl(vmfd, KVM_CREATE_DEVICE, &create_dev) >= 0);
2357 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2359 struct kvm_one_reg reg;
2363 reg.addr = (uintptr_t) source;
2364 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, ®);
2366 trace_kvm_failed_reg_set(id, strerror(-r));
2371 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2373 struct kvm_one_reg reg;
2377 reg.addr = (uintptr_t) target;
2378 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, ®);
2380 trace_kvm_failed_reg_get(id, strerror(-r));
2385 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2387 AccelClass *ac = ACCEL_CLASS(oc);
2389 ac->init_machine = kvm_init;
2390 ac->allowed = &kvm_allowed;
2393 static const TypeInfo kvm_accel_type = {
2394 .name = TYPE_KVM_ACCEL,
2395 .parent = TYPE_ACCEL,
2396 .class_init = kvm_accel_class_init,
2397 .instance_size = sizeof(KVMState),
2400 static void kvm_type_init(void)
2402 type_register_static(&kvm_accel_type);
2405 type_init(kvm_type_init);