Merge remote-tracking branch 'remotes/riku/tags/pull-linux-user-20150323' into staging
[qemu/ar7.git] / kvm-all.c
blob335438adb526c133b93cac095b76f973836aa0e3
1 /*
2 * QEMU KVM support
4 * Copyright IBM, Corp. 2008
5 * Red Hat, Inc. 2008
7 * Authors:
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Glauber Costa <gcosta@redhat.com>
11 * This work is licensed under the terms of the GNU GPL, version 2 or later.
12 * See the COPYING file in the top-level directory.
16 #include <sys/types.h>
17 #include <sys/ioctl.h>
18 #include <sys/mman.h>
19 #include <stdarg.h>
21 #include <linux/kvm.h>
23 #include "qemu-common.h"
24 #include "qemu/atomic.h"
25 #include "qemu/option.h"
26 #include "qemu/config-file.h"
27 #include "sysemu/sysemu.h"
28 #include "sysemu/accel.h"
29 #include "hw/hw.h"
30 #include "hw/pci/msi.h"
31 #include "hw/s390x/adapter.h"
32 #include "exec/gdbstub.h"
33 #include "sysemu/kvm.h"
34 #include "qemu/bswap.h"
35 #include "exec/memory.h"
36 #include "exec/ram_addr.h"
37 #include "exec/address-spaces.h"
38 #include "qemu/event_notifier.h"
39 #include "trace.h"
41 #include "hw/boards.h"
43 /* This check must be after config-host.h is included */
44 #ifdef CONFIG_EVENTFD
45 #include <sys/eventfd.h>
46 #endif
48 /* KVM uses PAGE_SIZE in its definition of COALESCED_MMIO_MAX */
49 #define PAGE_SIZE TARGET_PAGE_SIZE
51 //#define DEBUG_KVM
53 #ifdef DEBUG_KVM
54 #define DPRINTF(fmt, ...) \
55 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
56 #else
57 #define DPRINTF(fmt, ...) \
58 do { } while (0)
59 #endif
61 #define KVM_MSI_HASHTAB_SIZE 256
63 typedef struct KVMSlot
65 hwaddr start_addr;
66 ram_addr_t memory_size;
67 void *ram;
68 int slot;
69 int flags;
70 } KVMSlot;
72 typedef struct kvm_dirty_log KVMDirtyLog;
74 struct KVMState
76 AccelState parent_obj;
78 KVMSlot *slots;
79 int nr_slots;
80 int fd;
81 int vmfd;
82 int coalesced_mmio;
83 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
84 bool coalesced_flush_in_progress;
85 int broken_set_mem_region;
86 int migration_log;
87 int vcpu_events;
88 int robust_singlestep;
89 int debugregs;
90 #ifdef KVM_CAP_SET_GUEST_DEBUG
91 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
92 #endif
93 int pit_state2;
94 int xsave, xcrs;
95 int many_ioeventfds;
96 int intx_set_mask;
97 /* The man page (and posix) say ioctl numbers are signed int, but
98 * they're not. Linux, glibc and *BSD all treat ioctl numbers as
99 * unsigned, and treating them as signed here can break things */
100 unsigned irq_set_ioctl;
101 unsigned int sigmask_len;
102 #ifdef KVM_CAP_IRQ_ROUTING
103 struct kvm_irq_routing *irq_routes;
104 int nr_allocated_irq_routes;
105 uint32_t *used_gsi_bitmap;
106 unsigned int gsi_count;
107 QTAILQ_HEAD(msi_hashtab, KVMMSIRoute) msi_hashtab[KVM_MSI_HASHTAB_SIZE];
108 bool direct_msi;
109 #endif
112 #define TYPE_KVM_ACCEL ACCEL_CLASS_NAME("kvm")
114 #define KVM_STATE(obj) \
115 OBJECT_CHECK(KVMState, (obj), TYPE_KVM_ACCEL)
117 KVMState *kvm_state;
118 bool kvm_kernel_irqchip;
119 bool kvm_async_interrupts_allowed;
120 bool kvm_halt_in_kernel_allowed;
121 bool kvm_eventfds_allowed;
122 bool kvm_irqfds_allowed;
123 bool kvm_resamplefds_allowed;
124 bool kvm_msi_via_irqfd_allowed;
125 bool kvm_gsi_routing_allowed;
126 bool kvm_gsi_direct_mapping;
127 bool kvm_allowed;
128 bool kvm_readonly_mem_allowed;
129 bool kvm_vm_attributes_allowed;
131 static const KVMCapabilityInfo kvm_required_capabilites[] = {
132 KVM_CAP_INFO(USER_MEMORY),
133 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
134 KVM_CAP_LAST_INFO
137 static KVMSlot *kvm_get_free_slot(KVMState *s)
139 int i;
141 for (i = 0; i < s->nr_slots; i++) {
142 if (s->slots[i].memory_size == 0) {
143 return &s->slots[i];
147 return NULL;
150 bool kvm_has_free_slot(MachineState *ms)
152 return kvm_get_free_slot(KVM_STATE(ms->accelerator));
155 static KVMSlot *kvm_alloc_slot(KVMState *s)
157 KVMSlot *slot = kvm_get_free_slot(s);
159 if (slot) {
160 return slot;
163 fprintf(stderr, "%s: no free slot available\n", __func__);
164 abort();
167 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
168 hwaddr start_addr,
169 hwaddr end_addr)
171 int i;
173 for (i = 0; i < s->nr_slots; i++) {
174 KVMSlot *mem = &s->slots[i];
176 if (start_addr == mem->start_addr &&
177 end_addr == mem->start_addr + mem->memory_size) {
178 return mem;
182 return NULL;
186 * Find overlapping slot with lowest start address
188 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
189 hwaddr start_addr,
190 hwaddr end_addr)
192 KVMSlot *found = NULL;
193 int i;
195 for (i = 0; i < s->nr_slots; i++) {
196 KVMSlot *mem = &s->slots[i];
198 if (mem->memory_size == 0 ||
199 (found && found->start_addr < mem->start_addr)) {
200 continue;
203 if (end_addr > mem->start_addr &&
204 start_addr < mem->start_addr + mem->memory_size) {
205 found = mem;
209 return found;
212 int kvm_physical_memory_addr_from_host(KVMState *s, void *ram,
213 hwaddr *phys_addr)
215 int i;
217 for (i = 0; i < s->nr_slots; i++) {
218 KVMSlot *mem = &s->slots[i];
220 if (ram >= mem->ram && ram < mem->ram + mem->memory_size) {
221 *phys_addr = mem->start_addr + (ram - mem->ram);
222 return 1;
226 return 0;
229 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
231 struct kvm_userspace_memory_region mem;
233 mem.slot = slot->slot;
234 mem.guest_phys_addr = slot->start_addr;
235 mem.userspace_addr = (unsigned long)slot->ram;
236 mem.flags = slot->flags;
237 if (s->migration_log) {
238 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
241 if (slot->memory_size && mem.flags & KVM_MEM_READONLY) {
242 /* Set the slot size to 0 before setting the slot to the desired
243 * value. This is needed based on KVM commit 75d61fbc. */
244 mem.memory_size = 0;
245 kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
247 mem.memory_size = slot->memory_size;
248 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
251 int kvm_init_vcpu(CPUState *cpu)
253 KVMState *s = kvm_state;
254 long mmap_size;
255 int ret;
257 DPRINTF("kvm_init_vcpu\n");
259 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, (void *)kvm_arch_vcpu_id(cpu));
260 if (ret < 0) {
261 DPRINTF("kvm_create_vcpu failed\n");
262 goto err;
265 cpu->kvm_fd = ret;
266 cpu->kvm_state = s;
267 cpu->kvm_vcpu_dirty = true;
269 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
270 if (mmap_size < 0) {
271 ret = mmap_size;
272 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
273 goto err;
276 cpu->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
277 cpu->kvm_fd, 0);
278 if (cpu->kvm_run == MAP_FAILED) {
279 ret = -errno;
280 DPRINTF("mmap'ing vcpu state failed\n");
281 goto err;
284 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
285 s->coalesced_mmio_ring =
286 (void *)cpu->kvm_run + s->coalesced_mmio * PAGE_SIZE;
289 ret = kvm_arch_init_vcpu(cpu);
290 err:
291 return ret;
295 * dirty pages logging control
298 static int kvm_mem_flags(KVMState *s, bool log_dirty, bool readonly)
300 int flags = 0;
301 flags = log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
302 if (readonly && kvm_readonly_mem_allowed) {
303 flags |= KVM_MEM_READONLY;
305 return flags;
308 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
310 KVMState *s = kvm_state;
311 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
312 int old_flags;
314 old_flags = mem->flags;
316 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty, false);
317 mem->flags = flags;
319 /* If nothing changed effectively, no need to issue ioctl */
320 if (s->migration_log) {
321 flags |= KVM_MEM_LOG_DIRTY_PAGES;
324 if (flags == old_flags) {
325 return 0;
328 return kvm_set_user_memory_region(s, mem);
331 static int kvm_dirty_pages_log_change(hwaddr phys_addr,
332 ram_addr_t size, bool log_dirty)
334 KVMState *s = kvm_state;
335 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
337 if (mem == NULL) {
338 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
339 TARGET_FMT_plx "\n", __func__, phys_addr,
340 (hwaddr)(phys_addr + size - 1));
341 return -EINVAL;
343 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
346 static void kvm_log_start(MemoryListener *listener,
347 MemoryRegionSection *section)
349 int r;
351 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
352 int128_get64(section->size), true);
353 if (r < 0) {
354 abort();
358 static void kvm_log_stop(MemoryListener *listener,
359 MemoryRegionSection *section)
361 int r;
363 r = kvm_dirty_pages_log_change(section->offset_within_address_space,
364 int128_get64(section->size), false);
365 if (r < 0) {
366 abort();
370 static int kvm_set_migration_log(bool enable)
372 KVMState *s = kvm_state;
373 KVMSlot *mem;
374 int i, err;
376 s->migration_log = enable;
378 for (i = 0; i < s->nr_slots; i++) {
379 mem = &s->slots[i];
381 if (!mem->memory_size) {
382 continue;
384 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
385 continue;
387 err = kvm_set_user_memory_region(s, mem);
388 if (err) {
389 return err;
392 return 0;
395 /* get kvm's dirty pages bitmap and update qemu's */
396 static int kvm_get_dirty_pages_log_range(MemoryRegionSection *section,
397 unsigned long *bitmap)
399 ram_addr_t start = section->offset_within_region + section->mr->ram_addr;
400 ram_addr_t pages = int128_get64(section->size) / getpagesize();
402 cpu_physical_memory_set_dirty_lebitmap(bitmap, start, pages);
403 return 0;
406 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
409 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
410 * This function updates qemu's dirty bitmap using
411 * memory_region_set_dirty(). This means all bits are set
412 * to dirty.
414 * @start_add: start of logged region.
415 * @end_addr: end of logged region.
417 static int kvm_physical_sync_dirty_bitmap(MemoryRegionSection *section)
419 KVMState *s = kvm_state;
420 unsigned long size, allocated_size = 0;
421 KVMDirtyLog d = {};
422 KVMSlot *mem;
423 int ret = 0;
424 hwaddr start_addr = section->offset_within_address_space;
425 hwaddr end_addr = start_addr + int128_get64(section->size);
427 d.dirty_bitmap = NULL;
428 while (start_addr < end_addr) {
429 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
430 if (mem == NULL) {
431 break;
434 /* XXX bad kernel interface alert
435 * For dirty bitmap, kernel allocates array of size aligned to
436 * bits-per-long. But for case when the kernel is 64bits and
437 * the userspace is 32bits, userspace can't align to the same
438 * bits-per-long, since sizeof(long) is different between kernel
439 * and user space. This way, userspace will provide buffer which
440 * may be 4 bytes less than the kernel will use, resulting in
441 * userspace memory corruption (which is not detectable by valgrind
442 * too, in most cases).
443 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
444 * a hope that sizeof(long) wont become >8 any time soon.
446 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
447 /*HOST_LONG_BITS*/ 64) / 8;
448 if (!d.dirty_bitmap) {
449 d.dirty_bitmap = g_malloc(size);
450 } else if (size > allocated_size) {
451 d.dirty_bitmap = g_realloc(d.dirty_bitmap, size);
453 allocated_size = size;
454 memset(d.dirty_bitmap, 0, allocated_size);
456 d.slot = mem->slot;
458 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
459 DPRINTF("ioctl failed %d\n", errno);
460 ret = -1;
461 break;
464 kvm_get_dirty_pages_log_range(section, d.dirty_bitmap);
465 start_addr = mem->start_addr + mem->memory_size;
467 g_free(d.dirty_bitmap);
469 return ret;
472 static void kvm_coalesce_mmio_region(MemoryListener *listener,
473 MemoryRegionSection *secion,
474 hwaddr start, hwaddr size)
476 KVMState *s = kvm_state;
478 if (s->coalesced_mmio) {
479 struct kvm_coalesced_mmio_zone zone;
481 zone.addr = start;
482 zone.size = size;
483 zone.pad = 0;
485 (void)kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
489 static void kvm_uncoalesce_mmio_region(MemoryListener *listener,
490 MemoryRegionSection *secion,
491 hwaddr start, hwaddr size)
493 KVMState *s = kvm_state;
495 if (s->coalesced_mmio) {
496 struct kvm_coalesced_mmio_zone zone;
498 zone.addr = start;
499 zone.size = size;
500 zone.pad = 0;
502 (void)kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
506 int kvm_check_extension(KVMState *s, unsigned int extension)
508 int ret;
510 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
511 if (ret < 0) {
512 ret = 0;
515 return ret;
518 int kvm_vm_check_extension(KVMState *s, unsigned int extension)
520 int ret;
522 ret = kvm_vm_ioctl(s, KVM_CHECK_EXTENSION, extension);
523 if (ret < 0) {
524 /* VM wide version not implemented, use global one instead */
525 ret = kvm_check_extension(s, extension);
528 return ret;
531 static uint32_t adjust_ioeventfd_endianness(uint32_t val, uint32_t size)
533 #if defined(HOST_WORDS_BIGENDIAN) != defined(TARGET_WORDS_BIGENDIAN)
534 /* The kernel expects ioeventfd values in HOST_WORDS_BIGENDIAN
535 * endianness, but the memory core hands them in target endianness.
536 * For example, PPC is always treated as big-endian even if running
537 * on KVM and on PPC64LE. Correct here.
539 switch (size) {
540 case 2:
541 val = bswap16(val);
542 break;
543 case 4:
544 val = bswap32(val);
545 break;
547 #endif
548 return val;
551 static int kvm_set_ioeventfd_mmio(int fd, hwaddr addr, uint32_t val,
552 bool assign, uint32_t size, bool datamatch)
554 int ret;
555 struct kvm_ioeventfd iofd;
557 iofd.datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0;
558 iofd.addr = addr;
559 iofd.len = size;
560 iofd.flags = 0;
561 iofd.fd = fd;
563 if (!kvm_enabled()) {
564 return -ENOSYS;
567 if (datamatch) {
568 iofd.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
570 if (!assign) {
571 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
574 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
576 if (ret < 0) {
577 return -errno;
580 return 0;
583 static int kvm_set_ioeventfd_pio(int fd, uint16_t addr, uint16_t val,
584 bool assign, uint32_t size, bool datamatch)
586 struct kvm_ioeventfd kick = {
587 .datamatch = datamatch ? adjust_ioeventfd_endianness(val, size) : 0,
588 .addr = addr,
589 .flags = KVM_IOEVENTFD_FLAG_PIO,
590 .len = size,
591 .fd = fd,
593 int r;
594 if (!kvm_enabled()) {
595 return -ENOSYS;
597 if (datamatch) {
598 kick.flags |= KVM_IOEVENTFD_FLAG_DATAMATCH;
600 if (!assign) {
601 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
603 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
604 if (r < 0) {
605 return r;
607 return 0;
611 static int kvm_check_many_ioeventfds(void)
613 /* Userspace can use ioeventfd for io notification. This requires a host
614 * that supports eventfd(2) and an I/O thread; since eventfd does not
615 * support SIGIO it cannot interrupt the vcpu.
617 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
618 * can avoid creating too many ioeventfds.
620 #if defined(CONFIG_EVENTFD)
621 int ioeventfds[7];
622 int i, ret = 0;
623 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
624 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
625 if (ioeventfds[i] < 0) {
626 break;
628 ret = kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, true, 2, true);
629 if (ret < 0) {
630 close(ioeventfds[i]);
631 break;
635 /* Decide whether many devices are supported or not */
636 ret = i == ARRAY_SIZE(ioeventfds);
638 while (i-- > 0) {
639 kvm_set_ioeventfd_pio(ioeventfds[i], 0, i, false, 2, true);
640 close(ioeventfds[i]);
642 return ret;
643 #else
644 return 0;
645 #endif
648 static const KVMCapabilityInfo *
649 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
651 while (list->name) {
652 if (!kvm_check_extension(s, list->value)) {
653 return list;
655 list++;
657 return NULL;
660 static void kvm_set_phys_mem(MemoryRegionSection *section, bool add)
662 KVMState *s = kvm_state;
663 KVMSlot *mem, old;
664 int err;
665 MemoryRegion *mr = section->mr;
666 bool log_dirty = memory_region_is_logging(mr);
667 bool writeable = !mr->readonly && !mr->rom_device;
668 bool readonly_flag = mr->readonly || memory_region_is_romd(mr);
669 hwaddr start_addr = section->offset_within_address_space;
670 ram_addr_t size = int128_get64(section->size);
671 void *ram = NULL;
672 unsigned delta;
674 /* kvm works in page size chunks, but the function may be called
675 with sub-page size and unaligned start address. Pad the start
676 address to next and truncate size to previous page boundary. */
677 delta = (TARGET_PAGE_SIZE - (start_addr & ~TARGET_PAGE_MASK));
678 delta &= ~TARGET_PAGE_MASK;
679 if (delta > size) {
680 return;
682 start_addr += delta;
683 size -= delta;
684 size &= TARGET_PAGE_MASK;
685 if (!size || (start_addr & ~TARGET_PAGE_MASK)) {
686 return;
689 if (!memory_region_is_ram(mr)) {
690 if (writeable || !kvm_readonly_mem_allowed) {
691 return;
692 } else if (!mr->romd_mode) {
693 /* If the memory device is not in romd_mode, then we actually want
694 * to remove the kvm memory slot so all accesses will trap. */
695 add = false;
699 ram = memory_region_get_ram_ptr(mr) + section->offset_within_region + delta;
701 while (1) {
702 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
703 if (!mem) {
704 break;
707 if (add && start_addr >= mem->start_addr &&
708 (start_addr + size <= mem->start_addr + mem->memory_size) &&
709 (ram - start_addr == mem->ram - mem->start_addr)) {
710 /* The new slot fits into the existing one and comes with
711 * identical parameters - update flags and done. */
712 kvm_slot_dirty_pages_log_change(mem, log_dirty);
713 return;
716 old = *mem;
718 if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
719 kvm_physical_sync_dirty_bitmap(section);
722 /* unregister the overlapping slot */
723 mem->memory_size = 0;
724 err = kvm_set_user_memory_region(s, mem);
725 if (err) {
726 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
727 __func__, strerror(-err));
728 abort();
731 /* Workaround for older KVM versions: we can't join slots, even not by
732 * unregistering the previous ones and then registering the larger
733 * slot. We have to maintain the existing fragmentation. Sigh.
735 * This workaround assumes that the new slot starts at the same
736 * address as the first existing one. If not or if some overlapping
737 * slot comes around later, we will fail (not seen in practice so far)
738 * - and actually require a recent KVM version. */
739 if (s->broken_set_mem_region &&
740 old.start_addr == start_addr && old.memory_size < size && add) {
741 mem = kvm_alloc_slot(s);
742 mem->memory_size = old.memory_size;
743 mem->start_addr = old.start_addr;
744 mem->ram = old.ram;
745 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
747 err = kvm_set_user_memory_region(s, mem);
748 if (err) {
749 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
750 strerror(-err));
751 abort();
754 start_addr += old.memory_size;
755 ram += old.memory_size;
756 size -= old.memory_size;
757 continue;
760 /* register prefix slot */
761 if (old.start_addr < start_addr) {
762 mem = kvm_alloc_slot(s);
763 mem->memory_size = start_addr - old.start_addr;
764 mem->start_addr = old.start_addr;
765 mem->ram = old.ram;
766 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
768 err = kvm_set_user_memory_region(s, mem);
769 if (err) {
770 fprintf(stderr, "%s: error registering prefix slot: %s\n",
771 __func__, strerror(-err));
772 #ifdef TARGET_PPC
773 fprintf(stderr, "%s: This is probably because your kernel's " \
774 "PAGE_SIZE is too big. Please try to use 4k " \
775 "PAGE_SIZE!\n", __func__);
776 #endif
777 abort();
781 /* register suffix slot */
782 if (old.start_addr + old.memory_size > start_addr + size) {
783 ram_addr_t size_delta;
785 mem = kvm_alloc_slot(s);
786 mem->start_addr = start_addr + size;
787 size_delta = mem->start_addr - old.start_addr;
788 mem->memory_size = old.memory_size - size_delta;
789 mem->ram = old.ram + size_delta;
790 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
792 err = kvm_set_user_memory_region(s, mem);
793 if (err) {
794 fprintf(stderr, "%s: error registering suffix slot: %s\n",
795 __func__, strerror(-err));
796 abort();
801 /* in case the KVM bug workaround already "consumed" the new slot */
802 if (!size) {
803 return;
805 if (!add) {
806 return;
808 mem = kvm_alloc_slot(s);
809 mem->memory_size = size;
810 mem->start_addr = start_addr;
811 mem->ram = ram;
812 mem->flags = kvm_mem_flags(s, log_dirty, readonly_flag);
814 err = kvm_set_user_memory_region(s, mem);
815 if (err) {
816 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
817 strerror(-err));
818 abort();
822 static void kvm_region_add(MemoryListener *listener,
823 MemoryRegionSection *section)
825 memory_region_ref(section->mr);
826 kvm_set_phys_mem(section, true);
829 static void kvm_region_del(MemoryListener *listener,
830 MemoryRegionSection *section)
832 kvm_set_phys_mem(section, false);
833 memory_region_unref(section->mr);
836 static void kvm_log_sync(MemoryListener *listener,
837 MemoryRegionSection *section)
839 int r;
841 r = kvm_physical_sync_dirty_bitmap(section);
842 if (r < 0) {
843 abort();
847 static void kvm_log_global_start(struct MemoryListener *listener)
849 int r;
851 r = kvm_set_migration_log(1);
852 assert(r >= 0);
855 static void kvm_log_global_stop(struct MemoryListener *listener)
857 int r;
859 r = kvm_set_migration_log(0);
860 assert(r >= 0);
863 static void kvm_mem_ioeventfd_add(MemoryListener *listener,
864 MemoryRegionSection *section,
865 bool match_data, uint64_t data,
866 EventNotifier *e)
868 int fd = event_notifier_get_fd(e);
869 int r;
871 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
872 data, true, int128_get64(section->size),
873 match_data);
874 if (r < 0) {
875 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
876 __func__, strerror(-r));
877 abort();
881 static void kvm_mem_ioeventfd_del(MemoryListener *listener,
882 MemoryRegionSection *section,
883 bool match_data, uint64_t data,
884 EventNotifier *e)
886 int fd = event_notifier_get_fd(e);
887 int r;
889 r = kvm_set_ioeventfd_mmio(fd, section->offset_within_address_space,
890 data, false, int128_get64(section->size),
891 match_data);
892 if (r < 0) {
893 abort();
897 static void kvm_io_ioeventfd_add(MemoryListener *listener,
898 MemoryRegionSection *section,
899 bool match_data, uint64_t data,
900 EventNotifier *e)
902 int fd = event_notifier_get_fd(e);
903 int r;
905 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
906 data, true, int128_get64(section->size),
907 match_data);
908 if (r < 0) {
909 fprintf(stderr, "%s: error adding ioeventfd: %s\n",
910 __func__, strerror(-r));
911 abort();
915 static void kvm_io_ioeventfd_del(MemoryListener *listener,
916 MemoryRegionSection *section,
917 bool match_data, uint64_t data,
918 EventNotifier *e)
921 int fd = event_notifier_get_fd(e);
922 int r;
924 r = kvm_set_ioeventfd_pio(fd, section->offset_within_address_space,
925 data, false, int128_get64(section->size),
926 match_data);
927 if (r < 0) {
928 abort();
932 static MemoryListener kvm_memory_listener = {
933 .region_add = kvm_region_add,
934 .region_del = kvm_region_del,
935 .log_start = kvm_log_start,
936 .log_stop = kvm_log_stop,
937 .log_sync = kvm_log_sync,
938 .log_global_start = kvm_log_global_start,
939 .log_global_stop = kvm_log_global_stop,
940 .eventfd_add = kvm_mem_ioeventfd_add,
941 .eventfd_del = kvm_mem_ioeventfd_del,
942 .coalesced_mmio_add = kvm_coalesce_mmio_region,
943 .coalesced_mmio_del = kvm_uncoalesce_mmio_region,
944 .priority = 10,
947 static MemoryListener kvm_io_listener = {
948 .eventfd_add = kvm_io_ioeventfd_add,
949 .eventfd_del = kvm_io_ioeventfd_del,
950 .priority = 10,
953 static void kvm_handle_interrupt(CPUState *cpu, int mask)
955 cpu->interrupt_request |= mask;
957 if (!qemu_cpu_is_self(cpu)) {
958 qemu_cpu_kick(cpu);
962 int kvm_set_irq(KVMState *s, int irq, int level)
964 struct kvm_irq_level event;
965 int ret;
967 assert(kvm_async_interrupts_enabled());
969 event.level = level;
970 event.irq = irq;
971 ret = kvm_vm_ioctl(s, s->irq_set_ioctl, &event);
972 if (ret < 0) {
973 perror("kvm_set_irq");
974 abort();
977 return (s->irq_set_ioctl == KVM_IRQ_LINE) ? 1 : event.status;
980 #ifdef KVM_CAP_IRQ_ROUTING
981 typedef struct KVMMSIRoute {
982 struct kvm_irq_routing_entry kroute;
983 QTAILQ_ENTRY(KVMMSIRoute) entry;
984 } KVMMSIRoute;
986 static void set_gsi(KVMState *s, unsigned int gsi)
988 s->used_gsi_bitmap[gsi / 32] |= 1U << (gsi % 32);
991 static void clear_gsi(KVMState *s, unsigned int gsi)
993 s->used_gsi_bitmap[gsi / 32] &= ~(1U << (gsi % 32));
996 void kvm_init_irq_routing(KVMState *s)
998 int gsi_count, i;
1000 gsi_count = kvm_check_extension(s, KVM_CAP_IRQ_ROUTING) - 1;
1001 if (gsi_count > 0) {
1002 unsigned int gsi_bits, i;
1004 /* Round up so we can search ints using ffs */
1005 gsi_bits = ALIGN(gsi_count, 32);
1006 s->used_gsi_bitmap = g_malloc0(gsi_bits / 8);
1007 s->gsi_count = gsi_count;
1009 /* Mark any over-allocated bits as already in use */
1010 for (i = gsi_count; i < gsi_bits; i++) {
1011 set_gsi(s, i);
1015 s->irq_routes = g_malloc0(sizeof(*s->irq_routes));
1016 s->nr_allocated_irq_routes = 0;
1018 if (!s->direct_msi) {
1019 for (i = 0; i < KVM_MSI_HASHTAB_SIZE; i++) {
1020 QTAILQ_INIT(&s->msi_hashtab[i]);
1024 kvm_arch_init_irq_routing(s);
1027 void kvm_irqchip_commit_routes(KVMState *s)
1029 int ret;
1031 s->irq_routes->flags = 0;
1032 ret = kvm_vm_ioctl(s, KVM_SET_GSI_ROUTING, s->irq_routes);
1033 assert(ret == 0);
1036 static void kvm_add_routing_entry(KVMState *s,
1037 struct kvm_irq_routing_entry *entry)
1039 struct kvm_irq_routing_entry *new;
1040 int n, size;
1042 if (s->irq_routes->nr == s->nr_allocated_irq_routes) {
1043 n = s->nr_allocated_irq_routes * 2;
1044 if (n < 64) {
1045 n = 64;
1047 size = sizeof(struct kvm_irq_routing);
1048 size += n * sizeof(*new);
1049 s->irq_routes = g_realloc(s->irq_routes, size);
1050 s->nr_allocated_irq_routes = n;
1052 n = s->irq_routes->nr++;
1053 new = &s->irq_routes->entries[n];
1055 *new = *entry;
1057 set_gsi(s, entry->gsi);
1060 static int kvm_update_routing_entry(KVMState *s,
1061 struct kvm_irq_routing_entry *new_entry)
1063 struct kvm_irq_routing_entry *entry;
1064 int n;
1066 for (n = 0; n < s->irq_routes->nr; n++) {
1067 entry = &s->irq_routes->entries[n];
1068 if (entry->gsi != new_entry->gsi) {
1069 continue;
1072 if(!memcmp(entry, new_entry, sizeof *entry)) {
1073 return 0;
1076 *entry = *new_entry;
1078 kvm_irqchip_commit_routes(s);
1080 return 0;
1083 return -ESRCH;
1086 void kvm_irqchip_add_irq_route(KVMState *s, int irq, int irqchip, int pin)
1088 struct kvm_irq_routing_entry e = {};
1090 assert(pin < s->gsi_count);
1092 e.gsi = irq;
1093 e.type = KVM_IRQ_ROUTING_IRQCHIP;
1094 e.flags = 0;
1095 e.u.irqchip.irqchip = irqchip;
1096 e.u.irqchip.pin = pin;
1097 kvm_add_routing_entry(s, &e);
1100 void kvm_irqchip_release_virq(KVMState *s, int virq)
1102 struct kvm_irq_routing_entry *e;
1103 int i;
1105 if (kvm_gsi_direct_mapping()) {
1106 return;
1109 for (i = 0; i < s->irq_routes->nr; i++) {
1110 e = &s->irq_routes->entries[i];
1111 if (e->gsi == virq) {
1112 s->irq_routes->nr--;
1113 *e = s->irq_routes->entries[s->irq_routes->nr];
1116 clear_gsi(s, virq);
1119 static unsigned int kvm_hash_msi(uint32_t data)
1121 /* This is optimized for IA32 MSI layout. However, no other arch shall
1122 * repeat the mistake of not providing a direct MSI injection API. */
1123 return data & 0xff;
1126 static void kvm_flush_dynamic_msi_routes(KVMState *s)
1128 KVMMSIRoute *route, *next;
1129 unsigned int hash;
1131 for (hash = 0; hash < KVM_MSI_HASHTAB_SIZE; hash++) {
1132 QTAILQ_FOREACH_SAFE(route, &s->msi_hashtab[hash], entry, next) {
1133 kvm_irqchip_release_virq(s, route->kroute.gsi);
1134 QTAILQ_REMOVE(&s->msi_hashtab[hash], route, entry);
1135 g_free(route);
1140 static int kvm_irqchip_get_virq(KVMState *s)
1142 uint32_t *word = s->used_gsi_bitmap;
1143 int max_words = ALIGN(s->gsi_count, 32) / 32;
1144 int i, bit;
1145 bool retry = true;
1147 again:
1148 /* Return the lowest unused GSI in the bitmap */
1149 for (i = 0; i < max_words; i++) {
1150 bit = ffs(~word[i]);
1151 if (!bit) {
1152 continue;
1155 return bit - 1 + i * 32;
1157 if (!s->direct_msi && retry) {
1158 retry = false;
1159 kvm_flush_dynamic_msi_routes(s);
1160 goto again;
1162 return -ENOSPC;
1166 static KVMMSIRoute *kvm_lookup_msi_route(KVMState *s, MSIMessage msg)
1168 unsigned int hash = kvm_hash_msi(msg.data);
1169 KVMMSIRoute *route;
1171 QTAILQ_FOREACH(route, &s->msi_hashtab[hash], entry) {
1172 if (route->kroute.u.msi.address_lo == (uint32_t)msg.address &&
1173 route->kroute.u.msi.address_hi == (msg.address >> 32) &&
1174 route->kroute.u.msi.data == le32_to_cpu(msg.data)) {
1175 return route;
1178 return NULL;
1181 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1183 struct kvm_msi msi;
1184 KVMMSIRoute *route;
1186 if (s->direct_msi) {
1187 msi.address_lo = (uint32_t)msg.address;
1188 msi.address_hi = msg.address >> 32;
1189 msi.data = le32_to_cpu(msg.data);
1190 msi.flags = 0;
1191 memset(msi.pad, 0, sizeof(msi.pad));
1193 return kvm_vm_ioctl(s, KVM_SIGNAL_MSI, &msi);
1196 route = kvm_lookup_msi_route(s, msg);
1197 if (!route) {
1198 int virq;
1200 virq = kvm_irqchip_get_virq(s);
1201 if (virq < 0) {
1202 return virq;
1205 route = g_malloc0(sizeof(KVMMSIRoute));
1206 route->kroute.gsi = virq;
1207 route->kroute.type = KVM_IRQ_ROUTING_MSI;
1208 route->kroute.flags = 0;
1209 route->kroute.u.msi.address_lo = (uint32_t)msg.address;
1210 route->kroute.u.msi.address_hi = msg.address >> 32;
1211 route->kroute.u.msi.data = le32_to_cpu(msg.data);
1213 kvm_add_routing_entry(s, &route->kroute);
1214 kvm_irqchip_commit_routes(s);
1216 QTAILQ_INSERT_TAIL(&s->msi_hashtab[kvm_hash_msi(msg.data)], route,
1217 entry);
1220 assert(route->kroute.type == KVM_IRQ_ROUTING_MSI);
1222 return kvm_set_irq(s, route->kroute.gsi, 1);
1225 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1227 struct kvm_irq_routing_entry kroute = {};
1228 int virq;
1230 if (kvm_gsi_direct_mapping()) {
1231 return msg.data & 0xffff;
1234 if (!kvm_gsi_routing_enabled()) {
1235 return -ENOSYS;
1238 virq = kvm_irqchip_get_virq(s);
1239 if (virq < 0) {
1240 return virq;
1243 kroute.gsi = virq;
1244 kroute.type = KVM_IRQ_ROUTING_MSI;
1245 kroute.flags = 0;
1246 kroute.u.msi.address_lo = (uint32_t)msg.address;
1247 kroute.u.msi.address_hi = msg.address >> 32;
1248 kroute.u.msi.data = le32_to_cpu(msg.data);
1249 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1250 kvm_irqchip_release_virq(s, virq);
1251 return -EINVAL;
1254 kvm_add_routing_entry(s, &kroute);
1255 kvm_irqchip_commit_routes(s);
1257 return virq;
1260 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1262 struct kvm_irq_routing_entry kroute = {};
1264 if (kvm_gsi_direct_mapping()) {
1265 return 0;
1268 if (!kvm_irqchip_in_kernel()) {
1269 return -ENOSYS;
1272 kroute.gsi = virq;
1273 kroute.type = KVM_IRQ_ROUTING_MSI;
1274 kroute.flags = 0;
1275 kroute.u.msi.address_lo = (uint32_t)msg.address;
1276 kroute.u.msi.address_hi = msg.address >> 32;
1277 kroute.u.msi.data = le32_to_cpu(msg.data);
1278 if (kvm_arch_fixup_msi_route(&kroute, msg.address, msg.data)) {
1279 return -EINVAL;
1282 return kvm_update_routing_entry(s, &kroute);
1285 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int rfd, int virq,
1286 bool assign)
1288 struct kvm_irqfd irqfd = {
1289 .fd = fd,
1290 .gsi = virq,
1291 .flags = assign ? 0 : KVM_IRQFD_FLAG_DEASSIGN,
1294 if (rfd != -1) {
1295 irqfd.flags |= KVM_IRQFD_FLAG_RESAMPLE;
1296 irqfd.resamplefd = rfd;
1299 if (!kvm_irqfds_enabled()) {
1300 return -ENOSYS;
1303 return kvm_vm_ioctl(s, KVM_IRQFD, &irqfd);
1306 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1308 struct kvm_irq_routing_entry kroute = {};
1309 int virq;
1311 if (!kvm_gsi_routing_enabled()) {
1312 return -ENOSYS;
1315 virq = kvm_irqchip_get_virq(s);
1316 if (virq < 0) {
1317 return virq;
1320 kroute.gsi = virq;
1321 kroute.type = KVM_IRQ_ROUTING_S390_ADAPTER;
1322 kroute.flags = 0;
1323 kroute.u.adapter.summary_addr = adapter->summary_addr;
1324 kroute.u.adapter.ind_addr = adapter->ind_addr;
1325 kroute.u.adapter.summary_offset = adapter->summary_offset;
1326 kroute.u.adapter.ind_offset = adapter->ind_offset;
1327 kroute.u.adapter.adapter_id = adapter->adapter_id;
1329 kvm_add_routing_entry(s, &kroute);
1330 kvm_irqchip_commit_routes(s);
1332 return virq;
1335 #else /* !KVM_CAP_IRQ_ROUTING */
1337 void kvm_init_irq_routing(KVMState *s)
1341 void kvm_irqchip_release_virq(KVMState *s, int virq)
1345 int kvm_irqchip_send_msi(KVMState *s, MSIMessage msg)
1347 abort();
1350 int kvm_irqchip_add_msi_route(KVMState *s, MSIMessage msg)
1352 return -ENOSYS;
1355 int kvm_irqchip_add_adapter_route(KVMState *s, AdapterInfo *adapter)
1357 return -ENOSYS;
1360 static int kvm_irqchip_assign_irqfd(KVMState *s, int fd, int virq, bool assign)
1362 abort();
1365 int kvm_irqchip_update_msi_route(KVMState *s, int virq, MSIMessage msg)
1367 return -ENOSYS;
1369 #endif /* !KVM_CAP_IRQ_ROUTING */
1371 int kvm_irqchip_add_irqfd_notifier(KVMState *s, EventNotifier *n,
1372 EventNotifier *rn, int virq)
1374 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n),
1375 rn ? event_notifier_get_fd(rn) : -1, virq, true);
1378 int kvm_irqchip_remove_irqfd_notifier(KVMState *s, EventNotifier *n, int virq)
1380 return kvm_irqchip_assign_irqfd(s, event_notifier_get_fd(n), -1, virq,
1381 false);
1384 static int kvm_irqchip_create(MachineState *machine, KVMState *s)
1386 int ret;
1388 if (!machine_kernel_irqchip_allowed(machine) ||
1389 (!kvm_check_extension(s, KVM_CAP_IRQCHIP) &&
1390 (kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0) < 0))) {
1391 return 0;
1394 /* First probe and see if there's a arch-specific hook to create the
1395 * in-kernel irqchip for us */
1396 ret = kvm_arch_irqchip_create(s);
1397 if (ret < 0) {
1398 return ret;
1399 } else if (ret == 0) {
1400 ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
1401 if (ret < 0) {
1402 fprintf(stderr, "Create kernel irqchip failed\n");
1403 return ret;
1407 kvm_kernel_irqchip = true;
1408 /* If we have an in-kernel IRQ chip then we must have asynchronous
1409 * interrupt delivery (though the reverse is not necessarily true)
1411 kvm_async_interrupts_allowed = true;
1412 kvm_halt_in_kernel_allowed = true;
1414 kvm_init_irq_routing(s);
1416 return 0;
1419 /* Find number of supported CPUs using the recommended
1420 * procedure from the kernel API documentation to cope with
1421 * older kernels that may be missing capabilities.
1423 static int kvm_recommended_vcpus(KVMState *s)
1425 int ret = kvm_check_extension(s, KVM_CAP_NR_VCPUS);
1426 return (ret) ? ret : 4;
1429 static int kvm_max_vcpus(KVMState *s)
1431 int ret = kvm_check_extension(s, KVM_CAP_MAX_VCPUS);
1432 return (ret) ? ret : kvm_recommended_vcpus(s);
1435 static int kvm_init(MachineState *ms)
1437 MachineClass *mc = MACHINE_GET_CLASS(ms);
1438 static const char upgrade_note[] =
1439 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
1440 "(see http://sourceforge.net/projects/kvm).\n";
1441 struct {
1442 const char *name;
1443 int num;
1444 } num_cpus[] = {
1445 { "SMP", smp_cpus },
1446 { "hotpluggable", max_cpus },
1447 { NULL, }
1448 }, *nc = num_cpus;
1449 int soft_vcpus_limit, hard_vcpus_limit;
1450 KVMState *s;
1451 const KVMCapabilityInfo *missing_cap;
1452 int ret;
1453 int i, type = 0;
1454 const char *kvm_type;
1456 s = KVM_STATE(ms->accelerator);
1459 * On systems where the kernel can support different base page
1460 * sizes, host page size may be different from TARGET_PAGE_SIZE,
1461 * even with KVM. TARGET_PAGE_SIZE is assumed to be the minimum
1462 * page size for the system though.
1464 assert(TARGET_PAGE_SIZE <= getpagesize());
1465 page_size_init();
1467 s->sigmask_len = 8;
1469 #ifdef KVM_CAP_SET_GUEST_DEBUG
1470 QTAILQ_INIT(&s->kvm_sw_breakpoints);
1471 #endif
1472 s->vmfd = -1;
1473 s->fd = qemu_open("/dev/kvm", O_RDWR);
1474 if (s->fd == -1) {
1475 fprintf(stderr, "Could not access KVM kernel module: %m\n");
1476 ret = -errno;
1477 goto err;
1480 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
1481 if (ret < KVM_API_VERSION) {
1482 if (ret >= 0) {
1483 ret = -EINVAL;
1485 fprintf(stderr, "kvm version too old\n");
1486 goto err;
1489 if (ret > KVM_API_VERSION) {
1490 ret = -EINVAL;
1491 fprintf(stderr, "kvm version not supported\n");
1492 goto err;
1495 s->nr_slots = kvm_check_extension(s, KVM_CAP_NR_MEMSLOTS);
1497 /* If unspecified, use the default value */
1498 if (!s->nr_slots) {
1499 s->nr_slots = 32;
1502 s->slots = g_malloc0(s->nr_slots * sizeof(KVMSlot));
1504 for (i = 0; i < s->nr_slots; i++) {
1505 s->slots[i].slot = i;
1508 /* check the vcpu limits */
1509 soft_vcpus_limit = kvm_recommended_vcpus(s);
1510 hard_vcpus_limit = kvm_max_vcpus(s);
1512 while (nc->name) {
1513 if (nc->num > soft_vcpus_limit) {
1514 fprintf(stderr,
1515 "Warning: Number of %s cpus requested (%d) exceeds "
1516 "the recommended cpus supported by KVM (%d)\n",
1517 nc->name, nc->num, soft_vcpus_limit);
1519 if (nc->num > hard_vcpus_limit) {
1520 fprintf(stderr, "Number of %s cpus requested (%d) exceeds "
1521 "the maximum cpus supported by KVM (%d)\n",
1522 nc->name, nc->num, hard_vcpus_limit);
1523 exit(1);
1526 nc++;
1529 kvm_type = qemu_opt_get(qemu_get_machine_opts(), "kvm-type");
1530 if (mc->kvm_type) {
1531 type = mc->kvm_type(kvm_type);
1532 } else if (kvm_type) {
1533 ret = -EINVAL;
1534 fprintf(stderr, "Invalid argument kvm-type=%s\n", kvm_type);
1535 goto err;
1538 do {
1539 ret = kvm_ioctl(s, KVM_CREATE_VM, type);
1540 } while (ret == -EINTR);
1542 if (ret < 0) {
1543 fprintf(stderr, "ioctl(KVM_CREATE_VM) failed: %d %s\n", -ret,
1544 strerror(-ret));
1546 #ifdef TARGET_S390X
1547 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
1548 "your host kernel command line\n");
1549 #endif
1550 goto err;
1553 s->vmfd = ret;
1554 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
1555 if (!missing_cap) {
1556 missing_cap =
1557 kvm_check_extension_list(s, kvm_arch_required_capabilities);
1559 if (missing_cap) {
1560 ret = -EINVAL;
1561 fprintf(stderr, "kvm does not support %s\n%s",
1562 missing_cap->name, upgrade_note);
1563 goto err;
1566 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
1568 s->broken_set_mem_region = 1;
1569 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
1570 if (ret > 0) {
1571 s->broken_set_mem_region = 0;
1574 #ifdef KVM_CAP_VCPU_EVENTS
1575 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
1576 #endif
1578 s->robust_singlestep =
1579 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
1581 #ifdef KVM_CAP_DEBUGREGS
1582 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
1583 #endif
1585 #ifdef KVM_CAP_XSAVE
1586 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
1587 #endif
1589 #ifdef KVM_CAP_XCRS
1590 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
1591 #endif
1593 #ifdef KVM_CAP_PIT_STATE2
1594 s->pit_state2 = kvm_check_extension(s, KVM_CAP_PIT_STATE2);
1595 #endif
1597 #ifdef KVM_CAP_IRQ_ROUTING
1598 s->direct_msi = (kvm_check_extension(s, KVM_CAP_SIGNAL_MSI) > 0);
1599 #endif
1601 s->intx_set_mask = kvm_check_extension(s, KVM_CAP_PCI_2_3);
1603 s->irq_set_ioctl = KVM_IRQ_LINE;
1604 if (kvm_check_extension(s, KVM_CAP_IRQ_INJECT_STATUS)) {
1605 s->irq_set_ioctl = KVM_IRQ_LINE_STATUS;
1608 #ifdef KVM_CAP_READONLY_MEM
1609 kvm_readonly_mem_allowed =
1610 (kvm_check_extension(s, KVM_CAP_READONLY_MEM) > 0);
1611 #endif
1613 kvm_eventfds_allowed =
1614 (kvm_check_extension(s, KVM_CAP_IOEVENTFD) > 0);
1616 kvm_irqfds_allowed =
1617 (kvm_check_extension(s, KVM_CAP_IRQFD) > 0);
1619 kvm_resamplefds_allowed =
1620 (kvm_check_extension(s, KVM_CAP_IRQFD_RESAMPLE) > 0);
1622 kvm_vm_attributes_allowed =
1623 (kvm_check_extension(s, KVM_CAP_VM_ATTRIBUTES) > 0);
1625 ret = kvm_arch_init(ms, s);
1626 if (ret < 0) {
1627 goto err;
1630 ret = kvm_irqchip_create(ms, s);
1631 if (ret < 0) {
1632 goto err;
1635 kvm_state = s;
1636 memory_listener_register(&kvm_memory_listener, &address_space_memory);
1637 memory_listener_register(&kvm_io_listener, &address_space_io);
1639 s->many_ioeventfds = kvm_check_many_ioeventfds();
1641 cpu_interrupt_handler = kvm_handle_interrupt;
1643 return 0;
1645 err:
1646 assert(ret < 0);
1647 if (s->vmfd >= 0) {
1648 close(s->vmfd);
1650 if (s->fd != -1) {
1651 close(s->fd);
1653 g_free(s->slots);
1655 return ret;
1658 void kvm_set_sigmask_len(KVMState *s, unsigned int sigmask_len)
1660 s->sigmask_len = sigmask_len;
1663 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
1664 uint32_t count)
1666 int i;
1667 uint8_t *ptr = data;
1669 for (i = 0; i < count; i++) {
1670 address_space_rw(&address_space_io, port, ptr, size,
1671 direction == KVM_EXIT_IO_OUT);
1672 ptr += size;
1676 static int kvm_handle_internal_error(CPUState *cpu, struct kvm_run *run)
1678 fprintf(stderr, "KVM internal error. Suberror: %d\n",
1679 run->internal.suberror);
1681 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
1682 int i;
1684 for (i = 0; i < run->internal.ndata; ++i) {
1685 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
1686 i, (uint64_t)run->internal.data[i]);
1689 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
1690 fprintf(stderr, "emulation failure\n");
1691 if (!kvm_arch_stop_on_emulation_error(cpu)) {
1692 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1693 return EXCP_INTERRUPT;
1696 /* FIXME: Should trigger a qmp message to let management know
1697 * something went wrong.
1699 return -1;
1702 void kvm_flush_coalesced_mmio_buffer(void)
1704 KVMState *s = kvm_state;
1706 if (s->coalesced_flush_in_progress) {
1707 return;
1710 s->coalesced_flush_in_progress = true;
1712 if (s->coalesced_mmio_ring) {
1713 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
1714 while (ring->first != ring->last) {
1715 struct kvm_coalesced_mmio *ent;
1717 ent = &ring->coalesced_mmio[ring->first];
1719 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
1720 smp_wmb();
1721 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
1725 s->coalesced_flush_in_progress = false;
1728 static void do_kvm_cpu_synchronize_state(void *arg)
1730 CPUState *cpu = arg;
1732 if (!cpu->kvm_vcpu_dirty) {
1733 kvm_arch_get_registers(cpu);
1734 cpu->kvm_vcpu_dirty = true;
1738 void kvm_cpu_synchronize_state(CPUState *cpu)
1740 if (!cpu->kvm_vcpu_dirty) {
1741 run_on_cpu(cpu, do_kvm_cpu_synchronize_state, cpu);
1745 static void do_kvm_cpu_synchronize_post_reset(void *arg)
1747 CPUState *cpu = arg;
1749 kvm_arch_put_registers(cpu, KVM_PUT_RESET_STATE);
1750 cpu->kvm_vcpu_dirty = false;
1753 void kvm_cpu_synchronize_post_reset(CPUState *cpu)
1755 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_reset, cpu);
1758 static void do_kvm_cpu_synchronize_post_init(void *arg)
1760 CPUState *cpu = arg;
1762 kvm_arch_put_registers(cpu, KVM_PUT_FULL_STATE);
1763 cpu->kvm_vcpu_dirty = false;
1766 void kvm_cpu_synchronize_post_init(CPUState *cpu)
1768 run_on_cpu(cpu, do_kvm_cpu_synchronize_post_init, cpu);
1771 void kvm_cpu_clean_state(CPUState *cpu)
1773 cpu->kvm_vcpu_dirty = false;
1776 int kvm_cpu_exec(CPUState *cpu)
1778 struct kvm_run *run = cpu->kvm_run;
1779 int ret, run_ret;
1781 DPRINTF("kvm_cpu_exec()\n");
1783 if (kvm_arch_process_async_events(cpu)) {
1784 cpu->exit_request = 0;
1785 return EXCP_HLT;
1788 do {
1789 if (cpu->kvm_vcpu_dirty) {
1790 kvm_arch_put_registers(cpu, KVM_PUT_RUNTIME_STATE);
1791 cpu->kvm_vcpu_dirty = false;
1794 kvm_arch_pre_run(cpu, run);
1795 if (cpu->exit_request) {
1796 DPRINTF("interrupt exit requested\n");
1798 * KVM requires us to reenter the kernel after IO exits to complete
1799 * instruction emulation. This self-signal will ensure that we
1800 * leave ASAP again.
1802 qemu_cpu_kick_self();
1804 qemu_mutex_unlock_iothread();
1806 run_ret = kvm_vcpu_ioctl(cpu, KVM_RUN, 0);
1808 qemu_mutex_lock_iothread();
1809 kvm_arch_post_run(cpu, run);
1811 if (run_ret < 0) {
1812 if (run_ret == -EINTR || run_ret == -EAGAIN) {
1813 DPRINTF("io window exit\n");
1814 ret = EXCP_INTERRUPT;
1815 break;
1817 fprintf(stderr, "error: kvm run failed %s\n",
1818 strerror(-run_ret));
1819 ret = -1;
1820 break;
1823 trace_kvm_run_exit(cpu->cpu_index, run->exit_reason);
1824 switch (run->exit_reason) {
1825 case KVM_EXIT_IO:
1826 DPRINTF("handle_io\n");
1827 kvm_handle_io(run->io.port,
1828 (uint8_t *)run + run->io.data_offset,
1829 run->io.direction,
1830 run->io.size,
1831 run->io.count);
1832 ret = 0;
1833 break;
1834 case KVM_EXIT_MMIO:
1835 DPRINTF("handle_mmio\n");
1836 cpu_physical_memory_rw(run->mmio.phys_addr,
1837 run->mmio.data,
1838 run->mmio.len,
1839 run->mmio.is_write);
1840 ret = 0;
1841 break;
1842 case KVM_EXIT_IRQ_WINDOW_OPEN:
1843 DPRINTF("irq_window_open\n");
1844 ret = EXCP_INTERRUPT;
1845 break;
1846 case KVM_EXIT_SHUTDOWN:
1847 DPRINTF("shutdown\n");
1848 qemu_system_reset_request();
1849 ret = EXCP_INTERRUPT;
1850 break;
1851 case KVM_EXIT_UNKNOWN:
1852 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1853 (uint64_t)run->hw.hardware_exit_reason);
1854 ret = -1;
1855 break;
1856 case KVM_EXIT_INTERNAL_ERROR:
1857 ret = kvm_handle_internal_error(cpu, run);
1858 break;
1859 case KVM_EXIT_SYSTEM_EVENT:
1860 switch (run->system_event.type) {
1861 case KVM_SYSTEM_EVENT_SHUTDOWN:
1862 qemu_system_shutdown_request();
1863 ret = EXCP_INTERRUPT;
1864 break;
1865 case KVM_SYSTEM_EVENT_RESET:
1866 qemu_system_reset_request();
1867 ret = EXCP_INTERRUPT;
1868 break;
1869 default:
1870 DPRINTF("kvm_arch_handle_exit\n");
1871 ret = kvm_arch_handle_exit(cpu, run);
1872 break;
1874 break;
1875 default:
1876 DPRINTF("kvm_arch_handle_exit\n");
1877 ret = kvm_arch_handle_exit(cpu, run);
1878 break;
1880 } while (ret == 0);
1882 if (ret < 0) {
1883 cpu_dump_state(cpu, stderr, fprintf, CPU_DUMP_CODE);
1884 vm_stop(RUN_STATE_INTERNAL_ERROR);
1887 cpu->exit_request = 0;
1888 return ret;
1891 int kvm_ioctl(KVMState *s, int type, ...)
1893 int ret;
1894 void *arg;
1895 va_list ap;
1897 va_start(ap, type);
1898 arg = va_arg(ap, void *);
1899 va_end(ap);
1901 trace_kvm_ioctl(type, arg);
1902 ret = ioctl(s->fd, type, arg);
1903 if (ret == -1) {
1904 ret = -errno;
1906 return ret;
1909 int kvm_vm_ioctl(KVMState *s, int type, ...)
1911 int ret;
1912 void *arg;
1913 va_list ap;
1915 va_start(ap, type);
1916 arg = va_arg(ap, void *);
1917 va_end(ap);
1919 trace_kvm_vm_ioctl(type, arg);
1920 ret = ioctl(s->vmfd, type, arg);
1921 if (ret == -1) {
1922 ret = -errno;
1924 return ret;
1927 int kvm_vcpu_ioctl(CPUState *cpu, int type, ...)
1929 int ret;
1930 void *arg;
1931 va_list ap;
1933 va_start(ap, type);
1934 arg = va_arg(ap, void *);
1935 va_end(ap);
1937 trace_kvm_vcpu_ioctl(cpu->cpu_index, type, arg);
1938 ret = ioctl(cpu->kvm_fd, type, arg);
1939 if (ret == -1) {
1940 ret = -errno;
1942 return ret;
1945 int kvm_device_ioctl(int fd, int type, ...)
1947 int ret;
1948 void *arg;
1949 va_list ap;
1951 va_start(ap, type);
1952 arg = va_arg(ap, void *);
1953 va_end(ap);
1955 trace_kvm_device_ioctl(fd, type, arg);
1956 ret = ioctl(fd, type, arg);
1957 if (ret == -1) {
1958 ret = -errno;
1960 return ret;
1963 int kvm_vm_check_attr(KVMState *s, uint32_t group, uint64_t attr)
1965 int ret;
1966 struct kvm_device_attr attribute = {
1967 .group = group,
1968 .attr = attr,
1971 if (!kvm_vm_attributes_allowed) {
1972 return 0;
1975 ret = kvm_vm_ioctl(s, KVM_HAS_DEVICE_ATTR, &attribute);
1976 /* kvm returns 0 on success for HAS_DEVICE_ATTR */
1977 return ret ? 0 : 1;
1980 int kvm_has_sync_mmu(void)
1982 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1985 int kvm_has_vcpu_events(void)
1987 return kvm_state->vcpu_events;
1990 int kvm_has_robust_singlestep(void)
1992 return kvm_state->robust_singlestep;
1995 int kvm_has_debugregs(void)
1997 return kvm_state->debugregs;
2000 int kvm_has_xsave(void)
2002 return kvm_state->xsave;
2005 int kvm_has_xcrs(void)
2007 return kvm_state->xcrs;
2010 int kvm_has_pit_state2(void)
2012 return kvm_state->pit_state2;
2015 int kvm_has_many_ioeventfds(void)
2017 if (!kvm_enabled()) {
2018 return 0;
2020 return kvm_state->many_ioeventfds;
2023 int kvm_has_gsi_routing(void)
2025 #ifdef KVM_CAP_IRQ_ROUTING
2026 return kvm_check_extension(kvm_state, KVM_CAP_IRQ_ROUTING);
2027 #else
2028 return false;
2029 #endif
2032 int kvm_has_intx_set_mask(void)
2034 return kvm_state->intx_set_mask;
2037 void kvm_setup_guest_memory(void *start, size_t size)
2039 if (!kvm_has_sync_mmu()) {
2040 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
2042 if (ret) {
2043 perror("qemu_madvise");
2044 fprintf(stderr,
2045 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
2046 exit(1);
2051 #ifdef KVM_CAP_SET_GUEST_DEBUG
2052 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *cpu,
2053 target_ulong pc)
2055 struct kvm_sw_breakpoint *bp;
2057 QTAILQ_FOREACH(bp, &cpu->kvm_state->kvm_sw_breakpoints, entry) {
2058 if (bp->pc == pc) {
2059 return bp;
2062 return NULL;
2065 int kvm_sw_breakpoints_active(CPUState *cpu)
2067 return !QTAILQ_EMPTY(&cpu->kvm_state->kvm_sw_breakpoints);
2070 struct kvm_set_guest_debug_data {
2071 struct kvm_guest_debug dbg;
2072 CPUState *cpu;
2073 int err;
2076 static void kvm_invoke_set_guest_debug(void *data)
2078 struct kvm_set_guest_debug_data *dbg_data = data;
2080 dbg_data->err = kvm_vcpu_ioctl(dbg_data->cpu, KVM_SET_GUEST_DEBUG,
2081 &dbg_data->dbg);
2084 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2086 struct kvm_set_guest_debug_data data;
2088 data.dbg.control = reinject_trap;
2090 if (cpu->singlestep_enabled) {
2091 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
2093 kvm_arch_update_guest_debug(cpu, &data.dbg);
2094 data.cpu = cpu;
2096 run_on_cpu(cpu, kvm_invoke_set_guest_debug, &data);
2097 return data.err;
2100 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2101 target_ulong len, int type)
2103 struct kvm_sw_breakpoint *bp;
2104 int err;
2106 if (type == GDB_BREAKPOINT_SW) {
2107 bp = kvm_find_sw_breakpoint(cpu, addr);
2108 if (bp) {
2109 bp->use_count++;
2110 return 0;
2113 bp = g_malloc(sizeof(struct kvm_sw_breakpoint));
2114 bp->pc = addr;
2115 bp->use_count = 1;
2116 err = kvm_arch_insert_sw_breakpoint(cpu, bp);
2117 if (err) {
2118 g_free(bp);
2119 return err;
2122 QTAILQ_INSERT_HEAD(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2123 } else {
2124 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
2125 if (err) {
2126 return err;
2130 CPU_FOREACH(cpu) {
2131 err = kvm_update_guest_debug(cpu, 0);
2132 if (err) {
2133 return err;
2136 return 0;
2139 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2140 target_ulong len, int type)
2142 struct kvm_sw_breakpoint *bp;
2143 int err;
2145 if (type == GDB_BREAKPOINT_SW) {
2146 bp = kvm_find_sw_breakpoint(cpu, addr);
2147 if (!bp) {
2148 return -ENOENT;
2151 if (bp->use_count > 1) {
2152 bp->use_count--;
2153 return 0;
2156 err = kvm_arch_remove_sw_breakpoint(cpu, bp);
2157 if (err) {
2158 return err;
2161 QTAILQ_REMOVE(&cpu->kvm_state->kvm_sw_breakpoints, bp, entry);
2162 g_free(bp);
2163 } else {
2164 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
2165 if (err) {
2166 return err;
2170 CPU_FOREACH(cpu) {
2171 err = kvm_update_guest_debug(cpu, 0);
2172 if (err) {
2173 return err;
2176 return 0;
2179 void kvm_remove_all_breakpoints(CPUState *cpu)
2181 struct kvm_sw_breakpoint *bp, *next;
2182 KVMState *s = cpu->kvm_state;
2183 CPUState *tmpcpu;
2185 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
2186 if (kvm_arch_remove_sw_breakpoint(cpu, bp) != 0) {
2187 /* Try harder to find a CPU that currently sees the breakpoint. */
2188 CPU_FOREACH(tmpcpu) {
2189 if (kvm_arch_remove_sw_breakpoint(tmpcpu, bp) == 0) {
2190 break;
2194 QTAILQ_REMOVE(&s->kvm_sw_breakpoints, bp, entry);
2195 g_free(bp);
2197 kvm_arch_remove_all_hw_breakpoints();
2199 CPU_FOREACH(cpu) {
2200 kvm_update_guest_debug(cpu, 0);
2204 #else /* !KVM_CAP_SET_GUEST_DEBUG */
2206 int kvm_update_guest_debug(CPUState *cpu, unsigned long reinject_trap)
2208 return -EINVAL;
2211 int kvm_insert_breakpoint(CPUState *cpu, target_ulong addr,
2212 target_ulong len, int type)
2214 return -EINVAL;
2217 int kvm_remove_breakpoint(CPUState *cpu, target_ulong addr,
2218 target_ulong len, int type)
2220 return -EINVAL;
2223 void kvm_remove_all_breakpoints(CPUState *cpu)
2226 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
2228 int kvm_set_signal_mask(CPUState *cpu, const sigset_t *sigset)
2230 KVMState *s = kvm_state;
2231 struct kvm_signal_mask *sigmask;
2232 int r;
2234 if (!sigset) {
2235 return kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, NULL);
2238 sigmask = g_malloc(sizeof(*sigmask) + sizeof(*sigset));
2240 sigmask->len = s->sigmask_len;
2241 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
2242 r = kvm_vcpu_ioctl(cpu, KVM_SET_SIGNAL_MASK, sigmask);
2243 g_free(sigmask);
2245 return r;
2247 int kvm_on_sigbus_vcpu(CPUState *cpu, int code, void *addr)
2249 return kvm_arch_on_sigbus_vcpu(cpu, code, addr);
2252 int kvm_on_sigbus(int code, void *addr)
2254 return kvm_arch_on_sigbus(code, addr);
2257 int kvm_create_device(KVMState *s, uint64_t type, bool test)
2259 int ret;
2260 struct kvm_create_device create_dev;
2262 create_dev.type = type;
2263 create_dev.fd = -1;
2264 create_dev.flags = test ? KVM_CREATE_DEVICE_TEST : 0;
2266 if (!kvm_check_extension(s, KVM_CAP_DEVICE_CTRL)) {
2267 return -ENOTSUP;
2270 ret = kvm_vm_ioctl(s, KVM_CREATE_DEVICE, &create_dev);
2271 if (ret) {
2272 return ret;
2275 return test ? 0 : create_dev.fd;
2278 int kvm_set_one_reg(CPUState *cs, uint64_t id, void *source)
2280 struct kvm_one_reg reg;
2281 int r;
2283 reg.id = id;
2284 reg.addr = (uintptr_t) source;
2285 r = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
2286 if (r) {
2287 trace_kvm_failed_reg_set(id, strerror(r));
2289 return r;
2292 int kvm_get_one_reg(CPUState *cs, uint64_t id, void *target)
2294 struct kvm_one_reg reg;
2295 int r;
2297 reg.id = id;
2298 reg.addr = (uintptr_t) target;
2299 r = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
2300 if (r) {
2301 trace_kvm_failed_reg_get(id, strerror(r));
2303 return r;
2306 static void kvm_accel_class_init(ObjectClass *oc, void *data)
2308 AccelClass *ac = ACCEL_CLASS(oc);
2309 ac->name = "KVM";
2310 ac->init_machine = kvm_init;
2311 ac->allowed = &kvm_allowed;
2314 static const TypeInfo kvm_accel_type = {
2315 .name = TYPE_KVM_ACCEL,
2316 .parent = TYPE_ACCEL,
2317 .class_init = kvm_accel_class_init,
2318 .instance_size = sizeof(KVMState),
2321 static void kvm_type_init(void)
2323 type_register_static(&kvm_accel_type);
2326 type_init(kvm_type_init);