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