m48t59: add a Nvram interface
[qemu/ar7.git] / kvm-all.c
blob05a79c20e0bbabfd8e46515001e9b30d6f7e0a0e
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;
130 static const KVMCapabilityInfo kvm_required_capabilites[] = {
131 KVM_CAP_INFO(USER_MEMORY),
132 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
133 KVM_CAP_LAST_INFO
136 static KVMSlot *kvm_get_free_slot(KVMState *s)
138 int i;
140 for (i = 0; i < s->nr_slots; i++) {
141 if (s->slots[i].memory_size == 0) {
142 return &s->slots[i];
146 return NULL;
149 bool kvm_has_free_slot(MachineState *ms)
151 return kvm_get_free_slot(KVM_STATE(ms->accelerator));
154 static KVMSlot *kvm_alloc_slot(KVMState *s)
156 KVMSlot *slot = kvm_get_free_slot(s);
158 if (slot) {
159 return slot;
162 fprintf(stderr, "%s: no free slot available\n", __func__);
163 abort();
166 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
167 hwaddr start_addr,
168 hwaddr end_addr)
170 int i;
172 for (i = 0; i < s->nr_slots; i++) {
173 KVMSlot *mem = &s->slots[i];
175 if (start_addr == mem->start_addr &&
176 end_addr == mem->start_addr + mem->memory_size) {
177 return mem;
181 return NULL;
185 * Find overlapping slot with lowest start address
187 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
188 hwaddr start_addr,
189 hwaddr end_addr)
191 KVMSlot *found = NULL;
192 int i;
194 for (i = 0; i < s->nr_slots; i++) {
195 KVMSlot *mem = &s->slots[i];
197 if (mem->memory_size == 0 ||
198 (found && found->start_addr < mem->start_addr)) {
199 continue;
202 if (end_addr > mem->start_addr &&
203 start_addr < mem->start_addr + mem->memory_size) {
204 found = mem;
208 return found;
211 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
212 hwaddr *phys_addr)
214 int i;
216 for (i = 0; i < s->nr_slots; i++) {
217 KVMSlot *mem = &s->slots[i];
219 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
220 *phys_addr = mem->start_addr + (ram - mem->ram);
221 return 1;
225 return 0;
228 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
230 struct kvm_userspace_memory_region mem;
232 mem.slot = slot->slot;
233 mem.guest_phys_addr = slot->start_addr;
234 mem.userspace_addr = (unsigned long)slot->ram;
235 mem.flags = slot->flags;
236 if (s->migration_log) {
237 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
240 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
241 /* Set the slot size to 0 before setting the slot to the desired
242 * value. This is needed based on KVM commit 75d61fbc. */
243 mem.memory_size = 0;
244 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
246 mem.memory_size = slot->memory_size;
247 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
250 int kvm_init_vcpu(CPUState *cpu)
252 KVMState *s = kvm_state;
253 long mmap_size;
254 int ret;
256 DPRINTF("kvm_init_vcpu\n");
258 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
259 if (ret < 0) {
260 DPRINTF("kvm_create_vcpu failed\n");
261 goto err;
264 cpu->kvm_fd = ret;
265 cpu->kvm_state = s;
266 cpu->kvm_vcpu_dirty = true;
268 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
269 if (mmap_size < 0) {
270 ret = mmap_size;
271 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
272 goto err;
275 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
276 cpu->kvm_fd, 0);
277 if (cpu->kvm_run == MAP_FAILED) {
278 ret = -errno;
279 DPRINTF("mmap'ing vcpu state failed\n");
280 goto err;
283 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
284 s->coalesced_mmio_ring =
285 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
288 ret = kvm_arch_init_vcpu(cpu);
289 err:
290 return ret;
294 * dirty pages logging control
297 static int kvm_mem_flags(KVMState *s, bool log_dirty, bool readonly)
299 int flags = 0;
300 flags = log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
301 if (readonly && kvm_readonly_mem_allowed) {
302 flags |= KVM_MEM_READONLY;
304 return flags;
307 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
309 KVMState *s = kvm_state;
310 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
311 int old_flags;
313 old_flags = mem->flags;
315 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
316 mem->flags = flags;
318 /* If nothing changed effectively, no need to issue ioctl */
319 if (s->migration_log) {
320 flags |= KVM_MEM_LOG_DIRTY_PAGES;
323 if (flags == old_flags) {
324 return 0;
327 return kvm_set_user_memory_region(s, mem);
330 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
331 ram_addr_t size, bool log_dirty)
333 KVMState *s = kvm_state;
334 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
336 if (mem == NULL) {
337 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
338 TARGET_FMT_plx "\n", __func__, phys_addr,
339 (hwaddr)(phys_addr + size - 1));
340 return -EINVAL;
342 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
345 static void kvm_log_start(MemoryListener *listener,
346 MemoryRegionSection *section)
348 int r;
350 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
351 int128_get64(section->size), true);
352 if (r < 0) {
353 abort();
357 static void kvm_log_stop(MemoryListener *listener,
358 MemoryRegionSection *section)
360 int r;
362 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
363 int128_get64(section->size), false);
364 if (r < 0) {
365 abort();
369 static int kvm_set_migration_log(int enable)
371 KVMState *s = kvm_state;
372 KVMSlot *mem;
373 int i, err;
375 s->migration_log = enable;
377 for (i = 0; i < s->nr_slots; i++) {
378 mem = &s->slots[i];
380 if (!mem->memory_size) {
381 continue;
383 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
384 continue;
386 err = kvm_set_user_memory_region(s, mem);
387 if (err) {
388 return err;
391 return 0;
394 /* get kvm's dirty pages bitmap and update qemu's */
395 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
396 unsigned long *bitmap)
398 ram_addr_t start = section->offset_within_region + section->mr->ram_addr;
399 ram_addr_t pages = int128_get64(section->size) / getpagesize();
401 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
402 return 0;
405 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
408 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
409 * This function updates qemu's dirty bitmap using
410 * memory_region_set_dirty(). This means all bits are set
411 * to dirty.
413 * @start_add: start of logged region.
414 * @end_addr: end of logged region.
416 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
418 KVMState *s = kvm_state;
419 unsigned long size, allocated_size = 0;
420 KVMDirtyLog d = {};
421 KVMSlot *mem;
422 int ret = 0;
423 hwaddr start_addr = section->offset_within_address_space;
424 hwaddr end_addr = start_addr + int128_get64(section->size);
426 d.dirty_bitmap = NULL;
427 while (start_addr < end_addr) {
428 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
429 if (mem == NULL) {
430 break;
433 /* XXX bad kernel interface alert
434 * For dirty bitmap, kernel allocates array of size aligned to
435 * bits-per-long. But for case when the kernel is 64bits and
436 * the userspace is 32bits, userspace can't align to the same
437 * bits-per-long, since sizeof(long) is different between kernel
438 * and user space. This way, userspace will provide buffer which
439 * may be 4 bytes less than the kernel will use, resulting in
440 * userspace memory corruption (which is not detectable by valgrind
441 * too, in most cases).
442 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
443 * a hope that sizeof(long) wont become >8 any time soon.
445 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
446 /*HOST_LONG_BITS*/ 64) / 8;
447 if (!d.dirty_bitmap) {
448 d.dirty_bitmap = g_malloc(size);
449 } else if (size > allocated_size) {
450 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
452 allocated_size = size;
453 memset(d.dirty_bitmap, 0, allocated_size);
455 d.slot = mem->slot;
457 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
458 DPRINTF("ioctl failed %d\n", errno);
459 ret = -1;
460 break;
463 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
464 start_addr = mem->start_addr + mem->memory_size;
466 g_free(d.dirty_bitmap);
468 return ret;
471 static void kvm_coalesce_mmio_region(MemoryListener *listener,
472 MemoryRegionSection *secion,
473 hwaddr start, hwaddr size)
475 KVMState *s = kvm_state;
477 if (s->coalesced_mmio) {
478 struct kvm_coalesced_mmio_zone zone;
480 zone.addr = start;
481 zone.size = size;
482 zone.pad = 0;
484 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
488 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
489 MemoryRegionSection *secion,
490 hwaddr start, hwaddr size)
492 KVMState *s = kvm_state;
494 if (s->coalesced_mmio) {
495 struct kvm_coalesced_mmio_zone zone;
497 zone.addr = start;
498 zone.size = size;
499 zone.pad = 0;
501 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
505 int kvm_check_extension(KVMState *s, unsigned int extension)
507 int ret;
509 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
510 if (ret < 0) {
511 ret = 0;
514 return ret;
517 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
519 int ret;
521 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
522 if (ret < 0) {
523 /* VM wide version not implemented, use global one instead */
524 ret = kvm_check_extension(s, extension);
527 return ret;
530 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
531 bool assign, uint32_t size, bool datamatch)
533 int ret;
534 struct kvm_ioeventfd iofd;
536 iofd.datamatch = datamatch ? val : 0;
537 iofd.addr = addr;
538 iofd.len = size;
539 iofd.flags = 0;
540 iofd.fd = fd;
542 if (!kvm_enabled()) {
543 return -ENOSYS;
546 if (datamatch) {
547 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
549 if (!assign) {
550 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
553 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
555 if (ret < 0) {
556 return -errno;
559 return 0;
562 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
563 bool assign, uint32_t size, bool datamatch)
565 struct kvm_ioeventfd kick = {
566 .datamatch = datamatch ? val : 0,
567 .addr = addr,
568 .flags = KVM_IOEVENTFD_FLAG_PIO,
569 .len = size,
570 .fd = fd,
572 int r;
573 if (!kvm_enabled()) {
574 return -ENOSYS;
576 if (datamatch) {
577 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
579 if (!assign) {
580 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
582 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
583 if (r < 0) {
584 return r;
586 return 0;
590 static int kvm_check_many_ioeventfds(void)
592 /* Userspace can use ioeventfd for io notification. This requires a host
593 * that supports eventfd(2) and an I/O thread; since eventfd does not
594 * support SIGIO it cannot interrupt the vcpu.
596 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
597 * can avoid creating too many ioeventfds.
599 #if defined(CONFIG_EVENTFD)
600 int ioeventfds[7];
601 int i, ret = 0;
602 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
603 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
604 if (ioeventfds[i] < 0) {
605 break;
607 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
608 if (ret < 0) {
609 close(ioeventfds[i]);
610 break;
614 /* Decide whether many devices are supported or not */
615 ret = i == ARRAY_SIZE(ioeventfds);
617 while (i-- > 0) {
618 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
619 close(ioeventfds[i]);
621 return ret;
622 #else
623 return 0;
624 #endif
627 static const KVMCapabilityInfo *
628 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
630 while (list->name) {
631 if (!kvm_check_extension(s, list->value)) {
632 return list;
634 list++;
636 return NULL;
639 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
641 KVMState *s = kvm_state;
642 KVMSlot *mem, old;
643 int err;
644 MemoryRegion *mr = section->mr;
645 bool log_dirty = memory_region_is_logging(mr);
646 bool writeable = !mr->readonly && !mr->rom_device;
647 bool readonly_flag = mr->readonly || memory_region_is_romd(mr);
648 hwaddr start_addr = section->offset_within_address_space;
649 ram_addr_t size = int128_get64(section->size);
650 void *ram = NULL;
651 unsigned delta;
653 /* kvm works in page size chunks, but the function may be called
654 with sub-page size and unaligned start address. Pad the start
655 address to next and truncate size to previous page boundary. */
656 delta = (TARGET_PAGE_SIZE - (start_addr & ~TARGET_PAGE_MASK));
657 delta &= ~TARGET_PAGE_MASK;
658 if (delta > size) {
659 return;
661 start_addr += delta;
662 size -= delta;
663 size &= TARGET_PAGE_MASK;
664 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
665 return;
668 if (!memory_region_is_ram(mr)) {
669 if (writeable || !kvm_readonly_mem_allowed) {
670 return;
671 } else if (!mr->romd_mode) {
672 /* If the memory device is not in romd_mode, then we actually want
673 * to remove the kvm memory slot so all accesses will trap. */
674 add = false;
678 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
680 while (1) {
681 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
682 if (!mem) {
683 break;
686 if (add && start_addr >= mem->start_addr &&
687 (start_addr + size <= mem->start_addr + mem->memory_size) &&
688 (ram - start_addr == mem->ram - mem->start_addr)) {
689 /* The new slot fits into the existing one and comes with
690 * identical parameters - update flags and done. */
691 kvm_slot_dirty_pages_log_change(mem, log_dirty);
692 return;
695 old = *mem;
697 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
698 kvm_physical_sync_dirty_bitmap(section);
701 /* unregister the overlapping slot */
702 mem->memory_size = 0;
703 err = kvm_set_user_memory_region(s, mem);
704 if (err) {
705 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
706 __func__, strerror(-err));
707 abort();
710 /* Workaround for older KVM versions: we can't join slots, even not by
711 * unregistering the previous ones and then registering the larger
712 * slot. We have to maintain the existing fragmentation. Sigh.
714 * This workaround assumes that the new slot starts at the same
715 * address as the first existing one. If not or if some overlapping
716 * slot comes around later, we will fail (not seen in practice so far)
717 * - and actually require a recent KVM version. */
718 if (s->broken_set_mem_region &&
719 old.start_addr == start_addr && old.memory_size < size && add) {
720 mem = kvm_alloc_slot(s);
721 mem->memory_size = old.memory_size;
722 mem->start_addr = old.start_addr;
723 mem->ram = old.ram;
724 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
726 err = kvm_set_user_memory_region(s, mem);
727 if (err) {
728 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
729 strerror(-err));
730 abort();
733 start_addr += old.memory_size;
734 ram += old.memory_size;
735 size -= old.memory_size;
736 continue;
739 /* register prefix slot */
740 if (old.start_addr < start_addr) {
741 mem = kvm_alloc_slot(s);
742 mem->memory_size = start_addr - old.start_addr;
743 mem->start_addr = old.start_addr;
744 mem->ram = old.ram;
745 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
747 err = kvm_set_user_memory_region(s, mem);
748 if (err) {
749 fprintf(stderr, "%s: error registering prefix slot: %s\n",
750 __func__, strerror(-err));
751 #ifdef TARGET_PPC
752 fprintf(stderr, "%s: This is probably because your kernel's " \
753 "PAGE_SIZE is too big. Please try to use 4k " \
754 "PAGE_SIZE!\n", __func__);
755 #endif
756 abort();
760 /* register suffix slot */
761 if (old.start_addr + old.memory_size > start_addr + size) {
762 ram_addr_t size_delta;
764 mem = kvm_alloc_slot(s);
765 mem->start_addr = start_addr + size;
766 size_delta = mem->start_addr - old.start_addr;
767 mem->memory_size = old.memory_size - size_delta;
768 mem->ram = old.ram + size_delta;
769 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
771 err = kvm_set_user_memory_region(s, mem);
772 if (err) {
773 fprintf(stderr, "%s: error registering suffix slot: %s\n",
774 __func__, strerror(-err));
775 abort();
780 /* in case the KVM bug workaround already "consumed" the new slot */
781 if (!size) {
782 return;
784 if (!add) {
785 return;
787 mem = kvm_alloc_slot(s);
788 mem->memory_size = size;
789 mem->start_addr = start_addr;
790 mem->ram = ram;
791 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
793 err = kvm_set_user_memory_region(s, mem);
794 if (err) {
795 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
796 strerror(-err));
797 abort();
801 static void kvm_region_add(MemoryListener *listener,
802 MemoryRegionSection *section)
804 memory_region_ref(section->mr);
805 kvm_set_phys_mem(section, true);
808 static void kvm_region_del(MemoryListener *listener,
809 MemoryRegionSection *section)
811 kvm_set_phys_mem(section, false);
812 memory_region_unref(section->mr);
815 static void kvm_log_sync(MemoryListener *listener,
816 MemoryRegionSection *section)
818 int r;
820 r = kvm_physical_sync_dirty_bitmap(section);
821 if (r < 0) {
822 abort();
826 static void kvm_log_global_start(struct MemoryListener *listener)
828 int r;
830 r = kvm_set_migration_log(1);
831 assert(r >= 0);
834 static void kvm_log_global_stop(struct MemoryListener *listener)
836 int r;
838 r = kvm_set_migration_log(0);
839 assert(r >= 0);
842 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
843 MemoryRegionSection *section,
844 bool match_data, uint64_t data,
845 EventNotifier *e)
847 int fd = event_notifier_get_fd(e);
848 int r;
850 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
851 data, true, int128_get64(section->size),
852 match_data);
853 if (r < 0) {
854 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
855 __func__, strerror(-r));
856 abort();
860 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
861 MemoryRegionSection *section,
862 bool match_data, uint64_t data,
863 EventNotifier *e)
865 int fd = event_notifier_get_fd(e);
866 int r;
868 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
869 data, false, int128_get64(section->size),
870 match_data);
871 if (r < 0) {
872 abort();
876 static void kvm_io_ioeventfd_add(MemoryListener *listener,
877 MemoryRegionSection *section,
878 bool match_data, uint64_t data,
879 EventNotifier *e)
881 int fd = event_notifier_get_fd(e);
882 int r;
884 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
885 data, true, int128_get64(section->size),
886 match_data);
887 if (r < 0) {
888 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
889 __func__, strerror(-r));
890 abort();
894 static void kvm_io_ioeventfd_del(MemoryListener *listener,
895 MemoryRegionSection *section,
896 bool match_data, uint64_t data,
897 EventNotifier *e)
900 int fd = event_notifier_get_fd(e);
901 int r;
903 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
904 data, false, int128_get64(section->size),
905 match_data);
906 if (r < 0) {
907 abort();
911 static MemoryListener kvm_memory_listener = {
912 .region_add = kvm_region_add,
913 .region_del = kvm_region_del,
914 .log_start = kvm_log_start,
915 .log_stop = kvm_log_stop,
916 .log_sync = kvm_log_sync,
917 .log_global_start = kvm_log_global_start,
918 .log_global_stop = kvm_log_global_stop,
919 .eventfd_add = kvm_mem_ioeventfd_add,
920 .eventfd_del = kvm_mem_ioeventfd_del,
921 .coalesced_mmio_add = kvm_coalesce_mmio_region,
922 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
923 .priority = 10,
926 static MemoryListener kvm_io_listener = {
927 .eventfd_add = kvm_io_ioeventfd_add,
928 .eventfd_del = kvm_io_ioeventfd_del,
929 .priority = 10,
932 static void kvm_handle_interrupt(CPUState *cpu, int mask)
934 cpu->interrupt_request |= mask;
936 if (!qemu_cpu_is_self(cpu)) {
937 qemu_cpu_kick(cpu);
941 int kvm_set_irq(KVMState *s, int irq, int level)
943 struct kvm_irq_level event;
944 int ret;
946 assert(kvm_async_interrupts_enabled());
948 event.level = level;
949 event.irq = irq;
950 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
951 if (ret < 0) {
952 perror("kvm_set_irq");
953 abort();
956 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
959 #ifdef KVM_CAP_IRQ_ROUTING
960 typedef struct KVMMSIRoute {
961 struct kvm_irq_routing_entry kroute;
962 QTAILQ_ENTRY(KVMMSIRoute) entry;
963 } KVMMSIRoute;
965 static void set_gsi(KVMState *s, unsigned int gsi)
967 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
970 static void clear_gsi(KVMState *s, unsigned int gsi)
972 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
975 void kvm_init_irq_routing(KVMState *s)
977 int gsi_count, i;
979 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
980 if (gsi_count > 0) {
981 unsigned int gsi_bits, i;
983 /* Round up so we can search ints using ffs */
984 gsi_bits = ALIGN(gsi_count, 32);
985 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
986 s->gsi_count = gsi_count;
988 /* Mark any over-allocated bits as already in use */
989 for (i = gsi_count; i < gsi_bits; i++) {
990 set_gsi(s, i);
994 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
995 s->nr_allocated_irq_routes = 0;
997 if (!s->direct_msi) {
998 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
999 QTAILQ_INIT(&s->msi_hashtab[i]);
1003 kvm_arch_init_irq_routing(s);
1006 void kvm_irqchip_commit_routes(KVMState *s)
1008 int ret;
1010 s->irq_routes->flags = 0;
1011 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1012 assert(ret == 0);
1015 static void kvm_add_routing_entry(KVMState *s,
1016 struct kvm_irq_routing_entry *entry)
1018 struct kvm_irq_routing_entry *new;
1019 int n, size;
1021 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1022 n = s->nr_allocated_irq_routes * 2;
1023 if (n < 64) {
1024 n = 64;
1026 size = sizeof(struct kvm_irq_routing);
1027 size += n * sizeof(*new);
1028 s->irq_routes = g_realloc(s->irq_routes, size);
1029 s->nr_allocated_irq_routes = n;
1031 n = s->irq_routes->nr++;
1032 new = &s->irq_routes->entries[n];
1034 *new = *entry;
1036 set_gsi(s, entry->gsi);
1039 static int kvm_update_routing_entry(KVMState *s,
1040 struct kvm_irq_routing_entry *new_entry)
1042 struct kvm_irq_routing_entry *entry;
1043 int n;
1045 for (n = 0; n < s->irq_routes->nr; n++) {
1046 entry = &s->irq_routes->entries[n];
1047 if (entry->gsi != new_entry->gsi) {
1048 continue;
1051 if(!memcmp(entry, new_entry, sizeof *entry)) {
1052 return 0;
1055 *entry = *new_entry;
1057 kvm_irqchip_commit_routes(s);
1059 return 0;
1062 return -ESRCH;
1065 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1067 struct kvm_irq_routing_entry e = {};
1069 assert(pin < s->gsi_count);
1071 e.gsi = irq;
1072 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1073 e.flags = 0;
1074 e.u.irqchip.irqchip = irqchip;
1075 e.u.irqchip.pin = pin;
1076 kvm_add_routing_entry(s, &e);
1079 void kvm_irqchip_release_virq(KVMState *s, int virq)
1081 struct kvm_irq_routing_entry *e;
1082 int i;
1084 if (kvm_gsi_direct_mapping()) {
1085 return;
1088 for (i = 0; i < s->irq_routes->nr; i++) {
1089 e = &s->irq_routes->entries[i];
1090 if (e->gsi == virq) {
1091 s->irq_routes->nr--;
1092 *e = s->irq_routes->entries[s->irq_routes->nr];
1095 clear_gsi(s, virq);
1098 static unsigned int kvm_hash_msi(uint32_t data)
1100 /* This is optimized for IA32 MSI layout. However, no other arch shall
1101 * repeat the mistake of not providing a direct MSI injection API. */
1102 return data & 0xff;
1105 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1107 KVMMSIRoute *route, *next;
1108 unsigned int hash;
1110 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1111 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1112 kvm_irqchip_release_virq(s, route->kroute.gsi);
1113 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1114 g_free(route);
1119 static int kvm_irqchip_get_virq(KVMState *s)
1121 uint32_t *word = s->used_gsi_bitmap;
1122 int max_words = ALIGN(s->gsi_count, 32) / 32;
1123 int i, bit;
1124 bool retry = true;
1126 again:
1127 /* Return the lowest unused GSI in the bitmap */
1128 for (i = 0; i < max_words; i++) {
1129 bit = ffs(~word[i]);
1130 if (!bit) {
1131 continue;
1134 return bit - 1 + i * 32;
1136 if (!s->direct_msi && retry) {
1137 retry = false;
1138 kvm_flush_dynamic_msi_routes(s);
1139 goto again;
1141 return -ENOSPC;
1145 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1147 unsigned int hash = kvm_hash_msi(msg.data);
1148 KVMMSIRoute *route;
1150 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1151 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1152 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1153 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1154 return route;
1157 return NULL;
1160 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1162 struct kvm_msi msi;
1163 KVMMSIRoute *route;
1165 if (s->direct_msi) {
1166 msi.address_lo = (uint32_t)msg.address;
1167 msi.address_hi = msg.address >> 32;
1168 msi.data = le32_to_cpu(msg.data);
1169 msi.flags = 0;
1170 memset(msi.pad, 0, sizeof(msi.pad));
1172 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1175 route = kvm_lookup_msi_route(s, msg);
1176 if (!route) {
1177 int virq;
1179 virq = kvm_irqchip_get_virq(s);
1180 if (virq < 0) {
1181 return virq;
1184 route = g_malloc0(sizeof(KVMMSIRoute));
1185 route->kroute.gsi = virq;
1186 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1187 route->kroute.flags = 0;
1188 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1189 route->kroute.u.msi.address_hi = msg.address >> 32;
1190 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1192 kvm_add_routing_entry(s, &route->kroute);
1193 kvm_irqchip_commit_routes(s);
1195 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1196 entry);
1199 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1201 return kvm_set_irq(s, route->kroute.gsi, 1);
1204 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1206 struct kvm_irq_routing_entry kroute = {};
1207 int virq;
1209 if (kvm_gsi_direct_mapping()) {
1210 return msg.data & 0xffff;
1213 if (!kvm_gsi_routing_enabled()) {
1214 return -ENOSYS;
1217 virq = kvm_irqchip_get_virq(s);
1218 if (virq < 0) {
1219 return virq;
1222 kroute.gsi = virq;
1223 kroute.type = KVM_IRQ_ROUTING_MSI;
1224 kroute.flags = 0;
1225 kroute.u.msi.address_lo = (uint32_t)msg.address;
1226 kroute.u.msi.address_hi = msg.address >> 32;
1227 kroute.u.msi.data = le32_to_cpu(msg.data);
1228 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1229 kvm_irqchip_release_virq(s, virq);
1230 return -EINVAL;
1233 kvm_add_routing_entry(s, &kroute);
1234 kvm_irqchip_commit_routes(s);
1236 return virq;
1239 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1241 struct kvm_irq_routing_entry kroute = {};
1243 if (kvm_gsi_direct_mapping()) {
1244 return 0;
1247 if (!kvm_irqchip_in_kernel()) {
1248 return -ENOSYS;
1251 kroute.gsi = virq;
1252 kroute.type = KVM_IRQ_ROUTING_MSI;
1253 kroute.flags = 0;
1254 kroute.u.msi.address_lo = (uint32_t)msg.address;
1255 kroute.u.msi.address_hi = msg.address >> 32;
1256 kroute.u.msi.data = le32_to_cpu(msg.data);
1257 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1258 return -EINVAL;
1261 return kvm_update_routing_entry(s, &kroute);
1264 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1265 bool assign)
1267 struct kvm_irqfd irqfd = {
1268 .fd = fd,
1269 .gsi = virq,
1270 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1273 if (rfd != -1) {
1274 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1275 irqfd.resamplefd = rfd;
1278 if (!kvm_irqfds_enabled()) {
1279 return -ENOSYS;
1282 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1285 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1287 struct kvm_irq_routing_entry kroute = {};
1288 int virq;
1290 if (!kvm_gsi_routing_enabled()) {
1291 return -ENOSYS;
1294 virq = kvm_irqchip_get_virq(s);
1295 if (virq < 0) {
1296 return virq;
1299 kroute.gsi = virq;
1300 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1301 kroute.flags = 0;
1302 kroute.u.adapter.summary_addr = adapter->summary_addr;
1303 kroute.u.adapter.ind_addr = adapter->ind_addr;
1304 kroute.u.adapter.summary_offset = adapter->summary_offset;
1305 kroute.u.adapter.ind_offset = adapter->ind_offset;
1306 kroute.u.adapter.adapter_id = adapter->adapter_id;
1308 kvm_add_routing_entry(s, &kroute);
1309 kvm_irqchip_commit_routes(s);
1311 return virq;
1314 #else /* !KVM_CAP_IRQ_ROUTING */
1316 void kvm_init_irq_routing(KVMState *s)
1320 void kvm_irqchip_release_virq(KVMState *s, int virq)
1324 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1326 abort();
1329 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1331 return -ENOSYS;
1334 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1336 return -ENOSYS;
1339 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1341 abort();
1344 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1346 return -ENOSYS;
1348 #endif /* !KVM_CAP_IRQ_ROUTING */
1350 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1351 EventNotifier *rn, int virq)
1353 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1354 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1357 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1359 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1360 false);
1363 static int kvm_irqchip_create(KVMState *s)
1365 int ret;
1367 if (!qemu_opt_get_bool(qemu_get_machine_opts(), "kernel_irqchip", true) ||
1368 (!kvm_check_extension(s, KVM_CAP_IRQCHIP) &&
1369 (kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0) < 0))) {
1370 return 0;
1373 /* First probe and see if there's a arch-specific hook to create the
1374 * in-kernel irqchip for us */
1375 ret = kvm_arch_irqchip_create(s);
1376 if (ret < 0) {
1377 return ret;
1378 } else if (ret == 0) {
1379 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1380 if (ret < 0) {
1381 fprintf(stderr, "Create kernel irqchip failed\n");
1382 return ret;
1386 kvm_kernel_irqchip = true;
1387 /* If we have an in-kernel IRQ chip then we must have asynchronous
1388 * interrupt delivery (though the reverse is not necessarily true)
1390 kvm_async_interrupts_allowed = true;
1391 kvm_halt_in_kernel_allowed = true;
1393 kvm_init_irq_routing(s);
1395 return 0;
1398 /* Find number of supported CPUs using the recommended
1399 * procedure from the kernel API documentation to cope with
1400 * older kernels that may be missing capabilities.
1402 static int kvm_recommended_vcpus(KVMState *s)
1404 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1405 return (ret) ? ret : 4;
1408 static int kvm_max_vcpus(KVMState *s)
1410 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1411 return (ret) ? ret : kvm_recommended_vcpus(s);
1414 static int kvm_init(MachineState *ms)
1416 MachineClass *mc = MACHINE_GET_CLASS(ms);
1417 static const char upgrade_note[] =
1418 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1419 "(see http://sourceforge.net/projects/kvm).\n";
1420 struct {
1421 const char *name;
1422 int num;
1423 } num_cpus[] = {
1424 { "SMP", smp_cpus },
1425 { "hotpluggable", max_cpus },
1426 { NULL, }
1427 }, *nc = num_cpus;
1428 int soft_vcpus_limit, hard_vcpus_limit;
1429 KVMState *s;
1430 const KVMCapabilityInfo *missing_cap;
1431 int ret;
1432 int i, type = 0;
1433 const char *kvm_type;
1435 s = KVM_STATE(ms->accelerator);
1438 * On systems where the kernel can support different base page
1439 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1440 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1441 * page size for the system though.
1443 assert(TARGET_PAGE_SIZE <= getpagesize());
1444 page_size_init();
1446 s->sigmask_len = 8;
1448 #ifdef KVM_CAP_SET_GUEST_DEBUG
1449 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1450 #endif
1451 s->vmfd = -1;
1452 s->fd = qemu_open("/dev/kvm", O_RDWR);
1453 if (s->fd == -1) {
1454 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1455 ret = -errno;
1456 goto err;
1459 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1460 if (ret < KVM_API_VERSION) {
1461 if (ret >= 0) {
1462 ret = -EINVAL;
1464 fprintf(stderr, "kvm version too old\n");
1465 goto err;
1468 if (ret > KVM_API_VERSION) {
1469 ret = -EINVAL;
1470 fprintf(stderr, "kvm version not supported\n");
1471 goto err;
1474 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1476 /* If unspecified, use the default value */
1477 if (!s->nr_slots) {
1478 s->nr_slots = 32;
1481 s->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1483 for (i = 0; i < s->nr_slots; i++) {
1484 s->slots[i].slot = i;
1487 /* check the vcpu limits */
1488 soft_vcpus_limit = kvm_recommended_vcpus(s);
1489 hard_vcpus_limit = kvm_max_vcpus(s);
1491 while (nc->name) {
1492 if (nc->num > soft_vcpus_limit) {
1493 fprintf(stderr,
1494 "Warning: Number of %s cpus requested (%d) exceeds "
1495 "the recommended cpus supported by KVM (%d)\n",
1496 nc->name, nc->num, soft_vcpus_limit);
1498 if (nc->num > hard_vcpus_limit) {
1499 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1500 "the maximum cpus supported by KVM (%d)\n",
1501 nc->name, nc->num, hard_vcpus_limit);
1502 exit(1);
1505 nc++;
1508 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1509 if (mc->kvm_type) {
1510 type = mc->kvm_type(kvm_type);
1511 } else if (kvm_type) {
1512 ret = -EINVAL;
1513 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1514 goto err;
1517 do {
1518 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1519 } while (ret == -EINTR);
1521 if (ret < 0) {
1522 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1523 strerror(-ret));
1525 #ifdef TARGET_S390X
1526 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1527 "your host kernel command line\n");
1528 #endif
1529 goto err;
1532 s->vmfd = ret;
1533 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1534 if (!missing_cap) {
1535 missing_cap =
1536 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1538 if (missing_cap) {
1539 ret = -EINVAL;
1540 fprintf(stderr, "kvm does not support %s\n%s",
1541 missing_cap->name, upgrade_note);
1542 goto err;
1545 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1547 s->broken_set_mem_region = 1;
1548 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1549 if (ret > 0) {
1550 s->broken_set_mem_region = 0;
1553 #ifdef KVM_CAP_VCPU_EVENTS
1554 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1555 #endif
1557 s->robust_singlestep =
1558 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1560 #ifdef KVM_CAP_DEBUGREGS
1561 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1562 #endif
1564 #ifdef KVM_CAP_XSAVE
1565 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1566 #endif
1568 #ifdef KVM_CAP_XCRS
1569 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1570 #endif
1572 #ifdef KVM_CAP_PIT_STATE2
1573 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1574 #endif
1576 #ifdef KVM_CAP_IRQ_ROUTING
1577 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1578 #endif
1580 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1582 s->irq_set_ioctl = KVM_IRQ_LINE;
1583 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1584 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1587 #ifdef KVM_CAP_READONLY_MEM
1588 kvm_readonly_mem_allowed =
1589 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1590 #endif
1592 kvm_eventfds_allowed =
1593 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1595 kvm_irqfds_allowed =
1596 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1598 kvm_resamplefds_allowed =
1599 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1601 ret = kvm_arch_init(s);
1602 if (ret < 0) {
1603 goto err;
1606 ret = kvm_irqchip_create(s);
1607 if (ret < 0) {
1608 goto err;
1611 kvm_state = s;
1612 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1613 memory_listener_register(&kvm_io_listener, &address_space_io);
1615 s->many_ioeventfds = kvm_check_many_ioeventfds();
1617 cpu_interrupt_handler = kvm_handle_interrupt;
1619 return 0;
1621 err:
1622 assert(ret < 0);
1623 if (s->vmfd >= 0) {
1624 close(s->vmfd);
1626 if (s->fd != -1) {
1627 close(s->fd);
1629 g_free(s->slots);
1631 return ret;
1634 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1636 s->sigmask_len = sigmask_len;
1639 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1640 uint32_t count)
1642 int i;
1643 uint8_t *ptr = data;
1645 for (i = 0; i < count; i++) {
1646 address_space_rw(&address_space_io, port, ptr, size,
1647 direction == KVM_EXIT_IO_OUT);
1648 ptr += size;
1652 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1654 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1655 run->internal.suberror);
1657 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1658 int i;
1660 for (i = 0; i < run->internal.ndata; ++i) {
1661 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1662 i, (uint64_t)run->internal.data[i]);
1665 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1666 fprintf(stderr, "emulation failure\n");
1667 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1668 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1669 return EXCP_INTERRUPT;
1672 /* FIXME: Should trigger a qmp message to let management know
1673 * something went wrong.
1675 return -1;
1678 void kvm_flush_coalesced_mmio_buffer(void)
1680 KVMState *s = kvm_state;
1682 if (s->coalesced_flush_in_progress) {
1683 return;
1686 s->coalesced_flush_in_progress = true;
1688 if (s->coalesced_mmio_ring) {
1689 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1690 while (ring->first != ring->last) {
1691 struct kvm_coalesced_mmio *ent;
1693 ent = &ring->coalesced_mmio[ring->first];
1695 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1696 smp_wmb();
1697 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1701 s->coalesced_flush_in_progress = false;
1704 static void do_kvm_cpu_synchronize_state(void *arg)
1706 CPUState *cpu = arg;
1708 if (!cpu->kvm_vcpu_dirty) {
1709 kvm_arch_get_registers(cpu);
1710 cpu->kvm_vcpu_dirty = true;
1714 void kvm_cpu_synchronize_state(CPUState *cpu)
1716 if (!cpu->kvm_vcpu_dirty) {
1717 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1721 static void do_kvm_cpu_synchronize_post_reset(void *arg)
1723 CPUState *cpu = arg;
1725 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1726 cpu->kvm_vcpu_dirty = false;
1729 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1731 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, cpu);
1734 static void do_kvm_cpu_synchronize_post_init(void *arg)
1736 CPUState *cpu = arg;
1738 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1739 cpu->kvm_vcpu_dirty = false;
1742 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1744 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, cpu);
1747 void kvm_cpu_clean_state(CPUState *cpu)
1749 cpu->kvm_vcpu_dirty = false;
1752 int kvm_cpu_exec(CPUState *cpu)
1754 struct kvm_run *run = cpu->kvm_run;
1755 int ret, run_ret;
1757 DPRINTF("kvm_cpu_exec()\n");
1759 if (kvm_arch_process_async_events(cpu)) {
1760 cpu->exit_request = 0;
1761 return EXCP_HLT;
1764 do {
1765 if (cpu->kvm_vcpu_dirty) {
1766 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1767 cpu->kvm_vcpu_dirty = false;
1770 kvm_arch_pre_run(cpu, run);
1771 if (cpu->exit_request) {
1772 DPRINTF("interrupt exit requested\n");
1774 * KVM requires us to reenter the kernel after IO exits to complete
1775 * instruction emulation. This self-signal will ensure that we
1776 * leave ASAP again.
1778 qemu_cpu_kick_self();
1780 qemu_mutex_unlock_iothread();
1782 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1784 qemu_mutex_lock_iothread();
1785 kvm_arch_post_run(cpu, run);
1787 if (run_ret < 0) {
1788 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1789 DPRINTF("io window exit\n");
1790 ret = EXCP_INTERRUPT;
1791 break;
1793 fprintf(stderr, "error: kvm run failed %s\n",
1794 strerror(-run_ret));
1795 ret = -1;
1796 break;
1799 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1800 switch (run->exit_reason) {
1801 case KVM_EXIT_IO:
1802 DPRINTF("handle_io\n");
1803 kvm_handle_io(run->io.port,
1804 (uint8_t *)run + run->io.data_offset,
1805 run->io.direction,
1806 run->io.size,
1807 run->io.count);
1808 ret = 0;
1809 break;
1810 case KVM_EXIT_MMIO:
1811 DPRINTF("handle_mmio\n");
1812 cpu_physical_memory_rw(run->mmio.phys_addr,
1813 run->mmio.data,
1814 run->mmio.len,
1815 run->mmio.is_write);
1816 ret = 0;
1817 break;
1818 case KVM_EXIT_IRQ_WINDOW_OPEN:
1819 DPRINTF("irq_window_open\n");
1820 ret = EXCP_INTERRUPT;
1821 break;
1822 case KVM_EXIT_SHUTDOWN:
1823 DPRINTF("shutdown\n");
1824 qemu_system_reset_request();
1825 ret = EXCP_INTERRUPT;
1826 break;
1827 case KVM_EXIT_UNKNOWN:
1828 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1829 (uint64_t)run->hw.hardware_exit_reason);
1830 ret = -1;
1831 break;
1832 case KVM_EXIT_INTERNAL_ERROR:
1833 ret = kvm_handle_internal_error(cpu, run);
1834 break;
1835 case KVM_EXIT_SYSTEM_EVENT:
1836 switch (run->system_event.type) {
1837 case KVM_SYSTEM_EVENT_SHUTDOWN:
1838 qemu_system_shutdown_request();
1839 ret = EXCP_INTERRUPT;
1840 break;
1841 case KVM_SYSTEM_EVENT_RESET:
1842 qemu_system_reset_request();
1843 ret = EXCP_INTERRUPT;
1844 break;
1845 default:
1846 DPRINTF("kvm_arch_handle_exit\n");
1847 ret = kvm_arch_handle_exit(cpu, run);
1848 break;
1850 break;
1851 default:
1852 DPRINTF("kvm_arch_handle_exit\n");
1853 ret = kvm_arch_handle_exit(cpu, run);
1854 break;
1856 } while (ret == 0);
1858 if (ret < 0) {
1859 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1860 vm_stop(RUN_STATE_INTERNAL_ERROR);
1863 cpu->exit_request = 0;
1864 return ret;
1867 int kvm_ioctl(KVMState *s, int type, ...)
1869 int ret;
1870 void *arg;
1871 va_list ap;
1873 va_start(ap, type);
1874 arg = va_arg(ap, void *);
1875 va_end(ap);
1877 trace_kvm_ioctl(type, arg);
1878 ret = ioctl(s->fd, type, arg);
1879 if (ret == -1) {
1880 ret = -errno;
1882 return ret;
1885 int kvm_vm_ioctl(KVMState *s, int type, ...)
1887 int ret;
1888 void *arg;
1889 va_list ap;
1891 va_start(ap, type);
1892 arg = va_arg(ap, void *);
1893 va_end(ap);
1895 trace_kvm_vm_ioctl(type, arg);
1896 ret = ioctl(s->vmfd, type, arg);
1897 if (ret == -1) {
1898 ret = -errno;
1900 return ret;
1903 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1905 int ret;
1906 void *arg;
1907 va_list ap;
1909 va_start(ap, type);
1910 arg = va_arg(ap, void *);
1911 va_end(ap);
1913 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1914 ret = ioctl(cpu->kvm_fd, type, arg);
1915 if (ret == -1) {
1916 ret = -errno;
1918 return ret;
1921 int kvm_device_ioctl(int fd, int type, ...)
1923 int ret;
1924 void *arg;
1925 va_list ap;
1927 va_start(ap, type);
1928 arg = va_arg(ap, void *);
1929 va_end(ap);
1931 trace_kvm_device_ioctl(fd, type, arg);
1932 ret = ioctl(fd, type, arg);
1933 if (ret == -1) {
1934 ret = -errno;
1936 return ret;
1939 int kvm_has_sync_mmu(void)
1941 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1944 int kvm_has_vcpu_events(void)
1946 return kvm_state->vcpu_events;
1949 int kvm_has_robust_singlestep(void)
1951 return kvm_state->robust_singlestep;
1954 int kvm_has_debugregs(void)
1956 return kvm_state->debugregs;
1959 int kvm_has_xsave(void)
1961 return kvm_state->xsave;
1964 int kvm_has_xcrs(void)
1966 return kvm_state->xcrs;
1969 int kvm_has_pit_state2(void)
1971 return kvm_state->pit_state2;
1974 int kvm_has_many_ioeventfds(void)
1976 if (!kvm_enabled()) {
1977 return 0;
1979 return kvm_state->many_ioeventfds;
1982 int kvm_has_gsi_routing(void)
1984 #ifdef KVM_CAP_IRQ_ROUTING
1985 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1986 #else
1987 return false;
1988 #endif
1991 int kvm_has_intx_set_mask(void)
1993 return kvm_state->intx_set_mask;
1996 void kvm_setup_guest_memory(void *start, size_t size)
1998 if (!kvm_has_sync_mmu()) {
1999 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
2001 if (ret) {
2002 perror("qemu_madvise");
2003 fprintf(stderr,
2004 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2005 exit(1);
2010 #ifdef KVM_CAP_SET_GUEST_DEBUG
2011 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2012 target_ulong pc)
2014 struct kvm_sw_breakpoint *bp;
2016 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2017 if (bp->pc == pc) {
2018 return bp;
2021 return NULL;
2024 int kvm_sw_breakpoints_active(CPUState *cpu)
2026 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2029 struct kvm_set_guest_debug_data {
2030 struct kvm_guest_debug dbg;
2031 CPUState *cpu;
2032 int err;
2035 static void kvm_invoke_set_guest_debug(void *data)
2037 struct kvm_set_guest_debug_data *dbg_data = data;
2039 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
2040 &dbg_data->dbg);
2043 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2045 struct kvm_set_guest_debug_data data;
2047 data.dbg.control = reinject_trap;
2049 if (cpu->singlestep_enabled) {
2050 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2052 kvm_arch_update_guest_debug(cpu, &data.dbg);
2053 data.cpu = cpu;
2055 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
2056 return data.err;
2059 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2060 target_ulong len, int type)
2062 struct kvm_sw_breakpoint *bp;
2063 int err;
2065 if (type == GDB_BREAKPOINT_SW) {
2066 bp = kvm_find_sw_breakpoint(cpu, addr);
2067 if (bp) {
2068 bp->use_count++;
2069 return 0;
2072 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2073 bp->pc = addr;
2074 bp->use_count = 1;
2075 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2076 if (err) {
2077 g_free(bp);
2078 return err;
2081 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2082 } else {
2083 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2084 if (err) {
2085 return err;
2089 CPU_FOREACH(cpu) {
2090 err = kvm_update_guest_debug(cpu, 0);
2091 if (err) {
2092 return err;
2095 return 0;
2098 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2099 target_ulong len, int type)
2101 struct kvm_sw_breakpoint *bp;
2102 int err;
2104 if (type == GDB_BREAKPOINT_SW) {
2105 bp = kvm_find_sw_breakpoint(cpu, addr);
2106 if (!bp) {
2107 return -ENOENT;
2110 if (bp->use_count > 1) {
2111 bp->use_count--;
2112 return 0;
2115 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2116 if (err) {
2117 return err;
2120 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2121 g_free(bp);
2122 } else {
2123 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2124 if (err) {
2125 return err;
2129 CPU_FOREACH(cpu) {
2130 err = kvm_update_guest_debug(cpu, 0);
2131 if (err) {
2132 return err;
2135 return 0;
2138 void kvm_remove_all_breakpoints(CPUState *cpu)
2140 struct kvm_sw_breakpoint *bp, *next;
2141 KVMState *s = cpu->kvm_state;
2142 CPUState *tmpcpu;
2144 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2145 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2146 /* Try harder to find a CPU that currently sees the breakpoint. */
2147 CPU_FOREACH(tmpcpu) {
2148 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2149 break;
2153 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2154 g_free(bp);
2156 kvm_arch_remove_all_hw_breakpoints();
2158 CPU_FOREACH(cpu) {
2159 kvm_update_guest_debug(cpu, 0);
2163 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2165 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2167 return -EINVAL;
2170 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2171 target_ulong len, int type)
2173 return -EINVAL;
2176 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2177 target_ulong len, int type)
2179 return -EINVAL;
2182 void kvm_remove_all_breakpoints(CPUState *cpu)
2185 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2187 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2189 KVMState *s = kvm_state;
2190 struct kvm_signal_mask *sigmask;
2191 int r;
2193 if (!sigset) {
2194 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2197 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2199 sigmask->len = s->sigmask_len;
2200 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2201 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2202 g_free(sigmask);
2204 return r;
2206 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2208 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2211 int kvm_on_sigbus(int code, void *addr)
2213 return kvm_arch_on_sigbus(code, addr);
2216 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2218 int ret;
2219 struct kvm_create_device create_dev;
2221 create_dev.type = type;
2222 create_dev.fd = -1;
2223 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2225 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2226 return -ENOTSUP;
2229 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2230 if (ret) {
2231 return ret;
2234 return test ? 0 : create_dev.fd;
2237 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2239 struct kvm_one_reg reg;
2240 int r;
2242 reg.id = id;
2243 reg.addr = (uintptr_t) source;
2244 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2245 if (r) {
2246 trace_kvm_failed_reg_set(id, strerror(r));
2248 return r;
2251 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2253 struct kvm_one_reg reg;
2254 int r;
2256 reg.id = id;
2257 reg.addr = (uintptr_t) target;
2258 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2259 if (r) {
2260 trace_kvm_failed_reg_get(id, strerror(r));
2262 return r;
2265 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2267 AccelClass *ac = ACCEL_CLASS(oc);
2268 ac->name = "KVM";
2269 ac->init_machine = kvm_init;
2270 ac->allowed = &kvm_allowed;
2273 static const TypeInfo kvm_accel_type = {
2274 .name = TYPE_KVM_ACCEL,
2275 .parent = TYPE_ACCEL,
2276 .class_init = kvm_accel_class_init,
2277 .instance_size = sizeof(KVMState),
2280 static void kvm_type_init(void)
2282 type_register_static(&kvm_accel_type);
2285 type_init(kvm_type_init);