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     ARMCPU *cpu = ARM_CPU(obj);
 198     CPUARMState *env = &cpu->env;
 199
 200     if (arm_feature(env, ARM_FEATURE_GENERIC_TIMER)) {
 201         cpu->kvm_adjvtime = true;
 202         object_property_add_bool(obj, "kvm-no-adjvtime", kvm_no_adjvtime_get,
 203                                  kvm_no_adjvtime_set);
 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.");
 208     }
 209 }
 210
 211 bool kvm_arm_pmu_supported(void)
 212 {
 213     return kvm_check_extension(kvm_state, KVM_CAP_ARM_PMU_V3);
 214 }
 215
 216 int kvm_arm_get_max_vm_ipa_size(MachineState *ms)
 217 {
 218     KVMState *s = KVM_STATE(ms->accelerator);
 219     int ret;
 220
 221     ret = kvm_check_extension(s, KVM_CAP_ARM_VM_IPA_SIZE);
 222     return ret > 0 ? ret : 40;
 223 }
 224
 225 int kvm_arch_init(MachineState *ms, KVMState *s)
 226 {
 227     int ret = 0;
 228     /* For ARM interrupt delivery is always asynchronous,
 229      * whether we are using an in-kernel VGIC or not.
 230      */
 231     kvm_async_interrupts_allowed = true;
 232
 233     /*
 234      * PSCI wakes up secondary cores, so we always need to
 235      * have vCPUs waiting in kernel space
 236      */
 237     kvm_halt_in_kernel_allowed = true;
 238
 239     cap_has_mp_state = kvm_check_extension(s, KVM_CAP_MP_STATE);
 240
 241     if (ms->smp.cpus > 256 &&
 242         !kvm_check_extension(s, KVM_CAP_ARM_IRQ_LINE_LAYOUT_2)) {
 243         error_report("Using more than 256 vcpus requires a host kernel "
 244                      "with KVM_CAP_ARM_IRQ_LINE_LAYOUT_2");
 245         ret = -EINVAL;
 246     }
 247
 248     return ret;
 249 }
 250
 251 unsigned long kvm_arch_vcpu_id(CPUState *cpu)
 252 {
 253     return cpu->cpu_index;
 254 }
 255
 256 /* We track all the KVM devices which need their memory addresses
 257  * passing to the kernel in a list of these structures.
 258  * When board init is complete we run through the list and
 259  * tell the kernel the base addresses of the memory regions.
 260  * We use a MemoryListener to track mapping and unmapping of
 261  * the regions during board creation, so the board models don't
 262  * need to do anything special for the KVM case.
 263  *
 264  * Sometimes the address must be OR'ed with some other fields
 265  * (for example for KVM_VGIC_V3_ADDR_TYPE_REDIST_REGION).
 266  * @kda_addr_ormask aims at storing the value of those fields.
 267  */
 268 typedef struct KVMDevice {
 269     struct kvm_arm_device_addr kda;
 270     struct kvm_device_attr kdattr;
 271     uint64_t kda_addr_ormask;
 272     MemoryRegion *mr;
 273     QSLIST_ENTRY(KVMDevice) entries;
 274     int dev_fd;
 275 } KVMDevice;
 276
 277 static QSLIST_HEAD(, KVMDevice) kvm_devices_head;
 278
 279 static void kvm_arm_devlistener_add(MemoryListener *listener,
 280                                     MemoryRegionSection *section)
 281 {
 282     KVMDevice *kd;
 283
 284     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 285         if (section->mr == kd->mr) {
 286             kd->kda.addr = section->offset_within_address_space;
 287         }
 288     }
 289 }
 290
 291 static void kvm_arm_devlistener_del(MemoryListener *listener,
 292                                     MemoryRegionSection *section)
 293 {
 294     KVMDevice *kd;
 295
 296     QSLIST_FOREACH(kd, &kvm_devices_head, entries) {
 297         if (section->mr == kd->mr) {
 298             kd->kda.addr = -1;
 299         }
 300     }
 301 }
 302
 303 static MemoryListener devlistener = {
 304     .region_add = kvm_arm_devlistener_add,
 305     .region_del = kvm_arm_devlistener_del,
 306 };
 307
 308 static void kvm_arm_set_device_addr(KVMDevice *kd)
 309 {
 310     struct kvm_device_attr *attr = &kd->kdattr;
 311     int ret;
 312
 313     /* If the device control API is available and we have a device fd on the
 314      * KVMDevice struct, let's use the newer API
 315      */
 316     if (kd->dev_fd >= 0) {
 317         uint64_t addr = kd->kda.addr;
 318
 319         addr |= kd->kda_addr_ormask;
 320         attr->addr = (uintptr_t)&addr;
 321         ret = kvm_device_ioctl(kd->dev_fd, KVM_SET_DEVICE_ATTR, attr);
 322     } else {
 323         ret = kvm_vm_ioctl(kvm_state, KVM_ARM_SET_DEVICE_ADDR, &kd->kda);
 324     }
 325
 326     if (ret < 0) {
 327         fprintf(stderr, "Failed to set device address: %s\n",
 328                 strerror(-ret));
 329         abort();
 330     }
 331 }
 332
 333 static void kvm_arm_machine_init_done(Notifier *notifier, void *data)
 334 {
 335     KVMDevice *kd, *tkd;
 336
 337     QSLIST_FOREACH_SAFE(kd, &kvm_devices_head, entries, tkd) {
 338         if (kd->kda.addr != -1) {
 339             kvm_arm_set_device_addr(kd);
 340         }
 341         memory_region_unref(kd->mr);
 342         QSLIST_REMOVE_HEAD(&kvm_devices_head, entries);
 343         g_free(kd);
 344     }
 345     memory_listener_unregister(&devlistener);
 346 }
 347
 348 static Notifier notify = {
 349     .notify = kvm_arm_machine_init_done,
 350 };
 351
 352 void kvm_arm_register_device(MemoryRegion *mr, uint64_t devid, uint64_t group,
 353                              uint64_t attr, int dev_fd, uint64_t addr_ormask)
 354 {
 355     KVMDevice *kd;
 356
 357     if (!kvm_irqchip_in_kernel()) {
 358         return;
 359     }
 360
 361     if (QSLIST_EMPTY(&kvm_devices_head)) {
 362         memory_listener_register(&devlistener, &address_space_memory);
 363         qemu_add_machine_init_done_notifier(&notify);
 364     }
 365     kd = g_new0(KVMDevice, 1);
 366     kd->mr = mr;
 367     kd->kda.id = devid;
 368     kd->kda.addr = -1;
 369     kd->kdattr.flags = 0;
 370     kd->kdattr.group = group;
 371     kd->kdattr.attr = attr;
 372     kd->dev_fd = dev_fd;
 373     kd->kda_addr_ormask = addr_ormask;
 374     QSLIST_INSERT_HEAD(&kvm_devices_head, kd, entries);
 375     memory_region_ref(kd->mr);
 376 }
 377
 378 static int compare_u64(const void *a, const void *b)
 379 {
 380     if (*(uint64_t *)a > *(uint64_t *)b) {
 381         return 1;
 382     }
 383     if (*(uint64_t *)a < *(uint64_t *)b) {
 384         return -1;
 385     }
 386     return 0;
 387 }
 388
 389 /*
 390  * cpreg_values are sorted in ascending order by KVM register ID
 391  * (see kvm_arm_init_cpreg_list). This allows us to cheaply find
 392  * the storage for a KVM register by ID with a binary search.
 393  */
 394 static uint64_t *kvm_arm_get_cpreg_ptr(ARMCPU *cpu, uint64_t regidx)
 395 {
 396     uint64_t *res;
 397
 398     res = bsearch(&regidx, cpu->cpreg_indexes, cpu->cpreg_array_len,
 399                   sizeof(uint64_t), compare_u64);
 400     assert(res);
 401
 402     return &cpu->cpreg_values[res - cpu->cpreg_indexes];
 403 }
 404
 405 /* Initialize the ARMCPU cpreg list according to the kernel's
 406  * definition of what CPU registers it knows about (and throw away
 407  * the previous TCG-created cpreg list).
 408  */
 409 int kvm_arm_init_cpreg_list(ARMCPU *cpu)
 410 {
 411     struct kvm_reg_list rl;
 412     struct kvm_reg_list *rlp;
 413     int i, ret, arraylen;
 414     CPUState *cs = CPU(cpu);
 415
 416     rl.n = 0;
 417     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, &rl);
 418     if (ret != -E2BIG) {
 419         return ret;
 420     }
 421     rlp = g_malloc(sizeof(struct kvm_reg_list) + rl.n * sizeof(uint64_t));
 422     rlp->n = rl.n;
 423     ret = kvm_vcpu_ioctl(cs, KVM_GET_REG_LIST, rlp);
 424     if (ret) {
 425         goto out;
 426     }
 427     /* Sort the list we get back from the kernel, since cpreg_tuples
 428      * must be in strictly ascending order.
 429      */
 430     qsort(&rlp->reg, rlp->n, sizeof(rlp->reg[0]), compare_u64);
 431
 432     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 433         if (!kvm_arm_reg_syncs_via_cpreg_list(rlp->reg[i])) {
 434             continue;
 435         }
 436         switch (rlp->reg[i] & KVM_REG_SIZE_MASK) {
 437         case KVM_REG_SIZE_U32:
 438         case KVM_REG_SIZE_U64:
 439             break;
 440         default:
 441             fprintf(stderr, "Can't handle size of register in kernel list\n");
 442             ret = -EINVAL;
 443             goto out;
 444         }
 445
 446         arraylen++;
 447     }
 448
 449     cpu->cpreg_indexes = g_renew(uint64_t, cpu->cpreg_indexes, arraylen);
 450     cpu->cpreg_values = g_renew(uint64_t, cpu->cpreg_values, arraylen);
 451     cpu->cpreg_vmstate_indexes = g_renew(uint64_t, cpu->cpreg_vmstate_indexes,
 452                                          arraylen);
 453     cpu->cpreg_vmstate_values = g_renew(uint64_t, cpu->cpreg_vmstate_values,
 454                                         arraylen);
 455     cpu->cpreg_array_len = arraylen;
 456     cpu->cpreg_vmstate_array_len = arraylen;
 457
 458     for (i = 0, arraylen = 0; i < rlp->n; i++) {
 459         uint64_t regidx = rlp->reg[i];
 460         if (!kvm_arm_reg_syncs_via_cpreg_list(regidx)) {
 461             continue;
 462         }
 463         cpu->cpreg_indexes[arraylen] = regidx;
 464         arraylen++;
 465     }
 466     assert(cpu->cpreg_array_len == arraylen);
 467
 468     if (!write_kvmstate_to_list(cpu)) {
 469         /* Shouldn't happen unless kernel is inconsistent about
 470          * what registers exist.
 471          */
 472         fprintf(stderr, "Initial read of kernel register state failed\n");
 473         ret = -EINVAL;
 474         goto out;
 475     }
 476
 477 out:
 478     g_free(rlp);
 479     return ret;
 480 }
 481
 482 bool write_kvmstate_to_list(ARMCPU *cpu)
 483 {
 484     CPUState *cs = CPU(cpu);
 485     int i;
 486     bool ok = true;
 487
 488     for (i = 0; i < cpu->cpreg_array_len; i++) {
 489         struct kvm_one_reg r;
 490         uint64_t regidx = cpu->cpreg_indexes[i];
 491         uint32_t v32;
 492         int ret;
 493
 494         r.id = regidx;
 495
 496         switch (regidx & KVM_REG_SIZE_MASK) {
 497         case KVM_REG_SIZE_U32:
 498             r.addr = (uintptr_t)&v32;
 499             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 500             if (!ret) {
 501                 cpu->cpreg_values[i] = v32;
 502             }
 503             break;
 504         case KVM_REG_SIZE_U64:
 505             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 506             ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &r);
 507             break;
 508         default:
 509             abort();
 510         }
 511         if (ret) {
 512             ok = false;
 513         }
 514     }
 515     return ok;
 516 }
 517
 518 bool write_list_to_kvmstate(ARMCPU *cpu, int level)
 519 {
 520     CPUState *cs = CPU(cpu);
 521     int i;
 522     bool ok = true;
 523
 524     for (i = 0; i < cpu->cpreg_array_len; i++) {
 525         struct kvm_one_reg r;
 526         uint64_t regidx = cpu->cpreg_indexes[i];
 527         uint32_t v32;
 528         int ret;
 529
 530         if (kvm_arm_cpreg_level(regidx) > level) {
 531             continue;
 532         }
 533
 534         r.id = regidx;
 535         switch (regidx & KVM_REG_SIZE_MASK) {
 536         case KVM_REG_SIZE_U32:
 537             v32 = cpu->cpreg_values[i];
 538             r.addr = (uintptr_t)&v32;
 539             break;
 540         case KVM_REG_SIZE_U64:
 541             r.addr = (uintptr_t)(cpu->cpreg_values + i);
 542             break;
 543         default:
 544             abort();
 545         }
 546         ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &r);
 547         if (ret) {
 548             /* We might fail for "unknown register" and also for
 549              * "you tried to set a register which is constant with
 550              * a different value from what it actually contains".
 551              */
 552             ok = false;
 553         }
 554     }
 555     return ok;
 556 }
 557
 558 void kvm_arm_cpu_pre_save(ARMCPU *cpu)
 559 {
 560     /* KVM virtual time adjustment */
 561     if (cpu->kvm_vtime_dirty) {
 562         *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT) = cpu->kvm_vtime;
 563     }
 564 }
 565
 566 void kvm_arm_cpu_post_load(ARMCPU *cpu)
 567 {
 568     /* KVM virtual time adjustment */
 569     if (cpu->kvm_adjvtime) {
 570         cpu->kvm_vtime = *kvm_arm_get_cpreg_ptr(cpu, KVM_REG_ARM_TIMER_CNT);
 571         cpu->kvm_vtime_dirty = true;
 572     }
 573 }
 574
 575 void kvm_arm_reset_vcpu(ARMCPU *cpu)
 576 {
 577     int ret;
 578
 579     /* Re-init VCPU so that all registers are set to
 580      * their respective reset values.
 581      */
 582     ret = kvm_arm_vcpu_init(CPU(cpu));
 583     if (ret < 0) {
 584         fprintf(stderr, "kvm_arm_vcpu_init failed: %s\n", strerror(-ret));
 585         abort();
 586     }
 587     if (!write_kvmstate_to_list(cpu)) {
 588         fprintf(stderr, "write_kvmstate_to_list failed\n");
 589         abort();
 590     }
 591     /*
 592      * Sync the reset values also into the CPUState. This is necessary
 593      * because the next thing we do will be a kvm_arch_put_registers()
 594      * which will update the list values from the CPUState before copying
 595      * the list values back to KVM. It's OK to ignore failure returns here
 596      * for the same reason we do so in kvm_arch_get_registers().
 597      */
 598     write_list_to_cpustate(cpu);
 599 }
 600
 601 /*
 602  * Update KVM's MP_STATE based on what QEMU thinks it is
 603  */
 604 int kvm_arm_sync_mpstate_to_kvm(ARMCPU *cpu)
 605 {
 606     if (cap_has_mp_state) {
 607         struct kvm_mp_state mp_state = {
 608             .mp_state = (cpu->power_state == PSCI_OFF) ?
 609             KVM_MP_STATE_STOPPED : KVM_MP_STATE_RUNNABLE
 610         };
 611         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_MP_STATE, &mp_state);
 612         if (ret) {
 613             fprintf(stderr, "%s: failed to set MP_STATE %d/%s\n",
 614                     __func__, ret, strerror(-ret));
 615             return -1;
 616         }
 617     }
 618
 619     return 0;
 620 }
 621
 622 /*
 623  * Sync the KVM MP_STATE into QEMU
 624  */
 625 int kvm_arm_sync_mpstate_to_qemu(ARMCPU *cpu)
 626 {
 627     if (cap_has_mp_state) {
 628         struct kvm_mp_state mp_state;
 629         int ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_MP_STATE, &mp_state);
 630         if (ret) {
 631             fprintf(stderr, "%s: failed to get MP_STATE %d/%s\n",
 632                     __func__, ret, strerror(-ret));
 633             abort();
 634         }
 635         cpu->power_state = (mp_state.mp_state == KVM_MP_STATE_STOPPED) ?
 636             PSCI_OFF : PSCI_ON;
 637     }
 638
 639     return 0;
 640 }
 641
 642 void kvm_arm_get_virtual_time(CPUState *cs)
 643 {
 644     ARMCPU *cpu = ARM_CPU(cs);
 645     struct kvm_one_reg reg = {
 646         .id = KVM_REG_ARM_TIMER_CNT,
 647         .addr = (uintptr_t)&cpu->kvm_vtime,
 648     };
 649     int ret;
 650
 651     if (cpu->kvm_vtime_dirty) {
 652         return;
 653     }
 654
 655     ret = kvm_vcpu_ioctl(cs, KVM_GET_ONE_REG, &reg);
 656     if (ret) {
 657         error_report("Failed to get KVM_REG_ARM_TIMER_CNT");
 658         abort();
 659     }
 660
 661     cpu->kvm_vtime_dirty = true;
 662 }
 663
 664 void kvm_arm_put_virtual_time(CPUState *cs)
 665 {
 666     ARMCPU *cpu = ARM_CPU(cs);
 667     struct kvm_one_reg reg = {
 668         .id = KVM_REG_ARM_TIMER_CNT,
 669         .addr = (uintptr_t)&cpu->kvm_vtime,
 670     };
 671     int ret;
 672
 673     if (!cpu->kvm_vtime_dirty) {
 674         return;
 675     }
 676
 677     ret = kvm_vcpu_ioctl(cs, KVM_SET_ONE_REG, &reg);
 678     if (ret) {
 679         error_report("Failed to set KVM_REG_ARM_TIMER_CNT");
 680         abort();
 681     }
 682
 683     cpu->kvm_vtime_dirty = false;
 684 }
 685
 686 int kvm_put_vcpu_events(ARMCPU *cpu)
 687 {
 688     CPUARMState *env = &cpu->env;
 689     struct kvm_vcpu_events events;
 690     int ret;
 691
 692     if (!kvm_has_vcpu_events()) {
 693         return 0;
 694     }
 695
 696     memset(&events, 0, sizeof(events));
 697     events.exception.serror_pending = env->serror.pending;
 698
 699     /* Inject SError to guest with specified syndrome if host kernel
 700      * supports it, otherwise inject SError without syndrome.
 701      */
 702     if (cap_has_inject_serror_esr) {
 703         events.exception.serror_has_esr = env->serror.has_esr;
 704         events.exception.serror_esr = env->serror.esr;
 705     }
 706
 707     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_SET_VCPU_EVENTS, &events);
 708     if (ret) {
 709         error_report("failed to put vcpu events");
 710     }
 711
 712     return ret;
 713 }
 714
 715 int kvm_get_vcpu_events(ARMCPU *cpu)
 716 {
 717     CPUARMState *env = &cpu->env;
 718     struct kvm_vcpu_events events;
 719     int ret;
 720
 721     if (!kvm_has_vcpu_events()) {
 722         return 0;
 723     }
 724
 725     memset(&events, 0, sizeof(events));
 726     ret = kvm_vcpu_ioctl(CPU(cpu), KVM_GET_VCPU_EVENTS, &events);
 727     if (ret) {
 728         error_report("failed to get vcpu events");
 729         return ret;
 730     }
 731
 732     env->serror.pending = events.exception.serror_pending;
 733     env->serror.has_esr = events.exception.serror_has_esr;
 734     env->serror.esr = events.exception.serror_esr;
 735
 736     return 0;
 737 }
 738
 739 void kvm_arch_pre_run(CPUState *cs, struct kvm_run *run)
 740 {
 741 }
 742
 743 MemTxAttrs kvm_arch_post_run(CPUState *cs, struct kvm_run *run)
 744 {
 745     ARMCPU *cpu;
 746     uint32_t switched_level;
 747
 748     if (kvm_irqchip_in_kernel()) {
 749         /*
 750          * We only need to sync timer states with user-space interrupt
 751          * controllers, so return early and save cycles if we don't.
 752          */
 753         return MEMTXATTRS_UNSPECIFIED;
 754     }
 755
 756     cpu = ARM_CPU(cs);
 757
 758     /* Synchronize our shadowed in-kernel device irq lines with the kvm ones */
 759     if (run->s.regs.device_irq_level != cpu->device_irq_level) {
 760         switched_level = cpu->device_irq_level ^ run->s.regs.device_irq_level;
 761
 762         qemu_mutex_lock_iothread();
 763
 764         if (switched_level & KVM_ARM_DEV_EL1_VTIMER) {
 765             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_VIRT],
 766                          !!(run->s.regs.device_irq_level &
 767                             KVM_ARM_DEV_EL1_VTIMER));
 768             switched_level &= ~KVM_ARM_DEV_EL1_VTIMER;
 769         }
 770
 771         if (switched_level & KVM_ARM_DEV_EL1_PTIMER) {
 772             qemu_set_irq(cpu->gt_timer_outputs[GTIMER_PHYS],
 773                          !!(run->s.regs.device_irq_level &
 774                             KVM_ARM_DEV_EL1_PTIMER));
 775             switched_level &= ~KVM_ARM_DEV_EL1_PTIMER;
 776         }
 777
 778         if (switched_level & KVM_ARM_DEV_PMU) {
 779             qemu_set_irq(cpu->pmu_interrupt,
 780                          !!(run->s.regs.device_irq_level & KVM_ARM_DEV_PMU));
 781             switched_level &= ~KVM_ARM_DEV_PMU;
 782         }
 783
 784         if (switched_level) {
 785             qemu_log_mask(LOG_UNIMP, "%s: unhandled in-kernel device IRQ %x\n",
 786                           __func__, switched_level);
 787         }
 788
 789         /* We also mark unknown levels as processed to not waste cycles */
 790         cpu->device_irq_level = run->s.regs.device_irq_level;
 791         qemu_mutex_unlock_iothread();
 792     }
 793
 794     return MEMTXATTRS_UNSPECIFIED;
 795 }
 796
 797 void kvm_arm_vm_state_change(void *opaque, int running, RunState state)
 798 {
 799     CPUState *cs = opaque;
 800     ARMCPU *cpu = ARM_CPU(cs);
 801
 802     if (running) {
 803         if (cpu->kvm_adjvtime) {
 804             kvm_arm_put_virtual_time(cs);
 805         }
 806     } else {
 807         if (cpu->kvm_adjvtime) {
 808             kvm_arm_get_virtual_time(cs);
 809         }
 810     }
 811 }
 812
 813 int kvm_arch_handle_exit(CPUState *cs, struct kvm_run *run)
 814 {
 815     int ret = 0;
 816
 817     switch (run->exit_reason) {
 818     case KVM_EXIT_DEBUG:
 819         if (kvm_arm_handle_debug(cs, &run->debug.arch)) {
 820             ret = EXCP_DEBUG;
 821         } /* otherwise return to guest */
 822         break;
 823     default:
 824         qemu_log_mask(LOG_UNIMP, "%s: un-handled exit reason %d\n",
 825                       __func__, run->exit_reason);
 826         break;
 827     }
 828     return ret;
 829 }
 830
 831 bool kvm_arch_stop_on_emulation_error(CPUState *cs)
 832 {
 833     return true;
 834 }
 835
 836 int kvm_arch_process_async_events(CPUState *cs)
 837 {
 838     return 0;
 839 }
 840
 841 /* The #ifdef protections are until 32bit headers are imported and can
 842  * be removed once both 32 and 64 bit reach feature parity.
 843  */
 844 void kvm_arch_update_guest_debug(CPUState *cs, struct kvm_guest_debug *dbg)
 845 {
 846 #ifdef KVM_GUESTDBG_USE_SW_BP
 847     if (kvm_sw_breakpoints_active(cs)) {
 848         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_SW_BP;
 849     }
 850 #endif
 851 #ifdef KVM_GUESTDBG_USE_HW
 852     if (kvm_arm_hw_debug_active(cs)) {
 853         dbg->control |= KVM_GUESTDBG_ENABLE | KVM_GUESTDBG_USE_HW;
 854         kvm_arm_copy_hw_debug_data(&dbg->arch);
 855     }
 856 #endif
 857 }
 858
 859 void kvm_arch_init_irq_routing(KVMState *s)
 860 {
 861 }
 862
 863 int kvm_arch_irqchip_create(KVMState *s)
 864 {
 865     if (kvm_kernel_irqchip_split()) {
 866         perror("-machine kernel_irqchip=split is not supported on ARM.");
 867         exit(1);
 868     }
 869
 870     /* If we can create the VGIC using the newer device control API, we
 871      * let the device do this when it initializes itself, otherwise we
 872      * fall back to the old API */
 873     return kvm_check_extension(s, KVM_CAP_DEVICE_CTRL);
 874 }
 875
 876 int kvm_arm_vgic_probe(void)
 877 {
 878     int val = 0;
 879
 880     if (kvm_create_device(kvm_state,
 881                           KVM_DEV_TYPE_ARM_VGIC_V3, true) == 0) {
 882         val |= KVM_ARM_VGIC_V3;
 883     }
 884     if (kvm_create_device(kvm_state,
 885                           KVM_DEV_TYPE_ARM_VGIC_V2, true) == 0) {
 886         val |= KVM_ARM_VGIC_V2;
 887     }
 888     return val;
 889 }
 890
 891 int kvm_arm_set_irq(int cpu, int irqtype, int irq, int level)
 892 {
 893     int kvm_irq = (irqtype << KVM_ARM_IRQ_TYPE_SHIFT) | irq;
 894     int cpu_idx1 = cpu % 256;
 895     int cpu_idx2 = cpu / 256;
 896
 897     kvm_irq |= (cpu_idx1 << KVM_ARM_IRQ_VCPU_SHIFT) |
 898                (cpu_idx2 << KVM_ARM_IRQ_VCPU2_SHIFT);
 899
 900     return kvm_set_irq(kvm_state, kvm_irq, !!level);
 901 }
 902
 903 int kvm_arch_fixup_msi_route(struct kvm_irq_routing_entry *route,
 904                              uint64_t address, uint32_t data, PCIDevice *dev)
 905 {
 906     AddressSpace *as = pci_device_iommu_address_space(dev);
 907     hwaddr xlat, len, doorbell_gpa;
 908     MemoryRegionSection mrs;
 909     MemoryRegion *mr;
 910     int ret = 1;
 911
 912     if (as == &address_space_memory) {
 913         return 0;
 914     }
 915
 916     /* MSI doorbell address is translated by an IOMMU */
 917
 918     rcu_read_lock();
 919     mr = address_space_translate(as, address, &xlat, &len, true,
 920                                  MEMTXATTRS_UNSPECIFIED);
 921     if (!mr) {
 922         goto unlock;
 923     }
 924     mrs = memory_region_find(mr, xlat, 1);
 925     if (!mrs.mr) {
 926         goto unlock;
 927     }
 928
 929     doorbell_gpa = mrs.offset_within_address_space;
 930     memory_region_unref(mrs.mr);
 931
 932     route->u.msi.address_lo = doorbell_gpa;
 933     route->u.msi.address_hi = doorbell_gpa >> 32;
 934
 935     trace_kvm_arm_fixup_msi_route(address, doorbell_gpa);
 936
 937     ret = 0;
 938
 939 unlock:
 940     rcu_read_unlock();
 941     return ret;
 942 }
 943
 944 int kvm_arch_add_msi_route_post(struct kvm_irq_routing_entry *route,
 945                                 int vector, PCIDevice *dev)
 946 {
 947     return 0;
 948 }
 949
 950 int kvm_arch_release_virq_post(int virq)
 951 {
 952     return 0;
 953 }
 954
 955 int kvm_arch_msi_data_to_gsi(uint32_t data)
 956 {
 957     return (data - 32) & 0xffff;
 958 }