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 static void on_vcpu(CPUState *env, void (*func)(void *data), void *data)
158 if (env == cpu_single_env) {
165 int kvm_init_vcpu(CPUState *env)
167 KVMState *s = kvm_state;
171 dprintf("kvm_init_vcpu\n");
173 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
175 dprintf("kvm_create_vcpu failed\n");
182 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
184 dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
188 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
190 if (env->kvm_run == MAP_FAILED) {
192 dprintf("mmap'ing vcpu state failed\n");
196 ret = kvm_arch_init_vcpu(env);
198 qemu_register_reset(kvm_reset_vcpu, env);
199 ret = kvm_arch_put_registers(env);
205 int kvm_put_mp_state(CPUState *env)
207 struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
209 return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
212 int kvm_get_mp_state(CPUState *env)
214 struct kvm_mp_state mp_state;
217 ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
221 env->mp_state = mp_state.mp_state;
226 * dirty pages logging control
228 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
229 ram_addr_t size, int flags, int mask)
231 KVMState *s = kvm_state;
232 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
236 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
237 TARGET_FMT_plx "\n", __func__, phys_addr,
238 (target_phys_addr_t)(phys_addr + size - 1));
242 old_flags = mem->flags;
244 flags = (mem->flags & ~mask) | flags;
247 /* If nothing changed effectively, no need to issue ioctl */
248 if (s->migration_log) {
249 flags |= KVM_MEM_LOG_DIRTY_PAGES;
251 if (flags == old_flags) {
255 return kvm_set_user_memory_region(s, mem);
258 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
260 return kvm_dirty_pages_log_change(phys_addr, size,
261 KVM_MEM_LOG_DIRTY_PAGES,
262 KVM_MEM_LOG_DIRTY_PAGES);
265 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
267 return kvm_dirty_pages_log_change(phys_addr, size,
269 KVM_MEM_LOG_DIRTY_PAGES);
272 int kvm_set_migration_log(int enable)
274 KVMState *s = kvm_state;
278 s->migration_log = enable;
280 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
283 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
286 err = kvm_set_user_memory_region(s, mem);
295 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
296 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
297 * This means all bits are set to dirty.
299 * @start_add: start of logged region.
300 * @end_addr: end of logged region.
302 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
303 target_phys_addr_t end_addr)
305 KVMState *s = kvm_state;
306 unsigned long size, allocated_size = 0;
307 target_phys_addr_t phys_addr;
314 d.dirty_bitmap = NULL;
315 while (start_addr < end_addr) {
316 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
321 /* We didn't activate dirty logging? Don't care then. */
322 if(!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES)) {
326 size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
327 if (!d.dirty_bitmap) {
328 d.dirty_bitmap = qemu_malloc(size);
329 } else if (size > allocated_size) {
330 d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
332 allocated_size = size;
333 memset(d.dirty_bitmap, 0, allocated_size);
337 r = kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d);
339 dprintf("ioctl failed %d\n", errno);
344 for (phys_addr = mem->start_addr, addr = mem->phys_offset;
345 phys_addr < mem->start_addr + mem->memory_size;
346 phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
347 unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
348 unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
349 unsigned word = nr / (sizeof(*bitmap) * 8);
350 unsigned bit = nr % (sizeof(*bitmap) * 8);
352 if ((bitmap[word] >> bit) & 1) {
353 cpu_physical_memory_set_dirty(addr);
355 /* When our KVM implementation doesn't know about dirty logging
356 * we can just assume it's always dirty and be fine. */
357 cpu_physical_memory_set_dirty(addr);
360 start_addr = phys_addr;
362 qemu_free(d.dirty_bitmap);
367 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
370 #ifdef KVM_CAP_COALESCED_MMIO
371 KVMState *s = kvm_state;
373 if (s->coalesced_mmio) {
374 struct kvm_coalesced_mmio_zone zone;
379 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
386 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
389 #ifdef KVM_CAP_COALESCED_MMIO
390 KVMState *s = kvm_state;
392 if (s->coalesced_mmio) {
393 struct kvm_coalesced_mmio_zone zone;
398 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
405 int kvm_check_extension(KVMState *s, unsigned int extension)
409 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
417 int kvm_init(int smp_cpus)
419 static const char upgrade_note[] =
420 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
421 "(see http://sourceforge.net/projects/kvm).\n";
427 fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
431 s = qemu_mallocz(sizeof(KVMState));
433 #ifdef KVM_CAP_SET_GUEST_DEBUG
434 TAILQ_INIT(&s->kvm_sw_breakpoints);
436 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
437 s->slots[i].slot = i;
440 s->fd = open("/dev/kvm", O_RDWR);
442 fprintf(stderr, "Could not access KVM kernel module: %m\n");
447 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
448 if (ret < KVM_API_VERSION) {
451 fprintf(stderr, "kvm version too old\n");
455 if (ret > KVM_API_VERSION) {
457 fprintf(stderr, "kvm version not supported\n");
461 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
465 /* initially, KVM allocated its own memory and we had to jump through
466 * hooks to make phys_ram_base point to this. Modern versions of KVM
467 * just use a user allocated buffer so we can use regular pages
468 * unmodified. Make sure we have a sufficiently modern version of KVM.
470 if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
472 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n%s",
477 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
478 * destroyed properly. Since we rely on this capability, refuse to work
479 * with any kernel without this capability. */
480 if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
484 "KVM kernel module broken (DESTROY_MEMORY_REGION).\n%s",
489 #ifdef KVM_CAP_COALESCED_MMIO
490 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
492 s->coalesced_mmio = 0;
495 s->broken_set_mem_region = 1;
496 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
497 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
499 s->broken_set_mem_region = 0;
503 ret = kvm_arch_init(s, smp_cpus);
523 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
524 int direction, int size, uint32_t count)
529 for (i = 0; i < count; i++) {
530 if (direction == KVM_EXIT_IO_IN) {
533 stb_p(ptr, cpu_inb(env, port));
536 stw_p(ptr, cpu_inw(env, port));
539 stl_p(ptr, cpu_inl(env, port));
545 cpu_outb(env, port, ldub_p(ptr));
548 cpu_outw(env, port, lduw_p(ptr));
551 cpu_outl(env, port, ldl_p(ptr));
562 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
564 #ifdef KVM_CAP_COALESCED_MMIO
565 KVMState *s = kvm_state;
566 if (s->coalesced_mmio) {
567 struct kvm_coalesced_mmio_ring *ring;
569 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
570 while (ring->first != ring->last) {
571 struct kvm_coalesced_mmio *ent;
573 ent = &ring->coalesced_mmio[ring->first];
575 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
576 /* FIXME smp_wmb() */
577 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
583 int kvm_cpu_exec(CPUState *env)
585 struct kvm_run *run = env->kvm_run;
588 dprintf("kvm_cpu_exec()\n");
591 if (env->exit_request) {
592 dprintf("interrupt exit requested\n");
597 kvm_arch_pre_run(env, run);
598 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
599 kvm_arch_post_run(env, run);
601 if (ret == -EINTR || ret == -EAGAIN) {
602 dprintf("io window exit\n");
608 dprintf("kvm run failed %s\n", strerror(-ret));
612 kvm_run_coalesced_mmio(env, run);
614 ret = 0; /* exit loop */
615 switch (run->exit_reason) {
617 dprintf("handle_io\n");
618 ret = kvm_handle_io(env, run->io.port,
619 (uint8_t *)run + run->io.data_offset,
625 dprintf("handle_mmio\n");
626 cpu_physical_memory_rw(run->mmio.phys_addr,
632 case KVM_EXIT_IRQ_WINDOW_OPEN:
633 dprintf("irq_window_open\n");
635 case KVM_EXIT_SHUTDOWN:
636 dprintf("shutdown\n");
637 qemu_system_reset_request();
640 case KVM_EXIT_UNKNOWN:
641 dprintf("kvm_exit_unknown\n");
643 case KVM_EXIT_FAIL_ENTRY:
644 dprintf("kvm_exit_fail_entry\n");
646 case KVM_EXIT_EXCEPTION:
647 dprintf("kvm_exit_exception\n");
650 dprintf("kvm_exit_debug\n");
651 #ifdef KVM_CAP_SET_GUEST_DEBUG
652 if (kvm_arch_debug(&run->debug.arch)) {
653 gdb_set_stop_cpu(env);
655 env->exception_index = EXCP_DEBUG;
658 /* re-enter, this exception was guest-internal */
660 #endif /* KVM_CAP_SET_GUEST_DEBUG */
663 dprintf("kvm_arch_handle_exit\n");
664 ret = kvm_arch_handle_exit(env, run);
669 if (env->exit_request) {
670 env->exit_request = 0;
671 env->exception_index = EXCP_INTERRUPT;
677 void kvm_set_phys_mem(target_phys_addr_t start_addr,
679 ram_addr_t phys_offset)
681 KVMState *s = kvm_state;
682 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
686 if (start_addr & ~TARGET_PAGE_MASK) {
687 if (flags >= IO_MEM_UNASSIGNED) {
688 if (!kvm_lookup_overlapping_slot(s, start_addr,
689 start_addr + size)) {
692 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
694 fprintf(stderr, "Only page-aligned memory slots supported\n");
699 /* KVM does not support read-only slots */
700 phys_offset &= ~IO_MEM_ROM;
703 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
708 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
709 (start_addr + size <= mem->start_addr + mem->memory_size) &&
710 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
711 /* The new slot fits into the existing one and comes with
712 * identical parameters - nothing to be done. */
718 /* unregister the overlapping slot */
719 mem->memory_size = 0;
720 err = kvm_set_user_memory_region(s, mem);
722 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
723 __func__, strerror(-err));
727 /* Workaround for older KVM versions: we can't join slots, even not by
728 * unregistering the previous ones and then registering the larger
729 * slot. We have to maintain the existing fragmentation. Sigh.
731 * This workaround assumes that the new slot starts at the same
732 * address as the first existing one. If not or if some overlapping
733 * slot comes around later, we will fail (not seen in practice so far)
734 * - and actually require a recent KVM version. */
735 if (s->broken_set_mem_region &&
736 old.start_addr == start_addr && old.memory_size < size &&
737 flags < IO_MEM_UNASSIGNED) {
738 mem = kvm_alloc_slot(s);
739 mem->memory_size = old.memory_size;
740 mem->start_addr = old.start_addr;
741 mem->phys_offset = old.phys_offset;
744 err = kvm_set_user_memory_region(s, mem);
746 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
751 start_addr += old.memory_size;
752 phys_offset += old.memory_size;
753 size -= old.memory_size;
757 /* register prefix slot */
758 if (old.start_addr < start_addr) {
759 mem = kvm_alloc_slot(s);
760 mem->memory_size = start_addr - old.start_addr;
761 mem->start_addr = old.start_addr;
762 mem->phys_offset = old.phys_offset;
765 err = kvm_set_user_memory_region(s, mem);
767 fprintf(stderr, "%s: error registering prefix slot: %s\n",
768 __func__, strerror(-err));
773 /* register suffix slot */
774 if (old.start_addr + old.memory_size > start_addr + size) {
775 ram_addr_t size_delta;
777 mem = kvm_alloc_slot(s);
778 mem->start_addr = start_addr + size;
779 size_delta = mem->start_addr - old.start_addr;
780 mem->memory_size = old.memory_size - size_delta;
781 mem->phys_offset = old.phys_offset + size_delta;
784 err = kvm_set_user_memory_region(s, mem);
786 fprintf(stderr, "%s: error registering suffix slot: %s\n",
787 __func__, strerror(-err));
793 /* in case the KVM bug workaround already "consumed" the new slot */
797 /* KVM does not need to know about this memory */
798 if (flags >= IO_MEM_UNASSIGNED)
801 mem = kvm_alloc_slot(s);
802 mem->memory_size = size;
803 mem->start_addr = start_addr;
804 mem->phys_offset = phys_offset;
807 err = kvm_set_user_memory_region(s, mem);
809 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
815 int kvm_ioctl(KVMState *s, int type, ...)
822 arg = va_arg(ap, void *);
825 ret = ioctl(s->fd, type, arg);
832 int kvm_vm_ioctl(KVMState *s, int type, ...)
839 arg = va_arg(ap, void *);
842 ret = ioctl(s->vmfd, type, arg);
849 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
856 arg = va_arg(ap, void *);
859 ret = ioctl(env->kvm_fd, type, arg);
866 int kvm_has_sync_mmu(void)
868 #ifdef KVM_CAP_SYNC_MMU
869 KVMState *s = kvm_state;
871 return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
877 void kvm_setup_guest_memory(void *start, size_t size)
879 if (!kvm_has_sync_mmu()) {
881 int ret = madvise(start, size, MADV_DONTFORK);
889 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
895 #ifdef KVM_CAP_SET_GUEST_DEBUG
896 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
899 struct kvm_sw_breakpoint *bp;
901 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
908 int kvm_sw_breakpoints_active(CPUState *env)
910 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
913 struct kvm_set_guest_debug_data {
914 struct kvm_guest_debug dbg;
919 static void kvm_invoke_set_guest_debug(void *data)
921 struct kvm_set_guest_debug_data *dbg_data = data;
922 dbg_data->err = kvm_vcpu_ioctl(dbg_data->env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
925 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
927 struct kvm_set_guest_debug_data data;
929 data.dbg.control = 0;
930 if (env->singlestep_enabled)
931 data.dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
933 kvm_arch_update_guest_debug(env, &data.dbg);
934 data.dbg.control |= reinject_trap;
937 on_vcpu(env, kvm_invoke_set_guest_debug, &data);
941 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
942 target_ulong len, int type)
944 struct kvm_sw_breakpoint *bp;
948 if (type == GDB_BREAKPOINT_SW) {
949 bp = kvm_find_sw_breakpoint(current_env, addr);
955 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
961 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
967 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
970 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
975 for (env = first_cpu; env != NULL; env = env->next_cpu) {
976 err = kvm_update_guest_debug(env, 0);
983 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
984 target_ulong len, int type)
986 struct kvm_sw_breakpoint *bp;
990 if (type == GDB_BREAKPOINT_SW) {
991 bp = kvm_find_sw_breakpoint(current_env, addr);
995 if (bp->use_count > 1) {
1000 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1004 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1007 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1012 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1013 err = kvm_update_guest_debug(env, 0);
1020 void kvm_remove_all_breakpoints(CPUState *current_env)
1022 struct kvm_sw_breakpoint *bp, *next;
1023 KVMState *s = current_env->kvm_state;
1026 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1027 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1028 /* Try harder to find a CPU that currently sees the breakpoint. */
1029 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1030 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1035 kvm_arch_remove_all_hw_breakpoints();
1037 for (env = first_cpu; env != NULL; env = env->next_cpu)
1038 kvm_update_guest_debug(env, 0);
1041 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1043 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1048 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1049 target_ulong len, int type)
1054 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1055 target_ulong len, int type)
1060 void kvm_remove_all_breakpoints(CPUState *current_env)
1063 #endif /* !KVM_CAP_SET_GUEST_DEBUG */