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-common.h"
  17 #include "qemu/timer.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/main-loop.h"
  20 #include "qom/object.h"
  21 #include "qapi/error.h"
  22 #include "sysemu/sysemu.h"
  23 #include "sysemu/kvm.h"
  24 #include "sysemu/kvm_int.h"
  25 #include "kvm_arm.h"
  26 #include "cpu.h"
  27 #include "trace.h"
  28 #include "internals.h"
  29 #include "hw/pci/pci.h"
  30 #include "exec/memattrs.h"
  31 #include "exec/address-spaces.h"
  32 #include "hw/boards.h"
  33 #include "hw/irq.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
  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
  73     kvmfd = qemu_open("/dev/kvm", O_RDWR);
  74     if (kvmfd < 0) {
  75         goto err;
  76     }
  77     vmfd = ioctl(kvmfd, KVM_CREATE_VM, 0);
  78     if (vmfd < 0) {
  79         goto err;
  80     }
  81     cpufd = ioctl(vmfd, KVM_CREATE_VCPU, 0);
  82     if (cpufd < 0) {
  83         goto err;
  84     }
  85
  86     if (!init) {
  87         /* Caller doesn't want the VCPU to be initialized, so skip it */
  88         goto finish;
  89     }
  90
  91     if (init->target == -1) {
  92         struct kvm_vcpu_init preferred;
  93
  94         ret = ioctl(vmfd, KVM_ARM_PREFERRED_TARGET, &preferred);
  95         if (!ret) {
  96             init->target = preferred.target;
  97         }
  98     }
  99     if (ret >= 0) {
 100         ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, init);
 101         if (ret < 0) {
 102             goto err;
 103         }
 104     } else if (cpus_to_try) {
 105         /* Old kernel which doesn't know about the
 106          * PREFERRED_TARGET ioctl: we know it will only support
 107          * creating one kind of guest CPU which is its preferred
 108          * CPU type.
 109          */
 110         struct kvm_vcpu_init try;
 111
 112         while (*cpus_to_try != QEMU_KVM_ARM_TARGET_NONE) {
 113             try.target = *cpus_to_try++;
 114             memcpy(try.features, init->features, sizeof(init->features));
 115             ret = ioctl(cpufd, KVM_ARM_VCPU_INIT, &try);
 116             if (ret >= 0) {
 117                 break;
 118             }
 119         }
 120         if (ret < 0) {
 121             goto err;
 122         }
 123         init->target = try.target;
 124     } else {
 125         /* Treat a NULL cpus_to_try argument the same as an empty
 126          * list, which means we will fail the call since this must
 127          * be an old kernel which doesn't support PREFERRED_TARGET.
 128          */
 129         goto err;
 130     }
 131
 132 finish:
 133     fdarray[0] = kvmfd;
 134     fdarray[1] = vmfd;
 135     fdarray[2] = cpufd;
 136
 137     return true;
 138
 139 err:
 140     if (cpufd >= 0) {
 141         close(cpufd);
 142     }
 143     if (vmfd >= 0) {
 144         close(vmfd);
 145     }
 146     if (kvmfd >= 0) {
 147         close(kvmfd);
 148     }
 149
 150     return false;
 151 }
 152
 153 void kvm_arm_destroy_scratch_host_vcpu(int *fdarray)
 154 {
 155     int i;
 156
 157     for (i = 2; i >= 0; i--) {
 158         close(fdarray[i]);
 159     }
 160 }
 161
 162 void kvm_arm_set_cpu_features_from_host(ARMCPU *cpu)
 163 {
 164     CPUARMState *env = &cpu->env;
 165
 166     if (!arm_host_cpu_features.dtb_compatible) {
 167         if (!kvm_enabled() ||
 168             !kvm_arm_get_host_cpu_features(&arm_host_cpu_features)) {
 169             /* We can't report this error yet, so flag that we need to
 170              * in arm_cpu_realizefn().
 171              */
 172             cpu->kvm_target = QEMU_KVM_ARM_TARGET_NONE;
 173             cpu->host_cpu_probe_failed = true;
 174             return;
 175         }
 176     }
 177
 178     cpu->kvm_target = arm_host_cpu_features.target;
 179     cpu->dtb_compatible = arm_host_cpu_features.dtb_compatible;
 180     cpu->isar = arm_host_cpu_features.isar;
 181     env->features = arm_host_cpu_features.features;
 182 }
 183
 184 static bool kvm_no_adjvtime_get(Object *obj, Error **errp)
 185 {
 186     return !ARM_CPU(obj)->kvm_adjvtime;
 187 }
 188
 189 static void kvm_no_adjvtime_set(Object *obj, bool value, Error **errp)
 190 {
 191     ARM_CPU(obj)->kvm_adjvtime = !value;
 192 }
 193
 194 /* KVM VCPU properties should be prefixed with "kvm-". */
 195 void kvm_arm_add_vcpu_properties(Object *obj)
 196 {
 197     if (!kvm_enabled()) {
 198         return;
 199     }
 200
 201     ARM_CPU(obj)->kvm_adjvtime = true;
 202     object_property_add_bool(obj, "kvm-no-adjvtime", kvm_no_adjvtime_get,
 203                              kvm_no_adjvtime_set, &error_abort);
 204     object_property_set_description(obj, "kvm-no-adjvtime",
 205                                     "Set on to disable the adjustment of "
 206                                     "the virtual counter. VM stopped time "
 207                                     "will be counted.", &error_abort);
 208 }
 209
 210 bool kvm_arm_pmu_supported(CPUState *cpu)
 211 {
 212     return kvm_check_extension(cpu->kvm_state, KVM_CAP_ARM_PMU_V3);
 213 }
 214
 215 int kvm_arm_get_max_vm_ipa_size(MachineState *ms)
 216 {
 217     KVMState *s = KVM_STATE(ms->accelerator);
 218     int ret;
 219
 220     ret = kvm_check_extension(s, KVM_CAP_ARM_VM_IPA_SIZE);
 221     return ret > 0 ? ret : 40;
 222 }
 223
 224 int kvm_arch_init(MachineState *ms, KVMState *s)
 225 {
 226     int ret = 0;
 227     /* For ARM interrupt delivery is always asynchronous,
 228      * whether we are using an in-kernel VGIC or not.
 229      */
 230     kvm_async_interrupts_allowed = true;
 231
 232     /*
 233      * PSCI wakes up secondary cores, so we always need to
 234      * have vCPUs waiting in kernel space
 235      */
 236     kvm_halt_in_kernel_allowed = true;
 237
 238     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 239
 240     if (ms->smp.cpus > 256 &&
 241         !kvm_check_extension(s, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2)) {
 242         error_report("Using more than 256 vcpus requires a host kernel "
 243                      "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
 244         ret = -EINVAL;
 245     }
 246
 247     return ret;
 248 }
 249
 250 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 251 {
 252     return cpu->cpu_index;
 253 }
 254
 255 /* We track all the KVM devices which need their memory addresses
 256  * passing to the kernel in a list of these structures.
 257  * When board init is complete we run through the list and
 258  * tell the kernel the base addresses of the memory regions.
 259  * We use a MemoryListener to track mapping and unmapping of
 260  * the regions during board creation, so the board models don't
 261  * need to do anything special for the KVM case.
 262  *
 263  * Sometimes the address must be OR'ed with some other fields
 264  * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
 265  * @kda_addr_ormask aims at storing the value of those fields.
 266  */
 267 typedef struct KVMDevice {
 268     struct kvm_arm_device_addr kda;
 269     struct kvm_device_attr kdattr;
 270     uint64_t kda_addr_ormask;
 271     MemoryRegion *mr;
 272     QSLIST_ENTRY(KVMDevice) entries;
 273     int dev_fd;
 274 } KVMDevice;
 275
 276 static QSLIST_HEAD(, KVMDevice) kvm_devices_head;
 277
 278 static void kvm_arm_devlistener_add(MemoryListener *listener,
 279                                     MemoryRegionSection *section)
 280 {
 281     KVMDevice *kd;
 282
 283     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 284         if (section->mr == kd->mr) {
 285             kd->kda.addr = section->offset_within_address_space;
 286         }
 287     }
 288 }
 289
 290 static void kvm_arm_devlistener_del(MemoryListener *listener,
 291                                     MemoryRegionSection *section)
 292 {
 293     KVMDevice *kd;
 294
 295     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 296         if (section->mr == kd->mr) {
 297             kd->kda.addr = -1;
 298         }
 299     }
 300 }
 301
 302 static MemoryListener devlistener = {
 303     .region_add = kvm_arm_devlistener_add,
 304     .region_del = kvm_arm_devlistener_del,
 305 };
 306
 307 static void kvm_arm_set_device_addr(KVMDevice *kd)
 308 {
 309     struct kvm_device_attr *attr = &kd->kdattr;
 310     int ret;
 311
 312     /* If the device control API is available and we have a device fd on the
 313      * KVMDevice struct, let's use the newer API
 314      */
 315     if (kd->dev_fd >= 0) {
 316         uint64_t addr = kd->kda.addr;
 317
 318         addr |= kd->kda_addr_ormask;
 319         attr->addr = (uintptr_t)&addr;
 320         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 321     } else {
 322         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 323     }
 324
 325     if (ret < 0) {
 326         fprintf(stderr, "Failed to set device address: %s\n",
 327                 strerror(-ret));
 328         abort();
 329     }
 330 }
 331
 332 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 333 {
 334     KVMDevice *kd, *tkd;
 335
 336     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 337         if (kd->kda.addr != -1) {
 338             kvm_arm_set_device_addr(kd);
 339         }
 340         memory_region_unref(kd->mr);
 341         QSLIST_REMOVE_HEAD(&kvm_devices_head, entries);
 342         g_free(kd);
 343     }
 344     memory_listener_unregister(&devlistener);
 345 }
 346
 347 static Notifier notify = {
 348     .notify = kvm_arm_machine_init_done,
 349 };
 350
 351 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 352                              uint64_t attr, int dev_fd, uint64_t addr_ormask)
 353 {
 354     KVMDevice *kd;
 355
 356     if (!kvm_irqchip_in_kernel()) {
 357         return;
 358     }
 359
 360     if (QSLIST_EMPTY(&kvm_devices_head)) {
 361         memory_listener_register(&devlistener, &address_space_memory);
 362         qemu_add_machine_init_done_notifier(&notify);
 363     }
 364     kd = g_new0(KVMDevice, 1);
 365     kd->mr = mr;
 366     kd->kda.id = devid;
 367     kd->kda.addr = -1;
 368     kd->kdattr.flags = 0;
 369     kd->kdattr.group = group;
 370     kd->kdattr.attr = attr;
 371     kd->dev_fd = dev_fd;
 372     kd->kda_addr_ormask = addr_ormask;
 373     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 374     memory_region_ref(kd->mr);
 375 }
 376
 377 static int compare_u64(const void *a, const void *b)
 378 {
 379     if (*(uint64_t *)a > *(uint64_t *)b) {
 380         return 1;
 381     }
 382     if (*(uint64_t *)a < *(uint64_t *)b) {
 383         return -1;
 384     }
 385     return 0;
 386 }
 387
 388 /*
 389  * cpreg_values are sorted in ascending order by KVM register ID
 390  * (see kvm_arm_init_cpreg_list). This allows us to cheaply find
 391  * the storage for a KVM register by ID with a binary search.
 392  */
 393 static uint64_t *kvm_arm_get_cpreg_ptr(ARMCPU *cpu, uint64_t regidx)
 394 {
 395     uint64_t *res;
 396
 397     res = bsearch(&regidx, cpu->cpreg_indexes, cpu->cpreg_array_len,
 398                   sizeof(uint64_t), compare_u64);
 399     assert(res);
 400
 401     return &cpu->cpreg_values[res - cpu->cpreg_indexes];
 402 }
 403
 404 /* Initialize the ARMCPU cpreg list according to the kernel's
 405  * definition of what CPU registers it knows about (and throw away
 406  * the previous TCG-created cpreg list).
 407  */
 408 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 409 {
 410     struct kvm_reg_list rl;
 411     struct kvm_reg_list *rlp;
 412     int i, ret, arraylen;
 413     CPUState *cs = CPU(cpu);
 414
 415     rl.n = 0;
 416     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 417     if (ret != -E2BIG) {
 418         return ret;
 419     }
 420     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 421     rlp->n = rl.n;
 422     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 423     if (ret) {
 424         goto out;
 425     }
 426     /* Sort the list we get back from the kernel, since cpreg_tuples
 427      * must be in strictly ascending order.
 428      */
 429     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 430
 431     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 432         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 433             continue;
 434         }
 435         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 436         case KVM_REG_SIZE_U32:
 437         case KVM_REG_SIZE_U64:
 438             break;
 439         default:
 440             fprintf(stderr, "Can't handle size of register in kernel list\n");
 441             ret = -EINVAL;
 442             goto out;
 443         }
 444
 445         arraylen++;
 446     }
 447
 448     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 449     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 450     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 451                                          arraylen);
 452     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 453                                         arraylen);
 454     cpu->cpreg_array_len = arraylen;
 455     cpu->cpreg_vmstate_array_len = arraylen;
 456
 457     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 458         uint64_t regidx = rlp->reg[i];
 459         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 460             continue;
 461         }
 462         cpu->cpreg_indexes[arraylen] = regidx;
 463         arraylen++;
 464     }
 465     assert(cpu->cpreg_array_len == arraylen);
 466
 467     if (!write_kvmstate_to_list(cpu)) {
 468         /* Shouldn't happen unless kernel is inconsistent about
 469          * what registers exist.
 470          */
 471         fprintf(stderr, "Initial read of kernel register state failed\n");
 472         ret = -EINVAL;
 473         goto out;
 474     }
 475
 476 out:
 477     g_free(rlp);
 478     return ret;
 479 }
 480
 481 bool write_kvmstate_to_list(ARMCPU *cpu)
 482 {
 483     CPUState *cs = CPU(cpu);
 484     int i;
 485     bool ok = true;
 486
 487     for (i = 0; i < cpu->cpreg_array_len; i++) {
 488         struct kvm_one_reg r;
 489         uint64_t regidx = cpu->cpreg_indexes[i];
 490         uint32_t v32;
 491         int ret;
 492
 493         r.id = regidx;
 494
 495         switch (regidx & KVM_REG_SIZE_MASK) {
 496         case KVM_REG_SIZE_U32:
 497             r.addr = (uintptr_t)&v32;
 498             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 499             if (!ret) {
 500                 cpu->cpreg_values[i] = v32;
 501             }
 502             break;
 503         case KVM_REG_SIZE_U64:
 504             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 505             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 506             break;
 507         default:
 508             abort();
 509         }
 510         if (ret) {
 511             ok = false;
 512         }
 513     }
 514     return ok;
 515 }
 516
 517 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 518 {
 519     CPUState *cs = CPU(cpu);
 520     int i;
 521     bool ok = true;
 522
 523     for (i = 0; i < cpu->cpreg_array_len; i++) {
 524         struct kvm_one_reg r;
 525         uint64_t regidx = cpu->cpreg_indexes[i];
 526         uint32_t v32;
 527         int ret;
 528
 529         if (kvm_arm_cpreg_level(regidx) > level) {
 530             continue;
 531         }
 532
 533         r.id = regidx;
 534         switch (regidx & KVM_REG_SIZE_MASK) {
 535         case KVM_REG_SIZE_U32:
 536             v32 = cpu->cpreg_values[i];
 537             r.addr = (uintptr_t)&v32;
 538             break;
 539         case KVM_REG_SIZE_U64:
 540             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 541             break;
 542         default:
 543             abort();
 544         }
 545         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 546         if (ret) {
 547             /* We might fail for "unknown register" and also for
 548              * "you tried to set a register which is constant with
 549              * a different value from what it actually contains".
 550              */
 551             ok = false;
 552         }
 553     }
 554     return ok;
 555 }
 556
 557 void kvm_arm_cpu_pre_save(ARMCPU *cpu)
 558 {
 559     /* KVM virtual time adjustment */
 560     if (cpu->kvm_vtime_dirty) {
 561         *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT) = cpu->kvm_vtime;
 562     }
 563 }
 564
 565 void kvm_arm_cpu_post_load(ARMCPU *cpu)
 566 {
 567     /* KVM virtual time adjustment */
 568     if (cpu->kvm_adjvtime) {
 569         cpu->kvm_vtime = *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT);
 570         cpu->kvm_vtime_dirty = true;
 571     }
 572 }
 573
 574 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 575 {
 576     int ret;
 577
 578     /* Re-init VCPU so that all registers are set to
 579      * their respective reset values.
 580      */
 581     ret = kvm_arm_vcpu_init(CPU(cpu));
 582     if (ret < 0) {
 583         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 584         abort();
 585     }
 586     if (!write_kvmstate_to_list(cpu)) {
 587         fprintf(stderr, "write_kvmstate_to_list failed\n");
 588         abort();
 589     }
 590     /*
 591      * Sync the reset values also into the CPUState. This is necessary
 592      * because the next thing we do will be a kvm_arch_put_registers()
 593      * which will update the list values from the CPUState before copying
 594      * the list values back to KVM. It's OK to ignore failure returns here
 595      * for the same reason we do so in kvm_arch_get_registers().
 596      */
 597     write_list_to_cpustate(cpu);
 598 }
 599
 600 /*
 601  * Update KVM's MP_STATE based on what QEMU thinks it is
 602  */
 603 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 604 {
 605     if (cap_has_mp_state) {
 606         struct kvm_mp_state mp_state = {
 607             .mp_state = (cpu->power_state == PSCI_OFF) ?
 608             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 609         };
 610         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 611         if (ret) {
 612             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 613                     __func__, ret, strerror(-ret));
 614             return -1;
 615         }
 616     }
 617
 618     return 0;
 619 }
 620
 621 /*
 622  * Sync the KVM MP_STATE into QEMU
 623  */
 624 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 625 {
 626     if (cap_has_mp_state) {
 627         struct kvm_mp_state mp_state;
 628         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 629         if (ret) {
 630             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 631                     __func__, ret, strerror(-ret));
 632             abort();
 633         }
 634         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
 635             PSCI_OFF : PSCI_ON;
 636     }
 637
 638     return 0;
 639 }
 640
 641 void kvm_arm_get_virtual_time(CPUState *cs)
 642 {
 643     ARMCPU *cpu = ARM_CPU(cs);
 644     struct kvm_one_reg reg = {
 645         .id = KVM_REG_ARM_TIMER_CNT,
 646         .addr = (uintptr_t)&cpu->kvm_vtime,
 647     };
 648     int ret;
 649
 650     if (cpu->kvm_vtime_dirty) {
 651         return;
 652     }
 653
 654     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
 655     if (ret) {
 656         error_report("Failed to get KVM_REG_ARM_TIMER_CNT");
 657         abort();
 658     }
 659
 660     cpu->kvm_vtime_dirty = true;
 661 }
 662
 663 void kvm_arm_put_virtual_time(CPUState *cs)
 664 {
 665     ARMCPU *cpu = ARM_CPU(cs);
 666     struct kvm_one_reg reg = {
 667         .id = KVM_REG_ARM_TIMER_CNT,
 668         .addr = (uintptr_t)&cpu->kvm_vtime,
 669     };
 670     int ret;
 671
 672     if (!cpu->kvm_vtime_dirty) {
 673         return;
 674     }
 675
 676     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
 677     if (ret) {
 678         error_report("Failed to set KVM_REG_ARM_TIMER_CNT");
 679         abort();
 680     }
 681
 682     cpu->kvm_vtime_dirty = false;
 683 }
 684
 685 int kvm_put_vcpu_events(ARMCPU *cpu)
 686 {
 687     CPUARMState *env = &cpu->env;
 688     struct kvm_vcpu_events events;
 689     int ret;
 690
 691     if (!kvm_has_vcpu_events()) {
 692         return 0;
 693     }
 694
 695     memset(&events, 0, sizeof(events));
 696     events.exception.serror_pending = env->serror.pending;
 697
 698     /* Inject SError to guest with specified syndrome if host kernel
 699      * supports it, otherwise inject SError without syndrome.
 700      */
 701     if (cap_has_inject_serror_esr) {
 702         events.exception.serror_has_esr = env->serror.has_esr;
 703         events.exception.serror_esr = env->serror.esr;
 704     }
 705
 706     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_VCPU_EVENTS, &events);
 707     if (ret) {
 708         error_report("failed to put vcpu events");
 709     }
 710
 711     return ret;
 712 }
 713
 714 int kvm_get_vcpu_events(ARMCPU *cpu)
 715 {
 716     CPUARMState *env = &cpu->env;
 717     struct kvm_vcpu_events events;
 718     int ret;
 719
 720     if (!kvm_has_vcpu_events()) {
 721         return 0;
 722     }
 723
 724     memset(&events, 0, sizeof(events));
 725     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_VCPU_EVENTS, &events);
 726     if (ret) {
 727         error_report("failed to get vcpu events");
 728         return ret;
 729     }
 730
 731     env->serror.pending = events.exception.serror_pending;
 732     env->serror.has_esr = events.exception.serror_has_esr;
 733     env->serror.esr = events.exception.serror_esr;
 734
 735     return 0;
 736 }
 737
 738 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 739 {
 740 }
 741
 742 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 743 {
 744     ARMCPU *cpu;
 745     uint32_t switched_level;
 746
 747     if (kvm_irqchip_in_kernel()) {
 748         /*
 749          * We only need to sync timer states with user-space interrupt
 750          * controllers, so return early and save cycles if we don't.
 751          */
 752         return MEMTXATTRS_UNSPECIFIED;
 753     }
 754
 755     cpu = ARM_CPU(cs);
 756
 757     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
 758     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
 759         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
 760
 761         qemu_mutex_lock_iothread();
 762
 763         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
 764             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
 765                          !!(run->s.regs.device_irq_level &
 766                             KVM_ARM_DEV_EL1_VTIMER));
 767             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
 768         }
 769
 770         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
 771             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
 772                          !!(run->s.regs.device_irq_level &
 773                             KVM_ARM_DEV_EL1_PTIMER));
 774             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
 775         }
 776
 777         if (switched_level & KVM_ARM_DEV_PMU) {
 778             qemu_set_irq(cpu->pmu_interrupt,
 779                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
 780             switched_level &= ~KVM_ARM_DEV_PMU;
 781         }
 782
 783         if (switched_level) {
 784             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
 785                           __func__, switched_level);
 786         }
 787
 788         /* We also mark unknown levels as processed to not waste cycles */
 789         cpu->device_irq_level = run->s.regs.device_irq_level;
 790         qemu_mutex_unlock_iothread();
 791     }
 792
 793     return MEMTXATTRS_UNSPECIFIED;
 794 }
 795
 796 void kvm_arm_vm_state_change(void *opaque, int running, RunState state)
 797 {
 798     CPUState *cs = opaque;
 799     ARMCPU *cpu = ARM_CPU(cs);
 800
 801     if (running) {
 802         if (cpu->kvm_adjvtime) {
 803             kvm_arm_put_virtual_time(cs);
 804         }
 805     } else {
 806         if (cpu->kvm_adjvtime) {
 807             kvm_arm_get_virtual_time(cs);
 808         }
 809     }
 810 }
 811
 812 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 813 {
 814     int ret = 0;
 815
 816     switch (run->exit_reason) {
 817     case KVM_EXIT_DEBUG:
 818         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 819             ret = EXCP_DEBUG;
 820         } /* otherwise return to guest */
 821         break;
 822     default:
 823         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 824                       __func__, run->exit_reason);
 825         break;
 826     }
 827     return ret;
 828 }
 829
 830 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 831 {
 832     return true;
 833 }
 834
 835 int kvm_arch_process_async_events(CPUState *cs)
 836 {
 837     return 0;
 838 }
 839
 840 /* The #ifdef protections are until 32bit headers are imported and can
 841  * be removed once both 32 and 64 bit reach feature parity.
 842  */
 843 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 844 {
 845 #ifdef KVM_GUESTDBG_USE_SW_BP
 846     if (kvm_sw_breakpoints_active(cs)) {
 847         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 848     }
 849 #endif
 850 #ifdef KVM_GUESTDBG_USE_HW
 851     if (kvm_arm_hw_debug_active(cs)) {
 852         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 853         kvm_arm_copy_hw_debug_data(&dbg->arch);
 854     }
 855 #endif
 856 }
 857
 858 void kvm_arch_init_irq_routing(KVMState *s)
 859 {
 860 }
 861
 862 int kvm_arch_irqchip_create(KVMState *s)
 863 {
 864     if (kvm_kernel_irqchip_split()) {
 865         perror("-machine kernel_irqchip=split is not supported on ARM.");
 866         exit(1);
 867     }
 868
 869     /* If we can create the VGIC using the newer device control API, we
 870      * let the device do this when it initializes itself, otherwise we
 871      * fall back to the old API */
 872     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 873 }
 874
 875 int kvm_arm_vgic_probe(void)
 876 {
 877     if (kvm_create_device(kvm_state,
 878                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 879         return 3;
 880     } else if (kvm_create_device(kvm_state,
 881                                  KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 882         return 2;
 883     } else {
 884         return 0;
 885     }
 886 }
 887
 888 int kvm_arm_set_irq(int cpu, int irqtype, int irq, int level)
 889 {
 890     int kvm_irq = (irqtype << KVM_ARM_IRQ_TYPE_SHIFT) | irq;
 891     int cpu_idx1 = cpu % 256;
 892     int cpu_idx2 = cpu / 256;
 893
 894     kvm_irq |= (cpu_idx1 << KVM_ARM_IRQ_VCPU_SHIFT) |
 895                (cpu_idx2 << KVM_ARM_IRQ_VCPU2_SHIFT);
 896
 897     return kvm_set_irq(kvm_state, kvm_irq, !!level);
 898 }
 899
 900 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
 901                              uint64_t address, uint32_t data, PCIDevice *dev)
 902 {
 903     AddressSpace *as = pci_device_iommu_address_space(dev);
 904     hwaddr xlat, len, doorbell_gpa;
 905     MemoryRegionSection mrs;
 906     MemoryRegion *mr;
 907     int ret = 1;
 908
 909     if (as == &address_space_memory) {
 910         return 0;
 911     }
 912
 913     /* MSI doorbell address is translated by an IOMMU */
 914
 915     rcu_read_lock();
 916     mr = address_space_translate(as, address, &xlat, &len, true,
 917                                  MEMTXATTRS_UNSPECIFIED);
 918     if (!mr) {
 919         goto unlock;
 920     }
 921     mrs = memory_region_find(mr, xlat, 1);
 922     if (!mrs.mr) {
 923         goto unlock;
 924     }
 925
 926     doorbell_gpa = mrs.offset_within_address_space;
 927     memory_region_unref(mrs.mr);
 928
 929     route->u.msi.address_lo = doorbell_gpa;
 930     route->u.msi.address_hi = doorbell_gpa >> 32;
 931
 932     trace_kvm_arm_fixup_msi_route(address, doorbell_gpa);
 933
 934     ret = 0;
 935
 936 unlock:
 937     rcu_read_unlock();
 938     return ret;
 939 }
 940
 941 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
 942                                 int vector, PCIDevice *dev)
 943 {
 944     return 0;
 945 }
 946
 947 int kvm_arch_release_virq_post(int virq)
 948 {
 949     return 0;
 950 }
 951
 952 int kvm_arch_msi_data_to_gsi(uint32_t data)
 953 {
 954     return (data - 32) & 0xffff;
 955 }