target/arm/kvm.c

   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/timer.h"
  17 #include "qemu/error-report.h"
  18 #include "qemu/main-loop.h"
  19 #include "qom/object.h"
  20 #include "qapi/error.h"
  21 #include "sysemu/sysemu.h"
  22 #include "sysemu/kvm.h"
  23 #include "sysemu/kvm_int.h"
  24 #include "kvm_arm.h"
  25 #include "cpu.h"
  26 #include "trace.h"
  27 #include "internals.h"
  28 #include "hw/pci/pci.h"
  29 #include "exec/memattrs.h"
  30 #include "exec/address-spaces.h"
  31 #include "hw/boards.h"
  32 #include "hw/irq.h"
  33 #include "qapi/visitor.h"
  34 #include "qemu/log.h"
  35
  36 const KVMCapabilityInfo kvm_arch_required_capabilities[] = {
  37     KVM_CAP_LAST_INFO
  38 };
  39
  40 static bool cap_has_mp_state;
  41 static bool cap_has_inject_serror_esr;
  42 static bool cap_has_inject_ext_dabt;
  43
  44 static ARMHostCPUFeatures arm_host_cpu_features;
  45
  46 int kvm_arm_vcpu_init(CPUState *cs)
  47 {
  48     ARMCPU *cpu = ARM_CPU(cs);
  49     struct kvm_vcpu_init init;
  50
  51     init.target = cpu->kvm_target;
  52     memcpy(init.features, cpu->kvm_init_features, sizeof(init.features));
  53
  54     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_INIT, &init);
  55 }
  56
  57 int kvm_arm_vcpu_finalize(CPUState *cs, int feature)
  58 {
  59     return kvm_vcpu_ioctl(cs, KVM_ARM_VCPU_FINALIZE, &feature);
  60 }
  61
  62 void kvm_arm_init_serror_injection(CPUState *cs)
  63 {
  64     cap_has_inject_serror_esr = kvm_check_extension(cs->kvm_state,
  65                                     KVM_CAP_ARM_INJECT_SERROR_ESR);
  66 }
  67
  68 bool kvm_arm_create_scratch_host_vcpu(const uint32_t *cpus_to_try,
  69                                       int *fdarray,
  70                                       struct kvm_vcpu_init *init)
  71 {
  72     int ret = 0, kvmfd = -1, vmfd = -1, cpufd = -1;
  73     int max_vm_pa_size;
  74
  75     kvmfd = qemu_open_old("/dev/kvm", O_RDWR);
  76     if (kvmfd < 0) {
  77         goto err;
  78     }
  79     max_vm_pa_size = ioctl(kvmfd, KVM_CHECK_EXTENSION, KVM_CAP_ARM_VM_IPA_SIZE);
  80     if (max_vm_pa_size < 0) {
  81         max_vm_pa_size = 0;
  82     }
  83     do {
  84         vmfd = ioctl(kvmfd, KVM_CREATE_VM, max_vm_pa_size);
  85     } while (vmfd == -1 && errno == EINTR);
  86     if (vmfd < 0) {
  87         goto err;
  88     }
  89     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
  90     if (cpufd < 0) {
  91         goto err;
  92     }
  93
  94     if (!init) {
  95         /* Caller doesn't want the VCPU to be initialized, so skip it */
  96         goto finish;
  97     }
  98
  99     if (init->target == -1) {
 100         struct kvm_vcpu_init preferred;
 101
 102         ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, &preferred);
 103         if (!ret) {
 104             init->target = preferred.target;
 105         }
 106     }
 107     if (ret >= 0) {
 108         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
 109         if (ret < 0) {
 110             goto err;
 111         }
 112     } else if (cpus_to_try) {
 113         /* Old kernel which doesn't know about the
 114          * PREFERRED_TARGET ioctl: we know it will only support
 115          * creating one kind of guest CPU which is its preferred
 116          * CPU type.
 117          */
 118         struct kvm_vcpu_init try;
 119
 120         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
 121             try.target = *cpus_to_try++;
 122             memcpy(try.features, init->features, sizeof(init->features));
 123             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, &try);
 124             if (ret >= 0) {
 125                 break;
 126             }
 127         }
 128         if (ret < 0) {
 129             goto err;
 130         }
 131         init->target = try.target;
 132     } else {
 133         /* Treat a NULL cpus_to_try argument the same as an empty
 134          * list, which means we will fail the call since this must
 135          * be an old kernel which doesn't support PREFERRED_TARGET.
 136          */
 137         goto err;
 138     }
 139
 140 finish:
 141     fdarray[0] = kvmfd;
 142     fdarray[1] = vmfd;
 143     fdarray[2] = cpufd;
 144
 145     return true;
 146
 147 err:
 148     if (cpufd >= 0) {
 149         close(cpufd);
 150     }
 151     if (vmfd >= 0) {
 152         close(vmfd);
 153     }
 154     if (kvmfd >= 0) {
 155         close(kvmfd);
 156     }
 157
 158     return false;
 159 }
 160
 161 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 162 {
 163     int i;
 164
 165     for (i = 2; i >= 0; i--) {
 166         close(fdarray[i]);
 167     }
 168 }
 169
 170 void kvm_arm_set_cpu_features_from_host(ARMCPU *cpu)
 171 {
 172     CPUARMState *env = &cpu->env;
 173
 174     if (!arm_host_cpu_features.dtb_compatible) {
 175         if (!kvm_enabled() ||
 176             !kvm_arm_get_host_cpu_features(&arm_host_cpu_features)) {
 177             /* We can't report this error yet, so flag that we need to
 178              * in arm_cpu_realizefn().
 179              */
 180             cpu->kvm_target = QEMU_KVM_ARM_TARGET_NONE;
 181             cpu->host_cpu_probe_failed = true;
 182             return;
 183         }
 184     }
 185
 186     cpu->kvm_target = arm_host_cpu_features.target;
 187     cpu->dtb_compatible = arm_host_cpu_features.dtb_compatible;
 188     cpu->isar = arm_host_cpu_features.isar;
 189     env->features = arm_host_cpu_features.features;
 190 }
 191
 192 static bool kvm_no_adjvtime_get(Object *obj, Error **errp)
 193 {
 194     return !ARM_CPU(obj)->kvm_adjvtime;
 195 }
 196
 197 static void kvm_no_adjvtime_set(Object *obj, bool value, Error **errp)
 198 {
 199     ARM_CPU(obj)->kvm_adjvtime = !value;
 200 }
 201
 202 static bool kvm_steal_time_get(Object *obj, Error **errp)
 203 {
 204     return ARM_CPU(obj)->kvm_steal_time != ON_OFF_AUTO_OFF;
 205 }
 206
 207 static void kvm_steal_time_set(Object *obj, bool value, Error **errp)
 208 {
 209     ARM_CPU(obj)->kvm_steal_time = value ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
 210 }
 211
 212 /* KVM VCPU properties should be prefixed with "kvm-". */
 213 void kvm_arm_add_vcpu_properties(Object *obj)
 214 {
 215     ARMCPU *cpu = ARM_CPU(obj);
 216     CPUARMState *env = &cpu->env;
 217
 218     if (arm_feature(env, ARM_FEATURE_GENERIC_TIMER)) {
 219         cpu->kvm_adjvtime = true;
 220         object_property_add_bool(obj, "kvm-no-adjvtime", kvm_no_adjvtime_get,
 221                                  kvm_no_adjvtime_set);
 222         object_property_set_description(obj, "kvm-no-adjvtime",
 223                                         "Set on to disable the adjustment of "
 224                                         "the virtual counter. VM stopped time "
 225                                         "will be counted.");
 226     }
 227
 228     cpu->kvm_steal_time = ON_OFF_AUTO_AUTO;
 229     object_property_add_bool(obj, "kvm-steal-time", kvm_steal_time_get,
 230                              kvm_steal_time_set);
 231     object_property_set_description(obj, "kvm-steal-time",
 232                                     "Set off to disable KVM steal time.");
 233 }
 234
 235 bool kvm_arm_pmu_supported(void)
 236 {
 237     return kvm_check_extension(kvm_state, KVM_CAP_ARM_PMU_V3);
 238 }
 239
 240 int kvm_arm_get_max_vm_ipa_size(MachineState *ms, bool *fixed_ipa)
 241 {
 242     KVMState *s = KVM_STATE(ms->accelerator);
 243     int ret;
 244
 245     ret = kvm_check_extension(s, KVM_CAP_ARM_VM_IPA_SIZE);
 246     *fixed_ipa = ret <= 0;
 247
 248     return ret > 0 ? ret : 40;
 249 }
 250
 251 int kvm_arch_get_default_type(MachineState *ms)
 252 {
 253     bool fixed_ipa;
 254     int size = kvm_arm_get_max_vm_ipa_size(ms, &fixed_ipa);
 255     return fixed_ipa ? 0 : size;
 256 }
 257
 258 int kvm_arch_init(MachineState *ms, KVMState *s)
 259 {
 260     int ret = 0;
 261     /* For ARM interrupt delivery is always asynchronous,
 262      * whether we are using an in-kernel VGIC or not.
 263      */
 264     kvm_async_interrupts_allowed = true;
 265
 266     /*
 267      * PSCI wakes up secondary cores, so we always need to
 268      * have vCPUs waiting in kernel space
 269      */
 270     kvm_halt_in_kernel_allowed = true;
 271
 272     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 273
 274     if (ms->smp.cpus > 256 &&
 275         !kvm_check_extension(s, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2)) {
 276         error_report("Using more than 256 vcpus requires a host kernel "
 277                      "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
 278         ret = -EINVAL;
 279     }
 280
 281     if (kvm_check_extension(s, KVM_CAP_ARM_NISV_TO_USER)) {
 282         if (kvm_vm_enable_cap(s, KVM_CAP_ARM_NISV_TO_USER, 0)) {
 283             error_report("Failed to enable KVM_CAP_ARM_NISV_TO_USER cap");
 284         } else {
 285             /* Set status for supporting the external dabt injection */
 286             cap_has_inject_ext_dabt = kvm_check_extension(s,
 287                                     KVM_CAP_ARM_INJECT_EXT_DABT);
 288         }
 289     }
 290
 291     if (s->kvm_eager_split_size) {
 292         uint32_t sizes;
 293
 294         sizes = kvm_vm_check_extension(s, KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES);
 295         if (!sizes) {
 296             s->kvm_eager_split_size = 0;
 297             warn_report("Eager Page Split support not available");
 298         } else if (!(s->kvm_eager_split_size & sizes)) {
 299             error_report("Eager Page Split requested chunk size not valid");
 300             ret = -EINVAL;
 301         } else {
 302             ret = kvm_vm_enable_cap(s, KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE, 0,
 303                                     s->kvm_eager_split_size);
 304             if (ret < 0) {
 305                 error_report("Enabling of Eager Page Split failed: %s",
 306                              strerror(-ret));
 307             }
 308         }
 309     }
 310
 311     kvm_arm_init_debug(s);
 312
 313     return ret;
 314 }
 315
 316 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 317 {
 318     return cpu->cpu_index;
 319 }
 320
 321 /* We track all the KVM devices which need their memory addresses
 322  * passing to the kernel in a list of these structures.
 323  * When board init is complete we run through the list and
 324  * tell the kernel the base addresses of the memory regions.
 325  * We use a MemoryListener to track mapping and unmapping of
 326  * the regions during board creation, so the board models don't
 327  * need to do anything special for the KVM case.
 328  *
 329  * Sometimes the address must be OR'ed with some other fields
 330  * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
 331  * @kda_addr_ormask aims at storing the value of those fields.
 332  */
 333 typedef struct KVMDevice {
 334     struct kvm_arm_device_addr kda;
 335     struct kvm_device_attr kdattr;
 336     uint64_t kda_addr_ormask;
 337     MemoryRegion *mr;
 338     QSLIST_ENTRY(KVMDevice) entries;
 339     int dev_fd;
 340 } KVMDevice;
 341
 342 static QSLIST_HEAD(, KVMDevice) kvm_devices_head;
 343
 344 static void kvm_arm_devlistener_add(MemoryListener *listener,
 345                                     MemoryRegionSection *section)
 346 {
 347     KVMDevice *kd;
 348
 349     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 350         if (section->mr == kd->mr) {
 351             kd->kda.addr = section->offset_within_address_space;
 352         }
 353     }
 354 }
 355
 356 static void kvm_arm_devlistener_del(MemoryListener *listener,
 357                                     MemoryRegionSection *section)
 358 {
 359     KVMDevice *kd;
 360
 361     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 362         if (section->mr == kd->mr) {
 363             kd->kda.addr = -1;
 364         }
 365     }
 366 }
 367
 368 static MemoryListener devlistener = {
 369     .name = "kvm-arm",
 370     .region_add = kvm_arm_devlistener_add,
 371     .region_del = kvm_arm_devlistener_del,
 372     .priority = MEMORY_LISTENER_PRIORITY_MIN,
 373 };
 374
 375 static void kvm_arm_set_device_addr(KVMDevice *kd)
 376 {
 377     struct kvm_device_attr *attr = &kd->kdattr;
 378     int ret;
 379
 380     /* If the device control API is available and we have a device fd on the
 381      * KVMDevice struct, let's use the newer API
 382      */
 383     if (kd->dev_fd >= 0) {
 384         uint64_t addr = kd->kda.addr;
 385
 386         addr |= kd->kda_addr_ormask;
 387         attr->addr = (uintptr_t)&addr;
 388         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 389     } else {
 390         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 391     }
 392
 393     if (ret < 0) {
 394         fprintf(stderr, "Failed to set device address: %s\n",
 395                 strerror(-ret));
 396         abort();
 397     }
 398 }
 399
 400 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 401 {
 402     KVMDevice *kd, *tkd;
 403
 404     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 405         if (kd->kda.addr != -1) {
 406             kvm_arm_set_device_addr(kd);
 407         }
 408         memory_region_unref(kd->mr);
 409         QSLIST_REMOVE_HEAD(&kvm_devices_head, entries);
 410         g_free(kd);
 411     }
 412     memory_listener_unregister(&devlistener);
 413 }
 414
 415 static Notifier notify = {
 416     .notify = kvm_arm_machine_init_done,
 417 };
 418
 419 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 420                              uint64_t attr, int dev_fd, uint64_t addr_ormask)
 421 {
 422     KVMDevice *kd;
 423
 424     if (!kvm_irqchip_in_kernel()) {
 425         return;
 426     }
 427
 428     if (QSLIST_EMPTY(&kvm_devices_head)) {
 429         memory_listener_register(&devlistener, &address_space_memory);
 430         qemu_add_machine_init_done_notifier(&notify);
 431     }
 432     kd = g_new0(KVMDevice, 1);
 433     kd->mr = mr;
 434     kd->kda.id = devid;
 435     kd->kda.addr = -1;
 436     kd->kdattr.flags = 0;
 437     kd->kdattr.group = group;
 438     kd->kdattr.attr = attr;
 439     kd->dev_fd = dev_fd;
 440     kd->kda_addr_ormask = addr_ormask;
 441     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 442     memory_region_ref(kd->mr);
 443 }
 444
 445 static int compare_u64(const void *a, const void *b)
 446 {
 447     if (*(uint64_t *)a > *(uint64_t *)b) {
 448         return 1;
 449     }
 450     if (*(uint64_t *)a < *(uint64_t *)b) {
 451         return -1;
 452     }
 453     return 0;
 454 }
 455
 456 /*
 457  * cpreg_values are sorted in ascending order by KVM register ID
 458  * (see kvm_arm_init_cpreg_list). This allows us to cheaply find
 459  * the storage for a KVM register by ID with a binary search.
 460  */
 461 static uint64_t *kvm_arm_get_cpreg_ptr(ARMCPU *cpu, uint64_t regidx)
 462 {
 463     uint64_t *res;
 464
 465     res = bsearch(&regidx, cpu->cpreg_indexes, cpu->cpreg_array_len,
 466                   sizeof(uint64_t), compare_u64);
 467     assert(res);
 468
 469     return &cpu->cpreg_values[res - cpu->cpreg_indexes];
 470 }
 471
 472 /* Initialize the ARMCPU cpreg list according to the kernel's
 473  * definition of what CPU registers it knows about (and throw away
 474  * the previous TCG-created cpreg list).
 475  */
 476 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 477 {
 478     struct kvm_reg_list rl;
 479     struct kvm_reg_list *rlp;
 480     int i, ret, arraylen;
 481     CPUState *cs = CPU(cpu);
 482
 483     rl.n = 0;
 484     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 485     if (ret != -E2BIG) {
 486         return ret;
 487     }
 488     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 489     rlp->n = rl.n;
 490     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 491     if (ret) {
 492         goto out;
 493     }
 494     /* Sort the list we get back from the kernel, since cpreg_tuples
 495      * must be in strictly ascending order.
 496      */
 497     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 498
 499     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 500         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 501             continue;
 502         }
 503         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 504         case KVM_REG_SIZE_U32:
 505         case KVM_REG_SIZE_U64:
 506             break;
 507         default:
 508             fprintf(stderr, "Can't handle size of register in kernel list\n");
 509             ret = -EINVAL;
 510             goto out;
 511         }
 512
 513         arraylen++;
 514     }
 515
 516     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 517     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 518     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 519                                          arraylen);
 520     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 521                                         arraylen);
 522     cpu->cpreg_array_len = arraylen;
 523     cpu->cpreg_vmstate_array_len = arraylen;
 524
 525     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 526         uint64_t regidx = rlp->reg[i];
 527         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 528             continue;
 529         }
 530         cpu->cpreg_indexes[arraylen] = regidx;
 531         arraylen++;
 532     }
 533     assert(cpu->cpreg_array_len == arraylen);
 534
 535     if (!write_kvmstate_to_list(cpu)) {
 536         /* Shouldn't happen unless kernel is inconsistent about
 537          * what registers exist.
 538          */
 539         fprintf(stderr, "Initial read of kernel register state failed\n");
 540         ret = -EINVAL;
 541         goto out;
 542     }
 543
 544 out:
 545     g_free(rlp);
 546     return ret;
 547 }
 548
 549 bool write_kvmstate_to_list(ARMCPU *cpu)
 550 {
 551     CPUState *cs = CPU(cpu);
 552     int i;
 553     bool ok = true;
 554
 555     for (i = 0; i < cpu->cpreg_array_len; i++) {
 556         uint64_t regidx = cpu->cpreg_indexes[i];
 557         uint32_t v32;
 558         int ret;
 559
 560         switch (regidx & KVM_REG_SIZE_MASK) {
 561         case KVM_REG_SIZE_U32:
 562             ret = kvm_get_one_reg(cs, regidx, &v32);
 563             if (!ret) {
 564                 cpu->cpreg_values[i] = v32;
 565             }
 566             break;
 567         case KVM_REG_SIZE_U64:
 568             ret = kvm_get_one_reg(cs, regidx, cpu->cpreg_values + i);
 569             break;
 570         default:
 571             g_assert_not_reached();
 572         }
 573         if (ret) {
 574             ok = false;
 575         }
 576     }
 577     return ok;
 578 }
 579
 580 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 581 {
 582     CPUState *cs = CPU(cpu);
 583     int i;
 584     bool ok = true;
 585
 586     for (i = 0; i < cpu->cpreg_array_len; i++) {
 587         uint64_t regidx = cpu->cpreg_indexes[i];
 588         uint32_t v32;
 589         int ret;
 590
 591         if (kvm_arm_cpreg_level(regidx) > level) {
 592             continue;
 593         }
 594
 595         switch (regidx & KVM_REG_SIZE_MASK) {
 596         case KVM_REG_SIZE_U32:
 597             v32 = cpu->cpreg_values[i];
 598             ret = kvm_set_one_reg(cs, regidx, &v32);
 599             break;
 600         case KVM_REG_SIZE_U64:
 601             ret = kvm_set_one_reg(cs, regidx, cpu->cpreg_values + i);
 602             break;
 603         default:
 604             g_assert_not_reached();
 605         }
 606         if (ret) {
 607             /* We might fail for "unknown register" and also for
 608              * "you tried to set a register which is constant with
 609              * a different value from what it actually contains".
 610              */
 611             ok = false;
 612         }
 613     }
 614     return ok;
 615 }
 616
 617 void kvm_arm_cpu_pre_save(ARMCPU *cpu)
 618 {
 619     /* KVM virtual time adjustment */
 620     if (cpu->kvm_vtime_dirty) {
 621         *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT) = cpu->kvm_vtime;
 622     }
 623 }
 624
 625 void kvm_arm_cpu_post_load(ARMCPU *cpu)
 626 {
 627     /* KVM virtual time adjustment */
 628     if (cpu->kvm_adjvtime) {
 629         cpu->kvm_vtime = *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT);
 630         cpu->kvm_vtime_dirty = true;
 631     }
 632 }
 633
 634 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 635 {
 636     int ret;
 637
 638     /* Re-init VCPU so that all registers are set to
 639      * their respective reset values.
 640      */
 641     ret = kvm_arm_vcpu_init(CPU(cpu));
 642     if (ret < 0) {
 643         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 644         abort();
 645     }
 646     if (!write_kvmstate_to_list(cpu)) {
 647         fprintf(stderr, "write_kvmstate_to_list failed\n");
 648         abort();
 649     }
 650     /*
 651      * Sync the reset values also into the CPUState. This is necessary
 652      * because the next thing we do will be a kvm_arch_put_registers()
 653      * which will update the list values from the CPUState before copying
 654      * the list values back to KVM. It's OK to ignore failure returns here
 655      * for the same reason we do so in kvm_arch_get_registers().
 656      */
 657     write_list_to_cpustate(cpu);
 658 }
 659
 660 /*
 661  * Update KVM's MP_STATE based on what QEMU thinks it is
 662  */
 663 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 664 {
 665     if (cap_has_mp_state) {
 666         struct kvm_mp_state mp_state = {
 667             .mp_state = (cpu->power_state == PSCI_OFF) ?
 668             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 669         };
 670         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 671         if (ret) {
 672             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 673                     __func__, ret, strerror(-ret));
 674             return -1;
 675         }
 676     }
 677
 678     return 0;
 679 }
 680
 681 /*
 682  * Sync the KVM MP_STATE into QEMU
 683  */
 684 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 685 {
 686     if (cap_has_mp_state) {
 687         struct kvm_mp_state mp_state;
 688         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 689         if (ret) {
 690             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 691                     __func__, ret, strerror(-ret));
 692             abort();
 693         }
 694         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
 695             PSCI_OFF : PSCI_ON;
 696     }
 697
 698     return 0;
 699 }
 700
 701 void kvm_arm_get_virtual_time(CPUState *cs)
 702 {
 703     ARMCPU *cpu = ARM_CPU(cs);
 704     int ret;
 705
 706     if (cpu->kvm_vtime_dirty) {
 707         return;
 708     }
 709
 710     ret = kvm_get_one_reg(cs, KVM_REG_ARM_TIMER_CNT, &cpu->kvm_vtime);
 711     if (ret) {
 712         error_report("Failed to get KVM_REG_ARM_TIMER_CNT");
 713         abort();
 714     }
 715
 716     cpu->kvm_vtime_dirty = true;
 717 }
 718
 719 void kvm_arm_put_virtual_time(CPUState *cs)
 720 {
 721     ARMCPU *cpu = ARM_CPU(cs);
 722     int ret;
 723
 724     if (!cpu->kvm_vtime_dirty) {
 725         return;
 726     }
 727
 728     ret = kvm_set_one_reg(cs, KVM_REG_ARM_TIMER_CNT, &cpu->kvm_vtime);
 729     if (ret) {
 730         error_report("Failed to set KVM_REG_ARM_TIMER_CNT");
 731         abort();
 732     }
 733
 734     cpu->kvm_vtime_dirty = false;
 735 }
 736
 737 int kvm_put_vcpu_events(ARMCPU *cpu)
 738 {
 739     CPUARMState *env = &cpu->env;
 740     struct kvm_vcpu_events events;
 741     int ret;
 742
 743     if (!kvm_has_vcpu_events()) {
 744         return 0;
 745     }
 746
 747     memset(&events, 0, sizeof(events));
 748     events.exception.serror_pending = env->serror.pending;
 749
 750     /* Inject SError to guest with specified syndrome if host kernel
 751      * supports it, otherwise inject SError without syndrome.
 752      */
 753     if (cap_has_inject_serror_esr) {
 754         events.exception.serror_has_esr = env->serror.has_esr;
 755         events.exception.serror_esr = env->serror.esr;
 756     }
 757
 758     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_VCPU_EVENTS, &events);
 759     if (ret) {
 760         error_report("failed to put vcpu events");
 761     }
 762
 763     return ret;
 764 }
 765
 766 int kvm_get_vcpu_events(ARMCPU *cpu)
 767 {
 768     CPUARMState *env = &cpu->env;
 769     struct kvm_vcpu_events events;
 770     int ret;
 771
 772     if (!kvm_has_vcpu_events()) {
 773         return 0;
 774     }
 775
 776     memset(&events, 0, sizeof(events));
 777     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_VCPU_EVENTS, &events);
 778     if (ret) {
 779         error_report("failed to get vcpu events");
 780         return ret;
 781     }
 782
 783     env->serror.pending = events.exception.serror_pending;
 784     env->serror.has_esr = events.exception.serror_has_esr;
 785     env->serror.esr = events.exception.serror_esr;
 786
 787     return 0;
 788 }
 789
 790 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 791 {
 792     ARMCPU *cpu = ARM_CPU(cs);
 793     CPUARMState *env = &cpu->env;
 794
 795     if (unlikely(env->ext_dabt_raised)) {
 796         /*
 797          * Verifying that the ext DABT has been properly injected,
 798          * otherwise risking indefinitely re-running the faulting instruction
 799          * Covering a very narrow case for kernels 5.5..5.5.4
 800          * when injected abort was misconfigured to be
 801          * an IMPLEMENTATION DEFINED exception (for 32-bit EL1)
 802          */
 803         if (!arm_feature(env, ARM_FEATURE_AARCH64) &&
 804             unlikely(!kvm_arm_verify_ext_dabt_pending(cs))) {
 805
 806             error_report("Data abort exception with no valid ISS generated by "
 807                    "guest memory access. KVM unable to emulate faulting "
 808                    "instruction. Failed to inject an external data abort "
 809                    "into the guest.");
 810             abort();
 811        }
 812        /* Clear the status */
 813        env->ext_dabt_raised = 0;
 814     }
 815 }
 816
 817 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 818 {
 819     ARMCPU *cpu;
 820     uint32_t switched_level;
 821
 822     if (kvm_irqchip_in_kernel()) {
 823         /*
 824          * We only need to sync timer states with user-space interrupt
 825          * controllers, so return early and save cycles if we don't.
 826          */
 827         return MEMTXATTRS_UNSPECIFIED;
 828     }
 829
 830     cpu = ARM_CPU(cs);
 831
 832     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
 833     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
 834         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
 835
 836         qemu_mutex_lock_iothread();
 837
 838         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
 839             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
 840                          !!(run->s.regs.device_irq_level &
 841                             KVM_ARM_DEV_EL1_VTIMER));
 842             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
 843         }
 844
 845         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
 846             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
 847                          !!(run->s.regs.device_irq_level &
 848                             KVM_ARM_DEV_EL1_PTIMER));
 849             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
 850         }
 851
 852         if (switched_level & KVM_ARM_DEV_PMU) {
 853             qemu_set_irq(cpu->pmu_interrupt,
 854                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
 855             switched_level &= ~KVM_ARM_DEV_PMU;
 856         }
 857
 858         if (switched_level) {
 859             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
 860                           __func__, switched_level);
 861         }
 862
 863         /* We also mark unknown levels as processed to not waste cycles */
 864         cpu->device_irq_level = run->s.regs.device_irq_level;
 865         qemu_mutex_unlock_iothread();
 866     }
 867
 868     return MEMTXATTRS_UNSPECIFIED;
 869 }
 870
 871 void kvm_arm_vm_state_change(void *opaque, bool running, RunState state)
 872 {
 873     CPUState *cs = opaque;
 874     ARMCPU *cpu = ARM_CPU(cs);
 875
 876     if (running) {
 877         if (cpu->kvm_adjvtime) {
 878             kvm_arm_put_virtual_time(cs);
 879         }
 880     } else {
 881         if (cpu->kvm_adjvtime) {
 882             kvm_arm_get_virtual_time(cs);
 883         }
 884     }
 885 }
 886
 887 /**
 888  * kvm_arm_handle_dabt_nisv:
 889  * @cs: CPUState
 890  * @esr_iss: ISS encoding (limited) for the exception from Data Abort
 891  *           ISV bit set to '0b0' -> no valid instruction syndrome
 892  * @fault_ipa: faulting address for the synchronous data abort
 893  *
 894  * Returns: 0 if the exception has been handled, < 0 otherwise
 895  */
 896 static int kvm_arm_handle_dabt_nisv(CPUState *cs, uint64_t esr_iss,
 897                                     uint64_t fault_ipa)
 898 {
 899     ARMCPU *cpu = ARM_CPU(cs);
 900     CPUARMState *env = &cpu->env;
 901     /*
 902      * Request KVM to inject the external data abort into the guest
 903      */
 904     if (cap_has_inject_ext_dabt) {
 905         struct kvm_vcpu_events events = { };
 906         /*
 907          * The external data abort event will be handled immediately by KVM
 908          * using the address fault that triggered the exit on given VCPU.
 909          * Requesting injection of the external data abort does not rely
 910          * on any other VCPU state. Therefore, in this particular case, the VCPU
 911          * synchronization can be exceptionally skipped.
 912          */
 913         events.exception.ext_dabt_pending = 1;
 914         /* KVM_CAP_ARM_INJECT_EXT_DABT implies KVM_CAP_VCPU_EVENTS */
 915         if (!kvm_vcpu_ioctl(cs, KVM_SET_VCPU_EVENTS, &events)) {
 916             env->ext_dabt_raised = 1;
 917             return 0;
 918         }
 919     } else {
 920         error_report("Data abort exception triggered by guest memory access "
 921                      "at physical address: 0x"  TARGET_FMT_lx,
 922                      (target_ulong)fault_ipa);
 923         error_printf("KVM unable to emulate faulting instruction.\n");
 924     }
 925     return -1;
 926 }
 927
 928 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 929 {
 930     int ret = 0;
 931
 932     switch (run->exit_reason) {
 933     case KVM_EXIT_DEBUG:
 934         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 935             ret = EXCP_DEBUG;
 936         } /* otherwise return to guest */
 937         break;
 938     case KVM_EXIT_ARM_NISV:
 939         /* External DABT with no valid iss to decode */
 940         ret = kvm_arm_handle_dabt_nisv(cs, run->arm_nisv.esr_iss,
 941                                        run->arm_nisv.fault_ipa);
 942         break;
 943     default:
 944         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 945                       __func__, run->exit_reason);
 946         break;
 947     }
 948     return ret;
 949 }
 950
 951 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 952 {
 953     return true;
 954 }
 955
 956 int kvm_arch_process_async_events(CPUState *cs)
 957 {
 958     return 0;
 959 }
 960
 961 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 962 {
 963     if (kvm_sw_breakpoints_active(cs)) {
 964         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 965     }
 966     if (kvm_arm_hw_debug_active(cs)) {
 967         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 968         kvm_arm_copy_hw_debug_data(&dbg->arch);
 969     }
 970 }
 971
 972 void kvm_arch_init_irq_routing(KVMState *s)
 973 {
 974 }
 975
 976 int kvm_arch_irqchip_create(KVMState *s)
 977 {
 978     if (kvm_kernel_irqchip_split()) {
 979         error_report("-machine kernel_irqchip=split is not supported on ARM.");
 980         exit(1);
 981     }
 982
 983     /* If we can create the VGIC using the newer device control API, we
 984      * let the device do this when it initializes itself, otherwise we
 985      * fall back to the old API */
 986     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 987 }
 988
 989 int kvm_arm_vgic_probe(void)
 990 {
 991     int val = 0;
 992
 993     if (kvm_create_device(kvm_state,
 994                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 995         val |= KVM_ARM_VGIC_V3;
 996     }
 997     if (kvm_create_device(kvm_state,
 998                           KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 999         val |= KVM_ARM_VGIC_V2;
1000     }
1001     return val;
1002 }
1003
1004 int kvm_arm_set_irq(int cpu, int irqtype, int irq, int level)
1005 {
1006     int kvm_irq = (irqtype << KVM_ARM_IRQ_TYPE_SHIFT) | irq;
1007     int cpu_idx1 = cpu % 256;
1008     int cpu_idx2 = cpu / 256;
1009
1010     kvm_irq |= (cpu_idx1 << KVM_ARM_IRQ_VCPU_SHIFT) |
1011                (cpu_idx2 << KVM_ARM_IRQ_VCPU2_SHIFT);
1012
1013     return kvm_set_irq(kvm_state, kvm_irq, !!level);
1014 }
1015
1016 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
1017                              uint64_t address, uint32_t data, PCIDevice *dev)
1018 {
1019     AddressSpace *as = pci_device_iommu_address_space(dev);
1020     hwaddr xlat, len, doorbell_gpa;
1021     MemoryRegionSection mrs;
1022     MemoryRegion *mr;
1023
1024     if (as == &address_space_memory) {
1025         return 0;
1026     }
1027
1028     /* MSI doorbell address is translated by an IOMMU */
1029
1030     RCU_READ_LOCK_GUARD();
1031
1032     mr = address_space_translate(as, address, &xlat, &len, true,
1033                                  MEMTXATTRS_UNSPECIFIED);
1034
1035     if (!mr) {
1036         return 1;
1037     }
1038
1039     mrs = memory_region_find(mr, xlat, 1);
1040
1041     if (!mrs.mr) {
1042         return 1;
1043     }
1044
1045     doorbell_gpa = mrs.offset_within_address_space;
1046     memory_region_unref(mrs.mr);
1047
1048     route->u.msi.address_lo = doorbell_gpa;
1049     route->u.msi.address_hi = doorbell_gpa >> 32;
1050
1051     trace_kvm_arm_fixup_msi_route(address, doorbell_gpa);
1052
1053     return 0;
1054 }
1055
1056 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
1057                                 int vector, PCIDevice *dev)
1058 {
1059     return 0;
1060 }
1061
1062 int kvm_arch_release_virq_post(int virq)
1063 {
1064     return 0;
1065 }
1066
1067 int kvm_arch_msi_data_to_gsi(uint32_t data)
1068 {
1069     return (data - 32) & 0xffff;
1070 }
1071
1072 bool kvm_arch_cpu_check_are_resettable(void)
1073 {
1074     return true;
1075 }
1076
1077 static void kvm_arch_get_eager_split_size(Object *obj, Visitor *v,
1078                                           const char *name, void *opaque,
1079                                           Error **errp)
1080 {
1081     KVMState *s = KVM_STATE(obj);
1082     uint64_t value = s->kvm_eager_split_size;
1083
1084     visit_type_size(v, name, &value, errp);
1085 }
1086
1087 static void kvm_arch_set_eager_split_size(Object *obj, Visitor *v,
1088                                           const char *name, void *opaque,
1089                                           Error **errp)
1090 {
1091     KVMState *s = KVM_STATE(obj);
1092     uint64_t value;
1093
1094     if (s->fd != -1) {
1095         error_setg(errp, "Unable to set early-split-size after KVM has been initialized");
1096         return;
1097     }
1098
1099     if (!visit_type_size(v, name, &value, errp)) {
1100         return;
1101     }
1102
1103     if (value && !is_power_of_2(value)) {
1104         error_setg(errp, "early-split-size must be a power of two");
1105         return;
1106     }
1107
1108     s->kvm_eager_split_size = value;
1109 }
1110
1111 void kvm_arch_accel_class_init(ObjectClass *oc)
1112 {
1113     object_class_property_add(oc, "eager-split-size", "size",
1114                               kvm_arch_get_eager_split_size,
1115                               kvm_arch_set_eager_split_size, NULL, NULL);
1116
1117     object_class_property_set_description(oc, "eager-split-size",
1118         "Eager Page Split chunk size for hugepages. (default: 0, disabled)");
1119 }