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"
23 #include "internals.h"
24 #include "hw/arm/arm.h"
25 #include "exec/memattrs.h"
26 #include "exec/address-spaces.h"
27 #include "hw/boards.h"
30 const KVMCapabilityInfo kvm_arch_required_capabilities
[] = {
34 static bool cap_has_mp_state
;
36 static ARMHostCPUFeatures arm_host_cpu_features
;
38 int kvm_arm_vcpu_init(CPUState
*cs
)
40 ARMCPU
*cpu
= ARM_CPU(cs
);
41 struct kvm_vcpu_init init
;
43 init
.target
= cpu
->kvm_target
;
44 memcpy(init
.features
, cpu
->kvm_init_features
, sizeof(init
.features
));
46 return kvm_vcpu_ioctl(cs
, KVM_ARM_VCPU_INIT
, &init
);
49 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try
,
51 struct kvm_vcpu_init
*init
)
53 int ret
, kvmfd
= -1, vmfd
= -1, cpufd
= -1;
55 kvmfd
= qemu_open("/dev/kvm", O_RDWR
);
59 vmfd
= ioctl(kvmfd
, KVM_CREATE_VM
, 0);
63 cpufd
= ioctl(vmfd
, KVM_CREATE_VCPU
, 0);
69 /* Caller doesn't want the VCPU to be initialized, so skip it */
73 ret
= ioctl(vmfd
, KVM_ARM_PREFERRED_TARGET
, init
);
75 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
79 } else if (cpus_to_try
) {
80 /* Old kernel which doesn't know about the
81 * PREFERRED_TARGET ioctl: we know it will only support
82 * creating one kind of guest CPU which is its preferred
85 while (*cpus_to_try
!= QEMU_KVM_ARM_TARGET_NONE
) {
86 init
->target
= *cpus_to_try
++;
87 memset(init
->features
, 0, sizeof(init
->features
));
88 ret
= ioctl(cpufd
, KVM_ARM_VCPU_INIT
, init
);
97 /* Treat a NULL cpus_to_try argument the same as an empty
98 * list, which means we will fail the call since this must
99 * be an old kernel which doesn't support PREFERRED_TARGET.
125 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray
)
129 for (i
= 2; i
>= 0; i
--) {
134 void kvm_arm_set_cpu_features_from_host(ARMCPU
*cpu
)
136 CPUARMState
*env
= &cpu
->env
;
138 if (!arm_host_cpu_features
.dtb_compatible
) {
139 if (!kvm_enabled() ||
140 !kvm_arm_get_host_cpu_features(&arm_host_cpu_features
)) {
141 /* We can't report this error yet, so flag that we need to
142 * in arm_cpu_realizefn().
144 cpu
->kvm_target
= QEMU_KVM_ARM_TARGET_NONE
;
145 cpu
->host_cpu_probe_failed
= true;
150 cpu
->kvm_target
= arm_host_cpu_features
.target
;
151 cpu
->dtb_compatible
= arm_host_cpu_features
.dtb_compatible
;
152 env
->features
= arm_host_cpu_features
.features
;
155 int kvm_arch_init(MachineState
*ms
, KVMState
*s
)
157 /* For ARM interrupt delivery is always asynchronous,
158 * whether we are using an in-kernel VGIC or not.
160 kvm_async_interrupts_allowed
= true;
163 * PSCI wakes up secondary cores, so we always need to
164 * have vCPUs waiting in kernel space
166 kvm_halt_in_kernel_allowed
= true;
168 cap_has_mp_state
= kvm_check_extension(s
, KVM_CAP_MP_STATE
);
173 unsigned long kvm_arch_vcpu_id(CPUState
*cpu
)
175 return cpu
->cpu_index
;
178 /* We track all the KVM devices which need their memory addresses
179 * passing to the kernel in a list of these structures.
180 * When board init is complete we run through the list and
181 * tell the kernel the base addresses of the memory regions.
182 * We use a MemoryListener to track mapping and unmapping of
183 * the regions during board creation, so the board models don't
184 * need to do anything special for the KVM case.
186 typedef struct KVMDevice
{
187 struct kvm_arm_device_addr kda
;
188 struct kvm_device_attr kdattr
;
190 QSLIST_ENTRY(KVMDevice
) entries
;
194 static QSLIST_HEAD(kvm_devices_head
, KVMDevice
) kvm_devices_head
;
196 static void kvm_arm_devlistener_add(MemoryListener
*listener
,
197 MemoryRegionSection
*section
)
201 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
202 if (section
->mr
== kd
->mr
) {
203 kd
->kda
.addr
= section
->offset_within_address_space
;
208 static void kvm_arm_devlistener_del(MemoryListener
*listener
,
209 MemoryRegionSection
*section
)
213 QSLIST_FOREACH(kd
, &kvm_devices_head
, entries
) {
214 if (section
->mr
== kd
->mr
) {
220 static MemoryListener devlistener
= {
221 .region_add
= kvm_arm_devlistener_add
,
222 .region_del
= kvm_arm_devlistener_del
,
225 static void kvm_arm_set_device_addr(KVMDevice
*kd
)
227 struct kvm_device_attr
*attr
= &kd
->kdattr
;
230 /* If the device control API is available and we have a device fd on the
231 * KVMDevice struct, let's use the newer API
233 if (kd
->dev_fd
>= 0) {
234 uint64_t addr
= kd
->kda
.addr
;
235 attr
->addr
= (uintptr_t)&addr
;
236 ret
= kvm_device_ioctl(kd
->dev_fd
, KVM_SET_DEVICE_ATTR
, attr
);
238 ret
= kvm_vm_ioctl(kvm_state
, KVM_ARM_SET_DEVICE_ADDR
, &kd
->kda
);
242 fprintf(stderr
, "Failed to set device address: %s\n",
248 static void kvm_arm_machine_init_done(Notifier
*notifier
, void *data
)
252 QSLIST_FOREACH_SAFE(kd
, &kvm_devices_head
, entries
, tkd
) {
253 if (kd
->kda
.addr
!= -1) {
254 kvm_arm_set_device_addr(kd
);
256 memory_region_unref(kd
->mr
);
259 memory_listener_unregister(&devlistener
);
262 static Notifier notify
= {
263 .notify
= kvm_arm_machine_init_done
,
266 void kvm_arm_register_device(MemoryRegion
*mr
, uint64_t devid
, uint64_t group
,
267 uint64_t attr
, int dev_fd
)
271 if (!kvm_irqchip_in_kernel()) {
275 if (QSLIST_EMPTY(&kvm_devices_head
)) {
276 memory_listener_register(&devlistener
, &address_space_memory
);
277 qemu_add_machine_init_done_notifier(¬ify
);
279 kd
= g_new0(KVMDevice
, 1);
283 kd
->kdattr
.flags
= 0;
284 kd
->kdattr
.group
= group
;
285 kd
->kdattr
.attr
= attr
;
287 QSLIST_INSERT_HEAD(&kvm_devices_head
, kd
, entries
);
288 memory_region_ref(kd
->mr
);
291 static int compare_u64(const void *a
, const void *b
)
293 if (*(uint64_t *)a
> *(uint64_t *)b
) {
296 if (*(uint64_t *)a
< *(uint64_t *)b
) {
302 /* Initialize the CPUState's cpreg list according to the kernel's
303 * definition of what CPU registers it knows about (and throw away
304 * the previous TCG-created cpreg list).
306 int kvm_arm_init_cpreg_list(ARMCPU
*cpu
)
308 struct kvm_reg_list rl
;
309 struct kvm_reg_list
*rlp
;
310 int i
, ret
, arraylen
;
311 CPUState
*cs
= CPU(cpu
);
314 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, &rl
);
318 rlp
= g_malloc(sizeof(struct kvm_reg_list
) + rl
.n
* sizeof(uint64_t));
320 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_REG_LIST
, rlp
);
324 /* Sort the list we get back from the kernel, since cpreg_tuples
325 * must be in strictly ascending order.
327 qsort(&rlp
->reg
, rlp
->n
, sizeof(rlp
->reg
[0]), compare_u64
);
329 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
330 if (!kvm_arm_reg_syncs_via_cpreg_list(rlp
->reg
[i
])) {
333 switch (rlp
->reg
[i
] & KVM_REG_SIZE_MASK
) {
334 case KVM_REG_SIZE_U32
:
335 case KVM_REG_SIZE_U64
:
338 fprintf(stderr
, "Can't handle size of register in kernel list\n");
346 cpu
->cpreg_indexes
= g_renew(uint64_t, cpu
->cpreg_indexes
, arraylen
);
347 cpu
->cpreg_values
= g_renew(uint64_t, cpu
->cpreg_values
, arraylen
);
348 cpu
->cpreg_vmstate_indexes
= g_renew(uint64_t, cpu
->cpreg_vmstate_indexes
,
350 cpu
->cpreg_vmstate_values
= g_renew(uint64_t, cpu
->cpreg_vmstate_values
,
352 cpu
->cpreg_array_len
= arraylen
;
353 cpu
->cpreg_vmstate_array_len
= arraylen
;
355 for (i
= 0, arraylen
= 0; i
< rlp
->n
; i
++) {
356 uint64_t regidx
= rlp
->reg
[i
];
357 if (!kvm_arm_reg_syncs_via_cpreg_list(regidx
)) {
360 cpu
->cpreg_indexes
[arraylen
] = regidx
;
363 assert(cpu
->cpreg_array_len
== arraylen
);
365 if (!write_kvmstate_to_list(cpu
)) {
366 /* Shouldn't happen unless kernel is inconsistent about
367 * what registers exist.
369 fprintf(stderr
, "Initial read of kernel register state failed\n");
379 bool write_kvmstate_to_list(ARMCPU
*cpu
)
381 CPUState
*cs
= CPU(cpu
);
385 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
386 struct kvm_one_reg r
;
387 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
393 switch (regidx
& KVM_REG_SIZE_MASK
) {
394 case KVM_REG_SIZE_U32
:
395 r
.addr
= (uintptr_t)&v32
;
396 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
398 cpu
->cpreg_values
[i
] = v32
;
401 case KVM_REG_SIZE_U64
:
402 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
403 ret
= kvm_vcpu_ioctl(cs
, KVM_GET_ONE_REG
, &r
);
415 bool write_list_to_kvmstate(ARMCPU
*cpu
, int level
)
417 CPUState
*cs
= CPU(cpu
);
421 for (i
= 0; i
< cpu
->cpreg_array_len
; i
++) {
422 struct kvm_one_reg r
;
423 uint64_t regidx
= cpu
->cpreg_indexes
[i
];
427 if (kvm_arm_cpreg_level(regidx
) > level
) {
432 switch (regidx
& KVM_REG_SIZE_MASK
) {
433 case KVM_REG_SIZE_U32
:
434 v32
= cpu
->cpreg_values
[i
];
435 r
.addr
= (uintptr_t)&v32
;
437 case KVM_REG_SIZE_U64
:
438 r
.addr
= (uintptr_t)(cpu
->cpreg_values
+ i
);
443 ret
= kvm_vcpu_ioctl(cs
, KVM_SET_ONE_REG
, &r
);
445 /* We might fail for "unknown register" and also for
446 * "you tried to set a register which is constant with
447 * a different value from what it actually contains".
455 void kvm_arm_reset_vcpu(ARMCPU
*cpu
)
459 /* Re-init VCPU so that all registers are set to
460 * their respective reset values.
462 ret
= kvm_arm_vcpu_init(CPU(cpu
));
464 fprintf(stderr
, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret
));
467 if (!write_kvmstate_to_list(cpu
)) {
468 fprintf(stderr
, "write_kvmstate_to_list failed\n");
474 * Update KVM's MP_STATE based on what QEMU thinks it is
476 int kvm_arm_sync_mpstate_to_kvm(ARMCPU
*cpu
)
478 if (cap_has_mp_state
) {
479 struct kvm_mp_state mp_state
= {
480 .mp_state
= (cpu
->power_state
== PSCI_OFF
) ?
481 KVM_MP_STATE_STOPPED
: KVM_MP_STATE_RUNNABLE
483 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_SET_MP_STATE
, &mp_state
);
485 fprintf(stderr
, "%s: failed to set MP_STATE %d/%s\n",
486 __func__
, ret
, strerror(-ret
));
495 * Sync the KVM MP_STATE into QEMU
497 int kvm_arm_sync_mpstate_to_qemu(ARMCPU
*cpu
)
499 if (cap_has_mp_state
) {
500 struct kvm_mp_state mp_state
;
501 int ret
= kvm_vcpu_ioctl(CPU(cpu
), KVM_GET_MP_STATE
, &mp_state
);
503 fprintf(stderr
, "%s: failed to get MP_STATE %d/%s\n",
504 __func__
, ret
, strerror(-ret
));
507 cpu
->power_state
= (mp_state
.mp_state
== KVM_MP_STATE_STOPPED
) ?
514 void kvm_arch_pre_run(CPUState
*cs
, struct kvm_run
*run
)
518 MemTxAttrs
kvm_arch_post_run(CPUState
*cs
, struct kvm_run
*run
)
521 uint32_t switched_level
;
523 if (kvm_irqchip_in_kernel()) {
525 * We only need to sync timer states with user-space interrupt
526 * controllers, so return early and save cycles if we don't.
528 return MEMTXATTRS_UNSPECIFIED
;
533 /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
534 if (run
->s
.regs
.device_irq_level
!= cpu
->device_irq_level
) {
535 switched_level
= cpu
->device_irq_level
^ run
->s
.regs
.device_irq_level
;
537 qemu_mutex_lock_iothread();
539 if (switched_level
& KVM_ARM_DEV_EL1_VTIMER
) {
540 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_VIRT
],
541 !!(run
->s
.regs
.device_irq_level
&
542 KVM_ARM_DEV_EL1_VTIMER
));
543 switched_level
&= ~KVM_ARM_DEV_EL1_VTIMER
;
546 if (switched_level
& KVM_ARM_DEV_EL1_PTIMER
) {
547 qemu_set_irq(cpu
->gt_timer_outputs
[GTIMER_PHYS
],
548 !!(run
->s
.regs
.device_irq_level
&
549 KVM_ARM_DEV_EL1_PTIMER
));
550 switched_level
&= ~KVM_ARM_DEV_EL1_PTIMER
;
553 if (switched_level
& KVM_ARM_DEV_PMU
) {
554 qemu_set_irq(cpu
->pmu_interrupt
,
555 !!(run
->s
.regs
.device_irq_level
& KVM_ARM_DEV_PMU
));
556 switched_level
&= ~KVM_ARM_DEV_PMU
;
559 if (switched_level
) {
560 qemu_log_mask(LOG_UNIMP
, "%s: unhandled in-kernel device IRQ %x\n",
561 __func__
, switched_level
);
564 /* We also mark unknown levels as processed to not waste cycles */
565 cpu
->device_irq_level
= run
->s
.regs
.device_irq_level
;
566 qemu_mutex_unlock_iothread();
569 return MEMTXATTRS_UNSPECIFIED
;
573 int kvm_arch_handle_exit(CPUState
*cs
, struct kvm_run
*run
)
577 switch (run
->exit_reason
) {
579 if (kvm_arm_handle_debug(cs
, &run
->debug
.arch
)) {
581 } /* otherwise return to guest */
584 qemu_log_mask(LOG_UNIMP
, "%s: un-handled exit reason %d\n",
585 __func__
, run
->exit_reason
);
591 bool kvm_arch_stop_on_emulation_error(CPUState
*cs
)
596 int kvm_arch_process_async_events(CPUState
*cs
)
601 /* The #ifdef protections are until 32bit headers are imported and can
602 * be removed once both 32 and 64 bit reach feature parity.
604 void kvm_arch_update_guest_debug(CPUState
*cs
, struct kvm_guest_debug
*dbg
)
606 #ifdef KVM_GUESTDBG_USE_SW_BP
607 if (kvm_sw_breakpoints_active(cs
)) {
608 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_SW_BP
;
611 #ifdef KVM_GUESTDBG_USE_HW
612 if (kvm_arm_hw_debug_active(cs
)) {
613 dbg
->control
|= KVM_GUESTDBG_ENABLE
| KVM_GUESTDBG_USE_HW
;
614 kvm_arm_copy_hw_debug_data(&dbg
->arch
);
619 void kvm_arch_init_irq_routing(KVMState
*s
)
623 int kvm_arch_irqchip_create(MachineState
*ms
, KVMState
*s
)
625 if (machine_kernel_irqchip_split(ms
)) {
626 perror("-machine kernel_irqchip=split is not supported on ARM.");
630 /* If we can create the VGIC using the newer device control API, we
631 * let the device do this when it initializes itself, otherwise we
632 * fall back to the old API */
633 return kvm_check_extension(s
, KVM_CAP_DEVICE_CTRL
);
636 int kvm_arm_vgic_probe(void)
638 if (kvm_create_device(kvm_state
,
639 KVM_DEV_TYPE_ARM_VGIC_V3
, true) == 0) {
641 } else if (kvm_create_device(kvm_state
,
642 KVM_DEV_TYPE_ARM_VGIC_V2
, true) == 0) {
649 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry
*route
,
650 uint64_t address
, uint32_t data
, PCIDevice
*dev
)
655 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry
*route
,
656 int vector
, PCIDevice
*dev
)
661 int kvm_arch_release_virq_post(int virq
)
666 int kvm_arch_msi_data_to_gsi(uint32_t data
)
668 return (data
- 32) & 0xffff;