ditto hw/xilinx_uartlite.c
[qemu/aliguori.git] / kvm-all.c
blobb9c172b66437e2665108232613099a78df3d468d
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 "sysemu.h"
26 #include "hw/hw.h"
27 #include "gdbstub.h"
28 #include "kvm.h"
29 #include "bswap.h"
31 /* This check must be after config-host.h is included */
32 #ifdef CONFIG_EVENTFD
33 #include <sys/eventfd.h>
34 #endif
36 /* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
37 #define PAGE_SIZE TARGET_PAGE_SIZE
39 //#define DEBUG_KVM
41 #ifdef DEBUG_KVM
42 #define DPRINTF(fmt, ...) \
43 do { fprintf(stderr, fmt, ## __VA_ARGS__); } while (0)
44 #else
45 #define DPRINTF(fmt, ...) \
46 do { } while (0)
47 #endif
49 typedef struct KVMSlot
51 target_phys_addr_t start_addr;
52 ram_addr_t memory_size;
53 ram_addr_t phys_offset;
54 int slot;
55 int flags;
56 } KVMSlot;
58 typedef struct kvm_dirty_log KVMDirtyLog;
60 struct KVMState
62 KVMSlot slots[32];
63 int fd;
64 int vmfd;
65 int coalesced_mmio;
66 struct kvm_coalesced_mmio_ring *coalesced_mmio_ring;
67 int broken_set_mem_region;
68 int migration_log;
69 int vcpu_events;
70 int robust_singlestep;
71 int debugregs;
72 #ifdef KVM_CAP_SET_GUEST_DEBUG
73 struct kvm_sw_breakpoint_head kvm_sw_breakpoints;
74 #endif
75 int irqchip_in_kernel;
76 int pit_in_kernel;
77 int xsave, xcrs;
78 int many_ioeventfds;
81 KVMState *kvm_state;
83 static const KVMCapabilityInfo kvm_required_capabilites[] = {
84 KVM_CAP_INFO(USER_MEMORY),
85 KVM_CAP_INFO(DESTROY_MEMORY_REGION_WORKS),
86 KVM_CAP_LAST_INFO
89 static KVMSlot *kvm_alloc_slot(KVMState *s)
91 int i;
93 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
94 if (s->slots[i].memory_size == 0) {
95 return &s->slots[i];
99 fprintf(stderr, "%s: no free slot available\n", __func__);
100 abort();
103 static KVMSlot *kvm_lookup_matching_slot(KVMState *s,
104 target_phys_addr_t start_addr,
105 target_phys_addr_t end_addr)
107 int i;
109 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
110 KVMSlot *mem = &s->slots[i];
112 if (start_addr == mem->start_addr &&
113 end_addr == mem->start_addr + mem->memory_size) {
114 return mem;
118 return NULL;
122 * Find overlapping slot with lowest start address
124 static KVMSlot *kvm_lookup_overlapping_slot(KVMState *s,
125 target_phys_addr_t start_addr,
126 target_phys_addr_t end_addr)
128 KVMSlot *found = NULL;
129 int i;
131 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
132 KVMSlot *mem = &s->slots[i];
134 if (mem->memory_size == 0 ||
135 (found && found->start_addr < mem->start_addr)) {
136 continue;
139 if (end_addr > mem->start_addr &&
140 start_addr < mem->start_addr + mem->memory_size) {
141 found = mem;
145 return found;
148 int kvm_physical_memory_addr_from_ram(KVMState *s, ram_addr_t ram_addr,
149 target_phys_addr_t *phys_addr)
151 int i;
153 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
154 KVMSlot *mem = &s->slots[i];
156 if (ram_addr >= mem->phys_offset &&
157 ram_addr < mem->phys_offset + mem->memory_size) {
158 *phys_addr = mem->start_addr + (ram_addr - mem->phys_offset);
159 return 1;
163 return 0;
166 static int kvm_set_user_memory_region(KVMState *s, KVMSlot *slot)
168 struct kvm_userspace_memory_region mem;
170 mem.slot = slot->slot;
171 mem.guest_phys_addr = slot->start_addr;
172 mem.memory_size = slot->memory_size;
173 mem.userspace_addr = (unsigned long)qemu_safe_ram_ptr(slot->phys_offset);
174 mem.flags = slot->flags;
175 if (s->migration_log) {
176 mem.flags |= KVM_MEM_LOG_DIRTY_PAGES;
178 return kvm_vm_ioctl(s, KVM_SET_USER_MEMORY_REGION, &mem);
181 static void kvm_reset_vcpu(void *opaque)
183 CPUState *env = opaque;
185 kvm_arch_reset_vcpu(env);
188 int kvm_irqchip_in_kernel(void)
190 return kvm_state->irqchip_in_kernel;
193 int kvm_pit_in_kernel(void)
195 return kvm_state->pit_in_kernel;
198 int kvm_init_vcpu(CPUState *env)
200 KVMState *s = kvm_state;
201 long mmap_size;
202 int ret;
204 DPRINTF("kvm_init_vcpu\n");
206 ret = kvm_vm_ioctl(s, KVM_CREATE_VCPU, env->cpu_index);
207 if (ret < 0) {
208 DPRINTF("kvm_create_vcpu failed\n");
209 goto err;
212 env->kvm_fd = ret;
213 env->kvm_state = s;
214 env->kvm_vcpu_dirty = 1;
216 mmap_size = kvm_ioctl(s, KVM_GET_VCPU_MMAP_SIZE, 0);
217 if (mmap_size < 0) {
218 ret = mmap_size;
219 DPRINTF("KVM_GET_VCPU_MMAP_SIZE failed\n");
220 goto err;
223 env->kvm_run = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, MAP_SHARED,
224 env->kvm_fd, 0);
225 if (env->kvm_run == MAP_FAILED) {
226 ret = -errno;
227 DPRINTF("mmap'ing vcpu state failed\n");
228 goto err;
231 if (s->coalesced_mmio && !s->coalesced_mmio_ring) {
232 s->coalesced_mmio_ring =
233 (void *)env->kvm_run + s->coalesced_mmio * PAGE_SIZE;
236 ret = kvm_arch_init_vcpu(env);
237 if (ret == 0) {
238 qemu_register_reset(kvm_reset_vcpu, env);
239 kvm_arch_reset_vcpu(env);
241 err:
242 return ret;
246 * dirty pages logging control
249 static int kvm_mem_flags(KVMState *s, bool log_dirty)
251 return log_dirty ? KVM_MEM_LOG_DIRTY_PAGES : 0;
254 static int kvm_slot_dirty_pages_log_change(KVMSlot *mem, bool log_dirty)
256 KVMState *s = kvm_state;
257 int flags, mask = KVM_MEM_LOG_DIRTY_PAGES;
258 int old_flags;
260 old_flags = mem->flags;
262 flags = (mem->flags & ~mask) | kvm_mem_flags(s, log_dirty);
263 mem->flags = flags;
265 /* If nothing changed effectively, no need to issue ioctl */
266 if (s->migration_log) {
267 flags |= KVM_MEM_LOG_DIRTY_PAGES;
270 if (flags == old_flags) {
271 return 0;
274 return kvm_set_user_memory_region(s, mem);
277 static int kvm_dirty_pages_log_change(target_phys_addr_t phys_addr,
278 ram_addr_t size, bool log_dirty)
280 KVMState *s = kvm_state;
281 KVMSlot *mem = kvm_lookup_matching_slot(s, phys_addr, phys_addr + size);
283 if (mem == NULL) {
284 fprintf(stderr, "BUG: %s: invalid parameters " TARGET_FMT_plx "-"
285 TARGET_FMT_plx "\n", __func__, phys_addr,
286 (target_phys_addr_t)(phys_addr + size - 1));
287 return -EINVAL;
289 return kvm_slot_dirty_pages_log_change(mem, log_dirty);
292 static int kvm_log_start(CPUPhysMemoryClient *client,
293 target_phys_addr_t phys_addr, ram_addr_t size)
295 return kvm_dirty_pages_log_change(phys_addr, size, true);
298 static int kvm_log_stop(CPUPhysMemoryClient *client,
299 target_phys_addr_t phys_addr, ram_addr_t size)
301 return kvm_dirty_pages_log_change(phys_addr, size, false);
304 static int kvm_set_migration_log(int enable)
306 KVMState *s = kvm_state;
307 KVMSlot *mem;
308 int i, err;
310 s->migration_log = enable;
312 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
313 mem = &s->slots[i];
315 if (!mem->memory_size) {
316 continue;
318 if (!!(mem->flags & KVM_MEM_LOG_DIRTY_PAGES) == enable) {
319 continue;
321 err = kvm_set_user_memory_region(s, mem);
322 if (err) {
323 return err;
326 return 0;
329 /* get kvm's dirty pages bitmap and update qemu's */
330 static int kvm_get_dirty_pages_log_range(unsigned long start_addr,
331 unsigned long *bitmap,
332 unsigned long offset,
333 unsigned long mem_size)
335 unsigned int i, j;
336 unsigned long page_number, addr, addr1, c;
337 ram_addr_t ram_addr;
338 unsigned int len = ((mem_size / TARGET_PAGE_SIZE) + HOST_LONG_BITS - 1) /
339 HOST_LONG_BITS;
342 * bitmap-traveling is faster than memory-traveling (for addr...)
343 * especially when most of the memory is not dirty.
345 for (i = 0; i < len; i++) {
346 if (bitmap[i] != 0) {
347 c = leul_to_cpu(bitmap[i]);
348 do {
349 j = ffsl(c) - 1;
350 c &= ~(1ul << j);
351 page_number = i * HOST_LONG_BITS + j;
352 addr1 = page_number * TARGET_PAGE_SIZE;
353 addr = offset + addr1;
354 ram_addr = cpu_get_physical_page_desc(addr);
355 cpu_physical_memory_set_dirty(ram_addr);
356 } while (c != 0);
359 return 0;
362 #define ALIGN(x, y) (((x)+(y)-1) & ~((y)-1))
365 * kvm_physical_sync_dirty_bitmap - Grab dirty bitmap from kernel space
366 * This function updates qemu's dirty bitmap using cpu_physical_memory_set_dirty().
367 * This means all bits are set to dirty.
369 * @start_add: start of logged region.
370 * @end_addr: end of logged region.
372 static int kvm_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
373 target_phys_addr_t end_addr)
375 KVMState *s = kvm_state;
376 unsigned long size, allocated_size = 0;
377 KVMDirtyLog d;
378 KVMSlot *mem;
379 int ret = 0;
381 d.dirty_bitmap = NULL;
382 while (start_addr < end_addr) {
383 mem = kvm_lookup_overlapping_slot(s, start_addr, end_addr);
384 if (mem == NULL) {
385 break;
388 /* XXX bad kernel interface alert
389 * For dirty bitmap, kernel allocates array of size aligned to
390 * bits-per-long. But for case when the kernel is 64bits and
391 * the userspace is 32bits, userspace can't align to the same
392 * bits-per-long, since sizeof(long) is different between kernel
393 * and user space. This way, userspace will provide buffer which
394 * may be 4 bytes less than the kernel will use, resulting in
395 * userspace memory corruption (which is not detectable by valgrind
396 * too, in most cases).
397 * So for now, let's align to 64 instead of HOST_LONG_BITS here, in
398 * a hope that sizeof(long) wont become >8 any time soon.
400 size = ALIGN(((mem->memory_size) >> TARGET_PAGE_BITS),
401 /*HOST_LONG_BITS*/ 64) / 8;
402 if (!d.dirty_bitmap) {
403 d.dirty_bitmap = qemu_malloc(size);
404 } else if (size > allocated_size) {
405 d.dirty_bitmap = qemu_realloc(d.dirty_bitmap, size);
407 allocated_size = size;
408 memset(d.dirty_bitmap, 0, allocated_size);
410 d.slot = mem->slot;
412 if (kvm_vm_ioctl(s, KVM_GET_DIRTY_LOG, &d) == -1) {
413 DPRINTF("ioctl failed %d\n", errno);
414 ret = -1;
415 break;
418 kvm_get_dirty_pages_log_range(mem->start_addr, d.dirty_bitmap,
419 mem->start_addr, mem->memory_size);
420 start_addr = mem->start_addr + mem->memory_size;
422 qemu_free(d.dirty_bitmap);
424 return ret;
427 int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
429 int ret = -ENOSYS;
430 KVMState *s = kvm_state;
432 if (s->coalesced_mmio) {
433 struct kvm_coalesced_mmio_zone zone;
435 zone.addr = start;
436 zone.size = size;
438 ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
441 return ret;
444 int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
446 int ret = -ENOSYS;
447 KVMState *s = kvm_state;
449 if (s->coalesced_mmio) {
450 struct kvm_coalesced_mmio_zone zone;
452 zone.addr = start;
453 zone.size = size;
455 ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
458 return ret;
461 int kvm_check_extension(KVMState *s, unsigned int extension)
463 int ret;
465 ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, extension);
466 if (ret < 0) {
467 ret = 0;
470 return ret;
473 static int kvm_check_many_ioeventfds(void)
475 /* Userspace can use ioeventfd for io notification. This requires a host
476 * that supports eventfd(2) and an I/O thread; since eventfd does not
477 * support SIGIO it cannot interrupt the vcpu.
479 * Older kernels have a 6 device limit on the KVM io bus. Find out so we
480 * can avoid creating too many ioeventfds.
482 #if defined(CONFIG_EVENTFD) && defined(CONFIG_IOTHREAD)
483 int ioeventfds[7];
484 int i, ret = 0;
485 for (i = 0; i < ARRAY_SIZE(ioeventfds); i++) {
486 ioeventfds[i] = eventfd(0, EFD_CLOEXEC);
487 if (ioeventfds[i] < 0) {
488 break;
490 ret = kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, true);
491 if (ret < 0) {
492 close(ioeventfds[i]);
493 break;
497 /* Decide whether many devices are supported or not */
498 ret = i == ARRAY_SIZE(ioeventfds);
500 while (i-- > 0) {
501 kvm_set_ioeventfd_pio_word(ioeventfds[i], 0, i, false);
502 close(ioeventfds[i]);
504 return ret;
505 #else
506 return 0;
507 #endif
510 static const KVMCapabilityInfo *
511 kvm_check_extension_list(KVMState *s, const KVMCapabilityInfo *list)
513 while (list->name) {
514 if (!kvm_check_extension(s, list->value)) {
515 return list;
517 list++;
519 return NULL;
522 static void kvm_set_phys_mem(target_phys_addr_t start_addr, ram_addr_t size,
523 ram_addr_t phys_offset, bool log_dirty)
525 KVMState *s = kvm_state;
526 ram_addr_t flags = phys_offset & ~TARGET_PAGE_MASK;
527 KVMSlot *mem, old;
528 int err;
530 /* kvm works in page size chunks, but the function may be called
531 with sub-page size and unaligned start address. */
532 size = TARGET_PAGE_ALIGN(size);
533 start_addr = TARGET_PAGE_ALIGN(start_addr);
535 /* KVM does not support read-only slots */
536 phys_offset &= ~IO_MEM_ROM;
538 while (1) {
539 mem = kvm_lookup_overlapping_slot(s, start_addr, start_addr + size);
540 if (!mem) {
541 break;
544 if (flags < IO_MEM_UNASSIGNED && start_addr >= mem->start_addr &&
545 (start_addr + size <= mem->start_addr + mem->memory_size) &&
546 (phys_offset - start_addr == mem->phys_offset - mem->start_addr)) {
547 /* The new slot fits into the existing one and comes with
548 * identical parameters - update flags and done. */
549 kvm_slot_dirty_pages_log_change(mem, log_dirty);
550 return;
553 old = *mem;
555 /* unregister the overlapping slot */
556 mem->memory_size = 0;
557 err = kvm_set_user_memory_region(s, mem);
558 if (err) {
559 fprintf(stderr, "%s: error unregistering overlapping slot: %s\n",
560 __func__, strerror(-err));
561 abort();
564 /* Workaround for older KVM versions: we can't join slots, even not by
565 * unregistering the previous ones and then registering the larger
566 * slot. We have to maintain the existing fragmentation. Sigh.
568 * This workaround assumes that the new slot starts at the same
569 * address as the first existing one. If not or if some overlapping
570 * slot comes around later, we will fail (not seen in practice so far)
571 * - and actually require a recent KVM version. */
572 if (s->broken_set_mem_region &&
573 old.start_addr == start_addr && old.memory_size < size &&
574 flags < IO_MEM_UNASSIGNED) {
575 mem = kvm_alloc_slot(s);
576 mem->memory_size = old.memory_size;
577 mem->start_addr = old.start_addr;
578 mem->phys_offset = old.phys_offset;
579 mem->flags = kvm_mem_flags(s, log_dirty);
581 err = kvm_set_user_memory_region(s, mem);
582 if (err) {
583 fprintf(stderr, "%s: error updating slot: %s\n", __func__,
584 strerror(-err));
585 abort();
588 start_addr += old.memory_size;
589 phys_offset += old.memory_size;
590 size -= old.memory_size;
591 continue;
594 /* register prefix slot */
595 if (old.start_addr < start_addr) {
596 mem = kvm_alloc_slot(s);
597 mem->memory_size = start_addr - old.start_addr;
598 mem->start_addr = old.start_addr;
599 mem->phys_offset = old.phys_offset;
600 mem->flags = kvm_mem_flags(s, log_dirty);
602 err = kvm_set_user_memory_region(s, mem);
603 if (err) {
604 fprintf(stderr, "%s: error registering prefix slot: %s\n",
605 __func__, strerror(-err));
606 #ifdef TARGET_PPC
607 fprintf(stderr, "%s: This is probably because your kernel's " \
608 "PAGE_SIZE is too big. Please try to use 4k " \
609 "PAGE_SIZE!\n", __func__);
610 #endif
611 abort();
615 /* register suffix slot */
616 if (old.start_addr + old.memory_size > start_addr + size) {
617 ram_addr_t size_delta;
619 mem = kvm_alloc_slot(s);
620 mem->start_addr = start_addr + size;
621 size_delta = mem->start_addr - old.start_addr;
622 mem->memory_size = old.memory_size - size_delta;
623 mem->phys_offset = old.phys_offset + size_delta;
624 mem->flags = kvm_mem_flags(s, log_dirty);
626 err = kvm_set_user_memory_region(s, mem);
627 if (err) {
628 fprintf(stderr, "%s: error registering suffix slot: %s\n",
629 __func__, strerror(-err));
630 abort();
635 /* in case the KVM bug workaround already "consumed" the new slot */
636 if (!size) {
637 return;
639 /* KVM does not need to know about this memory */
640 if (flags >= IO_MEM_UNASSIGNED) {
641 return;
643 mem = kvm_alloc_slot(s);
644 mem->memory_size = size;
645 mem->start_addr = start_addr;
646 mem->phys_offset = phys_offset;
647 mem->flags = kvm_mem_flags(s, log_dirty);
649 err = kvm_set_user_memory_region(s, mem);
650 if (err) {
651 fprintf(stderr, "%s: error registering slot: %s\n", __func__,
652 strerror(-err));
653 abort();
657 static void kvm_client_set_memory(struct CPUPhysMemoryClient *client,
658 target_phys_addr_t start_addr,
659 ram_addr_t size, ram_addr_t phys_offset,
660 bool log_dirty)
662 kvm_set_phys_mem(start_addr, size, phys_offset, log_dirty);
665 static int kvm_client_sync_dirty_bitmap(struct CPUPhysMemoryClient *client,
666 target_phys_addr_t start_addr,
667 target_phys_addr_t end_addr)
669 return kvm_physical_sync_dirty_bitmap(start_addr, end_addr);
672 static int kvm_client_migration_log(struct CPUPhysMemoryClient *client,
673 int enable)
675 return kvm_set_migration_log(enable);
678 static CPUPhysMemoryClient kvm_cpu_phys_memory_client = {
679 .set_memory = kvm_client_set_memory,
680 .sync_dirty_bitmap = kvm_client_sync_dirty_bitmap,
681 .migration_log = kvm_client_migration_log,
682 .log_start = kvm_log_start,
683 .log_stop = kvm_log_stop,
686 static void kvm_handle_interrupt(CPUState *env, int mask)
688 env->interrupt_request |= mask;
690 if (!qemu_cpu_is_self(env)) {
691 qemu_cpu_kick(env);
695 int kvm_init(void)
697 static const char upgrade_note[] =
698 "Please upgrade to at least kernel 2.6.29 or recent kvm-kmod\n"
699 "(see http://sourceforge.net/projects/kvm).\n";
700 KVMState *s;
701 const KVMCapabilityInfo *missing_cap;
702 int ret;
703 int i;
705 s = qemu_mallocz(sizeof(KVMState));
707 #ifdef KVM_CAP_SET_GUEST_DEBUG
708 QTAILQ_INIT(&s->kvm_sw_breakpoints);
709 #endif
710 for (i = 0; i < ARRAY_SIZE(s->slots); i++) {
711 s->slots[i].slot = i;
713 s->vmfd = -1;
714 s->fd = qemu_open("/dev/kvm", O_RDWR);
715 if (s->fd == -1) {
716 fprintf(stderr, "Could not access KVM kernel module: %m\n");
717 ret = -errno;
718 goto err;
721 ret = kvm_ioctl(s, KVM_GET_API_VERSION, 0);
722 if (ret < KVM_API_VERSION) {
723 if (ret > 0) {
724 ret = -EINVAL;
726 fprintf(stderr, "kvm version too old\n");
727 goto err;
730 if (ret > KVM_API_VERSION) {
731 ret = -EINVAL;
732 fprintf(stderr, "kvm version not supported\n");
733 goto err;
736 s->vmfd = kvm_ioctl(s, KVM_CREATE_VM, 0);
737 if (s->vmfd < 0) {
738 #ifdef TARGET_S390X
739 fprintf(stderr, "Please add the 'switch_amode' kernel parameter to "
740 "your host kernel command line\n");
741 #endif
742 goto err;
745 missing_cap = kvm_check_extension_list(s, kvm_required_capabilites);
746 if (!missing_cap) {
747 missing_cap =
748 kvm_check_extension_list(s, kvm_arch_required_capabilities);
750 if (missing_cap) {
751 ret = -EINVAL;
752 fprintf(stderr, "kvm does not support %s\n%s",
753 missing_cap->name, upgrade_note);
754 goto err;
757 s->coalesced_mmio = kvm_check_extension(s, KVM_CAP_COALESCED_MMIO);
759 s->broken_set_mem_region = 1;
760 ret = kvm_check_extension(s, KVM_CAP_JOIN_MEMORY_REGIONS_WORKS);
761 if (ret > 0) {
762 s->broken_set_mem_region = 0;
765 #ifdef KVM_CAP_VCPU_EVENTS
766 s->vcpu_events = kvm_check_extension(s, KVM_CAP_VCPU_EVENTS);
767 #endif
769 s->robust_singlestep =
770 kvm_check_extension(s, KVM_CAP_X86_ROBUST_SINGLESTEP);
772 #ifdef KVM_CAP_DEBUGREGS
773 s->debugregs = kvm_check_extension(s, KVM_CAP_DEBUGREGS);
774 #endif
776 #ifdef KVM_CAP_XSAVE
777 s->xsave = kvm_check_extension(s, KVM_CAP_XSAVE);
778 #endif
780 #ifdef KVM_CAP_XCRS
781 s->xcrs = kvm_check_extension(s, KVM_CAP_XCRS);
782 #endif
784 ret = kvm_arch_init(s);
785 if (ret < 0) {
786 goto err;
789 kvm_state = s;
790 cpu_register_phys_memory_client(&kvm_cpu_phys_memory_client);
792 s->many_ioeventfds = kvm_check_many_ioeventfds();
794 cpu_interrupt_handler = kvm_handle_interrupt;
796 return 0;
798 err:
799 if (s) {
800 if (s->vmfd != -1) {
801 close(s->vmfd);
803 if (s->fd != -1) {
804 close(s->fd);
807 qemu_free(s);
809 return ret;
812 static void kvm_handle_io(uint16_t port, void *data, int direction, int size,
813 uint32_t count)
815 int i;
816 uint8_t *ptr = data;
818 for (i = 0; i < count; i++) {
819 if (direction == KVM_EXIT_IO_IN) {
820 switch (size) {
821 case 1:
822 stb_p(ptr, cpu_inb(port));
823 break;
824 case 2:
825 stw_p(ptr, cpu_inw(port));
826 break;
827 case 4:
828 stl_p(ptr, cpu_inl(port));
829 break;
831 } else {
832 switch (size) {
833 case 1:
834 cpu_outb(port, ldub_p(ptr));
835 break;
836 case 2:
837 cpu_outw(port, lduw_p(ptr));
838 break;
839 case 4:
840 cpu_outl(port, ldl_p(ptr));
841 break;
845 ptr += size;
849 static int kvm_handle_internal_error(CPUState *env, struct kvm_run *run)
851 fprintf(stderr, "KVM internal error.");
852 if (kvm_check_extension(kvm_state, KVM_CAP_INTERNAL_ERROR_DATA)) {
853 int i;
855 fprintf(stderr, " Suberror: %d\n", run->internal.suberror);
856 for (i = 0; i < run->internal.ndata; ++i) {
857 fprintf(stderr, "extra data[%d]: %"PRIx64"\n",
858 i, (uint64_t)run->internal.data[i]);
860 } else {
861 fprintf(stderr, "\n");
863 if (run->internal.suberror == KVM_INTERNAL_ERROR_EMULATION) {
864 fprintf(stderr, "emulation failure\n");
865 if (!kvm_arch_stop_on_emulation_error(env)) {
866 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
867 return EXCP_INTERRUPT;
870 /* FIXME: Should trigger a qmp message to let management know
871 * something went wrong.
873 return -1;
876 void kvm_flush_coalesced_mmio_buffer(void)
878 KVMState *s = kvm_state;
879 if (s->coalesced_mmio_ring) {
880 struct kvm_coalesced_mmio_ring *ring = s->coalesced_mmio_ring;
881 while (ring->first != ring->last) {
882 struct kvm_coalesced_mmio *ent;
884 ent = &ring->coalesced_mmio[ring->first];
886 cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
887 smp_wmb();
888 ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
893 static void do_kvm_cpu_synchronize_state(void *_env)
895 CPUState *env = _env;
897 if (!env->kvm_vcpu_dirty) {
898 kvm_arch_get_registers(env);
899 env->kvm_vcpu_dirty = 1;
903 void kvm_cpu_synchronize_state(CPUState *env)
905 if (!env->kvm_vcpu_dirty) {
906 run_on_cpu(env, do_kvm_cpu_synchronize_state, env);
910 void kvm_cpu_synchronize_post_reset(CPUState *env)
912 kvm_arch_put_registers(env, KVM_PUT_RESET_STATE);
913 env->kvm_vcpu_dirty = 0;
916 void kvm_cpu_synchronize_post_init(CPUState *env)
918 kvm_arch_put_registers(env, KVM_PUT_FULL_STATE);
919 env->kvm_vcpu_dirty = 0;
922 int kvm_cpu_exec(CPUState *env)
924 struct kvm_run *run = env->kvm_run;
925 int ret, run_ret;
927 DPRINTF("kvm_cpu_exec()\n");
929 if (kvm_arch_process_async_events(env)) {
930 env->exit_request = 0;
931 return EXCP_HLT;
934 cpu_single_env = env;
936 do {
937 if (env->kvm_vcpu_dirty) {
938 kvm_arch_put_registers(env, KVM_PUT_RUNTIME_STATE);
939 env->kvm_vcpu_dirty = 0;
942 kvm_arch_pre_run(env, run);
943 if (env->exit_request) {
944 DPRINTF("interrupt exit requested\n");
946 * KVM requires us to reenter the kernel after IO exits to complete
947 * instruction emulation. This self-signal will ensure that we
948 * leave ASAP again.
950 qemu_cpu_kick_self();
952 cpu_single_env = NULL;
953 qemu_mutex_unlock_iothread();
955 run_ret = kvm_vcpu_ioctl(env, KVM_RUN, 0);
957 qemu_mutex_lock_iothread();
958 cpu_single_env = env;
959 kvm_arch_post_run(env, run);
961 kvm_flush_coalesced_mmio_buffer();
963 if (run_ret < 0) {
964 if (run_ret == -EINTR || run_ret == -EAGAIN) {
965 DPRINTF("io window exit\n");
966 ret = EXCP_INTERRUPT;
967 break;
969 DPRINTF("kvm run failed %s\n", strerror(-run_ret));
970 abort();
973 switch (run->exit_reason) {
974 case KVM_EXIT_IO:
975 DPRINTF("handle_io\n");
976 kvm_handle_io(run->io.port,
977 (uint8_t *)run + run->io.data_offset,
978 run->io.direction,
979 run->io.size,
980 run->io.count);
981 ret = 0;
982 break;
983 case KVM_EXIT_MMIO:
984 DPRINTF("handle_mmio\n");
985 cpu_physical_memory_rw(run->mmio.phys_addr,
986 run->mmio.data,
987 run->mmio.len,
988 run->mmio.is_write);
989 ret = 0;
990 break;
991 case KVM_EXIT_IRQ_WINDOW_OPEN:
992 DPRINTF("irq_window_open\n");
993 ret = EXCP_INTERRUPT;
994 break;
995 case KVM_EXIT_SHUTDOWN:
996 DPRINTF("shutdown\n");
997 qemu_system_reset_request();
998 ret = EXCP_INTERRUPT;
999 break;
1000 case KVM_EXIT_UNKNOWN:
1001 fprintf(stderr, "KVM: unknown exit, hardware reason %" PRIx64 "\n",
1002 (uint64_t)run->hw.hardware_exit_reason);
1003 ret = -1;
1004 break;
1005 case KVM_EXIT_INTERNAL_ERROR:
1006 ret = kvm_handle_internal_error(env, run);
1007 break;
1008 default:
1009 DPRINTF("kvm_arch_handle_exit\n");
1010 ret = kvm_arch_handle_exit(env, run);
1011 break;
1013 } while (ret == 0);
1015 if (ret < 0) {
1016 cpu_dump_state(env, stderr, fprintf, CPU_DUMP_CODE);
1017 vm_stop(VMSTOP_PANIC);
1020 env->exit_request = 0;
1021 cpu_single_env = NULL;
1022 return ret;
1025 int kvm_ioctl(KVMState *s, int type, ...)
1027 int ret;
1028 void *arg;
1029 va_list ap;
1031 va_start(ap, type);
1032 arg = va_arg(ap, void *);
1033 va_end(ap);
1035 ret = ioctl(s->fd, type, arg);
1036 if (ret == -1) {
1037 ret = -errno;
1039 return ret;
1042 int kvm_vm_ioctl(KVMState *s, int type, ...)
1044 int ret;
1045 void *arg;
1046 va_list ap;
1048 va_start(ap, type);
1049 arg = va_arg(ap, void *);
1050 va_end(ap);
1052 ret = ioctl(s->vmfd, type, arg);
1053 if (ret == -1) {
1054 ret = -errno;
1056 return ret;
1059 int kvm_vcpu_ioctl(CPUState *env, int type, ...)
1061 int ret;
1062 void *arg;
1063 va_list ap;
1065 va_start(ap, type);
1066 arg = va_arg(ap, void *);
1067 va_end(ap);
1069 ret = ioctl(env->kvm_fd, type, arg);
1070 if (ret == -1) {
1071 ret = -errno;
1073 return ret;
1076 int kvm_has_sync_mmu(void)
1078 return kvm_check_extension(kvm_state, KVM_CAP_SYNC_MMU);
1081 int kvm_has_vcpu_events(void)
1083 return kvm_state->vcpu_events;
1086 int kvm_has_robust_singlestep(void)
1088 return kvm_state->robust_singlestep;
1091 int kvm_has_debugregs(void)
1093 return kvm_state->debugregs;
1096 int kvm_has_xsave(void)
1098 return kvm_state->xsave;
1101 int kvm_has_xcrs(void)
1103 return kvm_state->xcrs;
1106 int kvm_has_many_ioeventfds(void)
1108 if (!kvm_enabled()) {
1109 return 0;
1111 return kvm_state->many_ioeventfds;
1114 void kvm_setup_guest_memory(void *start, size_t size)
1116 if (!kvm_has_sync_mmu()) {
1117 int ret = qemu_madvise(start, size, QEMU_MADV_DONTFORK);
1119 if (ret) {
1120 perror("qemu_madvise");
1121 fprintf(stderr,
1122 "Need MADV_DONTFORK in absence of synchronous KVM MMU\n");
1123 exit(1);
1128 #ifdef KVM_CAP_SET_GUEST_DEBUG
1129 struct kvm_sw_breakpoint *kvm_find_sw_breakpoint(CPUState *env,
1130 target_ulong pc)
1132 struct kvm_sw_breakpoint *bp;
1134 QTAILQ_FOREACH(bp, &env->kvm_state->kvm_sw_breakpoints, entry) {
1135 if (bp->pc == pc) {
1136 return bp;
1139 return NULL;
1142 int kvm_sw_breakpoints_active(CPUState *env)
1144 return !QTAILQ_EMPTY(&env->kvm_state->kvm_sw_breakpoints);
1147 struct kvm_set_guest_debug_data {
1148 struct kvm_guest_debug dbg;
1149 CPUState *env;
1150 int err;
1153 static void kvm_invoke_set_guest_debug(void *data)
1155 struct kvm_set_guest_debug_data *dbg_data = data;
1156 CPUState *env = dbg_data->env;
1158 dbg_data->err = kvm_vcpu_ioctl(env, KVM_SET_GUEST_DEBUG, &dbg_data->dbg);
1161 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1163 struct kvm_set_guest_debug_data data;
1165 data.dbg.control = reinject_trap;
1167 if (env->singlestep_enabled) {
1168 data.dbg.control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_SINGLESTEP;
1170 kvm_arch_update_guest_debug(env, &data.dbg);
1171 data.env = env;
1173 run_on_cpu(env, kvm_invoke_set_guest_debug, &data);
1174 return data.err;
1177 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1178 target_ulong len, int type)
1180 struct kvm_sw_breakpoint *bp;
1181 CPUState *env;
1182 int err;
1184 if (type == GDB_BREAKPOINT_SW) {
1185 bp = kvm_find_sw_breakpoint(current_env, addr);
1186 if (bp) {
1187 bp->use_count++;
1188 return 0;
1191 bp = qemu_malloc(sizeof(struct kvm_sw_breakpoint));
1192 if (!bp) {
1193 return -ENOMEM;
1196 bp->pc = addr;
1197 bp->use_count = 1;
1198 err = kvm_arch_insert_sw_breakpoint(current_env, bp);
1199 if (err) {
1200 qemu_free(bp);
1201 return err;
1204 QTAILQ_INSERT_HEAD(&current_env->kvm_state->kvm_sw_breakpoints,
1205 bp, entry);
1206 } else {
1207 err = kvm_arch_insert_hw_breakpoint(addr, len, type);
1208 if (err) {
1209 return err;
1213 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1214 err = kvm_update_guest_debug(env, 0);
1215 if (err) {
1216 return err;
1219 return 0;
1222 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1223 target_ulong len, int type)
1225 struct kvm_sw_breakpoint *bp;
1226 CPUState *env;
1227 int err;
1229 if (type == GDB_BREAKPOINT_SW) {
1230 bp = kvm_find_sw_breakpoint(current_env, addr);
1231 if (!bp) {
1232 return -ENOENT;
1235 if (bp->use_count > 1) {
1236 bp->use_count--;
1237 return 0;
1240 err = kvm_arch_remove_sw_breakpoint(current_env, bp);
1241 if (err) {
1242 return err;
1245 QTAILQ_REMOVE(&current_env->kvm_state->kvm_sw_breakpoints, bp, entry);
1246 qemu_free(bp);
1247 } else {
1248 err = kvm_arch_remove_hw_breakpoint(addr, len, type);
1249 if (err) {
1250 return err;
1254 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1255 err = kvm_update_guest_debug(env, 0);
1256 if (err) {
1257 return err;
1260 return 0;
1263 void kvm_remove_all_breakpoints(CPUState *current_env)
1265 struct kvm_sw_breakpoint *bp, *next;
1266 KVMState *s = current_env->kvm_state;
1267 CPUState *env;
1269 QTAILQ_FOREACH_SAFE(bp, &s->kvm_sw_breakpoints, entry, next) {
1270 if (kvm_arch_remove_sw_breakpoint(current_env, bp) != 0) {
1271 /* Try harder to find a CPU that currently sees the breakpoint. */
1272 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1273 if (kvm_arch_remove_sw_breakpoint(env, bp) == 0) {
1274 break;
1279 kvm_arch_remove_all_hw_breakpoints();
1281 for (env = first_cpu; env != NULL; env = env->next_cpu) {
1282 kvm_update_guest_debug(env, 0);
1286 #else /* !KVM_CAP_SET_GUEST_DEBUG */
1288 int kvm_update_guest_debug(CPUState *env, unsigned long reinject_trap)
1290 return -EINVAL;
1293 int kvm_insert_breakpoint(CPUState *current_env, target_ulong addr,
1294 target_ulong len, int type)
1296 return -EINVAL;
1299 int kvm_remove_breakpoint(CPUState *current_env, target_ulong addr,
1300 target_ulong len, int type)
1302 return -EINVAL;
1305 void kvm_remove_all_breakpoints(CPUState *current_env)
1308 #endif /* !KVM_CAP_SET_GUEST_DEBUG */
1310 int kvm_set_signal_mask(CPUState *env, const sigset_t *sigset)
1312 struct kvm_signal_mask *sigmask;
1313 int r;
1315 if (!sigset) {
1316 return kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, NULL);
1319 sigmask = qemu_malloc(sizeof(*sigmask) + sizeof(*sigset));
1321 sigmask->len = 8;
1322 memcpy(sigmask->sigset, sigset, sizeof(*sigset));
1323 r = kvm_vcpu_ioctl(env, KVM_SET_SIGNAL_MASK, sigmask);
1324 qemu_free(sigmask);
1326 return r;
1329 int kvm_set_ioeventfd_mmio_long(int fd, uint32_t addr, uint32_t val, bool assign)
1331 int ret;
1332 struct kvm_ioeventfd iofd;
1334 iofd.datamatch = val;
1335 iofd.addr = addr;
1336 iofd.len = 4;
1337 iofd.flags = KVM_IOEVENTFD_FLAG_DATAMATCH;
1338 iofd.fd = fd;
1340 if (!kvm_enabled()) {
1341 return -ENOSYS;
1344 if (!assign) {
1345 iofd.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1348 ret = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &iofd);
1350 if (ret < 0) {
1351 return -errno;
1354 return 0;
1357 int kvm_set_ioeventfd_pio_word(int fd, uint16_t addr, uint16_t val, bool assign)
1359 struct kvm_ioeventfd kick = {
1360 .datamatch = val,
1361 .addr = addr,
1362 .len = 2,
1363 .flags = KVM_IOEVENTFD_FLAG_DATAMATCH | KVM_IOEVENTFD_FLAG_PIO,
1364 .fd = fd,
1366 int r;
1367 if (!kvm_enabled()) {
1368 return -ENOSYS;
1370 if (!assign) {
1371 kick.flags |= KVM_IOEVENTFD_FLAG_DEASSIGN;
1373 r = kvm_vm_ioctl(kvm_state, KVM_IOEVENTFD, &kick);
1374 if (r < 0) {
1375 return r;
1377 return 0;
1380 int kvm_on_sigbus_vcpu(CPUState *env, int code, void *addr)
1382 return kvm_arch_on_sigbus_vcpu(env, code, addr);
1385 int kvm_on_sigbus(int code, void *addr)
1387 return kvm_arch_on_sigbus(code, addr);