2 * ARM implementation of KVM hooks
4 * Copyright Christoffer Dall 2009-2010
6 * This work is licensed under the terms of the GNU GPL, version 2 or later.
7 * See the COPYING file in the top-level directory.
11 #include "qemu/osdep.h"
12 #include <sys/ioctl.h>
14 #include <linux/kvm.h>
16 #include "qemu-common.h"
17 #include "qemu/timer.h"
18 #include "qemu/error-report.h"
19 #include "sysemu/sysemu.h"
20 #include "sysemu/kvm.h"
21 #include "sysemu/kvm_int.h"
25 #include "internals.h"
26 #include "hw/arm/arm.h"
27 #include "hw/pci/pci.h"
28 #include "exec/memattrs.h"
29 #include "exec/address-spaces.h"
30 #include "hw/boards.h"
33 const KVMCapabilityInfo kvm_arch_required_capabilities
[] = {
37 static bool cap_has_mp_state
;
38 static bool cap_has_inject_serror_esr
;
40 static ARMHostCPUFeatures arm_host_cpu_features
;
42 int kvm_arm_vcpu_init(CPUState
*cs
)
44 ARMCPU
*cpu
= ARM_CPU(cs
);
45 struct kvm_vcpu_init init
;
47 init
.target
= cpu
->kvm_target
;
48 memcpy(init
.features
, cpu
->kvm_init_features
, sizeof(init
.features
));
50 return kvm_vcpu_ioctl(cs
, KVM_ARM_VCPU_INIT
, &init
);
53 void kvm_arm_init_serror_injection(CPUState
*cs
)
55 cap_has_inject_serror_esr
= kvm_check_extension(cs
->kvm_state
,
56 KVM_CAP_ARM_INJECT_SERROR_ESR
);
59 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try
,
61 struct kvm_vcpu_init
*init
)
63 int ret
, kvmfd
= -1, vmfd
= -1, cpufd
= -1;
65 kvmfd
= qemu_open("/dev/kvm", O_RDWR
);
69 vmfd
= ioctl(kvmfd
, KVM_CREATE_VM
, 0);
73 cpufd
= ioctl(vmfd
, KVM_CREATE_VCPU
, 0);
79 /* Caller doesn't want the VCPU to be initialized, so skip it */
83 ret
= ioctl(vmfd
, KVM_ARM_PREFERRED_TARGET
, init
);
85 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
89 } else if (cpus_to_try
) {
90 /* Old kernel which doesn't know about the
91 * PREFERRED_TARGET ioctl: we know it will only support
92 * creating one kind of guest CPU which is its preferred
95 while (*cpus_to_try
!= QEMU_KVM_ARM_TARGET_NONE
) {
96 init
->target
= *cpus_to_try
++;
97 memset(init
->features
, 0, sizeof(init
->features
));
98 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
107 /* Treat a NULL cpus_to_try argument the same as an empty
108 * list, which means we will fail the call since this must
109 * be an old kernel which doesn't support PREFERRED_TARGET.
135 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray
)
139 for (i
= 2; i
>= 0; i
--) {
144 void kvm_arm_set_cpu_features_from_host(ARMCPU
*cpu
)
146 CPUARMState
*env
= &cpu
->env
;
148 if (!arm_host_cpu_features
.dtb_compatible
) {
149 if (!kvm_enabled() ||
150 !kvm_arm_get_host_cpu_features(&arm_host_cpu_features
)) {
151 /* We can't report this error yet, so flag that we need to
152 * in arm_cpu_realizefn().
154 cpu
->kvm_target
= QEMU_KVM_ARM_TARGET_NONE
;
155 cpu
->host_cpu_probe_failed
= true;
160 cpu
->kvm_target
= arm_host_cpu_features
.target
;
161 cpu
->dtb_compatible
= arm_host_cpu_features
.dtb_compatible
;
162 cpu
->isar
= arm_host_cpu_features
.isar
;
163 env
->features
= arm_host_cpu_features
.features
;
166 int kvm_arm_get_max_vm_ipa_size(MachineState
*ms
)
168 KVMState
*s
= KVM_STATE(ms
->accelerator
);
171 ret
= kvm_check_extension(s
, KVM_CAP_ARM_VM_IPA_SIZE
);
172 return ret
> 0 ? ret
: 40;
175 int kvm_arch_init(MachineState
*ms
, KVMState
*s
)
177 /* For ARM interrupt delivery is always asynchronous,
178 * whether we are using an in-kernel VGIC or not.
180 kvm_async_interrupts_allowed
= true;
183 * PSCI wakes up secondary cores, so we always need to
184 * have vCPUs waiting in kernel space
186 kvm_halt_in_kernel_allowed
= true;
188 cap_has_mp_state
= kvm_check_extension(s
, KVM_CAP_MP_STATE
);
193 unsigned long kvm_arch_vcpu_id(CPUState
*cpu
)
195 return cpu
->cpu_index
;
198 /* We track all the KVM devices which need their memory addresses
199 * passing to the kernel in a list of these structures.
200 * When board init is complete we run through the list and
201 * tell the kernel the base addresses of the memory regions.
202 * We use a MemoryListener to track mapping and unmapping of
203 * the regions during board creation, so the board models don't
204 * need to do anything special for the KVM case.
206 * Sometimes the address must be OR'ed with some other fields
207 * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
208 * @kda_addr_ormask aims at storing the value of those fields.
210 typedef struct KVMDevice
{
211 struct kvm_arm_device_addr kda
;
212 struct kvm_device_attr kdattr
;
213 uint64_t kda_addr_ormask
;
215 QSLIST_ENTRY(KVMDevice
) entries
;
219 static QSLIST_HEAD(, KVMDevice
) kvm_devices_head
;
221 static void kvm_arm_devlistener_add(MemoryListener
*listener
,
222 MemoryRegionSection
*section
)
226 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
227 if (section
->mr
== kd
->mr
) {
228 kd
->kda
.addr
= section
->offset_within_address_space
;
233 static void kvm_arm_devlistener_del(MemoryListener
*listener
,
234 MemoryRegionSection
*section
)
238 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
239 if (section
->mr
== kd
->mr
) {
245 static MemoryListener devlistener
= {
246 .region_add
= kvm_arm_devlistener_add
,
247 .region_del
= kvm_arm_devlistener_del
,
250 static void kvm_arm_set_device_addr(KVMDevice
*kd
)
252 struct kvm_device_attr
*attr
= &kd
->kdattr
;
255 /* If the device control API is available and we have a device fd on the
256 * KVMDevice struct, let's use the newer API
258 if (kd
->dev_fd
>= 0) {
259 uint64_t addr
= kd
->kda
.addr
;
261 addr
|= kd
->kda_addr_ormask
;
262 attr
->addr
= (uintptr_t)&addr
;
263 ret
= kvm_device_ioctl(kd
->dev_fd
, KVM_SET_DEVICE_ATTR
, attr
);
265 ret
= kvm_vm_ioctl(kvm_state
, KVM_ARM_SET_DEVICE_ADDR
, &kd
->kda
);
269 fprintf(stderr
, "Failed to set device address: %s\n",
275 static void kvm_arm_machine_init_done(Notifier
*notifier
, void *data
)
279 QSLIST_FOREACH_SAFE(kd
, &kvm_devices_head
, entries
, tkd
) {
280 if (kd
->kda
.addr
!= -1) {
281 kvm_arm_set_device_addr(kd
);
283 memory_region_unref(kd
->mr
);
284 QSLIST_REMOVE_HEAD(&kvm_devices_head
, entries
);
287 memory_listener_unregister(&devlistener
);
290 static Notifier notify
= {
291 .notify
= kvm_arm_machine_init_done
,
294 void kvm_arm_register_device(MemoryRegion
*mr
, uint64_t devid
, uint64_t group
,
295 uint64_t attr
, int dev_fd
, uint64_t addr_ormask
)
299 if (!kvm_irqchip_in_kernel()) {
303 if (QSLIST_EMPTY(&kvm_devices_head
)) {
304 memory_listener_register(&devlistener
, &address_space_memory
);
305 qemu_add_machine_init_done_notifier(¬ify
);
307 kd
= g_new0(KVMDevice
, 1);
311 kd
->kdattr
.flags
= 0;
312 kd
->kdattr
.group
= group
;
313 kd
->kdattr
.attr
= attr
;
315 kd
->kda_addr_ormask
= addr_ormask
;
316 QSLIST_INSERT_HEAD(&kvm_devices_head
, kd
, entries
);
317 memory_region_ref(kd
->mr
);
320 static int compare_u64(const void *a
, const void *b
)
322 if (*(uint64_t *)a
> *(uint64_t *)b
) {
325 if (*(uint64_t *)a
< *(uint64_t *)b
) {
331 /* Initialize the ARMCPU cpreg list according to the kernel's
332 * definition of what CPU registers it knows about (and throw away
333 * the previous TCG-created cpreg list).
335 int kvm_arm_init_cpreg_list(ARMCPU
*cpu
)
337 struct kvm_reg_list rl
;
338 struct kvm_reg_list
*rlp
;
339 int i
, ret
, arraylen
;
340 CPUState
*cs
= CPU(cpu
);
343 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, &rl
);
347 rlp
= g_malloc(sizeof(struct kvm_reg_list
) + rl
.n
* sizeof(uint64_t));
349 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, rlp
);
353 /* Sort the list we get back from the kernel, since cpreg_tuples
354 * must be in strictly ascending order.
356 qsort(&rlp
->reg
, rlp
->n
, sizeof(rlp
->reg
[0]), compare_u64
);
358 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
359 if (!kvm_arm_reg_syncs_via_cpreg_list(rlp
->reg
[i
])) {
362 switch (rlp
->reg
[i
] & KVM_REG_SIZE_MASK
) {
363 case KVM_REG_SIZE_U32
:
364 case KVM_REG_SIZE_U64
:
367 fprintf(stderr
, "Can't handle size of register in kernel list\n");
375 cpu
->cpreg_indexes
= g_renew(uint64_t, cpu
->cpreg_indexes
, arraylen
);
376 cpu
->cpreg_values
= g_renew(uint64_t, cpu
->cpreg_values
, arraylen
);
377 cpu
->cpreg_vmstate_indexes
= g_renew(uint64_t, cpu
->cpreg_vmstate_indexes
,
379 cpu
->cpreg_vmstate_values
= g_renew(uint64_t, cpu
->cpreg_vmstate_values
,
381 cpu
->cpreg_array_len
= arraylen
;
382 cpu
->cpreg_vmstate_array_len
= arraylen
;
384 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
385 uint64_t regidx
= rlp
->reg
[i
];
386 if (!kvm_arm_reg_syncs_via_cpreg_list(regidx
)) {
389 cpu
->cpreg_indexes
[arraylen
] = regidx
;
392 assert(cpu
->cpreg_array_len
== arraylen
);
394 if (!write_kvmstate_to_list(cpu
)) {
395 /* Shouldn't happen unless kernel is inconsistent about
396 * what registers exist.
398 fprintf(stderr
, "Initial read of kernel register state failed\n");
408 bool write_kvmstate_to_list(ARMCPU
*cpu
)
410 CPUState
*cs
= CPU(cpu
);
414 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
415 struct kvm_one_reg r
;
416 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
422 switch (regidx
& KVM_REG_SIZE_MASK
) {
423 case KVM_REG_SIZE_U32
:
424 r
.addr
= (uintptr_t)&v32
;
425 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
427 cpu
->cpreg_values
[i
] = v32
;
430 case KVM_REG_SIZE_U64
:
431 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
432 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
444 bool write_list_to_kvmstate(ARMCPU
*cpu
, int level
)
446 CPUState
*cs
= CPU(cpu
);
450 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
451 struct kvm_one_reg r
;
452 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
456 if (kvm_arm_cpreg_level(regidx
) > level
) {
461 switch (regidx
& KVM_REG_SIZE_MASK
) {
462 case KVM_REG_SIZE_U32
:
463 v32
= cpu
->cpreg_values
[i
];
464 r
.addr
= (uintptr_t)&v32
;
466 case KVM_REG_SIZE_U64
:
467 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
472 ret
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, &r
);
474 /* We might fail for "unknown register" and also for
475 * "you tried to set a register which is constant with
476 * a different value from what it actually contains".
484 void kvm_arm_reset_vcpu(ARMCPU
*cpu
)
488 /* Re-init VCPU so that all registers are set to
489 * their respective reset values.
491 ret
= kvm_arm_vcpu_init(CPU(cpu
));
493 fprintf(stderr
, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret
));
496 if (!write_kvmstate_to_list(cpu
)) {
497 fprintf(stderr
, "write_kvmstate_to_list failed\n");
501 * Sync the reset values also into the CPUState. This is necessary
502 * because the next thing we do will be a kvm_arch_put_registers()
503 * which will update the list values from the CPUState before copying
504 * the list values back to KVM. It's OK to ignore failure returns here
505 * for the same reason we do so in kvm_arch_get_registers().
507 write_list_to_cpustate(cpu
);
511 * Update KVM's MP_STATE based on what QEMU thinks it is
513 int kvm_arm_sync_mpstate_to_kvm(ARMCPU
*cpu
)
515 if (cap_has_mp_state
) {
516 struct kvm_mp_state mp_state
= {
517 .mp_state
= (cpu
->power_state
== PSCI_OFF
) ?
518 KVM_MP_STATE_STOPPED
: KVM_MP_STATE_RUNNABLE
520 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_MP_STATE
, &mp_state
);
522 fprintf(stderr
, "%s: failed to set MP_STATE %d/%s\n",
523 __func__
, ret
, strerror(-ret
));
532 * Sync the KVM MP_STATE into QEMU
534 int kvm_arm_sync_mpstate_to_qemu(ARMCPU
*cpu
)
536 if (cap_has_mp_state
) {
537 struct kvm_mp_state mp_state
;
538 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_MP_STATE
, &mp_state
);
540 fprintf(stderr
, "%s: failed to get MP_STATE %d/%s\n",
541 __func__
, ret
, strerror(-ret
));
544 cpu
->power_state
= (mp_state
.mp_state
== KVM_MP_STATE_STOPPED
) ?
551 int kvm_put_vcpu_events(ARMCPU
*cpu
)
553 CPUARMState
*env
= &cpu
->env
;
554 struct kvm_vcpu_events events
;
557 if (!kvm_has_vcpu_events()) {
561 memset(&events
, 0, sizeof(events
));
562 events
.exception
.serror_pending
= env
->serror
.pending
;
564 /* Inject SError to guest with specified syndrome if host kernel
565 * supports it, otherwise inject SError without syndrome.
567 if (cap_has_inject_serror_esr
) {
568 events
.exception
.serror_has_esr
= env
->serror
.has_esr
;
569 events
.exception
.serror_esr
= env
->serror
.esr
;
572 ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_VCPU_EVENTS
, &events
);
574 error_report("failed to put vcpu events");
580 int kvm_get_vcpu_events(ARMCPU
*cpu
)
582 CPUARMState
*env
= &cpu
->env
;
583 struct kvm_vcpu_events events
;
586 if (!kvm_has_vcpu_events()) {
590 memset(&events
, 0, sizeof(events
));
591 ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_VCPU_EVENTS
, &events
);
593 error_report("failed to get vcpu events");
597 env
->serror
.pending
= events
.exception
.serror_pending
;
598 env
->serror
.has_esr
= events
.exception
.serror_has_esr
;
599 env
->serror
.esr
= events
.exception
.serror_esr
;
604 void kvm_arch_pre_run(CPUState
*cs
, struct kvm_run
*run
)
608 MemTxAttrs
kvm_arch_post_run(CPUState
*cs
, struct kvm_run
*run
)
611 uint32_t switched_level
;
613 if (kvm_irqchip_in_kernel()) {
615 * We only need to sync timer states with user-space interrupt
616 * controllers, so return early and save cycles if we don't.
618 return MEMTXATTRS_UNSPECIFIED
;
623 /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
624 if (run
->s
.regs
.device_irq_level
!= cpu
->device_irq_level
) {
625 switched_level
= cpu
->device_irq_level
^ run
->s
.regs
.device_irq_level
;
627 qemu_mutex_lock_iothread();
629 if (switched_level
& KVM_ARM_DEV_EL1_VTIMER
) {
630 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_VIRT
],
631 !!(run
->s
.regs
.device_irq_level
&
632 KVM_ARM_DEV_EL1_VTIMER
));
633 switched_level
&= ~KVM_ARM_DEV_EL1_VTIMER
;
636 if (switched_level
& KVM_ARM_DEV_EL1_PTIMER
) {
637 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_PHYS
],
638 !!(run
->s
.regs
.device_irq_level
&
639 KVM_ARM_DEV_EL1_PTIMER
));
640 switched_level
&= ~KVM_ARM_DEV_EL1_PTIMER
;
643 if (switched_level
& KVM_ARM_DEV_PMU
) {
644 qemu_set_irq(cpu
->pmu_interrupt
,
645 !!(run
->s
.regs
.device_irq_level
& KVM_ARM_DEV_PMU
));
646 switched_level
&= ~KVM_ARM_DEV_PMU
;
649 if (switched_level
) {
650 qemu_log_mask(LOG_UNIMP
, "%s: unhandled in-kernel device IRQ %x\n",
651 __func__
, switched_level
);
654 /* We also mark unknown levels as processed to not waste cycles */
655 cpu
->device_irq_level
= run
->s
.regs
.device_irq_level
;
656 qemu_mutex_unlock_iothread();
659 return MEMTXATTRS_UNSPECIFIED
;
663 int kvm_arch_handle_exit(CPUState
*cs
, struct kvm_run
*run
)
667 switch (run
->exit_reason
) {
669 if (kvm_arm_handle_debug(cs
, &run
->debug
.arch
)) {
671 } /* otherwise return to guest */
674 qemu_log_mask(LOG_UNIMP
, "%s: un-handled exit reason %d\n",
675 __func__
, run
->exit_reason
);
681 bool kvm_arch_stop_on_emulation_error(CPUState
*cs
)
686 int kvm_arch_process_async_events(CPUState
*cs
)
691 /* The #ifdef protections are until 32bit headers are imported and can
692 * be removed once both 32 and 64 bit reach feature parity.
694 void kvm_arch_update_guest_debug(CPUState
*cs
, struct kvm_guest_debug
*dbg
)
696 #ifdef KVM_GUESTDBG_USE_SW_BP
697 if (kvm_sw_breakpoints_active(cs
)) {
698 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_SW_BP
;
701 #ifdef KVM_GUESTDBG_USE_HW
702 if (kvm_arm_hw_debug_active(cs
)) {
703 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_HW
;
704 kvm_arm_copy_hw_debug_data(&dbg
->arch
);
709 void kvm_arch_init_irq_routing(KVMState
*s
)
713 int kvm_arch_irqchip_create(MachineState
*ms
, KVMState
*s
)
715 if (machine_kernel_irqchip_split(ms
)) {
716 perror("-machine kernel_irqchip=split is not supported on ARM.");
720 /* If we can create the VGIC using the newer device control API, we
721 * let the device do this when it initializes itself, otherwise we
722 * fall back to the old API */
723 return kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
);
726 int kvm_arm_vgic_probe(void)
728 if (kvm_create_device(kvm_state
,
729 KVM_DEV_TYPE_ARM_VGIC_V3
, true) == 0) {
731 } else if (kvm_create_device(kvm_state
,
732 KVM_DEV_TYPE_ARM_VGIC_V2
, true) == 0) {
739 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry
*route
,
740 uint64_t address
, uint32_t data
, PCIDevice
*dev
)
742 AddressSpace
*as
= pci_device_iommu_address_space(dev
);
743 hwaddr xlat
, len
, doorbell_gpa
;
744 MemoryRegionSection mrs
;
748 if (as
== &address_space_memory
) {
752 /* MSI doorbell address is translated by an IOMMU */
755 mr
= address_space_translate(as
, address
, &xlat
, &len
, true,
756 MEMTXATTRS_UNSPECIFIED
);
760 mrs
= memory_region_find(mr
, xlat
, 1);
765 doorbell_gpa
= mrs
.offset_within_address_space
;
766 memory_region_unref(mrs
.mr
);
768 route
->u
.msi
.address_lo
= doorbell_gpa
;
769 route
->u
.msi
.address_hi
= doorbell_gpa
>> 32;
771 trace_kvm_arm_fixup_msi_route(address
, doorbell_gpa
);
780 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry
*route
,
781 int vector
, PCIDevice
*dev
)
786 int kvm_arch_release_virq_post(int virq
)
791 int kvm_arch_msi_data_to_gsi(uint32_t data
)
793 return (data
- 32) & 0xffff;