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"
25 #include "qemu-option.h"
26 #include "qemu-config.h"
34 #include "exec-memory.h"
36 /* This check must be after config-host.h is included */
38 #include <sys/eventfd.h>
41 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
42 #define PAGE_SIZE TARGET_PAGE_SIZE
47 #define DPRINTF(fmt, ...) \
48 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
50 #define DPRINTF(fmt, ...) \
54 #define KVM_MSI_HASHTAB_SIZE 256
56 typedef struct KVMSlot
58 target_phys_addr_t start_addr;
59 ram_addr_t memory_size;
65 typedef struct kvm_dirty_log KVMDirtyLog;
73 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
74 bool coalesced_flush_in_progress;
75 int broken_set_mem_region;
78 int robust_singlestep;
80 #ifdef KVM_CAP_SET_GUEST_DEBUG
81 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
86 /* The man page (and posix) say ioctl numbers are signed int, but
87 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
88 * unsigned, and treating them as signed here can break things */
89 unsigned irqchip_inject_ioctl;
90 #ifdef KVM_CAP_IRQ_ROUTING
91 struct kvm_irq_routing *irq_routes;
92 int nr_allocated_irq_routes;
93 uint32_t *used_gsi_bitmap;
94 unsigned int gsi_count;
95 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
101 bool kvm_kernel_irqchip;
103 static const KVMCapabilityInfo kvm_required_capabilites[] = {
104 KVM_CAP_INFO(USER_MEMORY),
105 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
109 static KVMSlot *kvm_alloc_slot(KVMState *s)
113 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
114 if (s->slots[i].memory_size == 0) {
119 fprintf(stderr, "%s: no free slot available\n", __func__);
123 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
124 target_phys_addr_t start_addr,
125 target_phys_addr_t end_addr)
129 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
130 KVMSlot *mem = &s->slots[i];
132 if (start_addr == mem->start_addr &&
133 end_addr == mem->start_addr + mem->memory_size) {
142 * Find overlapping slot with lowest start address
144 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
145 target_phys_addr_t start_addr,
146 target_phys_addr_t end_addr)
148 KVMSlot *found = NULL;
151 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
152 KVMSlot *mem = &s->slots[i];
154 if (mem->memory_size == 0 ||
155 (found && found->start_addr < mem->start_addr)) {
159 if (end_addr > mem->start_addr &&
160 start_addr < mem->start_addr + mem->memory_size) {
168 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
169 target_phys_addr_t *phys_addr)
173 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
174 KVMSlot *mem = &s->slots[i];
176 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
177 *phys_addr = mem->start_addr + (ram - mem->ram);
185 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
187 struct kvm_userspace_memory_region mem;
189 mem.slot = slot->slot;
190 mem.guest_phys_addr = slot->start_addr;
191 mem.memory_size = slot->memory_size;
192 mem.userspace_addr = (unsigned long)slot->ram;
193 mem.flags = slot->flags;
194 if (s->migration_log) {
195 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
197 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
200 static void kvm_reset_vcpu(void *opaque)
202 CPUArchState *env = opaque;
204 kvm_arch_reset_vcpu(env);
207 int kvm_init_vcpu(CPUArchState *env)
209 KVMState *s = kvm_state;
213 DPRINTF("kvm_init_vcpu\n");
215 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
217 DPRINTF("kvm_create_vcpu failed\n");
223 env->kvm_vcpu_dirty = 1;
225 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
228 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
232 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
234 if (env->kvm_run == MAP_FAILED) {
236 DPRINTF("mmap'ing vcpu state failed\n");
240 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
241 s->coalesced_mmio_ring =
242 (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
245 ret = kvm_arch_init_vcpu(env);
247 qemu_register_reset(kvm_reset_vcpu, env);
248 kvm_arch_reset_vcpu(env);
255 * dirty pages logging control
258 static int kvm_mem_flags(KVMState *s, bool log_dirty)
260 return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
263 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
265 KVMState *s = kvm_state;
266 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
269 old_flags = mem->flags;
271 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
274 /* If nothing changed effectively, no need to issue ioctl */
275 if (s->migration_log) {
276 flags |= KVM_MEM_LOG_DIRTY_PAGES;
279 if (flags == old_flags) {
283 return kvm_set_user_memory_region(s, mem);
286 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
287 ram_addr_t size, bool log_dirty)
289 KVMState *s = kvm_state;
290 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
293 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
294 TARGET_FMT_plx "\n", __func__, phys_addr,
295 (target_phys_addr_t)(phys_addr + size - 1));
298 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
301 static void kvm_log_start(MemoryListener *listener,
302 MemoryRegionSection *section)
306 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
307 section->size, true);
313 static void kvm_log_stop(MemoryListener *listener,
314 MemoryRegionSection *section)
318 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
319 section->size, false);
325 static int kvm_set_migration_log(int enable)
327 KVMState *s = kvm_state;
331 s->migration_log = enable;
333 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
336 if (!mem->memory_size) {
339 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
342 err = kvm_set_user_memory_region(s, mem);
350 /* get kvm's dirty pages bitmap and update qemu's */
351 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
352 unsigned long *bitmap)
355 unsigned long page_number, c;
356 target_phys_addr_t addr, addr1;
357 unsigned int len = ((section->size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
358 unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
361 * bitmap-traveling is faster than memory-traveling (for addr...)
362 * especially when most of the memory is not dirty.
364 for (i = 0; i < len; i++) {
365 if (bitmap[i] != 0) {
366 c = leul_to_cpu(bitmap[i]);
370 page_number = (i * HOST_LONG_BITS + j) * hpratio;
371 addr1 = page_number * TARGET_PAGE_SIZE;
372 addr = section->offset_within_region + addr1;
373 memory_region_set_dirty(section->mr, addr,
374 TARGET_PAGE_SIZE * hpratio);
381 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
384 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
385 * This function updates qemu's dirty bitmap using
386 * memory_region_set_dirty(). This means all bits are set
389 * @start_add: start of logged region.
390 * @end_addr: end of logged region.
392 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
394 KVMState *s = kvm_state;
395 unsigned long size, allocated_size = 0;
399 target_phys_addr_t start_addr = section->offset_within_address_space;
400 target_phys_addr_t end_addr = start_addr + section->size;
402 d.dirty_bitmap = NULL;
403 while (start_addr < end_addr) {
404 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
409 /* XXX bad kernel interface alert
410 * For dirty bitmap, kernel allocates array of size aligned to
411 * bits-per-long. But for case when the kernel is 64bits and
412 * the userspace is 32bits, userspace can't align to the same
413 * bits-per-long, since sizeof(long) is different between kernel
414 * and user space. This way, userspace will provide buffer which
415 * may be 4 bytes less than the kernel will use, resulting in
416 * userspace memory corruption (which is not detectable by valgrind
417 * too, in most cases).
418 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
419 * a hope that sizeof(long) wont become >8 any time soon.
421 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
422 /*HOST_LONG_BITS*/ 64) / 8;
423 if (!d.dirty_bitmap) {
424 d.dirty_bitmap = g_malloc(size);
425 } else if (size > allocated_size) {
426 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
428 allocated_size = size;
429 memset(d.dirty_bitmap, 0, allocated_size);
433 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
434 DPRINTF("ioctl failed %d\n", errno);
439 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
440 start_addr = mem->start_addr + mem->memory_size;
442 g_free(d.dirty_bitmap);
447 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
450 KVMState *s = kvm_state;
452 if (s->coalesced_mmio) {
453 struct kvm_coalesced_mmio_zone zone;
459 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
465 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
468 KVMState *s = kvm_state;
470 if (s->coalesced_mmio) {
471 struct kvm_coalesced_mmio_zone zone;
477 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
483 int kvm_check_extension(KVMState *s, unsigned int extension)
487 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
495 static int kvm_check_many_ioeventfds(void)
497 /* Userspace can use ioeventfd for io notification. This requires a host
498 * that supports eventfd(2) and an I/O thread; since eventfd does not
499 * support SIGIO it cannot interrupt the vcpu.
501 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
502 * can avoid creating too many ioeventfds.
504 #if defined(CONFIG_EVENTFD)
507 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
508 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
509 if (ioeventfds[i] < 0) {
512 ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
514 close(ioeventfds[i]);
519 /* Decide whether many devices are supported or not */
520 ret = i == ARRAY_SIZE(ioeventfds);
523 kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
524 close(ioeventfds[i]);
532 static const KVMCapabilityInfo *
533 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
536 if (!kvm_check_extension(s, list->value)) {
544 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
546 KVMState *s = kvm_state;
549 MemoryRegion *mr = section->mr;
550 bool log_dirty = memory_region_is_logging(mr);
551 target_phys_addr_t start_addr = section->offset_within_address_space;
552 ram_addr_t size = section->size;
556 /* kvm works in page size chunks, but the function may be called
557 with sub-page size and unaligned start address. */
558 delta = TARGET_PAGE_ALIGN(size) - size;
564 size &= TARGET_PAGE_MASK;
565 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
569 if (!memory_region_is_ram(mr)) {
573 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
576 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
581 if (add && start_addr >= mem->start_addr &&
582 (start_addr + size <= mem->start_addr + mem->memory_size) &&
583 (ram - start_addr == mem->ram - mem->start_addr)) {
584 /* The new slot fits into the existing one and comes with
585 * identical parameters - update flags and done. */
586 kvm_slot_dirty_pages_log_change(mem, log_dirty);
592 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
593 kvm_physical_sync_dirty_bitmap(section);
596 /* unregister the overlapping slot */
597 mem->memory_size = 0;
598 err = kvm_set_user_memory_region(s, mem);
600 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
601 __func__, strerror(-err));
605 /* Workaround for older KVM versions: we can't join slots, even not by
606 * unregistering the previous ones and then registering the larger
607 * slot. We have to maintain the existing fragmentation. Sigh.
609 * This workaround assumes that the new slot starts at the same
610 * address as the first existing one. If not or if some overlapping
611 * slot comes around later, we will fail (not seen in practice so far)
612 * - and actually require a recent KVM version. */
613 if (s->broken_set_mem_region &&
614 old.start_addr == start_addr && old.memory_size < size && add) {
615 mem = kvm_alloc_slot(s);
616 mem->memory_size = old.memory_size;
617 mem->start_addr = old.start_addr;
619 mem->flags = kvm_mem_flags(s, log_dirty);
621 err = kvm_set_user_memory_region(s, mem);
623 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
628 start_addr += old.memory_size;
629 ram += old.memory_size;
630 size -= old.memory_size;
634 /* register prefix slot */
635 if (old.start_addr < start_addr) {
636 mem = kvm_alloc_slot(s);
637 mem->memory_size = start_addr - old.start_addr;
638 mem->start_addr = old.start_addr;
640 mem->flags = kvm_mem_flags(s, log_dirty);
642 err = kvm_set_user_memory_region(s, mem);
644 fprintf(stderr, "%s: error registering prefix slot: %s\n",
645 __func__, strerror(-err));
647 fprintf(stderr, "%s: This is probably because your kernel's " \
648 "PAGE_SIZE is too big. Please try to use 4k " \
649 "PAGE_SIZE!\n", __func__);
655 /* register suffix slot */
656 if (old.start_addr + old.memory_size > start_addr + size) {
657 ram_addr_t size_delta;
659 mem = kvm_alloc_slot(s);
660 mem->start_addr = start_addr + size;
661 size_delta = mem->start_addr - old.start_addr;
662 mem->memory_size = old.memory_size - size_delta;
663 mem->ram = old.ram + size_delta;
664 mem->flags = kvm_mem_flags(s, log_dirty);
666 err = kvm_set_user_memory_region(s, mem);
668 fprintf(stderr, "%s: error registering suffix slot: %s\n",
669 __func__, strerror(-err));
675 /* in case the KVM bug workaround already "consumed" the new slot */
682 mem = kvm_alloc_slot(s);
683 mem->memory_size = size;
684 mem->start_addr = start_addr;
686 mem->flags = kvm_mem_flags(s, log_dirty);
688 err = kvm_set_user_memory_region(s, mem);
690 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
696 static void kvm_begin(MemoryListener *listener)
700 static void kvm_commit(MemoryListener *listener)
704 static void kvm_region_add(MemoryListener *listener,
705 MemoryRegionSection *section)
707 kvm_set_phys_mem(section, true);
710 static void kvm_region_del(MemoryListener *listener,
711 MemoryRegionSection *section)
713 kvm_set_phys_mem(section, false);
716 static void kvm_region_nop(MemoryListener *listener,
717 MemoryRegionSection *section)
721 static void kvm_log_sync(MemoryListener *listener,
722 MemoryRegionSection *section)
726 r = kvm_physical_sync_dirty_bitmap(section);
732 static void kvm_log_global_start(struct MemoryListener *listener)
736 r = kvm_set_migration_log(1);
740 static void kvm_log_global_stop(struct MemoryListener *listener)
744 r = kvm_set_migration_log(0);
748 static void kvm_mem_ioeventfd_add(MemoryRegionSection *section,
749 bool match_data, uint64_t data, int fd)
753 assert(match_data && section->size <= 8);
755 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
756 data, true, section->size);
762 static void kvm_mem_ioeventfd_del(MemoryRegionSection *section,
763 bool match_data, uint64_t data, int fd)
767 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
768 data, false, section->size);
774 static void kvm_io_ioeventfd_add(MemoryRegionSection *section,
775 bool match_data, uint64_t data, int fd)
779 assert(match_data && section->size == 2);
781 r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
788 static void kvm_io_ioeventfd_del(MemoryRegionSection *section,
789 bool match_data, uint64_t data, int fd)
794 r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
801 static void kvm_eventfd_add(MemoryListener *listener,
802 MemoryRegionSection *section,
803 bool match_data, uint64_t data, int fd)
805 if (section->address_space == get_system_memory()) {
806 kvm_mem_ioeventfd_add(section, match_data, data, fd);
808 kvm_io_ioeventfd_add(section, match_data, data, fd);
812 static void kvm_eventfd_del(MemoryListener *listener,
813 MemoryRegionSection *section,
814 bool match_data, uint64_t data, int fd)
816 if (section->address_space == get_system_memory()) {
817 kvm_mem_ioeventfd_del(section, match_data, data, fd);
819 kvm_io_ioeventfd_del(section, match_data, data, fd);
823 static MemoryListener kvm_memory_listener = {
825 .commit = kvm_commit,
826 .region_add = kvm_region_add,
827 .region_del = kvm_region_del,
828 .region_nop = kvm_region_nop,
829 .log_start = kvm_log_start,
830 .log_stop = kvm_log_stop,
831 .log_sync = kvm_log_sync,
832 .log_global_start = kvm_log_global_start,
833 .log_global_stop = kvm_log_global_stop,
834 .eventfd_add = kvm_eventfd_add,
835 .eventfd_del = kvm_eventfd_del,
839 static void kvm_handle_interrupt(CPUArchState *env, int mask)
841 env->interrupt_request |= mask;
843 if (!qemu_cpu_is_self(env)) {
848 int kvm_irqchip_set_irq(KVMState *s, int irq, int level)
850 struct kvm_irq_level event;
853 assert(kvm_irqchip_in_kernel());
857 ret = kvm_vm_ioctl(s, s->irqchip_inject_ioctl, &event);
859 perror("kvm_set_irqchip_line");
863 return (s->irqchip_inject_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
866 #ifdef KVM_CAP_IRQ_ROUTING
867 typedef struct KVMMSIRoute {
868 struct kvm_irq_routing_entry kroute;
869 QTAILQ_ENTRY(KVMMSIRoute) entry;
872 static void set_gsi(KVMState *s, unsigned int gsi)
874 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
877 static void clear_gsi(KVMState *s, unsigned int gsi)
879 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
882 static void kvm_init_irq_routing(KVMState *s)
886 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
888 unsigned int gsi_bits, i;
890 /* Round up so we can search ints using ffs */
891 gsi_bits = ALIGN(gsi_count, 32);
892 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
893 s->gsi_count = gsi_count;
895 /* Mark any over-allocated bits as already in use */
896 for (i = gsi_count; i < gsi_bits; i++) {
901 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
902 s->nr_allocated_irq_routes = 0;
904 if (!s->direct_msi) {
905 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
906 QTAILQ_INIT(&s->msi_hashtab[i]);
910 kvm_arch_init_irq_routing(s);
913 static void kvm_irqchip_commit_routes(KVMState *s)
917 s->irq_routes->flags = 0;
918 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
922 static void kvm_add_routing_entry(KVMState *s,
923 struct kvm_irq_routing_entry *entry)
925 struct kvm_irq_routing_entry *new;
928 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
929 n = s->nr_allocated_irq_routes * 2;
933 size = sizeof(struct kvm_irq_routing);
934 size += n * sizeof(*new);
935 s->irq_routes = g_realloc(s->irq_routes, size);
936 s->nr_allocated_irq_routes = n;
938 n = s->irq_routes->nr++;
939 new = &s->irq_routes->entries[n];
940 memset(new, 0, sizeof(*new));
941 new->gsi = entry->gsi;
942 new->type = entry->type;
943 new->flags = entry->flags;
946 set_gsi(s, entry->gsi);
948 kvm_irqchip_commit_routes(s);
951 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
953 struct kvm_irq_routing_entry e;
955 assert(pin < s->gsi_count);
958 e.type = KVM_IRQ_ROUTING_IRQCHIP;
960 e.u.irqchip.irqchip = irqchip;
961 e.u.irqchip.pin = pin;
962 kvm_add_routing_entry(s, &e);
965 void kvm_irqchip_release_virq(KVMState *s, int virq)
967 struct kvm_irq_routing_entry *e;
970 for (i = 0; i < s->irq_routes->nr; i++) {
971 e = &s->irq_routes->entries[i];
972 if (e->gsi == virq) {
974 *e = s->irq_routes->entries[s->irq_routes->nr];
979 kvm_irqchip_commit_routes(s);
982 static unsigned int kvm_hash_msi(uint32_t data)
984 /* This is optimized for IA32 MSI layout. However, no other arch shall
985 * repeat the mistake of not providing a direct MSI injection API. */
989 static void kvm_flush_dynamic_msi_routes(KVMState *s)
991 KVMMSIRoute *route, *next;
994 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
995 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
996 kvm_irqchip_release_virq(s, route->kroute.gsi);
997 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1003 static int kvm_irqchip_get_virq(KVMState *s)
1005 uint32_t *word = s->used_gsi_bitmap;
1006 int max_words = ALIGN(s->gsi_count, 32) / 32;
1011 /* Return the lowest unused GSI in the bitmap */
1012 for (i = 0; i < max_words; i++) {
1013 bit = ffs(~word[i]);
1018 return bit - 1 + i * 32;
1020 if (!s->direct_msi && retry) {
1022 kvm_flush_dynamic_msi_routes(s);
1029 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1031 unsigned int hash = kvm_hash_msi(msg.data);
1034 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1035 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1036 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1037 route->kroute.u.msi.data == msg.data) {
1044 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1049 if (s->direct_msi) {
1050 msi.address_lo = (uint32_t)msg.address;
1051 msi.address_hi = msg.address >> 32;
1052 msi.data = msg.data;
1054 memset(msi.pad, 0, sizeof(msi.pad));
1056 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1059 route = kvm_lookup_msi_route(s, msg);
1063 virq = kvm_irqchip_get_virq(s);
1068 route = g_malloc(sizeof(KVMMSIRoute));
1069 route->kroute.gsi = virq;
1070 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1071 route->kroute.flags = 0;
1072 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1073 route->kroute.u.msi.address_hi = msg.address >> 32;
1074 route->kroute.u.msi.data = msg.data;
1076 kvm_add_routing_entry(s, &route->kroute);
1078 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1082 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1084 return kvm_irqchip_set_irq(s, route->kroute.gsi, 1);
1087 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1089 struct kvm_irq_routing_entry kroute;
1092 if (!kvm_irqchip_in_kernel()) {
1096 virq = kvm_irqchip_get_virq(s);
1102 kroute.type = KVM_IRQ_ROUTING_MSI;
1104 kroute.u.msi.address_lo = (uint32_t)msg.address;
1105 kroute.u.msi.address_hi = msg.address >> 32;
1106 kroute.u.msi.data = msg.data;
1108 kvm_add_routing_entry(s, &kroute);
1113 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1115 struct kvm_irqfd irqfd = {
1118 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1121 if (!kvm_irqchip_in_kernel()) {
1125 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1128 #else /* !KVM_CAP_IRQ_ROUTING */
1130 static void kvm_init_irq_routing(KVMState *s)
1134 void kvm_irqchip_release_virq(KVMState *s, int virq)
1138 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1143 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1148 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1152 #endif /* !KVM_CAP_IRQ_ROUTING */
1154 int kvm_irqchip_add_irqfd(KVMState *s, int fd, int virq)
1156 return kvm_irqchip_assign_irqfd(s, fd, virq, true);
1159 int kvm_irqchip_remove_irqfd(KVMState *s, int fd, int virq)
1161 return kvm_irqchip_assign_irqfd(s, fd, virq, false);
1164 static int kvm_irqchip_create(KVMState *s)
1166 QemuOptsList *list = qemu_find_opts("machine");
1169 if (QTAILQ_EMPTY(&list->head) ||
1170 !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1171 "kernel_irqchip", true) ||
1172 !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1176 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1178 fprintf(stderr, "Create kernel irqchip failed\n");
1182 s->irqchip_inject_ioctl = KVM_IRQ_LINE;
1183 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1184 s->irqchip_inject_ioctl = KVM_IRQ_LINE_STATUS;
1186 kvm_kernel_irqchip = true;
1188 kvm_init_irq_routing(s);
1195 static const char upgrade_note[] =
1196 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1197 "(see http://sourceforge.net/projects/kvm).\n";
1199 const KVMCapabilityInfo *missing_cap;
1203 s = g_malloc0(sizeof(KVMState));
1206 * On systems where the kernel can support different base page
1207 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1208 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1209 * page size for the system though.
1211 assert(TARGET_PAGE_SIZE <= getpagesize());
1213 #ifdef KVM_CAP_SET_GUEST_DEBUG
1214 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1216 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1217 s->slots[i].slot = i;
1220 s->fd = qemu_open("/dev/kvm", O_RDWR);
1222 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1227 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1228 if (ret < KVM_API_VERSION) {
1232 fprintf(stderr, "kvm version too old\n");
1236 if (ret > KVM_API_VERSION) {
1238 fprintf(stderr, "kvm version not supported\n");
1242 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1245 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1246 "your host kernel command line\n");
1252 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1255 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1259 fprintf(stderr, "kvm does not support %s\n%s",
1260 missing_cap->name, upgrade_note);
1264 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1266 s->broken_set_mem_region = 1;
1267 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1269 s->broken_set_mem_region = 0;
1272 #ifdef KVM_CAP_VCPU_EVENTS
1273 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1276 s->robust_singlestep =
1277 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1279 #ifdef KVM_CAP_DEBUGREGS
1280 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1283 #ifdef KVM_CAP_XSAVE
1284 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1288 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1291 #ifdef KVM_CAP_PIT_STATE2
1292 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1295 #ifdef KVM_CAP_IRQ_ROUTING
1296 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1299 ret = kvm_arch_init(s);
1304 ret = kvm_irqchip_create(s);
1310 memory_listener_register(&kvm_memory_listener, NULL);
1312 s->many_ioeventfds = kvm_check_many_ioeventfds();
1314 cpu_interrupt_handler = kvm_handle_interrupt;
1332 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1336 uint8_t *ptr = data;
1338 for (i = 0; i < count; i++) {
1339 if (direction == KVM_EXIT_IO_IN) {
1342 stb_p(ptr, cpu_inb(port));
1345 stw_p(ptr, cpu_inw(port));
1348 stl_p(ptr, cpu_inl(port));
1354 cpu_outb(port, ldub_p(ptr));
1357 cpu_outw(port, lduw_p(ptr));
1360 cpu_outl(port, ldl_p(ptr));
1369 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1371 fprintf(stderr, "KVM internal error.");
1372 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1375 fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1376 for (i = 0; i < run->internal.ndata; ++i) {
1377 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1378 i, (uint64_t)run->internal.data[i]);
1381 fprintf(stderr, "\n");
1383 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1384 fprintf(stderr, "emulation failure\n");
1385 if (!kvm_arch_stop_on_emulation_error(env)) {
1386 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1387 return EXCP_INTERRUPT;
1390 /* FIXME: Should trigger a qmp message to let management know
1391 * something went wrong.
1396 void kvm_flush_coalesced_mmio_buffer(void)
1398 KVMState *s = kvm_state;
1400 if (s->coalesced_flush_in_progress) {
1404 s->coalesced_flush_in_progress = true;
1406 if (s->coalesced_mmio_ring) {
1407 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1408 while (ring->first != ring->last) {
1409 struct kvm_coalesced_mmio *ent;
1411 ent = &ring->coalesced_mmio[ring->first];
1413 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1415 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1419 s->coalesced_flush_in_progress = false;
1422 static void do_kvm_cpu_synchronize_state(void *_env)
1424 CPUArchState *env = _env;
1426 if (!env->kvm_vcpu_dirty) {
1427 kvm_arch_get_registers(env);
1428 env->kvm_vcpu_dirty = 1;
1432 void kvm_cpu_synchronize_state(CPUArchState *env)
1434 if (!env->kvm_vcpu_dirty) {
1435 run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
1439 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1441 kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
1442 env->kvm_vcpu_dirty = 0;
1445 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1447 kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
1448 env->kvm_vcpu_dirty = 0;
1451 int kvm_cpu_exec(CPUArchState *env)
1453 struct kvm_run *run = env->kvm_run;
1456 DPRINTF("kvm_cpu_exec()\n");
1458 if (kvm_arch_process_async_events(env)) {
1459 env->exit_request = 0;
1464 if (env->kvm_vcpu_dirty) {
1465 kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
1466 env->kvm_vcpu_dirty = 0;
1469 kvm_arch_pre_run(env, run);
1470 if (env->exit_request) {
1471 DPRINTF("interrupt exit requested\n");
1473 * KVM requires us to reenter the kernel after IO exits to complete
1474 * instruction emulation. This self-signal will ensure that we
1477 qemu_cpu_kick_self();
1479 qemu_mutex_unlock_iothread();
1481 run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
1483 qemu_mutex_lock_iothread();
1484 kvm_arch_post_run(env, run);
1486 kvm_flush_coalesced_mmio_buffer();
1489 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1490 DPRINTF("io window exit\n");
1491 ret = EXCP_INTERRUPT;
1494 fprintf(stderr, "error: kvm run failed %s\n",
1495 strerror(-run_ret));
1499 switch (run->exit_reason) {
1501 DPRINTF("handle_io\n");
1502 kvm_handle_io(run->io.port,
1503 (uint8_t *)run + run->io.data_offset,
1510 DPRINTF("handle_mmio\n");
1511 cpu_physical_memory_rw(run->mmio.phys_addr,
1514 run->mmio.is_write);
1517 case KVM_EXIT_IRQ_WINDOW_OPEN:
1518 DPRINTF("irq_window_open\n");
1519 ret = EXCP_INTERRUPT;
1521 case KVM_EXIT_SHUTDOWN:
1522 DPRINTF("shutdown\n");
1523 qemu_system_reset_request();
1524 ret = EXCP_INTERRUPT;
1526 case KVM_EXIT_UNKNOWN:
1527 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1528 (uint64_t)run->hw.hardware_exit_reason);
1531 case KVM_EXIT_INTERNAL_ERROR:
1532 ret = kvm_handle_internal_error(env, run);
1535 DPRINTF("kvm_arch_handle_exit\n");
1536 ret = kvm_arch_handle_exit(env, run);
1542 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1543 vm_stop(RUN_STATE_INTERNAL_ERROR);
1546 env->exit_request = 0;
1550 int kvm_ioctl(KVMState *s, int type, ...)
1557 arg = va_arg(ap, void *);
1560 ret = ioctl(s->fd, type, arg);
1567 int kvm_vm_ioctl(KVMState *s, int type, ...)
1574 arg = va_arg(ap, void *);
1577 ret = ioctl(s->vmfd, type, arg);
1584 int kvm_vcpu_ioctl(CPUArchState *env, int type, ...)
1591 arg = va_arg(ap, void *);
1594 ret = ioctl(env->kvm_fd, type, arg);
1601 int kvm_has_sync_mmu(void)
1603 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1606 int kvm_has_vcpu_events(void)
1608 return kvm_state->vcpu_events;
1611 int kvm_has_robust_singlestep(void)
1613 return kvm_state->robust_singlestep;
1616 int kvm_has_debugregs(void)
1618 return kvm_state->debugregs;
1621 int kvm_has_xsave(void)
1623 return kvm_state->xsave;
1626 int kvm_has_xcrs(void)
1628 return kvm_state->xcrs;
1631 int kvm_has_pit_state2(void)
1633 return kvm_state->pit_state2;
1636 int kvm_has_many_ioeventfds(void)
1638 if (!kvm_enabled()) {
1641 return kvm_state->many_ioeventfds;
1644 int kvm_has_gsi_routing(void)
1646 #ifdef KVM_CAP_IRQ_ROUTING
1647 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1653 int kvm_allows_irq0_override(void)
1655 return !kvm_irqchip_in_kernel() || kvm_has_gsi_routing();
1658 void *kvm_vmalloc(ram_addr_t size)
1663 mem = kvm_arch_vmalloc(size);
1668 return qemu_vmalloc(size);
1671 void kvm_setup_guest_memory(void *start, size_t size)
1673 if (!kvm_has_sync_mmu()) {
1674 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1677 perror("qemu_madvise");
1679 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1685 #ifdef KVM_CAP_SET_GUEST_DEBUG
1686 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUArchState *env,
1689 struct kvm_sw_breakpoint *bp;
1691 QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1699 int kvm_sw_breakpoints_active(CPUArchState *env)
1701 return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1704 struct kvm_set_guest_debug_data {
1705 struct kvm_guest_debug dbg;
1710 static void kvm_invoke_set_guest_debug(void *data)
1712 struct kvm_set_guest_debug_data *dbg_data = data;
1713 CPUArchState *env = dbg_data->env;
1715 dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1718 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1720 struct kvm_set_guest_debug_data data;
1722 data.dbg.control = reinject_trap;
1724 if (env->singlestep_enabled) {
1725 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1727 kvm_arch_update_guest_debug(env, &data.dbg);
1730 run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1734 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1735 target_ulong len, int type)
1737 struct kvm_sw_breakpoint *bp;
1741 if (type == GDB_BREAKPOINT_SW) {
1742 bp = kvm_find_sw_breakpoint(current_env, addr);
1748 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1755 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1761 QTAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
1764 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1770 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1771 err = kvm_update_guest_debug(env, 0);
1779 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1780 target_ulong len, int type)
1782 struct kvm_sw_breakpoint *bp;
1786 if (type == GDB_BREAKPOINT_SW) {
1787 bp = kvm_find_sw_breakpoint(current_env, addr);
1792 if (bp->use_count > 1) {
1797 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1802 QTAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1805 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1811 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1812 err = kvm_update_guest_debug(env, 0);
1820 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1822 struct kvm_sw_breakpoint *bp, *next;
1823 KVMState *s = current_env->kvm_state;
1826 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1827 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1828 /* Try harder to find a CPU that currently sees the breakpoint. */
1829 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1830 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1836 kvm_arch_remove_all_hw_breakpoints();
1838 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1839 kvm_update_guest_debug(env, 0);
1843 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1845 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1850 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1851 target_ulong len, int type)
1856 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1857 target_ulong len, int type)
1862 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1865 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1867 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1869 struct kvm_signal_mask *sigmask;
1873 return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1876 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1879 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1880 r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1886 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1890 struct kvm_ioeventfd iofd;
1892 iofd.datamatch = val;
1895 iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1898 if (!kvm_enabled()) {
1903 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1906 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1915 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1917 struct kvm_ioeventfd kick = {
1921 .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1925 if (!kvm_enabled()) {
1929 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1931 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1938 int kvm_on_sigbus_vcpu(CPUArchState *env, int code, void *addr)
1940 return kvm_arch_on_sigbus_vcpu(env, code, addr);
1943 int kvm_on_sigbus(int code, void *addr)
1945 return kvm_arch_on_sigbus(code, addr);