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 "qemu/main-loop.h"
20 #include "sysemu/sysemu.h"
21 #include "sysemu/kvm.h"
22 #include "sysemu/kvm_int.h"
26 #include "internals.h"
27 #include "hw/pci/pci.h"
28 #include "exec/memattrs.h"
29 #include "exec/address-spaces.h"
30 #include "hw/boards.h"
34 const KVMCapabilityInfo kvm_arch_required_capabilities
[] = {
38 static bool cap_has_mp_state
;
39 static bool cap_has_inject_serror_esr
;
41 static ARMHostCPUFeatures arm_host_cpu_features
;
43 int kvm_arm_vcpu_init(CPUState
*cs
)
45 ARMCPU
*cpu
= ARM_CPU(cs
);
46 struct kvm_vcpu_init init
;
48 init
.target
= cpu
->kvm_target
;
49 memcpy(init
.features
, cpu
->kvm_init_features
, sizeof(init
.features
));
51 return kvm_vcpu_ioctl(cs
, KVM_ARM_VCPU_INIT
, &init
);
54 void kvm_arm_init_serror_injection(CPUState
*cs
)
56 cap_has_inject_serror_esr
= kvm_check_extension(cs
->kvm_state
,
57 KVM_CAP_ARM_INJECT_SERROR_ESR
);
60 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try
,
62 struct kvm_vcpu_init
*init
)
64 int ret
, kvmfd
= -1, vmfd
= -1, cpufd
= -1;
66 kvmfd
= qemu_open("/dev/kvm", O_RDWR
);
70 vmfd
= ioctl(kvmfd
, KVM_CREATE_VM
, 0);
74 cpufd
= ioctl(vmfd
, KVM_CREATE_VCPU
, 0);
80 /* Caller doesn't want the VCPU to be initialized, so skip it */
84 ret
= ioctl(vmfd
, KVM_ARM_PREFERRED_TARGET
, init
);
86 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
90 } else if (cpus_to_try
) {
91 /* Old kernel which doesn't know about the
92 * PREFERRED_TARGET ioctl: we know it will only support
93 * creating one kind of guest CPU which is its preferred
96 while (*cpus_to_try
!= QEMU_KVM_ARM_TARGET_NONE
) {
97 init
->target
= *cpus_to_try
++;
98 memset(init
->features
, 0, sizeof(init
->features
));
99 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
108 /* Treat a NULL cpus_to_try argument the same as an empty
109 * list, which means we will fail the call since this must
110 * be an old kernel which doesn't support PREFERRED_TARGET.
136 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray
)
140 for (i
= 2; i
>= 0; i
--) {
145 void kvm_arm_set_cpu_features_from_host(ARMCPU
*cpu
)
147 CPUARMState
*env
= &cpu
->env
;
149 if (!arm_host_cpu_features
.dtb_compatible
) {
150 if (!kvm_enabled() ||
151 !kvm_arm_get_host_cpu_features(&arm_host_cpu_features
)) {
152 /* We can't report this error yet, so flag that we need to
153 * in arm_cpu_realizefn().
155 cpu
->kvm_target
= QEMU_KVM_ARM_TARGET_NONE
;
156 cpu
->host_cpu_probe_failed
= true;
161 cpu
->kvm_target
= arm_host_cpu_features
.target
;
162 cpu
->dtb_compatible
= arm_host_cpu_features
.dtb_compatible
;
163 cpu
->isar
= arm_host_cpu_features
.isar
;
164 env
->features
= arm_host_cpu_features
.features
;
167 bool kvm_arm_pmu_supported(CPUState
*cpu
)
169 KVMState
*s
= KVM_STATE(current_machine
->accelerator
);
171 return kvm_check_extension(s
, KVM_CAP_ARM_PMU_V3
);
174 int kvm_arm_get_max_vm_ipa_size(MachineState
*ms
)
176 KVMState
*s
= KVM_STATE(ms
->accelerator
);
179 ret
= kvm_check_extension(s
, KVM_CAP_ARM_VM_IPA_SIZE
);
180 return ret
> 0 ? ret
: 40;
183 int kvm_arch_init(MachineState
*ms
, KVMState
*s
)
186 /* For ARM interrupt delivery is always asynchronous,
187 * whether we are using an in-kernel VGIC or not.
189 kvm_async_interrupts_allowed
= true;
192 * PSCI wakes up secondary cores, so we always need to
193 * have vCPUs waiting in kernel space
195 kvm_halt_in_kernel_allowed
= true;
197 cap_has_mp_state
= kvm_check_extension(s
, KVM_CAP_MP_STATE
);
199 if (ms
->smp
.cpus
> 256 &&
200 !kvm_check_extension(s
, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2
)) {
201 error_report("Using more than 256 vcpus requires a host kernel "
202 "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
209 unsigned long kvm_arch_vcpu_id(CPUState
*cpu
)
211 return cpu
->cpu_index
;
214 /* We track all the KVM devices which need their memory addresses
215 * passing to the kernel in a list of these structures.
216 * When board init is complete we run through the list and
217 * tell the kernel the base addresses of the memory regions.
218 * We use a MemoryListener to track mapping and unmapping of
219 * the regions during board creation, so the board models don't
220 * need to do anything special for the KVM case.
222 * Sometimes the address must be OR'ed with some other fields
223 * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
224 * @kda_addr_ormask aims at storing the value of those fields.
226 typedef struct KVMDevice
{
227 struct kvm_arm_device_addr kda
;
228 struct kvm_device_attr kdattr
;
229 uint64_t kda_addr_ormask
;
231 QSLIST_ENTRY(KVMDevice
) entries
;
235 static QSLIST_HEAD(, KVMDevice
) kvm_devices_head
;
237 static void kvm_arm_devlistener_add(MemoryListener
*listener
,
238 MemoryRegionSection
*section
)
242 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
243 if (section
->mr
== kd
->mr
) {
244 kd
->kda
.addr
= section
->offset_within_address_space
;
249 static void kvm_arm_devlistener_del(MemoryListener
*listener
,
250 MemoryRegionSection
*section
)
254 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
255 if (section
->mr
== kd
->mr
) {
261 static MemoryListener devlistener
= {
262 .region_add
= kvm_arm_devlistener_add
,
263 .region_del
= kvm_arm_devlistener_del
,
266 static void kvm_arm_set_device_addr(KVMDevice
*kd
)
268 struct kvm_device_attr
*attr
= &kd
->kdattr
;
271 /* If the device control API is available and we have a device fd on the
272 * KVMDevice struct, let's use the newer API
274 if (kd
->dev_fd
>= 0) {
275 uint64_t addr
= kd
->kda
.addr
;
277 addr
|= kd
->kda_addr_ormask
;
278 attr
->addr
= (uintptr_t)&addr
;
279 ret
= kvm_device_ioctl(kd
->dev_fd
, KVM_SET_DEVICE_ATTR
, attr
);
281 ret
= kvm_vm_ioctl(kvm_state
, KVM_ARM_SET_DEVICE_ADDR
, &kd
->kda
);
285 fprintf(stderr
, "Failed to set device address: %s\n",
291 static void kvm_arm_machine_init_done(Notifier
*notifier
, void *data
)
295 QSLIST_FOREACH_SAFE(kd
, &kvm_devices_head
, entries
, tkd
) {
296 if (kd
->kda
.addr
!= -1) {
297 kvm_arm_set_device_addr(kd
);
299 memory_region_unref(kd
->mr
);
300 QSLIST_REMOVE_HEAD(&kvm_devices_head
, entries
);
303 memory_listener_unregister(&devlistener
);
306 static Notifier notify
= {
307 .notify
= kvm_arm_machine_init_done
,
310 void kvm_arm_register_device(MemoryRegion
*mr
, uint64_t devid
, uint64_t group
,
311 uint64_t attr
, int dev_fd
, uint64_t addr_ormask
)
315 if (!kvm_irqchip_in_kernel()) {
319 if (QSLIST_EMPTY(&kvm_devices_head
)) {
320 memory_listener_register(&devlistener
, &address_space_memory
);
321 qemu_add_machine_init_done_notifier(¬ify
);
323 kd
= g_new0(KVMDevice
, 1);
327 kd
->kdattr
.flags
= 0;
328 kd
->kdattr
.group
= group
;
329 kd
->kdattr
.attr
= attr
;
331 kd
->kda_addr_ormask
= addr_ormask
;
332 QSLIST_INSERT_HEAD(&kvm_devices_head
, kd
, entries
);
333 memory_region_ref(kd
->mr
);
336 static int compare_u64(const void *a
, const void *b
)
338 if (*(uint64_t *)a
> *(uint64_t *)b
) {
341 if (*(uint64_t *)a
< *(uint64_t *)b
) {
347 /* Initialize the ARMCPU cpreg list according to the kernel's
348 * definition of what CPU registers it knows about (and throw away
349 * the previous TCG-created cpreg list).
351 int kvm_arm_init_cpreg_list(ARMCPU
*cpu
)
353 struct kvm_reg_list rl
;
354 struct kvm_reg_list
*rlp
;
355 int i
, ret
, arraylen
;
356 CPUState
*cs
= CPU(cpu
);
359 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, &rl
);
363 rlp
= g_malloc(sizeof(struct kvm_reg_list
) + rl
.n
* sizeof(uint64_t));
365 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, rlp
);
369 /* Sort the list we get back from the kernel, since cpreg_tuples
370 * must be in strictly ascending order.
372 qsort(&rlp
->reg
, rlp
->n
, sizeof(rlp
->reg
[0]), compare_u64
);
374 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
375 if (!kvm_arm_reg_syncs_via_cpreg_list(rlp
->reg
[i
])) {
378 switch (rlp
->reg
[i
] & KVM_REG_SIZE_MASK
) {
379 case KVM_REG_SIZE_U32
:
380 case KVM_REG_SIZE_U64
:
383 fprintf(stderr
, "Can't handle size of register in kernel list\n");
391 cpu
->cpreg_indexes
= g_renew(uint64_t, cpu
->cpreg_indexes
, arraylen
);
392 cpu
->cpreg_values
= g_renew(uint64_t, cpu
->cpreg_values
, arraylen
);
393 cpu
->cpreg_vmstate_indexes
= g_renew(uint64_t, cpu
->cpreg_vmstate_indexes
,
395 cpu
->cpreg_vmstate_values
= g_renew(uint64_t, cpu
->cpreg_vmstate_values
,
397 cpu
->cpreg_array_len
= arraylen
;
398 cpu
->cpreg_vmstate_array_len
= arraylen
;
400 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
401 uint64_t regidx
= rlp
->reg
[i
];
402 if (!kvm_arm_reg_syncs_via_cpreg_list(regidx
)) {
405 cpu
->cpreg_indexes
[arraylen
] = regidx
;
408 assert(cpu
->cpreg_array_len
== arraylen
);
410 if (!write_kvmstate_to_list(cpu
)) {
411 /* Shouldn't happen unless kernel is inconsistent about
412 * what registers exist.
414 fprintf(stderr
, "Initial read of kernel register state failed\n");
424 bool write_kvmstate_to_list(ARMCPU
*cpu
)
426 CPUState
*cs
= CPU(cpu
);
430 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
431 struct kvm_one_reg r
;
432 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
438 switch (regidx
& KVM_REG_SIZE_MASK
) {
439 case KVM_REG_SIZE_U32
:
440 r
.addr
= (uintptr_t)&v32
;
441 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
443 cpu
->cpreg_values
[i
] = v32
;
446 case KVM_REG_SIZE_U64
:
447 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
448 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
460 bool write_list_to_kvmstate(ARMCPU
*cpu
, int level
)
462 CPUState
*cs
= CPU(cpu
);
466 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
467 struct kvm_one_reg r
;
468 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
472 if (kvm_arm_cpreg_level(regidx
) > level
) {
477 switch (regidx
& KVM_REG_SIZE_MASK
) {
478 case KVM_REG_SIZE_U32
:
479 v32
= cpu
->cpreg_values
[i
];
480 r
.addr
= (uintptr_t)&v32
;
482 case KVM_REG_SIZE_U64
:
483 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
488 ret
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, &r
);
490 /* We might fail for "unknown register" and also for
491 * "you tried to set a register which is constant with
492 * a different value from what it actually contains".
500 void kvm_arm_reset_vcpu(ARMCPU
*cpu
)
504 /* Re-init VCPU so that all registers are set to
505 * their respective reset values.
507 ret
= kvm_arm_vcpu_init(CPU(cpu
));
509 fprintf(stderr
, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret
));
512 if (!write_kvmstate_to_list(cpu
)) {
513 fprintf(stderr
, "write_kvmstate_to_list failed\n");
517 * Sync the reset values also into the CPUState. This is necessary
518 * because the next thing we do will be a kvm_arch_put_registers()
519 * which will update the list values from the CPUState before copying
520 * the list values back to KVM. It's OK to ignore failure returns here
521 * for the same reason we do so in kvm_arch_get_registers().
523 write_list_to_cpustate(cpu
);
527 * Update KVM's MP_STATE based on what QEMU thinks it is
529 int kvm_arm_sync_mpstate_to_kvm(ARMCPU
*cpu
)
531 if (cap_has_mp_state
) {
532 struct kvm_mp_state mp_state
= {
533 .mp_state
= (cpu
->power_state
== PSCI_OFF
) ?
534 KVM_MP_STATE_STOPPED
: KVM_MP_STATE_RUNNABLE
536 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_MP_STATE
, &mp_state
);
538 fprintf(stderr
, "%s: failed to set MP_STATE %d/%s\n",
539 __func__
, ret
, strerror(-ret
));
548 * Sync the KVM MP_STATE into QEMU
550 int kvm_arm_sync_mpstate_to_qemu(ARMCPU
*cpu
)
552 if (cap_has_mp_state
) {
553 struct kvm_mp_state mp_state
;
554 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_MP_STATE
, &mp_state
);
556 fprintf(stderr
, "%s: failed to get MP_STATE %d/%s\n",
557 __func__
, ret
, strerror(-ret
));
560 cpu
->power_state
= (mp_state
.mp_state
== KVM_MP_STATE_STOPPED
) ?
567 int kvm_put_vcpu_events(ARMCPU
*cpu
)
569 CPUARMState
*env
= &cpu
->env
;
570 struct kvm_vcpu_events events
;
573 if (!kvm_has_vcpu_events()) {
577 memset(&events
, 0, sizeof(events
));
578 events
.exception
.serror_pending
= env
->serror
.pending
;
580 /* Inject SError to guest with specified syndrome if host kernel
581 * supports it, otherwise inject SError without syndrome.
583 if (cap_has_inject_serror_esr
) {
584 events
.exception
.serror_has_esr
= env
->serror
.has_esr
;
585 events
.exception
.serror_esr
= env
->serror
.esr
;
588 ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_VCPU_EVENTS
, &events
);
590 error_report("failed to put vcpu events");
596 int kvm_get_vcpu_events(ARMCPU
*cpu
)
598 CPUARMState
*env
= &cpu
->env
;
599 struct kvm_vcpu_events events
;
602 if (!kvm_has_vcpu_events()) {
606 memset(&events
, 0, sizeof(events
));
607 ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_VCPU_EVENTS
, &events
);
609 error_report("failed to get vcpu events");
613 env
->serror
.pending
= events
.exception
.serror_pending
;
614 env
->serror
.has_esr
= events
.exception
.serror_has_esr
;
615 env
->serror
.esr
= events
.exception
.serror_esr
;
620 void kvm_arch_pre_run(CPUState
*cs
, struct kvm_run
*run
)
624 MemTxAttrs
kvm_arch_post_run(CPUState
*cs
, struct kvm_run
*run
)
627 uint32_t switched_level
;
629 if (kvm_irqchip_in_kernel()) {
631 * We only need to sync timer states with user-space interrupt
632 * controllers, so return early and save cycles if we don't.
634 return MEMTXATTRS_UNSPECIFIED
;
639 /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
640 if (run
->s
.regs
.device_irq_level
!= cpu
->device_irq_level
) {
641 switched_level
= cpu
->device_irq_level
^ run
->s
.regs
.device_irq_level
;
643 qemu_mutex_lock_iothread();
645 if (switched_level
& KVM_ARM_DEV_EL1_VTIMER
) {
646 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_VIRT
],
647 !!(run
->s
.regs
.device_irq_level
&
648 KVM_ARM_DEV_EL1_VTIMER
));
649 switched_level
&= ~KVM_ARM_DEV_EL1_VTIMER
;
652 if (switched_level
& KVM_ARM_DEV_EL1_PTIMER
) {
653 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_PHYS
],
654 !!(run
->s
.regs
.device_irq_level
&
655 KVM_ARM_DEV_EL1_PTIMER
));
656 switched_level
&= ~KVM_ARM_DEV_EL1_PTIMER
;
659 if (switched_level
& KVM_ARM_DEV_PMU
) {
660 qemu_set_irq(cpu
->pmu_interrupt
,
661 !!(run
->s
.regs
.device_irq_level
& KVM_ARM_DEV_PMU
));
662 switched_level
&= ~KVM_ARM_DEV_PMU
;
665 if (switched_level
) {
666 qemu_log_mask(LOG_UNIMP
, "%s: unhandled in-kernel device IRQ %x\n",
667 __func__
, switched_level
);
670 /* We also mark unknown levels as processed to not waste cycles */
671 cpu
->device_irq_level
= run
->s
.regs
.device_irq_level
;
672 qemu_mutex_unlock_iothread();
675 return MEMTXATTRS_UNSPECIFIED
;
679 int kvm_arch_handle_exit(CPUState
*cs
, struct kvm_run
*run
)
683 switch (run
->exit_reason
) {
685 if (kvm_arm_handle_debug(cs
, &run
->debug
.arch
)) {
687 } /* otherwise return to guest */
690 qemu_log_mask(LOG_UNIMP
, "%s: un-handled exit reason %d\n",
691 __func__
, run
->exit_reason
);
697 bool kvm_arch_stop_on_emulation_error(CPUState
*cs
)
702 int kvm_arch_process_async_events(CPUState
*cs
)
707 /* The #ifdef protections are until 32bit headers are imported and can
708 * be removed once both 32 and 64 bit reach feature parity.
710 void kvm_arch_update_guest_debug(CPUState
*cs
, struct kvm_guest_debug
*dbg
)
712 #ifdef KVM_GUESTDBG_USE_SW_BP
713 if (kvm_sw_breakpoints_active(cs
)) {
714 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_SW_BP
;
717 #ifdef KVM_GUESTDBG_USE_HW
718 if (kvm_arm_hw_debug_active(cs
)) {
719 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_HW
;
720 kvm_arm_copy_hw_debug_data(&dbg
->arch
);
725 void kvm_arch_init_irq_routing(KVMState
*s
)
729 int kvm_arch_irqchip_create(MachineState
*ms
, KVMState
*s
)
731 if (machine_kernel_irqchip_split(ms
)) {
732 perror("-machine kernel_irqchip=split is not supported on ARM.");
736 /* If we can create the VGIC using the newer device control API, we
737 * let the device do this when it initializes itself, otherwise we
738 * fall back to the old API */
739 return kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
);
742 int kvm_arm_vgic_probe(void)
744 if (kvm_create_device(kvm_state
,
745 KVM_DEV_TYPE_ARM_VGIC_V3
, true) == 0) {
747 } else if (kvm_create_device(kvm_state
,
748 KVM_DEV_TYPE_ARM_VGIC_V2
, true) == 0) {
755 int kvm_arm_set_irq(int cpu
, int irqtype
, int irq
, int level
)
757 int kvm_irq
= (irqtype
<< KVM_ARM_IRQ_TYPE_SHIFT
) | irq
;
758 int cpu_idx1
= cpu
% 256;
759 int cpu_idx2
= cpu
/ 256;
761 kvm_irq
|= (cpu_idx1
<< KVM_ARM_IRQ_VCPU_SHIFT
) |
762 (cpu_idx2
<< KVM_ARM_IRQ_VCPU2_SHIFT
);
764 return kvm_set_irq(kvm_state
, kvm_irq
, !!level
);
767 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry
*route
,
768 uint64_t address
, uint32_t data
, PCIDevice
*dev
)
770 AddressSpace
*as
= pci_device_iommu_address_space(dev
);
771 hwaddr xlat
, len
, doorbell_gpa
;
772 MemoryRegionSection mrs
;
776 if (as
== &address_space_memory
) {
780 /* MSI doorbell address is translated by an IOMMU */
783 mr
= address_space_translate(as
, address
, &xlat
, &len
, true,
784 MEMTXATTRS_UNSPECIFIED
);
788 mrs
= memory_region_find(mr
, xlat
, 1);
793 doorbell_gpa
= mrs
.offset_within_address_space
;
794 memory_region_unref(mrs
.mr
);
796 route
->u
.msi
.address_lo
= doorbell_gpa
;
797 route
->u
.msi
.address_hi
= doorbell_gpa
>> 32;
799 trace_kvm_arm_fixup_msi_route(address
, doorbell_gpa
);
808 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry
*route
,
809 int vector
, PCIDevice
*dev
)
814 int kvm_arch_release_virq_post(int virq
)
819 int kvm_arch_msi_data_to_gsi(uint32_t data
)
821 return (data
- 32) & 0xffff;