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"
28 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
29 #define PAGE_SIZE TARGET_PAGE_SIZE
34 #define dprintf(fmt, ...) \
35 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
37 #define dprintf(fmt, ...) \
41 typedef struct KVMSlot
43 target_phys_addr_t start_addr;
44 ram_addr_t memory_size;
45 ram_addr_t phys_offset;
50 typedef struct kvm_dirty_log KVMDirtyLog;
60 int broken_set_mem_region;
62 #ifdef KVM_CAP_SET_GUEST_DEBUG
63 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
67 static KVMState *kvm_state;
69 static KVMSlot *kvm_alloc_slot(KVMState *s)
73 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
74 /* KVM private memory slots */
77 if (s->slots[i].memory_size == 0)
81 fprintf(stderr, "%s: no free slot available\n", __func__);
85 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
86 target_phys_addr_t start_addr,
87 target_phys_addr_t end_addr)
91 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
92 KVMSlot *mem = &s->slots[i];
94 if (start_addr == mem->start_addr &&
95 end_addr == mem->start_addr + mem->memory_size) {
104 * Find overlapping slot with lowest start address
106 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
107 target_phys_addr_t start_addr,
108 target_phys_addr_t end_addr)
110 KVMSlot *found = NULL;
113 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
114 KVMSlot *mem = &s->slots[i];
116 if (mem->memory_size == 0 ||
117 (found && found->start_addr < mem->start_addr)) {
121 if (end_addr > mem->start_addr &&
122 start_addr < mem->start_addr + mem->memory_size) {
130 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
132 struct kvm_userspace_memory_region mem;
134 mem.slot = slot->slot;
135 mem.guest_phys_addr = slot->start_addr;
136 mem.memory_size = slot->memory_size;
137 mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
138 mem.flags = slot->flags;
139 if (s->migration_log) {
140 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
142 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
146 int kvm_init_vcpu(CPUState *env)
148 KVMState *s = kvm_state;
152 dprintf("kvm_init_vcpu\n");
154 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
156 dprintf("kvm_create_vcpu failed\n");
163 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
165 dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
169 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
171 if (env->kvm_run == MAP_FAILED) {
173 dprintf("mmap'ing vcpu state failed\n");
177 ret = kvm_arch_init_vcpu(env);
183 int kvm_sync_vcpus(void)
187 for (env = first_cpu; env != NULL; env = env->next_cpu) {
190 ret = kvm_arch_put_registers(env);
199 * dirty pages logging control
201 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
202 ram_addr_t size, int flags, int mask)
204 KVMState *s = kvm_state;
205 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
209 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
210 TARGET_FMT_plx "\n", __func__, phys_addr,
211 phys_addr + size - 1);
215 old_flags = mem->flags;
217 flags = (mem->flags & ~mask) | flags;
220 /* If nothing changed effectively, no need to issue ioctl */
221 if (s->migration_log) {
222 flags |= KVM_MEM_LOG_DIRTY_PAGES;
224 if (flags == old_flags) {
228 return kvm_set_user_memory_region(s, mem);
231 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
233 return kvm_dirty_pages_log_change(phys_addr, size,
234 KVM_MEM_LOG_DIRTY_PAGES,
235 KVM_MEM_LOG_DIRTY_PAGES);
238 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
240 return kvm_dirty_pages_log_change(phys_addr, size,
242 KVM_MEM_LOG_DIRTY_PAGES);
245 int kvm_set_migration_log(int enable)
247 KVMState *s = kvm_state;
251 s->migration_log = enable;
253 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
256 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
259 err = kvm_set_user_memory_region(s, mem);
268 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
269 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
270 * This means all bits are set to dirty.
272 * @start_add: start of logged region.
273 * @end_addr: end of logged region.
275 int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
276 target_phys_addr_t end_addr)
278 KVMState *s = kvm_state;
279 unsigned long size, allocated_size = 0;
280 target_phys_addr_t phys_addr;
286 d.dirty_bitmap = NULL;
287 while (start_addr < end_addr) {
288 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
293 size = ((mem->memory_size >> TARGET_PAGE_BITS) + 7) / 8;
294 if (!d.dirty_bitmap) {
295 d.dirty_bitmap = qemu_malloc(size);
296 } else if (size > allocated_size) {
297 d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
299 allocated_size = size;
300 memset(d.dirty_bitmap, 0, allocated_size);
304 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
305 dprintf("ioctl failed %d\n", errno);
310 for (phys_addr = mem->start_addr, addr = mem->phys_offset;
311 phys_addr < mem->start_addr + mem->memory_size;
312 phys_addr += TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
313 unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
314 unsigned nr = (phys_addr - mem->start_addr) >> TARGET_PAGE_BITS;
315 unsigned word = nr / (sizeof(*bitmap) * 8);
316 unsigned bit = nr % (sizeof(*bitmap) * 8);
318 if ((bitmap[word] >> bit) & 1) {
319 cpu_physical_memory_set_dirty(addr);
322 start_addr = phys_addr;
324 qemu_free(d.dirty_bitmap);
329 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
332 #ifdef KVM_CAP_COALESCED_MMIO
333 KVMState *s = kvm_state;
335 if (s->coalesced_mmio) {
336 struct kvm_coalesced_mmio_zone zone;
341 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
348 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
351 #ifdef KVM_CAP_COALESCED_MMIO
352 KVMState *s = kvm_state;
354 if (s->coalesced_mmio) {
355 struct kvm_coalesced_mmio_zone zone;
360 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
367 int kvm_check_extension(KVMState *s, unsigned int extension)
371 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
379 int kvm_init(int smp_cpus)
386 fprintf(stderr, "No SMP KVM support, use '-smp 1'\n");
390 s = qemu_mallocz(sizeof(KVMState));
392 #ifdef KVM_CAP_SET_GUEST_DEBUG
393 TAILQ_INIT(&s->kvm_sw_breakpoints);
395 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
396 s->slots[i].slot = i;
399 s->fd = open("/dev/kvm", O_RDWR);
401 fprintf(stderr, "Could not access KVM kernel module: %m\n");
406 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
407 if (ret < KVM_API_VERSION) {
410 fprintf(stderr, "kvm version too old\n");
414 if (ret > KVM_API_VERSION) {
416 fprintf(stderr, "kvm version not supported\n");
420 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
424 /* initially, KVM allocated its own memory and we had to jump through
425 * hooks to make phys_ram_base point to this. Modern versions of KVM
426 * just use a user allocated buffer so we can use regular pages
427 * unmodified. Make sure we have a sufficiently modern version of KVM.
429 if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
431 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
435 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
436 * destroyed properly. Since we rely on this capability, refuse to work
437 * with any kernel without this capability. */
438 if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
442 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
443 "Please upgrade to at least kvm-81.\n");
447 #ifdef KVM_CAP_COALESCED_MMIO
448 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
450 s->coalesced_mmio = 0;
453 s->broken_set_mem_region = 1;
454 #ifdef KVM_CAP_JOIN_MEMORY_REGIONS_WORKS
455 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
457 s->broken_set_mem_region = 0;
461 ret = kvm_arch_init(s, smp_cpus);
481 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
482 int direction, int size, uint32_t count)
487 for (i = 0; i < count; i++) {
488 if (direction == KVM_EXIT_IO_IN) {
491 stb_p(ptr, cpu_inb(env, port));
494 stw_p(ptr, cpu_inw(env, port));
497 stl_p(ptr, cpu_inl(env, port));
503 cpu_outb(env, port, ldub_p(ptr));
506 cpu_outw(env, port, lduw_p(ptr));
509 cpu_outl(env, port, ldl_p(ptr));
520 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
522 #ifdef KVM_CAP_COALESCED_MMIO
523 KVMState *s = kvm_state;
524 if (s->coalesced_mmio) {
525 struct kvm_coalesced_mmio_ring *ring;
527 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
528 while (ring->first != ring->last) {
529 struct kvm_coalesced_mmio *ent;
531 ent = &ring->coalesced_mmio[ring->first];
533 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
534 /* FIXME smp_wmb() */
535 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
541 int kvm_cpu_exec(CPUState *env)
543 struct kvm_run *run = env->kvm_run;
546 dprintf("kvm_cpu_exec()\n");
549 kvm_arch_pre_run(env, run);
551 if (env->exit_request) {
552 dprintf("interrupt exit requested\n");
557 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
558 kvm_arch_post_run(env, run);
560 if (ret == -EINTR || ret == -EAGAIN) {
561 dprintf("io window exit\n");
567 dprintf("kvm run failed %s\n", strerror(-ret));
571 kvm_run_coalesced_mmio(env, run);
573 ret = 0; /* exit loop */
574 switch (run->exit_reason) {
576 dprintf("handle_io\n");
577 ret = kvm_handle_io(env, run->io.port,
578 (uint8_t *)run + run->io.data_offset,
584 dprintf("handle_mmio\n");
585 cpu_physical_memory_rw(run->mmio.phys_addr,
591 case KVM_EXIT_IRQ_WINDOW_OPEN:
592 dprintf("irq_window_open\n");
594 case KVM_EXIT_SHUTDOWN:
595 dprintf("shutdown\n");
596 qemu_system_reset_request();
599 case KVM_EXIT_UNKNOWN:
600 dprintf("kvm_exit_unknown\n");
602 case KVM_EXIT_FAIL_ENTRY:
603 dprintf("kvm_exit_fail_entry\n");
605 case KVM_EXIT_EXCEPTION:
606 dprintf("kvm_exit_exception\n");
609 dprintf("kvm_exit_debug\n");
610 #ifdef KVM_CAP_SET_GUEST_DEBUG
611 if (kvm_arch_debug(&run->debug.arch)) {
612 gdb_set_stop_cpu(env);
614 env->exception_index = EXCP_DEBUG;
617 /* re-enter, this exception was guest-internal */
619 #endif /* KVM_CAP_SET_GUEST_DEBUG */
622 dprintf("kvm_arch_handle_exit\n");
623 ret = kvm_arch_handle_exit(env, run);
628 if (env->exit_request) {
629 env->exit_request = 0;
630 env->exception_index = EXCP_INTERRUPT;
636 void kvm_set_phys_mem(target_phys_addr_t start_addr,
638 ram_addr_t phys_offset)
640 KVMState *s = kvm_state;
641 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
645 if (start_addr & ~TARGET_PAGE_MASK) {
646 if (flags >= IO_MEM_UNASSIGNED) {
647 if (!kvm_lookup_overlapping_slot(s, start_addr,
648 start_addr + size)) {
651 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
653 fprintf(stderr, "Only page-aligned memory slots supported\n");
658 /* KVM does not support read-only slots */
659 phys_offset &= ~IO_MEM_ROM;
662 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
667 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
668 (start_addr + size <= mem->start_addr + mem->memory_size) &&
669 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
670 /* The new slot fits into the existing one and comes with
671 * identical parameters - nothing to be done. */
677 /* unregister the overlapping slot */
678 mem->memory_size = 0;
679 err = kvm_set_user_memory_region(s, mem);
681 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
682 __func__, strerror(-err));
686 /* Workaround for older KVM versions: we can't join slots, even not by
687 * unregistering the previous ones and then registering the larger
688 * slot. We have to maintain the existing fragmentation. Sigh.
690 * This workaround assumes that the new slot starts at the same
691 * address as the first existing one. If not or if some overlapping
692 * slot comes around later, we will fail (not seen in practice so far)
693 * - and actually require a recent KVM version. */
694 if (s->broken_set_mem_region &&
695 old.start_addr == start_addr && old.memory_size < size &&
696 flags < IO_MEM_UNASSIGNED) {
697 mem = kvm_alloc_slot(s);
698 mem->memory_size = old.memory_size;
699 mem->start_addr = old.start_addr;
700 mem->phys_offset = old.phys_offset;
703 err = kvm_set_user_memory_region(s, mem);
705 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
710 start_addr += old.memory_size;
711 phys_offset += old.memory_size;
712 size -= old.memory_size;
716 /* register prefix slot */
717 if (old.start_addr < start_addr) {
718 mem = kvm_alloc_slot(s);
719 mem->memory_size = start_addr - old.start_addr;
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 registering prefix slot: %s\n",
727 __func__, strerror(-err));
732 /* register suffix slot */
733 if (old.start_addr + old.memory_size > start_addr + size) {
734 ram_addr_t size_delta;
736 mem = kvm_alloc_slot(s);
737 mem->start_addr = start_addr + size;
738 size_delta = mem->start_addr - old.start_addr;
739 mem->memory_size = old.memory_size - size_delta;
740 mem->phys_offset = old.phys_offset + size_delta;
743 err = kvm_set_user_memory_region(s, mem);
745 fprintf(stderr, "%s: error registering suffix slot: %s\n",
746 __func__, strerror(-err));
752 /* in case the KVM bug workaround already "consumed" the new slot */
756 /* KVM does not need to know about this memory */
757 if (flags >= IO_MEM_UNASSIGNED)
760 mem = kvm_alloc_slot(s);
761 mem->memory_size = size;
762 mem->start_addr = start_addr;
763 mem->phys_offset = phys_offset;
766 err = kvm_set_user_memory_region(s, mem);
768 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
774 int kvm_ioctl(KVMState *s, int type, ...)
781 arg = va_arg(ap, void *);
784 ret = ioctl(s->fd, type, arg);
791 int kvm_vm_ioctl(KVMState *s, int type, ...)
798 arg = va_arg(ap, void *);
801 ret = ioctl(s->vmfd, type, arg);
808 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
815 arg = va_arg(ap, void *);
818 ret = ioctl(env->kvm_fd, type, arg);
825 int kvm_has_sync_mmu(void)
827 #ifdef KVM_CAP_SYNC_MMU
828 KVMState *s = kvm_state;
830 return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
836 void kvm_setup_guest_memory(void *start, size_t size)
838 if (!kvm_has_sync_mmu()) {
840 int ret = madvise(start, size, MADV_DONTFORK);
848 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
854 #ifdef KVM_CAP_SET_GUEST_DEBUG
855 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
858 struct kvm_sw_breakpoint *bp;
860 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
867 int kvm_sw_breakpoints_active(CPUState *env)
869 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
872 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
874 struct kvm_guest_debug dbg;
877 if (env->singlestep_enabled)
878 dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
880 kvm_arch_update_guest_debug(env, &dbg);
881 dbg.control |= reinject_trap;
883 return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
886 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
887 target_ulong len, int type)
889 struct kvm_sw_breakpoint *bp;
893 if (type == GDB_BREAKPOINT_SW) {
894 bp = kvm_find_sw_breakpoint(current_env, addr);
900 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
906 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
912 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
915 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
920 for (env = first_cpu; env != NULL; env = env->next_cpu) {
921 err = kvm_update_guest_debug(env, 0);
928 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
929 target_ulong len, int type)
931 struct kvm_sw_breakpoint *bp;
935 if (type == GDB_BREAKPOINT_SW) {
936 bp = kvm_find_sw_breakpoint(current_env, addr);
940 if (bp->use_count > 1) {
945 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
949 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
952 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
957 for (env = first_cpu; env != NULL; env = env->next_cpu) {
958 err = kvm_update_guest_debug(env, 0);
965 void kvm_remove_all_breakpoints(CPUState *current_env)
967 struct kvm_sw_breakpoint *bp, *next;
968 KVMState *s = current_env->kvm_state;
971 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
972 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
973 /* Try harder to find a CPU that currently sees the breakpoint. */
974 for (env = first_cpu; env != NULL; env = env->next_cpu) {
975 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
980 kvm_arch_remove_all_hw_breakpoints();
982 for (env = first_cpu; env != NULL; env = env->next_cpu)
983 kvm_update_guest_debug(env, 0);
986 #else /* !KVM_CAP_SET_GUEST_DEBUG */
988 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
993 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
994 target_ulong len, int type)
999 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1000 target_ulong len, int type)
1005 void kvm_remove_all_breakpoints(CPUState *current_env)
1008 #endif /* !KVM_CAP_SET_GUEST_DEBUG */