4 * Copyright IBM, Corp. 2008
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 #ifdef KVM_CAP_SET_GUEST_DEBUG
61 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
65 static KVMState *kvm_state;
67 static KVMSlot *kvm_alloc_slot(KVMState *s)
71 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
72 /* KVM private memory slots */
75 if (s->slots[i].memory_size == 0)
79 fprintf(stderr, "%s: no free slot available\n", __func__);
83 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
84 target_phys_addr_t start_addr,
85 target_phys_addr_t end_addr)
89 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
90 KVMSlot *mem = &s->slots[i];
92 if (start_addr == mem->start_addr &&
93 end_addr == mem->start_addr + mem->memory_size) {
102 * Find overlapping slot with lowest start address
104 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
105 target_phys_addr_t start_addr,
106 target_phys_addr_t end_addr)
108 KVMSlot *found = NULL;
111 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
112 KVMSlot *mem = &s->slots[i];
114 if (mem->memory_size == 0 ||
115 (found && found->start_addr < mem->start_addr)) {
119 if (end_addr > mem->start_addr &&
120 start_addr < mem->start_addr + mem->memory_size) {
128 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
130 struct kvm_userspace_memory_region mem;
132 mem.slot = slot->slot;
133 mem.guest_phys_addr = slot->start_addr;
134 mem.memory_size = slot->memory_size;
135 mem.userspace_addr = (unsigned long)qemu_get_ram_ptr(slot->phys_offset);
136 mem.flags = slot->flags;
138 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
142 int kvm_init_vcpu(CPUState *env)
144 KVMState *s = kvm_state;
148 dprintf("kvm_init_vcpu\n");
150 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
152 dprintf("kvm_create_vcpu failed\n");
159 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
161 dprintf("KVM_GET_VCPU_MMAP_SIZE failed\n");
165 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
167 if (env->kvm_run == MAP_FAILED) {
169 dprintf("mmap'ing vcpu state failed\n");
173 ret = kvm_arch_init_vcpu(env);
179 int kvm_sync_vcpus(void)
183 for (env = first_cpu; env != NULL; env = env->next_cpu) {
186 ret = kvm_arch_put_registers(env);
195 * dirty pages logging control
197 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
198 ram_addr_t size, unsigned flags,
201 KVMState *s = kvm_state;
202 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
204 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
205 TARGET_FMT_plx "\n", __func__, phys_addr,
206 phys_addr + size - 1);
210 flags = (mem->flags & ~mask) | flags;
211 /* Nothing changed, no need to issue ioctl */
212 if (flags == mem->flags)
217 return kvm_set_user_memory_region(s, mem);
220 int kvm_log_start(target_phys_addr_t phys_addr, ram_addr_t size)
222 return kvm_dirty_pages_log_change(phys_addr, size,
223 KVM_MEM_LOG_DIRTY_PAGES,
224 KVM_MEM_LOG_DIRTY_PAGES);
227 int kvm_log_stop(target_phys_addr_t phys_addr, ram_addr_t size)
229 return kvm_dirty_pages_log_change(phys_addr, size,
231 KVM_MEM_LOG_DIRTY_PAGES);
235 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
236 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
237 * This means all bits are set to dirty.
239 * @start_add: start of logged region.
240 * @end_addr: end of logged region.
242 void kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
243 target_phys_addr_t end_addr)
245 KVMState *s = kvm_state;
247 KVMSlot *mem = kvm_lookup_matching_slot(s, start_addr, end_addr);
248 unsigned long alloc_size;
250 target_phys_addr_t phys_addr = start_addr;
252 dprintf("sync addr: " TARGET_FMT_lx " into %lx\n", start_addr,
255 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
256 TARGET_FMT_plx "\n", __func__, phys_addr, end_addr - 1);
260 alloc_size = mem->memory_size >> TARGET_PAGE_BITS / sizeof(d.dirty_bitmap);
261 d.dirty_bitmap = qemu_mallocz(alloc_size);
264 dprintf("slot %d, phys_addr %llx, uaddr: %llx\n",
265 d.slot, mem->start_addr, mem->phys_offset);
267 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
268 dprintf("ioctl failed %d\n", errno);
272 phys_addr = start_addr;
273 for (addr = mem->phys_offset; phys_addr < end_addr; phys_addr+= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
274 unsigned long *bitmap = (unsigned long *)d.dirty_bitmap;
275 unsigned nr = (phys_addr - start_addr) >> TARGET_PAGE_BITS;
276 unsigned word = nr / (sizeof(*bitmap) * 8);
277 unsigned bit = nr % (sizeof(*bitmap) * 8);
278 if ((bitmap[word] >> bit) & 1)
279 cpu_physical_memory_set_dirty(addr);
282 qemu_free(d.dirty_bitmap);
285 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
288 #ifdef KVM_CAP_COALESCED_MMIO
289 KVMState *s = kvm_state;
291 if (s->coalesced_mmio) {
292 struct kvm_coalesced_mmio_zone zone;
297 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
304 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
307 #ifdef KVM_CAP_COALESCED_MMIO
308 KVMState *s = kvm_state;
310 if (s->coalesced_mmio) {
311 struct kvm_coalesced_mmio_zone zone;
316 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
323 int kvm_check_extension(KVMState *s, unsigned int extension)
327 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
335 int kvm_init(int smp_cpus)
344 s = qemu_mallocz(sizeof(KVMState));
346 #ifdef KVM_CAP_SET_GUEST_DEBUG
347 TAILQ_INIT(&s->kvm_sw_breakpoints);
349 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
350 s->slots[i].slot = i;
353 s->fd = open("/dev/kvm", O_RDWR);
355 fprintf(stderr, "Could not access KVM kernel module: %m\n");
360 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
361 if (ret < KVM_API_VERSION) {
364 fprintf(stderr, "kvm version too old\n");
368 if (ret > KVM_API_VERSION) {
370 fprintf(stderr, "kvm version not supported\n");
374 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
378 /* initially, KVM allocated its own memory and we had to jump through
379 * hooks to make phys_ram_base point to this. Modern versions of KVM
380 * just use a user allocated buffer so we can use regular pages
381 * unmodified. Make sure we have a sufficiently modern version of KVM.
383 if (!kvm_check_extension(s, KVM_CAP_USER_MEMORY)) {
385 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
389 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
390 * destroyed properly. Since we rely on this capability, refuse to work
391 * with any kernel without this capability. */
392 if (!kvm_check_extension(s, KVM_CAP_DESTROY_MEMORY_REGION_WORKS)) {
396 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
397 "Please upgrade to at least kvm-81.\n");
401 #ifdef KVM_CAP_COALESCED_MMIO
402 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
404 s->coalesced_mmio = 0;
407 ret = kvm_arch_init(s, smp_cpus);
427 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
428 int direction, int size, uint32_t count)
433 for (i = 0; i < count; i++) {
434 if (direction == KVM_EXIT_IO_IN) {
437 stb_p(ptr, cpu_inb(env, port));
440 stw_p(ptr, cpu_inw(env, port));
443 stl_p(ptr, cpu_inl(env, port));
449 cpu_outb(env, port, ldub_p(ptr));
452 cpu_outw(env, port, lduw_p(ptr));
455 cpu_outl(env, port, ldl_p(ptr));
466 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
468 #ifdef KVM_CAP_COALESCED_MMIO
469 KVMState *s = kvm_state;
470 if (s->coalesced_mmio) {
471 struct kvm_coalesced_mmio_ring *ring;
473 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
474 while (ring->first != ring->last) {
475 struct kvm_coalesced_mmio *ent;
477 ent = &ring->coalesced_mmio[ring->first];
479 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
480 /* FIXME smp_wmb() */
481 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
487 int kvm_cpu_exec(CPUState *env)
489 struct kvm_run *run = env->kvm_run;
492 dprintf("kvm_cpu_exec()\n");
495 kvm_arch_pre_run(env, run);
497 if (env->exit_request) {
498 dprintf("interrupt exit requested\n");
503 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
504 kvm_arch_post_run(env, run);
506 if (ret == -EINTR || ret == -EAGAIN) {
507 dprintf("io window exit\n");
513 dprintf("kvm run failed %s\n", strerror(-ret));
517 kvm_run_coalesced_mmio(env, run);
519 ret = 0; /* exit loop */
520 switch (run->exit_reason) {
522 dprintf("handle_io\n");
523 ret = kvm_handle_io(env, run->io.port,
524 (uint8_t *)run + run->io.data_offset,
530 dprintf("handle_mmio\n");
531 cpu_physical_memory_rw(run->mmio.phys_addr,
537 case KVM_EXIT_IRQ_WINDOW_OPEN:
538 dprintf("irq_window_open\n");
540 case KVM_EXIT_SHUTDOWN:
541 dprintf("shutdown\n");
542 qemu_system_reset_request();
545 case KVM_EXIT_UNKNOWN:
546 dprintf("kvm_exit_unknown\n");
548 case KVM_EXIT_FAIL_ENTRY:
549 dprintf("kvm_exit_fail_entry\n");
551 case KVM_EXIT_EXCEPTION:
552 dprintf("kvm_exit_exception\n");
555 dprintf("kvm_exit_debug\n");
556 #ifdef KVM_CAP_SET_GUEST_DEBUG
557 if (kvm_arch_debug(&run->debug.arch)) {
558 gdb_set_stop_cpu(env);
560 env->exception_index = EXCP_DEBUG;
563 /* re-enter, this exception was guest-internal */
565 #endif /* KVM_CAP_SET_GUEST_DEBUG */
568 dprintf("kvm_arch_handle_exit\n");
569 ret = kvm_arch_handle_exit(env, run);
574 if (env->exit_request) {
575 env->exit_request = 0;
576 env->exception_index = EXCP_INTERRUPT;
582 void kvm_set_phys_mem(target_phys_addr_t start_addr,
584 ram_addr_t phys_offset)
586 KVMState *s = kvm_state;
587 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
591 if (start_addr & ~TARGET_PAGE_MASK) {
592 if (flags >= IO_MEM_UNASSIGNED) {
593 if (!kvm_lookup_overlapping_slot(s, start_addr,
594 start_addr + size)) {
597 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
599 fprintf(stderr, "Only page-aligned memory slots supported\n");
604 /* KVM does not support read-only slots */
605 phys_offset &= ~IO_MEM_ROM;
608 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
613 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
614 (start_addr + size <= mem->start_addr + mem->memory_size) &&
615 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
616 /* The new slot fits into the existing one and comes with
617 * identical parameters - nothing to be done. */
623 /* unregister the overlapping slot */
624 mem->memory_size = 0;
625 err = kvm_set_user_memory_region(s, mem);
627 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
628 __func__, strerror(-err));
632 /* Workaround for older KVM versions: we can't join slots, even not by
633 * unregistering the previous ones and then registering the larger
634 * slot. We have to maintain the existing fragmentation. Sigh.
636 * This workaround assumes that the new slot starts at the same
637 * address as the first existing one. If not or if some overlapping
638 * slot comes around later, we will fail (not seen in practice so far)
639 * - and actually require a recent KVM version. */
640 if (old.start_addr == start_addr && old.memory_size < size &&
641 flags < IO_MEM_UNASSIGNED) {
642 mem = kvm_alloc_slot(s);
643 mem->memory_size = old.memory_size;
644 mem->start_addr = old.start_addr;
645 mem->phys_offset = old.phys_offset;
648 err = kvm_set_user_memory_region(s, mem);
650 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
655 start_addr += old.memory_size;
656 phys_offset += old.memory_size;
657 size -= old.memory_size;
661 /* register prefix slot */
662 if (old.start_addr < start_addr) {
663 mem = kvm_alloc_slot(s);
664 mem->memory_size = start_addr - old.start_addr;
665 mem->start_addr = old.start_addr;
666 mem->phys_offset = old.phys_offset;
669 err = kvm_set_user_memory_region(s, mem);
671 fprintf(stderr, "%s: error registering prefix slot: %s\n",
672 __func__, strerror(-err));
677 /* register suffix slot */
678 if (old.start_addr + old.memory_size > start_addr + size) {
679 ram_addr_t size_delta;
681 mem = kvm_alloc_slot(s);
682 mem->start_addr = start_addr + size;
683 size_delta = mem->start_addr - old.start_addr;
684 mem->memory_size = old.memory_size - size_delta;
685 mem->phys_offset = old.phys_offset + size_delta;
688 err = kvm_set_user_memory_region(s, mem);
690 fprintf(stderr, "%s: error registering suffix slot: %s\n",
691 __func__, strerror(-err));
697 /* in case the KVM bug workaround already "consumed" the new slot */
701 /* KVM does not need to know about this memory */
702 if (flags >= IO_MEM_UNASSIGNED)
705 mem = kvm_alloc_slot(s);
706 mem->memory_size = size;
707 mem->start_addr = start_addr;
708 mem->phys_offset = phys_offset;
711 err = kvm_set_user_memory_region(s, mem);
713 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
719 int kvm_ioctl(KVMState *s, int type, ...)
726 arg = va_arg(ap, void *);
729 ret = ioctl(s->fd, type, arg);
736 int kvm_vm_ioctl(KVMState *s, int type, ...)
743 arg = va_arg(ap, void *);
746 ret = ioctl(s->vmfd, type, arg);
753 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
760 arg = va_arg(ap, void *);
763 ret = ioctl(env->kvm_fd, type, arg);
770 int kvm_has_sync_mmu(void)
772 #ifdef KVM_CAP_SYNC_MMU
773 KVMState *s = kvm_state;
775 return kvm_check_extension(s, KVM_CAP_SYNC_MMU);
781 void kvm_setup_guest_memory(void *start, size_t size)
783 if (!kvm_has_sync_mmu()) {
785 int ret = madvise(start, size, MADV_DONTFORK);
793 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
799 #ifdef KVM_CAP_SET_GUEST_DEBUG
800 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
803 struct kvm_sw_breakpoint *bp;
805 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
812 int kvm_sw_breakpoints_active(CPUState *env)
814 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
817 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
819 struct kvm_guest_debug dbg;
822 if (env->singlestep_enabled)
823 dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
825 kvm_arch_update_guest_debug(env, &dbg);
826 dbg.control |= reinject_trap;
828 return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
831 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
832 target_ulong len, int type)
834 struct kvm_sw_breakpoint *bp;
838 if (type == GDB_BREAKPOINT_SW) {
839 bp = kvm_find_sw_breakpoint(current_env, addr);
845 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
851 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
857 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
860 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
865 for (env = first_cpu; env != NULL; env = env->next_cpu) {
866 err = kvm_update_guest_debug(env, 0);
873 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
874 target_ulong len, int type)
876 struct kvm_sw_breakpoint *bp;
880 if (type == GDB_BREAKPOINT_SW) {
881 bp = kvm_find_sw_breakpoint(current_env, addr);
885 if (bp->use_count > 1) {
890 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
894 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
897 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
902 for (env = first_cpu; env != NULL; env = env->next_cpu) {
903 err = kvm_update_guest_debug(env, 0);
910 void kvm_remove_all_breakpoints(CPUState *current_env)
912 struct kvm_sw_breakpoint *bp, *next;
913 KVMState *s = current_env->kvm_state;
916 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
917 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
918 /* Try harder to find a CPU that currently sees the breakpoint. */
919 for (env = first_cpu; env != NULL; env = env->next_cpu) {
920 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
925 kvm_arch_remove_all_hw_breakpoints();
927 for (env = first_cpu; env != NULL; env = env->next_cpu)
928 kvm_update_guest_debug(env, 0);
931 #else /* !KVM_CAP_SET_GUEST_DEBUG */
933 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
938 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
939 target_ulong len, int type)
944 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
945 target_ulong len, int type)
950 void kvm_remove_all_breakpoints(CPUState *current_env)
953 #endif /* !KVM_CAP_SET_GUEST_DEBUG */