hw/spitz: Don't prematurely explode QEMUMachineInitArgs
[qemu/pbrook.git] / kvm-all.c
blob961e1dbb710cb1e9934a588e6b9d87b8529446d6
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-barrier.h"
25 #include "qemu-option.h"
26 #include "qemu-config.h"
27 #include "sysemu.h"
28 #include "hw/hw.h"
29 #include "hw/msi.h"
30 #include "gdbstub.h"
31 #include "kvm.h"
32 #include "bswap.h"
33 #include "memory.h"
34 #include "exec-memory.h"
35 #include "event_notifier.h"
37 /* This check must be after config-host.h is included */
38 #ifdef CONFIG_EVENTFD
39 #include <sys/eventfd.h>
40 #endif
42 #ifdef CONFIG_VALGRIND_H
43 #include <valgrind/memcheck.h>
44 #endif
46 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
47 #define PAGE_SIZE TARGET_PAGE_SIZE
49 //#define DEBUG_KVM
51 #ifdef DEBUG_KVM
52 #define DPRINTF(fmt, ...) \
53 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
54 #else
55 #define DPRINTF(fmt, ...) \
56 do { } while (0)
57 #endif
59 #define KVM_MSI_HASHTAB_SIZE 256
61 typedef struct KVMSlot
63 hwaddr start_addr;
64 ram_addr_t memory_size;
65 void *ram;
66 int slot;
67 int flags;
68 } KVMSlot;
70 typedef struct kvm_dirty_log KVMDirtyLog;
72 struct KVMState
74 KVMSlot slots[32];
75 int fd;
76 int vmfd;
77 int coalesced_mmio;
78 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
79 bool coalesced_flush_in_progress;
80 int broken_set_mem_region;
81 int migration_log;
82 int vcpu_events;
83 int robust_singlestep;
84 int debugregs;
85 #ifdef KVM_CAP_SET_GUEST_DEBUG
86 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
87 #endif
88 int pit_state2;
89 int xsave, xcrs;
90 int many_ioeventfds;
91 int intx_set_mask;
92 /* The man page (and posix) say ioctl numbers are signed int, but
93 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
94 * unsigned, and treating them as signed here can break things */
95 unsigned irq_set_ioctl;
96 #ifdef KVM_CAP_IRQ_ROUTING
97 struct kvm_irq_routing *irq_routes;
98 int nr_allocated_irq_routes;
99 uint32_t *used_gsi_bitmap;
100 unsigned int gsi_count;
101 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
102 bool direct_msi;
103 #endif
106 KVMState *kvm_state;
107 bool kvm_kernel_irqchip;
108 bool kvm_async_interrupts_allowed;
109 bool kvm_irqfds_allowed;
110 bool kvm_msi_via_irqfd_allowed;
111 bool kvm_gsi_routing_allowed;
113 static const KVMCapabilityInfo kvm_required_capabilites[] = {
114 KVM_CAP_INFO(USER_MEMORY),
115 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
116 KVM_CAP_LAST_INFO
119 static KVMSlot *kvm_alloc_slot(KVMState *s)
121 int i;
123 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
124 if (s->slots[i].memory_size == 0) {
125 return &s->slots[i];
129 fprintf(stderr, "%s: no free slot available\n", __func__);
130 abort();
133 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
134 hwaddr start_addr,
135 hwaddr end_addr)
137 int i;
139 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
140 KVMSlot *mem = &s->slots[i];
142 if (start_addr == mem->start_addr &&
143 end_addr == mem->start_addr + mem->memory_size) {
144 return mem;
148 return NULL;
152 * Find overlapping slot with lowest start address
154 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
155 hwaddr start_addr,
156 hwaddr end_addr)
158 KVMSlot *found = NULL;
159 int i;
161 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
162 KVMSlot *mem = &s->slots[i];
164 if (mem->memory_size == 0 ||
165 (found && found->start_addr < mem->start_addr)) {
166 continue;
169 if (end_addr > mem->start_addr &&
170 start_addr < mem->start_addr + mem->memory_size) {
171 found = mem;
175 return found;
178 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
179 hwaddr *phys_addr)
181 int i;
183 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
184 KVMSlot *mem = &s->slots[i];
186 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
187 *phys_addr = mem->start_addr + (ram - mem->ram);
188 return 1;
192 return 0;
195 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
197 struct kvm_userspace_memory_region mem;
199 mem.slot = slot->slot;
200 mem.guest_phys_addr = slot->start_addr;
201 mem.memory_size = slot->memory_size;
202 mem.userspace_addr = (unsigned long)slot->ram;
203 mem.flags = slot->flags;
204 if (s->migration_log) {
205 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
207 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
210 static void kvm_reset_vcpu(void *opaque)
212 CPUArchState *env = opaque;
214 kvm_arch_reset_vcpu(env);
217 int kvm_init_vcpu(CPUArchState *env)
219 KVMState *s = kvm_state;
220 long mmap_size;
221 int ret;
223 DPRINTF("kvm_init_vcpu\n");
225 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
226 if (ret < 0) {
227 DPRINTF("kvm_create_vcpu failed\n");
228 goto err;
231 env->kvm_fd = ret;
232 env->kvm_state = s;
233 env->kvm_vcpu_dirty = 1;
235 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
236 if (mmap_size < 0) {
237 ret = mmap_size;
238 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
239 goto err;
242 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
243 env->kvm_fd, 0);
244 if (env->kvm_run == MAP_FAILED) {
245 ret = -errno;
246 DPRINTF("mmap'ing vcpu state failed\n");
247 goto err;
250 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
251 s->coalesced_mmio_ring =
252 (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
255 ret = kvm_arch_init_vcpu(env);
256 if (ret == 0) {
257 qemu_register_reset(kvm_reset_vcpu, env);
258 kvm_arch_reset_vcpu(env);
260 err:
261 return ret;
265 * dirty pages logging control
268 static int kvm_mem_flags(KVMState *s, bool log_dirty)
270 return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
273 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
275 KVMState *s = kvm_state;
276 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
277 int old_flags;
279 old_flags = mem->flags;
281 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
282 mem->flags = flags;
284 /* If nothing changed effectively, no need to issue ioctl */
285 if (s->migration_log) {
286 flags |= KVM_MEM_LOG_DIRTY_PAGES;
289 if (flags == old_flags) {
290 return 0;
293 return kvm_set_user_memory_region(s, mem);
296 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
297 ram_addr_t size, bool log_dirty)
299 KVMState *s = kvm_state;
300 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
302 if (mem == NULL) {
303 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
304 TARGET_FMT_plx "\n", __func__, phys_addr,
305 (hwaddr)(phys_addr + size - 1));
306 return -EINVAL;
308 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
311 static void kvm_log_start(MemoryListener *listener,
312 MemoryRegionSection *section)
314 int r;
316 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
317 section->size, true);
318 if (r < 0) {
319 abort();
323 static void kvm_log_stop(MemoryListener *listener,
324 MemoryRegionSection *section)
326 int r;
328 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
329 section->size, false);
330 if (r < 0) {
331 abort();
335 static int kvm_set_migration_log(int enable)
337 KVMState *s = kvm_state;
338 KVMSlot *mem;
339 int i, err;
341 s->migration_log = enable;
343 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
344 mem = &s->slots[i];
346 if (!mem->memory_size) {
347 continue;
349 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
350 continue;
352 err = kvm_set_user_memory_region(s, mem);
353 if (err) {
354 return err;
357 return 0;
360 /* get kvm's dirty pages bitmap and update qemu's */
361 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
362 unsigned long *bitmap)
364 unsigned int i, j;
365 unsigned long page_number, c;
366 hwaddr addr, addr1;
367 unsigned int len = ((section->size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) / HOST_LONG_BITS;
368 unsigned long hpratio = getpagesize() / TARGET_PAGE_SIZE;
371 * bitmap-traveling is faster than memory-traveling (for addr...)
372 * especially when most of the memory is not dirty.
374 for (i = 0; i < len; i++) {
375 if (bitmap[i] != 0) {
376 c = leul_to_cpu(bitmap[i]);
377 do {
378 j = ffsl(c) - 1;
379 c &= ~(1ul << j);
380 page_number = (i * HOST_LONG_BITS + j) * hpratio;
381 addr1 = page_number * TARGET_PAGE_SIZE;
382 addr = section->offset_within_region + addr1;
383 memory_region_set_dirty(section->mr, addr,
384 TARGET_PAGE_SIZE * hpratio);
385 } while (c != 0);
388 return 0;
391 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
394 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
395 * This function updates qemu's dirty bitmap using
396 * memory_region_set_dirty(). This means all bits are set
397 * to dirty.
399 * @start_add: start of logged region.
400 * @end_addr: end of logged region.
402 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
404 KVMState *s = kvm_state;
405 unsigned long size, allocated_size = 0;
406 KVMDirtyLog d;
407 KVMSlot *mem;
408 int ret = 0;
409 hwaddr start_addr = section->offset_within_address_space;
410 hwaddr end_addr = start_addr + section->size;
412 d.dirty_bitmap = NULL;
413 while (start_addr < end_addr) {
414 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
415 if (mem == NULL) {
416 break;
419 /* XXX bad kernel interface alert
420 * For dirty bitmap, kernel allocates array of size aligned to
421 * bits-per-long. But for case when the kernel is 64bits and
422 * the userspace is 32bits, userspace can't align to the same
423 * bits-per-long, since sizeof(long) is different between kernel
424 * and user space. This way, userspace will provide buffer which
425 * may be 4 bytes less than the kernel will use, resulting in
426 * userspace memory corruption (which is not detectable by valgrind
427 * too, in most cases).
428 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
429 * a hope that sizeof(long) wont become >8 any time soon.
431 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
432 /*HOST_LONG_BITS*/ 64) / 8;
433 if (!d.dirty_bitmap) {
434 d.dirty_bitmap = g_malloc(size);
435 } else if (size > allocated_size) {
436 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
438 allocated_size = size;
439 memset(d.dirty_bitmap, 0, allocated_size);
441 d.slot = mem->slot;
443 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
444 DPRINTF("ioctl failed %d\n", errno);
445 ret = -1;
446 break;
449 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
450 start_addr = mem->start_addr + mem->memory_size;
452 g_free(d.dirty_bitmap);
454 return ret;
457 static void kvm_coalesce_mmio_region(MemoryListener *listener,
458 MemoryRegionSection *secion,
459 hwaddr start, hwaddr size)
461 KVMState *s = kvm_state;
463 if (s->coalesced_mmio) {
464 struct kvm_coalesced_mmio_zone zone;
466 zone.addr = start;
467 zone.size = size;
468 zone.pad = 0;
470 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
474 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
475 MemoryRegionSection *secion,
476 hwaddr start, hwaddr size)
478 KVMState *s = kvm_state;
480 if (s->coalesced_mmio) {
481 struct kvm_coalesced_mmio_zone zone;
483 zone.addr = start;
484 zone.size = size;
485 zone.pad = 0;
487 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
491 int kvm_check_extension(KVMState *s, unsigned int extension)
493 int ret;
495 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
496 if (ret < 0) {
497 ret = 0;
500 return ret;
503 static int kvm_check_many_ioeventfds(void)
505 /* Userspace can use ioeventfd for io notification. This requires a host
506 * that supports eventfd(2) and an I/O thread; since eventfd does not
507 * support SIGIO it cannot interrupt the vcpu.
509 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
510 * can avoid creating too many ioeventfds.
512 #if defined(CONFIG_EVENTFD)
513 int ioeventfds[7];
514 int i, ret = 0;
515 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
516 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
517 if (ioeventfds[i] < 0) {
518 break;
520 ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
521 if (ret < 0) {
522 close(ioeventfds[i]);
523 break;
527 /* Decide whether many devices are supported or not */
528 ret = i == ARRAY_SIZE(ioeventfds);
530 while (i-- > 0) {
531 kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
532 close(ioeventfds[i]);
534 return ret;
535 #else
536 return 0;
537 #endif
540 static const KVMCapabilityInfo *
541 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
543 while (list->name) {
544 if (!kvm_check_extension(s, list->value)) {
545 return list;
547 list++;
549 return NULL;
552 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
554 KVMState *s = kvm_state;
555 KVMSlot *mem, old;
556 int err;
557 MemoryRegion *mr = section->mr;
558 bool log_dirty = memory_region_is_logging(mr);
559 hwaddr start_addr = section->offset_within_address_space;
560 ram_addr_t size = section->size;
561 void *ram = NULL;
562 unsigned delta;
564 /* kvm works in page size chunks, but the function may be called
565 with sub-page size and unaligned start address. */
566 delta = TARGET_PAGE_ALIGN(size) - size;
567 if (delta > size) {
568 return;
570 start_addr += delta;
571 size -= delta;
572 size &= TARGET_PAGE_MASK;
573 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
574 return;
577 if (!memory_region_is_ram(mr)) {
578 return;
581 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
583 while (1) {
584 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
585 if (!mem) {
586 break;
589 if (add && start_addr >= mem->start_addr &&
590 (start_addr + size <= mem->start_addr + mem->memory_size) &&
591 (ram - start_addr == mem->ram - mem->start_addr)) {
592 /* The new slot fits into the existing one and comes with
593 * identical parameters - update flags and done. */
594 kvm_slot_dirty_pages_log_change(mem, log_dirty);
595 return;
598 old = *mem;
600 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
601 kvm_physical_sync_dirty_bitmap(section);
604 /* unregister the overlapping slot */
605 mem->memory_size = 0;
606 err = kvm_set_user_memory_region(s, mem);
607 if (err) {
608 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
609 __func__, strerror(-err));
610 abort();
613 /* Workaround for older KVM versions: we can't join slots, even not by
614 * unregistering the previous ones and then registering the larger
615 * slot. We have to maintain the existing fragmentation. Sigh.
617 * This workaround assumes that the new slot starts at the same
618 * address as the first existing one. If not or if some overlapping
619 * slot comes around later, we will fail (not seen in practice so far)
620 * - and actually require a recent KVM version. */
621 if (s->broken_set_mem_region &&
622 old.start_addr == start_addr && old.memory_size < size && add) {
623 mem = kvm_alloc_slot(s);
624 mem->memory_size = old.memory_size;
625 mem->start_addr = old.start_addr;
626 mem->ram = old.ram;
627 mem->flags = kvm_mem_flags(s, log_dirty);
629 err = kvm_set_user_memory_region(s, mem);
630 if (err) {
631 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
632 strerror(-err));
633 abort();
636 start_addr += old.memory_size;
637 ram += old.memory_size;
638 size -= old.memory_size;
639 continue;
642 /* register prefix slot */
643 if (old.start_addr < start_addr) {
644 mem = kvm_alloc_slot(s);
645 mem->memory_size = start_addr - old.start_addr;
646 mem->start_addr = old.start_addr;
647 mem->ram = old.ram;
648 mem->flags = kvm_mem_flags(s, log_dirty);
650 err = kvm_set_user_memory_region(s, mem);
651 if (err) {
652 fprintf(stderr, "%s: error registering prefix slot: %s\n",
653 __func__, strerror(-err));
654 #ifdef TARGET_PPC
655 fprintf(stderr, "%s: This is probably because your kernel's " \
656 "PAGE_SIZE is too big. Please try to use 4k " \
657 "PAGE_SIZE!\n", __func__);
658 #endif
659 abort();
663 /* register suffix slot */
664 if (old.start_addr + old.memory_size > start_addr + size) {
665 ram_addr_t size_delta;
667 mem = kvm_alloc_slot(s);
668 mem->start_addr = start_addr + size;
669 size_delta = mem->start_addr - old.start_addr;
670 mem->memory_size = old.memory_size - size_delta;
671 mem->ram = old.ram + size_delta;
672 mem->flags = kvm_mem_flags(s, log_dirty);
674 err = kvm_set_user_memory_region(s, mem);
675 if (err) {
676 fprintf(stderr, "%s: error registering suffix slot: %s\n",
677 __func__, strerror(-err));
678 abort();
683 /* in case the KVM bug workaround already "consumed" the new slot */
684 if (!size) {
685 return;
687 if (!add) {
688 return;
690 mem = kvm_alloc_slot(s);
691 mem->memory_size = size;
692 mem->start_addr = start_addr;
693 mem->ram = ram;
694 mem->flags = kvm_mem_flags(s, log_dirty);
696 err = kvm_set_user_memory_region(s, mem);
697 if (err) {
698 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
699 strerror(-err));
700 abort();
704 static void kvm_region_add(MemoryListener *listener,
705 MemoryRegionSection *section)
707 kvm_set_phys_mem(section, true);
710 static void kvm_region_del(MemoryListener *listener,
711 MemoryRegionSection *section)
713 kvm_set_phys_mem(section, false);
716 static void kvm_log_sync(MemoryListener *listener,
717 MemoryRegionSection *section)
719 int r;
721 r = kvm_physical_sync_dirty_bitmap(section);
722 if (r < 0) {
723 abort();
727 static void kvm_log_global_start(struct MemoryListener *listener)
729 int r;
731 r = kvm_set_migration_log(1);
732 assert(r >= 0);
735 static void kvm_log_global_stop(struct MemoryListener *listener)
737 int r;
739 r = kvm_set_migration_log(0);
740 assert(r >= 0);
743 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
744 MemoryRegionSection *section,
745 bool match_data, uint64_t data,
746 EventNotifier *e)
748 int fd = event_notifier_get_fd(e);
749 int r;
751 assert(match_data && section->size <= 8);
753 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
754 data, true, section->size);
755 if (r < 0) {
756 abort();
760 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
761 MemoryRegionSection *section,
762 bool match_data, uint64_t data,
763 EventNotifier *e)
765 int fd = event_notifier_get_fd(e);
766 int r;
768 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
769 data, false, section->size);
770 if (r < 0) {
771 abort();
775 static void kvm_io_ioeventfd_add(MemoryListener *listener,
776 MemoryRegionSection *section,
777 bool match_data, uint64_t data,
778 EventNotifier *e)
780 int fd = event_notifier_get_fd(e);
781 int r;
783 assert(match_data && section->size == 2);
785 r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
786 data, true);
787 if (r < 0) {
788 abort();
792 static void kvm_io_ioeventfd_del(MemoryListener *listener,
793 MemoryRegionSection *section,
794 bool match_data, uint64_t data,
795 EventNotifier *e)
798 int fd = event_notifier_get_fd(e);
799 int r;
801 r = kvm_set_ioeventfd_pio_word(fd, section->offset_within_address_space,
802 data, false);
803 if (r < 0) {
804 abort();
808 static MemoryListener kvm_memory_listener = {
809 .region_add = kvm_region_add,
810 .region_del = kvm_region_del,
811 .log_start = kvm_log_start,
812 .log_stop = kvm_log_stop,
813 .log_sync = kvm_log_sync,
814 .log_global_start = kvm_log_global_start,
815 .log_global_stop = kvm_log_global_stop,
816 .eventfd_add = kvm_mem_ioeventfd_add,
817 .eventfd_del = kvm_mem_ioeventfd_del,
818 .coalesced_mmio_add = kvm_coalesce_mmio_region,
819 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
820 .priority = 10,
823 static MemoryListener kvm_io_listener = {
824 .eventfd_add = kvm_io_ioeventfd_add,
825 .eventfd_del = kvm_io_ioeventfd_del,
826 .priority = 10,
829 static void kvm_handle_interrupt(CPUArchState *env, int mask)
831 env->interrupt_request |= mask;
833 if (!qemu_cpu_is_self(env)) {
834 qemu_cpu_kick(env);
838 int kvm_set_irq(KVMState *s, int irq, int level)
840 struct kvm_irq_level event;
841 int ret;
843 assert(kvm_async_interrupts_enabled());
845 event.level = level;
846 event.irq = irq;
847 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
848 if (ret < 0) {
849 perror("kvm_set_irq");
850 abort();
853 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
856 #ifdef KVM_CAP_IRQ_ROUTING
857 typedef struct KVMMSIRoute {
858 struct kvm_irq_routing_entry kroute;
859 QTAILQ_ENTRY(KVMMSIRoute) entry;
860 } KVMMSIRoute;
862 static void set_gsi(KVMState *s, unsigned int gsi)
864 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
867 static void clear_gsi(KVMState *s, unsigned int gsi)
869 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
872 static void kvm_init_irq_routing(KVMState *s)
874 int gsi_count, i;
876 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING);
877 if (gsi_count > 0) {
878 unsigned int gsi_bits, i;
880 /* Round up so we can search ints using ffs */
881 gsi_bits = ALIGN(gsi_count, 32);
882 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
883 s->gsi_count = gsi_count;
885 /* Mark any over-allocated bits as already in use */
886 for (i = gsi_count; i < gsi_bits; i++) {
887 set_gsi(s, i);
891 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
892 s->nr_allocated_irq_routes = 0;
894 if (!s->direct_msi) {
895 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
896 QTAILQ_INIT(&s->msi_hashtab[i]);
900 kvm_arch_init_irq_routing(s);
903 static void kvm_irqchip_commit_routes(KVMState *s)
905 int ret;
907 s->irq_routes->flags = 0;
908 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
909 assert(ret == 0);
912 static void kvm_add_routing_entry(KVMState *s,
913 struct kvm_irq_routing_entry *entry)
915 struct kvm_irq_routing_entry *new;
916 int n, size;
918 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
919 n = s->nr_allocated_irq_routes * 2;
920 if (n < 64) {
921 n = 64;
923 size = sizeof(struct kvm_irq_routing);
924 size += n * sizeof(*new);
925 s->irq_routes = g_realloc(s->irq_routes, size);
926 s->nr_allocated_irq_routes = n;
928 n = s->irq_routes->nr++;
929 new = &s->irq_routes->entries[n];
930 memset(new, 0, sizeof(*new));
931 new->gsi = entry->gsi;
932 new->type = entry->type;
933 new->flags = entry->flags;
934 new->u = entry->u;
936 set_gsi(s, entry->gsi);
938 kvm_irqchip_commit_routes(s);
941 static int kvm_update_routing_entry(KVMState *s,
942 struct kvm_irq_routing_entry *new_entry)
944 struct kvm_irq_routing_entry *entry;
945 int n;
947 for (n = 0; n < s->irq_routes->nr; n++) {
948 entry = &s->irq_routes->entries[n];
949 if (entry->gsi != new_entry->gsi) {
950 continue;
953 entry->type = new_entry->type;
954 entry->flags = new_entry->flags;
955 entry->u = new_entry->u;
957 kvm_irqchip_commit_routes(s);
959 return 0;
962 return -ESRCH;
965 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
967 struct kvm_irq_routing_entry e;
969 assert(pin < s->gsi_count);
971 e.gsi = irq;
972 e.type = KVM_IRQ_ROUTING_IRQCHIP;
973 e.flags = 0;
974 e.u.irqchip.irqchip = irqchip;
975 e.u.irqchip.pin = pin;
976 kvm_add_routing_entry(s, &e);
979 void kvm_irqchip_release_virq(KVMState *s, int virq)
981 struct kvm_irq_routing_entry *e;
982 int i;
984 for (i = 0; i < s->irq_routes->nr; i++) {
985 e = &s->irq_routes->entries[i];
986 if (e->gsi == virq) {
987 s->irq_routes->nr--;
988 *e = s->irq_routes->entries[s->irq_routes->nr];
991 clear_gsi(s, virq);
993 kvm_irqchip_commit_routes(s);
996 static unsigned int kvm_hash_msi(uint32_t data)
998 /* This is optimized for IA32 MSI layout. However, no other arch shall
999 * repeat the mistake of not providing a direct MSI injection API. */
1000 return data & 0xff;
1003 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1005 KVMMSIRoute *route, *next;
1006 unsigned int hash;
1008 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1009 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1010 kvm_irqchip_release_virq(s, route->kroute.gsi);
1011 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1012 g_free(route);
1017 static int kvm_irqchip_get_virq(KVMState *s)
1019 uint32_t *word = s->used_gsi_bitmap;
1020 int max_words = ALIGN(s->gsi_count, 32) / 32;
1021 int i, bit;
1022 bool retry = true;
1024 again:
1025 /* Return the lowest unused GSI in the bitmap */
1026 for (i = 0; i < max_words; i++) {
1027 bit = ffs(~word[i]);
1028 if (!bit) {
1029 continue;
1032 return bit - 1 + i * 32;
1034 if (!s->direct_msi && retry) {
1035 retry = false;
1036 kvm_flush_dynamic_msi_routes(s);
1037 goto again;
1039 return -ENOSPC;
1043 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1045 unsigned int hash = kvm_hash_msi(msg.data);
1046 KVMMSIRoute *route;
1048 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1049 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1050 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1051 route->kroute.u.msi.data == msg.data) {
1052 return route;
1055 return NULL;
1058 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1060 struct kvm_msi msi;
1061 KVMMSIRoute *route;
1063 if (s->direct_msi) {
1064 msi.address_lo = (uint32_t)msg.address;
1065 msi.address_hi = msg.address >> 32;
1066 msi.data = msg.data;
1067 msi.flags = 0;
1068 memset(msi.pad, 0, sizeof(msi.pad));
1070 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1073 route = kvm_lookup_msi_route(s, msg);
1074 if (!route) {
1075 int virq;
1077 virq = kvm_irqchip_get_virq(s);
1078 if (virq < 0) {
1079 return virq;
1082 route = g_malloc(sizeof(KVMMSIRoute));
1083 route->kroute.gsi = virq;
1084 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1085 route->kroute.flags = 0;
1086 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1087 route->kroute.u.msi.address_hi = msg.address >> 32;
1088 route->kroute.u.msi.data = msg.data;
1090 kvm_add_routing_entry(s, &route->kroute);
1092 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1093 entry);
1096 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1098 return kvm_set_irq(s, route->kroute.gsi, 1);
1101 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1103 struct kvm_irq_routing_entry kroute;
1104 int virq;
1106 if (!kvm_gsi_routing_enabled()) {
1107 return -ENOSYS;
1110 virq = kvm_irqchip_get_virq(s);
1111 if (virq < 0) {
1112 return virq;
1115 kroute.gsi = virq;
1116 kroute.type = KVM_IRQ_ROUTING_MSI;
1117 kroute.flags = 0;
1118 kroute.u.msi.address_lo = (uint32_t)msg.address;
1119 kroute.u.msi.address_hi = msg.address >> 32;
1120 kroute.u.msi.data = msg.data;
1122 kvm_add_routing_entry(s, &kroute);
1124 return virq;
1127 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1129 struct kvm_irq_routing_entry kroute;
1131 if (!kvm_irqchip_in_kernel()) {
1132 return -ENOSYS;
1135 kroute.gsi = virq;
1136 kroute.type = KVM_IRQ_ROUTING_MSI;
1137 kroute.flags = 0;
1138 kroute.u.msi.address_lo = (uint32_t)msg.address;
1139 kroute.u.msi.address_hi = msg.address >> 32;
1140 kroute.u.msi.data = msg.data;
1142 return kvm_update_routing_entry(s, &kroute);
1145 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1147 struct kvm_irqfd irqfd = {
1148 .fd = fd,
1149 .gsi = virq,
1150 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1153 if (!kvm_irqfds_enabled()) {
1154 return -ENOSYS;
1157 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1160 #else /* !KVM_CAP_IRQ_ROUTING */
1162 static void kvm_init_irq_routing(KVMState *s)
1166 void kvm_irqchip_release_virq(KVMState *s, int virq)
1170 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1172 abort();
1175 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1177 return -ENOSYS;
1180 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1182 abort();
1184 #endif /* !KVM_CAP_IRQ_ROUTING */
1186 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1188 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, true);
1191 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1193 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), virq, false);
1196 static int kvm_irqchip_create(KVMState *s)
1198 QemuOptsList *list = qemu_find_opts("machine");
1199 int ret;
1201 if (QTAILQ_EMPTY(&list->head) ||
1202 !qemu_opt_get_bool(QTAILQ_FIRST(&list->head),
1203 "kernel_irqchip", true) ||
1204 !kvm_check_extension(s, KVM_CAP_IRQCHIP)) {
1205 return 0;
1208 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1209 if (ret < 0) {
1210 fprintf(stderr, "Create kernel irqchip failed\n");
1211 return ret;
1214 kvm_kernel_irqchip = true;
1215 /* If we have an in-kernel IRQ chip then we must have asynchronous
1216 * interrupt delivery (though the reverse is not necessarily true)
1218 kvm_async_interrupts_allowed = true;
1220 kvm_init_irq_routing(s);
1222 return 0;
1225 static int kvm_max_vcpus(KVMState *s)
1227 int ret;
1229 /* Find number of supported CPUs using the recommended
1230 * procedure from the kernel API documentation to cope with
1231 * older kernels that may be missing capabilities.
1233 ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1234 if (ret) {
1235 return ret;
1237 ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1238 if (ret) {
1239 return ret;
1242 return 4;
1245 int kvm_init(void)
1247 static const char upgrade_note[] =
1248 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1249 "(see http://sourceforge.net/projects/kvm).\n";
1250 KVMState *s;
1251 const KVMCapabilityInfo *missing_cap;
1252 int ret;
1253 int i;
1254 int max_vcpus;
1256 s = g_malloc0(sizeof(KVMState));
1259 * On systems where the kernel can support different base page
1260 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1261 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1262 * page size for the system though.
1264 assert(TARGET_PAGE_SIZE <= getpagesize());
1266 #ifdef KVM_CAP_SET_GUEST_DEBUG
1267 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1268 #endif
1269 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
1270 s->slots[i].slot = i;
1272 s->vmfd = -1;
1273 s->fd = qemu_open("/dev/kvm", O_RDWR);
1274 if (s->fd == -1) {
1275 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1276 ret = -errno;
1277 goto err;
1280 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1281 if (ret < KVM_API_VERSION) {
1282 if (ret > 0) {
1283 ret = -EINVAL;
1285 fprintf(stderr, "kvm version too old\n");
1286 goto err;
1289 if (ret > KVM_API_VERSION) {
1290 ret = -EINVAL;
1291 fprintf(stderr, "kvm version not supported\n");
1292 goto err;
1295 max_vcpus = kvm_max_vcpus(s);
1296 if (smp_cpus > max_vcpus) {
1297 ret = -EINVAL;
1298 fprintf(stderr, "Number of SMP cpus requested (%d) exceeds max cpus "
1299 "supported by KVM (%d)\n", smp_cpus, max_vcpus);
1300 goto err;
1303 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
1304 if (s->vmfd < 0) {
1305 #ifdef TARGET_S390X
1306 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1307 "your host kernel command line\n");
1308 #endif
1309 ret = s->vmfd;
1310 goto err;
1313 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1314 if (!missing_cap) {
1315 missing_cap =
1316 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1318 if (missing_cap) {
1319 ret = -EINVAL;
1320 fprintf(stderr, "kvm does not support %s\n%s",
1321 missing_cap->name, upgrade_note);
1322 goto err;
1325 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1327 s->broken_set_mem_region = 1;
1328 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1329 if (ret > 0) {
1330 s->broken_set_mem_region = 0;
1333 #ifdef KVM_CAP_VCPU_EVENTS
1334 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1335 #endif
1337 s->robust_singlestep =
1338 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1340 #ifdef KVM_CAP_DEBUGREGS
1341 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1342 #endif
1344 #ifdef KVM_CAP_XSAVE
1345 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1346 #endif
1348 #ifdef KVM_CAP_XCRS
1349 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1350 #endif
1352 #ifdef KVM_CAP_PIT_STATE2
1353 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1354 #endif
1356 #ifdef KVM_CAP_IRQ_ROUTING
1357 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1358 #endif
1360 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1362 s->irq_set_ioctl = KVM_IRQ_LINE;
1363 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1364 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1367 ret = kvm_arch_init(s);
1368 if (ret < 0) {
1369 goto err;
1372 ret = kvm_irqchip_create(s);
1373 if (ret < 0) {
1374 goto err;
1377 kvm_state = s;
1378 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1379 memory_listener_register(&kvm_io_listener, &address_space_io);
1381 s->many_ioeventfds = kvm_check_many_ioeventfds();
1383 cpu_interrupt_handler = kvm_handle_interrupt;
1385 return 0;
1387 err:
1388 if (s->vmfd >= 0) {
1389 close(s->vmfd);
1391 if (s->fd != -1) {
1392 close(s->fd);
1394 g_free(s);
1396 return ret;
1399 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1400 uint32_t count)
1402 int i;
1403 uint8_t *ptr = data;
1405 for (i = 0; i < count; i++) {
1406 if (direction == KVM_EXIT_IO_IN) {
1407 switch (size) {
1408 case 1:
1409 stb_p(ptr, cpu_inb(port));
1410 break;
1411 case 2:
1412 stw_p(ptr, cpu_inw(port));
1413 break;
1414 case 4:
1415 stl_p(ptr, cpu_inl(port));
1416 break;
1418 } else {
1419 switch (size) {
1420 case 1:
1421 cpu_outb(port, ldub_p(ptr));
1422 break;
1423 case 2:
1424 cpu_outw(port, lduw_p(ptr));
1425 break;
1426 case 4:
1427 cpu_outl(port, ldl_p(ptr));
1428 break;
1432 ptr += size;
1436 static int kvm_handle_internal_error(CPUArchState *env, struct kvm_run *run)
1438 fprintf(stderr, "KVM internal error.");
1439 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1440 int i;
1442 fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
1443 for (i = 0; i < run->internal.ndata; ++i) {
1444 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1445 i, (uint64_t)run->internal.data[i]);
1447 } else {
1448 fprintf(stderr, "\n");
1450 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1451 fprintf(stderr, "emulation failure\n");
1452 if (!kvm_arch_stop_on_emulation_error(env)) {
1453 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1454 return EXCP_INTERRUPT;
1457 /* FIXME: Should trigger a qmp message to let management know
1458 * something went wrong.
1460 return -1;
1463 void kvm_flush_coalesced_mmio_buffer(void)
1465 KVMState *s = kvm_state;
1467 if (s->coalesced_flush_in_progress) {
1468 return;
1471 s->coalesced_flush_in_progress = true;
1473 if (s->coalesced_mmio_ring) {
1474 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1475 while (ring->first != ring->last) {
1476 struct kvm_coalesced_mmio *ent;
1478 ent = &ring->coalesced_mmio[ring->first];
1480 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1481 smp_wmb();
1482 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1486 s->coalesced_flush_in_progress = false;
1489 static void do_kvm_cpu_synchronize_state(void *_env)
1491 CPUArchState *env = _env;
1493 if (!env->kvm_vcpu_dirty) {
1494 kvm_arch_get_registers(env);
1495 env->kvm_vcpu_dirty = 1;
1499 void kvm_cpu_synchronize_state(CPUArchState *env)
1501 if (!env->kvm_vcpu_dirty) {
1502 run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
1506 void kvm_cpu_synchronize_post_reset(CPUArchState *env)
1508 kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
1509 env->kvm_vcpu_dirty = 0;
1512 void kvm_cpu_synchronize_post_init(CPUArchState *env)
1514 kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
1515 env->kvm_vcpu_dirty = 0;
1518 int kvm_cpu_exec(CPUArchState *env)
1520 struct kvm_run *run = env->kvm_run;
1521 int ret, run_ret;
1523 DPRINTF("kvm_cpu_exec()\n");
1525 if (kvm_arch_process_async_events(env)) {
1526 env->exit_request = 0;
1527 return EXCP_HLT;
1530 do {
1531 if (env->kvm_vcpu_dirty) {
1532 kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
1533 env->kvm_vcpu_dirty = 0;
1536 kvm_arch_pre_run(env, run);
1537 if (env->exit_request) {
1538 DPRINTF("interrupt exit requested\n");
1540 * KVM requires us to reenter the kernel after IO exits to complete
1541 * instruction emulation. This self-signal will ensure that we
1542 * leave ASAP again.
1544 qemu_cpu_kick_self();
1546 qemu_mutex_unlock_iothread();
1548 run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
1550 qemu_mutex_lock_iothread();
1551 kvm_arch_post_run(env, run);
1553 if (run_ret < 0) {
1554 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1555 DPRINTF("io window exit\n");
1556 ret = EXCP_INTERRUPT;
1557 break;
1559 fprintf(stderr, "error: kvm run failed %s\n",
1560 strerror(-run_ret));
1561 abort();
1564 switch (run->exit_reason) {
1565 case KVM_EXIT_IO:
1566 DPRINTF("handle_io\n");
1567 kvm_handle_io(run->io.port,
1568 (uint8_t *)run + run->io.data_offset,
1569 run->io.direction,
1570 run->io.size,
1571 run->io.count);
1572 ret = 0;
1573 break;
1574 case KVM_EXIT_MMIO:
1575 DPRINTF("handle_mmio\n");
1576 cpu_physical_memory_rw(run->mmio.phys_addr,
1577 run->mmio.data,
1578 run->mmio.len,
1579 run->mmio.is_write);
1580 ret = 0;
1581 break;
1582 case KVM_EXIT_IRQ_WINDOW_OPEN:
1583 DPRINTF("irq_window_open\n");
1584 ret = EXCP_INTERRUPT;
1585 break;
1586 case KVM_EXIT_SHUTDOWN:
1587 DPRINTF("shutdown\n");
1588 qemu_system_reset_request();
1589 ret = EXCP_INTERRUPT;
1590 break;
1591 case KVM_EXIT_UNKNOWN:
1592 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1593 (uint64_t)run->hw.hardware_exit_reason);
1594 ret = -1;
1595 break;
1596 case KVM_EXIT_INTERNAL_ERROR:
1597 ret = kvm_handle_internal_error(env, run);
1598 break;
1599 default:
1600 DPRINTF("kvm_arch_handle_exit\n");
1601 ret = kvm_arch_handle_exit(env, run);
1602 break;
1604 } while (ret == 0);
1606 if (ret < 0) {
1607 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1608 vm_stop(RUN_STATE_INTERNAL_ERROR);
1611 env->exit_request = 0;
1612 return ret;
1615 int kvm_ioctl(KVMState *s, int type, ...)
1617 int ret;
1618 void *arg;
1619 va_list ap;
1621 va_start(ap, type);
1622 arg = va_arg(ap, void *);
1623 va_end(ap);
1625 ret = ioctl(s->fd, type, arg);
1626 if (ret == -1) {
1627 ret = -errno;
1629 return ret;
1632 int kvm_vm_ioctl(KVMState *s, int type, ...)
1634 int ret;
1635 void *arg;
1636 va_list ap;
1638 va_start(ap, type);
1639 arg = va_arg(ap, void *);
1640 va_end(ap);
1642 ret = ioctl(s->vmfd, type, arg);
1643 if (ret == -1) {
1644 ret = -errno;
1646 return ret;
1649 int kvm_vcpu_ioctl(CPUArchState *env, int type, ...)
1651 int ret;
1652 void *arg;
1653 va_list ap;
1655 va_start(ap, type);
1656 arg = va_arg(ap, void *);
1657 va_end(ap);
1659 ret = ioctl(env->kvm_fd, type, arg);
1660 if (ret == -1) {
1661 ret = -errno;
1663 return ret;
1666 int kvm_has_sync_mmu(void)
1668 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1671 int kvm_has_vcpu_events(void)
1673 return kvm_state->vcpu_events;
1676 int kvm_has_robust_singlestep(void)
1678 return kvm_state->robust_singlestep;
1681 int kvm_has_debugregs(void)
1683 return kvm_state->debugregs;
1686 int kvm_has_xsave(void)
1688 return kvm_state->xsave;
1691 int kvm_has_xcrs(void)
1693 return kvm_state->xcrs;
1696 int kvm_has_pit_state2(void)
1698 return kvm_state->pit_state2;
1701 int kvm_has_many_ioeventfds(void)
1703 if (!kvm_enabled()) {
1704 return 0;
1706 return kvm_state->many_ioeventfds;
1709 int kvm_has_gsi_routing(void)
1711 #ifdef KVM_CAP_IRQ_ROUTING
1712 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
1713 #else
1714 return false;
1715 #endif
1718 int kvm_has_intx_set_mask(void)
1720 return kvm_state->intx_set_mask;
1723 void *kvm_vmalloc(ram_addr_t size)
1725 #ifdef TARGET_S390X
1726 void *mem;
1728 mem = kvm_arch_vmalloc(size);
1729 if (mem) {
1730 return mem;
1732 #endif
1733 return qemu_vmalloc(size);
1736 void kvm_setup_guest_memory(void *start, size_t size)
1738 #ifdef CONFIG_VALGRIND_H
1739 VALGRIND_MAKE_MEM_DEFINED(start, size);
1740 #endif
1741 if (!kvm_has_sync_mmu()) {
1742 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1744 if (ret) {
1745 perror("qemu_madvise");
1746 fprintf(stderr,
1747 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1748 exit(1);
1753 #ifdef KVM_CAP_SET_GUEST_DEBUG
1754 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUArchState *env,
1755 target_ulong pc)
1757 struct kvm_sw_breakpoint *bp;
1759 QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1760 if (bp->pc == pc) {
1761 return bp;
1764 return NULL;
1767 int kvm_sw_breakpoints_active(CPUArchState *env)
1769 return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1772 struct kvm_set_guest_debug_data {
1773 struct kvm_guest_debug dbg;
1774 CPUArchState *env;
1775 int err;
1778 static void kvm_invoke_set_guest_debug(void *data)
1780 struct kvm_set_guest_debug_data *dbg_data = data;
1781 CPUArchState *env = dbg_data->env;
1783 dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1786 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1788 struct kvm_set_guest_debug_data data;
1790 data.dbg.control = reinject_trap;
1792 if (env->singlestep_enabled) {
1793 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1795 kvm_arch_update_guest_debug(env, &data.dbg);
1796 data.env = env;
1798 run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1799 return data.err;
1802 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1803 target_ulong len, int type)
1805 struct kvm_sw_breakpoint *bp;
1806 CPUArchState *env;
1807 int err;
1809 if (type == GDB_BREAKPOINT_SW) {
1810 bp = kvm_find_sw_breakpoint(current_env, addr);
1811 if (bp) {
1812 bp->use_count++;
1813 return 0;
1816 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
1817 if (!bp) {
1818 return -ENOMEM;
1821 bp->pc = addr;
1822 bp->use_count = 1;
1823 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1824 if (err) {
1825 g_free(bp);
1826 return err;
1829 QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1830 bp, entry);
1831 } else {
1832 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1833 if (err) {
1834 return err;
1838 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1839 err = kvm_update_guest_debug(env, 0);
1840 if (err) {
1841 return err;
1844 return 0;
1847 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1848 target_ulong len, int type)
1850 struct kvm_sw_breakpoint *bp;
1851 CPUArchState *env;
1852 int err;
1854 if (type == GDB_BREAKPOINT_SW) {
1855 bp = kvm_find_sw_breakpoint(current_env, addr);
1856 if (!bp) {
1857 return -ENOENT;
1860 if (bp->use_count > 1) {
1861 bp->use_count--;
1862 return 0;
1865 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1866 if (err) {
1867 return err;
1870 QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1871 g_free(bp);
1872 } else {
1873 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1874 if (err) {
1875 return err;
1879 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1880 err = kvm_update_guest_debug(env, 0);
1881 if (err) {
1882 return err;
1885 return 0;
1888 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1890 struct kvm_sw_breakpoint *bp, *next;
1891 KVMState *s = current_env->kvm_state;
1892 CPUArchState *env;
1894 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1895 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1896 /* Try harder to find a CPU that currently sees the breakpoint. */
1897 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1898 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1899 break;
1904 kvm_arch_remove_all_hw_breakpoints();
1906 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1907 kvm_update_guest_debug(env, 0);
1911 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1913 int kvm_update_guest_debug(CPUArchState *env, unsigned long reinject_trap)
1915 return -EINVAL;
1918 int kvm_insert_breakpoint(CPUArchState *current_env, target_ulong addr,
1919 target_ulong len, int type)
1921 return -EINVAL;
1924 int kvm_remove_breakpoint(CPUArchState *current_env, target_ulong addr,
1925 target_ulong len, int type)
1927 return -EINVAL;
1930 void kvm_remove_all_breakpoints(CPUArchState *current_env)
1933 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1935 int kvm_set_signal_mask(CPUArchState *env, const sigset_t *sigset)
1937 struct kvm_signal_mask *sigmask;
1938 int r;
1940 if (!sigset) {
1941 return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1944 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
1946 sigmask->len = 8;
1947 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1948 r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1949 g_free(sigmask);
1951 return r;
1954 int kvm_set_ioeventfd_mmio(int fd, uint32_t addr, uint32_t val, bool assign,
1955 uint32_t size)
1957 int ret;
1958 struct kvm_ioeventfd iofd;
1960 iofd.datamatch = val;
1961 iofd.addr = addr;
1962 iofd.len = size;
1963 iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1964 iofd.fd = fd;
1966 if (!kvm_enabled()) {
1967 return -ENOSYS;
1970 if (!assign) {
1971 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1974 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1976 if (ret < 0) {
1977 return -errno;
1980 return 0;
1983 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1985 struct kvm_ioeventfd kick = {
1986 .datamatch = val,
1987 .addr = addr,
1988 .len = 2,
1989 .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1990 .fd = fd,
1992 int r;
1993 if (!kvm_enabled()) {
1994 return -ENOSYS;
1996 if (!assign) {
1997 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1999 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
2000 if (r < 0) {
2001 return r;
2003 return 0;
2006 int kvm_on_sigbus_vcpu(CPUArchState *env, int code, void *addr)
2008 return kvm_arch_on_sigbus_vcpu(env, code, addr);
2011 int kvm_on_sigbus(int code, void *addr)
2013 return kvm_arch_on_sigbus(code, addr);