4 * Copyright IBM, Corp. 2008
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
16 #include <sys/types.h>
17 #include <sys/ioctl.h>
21 #include <linux/kvm.h>
23 #include "qemu-common.h"
24 #include "qemu-barrier.h"
32 #include "exec-memory.h"
34 /* This check must be after config-host.h is included */
36 #include <sys/eventfd.h>
39 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
40 #define PAGE_SIZE TARGET_PAGE_SIZE
45 #define DPRINTF(fmt, ...) \
46 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
48 #define DPRINTF(fmt, ...) \
52 #define KVM_MSI_HASHTAB_SIZE 256
54 typedef struct KVMSlot
56 target_phys_addr_t start_addr;
57 ram_addr_t memory_size;
63 typedef struct kvm_dirty_log KVMDirtyLog;
65 typedef struct KVMMSIRoute {
66 struct kvm_irq_routing_entry kroute;
67 QTAILQ_ENTRY(KVMMSIRoute) entry;
76 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
77 bool coalesced_flush_in_progress;
78 int broken_set_mem_region;
81 int robust_singlestep;
83 #ifdef KVM_CAP_SET_GUEST_DEBUG
84 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
89 /* The man page (and posix) say ioctl numbers are signed int, but
90 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
91 * unsigned, and treating them as signed here can break things */
92 unsigned irqchip_inject_ioctl;
93 #ifdef KVM_CAP_IRQ_ROUTING
94 struct kvm_irq_routing *irq_routes;
95 int nr_allocated_irq_routes;
96 uint32_t *used_gsi_bitmap;
97 unsigned int gsi_count;
98 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
104 bool kvm_kernel_irqchip;
106 static const KVMCapabilityInfo kvm_required_capabilites[] = {
107 KVM_CAP_INFO(USER_MEMORY),
108 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
112 static KVMSlot *kvm_alloc_slot(KVMState *s)
116 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
117 if (s->slots[i].memory_size == 0) {
122 fprintf(stderr, "%s: no free slot available\n", __func__);
126 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
127 target_phys_addr_t start_addr,
128 target_phys_addr_t end_addr)
132 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
133 KVMSlot *mem = &s->slots[i];
135 if (start_addr == mem->start_addr &&
136 end_addr == mem->start_addr + mem->memory_size) {
145 * Find overlapping slot with lowest start address
147 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
148 target_phys_addr_t start_addr,
149 target_phys_addr_t end_addr)
151 KVMSlot *found = NULL;
154 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
155 KVMSlot *mem = &s->slots[i];
157 if (mem->memory_size == 0 ||
158 (found && found->start_addr < mem->start_addr)) {
162 if (end_addr > mem->start_addr &&
163 start_addr < mem->start_addr + mem->memory_size) {
171 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
172 target_phys_addr_t *phys_addr)
176 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
177 KVMSlot *mem = &s->slots[i];
179 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
180 *phys_addr = mem->start_addr + (ram - mem->ram);
188 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
190 struct kvm_userspace_memory_region mem;
192 mem.slot = slot->slot;
193 mem.guest_phys_addr = slot->start_addr;
194 mem.memory_size = slot->memory_size;
195 mem.userspace_addr = (unsigned long)slot->ram;
196 mem.flags = slot->flags;
197 if (s->migration_log) {
198 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
200 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
203 static void kvm_reset_vcpu(void *opaque)
205 CPUArchState *env = opaque;
207 kvm_arch_reset_vcpu(env);
210 int kvm_init_vcpu(CPUArchState *env)
212 KVMState *s = kvm_state;
216 DPRINTF("kvm_init_vcpu\n");
218 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
220 DPRINTF("kvm_create_vcpu failed\n");
226 env->kvm_vcpu_dirty = 1;
228 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
231 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
235 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
237 if (env->kvm_run == MAP_FAILED) {
239 DPRINTF("mmap'ing vcpu state failed\n");
243 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
244 s->coalesced_mmio_ring =
245 (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
248 ret = kvm_arch_init_vcpu(env);
250 qemu_register_reset(kvm_reset_vcpu, env);
251 kvm_arch_reset_vcpu(env);
258 * dirty pages logging control
261 static int kvm_mem_flags(KVMState *s, bool log_dirty)
263 return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
266 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
268 KVMState *s = kvm_state;
269 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
272 old_flags = mem->flags;
274 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
277 /* If nothing changed effectively, no need to issue ioctl */
278 if (s->migration_log) {
279 flags |= KVM_MEM_LOG_DIRTY_PAGES;
282 if (flags == old_flags) {
286 return kvm_set_user_memory_region(s, mem);
289 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
290 ram_addr_t size, bool log_dirty)
292 KVMState *s = kvm_state;
293 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
296 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
297 TARGET_FMT_plx "\n", __func__, phys_addr,
298 (target_phys_addr_t)(phys_addr + size - 1));
301 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
304 static void kvm_log_start(MemoryListener *listener,
305 MemoryRegionSection *section)
309 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
310 section->size, true);
316 static void kvm_log_stop(MemoryListener *listener,
317 MemoryRegionSection *section)
321 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
322 section->size, false);
328 static int kvm_set_migration_log(int enable)
330 KVMState *s = kvm_state;
334 s->migration_log = enable;
336 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
339 if (!mem->memory_size) {
342 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
345 err = kvm_set_user_memory_region(s, mem);
353 /* get kvm's dirty pages bitmap and update qemu's */
354 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
355 unsigned long *bitmap)
358 unsigned long page_number, c;
359 target_phys_addr_t addr, addr1;
360 unsigned int len = ((section->size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
361 unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
364 * bitmap-traveling is faster than memory-traveling (for addr...)
365 * especially when most of the memory is not dirty.
367 for (i = 0; i < len; i++) {
368 if (bitmap[i] != 0) {
369 c = leul_to_cpu(bitmap[i]);
373 page_number = (i * HOST_LONG_BITS + j) * hpratio;
374 addr1 = page_number * TARGET_PAGE_SIZE;
375 addr = section->offset_within_region + addr1;
376 memory_region_set_dirty(section->mr, addr,
377 TARGET_PAGE_SIZE * hpratio);
384 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
387 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
388 * This function updates qemu's dirty bitmap using
389 * memory_region_set_dirty(). This means all bits are set
392 * @start_add: start of logged region.
393 * @end_addr: end of logged region.
395 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
397 KVMState *s = kvm_state;
398 unsigned long size, allocated_size = 0;
402 target_phys_addr_t start_addr = section->offset_within_address_space;
403 target_phys_addr_t end_addr = start_addr + section->size;
405 d.dirty_bitmap = NULL;
406 while (start_addr < end_addr) {
407 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
412 /* XXX bad kernel interface alert
413 * For dirty bitmap, kernel allocates array of size aligned to
414 * bits-per-long. But for case when the kernel is 64bits and
415 * the userspace is 32bits, userspace can't align to the same
416 * bits-per-long, since sizeof(long) is different between kernel
417 * and user space. This way, userspace will provide buffer which
418 * may be 4 bytes less than the kernel will use, resulting in
419 * userspace memory corruption (which is not detectable by valgrind
420 * too, in most cases).
421 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
422 * a hope that sizeof(long) wont become >8 any time soon.
424 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
425 /*HOST_LONG_BITS*/ 64) / 8;
426 if (!d.dirty_bitmap) {
427 d.dirty_bitmap = g_malloc(size);
428 } else if (size > allocated_size) {
429 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
431 allocated_size = size;
432 memset(d.dirty_bitmap, 0, allocated_size);
436 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
437 DPRINTF("ioctl failed %d\n", errno);
442 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
443 start_addr = mem->start_addr + mem->memory_size;
445 g_free(d.dirty_bitmap);
450 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
453 KVMState *s = kvm_state;
455 if (s->coalesced_mmio) {
456 struct kvm_coalesced_mmio_zone zone;
462 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
468 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
471 KVMState *s = kvm_state;
473 if (s->coalesced_mmio) {
474 struct kvm_coalesced_mmio_zone zone;
480 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
486 int kvm_check_extension(KVMState *s, unsigned int extension)
490 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
498 static int kvm_check_many_ioeventfds(void)
500 /* Userspace can use ioeventfd for io notification. This requires a host
501 * that supports eventfd(2) and an I/O thread; since eventfd does not
502 * support SIGIO it cannot interrupt the vcpu.
504 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
505 * can avoid creating too many ioeventfds.
507 #if defined(CONFIG_EVENTFD)
510 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
511 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
512 if (ioeventfds[i] < 0) {
515 ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
517 close(ioeventfds[i]);
522 /* Decide whether many devices are supported or not */
523 ret = i == ARRAY_SIZE(ioeventfds);
526 kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
527 close(ioeventfds[i]);
535 static const KVMCapabilityInfo *
536 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
539 if (!kvm_check_extension(s, list->value)) {
547 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
549 KVMState *s = kvm_state;
552 MemoryRegion *mr = section->mr;
553 bool log_dirty = memory_region_is_logging(mr);
554 target_phys_addr_t start_addr = section->offset_within_address_space;
555 ram_addr_t size = section->size;
559 /* kvm works in page size chunks, but the function may be called
560 with sub-page size and unaligned start address. */
561 delta = TARGET_PAGE_ALIGN(size) - size;
567 size &= TARGET_PAGE_MASK;
568 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
572 if (!memory_region_is_ram(mr)) {
576 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
579 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
584 if (add && start_addr >= mem->start_addr &&
585 (start_addr + size <= mem->start_addr + mem->memory_size) &&
586 (ram - start_addr == mem->ram - mem->start_addr)) {
587 /* The new slot fits into the existing one and comes with
588 * identical parameters - update flags and done. */
589 kvm_slot_dirty_pages_log_change(mem, log_dirty);
595 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
596 kvm_physical_sync_dirty_bitmap(section);
599 /* unregister the overlapping slot */
600 mem->memory_size = 0;
601 err = kvm_set_user_memory_region(s, mem);
603 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
604 __func__, strerror(-err));
608 /* Workaround for older KVM versions: we can't join slots, even not by
609 * unregistering the previous ones and then registering the larger
610 * slot. We have to maintain the existing fragmentation. Sigh.
612 * This workaround assumes that the new slot starts at the same
613 * address as the first existing one. If not or if some overlapping
614 * slot comes around later, we will fail (not seen in practice so far)
615 * - and actually require a recent KVM version. */
616 if (s->broken_set_mem_region &&
617 old.start_addr == start_addr && old.memory_size < size && add) {
618 mem = kvm_alloc_slot(s);
619 mem->memory_size = old.memory_size;
620 mem->start_addr = old.start_addr;
622 mem->flags = kvm_mem_flags(s, log_dirty);
624 err = kvm_set_user_memory_region(s, mem);
626 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
631 start_addr += old.memory_size;
632 ram += old.memory_size;
633 size -= old.memory_size;
637 /* register prefix slot */
638 if (old.start_addr < start_addr) {
639 mem = kvm_alloc_slot(s);
640 mem->memory_size = start_addr - old.start_addr;
641 mem->start_addr = old.start_addr;
643 mem->flags = kvm_mem_flags(s, log_dirty);
645 err = kvm_set_user_memory_region(s, mem);
647 fprintf(stderr, "%s: error registering prefix slot: %s\n",
648 __func__, strerror(-err));
650 fprintf(stderr, "%s: This is probably because your kernel's " \
651 "PAGE_SIZE is too big. Please try to use 4k " \
652 "PAGE_SIZE!\n", __func__);
658 /* register suffix slot */
659 if (old.start_addr + old.memory_size > start_addr + size) {
660 ram_addr_t size_delta;
662 mem = kvm_alloc_slot(s);
663 mem->start_addr = start_addr + size;
664 size_delta = mem->start_addr - old.start_addr;
665 mem->memory_size = old.memory_size - size_delta;
666 mem->ram = old.ram + size_delta;
667 mem->flags = kvm_mem_flags(s, log_dirty);
669 err = kvm_set_user_memory_region(s, mem);
671 fprintf(stderr, "%s: error registering suffix slot: %s\n",
672 __func__, strerror(-err));
678 /* in case the KVM bug workaround already "consumed" the new slot */
685 mem = kvm_alloc_slot(s);
686 mem->memory_size = size;
687 mem->start_addr = start_addr;
689 mem->flags = kvm_mem_flags(s, log_dirty);
691 err = kvm_set_user_memory_region(s, mem);
693 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
699 static void kvm_begin(MemoryListener *listener)
703 static void kvm_commit(MemoryListener *listener)
707 static void kvm_region_add(MemoryListener *listener,
708 MemoryRegionSection *section)
710 kvm_set_phys_mem(section, true);
713 static void kvm_region_del(MemoryListener *listener,
714 MemoryRegionSection *section)
716 kvm_set_phys_mem(section, false);
719 static void kvm_region_nop(MemoryListener *listener,
720 MemoryRegionSection *section)
724 static void kvm_log_sync(MemoryListener *listener,
725 MemoryRegionSection *section)
729 r = kvm_physical_sync_dirty_bitmap(section);
735 static void kvm_log_global_start(struct MemoryListener *listener)
739 r = kvm_set_migration_log(1);
743 static void kvm_log_global_stop(struct MemoryListener *listener)
747 r = kvm_set_migration_log(0);
751 static void kvm_mem_ioeventfd_add(MemoryRegionSection *section,
752 bool match_data, uint64_t data, int fd)
756 assert(match_data && section->size <= 8);
758 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
759 data, true, section->size);
765 static void kvm_mem_ioeventfd_del(MemoryRegionSection *section,
766 bool match_data, uint64_t data, int fd)
770 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
771 data, false, section->size);
777 static void kvm_io_ioeventfd_add(MemoryRegionSection *section,
778 bool match_data, uint64_t data, int fd)
782 assert(match_data && section->size == 2);
784 r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
791 static void kvm_io_ioeventfd_del(MemoryRegionSection *section,
792 bool match_data, uint64_t data, int fd)
797 r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
804 static void kvm_eventfd_add(MemoryListener *listener,
805 MemoryRegionSection *section,
806 bool match_data, uint64_t data, int fd)
808 if (section->address_space == get_system_memory()) {
809 kvm_mem_ioeventfd_add(section, match_data, data, fd);
811 kvm_io_ioeventfd_add(section, match_data, data, fd);
815 static void kvm_eventfd_del(MemoryListener *listener,
816 MemoryRegionSection *section,
817 bool match_data, uint64_t data, int fd)
819 if (section->address_space == get_system_memory()) {
820 kvm_mem_ioeventfd_del(section, match_data, data, fd);
822 kvm_io_ioeventfd_del(section, match_data, data, fd);
826 static MemoryListener kvm_memory_listener = {
828 .commit = kvm_commit,
829 .region_add = kvm_region_add,
830 .region_del = kvm_region_del,
831 .region_nop = kvm_region_nop,
832 .log_start = kvm_log_start,
833 .log_stop = kvm_log_stop,
834 .log_sync = kvm_log_sync,
835 .log_global_start = kvm_log_global_start,
836 .log_global_stop = kvm_log_global_stop,
837 .eventfd_add = kvm_eventfd_add,
838 .eventfd_del = kvm_eventfd_del,
842 static void kvm_handle_interrupt(CPUArchState *env, int mask)
844 env->interrupt_request |= mask;
846 if (!qemu_cpu_is_self(env)) {
851 int kvm_irqchip_set_irq(KVMState *s, int irq, int level)
853 struct kvm_irq_level event;
856 assert(kvm_irqchip_in_kernel());
860 ret = kvm_vm_ioctl(s, s->irqchip_inject_ioctl, &event);
862 perror("kvm_set_irqchip_line");
866 return (s->irqchip_inject_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
869 #ifdef KVM_CAP_IRQ_ROUTING
870 static void set_gsi(KVMState *s, unsigned int gsi)
872 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
875 static void clear_gsi(KVMState *s, unsigned int gsi)
877 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
880 static void kvm_init_irq_routing(KVMState *s)
884 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
886 unsigned int gsi_bits, i;
888 /* Round up so we can search ints using ffs */
889 gsi_bits = ALIGN(gsi_count, 32);
890 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
891 s->gsi_count = gsi_count;
893 /* Mark any over-allocated bits as already in use */
894 for (i = gsi_count; i < gsi_bits; i++) {
899 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
900 s->nr_allocated_irq_routes = 0;
902 if (!s->direct_msi) {
903 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
904 QTAILQ_INIT(&s->msi_hashtab[i]);
908 kvm_arch_init_irq_routing(s);
911 static void kvm_add_routing_entry(KVMState *s,
912 struct kvm_irq_routing_entry *entry)
914 struct kvm_irq_routing_entry *new;
917 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
918 n = s->nr_allocated_irq_routes * 2;
922 size = sizeof(struct kvm_irq_routing);
923 size += n * sizeof(*new);
924 s->irq_routes = g_realloc(s->irq_routes, size);
925 s->nr_allocated_irq_routes = n;
927 n = s->irq_routes->nr++;
928 new = &s->irq_routes->entries[n];
929 memset(new, 0, sizeof(*new));
930 new->gsi = entry->gsi;
931 new->type = entry->type;
932 new->flags = entry->flags;
935 set_gsi(s, entry->gsi);
938 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
940 struct kvm_irq_routing_entry e;
942 assert(pin < s->gsi_count);
945 e.type = KVM_IRQ_ROUTING_IRQCHIP;
947 e.u.irqchip.irqchip = irqchip;
948 e.u.irqchip.pin = pin;
949 kvm_add_routing_entry(s, &e);
952 int kvm_irqchip_commit_routes(KVMState *s)
954 s->irq_routes->flags = 0;
955 return kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
958 void kvm_irqchip_release_virq(KVMState *s, int virq)
960 struct kvm_irq_routing_entry *e;
963 for (i = 0; i < s->irq_routes->nr; i++) {
964 e = &s->irq_routes->entries[i];
965 if (e->gsi == virq) {
967 *e = s->irq_routes->entries[s->irq_routes->nr];
973 static unsigned int kvm_hash_msi(uint32_t data)
975 /* This is optimized for IA32 MSI layout. However, no other arch shall
976 * repeat the mistake of not providing a direct MSI injection API. */
980 static void kvm_flush_dynamic_msi_routes(KVMState *s)
982 KVMMSIRoute *route, *next;
985 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
986 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
987 kvm_irqchip_release_virq(s, route->kroute.gsi);
988 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
994 static int kvm_irqchip_get_virq(KVMState *s)
996 uint32_t *word = s->used_gsi_bitmap;
997 int max_words = ALIGN(s->gsi_count, 32) / 32;
1002 /* Return the lowest unused GSI in the bitmap */
1003 for (i = 0; i < max_words; i++) {
1004 bit = ffs(~word[i]);
1009 return bit - 1 + i * 32;
1011 if (!s->direct_msi && retry) {
1013 kvm_flush_dynamic_msi_routes(s);
1020 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1022 unsigned int hash = kvm_hash_msi(msg.data);
1025 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1026 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1027 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1028 route->kroute.u.msi.data == msg.data) {
1035 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1040 if (s->direct_msi) {
1041 msi.address_lo = (uint32_t)msg.address;
1042 msi.address_hi = msg.address >> 32;
1043 msi.data = msg.data;
1045 memset(msi.pad, 0, sizeof(msi.pad));
1047 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1050 route = kvm_lookup_msi_route(s, msg);
1054 virq = kvm_irqchip_get_virq(s);
1059 route = g_malloc(sizeof(KVMMSIRoute));
1060 route->kroute.gsi = virq;
1061 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1062 route->kroute.flags = 0;
1063 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1064 route->kroute.u.msi.address_hi = msg.address >> 32;
1065 route->kroute.u.msi.data = msg.data;
1067 kvm_add_routing_entry(s, &route->kroute);
1069 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1072 ret = kvm_irqchip_commit_routes(s);
1078 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1080 return kvm_irqchip_set_irq(s, route->kroute.gsi, 1);
1083 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1085 struct kvm_irq_routing_entry kroute;
1088 if (!kvm_irqchip_in_kernel()) {
1092 virq = kvm_irqchip_get_virq(s);
1098 kroute.type = KVM_IRQ_ROUTING_MSI;
1100 kroute.u.msi.address_lo = (uint32_t)msg.address;
1101 kroute.u.msi.address_hi = msg.address >> 32;
1102 kroute.u.msi.data = msg.data;
1104 kvm_add_routing_entry(s, &kroute);
1109 #else /* !KVM_CAP_IRQ_ROUTING */
1111 static void kvm_init_irq_routing(KVMState *s)
1115 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1120 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1124 #endif /* !KVM_CAP_IRQ_ROUTING */
1126 static int kvm_irqchip_create(KVMState *s)
1128 QemuOptsList *list = qemu_find_opts("machine");
1131 if (QTAILQ_EMPTY(&list->head) ||
1132 !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1133 "kernel_irqchip", true) ||
1134 !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1138 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1140 fprintf(stderr, "Create kernel irqchip failed\n");
1144 s->irqchip_inject_ioctl = KVM_IRQ_LINE;
1145 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1146 s->irqchip_inject_ioctl = KVM_IRQ_LINE_STATUS;
1148 kvm_kernel_irqchip = true;
1150 kvm_init_irq_routing(s);
1157 static const char upgrade_note[] =
1158 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1159 "(see http://sourceforge.net/projects/kvm).\n";
1161 const KVMCapabilityInfo *missing_cap;
1165 s = g_malloc0(sizeof(KVMState));
1168 * On systems where the kernel can support different base page
1169 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1170 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1171 * page size for the system though.
1173 assert(TARGET_PAGE_SIZE <= getpagesize());
1175 #ifdef KVM_CAP_SET_GUEST_DEBUG
1176 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1178 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1179 s->slots[i].slot = i;
1182 s->fd = qemu_open("/dev/kvm", O_RDWR);
1184 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1189 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1190 if (ret < KVM_API_VERSION) {
1194 fprintf(stderr, "kvm version too old\n");
1198 if (ret > KVM_API_VERSION) {
1200 fprintf(stderr, "kvm version not supported\n");
1204 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1207 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1208 "your host kernel command line\n");
1214 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1217 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1221 fprintf(stderr, "kvm does not support %s\n%s",
1222 missing_cap->name, upgrade_note);
1226 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1228 s->broken_set_mem_region = 1;
1229 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1231 s->broken_set_mem_region = 0;
1234 #ifdef KVM_CAP_VCPU_EVENTS
1235 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1238 s->robust_singlestep =
1239 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1241 #ifdef KVM_CAP_DEBUGREGS
1242 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1245 #ifdef KVM_CAP_XSAVE
1246 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1250 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1253 #ifdef KVM_CAP_PIT_STATE2
1254 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1257 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1259 ret = kvm_arch_init(s);
1264 ret = kvm_irqchip_create(s);
1270 memory_listener_register(&kvm_memory_listener, NULL);
1272 s->many_ioeventfds = kvm_check_many_ioeventfds();
1274 cpu_interrupt_handler = kvm_handle_interrupt;
1292 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1296 uint8_t *ptr = data;
1298 for (i = 0; i < count; i++) {
1299 if (direction == KVM_EXIT_IO_IN) {
1302 stb_p(ptr, cpu_inb(port));
1305 stw_p(ptr, cpu_inw(port));
1308 stl_p(ptr, cpu_inl(port));
1314 cpu_outb(port, ldub_p(ptr));
1317 cpu_outw(port, lduw_p(ptr));
1320 cpu_outl(port, ldl_p(ptr));
1329 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1331 fprintf(stderr, "KVM internal error.");
1332 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1335 fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1336 for (i = 0; i < run->internal.ndata; ++i) {
1337 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1338 i, (uint64_t)run->internal.data[i]);
1341 fprintf(stderr, "\n");
1343 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1344 fprintf(stderr, "emulation failure\n");
1345 if (!kvm_arch_stop_on_emulation_error(env)) {
1346 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1347 return EXCP_INTERRUPT;
1350 /* FIXME: Should trigger a qmp message to let management know
1351 * something went wrong.
1356 void kvm_flush_coalesced_mmio_buffer(void)
1358 KVMState *s = kvm_state;
1360 if (s->coalesced_flush_in_progress) {
1364 s->coalesced_flush_in_progress = true;
1366 if (s->coalesced_mmio_ring) {
1367 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1368 while (ring->first != ring->last) {
1369 struct kvm_coalesced_mmio *ent;
1371 ent = &ring->coalesced_mmio[ring->first];
1373 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1375 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1379 s->coalesced_flush_in_progress = false;
1382 static void do_kvm_cpu_synchronize_state(void *_env)
1384 CPUArchState *env = _env;
1386 if (!env->kvm_vcpu_dirty) {
1387 kvm_arch_get_registers(env);
1388 env->kvm_vcpu_dirty = 1;
1392 void kvm_cpu_synchronize_state(CPUArchState *env)
1394 if (!env->kvm_vcpu_dirty) {
1395 run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
1399 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1401 kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
1402 env->kvm_vcpu_dirty = 0;
1405 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1407 kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
1408 env->kvm_vcpu_dirty = 0;
1411 int kvm_cpu_exec(CPUArchState *env)
1413 struct kvm_run *run = env->kvm_run;
1416 DPRINTF("kvm_cpu_exec()\n");
1418 if (kvm_arch_process_async_events(env)) {
1419 env->exit_request = 0;
1424 if (env->kvm_vcpu_dirty) {
1425 kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
1426 env->kvm_vcpu_dirty = 0;
1429 kvm_arch_pre_run(env, run);
1430 if (env->exit_request) {
1431 DPRINTF("interrupt exit requested\n");
1433 * KVM requires us to reenter the kernel after IO exits to complete
1434 * instruction emulation. This self-signal will ensure that we
1437 qemu_cpu_kick_self();
1439 qemu_mutex_unlock_iothread();
1441 run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
1443 qemu_mutex_lock_iothread();
1444 kvm_arch_post_run(env, run);
1446 kvm_flush_coalesced_mmio_buffer();
1449 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1450 DPRINTF("io window exit\n");
1451 ret = EXCP_INTERRUPT;
1454 fprintf(stderr, "error: kvm run failed %s\n",
1455 strerror(-run_ret));
1459 switch (run->exit_reason) {
1461 DPRINTF("handle_io\n");
1462 kvm_handle_io(run->io.port,
1463 (uint8_t *)run + run->io.data_offset,
1470 DPRINTF("handle_mmio\n");
1471 cpu_physical_memory_rw(run->mmio.phys_addr,
1474 run->mmio.is_write);
1477 case KVM_EXIT_IRQ_WINDOW_OPEN:
1478 DPRINTF("irq_window_open\n");
1479 ret = EXCP_INTERRUPT;
1481 case KVM_EXIT_SHUTDOWN:
1482 DPRINTF("shutdown\n");
1483 qemu_system_reset_request();
1484 ret = EXCP_INTERRUPT;
1486 case KVM_EXIT_UNKNOWN:
1487 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1488 (uint64_t)run->hw.hardware_exit_reason);
1491 case KVM_EXIT_INTERNAL_ERROR:
1492 ret = kvm_handle_internal_error(env, run);
1495 DPRINTF("kvm_arch_handle_exit\n");
1496 ret = kvm_arch_handle_exit(env, run);
1502 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1503 vm_stop(RUN_STATE_INTERNAL_ERROR);
1506 env->exit_request = 0;
1510 int kvm_ioctl(KVMState *s, int type, ...)
1517 arg = va_arg(ap, void *);
1520 ret = ioctl(s->fd, type, arg);
1527 int kvm_vm_ioctl(KVMState *s, int type, ...)
1534 arg = va_arg(ap, void *);
1537 ret = ioctl(s->vmfd, type, arg);
1544 int kvm_vcpu_ioctl(CPUArchState *env, int type, ...)
1551 arg = va_arg(ap, void *);
1554 ret = ioctl(env->kvm_fd, type, arg);
1561 int kvm_has_sync_mmu(void)
1563 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1566 int kvm_has_vcpu_events(void)
1568 return kvm_state->vcpu_events;
1571 int kvm_has_robust_singlestep(void)
1573 return kvm_state->robust_singlestep;
1576 int kvm_has_debugregs(void)
1578 return kvm_state->debugregs;
1581 int kvm_has_xsave(void)
1583 return kvm_state->xsave;
1586 int kvm_has_xcrs(void)
1588 return kvm_state->xcrs;
1591 int kvm_has_pit_state2(void)
1593 return kvm_state->pit_state2;
1596 int kvm_has_many_ioeventfds(void)
1598 if (!kvm_enabled()) {
1601 return kvm_state->many_ioeventfds;
1604 int kvm_has_gsi_routing(void)
1606 #ifdef KVM_CAP_IRQ_ROUTING
1607 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1613 int kvm_allows_irq0_override(void)
1615 return !kvm_irqchip_in_kernel() || kvm_has_gsi_routing();
1618 void kvm_setup_guest_memory(void *start, size_t size)
1620 if (!kvm_has_sync_mmu()) {
1621 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1624 perror("qemu_madvise");
1626 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1632 #ifdef KVM_CAP_SET_GUEST_DEBUG
1633 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUArchState *env,
1636 struct kvm_sw_breakpoint *bp;
1638 QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1646 int kvm_sw_breakpoints_active(CPUArchState *env)
1648 return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1651 struct kvm_set_guest_debug_data {
1652 struct kvm_guest_debug dbg;
1657 static void kvm_invoke_set_guest_debug(void *data)
1659 struct kvm_set_guest_debug_data *dbg_data = data;
1660 CPUArchState *env = dbg_data->env;
1662 dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1665 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1667 struct kvm_set_guest_debug_data data;
1669 data.dbg.control = reinject_trap;
1671 if (env->singlestep_enabled) {
1672 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1674 kvm_arch_update_guest_debug(env, &data.dbg);
1677 run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1681 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1682 target_ulong len, int type)
1684 struct kvm_sw_breakpoint *bp;
1688 if (type == GDB_BREAKPOINT_SW) {
1689 bp = kvm_find_sw_breakpoint(current_env, addr);
1695 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1702 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1708 QTAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
1711 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1717 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1718 err = kvm_update_guest_debug(env, 0);
1726 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1727 target_ulong len, int type)
1729 struct kvm_sw_breakpoint *bp;
1733 if (type == GDB_BREAKPOINT_SW) {
1734 bp = kvm_find_sw_breakpoint(current_env, addr);
1739 if (bp->use_count > 1) {
1744 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1749 QTAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1752 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1758 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1759 err = kvm_update_guest_debug(env, 0);
1767 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1769 struct kvm_sw_breakpoint *bp, *next;
1770 KVMState *s = current_env->kvm_state;
1773 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1774 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1775 /* Try harder to find a CPU that currently sees the breakpoint. */
1776 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1777 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1783 kvm_arch_remove_all_hw_breakpoints();
1785 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1786 kvm_update_guest_debug(env, 0);
1790 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1792 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1797 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1798 target_ulong len, int type)
1803 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1804 target_ulong len, int type)
1809 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1812 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1814 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1816 struct kvm_signal_mask *sigmask;
1820 return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1823 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1826 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1827 r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1833 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1837 struct kvm_ioeventfd iofd;
1839 iofd.datamatch = val;
1842 iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1845 if (!kvm_enabled()) {
1850 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1853 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1862 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1864 struct kvm_ioeventfd kick = {
1868 .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1872 if (!kvm_enabled()) {
1876 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1878 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1885 int kvm_on_sigbus_vcpu(CPUArchState *env, int code, void *addr)
1887 return kvm_arch_on_sigbus_vcpu(env, code, addr);
1890 int kvm_on_sigbus(int code, void *addr)
1892 return kvm_arch_on_sigbus(code, addr);