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 /* Check that cpu.h's idea of coprocessor fields matches KVM's */
27 #if (CP_REG_SIZE_SHIFT != KVM_REG_SIZE_SHIFT) || \
28 (CP_REG_SIZE_MASK != KVM_REG_SIZE_MASK) || \
29 (CP_REG_SIZE_U32 != KVM_REG_SIZE_U32) || \
30 (CP_REG_SIZE_U64 != KVM_REG_SIZE_U64) || \
31 (CP_REG_ARM != KVM_REG_ARM)
32 #error mismatch between cpu.h and KVM header definitions
35 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
39 int kvm_arch_init(KVMState *s)
41 /* For ARM interrupt delivery is always asynchronous,
42 * whether we are using an in-kernel VGIC or not.
44 kvm_async_interrupts_allowed = true;
48 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
50 return cpu->cpu_index;
53 static bool reg_syncs_via_tuple_list(uint64_t regidx)
55 /* Return true if the regidx is a register we should synchronize
56 * via the cpreg_tuples array (ie is not a core reg we sync by
57 * hand in kvm_arch_get/put_registers())
59 switch (regidx & KVM_REG_ARM_COPROC_MASK) {
60 case KVM_REG_ARM_CORE:
68 static int compare_u64(const void *a, const void *b)
70 if (*(uint64_t *)a > *(uint64_t *)b) {
73 if (*(uint64_t *)a < *(uint64_t *)b) {
79 int kvm_arch_init_vcpu(CPUState *cs)
81 struct kvm_vcpu_init init;
85 struct kvm_reg_list rl;
86 struct kvm_reg_list *rlp;
87 ARMCPU *cpu = ARM_CPU(cs);
89 init.target = KVM_ARM_TARGET_CORTEX_A15;
90 memset(init.features, 0, sizeof(init.features));
91 ret = kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
95 /* Query the kernel to make sure it supports 32 VFP
96 * registers: QEMU's "cortex-a15" CPU is always a
97 * VFP-D32 core. The simplest way to do this is just
98 * to attempt to read register d31.
100 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP | 31;
101 r.addr = (uintptr_t)(&v);
102 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
103 if (ret == -ENOENT) {
107 /* Populate the cpreg list based on the kernel's idea
108 * of what registers exist (and throw away the TCG-created list).
111 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
115 rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
117 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
121 /* Sort the list we get back from the kernel, since cpreg_tuples
122 * must be in strictly ascending order.
124 qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
126 for (i = 0, arraylen = 0; i < rlp->n; i++) {
127 if (!reg_syncs_via_tuple_list(rlp->reg[i])) {
130 switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
131 case KVM_REG_SIZE_U32:
132 case KVM_REG_SIZE_U64:
135 fprintf(stderr, "Can't handle size of register in kernel list\n");
143 cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
144 cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
145 cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
147 cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
149 cpu->cpreg_array_len = arraylen;
150 cpu->cpreg_vmstate_array_len = arraylen;
152 for (i = 0, arraylen = 0; i < rlp->n; i++) {
153 uint64_t regidx = rlp->reg[i];
154 if (!reg_syncs_via_tuple_list(regidx)) {
157 cpu->cpreg_indexes[arraylen] = regidx;
160 assert(cpu->cpreg_array_len == arraylen);
162 if (!write_kvmstate_to_list(cpu)) {
163 /* Shouldn't happen unless kernel is inconsistent about
164 * what registers exist.
166 fprintf(stderr, "Initial read of kernel register state failed\n");
171 /* Save a copy of the initial register values so that we can
172 * feed it back to the kernel on VCPU reset.
174 cpu->cpreg_reset_values = g_memdup(cpu->cpreg_values,
175 cpu->cpreg_array_len *
176 sizeof(cpu->cpreg_values[0]));
183 /* We track all the KVM devices which need their memory addresses
184 * passing to the kernel in a list of these structures.
185 * When board init is complete we run through the list and
186 * tell the kernel the base addresses of the memory regions.
187 * We use a MemoryListener to track mapping and unmapping of
188 * the regions during board creation, so the board models don't
189 * need to do anything special for the KVM case.
191 typedef struct KVMDevice {
192 struct kvm_arm_device_addr kda;
194 QSLIST_ENTRY(KVMDevice) entries;
197 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
199 static void kvm_arm_devlistener_add(MemoryListener *listener,
200 MemoryRegionSection *section)
204 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
205 if (section->mr == kd->mr) {
206 kd->kda.addr = section->offset_within_address_space;
211 static void kvm_arm_devlistener_del(MemoryListener *listener,
212 MemoryRegionSection *section)
216 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
217 if (section->mr == kd->mr) {
223 static MemoryListener devlistener = {
224 .region_add = kvm_arm_devlistener_add,
225 .region_del = kvm_arm_devlistener_del,
228 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
232 memory_listener_unregister(&devlistener);
233 QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
234 if (kd->kda.addr != -1) {
235 if (kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR,
237 fprintf(stderr, "KVM_ARM_SET_DEVICE_ADDRESS failed: %s\n",
242 memory_region_unref(kd->mr);
247 static Notifier notify = {
248 .notify = kvm_arm_machine_init_done,
251 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid)
255 if (!kvm_irqchip_in_kernel()) {
259 if (QSLIST_EMPTY(&kvm_devices_head)) {
260 memory_listener_register(&devlistener, NULL);
261 qemu_add_machine_init_done_notifier(¬ify);
263 kd = g_new0(KVMDevice, 1);
267 QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
268 memory_region_ref(kd->mr);
271 bool write_kvmstate_to_list(ARMCPU *cpu)
273 CPUState *cs = CPU(cpu);
277 for (i = 0; i < cpu->cpreg_array_len; i++) {
278 struct kvm_one_reg r;
279 uint64_t regidx = cpu->cpreg_indexes[i];
285 switch (regidx & KVM_REG_SIZE_MASK) {
286 case KVM_REG_SIZE_U32:
287 r.addr = (uintptr_t)&v32;
288 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
290 cpu->cpreg_values[i] = v32;
293 case KVM_REG_SIZE_U64:
294 r.addr = (uintptr_t)(cpu->cpreg_values + i);
295 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
307 bool write_list_to_kvmstate(ARMCPU *cpu)
309 CPUState *cs = CPU(cpu);
313 for (i = 0; i < cpu->cpreg_array_len; i++) {
314 struct kvm_one_reg r;
315 uint64_t regidx = cpu->cpreg_indexes[i];
320 switch (regidx & KVM_REG_SIZE_MASK) {
321 case KVM_REG_SIZE_U32:
322 v32 = cpu->cpreg_values[i];
323 r.addr = (uintptr_t)&v32;
325 case KVM_REG_SIZE_U64:
326 r.addr = (uintptr_t)(cpu->cpreg_values + i);
331 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
333 /* We might fail for "unknown register" and also for
334 * "you tried to set a register which is constant with
335 * a different value from what it actually contains".
348 #define COREREG(KERNELNAME, QEMUFIELD) \
350 KVM_REG_ARM | KVM_REG_SIZE_U32 | \
351 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
352 offsetof(CPUARMState, QEMUFIELD) \
355 #define VFPSYSREG(R) \
357 KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP | \
358 KVM_REG_ARM_VFP_##R, \
359 offsetof(CPUARMState, vfp.xregs[ARM_VFP_##R]) \
362 static const Reg regs[] = {
363 /* R0_usr .. R14_usr */
364 COREREG(usr_regs.uregs[0], regs[0]),
365 COREREG(usr_regs.uregs[1], regs[1]),
366 COREREG(usr_regs.uregs[2], regs[2]),
367 COREREG(usr_regs.uregs[3], regs[3]),
368 COREREG(usr_regs.uregs[4], regs[4]),
369 COREREG(usr_regs.uregs[5], regs[5]),
370 COREREG(usr_regs.uregs[6], regs[6]),
371 COREREG(usr_regs.uregs[7], regs[7]),
372 COREREG(usr_regs.uregs[8], usr_regs[0]),
373 COREREG(usr_regs.uregs[9], usr_regs[1]),
374 COREREG(usr_regs.uregs[10], usr_regs[2]),
375 COREREG(usr_regs.uregs[11], usr_regs[3]),
376 COREREG(usr_regs.uregs[12], usr_regs[4]),
377 COREREG(usr_regs.uregs[13], banked_r13[0]),
378 COREREG(usr_regs.uregs[14], banked_r14[0]),
379 /* R13, R14, SPSR for SVC, ABT, UND, IRQ banks */
380 COREREG(svc_regs[0], banked_r13[1]),
381 COREREG(svc_regs[1], banked_r14[1]),
382 COREREG(svc_regs[2], banked_spsr[1]),
383 COREREG(abt_regs[0], banked_r13[2]),
384 COREREG(abt_regs[1], banked_r14[2]),
385 COREREG(abt_regs[2], banked_spsr[2]),
386 COREREG(und_regs[0], banked_r13[3]),
387 COREREG(und_regs[1], banked_r14[3]),
388 COREREG(und_regs[2], banked_spsr[3]),
389 COREREG(irq_regs[0], banked_r13[4]),
390 COREREG(irq_regs[1], banked_r14[4]),
391 COREREG(irq_regs[2], banked_spsr[4]),
392 /* R8_fiq .. R14_fiq and SPSR_fiq */
393 COREREG(fiq_regs[0], fiq_regs[0]),
394 COREREG(fiq_regs[1], fiq_regs[1]),
395 COREREG(fiq_regs[2], fiq_regs[2]),
396 COREREG(fiq_regs[3], fiq_regs[3]),
397 COREREG(fiq_regs[4], fiq_regs[4]),
398 COREREG(fiq_regs[5], banked_r13[5]),
399 COREREG(fiq_regs[6], banked_r14[5]),
400 COREREG(fiq_regs[7], banked_spsr[5]),
402 COREREG(usr_regs.uregs[15], regs[15]),
403 /* VFP system registers */
412 int kvm_arch_put_registers(CPUState *cs, int level)
414 ARMCPU *cpu = ARM_CPU(cs);
415 CPUARMState *env = &cpu->env;
416 struct kvm_one_reg r;
419 uint32_t cpsr, fpscr;
421 /* Make sure the banked regs are properly set */
422 mode = env->uncached_cpsr & CPSR_M;
423 bn = bank_number(mode);
424 if (mode == ARM_CPU_MODE_FIQ) {
425 memcpy(env->fiq_regs, env->regs + 8, 5 * sizeof(uint32_t));
427 memcpy(env->usr_regs, env->regs + 8, 5 * sizeof(uint32_t));
429 env->banked_r13[bn] = env->regs[13];
430 env->banked_r14[bn] = env->regs[14];
431 env->banked_spsr[bn] = env->spsr;
433 /* Now we can safely copy stuff down to the kernel */
434 for (i = 0; i < ARRAY_SIZE(regs); i++) {
436 r.addr = (uintptr_t)(env) + regs[i].offset;
437 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
443 /* Special cases which aren't a single CPUARMState field */
444 cpsr = cpsr_read(env);
445 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
446 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
447 r.addr = (uintptr_t)(&cpsr);
448 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
454 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
455 for (i = 0; i < 32; i++) {
456 r.addr = (uintptr_t)(&env->vfp.regs[i]);
457 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
464 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
465 KVM_REG_ARM_VFP_FPSCR;
466 fpscr = vfp_get_fpscr(env);
467 r.addr = (uintptr_t)&fpscr;
468 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
473 /* Note that we do not call write_cpustate_to_list()
474 * here, so we are only writing the tuple list back to
475 * KVM. This is safe because nothing can change the
476 * CPUARMState cp15 fields (in particular gdb accesses cannot)
477 * and so there are no changes to sync. In fact syncing would
478 * be wrong at this point: for a constant register where TCG and
479 * KVM disagree about its value, the preceding write_list_to_cpustate()
480 * would not have had any effect on the CPUARMState value (since the
481 * register is read-only), and a write_cpustate_to_list() here would
482 * then try to write the TCG value back into KVM -- this would either
483 * fail or incorrectly change the value the guest sees.
485 * If we ever want to allow the user to modify cp15 registers via
486 * the gdb stub, we would need to be more clever here (for instance
487 * tracking the set of registers kvm_arch_get_registers() successfully
488 * managed to update the CPUARMState with, and only allowing those
489 * to be written back up into the kernel).
491 if (!write_list_to_kvmstate(cpu)) {
498 int kvm_arch_get_registers(CPUState *cs)
500 ARMCPU *cpu = ARM_CPU(cs);
501 CPUARMState *env = &cpu->env;
502 struct kvm_one_reg r;
505 uint32_t cpsr, fpscr;
507 for (i = 0; i < ARRAY_SIZE(regs); i++) {
509 r.addr = (uintptr_t)(env) + regs[i].offset;
510 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
516 /* Special cases which aren't a single CPUARMState field */
517 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
518 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
519 r.addr = (uintptr_t)(&cpsr);
520 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
524 cpsr_write(env, cpsr, 0xffffffff);
526 /* Make sure the current mode regs are properly set */
527 mode = env->uncached_cpsr & CPSR_M;
528 bn = bank_number(mode);
529 if (mode == ARM_CPU_MODE_FIQ) {
530 memcpy(env->regs + 8, env->fiq_regs, 5 * sizeof(uint32_t));
532 memcpy(env->regs + 8, env->usr_regs, 5 * sizeof(uint32_t));
534 env->regs[13] = env->banked_r13[bn];
535 env->regs[14] = env->banked_r14[bn];
536 env->spsr = env->banked_spsr[bn];
539 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
540 for (i = 0; i < 32; i++) {
541 r.addr = (uintptr_t)(&env->vfp.regs[i]);
542 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
549 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
550 KVM_REG_ARM_VFP_FPSCR;
551 r.addr = (uintptr_t)&fpscr;
552 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
556 vfp_set_fpscr(env, fpscr);
558 if (!write_kvmstate_to_list(cpu)) {
561 /* Note that it's OK to have registers which aren't in CPUState,
562 * so we can ignore a failure return here.
564 write_list_to_cpustate(cpu);
569 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
573 void kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
577 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
582 void kvm_arch_reset_vcpu(CPUState *cs)
584 /* Feed the kernel back its initial register state */
585 ARMCPU *cpu = ARM_CPU(cs);
587 memmove(cpu->cpreg_values, cpu->cpreg_reset_values,
588 cpu->cpreg_array_len * sizeof(cpu->cpreg_values[0]));
590 if (!write_list_to_kvmstate(cpu)) {
595 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
600 int kvm_arch_process_async_events(CPUState *cs)
605 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
610 int kvm_arch_on_sigbus(int code, void *addr)
615 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
617 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
620 int kvm_arch_insert_sw_breakpoint(CPUState *cs,
621 struct kvm_sw_breakpoint *bp)
623 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
627 int kvm_arch_insert_hw_breakpoint(target_ulong addr,
628 target_ulong len, int type)
630 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
634 int kvm_arch_remove_hw_breakpoint(target_ulong addr,
635 target_ulong len, int type)
637 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
641 int kvm_arch_remove_sw_breakpoint(CPUState *cs,
642 struct kvm_sw_breakpoint *bp)
644 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
648 void kvm_arch_remove_all_hw_breakpoints(void)
650 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
653 void kvm_arch_init_irq_routing(KVMState *s)