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_init(int smp_cpus)
332 s = qemu_mallocz(sizeof(KVMState));
334 #ifdef KVM_CAP_SET_GUEST_DEBUG
335 TAILQ_INIT(&s->kvm_sw_breakpoints);
337 for (i = 0; i < ARRAY_SIZE(s->slots); i++)
338 s->slots[i].slot = i;
341 s->fd = open("/dev/kvm", O_RDWR);
343 fprintf(stderr, "Could not access KVM kernel module: %m\n");
348 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
349 if (ret < KVM_API_VERSION) {
352 fprintf(stderr, "kvm version too old\n");
356 if (ret > KVM_API_VERSION) {
358 fprintf(stderr, "kvm version not supported\n");
362 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
366 /* initially, KVM allocated its own memory and we had to jump through
367 * hooks to make phys_ram_base point to this. Modern versions of KVM
368 * just use a user allocated buffer so we can use regular pages
369 * unmodified. Make sure we have a sufficiently modern version of KVM.
371 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_USER_MEMORY);
375 fprintf(stderr, "kvm does not support KVM_CAP_USER_MEMORY\n");
379 /* There was a nasty bug in < kvm-80 that prevents memory slots from being
380 * destroyed properly. Since we rely on this capability, refuse to work
381 * with any kernel without this capability. */
382 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION,
383 KVM_CAP_DESTROY_MEMORY_REGION_WORKS);
389 "KVM kernel module broken (DESTROY_MEMORY_REGION)\n"
390 "Please upgrade to at least kvm-81.\n");
394 s->coalesced_mmio = 0;
395 #ifdef KVM_CAP_COALESCED_MMIO
396 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_COALESCED_MMIO);
398 s->coalesced_mmio = ret;
401 ret = kvm_arch_init(s, smp_cpus);
421 static int kvm_handle_io(CPUState *env, uint16_t port, void *data,
422 int direction, int size, uint32_t count)
427 for (i = 0; i < count; i++) {
428 if (direction == KVM_EXIT_IO_IN) {
431 stb_p(ptr, cpu_inb(env, port));
434 stw_p(ptr, cpu_inw(env, port));
437 stl_p(ptr, cpu_inl(env, port));
443 cpu_outb(env, port, ldub_p(ptr));
446 cpu_outw(env, port, lduw_p(ptr));
449 cpu_outl(env, port, ldl_p(ptr));
460 static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
462 #ifdef KVM_CAP_COALESCED_MMIO
463 KVMState *s = kvm_state;
464 if (s->coalesced_mmio) {
465 struct kvm_coalesced_mmio_ring *ring;
467 ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
468 while (ring->first != ring->last) {
469 struct kvm_coalesced_mmio *ent;
471 ent = &ring->coalesced_mmio[ring->first];
473 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
474 /* FIXME smp_wmb() */
475 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
481 int kvm_cpu_exec(CPUState *env)
483 struct kvm_run *run = env->kvm_run;
486 dprintf("kvm_cpu_exec()\n");
489 kvm_arch_pre_run(env, run);
491 if (env->exit_request) {
492 dprintf("interrupt exit requested\n");
497 ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
498 kvm_arch_post_run(env, run);
500 if (ret == -EINTR || ret == -EAGAIN) {
501 dprintf("io window exit\n");
507 dprintf("kvm run failed %s\n", strerror(-ret));
511 kvm_run_coalesced_mmio(env, run);
513 ret = 0; /* exit loop */
514 switch (run->exit_reason) {
516 dprintf("handle_io\n");
517 ret = kvm_handle_io(env, run->io.port,
518 (uint8_t *)run + run->io.data_offset,
524 dprintf("handle_mmio\n");
525 cpu_physical_memory_rw(run->mmio.phys_addr,
531 case KVM_EXIT_IRQ_WINDOW_OPEN:
532 dprintf("irq_window_open\n");
534 case KVM_EXIT_SHUTDOWN:
535 dprintf("shutdown\n");
536 qemu_system_reset_request();
539 case KVM_EXIT_UNKNOWN:
540 dprintf("kvm_exit_unknown\n");
542 case KVM_EXIT_FAIL_ENTRY:
543 dprintf("kvm_exit_fail_entry\n");
545 case KVM_EXIT_EXCEPTION:
546 dprintf("kvm_exit_exception\n");
549 dprintf("kvm_exit_debug\n");
550 #ifdef KVM_CAP_SET_GUEST_DEBUG
551 if (kvm_arch_debug(&run->debug.arch)) {
552 gdb_set_stop_cpu(env);
554 env->exception_index = EXCP_DEBUG;
557 /* re-enter, this exception was guest-internal */
559 #endif /* KVM_CAP_SET_GUEST_DEBUG */
562 dprintf("kvm_arch_handle_exit\n");
563 ret = kvm_arch_handle_exit(env, run);
568 if (env->exit_request) {
569 env->exit_request = 0;
570 env->exception_index = EXCP_INTERRUPT;
576 void kvm_set_phys_mem(target_phys_addr_t start_addr,
578 ram_addr_t phys_offset)
580 KVMState *s = kvm_state;
581 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
585 if (start_addr & ~TARGET_PAGE_MASK) {
586 if (flags >= IO_MEM_UNASSIGNED) {
587 if (!kvm_lookup_overlapping_slot(s, start_addr,
588 start_addr + size)) {
591 fprintf(stderr, "Unaligned split of a KVM memory slot\n");
593 fprintf(stderr, "Only page-aligned memory slots supported\n");
598 /* KVM does not support read-only slots */
599 phys_offset &= ~IO_MEM_ROM;
602 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
607 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
608 (start_addr + size <= mem->start_addr + mem->memory_size) &&
609 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
610 /* The new slot fits into the existing one and comes with
611 * identical parameters - nothing to be done. */
617 /* unregister the overlapping slot */
618 mem->memory_size = 0;
619 err = kvm_set_user_memory_region(s, mem);
621 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
622 __func__, strerror(-err));
626 /* Workaround for older KVM versions: we can't join slots, even not by
627 * unregistering the previous ones and then registering the larger
628 * slot. We have to maintain the existing fragmentation. Sigh.
630 * This workaround assumes that the new slot starts at the same
631 * address as the first existing one. If not or if some overlapping
632 * slot comes around later, we will fail (not seen in practice so far)
633 * - and actually require a recent KVM version. */
634 if (old.start_addr == start_addr && old.memory_size < size &&
635 flags < IO_MEM_UNASSIGNED) {
636 mem = kvm_alloc_slot(s);
637 mem->memory_size = old.memory_size;
638 mem->start_addr = old.start_addr;
639 mem->phys_offset = old.phys_offset;
642 err = kvm_set_user_memory_region(s, mem);
644 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
649 start_addr += old.memory_size;
650 phys_offset += old.memory_size;
651 size -= old.memory_size;
655 /* register prefix slot */
656 if (old.start_addr < start_addr) {
657 mem = kvm_alloc_slot(s);
658 mem->memory_size = start_addr - old.start_addr;
659 mem->start_addr = old.start_addr;
660 mem->phys_offset = old.phys_offset;
663 err = kvm_set_user_memory_region(s, mem);
665 fprintf(stderr, "%s: error registering prefix slot: %s\n",
666 __func__, strerror(-err));
671 /* register suffix slot */
672 if (old.start_addr + old.memory_size > start_addr + size) {
673 ram_addr_t size_delta;
675 mem = kvm_alloc_slot(s);
676 mem->start_addr = start_addr + size;
677 size_delta = mem->start_addr - old.start_addr;
678 mem->memory_size = old.memory_size - size_delta;
679 mem->phys_offset = old.phys_offset + size_delta;
682 err = kvm_set_user_memory_region(s, mem);
684 fprintf(stderr, "%s: error registering suffix slot: %s\n",
685 __func__, strerror(-err));
691 /* in case the KVM bug workaround already "consumed" the new slot */
695 /* KVM does not need to know about this memory */
696 if (flags >= IO_MEM_UNASSIGNED)
699 mem = kvm_alloc_slot(s);
700 mem->memory_size = size;
701 mem->start_addr = start_addr;
702 mem->phys_offset = phys_offset;
705 err = kvm_set_user_memory_region(s, mem);
707 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
713 int kvm_ioctl(KVMState *s, int type, ...)
720 arg = va_arg(ap, void *);
723 ret = ioctl(s->fd, type, arg);
730 int kvm_vm_ioctl(KVMState *s, int type, ...)
737 arg = va_arg(ap, void *);
740 ret = ioctl(s->vmfd, type, arg);
747 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
754 arg = va_arg(ap, void *);
757 ret = ioctl(env->kvm_fd, type, arg);
764 int kvm_has_sync_mmu(void)
766 #ifdef KVM_CAP_SYNC_MMU
767 KVMState *s = kvm_state;
769 if (kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_SYNC_MMU) > 0)
776 void kvm_setup_guest_memory(void *start, size_t size)
778 if (!kvm_has_sync_mmu()) {
780 int ret = madvise(start, size, MADV_DONTFORK);
788 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
794 #ifdef KVM_CAP_SET_GUEST_DEBUG
795 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
798 struct kvm_sw_breakpoint *bp;
800 TAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
807 int kvm_sw_breakpoints_active(CPUState *env)
809 return !TAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
812 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
814 struct kvm_guest_debug dbg;
817 if (env->singlestep_enabled)
818 dbg.control = KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
820 kvm_arch_update_guest_debug(env, &dbg);
821 dbg.control |= reinject_trap;
823 return kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg);
826 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
827 target_ulong len, int type)
829 struct kvm_sw_breakpoint *bp;
833 if (type == GDB_BREAKPOINT_SW) {
834 bp = kvm_find_sw_breakpoint(current_env, addr);
840 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
846 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
852 TAILQ_INSERT_HEAD(¤t_env->kvm_state->kvm_sw_breakpoints,
855 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
860 for (env = first_cpu; env != NULL; env = env->next_cpu) {
861 err = kvm_update_guest_debug(env, 0);
868 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
869 target_ulong len, int type)
871 struct kvm_sw_breakpoint *bp;
875 if (type == GDB_BREAKPOINT_SW) {
876 bp = kvm_find_sw_breakpoint(current_env, addr);
880 if (bp->use_count > 1) {
885 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
889 TAILQ_REMOVE(¤t_env->kvm_state->kvm_sw_breakpoints, bp, entry);
892 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
897 for (env = first_cpu; env != NULL; env = env->next_cpu) {
898 err = kvm_update_guest_debug(env, 0);
905 void kvm_remove_all_breakpoints(CPUState *current_env)
907 struct kvm_sw_breakpoint *bp, *next;
908 KVMState *s = current_env->kvm_state;
911 TAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
912 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
913 /* Try harder to find a CPU that currently sees the breakpoint. */
914 for (env = first_cpu; env != NULL; env = env->next_cpu) {
915 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0)
920 kvm_arch_remove_all_hw_breakpoints();
922 for (env = first_cpu; env != NULL; env = env->next_cpu)
923 kvm_update_guest_debug(env, 0);
926 #else /* !KVM_CAP_SET_GUEST_DEBUG */
928 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
933 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
934 target_ulong len, int type)
939 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
940 target_ulong len, int type)
945 void kvm_remove_all_breakpoints(CPUState *current_env)
948 #endif /* !KVM_CAP_SET_GUEST_DEBUG */