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_put_mp_state(CPUState *env)
186 struct kvm_mp_state mp_state = { .mp_state = env->mp_state };
188 return kvm_vcpu_ioctl(env, KVM_SET_MP_STATE, &mp_state);
191 int kvm_get_mp_state(CPUState *env)
193 struct kvm_mp_state mp_state;
196 ret = kvm_vcpu_ioctl(env, KVM_GET_MP_STATE, &mp_state);
200 env->mp_state = mp_state.mp_state;
204 int kvm_sync_vcpus(void)
208 for (env = first_cpu; env != NULL; env = env->next_cpu) {
211 ret = kvm_arch_put_registers(env);
220 * dirty pages logging control
222 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
223 ram_addr_t size, int flags, int mask)
225 KVMState *s = kvm_state;
226 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
230 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
231 TARGET_FMT_plx "\n", __func__, phys_addr,
232 phys_addr + size - 1);
236 old_flags = mem->flags;
238 flags = (mem->flags & ~mask) | flags;
241 /* If nothing changed effectively, no need to issue ioctl */
242 if (s->migration_log) {
243 flags |= KVM_MEM_LOG_DIRTY_PAGES;
245 if (flags == old_flags) {
249 return kvm_set_user_memory_region(s, mem);
252 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
254 return kvm_dirty_pages_log_change(phys_addr, size,
255 KVM_MEM_LOG_DIRTY_PAGES,
256 KVM_MEM_LOG_DIRTY_PAGES);
259 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
261 return kvm_dirty_pages_log_change(phys_addr, size,
263 KVM_MEM_LOG_DIRTY_PAGES);
266 int kvm_set_migration_log(int enable)
268 KVMState *s = kvm_state;
272 s->migration_log = enable;
274 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
277 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
280 err = kvm_set_user_memory_region(s, mem);
289 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
290 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
291 * This means all bits are set to dirty.
293 * @start_add: start of logged region.
294 * @end_addr: end of logged region.
296 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
297 target_phys_addr_t end_addr)
299 KVMState *s = kvm_state;
300 unsigned long size, allocated_size = 0;
301 target_phys_addr_t phys_addr;
307 d.dirty_bitmap = NULL;
308 while (start_addr < end_addr) {
309 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
314 size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
315 if (!d.dirty_bitmap) {
316 d.dirty_bitmap = qemu_malloc(size);
317 } else if (size > allocated_size) {
318 d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
320 allocated_size = size;
321 memset(d.dirty_bitmap, 0, allocated_size);
325 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
326 dprintf("ioctl failed %d\n", errno);
331 for (phys_addr = mem->start_addr, addr = mem->phys_offset;
332 phys_addr < mem->start_addr + mem->memory_size;
333 phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
334 unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
335 unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
336 unsigned word = nr / (sizeof(*bitmap) * 8);
337 unsigned bit = nr % (sizeof(*bitmap) * 8);
339 if ((bitmap[word] >> bit) & 1) {
340 cpu_physical_memory_set_dirty(addr);
343 start_addr = phys_addr;
345 qemu_free(d.dirty_bitmap);
350 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
353 #ifdef KVM_CAP_COALESCED_MMIO
354 KVMState *s = kvm_state;
356 if (s->coalesced_mmio) {
357 struct kvm_coalesced_mmio_zone zone;
362 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
369 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
372 #ifdef KVM_CAP_COALESCED_MMIO
373 KVMState *s = kvm_state;
375 if (s->coalesced_mmio) {
376 struct kvm_coalesced_mmio_zone zone;
381 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
388 int kvm_check_extension(KVMState *s, unsigned int extension)
392 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
400 static void kvm_reset_vcpus(void *opaque)
405 int kvm_init(int smp_cpus)
412 fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
416 s = qemu_mallocz(sizeof(KVMState));
418 #ifdef KVM_CAP_SET_GUEST_DEBUG
419 TAILQ_INIT(&s->kvm_sw_breakpoints);
421 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
422 s->slots[i].slot = i;
425 s->fd = open("/dev/kvm", O_RDWR);
427 fprintf(stderr, "Could not access KVM kernel module: %m\n");
432 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
433 if (ret < KVM_API_VERSION) {
436 fprintf(stderr, "kvm version too old\n");
440 if (ret > KVM_API_VERSION) {
442 fprintf(stderr, "kvm version not supported\n");
446 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
450 /* initially, KVM allocated its own memory and we had to jump through
451 * hooks to make phys_ram_base point to this. Modern versions of KVM
452 * just use a user allocated buffer so we can use regular pages
453 * unmodified. Make sure we have a sufficiently modern version of KVM.
455 if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
457 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
461 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
462 * destroyed properly. Since we rely on this capability, refuse to work
463 * with any kernel without this capability. */
464 if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
468 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
469 "Please upgrade to at least kvm-81.\n");
473 #ifdef KVM_CAP_COALESCED_MMIO
474 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
476 s->coalesced_mmio = 0;
479 s->broken_set_mem_region = 1;
480 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
481 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
483 s->broken_set_mem_region = 0;
487 ret = kvm_arch_init(s, smp_cpus);
491 qemu_register_reset(kvm_reset_vcpus, INT_MAX, NULL);
509 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
510 int direction, int size, uint32_t count)
515 for (i = 0; i < count; i++) {
516 if (direction == KVM_EXIT_IO_IN) {
519 stb_p(ptr, cpu_inb(env, port));
522 stw_p(ptr, cpu_inw(env, port));
525 stl_p(ptr, cpu_inl(env, port));
531 cpu_outb(env, port, ldub_p(ptr));
534 cpu_outw(env, port, lduw_p(ptr));
537 cpu_outl(env, port, ldl_p(ptr));
548 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
550 #ifdef KVM_CAP_COALESCED_MMIO
551 KVMState *s = kvm_state;
552 if (s->coalesced_mmio) {
553 struct kvm_coalesced_mmio_ring *ring;
555 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
556 while (ring->first != ring->last) {
557 struct kvm_coalesced_mmio *ent;
559 ent = &ring->coalesced_mmio[ring->first];
561 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
562 /* FIXME smp_wmb() */
563 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
569 int kvm_cpu_exec(CPUState *env)
571 struct kvm_run *run = env->kvm_run;
574 dprintf("kvm_cpu_exec()\n");
577 kvm_arch_pre_run(env, run);
579 if (env->exit_request) {
580 dprintf("interrupt exit requested\n");
585 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
586 kvm_arch_post_run(env, run);
588 if (ret == -EINTR || ret == -EAGAIN) {
589 dprintf("io window exit\n");
595 dprintf("kvm run failed %s\n", strerror(-ret));
599 kvm_run_coalesced_mmio(env, run);
601 ret = 0; /* exit loop */
602 switch (run->exit_reason) {
604 dprintf("handle_io\n");
605 ret = kvm_handle_io(env, run->io.port,
606 (uint8_t *)run + run->io.data_offset,
612 dprintf("handle_mmio\n");
613 cpu_physical_memory_rw(run->mmio.phys_addr,
619 case KVM_EXIT_IRQ_WINDOW_OPEN:
620 dprintf("irq_window_open\n");
622 case KVM_EXIT_SHUTDOWN:
623 dprintf("shutdown\n");
624 qemu_system_reset_request();
627 case KVM_EXIT_UNKNOWN:
628 dprintf("kvm_exit_unknown\n");
630 case KVM_EXIT_FAIL_ENTRY:
631 dprintf("kvm_exit_fail_entry\n");
633 case KVM_EXIT_EXCEPTION:
634 dprintf("kvm_exit_exception\n");
637 dprintf("kvm_exit_debug\n");
638 #ifdef KVM_CAP_SET_GUEST_DEBUG
639 if (kvm_arch_debug(&run->debug.arch)) {
640 gdb_set_stop_cpu(env);
642 env->exception_index = EXCP_DEBUG;
645 /* re-enter, this exception was guest-internal */
647 #endif /* KVM_CAP_SET_GUEST_DEBUG */
650 dprintf("kvm_arch_handle_exit\n");
651 ret = kvm_arch_handle_exit(env, run);
656 if (env->exit_request) {
657 env->exit_request = 0;
658 env->exception_index = EXCP_INTERRUPT;
664 void kvm_set_phys_mem(target_phys_addr_t start_addr,
666 ram_addr_t phys_offset)
668 KVMState *s = kvm_state;
669 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
673 if (start_addr & ~TARGET_PAGE_MASK) {
674 if (flags >= IO_MEM_UNASSIGNED) {
675 if (!kvm_lookup_overlapping_slot(s, start_addr,
676 start_addr + size)) {
679 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
681 fprintf(stderr, "Only page-aligned memory slots supported\n");
686 /* KVM does not support read-only slots */
687 phys_offset &= ~IO_MEM_ROM;
690 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
695 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
696 (start_addr + size <= mem->start_addr + mem->memory_size) &&
697 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
698 /* The new slot fits into the existing one and comes with
699 * identical parameters - nothing to be done. */
705 /* unregister the overlapping slot */
706 mem->memory_size = 0;
707 err = kvm_set_user_memory_region(s, mem);
709 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
710 __func__, strerror(-err));
714 /* Workaround for older KVM versions: we can't join slots, even not by
715 * unregistering the previous ones and then registering the larger
716 * slot. We have to maintain the existing fragmentation. Sigh.
718 * This workaround assumes that the new slot starts at the same
719 * address as the first existing one. If not or if some overlapping
720 * slot comes around later, we will fail (not seen in practice so far)
721 * - and actually require a recent KVM version. */
722 if (s->broken_set_mem_region &&
723 old.start_addr == start_addr && old.memory_size < size &&
724 flags < IO_MEM_UNASSIGNED) {
725 mem = kvm_alloc_slot(s);
726 mem->memory_size = old.memory_size;
727 mem->start_addr = old.start_addr;
728 mem->phys_offset = old.phys_offset;
731 err = kvm_set_user_memory_region(s, mem);
733 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
738 start_addr += old.memory_size;
739 phys_offset += old.memory_size;
740 size -= old.memory_size;
744 /* register prefix slot */
745 if (old.start_addr < start_addr) {
746 mem = kvm_alloc_slot(s);
747 mem->memory_size = start_addr - old.start_addr;
748 mem->start_addr = old.start_addr;
749 mem->phys_offset = old.phys_offset;
752 err = kvm_set_user_memory_region(s, mem);
754 fprintf(stderr, "%s: error registering prefix slot: %s\n",
755 __func__, strerror(-err));
760 /* register suffix slot */
761 if (old.start_addr + old.memory_size > start_addr + size) {
762 ram_addr_t size_delta;
764 mem = kvm_alloc_slot(s);
765 mem->start_addr = start_addr + size;
766 size_delta = mem->start_addr - old.start_addr;
767 mem->memory_size = old.memory_size - size_delta;
768 mem->phys_offset = old.phys_offset + size_delta;
771 err = kvm_set_user_memory_region(s, mem);
773 fprintf(stderr, "%s: error registering suffix slot: %s\n",
774 __func__, strerror(-err));
780 /* in case the KVM bug workaround already "consumed" the new slot */
784 /* KVM does not need to know about this memory */
785 if (flags >= IO_MEM_UNASSIGNED)
788 mem = kvm_alloc_slot(s);
789 mem->memory_size = size;
790 mem->start_addr = start_addr;
791 mem->phys_offset = phys_offset;
794 err = kvm_set_user_memory_region(s, mem);
796 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
802 int kvm_ioctl(KVMState *s, int type, ...)
809 arg = va_arg(ap, void *);
812 ret = ioctl(s->fd, type, arg);
819 int kvm_vm_ioctl(KVMState *s, int type, ...)
826 arg = va_arg(ap, void *);
829 ret = ioctl(s->vmfd, type, arg);
836 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
843 arg = va_arg(ap, void *);
846 ret = ioctl(env->kvm_fd, type, arg);
853 int kvm_has_sync_mmu(void)
855 #ifdef KVM_CAP_SYNC_MMU
856 KVMState *s = kvm_state;
858 return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
864 void kvm_setup_guest_memory(void *start, size_t size)
866 if (!kvm_has_sync_mmu()) {
868 int ret = madvise(start, size, MADV_DONTFORK);
876 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
882 #ifdef KVM_CAP_SET_GUEST_DEBUG
883 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
886 struct kvm_sw_breakpoint *bp;
888 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
895 int kvm_sw_breakpoints_active(CPUState *env)
897 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
900 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
902 struct kvm_guest_debug dbg;
905 if (env->singlestep_enabled)
906 dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
908 kvm_arch_update_guest_debug(env, &dbg);
909 dbg.control |= reinject_trap;
911 return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
914 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
915 target_ulong len, int type)
917 struct kvm_sw_breakpoint *bp;
921 if (type == GDB_BREAKPOINT_SW) {
922 bp = kvm_find_sw_breakpoint(current_env, addr);
928 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
934 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
940 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
943 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
948 for (env = first_cpu; env != NULL; env = env->next_cpu) {
949 err = kvm_update_guest_debug(env, 0);
956 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
957 target_ulong len, int type)
959 struct kvm_sw_breakpoint *bp;
963 if (type == GDB_BREAKPOINT_SW) {
964 bp = kvm_find_sw_breakpoint(current_env, addr);
968 if (bp->use_count > 1) {
973 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
977 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
980 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
985 for (env = first_cpu; env != NULL; env = env->next_cpu) {
986 err = kvm_update_guest_debug(env, 0);
993 void kvm_remove_all_breakpoints(CPUState *current_env)
995 struct kvm_sw_breakpoint *bp, *next;
996 KVMState *s = current_env->kvm_state;
999 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1000 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1001 /* Try harder to find a CPU that currently sees the breakpoint. */
1002 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1003 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
1008 kvm_arch_remove_all_hw_breakpoints();
1010 for (env = first_cpu; env != NULL; env = env->next_cpu)
1011 kvm_update_guest_debug(env, 0);
1014 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1016 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1021 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1022 target_ulong len, int type)
1027 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1028 target_ulong len, int type)
1033 void kvm_remove_all_breakpoints(CPUState *current_env)
1036 #endif /* !KVM_CAP_SET_GUEST_DEBUG */