usb: Fix usb-bt-dongle initialization.
[qemu-kvm.git] / kvm-all.c
blob3ae30eeb68d38634be4c6117201d6004b4a0871a
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 "hw/hw.h"
29 #include "hw/pci/msi.h"
30 #include "hw/s390x/adapter.h"
31 #include "exec/gdbstub.h"
32 #include "sysemu/kvm.h"
33 #include "qemu/bswap.h"
34 #include "exec/memory.h"
35 #include "exec/ram_addr.h"
36 #include "exec/address-spaces.h"
37 #include "qemu/event_notifier.h"
38 #include "trace.h"
40 #include "hw/boards.h"
42 /* This check must be after config-host.h is included */
43 #ifdef CONFIG_EVENTFD
44 #include <sys/eventfd.h>
45 #endif
47 #ifdef CONFIG_VALGRIND_H
48 #include <valgrind/memcheck.h>
49 #endif
51 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
52 #define PAGE_SIZE TARGET_PAGE_SIZE
54 //#define DEBUG_KVM
56 #ifdef DEBUG_KVM
57 #define DPRINTF(fmt, ...) \
58 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
59 #else
60 #define DPRINTF(fmt, ...) \
61 do { } while (0)
62 #endif
64 #define KVM_MSI_HASHTAB_SIZE 256
66 typedef struct KVMSlot
68 hwaddr start_addr;
69 ram_addr_t memory_size;
70 void *ram;
71 int slot;
72 int flags;
73 } KVMSlot;
75 typedef struct kvm_dirty_log KVMDirtyLog;
77 struct KVMState
79 KVMSlot *slots;
80 int nr_slots;
81 int fd;
82 int vmfd;
83 int coalesced_mmio;
84 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
85 bool coalesced_flush_in_progress;
86 int broken_set_mem_region;
87 int migration_log;
88 int vcpu_events;
89 int robust_singlestep;
90 int debugregs;
91 #ifdef KVM_CAP_SET_GUEST_DEBUG
92 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
93 #endif
94 int pit_state2;
95 int xsave, xcrs;
96 int many_ioeventfds;
97 int intx_set_mask;
98 /* The man page (and posix) say ioctl numbers are signed int, but
99 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
100 * unsigned, and treating them as signed here can break things */
101 unsigned irq_set_ioctl;
102 unsigned int sigmask_len;
103 #ifdef KVM_CAP_IRQ_ROUTING
104 struct kvm_irq_routing *irq_routes;
105 int nr_allocated_irq_routes;
106 uint32_t *used_gsi_bitmap;
107 unsigned int gsi_count;
108 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
109 bool direct_msi;
110 #endif
113 KVMState *kvm_state;
114 bool kvm_kernel_irqchip;
115 bool kvm_async_interrupts_allowed;
116 bool kvm_halt_in_kernel_allowed;
117 bool kvm_eventfds_allowed;
118 bool kvm_irqfds_allowed;
119 bool kvm_msi_via_irqfd_allowed;
120 bool kvm_gsi_routing_allowed;
121 bool kvm_gsi_direct_mapping;
122 bool kvm_allowed;
123 bool kvm_readonly_mem_allowed;
125 static const KVMCapabilityInfo kvm_required_capabilites[] = {
126 KVM_CAP_INFO(USER_MEMORY),
127 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
128 KVM_CAP_LAST_INFO
131 static KVMSlot *kvm_alloc_slot(KVMState *s)
133 int i;
135 for (i = 0; i < s->nr_slots; i++) {
136 if (s->slots[i].memory_size == 0) {
137 return &s->slots[i];
141 fprintf(stderr, "%s: no free slot available\n", __func__);
142 abort();
145 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
146 hwaddr start_addr,
147 hwaddr end_addr)
149 int i;
151 for (i = 0; i < s->nr_slots; i++) {
152 KVMSlot *mem = &s->slots[i];
154 if (start_addr == mem->start_addr &&
155 end_addr == mem->start_addr + mem->memory_size) {
156 return mem;
160 return NULL;
164 * Find overlapping slot with lowest start address
166 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
167 hwaddr start_addr,
168 hwaddr end_addr)
170 KVMSlot *found = NULL;
171 int i;
173 for (i = 0; i < s->nr_slots; i++) {
174 KVMSlot *mem = &s->slots[i];
176 if (mem->memory_size == 0 ||
177 (found && found->start_addr < mem->start_addr)) {
178 continue;
181 if (end_addr > mem->start_addr &&
182 start_addr < mem->start_addr + mem->memory_size) {
183 found = mem;
187 return found;
190 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
191 hwaddr *phys_addr)
193 int i;
195 for (i = 0; i < s->nr_slots; i++) {
196 KVMSlot *mem = &s->slots[i];
198 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
199 *phys_addr = mem->start_addr + (ram - mem->ram);
200 return 1;
204 return 0;
207 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
209 struct kvm_userspace_memory_region mem;
211 mem.slot = slot->slot;
212 mem.guest_phys_addr = slot->start_addr;
213 mem.userspace_addr = (unsigned long)slot->ram;
214 mem.flags = slot->flags;
215 if (s->migration_log) {
216 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
219 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
220 /* Set the slot size to 0 before setting the slot to the desired
221 * value. This is needed based on KVM commit 75d61fbc. */
222 mem.memory_size = 0;
223 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
225 mem.memory_size = slot->memory_size;
226 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
229 int kvm_init_vcpu(CPUState *cpu)
231 KVMState *s = kvm_state;
232 long mmap_size;
233 int ret;
235 DPRINTF("kvm_init_vcpu\n");
237 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
238 if (ret < 0) {
239 DPRINTF("kvm_create_vcpu failed\n");
240 goto err;
243 cpu->kvm_fd = ret;
244 cpu->kvm_state = s;
245 cpu->kvm_vcpu_dirty = true;
247 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
248 if (mmap_size < 0) {
249 ret = mmap_size;
250 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
251 goto err;
254 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
255 cpu->kvm_fd, 0);
256 if (cpu->kvm_run == MAP_FAILED) {
257 ret = -errno;
258 DPRINTF("mmap'ing vcpu state failed\n");
259 goto err;
262 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
263 s->coalesced_mmio_ring =
264 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
267 ret = kvm_arch_init_vcpu(cpu);
268 err:
269 return ret;
273 * dirty pages logging control
276 static int kvm_mem_flags(KVMState *s, bool log_dirty, bool readonly)
278 int flags = 0;
279 flags = log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
280 if (readonly && kvm_readonly_mem_allowed) {
281 flags |= KVM_MEM_READONLY;
283 return flags;
286 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
288 KVMState *s = kvm_state;
289 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
290 int old_flags;
292 old_flags = mem->flags;
294 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
295 mem->flags = flags;
297 /* If nothing changed effectively, no need to issue ioctl */
298 if (s->migration_log) {
299 flags |= KVM_MEM_LOG_DIRTY_PAGES;
302 if (flags == old_flags) {
303 return 0;
306 return kvm_set_user_memory_region(s, mem);
309 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
310 ram_addr_t size, bool log_dirty)
312 KVMState *s = kvm_state;
313 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
315 if (mem == NULL) {
316 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
317 TARGET_FMT_plx "\n", __func__, phys_addr,
318 (hwaddr)(phys_addr + size - 1));
319 return -EINVAL;
321 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
324 static void kvm_log_start(MemoryListener *listener,
325 MemoryRegionSection *section)
327 int r;
329 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
330 int128_get64(section->size), true);
331 if (r < 0) {
332 abort();
336 static void kvm_log_stop(MemoryListener *listener,
337 MemoryRegionSection *section)
339 int r;
341 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
342 int128_get64(section->size), false);
343 if (r < 0) {
344 abort();
348 static int kvm_set_migration_log(int enable)
350 KVMState *s = kvm_state;
351 KVMSlot *mem;
352 int i, err;
354 s->migration_log = enable;
356 for (i = 0; i < s->nr_slots; i++) {
357 mem = &s->slots[i];
359 if (!mem->memory_size) {
360 continue;
362 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
363 continue;
365 err = kvm_set_user_memory_region(s, mem);
366 if (err) {
367 return err;
370 return 0;
373 /* get kvm's dirty pages bitmap and update qemu's */
374 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
375 unsigned long *bitmap)
377 ram_addr_t start = section->offset_within_region + section->mr->ram_addr;
378 ram_addr_t pages = int128_get64(section->size) / getpagesize();
380 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
381 return 0;
384 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
387 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
388 * This function updates qemu's dirty bitmap using
389 * memory_region_set_dirty(). This means all bits are set
390 * to dirty.
392 * @start_add: start of logged region.
393 * @end_addr: end of logged region.
395 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
397 KVMState *s = kvm_state;
398 unsigned long size, allocated_size = 0;
399 KVMDirtyLog d;
400 KVMSlot *mem;
401 int ret = 0;
402 hwaddr start_addr = section->offset_within_address_space;
403 hwaddr end_addr = start_addr + int128_get64(section->size);
405 d.dirty_bitmap = NULL;
406 while (start_addr < end_addr) {
407 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
408 if (mem == NULL) {
409 break;
412 /* XXX bad kernel interface alert
413 * For dirty bitmap, kernel allocates array of size aligned to
414 * bits-per-long. But for case when the kernel is 64bits and
415 * the userspace is 32bits, userspace can't align to the same
416 * bits-per-long, since sizeof(long) is different between kernel
417 * and user space. This way, userspace will provide buffer which
418 * may be 4 bytes less than the kernel will use, resulting in
419 * userspace memory corruption (which is not detectable by valgrind
420 * too, in most cases).
421 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
422 * a hope that sizeof(long) wont become >8 any time soon.
424 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
425 /*HOST_LONG_BITS*/ 64) / 8;
426 if (!d.dirty_bitmap) {
427 d.dirty_bitmap = g_malloc(size);
428 } else if (size > allocated_size) {
429 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
431 allocated_size = size;
432 memset(d.dirty_bitmap, 0, allocated_size);
434 d.slot = mem->slot;
436 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
437 DPRINTF("ioctl failed %d\n", errno);
438 ret = -1;
439 break;
442 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
443 start_addr = mem->start_addr + mem->memory_size;
445 g_free(d.dirty_bitmap);
447 return ret;
450 static void kvm_coalesce_mmio_region(MemoryListener *listener,
451 MemoryRegionSection *secion,
452 hwaddr start, hwaddr size)
454 KVMState *s = kvm_state;
456 if (s->coalesced_mmio) {
457 struct kvm_coalesced_mmio_zone zone;
459 zone.addr = start;
460 zone.size = size;
461 zone.pad = 0;
463 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
467 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
468 MemoryRegionSection *secion,
469 hwaddr start, hwaddr size)
471 KVMState *s = kvm_state;
473 if (s->coalesced_mmio) {
474 struct kvm_coalesced_mmio_zone zone;
476 zone.addr = start;
477 zone.size = size;
478 zone.pad = 0;
480 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
484 int kvm_check_extension(KVMState *s, unsigned int extension)
486 int ret;
488 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
489 if (ret < 0) {
490 ret = 0;
493 return ret;
496 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
497 bool assign, uint32_t size, bool datamatch)
499 int ret;
500 struct kvm_ioeventfd iofd;
502 iofd.datamatch = datamatch ? val : 0;
503 iofd.addr = addr;
504 iofd.len = size;
505 iofd.flags = 0;
506 iofd.fd = fd;
508 if (!kvm_enabled()) {
509 return -ENOSYS;
512 if (datamatch) {
513 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
515 if (!assign) {
516 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
519 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
521 if (ret < 0) {
522 return -errno;
525 return 0;
528 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
529 bool assign, uint32_t size, bool datamatch)
531 struct kvm_ioeventfd kick = {
532 .datamatch = datamatch ? val : 0,
533 .addr = addr,
534 .flags = KVM_IOEVENTFD_FLAG_PIO,
535 .len = size,
536 .fd = fd,
538 int r;
539 if (!kvm_enabled()) {
540 return -ENOSYS;
542 if (datamatch) {
543 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
545 if (!assign) {
546 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
548 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
549 if (r < 0) {
550 return r;
552 return 0;
556 static int kvm_check_many_ioeventfds(void)
558 /* Userspace can use ioeventfd for io notification. This requires a host
559 * that supports eventfd(2) and an I/O thread; since eventfd does not
560 * support SIGIO it cannot interrupt the vcpu.
562 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
563 * can avoid creating too many ioeventfds.
565 #if defined(CONFIG_EVENTFD)
566 int ioeventfds[7];
567 int i, ret = 0;
568 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
569 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
570 if (ioeventfds[i] < 0) {
571 break;
573 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
574 if (ret < 0) {
575 close(ioeventfds[i]);
576 break;
580 /* Decide whether many devices are supported or not */
581 ret = i == ARRAY_SIZE(ioeventfds);
583 while (i-- > 0) {
584 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
585 close(ioeventfds[i]);
587 return ret;
588 #else
589 return 0;
590 #endif
593 static const KVMCapabilityInfo *
594 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
596 while (list->name) {
597 if (!kvm_check_extension(s, list->value)) {
598 return list;
600 list++;
602 return NULL;
605 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
607 KVMState *s = kvm_state;
608 KVMSlot *mem, old;
609 int err;
610 MemoryRegion *mr = section->mr;
611 bool log_dirty = memory_region_is_logging(mr);
612 bool writeable = !mr->readonly && !mr->rom_device;
613 bool readonly_flag = mr->readonly || memory_region_is_romd(mr);
614 hwaddr start_addr = section->offset_within_address_space;
615 ram_addr_t size = int128_get64(section->size);
616 void *ram = NULL;
617 unsigned delta;
619 /* kvm works in page size chunks, but the function may be called
620 with sub-page size and unaligned start address. */
621 delta = TARGET_PAGE_ALIGN(size) - size;
622 if (delta > size) {
623 return;
625 start_addr += delta;
626 size -= delta;
627 size &= TARGET_PAGE_MASK;
628 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
629 return;
632 if (!memory_region_is_ram(mr)) {
633 if (writeable || !kvm_readonly_mem_allowed) {
634 return;
635 } else if (!mr->romd_mode) {
636 /* If the memory device is not in romd_mode, then we actually want
637 * to remove the kvm memory slot so all accesses will trap. */
638 add = false;
642 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
644 while (1) {
645 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
646 if (!mem) {
647 break;
650 if (add && start_addr >= mem->start_addr &&
651 (start_addr + size <= mem->start_addr + mem->memory_size) &&
652 (ram - start_addr == mem->ram - mem->start_addr)) {
653 /* The new slot fits into the existing one and comes with
654 * identical parameters - update flags and done. */
655 kvm_slot_dirty_pages_log_change(mem, log_dirty);
656 return;
659 old = *mem;
661 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
662 kvm_physical_sync_dirty_bitmap(section);
665 /* unregister the overlapping slot */
666 mem->memory_size = 0;
667 err = kvm_set_user_memory_region(s, mem);
668 if (err) {
669 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
670 __func__, strerror(-err));
671 abort();
674 /* Workaround for older KVM versions: we can't join slots, even not by
675 * unregistering the previous ones and then registering the larger
676 * slot. We have to maintain the existing fragmentation. Sigh.
678 * This workaround assumes that the new slot starts at the same
679 * address as the first existing one. If not or if some overlapping
680 * slot comes around later, we will fail (not seen in practice so far)
681 * - and actually require a recent KVM version. */
682 if (s->broken_set_mem_region &&
683 old.start_addr == start_addr && old.memory_size < size && add) {
684 mem = kvm_alloc_slot(s);
685 mem->memory_size = old.memory_size;
686 mem->start_addr = old.start_addr;
687 mem->ram = old.ram;
688 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
690 err = kvm_set_user_memory_region(s, mem);
691 if (err) {
692 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
693 strerror(-err));
694 abort();
697 start_addr += old.memory_size;
698 ram += old.memory_size;
699 size -= old.memory_size;
700 continue;
703 /* register prefix slot */
704 if (old.start_addr < start_addr) {
705 mem = kvm_alloc_slot(s);
706 mem->memory_size = start_addr - old.start_addr;
707 mem->start_addr = old.start_addr;
708 mem->ram = old.ram;
709 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
711 err = kvm_set_user_memory_region(s, mem);
712 if (err) {
713 fprintf(stderr, "%s: error registering prefix slot: %s\n",
714 __func__, strerror(-err));
715 #ifdef TARGET_PPC
716 fprintf(stderr, "%s: This is probably because your kernel's " \
717 "PAGE_SIZE is too big. Please try to use 4k " \
718 "PAGE_SIZE!\n", __func__);
719 #endif
720 abort();
724 /* register suffix slot */
725 if (old.start_addr + old.memory_size > start_addr + size) {
726 ram_addr_t size_delta;
728 mem = kvm_alloc_slot(s);
729 mem->start_addr = start_addr + size;
730 size_delta = mem->start_addr - old.start_addr;
731 mem->memory_size = old.memory_size - size_delta;
732 mem->ram = old.ram + size_delta;
733 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
735 err = kvm_set_user_memory_region(s, mem);
736 if (err) {
737 fprintf(stderr, "%s: error registering suffix slot: %s\n",
738 __func__, strerror(-err));
739 abort();
744 /* in case the KVM bug workaround already "consumed" the new slot */
745 if (!size) {
746 return;
748 if (!add) {
749 return;
751 mem = kvm_alloc_slot(s);
752 mem->memory_size = size;
753 mem->start_addr = start_addr;
754 mem->ram = ram;
755 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
757 err = kvm_set_user_memory_region(s, mem);
758 if (err) {
759 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
760 strerror(-err));
761 abort();
765 static void kvm_region_add(MemoryListener *listener,
766 MemoryRegionSection *section)
768 memory_region_ref(section->mr);
769 kvm_set_phys_mem(section, true);
772 static void kvm_region_del(MemoryListener *listener,
773 MemoryRegionSection *section)
775 kvm_set_phys_mem(section, false);
776 memory_region_unref(section->mr);
779 static void kvm_log_sync(MemoryListener *listener,
780 MemoryRegionSection *section)
782 int r;
784 r = kvm_physical_sync_dirty_bitmap(section);
785 if (r < 0) {
786 abort();
790 static void kvm_log_global_start(struct MemoryListener *listener)
792 int r;
794 r = kvm_set_migration_log(1);
795 assert(r >= 0);
798 static void kvm_log_global_stop(struct MemoryListener *listener)
800 int r;
802 r = kvm_set_migration_log(0);
803 assert(r >= 0);
806 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
807 MemoryRegionSection *section,
808 bool match_data, uint64_t data,
809 EventNotifier *e)
811 int fd = event_notifier_get_fd(e);
812 int r;
814 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
815 data, true, int128_get64(section->size),
816 match_data);
817 if (r < 0) {
818 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
819 __func__, strerror(-r));
820 abort();
824 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
825 MemoryRegionSection *section,
826 bool match_data, uint64_t data,
827 EventNotifier *e)
829 int fd = event_notifier_get_fd(e);
830 int r;
832 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
833 data, false, int128_get64(section->size),
834 match_data);
835 if (r < 0) {
836 abort();
840 static void kvm_io_ioeventfd_add(MemoryListener *listener,
841 MemoryRegionSection *section,
842 bool match_data, uint64_t data,
843 EventNotifier *e)
845 int fd = event_notifier_get_fd(e);
846 int r;
848 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
849 data, true, int128_get64(section->size),
850 match_data);
851 if (r < 0) {
852 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
853 __func__, strerror(-r));
854 abort();
858 static void kvm_io_ioeventfd_del(MemoryListener *listener,
859 MemoryRegionSection *section,
860 bool match_data, uint64_t data,
861 EventNotifier *e)
864 int fd = event_notifier_get_fd(e);
865 int r;
867 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
868 data, false, int128_get64(section->size),
869 match_data);
870 if (r < 0) {
871 abort();
875 static MemoryListener kvm_memory_listener = {
876 .region_add = kvm_region_add,
877 .region_del = kvm_region_del,
878 .log_start = kvm_log_start,
879 .log_stop = kvm_log_stop,
880 .log_sync = kvm_log_sync,
881 .log_global_start = kvm_log_global_start,
882 .log_global_stop = kvm_log_global_stop,
883 .eventfd_add = kvm_mem_ioeventfd_add,
884 .eventfd_del = kvm_mem_ioeventfd_del,
885 .coalesced_mmio_add = kvm_coalesce_mmio_region,
886 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
887 .priority = 10,
890 static MemoryListener kvm_io_listener = {
891 .eventfd_add = kvm_io_ioeventfd_add,
892 .eventfd_del = kvm_io_ioeventfd_del,
893 .priority = 10,
896 static void kvm_handle_interrupt(CPUState *cpu, int mask)
898 cpu->interrupt_request |= mask;
900 if (!qemu_cpu_is_self(cpu)) {
901 qemu_cpu_kick(cpu);
905 int kvm_set_irq(KVMState *s, int irq, int level)
907 struct kvm_irq_level event;
908 int ret;
910 assert(kvm_async_interrupts_enabled());
912 event.level = level;
913 event.irq = irq;
914 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
915 if (ret < 0) {
916 perror("kvm_set_irq");
917 abort();
920 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
923 #ifdef KVM_CAP_IRQ_ROUTING
924 typedef struct KVMMSIRoute {
925 struct kvm_irq_routing_entry kroute;
926 QTAILQ_ENTRY(KVMMSIRoute) entry;
927 } KVMMSIRoute;
929 static void set_gsi(KVMState *s, unsigned int gsi)
931 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
934 static void clear_gsi(KVMState *s, unsigned int gsi)
936 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
939 void kvm_init_irq_routing(KVMState *s)
941 int gsi_count, i;
943 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
944 if (gsi_count > 0) {
945 unsigned int gsi_bits, i;
947 /* Round up so we can search ints using ffs */
948 gsi_bits = ALIGN(gsi_count, 32);
949 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
950 s->gsi_count = gsi_count;
952 /* Mark any over-allocated bits as already in use */
953 for (i = gsi_count; i < gsi_bits; i++) {
954 set_gsi(s, i);
958 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
959 s->nr_allocated_irq_routes = 0;
961 if (!s->direct_msi) {
962 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
963 QTAILQ_INIT(&s->msi_hashtab[i]);
967 kvm_arch_init_irq_routing(s);
970 void kvm_irqchip_commit_routes(KVMState *s)
972 int ret;
974 s->irq_routes->flags = 0;
975 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
976 assert(ret == 0);
979 static void kvm_add_routing_entry(KVMState *s,
980 struct kvm_irq_routing_entry *entry)
982 struct kvm_irq_routing_entry *new;
983 int n, size;
985 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
986 n = s->nr_allocated_irq_routes * 2;
987 if (n < 64) {
988 n = 64;
990 size = sizeof(struct kvm_irq_routing);
991 size += n * sizeof(*new);
992 s->irq_routes = g_realloc(s->irq_routes, size);
993 s->nr_allocated_irq_routes = n;
995 n = s->irq_routes->nr++;
996 new = &s->irq_routes->entries[n];
998 *new = *entry;
1000 set_gsi(s, entry->gsi);
1003 static int kvm_update_routing_entry(KVMState *s,
1004 struct kvm_irq_routing_entry *new_entry)
1006 struct kvm_irq_routing_entry *entry;
1007 int n;
1009 for (n = 0; n < s->irq_routes->nr; n++) {
1010 entry = &s->irq_routes->entries[n];
1011 if (entry->gsi != new_entry->gsi) {
1012 continue;
1015 if(!memcmp(entry, new_entry, sizeof *entry)) {
1016 return 0;
1019 *entry = *new_entry;
1021 kvm_irqchip_commit_routes(s);
1023 return 0;
1026 return -ESRCH;
1029 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1031 struct kvm_irq_routing_entry e = {};
1033 assert(pin < s->gsi_count);
1035 e.gsi = irq;
1036 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1037 e.flags = 0;
1038 e.u.irqchip.irqchip = irqchip;
1039 e.u.irqchip.pin = pin;
1040 kvm_add_routing_entry(s, &e);
1043 void kvm_irqchip_release_virq(KVMState *s, int virq)
1045 struct kvm_irq_routing_entry *e;
1046 int i;
1048 if (kvm_gsi_direct_mapping()) {
1049 return;
1052 for (i = 0; i < s->irq_routes->nr; i++) {
1053 e = &s->irq_routes->entries[i];
1054 if (e->gsi == virq) {
1055 s->irq_routes->nr--;
1056 *e = s->irq_routes->entries[s->irq_routes->nr];
1059 clear_gsi(s, virq);
1062 static unsigned int kvm_hash_msi(uint32_t data)
1064 /* This is optimized for IA32 MSI layout. However, no other arch shall
1065 * repeat the mistake of not providing a direct MSI injection API. */
1066 return data & 0xff;
1069 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1071 KVMMSIRoute *route, *next;
1072 unsigned int hash;
1074 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1075 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1076 kvm_irqchip_release_virq(s, route->kroute.gsi);
1077 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1078 g_free(route);
1083 static int kvm_irqchip_get_virq(KVMState *s)
1085 uint32_t *word = s->used_gsi_bitmap;
1086 int max_words = ALIGN(s->gsi_count, 32) / 32;
1087 int i, bit;
1088 bool retry = true;
1090 again:
1091 /* Return the lowest unused GSI in the bitmap */
1092 for (i = 0; i < max_words; i++) {
1093 bit = ffs(~word[i]);
1094 if (!bit) {
1095 continue;
1098 return bit - 1 + i * 32;
1100 if (!s->direct_msi && retry) {
1101 retry = false;
1102 kvm_flush_dynamic_msi_routes(s);
1103 goto again;
1105 return -ENOSPC;
1109 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1111 unsigned int hash = kvm_hash_msi(msg.data);
1112 KVMMSIRoute *route;
1114 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1115 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1116 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1117 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1118 return route;
1121 return NULL;
1124 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1126 struct kvm_msi msi;
1127 KVMMSIRoute *route;
1129 if (s->direct_msi) {
1130 msi.address_lo = (uint32_t)msg.address;
1131 msi.address_hi = msg.address >> 32;
1132 msi.data = le32_to_cpu(msg.data);
1133 msi.flags = 0;
1134 memset(msi.pad, 0, sizeof(msi.pad));
1136 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1139 route = kvm_lookup_msi_route(s, msg);
1140 if (!route) {
1141 int virq;
1143 virq = kvm_irqchip_get_virq(s);
1144 if (virq < 0) {
1145 return virq;
1148 route = g_malloc0(sizeof(KVMMSIRoute));
1149 route->kroute.gsi = virq;
1150 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1151 route->kroute.flags = 0;
1152 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);
1156 kvm_add_routing_entry(s, &route->kroute);
1157 kvm_irqchip_commit_routes(s);
1159 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1160 entry);
1163 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1165 return kvm_set_irq(s, route->kroute.gsi, 1);
1168 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1170 struct kvm_irq_routing_entry kroute = {};
1171 int virq;
1173 if (kvm_gsi_direct_mapping()) {
1174 return msg.data & 0xffff;
1177 if (!kvm_gsi_routing_enabled()) {
1178 return -ENOSYS;
1181 virq = kvm_irqchip_get_virq(s);
1182 if (virq < 0) {
1183 return virq;
1186 kroute.gsi = virq;
1187 kroute.type = KVM_IRQ_ROUTING_MSI;
1188 kroute.flags = 0;
1189 kroute.u.msi.address_lo = (uint32_t)msg.address;
1190 kroute.u.msi.address_hi = msg.address >> 32;
1191 kroute.u.msi.data = le32_to_cpu(msg.data);
1193 kvm_add_routing_entry(s, &kroute);
1194 kvm_irqchip_commit_routes(s);
1196 return virq;
1199 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1201 struct kvm_irq_routing_entry kroute = {};
1203 if (kvm_gsi_direct_mapping()) {
1204 return 0;
1207 if (!kvm_irqchip_in_kernel()) {
1208 return -ENOSYS;
1211 kroute.gsi = virq;
1212 kroute.type = KVM_IRQ_ROUTING_MSI;
1213 kroute.flags = 0;
1214 kroute.u.msi.address_lo = (uint32_t)msg.address;
1215 kroute.u.msi.address_hi = msg.address >> 32;
1216 kroute.u.msi.data = le32_to_cpu(msg.data);
1218 return kvm_update_routing_entry(s, &kroute);
1221 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1222 bool assign)
1224 struct kvm_irqfd irqfd = {
1225 .fd = fd,
1226 .gsi = virq,
1227 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1230 if (rfd != -1) {
1231 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1232 irqfd.resamplefd = rfd;
1235 if (!kvm_irqfds_enabled()) {
1236 return -ENOSYS;
1239 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1242 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1244 struct kvm_irq_routing_entry kroute;
1245 int virq;
1247 if (!kvm_gsi_routing_enabled()) {
1248 return -ENOSYS;
1251 virq = kvm_irqchip_get_virq(s);
1252 if (virq < 0) {
1253 return virq;
1256 kroute.gsi = virq;
1257 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1258 kroute.flags = 0;
1259 kroute.u.adapter.summary_addr = adapter->summary_addr;
1260 kroute.u.adapter.ind_addr = adapter->ind_addr;
1261 kroute.u.adapter.summary_offset = adapter->summary_offset;
1262 kroute.u.adapter.ind_offset = adapter->ind_offset;
1263 kroute.u.adapter.adapter_id = adapter->adapter_id;
1265 kvm_add_routing_entry(s, &kroute);
1266 kvm_irqchip_commit_routes(s);
1268 return virq;
1271 #else /* !KVM_CAP_IRQ_ROUTING */
1273 void kvm_init_irq_routing(KVMState *s)
1277 void kvm_irqchip_release_virq(KVMState *s, int virq)
1281 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1283 abort();
1286 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1288 return -ENOSYS;
1291 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1293 return -ENOSYS;
1296 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1298 abort();
1301 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1303 return -ENOSYS;
1305 #endif /* !KVM_CAP_IRQ_ROUTING */
1307 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1308 EventNotifier *rn, int virq)
1310 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1311 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1314 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1316 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1317 false);
1320 static int kvm_irqchip_create(KVMState *s)
1322 int ret;
1324 if (!qemu_opt_get_bool(qemu_get_machine_opts(), "kernel_irqchip", true) ||
1325 (!kvm_check_extension(s, KVM_CAP_IRQCHIP) &&
1326 (kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0) < 0))) {
1327 return 0;
1330 /* First probe and see if there's a arch-specific hook to create the
1331 * in-kernel irqchip for us */
1332 ret = kvm_arch_irqchip_create(s);
1333 if (ret < 0) {
1334 return ret;
1335 } else if (ret == 0) {
1336 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1337 if (ret < 0) {
1338 fprintf(stderr, "Create kernel irqchip failed\n");
1339 return ret;
1343 kvm_kernel_irqchip = true;
1344 /* If we have an in-kernel IRQ chip then we must have asynchronous
1345 * interrupt delivery (though the reverse is not necessarily true)
1347 kvm_async_interrupts_allowed = true;
1348 kvm_halt_in_kernel_allowed = true;
1350 kvm_init_irq_routing(s);
1352 return 0;
1355 /* Find number of supported CPUs using the recommended
1356 * procedure from the kernel API documentation to cope with
1357 * older kernels that may be missing capabilities.
1359 static int kvm_recommended_vcpus(KVMState *s)
1361 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1362 return (ret) ? ret : 4;
1365 static int kvm_max_vcpus(KVMState *s)
1367 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1368 return (ret) ? ret : kvm_recommended_vcpus(s);
1371 int kvm_init(MachineClass *mc)
1373 static const char upgrade_note[] =
1374 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1375 "(see http://sourceforge.net/projects/kvm).\n";
1376 struct {
1377 const char *name;
1378 int num;
1379 } num_cpus[] = {
1380 { "SMP", smp_cpus },
1381 { "hotpluggable", max_cpus },
1382 { NULL, }
1383 }, *nc = num_cpus;
1384 int soft_vcpus_limit, hard_vcpus_limit;
1385 KVMState *s;
1386 const KVMCapabilityInfo *missing_cap;
1387 int ret;
1388 int i, type = 0;
1389 const char *kvm_type;
1391 s = g_malloc0(sizeof(KVMState));
1394 * On systems where the kernel can support different base page
1395 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1396 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1397 * page size for the system though.
1399 assert(TARGET_PAGE_SIZE <= getpagesize());
1400 page_size_init();
1402 s->sigmask_len = 8;
1404 #ifdef KVM_CAP_SET_GUEST_DEBUG
1405 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1406 #endif
1407 s->vmfd = -1;
1408 s->fd = qemu_open("/dev/kvm", O_RDWR);
1409 if (s->fd == -1) {
1410 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1411 ret = -errno;
1412 goto err;
1415 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1416 if (ret < KVM_API_VERSION) {
1417 if (ret >= 0) {
1418 ret = -EINVAL;
1420 fprintf(stderr, "kvm version too old\n");
1421 goto err;
1424 if (ret > KVM_API_VERSION) {
1425 ret = -EINVAL;
1426 fprintf(stderr, "kvm version not supported\n");
1427 goto err;
1430 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1432 /* If unspecified, use the default value */
1433 if (!s->nr_slots) {
1434 s->nr_slots = 32;
1437 s->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1439 for (i = 0; i < s->nr_slots; i++) {
1440 s->slots[i].slot = i;
1443 /* check the vcpu limits */
1444 soft_vcpus_limit = kvm_recommended_vcpus(s);
1445 hard_vcpus_limit = kvm_max_vcpus(s);
1447 while (nc->name) {
1448 if (nc->num > soft_vcpus_limit) {
1449 fprintf(stderr,
1450 "Warning: Number of %s cpus requested (%d) exceeds "
1451 "the recommended cpus supported by KVM (%d)\n",
1452 nc->name, nc->num, soft_vcpus_limit);
1454 if (nc->num > hard_vcpus_limit) {
1455 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1456 "the maximum cpus supported by KVM (%d)\n",
1457 nc->name, nc->num, hard_vcpus_limit);
1458 exit(1);
1461 nc++;
1464 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1465 if (mc->kvm_type) {
1466 type = mc->kvm_type(kvm_type);
1467 } else if (kvm_type) {
1468 ret = -EINVAL;
1469 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1470 goto err;
1473 do {
1474 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1475 } while (ret == -EINTR);
1477 if (ret < 0) {
1478 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1479 strerror(-ret));
1481 #ifdef TARGET_S390X
1482 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1483 "your host kernel command line\n");
1484 #endif
1485 goto err;
1488 s->vmfd = ret;
1489 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1490 if (!missing_cap) {
1491 missing_cap =
1492 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1494 if (missing_cap) {
1495 ret = -EINVAL;
1496 fprintf(stderr, "kvm does not support %s\n%s",
1497 missing_cap->name, upgrade_note);
1498 goto err;
1501 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1503 s->broken_set_mem_region = 1;
1504 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1505 if (ret > 0) {
1506 s->broken_set_mem_region = 0;
1509 #ifdef KVM_CAP_VCPU_EVENTS
1510 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1511 #endif
1513 s->robust_singlestep =
1514 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1516 #ifdef KVM_CAP_DEBUGREGS
1517 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1518 #endif
1520 #ifdef KVM_CAP_XSAVE
1521 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1522 #endif
1524 #ifdef KVM_CAP_XCRS
1525 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1526 #endif
1528 #ifdef KVM_CAP_PIT_STATE2
1529 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1530 #endif
1532 #ifdef KVM_CAP_IRQ_ROUTING
1533 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1534 #endif
1536 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1538 s->irq_set_ioctl = KVM_IRQ_LINE;
1539 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1540 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1543 #ifdef KVM_CAP_READONLY_MEM
1544 kvm_readonly_mem_allowed =
1545 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1546 #endif
1548 kvm_eventfds_allowed =
1549 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1551 ret = kvm_arch_init(s);
1552 if (ret < 0) {
1553 goto err;
1556 ret = kvm_irqchip_create(s);
1557 if (ret < 0) {
1558 goto err;
1561 kvm_state = s;
1562 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1563 memory_listener_register(&kvm_io_listener, &address_space_io);
1565 s->many_ioeventfds = kvm_check_many_ioeventfds();
1567 cpu_interrupt_handler = kvm_handle_interrupt;
1569 return 0;
1571 err:
1572 assert(ret < 0);
1573 if (s->vmfd >= 0) {
1574 close(s->vmfd);
1576 if (s->fd != -1) {
1577 close(s->fd);
1579 g_free(s->slots);
1580 g_free(s);
1582 return ret;
1585 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1587 s->sigmask_len = sigmask_len;
1590 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1591 uint32_t count)
1593 int i;
1594 uint8_t *ptr = data;
1596 for (i = 0; i < count; i++) {
1597 address_space_rw(&address_space_io, port, ptr, size,
1598 direction == KVM_EXIT_IO_OUT);
1599 ptr += size;
1603 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1605 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1606 run->internal.suberror);
1608 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1609 int i;
1611 for (i = 0; i < run->internal.ndata; ++i) {
1612 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1613 i, (uint64_t)run->internal.data[i]);
1616 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1617 fprintf(stderr, "emulation failure\n");
1618 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1619 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1620 return EXCP_INTERRUPT;
1623 /* FIXME: Should trigger a qmp message to let management know
1624 * something went wrong.
1626 return -1;
1629 void kvm_flush_coalesced_mmio_buffer(void)
1631 KVMState *s = kvm_state;
1633 if (s->coalesced_flush_in_progress) {
1634 return;
1637 s->coalesced_flush_in_progress = true;
1639 if (s->coalesced_mmio_ring) {
1640 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1641 while (ring->first != ring->last) {
1642 struct kvm_coalesced_mmio *ent;
1644 ent = &ring->coalesced_mmio[ring->first];
1646 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1647 smp_wmb();
1648 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1652 s->coalesced_flush_in_progress = false;
1655 static void do_kvm_cpu_synchronize_state(void *arg)
1657 CPUState *cpu = arg;
1659 if (!cpu->kvm_vcpu_dirty) {
1660 kvm_arch_get_registers(cpu);
1661 cpu->kvm_vcpu_dirty = true;
1665 void kvm_cpu_synchronize_state(CPUState *cpu)
1667 if (!cpu->kvm_vcpu_dirty) {
1668 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1672 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1674 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1675 cpu->kvm_vcpu_dirty = false;
1678 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1680 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1681 cpu->kvm_vcpu_dirty = false;
1684 int kvm_cpu_exec(CPUState *cpu)
1686 struct kvm_run *run = cpu->kvm_run;
1687 int ret, run_ret;
1689 DPRINTF("kvm_cpu_exec()\n");
1691 if (kvm_arch_process_async_events(cpu)) {
1692 cpu->exit_request = 0;
1693 return EXCP_HLT;
1696 do {
1697 if (cpu->kvm_vcpu_dirty) {
1698 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1699 cpu->kvm_vcpu_dirty = false;
1702 kvm_arch_pre_run(cpu, run);
1703 if (cpu->exit_request) {
1704 DPRINTF("interrupt exit requested\n");
1706 * KVM requires us to reenter the kernel after IO exits to complete
1707 * instruction emulation. This self-signal will ensure that we
1708 * leave ASAP again.
1710 qemu_cpu_kick_self();
1712 qemu_mutex_unlock_iothread();
1714 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1716 qemu_mutex_lock_iothread();
1717 kvm_arch_post_run(cpu, run);
1719 if (run_ret < 0) {
1720 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1721 DPRINTF("io window exit\n");
1722 ret = EXCP_INTERRUPT;
1723 break;
1725 fprintf(stderr, "error: kvm run failed %s\n",
1726 strerror(-run_ret));
1727 abort();
1730 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1731 switch (run->exit_reason) {
1732 case KVM_EXIT_IO:
1733 DPRINTF("handle_io\n");
1734 kvm_handle_io(run->io.port,
1735 (uint8_t *)run + run->io.data_offset,
1736 run->io.direction,
1737 run->io.size,
1738 run->io.count);
1739 ret = 0;
1740 break;
1741 case KVM_EXIT_MMIO:
1742 DPRINTF("handle_mmio\n");
1743 cpu_physical_memory_rw(run->mmio.phys_addr,
1744 run->mmio.data,
1745 run->mmio.len,
1746 run->mmio.is_write);
1747 ret = 0;
1748 break;
1749 case KVM_EXIT_IRQ_WINDOW_OPEN:
1750 DPRINTF("irq_window_open\n");
1751 ret = EXCP_INTERRUPT;
1752 break;
1753 case KVM_EXIT_SHUTDOWN:
1754 DPRINTF("shutdown\n");
1755 qemu_system_reset_request();
1756 ret = EXCP_INTERRUPT;
1757 break;
1758 case KVM_EXIT_UNKNOWN:
1759 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1760 (uint64_t)run->hw.hardware_exit_reason);
1761 ret = -1;
1762 break;
1763 case KVM_EXIT_INTERNAL_ERROR:
1764 ret = kvm_handle_internal_error(cpu, run);
1765 break;
1766 case KVM_EXIT_SYSTEM_EVENT:
1767 switch (run->system_event.type) {
1768 case KVM_SYSTEM_EVENT_SHUTDOWN:
1769 qemu_system_shutdown_request();
1770 ret = EXCP_INTERRUPT;
1771 break;
1772 case KVM_SYSTEM_EVENT_RESET:
1773 qemu_system_reset_request();
1774 ret = EXCP_INTERRUPT;
1775 break;
1776 default:
1777 DPRINTF("kvm_arch_handle_exit\n");
1778 ret = kvm_arch_handle_exit(cpu, run);
1779 break;
1781 break;
1782 default:
1783 DPRINTF("kvm_arch_handle_exit\n");
1784 ret = kvm_arch_handle_exit(cpu, run);
1785 break;
1787 } while (ret == 0);
1789 if (ret < 0) {
1790 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1791 vm_stop(RUN_STATE_INTERNAL_ERROR);
1794 cpu->exit_request = 0;
1795 return ret;
1798 int kvm_ioctl(KVMState *s, int type, ...)
1800 int ret;
1801 void *arg;
1802 va_list ap;
1804 va_start(ap, type);
1805 arg = va_arg(ap, void *);
1806 va_end(ap);
1808 trace_kvm_ioctl(type, arg);
1809 ret = ioctl(s->fd, type, arg);
1810 if (ret == -1) {
1811 ret = -errno;
1813 return ret;
1816 int kvm_vm_ioctl(KVMState *s, int type, ...)
1818 int ret;
1819 void *arg;
1820 va_list ap;
1822 va_start(ap, type);
1823 arg = va_arg(ap, void *);
1824 va_end(ap);
1826 trace_kvm_vm_ioctl(type, arg);
1827 ret = ioctl(s->vmfd, type, arg);
1828 if (ret == -1) {
1829 ret = -errno;
1831 return ret;
1834 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1836 int ret;
1837 void *arg;
1838 va_list ap;
1840 va_start(ap, type);
1841 arg = va_arg(ap, void *);
1842 va_end(ap);
1844 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1845 ret = ioctl(cpu->kvm_fd, type, arg);
1846 if (ret == -1) {
1847 ret = -errno;
1849 return ret;
1852 int kvm_device_ioctl(int fd, int type, ...)
1854 int ret;
1855 void *arg;
1856 va_list ap;
1858 va_start(ap, type);
1859 arg = va_arg(ap, void *);
1860 va_end(ap);
1862 trace_kvm_device_ioctl(fd, type, arg);
1863 ret = ioctl(fd, type, arg);
1864 if (ret == -1) {
1865 ret = -errno;
1867 return ret;
1870 int kvm_has_sync_mmu(void)
1872 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1875 int kvm_has_vcpu_events(void)
1877 return kvm_state->vcpu_events;
1880 int kvm_has_robust_singlestep(void)
1882 return kvm_state->robust_singlestep;
1885 int kvm_has_debugregs(void)
1887 return kvm_state->debugregs;
1890 int kvm_has_xsave(void)
1892 return kvm_state->xsave;
1895 int kvm_has_xcrs(void)
1897 return kvm_state->xcrs;
1900 int kvm_has_pit_state2(void)
1902 return kvm_state->pit_state2;
1905 int kvm_has_many_ioeventfds(void)
1907 if (!kvm_enabled()) {
1908 return 0;
1910 return kvm_state->many_ioeventfds;
1913 int kvm_has_gsi_routing(void)
1915 #ifdef KVM_CAP_IRQ_ROUTING
1916 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1917 #else
1918 return false;
1919 #endif
1922 int kvm_has_intx_set_mask(void)
1924 return kvm_state->intx_set_mask;
1927 void kvm_setup_guest_memory(void *start, size_t size)
1929 #ifdef CONFIG_VALGRIND_H
1930 VALGRIND_MAKE_MEM_DEFINED(start, size);
1931 #endif
1932 if (!kvm_has_sync_mmu()) {
1933 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1935 if (ret) {
1936 perror("qemu_madvise");
1937 fprintf(stderr,
1938 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1939 exit(1);
1944 #ifdef KVM_CAP_SET_GUEST_DEBUG
1945 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
1946 target_ulong pc)
1948 struct kvm_sw_breakpoint *bp;
1950 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
1951 if (bp->pc == pc) {
1952 return bp;
1955 return NULL;
1958 int kvm_sw_breakpoints_active(CPUState *cpu)
1960 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
1963 struct kvm_set_guest_debug_data {
1964 struct kvm_guest_debug dbg;
1965 CPUState *cpu;
1966 int err;
1969 static void kvm_invoke_set_guest_debug(void *data)
1971 struct kvm_set_guest_debug_data *dbg_data = data;
1973 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
1974 &dbg_data->dbg);
1977 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
1979 struct kvm_set_guest_debug_data data;
1981 data.dbg.control = reinject_trap;
1983 if (cpu->singlestep_enabled) {
1984 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1986 kvm_arch_update_guest_debug(cpu, &data.dbg);
1987 data.cpu = cpu;
1989 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
1990 return data.err;
1993 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
1994 target_ulong len, int type)
1996 struct kvm_sw_breakpoint *bp;
1997 int err;
1999 if (type == GDB_BREAKPOINT_SW) {
2000 bp = kvm_find_sw_breakpoint(cpu, addr);
2001 if (bp) {
2002 bp->use_count++;
2003 return 0;
2006 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2007 if (!bp) {
2008 return -ENOMEM;
2011 bp->pc = addr;
2012 bp->use_count = 1;
2013 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2014 if (err) {
2015 g_free(bp);
2016 return err;
2019 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2020 } else {
2021 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2022 if (err) {
2023 return err;
2027 CPU_FOREACH(cpu) {
2028 err = kvm_update_guest_debug(cpu, 0);
2029 if (err) {
2030 return err;
2033 return 0;
2036 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2037 target_ulong len, int type)
2039 struct kvm_sw_breakpoint *bp;
2040 int err;
2042 if (type == GDB_BREAKPOINT_SW) {
2043 bp = kvm_find_sw_breakpoint(cpu, addr);
2044 if (!bp) {
2045 return -ENOENT;
2048 if (bp->use_count > 1) {
2049 bp->use_count--;
2050 return 0;
2053 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2054 if (err) {
2055 return err;
2058 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2059 g_free(bp);
2060 } else {
2061 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2062 if (err) {
2063 return err;
2067 CPU_FOREACH(cpu) {
2068 err = kvm_update_guest_debug(cpu, 0);
2069 if (err) {
2070 return err;
2073 return 0;
2076 void kvm_remove_all_breakpoints(CPUState *cpu)
2078 struct kvm_sw_breakpoint *bp, *next;
2079 KVMState *s = cpu->kvm_state;
2081 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2082 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2083 /* Try harder to find a CPU that currently sees the breakpoint. */
2084 CPU_FOREACH(cpu) {
2085 if (kvm_arch_remove_sw_breakpoint(cpu, bp) == 0) {
2086 break;
2090 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2091 g_free(bp);
2093 kvm_arch_remove_all_hw_breakpoints();
2095 CPU_FOREACH(cpu) {
2096 kvm_update_guest_debug(cpu, 0);
2100 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2102 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2104 return -EINVAL;
2107 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2108 target_ulong len, int type)
2110 return -EINVAL;
2113 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2114 target_ulong len, int type)
2116 return -EINVAL;
2119 void kvm_remove_all_breakpoints(CPUState *cpu)
2122 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2124 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2126 KVMState *s = kvm_state;
2127 struct kvm_signal_mask *sigmask;
2128 int r;
2130 if (!sigset) {
2131 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2134 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2136 sigmask->len = s->sigmask_len;
2137 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2138 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2139 g_free(sigmask);
2141 return r;
2143 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2145 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2148 int kvm_on_sigbus(int code, void *addr)
2150 return kvm_arch_on_sigbus(code, addr);
2153 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2155 int ret;
2156 struct kvm_create_device create_dev;
2158 create_dev.type = type;
2159 create_dev.fd = -1;
2160 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2162 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2163 return -ENOTSUP;
2166 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2167 if (ret) {
2168 return ret;
2171 return test ? 0 : create_dev.fd;
2174 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2176 struct kvm_one_reg reg;
2177 int r;
2179 reg.id = id;
2180 reg.addr = (uintptr_t) source;
2181 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2182 if (r) {
2183 trace_kvm_failed_reg_set(id, strerror(r));
2185 return r;
2188 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2190 struct kvm_one_reg reg;
2191 int r;
2193 reg.id = id;
2194 reg.addr = (uintptr_t) target;
2195 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2196 if (r) {
2197 trace_kvm_failed_reg_get(id, strerror(r));
2199 return r;