xref: /openbmc/qemu/target/arm/kvm.c (revision ab938ae4)
1 /*
2  * ARM implementation of KVM hooks
3  *
4  * Copyright Christoffer Dall 2009-2010
5  *
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.
8  *
9  */
10 
11 #include "qemu/osdep.h"
12 #include <sys/ioctl.h>
13 
14 #include <linux/kvm.h>
15 
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"
21 #include "kvm_arm.h"
22 #include "cpu.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"
28 #include "qemu/log.h"
29 
30 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
31     KVM_CAP_LAST_INFO
32 };
33 
34 static bool cap_has_mp_state;
35 
36 int kvm_arm_vcpu_init(CPUState *cs)
37 {
38     ARMCPU *cpu = ARM_CPU(cs);
39     struct kvm_vcpu_init init;
40 
41     init.target = cpu->kvm_target;
42     memcpy(init.features, cpu->kvm_init_features, sizeof(init.features));
43 
44     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
45 }
46 
47 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try,
48                                       int *fdarray,
49                                       struct kvm_vcpu_init *init)
50 {
51     int ret, kvmfd = -1, vmfd = -1, cpufd = -1;
52 
53     kvmfd = qemu_open("/dev/kvm", O_RDWR);
54     if (kvmfd < 0) {
55         goto err;
56     }
57     vmfd = ioctl(kvmfd, KVM_CREATE_VM, 0);
58     if (vmfd < 0) {
59         goto err;
60     }
61     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
62     if (cpufd < 0) {
63         goto err;
64     }
65 
66     if (!init) {
67         /* Caller doesn't want the VCPU to be initialized, so skip it */
68         goto finish;
69     }
70 
71     ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, init);
72     if (ret >= 0) {
73         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
74         if (ret < 0) {
75             goto err;
76         }
77     } else if (cpus_to_try) {
78         /* Old kernel which doesn't know about the
79          * PREFERRED_TARGET ioctl: we know it will only support
80          * creating one kind of guest CPU which is its preferred
81          * CPU type.
82          */
83         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
84             init->target = *cpus_to_try++;
85             memset(init->features, 0, sizeof(init->features));
86             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
87             if (ret >= 0) {
88                 break;
89             }
90         }
91         if (ret < 0) {
92             goto err;
93         }
94     } else {
95         /* Treat a NULL cpus_to_try argument the same as an empty
96          * list, which means we will fail the call since this must
97          * be an old kernel which doesn't support PREFERRED_TARGET.
98          */
99         goto err;
100     }
101 
102 finish:
103     fdarray[0] = kvmfd;
104     fdarray[1] = vmfd;
105     fdarray[2] = cpufd;
106 
107     return true;
108 
109 err:
110     if (cpufd >= 0) {
111         close(cpufd);
112     }
113     if (vmfd >= 0) {
114         close(vmfd);
115     }
116     if (kvmfd >= 0) {
117         close(kvmfd);
118     }
119 
120     return false;
121 }
122 
123 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
124 {
125     int i;
126 
127     for (i = 2; i >= 0; i--) {
128         close(fdarray[i]);
129     }
130 }
131 
132 static void kvm_arm_host_cpu_class_init(ObjectClass *oc, void *data)
133 {
134     ARMHostCPUClass *ahcc = ARM_HOST_CPU_CLASS(oc);
135 
136     /* All we really need to set up for the 'host' CPU
137      * is the feature bits -- we rely on the fact that the
138      * various ID register values in ARMCPU are only used for
139      * TCG CPUs.
140      */
141     if (!kvm_arm_get_host_cpu_features(ahcc)) {
142         fprintf(stderr, "Failed to retrieve host CPU features!\n");
143         abort();
144     }
145 }
146 
147 static void kvm_arm_host_cpu_initfn(Object *obj)
148 {
149     ARMHostCPUClass *ahcc = ARM_HOST_CPU_GET_CLASS(obj);
150     ARMCPU *cpu = ARM_CPU(obj);
151     CPUARMState *env = &cpu->env;
152 
153     cpu->kvm_target = ahcc->target;
154     cpu->dtb_compatible = ahcc->dtb_compatible;
155     env->features = ahcc->features;
156 }
157 
158 static const TypeInfo host_arm_cpu_type_info = {
159     .name = TYPE_ARM_HOST_CPU,
160 #ifdef TARGET_AARCH64
161     .parent = TYPE_AARCH64_CPU,
162 #else
163     .parent = TYPE_ARM_CPU,
164 #endif
165     .instance_init = kvm_arm_host_cpu_initfn,
166     .class_init = kvm_arm_host_cpu_class_init,
167     .class_size = sizeof(ARMHostCPUClass),
168 };
169 
170 int kvm_arch_init(MachineState *ms, KVMState *s)
171 {
172     /* For ARM interrupt delivery is always asynchronous,
173      * whether we are using an in-kernel VGIC or not.
174      */
175     kvm_async_interrupts_allowed = true;
176 
177     /*
178      * PSCI wakes up secondary cores, so we always need to
179      * have vCPUs waiting in kernel space
180      */
181     kvm_halt_in_kernel_allowed = true;
182 
183     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
184 
185     type_register_static(&host_arm_cpu_type_info);
186 
187     return 0;
188 }
189 
190 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
191 {
192     return cpu->cpu_index;
193 }
194 
195 /* We track all the KVM devices which need their memory addresses
196  * passing to the kernel in a list of these structures.
197  * When board init is complete we run through the list and
198  * tell the kernel the base addresses of the memory regions.
199  * We use a MemoryListener to track mapping and unmapping of
200  * the regions during board creation, so the board models don't
201  * need to do anything special for the KVM case.
202  */
203 typedef struct KVMDevice {
204     struct kvm_arm_device_addr kda;
205     struct kvm_device_attr kdattr;
206     MemoryRegion *mr;
207     QSLIST_ENTRY(KVMDevice) entries;
208     int dev_fd;
209 } KVMDevice;
210 
211 static QSLIST_HEAD(kvm_devices_head, KVMDevice) kvm_devices_head;
212 
213 static void kvm_arm_devlistener_add(MemoryListener *listener,
214                                     MemoryRegionSection *section)
215 {
216     KVMDevice *kd;
217 
218     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
219         if (section->mr == kd->mr) {
220             kd->kda.addr = section->offset_within_address_space;
221         }
222     }
223 }
224 
225 static void kvm_arm_devlistener_del(MemoryListener *listener,
226                                     MemoryRegionSection *section)
227 {
228     KVMDevice *kd;
229 
230     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
231         if (section->mr == kd->mr) {
232             kd->kda.addr = -1;
233         }
234     }
235 }
236 
237 static MemoryListener devlistener = {
238     .region_add = kvm_arm_devlistener_add,
239     .region_del = kvm_arm_devlistener_del,
240 };
241 
242 static void kvm_arm_set_device_addr(KVMDevice *kd)
243 {
244     struct kvm_device_attr *attr = &kd->kdattr;
245     int ret;
246 
247     /* If the device control API is available and we have a device fd on the
248      * KVMDevice struct, let's use the newer API
249      */
250     if (kd->dev_fd >= 0) {
251         uint64_t addr = kd->kda.addr;
252         attr->addr = (uintptr_t)&addr;
253         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
254     } else {
255         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
256     }
257 
258     if (ret < 0) {
259         fprintf(stderr, "Failed to set device address: %s\n",
260                 strerror(-ret));
261         abort();
262     }
263 }
264 
265 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
266 {
267     KVMDevice *kd, *tkd;
268 
269     memory_listener_unregister(&devlistener);
270     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
271         if (kd->kda.addr != -1) {
272             kvm_arm_set_device_addr(kd);
273         }
274         memory_region_unref(kd->mr);
275         g_free(kd);
276     }
277 }
278 
279 static Notifier notify = {
280     .notify = kvm_arm_machine_init_done,
281 };
282 
283 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
284                              uint64_t attr, int dev_fd)
285 {
286     KVMDevice *kd;
287 
288     if (!kvm_irqchip_in_kernel()) {
289         return;
290     }
291 
292     if (QSLIST_EMPTY(&kvm_devices_head)) {
293         memory_listener_register(&devlistener, &address_space_memory);
294         qemu_add_machine_init_done_notifier(&notify);
295     }
296     kd = g_new0(KVMDevice, 1);
297     kd->mr = mr;
298     kd->kda.id = devid;
299     kd->kda.addr = -1;
300     kd->kdattr.flags = 0;
301     kd->kdattr.group = group;
302     kd->kdattr.attr = attr;
303     kd->dev_fd = dev_fd;
304     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
305     memory_region_ref(kd->mr);
306 }
307 
308 static int compare_u64(const void *a, const void *b)
309 {
310     if (*(uint64_t *)a > *(uint64_t *)b) {
311         return 1;
312     }
313     if (*(uint64_t *)a < *(uint64_t *)b) {
314         return -1;
315     }
316     return 0;
317 }
318 
319 /* Initialize the CPUState's cpreg list according to the kernel's
320  * definition of what CPU registers it knows about (and throw away
321  * the previous TCG-created cpreg list).
322  */
323 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
324 {
325     struct kvm_reg_list rl;
326     struct kvm_reg_list *rlp;
327     int i, ret, arraylen;
328     CPUState *cs = CPU(cpu);
329 
330     rl.n = 0;
331     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
332     if (ret != -E2BIG) {
333         return ret;
334     }
335     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
336     rlp->n = rl.n;
337     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
338     if (ret) {
339         goto out;
340     }
341     /* Sort the list we get back from the kernel, since cpreg_tuples
342      * must be in strictly ascending order.
343      */
344     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
345 
346     for (i = 0, arraylen = 0; i < rlp->n; i++) {
347         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
348             continue;
349         }
350         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
351         case KVM_REG_SIZE_U32:
352         case KVM_REG_SIZE_U64:
353             break;
354         default:
355             fprintf(stderr, "Can't handle size of register in kernel list\n");
356             ret = -EINVAL;
357             goto out;
358         }
359 
360         arraylen++;
361     }
362 
363     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
364     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
365     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
366                                          arraylen);
367     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
368                                         arraylen);
369     cpu->cpreg_array_len = arraylen;
370     cpu->cpreg_vmstate_array_len = arraylen;
371 
372     for (i = 0, arraylen = 0; i < rlp->n; i++) {
373         uint64_t regidx = rlp->reg[i];
374         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
375             continue;
376         }
377         cpu->cpreg_indexes[arraylen] = regidx;
378         arraylen++;
379     }
380     assert(cpu->cpreg_array_len == arraylen);
381 
382     if (!write_kvmstate_to_list(cpu)) {
383         /* Shouldn't happen unless kernel is inconsistent about
384          * what registers exist.
385          */
386         fprintf(stderr, "Initial read of kernel register state failed\n");
387         ret = -EINVAL;
388         goto out;
389     }
390 
391 out:
392     g_free(rlp);
393     return ret;
394 }
395 
396 bool write_kvmstate_to_list(ARMCPU *cpu)
397 {
398     CPUState *cs = CPU(cpu);
399     int i;
400     bool ok = true;
401 
402     for (i = 0; i < cpu->cpreg_array_len; i++) {
403         struct kvm_one_reg r;
404         uint64_t regidx = cpu->cpreg_indexes[i];
405         uint32_t v32;
406         int ret;
407 
408         r.id = regidx;
409 
410         switch (regidx & KVM_REG_SIZE_MASK) {
411         case KVM_REG_SIZE_U32:
412             r.addr = (uintptr_t)&v32;
413             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
414             if (!ret) {
415                 cpu->cpreg_values[i] = v32;
416             }
417             break;
418         case KVM_REG_SIZE_U64:
419             r.addr = (uintptr_t)(cpu->cpreg_values + i);
420             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
421             break;
422         default:
423             abort();
424         }
425         if (ret) {
426             ok = false;
427         }
428     }
429     return ok;
430 }
431 
432 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
433 {
434     CPUState *cs = CPU(cpu);
435     int i;
436     bool ok = true;
437 
438     for (i = 0; i < cpu->cpreg_array_len; i++) {
439         struct kvm_one_reg r;
440         uint64_t regidx = cpu->cpreg_indexes[i];
441         uint32_t v32;
442         int ret;
443 
444         if (kvm_arm_cpreg_level(regidx) > level) {
445             continue;
446         }
447 
448         r.id = regidx;
449         switch (regidx & KVM_REG_SIZE_MASK) {
450         case KVM_REG_SIZE_U32:
451             v32 = cpu->cpreg_values[i];
452             r.addr = (uintptr_t)&v32;
453             break;
454         case KVM_REG_SIZE_U64:
455             r.addr = (uintptr_t)(cpu->cpreg_values + i);
456             break;
457         default:
458             abort();
459         }
460         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
461         if (ret) {
462             /* We might fail for "unknown register" and also for
463              * "you tried to set a register which is constant with
464              * a different value from what it actually contains".
465              */
466             ok = false;
467         }
468     }
469     return ok;
470 }
471 
472 void kvm_arm_reset_vcpu(ARMCPU *cpu)
473 {
474     int ret;
475 
476     /* Re-init VCPU so that all registers are set to
477      * their respective reset values.
478      */
479     ret = kvm_arm_vcpu_init(CPU(cpu));
480     if (ret < 0) {
481         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
482         abort();
483     }
484     if (!write_kvmstate_to_list(cpu)) {
485         fprintf(stderr, "write_kvmstate_to_list failed\n");
486         abort();
487     }
488 }
489 
490 /*
491  * Update KVM's MP_STATE based on what QEMU thinks it is
492  */
493 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
494 {
495     if (cap_has_mp_state) {
496         struct kvm_mp_state mp_state = {
497             .mp_state = (cpu->power_state == PSCI_OFF) ?
498             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
499         };
500         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
501         if (ret) {
502             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
503                     __func__, ret, strerror(-ret));
504             return -1;
505         }
506     }
507 
508     return 0;
509 }
510 
511 /*
512  * Sync the KVM MP_STATE into QEMU
513  */
514 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
515 {
516     if (cap_has_mp_state) {
517         struct kvm_mp_state mp_state;
518         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
519         if (ret) {
520             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
521                     __func__, ret, strerror(-ret));
522             abort();
523         }
524         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
525             PSCI_OFF : PSCI_ON;
526     }
527 
528     return 0;
529 }
530 
531 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
532 {
533 }
534 
535 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
536 {
537     ARMCPU *cpu;
538     uint32_t switched_level;
539 
540     if (kvm_irqchip_in_kernel()) {
541         /*
542          * We only need to sync timer states with user-space interrupt
543          * controllers, so return early and save cycles if we don't.
544          */
545         return MEMTXATTRS_UNSPECIFIED;
546     }
547 
548     cpu = ARM_CPU(cs);
549 
550     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
551     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
552         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
553 
554         qemu_mutex_lock_iothread();
555 
556         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
557             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
558                          !!(run->s.regs.device_irq_level &
559                             KVM_ARM_DEV_EL1_VTIMER));
560             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
561         }
562 
563         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
564             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
565                          !!(run->s.regs.device_irq_level &
566                             KVM_ARM_DEV_EL1_PTIMER));
567             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
568         }
569 
570         if (switched_level & KVM_ARM_DEV_PMU) {
571             qemu_set_irq(cpu->pmu_interrupt,
572                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
573             switched_level &= ~KVM_ARM_DEV_PMU;
574         }
575 
576         if (switched_level) {
577             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
578                           __func__, switched_level);
579         }
580 
581         /* We also mark unknown levels as processed to not waste cycles */
582         cpu->device_irq_level = run->s.regs.device_irq_level;
583         qemu_mutex_unlock_iothread();
584     }
585 
586     return MEMTXATTRS_UNSPECIFIED;
587 }
588 
589 
590 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
591 {
592     int ret = 0;
593 
594     switch (run->exit_reason) {
595     case KVM_EXIT_DEBUG:
596         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
597             ret = EXCP_DEBUG;
598         } /* otherwise return to guest */
599         break;
600     default:
601         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
602                       __func__, run->exit_reason);
603         break;
604     }
605     return ret;
606 }
607 
608 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
609 {
610     return true;
611 }
612 
613 int kvm_arch_process_async_events(CPUState *cs)
614 {
615     return 0;
616 }
617 
618 /* The #ifdef protections are until 32bit headers are imported and can
619  * be removed once both 32 and 64 bit reach feature parity.
620  */
621 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
622 {
623 #ifdef KVM_GUESTDBG_USE_SW_BP
624     if (kvm_sw_breakpoints_active(cs)) {
625         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
626     }
627 #endif
628 #ifdef KVM_GUESTDBG_USE_HW
629     if (kvm_arm_hw_debug_active(cs)) {
630         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
631         kvm_arm_copy_hw_debug_data(&dbg->arch);
632     }
633 #endif
634 }
635 
636 void kvm_arch_init_irq_routing(KVMState *s)
637 {
638 }
639 
640 int kvm_arch_irqchip_create(MachineState *ms, KVMState *s)
641 {
642      if (machine_kernel_irqchip_split(ms)) {
643          perror("-machine kernel_irqchip=split is not supported on ARM.");
644          exit(1);
645     }
646 
647     /* If we can create the VGIC using the newer device control API, we
648      * let the device do this when it initializes itself, otherwise we
649      * fall back to the old API */
650     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
651 }
652 
653 int kvm_arm_vgic_probe(void)
654 {
655     if (kvm_create_device(kvm_state,
656                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
657         return 3;
658     } else if (kvm_create_device(kvm_state,
659                                  KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
660         return 2;
661     } else {
662         return 0;
663     }
664 }
665 
666 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
667                              uint64_t address, uint32_t data, PCIDevice *dev)
668 {
669     return 0;
670 }
671 
672 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
673                                 int vector, PCIDevice *dev)
674 {
675     return 0;
676 }
677 
678 int kvm_arch_release_virq_post(int virq)
679 {
680     return 0;
681 }
682 
683 int kvm_arch_msi_data_to_gsi(uint32_t data)
684 {
685     return (data - 32) & 0xffff;
686 }
687