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 return *(uint64_t *)a - *(uint64_t *)b;
73 int kvm_arch_init_vcpu(CPUState *cs)
75 struct kvm_vcpu_init init;
79 struct kvm_reg_list rl;
80 struct kvm_reg_list *rlp;
81 ARMCPU *cpu = ARM_CPU(cs);
83 init.target = KVM_ARM_TARGET_CORTEX_A15;
84 memset(init.features, 0, sizeof(init.features));
85 ret = kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
89 /* Query the kernel to make sure it supports 32 VFP
90 * registers: QEMU's "cortex-a15" CPU is always a
91 * VFP-D32 core. The simplest way to do this is just
92 * to attempt to read register d31.
94 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP | 31;
95 r.addr = (uintptr_t)(&v);
96 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
101 /* Populate the cpreg list based on the kernel's idea
102 * of what registers exist (and throw away the TCG-created list).
105 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
109 rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
111 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
115 /* Sort the list we get back from the kernel, since cpreg_tuples
116 * must be in strictly ascending order.
118 qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
120 for (i = 0, arraylen = 0; i < rlp->n; i++) {
121 if (!reg_syncs_via_tuple_list(rlp->reg[i])) {
124 switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
125 case KVM_REG_SIZE_U32:
126 case KVM_REG_SIZE_U64:
129 fprintf(stderr, "Can't handle size of register in kernel list\n");
137 cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
138 cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
139 cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
141 cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
143 cpu->cpreg_array_len = arraylen;
144 cpu->cpreg_vmstate_array_len = arraylen;
146 for (i = 0, arraylen = 0; i < rlp->n; i++) {
147 uint64_t regidx = rlp->reg[i];
148 if (!reg_syncs_via_tuple_list(regidx)) {
151 cpu->cpreg_indexes[arraylen] = regidx;
154 assert(cpu->cpreg_array_len == arraylen);
156 if (!write_kvmstate_to_list(cpu)) {
157 /* Shouldn't happen unless kernel is inconsistent about
158 * what registers exist.
160 fprintf(stderr, "Initial read of kernel register state failed\n");
165 /* Save a copy of the initial register values so that we can
166 * feed it back to the kernel on VCPU reset.
168 cpu->cpreg_reset_values = g_memdup(cpu->cpreg_values,
169 cpu->cpreg_array_len *
170 sizeof(cpu->cpreg_values[0]));
177 /* We track all the KVM devices which need their memory addresses
178 * passing to the kernel in a list of these structures.
179 * When board init is complete we run through the list and
180 * tell the kernel the base addresses of the memory regions.
181 * We use a MemoryListener to track mapping and unmapping of
182 * the regions during board creation, so the board models don't
183 * need to do anything special for the KVM case.
185 typedef struct KVMDevice {
186 struct kvm_arm_device_addr kda;
188 QSLIST_ENTRY(KVMDevice) entries;
191 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
193 static void kvm_arm_devlistener_add(MemoryListener *listener,
194 MemoryRegionSection *section)
198 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
199 if (section->mr == kd->mr) {
200 kd->kda.addr = section->offset_within_address_space;
205 static void kvm_arm_devlistener_del(MemoryListener *listener,
206 MemoryRegionSection *section)
210 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
211 if (section->mr == kd->mr) {
217 static MemoryListener devlistener = {
218 .region_add = kvm_arm_devlistener_add,
219 .region_del = kvm_arm_devlistener_del,
222 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
226 memory_listener_unregister(&devlistener);
227 QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
228 if (kd->kda.addr != -1) {
229 if (kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR,
231 fprintf(stderr, "KVM_ARM_SET_DEVICE_ADDRESS failed: %s\n",
236 memory_region_unref(kd->mr);
241 static Notifier notify = {
242 .notify = kvm_arm_machine_init_done,
245 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid)
249 if (!kvm_irqchip_in_kernel()) {
253 if (QSLIST_EMPTY(&kvm_devices_head)) {
254 memory_listener_register(&devlistener, NULL);
255 qemu_add_machine_init_done_notifier(¬ify);
257 kd = g_new0(KVMDevice, 1);
261 QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
262 memory_region_ref(kd->mr);
265 bool write_kvmstate_to_list(ARMCPU *cpu)
267 CPUState *cs = CPU(cpu);
271 for (i = 0; i < cpu->cpreg_array_len; i++) {
272 struct kvm_one_reg r;
273 uint64_t regidx = cpu->cpreg_indexes[i];
279 switch (regidx & KVM_REG_SIZE_MASK) {
280 case KVM_REG_SIZE_U32:
281 r.addr = (uintptr_t)&v32;
282 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
284 cpu->cpreg_values[i] = v32;
287 case KVM_REG_SIZE_U64:
288 r.addr = (uintptr_t)(cpu->cpreg_values + i);
289 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
301 bool write_list_to_kvmstate(ARMCPU *cpu)
303 CPUState *cs = CPU(cpu);
307 for (i = 0; i < cpu->cpreg_array_len; i++) {
308 struct kvm_one_reg r;
309 uint64_t regidx = cpu->cpreg_indexes[i];
314 switch (regidx & KVM_REG_SIZE_MASK) {
315 case KVM_REG_SIZE_U32:
316 v32 = cpu->cpreg_values[i];
317 r.addr = (uintptr_t)&v32;
319 case KVM_REG_SIZE_U64:
320 r.addr = (uintptr_t)(cpu->cpreg_values + i);
325 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
327 /* We might fail for "unknown register" and also for
328 * "you tried to set a register which is constant with
329 * a different value from what it actually contains".
342 #define COREREG(KERNELNAME, QEMUFIELD) \
344 KVM_REG_ARM | KVM_REG_SIZE_U32 | \
345 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
346 offsetof(CPUARMState, QEMUFIELD) \
349 #define VFPSYSREG(R) \
351 KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP | \
352 KVM_REG_ARM_VFP_##R, \
353 offsetof(CPUARMState, vfp.xregs[ARM_VFP_##R]) \
356 static const Reg regs[] = {
357 /* R0_usr .. R14_usr */
358 COREREG(usr_regs.uregs[0], regs[0]),
359 COREREG(usr_regs.uregs[1], regs[1]),
360 COREREG(usr_regs.uregs[2], regs[2]),
361 COREREG(usr_regs.uregs[3], regs[3]),
362 COREREG(usr_regs.uregs[4], regs[4]),
363 COREREG(usr_regs.uregs[5], regs[5]),
364 COREREG(usr_regs.uregs[6], regs[6]),
365 COREREG(usr_regs.uregs[7], regs[7]),
366 COREREG(usr_regs.uregs[8], usr_regs[0]),
367 COREREG(usr_regs.uregs[9], usr_regs[1]),
368 COREREG(usr_regs.uregs[10], usr_regs[2]),
369 COREREG(usr_regs.uregs[11], usr_regs[3]),
370 COREREG(usr_regs.uregs[12], usr_regs[4]),
371 COREREG(usr_regs.uregs[13], banked_r13[0]),
372 COREREG(usr_regs.uregs[14], banked_r14[0]),
373 /* R13, R14, SPSR for SVC, ABT, UND, IRQ banks */
374 COREREG(svc_regs[0], banked_r13[1]),
375 COREREG(svc_regs[1], banked_r14[1]),
376 COREREG(svc_regs[2], banked_spsr[1]),
377 COREREG(abt_regs[0], banked_r13[2]),
378 COREREG(abt_regs[1], banked_r14[2]),
379 COREREG(abt_regs[2], banked_spsr[2]),
380 COREREG(und_regs[0], banked_r13[3]),
381 COREREG(und_regs[1], banked_r14[3]),
382 COREREG(und_regs[2], banked_spsr[3]),
383 COREREG(irq_regs[0], banked_r13[4]),
384 COREREG(irq_regs[1], banked_r14[4]),
385 COREREG(irq_regs[2], banked_spsr[4]),
386 /* R8_fiq .. R14_fiq and SPSR_fiq */
387 COREREG(fiq_regs[0], fiq_regs[0]),
388 COREREG(fiq_regs[1], fiq_regs[1]),
389 COREREG(fiq_regs[2], fiq_regs[2]),
390 COREREG(fiq_regs[3], fiq_regs[3]),
391 COREREG(fiq_regs[4], fiq_regs[4]),
392 COREREG(fiq_regs[5], banked_r13[5]),
393 COREREG(fiq_regs[6], banked_r14[5]),
394 COREREG(fiq_regs[7], banked_spsr[5]),
396 COREREG(usr_regs.uregs[15], regs[15]),
397 /* VFP system registers */
406 int kvm_arch_put_registers(CPUState *cs, int level)
408 ARMCPU *cpu = ARM_CPU(cs);
409 CPUARMState *env = &cpu->env;
410 struct kvm_one_reg r;
413 uint32_t cpsr, fpscr;
415 /* Make sure the banked regs are properly set */
416 mode = env->uncached_cpsr & CPSR_M;
417 bn = bank_number(mode);
418 if (mode == ARM_CPU_MODE_FIQ) {
419 memcpy(env->fiq_regs, env->regs + 8, 5 * sizeof(uint32_t));
421 memcpy(env->usr_regs, env->regs + 8, 5 * sizeof(uint32_t));
423 env->banked_r13[bn] = env->regs[13];
424 env->banked_r14[bn] = env->regs[14];
425 env->banked_spsr[bn] = env->spsr;
427 /* Now we can safely copy stuff down to the kernel */
428 for (i = 0; i < ARRAY_SIZE(regs); i++) {
430 r.addr = (uintptr_t)(env) + regs[i].offset;
431 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
437 /* Special cases which aren't a single CPUARMState field */
438 cpsr = cpsr_read(env);
439 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
440 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
441 r.addr = (uintptr_t)(&cpsr);
442 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
448 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
449 for (i = 0; i < 32; i++) {
450 r.addr = (uintptr_t)(&env->vfp.regs[i]);
451 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
458 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
459 KVM_REG_ARM_VFP_FPSCR;
460 fpscr = vfp_get_fpscr(env);
461 r.addr = (uintptr_t)&fpscr;
462 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
467 /* Note that we do not call write_cpustate_to_list()
468 * here, so we are only writing the tuple list back to
469 * KVM. This is safe because nothing can change the
470 * CPUARMState cp15 fields (in particular gdb accesses cannot)
471 * and so there are no changes to sync. In fact syncing would
472 * be wrong at this point: for a constant register where TCG and
473 * KVM disagree about its value, the preceding write_list_to_cpustate()
474 * would not have had any effect on the CPUARMState value (since the
475 * register is read-only), and a write_cpustate_to_list() here would
476 * then try to write the TCG value back into KVM -- this would either
477 * fail or incorrectly change the value the guest sees.
479 * If we ever want to allow the user to modify cp15 registers via
480 * the gdb stub, we would need to be more clever here (for instance
481 * tracking the set of registers kvm_arch_get_registers() successfully
482 * managed to update the CPUARMState with, and only allowing those
483 * to be written back up into the kernel).
485 if (!write_list_to_kvmstate(cpu)) {
492 int kvm_arch_get_registers(CPUState *cs)
494 ARMCPU *cpu = ARM_CPU(cs);
495 CPUARMState *env = &cpu->env;
496 struct kvm_one_reg r;
499 uint32_t cpsr, fpscr;
501 for (i = 0; i < ARRAY_SIZE(regs); i++) {
503 r.addr = (uintptr_t)(env) + regs[i].offset;
504 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
510 /* Special cases which aren't a single CPUARMState field */
511 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
512 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
513 r.addr = (uintptr_t)(&cpsr);
514 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
518 cpsr_write(env, cpsr, 0xffffffff);
520 /* Make sure the current mode regs are properly set */
521 mode = env->uncached_cpsr & CPSR_M;
522 bn = bank_number(mode);
523 if (mode == ARM_CPU_MODE_FIQ) {
524 memcpy(env->regs + 8, env->fiq_regs, 5 * sizeof(uint32_t));
526 memcpy(env->regs + 8, env->usr_regs, 5 * sizeof(uint32_t));
528 env->regs[13] = env->banked_r13[bn];
529 env->regs[14] = env->banked_r14[bn];
530 env->spsr = env->banked_spsr[bn];
533 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
534 for (i = 0; i < 32; i++) {
535 r.addr = (uintptr_t)(&env->vfp.regs[i]);
536 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
543 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
544 KVM_REG_ARM_VFP_FPSCR;
545 r.addr = (uintptr_t)&fpscr;
546 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
550 vfp_set_fpscr(env, fpscr);
552 if (!write_kvmstate_to_list(cpu)) {
555 /* Note that it's OK to have registers which aren't in CPUState,
556 * so we can ignore a failure return here.
558 write_list_to_cpustate(cpu);
563 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
567 void kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
571 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
576 void kvm_arch_reset_vcpu(CPUState *cs)
578 /* Feed the kernel back its initial register state */
579 ARMCPU *cpu = ARM_CPU(cs);
581 memmove(cpu->cpreg_values, cpu->cpreg_reset_values,
582 cpu->cpreg_array_len * sizeof(cpu->cpreg_values[0]));
584 if (!write_list_to_kvmstate(cpu)) {
589 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
594 int kvm_arch_process_async_events(CPUState *cs)
599 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
604 int kvm_arch_on_sigbus(int code, void *addr)
609 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
611 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
614 int kvm_arch_insert_sw_breakpoint(CPUState *cs,
615 struct kvm_sw_breakpoint *bp)
617 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
621 int kvm_arch_insert_hw_breakpoint(target_ulong addr,
622 target_ulong len, int type)
624 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
628 int kvm_arch_remove_hw_breakpoint(target_ulong addr,
629 target_ulong len, int type)
631 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
635 int kvm_arch_remove_sw_breakpoint(CPUState *cs,
636 struct kvm_sw_breakpoint *bp)
638 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
642 void kvm_arch_remove_all_hw_breakpoints(void)
644 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
647 void kvm_arch_init_irq_routing(KVMState *s)