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);
147 int kvm_init_vcpu(CPUState *env)
149 KVMState *s = kvm_state;
153 dprintf("kvm_init_vcpu\n");
155 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
157 dprintf("kvm_create_vcpu failed\n");
164 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
166 dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
170 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
172 if (env->kvm_run == MAP_FAILED) {
174 dprintf("mmap'ing vcpu state failed\n");
178 ret = kvm_arch_init_vcpu(env);
184 int kvm_sync_vcpus(void)
188 for (env = first_cpu; env != NULL; env = env->next_cpu) {
191 ret = kvm_arch_put_registers(env);
200 * dirty pages logging control
202 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
203 ram_addr_t size, int flags, int mask)
205 KVMState *s = kvm_state;
206 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
210 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
211 TARGET_FMT_plx "\n", __func__, phys_addr,
212 phys_addr + size - 1);
216 old_flags = mem->flags;
218 flags = (mem->flags & ~mask) | flags;
221 /* If nothing changed effectively, no need to issue ioctl */
222 if (s->migration_log) {
223 flags |= KVM_MEM_LOG_DIRTY_PAGES;
225 if (flags == old_flags) {
229 return kvm_set_user_memory_region(s, mem);
232 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
234 return kvm_dirty_pages_log_change(phys_addr, size,
235 KVM_MEM_LOG_DIRTY_PAGES,
236 KVM_MEM_LOG_DIRTY_PAGES);
239 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
241 return kvm_dirty_pages_log_change(phys_addr, size,
243 KVM_MEM_LOG_DIRTY_PAGES);
246 int kvm_set_migration_log(int enable)
248 KVMState *s = kvm_state;
252 s->migration_log = enable;
254 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
257 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
260 err = kvm_set_user_memory_region(s, mem);
269 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
270 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
271 * This means all bits are set to dirty.
273 * @start_add: start of logged region.
274 * @end_addr: end of logged region.
276 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
277 target_phys_addr_t end_addr)
279 KVMState *s = kvm_state;
280 unsigned long size, allocated_size = 0;
281 target_phys_addr_t phys_addr;
287 d.dirty_bitmap = NULL;
288 while (start_addr < end_addr) {
289 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
294 size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
295 if (!d.dirty_bitmap) {
296 d.dirty_bitmap = qemu_malloc(size);
297 } else if (size > allocated_size) {
298 d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
300 allocated_size = size;
301 memset(d.dirty_bitmap, 0, allocated_size);
305 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
306 dprintf("ioctl failed %d\n", errno);
311 for (phys_addr = mem->start_addr, addr = mem->phys_offset;
312 phys_addr < mem->start_addr + mem->memory_size;
313 phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
314 unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
315 unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
316 unsigned word = nr / (sizeof(*bitmap) * 8);
317 unsigned bit = nr % (sizeof(*bitmap) * 8);
319 if ((bitmap[word] >> bit) & 1) {
320 cpu_physical_memory_set_dirty(addr);
323 start_addr = phys_addr;
325 qemu_free(d.dirty_bitmap);
330 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
333 #ifdef KVM_CAP_COALESCED_MMIO
334 KVMState *s = kvm_state;
336 if (s->coalesced_mmio) {
337 struct kvm_coalesced_mmio_zone zone;
342 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
349 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
352 #ifdef KVM_CAP_COALESCED_MMIO
353 KVMState *s = kvm_state;
355 if (s->coalesced_mmio) {
356 struct kvm_coalesced_mmio_zone zone;
361 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
368 int kvm_check_extension(KVMState *s, unsigned int extension)
372 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
380 static void kvm_reset_vcpus(void *opaque)
385 int kvm_init(int smp_cpus)
392 fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
396 s = qemu_mallocz(sizeof(KVMState));
398 #ifdef KVM_CAP_SET_GUEST_DEBUG
399 TAILQ_INIT(&s->kvm_sw_breakpoints);
401 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
402 s->slots[i].slot = i;
405 s->fd = open("/dev/kvm", O_RDWR);
407 fprintf(stderr, "Could not access KVM kernel module: %m\n");
412 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
413 if (ret < KVM_API_VERSION) {
416 fprintf(stderr, "kvm version too old\n");
420 if (ret > KVM_API_VERSION) {
422 fprintf(stderr, "kvm version not supported\n");
426 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
430 /* initially, KVM allocated its own memory and we had to jump through
431 * hooks to make phys_ram_base point to this. Modern versions of KVM
432 * just use a user allocated buffer so we can use regular pages
433 * unmodified. Make sure we have a sufficiently modern version of KVM.
435 if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
437 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
441 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
442 * destroyed properly. Since we rely on this capability, refuse to work
443 * with any kernel without this capability. */
444 if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
448 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
449 "Please upgrade to at least kvm-81.\n");
453 #ifdef KVM_CAP_COALESCED_MMIO
454 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
456 s->coalesced_mmio = 0;
459 s->broken_set_mem_region = 1;
460 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
461 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
463 s->broken_set_mem_region = 0;
467 ret = kvm_arch_init(s, smp_cpus);
471 qemu_register_reset(kvm_reset_vcpus, INT_MAX, NULL);
489 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
490 int direction, int size, uint32_t count)
495 for (i = 0; i < count; i++) {
496 if (direction == KVM_EXIT_IO_IN) {
499 stb_p(ptr, cpu_inb(env, port));
502 stw_p(ptr, cpu_inw(env, port));
505 stl_p(ptr, cpu_inl(env, port));
511 cpu_outb(env, port, ldub_p(ptr));
514 cpu_outw(env, port, lduw_p(ptr));
517 cpu_outl(env, port, ldl_p(ptr));
528 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
530 #ifdef KVM_CAP_COALESCED_MMIO
531 KVMState *s = kvm_state;
532 if (s->coalesced_mmio) {
533 struct kvm_coalesced_mmio_ring *ring;
535 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
536 while (ring->first != ring->last) {
537 struct kvm_coalesced_mmio *ent;
539 ent = &ring->coalesced_mmio[ring->first];
541 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
542 /* FIXME smp_wmb() */
543 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
549 int kvm_cpu_exec(CPUState *env)
551 struct kvm_run *run = env->kvm_run;
554 dprintf("kvm_cpu_exec()\n");
557 kvm_arch_pre_run(env, run);
559 if (env->exit_request) {
560 dprintf("interrupt exit requested\n");
565 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
566 kvm_arch_post_run(env, run);
568 if (ret == -EINTR || ret == -EAGAIN) {
569 dprintf("io window exit\n");
575 dprintf("kvm run failed %s\n", strerror(-ret));
579 kvm_run_coalesced_mmio(env, run);
581 ret = 0; /* exit loop */
582 switch (run->exit_reason) {
584 dprintf("handle_io\n");
585 ret = kvm_handle_io(env, run->io.port,
586 (uint8_t *)run + run->io.data_offset,
592 dprintf("handle_mmio\n");
593 cpu_physical_memory_rw(run->mmio.phys_addr,
599 case KVM_EXIT_IRQ_WINDOW_OPEN:
600 dprintf("irq_window_open\n");
602 case KVM_EXIT_SHUTDOWN:
603 dprintf("shutdown\n");
604 qemu_system_reset_request();
607 case KVM_EXIT_UNKNOWN:
608 dprintf("kvm_exit_unknown\n");
610 case KVM_EXIT_FAIL_ENTRY:
611 dprintf("kvm_exit_fail_entry\n");
613 case KVM_EXIT_EXCEPTION:
614 dprintf("kvm_exit_exception\n");
617 dprintf("kvm_exit_debug\n");
618 #ifdef KVM_CAP_SET_GUEST_DEBUG
619 if (kvm_arch_debug(&run->debug.arch)) {
620 gdb_set_stop_cpu(env);
622 env->exception_index = EXCP_DEBUG;
625 /* re-enter, this exception was guest-internal */
627 #endif /* KVM_CAP_SET_GUEST_DEBUG */
630 dprintf("kvm_arch_handle_exit\n");
631 ret = kvm_arch_handle_exit(env, run);
636 if (env->exit_request) {
637 env->exit_request = 0;
638 env->exception_index = EXCP_INTERRUPT;
644 void kvm_set_phys_mem(target_phys_addr_t start_addr,
646 ram_addr_t phys_offset)
648 KVMState *s = kvm_state;
649 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
653 if (start_addr & ~TARGET_PAGE_MASK) {
654 if (flags >= IO_MEM_UNASSIGNED) {
655 if (!kvm_lookup_overlapping_slot(s, start_addr,
656 start_addr + size)) {
659 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
661 fprintf(stderr, "Only page-aligned memory slots supported\n");
666 /* KVM does not support read-only slots */
667 phys_offset &= ~IO_MEM_ROM;
670 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
675 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
676 (start_addr + size <= mem->start_addr + mem->memory_size) &&
677 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
678 /* The new slot fits into the existing one and comes with
679 * identical parameters - nothing to be done. */
685 /* unregister the overlapping slot */
686 mem->memory_size = 0;
687 err = kvm_set_user_memory_region(s, mem);
689 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
690 __func__, strerror(-err));
694 /* Workaround for older KVM versions: we can't join slots, even not by
695 * unregistering the previous ones and then registering the larger
696 * slot. We have to maintain the existing fragmentation. Sigh.
698 * This workaround assumes that the new slot starts at the same
699 * address as the first existing one. If not or if some overlapping
700 * slot comes around later, we will fail (not seen in practice so far)
701 * - and actually require a recent KVM version. */
702 if (s->broken_set_mem_region &&
703 old.start_addr == start_addr && old.memory_size < size &&
704 flags < IO_MEM_UNASSIGNED) {
705 mem = kvm_alloc_slot(s);
706 mem->memory_size = old.memory_size;
707 mem->start_addr = old.start_addr;
708 mem->phys_offset = old.phys_offset;
711 err = kvm_set_user_memory_region(s, mem);
713 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
718 start_addr += old.memory_size;
719 phys_offset += old.memory_size;
720 size -= old.memory_size;
724 /* register prefix slot */
725 if (old.start_addr < start_addr) {
726 mem = kvm_alloc_slot(s);
727 mem->memory_size = start_addr - old.start_addr;
728 mem->start_addr = old.start_addr;
729 mem->phys_offset = old.phys_offset;
732 err = kvm_set_user_memory_region(s, mem);
734 fprintf(stderr, "%s: error registering prefix slot: %s\n",
735 __func__, strerror(-err));
740 /* register suffix slot */
741 if (old.start_addr + old.memory_size > start_addr + size) {
742 ram_addr_t size_delta;
744 mem = kvm_alloc_slot(s);
745 mem->start_addr = start_addr + size;
746 size_delta = mem->start_addr - old.start_addr;
747 mem->memory_size = old.memory_size - size_delta;
748 mem->phys_offset = old.phys_offset + size_delta;
751 err = kvm_set_user_memory_region(s, mem);
753 fprintf(stderr, "%s: error registering suffix slot: %s\n",
754 __func__, strerror(-err));
760 /* in case the KVM bug workaround already "consumed" the new slot */
764 /* KVM does not need to know about this memory */
765 if (flags >= IO_MEM_UNASSIGNED)
768 mem = kvm_alloc_slot(s);
769 mem->memory_size = size;
770 mem->start_addr = start_addr;
771 mem->phys_offset = phys_offset;
774 err = kvm_set_user_memory_region(s, mem);
776 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
782 int kvm_ioctl(KVMState *s, int type, ...)
789 arg = va_arg(ap, void *);
792 ret = ioctl(s->fd, type, arg);
799 int kvm_vm_ioctl(KVMState *s, int type, ...)
806 arg = va_arg(ap, void *);
809 ret = ioctl(s->vmfd, type, arg);
816 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
823 arg = va_arg(ap, void *);
826 ret = ioctl(env->kvm_fd, type, arg);
833 int kvm_has_sync_mmu(void)
835 #ifdef KVM_CAP_SYNC_MMU
836 KVMState *s = kvm_state;
838 return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
844 void kvm_setup_guest_memory(void *start, size_t size)
846 if (!kvm_has_sync_mmu()) {
848 int ret = madvise(start, size, MADV_DONTFORK);
856 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
862 #ifdef KVM_CAP_SET_GUEST_DEBUG
863 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
866 struct kvm_sw_breakpoint *bp;
868 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
875 int kvm_sw_breakpoints_active(CPUState *env)
877 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
880 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
882 struct kvm_guest_debug dbg;
885 if (env->singlestep_enabled)
886 dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
888 kvm_arch_update_guest_debug(env, &dbg);
889 dbg.control |= reinject_trap;
891 return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
894 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
895 target_ulong len, int type)
897 struct kvm_sw_breakpoint *bp;
901 if (type == GDB_BREAKPOINT_SW) {
902 bp = kvm_find_sw_breakpoint(current_env, addr);
908 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
914 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
920 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
923 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
928 for (env = first_cpu; env != NULL; env = env->next_cpu) {
929 err = kvm_update_guest_debug(env, 0);
936 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
937 target_ulong len, int type)
939 struct kvm_sw_breakpoint *bp;
943 if (type == GDB_BREAKPOINT_SW) {
944 bp = kvm_find_sw_breakpoint(current_env, addr);
948 if (bp->use_count > 1) {
953 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
957 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
960 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
965 for (env = first_cpu; env != NULL; env = env->next_cpu) {
966 err = kvm_update_guest_debug(env, 0);
973 void kvm_remove_all_breakpoints(CPUState *current_env)
975 struct kvm_sw_breakpoint *bp, *next;
976 KVMState *s = current_env->kvm_state;
979 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
980 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
981 /* Try harder to find a CPU that currently sees the breakpoint. */
982 for (env = first_cpu; env != NULL; env = env->next_cpu) {
983 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
988 kvm_arch_remove_all_hw_breakpoints();
990 for (env = first_cpu; env != NULL; env = env->next_cpu)
991 kvm_update_guest_debug(env, 0);
994 #else /* !KVM_CAP_SET_GUEST_DEBUG */
996 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1001 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1002 target_ulong len, int type)
1007 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1008 target_ulong len, int type)
1013 void kvm_remove_all_breakpoints(CPUState *current_env)
1016 #endif /* !KVM_CAP_SET_GUEST_DEBUG */