ui: remove unused 'wiremode' variable in VncState struct
[qemu/ar7.git] / kvm-all.c
blob55025cc366992b2a1d4a21b426edffe3e37f0ff6
1 /*
2 * QEMU KVM support
4 * Copyright IBM, Corp. 2008
5 * Red Hat, Inc. 2008
7 * Authors:
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>
18 #include <sys/mman.h>
19 #include <stdarg.h>
21 #include <linux/kvm.h>
23 #include "qemu-common.h"
24 #include "qemu/atomic.h"
25 #include "qemu/option.h"
26 #include "qemu/config-file.h"
27 #include "sysemu/sysemu.h"
28 #include "sysemu/accel.h"
29 #include "hw/hw.h"
30 #include "hw/pci/msi.h"
31 #include "hw/s390x/adapter.h"
32 #include "exec/gdbstub.h"
33 #include "sysemu/kvm.h"
34 #include "qemu/bswap.h"
35 #include "exec/memory.h"
36 #include "exec/ram_addr.h"
37 #include "exec/address-spaces.h"
38 #include "qemu/event_notifier.h"
39 #include "trace.h"
41 #include "hw/boards.h"
43 /* This check must be after config-host.h is included */
44 #ifdef CONFIG_EVENTFD
45 #include <sys/eventfd.h>
46 #endif
48 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
49 #define PAGE_SIZE TARGET_PAGE_SIZE
51 //#define DEBUG_KVM
53 #ifdef DEBUG_KVM
54 #define DPRINTF(fmt, ...) \
55 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
56 #else
57 #define DPRINTF(fmt, ...) \
58 do { } while (0)
59 #endif
61 #define KVM_MSI_HASHTAB_SIZE 256
63 typedef struct KVMSlot
65 hwaddr start_addr;
66 ram_addr_t memory_size;
67 void *ram;
68 int slot;
69 int flags;
70 } KVMSlot;
72 typedef struct kvm_dirty_log KVMDirtyLog;
74 struct KVMState
76 AccelState parent_obj;
78 KVMSlot *slots;
79 int nr_slots;
80 int fd;
81 int vmfd;
82 int coalesced_mmio;
83 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
84 bool coalesced_flush_in_progress;
85 int broken_set_mem_region;
86 int migration_log;
87 int vcpu_events;
88 int robust_singlestep;
89 int debugregs;
90 #ifdef KVM_CAP_SET_GUEST_DEBUG
91 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
92 #endif
93 int pit_state2;
94 int xsave, xcrs;
95 int many_ioeventfds;
96 int intx_set_mask;
97 /* The man page (and posix) say ioctl numbers are signed int, but
98 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
99 * unsigned, and treating them as signed here can break things */
100 unsigned irq_set_ioctl;
101 unsigned int sigmask_len;
102 #ifdef KVM_CAP_IRQ_ROUTING
103 struct kvm_irq_routing *irq_routes;
104 int nr_allocated_irq_routes;
105 uint32_t *used_gsi_bitmap;
106 unsigned int gsi_count;
107 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
108 bool direct_msi;
109 #endif
112 #define TYPE_KVM_ACCEL ACCEL_CLASS_NAME("kvm")
114 #define KVM_STATE(obj) \
115 OBJECT_CHECK(KVMState, (obj), TYPE_KVM_ACCEL)
117 KVMState *kvm_state;
118 bool kvm_kernel_irqchip;
119 bool kvm_async_interrupts_allowed;
120 bool kvm_halt_in_kernel_allowed;
121 bool kvm_eventfds_allowed;
122 bool kvm_irqfds_allowed;
123 bool kvm_resamplefds_allowed;
124 bool kvm_msi_via_irqfd_allowed;
125 bool kvm_gsi_routing_allowed;
126 bool kvm_gsi_direct_mapping;
127 bool kvm_allowed;
128 bool kvm_readonly_mem_allowed;
129 bool kvm_vm_attributes_allowed;
131 static const KVMCapabilityInfo kvm_required_capabilites[] = {
132 KVM_CAP_INFO(USER_MEMORY),
133 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
134 KVM_CAP_LAST_INFO
137 static KVMSlot *kvm_get_free_slot(KVMState *s)
139 int i;
141 for (i = 0; i < s->nr_slots; i++) {
142 if (s->slots[i].memory_size == 0) {
143 return &s->slots[i];
147 return NULL;
150 bool kvm_has_free_slot(MachineState *ms)
152 return kvm_get_free_slot(KVM_STATE(ms->accelerator));
155 static KVMSlot *kvm_alloc_slot(KVMState *s)
157 KVMSlot *slot = kvm_get_free_slot(s);
159 if (slot) {
160 return slot;
163 fprintf(stderr, "%s: no free slot available\n", __func__);
164 abort();
167 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
168 hwaddr start_addr,
169 hwaddr end_addr)
171 int i;
173 for (i = 0; i < s->nr_slots; i++) {
174 KVMSlot *mem = &s->slots[i];
176 if (start_addr == mem->start_addr &&
177 end_addr == mem->start_addr + mem->memory_size) {
178 return mem;
182 return NULL;
186 * Find overlapping slot with lowest start address
188 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
189 hwaddr start_addr,
190 hwaddr end_addr)
192 KVMSlot *found = NULL;
193 int i;
195 for (i = 0; i < s->nr_slots; i++) {
196 KVMSlot *mem = &s->slots[i];
198 if (mem->memory_size == 0 ||
199 (found && found->start_addr < mem->start_addr)) {
200 continue;
203 if (end_addr > mem->start_addr &&
204 start_addr < mem->start_addr + mem->memory_size) {
205 found = mem;
209 return found;
212 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
213 hwaddr *phys_addr)
215 int i;
217 for (i = 0; i < s->nr_slots; i++) {
218 KVMSlot *mem = &s->slots[i];
220 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
221 *phys_addr = mem->start_addr + (ram - mem->ram);
222 return 1;
226 return 0;
229 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
231 struct kvm_userspace_memory_region mem;
233 mem.slot = slot->slot;
234 mem.guest_phys_addr = slot->start_addr;
235 mem.userspace_addr = (unsigned long)slot->ram;
236 mem.flags = slot->flags;
237 if (s->migration_log) {
238 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
241 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
242 /* Set the slot size to 0 before setting the slot to the desired
243 * value. This is needed based on KVM commit 75d61fbc. */
244 mem.memory_size = 0;
245 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
247 mem.memory_size = slot->memory_size;
248 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
251 int kvm_init_vcpu(CPUState *cpu)
253 KVMState *s = kvm_state;
254 long mmap_size;
255 int ret;
257 DPRINTF("kvm_init_vcpu\n");
259 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
260 if (ret < 0) {
261 DPRINTF("kvm_create_vcpu failed\n");
262 goto err;
265 cpu->kvm_fd = ret;
266 cpu->kvm_state = s;
267 cpu->kvm_vcpu_dirty = true;
269 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
270 if (mmap_size < 0) {
271 ret = mmap_size;
272 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
273 goto err;
276 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
277 cpu->kvm_fd, 0);
278 if (cpu->kvm_run == MAP_FAILED) {
279 ret = -errno;
280 DPRINTF("mmap'ing vcpu state failed\n");
281 goto err;
284 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
285 s->coalesced_mmio_ring =
286 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
289 ret = kvm_arch_init_vcpu(cpu);
290 err:
291 return ret;
295 * dirty pages logging control
298 static int kvm_mem_flags(KVMState *s, bool log_dirty, bool readonly)
300 int flags = 0;
301 flags = log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
302 if (readonly && kvm_readonly_mem_allowed) {
303 flags |= KVM_MEM_READONLY;
305 return flags;
308 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
310 KVMState *s = kvm_state;
311 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
312 int old_flags;
314 old_flags = mem->flags;
316 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
317 mem->flags = flags;
319 /* If nothing changed effectively, no need to issue ioctl */
320 if (s->migration_log) {
321 flags |= KVM_MEM_LOG_DIRTY_PAGES;
324 if (flags == old_flags) {
325 return 0;
328 return kvm_set_user_memory_region(s, mem);
331 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
332 ram_addr_t size, bool log_dirty)
334 KVMState *s = kvm_state;
335 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
337 if (mem == NULL) {
338 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
339 TARGET_FMT_plx "\n", __func__, phys_addr,
340 (hwaddr)(phys_addr + size - 1));
341 return -EINVAL;
343 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
346 static void kvm_log_start(MemoryListener *listener,
347 MemoryRegionSection *section)
349 int r;
351 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
352 int128_get64(section->size), true);
353 if (r < 0) {
354 abort();
358 static void kvm_log_stop(MemoryListener *listener,
359 MemoryRegionSection *section)
361 int r;
363 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
364 int128_get64(section->size), false);
365 if (r < 0) {
366 abort();
370 static int kvm_set_migration_log(bool enable)
372 KVMState *s = kvm_state;
373 KVMSlot *mem;
374 int i, err;
376 s->migration_log = enable;
378 for (i = 0; i < s->nr_slots; i++) {
379 mem = &s->slots[i];
381 if (!mem->memory_size) {
382 continue;
384 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
385 continue;
387 err = kvm_set_user_memory_region(s, mem);
388 if (err) {
389 return err;
392 return 0;
395 /* get kvm's dirty pages bitmap and update qemu's */
396 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
397 unsigned long *bitmap)
399 ram_addr_t start = section->offset_within_region + section->mr->ram_addr;
400 ram_addr_t pages = int128_get64(section->size) / getpagesize();
402 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
403 return 0;
406 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
409 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
410 * This function updates qemu's dirty bitmap using
411 * memory_region_set_dirty(). This means all bits are set
412 * to dirty.
414 * @start_add: start of logged region.
415 * @end_addr: end of logged region.
417 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
419 KVMState *s = kvm_state;
420 unsigned long size, allocated_size = 0;
421 KVMDirtyLog d = {};
422 KVMSlot *mem;
423 int ret = 0;
424 hwaddr start_addr = section->offset_within_address_space;
425 hwaddr end_addr = start_addr + int128_get64(section->size);
427 d.dirty_bitmap = NULL;
428 while (start_addr < end_addr) {
429 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
430 if (mem == NULL) {
431 break;
434 /* XXX bad kernel interface alert
435 * For dirty bitmap, kernel allocates array of size aligned to
436 * bits-per-long. But for case when the kernel is 64bits and
437 * the userspace is 32bits, userspace can't align to the same
438 * bits-per-long, since sizeof(long) is different between kernel
439 * and user space. This way, userspace will provide buffer which
440 * may be 4 bytes less than the kernel will use, resulting in
441 * userspace memory corruption (which is not detectable by valgrind
442 * too, in most cases).
443 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
444 * a hope that sizeof(long) wont become >8 any time soon.
446 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
447 /*HOST_LONG_BITS*/ 64) / 8;
448 if (!d.dirty_bitmap) {
449 d.dirty_bitmap = g_malloc(size);
450 } else if (size > allocated_size) {
451 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
453 allocated_size = size;
454 memset(d.dirty_bitmap, 0, allocated_size);
456 d.slot = mem->slot;
458 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
459 DPRINTF("ioctl failed %d\n", errno);
460 ret = -1;
461 break;
464 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
465 start_addr = mem->start_addr + mem->memory_size;
467 g_free(d.dirty_bitmap);
469 return ret;
472 static void kvm_coalesce_mmio_region(MemoryListener *listener,
473 MemoryRegionSection *secion,
474 hwaddr start, hwaddr size)
476 KVMState *s = kvm_state;
478 if (s->coalesced_mmio) {
479 struct kvm_coalesced_mmio_zone zone;
481 zone.addr = start;
482 zone.size = size;
483 zone.pad = 0;
485 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
489 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
490 MemoryRegionSection *secion,
491 hwaddr start, hwaddr size)
493 KVMState *s = kvm_state;
495 if (s->coalesced_mmio) {
496 struct kvm_coalesced_mmio_zone zone;
498 zone.addr = start;
499 zone.size = size;
500 zone.pad = 0;
502 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
506 int kvm_check_extension(KVMState *s, unsigned int extension)
508 int ret;
510 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
511 if (ret < 0) {
512 ret = 0;
515 return ret;
518 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
520 int ret;
522 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
523 if (ret < 0) {
524 /* VM wide version not implemented, use global one instead */
525 ret = kvm_check_extension(s, extension);
528 return ret;
531 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
532 bool assign, uint32_t size, bool datamatch)
534 int ret;
535 struct kvm_ioeventfd iofd;
537 iofd.datamatch = datamatch ? val : 0;
538 iofd.addr = addr;
539 iofd.len = size;
540 iofd.flags = 0;
541 iofd.fd = fd;
543 if (!kvm_enabled()) {
544 return -ENOSYS;
547 if (datamatch) {
548 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
550 if (!assign) {
551 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
554 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
556 if (ret < 0) {
557 return -errno;
560 return 0;
563 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
564 bool assign, uint32_t size, bool datamatch)
566 struct kvm_ioeventfd kick = {
567 .datamatch = datamatch ? val : 0,
568 .addr = addr,
569 .flags = KVM_IOEVENTFD_FLAG_PIO,
570 .len = size,
571 .fd = fd,
573 int r;
574 if (!kvm_enabled()) {
575 return -ENOSYS;
577 if (datamatch) {
578 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
580 if (!assign) {
581 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
583 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
584 if (r < 0) {
585 return r;
587 return 0;
591 static int kvm_check_many_ioeventfds(void)
593 /* Userspace can use ioeventfd for io notification. This requires a host
594 * that supports eventfd(2) and an I/O thread; since eventfd does not
595 * support SIGIO it cannot interrupt the vcpu.
597 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
598 * can avoid creating too many ioeventfds.
600 #if defined(CONFIG_EVENTFD)
601 int ioeventfds[7];
602 int i, ret = 0;
603 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
604 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
605 if (ioeventfds[i] < 0) {
606 break;
608 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
609 if (ret < 0) {
610 close(ioeventfds[i]);
611 break;
615 /* Decide whether many devices are supported or not */
616 ret = i == ARRAY_SIZE(ioeventfds);
618 while (i-- > 0) {
619 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
620 close(ioeventfds[i]);
622 return ret;
623 #else
624 return 0;
625 #endif
628 static const KVMCapabilityInfo *
629 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
631 while (list->name) {
632 if (!kvm_check_extension(s, list->value)) {
633 return list;
635 list++;
637 return NULL;
640 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
642 KVMState *s = kvm_state;
643 KVMSlot *mem, old;
644 int err;
645 MemoryRegion *mr = section->mr;
646 bool log_dirty = memory_region_is_logging(mr);
647 bool writeable = !mr->readonly && !mr->rom_device;
648 bool readonly_flag = mr->readonly || memory_region_is_romd(mr);
649 hwaddr start_addr = section->offset_within_address_space;
650 ram_addr_t size = int128_get64(section->size);
651 void *ram = NULL;
652 unsigned delta;
654 /* kvm works in page size chunks, but the function may be called
655 with sub-page size and unaligned start address. Pad the start
656 address to next and truncate size to previous page boundary. */
657 delta = (TARGET_PAGE_SIZE - (start_addr & ~TARGET_PAGE_MASK));
658 delta &= ~TARGET_PAGE_MASK;
659 if (delta > size) {
660 return;
662 start_addr += delta;
663 size -= delta;
664 size &= TARGET_PAGE_MASK;
665 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
666 return;
669 if (!memory_region_is_ram(mr)) {
670 if (writeable || !kvm_readonly_mem_allowed) {
671 return;
672 } else if (!mr->romd_mode) {
673 /* If the memory device is not in romd_mode, then we actually want
674 * to remove the kvm memory slot so all accesses will trap. */
675 add = false;
679 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
681 while (1) {
682 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
683 if (!mem) {
684 break;
687 if (add && start_addr >= mem->start_addr &&
688 (start_addr + size <= mem->start_addr + mem->memory_size) &&
689 (ram - start_addr == mem->ram - mem->start_addr)) {
690 /* The new slot fits into the existing one and comes with
691 * identical parameters - update flags and done. */
692 kvm_slot_dirty_pages_log_change(mem, log_dirty);
693 return;
696 old = *mem;
698 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
699 kvm_physical_sync_dirty_bitmap(section);
702 /* unregister the overlapping slot */
703 mem->memory_size = 0;
704 err = kvm_set_user_memory_region(s, mem);
705 if (err) {
706 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
707 __func__, strerror(-err));
708 abort();
711 /* Workaround for older KVM versions: we can't join slots, even not by
712 * unregistering the previous ones and then registering the larger
713 * slot. We have to maintain the existing fragmentation. Sigh.
715 * This workaround assumes that the new slot starts at the same
716 * address as the first existing one. If not or if some overlapping
717 * slot comes around later, we will fail (not seen in practice so far)
718 * - and actually require a recent KVM version. */
719 if (s->broken_set_mem_region &&
720 old.start_addr == start_addr && old.memory_size < size && add) {
721 mem = kvm_alloc_slot(s);
722 mem->memory_size = old.memory_size;
723 mem->start_addr = old.start_addr;
724 mem->ram = old.ram;
725 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
727 err = kvm_set_user_memory_region(s, mem);
728 if (err) {
729 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
730 strerror(-err));
731 abort();
734 start_addr += old.memory_size;
735 ram += old.memory_size;
736 size -= old.memory_size;
737 continue;
740 /* register prefix slot */
741 if (old.start_addr < start_addr) {
742 mem = kvm_alloc_slot(s);
743 mem->memory_size = start_addr - old.start_addr;
744 mem->start_addr = old.start_addr;
745 mem->ram = old.ram;
746 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
748 err = kvm_set_user_memory_region(s, mem);
749 if (err) {
750 fprintf(stderr, "%s: error registering prefix slot: %s\n",
751 __func__, strerror(-err));
752 #ifdef TARGET_PPC
753 fprintf(stderr, "%s: This is probably because your kernel's " \
754 "PAGE_SIZE is too big. Please try to use 4k " \
755 "PAGE_SIZE!\n", __func__);
756 #endif
757 abort();
761 /* register suffix slot */
762 if (old.start_addr + old.memory_size > start_addr + size) {
763 ram_addr_t size_delta;
765 mem = kvm_alloc_slot(s);
766 mem->start_addr = start_addr + size;
767 size_delta = mem->start_addr - old.start_addr;
768 mem->memory_size = old.memory_size - size_delta;
769 mem->ram = old.ram + size_delta;
770 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
772 err = kvm_set_user_memory_region(s, mem);
773 if (err) {
774 fprintf(stderr, "%s: error registering suffix slot: %s\n",
775 __func__, strerror(-err));
776 abort();
781 /* in case the KVM bug workaround already "consumed" the new slot */
782 if (!size) {
783 return;
785 if (!add) {
786 return;
788 mem = kvm_alloc_slot(s);
789 mem->memory_size = size;
790 mem->start_addr = start_addr;
791 mem->ram = ram;
792 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
794 err = kvm_set_user_memory_region(s, mem);
795 if (err) {
796 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
797 strerror(-err));
798 abort();
802 static void kvm_region_add(MemoryListener *listener,
803 MemoryRegionSection *section)
805 memory_region_ref(section->mr);
806 kvm_set_phys_mem(section, true);
809 static void kvm_region_del(MemoryListener *listener,
810 MemoryRegionSection *section)
812 kvm_set_phys_mem(section, false);
813 memory_region_unref(section->mr);
816 static void kvm_log_sync(MemoryListener *listener,
817 MemoryRegionSection *section)
819 int r;
821 r = kvm_physical_sync_dirty_bitmap(section);
822 if (r < 0) {
823 abort();
827 static void kvm_log_global_start(struct MemoryListener *listener)
829 int r;
831 r = kvm_set_migration_log(1);
832 assert(r >= 0);
835 static void kvm_log_global_stop(struct MemoryListener *listener)
837 int r;
839 r = kvm_set_migration_log(0);
840 assert(r >= 0);
843 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
844 MemoryRegionSection *section,
845 bool match_data, uint64_t data,
846 EventNotifier *e)
848 int fd = event_notifier_get_fd(e);
849 int r;
851 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
852 data, true, int128_get64(section->size),
853 match_data);
854 if (r < 0) {
855 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
856 __func__, strerror(-r));
857 abort();
861 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
862 MemoryRegionSection *section,
863 bool match_data, uint64_t data,
864 EventNotifier *e)
866 int fd = event_notifier_get_fd(e);
867 int r;
869 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
870 data, false, int128_get64(section->size),
871 match_data);
872 if (r < 0) {
873 abort();
877 static void kvm_io_ioeventfd_add(MemoryListener *listener,
878 MemoryRegionSection *section,
879 bool match_data, uint64_t data,
880 EventNotifier *e)
882 int fd = event_notifier_get_fd(e);
883 int r;
885 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
886 data, true, int128_get64(section->size),
887 match_data);
888 if (r < 0) {
889 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
890 __func__, strerror(-r));
891 abort();
895 static void kvm_io_ioeventfd_del(MemoryListener *listener,
896 MemoryRegionSection *section,
897 bool match_data, uint64_t data,
898 EventNotifier *e)
901 int fd = event_notifier_get_fd(e);
902 int r;
904 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
905 data, false, int128_get64(section->size),
906 match_data);
907 if (r < 0) {
908 abort();
912 static MemoryListener kvm_memory_listener = {
913 .region_add = kvm_region_add,
914 .region_del = kvm_region_del,
915 .log_start = kvm_log_start,
916 .log_stop = kvm_log_stop,
917 .log_sync = kvm_log_sync,
918 .log_global_start = kvm_log_global_start,
919 .log_global_stop = kvm_log_global_stop,
920 .eventfd_add = kvm_mem_ioeventfd_add,
921 .eventfd_del = kvm_mem_ioeventfd_del,
922 .coalesced_mmio_add = kvm_coalesce_mmio_region,
923 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
924 .priority = 10,
927 static MemoryListener kvm_io_listener = {
928 .eventfd_add = kvm_io_ioeventfd_add,
929 .eventfd_del = kvm_io_ioeventfd_del,
930 .priority = 10,
933 static void kvm_handle_interrupt(CPUState *cpu, int mask)
935 cpu->interrupt_request |= mask;
937 if (!qemu_cpu_is_self(cpu)) {
938 qemu_cpu_kick(cpu);
942 int kvm_set_irq(KVMState *s, int irq, int level)
944 struct kvm_irq_level event;
945 int ret;
947 assert(kvm_async_interrupts_enabled());
949 event.level = level;
950 event.irq = irq;
951 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
952 if (ret < 0) {
953 perror("kvm_set_irq");
954 abort();
957 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
960 #ifdef KVM_CAP_IRQ_ROUTING
961 typedef struct KVMMSIRoute {
962 struct kvm_irq_routing_entry kroute;
963 QTAILQ_ENTRY(KVMMSIRoute) entry;
964 } KVMMSIRoute;
966 static void set_gsi(KVMState *s, unsigned int gsi)
968 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
971 static void clear_gsi(KVMState *s, unsigned int gsi)
973 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
976 void kvm_init_irq_routing(KVMState *s)
978 int gsi_count, i;
980 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
981 if (gsi_count > 0) {
982 unsigned int gsi_bits, i;
984 /* Round up so we can search ints using ffs */
985 gsi_bits = ALIGN(gsi_count, 32);
986 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
987 s->gsi_count = gsi_count;
989 /* Mark any over-allocated bits as already in use */
990 for (i = gsi_count; i < gsi_bits; i++) {
991 set_gsi(s, i);
995 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
996 s->nr_allocated_irq_routes = 0;
998 if (!s->direct_msi) {
999 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1000 QTAILQ_INIT(&s->msi_hashtab[i]);
1004 kvm_arch_init_irq_routing(s);
1007 void kvm_irqchip_commit_routes(KVMState *s)
1009 int ret;
1011 s->irq_routes->flags = 0;
1012 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1013 assert(ret == 0);
1016 static void kvm_add_routing_entry(KVMState *s,
1017 struct kvm_irq_routing_entry *entry)
1019 struct kvm_irq_routing_entry *new;
1020 int n, size;
1022 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1023 n = s->nr_allocated_irq_routes * 2;
1024 if (n < 64) {
1025 n = 64;
1027 size = sizeof(struct kvm_irq_routing);
1028 size += n * sizeof(*new);
1029 s->irq_routes = g_realloc(s->irq_routes, size);
1030 s->nr_allocated_irq_routes = n;
1032 n = s->irq_routes->nr++;
1033 new = &s->irq_routes->entries[n];
1035 *new = *entry;
1037 set_gsi(s, entry->gsi);
1040 static int kvm_update_routing_entry(KVMState *s,
1041 struct kvm_irq_routing_entry *new_entry)
1043 struct kvm_irq_routing_entry *entry;
1044 int n;
1046 for (n = 0; n < s->irq_routes->nr; n++) {
1047 entry = &s->irq_routes->entries[n];
1048 if (entry->gsi != new_entry->gsi) {
1049 continue;
1052 if(!memcmp(entry, new_entry, sizeof *entry)) {
1053 return 0;
1056 *entry = *new_entry;
1058 kvm_irqchip_commit_routes(s);
1060 return 0;
1063 return -ESRCH;
1066 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1068 struct kvm_irq_routing_entry e = {};
1070 assert(pin < s->gsi_count);
1072 e.gsi = irq;
1073 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1074 e.flags = 0;
1075 e.u.irqchip.irqchip = irqchip;
1076 e.u.irqchip.pin = pin;
1077 kvm_add_routing_entry(s, &e);
1080 void kvm_irqchip_release_virq(KVMState *s, int virq)
1082 struct kvm_irq_routing_entry *e;
1083 int i;
1085 if (kvm_gsi_direct_mapping()) {
1086 return;
1089 for (i = 0; i < s->irq_routes->nr; i++) {
1090 e = &s->irq_routes->entries[i];
1091 if (e->gsi == virq) {
1092 s->irq_routes->nr--;
1093 *e = s->irq_routes->entries[s->irq_routes->nr];
1096 clear_gsi(s, virq);
1099 static unsigned int kvm_hash_msi(uint32_t data)
1101 /* This is optimized for IA32 MSI layout. However, no other arch shall
1102 * repeat the mistake of not providing a direct MSI injection API. */
1103 return data & 0xff;
1106 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1108 KVMMSIRoute *route, *next;
1109 unsigned int hash;
1111 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1112 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1113 kvm_irqchip_release_virq(s, route->kroute.gsi);
1114 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1115 g_free(route);
1120 static int kvm_irqchip_get_virq(KVMState *s)
1122 uint32_t *word = s->used_gsi_bitmap;
1123 int max_words = ALIGN(s->gsi_count, 32) / 32;
1124 int i, bit;
1125 bool retry = true;
1127 again:
1128 /* Return the lowest unused GSI in the bitmap */
1129 for (i = 0; i < max_words; i++) {
1130 bit = ffs(~word[i]);
1131 if (!bit) {
1132 continue;
1135 return bit - 1 + i * 32;
1137 if (!s->direct_msi && retry) {
1138 retry = false;
1139 kvm_flush_dynamic_msi_routes(s);
1140 goto again;
1142 return -ENOSPC;
1146 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1148 unsigned int hash = kvm_hash_msi(msg.data);
1149 KVMMSIRoute *route;
1151 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1152 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1153 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1154 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1155 return route;
1158 return NULL;
1161 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1163 struct kvm_msi msi;
1164 KVMMSIRoute *route;
1166 if (s->direct_msi) {
1167 msi.address_lo = (uint32_t)msg.address;
1168 msi.address_hi = msg.address >> 32;
1169 msi.data = le32_to_cpu(msg.data);
1170 msi.flags = 0;
1171 memset(msi.pad, 0, sizeof(msi.pad));
1173 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1176 route = kvm_lookup_msi_route(s, msg);
1177 if (!route) {
1178 int virq;
1180 virq = kvm_irqchip_get_virq(s);
1181 if (virq < 0) {
1182 return virq;
1185 route = g_malloc0(sizeof(KVMMSIRoute));
1186 route->kroute.gsi = virq;
1187 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1188 route->kroute.flags = 0;
1189 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1190 route->kroute.u.msi.address_hi = msg.address >> 32;
1191 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1193 kvm_add_routing_entry(s, &route->kroute);
1194 kvm_irqchip_commit_routes(s);
1196 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1197 entry);
1200 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1202 return kvm_set_irq(s, route->kroute.gsi, 1);
1205 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1207 struct kvm_irq_routing_entry kroute = {};
1208 int virq;
1210 if (kvm_gsi_direct_mapping()) {
1211 return msg.data & 0xffff;
1214 if (!kvm_gsi_routing_enabled()) {
1215 return -ENOSYS;
1218 virq = kvm_irqchip_get_virq(s);
1219 if (virq < 0) {
1220 return virq;
1223 kroute.gsi = virq;
1224 kroute.type = KVM_IRQ_ROUTING_MSI;
1225 kroute.flags = 0;
1226 kroute.u.msi.address_lo = (uint32_t)msg.address;
1227 kroute.u.msi.address_hi = msg.address >> 32;
1228 kroute.u.msi.data = le32_to_cpu(msg.data);
1229 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1230 kvm_irqchip_release_virq(s, virq);
1231 return -EINVAL;
1234 kvm_add_routing_entry(s, &kroute);
1235 kvm_irqchip_commit_routes(s);
1237 return virq;
1240 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1242 struct kvm_irq_routing_entry kroute = {};
1244 if (kvm_gsi_direct_mapping()) {
1245 return 0;
1248 if (!kvm_irqchip_in_kernel()) {
1249 return -ENOSYS;
1252 kroute.gsi = virq;
1253 kroute.type = KVM_IRQ_ROUTING_MSI;
1254 kroute.flags = 0;
1255 kroute.u.msi.address_lo = (uint32_t)msg.address;
1256 kroute.u.msi.address_hi = msg.address >> 32;
1257 kroute.u.msi.data = le32_to_cpu(msg.data);
1258 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1259 return -EINVAL;
1262 return kvm_update_routing_entry(s, &kroute);
1265 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1266 bool assign)
1268 struct kvm_irqfd irqfd = {
1269 .fd = fd,
1270 .gsi = virq,
1271 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1274 if (rfd != -1) {
1275 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1276 irqfd.resamplefd = rfd;
1279 if (!kvm_irqfds_enabled()) {
1280 return -ENOSYS;
1283 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1286 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1288 struct kvm_irq_routing_entry kroute = {};
1289 int virq;
1291 if (!kvm_gsi_routing_enabled()) {
1292 return -ENOSYS;
1295 virq = kvm_irqchip_get_virq(s);
1296 if (virq < 0) {
1297 return virq;
1300 kroute.gsi = virq;
1301 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1302 kroute.flags = 0;
1303 kroute.u.adapter.summary_addr = adapter->summary_addr;
1304 kroute.u.adapter.ind_addr = adapter->ind_addr;
1305 kroute.u.adapter.summary_offset = adapter->summary_offset;
1306 kroute.u.adapter.ind_offset = adapter->ind_offset;
1307 kroute.u.adapter.adapter_id = adapter->adapter_id;
1309 kvm_add_routing_entry(s, &kroute);
1310 kvm_irqchip_commit_routes(s);
1312 return virq;
1315 #else /* !KVM_CAP_IRQ_ROUTING */
1317 void kvm_init_irq_routing(KVMState *s)
1321 void kvm_irqchip_release_virq(KVMState *s, int virq)
1325 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1327 abort();
1330 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1332 return -ENOSYS;
1335 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1337 return -ENOSYS;
1340 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1342 abort();
1345 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1347 return -ENOSYS;
1349 #endif /* !KVM_CAP_IRQ_ROUTING */
1351 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1352 EventNotifier *rn, int virq)
1354 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1355 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1358 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1360 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1361 false);
1364 static int kvm_irqchip_create(MachineState *machine, KVMState *s)
1366 int ret;
1368 if (!machine_kernel_irqchip_allowed(machine) ||
1369 (!kvm_check_extension(s, KVM_CAP_IRQCHIP) &&
1370 (kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0) < 0))) {
1371 return 0;
1374 /* First probe and see if there's a arch-specific hook to create the
1375 * in-kernel irqchip for us */
1376 ret = kvm_arch_irqchip_create(s);
1377 if (ret < 0) {
1378 return ret;
1379 } else if (ret == 0) {
1380 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1381 if (ret < 0) {
1382 fprintf(stderr, "Create kernel irqchip failed\n");
1383 return ret;
1387 kvm_kernel_irqchip = true;
1388 /* If we have an in-kernel IRQ chip then we must have asynchronous
1389 * interrupt delivery (though the reverse is not necessarily true)
1391 kvm_async_interrupts_allowed = true;
1392 kvm_halt_in_kernel_allowed = true;
1394 kvm_init_irq_routing(s);
1396 return 0;
1399 /* Find number of supported CPUs using the recommended
1400 * procedure from the kernel API documentation to cope with
1401 * older kernels that may be missing capabilities.
1403 static int kvm_recommended_vcpus(KVMState *s)
1405 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1406 return (ret) ? ret : 4;
1409 static int kvm_max_vcpus(KVMState *s)
1411 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1412 return (ret) ? ret : kvm_recommended_vcpus(s);
1415 static int kvm_init(MachineState *ms)
1417 MachineClass *mc = MACHINE_GET_CLASS(ms);
1418 static const char upgrade_note[] =
1419 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1420 "(see http://sourceforge.net/projects/kvm).\n";
1421 struct {
1422 const char *name;
1423 int num;
1424 } num_cpus[] = {
1425 { "SMP", smp_cpus },
1426 { "hotpluggable", max_cpus },
1427 { NULL, }
1428 }, *nc = num_cpus;
1429 int soft_vcpus_limit, hard_vcpus_limit;
1430 KVMState *s;
1431 const KVMCapabilityInfo *missing_cap;
1432 int ret;
1433 int i, type = 0;
1434 const char *kvm_type;
1436 s = KVM_STATE(ms->accelerator);
1439 * On systems where the kernel can support different base page
1440 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1441 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1442 * page size for the system though.
1444 assert(TARGET_PAGE_SIZE <= getpagesize());
1445 page_size_init();
1447 s->sigmask_len = 8;
1449 #ifdef KVM_CAP_SET_GUEST_DEBUG
1450 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1451 #endif
1452 s->vmfd = -1;
1453 s->fd = qemu_open("/dev/kvm", O_RDWR);
1454 if (s->fd == -1) {
1455 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1456 ret = -errno;
1457 goto err;
1460 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1461 if (ret < KVM_API_VERSION) {
1462 if (ret >= 0) {
1463 ret = -EINVAL;
1465 fprintf(stderr, "kvm version too old\n");
1466 goto err;
1469 if (ret > KVM_API_VERSION) {
1470 ret = -EINVAL;
1471 fprintf(stderr, "kvm version not supported\n");
1472 goto err;
1475 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1477 /* If unspecified, use the default value */
1478 if (!s->nr_slots) {
1479 s->nr_slots = 32;
1482 s->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1484 for (i = 0; i < s->nr_slots; i++) {
1485 s->slots[i].slot = i;
1488 /* check the vcpu limits */
1489 soft_vcpus_limit = kvm_recommended_vcpus(s);
1490 hard_vcpus_limit = kvm_max_vcpus(s);
1492 while (nc->name) {
1493 if (nc->num > soft_vcpus_limit) {
1494 fprintf(stderr,
1495 "Warning: Number of %s cpus requested (%d) exceeds "
1496 "the recommended cpus supported by KVM (%d)\n",
1497 nc->name, nc->num, soft_vcpus_limit);
1499 if (nc->num > hard_vcpus_limit) {
1500 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1501 "the maximum cpus supported by KVM (%d)\n",
1502 nc->name, nc->num, hard_vcpus_limit);
1503 exit(1);
1506 nc++;
1509 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1510 if (mc->kvm_type) {
1511 type = mc->kvm_type(kvm_type);
1512 } else if (kvm_type) {
1513 ret = -EINVAL;
1514 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1515 goto err;
1518 do {
1519 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1520 } while (ret == -EINTR);
1522 if (ret < 0) {
1523 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1524 strerror(-ret));
1526 #ifdef TARGET_S390X
1527 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1528 "your host kernel command line\n");
1529 #endif
1530 goto err;
1533 s->vmfd = ret;
1534 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1535 if (!missing_cap) {
1536 missing_cap =
1537 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1539 if (missing_cap) {
1540 ret = -EINVAL;
1541 fprintf(stderr, "kvm does not support %s\n%s",
1542 missing_cap->name, upgrade_note);
1543 goto err;
1546 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1548 s->broken_set_mem_region = 1;
1549 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1550 if (ret > 0) {
1551 s->broken_set_mem_region = 0;
1554 #ifdef KVM_CAP_VCPU_EVENTS
1555 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1556 #endif
1558 s->robust_singlestep =
1559 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1561 #ifdef KVM_CAP_DEBUGREGS
1562 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1563 #endif
1565 #ifdef KVM_CAP_XSAVE
1566 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1567 #endif
1569 #ifdef KVM_CAP_XCRS
1570 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1571 #endif
1573 #ifdef KVM_CAP_PIT_STATE2
1574 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1575 #endif
1577 #ifdef KVM_CAP_IRQ_ROUTING
1578 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1579 #endif
1581 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1583 s->irq_set_ioctl = KVM_IRQ_LINE;
1584 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1585 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1588 #ifdef KVM_CAP_READONLY_MEM
1589 kvm_readonly_mem_allowed =
1590 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1591 #endif
1593 kvm_eventfds_allowed =
1594 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1596 kvm_irqfds_allowed =
1597 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1599 kvm_resamplefds_allowed =
1600 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1602 kvm_vm_attributes_allowed =
1603 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1605 ret = kvm_arch_init(ms, s);
1606 if (ret < 0) {
1607 goto err;
1610 ret = kvm_irqchip_create(ms, s);
1611 if (ret < 0) {
1612 goto err;
1615 kvm_state = s;
1616 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1617 memory_listener_register(&kvm_io_listener, &address_space_io);
1619 s->many_ioeventfds = kvm_check_many_ioeventfds();
1621 cpu_interrupt_handler = kvm_handle_interrupt;
1623 return 0;
1625 err:
1626 assert(ret < 0);
1627 if (s->vmfd >= 0) {
1628 close(s->vmfd);
1630 if (s->fd != -1) {
1631 close(s->fd);
1633 g_free(s->slots);
1635 return ret;
1638 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1640 s->sigmask_len = sigmask_len;
1643 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1644 uint32_t count)
1646 int i;
1647 uint8_t *ptr = data;
1649 for (i = 0; i < count; i++) {
1650 address_space_rw(&address_space_io, port, ptr, size,
1651 direction == KVM_EXIT_IO_OUT);
1652 ptr += size;
1656 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1658 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1659 run->internal.suberror);
1661 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1662 int i;
1664 for (i = 0; i < run->internal.ndata; ++i) {
1665 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1666 i, (uint64_t)run->internal.data[i]);
1669 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1670 fprintf(stderr, "emulation failure\n");
1671 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1672 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1673 return EXCP_INTERRUPT;
1676 /* FIXME: Should trigger a qmp message to let management know
1677 * something went wrong.
1679 return -1;
1682 void kvm_flush_coalesced_mmio_buffer(void)
1684 KVMState *s = kvm_state;
1686 if (s->coalesced_flush_in_progress) {
1687 return;
1690 s->coalesced_flush_in_progress = true;
1692 if (s->coalesced_mmio_ring) {
1693 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1694 while (ring->first != ring->last) {
1695 struct kvm_coalesced_mmio *ent;
1697 ent = &ring->coalesced_mmio[ring->first];
1699 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1700 smp_wmb();
1701 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1705 s->coalesced_flush_in_progress = false;
1708 static void do_kvm_cpu_synchronize_state(void *arg)
1710 CPUState *cpu = arg;
1712 if (!cpu->kvm_vcpu_dirty) {
1713 kvm_arch_get_registers(cpu);
1714 cpu->kvm_vcpu_dirty = true;
1718 void kvm_cpu_synchronize_state(CPUState *cpu)
1720 if (!cpu->kvm_vcpu_dirty) {
1721 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1725 static void do_kvm_cpu_synchronize_post_reset(void *arg)
1727 CPUState *cpu = arg;
1729 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1730 cpu->kvm_vcpu_dirty = false;
1733 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1735 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, cpu);
1738 static void do_kvm_cpu_synchronize_post_init(void *arg)
1740 CPUState *cpu = arg;
1742 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1743 cpu->kvm_vcpu_dirty = false;
1746 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1748 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, cpu);
1751 void kvm_cpu_clean_state(CPUState *cpu)
1753 cpu->kvm_vcpu_dirty = false;
1756 int kvm_cpu_exec(CPUState *cpu)
1758 struct kvm_run *run = cpu->kvm_run;
1759 int ret, run_ret;
1761 DPRINTF("kvm_cpu_exec()\n");
1763 if (kvm_arch_process_async_events(cpu)) {
1764 cpu->exit_request = 0;
1765 return EXCP_HLT;
1768 do {
1769 if (cpu->kvm_vcpu_dirty) {
1770 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1771 cpu->kvm_vcpu_dirty = false;
1774 kvm_arch_pre_run(cpu, run);
1775 if (cpu->exit_request) {
1776 DPRINTF("interrupt exit requested\n");
1778 * KVM requires us to reenter the kernel after IO exits to complete
1779 * instruction emulation. This self-signal will ensure that we
1780 * leave ASAP again.
1782 qemu_cpu_kick_self();
1784 qemu_mutex_unlock_iothread();
1786 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1788 qemu_mutex_lock_iothread();
1789 kvm_arch_post_run(cpu, run);
1791 if (run_ret < 0) {
1792 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1793 DPRINTF("io window exit\n");
1794 ret = EXCP_INTERRUPT;
1795 break;
1797 fprintf(stderr, "error: kvm run failed %s\n",
1798 strerror(-run_ret));
1799 ret = -1;
1800 break;
1803 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1804 switch (run->exit_reason) {
1805 case KVM_EXIT_IO:
1806 DPRINTF("handle_io\n");
1807 kvm_handle_io(run->io.port,
1808 (uint8_t *)run + run->io.data_offset,
1809 run->io.direction,
1810 run->io.size,
1811 run->io.count);
1812 ret = 0;
1813 break;
1814 case KVM_EXIT_MMIO:
1815 DPRINTF("handle_mmio\n");
1816 cpu_physical_memory_rw(run->mmio.phys_addr,
1817 run->mmio.data,
1818 run->mmio.len,
1819 run->mmio.is_write);
1820 ret = 0;
1821 break;
1822 case KVM_EXIT_IRQ_WINDOW_OPEN:
1823 DPRINTF("irq_window_open\n");
1824 ret = EXCP_INTERRUPT;
1825 break;
1826 case KVM_EXIT_SHUTDOWN:
1827 DPRINTF("shutdown\n");
1828 qemu_system_reset_request();
1829 ret = EXCP_INTERRUPT;
1830 break;
1831 case KVM_EXIT_UNKNOWN:
1832 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1833 (uint64_t)run->hw.hardware_exit_reason);
1834 ret = -1;
1835 break;
1836 case KVM_EXIT_INTERNAL_ERROR:
1837 ret = kvm_handle_internal_error(cpu, run);
1838 break;
1839 case KVM_EXIT_SYSTEM_EVENT:
1840 switch (run->system_event.type) {
1841 case KVM_SYSTEM_EVENT_SHUTDOWN:
1842 qemu_system_shutdown_request();
1843 ret = EXCP_INTERRUPT;
1844 break;
1845 case KVM_SYSTEM_EVENT_RESET:
1846 qemu_system_reset_request();
1847 ret = EXCP_INTERRUPT;
1848 break;
1849 default:
1850 DPRINTF("kvm_arch_handle_exit\n");
1851 ret = kvm_arch_handle_exit(cpu, run);
1852 break;
1854 break;
1855 default:
1856 DPRINTF("kvm_arch_handle_exit\n");
1857 ret = kvm_arch_handle_exit(cpu, run);
1858 break;
1860 } while (ret == 0);
1862 if (ret < 0) {
1863 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1864 vm_stop(RUN_STATE_INTERNAL_ERROR);
1867 cpu->exit_request = 0;
1868 return ret;
1871 int kvm_ioctl(KVMState *s, int type, ...)
1873 int ret;
1874 void *arg;
1875 va_list ap;
1877 va_start(ap, type);
1878 arg = va_arg(ap, void *);
1879 va_end(ap);
1881 trace_kvm_ioctl(type, arg);
1882 ret = ioctl(s->fd, type, arg);
1883 if (ret == -1) {
1884 ret = -errno;
1886 return ret;
1889 int kvm_vm_ioctl(KVMState *s, int type, ...)
1891 int ret;
1892 void *arg;
1893 va_list ap;
1895 va_start(ap, type);
1896 arg = va_arg(ap, void *);
1897 va_end(ap);
1899 trace_kvm_vm_ioctl(type, arg);
1900 ret = ioctl(s->vmfd, type, arg);
1901 if (ret == -1) {
1902 ret = -errno;
1904 return ret;
1907 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1909 int ret;
1910 void *arg;
1911 va_list ap;
1913 va_start(ap, type);
1914 arg = va_arg(ap, void *);
1915 va_end(ap);
1917 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1918 ret = ioctl(cpu->kvm_fd, type, arg);
1919 if (ret == -1) {
1920 ret = -errno;
1922 return ret;
1925 int kvm_device_ioctl(int fd, int type, ...)
1927 int ret;
1928 void *arg;
1929 va_list ap;
1931 va_start(ap, type);
1932 arg = va_arg(ap, void *);
1933 va_end(ap);
1935 trace_kvm_device_ioctl(fd, type, arg);
1936 ret = ioctl(fd, type, arg);
1937 if (ret == -1) {
1938 ret = -errno;
1940 return ret;
1943 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
1945 int ret;
1946 struct kvm_device_attr attribute = {
1947 .group = group,
1948 .attr = attr,
1951 if (!kvm_vm_attributes_allowed) {
1952 return 0;
1955 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
1956 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
1957 return ret ? 0 : 1;
1960 int kvm_has_sync_mmu(void)
1962 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1965 int kvm_has_vcpu_events(void)
1967 return kvm_state->vcpu_events;
1970 int kvm_has_robust_singlestep(void)
1972 return kvm_state->robust_singlestep;
1975 int kvm_has_debugregs(void)
1977 return kvm_state->debugregs;
1980 int kvm_has_xsave(void)
1982 return kvm_state->xsave;
1985 int kvm_has_xcrs(void)
1987 return kvm_state->xcrs;
1990 int kvm_has_pit_state2(void)
1992 return kvm_state->pit_state2;
1995 int kvm_has_many_ioeventfds(void)
1997 if (!kvm_enabled()) {
1998 return 0;
2000 return kvm_state->many_ioeventfds;
2003 int kvm_has_gsi_routing(void)
2005 #ifdef KVM_CAP_IRQ_ROUTING
2006 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2007 #else
2008 return false;
2009 #endif
2012 int kvm_has_intx_set_mask(void)
2014 return kvm_state->intx_set_mask;
2017 void kvm_setup_guest_memory(void *start, size_t size)
2019 if (!kvm_has_sync_mmu()) {
2020 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
2022 if (ret) {
2023 perror("qemu_madvise");
2024 fprintf(stderr,
2025 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2026 exit(1);
2031 #ifdef KVM_CAP_SET_GUEST_DEBUG
2032 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2033 target_ulong pc)
2035 struct kvm_sw_breakpoint *bp;
2037 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2038 if (bp->pc == pc) {
2039 return bp;
2042 return NULL;
2045 int kvm_sw_breakpoints_active(CPUState *cpu)
2047 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2050 struct kvm_set_guest_debug_data {
2051 struct kvm_guest_debug dbg;
2052 CPUState *cpu;
2053 int err;
2056 static void kvm_invoke_set_guest_debug(void *data)
2058 struct kvm_set_guest_debug_data *dbg_data = data;
2060 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
2061 &dbg_data->dbg);
2064 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2066 struct kvm_set_guest_debug_data data;
2068 data.dbg.control = reinject_trap;
2070 if (cpu->singlestep_enabled) {
2071 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2073 kvm_arch_update_guest_debug(cpu, &data.dbg);
2074 data.cpu = cpu;
2076 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
2077 return data.err;
2080 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2081 target_ulong len, int type)
2083 struct kvm_sw_breakpoint *bp;
2084 int err;
2086 if (type == GDB_BREAKPOINT_SW) {
2087 bp = kvm_find_sw_breakpoint(cpu, addr);
2088 if (bp) {
2089 bp->use_count++;
2090 return 0;
2093 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2094 bp->pc = addr;
2095 bp->use_count = 1;
2096 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2097 if (err) {
2098 g_free(bp);
2099 return err;
2102 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2103 } else {
2104 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2105 if (err) {
2106 return err;
2110 CPU_FOREACH(cpu) {
2111 err = kvm_update_guest_debug(cpu, 0);
2112 if (err) {
2113 return err;
2116 return 0;
2119 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2120 target_ulong len, int type)
2122 struct kvm_sw_breakpoint *bp;
2123 int err;
2125 if (type == GDB_BREAKPOINT_SW) {
2126 bp = kvm_find_sw_breakpoint(cpu, addr);
2127 if (!bp) {
2128 return -ENOENT;
2131 if (bp->use_count > 1) {
2132 bp->use_count--;
2133 return 0;
2136 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2137 if (err) {
2138 return err;
2141 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2142 g_free(bp);
2143 } else {
2144 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2145 if (err) {
2146 return err;
2150 CPU_FOREACH(cpu) {
2151 err = kvm_update_guest_debug(cpu, 0);
2152 if (err) {
2153 return err;
2156 return 0;
2159 void kvm_remove_all_breakpoints(CPUState *cpu)
2161 struct kvm_sw_breakpoint *bp, *next;
2162 KVMState *s = cpu->kvm_state;
2163 CPUState *tmpcpu;
2165 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2166 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2167 /* Try harder to find a CPU that currently sees the breakpoint. */
2168 CPU_FOREACH(tmpcpu) {
2169 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2170 break;
2174 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2175 g_free(bp);
2177 kvm_arch_remove_all_hw_breakpoints();
2179 CPU_FOREACH(cpu) {
2180 kvm_update_guest_debug(cpu, 0);
2184 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2186 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2188 return -EINVAL;
2191 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2192 target_ulong len, int type)
2194 return -EINVAL;
2197 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2198 target_ulong len, int type)
2200 return -EINVAL;
2203 void kvm_remove_all_breakpoints(CPUState *cpu)
2206 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2208 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2210 KVMState *s = kvm_state;
2211 struct kvm_signal_mask *sigmask;
2212 int r;
2214 if (!sigset) {
2215 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2218 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2220 sigmask->len = s->sigmask_len;
2221 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2222 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2223 g_free(sigmask);
2225 return r;
2227 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2229 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2232 int kvm_on_sigbus(int code, void *addr)
2234 return kvm_arch_on_sigbus(code, addr);
2237 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2239 int ret;
2240 struct kvm_create_device create_dev;
2242 create_dev.type = type;
2243 create_dev.fd = -1;
2244 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2246 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2247 return -ENOTSUP;
2250 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2251 if (ret) {
2252 return ret;
2255 return test ? 0 : create_dev.fd;
2258 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2260 struct kvm_one_reg reg;
2261 int r;
2263 reg.id = id;
2264 reg.addr = (uintptr_t) source;
2265 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2266 if (r) {
2267 trace_kvm_failed_reg_set(id, strerror(r));
2269 return r;
2272 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2274 struct kvm_one_reg reg;
2275 int r;
2277 reg.id = id;
2278 reg.addr = (uintptr_t) target;
2279 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2280 if (r) {
2281 trace_kvm_failed_reg_get(id, strerror(r));
2283 return r;
2286 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2288 AccelClass *ac = ACCEL_CLASS(oc);
2289 ac->name = "KVM";
2290 ac->init_machine = kvm_init;
2291 ac->allowed = &kvm_allowed;
2294 static const TypeInfo kvm_accel_type = {
2295 .name = TYPE_KVM_ACCEL,
2296 .parent = TYPE_ACCEL,
2297 .class_init = kvm_accel_class_init,
2298 .instance_size = sizeof(KVMState),
2301 static void kvm_type_init(void)
2303 type_register_static(&kvm_accel_type);
2306 type_init(kvm_type_init);