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 init.target = KVM_ARM_TARGET_CORTEX_A15;
81 memset(init.features, 0, sizeof(init.features));
82 ret = kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
86 /* Query the kernel to make sure it supports 32 VFP
87 * registers: QEMU's "cortex-a15" CPU is always a
88 * VFP-D32 core. The simplest way to do this is just
89 * to attempt to read register d31.
91 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP | 31;
92 r.addr = (uintptr_t)(&v);
93 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
98 /* Populate the cpreg list based on the kernel's idea
99 * of what registers exist (and throw away the TCG-created list).
102 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
106 rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
108 ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
112 /* Sort the list we get back from the kernel, since cpreg_tuples
113 * must be in strictly ascending order.
115 qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
117 for (i = 0, arraylen = 0; i < rlp->n; i++) {
118 if (!reg_syncs_via_tuple_list(rlp->reg[i])) {
121 switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
122 case KVM_REG_SIZE_U32:
123 case KVM_REG_SIZE_U64:
126 fprintf(stderr, "Can't handle size of register in kernel list\n");
134 cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
135 cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
136 cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
138 cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
140 cpu->cpreg_array_len = arraylen;
141 cpu->cpreg_vmstate_array_len = arraylen;
143 for (i = 0, arraylen = 0; i < rlp->n; i++) {
144 uint64_t regidx = rlp->reg[i];
145 if (!reg_syncs_via_tuple_list(regidx)) {
148 cpu->cpreg_indexes[arraylen] = regidx;
151 assert(cpu->cpreg_array_len == arraylen);
153 if (!write_kvmstate_to_list(cpu)) {
154 /* Shouldn't happen unless kernel is inconsistent about
155 * what registers exist.
157 fprintf(stderr, "Initial read of kernel register state failed\n");
162 /* Save a copy of the initial register values so that we can
163 * feed it back to the kernel on VCPU reset.
165 cpu->cpreg_reset_values = g_memdup(cpu->cpreg_values,
166 cpu->cpreg_array_len *
167 sizeof(cpu->cpreg_values[0]));
174 /* We track all the KVM devices which need their memory addresses
175 * passing to the kernel in a list of these structures.
176 * When board init is complete we run through the list and
177 * tell the kernel the base addresses of the memory regions.
178 * We use a MemoryListener to track mapping and unmapping of
179 * the regions during board creation, so the board models don't
180 * need to do anything special for the KVM case.
182 typedef struct KVMDevice {
183 struct kvm_arm_device_addr kda;
185 QSLIST_ENTRY(KVMDevice) entries;
188 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
190 static void kvm_arm_devlistener_add(MemoryListener *listener,
191 MemoryRegionSection *section)
195 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
196 if (section->mr == kd->mr) {
197 kd->kda.addr = section->offset_within_address_space;
202 static void kvm_arm_devlistener_del(MemoryListener *listener,
203 MemoryRegionSection *section)
207 QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
208 if (section->mr == kd->mr) {
214 static MemoryListener devlistener = {
215 .region_add = kvm_arm_devlistener_add,
216 .region_del = kvm_arm_devlistener_del,
219 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
223 memory_listener_unregister(&devlistener);
224 QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
225 if (kd->kda.addr != -1) {
226 if (kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR,
228 fprintf(stderr, "KVM_ARM_SET_DEVICE_ADDRESS failed: %s\n",
233 memory_region_unref(kd->mr);
238 static Notifier notify = {
239 .notify = kvm_arm_machine_init_done,
242 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid)
246 if (!kvm_irqchip_in_kernel()) {
250 if (QSLIST_EMPTY(&kvm_devices_head)) {
251 memory_listener_register(&devlistener, NULL);
252 qemu_add_machine_init_done_notifier(¬ify);
254 kd = g_new0(KVMDevice, 1);
258 QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
259 memory_region_ref(kd->mr);
262 bool write_kvmstate_to_list(ARMCPU *cpu)
264 CPUState *cs = CPU(cpu);
268 for (i = 0; i < cpu->cpreg_array_len; i++) {
269 struct kvm_one_reg r;
270 uint64_t regidx = cpu->cpreg_indexes[i];
276 switch (regidx & KVM_REG_SIZE_MASK) {
277 case KVM_REG_SIZE_U32:
278 r.addr = (uintptr_t)&v32;
279 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
281 cpu->cpreg_values[i] = v32;
284 case KVM_REG_SIZE_U64:
285 r.addr = (uintptr_t)(cpu->cpreg_values + i);
286 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
298 bool write_list_to_kvmstate(ARMCPU *cpu)
300 CPUState *cs = CPU(cpu);
304 for (i = 0; i < cpu->cpreg_array_len; i++) {
305 struct kvm_one_reg r;
306 uint64_t regidx = cpu->cpreg_indexes[i];
311 switch (regidx & KVM_REG_SIZE_MASK) {
312 case KVM_REG_SIZE_U32:
313 v32 = cpu->cpreg_values[i];
314 r.addr = (uintptr_t)&v32;
316 case KVM_REG_SIZE_U64:
317 r.addr = (uintptr_t)(cpu->cpreg_values + i);
322 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
324 /* We might fail for "unknown register" and also for
325 * "you tried to set a register which is constant with
326 * a different value from what it actually contains".
339 #define COREREG(KERNELNAME, QEMUFIELD) \
341 KVM_REG_ARM | KVM_REG_SIZE_U32 | \
342 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(KERNELNAME), \
343 offsetof(CPUARMState, QEMUFIELD) \
346 #define VFPSYSREG(R) \
348 KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP | \
349 KVM_REG_ARM_VFP_##R, \
350 offsetof(CPUARMState, vfp.xregs[ARM_VFP_##R]) \
353 static const Reg regs[] = {
354 /* R0_usr .. R14_usr */
355 COREREG(usr_regs.uregs[0], regs[0]),
356 COREREG(usr_regs.uregs[1], regs[1]),
357 COREREG(usr_regs.uregs[2], regs[2]),
358 COREREG(usr_regs.uregs[3], regs[3]),
359 COREREG(usr_regs.uregs[4], regs[4]),
360 COREREG(usr_regs.uregs[5], regs[5]),
361 COREREG(usr_regs.uregs[6], regs[6]),
362 COREREG(usr_regs.uregs[7], regs[7]),
363 COREREG(usr_regs.uregs[8], usr_regs[0]),
364 COREREG(usr_regs.uregs[9], usr_regs[1]),
365 COREREG(usr_regs.uregs[10], usr_regs[2]),
366 COREREG(usr_regs.uregs[11], usr_regs[3]),
367 COREREG(usr_regs.uregs[12], usr_regs[4]),
368 COREREG(usr_regs.uregs[13], banked_r13[0]),
369 COREREG(usr_regs.uregs[14], banked_r14[0]),
370 /* R13, R14, SPSR for SVC, ABT, UND, IRQ banks */
371 COREREG(svc_regs[0], banked_r13[1]),
372 COREREG(svc_regs[1], banked_r14[1]),
373 COREREG(svc_regs[2], banked_spsr[1]),
374 COREREG(abt_regs[0], banked_r13[2]),
375 COREREG(abt_regs[1], banked_r14[2]),
376 COREREG(abt_regs[2], banked_spsr[2]),
377 COREREG(und_regs[0], banked_r13[3]),
378 COREREG(und_regs[1], banked_r14[3]),
379 COREREG(und_regs[2], banked_spsr[3]),
380 COREREG(irq_regs[0], banked_r13[4]),
381 COREREG(irq_regs[1], banked_r14[4]),
382 COREREG(irq_regs[2], banked_spsr[4]),
383 /* R8_fiq .. R14_fiq and SPSR_fiq */
384 COREREG(fiq_regs[0], fiq_regs[0]),
385 COREREG(fiq_regs[1], fiq_regs[1]),
386 COREREG(fiq_regs[2], fiq_regs[2]),
387 COREREG(fiq_regs[3], fiq_regs[3]),
388 COREREG(fiq_regs[4], fiq_regs[4]),
389 COREREG(fiq_regs[5], banked_r13[5]),
390 COREREG(fiq_regs[6], banked_r14[5]),
391 COREREG(fiq_regs[7], banked_spsr[5]),
393 COREREG(usr_regs.uregs[15], regs[15]),
394 /* VFP system registers */
403 int kvm_arch_put_registers(CPUState *cs, int level)
405 ARMCPU *cpu = ARM_CPU(cs);
406 CPUARMState *env = &cpu->env;
407 struct kvm_one_reg r;
410 uint32_t cpsr, fpscr;
412 /* Make sure the banked regs are properly set */
413 mode = env->uncached_cpsr & CPSR_M;
414 bn = bank_number(mode);
415 if (mode == ARM_CPU_MODE_FIQ) {
416 memcpy(env->fiq_regs, env->regs + 8, 5 * sizeof(uint32_t));
418 memcpy(env->usr_regs, env->regs + 8, 5 * sizeof(uint32_t));
420 env->banked_r13[bn] = env->regs[13];
421 env->banked_r14[bn] = env->regs[14];
422 env->banked_spsr[bn] = env->spsr;
424 /* Now we can safely copy stuff down to the kernel */
425 for (i = 0; i < ARRAY_SIZE(regs); i++) {
427 r.addr = (uintptr_t)(env) + regs[i].offset;
428 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
434 /* Special cases which aren't a single CPUARMState field */
435 cpsr = cpsr_read(env);
436 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
437 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
438 r.addr = (uintptr_t)(&cpsr);
439 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
445 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
446 for (i = 0; i < 32; i++) {
447 r.addr = (uintptr_t)(&env->vfp.regs[i]);
448 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
455 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
456 KVM_REG_ARM_VFP_FPSCR;
457 fpscr = vfp_get_fpscr(env);
458 r.addr = (uintptr_t)&fpscr;
459 ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
464 /* Note that we do not call write_cpustate_to_list()
465 * here, so we are only writing the tuple list back to
466 * KVM. This is safe because nothing can change the
467 * CPUARMState cp15 fields (in particular gdb accesses cannot)
468 * and so there are no changes to sync. In fact syncing would
469 * be wrong at this point: for a constant register where TCG and
470 * KVM disagree about its value, the preceding write_list_to_cpustate()
471 * would not have had any effect on the CPUARMState value (since the
472 * register is read-only), and a write_cpustate_to_list() here would
473 * then try to write the TCG value back into KVM -- this would either
474 * fail or incorrectly change the value the guest sees.
476 * If we ever want to allow the user to modify cp15 registers via
477 * the gdb stub, we would need to be more clever here (for instance
478 * tracking the set of registers kvm_arch_get_registers() successfully
479 * managed to update the CPUARMState with, and only allowing those
480 * to be written back up into the kernel).
482 if (!write_list_to_kvmstate(cpu)) {
489 int kvm_arch_get_registers(CPUState *cs)
491 ARMCPU *cpu = ARM_CPU(cs);
492 CPUARMState *env = &cpu->env;
493 struct kvm_one_reg r;
496 uint32_t cpsr, fpscr;
498 for (i = 0; i < ARRAY_SIZE(regs); i++) {
500 r.addr = (uintptr_t)(env) + regs[i].offset;
501 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
507 /* Special cases which aren't a single CPUARMState field */
508 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 |
509 KVM_REG_ARM_CORE | KVM_REG_ARM_CORE_REG(usr_regs.ARM_cpsr);
510 r.addr = (uintptr_t)(&cpsr);
511 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
515 cpsr_write(env, cpsr, 0xffffffff);
517 /* Make sure the current mode regs are properly set */
518 mode = env->uncached_cpsr & CPSR_M;
519 bn = bank_number(mode);
520 if (mode == ARM_CPU_MODE_FIQ) {
521 memcpy(env->regs + 8, env->fiq_regs, 5 * sizeof(uint32_t));
523 memcpy(env->regs + 8, env->usr_regs, 5 * sizeof(uint32_t));
525 env->regs[13] = env->banked_r13[bn];
526 env->regs[14] = env->banked_r14[bn];
527 env->spsr = env->banked_spsr[bn];
530 r.id = KVM_REG_ARM | KVM_REG_SIZE_U64 | KVM_REG_ARM_VFP;
531 for (i = 0; i < 32; i++) {
532 r.addr = (uintptr_t)(&env->vfp.regs[i]);
533 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
540 r.id = KVM_REG_ARM | KVM_REG_SIZE_U32 | KVM_REG_ARM_VFP |
541 KVM_REG_ARM_VFP_FPSCR;
542 r.addr = (uintptr_t)&fpscr;
543 ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
547 vfp_set_fpscr(env, fpscr);
549 if (!write_kvmstate_to_list(cpu)) {
552 /* Note that it's OK to have registers which aren't in CPUState,
553 * so we can ignore a failure return here.
555 write_list_to_cpustate(cpu);
560 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
564 void kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
568 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
573 void kvm_arch_reset_vcpu(CPUState *cs)
575 /* Feed the kernel back its initial register state */
576 ARMCPU *cpu = ARM_CPU(cs);
578 memmove(cpu->cpreg_values, cpu->cpreg_reset_values,
579 cpu->cpreg_array_len * sizeof(cpu->cpreg_values[0]));
581 if (!write_list_to_kvmstate(cpu)) {
586 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
591 int kvm_arch_process_async_events(CPUState *cs)
596 int kvm_arch_on_sigbus_vcpu(CPUState *cs, int code, void *addr)
601 int kvm_arch_on_sigbus(int code, void *addr)
606 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
608 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
611 int kvm_arch_insert_sw_breakpoint(CPUState *cs,
612 struct kvm_sw_breakpoint *bp)
614 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
618 int kvm_arch_insert_hw_breakpoint(target_ulong addr,
619 target_ulong len, int type)
621 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
625 int kvm_arch_remove_hw_breakpoint(target_ulong addr,
626 target_ulong len, int type)
628 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
632 int kvm_arch_remove_sw_breakpoint(CPUState *cs,
633 struct kvm_sw_breakpoint *bp)
635 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
639 void kvm_arch_remove_all_hw_breakpoints(void)
641 qemu_log_mask(LOG_UNIMP, "%s: not implemented\n", __func__);
644 void kvm_arch_init_irq_routing(KVMState *s)