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.
12 #include <sys/types.h>
13 #include <sys/ioctl.h>
16 #include <linux/kvm.h>
18 #include "qemu-common.h"
19 #include "qemu/timer.h"
20 #include "sysemu/sysemu.h"
21 #include "sysemu/kvm.h"
24 #include "hw/arm/arm.h"
26 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
30 int kvm_arch_init(KVMState *s)
32 /* For ARM interrupt delivery is always asynchronous,
33 * whether we are using an in-kernel VGIC or not.
35 kvm_async_interrupts_allowed = true;
39 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
41 return cpu->cpu_index;
44 static bool reg_syncs_via_tuple_list(uint64_t regidx)
46 /* Return true if the regidx is a register we should synchronize
47 * via the cpreg_tuples array (ie is not a core reg we sync by
48 * hand in kvm_arch_get/put_registers())
50 switch (regidx & KVM_REG_ARM_COPROC_MASK) {
51 case KVM_REG_ARM_CORE:
59 static int compare_u64(const void *a, const void *b)
61 if (*(uint64_t *)a > *(uint64_t *)b) {
64 if (*(uint64_t *)a < *(uint64_t *)b) {
70 int kvm_arch_init_vcpu(CPUState *cs)
72 struct kvm_vcpu_init init;
76 struct kvm_reg_list rl;
77 struct kvm_reg_list *rlp;
78 ARMCPU *cpu = ARM_CPU(cs);
80 if (cpu->kvm_target == QEMU_KVM_ARM_TARGET_NONE) {
81 fprintf(stderr, "KVM is not supported for this guest CPU type\n");
85 init.target = cpu->kvm_target;
86 memset(init.features, 0, sizeof(init.features));
87 if (cpu->start_powered_off) {
88 init.features[0] = 1 << KVM_ARM_VCPU_POWER_OFF;
90 ret = kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
94 /* Query the kernel to make sure it supports 32 VFP
95 * registers: QEMU's "cortex-a15" CPU is always a
96 * VFP-D32 core. The simplest way to do this is just
97 * to attempt to read register d31.
99 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP | 31;
100 r.addr = (uintptr_t)(&v);
101 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
102 if (ret == -ENOENT) {
106 /* Populate the cpreg list based on the kernel's idea
107 * of what registers exist (and throw away the TCG-created list).
110 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
114 rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
116 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
120 /* Sort the list we get back from the kernel, since cpreg_tuples
121 * must be in strictly ascending order.
123 qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
125 for (i = 0, arraylen = 0; i < rlp->n; i++) {
126 if (!reg_syncs_via_tuple_list(rlp->reg[i])) {
129 switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
130 case KVM_REG_SIZE_U32:
131 case KVM_REG_SIZE_U64:
134 fprintf(stderr, "Can't handle size of register in kernel list\n");
142 cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
143 cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
144 cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
146 cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
148 cpu->cpreg_array_len = arraylen;
149 cpu->cpreg_vmstate_array_len = arraylen;
151 for (i = 0, arraylen = 0; i < rlp->n; i++) {
152 uint64_t regidx = rlp->reg[i];
153 if (!reg_syncs_via_tuple_list(regidx)) {
156 cpu->cpreg_indexes[arraylen] = regidx;
159 assert(cpu->cpreg_array_len == arraylen);
161 if (!write_kvmstate_to_list(cpu)) {
162 /* Shouldn't happen unless kernel is inconsistent about
163 * what registers exist.
165 fprintf(stderr, "Initial read of kernel register state failed\n");
170 /* Save a copy of the initial register values so that we can
171 * feed it back to the kernel on VCPU reset.
173 cpu->cpreg_reset_values = g_memdup(cpu->cpreg_values,
174 cpu->cpreg_array_len *
175 sizeof(cpu->cpreg_values[0]));
182 /* We track all the KVM devices which need their memory addresses
183 * passing to the kernel in a list of these structures.
184 * When board init is complete we run through the list and
185 * tell the kernel the base addresses of the memory regions.
186 * We use a MemoryListener to track mapping and unmapping of
187 * the regions during board creation, so the board models don't
188 * need to do anything special for the KVM case.
190 typedef struct KVMDevice {
191 struct kvm_arm_device_addr kda;
193 QSLIST_ENTRY(KVMDevice) entries;
196 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
198 static void kvm_arm_devlistener_add(MemoryListener *listener,
199 MemoryRegionSection *section)
203 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
204 if (section->mr == kd->mr) {
205 kd->kda.addr = section->offset_within_address_space;
210 static void kvm_arm_devlistener_del(MemoryListener *listener,
211 MemoryRegionSection *section)
215 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
216 if (section->mr == kd->mr) {
222 static MemoryListener devlistener = {
223 .region_add = kvm_arm_devlistener_add,
224 .region_del = kvm_arm_devlistener_del,
227 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
231 memory_listener_unregister(&devlistener);
232 QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
233 if (kd->kda.addr != -1) {
234 if (kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR,
236 fprintf(stderr, "KVM_ARM_SET_DEVICE_ADDRESS failed: %s\n",
241 memory_region_unref(kd->mr);
246 static Notifier notify = {
247 .notify = kvm_arm_machine_init_done,
250 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid)
254 if (!kvm_irqchip_in_kernel()) {
258 if (QSLIST_EMPTY(&kvm_devices_head)) {
259 memory_listener_register(&devlistener, NULL);
260 qemu_add_machine_init_done_notifier(¬ify);
262 kd = g_new0(KVMDevice, 1);
266 QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
267 memory_region_ref(kd->mr);
270 bool write_kvmstate_to_list(ARMCPU *cpu)
272 CPUState *cs = CPU(cpu);
276 for (i = 0; i < cpu->cpreg_array_len; i++) {
277 struct kvm_one_reg r;
278 uint64_t regidx = cpu->cpreg_indexes[i];
284 switch (regidx & KVM_REG_SIZE_MASK) {
285 case KVM_REG_SIZE_U32:
286 r.addr = (uintptr_t)&v32;
287 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
289 cpu->cpreg_values[i] = v32;
292 case KVM_REG_SIZE_U64:
293 r.addr = (uintptr_t)(cpu->cpreg_values + i);
294 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
306 bool write_list_to_kvmstate(ARMCPU *cpu)
308 CPUState *cs = CPU(cpu);
312 for (i = 0; i < cpu->cpreg_array_len; i++) {
313 struct kvm_one_reg r;
314 uint64_t regidx = cpu->cpreg_indexes[i];
319 switch (regidx & KVM_REG_SIZE_MASK) {
320 case KVM_REG_SIZE_U32:
321 v32 = cpu->cpreg_values[i];
322 r.addr = (uintptr_t)&v32;
324 case KVM_REG_SIZE_U64:
325 r.addr = (uintptr_t)(cpu->cpreg_values + i);
330 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
332 /* We might fail for "unknown register" and also for
333 * "you tried to set a register which is constant with
334 * a different value from what it actually contains".
347 #define COREREG(KERNELNAME, QEMUFIELD) \
349 KVM_REG_ARM | KVM_REG_SIZE_U32 | \
350 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
351 offsetof(CPUARMState, QEMUFIELD) \
354 #define VFPSYSREG(R) \
356 KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP | \
357 KVM_REG_ARM_VFP_##R, \
358 offsetof(CPUARMState, vfp.xregs[ARM_VFP_##R]) \
361 static const Reg regs[] = {
362 /* R0_usr .. R14_usr */
363 COREREG(usr_regs.uregs[0], regs[0]),
364 COREREG(usr_regs.uregs[1], regs[1]),
365 COREREG(usr_regs.uregs[2], regs[2]),
366 COREREG(usr_regs.uregs[3], regs[3]),
367 COREREG(usr_regs.uregs[4], regs[4]),
368 COREREG(usr_regs.uregs[5], regs[5]),
369 COREREG(usr_regs.uregs[6], regs[6]),
370 COREREG(usr_regs.uregs[7], regs[7]),
371 COREREG(usr_regs.uregs[8], usr_regs[0]),
372 COREREG(usr_regs.uregs[9], usr_regs[1]),
373 COREREG(usr_regs.uregs[10], usr_regs[2]),
374 COREREG(usr_regs.uregs[11], usr_regs[3]),
375 COREREG(usr_regs.uregs[12], usr_regs[4]),
376 COREREG(usr_regs.uregs[13], banked_r13[0]),
377 COREREG(usr_regs.uregs[14], banked_r14[0]),
378 /* R13, R14, SPSR for SVC, ABT, UND, IRQ banks */
379 COREREG(svc_regs[0], banked_r13[1]),
380 COREREG(svc_regs[1], banked_r14[1]),
381 COREREG(svc_regs[2], banked_spsr[1]),
382 COREREG(abt_regs[0], banked_r13[2]),
383 COREREG(abt_regs[1], banked_r14[2]),
384 COREREG(abt_regs[2], banked_spsr[2]),
385 COREREG(und_regs[0], banked_r13[3]),
386 COREREG(und_regs[1], banked_r14[3]),
387 COREREG(und_regs[2], banked_spsr[3]),
388 COREREG(irq_regs[0], banked_r13[4]),
389 COREREG(irq_regs[1], banked_r14[4]),
390 COREREG(irq_regs[2], banked_spsr[4]),
391 /* R8_fiq .. R14_fiq and SPSR_fiq */
392 COREREG(fiq_regs[0], fiq_regs[0]),
393 COREREG(fiq_regs[1], fiq_regs[1]),
394 COREREG(fiq_regs[2], fiq_regs[2]),
395 COREREG(fiq_regs[3], fiq_regs[3]),
396 COREREG(fiq_regs[4], fiq_regs[4]),
397 COREREG(fiq_regs[5], banked_r13[5]),
398 COREREG(fiq_regs[6], banked_r14[5]),
399 COREREG(fiq_regs[7], banked_spsr[5]),
401 COREREG(usr_regs.uregs[15], regs[15]),
402 /* VFP system registers */
411 int kvm_arch_put_registers(CPUState *cs, int level)
413 ARMCPU *cpu = ARM_CPU(cs);
414 CPUARMState *env = &cpu->env;
415 struct kvm_one_reg r;
418 uint32_t cpsr, fpscr;
420 /* Make sure the banked regs are properly set */
421 mode = env->uncached_cpsr & CPSR_M;
422 bn = bank_number(mode);
423 if (mode == ARM_CPU_MODE_FIQ) {
424 memcpy(env->fiq_regs, env->regs + 8, 5 * sizeof(uint32_t));
426 memcpy(env->usr_regs, env->regs + 8, 5 * sizeof(uint32_t));
428 env->banked_r13[bn] = env->regs[13];
429 env->banked_r14[bn] = env->regs[14];
430 env->banked_spsr[bn] = env->spsr;
432 /* Now we can safely copy stuff down to the kernel */
433 for (i = 0; i < ARRAY_SIZE(regs); i++) {
435 r.addr = (uintptr_t)(env) + regs[i].offset;
436 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
442 /* Special cases which aren't a single CPUARMState field */
443 cpsr = cpsr_read(env);
444 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
445 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
446 r.addr = (uintptr_t)(&cpsr);
447 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
453 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
454 for (i = 0; i < 32; i++) {
455 r.addr = (uintptr_t)(&env->vfp.regs[i]);
456 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
463 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
464 KVM_REG_ARM_VFP_FPSCR;
465 fpscr = vfp_get_fpscr(env);
466 r.addr = (uintptr_t)&fpscr;
467 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
472 /* Note that we do not call write_cpustate_to_list()
473 * here, so we are only writing the tuple list back to
474 * KVM. This is safe because nothing can change the
475 * CPUARMState cp15 fields (in particular gdb accesses cannot)
476 * and so there are no changes to sync. In fact syncing would
477 * be wrong at this point: for a constant register where TCG and
478 * KVM disagree about its value, the preceding write_list_to_cpustate()
479 * would not have had any effect on the CPUARMState value (since the
480 * register is read-only), and a write_cpustate_to_list() here would
481 * then try to write the TCG value back into KVM -- this would either
482 * fail or incorrectly change the value the guest sees.
484 * If we ever want to allow the user to modify cp15 registers via
485 * the gdb stub, we would need to be more clever here (for instance
486 * tracking the set of registers kvm_arch_get_registers() successfully
487 * managed to update the CPUARMState with, and only allowing those
488 * to be written back up into the kernel).
490 if (!write_list_to_kvmstate(cpu)) {
497 int kvm_arch_get_registers(CPUState *cs)
499 ARMCPU *cpu = ARM_CPU(cs);
500 CPUARMState *env = &cpu->env;
501 struct kvm_one_reg r;
504 uint32_t cpsr, fpscr;
506 for (i = 0; i < ARRAY_SIZE(regs); i++) {
508 r.addr = (uintptr_t)(env) + regs[i].offset;
509 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
515 /* Special cases which aren't a single CPUARMState field */
516 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
517 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
518 r.addr = (uintptr_t)(&cpsr);
519 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
523 cpsr_write(env, cpsr, 0xffffffff);
525 /* Make sure the current mode regs are properly set */
526 mode = env->uncached_cpsr & CPSR_M;
527 bn = bank_number(mode);
528 if (mode == ARM_CPU_MODE_FIQ) {
529 memcpy(env->regs + 8, env->fiq_regs, 5 * sizeof(uint32_t));
531 memcpy(env->regs + 8, env->usr_regs, 5 * sizeof(uint32_t));
533 env->regs[13] = env->banked_r13[bn];
534 env->regs[14] = env->banked_r14[bn];
535 env->spsr = env->banked_spsr[bn];
538 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
539 for (i = 0; i < 32; i++) {
540 r.addr = (uintptr_t)(&env->vfp.regs[i]);
541 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
548 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
549 KVM_REG_ARM_VFP_FPSCR;
550 r.addr = (uintptr_t)&fpscr;
551 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
555 vfp_set_fpscr(env, fpscr);
557 if (!write_kvmstate_to_list(cpu)) {
560 /* Note that it's OK to have registers which aren't in CPUState,
561 * so we can ignore a failure return here.
563 write_list_to_cpustate(cpu);
568 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
572 void kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
576 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
581 void kvm_arch_reset_vcpu(CPUState *cs)
583 /* Feed the kernel back its initial register state */
584 ARMCPU *cpu = ARM_CPU(cs);
586 memmove(cpu->cpreg_values, cpu->cpreg_reset_values,
587 cpu->cpreg_array_len * sizeof(cpu->cpreg_values[0]));
589 if (!write_list_to_kvmstate(cpu)) {
594 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
599 int kvm_arch_process_async_events(CPUState *cs)
604 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
609 int kvm_arch_on_sigbus(int code, void *addr)
614 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
616 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
619 int kvm_arch_insert_sw_breakpoint(CPUState *cs,
620 struct kvm_sw_breakpoint *bp)
622 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
626 int kvm_arch_insert_hw_breakpoint(target_ulong addr,
627 target_ulong len, int type)
629 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
633 int kvm_arch_remove_hw_breakpoint(target_ulong addr,
634 target_ulong len, int type)
636 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
640 int kvm_arch_remove_sw_breakpoint(CPUState *cs,
641 struct kvm_sw_breakpoint *bp)
643 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
647 void kvm_arch_remove_all_hw_breakpoints(void)
649 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
652 void kvm_arch_init_irq_routing(KVMState *s)