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/pci/pci.h"
27 #include "exec/memattrs.h"
28 #include "exec/address-spaces.h"
29 #include "hw/boards.h"
32 const KVMCapabilityInfo kvm_arch_required_capabilities
[] = {
36 static bool cap_has_mp_state
;
37 static bool cap_has_inject_serror_esr
;
39 static ARMHostCPUFeatures arm_host_cpu_features
;
41 int kvm_arm_vcpu_init(CPUState
*cs
)
43 ARMCPU
*cpu
= ARM_CPU(cs
);
44 struct kvm_vcpu_init init
;
46 init
.target
= cpu
->kvm_target
;
47 memcpy(init
.features
, cpu
->kvm_init_features
, sizeof(init
.features
));
49 return kvm_vcpu_ioctl(cs
, KVM_ARM_VCPU_INIT
, &init
);
52 void kvm_arm_init_serror_injection(CPUState
*cs
)
54 cap_has_inject_serror_esr
= kvm_check_extension(cs
->kvm_state
,
55 KVM_CAP_ARM_INJECT_SERROR_ESR
);
58 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try
,
60 struct kvm_vcpu_init
*init
)
62 int ret
, kvmfd
= -1, vmfd
= -1, cpufd
= -1;
64 kvmfd
= qemu_open("/dev/kvm", O_RDWR
);
68 vmfd
= ioctl(kvmfd
, KVM_CREATE_VM
, 0);
72 cpufd
= ioctl(vmfd
, KVM_CREATE_VCPU
, 0);
78 /* Caller doesn't want the VCPU to be initialized, so skip it */
82 ret
= ioctl(vmfd
, KVM_ARM_PREFERRED_TARGET
, init
);
84 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
88 } else if (cpus_to_try
) {
89 /* Old kernel which doesn't know about the
90 * PREFERRED_TARGET ioctl: we know it will only support
91 * creating one kind of guest CPU which is its preferred
94 while (*cpus_to_try
!= QEMU_KVM_ARM_TARGET_NONE
) {
95 init
->target
= *cpus_to_try
++;
96 memset(init
->features
, 0, sizeof(init
->features
));
97 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
106 /* Treat a NULL cpus_to_try argument the same as an empty
107 * list, which means we will fail the call since this must
108 * be an old kernel which doesn't support PREFERRED_TARGET.
134 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray
)
138 for (i
= 2; i
>= 0; i
--) {
143 void kvm_arm_set_cpu_features_from_host(ARMCPU
*cpu
)
145 CPUARMState
*env
= &cpu
->env
;
147 if (!arm_host_cpu_features
.dtb_compatible
) {
148 if (!kvm_enabled() ||
149 !kvm_arm_get_host_cpu_features(&arm_host_cpu_features
)) {
150 /* We can't report this error yet, so flag that we need to
151 * in arm_cpu_realizefn().
153 cpu
->kvm_target
= QEMU_KVM_ARM_TARGET_NONE
;
154 cpu
->host_cpu_probe_failed
= true;
159 cpu
->kvm_target
= arm_host_cpu_features
.target
;
160 cpu
->dtb_compatible
= arm_host_cpu_features
.dtb_compatible
;
161 cpu
->isar
= arm_host_cpu_features
.isar
;
162 env
->features
= arm_host_cpu_features
.features
;
165 int kvm_arm_get_max_vm_ipa_size(MachineState
*ms
)
167 KVMState
*s
= KVM_STATE(ms
->accelerator
);
170 ret
= kvm_check_extension(s
, KVM_CAP_ARM_VM_IPA_SIZE
);
171 return ret
> 0 ? ret
: 40;
174 int kvm_arch_init(MachineState
*ms
, KVMState
*s
)
176 /* For ARM interrupt delivery is always asynchronous,
177 * whether we are using an in-kernel VGIC or not.
179 kvm_async_interrupts_allowed
= true;
182 * PSCI wakes up secondary cores, so we always need to
183 * have vCPUs waiting in kernel space
185 kvm_halt_in_kernel_allowed
= true;
187 cap_has_mp_state
= kvm_check_extension(s
, KVM_CAP_MP_STATE
);
192 unsigned long kvm_arch_vcpu_id(CPUState
*cpu
)
194 return cpu
->cpu_index
;
197 /* We track all the KVM devices which need their memory addresses
198 * passing to the kernel in a list of these structures.
199 * When board init is complete we run through the list and
200 * tell the kernel the base addresses of the memory regions.
201 * We use a MemoryListener to track mapping and unmapping of
202 * the regions during board creation, so the board models don't
203 * need to do anything special for the KVM case.
205 * Sometimes the address must be OR'ed with some other fields
206 * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
207 * @kda_addr_ormask aims at storing the value of those fields.
209 typedef struct KVMDevice
{
210 struct kvm_arm_device_addr kda
;
211 struct kvm_device_attr kdattr
;
212 uint64_t kda_addr_ormask
;
214 QSLIST_ENTRY(KVMDevice
) entries
;
218 static QSLIST_HEAD(, KVMDevice
) kvm_devices_head
;
220 static void kvm_arm_devlistener_add(MemoryListener
*listener
,
221 MemoryRegionSection
*section
)
225 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
226 if (section
->mr
== kd
->mr
) {
227 kd
->kda
.addr
= section
->offset_within_address_space
;
232 static void kvm_arm_devlistener_del(MemoryListener
*listener
,
233 MemoryRegionSection
*section
)
237 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
238 if (section
->mr
== kd
->mr
) {
244 static MemoryListener devlistener
= {
245 .region_add
= kvm_arm_devlistener_add
,
246 .region_del
= kvm_arm_devlistener_del
,
249 static void kvm_arm_set_device_addr(KVMDevice
*kd
)
251 struct kvm_device_attr
*attr
= &kd
->kdattr
;
254 /* If the device control API is available and we have a device fd on the
255 * KVMDevice struct, let's use the newer API
257 if (kd
->dev_fd
>= 0) {
258 uint64_t addr
= kd
->kda
.addr
;
260 addr
|= kd
->kda_addr_ormask
;
261 attr
->addr
= (uintptr_t)&addr
;
262 ret
= kvm_device_ioctl(kd
->dev_fd
, KVM_SET_DEVICE_ATTR
, attr
);
264 ret
= kvm_vm_ioctl(kvm_state
, KVM_ARM_SET_DEVICE_ADDR
, &kd
->kda
);
268 fprintf(stderr
, "Failed to set device address: %s\n",
274 static void kvm_arm_machine_init_done(Notifier
*notifier
, void *data
)
278 QSLIST_FOREACH_SAFE(kd
, &kvm_devices_head
, entries
, tkd
) {
279 if (kd
->kda
.addr
!= -1) {
280 kvm_arm_set_device_addr(kd
);
282 memory_region_unref(kd
->mr
);
283 QSLIST_REMOVE_HEAD(&kvm_devices_head
, entries
);
286 memory_listener_unregister(&devlistener
);
289 static Notifier notify
= {
290 .notify
= kvm_arm_machine_init_done
,
293 void kvm_arm_register_device(MemoryRegion
*mr
, uint64_t devid
, uint64_t group
,
294 uint64_t attr
, int dev_fd
, uint64_t addr_ormask
)
298 if (!kvm_irqchip_in_kernel()) {
302 if (QSLIST_EMPTY(&kvm_devices_head
)) {
303 memory_listener_register(&devlistener
, &address_space_memory
);
304 qemu_add_machine_init_done_notifier(¬ify
);
306 kd
= g_new0(KVMDevice
, 1);
310 kd
->kdattr
.flags
= 0;
311 kd
->kdattr
.group
= group
;
312 kd
->kdattr
.attr
= attr
;
314 kd
->kda_addr_ormask
= addr_ormask
;
315 QSLIST_INSERT_HEAD(&kvm_devices_head
, kd
, entries
);
316 memory_region_ref(kd
->mr
);
319 static int compare_u64(const void *a
, const void *b
)
321 if (*(uint64_t *)a
> *(uint64_t *)b
) {
324 if (*(uint64_t *)a
< *(uint64_t *)b
) {
330 /* Initialize the ARMCPU cpreg list according to the kernel's
331 * definition of what CPU registers it knows about (and throw away
332 * the previous TCG-created cpreg list).
334 int kvm_arm_init_cpreg_list(ARMCPU
*cpu
)
336 struct kvm_reg_list rl
;
337 struct kvm_reg_list
*rlp
;
338 int i
, ret
, arraylen
;
339 CPUState
*cs
= CPU(cpu
);
342 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, &rl
);
346 rlp
= g_malloc(sizeof(struct kvm_reg_list
) + rl
.n
* sizeof(uint64_t));
348 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, rlp
);
352 /* Sort the list we get back from the kernel, since cpreg_tuples
353 * must be in strictly ascending order.
355 qsort(&rlp
->reg
, rlp
->n
, sizeof(rlp
->reg
[0]), compare_u64
);
357 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
358 if (!kvm_arm_reg_syncs_via_cpreg_list(rlp
->reg
[i
])) {
361 switch (rlp
->reg
[i
] & KVM_REG_SIZE_MASK
) {
362 case KVM_REG_SIZE_U32
:
363 case KVM_REG_SIZE_U64
:
366 fprintf(stderr
, "Can't handle size of register in kernel list\n");
374 cpu
->cpreg_indexes
= g_renew(uint64_t, cpu
->cpreg_indexes
, arraylen
);
375 cpu
->cpreg_values
= g_renew(uint64_t, cpu
->cpreg_values
, arraylen
);
376 cpu
->cpreg_vmstate_indexes
= g_renew(uint64_t, cpu
->cpreg_vmstate_indexes
,
378 cpu
->cpreg_vmstate_values
= g_renew(uint64_t, cpu
->cpreg_vmstate_values
,
380 cpu
->cpreg_array_len
= arraylen
;
381 cpu
->cpreg_vmstate_array_len
= arraylen
;
383 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
384 uint64_t regidx
= rlp
->reg
[i
];
385 if (!kvm_arm_reg_syncs_via_cpreg_list(regidx
)) {
388 cpu
->cpreg_indexes
[arraylen
] = regidx
;
391 assert(cpu
->cpreg_array_len
== arraylen
);
393 if (!write_kvmstate_to_list(cpu
)) {
394 /* Shouldn't happen unless kernel is inconsistent about
395 * what registers exist.
397 fprintf(stderr
, "Initial read of kernel register state failed\n");
407 bool write_kvmstate_to_list(ARMCPU
*cpu
)
409 CPUState
*cs
= CPU(cpu
);
413 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
414 struct kvm_one_reg r
;
415 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
421 switch (regidx
& KVM_REG_SIZE_MASK
) {
422 case KVM_REG_SIZE_U32
:
423 r
.addr
= (uintptr_t)&v32
;
424 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
426 cpu
->cpreg_values
[i
] = v32
;
429 case KVM_REG_SIZE_U64
:
430 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
431 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
443 bool write_list_to_kvmstate(ARMCPU
*cpu
, int level
)
445 CPUState
*cs
= CPU(cpu
);
449 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
450 struct kvm_one_reg r
;
451 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
455 if (kvm_arm_cpreg_level(regidx
) > level
) {
460 switch (regidx
& KVM_REG_SIZE_MASK
) {
461 case KVM_REG_SIZE_U32
:
462 v32
= cpu
->cpreg_values
[i
];
463 r
.addr
= (uintptr_t)&v32
;
465 case KVM_REG_SIZE_U64
:
466 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
471 ret
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, &r
);
473 /* We might fail for "unknown register" and also for
474 * "you tried to set a register which is constant with
475 * a different value from what it actually contains".
483 void kvm_arm_reset_vcpu(ARMCPU
*cpu
)
487 /* Re-init VCPU so that all registers are set to
488 * their respective reset values.
490 ret
= kvm_arm_vcpu_init(CPU(cpu
));
492 fprintf(stderr
, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret
));
495 if (!write_kvmstate_to_list(cpu
)) {
496 fprintf(stderr
, "write_kvmstate_to_list failed\n");
500 * Sync the reset values also into the CPUState. This is necessary
501 * because the next thing we do will be a kvm_arch_put_registers()
502 * which will update the list values from the CPUState before copying
503 * the list values back to KVM. It's OK to ignore failure returns here
504 * for the same reason we do so in kvm_arch_get_registers().
506 write_list_to_cpustate(cpu
);
510 * Update KVM's MP_STATE based on what QEMU thinks it is
512 int kvm_arm_sync_mpstate_to_kvm(ARMCPU
*cpu
)
514 if (cap_has_mp_state
) {
515 struct kvm_mp_state mp_state
= {
516 .mp_state
= (cpu
->power_state
== PSCI_OFF
) ?
517 KVM_MP_STATE_STOPPED
: KVM_MP_STATE_RUNNABLE
519 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_MP_STATE
, &mp_state
);
521 fprintf(stderr
, "%s: failed to set MP_STATE %d/%s\n",
522 __func__
, ret
, strerror(-ret
));
531 * Sync the KVM MP_STATE into QEMU
533 int kvm_arm_sync_mpstate_to_qemu(ARMCPU
*cpu
)
535 if (cap_has_mp_state
) {
536 struct kvm_mp_state mp_state
;
537 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_MP_STATE
, &mp_state
);
539 fprintf(stderr
, "%s: failed to get MP_STATE %d/%s\n",
540 __func__
, ret
, strerror(-ret
));
543 cpu
->power_state
= (mp_state
.mp_state
== KVM_MP_STATE_STOPPED
) ?
550 int kvm_put_vcpu_events(ARMCPU
*cpu
)
552 CPUARMState
*env
= &cpu
->env
;
553 struct kvm_vcpu_events events
;
556 if (!kvm_has_vcpu_events()) {
560 memset(&events
, 0, sizeof(events
));
561 events
.exception
.serror_pending
= env
->serror
.pending
;
563 /* Inject SError to guest with specified syndrome if host kernel
564 * supports it, otherwise inject SError without syndrome.
566 if (cap_has_inject_serror_esr
) {
567 events
.exception
.serror_has_esr
= env
->serror
.has_esr
;
568 events
.exception
.serror_esr
= env
->serror
.esr
;
571 ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_VCPU_EVENTS
, &events
);
573 error_report("failed to put vcpu events");
579 int kvm_get_vcpu_events(ARMCPU
*cpu
)
581 CPUARMState
*env
= &cpu
->env
;
582 struct kvm_vcpu_events events
;
585 if (!kvm_has_vcpu_events()) {
589 memset(&events
, 0, sizeof(events
));
590 ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_VCPU_EVENTS
, &events
);
592 error_report("failed to get vcpu events");
596 env
->serror
.pending
= events
.exception
.serror_pending
;
597 env
->serror
.has_esr
= events
.exception
.serror_has_esr
;
598 env
->serror
.esr
= events
.exception
.serror_esr
;
603 void kvm_arch_pre_run(CPUState
*cs
, struct kvm_run
*run
)
607 MemTxAttrs
kvm_arch_post_run(CPUState
*cs
, struct kvm_run
*run
)
610 uint32_t switched_level
;
612 if (kvm_irqchip_in_kernel()) {
614 * We only need to sync timer states with user-space interrupt
615 * controllers, so return early and save cycles if we don't.
617 return MEMTXATTRS_UNSPECIFIED
;
622 /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
623 if (run
->s
.regs
.device_irq_level
!= cpu
->device_irq_level
) {
624 switched_level
= cpu
->device_irq_level
^ run
->s
.regs
.device_irq_level
;
626 qemu_mutex_lock_iothread();
628 if (switched_level
& KVM_ARM_DEV_EL1_VTIMER
) {
629 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_VIRT
],
630 !!(run
->s
.regs
.device_irq_level
&
631 KVM_ARM_DEV_EL1_VTIMER
));
632 switched_level
&= ~KVM_ARM_DEV_EL1_VTIMER
;
635 if (switched_level
& KVM_ARM_DEV_EL1_PTIMER
) {
636 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_PHYS
],
637 !!(run
->s
.regs
.device_irq_level
&
638 KVM_ARM_DEV_EL1_PTIMER
));
639 switched_level
&= ~KVM_ARM_DEV_EL1_PTIMER
;
642 if (switched_level
& KVM_ARM_DEV_PMU
) {
643 qemu_set_irq(cpu
->pmu_interrupt
,
644 !!(run
->s
.regs
.device_irq_level
& KVM_ARM_DEV_PMU
));
645 switched_level
&= ~KVM_ARM_DEV_PMU
;
648 if (switched_level
) {
649 qemu_log_mask(LOG_UNIMP
, "%s: unhandled in-kernel device IRQ %x\n",
650 __func__
, switched_level
);
653 /* We also mark unknown levels as processed to not waste cycles */
654 cpu
->device_irq_level
= run
->s
.regs
.device_irq_level
;
655 qemu_mutex_unlock_iothread();
658 return MEMTXATTRS_UNSPECIFIED
;
662 int kvm_arch_handle_exit(CPUState
*cs
, struct kvm_run
*run
)
666 switch (run
->exit_reason
) {
668 if (kvm_arm_handle_debug(cs
, &run
->debug
.arch
)) {
670 } /* otherwise return to guest */
673 qemu_log_mask(LOG_UNIMP
, "%s: un-handled exit reason %d\n",
674 __func__
, run
->exit_reason
);
680 bool kvm_arch_stop_on_emulation_error(CPUState
*cs
)
685 int kvm_arch_process_async_events(CPUState
*cs
)
690 /* The #ifdef protections are until 32bit headers are imported and can
691 * be removed once both 32 and 64 bit reach feature parity.
693 void kvm_arch_update_guest_debug(CPUState
*cs
, struct kvm_guest_debug
*dbg
)
695 #ifdef KVM_GUESTDBG_USE_SW_BP
696 if (kvm_sw_breakpoints_active(cs
)) {
697 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_SW_BP
;
700 #ifdef KVM_GUESTDBG_USE_HW
701 if (kvm_arm_hw_debug_active(cs
)) {
702 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_HW
;
703 kvm_arm_copy_hw_debug_data(&dbg
->arch
);
708 void kvm_arch_init_irq_routing(KVMState
*s
)
712 int kvm_arch_irqchip_create(MachineState
*ms
, KVMState
*s
)
714 if (machine_kernel_irqchip_split(ms
)) {
715 perror("-machine kernel_irqchip=split is not supported on ARM.");
719 /* If we can create the VGIC using the newer device control API, we
720 * let the device do this when it initializes itself, otherwise we
721 * fall back to the old API */
722 return kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
);
725 int kvm_arm_vgic_probe(void)
727 if (kvm_create_device(kvm_state
,
728 KVM_DEV_TYPE_ARM_VGIC_V3
, true) == 0) {
730 } else if (kvm_create_device(kvm_state
,
731 KVM_DEV_TYPE_ARM_VGIC_V2
, true) == 0) {
738 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry
*route
,
739 uint64_t address
, uint32_t data
, PCIDevice
*dev
)
741 AddressSpace
*as
= pci_device_iommu_address_space(dev
);
742 hwaddr xlat
, len
, doorbell_gpa
;
743 MemoryRegionSection mrs
;
747 if (as
== &address_space_memory
) {
751 /* MSI doorbell address is translated by an IOMMU */
754 mr
= address_space_translate(as
, address
, &xlat
, &len
, true,
755 MEMTXATTRS_UNSPECIFIED
);
759 mrs
= memory_region_find(mr
, xlat
, 1);
764 doorbell_gpa
= mrs
.offset_within_address_space
;
765 memory_region_unref(mrs
.mr
);
767 route
->u
.msi
.address_lo
= doorbell_gpa
;
768 route
->u
.msi
.address_hi
= doorbell_gpa
>> 32;
770 trace_kvm_arm_fixup_msi_route(address
, doorbell_gpa
);
779 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry
*route
,
780 int vector
, PCIDevice
*dev
)
785 int kvm_arch_release_virq_post(int virq
)
790 int kvm_arch_msi_data_to_gsi(uint32_t data
)
792 return (data
- 32) & 0xffff;