4 * Copyright IBM, Corp. 2008
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Glauber Costa <gcosta@redhat.com>
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
16 #include <sys/types.h>
17 #include <sys/ioctl.h>
21 #include <linux/kvm.h>
23 #include "qemu-common.h"
29 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
30 #define PAGE_SIZE TARGET_PAGE_SIZE
35 #define dprintf(fmt, ...) \
36 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
38 #define dprintf(fmt, ...) \
42 typedef struct KVMSlot
44 target_phys_addr_t start_addr;
45 ram_addr_t memory_size;
46 ram_addr_t phys_offset;
51 typedef struct kvm_dirty_log KVMDirtyLog;
61 int broken_set_mem_region;
63 #ifdef KVM_CAP_SET_GUEST_DEBUG
64 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
68 static KVMState *kvm_state;
70 static KVMSlot *kvm_alloc_slot(KVMState *s)
74 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
75 /* KVM private memory slots */
78 if (s->slots[i].memory_size == 0)
82 fprintf(stderr, "%s: no free slot available\n", __func__);
86 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
87 target_phys_addr_t start_addr,
88 target_phys_addr_t end_addr)
92 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
93 KVMSlot *mem = &s->slots[i];
95 if (start_addr == mem->start_addr &&
96 end_addr == mem->start_addr + mem->memory_size) {
105 * Find overlapping slot with lowest start address
107 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
108 target_phys_addr_t start_addr,
109 target_phys_addr_t end_addr)
111 KVMSlot *found = NULL;
114 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
115 KVMSlot *mem = &s->slots[i];
117 if (mem->memory_size == 0 ||
118 (found && found->start_addr < mem->start_addr)) {
122 if (end_addr > mem->start_addr &&
123 start_addr < mem->start_addr + mem->memory_size) {
131 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
133 struct kvm_userspace_memory_region mem;
135 mem.slot = slot->slot;
136 mem.guest_phys_addr = slot->start_addr;
137 mem.memory_size = slot->memory_size;
138 mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
139 mem.flags = slot->flags;
140 if (s->migration_log) {
141 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
143 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
146 static void kvm_reset_vcpu(void *opaque)
148 CPUState *env = opaque;
150 if (kvm_arch_put_registers(env)) {
151 fprintf(stderr, "Fatal: kvm vcpu reset failed\n");
156 int kvm_init_vcpu(CPUState *env)
158 KVMState *s = kvm_state;
162 dprintf("kvm_init_vcpu\n");
164 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
166 dprintf("kvm_create_vcpu failed\n");
173 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
175 dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
179 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
181 if (env->kvm_run == MAP_FAILED) {
183 dprintf("mmap'ing vcpu state failed\n");
187 ret = kvm_arch_init_vcpu(env);
189 qemu_register_reset(kvm_reset_vcpu, env);
190 ret = kvm_arch_put_registers(env);
196 int kvm_put_mp_state(CPUState *env)
198 struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
200 return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
203 int kvm_get_mp_state(CPUState *env)
205 struct kvm_mp_state mp_state;
208 ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
212 env->mp_state = mp_state.mp_state;
217 * dirty pages logging control
219 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
220 ram_addr_t size, int flags, int mask)
222 KVMState *s = kvm_state;
223 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
227 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
228 TARGET_FMT_plx "\n", __func__, phys_addr,
229 phys_addr + size - 1);
233 old_flags = mem->flags;
235 flags = (mem->flags & ~mask) | flags;
238 /* If nothing changed effectively, no need to issue ioctl */
239 if (s->migration_log) {
240 flags |= KVM_MEM_LOG_DIRTY_PAGES;
242 if (flags == old_flags) {
246 return kvm_set_user_memory_region(s, mem);
249 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
251 return kvm_dirty_pages_log_change(phys_addr, size,
252 KVM_MEM_LOG_DIRTY_PAGES,
253 KVM_MEM_LOG_DIRTY_PAGES);
256 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
258 return kvm_dirty_pages_log_change(phys_addr, size,
260 KVM_MEM_LOG_DIRTY_PAGES);
263 int kvm_set_migration_log(int enable)
265 KVMState *s = kvm_state;
269 s->migration_log = enable;
271 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
274 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
277 err = kvm_set_user_memory_region(s, mem);
286 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
287 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
288 * This means all bits are set to dirty.
290 * @start_add: start of logged region.
291 * @end_addr: end of logged region.
293 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
294 target_phys_addr_t end_addr)
296 KVMState *s = kvm_state;
297 unsigned long size, allocated_size = 0;
298 target_phys_addr_t phys_addr;
304 d.dirty_bitmap = NULL;
305 while (start_addr < end_addr) {
306 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
311 size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
312 if (!d.dirty_bitmap) {
313 d.dirty_bitmap = qemu_malloc(size);
314 } else if (size > allocated_size) {
315 d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
317 allocated_size = size;
318 memset(d.dirty_bitmap, 0, allocated_size);
322 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
323 dprintf("ioctl failed %d\n", errno);
328 for (phys_addr = mem->start_addr, addr = mem->phys_offset;
329 phys_addr < mem->start_addr + mem->memory_size;
330 phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
331 unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
332 unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
333 unsigned word = nr / (sizeof(*bitmap) * 8);
334 unsigned bit = nr % (sizeof(*bitmap) * 8);
336 if ((bitmap[word] >> bit) & 1) {
337 cpu_physical_memory_set_dirty(addr);
340 start_addr = phys_addr;
342 qemu_free(d.dirty_bitmap);
347 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
350 #ifdef KVM_CAP_COALESCED_MMIO
351 KVMState *s = kvm_state;
353 if (s->coalesced_mmio) {
354 struct kvm_coalesced_mmio_zone zone;
359 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
366 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
369 #ifdef KVM_CAP_COALESCED_MMIO
370 KVMState *s = kvm_state;
372 if (s->coalesced_mmio) {
373 struct kvm_coalesced_mmio_zone zone;
378 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
385 int kvm_check_extension(KVMState *s, unsigned int extension)
389 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
397 int kvm_init(int smp_cpus)
399 static const char upgrade_note[] =
400 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
401 "(see http://sourceforge.net/projects/kvm).\n";
407 fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
411 s = qemu_mallocz(sizeof(KVMState));
413 #ifdef KVM_CAP_SET_GUEST_DEBUG
414 TAILQ_INIT(&s->kvm_sw_breakpoints);
416 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
417 s->slots[i].slot = i;
420 s->fd = open("/dev/kvm", O_RDWR);
422 fprintf(stderr, "Could not access KVM kernel module: %m\n");
427 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
428 if (ret < KVM_API_VERSION) {
431 fprintf(stderr, "kvm version too old\n");
435 if (ret > KVM_API_VERSION) {
437 fprintf(stderr, "kvm version not supported\n");
441 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
445 /* initially, KVM allocated its own memory and we had to jump through
446 * hooks to make phys_ram_base point to this. Modern versions of KVM
447 * just use a user allocated buffer so we can use regular pages
448 * unmodified. Make sure we have a sufficiently modern version of KVM.
450 if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
452 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
457 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
458 * destroyed properly. Since we rely on this capability, refuse to work
459 * with any kernel without this capability. */
460 if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
464 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
469 #ifdef KVM_CAP_COALESCED_MMIO
470 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
472 s->coalesced_mmio = 0;
475 s->broken_set_mem_region = 1;
476 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
477 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
479 s->broken_set_mem_region = 0;
483 ret = kvm_arch_init(s, smp_cpus);
503 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
504 int direction, int size, uint32_t count)
509 for (i = 0; i < count; i++) {
510 if (direction == KVM_EXIT_IO_IN) {
513 stb_p(ptr, cpu_inb(env, port));
516 stw_p(ptr, cpu_inw(env, port));
519 stl_p(ptr, cpu_inl(env, port));
525 cpu_outb(env, port, ldub_p(ptr));
528 cpu_outw(env, port, lduw_p(ptr));
531 cpu_outl(env, port, ldl_p(ptr));
542 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
544 #ifdef KVM_CAP_COALESCED_MMIO
545 KVMState *s = kvm_state;
546 if (s->coalesced_mmio) {
547 struct kvm_coalesced_mmio_ring *ring;
549 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
550 while (ring->first != ring->last) {
551 struct kvm_coalesced_mmio *ent;
553 ent = &ring->coalesced_mmio[ring->first];
555 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
556 /* FIXME smp_wmb() */
557 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
563 int kvm_cpu_exec(CPUState *env)
565 struct kvm_run *run = env->kvm_run;
568 dprintf("kvm_cpu_exec()\n");
571 if (env->exit_request) {
572 dprintf("interrupt exit requested\n");
577 kvm_arch_pre_run(env, run);
578 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
579 kvm_arch_post_run(env, run);
581 if (ret == -EINTR || ret == -EAGAIN) {
582 dprintf("io window exit\n");
588 dprintf("kvm run failed %s\n", strerror(-ret));
592 kvm_run_coalesced_mmio(env, run);
594 ret = 0; /* exit loop */
595 switch (run->exit_reason) {
597 dprintf("handle_io\n");
598 ret = kvm_handle_io(env, run->io.port,
599 (uint8_t *)run + run->io.data_offset,
605 dprintf("handle_mmio\n");
606 cpu_physical_memory_rw(run->mmio.phys_addr,
612 case KVM_EXIT_IRQ_WINDOW_OPEN:
613 dprintf("irq_window_open\n");
615 case KVM_EXIT_SHUTDOWN:
616 dprintf("shutdown\n");
617 qemu_system_reset_request();
620 case KVM_EXIT_UNKNOWN:
621 dprintf("kvm_exit_unknown\n");
623 case KVM_EXIT_FAIL_ENTRY:
624 dprintf("kvm_exit_fail_entry\n");
626 case KVM_EXIT_EXCEPTION:
627 dprintf("kvm_exit_exception\n");
630 dprintf("kvm_exit_debug\n");
631 #ifdef KVM_CAP_SET_GUEST_DEBUG
632 if (kvm_arch_debug(&run->debug.arch)) {
633 gdb_set_stop_cpu(env);
635 env->exception_index = EXCP_DEBUG;
638 /* re-enter, this exception was guest-internal */
640 #endif /* KVM_CAP_SET_GUEST_DEBUG */
643 dprintf("kvm_arch_handle_exit\n");
644 ret = kvm_arch_handle_exit(env, run);
649 if (env->exit_request) {
650 env->exit_request = 0;
651 env->exception_index = EXCP_INTERRUPT;
657 void kvm_set_phys_mem(target_phys_addr_t start_addr,
659 ram_addr_t phys_offset)
661 KVMState *s = kvm_state;
662 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
666 if (start_addr & ~TARGET_PAGE_MASK) {
667 if (flags >= IO_MEM_UNASSIGNED) {
668 if (!kvm_lookup_overlapping_slot(s, start_addr,
669 start_addr + size)) {
672 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
674 fprintf(stderr, "Only page-aligned memory slots supported\n");
679 /* KVM does not support read-only slots */
680 phys_offset &= ~IO_MEM_ROM;
683 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
688 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
689 (start_addr + size <= mem->start_addr + mem->memory_size) &&
690 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
691 /* The new slot fits into the existing one and comes with
692 * identical parameters - nothing to be done. */
698 /* unregister the overlapping slot */
699 mem->memory_size = 0;
700 err = kvm_set_user_memory_region(s, mem);
702 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
703 __func__, strerror(-err));
707 /* Workaround for older KVM versions: we can't join slots, even not by
708 * unregistering the previous ones and then registering the larger
709 * slot. We have to maintain the existing fragmentation. Sigh.
711 * This workaround assumes that the new slot starts at the same
712 * address as the first existing one. If not or if some overlapping
713 * slot comes around later, we will fail (not seen in practice so far)
714 * - and actually require a recent KVM version. */
715 if (s->broken_set_mem_region &&
716 old.start_addr == start_addr && old.memory_size < size &&
717 flags < IO_MEM_UNASSIGNED) {
718 mem = kvm_alloc_slot(s);
719 mem->memory_size = old.memory_size;
720 mem->start_addr = old.start_addr;
721 mem->phys_offset = old.phys_offset;
724 err = kvm_set_user_memory_region(s, mem);
726 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
731 start_addr += old.memory_size;
732 phys_offset += old.memory_size;
733 size -= old.memory_size;
737 /* register prefix slot */
738 if (old.start_addr < start_addr) {
739 mem = kvm_alloc_slot(s);
740 mem->memory_size = start_addr - old.start_addr;
741 mem->start_addr = old.start_addr;
742 mem->phys_offset = old.phys_offset;
745 err = kvm_set_user_memory_region(s, mem);
747 fprintf(stderr, "%s: error registering prefix slot: %s\n",
748 __func__, strerror(-err));
753 /* register suffix slot */
754 if (old.start_addr + old.memory_size > start_addr + size) {
755 ram_addr_t size_delta;
757 mem = kvm_alloc_slot(s);
758 mem->start_addr = start_addr + size;
759 size_delta = mem->start_addr - old.start_addr;
760 mem->memory_size = old.memory_size - size_delta;
761 mem->phys_offset = old.phys_offset + size_delta;
764 err = kvm_set_user_memory_region(s, mem);
766 fprintf(stderr, "%s: error registering suffix slot: %s\n",
767 __func__, strerror(-err));
773 /* in case the KVM bug workaround already "consumed" the new slot */
777 /* KVM does not need to know about this memory */
778 if (flags >= IO_MEM_UNASSIGNED)
781 mem = kvm_alloc_slot(s);
782 mem->memory_size = size;
783 mem->start_addr = start_addr;
784 mem->phys_offset = phys_offset;
787 err = kvm_set_user_memory_region(s, mem);
789 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
795 int kvm_ioctl(KVMState *s, int type, ...)
802 arg = va_arg(ap, void *);
805 ret = ioctl(s->fd, type, arg);
812 int kvm_vm_ioctl(KVMState *s, int type, ...)
819 arg = va_arg(ap, void *);
822 ret = ioctl(s->vmfd, type, arg);
829 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
836 arg = va_arg(ap, void *);
839 ret = ioctl(env->kvm_fd, type, arg);
846 int kvm_has_sync_mmu(void)
848 #ifdef KVM_CAP_SYNC_MMU
849 KVMState *s = kvm_state;
851 return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
857 void kvm_setup_guest_memory(void *start, size_t size)
859 if (!kvm_has_sync_mmu()) {
861 int ret = madvise(start, size, MADV_DONTFORK);
869 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
875 #ifdef KVM_CAP_SET_GUEST_DEBUG
876 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
879 struct kvm_sw_breakpoint *bp;
881 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
888 int kvm_sw_breakpoints_active(CPUState *env)
890 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
893 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
895 struct kvm_guest_debug dbg;
898 if (env->singlestep_enabled)
899 dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
901 kvm_arch_update_guest_debug(env, &dbg);
902 dbg.control |= reinject_trap;
904 return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
907 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
908 target_ulong len, int type)
910 struct kvm_sw_breakpoint *bp;
914 if (type == GDB_BREAKPOINT_SW) {
915 bp = kvm_find_sw_breakpoint(current_env, addr);
921 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
927 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
933 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
936 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
941 for (env = first_cpu; env != NULL; env = env->next_cpu) {
942 err = kvm_update_guest_debug(env, 0);
949 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
950 target_ulong len, int type)
952 struct kvm_sw_breakpoint *bp;
956 if (type == GDB_BREAKPOINT_SW) {
957 bp = kvm_find_sw_breakpoint(current_env, addr);
961 if (bp->use_count > 1) {
966 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
970 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
973 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
978 for (env = first_cpu; env != NULL; env = env->next_cpu) {
979 err = kvm_update_guest_debug(env, 0);
986 void kvm_remove_all_breakpoints(CPUState *current_env)
988 struct kvm_sw_breakpoint *bp, *next;
989 KVMState *s = current_env->kvm_state;
992 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
993 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
994 /* Try harder to find a CPU that currently sees the breakpoint. */
995 for (env = first_cpu; env != NULL; env = env->next_cpu) {
996 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1001 kvm_arch_remove_all_hw_breakpoints();
1003 for (env = first_cpu; env != NULL; env = env->next_cpu)
1004 kvm_update_guest_debug(env, 0);
1007 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1009 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1014 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1015 target_ulong len, int type)
1020 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1021 target_ulong len, int type)
1026 void kvm_remove_all_breakpoints(CPUState *current_env)
1029 #endif /* !KVM_CAP_SET_GUEST_DEBUG */